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
;
538 static int mb_check_counter
;
539 if (mb_check_counter
++ % 100 != 0)
544 buddy
= mb_find_buddy(e4b
, order
, &max
);
545 MB_CHECK_ASSERT(buddy
);
546 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
547 MB_CHECK_ASSERT(buddy2
);
548 MB_CHECK_ASSERT(buddy
!= buddy2
);
549 MB_CHECK_ASSERT(max
* 2 == max2
);
552 for (i
= 0; i
< max
; i
++) {
554 if (mb_test_bit(i
, buddy
)) {
555 /* only single bit in buddy2 may be 1 */
556 if (!mb_test_bit(i
<< 1, buddy2
)) {
558 mb_test_bit((i
<<1)+1, buddy2
));
559 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
561 mb_test_bit(i
<< 1, buddy2
));
566 /* both bits in buddy2 must be 0 */
567 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
568 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
570 for (j
= 0; j
< (1 << order
); j
++) {
571 k
= (i
* (1 << order
)) + j
;
573 !mb_test_bit(k
, EXT4_MB_BITMAP(e4b
)));
577 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
582 buddy
= mb_find_buddy(e4b
, 0, &max
);
583 for (i
= 0; i
< max
; i
++) {
584 if (!mb_test_bit(i
, buddy
)) {
585 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
593 /* check used bits only */
594 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
595 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
597 MB_CHECK_ASSERT(k
< max2
);
598 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
601 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
602 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
604 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
605 buddy
= mb_find_buddy(e4b
, 0, &max
);
606 list_for_each(cur
, &grp
->bb_prealloc_list
) {
607 ext4_group_t groupnr
;
608 struct ext4_prealloc_space
*pa
;
609 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
610 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
611 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
612 for (i
= 0; i
< pa
->pa_len
; i
++)
613 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
617 #undef MB_CHECK_ASSERT
618 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
619 __FILE__, __func__, __LINE__)
621 #define mb_check_buddy(e4b)
624 /* FIXME!! need more doc */
625 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
626 void *buddy
, unsigned first
, int len
,
627 struct ext4_group_info
*grp
)
629 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
632 unsigned short chunk
;
633 unsigned short border
;
635 BUG_ON(len
> EXT4_BLOCKS_PER_GROUP(sb
));
637 border
= 2 << sb
->s_blocksize_bits
;
640 /* find how many blocks can be covered since this position */
641 max
= ffs(first
| border
) - 1;
643 /* find how many blocks of power 2 we need to mark */
650 /* mark multiblock chunks only */
651 grp
->bb_counters
[min
]++;
653 mb_clear_bit(first
>> min
,
654 buddy
+ sbi
->s_mb_offsets
[min
]);
661 static void ext4_mb_generate_buddy(struct super_block
*sb
,
662 void *buddy
, void *bitmap
, ext4_group_t group
)
664 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
665 unsigned short max
= EXT4_BLOCKS_PER_GROUP(sb
);
666 unsigned short i
= 0;
667 unsigned short first
;
670 unsigned fragments
= 0;
671 unsigned long long period
= get_cycles();
673 /* initialize buddy from bitmap which is aggregation
674 * of on-disk bitmap and preallocations */
675 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
676 grp
->bb_first_free
= i
;
680 i
= mb_find_next_bit(bitmap
, max
, i
);
684 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
686 grp
->bb_counters
[0]++;
688 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
690 grp
->bb_fragments
= fragments
;
692 if (free
!= grp
->bb_free
) {
693 ext4_error(sb
, __func__
,
694 "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
695 group
, free
, grp
->bb_free
);
697 * If we intent to continue, we consider group descritor
698 * corrupt and update bb_free using bitmap value
703 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
705 period
= get_cycles() - period
;
706 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
707 EXT4_SB(sb
)->s_mb_buddies_generated
++;
708 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
709 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
712 /* The buddy information is attached the buddy cache inode
713 * for convenience. The information regarding each group
714 * is loaded via ext4_mb_load_buddy. The information involve
715 * block bitmap and buddy information. The information are
716 * stored in the inode as
719 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
722 * one block each for bitmap and buddy information.
723 * So for each group we take up 2 blocks. A page can
724 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
725 * So it can have information regarding groups_per_page which
726 * is blocks_per_page/2
729 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
736 ext4_group_t first_group
;
738 struct super_block
*sb
;
739 struct buffer_head
*bhs
;
740 struct buffer_head
**bh
;
745 mb_debug("init page %lu\n", page
->index
);
747 inode
= page
->mapping
->host
;
749 blocksize
= 1 << inode
->i_blkbits
;
750 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
752 groups_per_page
= blocks_per_page
>> 1;
753 if (groups_per_page
== 0)
756 /* allocate buffer_heads to read bitmaps */
757 if (groups_per_page
> 1) {
759 i
= sizeof(struct buffer_head
*) * groups_per_page
;
760 bh
= kzalloc(i
, GFP_NOFS
);
766 first_group
= page
->index
* blocks_per_page
/ 2;
768 /* read all groups the page covers into the cache */
769 for (i
= 0; i
< groups_per_page
; i
++) {
770 struct ext4_group_desc
*desc
;
772 if (first_group
+ i
>= EXT4_SB(sb
)->s_groups_count
)
776 desc
= ext4_get_group_desc(sb
, first_group
+ i
, NULL
);
781 bh
[i
] = sb_getblk(sb
, ext4_block_bitmap(sb
, desc
));
785 if (bh_uptodate_or_lock(bh
[i
]))
788 spin_lock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
789 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
790 ext4_init_block_bitmap(sb
, bh
[i
],
791 first_group
+ i
, desc
);
792 set_buffer_uptodate(bh
[i
]);
793 unlock_buffer(bh
[i
]);
794 spin_unlock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
797 spin_unlock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
799 bh
[i
]->b_end_io
= end_buffer_read_sync
;
800 submit_bh(READ
, bh
[i
]);
801 mb_debug("read bitmap for group %lu\n", first_group
+ i
);
804 /* wait for I/O completion */
805 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
806 wait_on_buffer(bh
[i
]);
809 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
810 if (!buffer_uptodate(bh
[i
]))
814 first_block
= page
->index
* blocks_per_page
;
815 for (i
= 0; i
< blocks_per_page
; i
++) {
817 struct ext4_group_info
*grinfo
;
819 group
= (first_block
+ i
) >> 1;
820 if (group
>= EXT4_SB(sb
)->s_groups_count
)
824 * data carry information regarding this
825 * particular group in the format specified
829 data
= page_address(page
) + (i
* blocksize
);
830 bitmap
= bh
[group
- first_group
]->b_data
;
833 * We place the buddy block and bitmap block
836 if ((first_block
+ i
) & 1) {
837 /* this is block of buddy */
838 BUG_ON(incore
== NULL
);
839 mb_debug("put buddy for group %u in page %lu/%x\n",
840 group
, page
->index
, i
* blocksize
);
841 memset(data
, 0xff, blocksize
);
842 grinfo
= ext4_get_group_info(sb
, group
);
843 grinfo
->bb_fragments
= 0;
844 memset(grinfo
->bb_counters
, 0,
845 sizeof(unsigned short)*(sb
->s_blocksize_bits
+2));
847 * incore got set to the group block bitmap below
849 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
852 /* this is block of bitmap */
853 BUG_ON(incore
!= NULL
);
854 mb_debug("put bitmap for group %u in page %lu/%x\n",
855 group
, page
->index
, i
* blocksize
);
857 /* see comments in ext4_mb_put_pa() */
858 ext4_lock_group(sb
, group
);
859 memcpy(data
, bitmap
, blocksize
);
861 /* mark all preallocated blks used in in-core bitmap */
862 ext4_mb_generate_from_pa(sb
, data
, group
);
863 ext4_unlock_group(sb
, group
);
865 /* set incore so that the buddy information can be
866 * generated using this
871 SetPageUptodate(page
);
875 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
883 static noinline_for_stack
int
884 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
885 struct ext4_buddy
*e4b
)
887 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
888 struct inode
*inode
= sbi
->s_buddy_cache
;
896 mb_debug("load group %lu\n", group
);
898 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
900 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
901 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
903 e4b
->bd_group
= group
;
904 e4b
->bd_buddy_page
= NULL
;
905 e4b
->bd_bitmap_page
= NULL
;
908 * the buddy cache inode stores the block bitmap
909 * and buddy information in consecutive blocks.
910 * So for each group we need two blocks.
913 pnum
= block
/ blocks_per_page
;
914 poff
= block
% blocks_per_page
;
916 /* we could use find_or_create_page(), but it locks page
917 * what we'd like to avoid in fast path ... */
918 page
= find_get_page(inode
->i_mapping
, pnum
);
919 if (page
== NULL
|| !PageUptodate(page
)) {
921 page_cache_release(page
);
922 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
924 BUG_ON(page
->mapping
!= inode
->i_mapping
);
925 if (!PageUptodate(page
)) {
926 ret
= ext4_mb_init_cache(page
, NULL
);
931 mb_cmp_bitmaps(e4b
, page_address(page
) +
932 (poff
* sb
->s_blocksize
));
937 if (page
== NULL
|| !PageUptodate(page
)) {
941 e4b
->bd_bitmap_page
= page
;
942 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
943 mark_page_accessed(page
);
946 pnum
= block
/ blocks_per_page
;
947 poff
= block
% blocks_per_page
;
949 page
= find_get_page(inode
->i_mapping
, pnum
);
950 if (page
== NULL
|| !PageUptodate(page
)) {
952 page_cache_release(page
);
953 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
955 BUG_ON(page
->mapping
!= inode
->i_mapping
);
956 if (!PageUptodate(page
)) {
957 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
966 if (page
== NULL
|| !PageUptodate(page
)) {
970 e4b
->bd_buddy_page
= page
;
971 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
972 mark_page_accessed(page
);
974 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
975 BUG_ON(e4b
->bd_buddy_page
== NULL
);
980 if (e4b
->bd_bitmap_page
)
981 page_cache_release(e4b
->bd_bitmap_page
);
982 if (e4b
->bd_buddy_page
)
983 page_cache_release(e4b
->bd_buddy_page
);
984 e4b
->bd_buddy
= NULL
;
985 e4b
->bd_bitmap
= NULL
;
989 static void ext4_mb_release_desc(struct ext4_buddy
*e4b
)
991 if (e4b
->bd_bitmap_page
)
992 page_cache_release(e4b
->bd_bitmap_page
);
993 if (e4b
->bd_buddy_page
)
994 page_cache_release(e4b
->bd_buddy_page
);
998 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1003 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1004 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1006 bb
= EXT4_MB_BUDDY(e4b
);
1007 while (order
<= e4b
->bd_blkbits
+ 1) {
1009 if (!mb_test_bit(block
, bb
)) {
1010 /* this block is part of buddy of order 'order' */
1013 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1019 static void mb_clear_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1025 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1026 /* fast path: clear whole word at once */
1027 addr
= bm
+ (cur
>> 3);
1032 mb_clear_bit_atomic(lock
, cur
, bm
);
1037 static void mb_set_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1043 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1044 /* fast path: set whole word at once */
1045 addr
= bm
+ (cur
>> 3);
1050 mb_set_bit_atomic(lock
, cur
, bm
);
1055 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1056 int first
, int count
)
1063 struct super_block
*sb
= e4b
->bd_sb
;
1065 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1066 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
1067 mb_check_buddy(e4b
);
1068 mb_free_blocks_double(inode
, e4b
, first
, count
);
1070 e4b
->bd_info
->bb_free
+= count
;
1071 if (first
< e4b
->bd_info
->bb_first_free
)
1072 e4b
->bd_info
->bb_first_free
= first
;
1074 /* let's maintain fragments counter */
1076 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1077 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1078 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1080 e4b
->bd_info
->bb_fragments
--;
1081 else if (!block
&& !max
)
1082 e4b
->bd_info
->bb_fragments
++;
1084 /* let's maintain buddy itself */
1085 while (count
-- > 0) {
1089 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1090 ext4_fsblk_t blocknr
;
1091 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
1094 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
1095 ext4_unlock_group(sb
, e4b
->bd_group
);
1096 ext4_error(sb
, __func__
, "double-free of inode"
1097 " %lu's block %llu(bit %u in group %lu)\n",
1098 inode
? inode
->i_ino
: 0, blocknr
, block
,
1100 ext4_lock_group(sb
, e4b
->bd_group
);
1102 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1103 e4b
->bd_info
->bb_counters
[order
]++;
1105 /* start of the buddy */
1106 buddy
= mb_find_buddy(e4b
, order
, &max
);
1110 if (mb_test_bit(block
, buddy
) ||
1111 mb_test_bit(block
+ 1, buddy
))
1114 /* both the buddies are free, try to coalesce them */
1115 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1121 /* for special purposes, we don't set
1122 * free bits in bitmap */
1123 mb_set_bit(block
, buddy
);
1124 mb_set_bit(block
+ 1, buddy
);
1126 e4b
->bd_info
->bb_counters
[order
]--;
1127 e4b
->bd_info
->bb_counters
[order
]--;
1131 e4b
->bd_info
->bb_counters
[order
]++;
1133 mb_clear_bit(block
, buddy2
);
1137 mb_check_buddy(e4b
);
1140 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1141 int needed
, struct ext4_free_extent
*ex
)
1148 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1151 buddy
= mb_find_buddy(e4b
, order
, &max
);
1152 BUG_ON(buddy
== NULL
);
1153 BUG_ON(block
>= max
);
1154 if (mb_test_bit(block
, buddy
)) {
1161 /* FIXME dorp order completely ? */
1162 if (likely(order
== 0)) {
1163 /* find actual order */
1164 order
= mb_find_order_for_block(e4b
, block
);
1165 block
= block
>> order
;
1168 ex
->fe_len
= 1 << order
;
1169 ex
->fe_start
= block
<< order
;
1170 ex
->fe_group
= e4b
->bd_group
;
1172 /* calc difference from given start */
1173 next
= next
- ex
->fe_start
;
1175 ex
->fe_start
+= next
;
1177 while (needed
> ex
->fe_len
&&
1178 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1180 if (block
+ 1 >= max
)
1183 next
= (block
+ 1) * (1 << order
);
1184 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1187 ord
= mb_find_order_for_block(e4b
, next
);
1190 block
= next
>> order
;
1191 ex
->fe_len
+= 1 << order
;
1194 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1198 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1204 int start
= ex
->fe_start
;
1205 int len
= ex
->fe_len
;
1210 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1211 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1212 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1213 mb_check_buddy(e4b
);
1214 mb_mark_used_double(e4b
, start
, len
);
1216 e4b
->bd_info
->bb_free
-= len
;
1217 if (e4b
->bd_info
->bb_first_free
== start
)
1218 e4b
->bd_info
->bb_first_free
+= len
;
1220 /* let's maintain fragments counter */
1222 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1223 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1224 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1226 e4b
->bd_info
->bb_fragments
++;
1227 else if (!mlen
&& !max
)
1228 e4b
->bd_info
->bb_fragments
--;
1230 /* let's maintain buddy itself */
1232 ord
= mb_find_order_for_block(e4b
, start
);
1234 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1235 /* the whole chunk may be allocated at once! */
1237 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1238 BUG_ON((start
>> ord
) >= max
);
1239 mb_set_bit(start
>> ord
, buddy
);
1240 e4b
->bd_info
->bb_counters
[ord
]--;
1247 /* store for history */
1249 ret
= len
| (ord
<< 16);
1251 /* we have to split large buddy */
1253 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1254 mb_set_bit(start
>> ord
, buddy
);
1255 e4b
->bd_info
->bb_counters
[ord
]--;
1258 cur
= (start
>> ord
) & ~1U;
1259 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1260 mb_clear_bit(cur
, buddy
);
1261 mb_clear_bit(cur
+ 1, buddy
);
1262 e4b
->bd_info
->bb_counters
[ord
]++;
1263 e4b
->bd_info
->bb_counters
[ord
]++;
1266 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b
->bd_sb
), ex
->fe_group
),
1267 EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1268 mb_check_buddy(e4b
);
1274 * Must be called under group lock!
1276 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1277 struct ext4_buddy
*e4b
)
1279 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1282 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1283 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1285 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1286 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1287 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1289 /* preallocation can change ac_b_ex, thus we store actually
1290 * allocated blocks for history */
1291 ac
->ac_f_ex
= ac
->ac_b_ex
;
1293 ac
->ac_status
= AC_STATUS_FOUND
;
1294 ac
->ac_tail
= ret
& 0xffff;
1295 ac
->ac_buddy
= ret
>> 16;
1297 /* XXXXXXX: SUCH A HORRIBLE **CK */
1299 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1300 get_page(ac
->ac_bitmap_page
);
1301 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1302 get_page(ac
->ac_buddy_page
);
1304 /* store last allocated for subsequent stream allocation */
1305 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1306 spin_lock(&sbi
->s_md_lock
);
1307 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1308 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1309 spin_unlock(&sbi
->s_md_lock
);
1314 * regular allocator, for general purposes allocation
1317 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1318 struct ext4_buddy
*e4b
,
1321 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1322 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1323 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1324 struct ext4_free_extent ex
;
1328 * We don't want to scan for a whole year
1330 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1331 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1332 ac
->ac_status
= AC_STATUS_BREAK
;
1337 * Haven't found good chunk so far, let's continue
1339 if (bex
->fe_len
< gex
->fe_len
)
1342 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1343 && bex
->fe_group
== e4b
->bd_group
) {
1344 /* recheck chunk's availability - we don't know
1345 * when it was found (within this lock-unlock
1347 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1348 if (max
>= gex
->fe_len
) {
1349 ext4_mb_use_best_found(ac
, e4b
);
1356 * The routine checks whether found extent is good enough. If it is,
1357 * then the extent gets marked used and flag is set to the context
1358 * to stop scanning. Otherwise, the extent is compared with the
1359 * previous found extent and if new one is better, then it's stored
1360 * in the context. Later, the best found extent will be used, if
1361 * mballoc can't find good enough extent.
1363 * FIXME: real allocation policy is to be designed yet!
1365 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1366 struct ext4_free_extent
*ex
,
1367 struct ext4_buddy
*e4b
)
1369 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1370 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1372 BUG_ON(ex
->fe_len
<= 0);
1373 BUG_ON(ex
->fe_len
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1374 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1375 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1380 * The special case - take what you catch first
1382 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1384 ext4_mb_use_best_found(ac
, e4b
);
1389 * Let's check whether the chuck is good enough
1391 if (ex
->fe_len
== gex
->fe_len
) {
1393 ext4_mb_use_best_found(ac
, e4b
);
1398 * If this is first found extent, just store it in the context
1400 if (bex
->fe_len
== 0) {
1406 * If new found extent is better, store it in the context
1408 if (bex
->fe_len
< gex
->fe_len
) {
1409 /* if the request isn't satisfied, any found extent
1410 * larger than previous best one is better */
1411 if (ex
->fe_len
> bex
->fe_len
)
1413 } else if (ex
->fe_len
> gex
->fe_len
) {
1414 /* if the request is satisfied, then we try to find
1415 * an extent that still satisfy the request, but is
1416 * smaller than previous one */
1417 if (ex
->fe_len
< bex
->fe_len
)
1421 ext4_mb_check_limits(ac
, e4b
, 0);
1424 static int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1425 struct ext4_buddy
*e4b
)
1427 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1428 ext4_group_t group
= ex
.fe_group
;
1432 BUG_ON(ex
.fe_len
<= 0);
1433 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1437 ext4_lock_group(ac
->ac_sb
, group
);
1438 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1442 ext4_mb_use_best_found(ac
, e4b
);
1445 ext4_unlock_group(ac
->ac_sb
, group
);
1446 ext4_mb_release_desc(e4b
);
1451 static int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1452 struct ext4_buddy
*e4b
)
1454 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1457 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1458 struct ext4_super_block
*es
= sbi
->s_es
;
1459 struct ext4_free_extent ex
;
1461 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1464 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1468 ext4_lock_group(ac
->ac_sb
, group
);
1469 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1470 ac
->ac_g_ex
.fe_len
, &ex
);
1472 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1475 start
= (e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
)) +
1476 ex
.fe_start
+ le32_to_cpu(es
->s_first_data_block
);
1477 /* use do_div to get remainder (would be 64-bit modulo) */
1478 if (do_div(start
, sbi
->s_stripe
) == 0) {
1481 ext4_mb_use_best_found(ac
, e4b
);
1483 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1484 BUG_ON(ex
.fe_len
<= 0);
1485 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1486 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1489 ext4_mb_use_best_found(ac
, e4b
);
1490 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1491 /* Sometimes, caller may want to merge even small
1492 * number of blocks to an existing extent */
1493 BUG_ON(ex
.fe_len
<= 0);
1494 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1495 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1498 ext4_mb_use_best_found(ac
, e4b
);
1500 ext4_unlock_group(ac
->ac_sb
, group
);
1501 ext4_mb_release_desc(e4b
);
1507 * The routine scans buddy structures (not bitmap!) from given order
1508 * to max order and tries to find big enough chunk to satisfy the req
1510 static void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1511 struct ext4_buddy
*e4b
)
1513 struct super_block
*sb
= ac
->ac_sb
;
1514 struct ext4_group_info
*grp
= e4b
->bd_info
;
1520 BUG_ON(ac
->ac_2order
<= 0);
1521 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1522 if (grp
->bb_counters
[i
] == 0)
1525 buddy
= mb_find_buddy(e4b
, i
, &max
);
1526 BUG_ON(buddy
== NULL
);
1528 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1533 ac
->ac_b_ex
.fe_len
= 1 << i
;
1534 ac
->ac_b_ex
.fe_start
= k
<< i
;
1535 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1537 ext4_mb_use_best_found(ac
, e4b
);
1539 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1541 if (EXT4_SB(sb
)->s_mb_stats
)
1542 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1549 * The routine scans the group and measures all found extents.
1550 * In order to optimize scanning, caller must pass number of
1551 * free blocks in the group, so the routine can know upper limit.
1553 static void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1554 struct ext4_buddy
*e4b
)
1556 struct super_block
*sb
= ac
->ac_sb
;
1557 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1558 struct ext4_free_extent ex
;
1562 free
= e4b
->bd_info
->bb_free
;
1565 i
= e4b
->bd_info
->bb_first_free
;
1567 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1568 i
= mb_find_next_zero_bit(bitmap
,
1569 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1570 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1572 * IF we have corrupt bitmap, we won't find any
1573 * free blocks even though group info says we
1574 * we have free blocks
1576 ext4_error(sb
, __func__
, "%d free blocks as per "
1577 "group info. But bitmap says 0\n",
1582 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1583 BUG_ON(ex
.fe_len
<= 0);
1584 if (free
< ex
.fe_len
) {
1585 ext4_error(sb
, __func__
, "%d free blocks as per "
1586 "group info. But got %d blocks\n",
1589 * The number of free blocks differs. This mostly
1590 * indicate that the bitmap is corrupt. So exit
1591 * without claiming the space.
1596 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1602 ext4_mb_check_limits(ac
, e4b
, 1);
1606 * This is a special case for storages like raid5
1607 * we try to find stripe-aligned chunks for stripe-size requests
1608 * XXX should do so at least for multiples of stripe size as well
1610 static void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1611 struct ext4_buddy
*e4b
)
1613 struct super_block
*sb
= ac
->ac_sb
;
1614 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1615 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1616 struct ext4_free_extent ex
;
1617 ext4_fsblk_t first_group_block
;
1622 BUG_ON(sbi
->s_stripe
== 0);
1624 /* find first stripe-aligned block in group */
1625 first_group_block
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
)
1626 + le32_to_cpu(sbi
->s_es
->s_first_data_block
);
1627 a
= first_group_block
+ sbi
->s_stripe
- 1;
1628 do_div(a
, sbi
->s_stripe
);
1629 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1631 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1632 if (!mb_test_bit(i
, bitmap
)) {
1633 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1634 if (max
>= sbi
->s_stripe
) {
1637 ext4_mb_use_best_found(ac
, e4b
);
1645 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1646 ext4_group_t group
, int cr
)
1648 unsigned free
, fragments
;
1650 struct ext4_group_desc
*desc
;
1651 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1653 BUG_ON(cr
< 0 || cr
>= 4);
1654 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp
));
1656 free
= grp
->bb_free
;
1657 fragments
= grp
->bb_fragments
;
1665 BUG_ON(ac
->ac_2order
== 0);
1666 /* If this group is uninitialized, skip it initially */
1667 desc
= ext4_get_group_desc(ac
->ac_sb
, group
, NULL
);
1668 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))
1671 bits
= ac
->ac_sb
->s_blocksize_bits
+ 1;
1672 for (i
= ac
->ac_2order
; i
<= bits
; i
++)
1673 if (grp
->bb_counters
[i
] > 0)
1677 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1681 if (free
>= ac
->ac_g_ex
.fe_len
)
1693 static noinline_for_stack
int
1694 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
1701 struct ext4_sb_info
*sbi
;
1702 struct super_block
*sb
;
1703 struct ext4_buddy e4b
;
1708 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1710 /* first, try the goal */
1711 err
= ext4_mb_find_by_goal(ac
, &e4b
);
1712 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
1715 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
1719 * ac->ac2_order is set only if the fe_len is a power of 2
1720 * if ac2_order is set we also set criteria to 0 so that we
1721 * try exact allocation using buddy.
1723 i
= fls(ac
->ac_g_ex
.fe_len
);
1726 * We search using buddy data only if the order of the request
1727 * is greater than equal to the sbi_s_mb_order2_reqs
1728 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1730 if (i
>= sbi
->s_mb_order2_reqs
) {
1732 * This should tell if fe_len is exactly power of 2
1734 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
1735 ac
->ac_2order
= i
- 1;
1738 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
1739 /* if stream allocation is enabled, use global goal */
1740 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
1741 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
1745 if (size
< sbi
->s_mb_stream_request
&&
1746 (ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1747 /* TBD: may be hot point */
1748 spin_lock(&sbi
->s_md_lock
);
1749 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
1750 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
1751 spin_unlock(&sbi
->s_md_lock
);
1753 /* Let's just scan groups to find more-less suitable blocks */
1754 cr
= ac
->ac_2order
? 0 : 1;
1756 * cr == 0 try to get exact allocation,
1757 * cr == 3 try to get anything
1760 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
1761 ac
->ac_criteria
= cr
;
1763 * searching for the right group start
1764 * from the goal value specified
1766 group
= ac
->ac_g_ex
.fe_group
;
1768 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; group
++, i
++) {
1769 struct ext4_group_info
*grp
;
1770 struct ext4_group_desc
*desc
;
1772 if (group
== EXT4_SB(sb
)->s_groups_count
)
1775 /* quick check to skip empty groups */
1776 grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1777 if (grp
->bb_free
== 0)
1781 * if the group is already init we check whether it is
1782 * a good group and if not we don't load the buddy
1784 if (EXT4_MB_GRP_NEED_INIT(grp
)) {
1786 * we need full data about the group
1787 * to make a good selection
1789 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
1792 ext4_mb_release_desc(&e4b
);
1796 * If the particular group doesn't satisfy our
1797 * criteria we continue with the next group
1799 if (!ext4_mb_good_group(ac
, group
, cr
))
1802 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
1806 ext4_lock_group(sb
, group
);
1807 if (!ext4_mb_good_group(ac
, group
, cr
)) {
1808 /* someone did allocation from this group */
1809 ext4_unlock_group(sb
, group
);
1810 ext4_mb_release_desc(&e4b
);
1814 ac
->ac_groups_scanned
++;
1815 desc
= ext4_get_group_desc(sb
, group
, NULL
);
1816 if (cr
== 0 || (desc
->bg_flags
&
1817 cpu_to_le16(EXT4_BG_BLOCK_UNINIT
) &&
1818 ac
->ac_2order
!= 0))
1819 ext4_mb_simple_scan_group(ac
, &e4b
);
1821 ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
)
1822 ext4_mb_scan_aligned(ac
, &e4b
);
1824 ext4_mb_complex_scan_group(ac
, &e4b
);
1826 ext4_unlock_group(sb
, group
);
1827 ext4_mb_release_desc(&e4b
);
1829 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
1834 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
1835 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1837 * We've been searching too long. Let's try to allocate
1838 * the best chunk we've found so far
1841 ext4_mb_try_best_found(ac
, &e4b
);
1842 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
1844 * Someone more lucky has already allocated it.
1845 * The only thing we can do is just take first
1847 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
1849 ac
->ac_b_ex
.fe_group
= 0;
1850 ac
->ac_b_ex
.fe_start
= 0;
1851 ac
->ac_b_ex
.fe_len
= 0;
1852 ac
->ac_status
= AC_STATUS_CONTINUE
;
1853 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
1855 atomic_inc(&sbi
->s_mb_lost_chunks
);
1863 #ifdef EXT4_MB_HISTORY
1864 struct ext4_mb_proc_session
{
1865 struct ext4_mb_history
*history
;
1866 struct super_block
*sb
;
1871 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session
*s
,
1872 struct ext4_mb_history
*hs
,
1875 if (hs
== s
->history
+ s
->max
)
1877 if (!first
&& hs
== s
->history
+ s
->start
)
1879 while (hs
->orig
.fe_len
== 0) {
1881 if (hs
== s
->history
+ s
->max
)
1883 if (hs
== s
->history
+ s
->start
)
1889 static void *ext4_mb_seq_history_start(struct seq_file
*seq
, loff_t
*pos
)
1891 struct ext4_mb_proc_session
*s
= seq
->private;
1892 struct ext4_mb_history
*hs
;
1896 return SEQ_START_TOKEN
;
1897 hs
= ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
1900 while (--l
&& (hs
= ext4_mb_history_skip_empty(s
, ++hs
, 0)) != NULL
);
1904 static void *ext4_mb_seq_history_next(struct seq_file
*seq
, void *v
,
1907 struct ext4_mb_proc_session
*s
= seq
->private;
1908 struct ext4_mb_history
*hs
= v
;
1911 if (v
== SEQ_START_TOKEN
)
1912 return ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
1914 return ext4_mb_history_skip_empty(s
, ++hs
, 0);
1917 static int ext4_mb_seq_history_show(struct seq_file
*seq
, void *v
)
1919 char buf
[25], buf2
[25], buf3
[25], *fmt
;
1920 struct ext4_mb_history
*hs
= v
;
1922 if (v
== SEQ_START_TOKEN
) {
1923 seq_printf(seq
, "%-5s %-8s %-23s %-23s %-23s %-5s "
1924 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
1925 "pid", "inode", "original", "goal", "result", "found",
1926 "grps", "cr", "flags", "merge", "tail", "broken");
1930 if (hs
->op
== EXT4_MB_HISTORY_ALLOC
) {
1931 fmt
= "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
1932 "%-5u %-5s %-5u %-6u\n";
1933 sprintf(buf2
, "%lu/%d/%u@%u", hs
->result
.fe_group
,
1934 hs
->result
.fe_start
, hs
->result
.fe_len
,
1935 hs
->result
.fe_logical
);
1936 sprintf(buf
, "%lu/%d/%u@%u", hs
->orig
.fe_group
,
1937 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
1938 hs
->orig
.fe_logical
);
1939 sprintf(buf3
, "%lu/%d/%u@%u", hs
->goal
.fe_group
,
1940 hs
->goal
.fe_start
, hs
->goal
.fe_len
,
1941 hs
->goal
.fe_logical
);
1942 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, buf3
, buf2
,
1943 hs
->found
, hs
->groups
, hs
->cr
, hs
->flags
,
1944 hs
->merged
? "M" : "", hs
->tail
,
1945 hs
->buddy
? 1 << hs
->buddy
: 0);
1946 } else if (hs
->op
== EXT4_MB_HISTORY_PREALLOC
) {
1947 fmt
= "%-5u %-8u %-23s %-23s %-23s\n";
1948 sprintf(buf2
, "%lu/%d/%u@%u", hs
->result
.fe_group
,
1949 hs
->result
.fe_start
, hs
->result
.fe_len
,
1950 hs
->result
.fe_logical
);
1951 sprintf(buf
, "%lu/%d/%u@%u", hs
->orig
.fe_group
,
1952 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
1953 hs
->orig
.fe_logical
);
1954 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, "", buf2
);
1955 } else if (hs
->op
== EXT4_MB_HISTORY_DISCARD
) {
1956 sprintf(buf2
, "%lu/%d/%u", hs
->result
.fe_group
,
1957 hs
->result
.fe_start
, hs
->result
.fe_len
);
1958 seq_printf(seq
, "%-5u %-8u %-23s discard\n",
1959 hs
->pid
, hs
->ino
, buf2
);
1960 } else if (hs
->op
== EXT4_MB_HISTORY_FREE
) {
1961 sprintf(buf2
, "%lu/%d/%u", hs
->result
.fe_group
,
1962 hs
->result
.fe_start
, hs
->result
.fe_len
);
1963 seq_printf(seq
, "%-5u %-8u %-23s free\n",
1964 hs
->pid
, hs
->ino
, buf2
);
1969 static void ext4_mb_seq_history_stop(struct seq_file
*seq
, void *v
)
1973 static struct seq_operations ext4_mb_seq_history_ops
= {
1974 .start
= ext4_mb_seq_history_start
,
1975 .next
= ext4_mb_seq_history_next
,
1976 .stop
= ext4_mb_seq_history_stop
,
1977 .show
= ext4_mb_seq_history_show
,
1980 static int ext4_mb_seq_history_open(struct inode
*inode
, struct file
*file
)
1982 struct super_block
*sb
= PDE(inode
)->data
;
1983 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1984 struct ext4_mb_proc_session
*s
;
1988 if (unlikely(sbi
->s_mb_history
== NULL
))
1990 s
= kmalloc(sizeof(*s
), GFP_KERNEL
);
1994 size
= sizeof(struct ext4_mb_history
) * sbi
->s_mb_history_max
;
1995 s
->history
= kmalloc(size
, GFP_KERNEL
);
1996 if (s
->history
== NULL
) {
2001 spin_lock(&sbi
->s_mb_history_lock
);
2002 memcpy(s
->history
, sbi
->s_mb_history
, size
);
2003 s
->max
= sbi
->s_mb_history_max
;
2004 s
->start
= sbi
->s_mb_history_cur
% s
->max
;
2005 spin_unlock(&sbi
->s_mb_history_lock
);
2007 rc
= seq_open(file
, &ext4_mb_seq_history_ops
);
2009 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2019 static int ext4_mb_seq_history_release(struct inode
*inode
, struct file
*file
)
2021 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2022 struct ext4_mb_proc_session
*s
= seq
->private;
2025 return seq_release(inode
, file
);
2028 static ssize_t
ext4_mb_seq_history_write(struct file
*file
,
2029 const char __user
*buffer
,
2030 size_t count
, loff_t
*ppos
)
2032 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2033 struct ext4_mb_proc_session
*s
= seq
->private;
2034 struct super_block
*sb
= s
->sb
;
2038 if (count
>= sizeof(str
)) {
2039 printk(KERN_ERR
"EXT4-fs: %s string too long, max %u bytes\n",
2040 "mb_history", (int)sizeof(str
));
2044 if (copy_from_user(str
, buffer
, count
))
2047 value
= simple_strtol(str
, NULL
, 0);
2050 EXT4_SB(sb
)->s_mb_history_filter
= value
;
2055 static struct file_operations ext4_mb_seq_history_fops
= {
2056 .owner
= THIS_MODULE
,
2057 .open
= ext4_mb_seq_history_open
,
2059 .write
= ext4_mb_seq_history_write
,
2060 .llseek
= seq_lseek
,
2061 .release
= ext4_mb_seq_history_release
,
2064 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2066 struct super_block
*sb
= seq
->private;
2067 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2070 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2074 return (void *) group
;
2077 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2079 struct super_block
*sb
= seq
->private;
2080 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2084 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2087 return (void *) group
;;
2090 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2092 struct super_block
*sb
= seq
->private;
2093 long group
= (long) v
;
2096 struct ext4_buddy e4b
;
2098 struct ext4_group_info info
;
2099 unsigned short counters
[16];
2104 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2105 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2106 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2107 "group", "free", "frags", "first",
2108 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2109 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2111 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2112 sizeof(struct ext4_group_info
);
2113 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2115 seq_printf(seq
, "#%-5lu: I/O error\n", group
);
2118 ext4_lock_group(sb
, group
);
2119 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2120 ext4_unlock_group(sb
, group
);
2121 ext4_mb_release_desc(&e4b
);
2123 seq_printf(seq
, "#%-5lu: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2124 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2125 for (i
= 0; i
<= 13; i
++)
2126 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2127 sg
.info
.bb_counters
[i
] : 0);
2128 seq_printf(seq
, " ]\n");
2133 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2137 static struct seq_operations ext4_mb_seq_groups_ops
= {
2138 .start
= ext4_mb_seq_groups_start
,
2139 .next
= ext4_mb_seq_groups_next
,
2140 .stop
= ext4_mb_seq_groups_stop
,
2141 .show
= ext4_mb_seq_groups_show
,
2144 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2146 struct super_block
*sb
= PDE(inode
)->data
;
2149 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2151 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2158 static struct file_operations ext4_mb_seq_groups_fops
= {
2159 .owner
= THIS_MODULE
,
2160 .open
= ext4_mb_seq_groups_open
,
2162 .llseek
= seq_lseek
,
2163 .release
= seq_release
,
2166 static void ext4_mb_history_release(struct super_block
*sb
)
2168 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2170 if (sbi
->s_proc
!= NULL
) {
2171 remove_proc_entry("mb_groups", sbi
->s_proc
);
2172 remove_proc_entry("mb_history", sbi
->s_proc
);
2174 kfree(sbi
->s_mb_history
);
2177 static void ext4_mb_history_init(struct super_block
*sb
)
2179 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2182 if (sbi
->s_proc
!= NULL
) {
2183 proc_create_data("mb_history", S_IRUGO
, sbi
->s_proc
,
2184 &ext4_mb_seq_history_fops
, sb
);
2185 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2186 &ext4_mb_seq_groups_fops
, sb
);
2189 sbi
->s_mb_history_max
= 1000;
2190 sbi
->s_mb_history_cur
= 0;
2191 spin_lock_init(&sbi
->s_mb_history_lock
);
2192 i
= sbi
->s_mb_history_max
* sizeof(struct ext4_mb_history
);
2193 sbi
->s_mb_history
= kzalloc(i
, GFP_KERNEL
);
2194 /* if we can't allocate history, then we simple won't use it */
2197 static noinline_for_stack
void
2198 ext4_mb_store_history(struct ext4_allocation_context
*ac
)
2200 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2201 struct ext4_mb_history h
;
2203 if (unlikely(sbi
->s_mb_history
== NULL
))
2206 if (!(ac
->ac_op
& sbi
->s_mb_history_filter
))
2210 h
.pid
= current
->pid
;
2211 h
.ino
= ac
->ac_inode
? ac
->ac_inode
->i_ino
: 0;
2212 h
.orig
= ac
->ac_o_ex
;
2213 h
.result
= ac
->ac_b_ex
;
2214 h
.flags
= ac
->ac_flags
;
2215 h
.found
= ac
->ac_found
;
2216 h
.groups
= ac
->ac_groups_scanned
;
2217 h
.cr
= ac
->ac_criteria
;
2218 h
.tail
= ac
->ac_tail
;
2219 h
.buddy
= ac
->ac_buddy
;
2221 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
) {
2222 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
2223 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
2225 h
.goal
= ac
->ac_g_ex
;
2226 h
.result
= ac
->ac_f_ex
;
2229 spin_lock(&sbi
->s_mb_history_lock
);
2230 memcpy(sbi
->s_mb_history
+ sbi
->s_mb_history_cur
, &h
, sizeof(h
));
2231 if (++sbi
->s_mb_history_cur
>= sbi
->s_mb_history_max
)
2232 sbi
->s_mb_history_cur
= 0;
2233 spin_unlock(&sbi
->s_mb_history_lock
);
2237 #define ext4_mb_history_release(sb)
2238 #define ext4_mb_history_init(sb)
2242 /* Create and initialize ext4_group_info data for the given group. */
2243 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2244 struct ext4_group_desc
*desc
)
2248 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2249 struct ext4_group_info
**meta_group_info
;
2252 * First check if this group is the first of a reserved block.
2253 * If it's true, we have to allocate a new table of pointers
2254 * to ext4_group_info structures
2256 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2257 metalen
= sizeof(*meta_group_info
) <<
2258 EXT4_DESC_PER_BLOCK_BITS(sb
);
2259 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2260 if (meta_group_info
== NULL
) {
2261 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2263 goto exit_meta_group_info
;
2265 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2270 * calculate needed size. if change bb_counters size,
2271 * don't forget about ext4_mb_generate_buddy()
2273 len
= offsetof(typeof(**meta_group_info
),
2274 bb_counters
[sb
->s_blocksize_bits
+ 2]);
2277 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2278 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2280 meta_group_info
[i
] = kzalloc(len
, GFP_KERNEL
);
2281 if (meta_group_info
[i
] == NULL
) {
2282 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2283 goto exit_group_info
;
2285 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2286 &(meta_group_info
[i
]->bb_state
));
2289 * initialize bb_free to be able to skip
2290 * empty groups without initialization
2292 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2293 meta_group_info
[i
]->bb_free
=
2294 ext4_free_blocks_after_init(sb
, group
, desc
);
2296 meta_group_info
[i
]->bb_free
=
2297 le16_to_cpu(desc
->bg_free_blocks_count
);
2300 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2304 struct buffer_head
*bh
;
2305 meta_group_info
[i
]->bb_bitmap
=
2306 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2307 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2308 bh
= ext4_read_block_bitmap(sb
, group
);
2310 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2319 /* If a meta_group_info table has been allocated, release it now */
2320 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0)
2321 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2322 exit_meta_group_info
:
2324 } /* ext4_mb_add_groupinfo */
2327 * Add a group to the existing groups.
2328 * This function is used for online resize
2330 int ext4_mb_add_more_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2331 struct ext4_group_desc
*desc
)
2333 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2334 struct inode
*inode
= sbi
->s_buddy_cache
;
2335 int blocks_per_page
;
2341 /* Add group based on group descriptor*/
2342 err
= ext4_mb_add_groupinfo(sb
, group
, desc
);
2347 * Cache pages containing dynamic mb_alloc datas (buddy and bitmap
2348 * datas) are set not up to date so that they will be re-initilaized
2349 * during the next call to ext4_mb_load_buddy
2352 /* Set buddy page as not up to date */
2353 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
2355 pnum
= block
/ blocks_per_page
;
2356 page
= find_get_page(inode
->i_mapping
, pnum
);
2358 ClearPageUptodate(page
);
2359 page_cache_release(page
);
2362 /* Set bitmap page as not up to date */
2364 pnum
= block
/ blocks_per_page
;
2365 page
= find_get_page(inode
->i_mapping
, pnum
);
2367 ClearPageUptodate(page
);
2368 page_cache_release(page
);
2375 * Update an existing group.
2376 * This function is used for online resize
2378 void ext4_mb_update_group_info(struct ext4_group_info
*grp
, ext4_grpblk_t add
)
2380 grp
->bb_free
+= add
;
2383 static int ext4_mb_init_backend(struct super_block
*sb
)
2387 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2388 struct ext4_super_block
*es
= sbi
->s_es
;
2389 int num_meta_group_infos
;
2390 int num_meta_group_infos_max
;
2392 struct ext4_group_info
**meta_group_info
;
2393 struct ext4_group_desc
*desc
;
2395 /* This is the number of blocks used by GDT */
2396 num_meta_group_infos
= (sbi
->s_groups_count
+ EXT4_DESC_PER_BLOCK(sb
) -
2397 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2400 * This is the total number of blocks used by GDT including
2401 * the number of reserved blocks for GDT.
2402 * The s_group_info array is allocated with this value
2403 * to allow a clean online resize without a complex
2404 * manipulation of pointer.
2405 * The drawback is the unused memory when no resize
2406 * occurs but it's very low in terms of pages
2407 * (see comments below)
2408 * Need to handle this properly when META_BG resizing is allowed
2410 num_meta_group_infos_max
= num_meta_group_infos
+
2411 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2414 * array_size is the size of s_group_info array. We round it
2415 * to the next power of two because this approximation is done
2416 * internally by kmalloc so we can have some more memory
2417 * for free here (e.g. may be used for META_BG resize).
2420 while (array_size
< sizeof(*sbi
->s_group_info
) *
2421 num_meta_group_infos_max
)
2422 array_size
= array_size
<< 1;
2423 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2424 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2425 * So a two level scheme suffices for now. */
2426 sbi
->s_group_info
= kmalloc(array_size
, GFP_KERNEL
);
2427 if (sbi
->s_group_info
== NULL
) {
2428 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2431 sbi
->s_buddy_cache
= new_inode(sb
);
2432 if (sbi
->s_buddy_cache
== NULL
) {
2433 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2436 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2438 metalen
= sizeof(*meta_group_info
) << EXT4_DESC_PER_BLOCK_BITS(sb
);
2439 for (i
= 0; i
< num_meta_group_infos
; i
++) {
2440 if ((i
+ 1) == num_meta_group_infos
)
2441 metalen
= sizeof(*meta_group_info
) *
2442 (sbi
->s_groups_count
-
2443 (i
<< EXT4_DESC_PER_BLOCK_BITS(sb
)));
2444 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2445 if (meta_group_info
== NULL
) {
2446 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2450 sbi
->s_group_info
[i
] = meta_group_info
;
2453 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2454 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2457 "EXT4-fs: can't read descriptor %lu\n", i
);
2460 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2468 kfree(ext4_get_group_info(sb
, i
));
2469 i
= num_meta_group_infos
;
2472 kfree(sbi
->s_group_info
[i
]);
2473 iput(sbi
->s_buddy_cache
);
2475 kfree(sbi
->s_group_info
);
2479 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2481 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2487 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned short);
2489 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2490 if (sbi
->s_mb_offsets
== NULL
) {
2493 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2494 if (sbi
->s_mb_maxs
== NULL
) {
2495 kfree(sbi
->s_mb_maxs
);
2499 /* order 0 is regular bitmap */
2500 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2501 sbi
->s_mb_offsets
[0] = 0;
2505 max
= sb
->s_blocksize
<< 2;
2507 sbi
->s_mb_offsets
[i
] = offset
;
2508 sbi
->s_mb_maxs
[i
] = max
;
2509 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2512 } while (i
<= sb
->s_blocksize_bits
+ 1);
2514 /* init file for buddy data */
2515 ret
= ext4_mb_init_backend(sb
);
2517 kfree(sbi
->s_mb_offsets
);
2518 kfree(sbi
->s_mb_maxs
);
2522 spin_lock_init(&sbi
->s_md_lock
);
2523 INIT_LIST_HEAD(&sbi
->s_active_transaction
);
2524 INIT_LIST_HEAD(&sbi
->s_closed_transaction
);
2525 INIT_LIST_HEAD(&sbi
->s_committed_transaction
);
2526 spin_lock_init(&sbi
->s_bal_lock
);
2528 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2529 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2530 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2531 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2532 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2533 sbi
->s_mb_history_filter
= EXT4_MB_HISTORY_DEFAULT
;
2534 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2536 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2537 if (sbi
->s_locality_groups
== NULL
) {
2538 kfree(sbi
->s_mb_offsets
);
2539 kfree(sbi
->s_mb_maxs
);
2542 for_each_possible_cpu(i
) {
2543 struct ext4_locality_group
*lg
;
2544 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2545 mutex_init(&lg
->lg_mutex
);
2546 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2547 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2548 spin_lock_init(&lg
->lg_prealloc_lock
);
2551 ext4_mb_init_per_dev_proc(sb
);
2552 ext4_mb_history_init(sb
);
2554 printk(KERN_INFO
"EXT4-fs: mballoc enabled\n");
2558 /* need to called with ext4 group lock (ext4_lock_group) */
2559 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2561 struct ext4_prealloc_space
*pa
;
2562 struct list_head
*cur
, *tmp
;
2565 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2566 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2567 list_del(&pa
->pa_group_list
);
2572 mb_debug("mballoc: %u PAs left\n", count
);
2576 int ext4_mb_release(struct super_block
*sb
)
2579 int num_meta_group_infos
;
2580 struct ext4_group_info
*grinfo
;
2581 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2583 /* release freed, non-committed blocks */
2584 spin_lock(&sbi
->s_md_lock
);
2585 list_splice_init(&sbi
->s_closed_transaction
,
2586 &sbi
->s_committed_transaction
);
2587 list_splice_init(&sbi
->s_active_transaction
,
2588 &sbi
->s_committed_transaction
);
2589 spin_unlock(&sbi
->s_md_lock
);
2590 ext4_mb_free_committed_blocks(sb
);
2592 if (sbi
->s_group_info
) {
2593 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2594 grinfo
= ext4_get_group_info(sb
, i
);
2596 kfree(grinfo
->bb_bitmap
);
2598 ext4_lock_group(sb
, i
);
2599 ext4_mb_cleanup_pa(grinfo
);
2600 ext4_unlock_group(sb
, i
);
2603 num_meta_group_infos
= (sbi
->s_groups_count
+
2604 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2605 EXT4_DESC_PER_BLOCK_BITS(sb
);
2606 for (i
= 0; i
< num_meta_group_infos
; i
++)
2607 kfree(sbi
->s_group_info
[i
]);
2608 kfree(sbi
->s_group_info
);
2610 kfree(sbi
->s_mb_offsets
);
2611 kfree(sbi
->s_mb_maxs
);
2612 if (sbi
->s_buddy_cache
)
2613 iput(sbi
->s_buddy_cache
);
2614 if (sbi
->s_mb_stats
) {
2616 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2617 atomic_read(&sbi
->s_bal_allocated
),
2618 atomic_read(&sbi
->s_bal_reqs
),
2619 atomic_read(&sbi
->s_bal_success
));
2621 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2622 "%u 2^N hits, %u breaks, %u lost\n",
2623 atomic_read(&sbi
->s_bal_ex_scanned
),
2624 atomic_read(&sbi
->s_bal_goals
),
2625 atomic_read(&sbi
->s_bal_2orders
),
2626 atomic_read(&sbi
->s_bal_breaks
),
2627 atomic_read(&sbi
->s_mb_lost_chunks
));
2629 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2630 sbi
->s_mb_buddies_generated
++,
2631 sbi
->s_mb_generation_time
);
2633 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2634 atomic_read(&sbi
->s_mb_preallocated
),
2635 atomic_read(&sbi
->s_mb_discarded
));
2638 free_percpu(sbi
->s_locality_groups
);
2639 ext4_mb_history_release(sb
);
2640 ext4_mb_destroy_per_dev_proc(sb
);
2645 static noinline_for_stack
void
2646 ext4_mb_free_committed_blocks(struct super_block
*sb
)
2648 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2653 struct ext4_free_metadata
*md
;
2654 struct ext4_buddy e4b
;
2656 if (list_empty(&sbi
->s_committed_transaction
))
2659 /* there is committed blocks to be freed yet */
2661 /* get next array of blocks */
2663 spin_lock(&sbi
->s_md_lock
);
2664 if (!list_empty(&sbi
->s_committed_transaction
)) {
2665 md
= list_entry(sbi
->s_committed_transaction
.next
,
2666 struct ext4_free_metadata
, list
);
2667 list_del(&md
->list
);
2669 spin_unlock(&sbi
->s_md_lock
);
2674 mb_debug("gonna free %u blocks in group %lu (0x%p):",
2675 md
->num
, md
->group
, md
);
2677 err
= ext4_mb_load_buddy(sb
, md
->group
, &e4b
);
2678 /* we expect to find existing buddy because it's pinned */
2681 /* there are blocks to put in buddy to make them really free */
2684 ext4_lock_group(sb
, md
->group
);
2685 for (i
= 0; i
< md
->num
; i
++) {
2686 mb_debug(" %u", md
->blocks
[i
]);
2687 mb_free_blocks(NULL
, &e4b
, md
->blocks
[i
], 1);
2690 ext4_unlock_group(sb
, md
->group
);
2692 /* balance refcounts from ext4_mb_free_metadata() */
2693 page_cache_release(e4b
.bd_buddy_page
);
2694 page_cache_release(e4b
.bd_bitmap_page
);
2697 ext4_mb_release_desc(&e4b
);
2701 mb_debug("freed %u blocks in %u structures\n", count
, count2
);
2704 #define EXT4_MB_STATS_NAME "stats"
2705 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2706 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2707 #define EXT4_MB_ORDER2_REQ "order2_req"
2708 #define EXT4_MB_STREAM_REQ "stream_req"
2709 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2711 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
)
2713 mode_t mode
= S_IFREG
| S_IRUGO
| S_IWUSR
;
2714 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2715 struct proc_dir_entry
*proc
;
2717 if (sbi
->s_proc
== NULL
)
2720 EXT4_PROC_HANDLER(EXT4_MB_STATS_NAME
, mb_stats
);
2721 EXT4_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME
, mb_max_to_scan
);
2722 EXT4_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME
, mb_min_to_scan
);
2723 EXT4_PROC_HANDLER(EXT4_MB_ORDER2_REQ
, mb_order2_reqs
);
2724 EXT4_PROC_HANDLER(EXT4_MB_STREAM_REQ
, mb_stream_request
);
2725 EXT4_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC
, mb_group_prealloc
);
2729 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2730 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2731 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2732 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2733 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2734 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2738 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
)
2740 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2742 if (sbi
->s_proc
== NULL
)
2745 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2746 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2747 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2748 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2749 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2750 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2755 int __init
init_ext4_mballoc(void)
2757 ext4_pspace_cachep
=
2758 kmem_cache_create("ext4_prealloc_space",
2759 sizeof(struct ext4_prealloc_space
),
2760 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2761 if (ext4_pspace_cachep
== NULL
)
2765 kmem_cache_create("ext4_alloc_context",
2766 sizeof(struct ext4_allocation_context
),
2767 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2768 if (ext4_ac_cachep
== NULL
) {
2769 kmem_cache_destroy(ext4_pspace_cachep
);
2775 void exit_ext4_mballoc(void)
2777 /* XXX: synchronize_rcu(); */
2778 kmem_cache_destroy(ext4_pspace_cachep
);
2779 kmem_cache_destroy(ext4_ac_cachep
);
2784 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2785 * Returns 0 if success or error code
2787 static noinline_for_stack
int
2788 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2789 handle_t
*handle
, unsigned long reserv_blks
)
2791 struct buffer_head
*bitmap_bh
= NULL
;
2792 struct ext4_super_block
*es
;
2793 struct ext4_group_desc
*gdp
;
2794 struct buffer_head
*gdp_bh
;
2795 struct ext4_sb_info
*sbi
;
2796 struct super_block
*sb
;
2800 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2801 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2809 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2813 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2818 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2822 ext4_debug("using block group %lu(%d)\n", ac
->ac_b_ex
.fe_group
,
2823 gdp
->bg_free_blocks_count
);
2825 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2829 block
= ac
->ac_b_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
2830 + ac
->ac_b_ex
.fe_start
2831 + le32_to_cpu(es
->s_first_data_block
);
2833 len
= ac
->ac_b_ex
.fe_len
;
2834 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, len
) ||
2835 in_range(ext4_inode_bitmap(sb
, gdp
), block
, len
) ||
2836 in_range(block
, ext4_inode_table(sb
, gdp
),
2837 EXT4_SB(sb
)->s_itb_per_group
) ||
2838 in_range(block
+ len
- 1, ext4_inode_table(sb
, gdp
),
2839 EXT4_SB(sb
)->s_itb_per_group
)) {
2840 ext4_error(sb
, __func__
,
2841 "Allocating block in system zone - block = %llu",
2843 /* File system mounted not to panic on error
2844 * Fix the bitmap and repeat the block allocation
2845 * We leak some of the blocks here.
2847 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
),
2848 bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2849 ac
->ac_b_ex
.fe_len
);
2850 err
= ext4_journal_dirty_metadata(handle
, bitmap_bh
);
2855 #ifdef AGGRESSIVE_CHECK
2858 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2859 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2860 bitmap_bh
->b_data
));
2864 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
), bitmap_bh
->b_data
,
2865 ac
->ac_b_ex
.fe_start
, ac
->ac_b_ex
.fe_len
);
2867 spin_lock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
2868 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2869 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2870 gdp
->bg_free_blocks_count
=
2871 cpu_to_le16(ext4_free_blocks_after_init(sb
,
2872 ac
->ac_b_ex
.fe_group
,
2875 le16_add_cpu(&gdp
->bg_free_blocks_count
, -ac
->ac_b_ex
.fe_len
);
2876 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
2877 spin_unlock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
2878 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
2880 * Now reduce the dirty block count also. Should not go negative
2882 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2883 /* release all the reserved blocks if non delalloc */
2884 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
, reserv_blks
);
2886 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
,
2887 ac
->ac_b_ex
.fe_len
);
2889 if (sbi
->s_log_groups_per_flex
) {
2890 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2891 ac
->ac_b_ex
.fe_group
);
2892 spin_lock(sb_bgl_lock(sbi
, flex_group
));
2893 sbi
->s_flex_groups
[flex_group
].free_blocks
-= ac
->ac_b_ex
.fe_len
;
2894 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
2897 err
= ext4_journal_dirty_metadata(handle
, bitmap_bh
);
2900 err
= ext4_journal_dirty_metadata(handle
, gdp_bh
);
2909 * here we normalize request for locality group
2910 * Group request are normalized to s_strip size if we set the same via mount
2911 * option. If not we set it to s_mb_group_prealloc which can be configured via
2912 * /proc/fs/ext4/<partition>/group_prealloc
2914 * XXX: should we try to preallocate more than the group has now?
2916 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2918 struct super_block
*sb
= ac
->ac_sb
;
2919 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2922 if (EXT4_SB(sb
)->s_stripe
)
2923 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
2925 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2926 mb_debug("#%u: goal %u blocks for locality group\n",
2927 current
->pid
, ac
->ac_g_ex
.fe_len
);
2931 * Normalization means making request better in terms of
2932 * size and alignment
2934 static noinline_for_stack
void
2935 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
2936 struct ext4_allocation_request
*ar
)
2940 loff_t size
, orig_size
, start_off
;
2941 ext4_lblk_t start
, orig_start
;
2942 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
2943 struct ext4_prealloc_space
*pa
;
2945 /* do normalize only data requests, metadata requests
2946 do not need preallocation */
2947 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
2950 /* sometime caller may want exact blocks */
2951 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2954 /* caller may indicate that preallocation isn't
2955 * required (it's a tail, for example) */
2956 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
2959 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
2960 ext4_mb_normalize_group_request(ac
);
2964 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
2966 /* first, let's learn actual file size
2967 * given current request is allocated */
2968 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
2969 size
= size
<< bsbits
;
2970 if (size
< i_size_read(ac
->ac_inode
))
2971 size
= i_size_read(ac
->ac_inode
);
2973 /* max size of free chunks */
2976 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2977 (req <= (size) || max <= (chunk_size))
2979 /* first, try to predict filesize */
2980 /* XXX: should this table be tunable? */
2982 if (size
<= 16 * 1024) {
2984 } else if (size
<= 32 * 1024) {
2986 } else if (size
<= 64 * 1024) {
2988 } else if (size
<= 128 * 1024) {
2990 } else if (size
<= 256 * 1024) {
2992 } else if (size
<= 512 * 1024) {
2994 } else if (size
<= 1024 * 1024) {
2996 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
2997 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2998 (21 - bsbits
)) << 21;
2999 size
= 2 * 1024 * 1024;
3000 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3001 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3002 (22 - bsbits
)) << 22;
3003 size
= 4 * 1024 * 1024;
3004 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3005 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3006 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3007 (23 - bsbits
)) << 23;
3008 size
= 8 * 1024 * 1024;
3010 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3011 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3013 orig_size
= size
= size
>> bsbits
;
3014 orig_start
= start
= start_off
>> bsbits
;
3016 /* don't cover already allocated blocks in selected range */
3017 if (ar
->pleft
&& start
<= ar
->lleft
) {
3018 size
-= ar
->lleft
+ 1 - start
;
3019 start
= ar
->lleft
+ 1;
3021 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3022 size
-= start
+ size
- ar
->lright
;
3026 /* check we don't cross already preallocated blocks */
3028 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3029 unsigned long pa_end
;
3033 spin_lock(&pa
->pa_lock
);
3034 if (pa
->pa_deleted
) {
3035 spin_unlock(&pa
->pa_lock
);
3039 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3041 /* PA must not overlap original request */
3042 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3043 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3045 /* skip PA normalized request doesn't overlap with */
3046 if (pa
->pa_lstart
>= end
) {
3047 spin_unlock(&pa
->pa_lock
);
3050 if (pa_end
<= start
) {
3051 spin_unlock(&pa
->pa_lock
);
3054 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3056 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3057 BUG_ON(pa_end
< start
);
3061 if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3062 BUG_ON(pa
->pa_lstart
> end
);
3063 end
= pa
->pa_lstart
;
3065 spin_unlock(&pa
->pa_lock
);
3070 /* XXX: extra loop to check we really don't overlap preallocations */
3072 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3073 unsigned long pa_end
;
3074 spin_lock(&pa
->pa_lock
);
3075 if (pa
->pa_deleted
== 0) {
3076 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3077 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3079 spin_unlock(&pa
->pa_lock
);
3083 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3084 start
> ac
->ac_o_ex
.fe_logical
) {
3085 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3086 (unsigned long) start
, (unsigned long) size
,
3087 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3089 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3090 start
> ac
->ac_o_ex
.fe_logical
);
3091 BUG_ON(size
<= 0 || size
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3093 /* now prepare goal request */
3095 /* XXX: is it better to align blocks WRT to logical
3096 * placement or satisfy big request as is */
3097 ac
->ac_g_ex
.fe_logical
= start
;
3098 ac
->ac_g_ex
.fe_len
= size
;
3100 /* define goal start in order to merge */
3101 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3102 /* merge to the right */
3103 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3104 &ac
->ac_f_ex
.fe_group
,
3105 &ac
->ac_f_ex
.fe_start
);
3106 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3108 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3109 /* merge to the left */
3110 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3111 &ac
->ac_f_ex
.fe_group
,
3112 &ac
->ac_f_ex
.fe_start
);
3113 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3116 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3117 (unsigned) orig_size
, (unsigned) start
);
3120 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3122 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3124 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3125 atomic_inc(&sbi
->s_bal_reqs
);
3126 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3127 if (ac
->ac_o_ex
.fe_len
>= ac
->ac_g_ex
.fe_len
)
3128 atomic_inc(&sbi
->s_bal_success
);
3129 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3130 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3131 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3132 atomic_inc(&sbi
->s_bal_goals
);
3133 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3134 atomic_inc(&sbi
->s_bal_breaks
);
3137 ext4_mb_store_history(ac
);
3141 * use blocks preallocated to inode
3143 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3144 struct ext4_prealloc_space
*pa
)
3150 /* found preallocated blocks, use them */
3151 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3152 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3154 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3155 &ac
->ac_b_ex
.fe_start
);
3156 ac
->ac_b_ex
.fe_len
= len
;
3157 ac
->ac_status
= AC_STATUS_FOUND
;
3160 BUG_ON(start
< pa
->pa_pstart
);
3161 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3162 BUG_ON(pa
->pa_free
< len
);
3165 mb_debug("use %llu/%u from inode pa %p\n", start
, len
, pa
);
3169 * use blocks preallocated to locality group
3171 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3172 struct ext4_prealloc_space
*pa
)
3174 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3176 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3177 &ac
->ac_b_ex
.fe_group
,
3178 &ac
->ac_b_ex
.fe_start
);
3179 ac
->ac_b_ex
.fe_len
= len
;
3180 ac
->ac_status
= AC_STATUS_FOUND
;
3183 /* we don't correct pa_pstart or pa_plen here to avoid
3184 * possible race when the group is being loaded concurrently
3185 * instead we correct pa later, after blocks are marked
3186 * in on-disk bitmap -- see ext4_mb_release_context()
3187 * Other CPUs are prevented from allocating from this pa by lg_mutex
3189 mb_debug("use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3193 * Return the prealloc space that have minimal distance
3194 * from the goal block. @cpa is the prealloc
3195 * space that is having currently known minimal distance
3196 * from the goal block.
3198 static struct ext4_prealloc_space
*
3199 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3200 struct ext4_prealloc_space
*pa
,
3201 struct ext4_prealloc_space
*cpa
)
3203 ext4_fsblk_t cur_distance
, new_distance
;
3206 atomic_inc(&pa
->pa_count
);
3209 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3210 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3212 if (cur_distance
< new_distance
)
3215 /* drop the previous reference */
3216 atomic_dec(&cpa
->pa_count
);
3217 atomic_inc(&pa
->pa_count
);
3222 * search goal blocks in preallocated space
3224 static noinline_for_stack
int
3225 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3228 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3229 struct ext4_locality_group
*lg
;
3230 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3231 ext4_fsblk_t goal_block
;
3233 /* only data can be preallocated */
3234 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3237 /* first, try per-file preallocation */
3239 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3241 /* all fields in this condition don't change,
3242 * so we can skip locking for them */
3243 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3244 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3247 /* found preallocated blocks, use them */
3248 spin_lock(&pa
->pa_lock
);
3249 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3250 atomic_inc(&pa
->pa_count
);
3251 ext4_mb_use_inode_pa(ac
, pa
);
3252 spin_unlock(&pa
->pa_lock
);
3253 ac
->ac_criteria
= 10;
3257 spin_unlock(&pa
->pa_lock
);
3261 /* can we use group allocation? */
3262 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3265 /* inode may have no locality group for some reason */
3269 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3270 if (order
> PREALLOC_TB_SIZE
- 1)
3271 /* The max size of hash table is PREALLOC_TB_SIZE */
3272 order
= PREALLOC_TB_SIZE
- 1;
3274 goal_block
= ac
->ac_g_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
) +
3275 ac
->ac_g_ex
.fe_start
+
3276 le32_to_cpu(EXT4_SB(ac
->ac_sb
)->s_es
->s_first_data_block
);
3278 * search for the prealloc space that is having
3279 * minimal distance from the goal block.
3281 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3283 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3285 spin_lock(&pa
->pa_lock
);
3286 if (pa
->pa_deleted
== 0 &&
3287 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3289 cpa
= ext4_mb_check_group_pa(goal_block
,
3292 spin_unlock(&pa
->pa_lock
);
3297 ext4_mb_use_group_pa(ac
, cpa
);
3298 ac
->ac_criteria
= 20;
3305 * the function goes through all preallocation in this group and marks them
3306 * used in in-core bitmap. buddy must be generated from this bitmap
3307 * Need to be called with ext4 group lock (ext4_lock_group)
3309 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3312 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3313 struct ext4_prealloc_space
*pa
;
3314 struct list_head
*cur
;
3315 ext4_group_t groupnr
;
3316 ext4_grpblk_t start
;
3317 int preallocated
= 0;
3321 /* all form of preallocation discards first load group,
3322 * so the only competing code is preallocation use.
3323 * we don't need any locking here
3324 * notice we do NOT ignore preallocations with pa_deleted
3325 * otherwise we could leave used blocks available for
3326 * allocation in buddy when concurrent ext4_mb_put_pa()
3327 * is dropping preallocation
3329 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3330 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3331 spin_lock(&pa
->pa_lock
);
3332 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3335 spin_unlock(&pa
->pa_lock
);
3336 if (unlikely(len
== 0))
3338 BUG_ON(groupnr
!= group
);
3339 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3340 bitmap
, start
, len
);
3341 preallocated
+= len
;
3344 mb_debug("prellocated %u for group %lu\n", preallocated
, group
);
3347 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3349 struct ext4_prealloc_space
*pa
;
3350 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3351 kmem_cache_free(ext4_pspace_cachep
, pa
);
3355 * drops a reference to preallocated space descriptor
3356 * if this was the last reference and the space is consumed
3358 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3359 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3363 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3366 /* in this short window concurrent discard can set pa_deleted */
3367 spin_lock(&pa
->pa_lock
);
3368 if (pa
->pa_deleted
== 1) {
3369 spin_unlock(&pa
->pa_lock
);
3374 spin_unlock(&pa
->pa_lock
);
3376 /* -1 is to protect from crossing allocation group */
3377 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
- 1, &grp
, NULL
);
3382 * P1 (buddy init) P2 (regular allocation)
3383 * find block B in PA
3384 * copy on-disk bitmap to buddy
3385 * mark B in on-disk bitmap
3386 * drop PA from group
3387 * mark all PAs in buddy
3389 * thus, P1 initializes buddy with B available. to prevent this
3390 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3393 ext4_lock_group(sb
, grp
);
3394 list_del(&pa
->pa_group_list
);
3395 ext4_unlock_group(sb
, grp
);
3397 spin_lock(pa
->pa_obj_lock
);
3398 list_del_rcu(&pa
->pa_inode_list
);
3399 spin_unlock(pa
->pa_obj_lock
);
3401 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3405 * creates new preallocated space for given inode
3407 static noinline_for_stack
int
3408 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3410 struct super_block
*sb
= ac
->ac_sb
;
3411 struct ext4_prealloc_space
*pa
;
3412 struct ext4_group_info
*grp
;
3413 struct ext4_inode_info
*ei
;
3415 /* preallocate only when found space is larger then requested */
3416 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3417 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3418 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3420 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3424 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3430 /* we can't allocate as much as normalizer wants.
3431 * so, found space must get proper lstart
3432 * to cover original request */
3433 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3434 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3436 /* we're limited by original request in that
3437 * logical block must be covered any way
3438 * winl is window we can move our chunk within */
3439 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3441 /* also, we should cover whole original request */
3442 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3444 /* the smallest one defines real window */
3445 win
= min(winl
, wins
);
3447 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3448 if (offs
&& offs
< win
)
3451 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3452 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3453 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3456 /* preallocation can change ac_b_ex, thus we store actually
3457 * allocated blocks for history */
3458 ac
->ac_f_ex
= ac
->ac_b_ex
;
3460 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3461 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3462 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3463 pa
->pa_free
= pa
->pa_len
;
3464 atomic_set(&pa
->pa_count
, 1);
3465 spin_lock_init(&pa
->pa_lock
);
3469 mb_debug("new inode pa %p: %llu/%u for %u\n", pa
,
3470 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3472 ext4_mb_use_inode_pa(ac
, pa
);
3473 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3475 ei
= EXT4_I(ac
->ac_inode
);
3476 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3478 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3479 pa
->pa_inode
= ac
->ac_inode
;
3481 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3482 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3483 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3485 spin_lock(pa
->pa_obj_lock
);
3486 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3487 spin_unlock(pa
->pa_obj_lock
);
3493 * creates new preallocated space for locality group inodes belongs to
3495 static noinline_for_stack
int
3496 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3498 struct super_block
*sb
= ac
->ac_sb
;
3499 struct ext4_locality_group
*lg
;
3500 struct ext4_prealloc_space
*pa
;
3501 struct ext4_group_info
*grp
;
3503 /* preallocate only when found space is larger then requested */
3504 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3505 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3506 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3508 BUG_ON(ext4_pspace_cachep
== NULL
);
3509 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3513 /* preallocation can change ac_b_ex, thus we store actually
3514 * allocated blocks for history */
3515 ac
->ac_f_ex
= ac
->ac_b_ex
;
3517 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3518 pa
->pa_lstart
= pa
->pa_pstart
;
3519 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3520 pa
->pa_free
= pa
->pa_len
;
3521 atomic_set(&pa
->pa_count
, 1);
3522 spin_lock_init(&pa
->pa_lock
);
3523 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3527 mb_debug("new group pa %p: %llu/%u for %u\n", pa
,
3528 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3530 ext4_mb_use_group_pa(ac
, pa
);
3531 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3533 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3537 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3538 pa
->pa_inode
= NULL
;
3540 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3541 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3542 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3545 * We will later add the new pa to the right bucket
3546 * after updating the pa_free in ext4_mb_release_context
3551 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3555 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3556 err
= ext4_mb_new_group_pa(ac
);
3558 err
= ext4_mb_new_inode_pa(ac
);
3563 * finds all unused blocks in on-disk bitmap, frees them in
3564 * in-core bitmap and buddy.
3565 * @pa must be unlinked from inode and group lists, so that
3566 * nobody else can find/use it.
3567 * the caller MUST hold group/inode locks.
3568 * TODO: optimize the case when there are no in-core structures yet
3570 static noinline_for_stack
int
3571 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3572 struct ext4_prealloc_space
*pa
,
3573 struct ext4_allocation_context
*ac
)
3575 struct super_block
*sb
= e4b
->bd_sb
;
3576 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3585 BUG_ON(pa
->pa_deleted
== 0);
3586 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3587 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3588 end
= bit
+ pa
->pa_len
;
3592 ac
->ac_inode
= pa
->pa_inode
;
3593 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3597 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3600 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3601 start
= group
* EXT4_BLOCKS_PER_GROUP(sb
) + bit
+
3602 le32_to_cpu(sbi
->s_es
->s_first_data_block
);
3603 mb_debug(" free preallocated %u/%u in group %u\n",
3604 (unsigned) start
, (unsigned) next
- bit
,
3609 ac
->ac_b_ex
.fe_group
= group
;
3610 ac
->ac_b_ex
.fe_start
= bit
;
3611 ac
->ac_b_ex
.fe_len
= next
- bit
;
3612 ac
->ac_b_ex
.fe_logical
= 0;
3613 ext4_mb_store_history(ac
);
3616 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3619 if (free
!= pa
->pa_free
) {
3620 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3621 pa
, (unsigned long) pa
->pa_lstart
,
3622 (unsigned long) pa
->pa_pstart
,
3623 (unsigned long) pa
->pa_len
);
3624 ext4_error(sb
, __func__
, "free %u, pa_free %u\n",
3627 * pa is already deleted so we use the value obtained
3628 * from the bitmap and continue.
3631 atomic_add(free
, &sbi
->s_mb_discarded
);
3636 static noinline_for_stack
int
3637 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3638 struct ext4_prealloc_space
*pa
,
3639 struct ext4_allocation_context
*ac
)
3641 struct super_block
*sb
= e4b
->bd_sb
;
3646 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3648 BUG_ON(pa
->pa_deleted
== 0);
3649 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3650 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3651 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3652 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3656 ac
->ac_inode
= NULL
;
3657 ac
->ac_b_ex
.fe_group
= group
;
3658 ac
->ac_b_ex
.fe_start
= bit
;
3659 ac
->ac_b_ex
.fe_len
= pa
->pa_len
;
3660 ac
->ac_b_ex
.fe_logical
= 0;
3661 ext4_mb_store_history(ac
);
3668 * releases all preallocations in given group
3670 * first, we need to decide discard policy:
3671 * - when do we discard
3673 * - how many do we discard
3674 * 1) how many requested
3676 static noinline_for_stack
int
3677 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3678 ext4_group_t group
, int needed
)
3680 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3681 struct buffer_head
*bitmap_bh
= NULL
;
3682 struct ext4_prealloc_space
*pa
, *tmp
;
3683 struct ext4_allocation_context
*ac
;
3684 struct list_head list
;
3685 struct ext4_buddy e4b
;
3690 mb_debug("discard preallocation for group %lu\n", group
);
3692 if (list_empty(&grp
->bb_prealloc_list
))
3695 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3696 if (bitmap_bh
== NULL
) {
3697 ext4_error(sb
, __func__
, "Error in reading block "
3698 "bitmap for %lu\n", group
);
3702 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3704 ext4_error(sb
, __func__
, "Error in loading buddy "
3705 "information for %lu\n", group
);
3711 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3713 INIT_LIST_HEAD(&list
);
3714 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3716 ext4_lock_group(sb
, group
);
3717 list_for_each_entry_safe(pa
, tmp
,
3718 &grp
->bb_prealloc_list
, pa_group_list
) {
3719 spin_lock(&pa
->pa_lock
);
3720 if (atomic_read(&pa
->pa_count
)) {
3721 spin_unlock(&pa
->pa_lock
);
3725 if (pa
->pa_deleted
) {
3726 spin_unlock(&pa
->pa_lock
);
3730 /* seems this one can be freed ... */
3733 /* we can trust pa_free ... */
3734 free
+= pa
->pa_free
;
3736 spin_unlock(&pa
->pa_lock
);
3738 list_del(&pa
->pa_group_list
);
3739 list_add(&pa
->u
.pa_tmp_list
, &list
);
3742 /* if we still need more blocks and some PAs were used, try again */
3743 if (free
< needed
&& busy
) {
3745 ext4_unlock_group(sb
, group
);
3747 * Yield the CPU here so that we don't get soft lockup
3748 * in non preempt case.
3754 /* found anything to free? */
3755 if (list_empty(&list
)) {
3760 /* now free all selected PAs */
3761 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3763 /* remove from object (inode or locality group) */
3764 spin_lock(pa
->pa_obj_lock
);
3765 list_del_rcu(&pa
->pa_inode_list
);
3766 spin_unlock(pa
->pa_obj_lock
);
3769 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
3771 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3773 list_del(&pa
->u
.pa_tmp_list
);
3774 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3778 ext4_unlock_group(sb
, group
);
3780 kmem_cache_free(ext4_ac_cachep
, ac
);
3781 ext4_mb_release_desc(&e4b
);
3787 * releases all non-used preallocated blocks for given inode
3789 * It's important to discard preallocations under i_data_sem
3790 * We don't want another block to be served from the prealloc
3791 * space when we are discarding the inode prealloc space.
3793 * FIXME!! Make sure it is valid at all the call sites
3795 void ext4_discard_preallocations(struct inode
*inode
)
3797 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3798 struct super_block
*sb
= inode
->i_sb
;
3799 struct buffer_head
*bitmap_bh
= NULL
;
3800 struct ext4_prealloc_space
*pa
, *tmp
;
3801 struct ext4_allocation_context
*ac
;
3802 ext4_group_t group
= 0;
3803 struct list_head list
;
3804 struct ext4_buddy e4b
;
3807 if (!S_ISREG(inode
->i_mode
)) {
3808 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3812 mb_debug("discard preallocation for inode %lu\n", inode
->i_ino
);
3814 INIT_LIST_HEAD(&list
);
3816 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3818 /* first, collect all pa's in the inode */
3819 spin_lock(&ei
->i_prealloc_lock
);
3820 while (!list_empty(&ei
->i_prealloc_list
)) {
3821 pa
= list_entry(ei
->i_prealloc_list
.next
,
3822 struct ext4_prealloc_space
, pa_inode_list
);
3823 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3824 spin_lock(&pa
->pa_lock
);
3825 if (atomic_read(&pa
->pa_count
)) {
3826 /* this shouldn't happen often - nobody should
3827 * use preallocation while we're discarding it */
3828 spin_unlock(&pa
->pa_lock
);
3829 spin_unlock(&ei
->i_prealloc_lock
);
3830 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
3832 schedule_timeout_uninterruptible(HZ
);
3836 if (pa
->pa_deleted
== 0) {
3838 spin_unlock(&pa
->pa_lock
);
3839 list_del_rcu(&pa
->pa_inode_list
);
3840 list_add(&pa
->u
.pa_tmp_list
, &list
);
3844 /* someone is deleting pa right now */
3845 spin_unlock(&pa
->pa_lock
);
3846 spin_unlock(&ei
->i_prealloc_lock
);
3848 /* we have to wait here because pa_deleted
3849 * doesn't mean pa is already unlinked from
3850 * the list. as we might be called from
3851 * ->clear_inode() the inode will get freed
3852 * and concurrent thread which is unlinking
3853 * pa from inode's list may access already
3854 * freed memory, bad-bad-bad */
3856 /* XXX: if this happens too often, we can
3857 * add a flag to force wait only in case
3858 * of ->clear_inode(), but not in case of
3859 * regular truncate */
3860 schedule_timeout_uninterruptible(HZ
);
3863 spin_unlock(&ei
->i_prealloc_lock
);
3865 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3866 BUG_ON(pa
->pa_linear
!= 0);
3867 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
3869 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3871 ext4_error(sb
, __func__
, "Error in loading buddy "
3872 "information for %lu\n", group
);
3876 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3877 if (bitmap_bh
== NULL
) {
3878 ext4_error(sb
, __func__
, "Error in reading block "
3879 "bitmap for %lu\n", group
);
3880 ext4_mb_release_desc(&e4b
);
3884 ext4_lock_group(sb
, group
);
3885 list_del(&pa
->pa_group_list
);
3886 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3887 ext4_unlock_group(sb
, group
);
3889 ext4_mb_release_desc(&e4b
);
3892 list_del(&pa
->u
.pa_tmp_list
);
3893 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3896 kmem_cache_free(ext4_ac_cachep
, ac
);
3900 * finds all preallocated spaces and return blocks being freed to them
3901 * if preallocated space becomes full (no block is used from the space)
3902 * then the function frees space in buddy
3903 * XXX: at the moment, truncate (which is the only way to free blocks)
3904 * discards all preallocations
3906 static void ext4_mb_return_to_preallocation(struct inode
*inode
,
3907 struct ext4_buddy
*e4b
,
3908 sector_t block
, int count
)
3910 BUG_ON(!list_empty(&EXT4_I(inode
)->i_prealloc_list
));
3913 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3915 struct super_block
*sb
= ac
->ac_sb
;
3918 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
3919 " Allocation context details:\n");
3920 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
3921 ac
->ac_status
, ac
->ac_flags
);
3922 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3923 "best %lu/%lu/%lu@%lu cr %d\n",
3924 (unsigned long)ac
->ac_o_ex
.fe_group
,
3925 (unsigned long)ac
->ac_o_ex
.fe_start
,
3926 (unsigned long)ac
->ac_o_ex
.fe_len
,
3927 (unsigned long)ac
->ac_o_ex
.fe_logical
,
3928 (unsigned long)ac
->ac_g_ex
.fe_group
,
3929 (unsigned long)ac
->ac_g_ex
.fe_start
,
3930 (unsigned long)ac
->ac_g_ex
.fe_len
,
3931 (unsigned long)ac
->ac_g_ex
.fe_logical
,
3932 (unsigned long)ac
->ac_b_ex
.fe_group
,
3933 (unsigned long)ac
->ac_b_ex
.fe_start
,
3934 (unsigned long)ac
->ac_b_ex
.fe_len
,
3935 (unsigned long)ac
->ac_b_ex
.fe_logical
,
3936 (int)ac
->ac_criteria
);
3937 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
3939 printk(KERN_ERR
"EXT4-fs: groups: \n");
3940 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; i
++) {
3941 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
3942 struct ext4_prealloc_space
*pa
;
3943 ext4_grpblk_t start
;
3944 struct list_head
*cur
;
3945 ext4_lock_group(sb
, i
);
3946 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3947 pa
= list_entry(cur
, struct ext4_prealloc_space
,
3949 spin_lock(&pa
->pa_lock
);
3950 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3952 spin_unlock(&pa
->pa_lock
);
3953 printk(KERN_ERR
"PA:%lu:%d:%u \n", i
,
3956 ext4_unlock_group(sb
, i
);
3958 if (grp
->bb_free
== 0)
3960 printk(KERN_ERR
"%lu: %d/%d \n",
3961 i
, grp
->bb_free
, grp
->bb_fragments
);
3963 printk(KERN_ERR
"\n");
3966 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3973 * We use locality group preallocation for small size file. The size of the
3974 * file is determined by the current size or the resulting size after
3975 * allocation which ever is larger
3977 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
3979 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
3981 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3982 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3985 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3988 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3989 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
3990 size
= max(size
, isize
);
3992 /* don't use group allocation for large files */
3993 if (size
>= sbi
->s_mb_stream_request
)
3996 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3999 BUG_ON(ac
->ac_lg
!= NULL
);
4001 * locality group prealloc space are per cpu. The reason for having
4002 * per cpu locality group is to reduce the contention between block
4003 * request from multiple CPUs.
4005 ac
->ac_lg
= per_cpu_ptr(sbi
->s_locality_groups
, raw_smp_processor_id());
4007 /* we're going to use group allocation */
4008 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4010 /* serialize all allocations in the group */
4011 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4014 static noinline_for_stack
int
4015 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4016 struct ext4_allocation_request
*ar
)
4018 struct super_block
*sb
= ar
->inode
->i_sb
;
4019 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4020 struct ext4_super_block
*es
= sbi
->s_es
;
4024 ext4_grpblk_t block
;
4026 /* we can't allocate > group size */
4029 /* just a dirty hack to filter too big requests */
4030 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4031 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4033 /* start searching from the goal */
4035 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4036 goal
>= ext4_blocks_count(es
))
4037 goal
= le32_to_cpu(es
->s_first_data_block
);
4038 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4040 /* set up allocation goals */
4041 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4042 ac
->ac_b_ex
.fe_group
= 0;
4043 ac
->ac_b_ex
.fe_start
= 0;
4044 ac
->ac_b_ex
.fe_len
= 0;
4045 ac
->ac_status
= AC_STATUS_CONTINUE
;
4046 ac
->ac_groups_scanned
= 0;
4047 ac
->ac_ex_scanned
= 0;
4050 ac
->ac_inode
= ar
->inode
;
4051 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4052 ac
->ac_o_ex
.fe_group
= group
;
4053 ac
->ac_o_ex
.fe_start
= block
;
4054 ac
->ac_o_ex
.fe_len
= len
;
4055 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4056 ac
->ac_g_ex
.fe_group
= group
;
4057 ac
->ac_g_ex
.fe_start
= block
;
4058 ac
->ac_g_ex
.fe_len
= len
;
4059 ac
->ac_f_ex
.fe_len
= 0;
4060 ac
->ac_flags
= ar
->flags
;
4062 ac
->ac_criteria
= 0;
4064 ac
->ac_bitmap_page
= NULL
;
4065 ac
->ac_buddy_page
= NULL
;
4068 /* we have to define context: we'll we work with a file or
4069 * locality group. this is a policy, actually */
4070 ext4_mb_group_or_file(ac
);
4072 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4073 "left: %u/%u, right %u/%u to %swritable\n",
4074 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4075 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4076 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4077 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4078 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4083 static noinline_for_stack
void
4084 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4085 struct ext4_locality_group
*lg
,
4086 int order
, int total_entries
)
4088 ext4_group_t group
= 0;
4089 struct ext4_buddy e4b
;
4090 struct list_head discard_list
;
4091 struct ext4_prealloc_space
*pa
, *tmp
;
4092 struct ext4_allocation_context
*ac
;
4094 mb_debug("discard locality group preallocation\n");
4096 INIT_LIST_HEAD(&discard_list
);
4097 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4099 spin_lock(&lg
->lg_prealloc_lock
);
4100 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4102 spin_lock(&pa
->pa_lock
);
4103 if (atomic_read(&pa
->pa_count
)) {
4105 * This is the pa that we just used
4106 * for block allocation. So don't
4109 spin_unlock(&pa
->pa_lock
);
4112 if (pa
->pa_deleted
) {
4113 spin_unlock(&pa
->pa_lock
);
4116 /* only lg prealloc space */
4117 BUG_ON(!pa
->pa_linear
);
4119 /* seems this one can be freed ... */
4121 spin_unlock(&pa
->pa_lock
);
4123 list_del_rcu(&pa
->pa_inode_list
);
4124 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4127 if (total_entries
<= 5) {
4129 * we want to keep only 5 entries
4130 * allowing it to grow to 8. This
4131 * mak sure we don't call discard
4132 * soon for this list.
4137 spin_unlock(&lg
->lg_prealloc_lock
);
4139 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4141 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4142 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4143 ext4_error(sb
, __func__
, "Error in loading buddy "
4144 "information for %lu\n", group
);
4147 ext4_lock_group(sb
, group
);
4148 list_del(&pa
->pa_group_list
);
4149 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
4150 ext4_unlock_group(sb
, group
);
4152 ext4_mb_release_desc(&e4b
);
4153 list_del(&pa
->u
.pa_tmp_list
);
4154 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4157 kmem_cache_free(ext4_ac_cachep
, ac
);
4161 * We have incremented pa_count. So it cannot be freed at this
4162 * point. Also we hold lg_mutex. So no parallel allocation is
4163 * possible from this lg. That means pa_free cannot be updated.
4165 * A parallel ext4_mb_discard_group_preallocations is possible.
4166 * which can cause the lg_prealloc_list to be updated.
4169 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4171 int order
, added
= 0, lg_prealloc_count
= 1;
4172 struct super_block
*sb
= ac
->ac_sb
;
4173 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4174 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4176 order
= fls(pa
->pa_free
) - 1;
4177 if (order
> PREALLOC_TB_SIZE
- 1)
4178 /* The max size of hash table is PREALLOC_TB_SIZE */
4179 order
= PREALLOC_TB_SIZE
- 1;
4180 /* Add the prealloc space to lg */
4182 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4184 spin_lock(&tmp_pa
->pa_lock
);
4185 if (tmp_pa
->pa_deleted
) {
4186 spin_unlock(&pa
->pa_lock
);
4189 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4190 /* Add to the tail of the previous entry */
4191 list_add_tail_rcu(&pa
->pa_inode_list
,
4192 &tmp_pa
->pa_inode_list
);
4195 * we want to count the total
4196 * number of entries in the list
4199 spin_unlock(&tmp_pa
->pa_lock
);
4200 lg_prealloc_count
++;
4203 list_add_tail_rcu(&pa
->pa_inode_list
,
4204 &lg
->lg_prealloc_list
[order
]);
4207 /* Now trim the list to be not more than 8 elements */
4208 if (lg_prealloc_count
> 8) {
4209 ext4_mb_discard_lg_preallocations(sb
, lg
,
4210 order
, lg_prealloc_count
);
4217 * release all resource we used in allocation
4219 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4221 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4223 if (pa
->pa_linear
) {
4224 /* see comment in ext4_mb_use_group_pa() */
4225 spin_lock(&pa
->pa_lock
);
4226 pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4227 pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4228 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4229 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4230 spin_unlock(&pa
->pa_lock
);
4232 * We want to add the pa to the right bucket.
4233 * Remove it from the list and while adding
4234 * make sure the list to which we are adding
4237 if (likely(pa
->pa_free
)) {
4238 spin_lock(pa
->pa_obj_lock
);
4239 list_del_rcu(&pa
->pa_inode_list
);
4240 spin_unlock(pa
->pa_obj_lock
);
4241 ext4_mb_add_n_trim(ac
);
4244 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4246 if (ac
->ac_bitmap_page
)
4247 page_cache_release(ac
->ac_bitmap_page
);
4248 if (ac
->ac_buddy_page
)
4249 page_cache_release(ac
->ac_buddy_page
);
4250 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4251 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4252 ext4_mb_collect_stats(ac
);
4256 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4262 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
&& needed
> 0; i
++) {
4263 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4272 * Main entry point into mballoc to allocate blocks
4273 * it tries to use preallocation first, then falls back
4274 * to usual allocation
4276 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4277 struct ext4_allocation_request
*ar
, int *errp
)
4280 struct ext4_allocation_context
*ac
= NULL
;
4281 struct ext4_sb_info
*sbi
;
4282 struct super_block
*sb
;
4283 ext4_fsblk_t block
= 0;
4284 unsigned long inquota
;
4285 unsigned long reserv_blks
= 0;
4287 sb
= ar
->inode
->i_sb
;
4290 if (!EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
) {
4292 * With delalloc we already reserved the blocks
4294 while (ar
->len
&& ext4_claim_free_blocks(sbi
, ar
->len
)) {
4295 /* let others to free the space */
4297 ar
->len
= ar
->len
>> 1;
4303 reserv_blks
= ar
->len
;
4305 while (ar
->len
&& DQUOT_ALLOC_BLOCK(ar
->inode
, ar
->len
)) {
4306 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4315 if (EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
)
4316 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4318 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4325 ext4_mb_poll_new_transaction(sb
, handle
);
4327 *errp
= ext4_mb_initialize_context(ac
, ar
);
4333 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4334 if (!ext4_mb_use_preallocated(ac
)) {
4335 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4336 ext4_mb_normalize_request(ac
, ar
);
4338 /* allocate space in core */
4339 ext4_mb_regular_allocator(ac
);
4341 /* as we've just preallocated more space than
4342 * user requested orinally, we store allocated
4343 * space in a special descriptor */
4344 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4345 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4346 ext4_mb_new_preallocation(ac
);
4349 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4350 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_blks
);
4351 if (*errp
== -EAGAIN
) {
4352 ac
->ac_b_ex
.fe_group
= 0;
4353 ac
->ac_b_ex
.fe_start
= 0;
4354 ac
->ac_b_ex
.fe_len
= 0;
4355 ac
->ac_status
= AC_STATUS_CONTINUE
;
4358 ac
->ac_b_ex
.fe_len
= 0;
4360 ext4_mb_show_ac(ac
);
4362 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4363 ar
->len
= ac
->ac_b_ex
.fe_len
;
4366 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4370 ac
->ac_b_ex
.fe_len
= 0;
4372 ext4_mb_show_ac(ac
);
4375 ext4_mb_release_context(ac
);
4378 kmem_cache_free(ext4_ac_cachep
, ac
);
4380 if (ar
->len
< inquota
)
4381 DQUOT_FREE_BLOCK(ar
->inode
, inquota
- ar
->len
);
4385 static void ext4_mb_poll_new_transaction(struct super_block
*sb
,
4388 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4390 if (sbi
->s_last_transaction
== handle
->h_transaction
->t_tid
)
4393 /* new transaction! time to close last one and free blocks for
4394 * committed transaction. we know that only transaction can be
4395 * active, so previos transaction can be being logged and we
4396 * know that transaction before previous is known to be already
4397 * logged. this means that now we may free blocks freed in all
4398 * transactions before previous one. hope I'm clear enough ... */
4400 spin_lock(&sbi
->s_md_lock
);
4401 if (sbi
->s_last_transaction
!= handle
->h_transaction
->t_tid
) {
4402 mb_debug("new transaction %lu, old %lu\n",
4403 (unsigned long) handle
->h_transaction
->t_tid
,
4404 (unsigned long) sbi
->s_last_transaction
);
4405 list_splice_init(&sbi
->s_closed_transaction
,
4406 &sbi
->s_committed_transaction
);
4407 list_splice_init(&sbi
->s_active_transaction
,
4408 &sbi
->s_closed_transaction
);
4409 sbi
->s_last_transaction
= handle
->h_transaction
->t_tid
;
4411 spin_unlock(&sbi
->s_md_lock
);
4413 ext4_mb_free_committed_blocks(sb
);
4416 static noinline_for_stack
int
4417 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4418 ext4_group_t group
, ext4_grpblk_t block
, int count
)
4420 struct ext4_group_info
*db
= e4b
->bd_info
;
4421 struct super_block
*sb
= e4b
->bd_sb
;
4422 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4423 struct ext4_free_metadata
*md
;
4426 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4427 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4429 ext4_lock_group(sb
, group
);
4430 for (i
= 0; i
< count
; i
++) {
4432 if (md
&& db
->bb_tid
!= handle
->h_transaction
->t_tid
) {
4433 db
->bb_md_cur
= NULL
;
4438 ext4_unlock_group(sb
, group
);
4439 md
= kmalloc(sizeof(*md
), GFP_NOFS
);
4445 ext4_lock_group(sb
, group
);
4446 if (db
->bb_md_cur
== NULL
) {
4447 spin_lock(&sbi
->s_md_lock
);
4448 list_add(&md
->list
, &sbi
->s_active_transaction
);
4449 spin_unlock(&sbi
->s_md_lock
);
4450 /* protect buddy cache from being freed,
4451 * otherwise we'll refresh it from
4452 * on-disk bitmap and lose not-yet-available
4454 page_cache_get(e4b
->bd_buddy_page
);
4455 page_cache_get(e4b
->bd_bitmap_page
);
4457 db
->bb_tid
= handle
->h_transaction
->t_tid
;
4458 mb_debug("new md 0x%p for group %lu\n",
4466 BUG_ON(md
->num
>= EXT4_BB_MAX_BLOCKS
);
4467 md
->blocks
[md
->num
] = block
+ i
;
4469 if (md
->num
== EXT4_BB_MAX_BLOCKS
) {
4470 /* no more space, put full container on a sb's list */
4471 db
->bb_md_cur
= NULL
;
4474 ext4_unlock_group(sb
, group
);
4479 * Main entry point into mballoc to free blocks
4481 void ext4_mb_free_blocks(handle_t
*handle
, struct inode
*inode
,
4482 unsigned long block
, unsigned long count
,
4483 int metadata
, unsigned long *freed
)
4485 struct buffer_head
*bitmap_bh
= NULL
;
4486 struct super_block
*sb
= inode
->i_sb
;
4487 struct ext4_allocation_context
*ac
= NULL
;
4488 struct ext4_group_desc
*gdp
;
4489 struct ext4_super_block
*es
;
4490 unsigned long overflow
;
4492 struct buffer_head
*gd_bh
;
4493 ext4_group_t block_group
;
4494 struct ext4_sb_info
*sbi
;
4495 struct ext4_buddy e4b
;
4501 ext4_mb_poll_new_transaction(sb
, handle
);
4504 es
= EXT4_SB(sb
)->s_es
;
4505 if (block
< le32_to_cpu(es
->s_first_data_block
) ||
4506 block
+ count
< block
||
4507 block
+ count
> ext4_blocks_count(es
)) {
4508 ext4_error(sb
, __func__
,
4509 "Freeing blocks not in datazone - "
4510 "block = %lu, count = %lu", block
, count
);
4514 ext4_debug("freeing block %lu\n", block
);
4516 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4518 ac
->ac_op
= EXT4_MB_HISTORY_FREE
;
4519 ac
->ac_inode
= inode
;
4525 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4528 * Check to see if we are freeing blocks across a group
4531 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4532 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4535 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4540 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4546 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4547 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4548 in_range(block
, ext4_inode_table(sb
, gdp
),
4549 EXT4_SB(sb
)->s_itb_per_group
) ||
4550 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4551 EXT4_SB(sb
)->s_itb_per_group
)) {
4553 ext4_error(sb
, __func__
,
4554 "Freeing blocks in system zone - "
4555 "Block = %lu, count = %lu", block
, count
);
4556 /* err = 0. ext4_std_error should be a no op */
4560 BUFFER_TRACE(bitmap_bh
, "getting write access");
4561 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4566 * We are about to modify some metadata. Call the journal APIs
4567 * to unshare ->b_data if a currently-committing transaction is
4570 BUFFER_TRACE(gd_bh
, "get_write_access");
4571 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4575 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4579 #ifdef AGGRESSIVE_CHECK
4582 for (i
= 0; i
< count
; i
++)
4583 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4586 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4589 /* We dirtied the bitmap block */
4590 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4591 err
= ext4_journal_dirty_metadata(handle
, bitmap_bh
);
4594 ac
->ac_b_ex
.fe_group
= block_group
;
4595 ac
->ac_b_ex
.fe_start
= bit
;
4596 ac
->ac_b_ex
.fe_len
= count
;
4597 ext4_mb_store_history(ac
);
4601 /* blocks being freed are metadata. these blocks shouldn't
4602 * be used until this transaction is committed */
4603 ext4_mb_free_metadata(handle
, &e4b
, block_group
, bit
, count
);
4605 ext4_lock_group(sb
, block_group
);
4606 mb_free_blocks(inode
, &e4b
, bit
, count
);
4607 ext4_mb_return_to_preallocation(inode
, &e4b
, block
, count
);
4608 ext4_unlock_group(sb
, block_group
);
4611 spin_lock(sb_bgl_lock(sbi
, block_group
));
4612 le16_add_cpu(&gdp
->bg_free_blocks_count
, count
);
4613 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4614 spin_unlock(sb_bgl_lock(sbi
, block_group
));
4615 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4617 if (sbi
->s_log_groups_per_flex
) {
4618 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4619 spin_lock(sb_bgl_lock(sbi
, flex_group
));
4620 sbi
->s_flex_groups
[flex_group
].free_blocks
+= count
;
4621 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
4624 ext4_mb_release_desc(&e4b
);
4628 /* And the group descriptor block */
4629 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4630 ret
= ext4_journal_dirty_metadata(handle
, gd_bh
);
4634 if (overflow
&& !err
) {
4643 ext4_std_error(sb
, err
);
4645 kmem_cache_free(ext4_ac_cachep
, ac
);