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 buddy][ group 0 bitmap] [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:
334 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
336 #if BITS_PER_LONG == 64
337 *bit
+= ((unsigned long) addr
& 7UL) << 3;
338 addr
= (void *) ((unsigned long) addr
& ~7UL);
339 #elif BITS_PER_LONG == 32
340 *bit
+= ((unsigned long) addr
& 3UL) << 3;
341 addr
= (void *) ((unsigned long) addr
& ~3UL);
343 #error "how many bits you are?!"
348 static inline int mb_test_bit(int bit
, void *addr
)
351 * ext4_test_bit on architecture like powerpc
352 * needs unsigned long aligned address
354 addr
= mb_correct_addr_and_bit(&bit
, addr
);
355 return ext4_test_bit(bit
, addr
);
358 static inline void mb_set_bit(int bit
, void *addr
)
360 addr
= mb_correct_addr_and_bit(&bit
, addr
);
361 ext4_set_bit(bit
, addr
);
364 static inline void mb_set_bit_atomic(spinlock_t
*lock
, int bit
, void *addr
)
366 addr
= mb_correct_addr_and_bit(&bit
, addr
);
367 ext4_set_bit_atomic(lock
, bit
, addr
);
370 static inline void mb_clear_bit(int bit
, void *addr
)
372 addr
= mb_correct_addr_and_bit(&bit
, addr
);
373 ext4_clear_bit(bit
, addr
);
376 static inline void mb_clear_bit_atomic(spinlock_t
*lock
, int bit
, void *addr
)
378 addr
= mb_correct_addr_and_bit(&bit
, addr
);
379 ext4_clear_bit_atomic(lock
, bit
, addr
);
382 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
384 int fix
= 0, ret
, tmpmax
;
385 addr
= mb_correct_addr_and_bit(&fix
, addr
);
389 ret
= ext4_find_next_zero_bit(addr
, tmpmax
, start
) - fix
;
395 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
397 int fix
= 0, ret
, tmpmax
;
398 addr
= mb_correct_addr_and_bit(&fix
, addr
);
402 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
408 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
412 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
415 if (order
> e4b
->bd_blkbits
+ 1) {
420 /* at order 0 we see each particular block */
421 *max
= 1 << (e4b
->bd_blkbits
+ 3);
423 return EXT4_MB_BITMAP(e4b
);
425 bb
= EXT4_MB_BUDDY(e4b
) + EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
426 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
432 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
433 int first
, int count
)
436 struct super_block
*sb
= e4b
->bd_sb
;
438 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
440 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
441 for (i
= 0; i
< count
; i
++) {
442 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
443 ext4_fsblk_t blocknr
;
444 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
445 blocknr
+= first
+ i
;
447 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
449 ext4_error(sb
, __func__
, "double-free of inode"
450 " %lu's block %llu(bit %u in group %lu)\n",
451 inode
? inode
->i_ino
: 0, blocknr
,
452 first
+ i
, e4b
->bd_group
);
454 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
458 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
462 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
464 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
465 for (i
= 0; i
< count
; i
++) {
466 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
467 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
471 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
473 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
474 unsigned char *b1
, *b2
;
476 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
477 b2
= (unsigned char *) bitmap
;
478 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
479 if (b1
[i
] != b2
[i
]) {
480 printk(KERN_ERR
"corruption in group %lu "
481 "at byte %u(%u): %x in copy != %x "
482 "on disk/prealloc\n",
483 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
491 static inline void mb_free_blocks_double(struct inode
*inode
,
492 struct ext4_buddy
*e4b
, int first
, int count
)
496 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
497 int first
, int count
)
501 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
507 #ifdef AGGRESSIVE_CHECK
509 #define MB_CHECK_ASSERT(assert) \
513 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
514 function, file, line, # assert); \
519 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
520 const char *function
, int line
)
522 struct super_block
*sb
= e4b
->bd_sb
;
523 int order
= e4b
->bd_blkbits
+ 1;
530 struct ext4_group_info
*grp
;
533 struct list_head
*cur
;
537 if (!test_opt(sb
, MBALLOC
))
541 static int mb_check_counter
;
542 if (mb_check_counter
++ % 100 != 0)
547 buddy
= mb_find_buddy(e4b
, order
, &max
);
548 MB_CHECK_ASSERT(buddy
);
549 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
550 MB_CHECK_ASSERT(buddy2
);
551 MB_CHECK_ASSERT(buddy
!= buddy2
);
552 MB_CHECK_ASSERT(max
* 2 == max2
);
555 for (i
= 0; i
< max
; i
++) {
557 if (mb_test_bit(i
, buddy
)) {
558 /* only single bit in buddy2 may be 1 */
559 if (!mb_test_bit(i
<< 1, buddy2
)) {
561 mb_test_bit((i
<<1)+1, buddy2
));
562 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
564 mb_test_bit(i
<< 1, buddy2
));
569 /* both bits in buddy2 must be 0 */
570 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
571 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
573 for (j
= 0; j
< (1 << order
); j
++) {
574 k
= (i
* (1 << order
)) + j
;
576 !mb_test_bit(k
, EXT4_MB_BITMAP(e4b
)));
580 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
585 buddy
= mb_find_buddy(e4b
, 0, &max
);
586 for (i
= 0; i
< max
; i
++) {
587 if (!mb_test_bit(i
, buddy
)) {
588 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
596 /* check used bits only */
597 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
598 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
600 MB_CHECK_ASSERT(k
< max2
);
601 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
604 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
605 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
607 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
608 buddy
= mb_find_buddy(e4b
, 0, &max
);
609 list_for_each(cur
, &grp
->bb_prealloc_list
) {
610 ext4_group_t groupnr
;
611 struct ext4_prealloc_space
*pa
;
612 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
613 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
614 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
615 for (i
= 0; i
< pa
->pa_len
; i
++)
616 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
620 #undef MB_CHECK_ASSERT
621 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
622 __FILE__, __func__, __LINE__)
624 #define mb_check_buddy(e4b)
627 /* FIXME!! need more doc */
628 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
629 void *buddy
, unsigned first
, int len
,
630 struct ext4_group_info
*grp
)
632 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
635 unsigned short chunk
;
636 unsigned short border
;
638 BUG_ON(len
> EXT4_BLOCKS_PER_GROUP(sb
));
640 border
= 2 << sb
->s_blocksize_bits
;
643 /* find how many blocks can be covered since this position */
644 max
= ffs(first
| border
) - 1;
646 /* find how many blocks of power 2 we need to mark */
653 /* mark multiblock chunks only */
654 grp
->bb_counters
[min
]++;
656 mb_clear_bit(first
>> min
,
657 buddy
+ sbi
->s_mb_offsets
[min
]);
664 static void ext4_mb_generate_buddy(struct super_block
*sb
,
665 void *buddy
, void *bitmap
, ext4_group_t group
)
667 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
668 unsigned short max
= EXT4_BLOCKS_PER_GROUP(sb
);
669 unsigned short i
= 0;
670 unsigned short first
;
673 unsigned fragments
= 0;
674 unsigned long long period
= get_cycles();
676 /* initialize buddy from bitmap which is aggregation
677 * of on-disk bitmap and preallocations */
678 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
679 grp
->bb_first_free
= i
;
683 i
= mb_find_next_bit(bitmap
, max
, i
);
687 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
689 grp
->bb_counters
[0]++;
691 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
693 grp
->bb_fragments
= fragments
;
695 if (free
!= grp
->bb_free
) {
696 ext4_error(sb
, __func__
,
697 "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
698 group
, free
, grp
->bb_free
);
700 * If we intent to continue, we consider group descritor
701 * corrupt and update bb_free using bitmap value
706 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
708 period
= get_cycles() - period
;
709 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
710 EXT4_SB(sb
)->s_mb_buddies_generated
++;
711 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
712 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
715 /* The buddy information is attached the buddy cache inode
716 * for convenience. The information regarding each group
717 * is loaded via ext4_mb_load_buddy. The information involve
718 * block bitmap and buddy information. The information are
719 * stored in the inode as
722 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
725 * one block each for bitmap and buddy information.
726 * So for each group we take up 2 blocks. A page can
727 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
728 * So it can have information regarding groups_per_page which
729 * is blocks_per_page/2
732 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
739 ext4_group_t first_group
;
741 struct super_block
*sb
;
742 struct buffer_head
*bhs
;
743 struct buffer_head
**bh
;
748 mb_debug("init page %lu\n", page
->index
);
750 inode
= page
->mapping
->host
;
752 blocksize
= 1 << inode
->i_blkbits
;
753 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
755 groups_per_page
= blocks_per_page
>> 1;
756 if (groups_per_page
== 0)
759 /* allocate buffer_heads to read bitmaps */
760 if (groups_per_page
> 1) {
762 i
= sizeof(struct buffer_head
*) * groups_per_page
;
763 bh
= kzalloc(i
, GFP_NOFS
);
769 first_group
= page
->index
* blocks_per_page
/ 2;
771 /* read all groups the page covers into the cache */
772 for (i
= 0; i
< groups_per_page
; i
++) {
773 struct ext4_group_desc
*desc
;
775 if (first_group
+ i
>= EXT4_SB(sb
)->s_groups_count
)
779 desc
= ext4_get_group_desc(sb
, first_group
+ i
, NULL
);
784 bh
[i
] = sb_getblk(sb
, ext4_block_bitmap(sb
, desc
));
788 if (bh_uptodate_or_lock(bh
[i
]))
791 spin_lock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
792 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
793 ext4_init_block_bitmap(sb
, bh
[i
],
794 first_group
+ i
, desc
);
795 set_buffer_uptodate(bh
[i
]);
796 unlock_buffer(bh
[i
]);
797 spin_unlock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
800 spin_unlock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
802 bh
[i
]->b_end_io
= end_buffer_read_sync
;
803 submit_bh(READ
, bh
[i
]);
804 mb_debug("read bitmap for group %lu\n", first_group
+ i
);
807 /* wait for I/O completion */
808 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
809 wait_on_buffer(bh
[i
]);
812 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
813 if (!buffer_uptodate(bh
[i
]))
817 first_block
= page
->index
* blocks_per_page
;
818 for (i
= 0; i
< blocks_per_page
; i
++) {
820 struct ext4_group_info
*grinfo
;
822 group
= (first_block
+ i
) >> 1;
823 if (group
>= EXT4_SB(sb
)->s_groups_count
)
827 * data carry information regarding this
828 * particular group in the format specified
832 data
= page_address(page
) + (i
* blocksize
);
833 bitmap
= bh
[group
- first_group
]->b_data
;
836 * We place the buddy block and bitmap block
839 if ((first_block
+ i
) & 1) {
840 /* this is block of buddy */
841 BUG_ON(incore
== NULL
);
842 mb_debug("put buddy for group %u in page %lu/%x\n",
843 group
, page
->index
, i
* blocksize
);
844 memset(data
, 0xff, blocksize
);
845 grinfo
= ext4_get_group_info(sb
, group
);
846 grinfo
->bb_fragments
= 0;
847 memset(grinfo
->bb_counters
, 0,
848 sizeof(unsigned short)*(sb
->s_blocksize_bits
+2));
850 * incore got set to the group block bitmap below
852 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
855 /* this is block of bitmap */
856 BUG_ON(incore
!= NULL
);
857 mb_debug("put bitmap for group %u in page %lu/%x\n",
858 group
, page
->index
, i
* blocksize
);
860 /* see comments in ext4_mb_put_pa() */
861 ext4_lock_group(sb
, group
);
862 memcpy(data
, bitmap
, blocksize
);
864 /* mark all preallocated blks used in in-core bitmap */
865 ext4_mb_generate_from_pa(sb
, data
, group
);
866 ext4_unlock_group(sb
, group
);
868 /* set incore so that the buddy information can be
869 * generated using this
874 SetPageUptodate(page
);
878 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
886 static noinline_for_stack
int
887 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
888 struct ext4_buddy
*e4b
)
890 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
891 struct inode
*inode
= sbi
->s_buddy_cache
;
899 mb_debug("load group %lu\n", group
);
901 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
903 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
904 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
906 e4b
->bd_group
= group
;
907 e4b
->bd_buddy_page
= NULL
;
908 e4b
->bd_bitmap_page
= NULL
;
911 * the buddy cache inode stores the block bitmap
912 * and buddy information in consecutive blocks.
913 * So for each group we need two blocks.
916 pnum
= block
/ blocks_per_page
;
917 poff
= block
% blocks_per_page
;
919 /* we could use find_or_create_page(), but it locks page
920 * what we'd like to avoid in fast path ... */
921 page
= find_get_page(inode
->i_mapping
, pnum
);
922 if (page
== NULL
|| !PageUptodate(page
)) {
924 page_cache_release(page
);
925 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
927 BUG_ON(page
->mapping
!= inode
->i_mapping
);
928 if (!PageUptodate(page
)) {
929 ret
= ext4_mb_init_cache(page
, NULL
);
934 mb_cmp_bitmaps(e4b
, page_address(page
) +
935 (poff
* sb
->s_blocksize
));
940 if (page
== NULL
|| !PageUptodate(page
)) {
944 e4b
->bd_bitmap_page
= page
;
945 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
946 mark_page_accessed(page
);
949 pnum
= block
/ blocks_per_page
;
950 poff
= block
% blocks_per_page
;
952 page
= find_get_page(inode
->i_mapping
, pnum
);
953 if (page
== NULL
|| !PageUptodate(page
)) {
955 page_cache_release(page
);
956 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
958 BUG_ON(page
->mapping
!= inode
->i_mapping
);
959 if (!PageUptodate(page
)) {
960 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
969 if (page
== NULL
|| !PageUptodate(page
)) {
973 e4b
->bd_buddy_page
= page
;
974 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
975 mark_page_accessed(page
);
977 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
978 BUG_ON(e4b
->bd_buddy_page
== NULL
);
983 if (e4b
->bd_bitmap_page
)
984 page_cache_release(e4b
->bd_bitmap_page
);
985 if (e4b
->bd_buddy_page
)
986 page_cache_release(e4b
->bd_buddy_page
);
987 e4b
->bd_buddy
= NULL
;
988 e4b
->bd_bitmap
= NULL
;
992 static void ext4_mb_release_desc(struct ext4_buddy
*e4b
)
994 if (e4b
->bd_bitmap_page
)
995 page_cache_release(e4b
->bd_bitmap_page
);
996 if (e4b
->bd_buddy_page
)
997 page_cache_release(e4b
->bd_buddy_page
);
1001 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1006 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1007 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1009 bb
= EXT4_MB_BUDDY(e4b
);
1010 while (order
<= e4b
->bd_blkbits
+ 1) {
1012 if (!mb_test_bit(block
, bb
)) {
1013 /* this block is part of buddy of order 'order' */
1016 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1022 static void mb_clear_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1028 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1029 /* fast path: clear whole word at once */
1030 addr
= bm
+ (cur
>> 3);
1035 mb_clear_bit_atomic(lock
, cur
, bm
);
1040 static void mb_set_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1046 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1047 /* fast path: set whole word at once */
1048 addr
= bm
+ (cur
>> 3);
1053 mb_set_bit_atomic(lock
, cur
, bm
);
1058 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1059 int first
, int count
)
1066 struct super_block
*sb
= e4b
->bd_sb
;
1068 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1069 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
1070 mb_check_buddy(e4b
);
1071 mb_free_blocks_double(inode
, e4b
, first
, count
);
1073 e4b
->bd_info
->bb_free
+= count
;
1074 if (first
< e4b
->bd_info
->bb_first_free
)
1075 e4b
->bd_info
->bb_first_free
= first
;
1077 /* let's maintain fragments counter */
1079 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1080 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1081 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1083 e4b
->bd_info
->bb_fragments
--;
1084 else if (!block
&& !max
)
1085 e4b
->bd_info
->bb_fragments
++;
1087 /* let's maintain buddy itself */
1088 while (count
-- > 0) {
1092 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1093 ext4_fsblk_t blocknr
;
1094 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
1097 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
1098 ext4_unlock_group(sb
, e4b
->bd_group
);
1099 ext4_error(sb
, __func__
, "double-free of inode"
1100 " %lu's block %llu(bit %u in group %lu)\n",
1101 inode
? inode
->i_ino
: 0, blocknr
, block
,
1103 ext4_lock_group(sb
, e4b
->bd_group
);
1105 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1106 e4b
->bd_info
->bb_counters
[order
]++;
1108 /* start of the buddy */
1109 buddy
= mb_find_buddy(e4b
, order
, &max
);
1113 if (mb_test_bit(block
, buddy
) ||
1114 mb_test_bit(block
+ 1, buddy
))
1117 /* both the buddies are free, try to coalesce them */
1118 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1124 /* for special purposes, we don't set
1125 * free bits in bitmap */
1126 mb_set_bit(block
, buddy
);
1127 mb_set_bit(block
+ 1, buddy
);
1129 e4b
->bd_info
->bb_counters
[order
]--;
1130 e4b
->bd_info
->bb_counters
[order
]--;
1134 e4b
->bd_info
->bb_counters
[order
]++;
1136 mb_clear_bit(block
, buddy2
);
1140 mb_check_buddy(e4b
);
1143 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1144 int needed
, struct ext4_free_extent
*ex
)
1151 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1154 buddy
= mb_find_buddy(e4b
, order
, &max
);
1155 BUG_ON(buddy
== NULL
);
1156 BUG_ON(block
>= max
);
1157 if (mb_test_bit(block
, buddy
)) {
1164 /* FIXME dorp order completely ? */
1165 if (likely(order
== 0)) {
1166 /* find actual order */
1167 order
= mb_find_order_for_block(e4b
, block
);
1168 block
= block
>> order
;
1171 ex
->fe_len
= 1 << order
;
1172 ex
->fe_start
= block
<< order
;
1173 ex
->fe_group
= e4b
->bd_group
;
1175 /* calc difference from given start */
1176 next
= next
- ex
->fe_start
;
1178 ex
->fe_start
+= next
;
1180 while (needed
> ex
->fe_len
&&
1181 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1183 if (block
+ 1 >= max
)
1186 next
= (block
+ 1) * (1 << order
);
1187 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1190 ord
= mb_find_order_for_block(e4b
, next
);
1193 block
= next
>> order
;
1194 ex
->fe_len
+= 1 << order
;
1197 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1201 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1207 int start
= ex
->fe_start
;
1208 int len
= ex
->fe_len
;
1213 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1214 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1215 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1216 mb_check_buddy(e4b
);
1217 mb_mark_used_double(e4b
, start
, len
);
1219 e4b
->bd_info
->bb_free
-= len
;
1220 if (e4b
->bd_info
->bb_first_free
== start
)
1221 e4b
->bd_info
->bb_first_free
+= len
;
1223 /* let's maintain fragments counter */
1225 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1226 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1227 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1229 e4b
->bd_info
->bb_fragments
++;
1230 else if (!mlen
&& !max
)
1231 e4b
->bd_info
->bb_fragments
--;
1233 /* let's maintain buddy itself */
1235 ord
= mb_find_order_for_block(e4b
, start
);
1237 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1238 /* the whole chunk may be allocated at once! */
1240 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1241 BUG_ON((start
>> ord
) >= max
);
1242 mb_set_bit(start
>> ord
, buddy
);
1243 e4b
->bd_info
->bb_counters
[ord
]--;
1250 /* store for history */
1252 ret
= len
| (ord
<< 16);
1254 /* we have to split large buddy */
1256 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1257 mb_set_bit(start
>> ord
, buddy
);
1258 e4b
->bd_info
->bb_counters
[ord
]--;
1261 cur
= (start
>> ord
) & ~1U;
1262 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1263 mb_clear_bit(cur
, buddy
);
1264 mb_clear_bit(cur
+ 1, buddy
);
1265 e4b
->bd_info
->bb_counters
[ord
]++;
1266 e4b
->bd_info
->bb_counters
[ord
]++;
1269 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b
->bd_sb
), ex
->fe_group
),
1270 EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1271 mb_check_buddy(e4b
);
1277 * Must be called under group lock!
1279 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1280 struct ext4_buddy
*e4b
)
1282 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1285 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1286 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1288 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1289 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1290 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1292 /* preallocation can change ac_b_ex, thus we store actually
1293 * allocated blocks for history */
1294 ac
->ac_f_ex
= ac
->ac_b_ex
;
1296 ac
->ac_status
= AC_STATUS_FOUND
;
1297 ac
->ac_tail
= ret
& 0xffff;
1298 ac
->ac_buddy
= ret
>> 16;
1300 /* XXXXXXX: SUCH A HORRIBLE **CK */
1302 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1303 get_page(ac
->ac_bitmap_page
);
1304 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1305 get_page(ac
->ac_buddy_page
);
1307 /* store last allocated for subsequent stream allocation */
1308 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1309 spin_lock(&sbi
->s_md_lock
);
1310 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1311 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1312 spin_unlock(&sbi
->s_md_lock
);
1317 * regular allocator, for general purposes allocation
1320 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1321 struct ext4_buddy
*e4b
,
1324 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1325 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1326 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1327 struct ext4_free_extent ex
;
1331 * We don't want to scan for a whole year
1333 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1334 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1335 ac
->ac_status
= AC_STATUS_BREAK
;
1340 * Haven't found good chunk so far, let's continue
1342 if (bex
->fe_len
< gex
->fe_len
)
1345 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1346 && bex
->fe_group
== e4b
->bd_group
) {
1347 /* recheck chunk's availability - we don't know
1348 * when it was found (within this lock-unlock
1350 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1351 if (max
>= gex
->fe_len
) {
1352 ext4_mb_use_best_found(ac
, e4b
);
1359 * The routine checks whether found extent is good enough. If it is,
1360 * then the extent gets marked used and flag is set to the context
1361 * to stop scanning. Otherwise, the extent is compared with the
1362 * previous found extent and if new one is better, then it's stored
1363 * in the context. Later, the best found extent will be used, if
1364 * mballoc can't find good enough extent.
1366 * FIXME: real allocation policy is to be designed yet!
1368 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1369 struct ext4_free_extent
*ex
,
1370 struct ext4_buddy
*e4b
)
1372 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1373 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1375 BUG_ON(ex
->fe_len
<= 0);
1376 BUG_ON(ex
->fe_len
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1377 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1378 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1383 * The special case - take what you catch first
1385 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1387 ext4_mb_use_best_found(ac
, e4b
);
1392 * Let's check whether the chuck is good enough
1394 if (ex
->fe_len
== gex
->fe_len
) {
1396 ext4_mb_use_best_found(ac
, e4b
);
1401 * If this is first found extent, just store it in the context
1403 if (bex
->fe_len
== 0) {
1409 * If new found extent is better, store it in the context
1411 if (bex
->fe_len
< gex
->fe_len
) {
1412 /* if the request isn't satisfied, any found extent
1413 * larger than previous best one is better */
1414 if (ex
->fe_len
> bex
->fe_len
)
1416 } else if (ex
->fe_len
> gex
->fe_len
) {
1417 /* if the request is satisfied, then we try to find
1418 * an extent that still satisfy the request, but is
1419 * smaller than previous one */
1420 if (ex
->fe_len
< bex
->fe_len
)
1424 ext4_mb_check_limits(ac
, e4b
, 0);
1427 static int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1428 struct ext4_buddy
*e4b
)
1430 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1431 ext4_group_t group
= ex
.fe_group
;
1435 BUG_ON(ex
.fe_len
<= 0);
1436 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1440 ext4_lock_group(ac
->ac_sb
, group
);
1441 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1445 ext4_mb_use_best_found(ac
, e4b
);
1448 ext4_unlock_group(ac
->ac_sb
, group
);
1449 ext4_mb_release_desc(e4b
);
1454 static int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1455 struct ext4_buddy
*e4b
)
1457 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1460 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1461 struct ext4_super_block
*es
= sbi
->s_es
;
1462 struct ext4_free_extent ex
;
1464 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1467 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1471 ext4_lock_group(ac
->ac_sb
, group
);
1472 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1473 ac
->ac_g_ex
.fe_len
, &ex
);
1475 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1478 start
= (e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
)) +
1479 ex
.fe_start
+ le32_to_cpu(es
->s_first_data_block
);
1480 /* use do_div to get remainder (would be 64-bit modulo) */
1481 if (do_div(start
, sbi
->s_stripe
) == 0) {
1484 ext4_mb_use_best_found(ac
, e4b
);
1486 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1487 BUG_ON(ex
.fe_len
<= 0);
1488 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1489 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1492 ext4_mb_use_best_found(ac
, e4b
);
1493 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1494 /* Sometimes, caller may want to merge even small
1495 * number of blocks to an existing extent */
1496 BUG_ON(ex
.fe_len
<= 0);
1497 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1498 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1501 ext4_mb_use_best_found(ac
, e4b
);
1503 ext4_unlock_group(ac
->ac_sb
, group
);
1504 ext4_mb_release_desc(e4b
);
1510 * The routine scans buddy structures (not bitmap!) from given order
1511 * to max order and tries to find big enough chunk to satisfy the req
1513 static void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1514 struct ext4_buddy
*e4b
)
1516 struct super_block
*sb
= ac
->ac_sb
;
1517 struct ext4_group_info
*grp
= e4b
->bd_info
;
1523 BUG_ON(ac
->ac_2order
<= 0);
1524 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1525 if (grp
->bb_counters
[i
] == 0)
1528 buddy
= mb_find_buddy(e4b
, i
, &max
);
1529 BUG_ON(buddy
== NULL
);
1531 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1536 ac
->ac_b_ex
.fe_len
= 1 << i
;
1537 ac
->ac_b_ex
.fe_start
= k
<< i
;
1538 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1540 ext4_mb_use_best_found(ac
, e4b
);
1542 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1544 if (EXT4_SB(sb
)->s_mb_stats
)
1545 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1552 * The routine scans the group and measures all found extents.
1553 * In order to optimize scanning, caller must pass number of
1554 * free blocks in the group, so the routine can know upper limit.
1556 static void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1557 struct ext4_buddy
*e4b
)
1559 struct super_block
*sb
= ac
->ac_sb
;
1560 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1561 struct ext4_free_extent ex
;
1565 free
= e4b
->bd_info
->bb_free
;
1568 i
= e4b
->bd_info
->bb_first_free
;
1570 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1571 i
= mb_find_next_zero_bit(bitmap
,
1572 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1573 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1575 * IF we have corrupt bitmap, we won't find any
1576 * free blocks even though group info says we
1577 * we have free blocks
1579 ext4_error(sb
, __func__
, "%d free blocks as per "
1580 "group info. But bitmap says 0\n",
1585 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1586 BUG_ON(ex
.fe_len
<= 0);
1587 if (free
< ex
.fe_len
) {
1588 ext4_error(sb
, __func__
, "%d free blocks as per "
1589 "group info. But got %d blocks\n",
1592 * The number of free blocks differs. This mostly
1593 * indicate that the bitmap is corrupt. So exit
1594 * without claiming the space.
1599 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1605 ext4_mb_check_limits(ac
, e4b
, 1);
1609 * This is a special case for storages like raid5
1610 * we try to find stripe-aligned chunks for stripe-size requests
1611 * XXX should do so at least for multiples of stripe size as well
1613 static void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1614 struct ext4_buddy
*e4b
)
1616 struct super_block
*sb
= ac
->ac_sb
;
1617 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1618 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1619 struct ext4_free_extent ex
;
1620 ext4_fsblk_t first_group_block
;
1625 BUG_ON(sbi
->s_stripe
== 0);
1627 /* find first stripe-aligned block in group */
1628 first_group_block
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
)
1629 + le32_to_cpu(sbi
->s_es
->s_first_data_block
);
1630 a
= first_group_block
+ sbi
->s_stripe
- 1;
1631 do_div(a
, sbi
->s_stripe
);
1632 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1634 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1635 if (!mb_test_bit(i
, bitmap
)) {
1636 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1637 if (max
>= sbi
->s_stripe
) {
1640 ext4_mb_use_best_found(ac
, e4b
);
1648 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1649 ext4_group_t group
, int cr
)
1651 unsigned free
, fragments
;
1653 struct ext4_group_desc
*desc
;
1654 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1656 BUG_ON(cr
< 0 || cr
>= 4);
1657 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp
));
1659 free
= grp
->bb_free
;
1660 fragments
= grp
->bb_fragments
;
1668 BUG_ON(ac
->ac_2order
== 0);
1669 /* If this group is uninitialized, skip it initially */
1670 desc
= ext4_get_group_desc(ac
->ac_sb
, group
, NULL
);
1671 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))
1674 bits
= ac
->ac_sb
->s_blocksize_bits
+ 1;
1675 for (i
= ac
->ac_2order
; i
<= bits
; i
++)
1676 if (grp
->bb_counters
[i
] > 0)
1680 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1684 if (free
>= ac
->ac_g_ex
.fe_len
)
1696 static noinline_for_stack
int
1697 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
1704 struct ext4_sb_info
*sbi
;
1705 struct super_block
*sb
;
1706 struct ext4_buddy e4b
;
1711 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1713 /* first, try the goal */
1714 err
= ext4_mb_find_by_goal(ac
, &e4b
);
1715 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
1718 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
1722 * ac->ac2_order is set only if the fe_len is a power of 2
1723 * if ac2_order is set we also set criteria to 0 so that we
1724 * try exact allocation using buddy.
1726 i
= fls(ac
->ac_g_ex
.fe_len
);
1729 * We search using buddy data only if the order of the request
1730 * is greater than equal to the sbi_s_mb_order2_reqs
1731 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1733 if (i
>= sbi
->s_mb_order2_reqs
) {
1735 * This should tell if fe_len is exactly power of 2
1737 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
1738 ac
->ac_2order
= i
- 1;
1741 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
1742 /* if stream allocation is enabled, use global goal */
1743 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
1744 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
1748 if (size
< sbi
->s_mb_stream_request
&&
1749 (ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1750 /* TBD: may be hot point */
1751 spin_lock(&sbi
->s_md_lock
);
1752 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
1753 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
1754 spin_unlock(&sbi
->s_md_lock
);
1756 /* Let's just scan groups to find more-less suitable blocks */
1757 cr
= ac
->ac_2order
? 0 : 1;
1759 * cr == 0 try to get exact allocation,
1760 * cr == 3 try to get anything
1763 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
1764 ac
->ac_criteria
= cr
;
1766 * searching for the right group start
1767 * from the goal value specified
1769 group
= ac
->ac_g_ex
.fe_group
;
1771 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; group
++, i
++) {
1772 struct ext4_group_info
*grp
;
1773 struct ext4_group_desc
*desc
;
1775 if (group
== EXT4_SB(sb
)->s_groups_count
)
1778 /* quick check to skip empty groups */
1779 grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1780 if (grp
->bb_free
== 0)
1784 * if the group is already init we check whether it is
1785 * a good group and if not we don't load the buddy
1787 if (EXT4_MB_GRP_NEED_INIT(grp
)) {
1789 * we need full data about the group
1790 * to make a good selection
1792 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
1795 ext4_mb_release_desc(&e4b
);
1799 * If the particular group doesn't satisfy our
1800 * criteria we continue with the next group
1802 if (!ext4_mb_good_group(ac
, group
, cr
))
1805 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
1809 ext4_lock_group(sb
, group
);
1810 if (!ext4_mb_good_group(ac
, group
, cr
)) {
1811 /* someone did allocation from this group */
1812 ext4_unlock_group(sb
, group
);
1813 ext4_mb_release_desc(&e4b
);
1817 ac
->ac_groups_scanned
++;
1818 desc
= ext4_get_group_desc(sb
, group
, NULL
);
1819 if (cr
== 0 || (desc
->bg_flags
&
1820 cpu_to_le16(EXT4_BG_BLOCK_UNINIT
) &&
1821 ac
->ac_2order
!= 0))
1822 ext4_mb_simple_scan_group(ac
, &e4b
);
1824 ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
)
1825 ext4_mb_scan_aligned(ac
, &e4b
);
1827 ext4_mb_complex_scan_group(ac
, &e4b
);
1829 ext4_unlock_group(sb
, group
);
1830 ext4_mb_release_desc(&e4b
);
1832 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
1837 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
1838 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1840 * We've been searching too long. Let's try to allocate
1841 * the best chunk we've found so far
1844 ext4_mb_try_best_found(ac
, &e4b
);
1845 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
1847 * Someone more lucky has already allocated it.
1848 * The only thing we can do is just take first
1850 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
1852 ac
->ac_b_ex
.fe_group
= 0;
1853 ac
->ac_b_ex
.fe_start
= 0;
1854 ac
->ac_b_ex
.fe_len
= 0;
1855 ac
->ac_status
= AC_STATUS_CONTINUE
;
1856 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
1858 atomic_inc(&sbi
->s_mb_lost_chunks
);
1866 #ifdef EXT4_MB_HISTORY
1867 struct ext4_mb_proc_session
{
1868 struct ext4_mb_history
*history
;
1869 struct super_block
*sb
;
1874 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session
*s
,
1875 struct ext4_mb_history
*hs
,
1878 if (hs
== s
->history
+ s
->max
)
1880 if (!first
&& hs
== s
->history
+ s
->start
)
1882 while (hs
->orig
.fe_len
== 0) {
1884 if (hs
== s
->history
+ s
->max
)
1886 if (hs
== s
->history
+ s
->start
)
1892 static void *ext4_mb_seq_history_start(struct seq_file
*seq
, loff_t
*pos
)
1894 struct ext4_mb_proc_session
*s
= seq
->private;
1895 struct ext4_mb_history
*hs
;
1899 return SEQ_START_TOKEN
;
1900 hs
= ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
1903 while (--l
&& (hs
= ext4_mb_history_skip_empty(s
, ++hs
, 0)) != NULL
);
1907 static void *ext4_mb_seq_history_next(struct seq_file
*seq
, void *v
,
1910 struct ext4_mb_proc_session
*s
= seq
->private;
1911 struct ext4_mb_history
*hs
= v
;
1914 if (v
== SEQ_START_TOKEN
)
1915 return ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
1917 return ext4_mb_history_skip_empty(s
, ++hs
, 0);
1920 static int ext4_mb_seq_history_show(struct seq_file
*seq
, void *v
)
1922 char buf
[25], buf2
[25], buf3
[25], *fmt
;
1923 struct ext4_mb_history
*hs
= v
;
1925 if (v
== SEQ_START_TOKEN
) {
1926 seq_printf(seq
, "%-5s %-8s %-23s %-23s %-23s %-5s "
1927 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
1928 "pid", "inode", "original", "goal", "result", "found",
1929 "grps", "cr", "flags", "merge", "tail", "broken");
1933 if (hs
->op
== EXT4_MB_HISTORY_ALLOC
) {
1934 fmt
= "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
1935 "%-5u %-5s %-5u %-6u\n";
1936 sprintf(buf2
, "%lu/%d/%u@%u", hs
->result
.fe_group
,
1937 hs
->result
.fe_start
, hs
->result
.fe_len
,
1938 hs
->result
.fe_logical
);
1939 sprintf(buf
, "%lu/%d/%u@%u", hs
->orig
.fe_group
,
1940 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
1941 hs
->orig
.fe_logical
);
1942 sprintf(buf3
, "%lu/%d/%u@%u", hs
->goal
.fe_group
,
1943 hs
->goal
.fe_start
, hs
->goal
.fe_len
,
1944 hs
->goal
.fe_logical
);
1945 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, buf3
, buf2
,
1946 hs
->found
, hs
->groups
, hs
->cr
, hs
->flags
,
1947 hs
->merged
? "M" : "", hs
->tail
,
1948 hs
->buddy
? 1 << hs
->buddy
: 0);
1949 } else if (hs
->op
== EXT4_MB_HISTORY_PREALLOC
) {
1950 fmt
= "%-5u %-8u %-23s %-23s %-23s\n";
1951 sprintf(buf2
, "%lu/%d/%u@%u", hs
->result
.fe_group
,
1952 hs
->result
.fe_start
, hs
->result
.fe_len
,
1953 hs
->result
.fe_logical
);
1954 sprintf(buf
, "%lu/%d/%u@%u", hs
->orig
.fe_group
,
1955 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
1956 hs
->orig
.fe_logical
);
1957 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, "", buf2
);
1958 } else if (hs
->op
== EXT4_MB_HISTORY_DISCARD
) {
1959 sprintf(buf2
, "%lu/%d/%u", hs
->result
.fe_group
,
1960 hs
->result
.fe_start
, hs
->result
.fe_len
);
1961 seq_printf(seq
, "%-5u %-8u %-23s discard\n",
1962 hs
->pid
, hs
->ino
, buf2
);
1963 } else if (hs
->op
== EXT4_MB_HISTORY_FREE
) {
1964 sprintf(buf2
, "%lu/%d/%u", hs
->result
.fe_group
,
1965 hs
->result
.fe_start
, hs
->result
.fe_len
);
1966 seq_printf(seq
, "%-5u %-8u %-23s free\n",
1967 hs
->pid
, hs
->ino
, buf2
);
1972 static void ext4_mb_seq_history_stop(struct seq_file
*seq
, void *v
)
1976 static struct seq_operations ext4_mb_seq_history_ops
= {
1977 .start
= ext4_mb_seq_history_start
,
1978 .next
= ext4_mb_seq_history_next
,
1979 .stop
= ext4_mb_seq_history_stop
,
1980 .show
= ext4_mb_seq_history_show
,
1983 static int ext4_mb_seq_history_open(struct inode
*inode
, struct file
*file
)
1985 struct super_block
*sb
= PDE(inode
)->data
;
1986 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1987 struct ext4_mb_proc_session
*s
;
1991 if (unlikely(sbi
->s_mb_history
== NULL
))
1993 s
= kmalloc(sizeof(*s
), GFP_KERNEL
);
1997 size
= sizeof(struct ext4_mb_history
) * sbi
->s_mb_history_max
;
1998 s
->history
= kmalloc(size
, GFP_KERNEL
);
1999 if (s
->history
== NULL
) {
2004 spin_lock(&sbi
->s_mb_history_lock
);
2005 memcpy(s
->history
, sbi
->s_mb_history
, size
);
2006 s
->max
= sbi
->s_mb_history_max
;
2007 s
->start
= sbi
->s_mb_history_cur
% s
->max
;
2008 spin_unlock(&sbi
->s_mb_history_lock
);
2010 rc
= seq_open(file
, &ext4_mb_seq_history_ops
);
2012 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2022 static int ext4_mb_seq_history_release(struct inode
*inode
, struct file
*file
)
2024 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2025 struct ext4_mb_proc_session
*s
= seq
->private;
2028 return seq_release(inode
, file
);
2031 static ssize_t
ext4_mb_seq_history_write(struct file
*file
,
2032 const char __user
*buffer
,
2033 size_t count
, loff_t
*ppos
)
2035 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2036 struct ext4_mb_proc_session
*s
= seq
->private;
2037 struct super_block
*sb
= s
->sb
;
2041 if (count
>= sizeof(str
)) {
2042 printk(KERN_ERR
"EXT4-fs: %s string too long, max %u bytes\n",
2043 "mb_history", (int)sizeof(str
));
2047 if (copy_from_user(str
, buffer
, count
))
2050 value
= simple_strtol(str
, NULL
, 0);
2053 EXT4_SB(sb
)->s_mb_history_filter
= value
;
2058 static struct file_operations ext4_mb_seq_history_fops
= {
2059 .owner
= THIS_MODULE
,
2060 .open
= ext4_mb_seq_history_open
,
2062 .write
= ext4_mb_seq_history_write
,
2063 .llseek
= seq_lseek
,
2064 .release
= ext4_mb_seq_history_release
,
2067 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2069 struct super_block
*sb
= seq
->private;
2070 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2073 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2077 return (void *) group
;
2080 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2082 struct super_block
*sb
= seq
->private;
2083 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2087 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2090 return (void *) group
;;
2093 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2095 struct super_block
*sb
= seq
->private;
2096 long group
= (long) v
;
2099 struct ext4_buddy e4b
;
2101 struct ext4_group_info info
;
2102 unsigned short counters
[16];
2107 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2108 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2109 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2110 "group", "free", "frags", "first",
2111 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2112 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2114 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2115 sizeof(struct ext4_group_info
);
2116 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2118 seq_printf(seq
, "#%-5lu: I/O error\n", group
);
2121 ext4_lock_group(sb
, group
);
2122 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2123 ext4_unlock_group(sb
, group
);
2124 ext4_mb_release_desc(&e4b
);
2126 seq_printf(seq
, "#%-5lu: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2127 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2128 for (i
= 0; i
<= 13; i
++)
2129 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2130 sg
.info
.bb_counters
[i
] : 0);
2131 seq_printf(seq
, " ]\n");
2136 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2140 static struct seq_operations ext4_mb_seq_groups_ops
= {
2141 .start
= ext4_mb_seq_groups_start
,
2142 .next
= ext4_mb_seq_groups_next
,
2143 .stop
= ext4_mb_seq_groups_stop
,
2144 .show
= ext4_mb_seq_groups_show
,
2147 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2149 struct super_block
*sb
= PDE(inode
)->data
;
2152 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2154 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2161 static struct file_operations ext4_mb_seq_groups_fops
= {
2162 .owner
= THIS_MODULE
,
2163 .open
= ext4_mb_seq_groups_open
,
2165 .llseek
= seq_lseek
,
2166 .release
= seq_release
,
2169 static void ext4_mb_history_release(struct super_block
*sb
)
2171 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2173 if (sbi
->s_proc
!= NULL
) {
2174 remove_proc_entry("mb_groups", sbi
->s_proc
);
2175 remove_proc_entry("mb_history", sbi
->s_proc
);
2177 kfree(sbi
->s_mb_history
);
2180 static void ext4_mb_history_init(struct super_block
*sb
)
2182 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2185 if (sbi
->s_proc
!= NULL
) {
2186 proc_create_data("mb_history", S_IRUGO
, sbi
->s_proc
,
2187 &ext4_mb_seq_history_fops
, sb
);
2188 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2189 &ext4_mb_seq_groups_fops
, sb
);
2192 sbi
->s_mb_history_max
= 1000;
2193 sbi
->s_mb_history_cur
= 0;
2194 spin_lock_init(&sbi
->s_mb_history_lock
);
2195 i
= sbi
->s_mb_history_max
* sizeof(struct ext4_mb_history
);
2196 sbi
->s_mb_history
= kzalloc(i
, GFP_KERNEL
);
2197 /* if we can't allocate history, then we simple won't use it */
2200 static noinline_for_stack
void
2201 ext4_mb_store_history(struct ext4_allocation_context
*ac
)
2203 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2204 struct ext4_mb_history h
;
2206 if (unlikely(sbi
->s_mb_history
== NULL
))
2209 if (!(ac
->ac_op
& sbi
->s_mb_history_filter
))
2213 h
.pid
= current
->pid
;
2214 h
.ino
= ac
->ac_inode
? ac
->ac_inode
->i_ino
: 0;
2215 h
.orig
= ac
->ac_o_ex
;
2216 h
.result
= ac
->ac_b_ex
;
2217 h
.flags
= ac
->ac_flags
;
2218 h
.found
= ac
->ac_found
;
2219 h
.groups
= ac
->ac_groups_scanned
;
2220 h
.cr
= ac
->ac_criteria
;
2221 h
.tail
= ac
->ac_tail
;
2222 h
.buddy
= ac
->ac_buddy
;
2224 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
) {
2225 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
2226 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
2228 h
.goal
= ac
->ac_g_ex
;
2229 h
.result
= ac
->ac_f_ex
;
2232 spin_lock(&sbi
->s_mb_history_lock
);
2233 memcpy(sbi
->s_mb_history
+ sbi
->s_mb_history_cur
, &h
, sizeof(h
));
2234 if (++sbi
->s_mb_history_cur
>= sbi
->s_mb_history_max
)
2235 sbi
->s_mb_history_cur
= 0;
2236 spin_unlock(&sbi
->s_mb_history_lock
);
2240 #define ext4_mb_history_release(sb)
2241 #define ext4_mb_history_init(sb)
2245 /* Create and initialize ext4_group_info data for the given group. */
2246 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2247 struct ext4_group_desc
*desc
)
2251 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2252 struct ext4_group_info
**meta_group_info
;
2255 * First check if this group is the first of a reserved block.
2256 * If it's true, we have to allocate a new table of pointers
2257 * to ext4_group_info structures
2259 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2260 metalen
= sizeof(*meta_group_info
) <<
2261 EXT4_DESC_PER_BLOCK_BITS(sb
);
2262 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2263 if (meta_group_info
== NULL
) {
2264 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2266 goto exit_meta_group_info
;
2268 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2273 * calculate needed size. if change bb_counters size,
2274 * don't forget about ext4_mb_generate_buddy()
2276 len
= offsetof(typeof(**meta_group_info
),
2277 bb_counters
[sb
->s_blocksize_bits
+ 2]);
2280 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2281 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2283 meta_group_info
[i
] = kzalloc(len
, GFP_KERNEL
);
2284 if (meta_group_info
[i
] == NULL
) {
2285 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2286 goto exit_group_info
;
2288 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2289 &(meta_group_info
[i
]->bb_state
));
2292 * initialize bb_free to be able to skip
2293 * empty groups without initialization
2295 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2296 meta_group_info
[i
]->bb_free
=
2297 ext4_free_blocks_after_init(sb
, group
, desc
);
2299 meta_group_info
[i
]->bb_free
=
2300 le16_to_cpu(desc
->bg_free_blocks_count
);
2303 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2307 struct buffer_head
*bh
;
2308 meta_group_info
[i
]->bb_bitmap
=
2309 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2310 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2311 bh
= ext4_read_block_bitmap(sb
, group
);
2313 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2322 /* If a meta_group_info table has been allocated, release it now */
2323 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0)
2324 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2325 exit_meta_group_info
:
2327 } /* ext4_mb_add_groupinfo */
2330 * Add a group to the existing groups.
2331 * This function is used for online resize
2333 int ext4_mb_add_more_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2334 struct ext4_group_desc
*desc
)
2336 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2337 struct inode
*inode
= sbi
->s_buddy_cache
;
2338 int blocks_per_page
;
2344 /* Add group based on group descriptor*/
2345 err
= ext4_mb_add_groupinfo(sb
, group
, desc
);
2350 * Cache pages containing dynamic mb_alloc datas (buddy and bitmap
2351 * datas) are set not up to date so that they will be re-initilaized
2352 * during the next call to ext4_mb_load_buddy
2355 /* Set buddy page as not up to date */
2356 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
2358 pnum
= block
/ blocks_per_page
;
2359 page
= find_get_page(inode
->i_mapping
, pnum
);
2361 ClearPageUptodate(page
);
2362 page_cache_release(page
);
2365 /* Set bitmap page as not up to date */
2367 pnum
= block
/ blocks_per_page
;
2368 page
= find_get_page(inode
->i_mapping
, pnum
);
2370 ClearPageUptodate(page
);
2371 page_cache_release(page
);
2378 * Update an existing group.
2379 * This function is used for online resize
2381 void ext4_mb_update_group_info(struct ext4_group_info
*grp
, ext4_grpblk_t add
)
2383 grp
->bb_free
+= add
;
2386 static int ext4_mb_init_backend(struct super_block
*sb
)
2390 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2391 struct ext4_super_block
*es
= sbi
->s_es
;
2392 int num_meta_group_infos
;
2393 int num_meta_group_infos_max
;
2395 struct ext4_group_info
**meta_group_info
;
2396 struct ext4_group_desc
*desc
;
2398 /* This is the number of blocks used by GDT */
2399 num_meta_group_infos
= (sbi
->s_groups_count
+ EXT4_DESC_PER_BLOCK(sb
) -
2400 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2403 * This is the total number of blocks used by GDT including
2404 * the number of reserved blocks for GDT.
2405 * The s_group_info array is allocated with this value
2406 * to allow a clean online resize without a complex
2407 * manipulation of pointer.
2408 * The drawback is the unused memory when no resize
2409 * occurs but it's very low in terms of pages
2410 * (see comments below)
2411 * Need to handle this properly when META_BG resizing is allowed
2413 num_meta_group_infos_max
= num_meta_group_infos
+
2414 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2417 * array_size is the size of s_group_info array. We round it
2418 * to the next power of two because this approximation is done
2419 * internally by kmalloc so we can have some more memory
2420 * for free here (e.g. may be used for META_BG resize).
2423 while (array_size
< sizeof(*sbi
->s_group_info
) *
2424 num_meta_group_infos_max
)
2425 array_size
= array_size
<< 1;
2426 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2427 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2428 * So a two level scheme suffices for now. */
2429 sbi
->s_group_info
= kmalloc(array_size
, GFP_KERNEL
);
2430 if (sbi
->s_group_info
== NULL
) {
2431 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2434 sbi
->s_buddy_cache
= new_inode(sb
);
2435 if (sbi
->s_buddy_cache
== NULL
) {
2436 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2439 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2441 metalen
= sizeof(*meta_group_info
) << EXT4_DESC_PER_BLOCK_BITS(sb
);
2442 for (i
= 0; i
< num_meta_group_infos
; i
++) {
2443 if ((i
+ 1) == num_meta_group_infos
)
2444 metalen
= sizeof(*meta_group_info
) *
2445 (sbi
->s_groups_count
-
2446 (i
<< EXT4_DESC_PER_BLOCK_BITS(sb
)));
2447 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2448 if (meta_group_info
== NULL
) {
2449 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2453 sbi
->s_group_info
[i
] = meta_group_info
;
2456 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2457 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2460 "EXT4-fs: can't read descriptor %lu\n", i
);
2463 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2471 kfree(ext4_get_group_info(sb
, i
));
2472 i
= num_meta_group_infos
;
2475 kfree(sbi
->s_group_info
[i
]);
2476 iput(sbi
->s_buddy_cache
);
2478 kfree(sbi
->s_group_info
);
2482 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2484 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2490 if (!test_opt(sb
, MBALLOC
))
2493 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned short);
2495 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2496 if (sbi
->s_mb_offsets
== NULL
) {
2497 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2500 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2501 if (sbi
->s_mb_maxs
== NULL
) {
2502 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2503 kfree(sbi
->s_mb_maxs
);
2507 /* order 0 is regular bitmap */
2508 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2509 sbi
->s_mb_offsets
[0] = 0;
2513 max
= sb
->s_blocksize
<< 2;
2515 sbi
->s_mb_offsets
[i
] = offset
;
2516 sbi
->s_mb_maxs
[i
] = max
;
2517 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2520 } while (i
<= sb
->s_blocksize_bits
+ 1);
2522 /* init file for buddy data */
2523 ret
= ext4_mb_init_backend(sb
);
2525 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2526 kfree(sbi
->s_mb_offsets
);
2527 kfree(sbi
->s_mb_maxs
);
2531 spin_lock_init(&sbi
->s_md_lock
);
2532 INIT_LIST_HEAD(&sbi
->s_active_transaction
);
2533 INIT_LIST_HEAD(&sbi
->s_closed_transaction
);
2534 INIT_LIST_HEAD(&sbi
->s_committed_transaction
);
2535 spin_lock_init(&sbi
->s_bal_lock
);
2537 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2538 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2539 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2540 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2541 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2542 sbi
->s_mb_history_filter
= EXT4_MB_HISTORY_DEFAULT
;
2543 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2545 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2546 if (sbi
->s_locality_groups
== NULL
) {
2547 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2548 kfree(sbi
->s_mb_offsets
);
2549 kfree(sbi
->s_mb_maxs
);
2552 for_each_possible_cpu(i
) {
2553 struct ext4_locality_group
*lg
;
2554 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2555 mutex_init(&lg
->lg_mutex
);
2556 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2557 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2558 spin_lock_init(&lg
->lg_prealloc_lock
);
2561 ext4_mb_init_per_dev_proc(sb
);
2562 ext4_mb_history_init(sb
);
2564 printk(KERN_INFO
"EXT4-fs: mballoc enabled\n");
2568 /* need to called with ext4 group lock (ext4_lock_group) */
2569 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2571 struct ext4_prealloc_space
*pa
;
2572 struct list_head
*cur
, *tmp
;
2575 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2576 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2577 list_del(&pa
->pa_group_list
);
2582 mb_debug("mballoc: %u PAs left\n", count
);
2586 int ext4_mb_release(struct super_block
*sb
)
2589 int num_meta_group_infos
;
2590 struct ext4_group_info
*grinfo
;
2591 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2593 if (!test_opt(sb
, MBALLOC
))
2596 /* release freed, non-committed blocks */
2597 spin_lock(&sbi
->s_md_lock
);
2598 list_splice_init(&sbi
->s_closed_transaction
,
2599 &sbi
->s_committed_transaction
);
2600 list_splice_init(&sbi
->s_active_transaction
,
2601 &sbi
->s_committed_transaction
);
2602 spin_unlock(&sbi
->s_md_lock
);
2603 ext4_mb_free_committed_blocks(sb
);
2605 if (sbi
->s_group_info
) {
2606 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2607 grinfo
= ext4_get_group_info(sb
, i
);
2609 kfree(grinfo
->bb_bitmap
);
2611 ext4_lock_group(sb
, i
);
2612 ext4_mb_cleanup_pa(grinfo
);
2613 ext4_unlock_group(sb
, i
);
2616 num_meta_group_infos
= (sbi
->s_groups_count
+
2617 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2618 EXT4_DESC_PER_BLOCK_BITS(sb
);
2619 for (i
= 0; i
< num_meta_group_infos
; i
++)
2620 kfree(sbi
->s_group_info
[i
]);
2621 kfree(sbi
->s_group_info
);
2623 kfree(sbi
->s_mb_offsets
);
2624 kfree(sbi
->s_mb_maxs
);
2625 if (sbi
->s_buddy_cache
)
2626 iput(sbi
->s_buddy_cache
);
2627 if (sbi
->s_mb_stats
) {
2629 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2630 atomic_read(&sbi
->s_bal_allocated
),
2631 atomic_read(&sbi
->s_bal_reqs
),
2632 atomic_read(&sbi
->s_bal_success
));
2634 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2635 "%u 2^N hits, %u breaks, %u lost\n",
2636 atomic_read(&sbi
->s_bal_ex_scanned
),
2637 atomic_read(&sbi
->s_bal_goals
),
2638 atomic_read(&sbi
->s_bal_2orders
),
2639 atomic_read(&sbi
->s_bal_breaks
),
2640 atomic_read(&sbi
->s_mb_lost_chunks
));
2642 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2643 sbi
->s_mb_buddies_generated
++,
2644 sbi
->s_mb_generation_time
);
2646 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2647 atomic_read(&sbi
->s_mb_preallocated
),
2648 atomic_read(&sbi
->s_mb_discarded
));
2651 free_percpu(sbi
->s_locality_groups
);
2652 ext4_mb_history_release(sb
);
2653 ext4_mb_destroy_per_dev_proc(sb
);
2658 static noinline_for_stack
void
2659 ext4_mb_free_committed_blocks(struct super_block
*sb
)
2661 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2666 struct ext4_free_metadata
*md
;
2667 struct ext4_buddy e4b
;
2669 if (list_empty(&sbi
->s_committed_transaction
))
2672 /* there is committed blocks to be freed yet */
2674 /* get next array of blocks */
2676 spin_lock(&sbi
->s_md_lock
);
2677 if (!list_empty(&sbi
->s_committed_transaction
)) {
2678 md
= list_entry(sbi
->s_committed_transaction
.next
,
2679 struct ext4_free_metadata
, list
);
2680 list_del(&md
->list
);
2682 spin_unlock(&sbi
->s_md_lock
);
2687 mb_debug("gonna free %u blocks in group %lu (0x%p):",
2688 md
->num
, md
->group
, md
);
2690 err
= ext4_mb_load_buddy(sb
, md
->group
, &e4b
);
2691 /* we expect to find existing buddy because it's pinned */
2694 /* there are blocks to put in buddy to make them really free */
2697 ext4_lock_group(sb
, md
->group
);
2698 for (i
= 0; i
< md
->num
; i
++) {
2699 mb_debug(" %u", md
->blocks
[i
]);
2700 mb_free_blocks(NULL
, &e4b
, md
->blocks
[i
], 1);
2703 ext4_unlock_group(sb
, md
->group
);
2705 /* balance refcounts from ext4_mb_free_metadata() */
2706 page_cache_release(e4b
.bd_buddy_page
);
2707 page_cache_release(e4b
.bd_bitmap_page
);
2710 ext4_mb_release_desc(&e4b
);
2714 mb_debug("freed %u blocks in %u structures\n", count
, count2
);
2717 #define EXT4_MB_STATS_NAME "stats"
2718 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2719 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2720 #define EXT4_MB_ORDER2_REQ "order2_req"
2721 #define EXT4_MB_STREAM_REQ "stream_req"
2722 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2724 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
)
2726 mode_t mode
= S_IFREG
| S_IRUGO
| S_IWUSR
;
2727 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2728 struct proc_dir_entry
*proc
;
2730 if (sbi
->s_proc
== NULL
)
2733 EXT4_PROC_HANDLER(EXT4_MB_STATS_NAME
, mb_stats
);
2734 EXT4_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME
, mb_max_to_scan
);
2735 EXT4_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME
, mb_min_to_scan
);
2736 EXT4_PROC_HANDLER(EXT4_MB_ORDER2_REQ
, mb_order2_reqs
);
2737 EXT4_PROC_HANDLER(EXT4_MB_STREAM_REQ
, mb_stream_request
);
2738 EXT4_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC
, mb_group_prealloc
);
2742 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2743 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2744 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2745 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2746 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2747 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2751 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
)
2753 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2755 if (sbi
->s_proc
== NULL
)
2758 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2759 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2760 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2761 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2762 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2763 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2768 int __init
init_ext4_mballoc(void)
2770 ext4_pspace_cachep
=
2771 kmem_cache_create("ext4_prealloc_space",
2772 sizeof(struct ext4_prealloc_space
),
2773 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2774 if (ext4_pspace_cachep
== NULL
)
2778 kmem_cache_create("ext4_alloc_context",
2779 sizeof(struct ext4_allocation_context
),
2780 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2781 if (ext4_ac_cachep
== NULL
) {
2782 kmem_cache_destroy(ext4_pspace_cachep
);
2788 void exit_ext4_mballoc(void)
2790 /* XXX: synchronize_rcu(); */
2791 kmem_cache_destroy(ext4_pspace_cachep
);
2792 kmem_cache_destroy(ext4_ac_cachep
);
2797 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2798 * Returns 0 if success or error code
2800 static noinline_for_stack
int
2801 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2802 handle_t
*handle
, unsigned long reserv_blks
)
2804 struct buffer_head
*bitmap_bh
= NULL
;
2805 struct ext4_super_block
*es
;
2806 struct ext4_group_desc
*gdp
;
2807 struct buffer_head
*gdp_bh
;
2808 struct ext4_sb_info
*sbi
;
2809 struct super_block
*sb
;
2813 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2814 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2822 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2826 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2831 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2835 ext4_debug("using block group %lu(%d)\n", ac
->ac_b_ex
.fe_group
,
2836 gdp
->bg_free_blocks_count
);
2838 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2842 block
= ac
->ac_b_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
2843 + ac
->ac_b_ex
.fe_start
2844 + le32_to_cpu(es
->s_first_data_block
);
2846 len
= ac
->ac_b_ex
.fe_len
;
2847 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, len
) ||
2848 in_range(ext4_inode_bitmap(sb
, gdp
), block
, len
) ||
2849 in_range(block
, ext4_inode_table(sb
, gdp
),
2850 EXT4_SB(sb
)->s_itb_per_group
) ||
2851 in_range(block
+ len
- 1, ext4_inode_table(sb
, gdp
),
2852 EXT4_SB(sb
)->s_itb_per_group
)) {
2853 ext4_error(sb
, __func__
,
2854 "Allocating block in system zone - block = %llu",
2856 /* File system mounted not to panic on error
2857 * Fix the bitmap and repeat the block allocation
2858 * We leak some of the blocks here.
2860 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
),
2861 bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2862 ac
->ac_b_ex
.fe_len
);
2863 err
= ext4_journal_dirty_metadata(handle
, bitmap_bh
);
2868 #ifdef AGGRESSIVE_CHECK
2871 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2872 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2873 bitmap_bh
->b_data
));
2877 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
), bitmap_bh
->b_data
,
2878 ac
->ac_b_ex
.fe_start
, ac
->ac_b_ex
.fe_len
);
2880 spin_lock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
2881 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2882 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2883 gdp
->bg_free_blocks_count
=
2884 cpu_to_le16(ext4_free_blocks_after_init(sb
,
2885 ac
->ac_b_ex
.fe_group
,
2888 le16_add_cpu(&gdp
->bg_free_blocks_count
, -ac
->ac_b_ex
.fe_len
);
2889 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
2890 spin_unlock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
2891 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
2893 * Now reduce the dirty block count also. Should not go negative
2895 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2896 /* release all the reserved blocks if non delalloc */
2897 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
, reserv_blks
);
2899 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
,
2900 ac
->ac_b_ex
.fe_len
);
2902 if (sbi
->s_log_groups_per_flex
) {
2903 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2904 ac
->ac_b_ex
.fe_group
);
2905 spin_lock(sb_bgl_lock(sbi
, flex_group
));
2906 sbi
->s_flex_groups
[flex_group
].free_blocks
-= ac
->ac_b_ex
.fe_len
;
2907 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
2910 err
= ext4_journal_dirty_metadata(handle
, bitmap_bh
);
2913 err
= ext4_journal_dirty_metadata(handle
, gdp_bh
);
2922 * here we normalize request for locality group
2923 * Group request are normalized to s_strip size if we set the same via mount
2924 * option. If not we set it to s_mb_group_prealloc which can be configured via
2925 * /proc/fs/ext4/<partition>/group_prealloc
2927 * XXX: should we try to preallocate more than the group has now?
2929 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2931 struct super_block
*sb
= ac
->ac_sb
;
2932 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2935 if (EXT4_SB(sb
)->s_stripe
)
2936 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
2938 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2939 mb_debug("#%u: goal %u blocks for locality group\n",
2940 current
->pid
, ac
->ac_g_ex
.fe_len
);
2944 * Normalization means making request better in terms of
2945 * size and alignment
2947 static noinline_for_stack
void
2948 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
2949 struct ext4_allocation_request
*ar
)
2953 loff_t size
, orig_size
, start_off
;
2954 ext4_lblk_t start
, orig_start
;
2955 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
2956 struct ext4_prealloc_space
*pa
;
2958 /* do normalize only data requests, metadata requests
2959 do not need preallocation */
2960 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
2963 /* sometime caller may want exact blocks */
2964 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2967 /* caller may indicate that preallocation isn't
2968 * required (it's a tail, for example) */
2969 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
2972 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
2973 ext4_mb_normalize_group_request(ac
);
2977 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
2979 /* first, let's learn actual file size
2980 * given current request is allocated */
2981 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
2982 size
= size
<< bsbits
;
2983 if (size
< i_size_read(ac
->ac_inode
))
2984 size
= i_size_read(ac
->ac_inode
);
2986 /* max size of free chunks */
2989 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2990 (req <= (size) || max <= (chunk_size))
2992 /* first, try to predict filesize */
2993 /* XXX: should this table be tunable? */
2995 if (size
<= 16 * 1024) {
2997 } else if (size
<= 32 * 1024) {
2999 } else if (size
<= 64 * 1024) {
3001 } else if (size
<= 128 * 1024) {
3003 } else if (size
<= 256 * 1024) {
3005 } else if (size
<= 512 * 1024) {
3007 } else if (size
<= 1024 * 1024) {
3009 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3010 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3011 (21 - bsbits
)) << 21;
3012 size
= 2 * 1024 * 1024;
3013 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3014 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3015 (22 - bsbits
)) << 22;
3016 size
= 4 * 1024 * 1024;
3017 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3018 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3019 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3020 (23 - bsbits
)) << 23;
3021 size
= 8 * 1024 * 1024;
3023 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3024 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3026 orig_size
= size
= size
>> bsbits
;
3027 orig_start
= start
= start_off
>> bsbits
;
3029 /* don't cover already allocated blocks in selected range */
3030 if (ar
->pleft
&& start
<= ar
->lleft
) {
3031 size
-= ar
->lleft
+ 1 - start
;
3032 start
= ar
->lleft
+ 1;
3034 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3035 size
-= start
+ size
- ar
->lright
;
3039 /* check we don't cross already preallocated blocks */
3041 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3042 unsigned long pa_end
;
3046 spin_lock(&pa
->pa_lock
);
3047 if (pa
->pa_deleted
) {
3048 spin_unlock(&pa
->pa_lock
);
3052 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3054 /* PA must not overlap original request */
3055 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3056 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3058 /* skip PA normalized request doesn't overlap with */
3059 if (pa
->pa_lstart
>= end
) {
3060 spin_unlock(&pa
->pa_lock
);
3063 if (pa_end
<= start
) {
3064 spin_unlock(&pa
->pa_lock
);
3067 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3069 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3070 BUG_ON(pa_end
< start
);
3074 if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3075 BUG_ON(pa
->pa_lstart
> end
);
3076 end
= pa
->pa_lstart
;
3078 spin_unlock(&pa
->pa_lock
);
3083 /* XXX: extra loop to check we really don't overlap preallocations */
3085 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3086 unsigned long pa_end
;
3087 spin_lock(&pa
->pa_lock
);
3088 if (pa
->pa_deleted
== 0) {
3089 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3090 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3092 spin_unlock(&pa
->pa_lock
);
3096 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3097 start
> ac
->ac_o_ex
.fe_logical
) {
3098 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3099 (unsigned long) start
, (unsigned long) size
,
3100 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3102 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3103 start
> ac
->ac_o_ex
.fe_logical
);
3104 BUG_ON(size
<= 0 || size
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3106 /* now prepare goal request */
3108 /* XXX: is it better to align blocks WRT to logical
3109 * placement or satisfy big request as is */
3110 ac
->ac_g_ex
.fe_logical
= start
;
3111 ac
->ac_g_ex
.fe_len
= size
;
3113 /* define goal start in order to merge */
3114 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3115 /* merge to the right */
3116 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3117 &ac
->ac_f_ex
.fe_group
,
3118 &ac
->ac_f_ex
.fe_start
);
3119 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3121 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3122 /* merge to the left */
3123 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3124 &ac
->ac_f_ex
.fe_group
,
3125 &ac
->ac_f_ex
.fe_start
);
3126 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3129 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3130 (unsigned) orig_size
, (unsigned) start
);
3133 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3135 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3137 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3138 atomic_inc(&sbi
->s_bal_reqs
);
3139 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3140 if (ac
->ac_o_ex
.fe_len
>= ac
->ac_g_ex
.fe_len
)
3141 atomic_inc(&sbi
->s_bal_success
);
3142 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3143 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3144 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3145 atomic_inc(&sbi
->s_bal_goals
);
3146 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3147 atomic_inc(&sbi
->s_bal_breaks
);
3150 ext4_mb_store_history(ac
);
3154 * use blocks preallocated to inode
3156 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3157 struct ext4_prealloc_space
*pa
)
3163 /* found preallocated blocks, use them */
3164 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3165 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3167 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3168 &ac
->ac_b_ex
.fe_start
);
3169 ac
->ac_b_ex
.fe_len
= len
;
3170 ac
->ac_status
= AC_STATUS_FOUND
;
3173 BUG_ON(start
< pa
->pa_pstart
);
3174 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3175 BUG_ON(pa
->pa_free
< len
);
3178 mb_debug("use %llu/%u from inode pa %p\n", start
, len
, pa
);
3182 * use blocks preallocated to locality group
3184 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3185 struct ext4_prealloc_space
*pa
)
3187 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3189 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3190 &ac
->ac_b_ex
.fe_group
,
3191 &ac
->ac_b_ex
.fe_start
);
3192 ac
->ac_b_ex
.fe_len
= len
;
3193 ac
->ac_status
= AC_STATUS_FOUND
;
3196 /* we don't correct pa_pstart or pa_plen here to avoid
3197 * possible race when the group is being loaded concurrently
3198 * instead we correct pa later, after blocks are marked
3199 * in on-disk bitmap -- see ext4_mb_release_context()
3200 * Other CPUs are prevented from allocating from this pa by lg_mutex
3202 mb_debug("use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3206 * Return the prealloc space that have minimal distance
3207 * from the goal block. @cpa is the prealloc
3208 * space that is having currently known minimal distance
3209 * from the goal block.
3211 static struct ext4_prealloc_space
*
3212 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3213 struct ext4_prealloc_space
*pa
,
3214 struct ext4_prealloc_space
*cpa
)
3216 ext4_fsblk_t cur_distance
, new_distance
;
3219 atomic_inc(&pa
->pa_count
);
3222 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3223 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3225 if (cur_distance
< new_distance
)
3228 /* drop the previous reference */
3229 atomic_dec(&cpa
->pa_count
);
3230 atomic_inc(&pa
->pa_count
);
3235 * search goal blocks in preallocated space
3237 static noinline_for_stack
int
3238 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3241 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3242 struct ext4_locality_group
*lg
;
3243 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3244 ext4_fsblk_t goal_block
;
3246 /* only data can be preallocated */
3247 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3250 /* first, try per-file preallocation */
3252 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3254 /* all fields in this condition don't change,
3255 * so we can skip locking for them */
3256 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3257 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3260 /* found preallocated blocks, use them */
3261 spin_lock(&pa
->pa_lock
);
3262 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3263 atomic_inc(&pa
->pa_count
);
3264 ext4_mb_use_inode_pa(ac
, pa
);
3265 spin_unlock(&pa
->pa_lock
);
3266 ac
->ac_criteria
= 10;
3270 spin_unlock(&pa
->pa_lock
);
3274 /* can we use group allocation? */
3275 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3278 /* inode may have no locality group for some reason */
3282 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3283 if (order
> PREALLOC_TB_SIZE
- 1)
3284 /* The max size of hash table is PREALLOC_TB_SIZE */
3285 order
= PREALLOC_TB_SIZE
- 1;
3287 goal_block
= ac
->ac_g_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
) +
3288 ac
->ac_g_ex
.fe_start
+
3289 le32_to_cpu(EXT4_SB(ac
->ac_sb
)->s_es
->s_first_data_block
);
3291 * search for the prealloc space that is having
3292 * minimal distance from the goal block.
3294 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3296 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3298 spin_lock(&pa
->pa_lock
);
3299 if (pa
->pa_deleted
== 0 &&
3300 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3302 cpa
= ext4_mb_check_group_pa(goal_block
,
3305 spin_unlock(&pa
->pa_lock
);
3310 ext4_mb_use_group_pa(ac
, cpa
);
3311 ac
->ac_criteria
= 20;
3318 * the function goes through all preallocation in this group and marks them
3319 * used in in-core bitmap. buddy must be generated from this bitmap
3320 * Need to be called with ext4 group lock (ext4_lock_group)
3322 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3325 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3326 struct ext4_prealloc_space
*pa
;
3327 struct list_head
*cur
;
3328 ext4_group_t groupnr
;
3329 ext4_grpblk_t start
;
3330 int preallocated
= 0;
3334 /* all form of preallocation discards first load group,
3335 * so the only competing code is preallocation use.
3336 * we don't need any locking here
3337 * notice we do NOT ignore preallocations with pa_deleted
3338 * otherwise we could leave used blocks available for
3339 * allocation in buddy when concurrent ext4_mb_put_pa()
3340 * is dropping preallocation
3342 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3343 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3344 spin_lock(&pa
->pa_lock
);
3345 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3348 spin_unlock(&pa
->pa_lock
);
3349 if (unlikely(len
== 0))
3351 BUG_ON(groupnr
!= group
);
3352 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3353 bitmap
, start
, len
);
3354 preallocated
+= len
;
3357 mb_debug("prellocated %u for group %lu\n", preallocated
, group
);
3360 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3362 struct ext4_prealloc_space
*pa
;
3363 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3364 kmem_cache_free(ext4_pspace_cachep
, pa
);
3368 * drops a reference to preallocated space descriptor
3369 * if this was the last reference and the space is consumed
3371 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3372 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3376 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3379 /* in this short window concurrent discard can set pa_deleted */
3380 spin_lock(&pa
->pa_lock
);
3381 if (pa
->pa_deleted
== 1) {
3382 spin_unlock(&pa
->pa_lock
);
3387 spin_unlock(&pa
->pa_lock
);
3389 /* -1 is to protect from crossing allocation group */
3390 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
- 1, &grp
, NULL
);
3395 * P1 (buddy init) P2 (regular allocation)
3396 * find block B in PA
3397 * copy on-disk bitmap to buddy
3398 * mark B in on-disk bitmap
3399 * drop PA from group
3400 * mark all PAs in buddy
3402 * thus, P1 initializes buddy with B available. to prevent this
3403 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3406 ext4_lock_group(sb
, grp
);
3407 list_del(&pa
->pa_group_list
);
3408 ext4_unlock_group(sb
, grp
);
3410 spin_lock(pa
->pa_obj_lock
);
3411 list_del_rcu(&pa
->pa_inode_list
);
3412 spin_unlock(pa
->pa_obj_lock
);
3414 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3418 * creates new preallocated space for given inode
3420 static noinline_for_stack
int
3421 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3423 struct super_block
*sb
= ac
->ac_sb
;
3424 struct ext4_prealloc_space
*pa
;
3425 struct ext4_group_info
*grp
;
3426 struct ext4_inode_info
*ei
;
3428 /* preallocate only when found space is larger then requested */
3429 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3430 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3431 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3433 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3437 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3443 /* we can't allocate as much as normalizer wants.
3444 * so, found space must get proper lstart
3445 * to cover original request */
3446 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3447 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3449 /* we're limited by original request in that
3450 * logical block must be covered any way
3451 * winl is window we can move our chunk within */
3452 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3454 /* also, we should cover whole original request */
3455 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3457 /* the smallest one defines real window */
3458 win
= min(winl
, wins
);
3460 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3461 if (offs
&& offs
< win
)
3464 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3465 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3466 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3469 /* preallocation can change ac_b_ex, thus we store actually
3470 * allocated blocks for history */
3471 ac
->ac_f_ex
= ac
->ac_b_ex
;
3473 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3474 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3475 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3476 pa
->pa_free
= pa
->pa_len
;
3477 atomic_set(&pa
->pa_count
, 1);
3478 spin_lock_init(&pa
->pa_lock
);
3482 mb_debug("new inode pa %p: %llu/%u for %u\n", pa
,
3483 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3485 ext4_mb_use_inode_pa(ac
, pa
);
3486 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3488 ei
= EXT4_I(ac
->ac_inode
);
3489 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3491 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3492 pa
->pa_inode
= ac
->ac_inode
;
3494 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3495 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3496 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3498 spin_lock(pa
->pa_obj_lock
);
3499 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3500 spin_unlock(pa
->pa_obj_lock
);
3506 * creates new preallocated space for locality group inodes belongs to
3508 static noinline_for_stack
int
3509 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3511 struct super_block
*sb
= ac
->ac_sb
;
3512 struct ext4_locality_group
*lg
;
3513 struct ext4_prealloc_space
*pa
;
3514 struct ext4_group_info
*grp
;
3516 /* preallocate only when found space is larger then requested */
3517 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3518 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3519 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3521 BUG_ON(ext4_pspace_cachep
== NULL
);
3522 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3526 /* preallocation can change ac_b_ex, thus we store actually
3527 * allocated blocks for history */
3528 ac
->ac_f_ex
= ac
->ac_b_ex
;
3530 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3531 pa
->pa_lstart
= pa
->pa_pstart
;
3532 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3533 pa
->pa_free
= pa
->pa_len
;
3534 atomic_set(&pa
->pa_count
, 1);
3535 spin_lock_init(&pa
->pa_lock
);
3536 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3540 mb_debug("new group pa %p: %llu/%u for %u\n", pa
,
3541 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3543 ext4_mb_use_group_pa(ac
, pa
);
3544 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3546 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3550 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3551 pa
->pa_inode
= NULL
;
3553 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3554 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3555 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3558 * We will later add the new pa to the right bucket
3559 * after updating the pa_free in ext4_mb_release_context
3564 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3568 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3569 err
= ext4_mb_new_group_pa(ac
);
3571 err
= ext4_mb_new_inode_pa(ac
);
3576 * finds all unused blocks in on-disk bitmap, frees them in
3577 * in-core bitmap and buddy.
3578 * @pa must be unlinked from inode and group lists, so that
3579 * nobody else can find/use it.
3580 * the caller MUST hold group/inode locks.
3581 * TODO: optimize the case when there are no in-core structures yet
3583 static noinline_for_stack
int
3584 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3585 struct ext4_prealloc_space
*pa
,
3586 struct ext4_allocation_context
*ac
)
3588 struct super_block
*sb
= e4b
->bd_sb
;
3589 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3598 BUG_ON(pa
->pa_deleted
== 0);
3599 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3600 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3601 end
= bit
+ pa
->pa_len
;
3605 ac
->ac_inode
= pa
->pa_inode
;
3606 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3610 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3613 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3614 start
= group
* EXT4_BLOCKS_PER_GROUP(sb
) + bit
+
3615 le32_to_cpu(sbi
->s_es
->s_first_data_block
);
3616 mb_debug(" free preallocated %u/%u in group %u\n",
3617 (unsigned) start
, (unsigned) next
- bit
,
3622 ac
->ac_b_ex
.fe_group
= group
;
3623 ac
->ac_b_ex
.fe_start
= bit
;
3624 ac
->ac_b_ex
.fe_len
= next
- bit
;
3625 ac
->ac_b_ex
.fe_logical
= 0;
3626 ext4_mb_store_history(ac
);
3629 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3632 if (free
!= pa
->pa_free
) {
3633 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3634 pa
, (unsigned long) pa
->pa_lstart
,
3635 (unsigned long) pa
->pa_pstart
,
3636 (unsigned long) pa
->pa_len
);
3637 ext4_error(sb
, __func__
, "free %u, pa_free %u\n",
3640 * pa is already deleted so we use the value obtained
3641 * from the bitmap and continue.
3644 atomic_add(free
, &sbi
->s_mb_discarded
);
3649 static noinline_for_stack
int
3650 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3651 struct ext4_prealloc_space
*pa
,
3652 struct ext4_allocation_context
*ac
)
3654 struct super_block
*sb
= e4b
->bd_sb
;
3659 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3661 BUG_ON(pa
->pa_deleted
== 0);
3662 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3663 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3664 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3665 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3669 ac
->ac_inode
= NULL
;
3670 ac
->ac_b_ex
.fe_group
= group
;
3671 ac
->ac_b_ex
.fe_start
= bit
;
3672 ac
->ac_b_ex
.fe_len
= pa
->pa_len
;
3673 ac
->ac_b_ex
.fe_logical
= 0;
3674 ext4_mb_store_history(ac
);
3681 * releases all preallocations in given group
3683 * first, we need to decide discard policy:
3684 * - when do we discard
3686 * - how many do we discard
3687 * 1) how many requested
3689 static noinline_for_stack
int
3690 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3691 ext4_group_t group
, int needed
)
3693 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3694 struct buffer_head
*bitmap_bh
= NULL
;
3695 struct ext4_prealloc_space
*pa
, *tmp
;
3696 struct ext4_allocation_context
*ac
;
3697 struct list_head list
;
3698 struct ext4_buddy e4b
;
3703 mb_debug("discard preallocation for group %lu\n", group
);
3705 if (list_empty(&grp
->bb_prealloc_list
))
3708 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3709 if (bitmap_bh
== NULL
) {
3710 ext4_error(sb
, __func__
, "Error in reading block "
3711 "bitmap for %lu\n", group
);
3715 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3717 ext4_error(sb
, __func__
, "Error in loading buddy "
3718 "information for %lu\n", group
);
3724 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3726 INIT_LIST_HEAD(&list
);
3727 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3729 ext4_lock_group(sb
, group
);
3730 list_for_each_entry_safe(pa
, tmp
,
3731 &grp
->bb_prealloc_list
, pa_group_list
) {
3732 spin_lock(&pa
->pa_lock
);
3733 if (atomic_read(&pa
->pa_count
)) {
3734 spin_unlock(&pa
->pa_lock
);
3738 if (pa
->pa_deleted
) {
3739 spin_unlock(&pa
->pa_lock
);
3743 /* seems this one can be freed ... */
3746 /* we can trust pa_free ... */
3747 free
+= pa
->pa_free
;
3749 spin_unlock(&pa
->pa_lock
);
3751 list_del(&pa
->pa_group_list
);
3752 list_add(&pa
->u
.pa_tmp_list
, &list
);
3755 /* if we still need more blocks and some PAs were used, try again */
3756 if (free
< needed
&& busy
) {
3758 ext4_unlock_group(sb
, group
);
3760 * Yield the CPU here so that we don't get soft lockup
3761 * in non preempt case.
3767 /* found anything to free? */
3768 if (list_empty(&list
)) {
3773 /* now free all selected PAs */
3774 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3776 /* remove from object (inode or locality group) */
3777 spin_lock(pa
->pa_obj_lock
);
3778 list_del_rcu(&pa
->pa_inode_list
);
3779 spin_unlock(pa
->pa_obj_lock
);
3782 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
3784 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3786 list_del(&pa
->u
.pa_tmp_list
);
3787 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3791 ext4_unlock_group(sb
, group
);
3793 kmem_cache_free(ext4_ac_cachep
, ac
);
3794 ext4_mb_release_desc(&e4b
);
3800 * releases all non-used preallocated blocks for given inode
3802 * It's important to discard preallocations under i_data_sem
3803 * We don't want another block to be served from the prealloc
3804 * space when we are discarding the inode prealloc space.
3806 * FIXME!! Make sure it is valid at all the call sites
3808 void ext4_mb_discard_inode_preallocations(struct inode
*inode
)
3810 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3811 struct super_block
*sb
= inode
->i_sb
;
3812 struct buffer_head
*bitmap_bh
= NULL
;
3813 struct ext4_prealloc_space
*pa
, *tmp
;
3814 struct ext4_allocation_context
*ac
;
3815 ext4_group_t group
= 0;
3816 struct list_head list
;
3817 struct ext4_buddy e4b
;
3820 if (!test_opt(sb
, MBALLOC
) || !S_ISREG(inode
->i_mode
)) {
3821 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3825 mb_debug("discard preallocation for inode %lu\n", inode
->i_ino
);
3827 INIT_LIST_HEAD(&list
);
3829 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3831 /* first, collect all pa's in the inode */
3832 spin_lock(&ei
->i_prealloc_lock
);
3833 while (!list_empty(&ei
->i_prealloc_list
)) {
3834 pa
= list_entry(ei
->i_prealloc_list
.next
,
3835 struct ext4_prealloc_space
, pa_inode_list
);
3836 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3837 spin_lock(&pa
->pa_lock
);
3838 if (atomic_read(&pa
->pa_count
)) {
3839 /* this shouldn't happen often - nobody should
3840 * use preallocation while we're discarding it */
3841 spin_unlock(&pa
->pa_lock
);
3842 spin_unlock(&ei
->i_prealloc_lock
);
3843 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
3845 schedule_timeout_uninterruptible(HZ
);
3849 if (pa
->pa_deleted
== 0) {
3851 spin_unlock(&pa
->pa_lock
);
3852 list_del_rcu(&pa
->pa_inode_list
);
3853 list_add(&pa
->u
.pa_tmp_list
, &list
);
3857 /* someone is deleting pa right now */
3858 spin_unlock(&pa
->pa_lock
);
3859 spin_unlock(&ei
->i_prealloc_lock
);
3861 /* we have to wait here because pa_deleted
3862 * doesn't mean pa is already unlinked from
3863 * the list. as we might be called from
3864 * ->clear_inode() the inode will get freed
3865 * and concurrent thread which is unlinking
3866 * pa from inode's list may access already
3867 * freed memory, bad-bad-bad */
3869 /* XXX: if this happens too often, we can
3870 * add a flag to force wait only in case
3871 * of ->clear_inode(), but not in case of
3872 * regular truncate */
3873 schedule_timeout_uninterruptible(HZ
);
3876 spin_unlock(&ei
->i_prealloc_lock
);
3878 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3879 BUG_ON(pa
->pa_linear
!= 0);
3880 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
3882 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3884 ext4_error(sb
, __func__
, "Error in loading buddy "
3885 "information for %lu\n", group
);
3889 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3890 if (bitmap_bh
== NULL
) {
3891 ext4_error(sb
, __func__
, "Error in reading block "
3892 "bitmap for %lu\n", group
);
3893 ext4_mb_release_desc(&e4b
);
3897 ext4_lock_group(sb
, group
);
3898 list_del(&pa
->pa_group_list
);
3899 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3900 ext4_unlock_group(sb
, group
);
3902 ext4_mb_release_desc(&e4b
);
3905 list_del(&pa
->u
.pa_tmp_list
);
3906 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3909 kmem_cache_free(ext4_ac_cachep
, ac
);
3913 * finds all preallocated spaces and return blocks being freed to them
3914 * if preallocated space becomes full (no block is used from the space)
3915 * then the function frees space in buddy
3916 * XXX: at the moment, truncate (which is the only way to free blocks)
3917 * discards all preallocations
3919 static void ext4_mb_return_to_preallocation(struct inode
*inode
,
3920 struct ext4_buddy
*e4b
,
3921 sector_t block
, int count
)
3923 BUG_ON(!list_empty(&EXT4_I(inode
)->i_prealloc_list
));
3926 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3928 struct super_block
*sb
= ac
->ac_sb
;
3931 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
3932 " Allocation context details:\n");
3933 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
3934 ac
->ac_status
, ac
->ac_flags
);
3935 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3936 "best %lu/%lu/%lu@%lu cr %d\n",
3937 (unsigned long)ac
->ac_o_ex
.fe_group
,
3938 (unsigned long)ac
->ac_o_ex
.fe_start
,
3939 (unsigned long)ac
->ac_o_ex
.fe_len
,
3940 (unsigned long)ac
->ac_o_ex
.fe_logical
,
3941 (unsigned long)ac
->ac_g_ex
.fe_group
,
3942 (unsigned long)ac
->ac_g_ex
.fe_start
,
3943 (unsigned long)ac
->ac_g_ex
.fe_len
,
3944 (unsigned long)ac
->ac_g_ex
.fe_logical
,
3945 (unsigned long)ac
->ac_b_ex
.fe_group
,
3946 (unsigned long)ac
->ac_b_ex
.fe_start
,
3947 (unsigned long)ac
->ac_b_ex
.fe_len
,
3948 (unsigned long)ac
->ac_b_ex
.fe_logical
,
3949 (int)ac
->ac_criteria
);
3950 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
3952 printk(KERN_ERR
"EXT4-fs: groups: \n");
3953 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; i
++) {
3954 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
3955 struct ext4_prealloc_space
*pa
;
3956 ext4_grpblk_t start
;
3957 struct list_head
*cur
;
3958 ext4_lock_group(sb
, i
);
3959 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3960 pa
= list_entry(cur
, struct ext4_prealloc_space
,
3962 spin_lock(&pa
->pa_lock
);
3963 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3965 spin_unlock(&pa
->pa_lock
);
3966 printk(KERN_ERR
"PA:%lu:%d:%u \n", i
,
3969 ext4_unlock_group(sb
, i
);
3971 if (grp
->bb_free
== 0)
3973 printk(KERN_ERR
"%lu: %d/%d \n",
3974 i
, grp
->bb_free
, grp
->bb_fragments
);
3976 printk(KERN_ERR
"\n");
3979 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3986 * We use locality group preallocation for small size file. The size of the
3987 * file is determined by the current size or the resulting size after
3988 * allocation which ever is larger
3990 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
3992 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
3994 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3995 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3998 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4001 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
4002 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
4003 size
= max(size
, isize
);
4005 /* don't use group allocation for large files */
4006 if (size
>= sbi
->s_mb_stream_request
)
4009 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4012 BUG_ON(ac
->ac_lg
!= NULL
);
4014 * locality group prealloc space are per cpu. The reason for having
4015 * per cpu locality group is to reduce the contention between block
4016 * request from multiple CPUs.
4018 ac
->ac_lg
= per_cpu_ptr(sbi
->s_locality_groups
, raw_smp_processor_id());
4020 /* we're going to use group allocation */
4021 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4023 /* serialize all allocations in the group */
4024 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4027 static noinline_for_stack
int
4028 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4029 struct ext4_allocation_request
*ar
)
4031 struct super_block
*sb
= ar
->inode
->i_sb
;
4032 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4033 struct ext4_super_block
*es
= sbi
->s_es
;
4037 ext4_grpblk_t block
;
4039 /* we can't allocate > group size */
4042 /* just a dirty hack to filter too big requests */
4043 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4044 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4046 /* start searching from the goal */
4048 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4049 goal
>= ext4_blocks_count(es
))
4050 goal
= le32_to_cpu(es
->s_first_data_block
);
4051 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4053 /* set up allocation goals */
4054 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4055 ac
->ac_b_ex
.fe_group
= 0;
4056 ac
->ac_b_ex
.fe_start
= 0;
4057 ac
->ac_b_ex
.fe_len
= 0;
4058 ac
->ac_status
= AC_STATUS_CONTINUE
;
4059 ac
->ac_groups_scanned
= 0;
4060 ac
->ac_ex_scanned
= 0;
4063 ac
->ac_inode
= ar
->inode
;
4064 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4065 ac
->ac_o_ex
.fe_group
= group
;
4066 ac
->ac_o_ex
.fe_start
= block
;
4067 ac
->ac_o_ex
.fe_len
= len
;
4068 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4069 ac
->ac_g_ex
.fe_group
= group
;
4070 ac
->ac_g_ex
.fe_start
= block
;
4071 ac
->ac_g_ex
.fe_len
= len
;
4072 ac
->ac_f_ex
.fe_len
= 0;
4073 ac
->ac_flags
= ar
->flags
;
4075 ac
->ac_criteria
= 0;
4077 ac
->ac_bitmap_page
= NULL
;
4078 ac
->ac_buddy_page
= NULL
;
4081 /* we have to define context: we'll we work with a file or
4082 * locality group. this is a policy, actually */
4083 ext4_mb_group_or_file(ac
);
4085 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4086 "left: %u/%u, right %u/%u to %swritable\n",
4087 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4088 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4089 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4090 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4091 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4096 static noinline_for_stack
void
4097 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4098 struct ext4_locality_group
*lg
,
4099 int order
, int total_entries
)
4101 ext4_group_t group
= 0;
4102 struct ext4_buddy e4b
;
4103 struct list_head discard_list
;
4104 struct ext4_prealloc_space
*pa
, *tmp
;
4105 struct ext4_allocation_context
*ac
;
4107 mb_debug("discard locality group preallocation\n");
4109 INIT_LIST_HEAD(&discard_list
);
4110 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4112 spin_lock(&lg
->lg_prealloc_lock
);
4113 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4115 spin_lock(&pa
->pa_lock
);
4116 if (atomic_read(&pa
->pa_count
)) {
4118 * This is the pa that we just used
4119 * for block allocation. So don't
4122 spin_unlock(&pa
->pa_lock
);
4125 if (pa
->pa_deleted
) {
4126 spin_unlock(&pa
->pa_lock
);
4129 /* only lg prealloc space */
4130 BUG_ON(!pa
->pa_linear
);
4132 /* seems this one can be freed ... */
4134 spin_unlock(&pa
->pa_lock
);
4136 list_del_rcu(&pa
->pa_inode_list
);
4137 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4140 if (total_entries
<= 5) {
4142 * we want to keep only 5 entries
4143 * allowing it to grow to 8. This
4144 * mak sure we don't call discard
4145 * soon for this list.
4150 spin_unlock(&lg
->lg_prealloc_lock
);
4152 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4154 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4155 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4156 ext4_error(sb
, __func__
, "Error in loading buddy "
4157 "information for %lu\n", group
);
4160 ext4_lock_group(sb
, group
);
4161 list_del(&pa
->pa_group_list
);
4162 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
4163 ext4_unlock_group(sb
, group
);
4165 ext4_mb_release_desc(&e4b
);
4166 list_del(&pa
->u
.pa_tmp_list
);
4167 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4170 kmem_cache_free(ext4_ac_cachep
, ac
);
4174 * We have incremented pa_count. So it cannot be freed at this
4175 * point. Also we hold lg_mutex. So no parallel allocation is
4176 * possible from this lg. That means pa_free cannot be updated.
4178 * A parallel ext4_mb_discard_group_preallocations is possible.
4179 * which can cause the lg_prealloc_list to be updated.
4182 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4184 int order
, added
= 0, lg_prealloc_count
= 1;
4185 struct super_block
*sb
= ac
->ac_sb
;
4186 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4187 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4189 order
= fls(pa
->pa_free
) - 1;
4190 if (order
> PREALLOC_TB_SIZE
- 1)
4191 /* The max size of hash table is PREALLOC_TB_SIZE */
4192 order
= PREALLOC_TB_SIZE
- 1;
4193 /* Add the prealloc space to lg */
4195 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4197 spin_lock(&tmp_pa
->pa_lock
);
4198 if (tmp_pa
->pa_deleted
) {
4199 spin_unlock(&pa
->pa_lock
);
4202 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4203 /* Add to the tail of the previous entry */
4204 list_add_tail_rcu(&pa
->pa_inode_list
,
4205 &tmp_pa
->pa_inode_list
);
4208 * we want to count the total
4209 * number of entries in the list
4212 spin_unlock(&tmp_pa
->pa_lock
);
4213 lg_prealloc_count
++;
4216 list_add_tail_rcu(&pa
->pa_inode_list
,
4217 &lg
->lg_prealloc_list
[order
]);
4220 /* Now trim the list to be not more than 8 elements */
4221 if (lg_prealloc_count
> 8) {
4222 ext4_mb_discard_lg_preallocations(sb
, lg
,
4223 order
, lg_prealloc_count
);
4230 * release all resource we used in allocation
4232 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4234 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4236 if (pa
->pa_linear
) {
4237 /* see comment in ext4_mb_use_group_pa() */
4238 spin_lock(&pa
->pa_lock
);
4239 pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4240 pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4241 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4242 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4243 spin_unlock(&pa
->pa_lock
);
4245 * We want to add the pa to the right bucket.
4246 * Remove it from the list and while adding
4247 * make sure the list to which we are adding
4250 if (likely(pa
->pa_free
)) {
4251 spin_lock(pa
->pa_obj_lock
);
4252 list_del_rcu(&pa
->pa_inode_list
);
4253 spin_unlock(pa
->pa_obj_lock
);
4254 ext4_mb_add_n_trim(ac
);
4257 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4259 if (ac
->ac_bitmap_page
)
4260 page_cache_release(ac
->ac_bitmap_page
);
4261 if (ac
->ac_buddy_page
)
4262 page_cache_release(ac
->ac_buddy_page
);
4263 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4264 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4265 ext4_mb_collect_stats(ac
);
4269 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4275 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
&& needed
> 0; i
++) {
4276 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4285 * Main entry point into mballoc to allocate blocks
4286 * it tries to use preallocation first, then falls back
4287 * to usual allocation
4289 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4290 struct ext4_allocation_request
*ar
, int *errp
)
4293 struct ext4_allocation_context
*ac
= NULL
;
4294 struct ext4_sb_info
*sbi
;
4295 struct super_block
*sb
;
4296 ext4_fsblk_t block
= 0;
4297 unsigned long inquota
;
4298 unsigned long reserv_blks
= 0;
4300 sb
= ar
->inode
->i_sb
;
4303 if (!test_opt(sb
, MBALLOC
)) {
4304 block
= ext4_old_new_blocks(handle
, ar
->inode
, ar
->goal
,
4308 if (!EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
) {
4310 * With delalloc we already reserved the blocks
4312 while (ar
->len
&& ext4_claim_free_blocks(sbi
, ar
->len
)) {
4313 /* let others to free the space */
4315 ar
->len
= ar
->len
>> 1;
4321 reserv_blks
= ar
->len
;
4323 while (ar
->len
&& DQUOT_ALLOC_BLOCK(ar
->inode
, ar
->len
)) {
4324 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4333 if (EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
)
4334 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4336 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4343 ext4_mb_poll_new_transaction(sb
, handle
);
4345 *errp
= ext4_mb_initialize_context(ac
, ar
);
4351 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4352 if (!ext4_mb_use_preallocated(ac
)) {
4353 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4354 ext4_mb_normalize_request(ac
, ar
);
4356 /* allocate space in core */
4357 ext4_mb_regular_allocator(ac
);
4359 /* as we've just preallocated more space than
4360 * user requested orinally, we store allocated
4361 * space in a special descriptor */
4362 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4363 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4364 ext4_mb_new_preallocation(ac
);
4367 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4368 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_blks
);
4369 if (*errp
== -EAGAIN
) {
4370 ac
->ac_b_ex
.fe_group
= 0;
4371 ac
->ac_b_ex
.fe_start
= 0;
4372 ac
->ac_b_ex
.fe_len
= 0;
4373 ac
->ac_status
= AC_STATUS_CONTINUE
;
4376 ac
->ac_b_ex
.fe_len
= 0;
4378 ext4_mb_show_ac(ac
);
4380 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4381 ar
->len
= ac
->ac_b_ex
.fe_len
;
4384 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4388 ac
->ac_b_ex
.fe_len
= 0;
4390 ext4_mb_show_ac(ac
);
4393 ext4_mb_release_context(ac
);
4396 kmem_cache_free(ext4_ac_cachep
, ac
);
4398 if (ar
->len
< inquota
)
4399 DQUOT_FREE_BLOCK(ar
->inode
, inquota
- ar
->len
);
4403 static void ext4_mb_poll_new_transaction(struct super_block
*sb
,
4406 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4408 if (sbi
->s_last_transaction
== handle
->h_transaction
->t_tid
)
4411 /* new transaction! time to close last one and free blocks for
4412 * committed transaction. we know that only transaction can be
4413 * active, so previos transaction can be being logged and we
4414 * know that transaction before previous is known to be already
4415 * logged. this means that now we may free blocks freed in all
4416 * transactions before previous one. hope I'm clear enough ... */
4418 spin_lock(&sbi
->s_md_lock
);
4419 if (sbi
->s_last_transaction
!= handle
->h_transaction
->t_tid
) {
4420 mb_debug("new transaction %lu, old %lu\n",
4421 (unsigned long) handle
->h_transaction
->t_tid
,
4422 (unsigned long) sbi
->s_last_transaction
);
4423 list_splice_init(&sbi
->s_closed_transaction
,
4424 &sbi
->s_committed_transaction
);
4425 list_splice_init(&sbi
->s_active_transaction
,
4426 &sbi
->s_closed_transaction
);
4427 sbi
->s_last_transaction
= handle
->h_transaction
->t_tid
;
4429 spin_unlock(&sbi
->s_md_lock
);
4431 ext4_mb_free_committed_blocks(sb
);
4434 static noinline_for_stack
int
4435 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4436 ext4_group_t group
, ext4_grpblk_t block
, int count
)
4438 struct ext4_group_info
*db
= e4b
->bd_info
;
4439 struct super_block
*sb
= e4b
->bd_sb
;
4440 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4441 struct ext4_free_metadata
*md
;
4444 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4445 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4447 ext4_lock_group(sb
, group
);
4448 for (i
= 0; i
< count
; i
++) {
4450 if (md
&& db
->bb_tid
!= handle
->h_transaction
->t_tid
) {
4451 db
->bb_md_cur
= NULL
;
4456 ext4_unlock_group(sb
, group
);
4457 md
= kmalloc(sizeof(*md
), GFP_NOFS
);
4463 ext4_lock_group(sb
, group
);
4464 if (db
->bb_md_cur
== NULL
) {
4465 spin_lock(&sbi
->s_md_lock
);
4466 list_add(&md
->list
, &sbi
->s_active_transaction
);
4467 spin_unlock(&sbi
->s_md_lock
);
4468 /* protect buddy cache from being freed,
4469 * otherwise we'll refresh it from
4470 * on-disk bitmap and lose not-yet-available
4472 page_cache_get(e4b
->bd_buddy_page
);
4473 page_cache_get(e4b
->bd_bitmap_page
);
4475 db
->bb_tid
= handle
->h_transaction
->t_tid
;
4476 mb_debug("new md 0x%p for group %lu\n",
4484 BUG_ON(md
->num
>= EXT4_BB_MAX_BLOCKS
);
4485 md
->blocks
[md
->num
] = block
+ i
;
4487 if (md
->num
== EXT4_BB_MAX_BLOCKS
) {
4488 /* no more space, put full container on a sb's list */
4489 db
->bb_md_cur
= NULL
;
4492 ext4_unlock_group(sb
, group
);
4497 * Main entry point into mballoc to free blocks
4499 void ext4_mb_free_blocks(handle_t
*handle
, struct inode
*inode
,
4500 unsigned long block
, unsigned long count
,
4501 int metadata
, unsigned long *freed
)
4503 struct buffer_head
*bitmap_bh
= NULL
;
4504 struct super_block
*sb
= inode
->i_sb
;
4505 struct ext4_allocation_context
*ac
= NULL
;
4506 struct ext4_group_desc
*gdp
;
4507 struct ext4_super_block
*es
;
4508 unsigned long overflow
;
4510 struct buffer_head
*gd_bh
;
4511 ext4_group_t block_group
;
4512 struct ext4_sb_info
*sbi
;
4513 struct ext4_buddy e4b
;
4519 ext4_mb_poll_new_transaction(sb
, handle
);
4522 es
= EXT4_SB(sb
)->s_es
;
4523 if (block
< le32_to_cpu(es
->s_first_data_block
) ||
4524 block
+ count
< block
||
4525 block
+ count
> ext4_blocks_count(es
)) {
4526 ext4_error(sb
, __func__
,
4527 "Freeing blocks not in datazone - "
4528 "block = %lu, count = %lu", block
, count
);
4532 ext4_debug("freeing block %lu\n", block
);
4534 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4536 ac
->ac_op
= EXT4_MB_HISTORY_FREE
;
4537 ac
->ac_inode
= inode
;
4543 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4546 * Check to see if we are freeing blocks across a group
4549 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4550 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4553 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4558 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4564 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4565 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4566 in_range(block
, ext4_inode_table(sb
, gdp
),
4567 EXT4_SB(sb
)->s_itb_per_group
) ||
4568 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4569 EXT4_SB(sb
)->s_itb_per_group
)) {
4571 ext4_error(sb
, __func__
,
4572 "Freeing blocks in system zone - "
4573 "Block = %lu, count = %lu", block
, count
);
4574 /* err = 0. ext4_std_error should be a no op */
4578 BUFFER_TRACE(bitmap_bh
, "getting write access");
4579 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4584 * We are about to modify some metadata. Call the journal APIs
4585 * to unshare ->b_data if a currently-committing transaction is
4588 BUFFER_TRACE(gd_bh
, "get_write_access");
4589 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4593 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4597 #ifdef AGGRESSIVE_CHECK
4600 for (i
= 0; i
< count
; i
++)
4601 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4604 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4607 /* We dirtied the bitmap block */
4608 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4609 err
= ext4_journal_dirty_metadata(handle
, bitmap_bh
);
4612 ac
->ac_b_ex
.fe_group
= block_group
;
4613 ac
->ac_b_ex
.fe_start
= bit
;
4614 ac
->ac_b_ex
.fe_len
= count
;
4615 ext4_mb_store_history(ac
);
4619 /* blocks being freed are metadata. these blocks shouldn't
4620 * be used until this transaction is committed */
4621 ext4_mb_free_metadata(handle
, &e4b
, block_group
, bit
, count
);
4623 ext4_lock_group(sb
, block_group
);
4624 mb_free_blocks(inode
, &e4b
, bit
, count
);
4625 ext4_mb_return_to_preallocation(inode
, &e4b
, block
, count
);
4626 ext4_unlock_group(sb
, block_group
);
4629 spin_lock(sb_bgl_lock(sbi
, block_group
));
4630 le16_add_cpu(&gdp
->bg_free_blocks_count
, count
);
4631 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4632 spin_unlock(sb_bgl_lock(sbi
, block_group
));
4633 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4635 if (sbi
->s_log_groups_per_flex
) {
4636 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4637 spin_lock(sb_bgl_lock(sbi
, flex_group
));
4638 sbi
->s_flex_groups
[flex_group
].free_blocks
+= count
;
4639 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
4642 ext4_mb_release_desc(&e4b
);
4646 /* And the group descriptor block */
4647 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4648 ret
= ext4_journal_dirty_metadata(handle
, gd_bh
);
4652 if (overflow
&& !err
) {
4661 ext4_std_error(sb
, err
);
4663 kmem_cache_free(ext4_ac_cachep
, ac
);