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)",
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",
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 (buffer_uptodate(bh
[i
]) &&
786 !(desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)))
790 spin_lock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
791 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
792 ext4_init_block_bitmap(sb
, bh
[i
],
793 first_group
+ i
, desc
);
794 set_buffer_uptodate(bh
[i
]);
795 unlock_buffer(bh
[i
]);
796 spin_unlock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
799 spin_unlock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
801 bh
[i
]->b_end_io
= end_buffer_read_sync
;
802 submit_bh(READ
, bh
[i
]);
803 mb_debug("read bitmap for group %lu\n", first_group
+ i
);
806 /* wait for I/O completion */
807 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
808 wait_on_buffer(bh
[i
]);
811 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
812 if (!buffer_uptodate(bh
[i
]))
816 first_block
= page
->index
* blocks_per_page
;
817 for (i
= 0; i
< blocks_per_page
; i
++) {
819 struct ext4_group_info
*grinfo
;
821 group
= (first_block
+ i
) >> 1;
822 if (group
>= EXT4_SB(sb
)->s_groups_count
)
826 * data carry information regarding this
827 * particular group in the format specified
831 data
= page_address(page
) + (i
* blocksize
);
832 bitmap
= bh
[group
- first_group
]->b_data
;
835 * We place the buddy block and bitmap block
838 if ((first_block
+ i
) & 1) {
839 /* this is block of buddy */
840 BUG_ON(incore
== NULL
);
841 mb_debug("put buddy for group %u in page %lu/%x\n",
842 group
, page
->index
, i
* blocksize
);
843 memset(data
, 0xff, blocksize
);
844 grinfo
= ext4_get_group_info(sb
, group
);
845 grinfo
->bb_fragments
= 0;
846 memset(grinfo
->bb_counters
, 0,
847 sizeof(unsigned short)*(sb
->s_blocksize_bits
+2));
849 * incore got set to the group block bitmap below
851 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
854 /* this is block of bitmap */
855 BUG_ON(incore
!= NULL
);
856 mb_debug("put bitmap for group %u in page %lu/%x\n",
857 group
, page
->index
, i
* blocksize
);
859 /* see comments in ext4_mb_put_pa() */
860 ext4_lock_group(sb
, group
);
861 memcpy(data
, bitmap
, blocksize
);
863 /* mark all preallocated blks used in in-core bitmap */
864 ext4_mb_generate_from_pa(sb
, data
, group
);
865 ext4_unlock_group(sb
, group
);
867 /* set incore so that the buddy information can be
868 * generated using this
873 SetPageUptodate(page
);
877 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
885 static noinline_for_stack
int
886 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
887 struct ext4_buddy
*e4b
)
889 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
890 struct inode
*inode
= sbi
->s_buddy_cache
;
898 mb_debug("load group %lu\n", group
);
900 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
902 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
903 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
905 e4b
->bd_group
= group
;
906 e4b
->bd_buddy_page
= NULL
;
907 e4b
->bd_bitmap_page
= NULL
;
910 * the buddy cache inode stores the block bitmap
911 * and buddy information in consecutive blocks.
912 * So for each group we need two blocks.
915 pnum
= block
/ blocks_per_page
;
916 poff
= block
% blocks_per_page
;
918 /* we could use find_or_create_page(), but it locks page
919 * what we'd like to avoid in fast path ... */
920 page
= find_get_page(inode
->i_mapping
, pnum
);
921 if (page
== NULL
|| !PageUptodate(page
)) {
923 page_cache_release(page
);
924 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
926 BUG_ON(page
->mapping
!= inode
->i_mapping
);
927 if (!PageUptodate(page
)) {
928 ret
= ext4_mb_init_cache(page
, NULL
);
933 mb_cmp_bitmaps(e4b
, page_address(page
) +
934 (poff
* sb
->s_blocksize
));
939 if (page
== NULL
|| !PageUptodate(page
)) {
943 e4b
->bd_bitmap_page
= page
;
944 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
945 mark_page_accessed(page
);
948 pnum
= block
/ blocks_per_page
;
949 poff
= block
% blocks_per_page
;
951 page
= find_get_page(inode
->i_mapping
, pnum
);
952 if (page
== NULL
|| !PageUptodate(page
)) {
954 page_cache_release(page
);
955 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
957 BUG_ON(page
->mapping
!= inode
->i_mapping
);
958 if (!PageUptodate(page
)) {
959 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
968 if (page
== NULL
|| !PageUptodate(page
)) {
972 e4b
->bd_buddy_page
= page
;
973 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
974 mark_page_accessed(page
);
976 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
977 BUG_ON(e4b
->bd_buddy_page
== NULL
);
982 if (e4b
->bd_bitmap_page
)
983 page_cache_release(e4b
->bd_bitmap_page
);
984 if (e4b
->bd_buddy_page
)
985 page_cache_release(e4b
->bd_buddy_page
);
986 e4b
->bd_buddy
= NULL
;
987 e4b
->bd_bitmap
= NULL
;
991 static void ext4_mb_release_desc(struct ext4_buddy
*e4b
)
993 if (e4b
->bd_bitmap_page
)
994 page_cache_release(e4b
->bd_bitmap_page
);
995 if (e4b
->bd_buddy_page
)
996 page_cache_release(e4b
->bd_buddy_page
);
1000 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1005 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1006 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1008 bb
= EXT4_MB_BUDDY(e4b
);
1009 while (order
<= e4b
->bd_blkbits
+ 1) {
1011 if (!mb_test_bit(block
, bb
)) {
1012 /* this block is part of buddy of order 'order' */
1015 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1021 static void mb_clear_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1027 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1028 /* fast path: clear whole word at once */
1029 addr
= bm
+ (cur
>> 3);
1034 mb_clear_bit_atomic(lock
, cur
, bm
);
1039 static void mb_set_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1045 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1046 /* fast path: set whole word at once */
1047 addr
= bm
+ (cur
>> 3);
1052 mb_set_bit_atomic(lock
, cur
, bm
);
1057 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1058 int first
, int count
)
1065 struct super_block
*sb
= e4b
->bd_sb
;
1067 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1068 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
1069 mb_check_buddy(e4b
);
1070 mb_free_blocks_double(inode
, e4b
, first
, count
);
1072 e4b
->bd_info
->bb_free
+= count
;
1073 if (first
< e4b
->bd_info
->bb_first_free
)
1074 e4b
->bd_info
->bb_first_free
= first
;
1076 /* let's maintain fragments counter */
1078 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1079 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1080 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1082 e4b
->bd_info
->bb_fragments
--;
1083 else if (!block
&& !max
)
1084 e4b
->bd_info
->bb_fragments
++;
1086 /* let's maintain buddy itself */
1087 while (count
-- > 0) {
1091 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1092 ext4_fsblk_t blocknr
;
1093 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
1096 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
1097 ext4_unlock_group(sb
, e4b
->bd_group
);
1098 ext4_error(sb
, __func__
, "double-free of inode"
1099 " %lu's block %llu(bit %u in group %lu)",
1100 inode
? inode
->i_ino
: 0, blocknr
, block
,
1102 ext4_lock_group(sb
, e4b
->bd_group
);
1104 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1105 e4b
->bd_info
->bb_counters
[order
]++;
1107 /* start of the buddy */
1108 buddy
= mb_find_buddy(e4b
, order
, &max
);
1112 if (mb_test_bit(block
, buddy
) ||
1113 mb_test_bit(block
+ 1, buddy
))
1116 /* both the buddies are free, try to coalesce them */
1117 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1123 /* for special purposes, we don't set
1124 * free bits in bitmap */
1125 mb_set_bit(block
, buddy
);
1126 mb_set_bit(block
+ 1, buddy
);
1128 e4b
->bd_info
->bb_counters
[order
]--;
1129 e4b
->bd_info
->bb_counters
[order
]--;
1133 e4b
->bd_info
->bb_counters
[order
]++;
1135 mb_clear_bit(block
, buddy2
);
1139 mb_check_buddy(e4b
);
1142 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1143 int needed
, struct ext4_free_extent
*ex
)
1150 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1153 buddy
= mb_find_buddy(e4b
, order
, &max
);
1154 BUG_ON(buddy
== NULL
);
1155 BUG_ON(block
>= max
);
1156 if (mb_test_bit(block
, buddy
)) {
1163 /* FIXME dorp order completely ? */
1164 if (likely(order
== 0)) {
1165 /* find actual order */
1166 order
= mb_find_order_for_block(e4b
, block
);
1167 block
= block
>> order
;
1170 ex
->fe_len
= 1 << order
;
1171 ex
->fe_start
= block
<< order
;
1172 ex
->fe_group
= e4b
->bd_group
;
1174 /* calc difference from given start */
1175 next
= next
- ex
->fe_start
;
1177 ex
->fe_start
+= next
;
1179 while (needed
> ex
->fe_len
&&
1180 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1182 if (block
+ 1 >= max
)
1185 next
= (block
+ 1) * (1 << order
);
1186 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1189 ord
= mb_find_order_for_block(e4b
, next
);
1192 block
= next
>> order
;
1193 ex
->fe_len
+= 1 << order
;
1196 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1200 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1206 int start
= ex
->fe_start
;
1207 int len
= ex
->fe_len
;
1212 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1213 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1214 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1215 mb_check_buddy(e4b
);
1216 mb_mark_used_double(e4b
, start
, len
);
1218 e4b
->bd_info
->bb_free
-= len
;
1219 if (e4b
->bd_info
->bb_first_free
== start
)
1220 e4b
->bd_info
->bb_first_free
+= len
;
1222 /* let's maintain fragments counter */
1224 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1225 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1226 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1228 e4b
->bd_info
->bb_fragments
++;
1229 else if (!mlen
&& !max
)
1230 e4b
->bd_info
->bb_fragments
--;
1232 /* let's maintain buddy itself */
1234 ord
= mb_find_order_for_block(e4b
, start
);
1236 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1237 /* the whole chunk may be allocated at once! */
1239 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1240 BUG_ON((start
>> ord
) >= max
);
1241 mb_set_bit(start
>> ord
, buddy
);
1242 e4b
->bd_info
->bb_counters
[ord
]--;
1249 /* store for history */
1251 ret
= len
| (ord
<< 16);
1253 /* we have to split large buddy */
1255 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1256 mb_set_bit(start
>> ord
, buddy
);
1257 e4b
->bd_info
->bb_counters
[ord
]--;
1260 cur
= (start
>> ord
) & ~1U;
1261 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1262 mb_clear_bit(cur
, buddy
);
1263 mb_clear_bit(cur
+ 1, buddy
);
1264 e4b
->bd_info
->bb_counters
[ord
]++;
1265 e4b
->bd_info
->bb_counters
[ord
]++;
1268 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b
->bd_sb
), ex
->fe_group
),
1269 EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1270 mb_check_buddy(e4b
);
1276 * Must be called under group lock!
1278 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1279 struct ext4_buddy
*e4b
)
1281 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1284 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1285 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1287 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1288 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1289 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1291 /* preallocation can change ac_b_ex, thus we store actually
1292 * allocated blocks for history */
1293 ac
->ac_f_ex
= ac
->ac_b_ex
;
1295 ac
->ac_status
= AC_STATUS_FOUND
;
1296 ac
->ac_tail
= ret
& 0xffff;
1297 ac
->ac_buddy
= ret
>> 16;
1299 /* XXXXXXX: SUCH A HORRIBLE **CK */
1301 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1302 get_page(ac
->ac_bitmap_page
);
1303 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1304 get_page(ac
->ac_buddy_page
);
1306 /* store last allocated for subsequent stream allocation */
1307 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1308 spin_lock(&sbi
->s_md_lock
);
1309 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1310 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1311 spin_unlock(&sbi
->s_md_lock
);
1316 * regular allocator, for general purposes allocation
1319 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1320 struct ext4_buddy
*e4b
,
1323 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1324 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1325 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1326 struct ext4_free_extent ex
;
1329 if (ac
->ac_status
== AC_STATUS_FOUND
)
1332 * We don't want to scan for a whole year
1334 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1335 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1336 ac
->ac_status
= AC_STATUS_BREAK
;
1341 * Haven't found good chunk so far, let's continue
1343 if (bex
->fe_len
< gex
->fe_len
)
1346 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1347 && bex
->fe_group
== e4b
->bd_group
) {
1348 /* recheck chunk's availability - we don't know
1349 * when it was found (within this lock-unlock
1351 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1352 if (max
>= gex
->fe_len
) {
1353 ext4_mb_use_best_found(ac
, e4b
);
1360 * The routine checks whether found extent is good enough. If it is,
1361 * then the extent gets marked used and flag is set to the context
1362 * to stop scanning. Otherwise, the extent is compared with the
1363 * previous found extent and if new one is better, then it's stored
1364 * in the context. Later, the best found extent will be used, if
1365 * mballoc can't find good enough extent.
1367 * FIXME: real allocation policy is to be designed yet!
1369 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1370 struct ext4_free_extent
*ex
,
1371 struct ext4_buddy
*e4b
)
1373 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1374 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1376 BUG_ON(ex
->fe_len
<= 0);
1377 BUG_ON(ex
->fe_len
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1378 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1379 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1384 * The special case - take what you catch first
1386 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1388 ext4_mb_use_best_found(ac
, e4b
);
1393 * Let's check whether the chuck is good enough
1395 if (ex
->fe_len
== gex
->fe_len
) {
1397 ext4_mb_use_best_found(ac
, e4b
);
1402 * If this is first found extent, just store it in the context
1404 if (bex
->fe_len
== 0) {
1410 * If new found extent is better, store it in the context
1412 if (bex
->fe_len
< gex
->fe_len
) {
1413 /* if the request isn't satisfied, any found extent
1414 * larger than previous best one is better */
1415 if (ex
->fe_len
> bex
->fe_len
)
1417 } else if (ex
->fe_len
> gex
->fe_len
) {
1418 /* if the request is satisfied, then we try to find
1419 * an extent that still satisfy the request, but is
1420 * smaller than previous one */
1421 if (ex
->fe_len
< bex
->fe_len
)
1425 ext4_mb_check_limits(ac
, e4b
, 0);
1428 static int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1429 struct ext4_buddy
*e4b
)
1431 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1432 ext4_group_t group
= ex
.fe_group
;
1436 BUG_ON(ex
.fe_len
<= 0);
1437 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1441 ext4_lock_group(ac
->ac_sb
, group
);
1442 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1446 ext4_mb_use_best_found(ac
, e4b
);
1449 ext4_unlock_group(ac
->ac_sb
, group
);
1450 ext4_mb_release_desc(e4b
);
1455 static int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1456 struct ext4_buddy
*e4b
)
1458 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1461 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1462 struct ext4_super_block
*es
= sbi
->s_es
;
1463 struct ext4_free_extent ex
;
1465 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1468 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1472 ext4_lock_group(ac
->ac_sb
, group
);
1473 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1474 ac
->ac_g_ex
.fe_len
, &ex
);
1476 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1479 start
= (e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
)) +
1480 ex
.fe_start
+ le32_to_cpu(es
->s_first_data_block
);
1481 /* use do_div to get remainder (would be 64-bit modulo) */
1482 if (do_div(start
, sbi
->s_stripe
) == 0) {
1485 ext4_mb_use_best_found(ac
, e4b
);
1487 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1488 BUG_ON(ex
.fe_len
<= 0);
1489 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1490 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1493 ext4_mb_use_best_found(ac
, e4b
);
1494 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1495 /* Sometimes, caller may want to merge even small
1496 * number of blocks to an existing extent */
1497 BUG_ON(ex
.fe_len
<= 0);
1498 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1499 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1502 ext4_mb_use_best_found(ac
, e4b
);
1504 ext4_unlock_group(ac
->ac_sb
, group
);
1505 ext4_mb_release_desc(e4b
);
1511 * The routine scans buddy structures (not bitmap!) from given order
1512 * to max order and tries to find big enough chunk to satisfy the req
1514 static void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1515 struct ext4_buddy
*e4b
)
1517 struct super_block
*sb
= ac
->ac_sb
;
1518 struct ext4_group_info
*grp
= e4b
->bd_info
;
1524 BUG_ON(ac
->ac_2order
<= 0);
1525 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1526 if (grp
->bb_counters
[i
] == 0)
1529 buddy
= mb_find_buddy(e4b
, i
, &max
);
1530 BUG_ON(buddy
== NULL
);
1532 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1537 ac
->ac_b_ex
.fe_len
= 1 << i
;
1538 ac
->ac_b_ex
.fe_start
= k
<< i
;
1539 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1541 ext4_mb_use_best_found(ac
, e4b
);
1543 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1545 if (EXT4_SB(sb
)->s_mb_stats
)
1546 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1553 * The routine scans the group and measures all found extents.
1554 * In order to optimize scanning, caller must pass number of
1555 * free blocks in the group, so the routine can know upper limit.
1557 static void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1558 struct ext4_buddy
*e4b
)
1560 struct super_block
*sb
= ac
->ac_sb
;
1561 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1562 struct ext4_free_extent ex
;
1566 free
= e4b
->bd_info
->bb_free
;
1569 i
= e4b
->bd_info
->bb_first_free
;
1571 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1572 i
= mb_find_next_zero_bit(bitmap
,
1573 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1574 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1576 * IF we have corrupt bitmap, we won't find any
1577 * free blocks even though group info says we
1578 * we have free blocks
1580 ext4_error(sb
, __func__
, "%d free blocks as per "
1581 "group info. But bitmap says 0",
1586 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1587 BUG_ON(ex
.fe_len
<= 0);
1588 if (free
< ex
.fe_len
) {
1589 ext4_error(sb
, __func__
, "%d free blocks as per "
1590 "group info. But got %d blocks",
1593 * The number of free blocks differs. This mostly
1594 * indicate that the bitmap is corrupt. So exit
1595 * without claiming the space.
1600 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1606 ext4_mb_check_limits(ac
, e4b
, 1);
1610 * This is a special case for storages like raid5
1611 * we try to find stripe-aligned chunks for stripe-size requests
1612 * XXX should do so at least for multiples of stripe size as well
1614 static void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1615 struct ext4_buddy
*e4b
)
1617 struct super_block
*sb
= ac
->ac_sb
;
1618 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1619 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1620 struct ext4_free_extent ex
;
1621 ext4_fsblk_t first_group_block
;
1626 BUG_ON(sbi
->s_stripe
== 0);
1628 /* find first stripe-aligned block in group */
1629 first_group_block
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
)
1630 + le32_to_cpu(sbi
->s_es
->s_first_data_block
);
1631 a
= first_group_block
+ sbi
->s_stripe
- 1;
1632 do_div(a
, sbi
->s_stripe
);
1633 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1635 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1636 if (!mb_test_bit(i
, bitmap
)) {
1637 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1638 if (max
>= sbi
->s_stripe
) {
1641 ext4_mb_use_best_found(ac
, e4b
);
1649 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1650 ext4_group_t group
, int cr
)
1652 unsigned free
, fragments
;
1654 struct ext4_group_desc
*desc
;
1655 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1657 BUG_ON(cr
< 0 || cr
>= 4);
1658 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp
));
1660 free
= grp
->bb_free
;
1661 fragments
= grp
->bb_fragments
;
1669 BUG_ON(ac
->ac_2order
== 0);
1670 /* If this group is uninitialized, skip it initially */
1671 desc
= ext4_get_group_desc(ac
->ac_sb
, group
, NULL
);
1672 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))
1675 bits
= ac
->ac_sb
->s_blocksize_bits
+ 1;
1676 for (i
= ac
->ac_2order
; i
<= bits
; i
++)
1677 if (grp
->bb_counters
[i
] > 0)
1681 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1685 if (free
>= ac
->ac_g_ex
.fe_len
)
1697 static noinline_for_stack
int
1698 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
1705 struct ext4_sb_info
*sbi
;
1706 struct super_block
*sb
;
1707 struct ext4_buddy e4b
;
1712 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1714 /* first, try the goal */
1715 err
= ext4_mb_find_by_goal(ac
, &e4b
);
1716 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
1719 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
1723 * ac->ac2_order is set only if the fe_len is a power of 2
1724 * if ac2_order is set we also set criteria to 0 so that we
1725 * try exact allocation using buddy.
1727 i
= fls(ac
->ac_g_ex
.fe_len
);
1730 * We search using buddy data only if the order of the request
1731 * is greater than equal to the sbi_s_mb_order2_reqs
1732 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1734 if (i
>= sbi
->s_mb_order2_reqs
) {
1736 * This should tell if fe_len is exactly power of 2
1738 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
1739 ac
->ac_2order
= i
- 1;
1742 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
1743 /* if stream allocation is enabled, use global goal */
1744 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
1745 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
1749 if (size
< sbi
->s_mb_stream_request
&&
1750 (ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1751 /* TBD: may be hot point */
1752 spin_lock(&sbi
->s_md_lock
);
1753 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
1754 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
1755 spin_unlock(&sbi
->s_md_lock
);
1757 /* Let's just scan groups to find more-less suitable blocks */
1758 cr
= ac
->ac_2order
? 0 : 1;
1760 * cr == 0 try to get exact allocation,
1761 * cr == 3 try to get anything
1764 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
1765 ac
->ac_criteria
= cr
;
1767 * searching for the right group start
1768 * from the goal value specified
1770 group
= ac
->ac_g_ex
.fe_group
;
1772 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; group
++, i
++) {
1773 struct ext4_group_info
*grp
;
1774 struct ext4_group_desc
*desc
;
1776 if (group
== EXT4_SB(sb
)->s_groups_count
)
1779 /* quick check to skip empty groups */
1780 grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1781 if (grp
->bb_free
== 0)
1785 * if the group is already init we check whether it is
1786 * a good group and if not we don't load the buddy
1788 if (EXT4_MB_GRP_NEED_INIT(grp
)) {
1790 * we need full data about the group
1791 * to make a good selection
1793 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
1796 ext4_mb_release_desc(&e4b
);
1800 * If the particular group doesn't satisfy our
1801 * criteria we continue with the next group
1803 if (!ext4_mb_good_group(ac
, group
, cr
))
1806 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
1810 ext4_lock_group(sb
, group
);
1811 if (!ext4_mb_good_group(ac
, group
, cr
)) {
1812 /* someone did allocation from this group */
1813 ext4_unlock_group(sb
, group
);
1814 ext4_mb_release_desc(&e4b
);
1818 ac
->ac_groups_scanned
++;
1819 desc
= ext4_get_group_desc(sb
, group
, NULL
);
1820 if (cr
== 0 || (desc
->bg_flags
&
1821 cpu_to_le16(EXT4_BG_BLOCK_UNINIT
) &&
1822 ac
->ac_2order
!= 0))
1823 ext4_mb_simple_scan_group(ac
, &e4b
);
1825 ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
)
1826 ext4_mb_scan_aligned(ac
, &e4b
);
1828 ext4_mb_complex_scan_group(ac
, &e4b
);
1830 ext4_unlock_group(sb
, group
);
1831 ext4_mb_release_desc(&e4b
);
1833 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
1838 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
1839 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1841 * We've been searching too long. Let's try to allocate
1842 * the best chunk we've found so far
1845 ext4_mb_try_best_found(ac
, &e4b
);
1846 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
1848 * Someone more lucky has already allocated it.
1849 * The only thing we can do is just take first
1851 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
1853 ac
->ac_b_ex
.fe_group
= 0;
1854 ac
->ac_b_ex
.fe_start
= 0;
1855 ac
->ac_b_ex
.fe_len
= 0;
1856 ac
->ac_status
= AC_STATUS_CONTINUE
;
1857 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
1859 atomic_inc(&sbi
->s_mb_lost_chunks
);
1867 #ifdef EXT4_MB_HISTORY
1868 struct ext4_mb_proc_session
{
1869 struct ext4_mb_history
*history
;
1870 struct super_block
*sb
;
1875 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session
*s
,
1876 struct ext4_mb_history
*hs
,
1879 if (hs
== s
->history
+ s
->max
)
1881 if (!first
&& hs
== s
->history
+ s
->start
)
1883 while (hs
->orig
.fe_len
== 0) {
1885 if (hs
== s
->history
+ s
->max
)
1887 if (hs
== s
->history
+ s
->start
)
1893 static void *ext4_mb_seq_history_start(struct seq_file
*seq
, loff_t
*pos
)
1895 struct ext4_mb_proc_session
*s
= seq
->private;
1896 struct ext4_mb_history
*hs
;
1900 return SEQ_START_TOKEN
;
1901 hs
= ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
1904 while (--l
&& (hs
= ext4_mb_history_skip_empty(s
, ++hs
, 0)) != NULL
);
1908 static void *ext4_mb_seq_history_next(struct seq_file
*seq
, void *v
,
1911 struct ext4_mb_proc_session
*s
= seq
->private;
1912 struct ext4_mb_history
*hs
= v
;
1915 if (v
== SEQ_START_TOKEN
)
1916 return ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
1918 return ext4_mb_history_skip_empty(s
, ++hs
, 0);
1921 static int ext4_mb_seq_history_show(struct seq_file
*seq
, void *v
)
1923 char buf
[25], buf2
[25], buf3
[25], *fmt
;
1924 struct ext4_mb_history
*hs
= v
;
1926 if (v
== SEQ_START_TOKEN
) {
1927 seq_printf(seq
, "%-5s %-8s %-23s %-23s %-23s %-5s "
1928 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
1929 "pid", "inode", "original", "goal", "result", "found",
1930 "grps", "cr", "flags", "merge", "tail", "broken");
1934 if (hs
->op
== EXT4_MB_HISTORY_ALLOC
) {
1935 fmt
= "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
1936 "%-5u %-5s %-5u %-6u\n";
1937 sprintf(buf2
, "%lu/%d/%u@%u", hs
->result
.fe_group
,
1938 hs
->result
.fe_start
, hs
->result
.fe_len
,
1939 hs
->result
.fe_logical
);
1940 sprintf(buf
, "%lu/%d/%u@%u", hs
->orig
.fe_group
,
1941 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
1942 hs
->orig
.fe_logical
);
1943 sprintf(buf3
, "%lu/%d/%u@%u", hs
->goal
.fe_group
,
1944 hs
->goal
.fe_start
, hs
->goal
.fe_len
,
1945 hs
->goal
.fe_logical
);
1946 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, buf3
, buf2
,
1947 hs
->found
, hs
->groups
, hs
->cr
, hs
->flags
,
1948 hs
->merged
? "M" : "", hs
->tail
,
1949 hs
->buddy
? 1 << hs
->buddy
: 0);
1950 } else if (hs
->op
== EXT4_MB_HISTORY_PREALLOC
) {
1951 fmt
= "%-5u %-8u %-23s %-23s %-23s\n";
1952 sprintf(buf2
, "%lu/%d/%u@%u", hs
->result
.fe_group
,
1953 hs
->result
.fe_start
, hs
->result
.fe_len
,
1954 hs
->result
.fe_logical
);
1955 sprintf(buf
, "%lu/%d/%u@%u", hs
->orig
.fe_group
,
1956 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
1957 hs
->orig
.fe_logical
);
1958 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, "", buf2
);
1959 } else if (hs
->op
== EXT4_MB_HISTORY_DISCARD
) {
1960 sprintf(buf2
, "%lu/%d/%u", hs
->result
.fe_group
,
1961 hs
->result
.fe_start
, hs
->result
.fe_len
);
1962 seq_printf(seq
, "%-5u %-8u %-23s discard\n",
1963 hs
->pid
, hs
->ino
, buf2
);
1964 } else if (hs
->op
== EXT4_MB_HISTORY_FREE
) {
1965 sprintf(buf2
, "%lu/%d/%u", hs
->result
.fe_group
,
1966 hs
->result
.fe_start
, hs
->result
.fe_len
);
1967 seq_printf(seq
, "%-5u %-8u %-23s free\n",
1968 hs
->pid
, hs
->ino
, buf2
);
1973 static void ext4_mb_seq_history_stop(struct seq_file
*seq
, void *v
)
1977 static struct seq_operations ext4_mb_seq_history_ops
= {
1978 .start
= ext4_mb_seq_history_start
,
1979 .next
= ext4_mb_seq_history_next
,
1980 .stop
= ext4_mb_seq_history_stop
,
1981 .show
= ext4_mb_seq_history_show
,
1984 static int ext4_mb_seq_history_open(struct inode
*inode
, struct file
*file
)
1986 struct super_block
*sb
= PDE(inode
)->data
;
1987 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1988 struct ext4_mb_proc_session
*s
;
1992 if (unlikely(sbi
->s_mb_history
== NULL
))
1994 s
= kmalloc(sizeof(*s
), GFP_KERNEL
);
1998 size
= sizeof(struct ext4_mb_history
) * sbi
->s_mb_history_max
;
1999 s
->history
= kmalloc(size
, GFP_KERNEL
);
2000 if (s
->history
== NULL
) {
2005 spin_lock(&sbi
->s_mb_history_lock
);
2006 memcpy(s
->history
, sbi
->s_mb_history
, size
);
2007 s
->max
= sbi
->s_mb_history_max
;
2008 s
->start
= sbi
->s_mb_history_cur
% s
->max
;
2009 spin_unlock(&sbi
->s_mb_history_lock
);
2011 rc
= seq_open(file
, &ext4_mb_seq_history_ops
);
2013 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2023 static int ext4_mb_seq_history_release(struct inode
*inode
, struct file
*file
)
2025 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2026 struct ext4_mb_proc_session
*s
= seq
->private;
2029 return seq_release(inode
, file
);
2032 static ssize_t
ext4_mb_seq_history_write(struct file
*file
,
2033 const char __user
*buffer
,
2034 size_t count
, loff_t
*ppos
)
2036 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2037 struct ext4_mb_proc_session
*s
= seq
->private;
2038 struct super_block
*sb
= s
->sb
;
2042 if (count
>= sizeof(str
)) {
2043 printk(KERN_ERR
"EXT4-fs: %s string too long, max %u bytes\n",
2044 "mb_history", (int)sizeof(str
));
2048 if (copy_from_user(str
, buffer
, count
))
2051 value
= simple_strtol(str
, NULL
, 0);
2054 EXT4_SB(sb
)->s_mb_history_filter
= value
;
2059 static struct file_operations ext4_mb_seq_history_fops
= {
2060 .owner
= THIS_MODULE
,
2061 .open
= ext4_mb_seq_history_open
,
2063 .write
= ext4_mb_seq_history_write
,
2064 .llseek
= seq_lseek
,
2065 .release
= ext4_mb_seq_history_release
,
2068 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2070 struct super_block
*sb
= seq
->private;
2071 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2074 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2078 return (void *) group
;
2081 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2083 struct super_block
*sb
= seq
->private;
2084 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2088 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2091 return (void *) group
;;
2094 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2096 struct super_block
*sb
= seq
->private;
2097 long group
= (long) v
;
2100 struct ext4_buddy e4b
;
2102 struct ext4_group_info info
;
2103 unsigned short counters
[16];
2108 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2109 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2110 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2111 "group", "free", "frags", "first",
2112 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2113 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2115 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2116 sizeof(struct ext4_group_info
);
2117 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2119 seq_printf(seq
, "#%-5lu: I/O error\n", group
);
2122 ext4_lock_group(sb
, group
);
2123 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2124 ext4_unlock_group(sb
, group
);
2125 ext4_mb_release_desc(&e4b
);
2127 seq_printf(seq
, "#%-5lu: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2128 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2129 for (i
= 0; i
<= 13; i
++)
2130 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2131 sg
.info
.bb_counters
[i
] : 0);
2132 seq_printf(seq
, " ]\n");
2137 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2141 static struct seq_operations ext4_mb_seq_groups_ops
= {
2142 .start
= ext4_mb_seq_groups_start
,
2143 .next
= ext4_mb_seq_groups_next
,
2144 .stop
= ext4_mb_seq_groups_stop
,
2145 .show
= ext4_mb_seq_groups_show
,
2148 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2150 struct super_block
*sb
= PDE(inode
)->data
;
2153 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2155 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2162 static struct file_operations ext4_mb_seq_groups_fops
= {
2163 .owner
= THIS_MODULE
,
2164 .open
= ext4_mb_seq_groups_open
,
2166 .llseek
= seq_lseek
,
2167 .release
= seq_release
,
2170 static void ext4_mb_history_release(struct super_block
*sb
)
2172 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2174 if (sbi
->s_proc
!= NULL
) {
2175 remove_proc_entry("mb_groups", sbi
->s_proc
);
2176 remove_proc_entry("mb_history", sbi
->s_proc
);
2178 kfree(sbi
->s_mb_history
);
2181 static void ext4_mb_history_init(struct super_block
*sb
)
2183 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2186 if (sbi
->s_proc
!= NULL
) {
2187 proc_create_data("mb_history", S_IRUGO
, sbi
->s_proc
,
2188 &ext4_mb_seq_history_fops
, sb
);
2189 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2190 &ext4_mb_seq_groups_fops
, sb
);
2193 sbi
->s_mb_history_max
= 1000;
2194 sbi
->s_mb_history_cur
= 0;
2195 spin_lock_init(&sbi
->s_mb_history_lock
);
2196 i
= sbi
->s_mb_history_max
* sizeof(struct ext4_mb_history
);
2197 sbi
->s_mb_history
= kzalloc(i
, GFP_KERNEL
);
2198 /* if we can't allocate history, then we simple won't use it */
2201 static noinline_for_stack
void
2202 ext4_mb_store_history(struct ext4_allocation_context
*ac
)
2204 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2205 struct ext4_mb_history h
;
2207 if (unlikely(sbi
->s_mb_history
== NULL
))
2210 if (!(ac
->ac_op
& sbi
->s_mb_history_filter
))
2214 h
.pid
= current
->pid
;
2215 h
.ino
= ac
->ac_inode
? ac
->ac_inode
->i_ino
: 0;
2216 h
.orig
= ac
->ac_o_ex
;
2217 h
.result
= ac
->ac_b_ex
;
2218 h
.flags
= ac
->ac_flags
;
2219 h
.found
= ac
->ac_found
;
2220 h
.groups
= ac
->ac_groups_scanned
;
2221 h
.cr
= ac
->ac_criteria
;
2222 h
.tail
= ac
->ac_tail
;
2223 h
.buddy
= ac
->ac_buddy
;
2225 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
) {
2226 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
2227 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
2229 h
.goal
= ac
->ac_g_ex
;
2230 h
.result
= ac
->ac_f_ex
;
2233 spin_lock(&sbi
->s_mb_history_lock
);
2234 memcpy(sbi
->s_mb_history
+ sbi
->s_mb_history_cur
, &h
, sizeof(h
));
2235 if (++sbi
->s_mb_history_cur
>= sbi
->s_mb_history_max
)
2236 sbi
->s_mb_history_cur
= 0;
2237 spin_unlock(&sbi
->s_mb_history_lock
);
2241 #define ext4_mb_history_release(sb)
2242 #define ext4_mb_history_init(sb)
2246 /* Create and initialize ext4_group_info data for the given group. */
2247 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2248 struct ext4_group_desc
*desc
)
2252 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2253 struct ext4_group_info
**meta_group_info
;
2256 * First check if this group is the first of a reserved block.
2257 * If it's true, we have to allocate a new table of pointers
2258 * to ext4_group_info structures
2260 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2261 metalen
= sizeof(*meta_group_info
) <<
2262 EXT4_DESC_PER_BLOCK_BITS(sb
);
2263 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2264 if (meta_group_info
== NULL
) {
2265 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2267 goto exit_meta_group_info
;
2269 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2274 * calculate needed size. if change bb_counters size,
2275 * don't forget about ext4_mb_generate_buddy()
2277 len
= offsetof(typeof(**meta_group_info
),
2278 bb_counters
[sb
->s_blocksize_bits
+ 2]);
2281 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2282 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2284 meta_group_info
[i
] = kzalloc(len
, GFP_KERNEL
);
2285 if (meta_group_info
[i
] == NULL
) {
2286 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2287 goto exit_group_info
;
2289 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2290 &(meta_group_info
[i
]->bb_state
));
2293 * initialize bb_free to be able to skip
2294 * empty groups without initialization
2296 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2297 meta_group_info
[i
]->bb_free
=
2298 ext4_free_blocks_after_init(sb
, group
, desc
);
2300 meta_group_info
[i
]->bb_free
=
2301 le16_to_cpu(desc
->bg_free_blocks_count
);
2304 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2305 meta_group_info
[i
]->bb_free_root
.rb_node
= NULL
;;
2309 struct buffer_head
*bh
;
2310 meta_group_info
[i
]->bb_bitmap
=
2311 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2312 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2313 bh
= ext4_read_block_bitmap(sb
, group
);
2315 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2324 /* If a meta_group_info table has been allocated, release it now */
2325 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0)
2326 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2327 exit_meta_group_info
:
2329 } /* ext4_mb_add_groupinfo */
2332 * Add a group to the existing groups.
2333 * This function is used for online resize
2335 int ext4_mb_add_more_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2336 struct ext4_group_desc
*desc
)
2338 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2339 struct inode
*inode
= sbi
->s_buddy_cache
;
2340 int blocks_per_page
;
2346 /* Add group based on group descriptor*/
2347 err
= ext4_mb_add_groupinfo(sb
, group
, desc
);
2352 * Cache pages containing dynamic mb_alloc datas (buddy and bitmap
2353 * datas) are set not up to date so that they will be re-initilaized
2354 * during the next call to ext4_mb_load_buddy
2357 /* Set buddy page as not up to date */
2358 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
2360 pnum
= block
/ blocks_per_page
;
2361 page
= find_get_page(inode
->i_mapping
, pnum
);
2363 ClearPageUptodate(page
);
2364 page_cache_release(page
);
2367 /* Set bitmap page as not up to date */
2369 pnum
= block
/ blocks_per_page
;
2370 page
= find_get_page(inode
->i_mapping
, pnum
);
2372 ClearPageUptodate(page
);
2373 page_cache_release(page
);
2380 * Update an existing group.
2381 * This function is used for online resize
2383 void ext4_mb_update_group_info(struct ext4_group_info
*grp
, ext4_grpblk_t add
)
2385 grp
->bb_free
+= add
;
2388 static int ext4_mb_init_backend(struct super_block
*sb
)
2392 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2393 struct ext4_super_block
*es
= sbi
->s_es
;
2394 int num_meta_group_infos
;
2395 int num_meta_group_infos_max
;
2397 struct ext4_group_info
**meta_group_info
;
2398 struct ext4_group_desc
*desc
;
2400 /* This is the number of blocks used by GDT */
2401 num_meta_group_infos
= (sbi
->s_groups_count
+ EXT4_DESC_PER_BLOCK(sb
) -
2402 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2405 * This is the total number of blocks used by GDT including
2406 * the number of reserved blocks for GDT.
2407 * The s_group_info array is allocated with this value
2408 * to allow a clean online resize without a complex
2409 * manipulation of pointer.
2410 * The drawback is the unused memory when no resize
2411 * occurs but it's very low in terms of pages
2412 * (see comments below)
2413 * Need to handle this properly when META_BG resizing is allowed
2415 num_meta_group_infos_max
= num_meta_group_infos
+
2416 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2419 * array_size is the size of s_group_info array. We round it
2420 * to the next power of two because this approximation is done
2421 * internally by kmalloc so we can have some more memory
2422 * for free here (e.g. may be used for META_BG resize).
2425 while (array_size
< sizeof(*sbi
->s_group_info
) *
2426 num_meta_group_infos_max
)
2427 array_size
= array_size
<< 1;
2428 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2429 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2430 * So a two level scheme suffices for now. */
2431 sbi
->s_group_info
= kmalloc(array_size
, GFP_KERNEL
);
2432 if (sbi
->s_group_info
== NULL
) {
2433 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2436 sbi
->s_buddy_cache
= new_inode(sb
);
2437 if (sbi
->s_buddy_cache
== NULL
) {
2438 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2441 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2443 metalen
= sizeof(*meta_group_info
) << EXT4_DESC_PER_BLOCK_BITS(sb
);
2444 for (i
= 0; i
< num_meta_group_infos
; i
++) {
2445 if ((i
+ 1) == num_meta_group_infos
)
2446 metalen
= sizeof(*meta_group_info
) *
2447 (sbi
->s_groups_count
-
2448 (i
<< EXT4_DESC_PER_BLOCK_BITS(sb
)));
2449 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2450 if (meta_group_info
== NULL
) {
2451 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2455 sbi
->s_group_info
[i
] = meta_group_info
;
2458 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2459 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2462 "EXT4-fs: can't read descriptor %lu\n", i
);
2465 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2473 kfree(ext4_get_group_info(sb
, i
));
2474 i
= num_meta_group_infos
;
2477 kfree(sbi
->s_group_info
[i
]);
2478 iput(sbi
->s_buddy_cache
);
2480 kfree(sbi
->s_group_info
);
2484 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2486 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2492 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned short);
2494 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2495 if (sbi
->s_mb_offsets
== NULL
) {
2499 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned int);
2500 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2501 if (sbi
->s_mb_maxs
== NULL
) {
2502 kfree(sbi
->s_mb_maxs
);
2506 /* order 0 is regular bitmap */
2507 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2508 sbi
->s_mb_offsets
[0] = 0;
2512 max
= sb
->s_blocksize
<< 2;
2514 sbi
->s_mb_offsets
[i
] = offset
;
2515 sbi
->s_mb_maxs
[i
] = max
;
2516 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2519 } while (i
<= sb
->s_blocksize_bits
+ 1);
2521 /* init file for buddy data */
2522 ret
= ext4_mb_init_backend(sb
);
2524 kfree(sbi
->s_mb_offsets
);
2525 kfree(sbi
->s_mb_maxs
);
2529 spin_lock_init(&sbi
->s_md_lock
);
2530 spin_lock_init(&sbi
->s_bal_lock
);
2532 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2533 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2534 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2535 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2536 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2537 sbi
->s_mb_history_filter
= EXT4_MB_HISTORY_DEFAULT
;
2538 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2540 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2541 if (sbi
->s_locality_groups
== NULL
) {
2542 kfree(sbi
->s_mb_offsets
);
2543 kfree(sbi
->s_mb_maxs
);
2546 for_each_possible_cpu(i
) {
2547 struct ext4_locality_group
*lg
;
2548 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2549 mutex_init(&lg
->lg_mutex
);
2550 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2551 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2552 spin_lock_init(&lg
->lg_prealloc_lock
);
2555 ext4_mb_init_per_dev_proc(sb
);
2556 ext4_mb_history_init(sb
);
2559 sbi
->s_journal
->j_commit_callback
= release_blocks_on_commit
;
2561 printk(KERN_INFO
"EXT4-fs: mballoc enabled\n");
2565 /* need to called with ext4 group lock (ext4_lock_group) */
2566 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2568 struct ext4_prealloc_space
*pa
;
2569 struct list_head
*cur
, *tmp
;
2572 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2573 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2574 list_del(&pa
->pa_group_list
);
2576 kmem_cache_free(ext4_pspace_cachep
, pa
);
2579 mb_debug("mballoc: %u PAs left\n", count
);
2583 int ext4_mb_release(struct super_block
*sb
)
2586 int num_meta_group_infos
;
2587 struct ext4_group_info
*grinfo
;
2588 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2590 if (sbi
->s_group_info
) {
2591 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2592 grinfo
= ext4_get_group_info(sb
, i
);
2594 kfree(grinfo
->bb_bitmap
);
2596 ext4_lock_group(sb
, i
);
2597 ext4_mb_cleanup_pa(grinfo
);
2598 ext4_unlock_group(sb
, i
);
2601 num_meta_group_infos
= (sbi
->s_groups_count
+
2602 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2603 EXT4_DESC_PER_BLOCK_BITS(sb
);
2604 for (i
= 0; i
< num_meta_group_infos
; i
++)
2605 kfree(sbi
->s_group_info
[i
]);
2606 kfree(sbi
->s_group_info
);
2608 kfree(sbi
->s_mb_offsets
);
2609 kfree(sbi
->s_mb_maxs
);
2610 if (sbi
->s_buddy_cache
)
2611 iput(sbi
->s_buddy_cache
);
2612 if (sbi
->s_mb_stats
) {
2614 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2615 atomic_read(&sbi
->s_bal_allocated
),
2616 atomic_read(&sbi
->s_bal_reqs
),
2617 atomic_read(&sbi
->s_bal_success
));
2619 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2620 "%u 2^N hits, %u breaks, %u lost\n",
2621 atomic_read(&sbi
->s_bal_ex_scanned
),
2622 atomic_read(&sbi
->s_bal_goals
),
2623 atomic_read(&sbi
->s_bal_2orders
),
2624 atomic_read(&sbi
->s_bal_breaks
),
2625 atomic_read(&sbi
->s_mb_lost_chunks
));
2627 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2628 sbi
->s_mb_buddies_generated
++,
2629 sbi
->s_mb_generation_time
);
2631 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2632 atomic_read(&sbi
->s_mb_preallocated
),
2633 atomic_read(&sbi
->s_mb_discarded
));
2636 free_percpu(sbi
->s_locality_groups
);
2637 ext4_mb_history_release(sb
);
2638 ext4_mb_destroy_per_dev_proc(sb
);
2644 * This function is called by the jbd2 layer once the commit has finished,
2645 * so we know we can free the blocks that were released with that commit.
2647 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
)
2649 struct super_block
*sb
= journal
->j_private
;
2650 struct ext4_buddy e4b
;
2651 struct ext4_group_info
*db
;
2652 int err
, count
= 0, count2
= 0;
2653 struct ext4_free_data
*entry
;
2654 ext4_fsblk_t discard_block
;
2655 struct list_head
*l
, *ltmp
;
2657 list_for_each_safe(l
, ltmp
, &txn
->t_private_list
) {
2658 entry
= list_entry(l
, struct ext4_free_data
, list
);
2660 mb_debug("gonna free %u blocks in group %lu (0x%p):",
2661 entry
->count
, entry
->group
, entry
);
2663 err
= ext4_mb_load_buddy(sb
, entry
->group
, &e4b
);
2664 /* we expect to find existing buddy because it's pinned */
2668 /* there are blocks to put in buddy to make them really free */
2669 count
+= entry
->count
;
2671 ext4_lock_group(sb
, entry
->group
);
2672 /* Take it out of per group rb tree */
2673 rb_erase(&entry
->node
, &(db
->bb_free_root
));
2674 mb_free_blocks(NULL
, &e4b
, entry
->start_blk
, entry
->count
);
2676 if (!db
->bb_free_root
.rb_node
) {
2677 /* No more items in the per group rb tree
2678 * balance refcounts from ext4_mb_free_metadata()
2680 page_cache_release(e4b
.bd_buddy_page
);
2681 page_cache_release(e4b
.bd_bitmap_page
);
2683 ext4_unlock_group(sb
, entry
->group
);
2684 discard_block
= (ext4_fsblk_t
) entry
->group
* EXT4_BLOCKS_PER_GROUP(sb
)
2686 + le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
2687 trace_mark(ext4_discard_blocks
, "dev %s blk %llu count %u", sb
->s_id
,
2688 (unsigned long long) discard_block
, entry
->count
);
2689 sb_issue_discard(sb
, discard_block
, entry
->count
);
2691 kmem_cache_free(ext4_free_ext_cachep
, entry
);
2692 ext4_mb_release_desc(&e4b
);
2695 mb_debug("freed %u blocks in %u structures\n", count
, count2
);
2698 #define EXT4_MB_STATS_NAME "stats"
2699 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2700 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2701 #define EXT4_MB_ORDER2_REQ "order2_req"
2702 #define EXT4_MB_STREAM_REQ "stream_req"
2703 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2705 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
)
2707 #ifdef CONFIG_PROC_FS
2708 mode_t mode
= S_IFREG
| S_IRUGO
| S_IWUSR
;
2709 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2710 struct proc_dir_entry
*proc
;
2712 if (sbi
->s_proc
== NULL
)
2715 EXT4_PROC_HANDLER(EXT4_MB_STATS_NAME
, mb_stats
);
2716 EXT4_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME
, mb_max_to_scan
);
2717 EXT4_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME
, mb_min_to_scan
);
2718 EXT4_PROC_HANDLER(EXT4_MB_ORDER2_REQ
, mb_order2_reqs
);
2719 EXT4_PROC_HANDLER(EXT4_MB_STREAM_REQ
, mb_stream_request
);
2720 EXT4_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC
, mb_group_prealloc
);
2724 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2725 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2726 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2727 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2728 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2729 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2736 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
)
2738 #ifdef CONFIG_PROC_FS
2739 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2741 if (sbi
->s_proc
== NULL
)
2744 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2745 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2746 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2747 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2748 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2749 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2754 int __init
init_ext4_mballoc(void)
2756 ext4_pspace_cachep
=
2757 kmem_cache_create("ext4_prealloc_space",
2758 sizeof(struct ext4_prealloc_space
),
2759 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2760 if (ext4_pspace_cachep
== NULL
)
2764 kmem_cache_create("ext4_alloc_context",
2765 sizeof(struct ext4_allocation_context
),
2766 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2767 if (ext4_ac_cachep
== NULL
) {
2768 kmem_cache_destroy(ext4_pspace_cachep
);
2772 ext4_free_ext_cachep
=
2773 kmem_cache_create("ext4_free_block_extents",
2774 sizeof(struct ext4_free_data
),
2775 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2776 if (ext4_free_ext_cachep
== NULL
) {
2777 kmem_cache_destroy(ext4_pspace_cachep
);
2778 kmem_cache_destroy(ext4_ac_cachep
);
2784 void exit_ext4_mballoc(void)
2786 /* XXX: synchronize_rcu(); */
2787 kmem_cache_destroy(ext4_pspace_cachep
);
2788 kmem_cache_destroy(ext4_ac_cachep
);
2789 kmem_cache_destroy(ext4_free_ext_cachep
);
2794 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2795 * Returns 0 if success or error code
2797 static noinline_for_stack
int
2798 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2799 handle_t
*handle
, unsigned long reserv_blks
)
2801 struct buffer_head
*bitmap_bh
= NULL
;
2802 struct ext4_super_block
*es
;
2803 struct ext4_group_desc
*gdp
;
2804 struct buffer_head
*gdp_bh
;
2805 struct ext4_sb_info
*sbi
;
2806 struct super_block
*sb
;
2810 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2811 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2819 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2823 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2828 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2832 ext4_debug("using block group %lu(%d)\n", ac
->ac_b_ex
.fe_group
,
2833 gdp
->bg_free_blocks_count
);
2835 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2839 block
= ac
->ac_b_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
2840 + ac
->ac_b_ex
.fe_start
2841 + le32_to_cpu(es
->s_first_data_block
);
2843 len
= ac
->ac_b_ex
.fe_len
;
2844 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, len
) ||
2845 in_range(ext4_inode_bitmap(sb
, gdp
), block
, len
) ||
2846 in_range(block
, ext4_inode_table(sb
, gdp
),
2847 EXT4_SB(sb
)->s_itb_per_group
) ||
2848 in_range(block
+ len
- 1, ext4_inode_table(sb
, gdp
),
2849 EXT4_SB(sb
)->s_itb_per_group
)) {
2850 ext4_error(sb
, __func__
,
2851 "Allocating block in system zone - block = %llu",
2853 /* File system mounted not to panic on error
2854 * Fix the bitmap and repeat the block allocation
2855 * We leak some of the blocks here.
2857 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
),
2858 bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2859 ac
->ac_b_ex
.fe_len
);
2860 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2865 #ifdef AGGRESSIVE_CHECK
2868 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2869 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2870 bitmap_bh
->b_data
));
2874 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
), bitmap_bh
->b_data
,
2875 ac
->ac_b_ex
.fe_start
, ac
->ac_b_ex
.fe_len
);
2877 spin_lock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
2878 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2879 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2880 gdp
->bg_free_blocks_count
=
2881 cpu_to_le16(ext4_free_blocks_after_init(sb
,
2882 ac
->ac_b_ex
.fe_group
,
2885 le16_add_cpu(&gdp
->bg_free_blocks_count
, -ac
->ac_b_ex
.fe_len
);
2886 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
2887 spin_unlock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
2888 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
2890 * Now reduce the dirty block count also. Should not go negative
2892 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2893 /* release all the reserved blocks if non delalloc */
2894 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
, reserv_blks
);
2896 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
,
2897 ac
->ac_b_ex
.fe_len
);
2899 if (sbi
->s_log_groups_per_flex
) {
2900 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2901 ac
->ac_b_ex
.fe_group
);
2902 spin_lock(sb_bgl_lock(sbi
, flex_group
));
2903 sbi
->s_flex_groups
[flex_group
].free_blocks
-= ac
->ac_b_ex
.fe_len
;
2904 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
2907 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2910 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
2919 * here we normalize request for locality group
2920 * Group request are normalized to s_strip size if we set the same via mount
2921 * option. If not we set it to s_mb_group_prealloc which can be configured via
2922 * /proc/fs/ext4/<partition>/group_prealloc
2924 * XXX: should we try to preallocate more than the group has now?
2926 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2928 struct super_block
*sb
= ac
->ac_sb
;
2929 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2932 if (EXT4_SB(sb
)->s_stripe
)
2933 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
2935 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2936 mb_debug("#%u: goal %u blocks for locality group\n",
2937 current
->pid
, ac
->ac_g_ex
.fe_len
);
2941 * Normalization means making request better in terms of
2942 * size and alignment
2944 static noinline_for_stack
void
2945 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
2946 struct ext4_allocation_request
*ar
)
2950 loff_t size
, orig_size
, start_off
;
2951 ext4_lblk_t start
, orig_start
;
2952 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
2953 struct ext4_prealloc_space
*pa
;
2955 /* do normalize only data requests, metadata requests
2956 do not need preallocation */
2957 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
2960 /* sometime caller may want exact blocks */
2961 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2964 /* caller may indicate that preallocation isn't
2965 * required (it's a tail, for example) */
2966 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
2969 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
2970 ext4_mb_normalize_group_request(ac
);
2974 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
2976 /* first, let's learn actual file size
2977 * given current request is allocated */
2978 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
2979 size
= size
<< bsbits
;
2980 if (size
< i_size_read(ac
->ac_inode
))
2981 size
= i_size_read(ac
->ac_inode
);
2983 /* max size of free chunks */
2986 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2987 (req <= (size) || max <= (chunk_size))
2989 /* first, try to predict filesize */
2990 /* XXX: should this table be tunable? */
2992 if (size
<= 16 * 1024) {
2994 } else if (size
<= 32 * 1024) {
2996 } else if (size
<= 64 * 1024) {
2998 } else if (size
<= 128 * 1024) {
3000 } else if (size
<= 256 * 1024) {
3002 } else if (size
<= 512 * 1024) {
3004 } else if (size
<= 1024 * 1024) {
3006 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3007 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3008 (21 - bsbits
)) << 21;
3009 size
= 2 * 1024 * 1024;
3010 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3011 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3012 (22 - bsbits
)) << 22;
3013 size
= 4 * 1024 * 1024;
3014 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3015 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3016 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3017 (23 - bsbits
)) << 23;
3018 size
= 8 * 1024 * 1024;
3020 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3021 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3023 orig_size
= size
= size
>> bsbits
;
3024 orig_start
= start
= start_off
>> bsbits
;
3026 /* don't cover already allocated blocks in selected range */
3027 if (ar
->pleft
&& start
<= ar
->lleft
) {
3028 size
-= ar
->lleft
+ 1 - start
;
3029 start
= ar
->lleft
+ 1;
3031 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3032 size
-= start
+ size
- ar
->lright
;
3036 /* check we don't cross already preallocated blocks */
3038 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3039 unsigned long pa_end
;
3043 spin_lock(&pa
->pa_lock
);
3044 if (pa
->pa_deleted
) {
3045 spin_unlock(&pa
->pa_lock
);
3049 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3051 /* PA must not overlap original request */
3052 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3053 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3055 /* skip PA normalized request doesn't overlap with */
3056 if (pa
->pa_lstart
>= end
) {
3057 spin_unlock(&pa
->pa_lock
);
3060 if (pa_end
<= start
) {
3061 spin_unlock(&pa
->pa_lock
);
3064 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3066 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3067 BUG_ON(pa_end
< start
);
3071 if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3072 BUG_ON(pa
->pa_lstart
> end
);
3073 end
= pa
->pa_lstart
;
3075 spin_unlock(&pa
->pa_lock
);
3080 /* XXX: extra loop to check we really don't overlap preallocations */
3082 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3083 unsigned long pa_end
;
3084 spin_lock(&pa
->pa_lock
);
3085 if (pa
->pa_deleted
== 0) {
3086 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3087 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3089 spin_unlock(&pa
->pa_lock
);
3093 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3094 start
> ac
->ac_o_ex
.fe_logical
) {
3095 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3096 (unsigned long) start
, (unsigned long) size
,
3097 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3099 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3100 start
> ac
->ac_o_ex
.fe_logical
);
3101 BUG_ON(size
<= 0 || size
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3103 /* now prepare goal request */
3105 /* XXX: is it better to align blocks WRT to logical
3106 * placement or satisfy big request as is */
3107 ac
->ac_g_ex
.fe_logical
= start
;
3108 ac
->ac_g_ex
.fe_len
= size
;
3110 /* define goal start in order to merge */
3111 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3112 /* merge to the right */
3113 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3114 &ac
->ac_f_ex
.fe_group
,
3115 &ac
->ac_f_ex
.fe_start
);
3116 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3118 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3119 /* merge to the left */
3120 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3121 &ac
->ac_f_ex
.fe_group
,
3122 &ac
->ac_f_ex
.fe_start
);
3123 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3126 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3127 (unsigned) orig_size
, (unsigned) start
);
3130 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3132 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3134 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3135 atomic_inc(&sbi
->s_bal_reqs
);
3136 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3137 if (ac
->ac_o_ex
.fe_len
>= ac
->ac_g_ex
.fe_len
)
3138 atomic_inc(&sbi
->s_bal_success
);
3139 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3140 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3141 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3142 atomic_inc(&sbi
->s_bal_goals
);
3143 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3144 atomic_inc(&sbi
->s_bal_breaks
);
3147 ext4_mb_store_history(ac
);
3151 * use blocks preallocated to inode
3153 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3154 struct ext4_prealloc_space
*pa
)
3160 /* found preallocated blocks, use them */
3161 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3162 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3164 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3165 &ac
->ac_b_ex
.fe_start
);
3166 ac
->ac_b_ex
.fe_len
= len
;
3167 ac
->ac_status
= AC_STATUS_FOUND
;
3170 BUG_ON(start
< pa
->pa_pstart
);
3171 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3172 BUG_ON(pa
->pa_free
< len
);
3175 mb_debug("use %llu/%u from inode pa %p\n", start
, len
, pa
);
3179 * use blocks preallocated to locality group
3181 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3182 struct ext4_prealloc_space
*pa
)
3184 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3186 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3187 &ac
->ac_b_ex
.fe_group
,
3188 &ac
->ac_b_ex
.fe_start
);
3189 ac
->ac_b_ex
.fe_len
= len
;
3190 ac
->ac_status
= AC_STATUS_FOUND
;
3193 /* we don't correct pa_pstart or pa_plen here to avoid
3194 * possible race when the group is being loaded concurrently
3195 * instead we correct pa later, after blocks are marked
3196 * in on-disk bitmap -- see ext4_mb_release_context()
3197 * Other CPUs are prevented from allocating from this pa by lg_mutex
3199 mb_debug("use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3203 * Return the prealloc space that have minimal distance
3204 * from the goal block. @cpa is the prealloc
3205 * space that is having currently known minimal distance
3206 * from the goal block.
3208 static struct ext4_prealloc_space
*
3209 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3210 struct ext4_prealloc_space
*pa
,
3211 struct ext4_prealloc_space
*cpa
)
3213 ext4_fsblk_t cur_distance
, new_distance
;
3216 atomic_inc(&pa
->pa_count
);
3219 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3220 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3222 if (cur_distance
< new_distance
)
3225 /* drop the previous reference */
3226 atomic_dec(&cpa
->pa_count
);
3227 atomic_inc(&pa
->pa_count
);
3232 * search goal blocks in preallocated space
3234 static noinline_for_stack
int
3235 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3238 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3239 struct ext4_locality_group
*lg
;
3240 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3241 ext4_fsblk_t goal_block
;
3243 /* only data can be preallocated */
3244 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3247 /* first, try per-file preallocation */
3249 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3251 /* all fields in this condition don't change,
3252 * so we can skip locking for them */
3253 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3254 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3257 /* found preallocated blocks, use them */
3258 spin_lock(&pa
->pa_lock
);
3259 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3260 atomic_inc(&pa
->pa_count
);
3261 ext4_mb_use_inode_pa(ac
, pa
);
3262 spin_unlock(&pa
->pa_lock
);
3263 ac
->ac_criteria
= 10;
3267 spin_unlock(&pa
->pa_lock
);
3271 /* can we use group allocation? */
3272 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3275 /* inode may have no locality group for some reason */
3279 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3280 if (order
> PREALLOC_TB_SIZE
- 1)
3281 /* The max size of hash table is PREALLOC_TB_SIZE */
3282 order
= PREALLOC_TB_SIZE
- 1;
3284 goal_block
= ac
->ac_g_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
) +
3285 ac
->ac_g_ex
.fe_start
+
3286 le32_to_cpu(EXT4_SB(ac
->ac_sb
)->s_es
->s_first_data_block
);
3288 * search for the prealloc space that is having
3289 * minimal distance from the goal block.
3291 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3293 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3295 spin_lock(&pa
->pa_lock
);
3296 if (pa
->pa_deleted
== 0 &&
3297 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3299 cpa
= ext4_mb_check_group_pa(goal_block
,
3302 spin_unlock(&pa
->pa_lock
);
3307 ext4_mb_use_group_pa(ac
, cpa
);
3308 ac
->ac_criteria
= 20;
3315 * the function goes through all preallocation in this group and marks them
3316 * used in in-core bitmap. buddy must be generated from this bitmap
3317 * Need to be called with ext4 group lock (ext4_lock_group)
3319 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3322 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3323 struct ext4_prealloc_space
*pa
;
3324 struct list_head
*cur
;
3325 ext4_group_t groupnr
;
3326 ext4_grpblk_t start
;
3327 int preallocated
= 0;
3331 /* all form of preallocation discards first load group,
3332 * so the only competing code is preallocation use.
3333 * we don't need any locking here
3334 * notice we do NOT ignore preallocations with pa_deleted
3335 * otherwise we could leave used blocks available for
3336 * allocation in buddy when concurrent ext4_mb_put_pa()
3337 * is dropping preallocation
3339 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3340 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3341 spin_lock(&pa
->pa_lock
);
3342 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3345 spin_unlock(&pa
->pa_lock
);
3346 if (unlikely(len
== 0))
3348 BUG_ON(groupnr
!= group
);
3349 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3350 bitmap
, start
, len
);
3351 preallocated
+= len
;
3354 mb_debug("prellocated %u for group %lu\n", preallocated
, group
);
3357 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3359 struct ext4_prealloc_space
*pa
;
3360 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3361 kmem_cache_free(ext4_pspace_cachep
, pa
);
3365 * drops a reference to preallocated space descriptor
3366 * if this was the last reference and the space is consumed
3368 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3369 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3373 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3376 /* in this short window concurrent discard can set pa_deleted */
3377 spin_lock(&pa
->pa_lock
);
3378 if (pa
->pa_deleted
== 1) {
3379 spin_unlock(&pa
->pa_lock
);
3384 spin_unlock(&pa
->pa_lock
);
3386 /* -1 is to protect from crossing allocation group */
3387 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
- 1, &grp
, NULL
);
3392 * P1 (buddy init) P2 (regular allocation)
3393 * find block B in PA
3394 * copy on-disk bitmap to buddy
3395 * mark B in on-disk bitmap
3396 * drop PA from group
3397 * mark all PAs in buddy
3399 * thus, P1 initializes buddy with B available. to prevent this
3400 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3403 ext4_lock_group(sb
, grp
);
3404 list_del(&pa
->pa_group_list
);
3405 ext4_unlock_group(sb
, grp
);
3407 spin_lock(pa
->pa_obj_lock
);
3408 list_del_rcu(&pa
->pa_inode_list
);
3409 spin_unlock(pa
->pa_obj_lock
);
3411 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3415 * creates new preallocated space for given inode
3417 static noinline_for_stack
int
3418 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3420 struct super_block
*sb
= ac
->ac_sb
;
3421 struct ext4_prealloc_space
*pa
;
3422 struct ext4_group_info
*grp
;
3423 struct ext4_inode_info
*ei
;
3425 /* preallocate only when found space is larger then requested */
3426 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3427 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3428 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3430 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3434 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3440 /* we can't allocate as much as normalizer wants.
3441 * so, found space must get proper lstart
3442 * to cover original request */
3443 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3444 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3446 /* we're limited by original request in that
3447 * logical block must be covered any way
3448 * winl is window we can move our chunk within */
3449 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3451 /* also, we should cover whole original request */
3452 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3454 /* the smallest one defines real window */
3455 win
= min(winl
, wins
);
3457 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3458 if (offs
&& offs
< win
)
3461 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3462 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3463 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3466 /* preallocation can change ac_b_ex, thus we store actually
3467 * allocated blocks for history */
3468 ac
->ac_f_ex
= ac
->ac_b_ex
;
3470 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3471 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3472 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3473 pa
->pa_free
= pa
->pa_len
;
3474 atomic_set(&pa
->pa_count
, 1);
3475 spin_lock_init(&pa
->pa_lock
);
3479 mb_debug("new inode pa %p: %llu/%u for %u\n", pa
,
3480 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3482 ext4_mb_use_inode_pa(ac
, pa
);
3483 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3485 ei
= EXT4_I(ac
->ac_inode
);
3486 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3488 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3489 pa
->pa_inode
= ac
->ac_inode
;
3491 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3492 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3493 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3495 spin_lock(pa
->pa_obj_lock
);
3496 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3497 spin_unlock(pa
->pa_obj_lock
);
3503 * creates new preallocated space for locality group inodes belongs to
3505 static noinline_for_stack
int
3506 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3508 struct super_block
*sb
= ac
->ac_sb
;
3509 struct ext4_locality_group
*lg
;
3510 struct ext4_prealloc_space
*pa
;
3511 struct ext4_group_info
*grp
;
3513 /* preallocate only when found space is larger then requested */
3514 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3515 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3516 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3518 BUG_ON(ext4_pspace_cachep
== NULL
);
3519 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3523 /* preallocation can change ac_b_ex, thus we store actually
3524 * allocated blocks for history */
3525 ac
->ac_f_ex
= ac
->ac_b_ex
;
3527 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3528 pa
->pa_lstart
= pa
->pa_pstart
;
3529 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3530 pa
->pa_free
= pa
->pa_len
;
3531 atomic_set(&pa
->pa_count
, 1);
3532 spin_lock_init(&pa
->pa_lock
);
3533 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3537 mb_debug("new group pa %p: %llu/%u for %u\n", pa
,
3538 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3540 ext4_mb_use_group_pa(ac
, pa
);
3541 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3543 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3547 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3548 pa
->pa_inode
= NULL
;
3550 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3551 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3552 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3555 * We will later add the new pa to the right bucket
3556 * after updating the pa_free in ext4_mb_release_context
3561 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3565 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3566 err
= ext4_mb_new_group_pa(ac
);
3568 err
= ext4_mb_new_inode_pa(ac
);
3573 * finds all unused blocks in on-disk bitmap, frees them in
3574 * in-core bitmap and buddy.
3575 * @pa must be unlinked from inode and group lists, so that
3576 * nobody else can find/use it.
3577 * the caller MUST hold group/inode locks.
3578 * TODO: optimize the case when there are no in-core structures yet
3580 static noinline_for_stack
int
3581 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3582 struct ext4_prealloc_space
*pa
,
3583 struct ext4_allocation_context
*ac
)
3585 struct super_block
*sb
= e4b
->bd_sb
;
3586 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3595 BUG_ON(pa
->pa_deleted
== 0);
3596 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3597 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3598 end
= bit
+ pa
->pa_len
;
3602 ac
->ac_inode
= pa
->pa_inode
;
3603 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3607 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3610 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3611 start
= group
* EXT4_BLOCKS_PER_GROUP(sb
) + bit
+
3612 le32_to_cpu(sbi
->s_es
->s_first_data_block
);
3613 mb_debug(" free preallocated %u/%u in group %u\n",
3614 (unsigned) start
, (unsigned) next
- bit
,
3619 ac
->ac_b_ex
.fe_group
= group
;
3620 ac
->ac_b_ex
.fe_start
= bit
;
3621 ac
->ac_b_ex
.fe_len
= next
- bit
;
3622 ac
->ac_b_ex
.fe_logical
= 0;
3623 ext4_mb_store_history(ac
);
3626 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3629 if (free
!= pa
->pa_free
) {
3630 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3631 pa
, (unsigned long) pa
->pa_lstart
,
3632 (unsigned long) pa
->pa_pstart
,
3633 (unsigned long) pa
->pa_len
);
3634 ext4_error(sb
, __func__
, "free %u, pa_free %u",
3637 * pa is already deleted so we use the value obtained
3638 * from the bitmap and continue.
3641 atomic_add(free
, &sbi
->s_mb_discarded
);
3646 static noinline_for_stack
int
3647 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3648 struct ext4_prealloc_space
*pa
,
3649 struct ext4_allocation_context
*ac
)
3651 struct super_block
*sb
= e4b
->bd_sb
;
3656 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3658 BUG_ON(pa
->pa_deleted
== 0);
3659 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3660 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3661 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3662 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3666 ac
->ac_inode
= NULL
;
3667 ac
->ac_b_ex
.fe_group
= group
;
3668 ac
->ac_b_ex
.fe_start
= bit
;
3669 ac
->ac_b_ex
.fe_len
= pa
->pa_len
;
3670 ac
->ac_b_ex
.fe_logical
= 0;
3671 ext4_mb_store_history(ac
);
3678 * releases all preallocations in given group
3680 * first, we need to decide discard policy:
3681 * - when do we discard
3683 * - how many do we discard
3684 * 1) how many requested
3686 static noinline_for_stack
int
3687 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3688 ext4_group_t group
, int needed
)
3690 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3691 struct buffer_head
*bitmap_bh
= NULL
;
3692 struct ext4_prealloc_space
*pa
, *tmp
;
3693 struct ext4_allocation_context
*ac
;
3694 struct list_head list
;
3695 struct ext4_buddy e4b
;
3700 mb_debug("discard preallocation for group %lu\n", group
);
3702 if (list_empty(&grp
->bb_prealloc_list
))
3705 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3706 if (bitmap_bh
== NULL
) {
3707 ext4_error(sb
, __func__
, "Error in reading block "
3708 "bitmap for %lu", group
);
3712 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3714 ext4_error(sb
, __func__
, "Error in loading buddy "
3715 "information for %lu", group
);
3721 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3723 INIT_LIST_HEAD(&list
);
3724 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3726 ext4_lock_group(sb
, group
);
3727 list_for_each_entry_safe(pa
, tmp
,
3728 &grp
->bb_prealloc_list
, pa_group_list
) {
3729 spin_lock(&pa
->pa_lock
);
3730 if (atomic_read(&pa
->pa_count
)) {
3731 spin_unlock(&pa
->pa_lock
);
3735 if (pa
->pa_deleted
) {
3736 spin_unlock(&pa
->pa_lock
);
3740 /* seems this one can be freed ... */
3743 /* we can trust pa_free ... */
3744 free
+= pa
->pa_free
;
3746 spin_unlock(&pa
->pa_lock
);
3748 list_del(&pa
->pa_group_list
);
3749 list_add(&pa
->u
.pa_tmp_list
, &list
);
3752 /* if we still need more blocks and some PAs were used, try again */
3753 if (free
< needed
&& busy
) {
3755 ext4_unlock_group(sb
, group
);
3757 * Yield the CPU here so that we don't get soft lockup
3758 * in non preempt case.
3764 /* found anything to free? */
3765 if (list_empty(&list
)) {
3770 /* now free all selected PAs */
3771 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3773 /* remove from object (inode or locality group) */
3774 spin_lock(pa
->pa_obj_lock
);
3775 list_del_rcu(&pa
->pa_inode_list
);
3776 spin_unlock(pa
->pa_obj_lock
);
3779 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
3781 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3783 list_del(&pa
->u
.pa_tmp_list
);
3784 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3788 ext4_unlock_group(sb
, group
);
3790 kmem_cache_free(ext4_ac_cachep
, ac
);
3791 ext4_mb_release_desc(&e4b
);
3797 * releases all non-used preallocated blocks for given inode
3799 * It's important to discard preallocations under i_data_sem
3800 * We don't want another block to be served from the prealloc
3801 * space when we are discarding the inode prealloc space.
3803 * FIXME!! Make sure it is valid at all the call sites
3805 void ext4_discard_preallocations(struct inode
*inode
)
3807 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3808 struct super_block
*sb
= inode
->i_sb
;
3809 struct buffer_head
*bitmap_bh
= NULL
;
3810 struct ext4_prealloc_space
*pa
, *tmp
;
3811 struct ext4_allocation_context
*ac
;
3812 ext4_group_t group
= 0;
3813 struct list_head list
;
3814 struct ext4_buddy e4b
;
3817 if (!S_ISREG(inode
->i_mode
)) {
3818 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3822 mb_debug("discard preallocation for inode %lu\n", inode
->i_ino
);
3824 INIT_LIST_HEAD(&list
);
3826 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3828 /* first, collect all pa's in the inode */
3829 spin_lock(&ei
->i_prealloc_lock
);
3830 while (!list_empty(&ei
->i_prealloc_list
)) {
3831 pa
= list_entry(ei
->i_prealloc_list
.next
,
3832 struct ext4_prealloc_space
, pa_inode_list
);
3833 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3834 spin_lock(&pa
->pa_lock
);
3835 if (atomic_read(&pa
->pa_count
)) {
3836 /* this shouldn't happen often - nobody should
3837 * use preallocation while we're discarding it */
3838 spin_unlock(&pa
->pa_lock
);
3839 spin_unlock(&ei
->i_prealloc_lock
);
3840 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
3842 schedule_timeout_uninterruptible(HZ
);
3846 if (pa
->pa_deleted
== 0) {
3848 spin_unlock(&pa
->pa_lock
);
3849 list_del_rcu(&pa
->pa_inode_list
);
3850 list_add(&pa
->u
.pa_tmp_list
, &list
);
3854 /* someone is deleting pa right now */
3855 spin_unlock(&pa
->pa_lock
);
3856 spin_unlock(&ei
->i_prealloc_lock
);
3858 /* we have to wait here because pa_deleted
3859 * doesn't mean pa is already unlinked from
3860 * the list. as we might be called from
3861 * ->clear_inode() the inode will get freed
3862 * and concurrent thread which is unlinking
3863 * pa from inode's list may access already
3864 * freed memory, bad-bad-bad */
3866 /* XXX: if this happens too often, we can
3867 * add a flag to force wait only in case
3868 * of ->clear_inode(), but not in case of
3869 * regular truncate */
3870 schedule_timeout_uninterruptible(HZ
);
3873 spin_unlock(&ei
->i_prealloc_lock
);
3875 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3876 BUG_ON(pa
->pa_linear
!= 0);
3877 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
3879 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3881 ext4_error(sb
, __func__
, "Error in loading buddy "
3882 "information for %lu", group
);
3886 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3887 if (bitmap_bh
== NULL
) {
3888 ext4_error(sb
, __func__
, "Error in reading block "
3889 "bitmap for %lu", group
);
3890 ext4_mb_release_desc(&e4b
);
3894 ext4_lock_group(sb
, group
);
3895 list_del(&pa
->pa_group_list
);
3896 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3897 ext4_unlock_group(sb
, group
);
3899 ext4_mb_release_desc(&e4b
);
3902 list_del(&pa
->u
.pa_tmp_list
);
3903 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3906 kmem_cache_free(ext4_ac_cachep
, ac
);
3910 * finds all preallocated spaces and return blocks being freed to them
3911 * if preallocated space becomes full (no block is used from the space)
3912 * then the function frees space in buddy
3913 * XXX: at the moment, truncate (which is the only way to free blocks)
3914 * discards all preallocations
3916 static void ext4_mb_return_to_preallocation(struct inode
*inode
,
3917 struct ext4_buddy
*e4b
,
3918 sector_t block
, int count
)
3920 BUG_ON(!list_empty(&EXT4_I(inode
)->i_prealloc_list
));
3923 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3925 struct super_block
*sb
= ac
->ac_sb
;
3928 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
3929 " Allocation context details:\n");
3930 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
3931 ac
->ac_status
, ac
->ac_flags
);
3932 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3933 "best %lu/%lu/%lu@%lu cr %d\n",
3934 (unsigned long)ac
->ac_o_ex
.fe_group
,
3935 (unsigned long)ac
->ac_o_ex
.fe_start
,
3936 (unsigned long)ac
->ac_o_ex
.fe_len
,
3937 (unsigned long)ac
->ac_o_ex
.fe_logical
,
3938 (unsigned long)ac
->ac_g_ex
.fe_group
,
3939 (unsigned long)ac
->ac_g_ex
.fe_start
,
3940 (unsigned long)ac
->ac_g_ex
.fe_len
,
3941 (unsigned long)ac
->ac_g_ex
.fe_logical
,
3942 (unsigned long)ac
->ac_b_ex
.fe_group
,
3943 (unsigned long)ac
->ac_b_ex
.fe_start
,
3944 (unsigned long)ac
->ac_b_ex
.fe_len
,
3945 (unsigned long)ac
->ac_b_ex
.fe_logical
,
3946 (int)ac
->ac_criteria
);
3947 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
3949 printk(KERN_ERR
"EXT4-fs: groups: \n");
3950 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; i
++) {
3951 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
3952 struct ext4_prealloc_space
*pa
;
3953 ext4_grpblk_t start
;
3954 struct list_head
*cur
;
3955 ext4_lock_group(sb
, i
);
3956 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3957 pa
= list_entry(cur
, struct ext4_prealloc_space
,
3959 spin_lock(&pa
->pa_lock
);
3960 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3962 spin_unlock(&pa
->pa_lock
);
3963 printk(KERN_ERR
"PA:%lu:%d:%u \n", i
,
3966 ext4_unlock_group(sb
, i
);
3968 if (grp
->bb_free
== 0)
3970 printk(KERN_ERR
"%lu: %d/%d \n",
3971 i
, grp
->bb_free
, grp
->bb_fragments
);
3973 printk(KERN_ERR
"\n");
3976 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3983 * We use locality group preallocation for small size file. The size of the
3984 * file is determined by the current size or the resulting size after
3985 * allocation which ever is larger
3987 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
3989 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
3991 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3992 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3995 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3998 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3999 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
4000 size
= max(size
, isize
);
4002 /* don't use group allocation for large files */
4003 if (size
>= sbi
->s_mb_stream_request
)
4006 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4009 BUG_ON(ac
->ac_lg
!= NULL
);
4011 * locality group prealloc space are per cpu. The reason for having
4012 * per cpu locality group is to reduce the contention between block
4013 * request from multiple CPUs.
4015 ac
->ac_lg
= per_cpu_ptr(sbi
->s_locality_groups
, raw_smp_processor_id());
4017 /* we're going to use group allocation */
4018 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4020 /* serialize all allocations in the group */
4021 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4024 static noinline_for_stack
int
4025 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4026 struct ext4_allocation_request
*ar
)
4028 struct super_block
*sb
= ar
->inode
->i_sb
;
4029 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4030 struct ext4_super_block
*es
= sbi
->s_es
;
4034 ext4_grpblk_t block
;
4036 /* we can't allocate > group size */
4039 /* just a dirty hack to filter too big requests */
4040 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4041 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4043 /* start searching from the goal */
4045 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4046 goal
>= ext4_blocks_count(es
))
4047 goal
= le32_to_cpu(es
->s_first_data_block
);
4048 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4050 /* set up allocation goals */
4051 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4052 ac
->ac_b_ex
.fe_group
= 0;
4053 ac
->ac_b_ex
.fe_start
= 0;
4054 ac
->ac_b_ex
.fe_len
= 0;
4055 ac
->ac_status
= AC_STATUS_CONTINUE
;
4056 ac
->ac_groups_scanned
= 0;
4057 ac
->ac_ex_scanned
= 0;
4060 ac
->ac_inode
= ar
->inode
;
4061 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4062 ac
->ac_o_ex
.fe_group
= group
;
4063 ac
->ac_o_ex
.fe_start
= block
;
4064 ac
->ac_o_ex
.fe_len
= len
;
4065 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4066 ac
->ac_g_ex
.fe_group
= group
;
4067 ac
->ac_g_ex
.fe_start
= block
;
4068 ac
->ac_g_ex
.fe_len
= len
;
4069 ac
->ac_f_ex
.fe_len
= 0;
4070 ac
->ac_flags
= ar
->flags
;
4072 ac
->ac_criteria
= 0;
4074 ac
->ac_bitmap_page
= NULL
;
4075 ac
->ac_buddy_page
= NULL
;
4078 /* we have to define context: we'll we work with a file or
4079 * locality group. this is a policy, actually */
4080 ext4_mb_group_or_file(ac
);
4082 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4083 "left: %u/%u, right %u/%u to %swritable\n",
4084 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4085 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4086 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4087 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4088 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4093 static noinline_for_stack
void
4094 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4095 struct ext4_locality_group
*lg
,
4096 int order
, int total_entries
)
4098 ext4_group_t group
= 0;
4099 struct ext4_buddy e4b
;
4100 struct list_head discard_list
;
4101 struct ext4_prealloc_space
*pa
, *tmp
;
4102 struct ext4_allocation_context
*ac
;
4104 mb_debug("discard locality group preallocation\n");
4106 INIT_LIST_HEAD(&discard_list
);
4107 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4109 spin_lock(&lg
->lg_prealloc_lock
);
4110 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4112 spin_lock(&pa
->pa_lock
);
4113 if (atomic_read(&pa
->pa_count
)) {
4115 * This is the pa that we just used
4116 * for block allocation. So don't
4119 spin_unlock(&pa
->pa_lock
);
4122 if (pa
->pa_deleted
) {
4123 spin_unlock(&pa
->pa_lock
);
4126 /* only lg prealloc space */
4127 BUG_ON(!pa
->pa_linear
);
4129 /* seems this one can be freed ... */
4131 spin_unlock(&pa
->pa_lock
);
4133 list_del_rcu(&pa
->pa_inode_list
);
4134 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4137 if (total_entries
<= 5) {
4139 * we want to keep only 5 entries
4140 * allowing it to grow to 8. This
4141 * mak sure we don't call discard
4142 * soon for this list.
4147 spin_unlock(&lg
->lg_prealloc_lock
);
4149 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4151 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4152 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4153 ext4_error(sb
, __func__
, "Error in loading buddy "
4154 "information for %lu", group
);
4157 ext4_lock_group(sb
, group
);
4158 list_del(&pa
->pa_group_list
);
4159 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
4160 ext4_unlock_group(sb
, group
);
4162 ext4_mb_release_desc(&e4b
);
4163 list_del(&pa
->u
.pa_tmp_list
);
4164 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4167 kmem_cache_free(ext4_ac_cachep
, ac
);
4171 * We have incremented pa_count. So it cannot be freed at this
4172 * point. Also we hold lg_mutex. So no parallel allocation is
4173 * possible from this lg. That means pa_free cannot be updated.
4175 * A parallel ext4_mb_discard_group_preallocations is possible.
4176 * which can cause the lg_prealloc_list to be updated.
4179 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4181 int order
, added
= 0, lg_prealloc_count
= 1;
4182 struct super_block
*sb
= ac
->ac_sb
;
4183 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4184 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4186 order
= fls(pa
->pa_free
) - 1;
4187 if (order
> PREALLOC_TB_SIZE
- 1)
4188 /* The max size of hash table is PREALLOC_TB_SIZE */
4189 order
= PREALLOC_TB_SIZE
- 1;
4190 /* Add the prealloc space to lg */
4192 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4194 spin_lock(&tmp_pa
->pa_lock
);
4195 if (tmp_pa
->pa_deleted
) {
4196 spin_unlock(&pa
->pa_lock
);
4199 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4200 /* Add to the tail of the previous entry */
4201 list_add_tail_rcu(&pa
->pa_inode_list
,
4202 &tmp_pa
->pa_inode_list
);
4205 * we want to count the total
4206 * number of entries in the list
4209 spin_unlock(&tmp_pa
->pa_lock
);
4210 lg_prealloc_count
++;
4213 list_add_tail_rcu(&pa
->pa_inode_list
,
4214 &lg
->lg_prealloc_list
[order
]);
4217 /* Now trim the list to be not more than 8 elements */
4218 if (lg_prealloc_count
> 8) {
4219 ext4_mb_discard_lg_preallocations(sb
, lg
,
4220 order
, lg_prealloc_count
);
4227 * release all resource we used in allocation
4229 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4231 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4233 if (pa
->pa_linear
) {
4234 /* see comment in ext4_mb_use_group_pa() */
4235 spin_lock(&pa
->pa_lock
);
4236 pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4237 pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4238 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4239 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4240 spin_unlock(&pa
->pa_lock
);
4242 * We want to add the pa to the right bucket.
4243 * Remove it from the list and while adding
4244 * make sure the list to which we are adding
4247 if (likely(pa
->pa_free
)) {
4248 spin_lock(pa
->pa_obj_lock
);
4249 list_del_rcu(&pa
->pa_inode_list
);
4250 spin_unlock(pa
->pa_obj_lock
);
4251 ext4_mb_add_n_trim(ac
);
4254 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4256 if (ac
->ac_bitmap_page
)
4257 page_cache_release(ac
->ac_bitmap_page
);
4258 if (ac
->ac_buddy_page
)
4259 page_cache_release(ac
->ac_buddy_page
);
4260 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4261 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4262 ext4_mb_collect_stats(ac
);
4266 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4272 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
&& needed
> 0; i
++) {
4273 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4282 * Main entry point into mballoc to allocate blocks
4283 * it tries to use preallocation first, then falls back
4284 * to usual allocation
4286 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4287 struct ext4_allocation_request
*ar
, int *errp
)
4290 struct ext4_allocation_context
*ac
= NULL
;
4291 struct ext4_sb_info
*sbi
;
4292 struct super_block
*sb
;
4293 ext4_fsblk_t block
= 0;
4294 unsigned long inquota
;
4295 unsigned long reserv_blks
= 0;
4297 sb
= ar
->inode
->i_sb
;
4300 if (!EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
) {
4302 * With delalloc we already reserved the blocks
4304 while (ar
->len
&& ext4_claim_free_blocks(sbi
, ar
->len
)) {
4305 /* let others to free the space */
4307 ar
->len
= ar
->len
>> 1;
4313 reserv_blks
= ar
->len
;
4315 while (ar
->len
&& DQUOT_ALLOC_BLOCK(ar
->inode
, ar
->len
)) {
4316 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4325 if (EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
)
4326 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4328 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4335 *errp
= ext4_mb_initialize_context(ac
, ar
);
4341 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4342 if (!ext4_mb_use_preallocated(ac
)) {
4343 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4344 ext4_mb_normalize_request(ac
, ar
);
4346 /* allocate space in core */
4347 ext4_mb_regular_allocator(ac
);
4349 /* as we've just preallocated more space than
4350 * user requested orinally, we store allocated
4351 * space in a special descriptor */
4352 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4353 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4354 ext4_mb_new_preallocation(ac
);
4357 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4358 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_blks
);
4359 if (*errp
== -EAGAIN
) {
4360 ac
->ac_b_ex
.fe_group
= 0;
4361 ac
->ac_b_ex
.fe_start
= 0;
4362 ac
->ac_b_ex
.fe_len
= 0;
4363 ac
->ac_status
= AC_STATUS_CONTINUE
;
4366 ac
->ac_b_ex
.fe_len
= 0;
4368 ext4_mb_show_ac(ac
);
4370 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4371 ar
->len
= ac
->ac_b_ex
.fe_len
;
4374 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4378 ac
->ac_b_ex
.fe_len
= 0;
4380 ext4_mb_show_ac(ac
);
4383 ext4_mb_release_context(ac
);
4386 kmem_cache_free(ext4_ac_cachep
, ac
);
4388 if (ar
->len
< inquota
)
4389 DQUOT_FREE_BLOCK(ar
->inode
, inquota
- ar
->len
);
4395 * We can merge two free data extents only if the physical blocks
4396 * are contiguous, AND the extents were freed by the same transaction,
4397 * AND the blocks are associated with the same group.
4399 static int can_merge(struct ext4_free_data
*entry1
,
4400 struct ext4_free_data
*entry2
)
4402 if ((entry1
->t_tid
== entry2
->t_tid
) &&
4403 (entry1
->group
== entry2
->group
) &&
4404 ((entry1
->start_blk
+ entry1
->count
) == entry2
->start_blk
))
4409 static noinline_for_stack
int
4410 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4411 ext4_group_t group
, ext4_grpblk_t block
, int count
)
4413 struct ext4_group_info
*db
= e4b
->bd_info
;
4414 struct super_block
*sb
= e4b
->bd_sb
;
4415 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4416 struct ext4_free_data
*entry
, *new_entry
;
4417 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4418 struct rb_node
*parent
= NULL
, *new_node
;
4420 BUG_ON(!ext4_handle_valid(handle
));
4421 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4422 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4424 new_entry
= kmem_cache_alloc(ext4_free_ext_cachep
, GFP_NOFS
);
4425 new_entry
->start_blk
= block
;
4426 new_entry
->group
= group
;
4427 new_entry
->count
= count
;
4428 new_entry
->t_tid
= handle
->h_transaction
->t_tid
;
4429 new_node
= &new_entry
->node
;
4431 ext4_lock_group(sb
, group
);
4433 /* first free block exent. We need to
4434 protect buddy cache from being freed,
4435 * otherwise we'll refresh it from
4436 * on-disk bitmap and lose not-yet-available
4438 page_cache_get(e4b
->bd_buddy_page
);
4439 page_cache_get(e4b
->bd_bitmap_page
);
4443 entry
= rb_entry(parent
, struct ext4_free_data
, node
);
4444 if (block
< entry
->start_blk
)
4446 else if (block
>= (entry
->start_blk
+ entry
->count
))
4447 n
= &(*n
)->rb_right
;
4449 ext4_unlock_group(sb
, group
);
4450 ext4_error(sb
, __func__
,
4451 "Double free of blocks %d (%d %d)",
4452 block
, entry
->start_blk
, entry
->count
);
4457 rb_link_node(new_node
, parent
, n
);
4458 rb_insert_color(new_node
, &db
->bb_free_root
);
4460 /* Now try to see the extent can be merged to left and right */
4461 node
= rb_prev(new_node
);
4463 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4464 if (can_merge(entry
, new_entry
)) {
4465 new_entry
->start_blk
= entry
->start_blk
;
4466 new_entry
->count
+= entry
->count
;
4467 rb_erase(node
, &(db
->bb_free_root
));
4468 spin_lock(&sbi
->s_md_lock
);
4469 list_del(&entry
->list
);
4470 spin_unlock(&sbi
->s_md_lock
);
4471 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4475 node
= rb_next(new_node
);
4477 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4478 if (can_merge(new_entry
, entry
)) {
4479 new_entry
->count
+= entry
->count
;
4480 rb_erase(node
, &(db
->bb_free_root
));
4481 spin_lock(&sbi
->s_md_lock
);
4482 list_del(&entry
->list
);
4483 spin_unlock(&sbi
->s_md_lock
);
4484 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4487 /* Add the extent to transaction's private list */
4488 spin_lock(&sbi
->s_md_lock
);
4489 list_add(&new_entry
->list
, &handle
->h_transaction
->t_private_list
);
4490 spin_unlock(&sbi
->s_md_lock
);
4491 ext4_unlock_group(sb
, group
);
4496 * Main entry point into mballoc to free blocks
4498 void ext4_mb_free_blocks(handle_t
*handle
, struct inode
*inode
,
4499 unsigned long block
, unsigned long count
,
4500 int metadata
, unsigned long *freed
)
4502 struct buffer_head
*bitmap_bh
= NULL
;
4503 struct super_block
*sb
= inode
->i_sb
;
4504 struct ext4_allocation_context
*ac
= NULL
;
4505 struct ext4_group_desc
*gdp
;
4506 struct ext4_super_block
*es
;
4507 unsigned long overflow
;
4509 struct buffer_head
*gd_bh
;
4510 ext4_group_t block_group
;
4511 struct ext4_sb_info
*sbi
;
4512 struct ext4_buddy e4b
;
4519 es
= EXT4_SB(sb
)->s_es
;
4520 if (block
< le32_to_cpu(es
->s_first_data_block
) ||
4521 block
+ count
< block
||
4522 block
+ count
> ext4_blocks_count(es
)) {
4523 ext4_error(sb
, __func__
,
4524 "Freeing blocks not in datazone - "
4525 "block = %lu, count = %lu", block
, count
);
4529 ext4_debug("freeing block %lu\n", block
);
4531 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4533 ac
->ac_op
= EXT4_MB_HISTORY_FREE
;
4534 ac
->ac_inode
= inode
;
4540 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4543 * Check to see if we are freeing blocks across a group
4546 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4547 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4550 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4555 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4561 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4562 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4563 in_range(block
, ext4_inode_table(sb
, gdp
),
4564 EXT4_SB(sb
)->s_itb_per_group
) ||
4565 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4566 EXT4_SB(sb
)->s_itb_per_group
)) {
4568 ext4_error(sb
, __func__
,
4569 "Freeing blocks in system zone - "
4570 "Block = %lu, count = %lu", block
, count
);
4571 /* err = 0. ext4_std_error should be a no op */
4575 BUFFER_TRACE(bitmap_bh
, "getting write access");
4576 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4581 * We are about to modify some metadata. Call the journal APIs
4582 * to unshare ->b_data if a currently-committing transaction is
4585 BUFFER_TRACE(gd_bh
, "get_write_access");
4586 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4590 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4594 #ifdef AGGRESSIVE_CHECK
4597 for (i
= 0; i
< count
; i
++)
4598 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4601 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4604 /* We dirtied the bitmap block */
4605 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4606 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4609 ac
->ac_b_ex
.fe_group
= block_group
;
4610 ac
->ac_b_ex
.fe_start
= bit
;
4611 ac
->ac_b_ex
.fe_len
= count
;
4612 ext4_mb_store_history(ac
);
4615 if (metadata
&& ext4_handle_valid(handle
)) {
4616 /* blocks being freed are metadata. these blocks shouldn't
4617 * be used until this transaction is committed */
4618 ext4_mb_free_metadata(handle
, &e4b
, block_group
, bit
, count
);
4620 ext4_lock_group(sb
, block_group
);
4621 mb_free_blocks(inode
, &e4b
, bit
, count
);
4622 ext4_mb_return_to_preallocation(inode
, &e4b
, block
, count
);
4623 ext4_unlock_group(sb
, block_group
);
4626 spin_lock(sb_bgl_lock(sbi
, block_group
));
4627 le16_add_cpu(&gdp
->bg_free_blocks_count
, count
);
4628 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4629 spin_unlock(sb_bgl_lock(sbi
, block_group
));
4630 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4632 if (sbi
->s_log_groups_per_flex
) {
4633 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4634 spin_lock(sb_bgl_lock(sbi
, flex_group
));
4635 sbi
->s_flex_groups
[flex_group
].free_blocks
+= count
;
4636 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
4639 ext4_mb_release_desc(&e4b
);
4643 /* And the group descriptor block */
4644 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4645 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4649 if (overflow
&& !err
) {
4658 ext4_std_error(sb
, err
);
4660 kmem_cache_free(ext4_ac_cachep
, ac
);