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
25 #include <linux/debugfs.h>
26 #include <trace/events/ext4.h>
30 * - test ext4_ext_search_left() and ext4_ext_search_right()
31 * - search for metadata in few groups
34 * - normalization should take into account whether file is still open
35 * - discard preallocations if no free space left (policy?)
36 * - don't normalize tails
38 * - reservation for superuser
41 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
42 * - track min/max extents in each group for better group selection
43 * - mb_mark_used() may allocate chunk right after splitting buddy
44 * - tree of groups sorted by number of free blocks
49 * The allocation request involve request for multiple number of blocks
50 * near to the goal(block) value specified.
52 * During initialization phase of the allocator we decide to use the
53 * group preallocation or inode preallocation depending on the size of
54 * the file. The size of the file could be the resulting file size we
55 * would have after allocation, or the current file size, which ever
56 * is larger. If the size is less than sbi->s_mb_stream_request we
57 * select to use the group preallocation. The default value of
58 * s_mb_stream_request is 16 blocks. This can also be tuned via
59 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
60 * terms of number of blocks.
62 * The main motivation for having small file use group preallocation is to
63 * ensure that we have small files closer together on the disk.
65 * First stage the allocator looks at the inode prealloc list,
66 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
67 * spaces for this particular inode. The inode prealloc space is
70 * pa_lstart -> the logical start block for this prealloc space
71 * pa_pstart -> the physical start block for this prealloc space
72 * pa_len -> lenght for this prealloc space
73 * pa_free -> free space available in this prealloc space
75 * The inode preallocation space is used looking at the _logical_ start
76 * block. If only the logical file block falls within the range of prealloc
77 * space we will consume the particular prealloc space. This make sure that
78 * that the we have contiguous physical blocks representing the file blocks
80 * The important thing to be noted in case of inode prealloc space is that
81 * we don't modify the values associated to inode prealloc space except
84 * If we are not able to find blocks in the inode prealloc space and if we
85 * have the group allocation flag set then we look at the locality group
86 * prealloc space. These are per CPU prealloc list repreasented as
88 * ext4_sb_info.s_locality_groups[smp_processor_id()]
90 * The reason for having a per cpu locality group is to reduce the contention
91 * between CPUs. It is possible to get scheduled at this point.
93 * The locality group prealloc space is used looking at whether we have
94 * enough free space (pa_free) withing the prealloc space.
96 * If we can't allocate blocks via inode prealloc or/and locality group
97 * prealloc then we look at the buddy cache. The buddy cache is represented
98 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
99 * mapped to the buddy and bitmap information regarding different
100 * groups. The buddy information is attached to buddy cache inode so that
101 * we can access them through the page cache. The information regarding
102 * each group is loaded via ext4_mb_load_buddy. The information involve
103 * block bitmap and buddy information. The information are stored in the
107 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
110 * one block each for bitmap and buddy information. So for each group we
111 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
112 * blocksize) blocks. So it can have information regarding groups_per_page
113 * which is blocks_per_page/2
115 * The buddy cache inode is not stored on disk. The inode is thrown
116 * away when the filesystem is unmounted.
118 * We look for count number of blocks in the buddy cache. If we were able
119 * to locate that many free blocks we return with additional information
120 * regarding rest of the contiguous physical block available
122 * Before allocating blocks via buddy cache we normalize the request
123 * blocks. This ensure we ask for more blocks that we needed. The extra
124 * blocks that we get after allocation is added to the respective prealloc
125 * list. In case of inode preallocation we follow a list of heuristics
126 * based on file size. This can be found in ext4_mb_normalize_request. If
127 * we are doing a group prealloc we try to normalize the request to
128 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
129 * 512 blocks. This can be tuned via
130 * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
131 * terms of number of blocks. If we have mounted the file system with -O
132 * stripe=<value> option the group prealloc request is normalized to the
133 * stripe value (sbi->s_stripe)
135 * The regular allocator(using the buddy cache) supports few tunables.
137 * /sys/fs/ext4/<partition>/mb_min_to_scan
138 * /sys/fs/ext4/<partition>/mb_max_to_scan
139 * /sys/fs/ext4/<partition>/mb_order2_req
141 * The regular allocator uses buddy scan only if the request len is power of
142 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
143 * value of s_mb_order2_reqs can be tuned via
144 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
145 * stripe size (sbi->s_stripe), we try to search for contiguous block in
146 * stripe size. This should result in better allocation on RAID setups. If
147 * not, we search in the specific group using bitmap for best extents. The
148 * tunable min_to_scan and max_to_scan control the behaviour here.
149 * min_to_scan indicate how long the mballoc __must__ look for a best
150 * extent and max_to_scan indicates how long the mballoc __can__ look for a
151 * best extent in the found extents. Searching for the blocks starts with
152 * the group specified as the goal value in allocation context via
153 * ac_g_ex. Each group is first checked based on the criteria whether it
154 * can used for allocation. ext4_mb_good_group explains how the groups are
157 * Both the prealloc space are getting populated as above. So for the first
158 * request we will hit the buddy cache which will result in this prealloc
159 * space getting filled. The prealloc space is then later used for the
160 * subsequent request.
164 * mballoc operates on the following data:
166 * - in-core buddy (actually includes buddy and bitmap)
167 * - preallocation descriptors (PAs)
169 * there are two types of preallocations:
171 * assiged to specific inode and can be used for this inode only.
172 * it describes part of inode's space preallocated to specific
173 * physical blocks. any block from that preallocated can be used
174 * independent. the descriptor just tracks number of blocks left
175 * unused. so, before taking some block from descriptor, one must
176 * make sure corresponded logical block isn't allocated yet. this
177 * also means that freeing any block within descriptor's range
178 * must discard all preallocated blocks.
180 * assigned to specific locality group which does not translate to
181 * permanent set of inodes: inode can join and leave group. space
182 * from this type of preallocation can be used for any inode. thus
183 * it's consumed from the beginning to the end.
185 * relation between them can be expressed as:
186 * in-core buddy = on-disk bitmap + preallocation descriptors
188 * this mean blocks mballoc considers used are:
189 * - allocated blocks (persistent)
190 * - preallocated blocks (non-persistent)
192 * consistency in mballoc world means that at any time a block is either
193 * free or used in ALL structures. notice: "any time" should not be read
194 * literally -- time is discrete and delimited by locks.
196 * to keep it simple, we don't use block numbers, instead we count number of
197 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
199 * all operations can be expressed as:
200 * - init buddy: buddy = on-disk + PAs
201 * - new PA: buddy += N; PA = N
202 * - use inode PA: on-disk += N; PA -= N
203 * - discard inode PA buddy -= on-disk - PA; PA = 0
204 * - use locality group PA on-disk += N; PA -= N
205 * - discard locality group PA buddy -= PA; PA = 0
206 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
207 * is used in real operation because we can't know actual used
208 * bits from PA, only from on-disk bitmap
210 * if we follow this strict logic, then all operations above should be atomic.
211 * given some of them can block, we'd have to use something like semaphores
212 * killing performance on high-end SMP hardware. let's try to relax it using
213 * the following knowledge:
214 * 1) if buddy is referenced, it's already initialized
215 * 2) while block is used in buddy and the buddy is referenced,
216 * nobody can re-allocate that block
217 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
218 * bit set and PA claims same block, it's OK. IOW, one can set bit in
219 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
222 * so, now we're building a concurrency table:
225 * blocks for PA are allocated in the buddy, buddy must be referenced
226 * until PA is linked to allocation group to avoid concurrent buddy init
228 * we need to make sure that either on-disk bitmap or PA has uptodate data
229 * given (3) we care that PA-=N operation doesn't interfere with init
231 * the simplest way would be to have buddy initialized by the discard
232 * - use locality group PA
233 * again PA-=N must be serialized with init
234 * - discard locality group PA
235 * the simplest way would be to have buddy initialized by the discard
238 * i_data_sem serializes them
240 * discard process must wait until PA isn't used by another process
241 * - use locality group PA
242 * some mutex should serialize them
243 * - discard locality group PA
244 * discard process must wait until PA isn't used by another process
247 * i_data_sem or another mutex should serializes them
249 * discard process must wait until PA isn't used by another process
250 * - use locality group PA
251 * nothing wrong here -- they're different PAs covering different blocks
252 * - discard locality group PA
253 * discard process must wait until PA isn't used by another process
255 * now we're ready to make few consequences:
256 * - PA is referenced and while it is no discard is possible
257 * - PA is referenced until block isn't marked in on-disk bitmap
258 * - PA changes only after on-disk bitmap
259 * - discard must not compete with init. either init is done before
260 * any discard or they're serialized somehow
261 * - buddy init as sum of on-disk bitmap and PAs is done atomically
263 * a special case when we've used PA to emptiness. no need to modify buddy
264 * in this case, but we should care about concurrent init
269 * Logic in few words:
274 * mark bits in on-disk bitmap
277 * - use preallocation:
278 * find proper PA (per-inode or group)
280 * mark bits in on-disk bitmap
286 * mark bits in on-disk bitmap
289 * - discard preallocations in group:
291 * move them onto local list
292 * load on-disk bitmap
294 * remove PA from object (inode or locality group)
295 * mark free blocks in-core
297 * - discard inode's preallocations:
304 * - bitlock on a group (group)
305 * - object (inode/locality) (object)
316 * - release consumed pa:
321 * - generate in-core bitmap:
325 * - discard all for given object (inode, locality group):
330 * - discard all for given group:
337 static struct kmem_cache
*ext4_pspace_cachep
;
338 static struct kmem_cache
*ext4_ac_cachep
;
339 static struct kmem_cache
*ext4_free_ext_cachep
;
340 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
342 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
344 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
);
346 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
348 #if BITS_PER_LONG == 64
349 *bit
+= ((unsigned long) addr
& 7UL) << 3;
350 addr
= (void *) ((unsigned long) addr
& ~7UL);
351 #elif BITS_PER_LONG == 32
352 *bit
+= ((unsigned long) addr
& 3UL) << 3;
353 addr
= (void *) ((unsigned long) addr
& ~3UL);
355 #error "how many bits you are?!"
360 static inline int mb_test_bit(int bit
, void *addr
)
363 * ext4_test_bit on architecture like powerpc
364 * needs unsigned long aligned address
366 addr
= mb_correct_addr_and_bit(&bit
, addr
);
367 return ext4_test_bit(bit
, addr
);
370 static inline void mb_set_bit(int bit
, void *addr
)
372 addr
= mb_correct_addr_and_bit(&bit
, addr
);
373 ext4_set_bit(bit
, addr
);
376 static inline void mb_clear_bit(int bit
, void *addr
)
378 addr
= mb_correct_addr_and_bit(&bit
, addr
);
379 ext4_clear_bit(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 assert_spin_locked(ext4_group_lock_ptr(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
;
445 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
446 blocknr
+= first
+ i
;
447 ext4_grp_locked_error(sb
, e4b
->bd_group
,
448 __func__
, "double-free of inode"
449 " %lu's block %llu(bit %u in group %u)",
450 inode
? inode
->i_ino
: 0, blocknr
,
451 first
+ i
, e4b
->bd_group
);
453 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
457 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
461 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
463 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
464 for (i
= 0; i
< count
; i
++) {
465 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
466 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
470 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
472 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
473 unsigned char *b1
, *b2
;
475 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
476 b2
= (unsigned char *) bitmap
;
477 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
478 if (b1
[i
] != b2
[i
]) {
479 printk(KERN_ERR
"corruption in group %u "
480 "at byte %u(%u): %x in copy != %x "
481 "on disk/prealloc\n",
482 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
490 static inline void mb_free_blocks_double(struct inode
*inode
,
491 struct ext4_buddy
*e4b
, int first
, int count
)
495 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
496 int first
, int count
)
500 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
506 #ifdef AGGRESSIVE_CHECK
508 #define MB_CHECK_ASSERT(assert) \
512 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
513 function, file, line, # assert); \
518 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
519 const char *function
, int line
)
521 struct super_block
*sb
= e4b
->bd_sb
;
522 int order
= e4b
->bd_blkbits
+ 1;
529 struct ext4_group_info
*grp
;
532 struct list_head
*cur
;
537 static int mb_check_counter
;
538 if (mb_check_counter
++ % 100 != 0)
543 buddy
= mb_find_buddy(e4b
, order
, &max
);
544 MB_CHECK_ASSERT(buddy
);
545 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
546 MB_CHECK_ASSERT(buddy2
);
547 MB_CHECK_ASSERT(buddy
!= buddy2
);
548 MB_CHECK_ASSERT(max
* 2 == max2
);
551 for (i
= 0; i
< max
; i
++) {
553 if (mb_test_bit(i
, buddy
)) {
554 /* only single bit in buddy2 may be 1 */
555 if (!mb_test_bit(i
<< 1, buddy2
)) {
557 mb_test_bit((i
<<1)+1, buddy2
));
558 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
560 mb_test_bit(i
<< 1, buddy2
));
565 /* both bits in buddy2 must be 0 */
566 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
567 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
569 for (j
= 0; j
< (1 << order
); j
++) {
570 k
= (i
* (1 << order
)) + j
;
572 !mb_test_bit(k
, EXT4_MB_BITMAP(e4b
)));
576 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
581 buddy
= mb_find_buddy(e4b
, 0, &max
);
582 for (i
= 0; i
< max
; i
++) {
583 if (!mb_test_bit(i
, buddy
)) {
584 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
592 /* check used bits only */
593 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
594 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
596 MB_CHECK_ASSERT(k
< max2
);
597 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
600 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
601 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
603 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
604 buddy
= mb_find_buddy(e4b
, 0, &max
);
605 list_for_each(cur
, &grp
->bb_prealloc_list
) {
606 ext4_group_t groupnr
;
607 struct ext4_prealloc_space
*pa
;
608 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
609 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
610 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
611 for (i
= 0; i
< pa
->pa_len
; i
++)
612 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
616 #undef MB_CHECK_ASSERT
617 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
618 __FILE__, __func__, __LINE__)
620 #define mb_check_buddy(e4b)
623 /* FIXME!! need more doc */
624 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
625 void *buddy
, ext4_grpblk_t first
, ext4_grpblk_t len
,
626 struct ext4_group_info
*grp
)
628 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
632 unsigned short border
;
634 BUG_ON(len
> EXT4_BLOCKS_PER_GROUP(sb
));
636 border
= 2 << sb
->s_blocksize_bits
;
639 /* find how many blocks can be covered since this position */
640 max
= ffs(first
| border
) - 1;
642 /* find how many blocks of power 2 we need to mark */
649 /* mark multiblock chunks only */
650 grp
->bb_counters
[min
]++;
652 mb_clear_bit(first
>> min
,
653 buddy
+ sbi
->s_mb_offsets
[min
]);
660 static noinline_for_stack
661 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 ext4_grpblk_t max
= EXT4_BLOCKS_PER_GROUP(sb
);
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_grp_locked_error(sb
, group
, __func__
,
694 "EXT4-fs: group %u: %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 bitmap][ group 0 buddy] [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
)
731 ext4_group_t ngroups
;
737 ext4_group_t first_group
;
739 struct super_block
*sb
;
740 struct buffer_head
*bhs
;
741 struct buffer_head
**bh
;
746 mb_debug(1, "init page %lu\n", page
->index
);
748 inode
= page
->mapping
->host
;
750 ngroups
= ext4_get_groups_count(sb
);
751 blocksize
= 1 << inode
->i_blkbits
;
752 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
754 groups_per_page
= blocks_per_page
>> 1;
755 if (groups_per_page
== 0)
758 /* allocate buffer_heads to read bitmaps */
759 if (groups_per_page
> 1) {
761 i
= sizeof(struct buffer_head
*) * groups_per_page
;
762 bh
= kzalloc(i
, GFP_NOFS
);
768 first_group
= page
->index
* blocks_per_page
/ 2;
770 /* read all groups the page covers into the cache */
771 for (i
= 0; i
< groups_per_page
; i
++) {
772 struct ext4_group_desc
*desc
;
774 if (first_group
+ i
>= ngroups
)
778 desc
= ext4_get_group_desc(sb
, first_group
+ i
, NULL
);
783 bh
[i
] = sb_getblk(sb
, ext4_block_bitmap(sb
, desc
));
787 if (bitmap_uptodate(bh
[i
]))
791 if (bitmap_uptodate(bh
[i
])) {
792 unlock_buffer(bh
[i
]);
795 ext4_lock_group(sb
, first_group
+ i
);
796 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
797 ext4_init_block_bitmap(sb
, bh
[i
],
798 first_group
+ i
, desc
);
799 set_bitmap_uptodate(bh
[i
]);
800 set_buffer_uptodate(bh
[i
]);
801 ext4_unlock_group(sb
, first_group
+ i
);
802 unlock_buffer(bh
[i
]);
805 ext4_unlock_group(sb
, first_group
+ i
);
806 if (buffer_uptodate(bh
[i
])) {
808 * if not uninit if bh is uptodate,
809 * bitmap is also uptodate
811 set_bitmap_uptodate(bh
[i
]);
812 unlock_buffer(bh
[i
]);
817 * submit the buffer_head for read. We can
818 * safely mark the bitmap as uptodate now.
819 * We do it here so the bitmap uptodate bit
820 * get set with buffer lock held.
822 set_bitmap_uptodate(bh
[i
]);
823 bh
[i
]->b_end_io
= end_buffer_read_sync
;
824 submit_bh(READ
, bh
[i
]);
825 mb_debug(1, "read bitmap for group %u\n", first_group
+ i
);
828 /* wait for I/O completion */
829 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
830 wait_on_buffer(bh
[i
]);
833 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
834 if (!buffer_uptodate(bh
[i
]))
838 first_block
= page
->index
* blocks_per_page
;
840 memset(page_address(page
), 0xff, PAGE_CACHE_SIZE
);
841 for (i
= 0; i
< blocks_per_page
; i
++) {
843 struct ext4_group_info
*grinfo
;
845 group
= (first_block
+ i
) >> 1;
846 if (group
>= ngroups
)
850 * data carry information regarding this
851 * particular group in the format specified
855 data
= page_address(page
) + (i
* blocksize
);
856 bitmap
= bh
[group
- first_group
]->b_data
;
859 * We place the buddy block and bitmap block
862 if ((first_block
+ i
) & 1) {
863 /* this is block of buddy */
864 BUG_ON(incore
== NULL
);
865 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
866 group
, page
->index
, i
* blocksize
);
867 grinfo
= ext4_get_group_info(sb
, group
);
868 grinfo
->bb_fragments
= 0;
869 memset(grinfo
->bb_counters
, 0,
870 sizeof(*grinfo
->bb_counters
) *
871 (sb
->s_blocksize_bits
+2));
873 * incore got set to the group block bitmap below
875 ext4_lock_group(sb
, group
);
876 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
877 ext4_unlock_group(sb
, group
);
880 /* this is block of bitmap */
881 BUG_ON(incore
!= NULL
);
882 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
883 group
, page
->index
, i
* blocksize
);
885 /* see comments in ext4_mb_put_pa() */
886 ext4_lock_group(sb
, group
);
887 memcpy(data
, bitmap
, blocksize
);
889 /* mark all preallocated blks used in in-core bitmap */
890 ext4_mb_generate_from_pa(sb
, data
, group
);
891 ext4_mb_generate_from_freelist(sb
, data
, group
);
892 ext4_unlock_group(sb
, group
);
894 /* set incore so that the buddy information can be
895 * generated using this
900 SetPageUptodate(page
);
904 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
912 static noinline_for_stack
913 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
)
919 int block
, pnum
, poff
;
920 int num_grp_locked
= 0;
921 struct ext4_group_info
*this_grp
;
922 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
923 struct inode
*inode
= sbi
->s_buddy_cache
;
924 struct page
*page
= NULL
, *bitmap_page
= NULL
;
926 mb_debug(1, "init group %u\n", group
);
927 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
928 this_grp
= ext4_get_group_info(sb
, group
);
930 * This ensures that we don't reinit the buddy cache
931 * page which map to the group from which we are already
932 * allocating. If we are looking at the buddy cache we would
933 * have taken a reference using ext4_mb_load_buddy and that
934 * would have taken the alloc_sem lock.
936 num_grp_locked
= ext4_mb_get_buddy_cache_lock(sb
, group
);
937 if (!EXT4_MB_GRP_NEED_INIT(this_grp
)) {
939 * somebody initialized the group
940 * return without doing anything
946 * the buddy cache inode stores the block bitmap
947 * and buddy information in consecutive blocks.
948 * So for each group we need two blocks.
951 pnum
= block
/ blocks_per_page
;
952 poff
= block
% blocks_per_page
;
953 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
955 BUG_ON(page
->mapping
!= inode
->i_mapping
);
956 ret
= ext4_mb_init_cache(page
, NULL
);
963 if (page
== NULL
|| !PageUptodate(page
)) {
967 mark_page_accessed(page
);
969 bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
971 /* init buddy cache */
973 pnum
= block
/ blocks_per_page
;
974 poff
= block
% blocks_per_page
;
975 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
976 if (page
== bitmap_page
) {
978 * If both the bitmap and buddy are in
979 * the same page we don't need to force
984 BUG_ON(page
->mapping
!= inode
->i_mapping
);
985 ret
= ext4_mb_init_cache(page
, bitmap
);
992 if (page
== NULL
|| !PageUptodate(page
)) {
996 mark_page_accessed(page
);
998 ext4_mb_put_buddy_cache_lock(sb
, group
, num_grp_locked
);
1000 page_cache_release(bitmap_page
);
1002 page_cache_release(page
);
1006 static noinline_for_stack
int
1007 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1008 struct ext4_buddy
*e4b
)
1010 int blocks_per_page
;
1016 struct ext4_group_info
*grp
;
1017 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1018 struct inode
*inode
= sbi
->s_buddy_cache
;
1020 mb_debug(1, "load group %u\n", group
);
1022 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1023 grp
= ext4_get_group_info(sb
, group
);
1025 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1026 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
1028 e4b
->bd_group
= group
;
1029 e4b
->bd_buddy_page
= NULL
;
1030 e4b
->bd_bitmap_page
= NULL
;
1031 e4b
->alloc_semp
= &grp
->alloc_sem
;
1033 /* Take the read lock on the group alloc
1034 * sem. This would make sure a parallel
1035 * ext4_mb_init_group happening on other
1036 * groups mapped by the page is blocked
1037 * till we are done with allocation
1040 down_read(e4b
->alloc_semp
);
1042 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1043 /* we need to check for group need init flag
1044 * with alloc_semp held so that we can be sure
1045 * that new blocks didn't get added to the group
1046 * when we are loading the buddy cache
1048 up_read(e4b
->alloc_semp
);
1050 * we need full data about the group
1051 * to make a good selection
1053 ret
= ext4_mb_init_group(sb
, group
);
1056 goto repeat_load_buddy
;
1060 * the buddy cache inode stores the block bitmap
1061 * and buddy information in consecutive blocks.
1062 * So for each group we need two blocks.
1065 pnum
= block
/ blocks_per_page
;
1066 poff
= block
% blocks_per_page
;
1068 /* we could use find_or_create_page(), but it locks page
1069 * what we'd like to avoid in fast path ... */
1070 page
= find_get_page(inode
->i_mapping
, pnum
);
1071 if (page
== NULL
|| !PageUptodate(page
)) {
1074 * drop the page reference and try
1075 * to get the page with lock. If we
1076 * are not uptodate that implies
1077 * somebody just created the page but
1078 * is yet to initialize the same. So
1079 * wait for it to initialize.
1081 page_cache_release(page
);
1082 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1084 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1085 if (!PageUptodate(page
)) {
1086 ret
= ext4_mb_init_cache(page
, NULL
);
1091 mb_cmp_bitmaps(e4b
, page_address(page
) +
1092 (poff
* sb
->s_blocksize
));
1097 if (page
== NULL
|| !PageUptodate(page
)) {
1101 e4b
->bd_bitmap_page
= page
;
1102 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1103 mark_page_accessed(page
);
1106 pnum
= block
/ blocks_per_page
;
1107 poff
= block
% blocks_per_page
;
1109 page
= find_get_page(inode
->i_mapping
, pnum
);
1110 if (page
== NULL
|| !PageUptodate(page
)) {
1112 page_cache_release(page
);
1113 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1115 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1116 if (!PageUptodate(page
)) {
1117 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1126 if (page
== NULL
|| !PageUptodate(page
)) {
1130 e4b
->bd_buddy_page
= page
;
1131 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1132 mark_page_accessed(page
);
1134 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1135 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1140 if (e4b
->bd_bitmap_page
)
1141 page_cache_release(e4b
->bd_bitmap_page
);
1142 if (e4b
->bd_buddy_page
)
1143 page_cache_release(e4b
->bd_buddy_page
);
1144 e4b
->bd_buddy
= NULL
;
1145 e4b
->bd_bitmap
= NULL
;
1147 /* Done with the buddy cache */
1148 up_read(e4b
->alloc_semp
);
1152 static void ext4_mb_release_desc(struct ext4_buddy
*e4b
)
1154 if (e4b
->bd_bitmap_page
)
1155 page_cache_release(e4b
->bd_bitmap_page
);
1156 if (e4b
->bd_buddy_page
)
1157 page_cache_release(e4b
->bd_buddy_page
);
1158 /* Done with the buddy cache */
1159 if (e4b
->alloc_semp
)
1160 up_read(e4b
->alloc_semp
);
1164 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1169 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1170 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1172 bb
= EXT4_MB_BUDDY(e4b
);
1173 while (order
<= e4b
->bd_blkbits
+ 1) {
1175 if (!mb_test_bit(block
, bb
)) {
1176 /* this block is part of buddy of order 'order' */
1179 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1185 static void mb_clear_bits(void *bm
, int cur
, int len
)
1191 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1192 /* fast path: clear whole word at once */
1193 addr
= bm
+ (cur
>> 3);
1198 mb_clear_bit(cur
, bm
);
1203 static void mb_set_bits(void *bm
, int cur
, int len
)
1209 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1210 /* fast path: set whole word at once */
1211 addr
= bm
+ (cur
>> 3);
1216 mb_set_bit(cur
, bm
);
1221 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1222 int first
, int count
)
1229 struct super_block
*sb
= e4b
->bd_sb
;
1231 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1232 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1233 mb_check_buddy(e4b
);
1234 mb_free_blocks_double(inode
, e4b
, first
, count
);
1236 e4b
->bd_info
->bb_free
+= count
;
1237 if (first
< e4b
->bd_info
->bb_first_free
)
1238 e4b
->bd_info
->bb_first_free
= first
;
1240 /* let's maintain fragments counter */
1242 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1243 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1244 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1246 e4b
->bd_info
->bb_fragments
--;
1247 else if (!block
&& !max
)
1248 e4b
->bd_info
->bb_fragments
++;
1250 /* let's maintain buddy itself */
1251 while (count
-- > 0) {
1255 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1256 ext4_fsblk_t blocknr
;
1258 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1260 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1261 __func__
, "double-free of inode"
1262 " %lu's block %llu(bit %u in group %u)",
1263 inode
? inode
->i_ino
: 0, blocknr
, block
,
1266 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1267 e4b
->bd_info
->bb_counters
[order
]++;
1269 /* start of the buddy */
1270 buddy
= mb_find_buddy(e4b
, order
, &max
);
1274 if (mb_test_bit(block
, buddy
) ||
1275 mb_test_bit(block
+ 1, buddy
))
1278 /* both the buddies are free, try to coalesce them */
1279 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1285 /* for special purposes, we don't set
1286 * free bits in bitmap */
1287 mb_set_bit(block
, buddy
);
1288 mb_set_bit(block
+ 1, buddy
);
1290 e4b
->bd_info
->bb_counters
[order
]--;
1291 e4b
->bd_info
->bb_counters
[order
]--;
1295 e4b
->bd_info
->bb_counters
[order
]++;
1297 mb_clear_bit(block
, buddy2
);
1301 mb_check_buddy(e4b
);
1304 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1305 int needed
, struct ext4_free_extent
*ex
)
1312 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1315 buddy
= mb_find_buddy(e4b
, order
, &max
);
1316 BUG_ON(buddy
== NULL
);
1317 BUG_ON(block
>= max
);
1318 if (mb_test_bit(block
, buddy
)) {
1325 /* FIXME dorp order completely ? */
1326 if (likely(order
== 0)) {
1327 /* find actual order */
1328 order
= mb_find_order_for_block(e4b
, block
);
1329 block
= block
>> order
;
1332 ex
->fe_len
= 1 << order
;
1333 ex
->fe_start
= block
<< order
;
1334 ex
->fe_group
= e4b
->bd_group
;
1336 /* calc difference from given start */
1337 next
= next
- ex
->fe_start
;
1339 ex
->fe_start
+= next
;
1341 while (needed
> ex
->fe_len
&&
1342 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1344 if (block
+ 1 >= max
)
1347 next
= (block
+ 1) * (1 << order
);
1348 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1351 ord
= mb_find_order_for_block(e4b
, next
);
1354 block
= next
>> order
;
1355 ex
->fe_len
+= 1 << order
;
1358 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1362 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1368 int start
= ex
->fe_start
;
1369 int len
= ex
->fe_len
;
1374 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1375 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1376 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1377 mb_check_buddy(e4b
);
1378 mb_mark_used_double(e4b
, start
, len
);
1380 e4b
->bd_info
->bb_free
-= len
;
1381 if (e4b
->bd_info
->bb_first_free
== start
)
1382 e4b
->bd_info
->bb_first_free
+= len
;
1384 /* let's maintain fragments counter */
1386 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1387 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1388 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1390 e4b
->bd_info
->bb_fragments
++;
1391 else if (!mlen
&& !max
)
1392 e4b
->bd_info
->bb_fragments
--;
1394 /* let's maintain buddy itself */
1396 ord
= mb_find_order_for_block(e4b
, start
);
1398 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1399 /* the whole chunk may be allocated at once! */
1401 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1402 BUG_ON((start
>> ord
) >= max
);
1403 mb_set_bit(start
>> ord
, buddy
);
1404 e4b
->bd_info
->bb_counters
[ord
]--;
1411 /* store for history */
1413 ret
= len
| (ord
<< 16);
1415 /* we have to split large buddy */
1417 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1418 mb_set_bit(start
>> ord
, buddy
);
1419 e4b
->bd_info
->bb_counters
[ord
]--;
1422 cur
= (start
>> ord
) & ~1U;
1423 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1424 mb_clear_bit(cur
, buddy
);
1425 mb_clear_bit(cur
+ 1, buddy
);
1426 e4b
->bd_info
->bb_counters
[ord
]++;
1427 e4b
->bd_info
->bb_counters
[ord
]++;
1430 mb_set_bits(EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1431 mb_check_buddy(e4b
);
1437 * Must be called under group lock!
1439 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1440 struct ext4_buddy
*e4b
)
1442 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1445 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1446 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1448 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1449 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1450 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1452 /* preallocation can change ac_b_ex, thus we store actually
1453 * allocated blocks for history */
1454 ac
->ac_f_ex
= ac
->ac_b_ex
;
1456 ac
->ac_status
= AC_STATUS_FOUND
;
1457 ac
->ac_tail
= ret
& 0xffff;
1458 ac
->ac_buddy
= ret
>> 16;
1461 * take the page reference. We want the page to be pinned
1462 * so that we don't get a ext4_mb_init_cache_call for this
1463 * group until we update the bitmap. That would mean we
1464 * double allocate blocks. The reference is dropped
1465 * in ext4_mb_release_context
1467 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1468 get_page(ac
->ac_bitmap_page
);
1469 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1470 get_page(ac
->ac_buddy_page
);
1471 /* on allocation we use ac to track the held semaphore */
1472 ac
->alloc_semp
= e4b
->alloc_semp
;
1473 e4b
->alloc_semp
= NULL
;
1474 /* store last allocated for subsequent stream allocation */
1475 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1476 spin_lock(&sbi
->s_md_lock
);
1477 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1478 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1479 spin_unlock(&sbi
->s_md_lock
);
1484 * regular allocator, for general purposes allocation
1487 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1488 struct ext4_buddy
*e4b
,
1491 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1492 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1493 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1494 struct ext4_free_extent ex
;
1497 if (ac
->ac_status
== AC_STATUS_FOUND
)
1500 * We don't want to scan for a whole year
1502 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1503 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1504 ac
->ac_status
= AC_STATUS_BREAK
;
1509 * Haven't found good chunk so far, let's continue
1511 if (bex
->fe_len
< gex
->fe_len
)
1514 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1515 && bex
->fe_group
== e4b
->bd_group
) {
1516 /* recheck chunk's availability - we don't know
1517 * when it was found (within this lock-unlock
1519 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1520 if (max
>= gex
->fe_len
) {
1521 ext4_mb_use_best_found(ac
, e4b
);
1528 * The routine checks whether found extent is good enough. If it is,
1529 * then the extent gets marked used and flag is set to the context
1530 * to stop scanning. Otherwise, the extent is compared with the
1531 * previous found extent and if new one is better, then it's stored
1532 * in the context. Later, the best found extent will be used, if
1533 * mballoc can't find good enough extent.
1535 * FIXME: real allocation policy is to be designed yet!
1537 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1538 struct ext4_free_extent
*ex
,
1539 struct ext4_buddy
*e4b
)
1541 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1542 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1544 BUG_ON(ex
->fe_len
<= 0);
1545 BUG_ON(ex
->fe_len
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1546 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1547 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1552 * The special case - take what you catch first
1554 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1556 ext4_mb_use_best_found(ac
, e4b
);
1561 * Let's check whether the chuck is good enough
1563 if (ex
->fe_len
== gex
->fe_len
) {
1565 ext4_mb_use_best_found(ac
, e4b
);
1570 * If this is first found extent, just store it in the context
1572 if (bex
->fe_len
== 0) {
1578 * If new found extent is better, store it in the context
1580 if (bex
->fe_len
< gex
->fe_len
) {
1581 /* if the request isn't satisfied, any found extent
1582 * larger than previous best one is better */
1583 if (ex
->fe_len
> bex
->fe_len
)
1585 } else if (ex
->fe_len
> gex
->fe_len
) {
1586 /* if the request is satisfied, then we try to find
1587 * an extent that still satisfy the request, but is
1588 * smaller than previous one */
1589 if (ex
->fe_len
< bex
->fe_len
)
1593 ext4_mb_check_limits(ac
, e4b
, 0);
1596 static noinline_for_stack
1597 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1598 struct ext4_buddy
*e4b
)
1600 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1601 ext4_group_t group
= ex
.fe_group
;
1605 BUG_ON(ex
.fe_len
<= 0);
1606 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1610 ext4_lock_group(ac
->ac_sb
, group
);
1611 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1615 ext4_mb_use_best_found(ac
, e4b
);
1618 ext4_unlock_group(ac
->ac_sb
, group
);
1619 ext4_mb_release_desc(e4b
);
1624 static noinline_for_stack
1625 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1626 struct ext4_buddy
*e4b
)
1628 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1631 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1632 struct ext4_free_extent ex
;
1634 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1637 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1641 ext4_lock_group(ac
->ac_sb
, group
);
1642 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1643 ac
->ac_g_ex
.fe_len
, &ex
);
1645 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1648 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1650 /* use do_div to get remainder (would be 64-bit modulo) */
1651 if (do_div(start
, sbi
->s_stripe
) == 0) {
1654 ext4_mb_use_best_found(ac
, e4b
);
1656 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1657 BUG_ON(ex
.fe_len
<= 0);
1658 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1659 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1662 ext4_mb_use_best_found(ac
, e4b
);
1663 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1664 /* Sometimes, caller may want to merge even small
1665 * number of blocks to an existing extent */
1666 BUG_ON(ex
.fe_len
<= 0);
1667 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1668 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1671 ext4_mb_use_best_found(ac
, e4b
);
1673 ext4_unlock_group(ac
->ac_sb
, group
);
1674 ext4_mb_release_desc(e4b
);
1680 * The routine scans buddy structures (not bitmap!) from given order
1681 * to max order and tries to find big enough chunk to satisfy the req
1683 static noinline_for_stack
1684 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1685 struct ext4_buddy
*e4b
)
1687 struct super_block
*sb
= ac
->ac_sb
;
1688 struct ext4_group_info
*grp
= e4b
->bd_info
;
1694 BUG_ON(ac
->ac_2order
<= 0);
1695 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1696 if (grp
->bb_counters
[i
] == 0)
1699 buddy
= mb_find_buddy(e4b
, i
, &max
);
1700 BUG_ON(buddy
== NULL
);
1702 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1707 ac
->ac_b_ex
.fe_len
= 1 << i
;
1708 ac
->ac_b_ex
.fe_start
= k
<< i
;
1709 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1711 ext4_mb_use_best_found(ac
, e4b
);
1713 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1715 if (EXT4_SB(sb
)->s_mb_stats
)
1716 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1723 * The routine scans the group and measures all found extents.
1724 * In order to optimize scanning, caller must pass number of
1725 * free blocks in the group, so the routine can know upper limit.
1727 static noinline_for_stack
1728 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1729 struct ext4_buddy
*e4b
)
1731 struct super_block
*sb
= ac
->ac_sb
;
1732 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1733 struct ext4_free_extent ex
;
1737 free
= e4b
->bd_info
->bb_free
;
1740 i
= e4b
->bd_info
->bb_first_free
;
1742 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1743 i
= mb_find_next_zero_bit(bitmap
,
1744 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1745 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1747 * IF we have corrupt bitmap, we won't find any
1748 * free blocks even though group info says we
1749 * we have free blocks
1751 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1752 __func__
, "%d free blocks as per "
1753 "group info. But bitmap says 0",
1758 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1759 BUG_ON(ex
.fe_len
<= 0);
1760 if (free
< ex
.fe_len
) {
1761 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1762 __func__
, "%d free blocks as per "
1763 "group info. But got %d blocks",
1766 * The number of free blocks differs. This mostly
1767 * indicate that the bitmap is corrupt. So exit
1768 * without claiming the space.
1773 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1779 ext4_mb_check_limits(ac
, e4b
, 1);
1783 * This is a special case for storages like raid5
1784 * we try to find stripe-aligned chunks for stripe-size requests
1785 * XXX should do so at least for multiples of stripe size as well
1787 static noinline_for_stack
1788 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1789 struct ext4_buddy
*e4b
)
1791 struct super_block
*sb
= ac
->ac_sb
;
1792 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1793 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1794 struct ext4_free_extent ex
;
1795 ext4_fsblk_t first_group_block
;
1800 BUG_ON(sbi
->s_stripe
== 0);
1802 /* find first stripe-aligned block in group */
1803 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1805 a
= first_group_block
+ sbi
->s_stripe
- 1;
1806 do_div(a
, sbi
->s_stripe
);
1807 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1809 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1810 if (!mb_test_bit(i
, bitmap
)) {
1811 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1812 if (max
>= sbi
->s_stripe
) {
1815 ext4_mb_use_best_found(ac
, e4b
);
1823 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1824 ext4_group_t group
, int cr
)
1826 unsigned free
, fragments
;
1828 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
1829 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1831 BUG_ON(cr
< 0 || cr
>= 4);
1832 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp
));
1834 free
= grp
->bb_free
;
1835 fragments
= grp
->bb_fragments
;
1843 BUG_ON(ac
->ac_2order
== 0);
1845 /* Avoid using the first bg of a flexgroup for data files */
1846 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
1847 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
1848 ((group
% flex_size
) == 0))
1851 bits
= ac
->ac_sb
->s_blocksize_bits
+ 1;
1852 for (i
= ac
->ac_2order
; i
<= bits
; i
++)
1853 if (grp
->bb_counters
[i
] > 0)
1857 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1861 if (free
>= ac
->ac_g_ex
.fe_len
)
1874 * lock the group_info alloc_sem of all the groups
1875 * belonging to the same buddy cache page. This
1876 * make sure other parallel operation on the buddy
1877 * cache doesn't happen whild holding the buddy cache
1880 int ext4_mb_get_buddy_cache_lock(struct super_block
*sb
, ext4_group_t group
)
1884 int blocks_per_page
;
1885 int groups_per_page
;
1886 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
1887 ext4_group_t first_group
;
1888 struct ext4_group_info
*grp
;
1890 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1892 * the buddy cache inode stores the block bitmap
1893 * and buddy information in consecutive blocks.
1894 * So for each group we need two blocks.
1897 pnum
= block
/ blocks_per_page
;
1898 first_group
= pnum
* blocks_per_page
/ 2;
1900 groups_per_page
= blocks_per_page
>> 1;
1901 if (groups_per_page
== 0)
1902 groups_per_page
= 1;
1903 /* read all groups the page covers into the cache */
1904 for (i
= 0; i
< groups_per_page
; i
++) {
1906 if ((first_group
+ i
) >= ngroups
)
1908 grp
= ext4_get_group_info(sb
, first_group
+ i
);
1909 /* take all groups write allocation
1910 * semaphore. This make sure there is
1911 * no block allocation going on in any
1914 down_write_nested(&grp
->alloc_sem
, i
);
1919 void ext4_mb_put_buddy_cache_lock(struct super_block
*sb
,
1920 ext4_group_t group
, int locked_group
)
1924 int blocks_per_page
;
1925 ext4_group_t first_group
;
1926 struct ext4_group_info
*grp
;
1928 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1930 * the buddy cache inode stores the block bitmap
1931 * and buddy information in consecutive blocks.
1932 * So for each group we need two blocks.
1935 pnum
= block
/ blocks_per_page
;
1936 first_group
= pnum
* blocks_per_page
/ 2;
1937 /* release locks on all the groups */
1938 for (i
= 0; i
< locked_group
; i
++) {
1940 grp
= ext4_get_group_info(sb
, first_group
+ i
);
1941 /* take all groups write allocation
1942 * semaphore. This make sure there is
1943 * no block allocation going on in any
1946 up_write(&grp
->alloc_sem
);
1951 static noinline_for_stack
int
1952 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
1954 ext4_group_t ngroups
, group
, i
;
1958 struct ext4_sb_info
*sbi
;
1959 struct super_block
*sb
;
1960 struct ext4_buddy e4b
;
1964 ngroups
= ext4_get_groups_count(sb
);
1965 /* non-extent files are limited to low blocks/groups */
1966 if (!(EXT4_I(ac
->ac_inode
)->i_flags
& EXT4_EXTENTS_FL
))
1967 ngroups
= sbi
->s_blockfile_groups
;
1969 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1971 /* first, try the goal */
1972 err
= ext4_mb_find_by_goal(ac
, &e4b
);
1973 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
1976 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
1980 * ac->ac2_order is set only if the fe_len is a power of 2
1981 * if ac2_order is set we also set criteria to 0 so that we
1982 * try exact allocation using buddy.
1984 i
= fls(ac
->ac_g_ex
.fe_len
);
1987 * We search using buddy data only if the order of the request
1988 * is greater than equal to the sbi_s_mb_order2_reqs
1989 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1991 if (i
>= sbi
->s_mb_order2_reqs
) {
1993 * This should tell if fe_len is exactly power of 2
1995 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
1996 ac
->ac_2order
= i
- 1;
1999 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
2001 /* if stream allocation is enabled, use global goal */
2002 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2003 /* TBD: may be hot point */
2004 spin_lock(&sbi
->s_md_lock
);
2005 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2006 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2007 spin_unlock(&sbi
->s_md_lock
);
2010 /* Let's just scan groups to find more-less suitable blocks */
2011 cr
= ac
->ac_2order
? 0 : 1;
2013 * cr == 0 try to get exact allocation,
2014 * cr == 3 try to get anything
2017 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2018 ac
->ac_criteria
= cr
;
2020 * searching for the right group start
2021 * from the goal value specified
2023 group
= ac
->ac_g_ex
.fe_group
;
2025 for (i
= 0; i
< ngroups
; group
++, i
++) {
2026 struct ext4_group_info
*grp
;
2027 struct ext4_group_desc
*desc
;
2029 if (group
== ngroups
)
2032 /* quick check to skip empty groups */
2033 grp
= ext4_get_group_info(sb
, group
);
2034 if (grp
->bb_free
== 0)
2037 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2041 ext4_lock_group(sb
, group
);
2042 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2043 /* someone did allocation from this group */
2044 ext4_unlock_group(sb
, group
);
2045 ext4_mb_release_desc(&e4b
);
2049 ac
->ac_groups_scanned
++;
2050 desc
= ext4_get_group_desc(sb
, group
, NULL
);
2052 ext4_mb_simple_scan_group(ac
, &e4b
);
2054 ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
)
2055 ext4_mb_scan_aligned(ac
, &e4b
);
2057 ext4_mb_complex_scan_group(ac
, &e4b
);
2059 ext4_unlock_group(sb
, group
);
2060 ext4_mb_release_desc(&e4b
);
2062 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2067 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2068 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2070 * We've been searching too long. Let's try to allocate
2071 * the best chunk we've found so far
2074 ext4_mb_try_best_found(ac
, &e4b
);
2075 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2077 * Someone more lucky has already allocated it.
2078 * The only thing we can do is just take first
2080 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2082 ac
->ac_b_ex
.fe_group
= 0;
2083 ac
->ac_b_ex
.fe_start
= 0;
2084 ac
->ac_b_ex
.fe_len
= 0;
2085 ac
->ac_status
= AC_STATUS_CONTINUE
;
2086 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2088 atomic_inc(&sbi
->s_mb_lost_chunks
);
2096 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2098 struct super_block
*sb
= seq
->private;
2101 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2104 return (void *) ((unsigned long) group
);
2107 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2109 struct super_block
*sb
= seq
->private;
2113 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2116 return (void *) ((unsigned long) group
);
2119 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2121 struct super_block
*sb
= seq
->private;
2122 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2125 struct ext4_buddy e4b
;
2127 struct ext4_group_info info
;
2128 ext4_grpblk_t counters
[16];
2133 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2134 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2135 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2136 "group", "free", "frags", "first",
2137 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2138 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2140 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2141 sizeof(struct ext4_group_info
);
2142 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2144 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2147 ext4_lock_group(sb
, group
);
2148 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2149 ext4_unlock_group(sb
, group
);
2150 ext4_mb_release_desc(&e4b
);
2152 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2153 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2154 for (i
= 0; i
<= 13; i
++)
2155 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2156 sg
.info
.bb_counters
[i
] : 0);
2157 seq_printf(seq
, " ]\n");
2162 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2166 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2167 .start
= ext4_mb_seq_groups_start
,
2168 .next
= ext4_mb_seq_groups_next
,
2169 .stop
= ext4_mb_seq_groups_stop
,
2170 .show
= ext4_mb_seq_groups_show
,
2173 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2175 struct super_block
*sb
= PDE(inode
)->data
;
2178 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2180 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2187 static const struct file_operations ext4_mb_seq_groups_fops
= {
2188 .owner
= THIS_MODULE
,
2189 .open
= ext4_mb_seq_groups_open
,
2191 .llseek
= seq_lseek
,
2192 .release
= seq_release
,
2196 /* Create and initialize ext4_group_info data for the given group. */
2197 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2198 struct ext4_group_desc
*desc
)
2202 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2203 struct ext4_group_info
**meta_group_info
;
2206 * First check if this group is the first of a reserved block.
2207 * If it's true, we have to allocate a new table of pointers
2208 * to ext4_group_info structures
2210 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2211 metalen
= sizeof(*meta_group_info
) <<
2212 EXT4_DESC_PER_BLOCK_BITS(sb
);
2213 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2214 if (meta_group_info
== NULL
) {
2215 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2217 goto exit_meta_group_info
;
2219 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2224 * calculate needed size. if change bb_counters size,
2225 * don't forget about ext4_mb_generate_buddy()
2227 len
= offsetof(typeof(**meta_group_info
),
2228 bb_counters
[sb
->s_blocksize_bits
+ 2]);
2231 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2232 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2234 meta_group_info
[i
] = kzalloc(len
, GFP_KERNEL
);
2235 if (meta_group_info
[i
] == NULL
) {
2236 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2237 goto exit_group_info
;
2239 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2240 &(meta_group_info
[i
]->bb_state
));
2243 * initialize bb_free to be able to skip
2244 * empty groups without initialization
2246 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2247 meta_group_info
[i
]->bb_free
=
2248 ext4_free_blocks_after_init(sb
, group
, desc
);
2250 meta_group_info
[i
]->bb_free
=
2251 ext4_free_blks_count(sb
, desc
);
2254 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2255 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2256 meta_group_info
[i
]->bb_free_root
.rb_node
= NULL
;
2260 struct buffer_head
*bh
;
2261 meta_group_info
[i
]->bb_bitmap
=
2262 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2263 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2264 bh
= ext4_read_block_bitmap(sb
, group
);
2266 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2275 /* If a meta_group_info table has been allocated, release it now */
2276 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0)
2277 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2278 exit_meta_group_info
:
2280 } /* ext4_mb_add_groupinfo */
2282 static int ext4_mb_init_backend(struct super_block
*sb
)
2284 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2286 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2287 struct ext4_super_block
*es
= sbi
->s_es
;
2288 int num_meta_group_infos
;
2289 int num_meta_group_infos_max
;
2291 struct ext4_group_desc
*desc
;
2293 /* This is the number of blocks used by GDT */
2294 num_meta_group_infos
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) -
2295 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2298 * This is the total number of blocks used by GDT including
2299 * the number of reserved blocks for GDT.
2300 * The s_group_info array is allocated with this value
2301 * to allow a clean online resize without a complex
2302 * manipulation of pointer.
2303 * The drawback is the unused memory when no resize
2304 * occurs but it's very low in terms of pages
2305 * (see comments below)
2306 * Need to handle this properly when META_BG resizing is allowed
2308 num_meta_group_infos_max
= num_meta_group_infos
+
2309 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2312 * array_size is the size of s_group_info array. We round it
2313 * to the next power of two because this approximation is done
2314 * internally by kmalloc so we can have some more memory
2315 * for free here (e.g. may be used for META_BG resize).
2318 while (array_size
< sizeof(*sbi
->s_group_info
) *
2319 num_meta_group_infos_max
)
2320 array_size
= array_size
<< 1;
2321 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2322 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2323 * So a two level scheme suffices for now. */
2324 sbi
->s_group_info
= kmalloc(array_size
, GFP_KERNEL
);
2325 if (sbi
->s_group_info
== NULL
) {
2326 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2329 sbi
->s_buddy_cache
= new_inode(sb
);
2330 if (sbi
->s_buddy_cache
== NULL
) {
2331 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2334 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2335 for (i
= 0; i
< ngroups
; i
++) {
2336 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2339 "EXT4-fs: can't read descriptor %u\n", i
);
2342 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2350 kfree(ext4_get_group_info(sb
, i
));
2351 i
= num_meta_group_infos
;
2353 kfree(sbi
->s_group_info
[i
]);
2354 iput(sbi
->s_buddy_cache
);
2356 kfree(sbi
->s_group_info
);
2360 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2362 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2368 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2370 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2371 if (sbi
->s_mb_offsets
== NULL
) {
2375 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2376 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2377 if (sbi
->s_mb_maxs
== NULL
) {
2378 kfree(sbi
->s_mb_offsets
);
2382 /* order 0 is regular bitmap */
2383 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2384 sbi
->s_mb_offsets
[0] = 0;
2388 max
= sb
->s_blocksize
<< 2;
2390 sbi
->s_mb_offsets
[i
] = offset
;
2391 sbi
->s_mb_maxs
[i
] = max
;
2392 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2395 } while (i
<= sb
->s_blocksize_bits
+ 1);
2397 /* init file for buddy data */
2398 ret
= ext4_mb_init_backend(sb
);
2400 kfree(sbi
->s_mb_offsets
);
2401 kfree(sbi
->s_mb_maxs
);
2405 spin_lock_init(&sbi
->s_md_lock
);
2406 spin_lock_init(&sbi
->s_bal_lock
);
2408 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2409 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2410 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2411 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2412 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2413 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2415 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2416 if (sbi
->s_locality_groups
== NULL
) {
2417 kfree(sbi
->s_mb_offsets
);
2418 kfree(sbi
->s_mb_maxs
);
2421 for_each_possible_cpu(i
) {
2422 struct ext4_locality_group
*lg
;
2423 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2424 mutex_init(&lg
->lg_mutex
);
2425 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2426 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2427 spin_lock_init(&lg
->lg_prealloc_lock
);
2431 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2432 &ext4_mb_seq_groups_fops
, sb
);
2435 sbi
->s_journal
->j_commit_callback
= release_blocks_on_commit
;
2439 /* need to called with the ext4 group lock held */
2440 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2442 struct ext4_prealloc_space
*pa
;
2443 struct list_head
*cur
, *tmp
;
2446 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2447 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2448 list_del(&pa
->pa_group_list
);
2450 kmem_cache_free(ext4_pspace_cachep
, pa
);
2453 mb_debug(1, "mballoc: %u PAs left\n", count
);
2457 int ext4_mb_release(struct super_block
*sb
)
2459 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2461 int num_meta_group_infos
;
2462 struct ext4_group_info
*grinfo
;
2463 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2465 if (sbi
->s_group_info
) {
2466 for (i
= 0; i
< ngroups
; i
++) {
2467 grinfo
= ext4_get_group_info(sb
, i
);
2469 kfree(grinfo
->bb_bitmap
);
2471 ext4_lock_group(sb
, i
);
2472 ext4_mb_cleanup_pa(grinfo
);
2473 ext4_unlock_group(sb
, i
);
2476 num_meta_group_infos
= (ngroups
+
2477 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2478 EXT4_DESC_PER_BLOCK_BITS(sb
);
2479 for (i
= 0; i
< num_meta_group_infos
; i
++)
2480 kfree(sbi
->s_group_info
[i
]);
2481 kfree(sbi
->s_group_info
);
2483 kfree(sbi
->s_mb_offsets
);
2484 kfree(sbi
->s_mb_maxs
);
2485 if (sbi
->s_buddy_cache
)
2486 iput(sbi
->s_buddy_cache
);
2487 if (sbi
->s_mb_stats
) {
2489 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2490 atomic_read(&sbi
->s_bal_allocated
),
2491 atomic_read(&sbi
->s_bal_reqs
),
2492 atomic_read(&sbi
->s_bal_success
));
2494 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2495 "%u 2^N hits, %u breaks, %u lost\n",
2496 atomic_read(&sbi
->s_bal_ex_scanned
),
2497 atomic_read(&sbi
->s_bal_goals
),
2498 atomic_read(&sbi
->s_bal_2orders
),
2499 atomic_read(&sbi
->s_bal_breaks
),
2500 atomic_read(&sbi
->s_mb_lost_chunks
));
2502 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2503 sbi
->s_mb_buddies_generated
++,
2504 sbi
->s_mb_generation_time
);
2506 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2507 atomic_read(&sbi
->s_mb_preallocated
),
2508 atomic_read(&sbi
->s_mb_discarded
));
2511 free_percpu(sbi
->s_locality_groups
);
2513 remove_proc_entry("mb_groups", sbi
->s_proc
);
2519 * This function is called by the jbd2 layer once the commit has finished,
2520 * so we know we can free the blocks that were released with that commit.
2522 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
)
2524 struct super_block
*sb
= journal
->j_private
;
2525 struct ext4_buddy e4b
;
2526 struct ext4_group_info
*db
;
2527 int err
, count
= 0, count2
= 0;
2528 struct ext4_free_data
*entry
;
2529 struct list_head
*l
, *ltmp
;
2531 list_for_each_safe(l
, ltmp
, &txn
->t_private_list
) {
2532 entry
= list_entry(l
, struct ext4_free_data
, list
);
2534 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2535 entry
->count
, entry
->group
, entry
);
2537 err
= ext4_mb_load_buddy(sb
, entry
->group
, &e4b
);
2538 /* we expect to find existing buddy because it's pinned */
2542 /* there are blocks to put in buddy to make them really free */
2543 count
+= entry
->count
;
2545 ext4_lock_group(sb
, entry
->group
);
2546 /* Take it out of per group rb tree */
2547 rb_erase(&entry
->node
, &(db
->bb_free_root
));
2548 mb_free_blocks(NULL
, &e4b
, entry
->start_blk
, entry
->count
);
2550 if (!db
->bb_free_root
.rb_node
) {
2551 /* No more items in the per group rb tree
2552 * balance refcounts from ext4_mb_free_metadata()
2554 page_cache_release(e4b
.bd_buddy_page
);
2555 page_cache_release(e4b
.bd_bitmap_page
);
2557 ext4_unlock_group(sb
, entry
->group
);
2558 if (test_opt(sb
, DISCARD
)) {
2559 ext4_fsblk_t discard_block
;
2561 discard_block
= entry
->start_blk
+
2562 ext4_group_first_block_no(sb
, entry
->group
);
2563 trace_ext4_discard_blocks(sb
,
2564 (unsigned long long)discard_block
,
2566 sb_issue_discard(sb
, discard_block
, entry
->count
);
2568 kmem_cache_free(ext4_free_ext_cachep
, entry
);
2569 ext4_mb_release_desc(&e4b
);
2572 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2575 #ifdef CONFIG_EXT4_DEBUG
2576 u8 mb_enable_debug __read_mostly
;
2578 static struct dentry
*debugfs_dir
;
2579 static struct dentry
*debugfs_debug
;
2581 static void __init
ext4_create_debugfs_entry(void)
2583 debugfs_dir
= debugfs_create_dir("ext4", NULL
);
2585 debugfs_debug
= debugfs_create_u8("mballoc-debug",
2591 static void ext4_remove_debugfs_entry(void)
2593 debugfs_remove(debugfs_debug
);
2594 debugfs_remove(debugfs_dir
);
2599 static void __init
ext4_create_debugfs_entry(void)
2603 static void ext4_remove_debugfs_entry(void)
2609 int __init
init_ext4_mballoc(void)
2611 ext4_pspace_cachep
=
2612 kmem_cache_create("ext4_prealloc_space",
2613 sizeof(struct ext4_prealloc_space
),
2614 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2615 if (ext4_pspace_cachep
== NULL
)
2619 kmem_cache_create("ext4_alloc_context",
2620 sizeof(struct ext4_allocation_context
),
2621 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2622 if (ext4_ac_cachep
== NULL
) {
2623 kmem_cache_destroy(ext4_pspace_cachep
);
2627 ext4_free_ext_cachep
=
2628 kmem_cache_create("ext4_free_block_extents",
2629 sizeof(struct ext4_free_data
),
2630 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2631 if (ext4_free_ext_cachep
== NULL
) {
2632 kmem_cache_destroy(ext4_pspace_cachep
);
2633 kmem_cache_destroy(ext4_ac_cachep
);
2636 ext4_create_debugfs_entry();
2640 void exit_ext4_mballoc(void)
2643 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2644 * before destroying the slab cache.
2647 kmem_cache_destroy(ext4_pspace_cachep
);
2648 kmem_cache_destroy(ext4_ac_cachep
);
2649 kmem_cache_destroy(ext4_free_ext_cachep
);
2650 ext4_remove_debugfs_entry();
2655 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2656 * Returns 0 if success or error code
2658 static noinline_for_stack
int
2659 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2660 handle_t
*handle
, unsigned int reserv_blks
)
2662 struct buffer_head
*bitmap_bh
= NULL
;
2663 struct ext4_super_block
*es
;
2664 struct ext4_group_desc
*gdp
;
2665 struct buffer_head
*gdp_bh
;
2666 struct ext4_sb_info
*sbi
;
2667 struct super_block
*sb
;
2671 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2672 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2680 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2684 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2689 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2693 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2694 ext4_free_blks_count(sb
, gdp
));
2696 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2700 block
= ac
->ac_b_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
2701 + ac
->ac_b_ex
.fe_start
2702 + le32_to_cpu(es
->s_first_data_block
);
2704 len
= ac
->ac_b_ex
.fe_len
;
2705 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2706 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2707 "fs metadata\n", block
, block
+len
);
2708 /* File system mounted not to panic on error
2709 * Fix the bitmap and repeat the block allocation
2710 * We leak some of the blocks here.
2712 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2713 mb_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2714 ac
->ac_b_ex
.fe_len
);
2715 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2716 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2722 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2723 #ifdef AGGRESSIVE_CHECK
2726 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2727 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2728 bitmap_bh
->b_data
));
2732 mb_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,ac
->ac_b_ex
.fe_len
);
2733 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2734 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2735 ext4_free_blks_set(sb
, gdp
,
2736 ext4_free_blocks_after_init(sb
,
2737 ac
->ac_b_ex
.fe_group
, gdp
));
2739 len
= ext4_free_blks_count(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
2740 ext4_free_blks_set(sb
, gdp
, len
);
2741 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
2743 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2744 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
2746 * Now reduce the dirty block count also. Should not go negative
2748 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2749 /* release all the reserved blocks if non delalloc */
2750 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
, reserv_blks
);
2752 if (sbi
->s_log_groups_per_flex
) {
2753 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2754 ac
->ac_b_ex
.fe_group
);
2755 atomic_sub(ac
->ac_b_ex
.fe_len
,
2756 &sbi
->s_flex_groups
[flex_group
].free_blocks
);
2759 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2762 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
2771 * here we normalize request for locality group
2772 * Group request are normalized to s_strip size if we set the same via mount
2773 * option. If not we set it to s_mb_group_prealloc which can be configured via
2774 * /sys/fs/ext4/<partition>/mb_group_prealloc
2776 * XXX: should we try to preallocate more than the group has now?
2778 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2780 struct super_block
*sb
= ac
->ac_sb
;
2781 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2784 if (EXT4_SB(sb
)->s_stripe
)
2785 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
2787 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2788 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2789 current
->pid
, ac
->ac_g_ex
.fe_len
);
2793 * Normalization means making request better in terms of
2794 * size and alignment
2796 static noinline_for_stack
void
2797 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
2798 struct ext4_allocation_request
*ar
)
2802 loff_t size
, orig_size
, start_off
;
2803 ext4_lblk_t start
, orig_start
;
2804 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
2805 struct ext4_prealloc_space
*pa
;
2807 /* do normalize only data requests, metadata requests
2808 do not need preallocation */
2809 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
2812 /* sometime caller may want exact blocks */
2813 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2816 /* caller may indicate that preallocation isn't
2817 * required (it's a tail, for example) */
2818 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
2821 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
2822 ext4_mb_normalize_group_request(ac
);
2826 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
2828 /* first, let's learn actual file size
2829 * given current request is allocated */
2830 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
2831 size
= size
<< bsbits
;
2832 if (size
< i_size_read(ac
->ac_inode
))
2833 size
= i_size_read(ac
->ac_inode
);
2835 /* max size of free chunks */
2838 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2839 (req <= (size) || max <= (chunk_size))
2841 /* first, try to predict filesize */
2842 /* XXX: should this table be tunable? */
2844 if (size
<= 16 * 1024) {
2846 } else if (size
<= 32 * 1024) {
2848 } else if (size
<= 64 * 1024) {
2850 } else if (size
<= 128 * 1024) {
2852 } else if (size
<= 256 * 1024) {
2854 } else if (size
<= 512 * 1024) {
2856 } else if (size
<= 1024 * 1024) {
2858 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
2859 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2860 (21 - bsbits
)) << 21;
2861 size
= 2 * 1024 * 1024;
2862 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
2863 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2864 (22 - bsbits
)) << 22;
2865 size
= 4 * 1024 * 1024;
2866 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
2867 (8<<20)>>bsbits
, max
, 8 * 1024)) {
2868 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2869 (23 - bsbits
)) << 23;
2870 size
= 8 * 1024 * 1024;
2872 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
2873 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
2875 orig_size
= size
= size
>> bsbits
;
2876 orig_start
= start
= start_off
>> bsbits
;
2878 /* don't cover already allocated blocks in selected range */
2879 if (ar
->pleft
&& start
<= ar
->lleft
) {
2880 size
-= ar
->lleft
+ 1 - start
;
2881 start
= ar
->lleft
+ 1;
2883 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
2884 size
-= start
+ size
- ar
->lright
;
2888 /* check we don't cross already preallocated blocks */
2890 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
2895 spin_lock(&pa
->pa_lock
);
2896 if (pa
->pa_deleted
) {
2897 spin_unlock(&pa
->pa_lock
);
2901 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
2903 /* PA must not overlap original request */
2904 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
2905 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
2907 /* skip PAs this normalized request doesn't overlap with */
2908 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
2909 spin_unlock(&pa
->pa_lock
);
2912 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
2914 /* adjust start or end to be adjacent to this pa */
2915 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
2916 BUG_ON(pa_end
< start
);
2918 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
2919 BUG_ON(pa
->pa_lstart
> end
);
2920 end
= pa
->pa_lstart
;
2922 spin_unlock(&pa
->pa_lock
);
2927 /* XXX: extra loop to check we really don't overlap preallocations */
2929 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
2931 spin_lock(&pa
->pa_lock
);
2932 if (pa
->pa_deleted
== 0) {
2933 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
2934 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
2936 spin_unlock(&pa
->pa_lock
);
2940 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
2941 start
> ac
->ac_o_ex
.fe_logical
) {
2942 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
2943 (unsigned long) start
, (unsigned long) size
,
2944 (unsigned long) ac
->ac_o_ex
.fe_logical
);
2946 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
2947 start
> ac
->ac_o_ex
.fe_logical
);
2948 BUG_ON(size
<= 0 || size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
2950 /* now prepare goal request */
2952 /* XXX: is it better to align blocks WRT to logical
2953 * placement or satisfy big request as is */
2954 ac
->ac_g_ex
.fe_logical
= start
;
2955 ac
->ac_g_ex
.fe_len
= size
;
2957 /* define goal start in order to merge */
2958 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
2959 /* merge to the right */
2960 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
2961 &ac
->ac_f_ex
.fe_group
,
2962 &ac
->ac_f_ex
.fe_start
);
2963 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
2965 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
2966 /* merge to the left */
2967 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
2968 &ac
->ac_f_ex
.fe_group
,
2969 &ac
->ac_f_ex
.fe_start
);
2970 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
2973 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
2974 (unsigned) orig_size
, (unsigned) start
);
2977 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
2979 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2981 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
2982 atomic_inc(&sbi
->s_bal_reqs
);
2983 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
2984 if (ac
->ac_o_ex
.fe_len
>= ac
->ac_g_ex
.fe_len
)
2985 atomic_inc(&sbi
->s_bal_success
);
2986 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
2987 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
2988 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
2989 atomic_inc(&sbi
->s_bal_goals
);
2990 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
2991 atomic_inc(&sbi
->s_bal_breaks
);
2994 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
2995 trace_ext4_mballoc_alloc(ac
);
2997 trace_ext4_mballoc_prealloc(ac
);
3001 * Called on failure; free up any blocks from the inode PA for this
3002 * context. We don't need this for MB_GROUP_PA because we only change
3003 * pa_free in ext4_mb_release_context(), but on failure, we've already
3004 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3006 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3008 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3011 if (pa
&& pa
->pa_type
== MB_INODE_PA
) {
3012 len
= ac
->ac_b_ex
.fe_len
;
3019 * use blocks preallocated to inode
3021 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3022 struct ext4_prealloc_space
*pa
)
3028 /* found preallocated blocks, use them */
3029 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3030 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3032 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3033 &ac
->ac_b_ex
.fe_start
);
3034 ac
->ac_b_ex
.fe_len
= len
;
3035 ac
->ac_status
= AC_STATUS_FOUND
;
3038 BUG_ON(start
< pa
->pa_pstart
);
3039 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3040 BUG_ON(pa
->pa_free
< len
);
3043 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3047 * use blocks preallocated to locality group
3049 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3050 struct ext4_prealloc_space
*pa
)
3052 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3054 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3055 &ac
->ac_b_ex
.fe_group
,
3056 &ac
->ac_b_ex
.fe_start
);
3057 ac
->ac_b_ex
.fe_len
= len
;
3058 ac
->ac_status
= AC_STATUS_FOUND
;
3061 /* we don't correct pa_pstart or pa_plen here to avoid
3062 * possible race when the group is being loaded concurrently
3063 * instead we correct pa later, after blocks are marked
3064 * in on-disk bitmap -- see ext4_mb_release_context()
3065 * Other CPUs are prevented from allocating from this pa by lg_mutex
3067 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3071 * Return the prealloc space that have minimal distance
3072 * from the goal block. @cpa is the prealloc
3073 * space that is having currently known minimal distance
3074 * from the goal block.
3076 static struct ext4_prealloc_space
*
3077 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3078 struct ext4_prealloc_space
*pa
,
3079 struct ext4_prealloc_space
*cpa
)
3081 ext4_fsblk_t cur_distance
, new_distance
;
3084 atomic_inc(&pa
->pa_count
);
3087 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3088 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3090 if (cur_distance
< new_distance
)
3093 /* drop the previous reference */
3094 atomic_dec(&cpa
->pa_count
);
3095 atomic_inc(&pa
->pa_count
);
3100 * search goal blocks in preallocated space
3102 static noinline_for_stack
int
3103 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3106 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3107 struct ext4_locality_group
*lg
;
3108 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3109 ext4_fsblk_t goal_block
;
3111 /* only data can be preallocated */
3112 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3115 /* first, try per-file preallocation */
3117 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3119 /* all fields in this condition don't change,
3120 * so we can skip locking for them */
3121 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3122 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3125 /* non-extent files can't have physical blocks past 2^32 */
3126 if (!(EXT4_I(ac
->ac_inode
)->i_flags
& EXT4_EXTENTS_FL
) &&
3127 pa
->pa_pstart
+ pa
->pa_len
> EXT4_MAX_BLOCK_FILE_PHYS
)
3130 /* found preallocated blocks, use them */
3131 spin_lock(&pa
->pa_lock
);
3132 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3133 atomic_inc(&pa
->pa_count
);
3134 ext4_mb_use_inode_pa(ac
, pa
);
3135 spin_unlock(&pa
->pa_lock
);
3136 ac
->ac_criteria
= 10;
3140 spin_unlock(&pa
->pa_lock
);
3144 /* can we use group allocation? */
3145 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3148 /* inode may have no locality group for some reason */
3152 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3153 if (order
> PREALLOC_TB_SIZE
- 1)
3154 /* The max size of hash table is PREALLOC_TB_SIZE */
3155 order
= PREALLOC_TB_SIZE
- 1;
3157 goal_block
= ac
->ac_g_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
) +
3158 ac
->ac_g_ex
.fe_start
+
3159 le32_to_cpu(EXT4_SB(ac
->ac_sb
)->s_es
->s_first_data_block
);
3161 * search for the prealloc space that is having
3162 * minimal distance from the goal block.
3164 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3166 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3168 spin_lock(&pa
->pa_lock
);
3169 if (pa
->pa_deleted
== 0 &&
3170 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3172 cpa
= ext4_mb_check_group_pa(goal_block
,
3175 spin_unlock(&pa
->pa_lock
);
3180 ext4_mb_use_group_pa(ac
, cpa
);
3181 ac
->ac_criteria
= 20;
3188 * the function goes through all block freed in the group
3189 * but not yet committed and marks them used in in-core bitmap.
3190 * buddy must be generated from this bitmap
3191 * Need to be called with the ext4 group lock held
3193 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3197 struct ext4_group_info
*grp
;
3198 struct ext4_free_data
*entry
;
3200 grp
= ext4_get_group_info(sb
, group
);
3201 n
= rb_first(&(grp
->bb_free_root
));
3204 entry
= rb_entry(n
, struct ext4_free_data
, node
);
3205 mb_set_bits(bitmap
, entry
->start_blk
, entry
->count
);
3212 * the function goes through all preallocation in this group and marks them
3213 * used in in-core bitmap. buddy must be generated from this bitmap
3214 * Need to be called with ext4 group lock held
3216 static noinline_for_stack
3217 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3220 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3221 struct ext4_prealloc_space
*pa
;
3222 struct list_head
*cur
;
3223 ext4_group_t groupnr
;
3224 ext4_grpblk_t start
;
3225 int preallocated
= 0;
3229 /* all form of preallocation discards first load group,
3230 * so the only competing code is preallocation use.
3231 * we don't need any locking here
3232 * notice we do NOT ignore preallocations with pa_deleted
3233 * otherwise we could leave used blocks available for
3234 * allocation in buddy when concurrent ext4_mb_put_pa()
3235 * is dropping preallocation
3237 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3238 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3239 spin_lock(&pa
->pa_lock
);
3240 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3243 spin_unlock(&pa
->pa_lock
);
3244 if (unlikely(len
== 0))
3246 BUG_ON(groupnr
!= group
);
3247 mb_set_bits(bitmap
, start
, len
);
3248 preallocated
+= len
;
3251 mb_debug(1, "prellocated %u for group %u\n", preallocated
, group
);
3254 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3256 struct ext4_prealloc_space
*pa
;
3257 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3258 kmem_cache_free(ext4_pspace_cachep
, pa
);
3262 * drops a reference to preallocated space descriptor
3263 * if this was the last reference and the space is consumed
3265 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3266 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3269 ext4_fsblk_t grp_blk
;
3271 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3274 /* in this short window concurrent discard can set pa_deleted */
3275 spin_lock(&pa
->pa_lock
);
3276 if (pa
->pa_deleted
== 1) {
3277 spin_unlock(&pa
->pa_lock
);
3282 spin_unlock(&pa
->pa_lock
);
3284 grp_blk
= pa
->pa_pstart
;
3286 * If doing group-based preallocation, pa_pstart may be in the
3287 * next group when pa is used up
3289 if (pa
->pa_type
== MB_GROUP_PA
)
3292 ext4_get_group_no_and_offset(sb
, grp_blk
, &grp
, NULL
);
3297 * P1 (buddy init) P2 (regular allocation)
3298 * find block B in PA
3299 * copy on-disk bitmap to buddy
3300 * mark B in on-disk bitmap
3301 * drop PA from group
3302 * mark all PAs in buddy
3304 * thus, P1 initializes buddy with B available. to prevent this
3305 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3308 ext4_lock_group(sb
, grp
);
3309 list_del(&pa
->pa_group_list
);
3310 ext4_unlock_group(sb
, grp
);
3312 spin_lock(pa
->pa_obj_lock
);
3313 list_del_rcu(&pa
->pa_inode_list
);
3314 spin_unlock(pa
->pa_obj_lock
);
3316 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3320 * creates new preallocated space for given inode
3322 static noinline_for_stack
int
3323 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3325 struct super_block
*sb
= ac
->ac_sb
;
3326 struct ext4_prealloc_space
*pa
;
3327 struct ext4_group_info
*grp
;
3328 struct ext4_inode_info
*ei
;
3330 /* preallocate only when found space is larger then requested */
3331 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3332 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3333 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3335 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3339 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3345 /* we can't allocate as much as normalizer wants.
3346 * so, found space must get proper lstart
3347 * to cover original request */
3348 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3349 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3351 /* we're limited by original request in that
3352 * logical block must be covered any way
3353 * winl is window we can move our chunk within */
3354 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3356 /* also, we should cover whole original request */
3357 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3359 /* the smallest one defines real window */
3360 win
= min(winl
, wins
);
3362 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3363 if (offs
&& offs
< win
)
3366 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3367 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3368 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3371 /* preallocation can change ac_b_ex, thus we store actually
3372 * allocated blocks for history */
3373 ac
->ac_f_ex
= ac
->ac_b_ex
;
3375 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3376 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3377 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3378 pa
->pa_free
= pa
->pa_len
;
3379 atomic_set(&pa
->pa_count
, 1);
3380 spin_lock_init(&pa
->pa_lock
);
3381 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3382 INIT_LIST_HEAD(&pa
->pa_group_list
);
3384 pa
->pa_type
= MB_INODE_PA
;
3386 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3387 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3388 trace_ext4_mb_new_inode_pa(ac
, pa
);
3390 ext4_mb_use_inode_pa(ac
, pa
);
3391 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3393 ei
= EXT4_I(ac
->ac_inode
);
3394 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3396 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3397 pa
->pa_inode
= ac
->ac_inode
;
3399 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3400 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3401 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3403 spin_lock(pa
->pa_obj_lock
);
3404 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3405 spin_unlock(pa
->pa_obj_lock
);
3411 * creates new preallocated space for locality group inodes belongs to
3413 static noinline_for_stack
int
3414 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3416 struct super_block
*sb
= ac
->ac_sb
;
3417 struct ext4_locality_group
*lg
;
3418 struct ext4_prealloc_space
*pa
;
3419 struct ext4_group_info
*grp
;
3421 /* preallocate only when found space is larger then requested */
3422 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3423 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3424 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3426 BUG_ON(ext4_pspace_cachep
== NULL
);
3427 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3431 /* preallocation can change ac_b_ex, thus we store actually
3432 * allocated blocks for history */
3433 ac
->ac_f_ex
= ac
->ac_b_ex
;
3435 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3436 pa
->pa_lstart
= pa
->pa_pstart
;
3437 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3438 pa
->pa_free
= pa
->pa_len
;
3439 atomic_set(&pa
->pa_count
, 1);
3440 spin_lock_init(&pa
->pa_lock
);
3441 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3442 INIT_LIST_HEAD(&pa
->pa_group_list
);
3444 pa
->pa_type
= MB_GROUP_PA
;
3446 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3447 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3448 trace_ext4_mb_new_group_pa(ac
, pa
);
3450 ext4_mb_use_group_pa(ac
, pa
);
3451 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3453 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3457 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3458 pa
->pa_inode
= NULL
;
3460 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3461 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3462 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3465 * We will later add the new pa to the right bucket
3466 * after updating the pa_free in ext4_mb_release_context
3471 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3475 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3476 err
= ext4_mb_new_group_pa(ac
);
3478 err
= ext4_mb_new_inode_pa(ac
);
3483 * finds all unused blocks in on-disk bitmap, frees them in
3484 * in-core bitmap and buddy.
3485 * @pa must be unlinked from inode and group lists, so that
3486 * nobody else can find/use it.
3487 * the caller MUST hold group/inode locks.
3488 * TODO: optimize the case when there are no in-core structures yet
3490 static noinline_for_stack
int
3491 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3492 struct ext4_prealloc_space
*pa
,
3493 struct ext4_allocation_context
*ac
)
3495 struct super_block
*sb
= e4b
->bd_sb
;
3496 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3501 unsigned long long grp_blk_start
;
3506 BUG_ON(pa
->pa_deleted
== 0);
3507 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3508 grp_blk_start
= pa
->pa_pstart
- bit
;
3509 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3510 end
= bit
+ pa
->pa_len
;
3514 ac
->ac_inode
= pa
->pa_inode
;
3518 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3521 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3522 start
= ext4_group_first_block_no(sb
, group
) + bit
;
3523 mb_debug(1, " free preallocated %u/%u in group %u\n",
3524 (unsigned) start
, (unsigned) next
- bit
,
3529 ac
->ac_b_ex
.fe_group
= group
;
3530 ac
->ac_b_ex
.fe_start
= bit
;
3531 ac
->ac_b_ex
.fe_len
= next
- bit
;
3532 ac
->ac_b_ex
.fe_logical
= 0;
3533 trace_ext4_mballoc_discard(ac
);
3536 trace_ext4_mb_release_inode_pa(ac
, pa
, grp_blk_start
+ bit
,
3538 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3541 if (free
!= pa
->pa_free
) {
3542 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3543 pa
, (unsigned long) pa
->pa_lstart
,
3544 (unsigned long) pa
->pa_pstart
,
3545 (unsigned long) pa
->pa_len
);
3546 ext4_grp_locked_error(sb
, group
,
3547 __func__
, "free %u, pa_free %u",
3550 * pa is already deleted so we use the value obtained
3551 * from the bitmap and continue.
3554 atomic_add(free
, &sbi
->s_mb_discarded
);
3559 static noinline_for_stack
int
3560 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3561 struct ext4_prealloc_space
*pa
,
3562 struct ext4_allocation_context
*ac
)
3564 struct super_block
*sb
= e4b
->bd_sb
;
3568 trace_ext4_mb_release_group_pa(ac
, pa
);
3569 BUG_ON(pa
->pa_deleted
== 0);
3570 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3571 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3572 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3573 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3577 ac
->ac_inode
= NULL
;
3578 ac
->ac_b_ex
.fe_group
= group
;
3579 ac
->ac_b_ex
.fe_start
= bit
;
3580 ac
->ac_b_ex
.fe_len
= pa
->pa_len
;
3581 ac
->ac_b_ex
.fe_logical
= 0;
3582 trace_ext4_mballoc_discard(ac
);
3589 * releases all preallocations in given group
3591 * first, we need to decide discard policy:
3592 * - when do we discard
3594 * - how many do we discard
3595 * 1) how many requested
3597 static noinline_for_stack
int
3598 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3599 ext4_group_t group
, int needed
)
3601 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3602 struct buffer_head
*bitmap_bh
= NULL
;
3603 struct ext4_prealloc_space
*pa
, *tmp
;
3604 struct ext4_allocation_context
*ac
;
3605 struct list_head list
;
3606 struct ext4_buddy e4b
;
3611 mb_debug(1, "discard preallocation for group %u\n", group
);
3613 if (list_empty(&grp
->bb_prealloc_list
))
3616 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3617 if (bitmap_bh
== NULL
) {
3618 ext4_error(sb
, "Error reading block bitmap for %u", group
);
3622 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3624 ext4_error(sb
, "Error loading buddy information for %u", group
);
3630 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3632 INIT_LIST_HEAD(&list
);
3633 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3637 ext4_lock_group(sb
, group
);
3638 list_for_each_entry_safe(pa
, tmp
,
3639 &grp
->bb_prealloc_list
, pa_group_list
) {
3640 spin_lock(&pa
->pa_lock
);
3641 if (atomic_read(&pa
->pa_count
)) {
3642 spin_unlock(&pa
->pa_lock
);
3646 if (pa
->pa_deleted
) {
3647 spin_unlock(&pa
->pa_lock
);
3651 /* seems this one can be freed ... */
3654 /* we can trust pa_free ... */
3655 free
+= pa
->pa_free
;
3657 spin_unlock(&pa
->pa_lock
);
3659 list_del(&pa
->pa_group_list
);
3660 list_add(&pa
->u
.pa_tmp_list
, &list
);
3663 /* if we still need more blocks and some PAs were used, try again */
3664 if (free
< needed
&& busy
) {
3666 ext4_unlock_group(sb
, group
);
3668 * Yield the CPU here so that we don't get soft lockup
3669 * in non preempt case.
3675 /* found anything to free? */
3676 if (list_empty(&list
)) {
3681 /* now free all selected PAs */
3682 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3684 /* remove from object (inode or locality group) */
3685 spin_lock(pa
->pa_obj_lock
);
3686 list_del_rcu(&pa
->pa_inode_list
);
3687 spin_unlock(pa
->pa_obj_lock
);
3689 if (pa
->pa_type
== MB_GROUP_PA
)
3690 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
3692 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3694 list_del(&pa
->u
.pa_tmp_list
);
3695 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3699 ext4_unlock_group(sb
, group
);
3701 kmem_cache_free(ext4_ac_cachep
, ac
);
3702 ext4_mb_release_desc(&e4b
);
3708 * releases all non-used preallocated blocks for given inode
3710 * It's important to discard preallocations under i_data_sem
3711 * We don't want another block to be served from the prealloc
3712 * space when we are discarding the inode prealloc space.
3714 * FIXME!! Make sure it is valid at all the call sites
3716 void ext4_discard_preallocations(struct inode
*inode
)
3718 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3719 struct super_block
*sb
= inode
->i_sb
;
3720 struct buffer_head
*bitmap_bh
= NULL
;
3721 struct ext4_prealloc_space
*pa
, *tmp
;
3722 struct ext4_allocation_context
*ac
;
3723 ext4_group_t group
= 0;
3724 struct list_head list
;
3725 struct ext4_buddy e4b
;
3728 if (!S_ISREG(inode
->i_mode
)) {
3729 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3733 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
3734 trace_ext4_discard_preallocations(inode
);
3736 INIT_LIST_HEAD(&list
);
3738 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3741 ac
->ac_inode
= inode
;
3744 /* first, collect all pa's in the inode */
3745 spin_lock(&ei
->i_prealloc_lock
);
3746 while (!list_empty(&ei
->i_prealloc_list
)) {
3747 pa
= list_entry(ei
->i_prealloc_list
.next
,
3748 struct ext4_prealloc_space
, pa_inode_list
);
3749 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3750 spin_lock(&pa
->pa_lock
);
3751 if (atomic_read(&pa
->pa_count
)) {
3752 /* this shouldn't happen often - nobody should
3753 * use preallocation while we're discarding it */
3754 spin_unlock(&pa
->pa_lock
);
3755 spin_unlock(&ei
->i_prealloc_lock
);
3756 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
3758 schedule_timeout_uninterruptible(HZ
);
3762 if (pa
->pa_deleted
== 0) {
3764 spin_unlock(&pa
->pa_lock
);
3765 list_del_rcu(&pa
->pa_inode_list
);
3766 list_add(&pa
->u
.pa_tmp_list
, &list
);
3770 /* someone is deleting pa right now */
3771 spin_unlock(&pa
->pa_lock
);
3772 spin_unlock(&ei
->i_prealloc_lock
);
3774 /* we have to wait here because pa_deleted
3775 * doesn't mean pa is already unlinked from
3776 * the list. as we might be called from
3777 * ->clear_inode() the inode will get freed
3778 * and concurrent thread which is unlinking
3779 * pa from inode's list may access already
3780 * freed memory, bad-bad-bad */
3782 /* XXX: if this happens too often, we can
3783 * add a flag to force wait only in case
3784 * of ->clear_inode(), but not in case of
3785 * regular truncate */
3786 schedule_timeout_uninterruptible(HZ
);
3789 spin_unlock(&ei
->i_prealloc_lock
);
3791 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3792 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
3793 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
3795 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3797 ext4_error(sb
, "Error loading buddy information for %u",
3802 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3803 if (bitmap_bh
== NULL
) {
3804 ext4_error(sb
, "Error reading block bitmap for %u",
3806 ext4_mb_release_desc(&e4b
);
3810 ext4_lock_group(sb
, group
);
3811 list_del(&pa
->pa_group_list
);
3812 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3813 ext4_unlock_group(sb
, group
);
3815 ext4_mb_release_desc(&e4b
);
3818 list_del(&pa
->u
.pa_tmp_list
);
3819 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3822 kmem_cache_free(ext4_ac_cachep
, ac
);
3826 * finds all preallocated spaces and return blocks being freed to them
3827 * if preallocated space becomes full (no block is used from the space)
3828 * then the function frees space in buddy
3829 * XXX: at the moment, truncate (which is the only way to free blocks)
3830 * discards all preallocations
3832 static void ext4_mb_return_to_preallocation(struct inode
*inode
,
3833 struct ext4_buddy
*e4b
,
3834 sector_t block
, int count
)
3836 BUG_ON(!list_empty(&EXT4_I(inode
)->i_prealloc_list
));
3838 #ifdef CONFIG_EXT4_DEBUG
3839 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3841 struct super_block
*sb
= ac
->ac_sb
;
3842 ext4_group_t ngroups
, i
;
3844 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
3845 " Allocation context details:\n");
3846 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
3847 ac
->ac_status
, ac
->ac_flags
);
3848 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3849 "best %lu/%lu/%lu@%lu cr %d\n",
3850 (unsigned long)ac
->ac_o_ex
.fe_group
,
3851 (unsigned long)ac
->ac_o_ex
.fe_start
,
3852 (unsigned long)ac
->ac_o_ex
.fe_len
,
3853 (unsigned long)ac
->ac_o_ex
.fe_logical
,
3854 (unsigned long)ac
->ac_g_ex
.fe_group
,
3855 (unsigned long)ac
->ac_g_ex
.fe_start
,
3856 (unsigned long)ac
->ac_g_ex
.fe_len
,
3857 (unsigned long)ac
->ac_g_ex
.fe_logical
,
3858 (unsigned long)ac
->ac_b_ex
.fe_group
,
3859 (unsigned long)ac
->ac_b_ex
.fe_start
,
3860 (unsigned long)ac
->ac_b_ex
.fe_len
,
3861 (unsigned long)ac
->ac_b_ex
.fe_logical
,
3862 (int)ac
->ac_criteria
);
3863 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
3865 printk(KERN_ERR
"EXT4-fs: groups: \n");
3866 ngroups
= ext4_get_groups_count(sb
);
3867 for (i
= 0; i
< ngroups
; i
++) {
3868 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
3869 struct ext4_prealloc_space
*pa
;
3870 ext4_grpblk_t start
;
3871 struct list_head
*cur
;
3872 ext4_lock_group(sb
, i
);
3873 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3874 pa
= list_entry(cur
, struct ext4_prealloc_space
,
3876 spin_lock(&pa
->pa_lock
);
3877 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3879 spin_unlock(&pa
->pa_lock
);
3880 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
3883 ext4_unlock_group(sb
, i
);
3885 if (grp
->bb_free
== 0)
3887 printk(KERN_ERR
"%u: %d/%d \n",
3888 i
, grp
->bb_free
, grp
->bb_fragments
);
3890 printk(KERN_ERR
"\n");
3893 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3900 * We use locality group preallocation for small size file. The size of the
3901 * file is determined by the current size or the resulting size after
3902 * allocation which ever is larger
3904 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3906 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
3908 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3909 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3912 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3915 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3918 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3919 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
3922 if ((size
== isize
) &&
3923 !ext4_fs_is_busy(sbi
) &&
3924 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
3925 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
3929 /* don't use group allocation for large files */
3930 size
= max(size
, isize
);
3931 if (size
> sbi
->s_mb_stream_request
) {
3932 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
3936 BUG_ON(ac
->ac_lg
!= NULL
);
3938 * locality group prealloc space are per cpu. The reason for having
3939 * per cpu locality group is to reduce the contention between block
3940 * request from multiple CPUs.
3942 ac
->ac_lg
= __this_cpu_ptr(sbi
->s_locality_groups
);
3944 /* we're going to use group allocation */
3945 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
3947 /* serialize all allocations in the group */
3948 mutex_lock(&ac
->ac_lg
->lg_mutex
);
3951 static noinline_for_stack
int
3952 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
3953 struct ext4_allocation_request
*ar
)
3955 struct super_block
*sb
= ar
->inode
->i_sb
;
3956 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3957 struct ext4_super_block
*es
= sbi
->s_es
;
3961 ext4_grpblk_t block
;
3963 /* we can't allocate > group size */
3966 /* just a dirty hack to filter too big requests */
3967 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
3968 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
3970 /* start searching from the goal */
3972 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
3973 goal
>= ext4_blocks_count(es
))
3974 goal
= le32_to_cpu(es
->s_first_data_block
);
3975 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
3977 /* set up allocation goals */
3978 memset(ac
, 0, sizeof(struct ext4_allocation_context
));
3979 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
3980 ac
->ac_status
= AC_STATUS_CONTINUE
;
3982 ac
->ac_inode
= ar
->inode
;
3983 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
3984 ac
->ac_o_ex
.fe_group
= group
;
3985 ac
->ac_o_ex
.fe_start
= block
;
3986 ac
->ac_o_ex
.fe_len
= len
;
3987 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
3988 ac
->ac_g_ex
.fe_group
= group
;
3989 ac
->ac_g_ex
.fe_start
= block
;
3990 ac
->ac_g_ex
.fe_len
= len
;
3991 ac
->ac_flags
= ar
->flags
;
3993 /* we have to define context: we'll we work with a file or
3994 * locality group. this is a policy, actually */
3995 ext4_mb_group_or_file(ac
);
3997 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
3998 "left: %u/%u, right %u/%u to %swritable\n",
3999 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4000 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4001 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4002 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4003 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4008 static noinline_for_stack
void
4009 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4010 struct ext4_locality_group
*lg
,
4011 int order
, int total_entries
)
4013 ext4_group_t group
= 0;
4014 struct ext4_buddy e4b
;
4015 struct list_head discard_list
;
4016 struct ext4_prealloc_space
*pa
, *tmp
;
4017 struct ext4_allocation_context
*ac
;
4019 mb_debug(1, "discard locality group preallocation\n");
4021 INIT_LIST_HEAD(&discard_list
);
4022 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4026 spin_lock(&lg
->lg_prealloc_lock
);
4027 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4029 spin_lock(&pa
->pa_lock
);
4030 if (atomic_read(&pa
->pa_count
)) {
4032 * This is the pa that we just used
4033 * for block allocation. So don't
4036 spin_unlock(&pa
->pa_lock
);
4039 if (pa
->pa_deleted
) {
4040 spin_unlock(&pa
->pa_lock
);
4043 /* only lg prealloc space */
4044 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4046 /* seems this one can be freed ... */
4048 spin_unlock(&pa
->pa_lock
);
4050 list_del_rcu(&pa
->pa_inode_list
);
4051 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4054 if (total_entries
<= 5) {
4056 * we want to keep only 5 entries
4057 * allowing it to grow to 8. This
4058 * mak sure we don't call discard
4059 * soon for this list.
4064 spin_unlock(&lg
->lg_prealloc_lock
);
4066 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4068 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4069 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4070 ext4_error(sb
, "Error loading buddy information for %u",
4074 ext4_lock_group(sb
, group
);
4075 list_del(&pa
->pa_group_list
);
4076 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
4077 ext4_unlock_group(sb
, group
);
4079 ext4_mb_release_desc(&e4b
);
4080 list_del(&pa
->u
.pa_tmp_list
);
4081 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4084 kmem_cache_free(ext4_ac_cachep
, ac
);
4088 * We have incremented pa_count. So it cannot be freed at this
4089 * point. Also we hold lg_mutex. So no parallel allocation is
4090 * possible from this lg. That means pa_free cannot be updated.
4092 * A parallel ext4_mb_discard_group_preallocations is possible.
4093 * which can cause the lg_prealloc_list to be updated.
4096 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4098 int order
, added
= 0, lg_prealloc_count
= 1;
4099 struct super_block
*sb
= ac
->ac_sb
;
4100 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4101 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4103 order
= fls(pa
->pa_free
) - 1;
4104 if (order
> PREALLOC_TB_SIZE
- 1)
4105 /* The max size of hash table is PREALLOC_TB_SIZE */
4106 order
= PREALLOC_TB_SIZE
- 1;
4107 /* Add the prealloc space to lg */
4109 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4111 spin_lock(&tmp_pa
->pa_lock
);
4112 if (tmp_pa
->pa_deleted
) {
4113 spin_unlock(&tmp_pa
->pa_lock
);
4116 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4117 /* Add to the tail of the previous entry */
4118 list_add_tail_rcu(&pa
->pa_inode_list
,
4119 &tmp_pa
->pa_inode_list
);
4122 * we want to count the total
4123 * number of entries in the list
4126 spin_unlock(&tmp_pa
->pa_lock
);
4127 lg_prealloc_count
++;
4130 list_add_tail_rcu(&pa
->pa_inode_list
,
4131 &lg
->lg_prealloc_list
[order
]);
4134 /* Now trim the list to be not more than 8 elements */
4135 if (lg_prealloc_count
> 8) {
4136 ext4_mb_discard_lg_preallocations(sb
, lg
,
4137 order
, lg_prealloc_count
);
4144 * release all resource we used in allocation
4146 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4148 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4150 if (pa
->pa_type
== MB_GROUP_PA
) {
4151 /* see comment in ext4_mb_use_group_pa() */
4152 spin_lock(&pa
->pa_lock
);
4153 pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4154 pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4155 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4156 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4157 spin_unlock(&pa
->pa_lock
);
4161 up_read(ac
->alloc_semp
);
4164 * We want to add the pa to the right bucket.
4165 * Remove it from the list and while adding
4166 * make sure the list to which we are adding
4167 * doesn't grow big. We need to release
4168 * alloc_semp before calling ext4_mb_add_n_trim()
4170 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4171 spin_lock(pa
->pa_obj_lock
);
4172 list_del_rcu(&pa
->pa_inode_list
);
4173 spin_unlock(pa
->pa_obj_lock
);
4174 ext4_mb_add_n_trim(ac
);
4176 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4178 if (ac
->ac_bitmap_page
)
4179 page_cache_release(ac
->ac_bitmap_page
);
4180 if (ac
->ac_buddy_page
)
4181 page_cache_release(ac
->ac_buddy_page
);
4182 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4183 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4184 ext4_mb_collect_stats(ac
);
4188 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4190 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4194 trace_ext4_mb_discard_preallocations(sb
, needed
);
4195 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4196 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4205 * Main entry point into mballoc to allocate blocks
4206 * it tries to use preallocation first, then falls back
4207 * to usual allocation
4209 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4210 struct ext4_allocation_request
*ar
, int *errp
)
4213 struct ext4_allocation_context
*ac
= NULL
;
4214 struct ext4_sb_info
*sbi
;
4215 struct super_block
*sb
;
4216 ext4_fsblk_t block
= 0;
4217 unsigned int inquota
= 0;
4218 unsigned int reserv_blks
= 0;
4220 sb
= ar
->inode
->i_sb
;
4223 trace_ext4_request_blocks(ar
);
4226 * For delayed allocation, we could skip the ENOSPC and
4227 * EDQUOT check, as blocks and quotas have been already
4228 * reserved when data being copied into pagecache.
4230 if (EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
)
4231 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4233 /* Without delayed allocation we need to verify
4234 * there is enough free blocks to do block allocation
4235 * and verify allocation doesn't exceed the quota limits.
4237 while (ar
->len
&& ext4_claim_free_blocks(sbi
, ar
->len
)) {
4238 /* let others to free the space */
4240 ar
->len
= ar
->len
>> 1;
4246 reserv_blks
= ar
->len
;
4247 while (ar
->len
&& vfs_dq_alloc_block(ar
->inode
, ar
->len
)) {
4248 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4258 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4265 *errp
= ext4_mb_initialize_context(ac
, ar
);
4271 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4272 if (!ext4_mb_use_preallocated(ac
)) {
4273 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4274 ext4_mb_normalize_request(ac
, ar
);
4276 /* allocate space in core */
4277 ext4_mb_regular_allocator(ac
);
4279 /* as we've just preallocated more space than
4280 * user requested orinally, we store allocated
4281 * space in a special descriptor */
4282 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4283 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4284 ext4_mb_new_preallocation(ac
);
4286 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4287 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_blks
);
4288 if (*errp
== -EAGAIN
) {
4290 * drop the reference that we took
4291 * in ext4_mb_use_best_found
4293 ext4_mb_release_context(ac
);
4294 ac
->ac_b_ex
.fe_group
= 0;
4295 ac
->ac_b_ex
.fe_start
= 0;
4296 ac
->ac_b_ex
.fe_len
= 0;
4297 ac
->ac_status
= AC_STATUS_CONTINUE
;
4300 ext4_discard_allocated_blocks(ac
);
4301 ac
->ac_b_ex
.fe_len
= 0;
4303 ext4_mb_show_ac(ac
);
4305 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4306 ar
->len
= ac
->ac_b_ex
.fe_len
;
4309 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4313 ac
->ac_b_ex
.fe_len
= 0;
4315 ext4_mb_show_ac(ac
);
4318 ext4_mb_release_context(ac
);
4321 kmem_cache_free(ext4_ac_cachep
, ac
);
4323 if (inquota
&& ar
->len
< inquota
)
4324 vfs_dq_free_block(ar
->inode
, inquota
- ar
->len
);
4327 if (!EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
)
4328 /* release all the reserved blocks if non delalloc */
4329 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
,
4333 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4339 * We can merge two free data extents only if the physical blocks
4340 * are contiguous, AND the extents were freed by the same transaction,
4341 * AND the blocks are associated with the same group.
4343 static int can_merge(struct ext4_free_data
*entry1
,
4344 struct ext4_free_data
*entry2
)
4346 if ((entry1
->t_tid
== entry2
->t_tid
) &&
4347 (entry1
->group
== entry2
->group
) &&
4348 ((entry1
->start_blk
+ entry1
->count
) == entry2
->start_blk
))
4353 static noinline_for_stack
int
4354 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4355 struct ext4_free_data
*new_entry
)
4357 ext4_grpblk_t block
;
4358 struct ext4_free_data
*entry
;
4359 struct ext4_group_info
*db
= e4b
->bd_info
;
4360 struct super_block
*sb
= e4b
->bd_sb
;
4361 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4362 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4363 struct rb_node
*parent
= NULL
, *new_node
;
4365 BUG_ON(!ext4_handle_valid(handle
));
4366 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4367 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4369 new_node
= &new_entry
->node
;
4370 block
= new_entry
->start_blk
;
4373 /* first free block exent. We need to
4374 protect buddy cache from being freed,
4375 * otherwise we'll refresh it from
4376 * on-disk bitmap and lose not-yet-available
4378 page_cache_get(e4b
->bd_buddy_page
);
4379 page_cache_get(e4b
->bd_bitmap_page
);
4383 entry
= rb_entry(parent
, struct ext4_free_data
, node
);
4384 if (block
< entry
->start_blk
)
4386 else if (block
>= (entry
->start_blk
+ entry
->count
))
4387 n
= &(*n
)->rb_right
;
4389 ext4_grp_locked_error(sb
, e4b
->bd_group
, __func__
,
4390 "Double free of blocks %d (%d %d)",
4391 block
, entry
->start_blk
, entry
->count
);
4396 rb_link_node(new_node
, parent
, n
);
4397 rb_insert_color(new_node
, &db
->bb_free_root
);
4399 /* Now try to see the extent can be merged to left and right */
4400 node
= rb_prev(new_node
);
4402 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4403 if (can_merge(entry
, new_entry
)) {
4404 new_entry
->start_blk
= entry
->start_blk
;
4405 new_entry
->count
+= entry
->count
;
4406 rb_erase(node
, &(db
->bb_free_root
));
4407 spin_lock(&sbi
->s_md_lock
);
4408 list_del(&entry
->list
);
4409 spin_unlock(&sbi
->s_md_lock
);
4410 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4414 node
= rb_next(new_node
);
4416 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4417 if (can_merge(new_entry
, entry
)) {
4418 new_entry
->count
+= entry
->count
;
4419 rb_erase(node
, &(db
->bb_free_root
));
4420 spin_lock(&sbi
->s_md_lock
);
4421 list_del(&entry
->list
);
4422 spin_unlock(&sbi
->s_md_lock
);
4423 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4426 /* Add the extent to transaction's private list */
4427 spin_lock(&sbi
->s_md_lock
);
4428 list_add(&new_entry
->list
, &handle
->h_transaction
->t_private_list
);
4429 spin_unlock(&sbi
->s_md_lock
);
4434 * ext4_free_blocks() -- Free given blocks and update quota
4435 * @handle: handle for this transaction
4437 * @block: start physical block to free
4438 * @count: number of blocks to count
4439 * @metadata: Are these metadata blocks
4441 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4442 struct buffer_head
*bh
, ext4_fsblk_t block
,
4443 unsigned long count
, int flags
)
4445 struct buffer_head
*bitmap_bh
= NULL
;
4446 struct super_block
*sb
= inode
->i_sb
;
4447 struct ext4_allocation_context
*ac
= NULL
;
4448 struct ext4_group_desc
*gdp
;
4449 struct ext4_super_block
*es
;
4450 unsigned long freed
= 0;
4451 unsigned int overflow
;
4453 struct buffer_head
*gd_bh
;
4454 ext4_group_t block_group
;
4455 struct ext4_sb_info
*sbi
;
4456 struct ext4_buddy e4b
;
4462 BUG_ON(block
!= bh
->b_blocknr
);
4464 block
= bh
->b_blocknr
;
4468 es
= EXT4_SB(sb
)->s_es
;
4469 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4470 !ext4_data_block_valid(sbi
, block
, count
)) {
4471 ext4_error(sb
, "Freeing blocks not in datazone - "
4472 "block = %llu, count = %lu", block
, count
);
4476 ext4_debug("freeing block %llu\n", block
);
4477 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4479 if (flags
& EXT4_FREE_BLOCKS_FORGET
) {
4480 struct buffer_head
*tbh
= bh
;
4483 BUG_ON(bh
&& (count
> 1));
4485 for (i
= 0; i
< count
; i
++) {
4487 tbh
= sb_find_get_block(inode
->i_sb
,
4489 ext4_forget(handle
, flags
& EXT4_FREE_BLOCKS_METADATA
,
4490 inode
, tbh
, block
+ i
);
4495 * We need to make sure we don't reuse the freed block until
4496 * after the transaction is committed, which we can do by
4497 * treating the block as metadata, below. We make an
4498 * exception if the inode is to be written in writeback mode
4499 * since writeback mode has weak data consistency guarantees.
4501 if (!ext4_should_writeback_data(inode
))
4502 flags
|= EXT4_FREE_BLOCKS_METADATA
;
4504 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4506 ac
->ac_inode
= inode
;
4512 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4515 * Check to see if we are freeing blocks across a group
4518 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4519 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4522 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4527 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4533 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4534 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4535 in_range(block
, ext4_inode_table(sb
, gdp
),
4536 EXT4_SB(sb
)->s_itb_per_group
) ||
4537 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4538 EXT4_SB(sb
)->s_itb_per_group
)) {
4540 ext4_error(sb
, "Freeing blocks in system zone - "
4541 "Block = %llu, count = %lu", block
, count
);
4542 /* err = 0. ext4_std_error should be a no op */
4546 BUFFER_TRACE(bitmap_bh
, "getting write access");
4547 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4552 * We are about to modify some metadata. Call the journal APIs
4553 * to unshare ->b_data if a currently-committing transaction is
4556 BUFFER_TRACE(gd_bh
, "get_write_access");
4557 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4560 #ifdef AGGRESSIVE_CHECK
4563 for (i
= 0; i
< count
; i
++)
4564 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4568 ac
->ac_b_ex
.fe_group
= block_group
;
4569 ac
->ac_b_ex
.fe_start
= bit
;
4570 ac
->ac_b_ex
.fe_len
= count
;
4571 trace_ext4_mballoc_free(ac
);
4574 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4578 if ((flags
& EXT4_FREE_BLOCKS_METADATA
) && ext4_handle_valid(handle
)) {
4579 struct ext4_free_data
*new_entry
;
4581 * blocks being freed are metadata. these blocks shouldn't
4582 * be used until this transaction is committed
4584 new_entry
= kmem_cache_alloc(ext4_free_ext_cachep
, GFP_NOFS
);
4585 new_entry
->start_blk
= bit
;
4586 new_entry
->group
= block_group
;
4587 new_entry
->count
= count
;
4588 new_entry
->t_tid
= handle
->h_transaction
->t_tid
;
4590 ext4_lock_group(sb
, block_group
);
4591 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4592 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4594 /* need to update group_info->bb_free and bitmap
4595 * with group lock held. generate_buddy look at
4596 * them with group lock_held
4598 ext4_lock_group(sb
, block_group
);
4599 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4600 mb_free_blocks(inode
, &e4b
, bit
, count
);
4601 ext4_mb_return_to_preallocation(inode
, &e4b
, block
, count
);
4604 ret
= ext4_free_blks_count(sb
, gdp
) + count
;
4605 ext4_free_blks_set(sb
, gdp
, ret
);
4606 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4607 ext4_unlock_group(sb
, block_group
);
4608 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4610 if (sbi
->s_log_groups_per_flex
) {
4611 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4612 atomic_add(count
, &sbi
->s_flex_groups
[flex_group
].free_blocks
);
4615 ext4_mb_release_desc(&e4b
);
4619 /* We dirtied the bitmap block */
4620 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4621 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4623 /* And the group descriptor block */
4624 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4625 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4629 if (overflow
&& !err
) {
4638 vfs_dq_free_block(inode
, freed
);
4640 ext4_std_error(sb
, err
);
4642 kmem_cache_free(ext4_ac_cachep
, ac
);