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 <linux/slab.h>
27 #include <trace/events/ext4.h>
31 * - test ext4_ext_search_left() and ext4_ext_search_right()
32 * - search for metadata in few groups
35 * - normalization should take into account whether file is still open
36 * - discard preallocations if no free space left (policy?)
37 * - don't normalize tails
39 * - reservation for superuser
42 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
43 * - track min/max extents in each group for better group selection
44 * - mb_mark_used() may allocate chunk right after splitting buddy
45 * - tree of groups sorted by number of free blocks
50 * The allocation request involve request for multiple number of blocks
51 * near to the goal(block) value specified.
53 * During initialization phase of the allocator we decide to use the
54 * group preallocation or inode preallocation depending on the size of
55 * the file. The size of the file could be the resulting file size we
56 * would have after allocation, or the current file size, which ever
57 * is larger. If the size is less than sbi->s_mb_stream_request we
58 * select to use the group preallocation. The default value of
59 * s_mb_stream_request is 16 blocks. This can also be tuned via
60 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
61 * terms of number of blocks.
63 * The main motivation for having small file use group preallocation is to
64 * ensure that we have small files closer together on the disk.
66 * First stage the allocator looks at the inode prealloc list,
67 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
68 * spaces for this particular inode. The inode prealloc space is
71 * pa_lstart -> the logical start block for this prealloc space
72 * pa_pstart -> the physical start block for this prealloc space
73 * pa_len -> length for this prealloc space
74 * pa_free -> free space available in this prealloc space
76 * The inode preallocation space is used looking at the _logical_ start
77 * block. If only the logical file block falls within the range of prealloc
78 * space we will consume the particular prealloc space. This make sure that
79 * that the we have contiguous physical blocks representing the file blocks
81 * The important thing to be noted in case of inode prealloc space is that
82 * we don't modify the values associated to inode prealloc space except
85 * If we are not able to find blocks in the inode prealloc space and if we
86 * have the group allocation flag set then we look at the locality group
87 * prealloc space. These are per CPU prealloc list repreasented as
89 * ext4_sb_info.s_locality_groups[smp_processor_id()]
91 * The reason for having a per cpu locality group is to reduce the contention
92 * between CPUs. It is possible to get scheduled at this point.
94 * The locality group prealloc space is used looking at whether we have
95 * enough free space (pa_free) within the prealloc space.
97 * If we can't allocate blocks via inode prealloc or/and locality group
98 * prealloc then we look at the buddy cache. The buddy cache is represented
99 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
100 * mapped to the buddy and bitmap information regarding different
101 * groups. The buddy information is attached to buddy cache inode so that
102 * we can access them through the page cache. The information regarding
103 * each group is loaded via ext4_mb_load_buddy. The information involve
104 * block bitmap and buddy information. The information are stored in the
108 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
111 * one block each for bitmap and buddy information. So for each group we
112 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
113 * blocksize) blocks. So it can have information regarding groups_per_page
114 * which is blocks_per_page/2
116 * The buddy cache inode is not stored on disk. The inode is thrown
117 * away when the filesystem is unmounted.
119 * We look for count number of blocks in the buddy cache. If we were able
120 * to locate that many free blocks we return with additional information
121 * regarding rest of the contiguous physical block available
123 * Before allocating blocks via buddy cache we normalize the request
124 * blocks. This ensure we ask for more blocks that we needed. The extra
125 * blocks that we get after allocation is added to the respective prealloc
126 * list. In case of inode preallocation we follow a list of heuristics
127 * based on file size. This can be found in ext4_mb_normalize_request. If
128 * we are doing a group prealloc we try to normalize the request to
129 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
130 * 512 blocks. This can be tuned via
131 * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
132 * terms of number of blocks. If we have mounted the file system with -O
133 * stripe=<value> option the group prealloc request is normalized to the
134 * stripe value (sbi->s_stripe)
136 * The regular allocator(using the buddy cache) supports few tunables.
138 * /sys/fs/ext4/<partition>/mb_min_to_scan
139 * /sys/fs/ext4/<partition>/mb_max_to_scan
140 * /sys/fs/ext4/<partition>/mb_order2_req
142 * The regular allocator uses buddy scan only if the request len is power of
143 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
144 * value of s_mb_order2_reqs can be tuned via
145 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
146 * stripe size (sbi->s_stripe), we try to search for contiguous block in
147 * stripe size. This should result in better allocation on RAID setups. If
148 * not, we search in the specific group using bitmap for best extents. The
149 * tunable min_to_scan and max_to_scan control the behaviour here.
150 * min_to_scan indicate how long the mballoc __must__ look for a best
151 * extent and max_to_scan indicates how long the mballoc __can__ look for a
152 * best extent in the found extents. Searching for the blocks starts with
153 * the group specified as the goal value in allocation context via
154 * ac_g_ex. Each group is first checked based on the criteria whether it
155 * can used for allocation. ext4_mb_good_group explains how the groups are
158 * Both the prealloc space are getting populated as above. So for the first
159 * request we will hit the buddy cache which will result in this prealloc
160 * space getting filled. The prealloc space is then later used for the
161 * subsequent request.
165 * mballoc operates on the following data:
167 * - in-core buddy (actually includes buddy and bitmap)
168 * - preallocation descriptors (PAs)
170 * there are two types of preallocations:
172 * assiged to specific inode and can be used for this inode only.
173 * it describes part of inode's space preallocated to specific
174 * physical blocks. any block from that preallocated can be used
175 * independent. the descriptor just tracks number of blocks left
176 * unused. so, before taking some block from descriptor, one must
177 * make sure corresponded logical block isn't allocated yet. this
178 * also means that freeing any block within descriptor's range
179 * must discard all preallocated blocks.
181 * assigned to specific locality group which does not translate to
182 * permanent set of inodes: inode can join and leave group. space
183 * from this type of preallocation can be used for any inode. thus
184 * it's consumed from the beginning to the end.
186 * relation between them can be expressed as:
187 * in-core buddy = on-disk bitmap + preallocation descriptors
189 * this mean blocks mballoc considers used are:
190 * - allocated blocks (persistent)
191 * - preallocated blocks (non-persistent)
193 * consistency in mballoc world means that at any time a block is either
194 * free or used in ALL structures. notice: "any time" should not be read
195 * literally -- time is discrete and delimited by locks.
197 * to keep it simple, we don't use block numbers, instead we count number of
198 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
200 * all operations can be expressed as:
201 * - init buddy: buddy = on-disk + PAs
202 * - new PA: buddy += N; PA = N
203 * - use inode PA: on-disk += N; PA -= N
204 * - discard inode PA buddy -= on-disk - PA; PA = 0
205 * - use locality group PA on-disk += N; PA -= N
206 * - discard locality group PA buddy -= PA; PA = 0
207 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
208 * is used in real operation because we can't know actual used
209 * bits from PA, only from on-disk bitmap
211 * if we follow this strict logic, then all operations above should be atomic.
212 * given some of them can block, we'd have to use something like semaphores
213 * killing performance on high-end SMP hardware. let's try to relax it using
214 * the following knowledge:
215 * 1) if buddy is referenced, it's already initialized
216 * 2) while block is used in buddy and the buddy is referenced,
217 * nobody can re-allocate that block
218 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
219 * bit set and PA claims same block, it's OK. IOW, one can set bit in
220 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
223 * so, now we're building a concurrency table:
226 * blocks for PA are allocated in the buddy, buddy must be referenced
227 * until PA is linked to allocation group to avoid concurrent buddy init
229 * we need to make sure that either on-disk bitmap or PA has uptodate data
230 * given (3) we care that PA-=N operation doesn't interfere with init
232 * the simplest way would be to have buddy initialized by the discard
233 * - use locality group PA
234 * again PA-=N must be serialized with init
235 * - discard locality group PA
236 * the simplest way would be to have buddy initialized by the discard
239 * i_data_sem serializes them
241 * discard process must wait until PA isn't used by another process
242 * - use locality group PA
243 * some mutex should serialize them
244 * - discard locality group PA
245 * discard process must wait until PA isn't used by another process
248 * i_data_sem or another mutex should serializes them
250 * discard process must wait until PA isn't used by another process
251 * - use locality group PA
252 * nothing wrong here -- they're different PAs covering different blocks
253 * - discard locality group PA
254 * discard process must wait until PA isn't used by another process
256 * now we're ready to make few consequences:
257 * - PA is referenced and while it is no discard is possible
258 * - PA is referenced until block isn't marked in on-disk bitmap
259 * - PA changes only after on-disk bitmap
260 * - discard must not compete with init. either init is done before
261 * any discard or they're serialized somehow
262 * - buddy init as sum of on-disk bitmap and PAs is done atomically
264 * a special case when we've used PA to emptiness. no need to modify buddy
265 * in this case, but we should care about concurrent init
270 * Logic in few words:
275 * mark bits in on-disk bitmap
278 * - use preallocation:
279 * find proper PA (per-inode or group)
281 * mark bits in on-disk bitmap
287 * mark bits in on-disk bitmap
290 * - discard preallocations in group:
292 * move them onto local list
293 * load on-disk bitmap
295 * remove PA from object (inode or locality group)
296 * mark free blocks in-core
298 * - discard inode's preallocations:
305 * - bitlock on a group (group)
306 * - object (inode/locality) (object)
317 * - release consumed pa:
322 * - generate in-core bitmap:
326 * - discard all for given object (inode, locality group):
331 * - discard all for given group:
338 static struct kmem_cache
*ext4_pspace_cachep
;
339 static struct kmem_cache
*ext4_ac_cachep
;
340 static struct kmem_cache
*ext4_free_ext_cachep
;
342 /* We create slab caches for groupinfo data structures based on the
343 * superblock block size. There will be one per mounted filesystem for
344 * each unique s_blocksize_bits */
345 #define NR_GRPINFO_CACHES 8
346 static struct kmem_cache
*ext4_groupinfo_caches
[NR_GRPINFO_CACHES
];
348 static const char *ext4_groupinfo_slab_names
[NR_GRPINFO_CACHES
] = {
349 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
350 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
351 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
354 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
356 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
358 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
);
360 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
362 #if BITS_PER_LONG == 64
363 *bit
+= ((unsigned long) addr
& 7UL) << 3;
364 addr
= (void *) ((unsigned long) addr
& ~7UL);
365 #elif BITS_PER_LONG == 32
366 *bit
+= ((unsigned long) addr
& 3UL) << 3;
367 addr
= (void *) ((unsigned long) addr
& ~3UL);
369 #error "how many bits you are?!"
374 static inline int mb_test_bit(int bit
, void *addr
)
377 * ext4_test_bit on architecture like powerpc
378 * needs unsigned long aligned address
380 addr
= mb_correct_addr_and_bit(&bit
, addr
);
381 return ext4_test_bit(bit
, addr
);
384 static inline void mb_set_bit(int bit
, void *addr
)
386 addr
= mb_correct_addr_and_bit(&bit
, addr
);
387 ext4_set_bit(bit
, addr
);
390 static inline void mb_clear_bit(int bit
, void *addr
)
392 addr
= mb_correct_addr_and_bit(&bit
, addr
);
393 ext4_clear_bit(bit
, addr
);
396 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
398 int fix
= 0, ret
, tmpmax
;
399 addr
= mb_correct_addr_and_bit(&fix
, addr
);
403 ret
= ext4_find_next_zero_bit(addr
, tmpmax
, start
) - fix
;
409 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
411 int fix
= 0, ret
, tmpmax
;
412 addr
= mb_correct_addr_and_bit(&fix
, addr
);
416 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
422 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
426 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
429 if (order
> e4b
->bd_blkbits
+ 1) {
434 /* at order 0 we see each particular block */
436 *max
= 1 << (e4b
->bd_blkbits
+ 3);
437 return EXT4_MB_BITMAP(e4b
);
440 bb
= EXT4_MB_BUDDY(e4b
) + EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
441 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
447 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
448 int first
, int count
)
451 struct super_block
*sb
= e4b
->bd_sb
;
453 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
455 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
456 for (i
= 0; i
< count
; i
++) {
457 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
458 ext4_fsblk_t blocknr
;
460 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
461 blocknr
+= first
+ i
;
462 ext4_grp_locked_error(sb
, e4b
->bd_group
,
463 inode
? inode
->i_ino
: 0,
465 "freeing block already freed "
469 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
473 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
477 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
479 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
480 for (i
= 0; i
< count
; i
++) {
481 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
482 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
486 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
488 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
489 unsigned char *b1
, *b2
;
491 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
492 b2
= (unsigned char *) bitmap
;
493 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
494 if (b1
[i
] != b2
[i
]) {
495 printk(KERN_ERR
"corruption in group %u "
496 "at byte %u(%u): %x in copy != %x "
497 "on disk/prealloc\n",
498 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
506 static inline void mb_free_blocks_double(struct inode
*inode
,
507 struct ext4_buddy
*e4b
, int first
, int count
)
511 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
512 int first
, int count
)
516 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
522 #ifdef AGGRESSIVE_CHECK
524 #define MB_CHECK_ASSERT(assert) \
528 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
529 function, file, line, # assert); \
534 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
535 const char *function
, int line
)
537 struct super_block
*sb
= e4b
->bd_sb
;
538 int order
= e4b
->bd_blkbits
+ 1;
545 struct ext4_group_info
*grp
;
548 struct list_head
*cur
;
553 static int mb_check_counter
;
554 if (mb_check_counter
++ % 100 != 0)
559 buddy
= mb_find_buddy(e4b
, order
, &max
);
560 MB_CHECK_ASSERT(buddy
);
561 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
562 MB_CHECK_ASSERT(buddy2
);
563 MB_CHECK_ASSERT(buddy
!= buddy2
);
564 MB_CHECK_ASSERT(max
* 2 == max2
);
567 for (i
= 0; i
< max
; i
++) {
569 if (mb_test_bit(i
, buddy
)) {
570 /* only single bit in buddy2 may be 1 */
571 if (!mb_test_bit(i
<< 1, buddy2
)) {
573 mb_test_bit((i
<<1)+1, buddy2
));
574 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
576 mb_test_bit(i
<< 1, buddy2
));
581 /* both bits in buddy2 must be 0 */
582 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
583 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
585 for (j
= 0; j
< (1 << order
); j
++) {
586 k
= (i
* (1 << order
)) + j
;
588 !mb_test_bit(k
, EXT4_MB_BITMAP(e4b
)));
592 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
597 buddy
= mb_find_buddy(e4b
, 0, &max
);
598 for (i
= 0; i
< max
; i
++) {
599 if (!mb_test_bit(i
, buddy
)) {
600 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
608 /* check used bits only */
609 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
610 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
612 MB_CHECK_ASSERT(k
< max2
);
613 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
616 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
617 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
619 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
620 list_for_each(cur
, &grp
->bb_prealloc_list
) {
621 ext4_group_t groupnr
;
622 struct ext4_prealloc_space
*pa
;
623 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
624 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
625 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
626 for (i
= 0; i
< pa
->pa_len
; i
++)
627 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
631 #undef MB_CHECK_ASSERT
632 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
633 __FILE__, __func__, __LINE__)
635 #define mb_check_buddy(e4b)
639 * Divide blocks started from @first with length @len into
640 * smaller chunks with power of 2 blocks.
641 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
642 * then increase bb_counters[] for corresponded chunk size.
644 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
645 void *buddy
, ext4_grpblk_t first
, ext4_grpblk_t len
,
646 struct ext4_group_info
*grp
)
648 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
652 unsigned short border
;
654 BUG_ON(len
> EXT4_BLOCKS_PER_GROUP(sb
));
656 border
= 2 << sb
->s_blocksize_bits
;
659 /* find how many blocks can be covered since this position */
660 max
= ffs(first
| border
) - 1;
662 /* find how many blocks of power 2 we need to mark */
669 /* mark multiblock chunks only */
670 grp
->bb_counters
[min
]++;
672 mb_clear_bit(first
>> min
,
673 buddy
+ sbi
->s_mb_offsets
[min
]);
681 * Cache the order of the largest free extent we have available in this block
685 mb_set_largest_free_order(struct super_block
*sb
, struct ext4_group_info
*grp
)
690 grp
->bb_largest_free_order
= -1; /* uninit */
692 bits
= sb
->s_blocksize_bits
+ 1;
693 for (i
= bits
; i
>= 0; i
--) {
694 if (grp
->bb_counters
[i
] > 0) {
695 grp
->bb_largest_free_order
= i
;
701 static noinline_for_stack
702 void ext4_mb_generate_buddy(struct super_block
*sb
,
703 void *buddy
, void *bitmap
, ext4_group_t group
)
705 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
706 ext4_grpblk_t max
= EXT4_BLOCKS_PER_GROUP(sb
);
711 unsigned fragments
= 0;
712 unsigned long long period
= get_cycles();
714 /* initialize buddy from bitmap which is aggregation
715 * of on-disk bitmap and preallocations */
716 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
717 grp
->bb_first_free
= i
;
721 i
= mb_find_next_bit(bitmap
, max
, i
);
725 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
727 grp
->bb_counters
[0]++;
729 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
731 grp
->bb_fragments
= fragments
;
733 if (free
!= grp
->bb_free
) {
734 ext4_grp_locked_error(sb
, group
, 0, 0,
735 "%u blocks in bitmap, %u in gd",
738 * If we intent to continue, we consider group descritor
739 * corrupt and update bb_free using bitmap value
743 mb_set_largest_free_order(sb
, grp
);
745 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
747 period
= get_cycles() - period
;
748 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
749 EXT4_SB(sb
)->s_mb_buddies_generated
++;
750 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
751 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
754 /* The buddy information is attached the buddy cache inode
755 * for convenience. The information regarding each group
756 * is loaded via ext4_mb_load_buddy. The information involve
757 * block bitmap and buddy information. The information are
758 * stored in the inode as
761 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
764 * one block each for bitmap and buddy information.
765 * So for each group we take up 2 blocks. A page can
766 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
767 * So it can have information regarding groups_per_page which
768 * is blocks_per_page/2
770 * Locking note: This routine takes the block group lock of all groups
771 * for this page; do not hold this lock when calling this routine!
774 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
776 ext4_group_t ngroups
;
782 ext4_group_t first_group
;
784 struct super_block
*sb
;
785 struct buffer_head
*bhs
;
786 struct buffer_head
**bh
;
791 mb_debug(1, "init page %lu\n", page
->index
);
793 inode
= page
->mapping
->host
;
795 ngroups
= ext4_get_groups_count(sb
);
796 blocksize
= 1 << inode
->i_blkbits
;
797 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
799 groups_per_page
= blocks_per_page
>> 1;
800 if (groups_per_page
== 0)
803 /* allocate buffer_heads to read bitmaps */
804 if (groups_per_page
> 1) {
806 i
= sizeof(struct buffer_head
*) * groups_per_page
;
807 bh
= kzalloc(i
, GFP_NOFS
);
813 first_group
= page
->index
* blocks_per_page
/ 2;
815 /* read all groups the page covers into the cache */
816 for (i
= 0; i
< groups_per_page
; i
++) {
817 struct ext4_group_desc
*desc
;
819 if (first_group
+ i
>= ngroups
)
823 desc
= ext4_get_group_desc(sb
, first_group
+ i
, NULL
);
828 bh
[i
] = sb_getblk(sb
, ext4_block_bitmap(sb
, desc
));
832 if (bitmap_uptodate(bh
[i
]))
836 if (bitmap_uptodate(bh
[i
])) {
837 unlock_buffer(bh
[i
]);
840 ext4_lock_group(sb
, first_group
+ i
);
841 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
842 ext4_init_block_bitmap(sb
, bh
[i
],
843 first_group
+ i
, desc
);
844 set_bitmap_uptodate(bh
[i
]);
845 set_buffer_uptodate(bh
[i
]);
846 ext4_unlock_group(sb
, first_group
+ i
);
847 unlock_buffer(bh
[i
]);
850 ext4_unlock_group(sb
, first_group
+ i
);
851 if (buffer_uptodate(bh
[i
])) {
853 * if not uninit if bh is uptodate,
854 * bitmap is also uptodate
856 set_bitmap_uptodate(bh
[i
]);
857 unlock_buffer(bh
[i
]);
862 * submit the buffer_head for read. We can
863 * safely mark the bitmap as uptodate now.
864 * We do it here so the bitmap uptodate bit
865 * get set with buffer lock held.
867 set_bitmap_uptodate(bh
[i
]);
868 bh
[i
]->b_end_io
= end_buffer_read_sync
;
869 submit_bh(READ
, bh
[i
]);
870 mb_debug(1, "read bitmap for group %u\n", first_group
+ i
);
873 /* wait for I/O completion */
874 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
875 wait_on_buffer(bh
[i
]);
878 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
879 if (!buffer_uptodate(bh
[i
]))
883 first_block
= page
->index
* blocks_per_page
;
885 memset(page_address(page
), 0xff, PAGE_CACHE_SIZE
);
886 for (i
= 0; i
< blocks_per_page
; i
++) {
888 struct ext4_group_info
*grinfo
;
890 group
= (first_block
+ i
) >> 1;
891 if (group
>= ngroups
)
895 * data carry information regarding this
896 * particular group in the format specified
900 data
= page_address(page
) + (i
* blocksize
);
901 bitmap
= bh
[group
- first_group
]->b_data
;
904 * We place the buddy block and bitmap block
907 if ((first_block
+ i
) & 1) {
908 /* this is block of buddy */
909 BUG_ON(incore
== NULL
);
910 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
911 group
, page
->index
, i
* blocksize
);
912 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
913 grinfo
= ext4_get_group_info(sb
, group
);
914 grinfo
->bb_fragments
= 0;
915 memset(grinfo
->bb_counters
, 0,
916 sizeof(*grinfo
->bb_counters
) *
917 (sb
->s_blocksize_bits
+2));
919 * incore got set to the group block bitmap below
921 ext4_lock_group(sb
, group
);
922 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
923 ext4_unlock_group(sb
, group
);
926 /* this is block of bitmap */
927 BUG_ON(incore
!= NULL
);
928 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
929 group
, page
->index
, i
* blocksize
);
930 trace_ext4_mb_bitmap_load(sb
, group
);
932 /* see comments in ext4_mb_put_pa() */
933 ext4_lock_group(sb
, group
);
934 memcpy(data
, bitmap
, blocksize
);
936 /* mark all preallocated blks used in in-core bitmap */
937 ext4_mb_generate_from_pa(sb
, data
, group
);
938 ext4_mb_generate_from_freelist(sb
, data
, group
);
939 ext4_unlock_group(sb
, group
);
941 /* set incore so that the buddy information can be
942 * generated using this
947 SetPageUptodate(page
);
951 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
960 * lock the group_info alloc_sem of all the groups
961 * belonging to the same buddy cache page. This
962 * make sure other parallel operation on the buddy
963 * cache doesn't happen whild holding the buddy cache
966 static int ext4_mb_get_buddy_cache_lock(struct super_block
*sb
,
973 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
974 ext4_group_t first_group
;
975 struct ext4_group_info
*grp
;
977 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
979 * the buddy cache inode stores the block bitmap
980 * and buddy information in consecutive blocks.
981 * So for each group we need two blocks.
984 pnum
= block
/ blocks_per_page
;
985 first_group
= pnum
* blocks_per_page
/ 2;
987 groups_per_page
= blocks_per_page
>> 1;
988 if (groups_per_page
== 0)
990 /* read all groups the page covers into the cache */
991 for (i
= 0; i
< groups_per_page
; i
++) {
993 if ((first_group
+ i
) >= ngroups
)
995 grp
= ext4_get_group_info(sb
, first_group
+ i
);
996 /* take all groups write allocation
997 * semaphore. This make sure there is
998 * no block allocation going on in any
1001 down_write_nested(&grp
->alloc_sem
, i
);
1006 static void ext4_mb_put_buddy_cache_lock(struct super_block
*sb
,
1007 ext4_group_t group
, int locked_group
)
1011 int blocks_per_page
;
1012 ext4_group_t first_group
;
1013 struct ext4_group_info
*grp
;
1015 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1017 * the buddy cache inode stores the block bitmap
1018 * and buddy information in consecutive blocks.
1019 * So for each group we need two blocks.
1022 pnum
= block
/ blocks_per_page
;
1023 first_group
= pnum
* blocks_per_page
/ 2;
1024 /* release locks on all the groups */
1025 for (i
= 0; i
< locked_group
; i
++) {
1027 grp
= ext4_get_group_info(sb
, first_group
+ i
);
1028 /* take all groups write allocation
1029 * semaphore. This make sure there is
1030 * no block allocation going on in any
1033 up_write(&grp
->alloc_sem
);
1039 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1040 * block group lock of all groups for this page; do not hold the BG lock when
1041 * calling this routine!
1043 static noinline_for_stack
1044 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
)
1049 int blocks_per_page
;
1050 int block
, pnum
, poff
;
1051 int num_grp_locked
= 0;
1052 struct ext4_group_info
*this_grp
;
1053 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1054 struct inode
*inode
= sbi
->s_buddy_cache
;
1055 struct page
*page
= NULL
, *bitmap_page
= NULL
;
1057 mb_debug(1, "init group %u\n", group
);
1058 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1059 this_grp
= ext4_get_group_info(sb
, group
);
1061 * This ensures that we don't reinit the buddy cache
1062 * page which map to the group from which we are already
1063 * allocating. If we are looking at the buddy cache we would
1064 * have taken a reference using ext4_mb_load_buddy and that
1065 * would have taken the alloc_sem lock.
1067 num_grp_locked
= ext4_mb_get_buddy_cache_lock(sb
, group
);
1068 if (!EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1070 * somebody initialized the group
1071 * return without doing anything
1077 * the buddy cache inode stores the block bitmap
1078 * and buddy information in consecutive blocks.
1079 * So for each group we need two blocks.
1082 pnum
= block
/ blocks_per_page
;
1083 poff
= block
% blocks_per_page
;
1084 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1086 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1087 ret
= ext4_mb_init_cache(page
, NULL
);
1094 if (page
== NULL
|| !PageUptodate(page
)) {
1098 mark_page_accessed(page
);
1100 bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1102 /* init buddy cache */
1104 pnum
= block
/ blocks_per_page
;
1105 poff
= block
% blocks_per_page
;
1106 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1107 if (page
== bitmap_page
) {
1109 * If both the bitmap and buddy are in
1110 * the same page we don't need to force
1115 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1116 ret
= ext4_mb_init_cache(page
, bitmap
);
1123 if (page
== NULL
|| !PageUptodate(page
)) {
1127 mark_page_accessed(page
);
1129 ext4_mb_put_buddy_cache_lock(sb
, group
, num_grp_locked
);
1131 page_cache_release(bitmap_page
);
1133 page_cache_release(page
);
1138 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1139 * block group lock of all groups for this page; do not hold the BG lock when
1140 * calling this routine!
1142 static noinline_for_stack
int
1143 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1144 struct ext4_buddy
*e4b
)
1146 int blocks_per_page
;
1152 struct ext4_group_info
*grp
;
1153 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1154 struct inode
*inode
= sbi
->s_buddy_cache
;
1156 mb_debug(1, "load group %u\n", group
);
1158 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1159 grp
= ext4_get_group_info(sb
, group
);
1161 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1162 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
1164 e4b
->bd_group
= group
;
1165 e4b
->bd_buddy_page
= NULL
;
1166 e4b
->bd_bitmap_page
= NULL
;
1167 e4b
->alloc_semp
= &grp
->alloc_sem
;
1169 /* Take the read lock on the group alloc
1170 * sem. This would make sure a parallel
1171 * ext4_mb_init_group happening on other
1172 * groups mapped by the page is blocked
1173 * till we are done with allocation
1176 down_read(e4b
->alloc_semp
);
1178 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1179 /* we need to check for group need init flag
1180 * with alloc_semp held so that we can be sure
1181 * that new blocks didn't get added to the group
1182 * when we are loading the buddy cache
1184 up_read(e4b
->alloc_semp
);
1186 * we need full data about the group
1187 * to make a good selection
1189 ret
= ext4_mb_init_group(sb
, group
);
1192 goto repeat_load_buddy
;
1196 * the buddy cache inode stores the block bitmap
1197 * and buddy information in consecutive blocks.
1198 * So for each group we need two blocks.
1201 pnum
= block
/ blocks_per_page
;
1202 poff
= block
% blocks_per_page
;
1204 /* we could use find_or_create_page(), but it locks page
1205 * what we'd like to avoid in fast path ... */
1206 page
= find_get_page(inode
->i_mapping
, pnum
);
1207 if (page
== NULL
|| !PageUptodate(page
)) {
1210 * drop the page reference and try
1211 * to get the page with lock. If we
1212 * are not uptodate that implies
1213 * somebody just created the page but
1214 * is yet to initialize the same. So
1215 * wait for it to initialize.
1217 page_cache_release(page
);
1218 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1220 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1221 if (!PageUptodate(page
)) {
1222 ret
= ext4_mb_init_cache(page
, NULL
);
1227 mb_cmp_bitmaps(e4b
, page_address(page
) +
1228 (poff
* sb
->s_blocksize
));
1233 if (page
== NULL
|| !PageUptodate(page
)) {
1237 e4b
->bd_bitmap_page
= page
;
1238 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1239 mark_page_accessed(page
);
1242 pnum
= block
/ blocks_per_page
;
1243 poff
= block
% blocks_per_page
;
1245 page
= find_get_page(inode
->i_mapping
, pnum
);
1246 if (page
== NULL
|| !PageUptodate(page
)) {
1248 page_cache_release(page
);
1249 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1251 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1252 if (!PageUptodate(page
)) {
1253 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1262 if (page
== NULL
|| !PageUptodate(page
)) {
1266 e4b
->bd_buddy_page
= page
;
1267 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1268 mark_page_accessed(page
);
1270 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1271 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1277 page_cache_release(page
);
1278 if (e4b
->bd_bitmap_page
)
1279 page_cache_release(e4b
->bd_bitmap_page
);
1280 if (e4b
->bd_buddy_page
)
1281 page_cache_release(e4b
->bd_buddy_page
);
1282 e4b
->bd_buddy
= NULL
;
1283 e4b
->bd_bitmap
= NULL
;
1285 /* Done with the buddy cache */
1286 up_read(e4b
->alloc_semp
);
1290 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1292 if (e4b
->bd_bitmap_page
)
1293 page_cache_release(e4b
->bd_bitmap_page
);
1294 if (e4b
->bd_buddy_page
)
1295 page_cache_release(e4b
->bd_buddy_page
);
1296 /* Done with the buddy cache */
1297 if (e4b
->alloc_semp
)
1298 up_read(e4b
->alloc_semp
);
1302 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1307 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1308 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1310 bb
= EXT4_MB_BUDDY(e4b
);
1311 while (order
<= e4b
->bd_blkbits
+ 1) {
1313 if (!mb_test_bit(block
, bb
)) {
1314 /* this block is part of buddy of order 'order' */
1317 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1323 static void mb_clear_bits(void *bm
, int cur
, int len
)
1329 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1330 /* fast path: clear whole word at once */
1331 addr
= bm
+ (cur
>> 3);
1336 mb_clear_bit(cur
, bm
);
1341 static void mb_set_bits(void *bm
, int cur
, int len
)
1347 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1348 /* fast path: set whole word at once */
1349 addr
= bm
+ (cur
>> 3);
1354 mb_set_bit(cur
, bm
);
1359 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1360 int first
, int count
)
1367 struct super_block
*sb
= e4b
->bd_sb
;
1369 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1370 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1371 mb_check_buddy(e4b
);
1372 mb_free_blocks_double(inode
, e4b
, first
, count
);
1374 e4b
->bd_info
->bb_free
+= count
;
1375 if (first
< e4b
->bd_info
->bb_first_free
)
1376 e4b
->bd_info
->bb_first_free
= first
;
1378 /* let's maintain fragments counter */
1380 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1381 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1382 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1384 e4b
->bd_info
->bb_fragments
--;
1385 else if (!block
&& !max
)
1386 e4b
->bd_info
->bb_fragments
++;
1388 /* let's maintain buddy itself */
1389 while (count
-- > 0) {
1393 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1394 ext4_fsblk_t blocknr
;
1396 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1398 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1399 inode
? inode
->i_ino
: 0,
1401 "freeing already freed block "
1404 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1405 e4b
->bd_info
->bb_counters
[order
]++;
1407 /* start of the buddy */
1408 buddy
= mb_find_buddy(e4b
, order
, &max
);
1412 if (mb_test_bit(block
, buddy
) ||
1413 mb_test_bit(block
+ 1, buddy
))
1416 /* both the buddies are free, try to coalesce them */
1417 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1423 /* for special purposes, we don't set
1424 * free bits in bitmap */
1425 mb_set_bit(block
, buddy
);
1426 mb_set_bit(block
+ 1, buddy
);
1428 e4b
->bd_info
->bb_counters
[order
]--;
1429 e4b
->bd_info
->bb_counters
[order
]--;
1433 e4b
->bd_info
->bb_counters
[order
]++;
1435 mb_clear_bit(block
, buddy2
);
1439 mb_set_largest_free_order(sb
, e4b
->bd_info
);
1440 mb_check_buddy(e4b
);
1443 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1444 int needed
, struct ext4_free_extent
*ex
)
1451 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1454 buddy
= mb_find_buddy(e4b
, order
, &max
);
1455 BUG_ON(buddy
== NULL
);
1456 BUG_ON(block
>= max
);
1457 if (mb_test_bit(block
, buddy
)) {
1464 /* FIXME dorp order completely ? */
1465 if (likely(order
== 0)) {
1466 /* find actual order */
1467 order
= mb_find_order_for_block(e4b
, block
);
1468 block
= block
>> order
;
1471 ex
->fe_len
= 1 << order
;
1472 ex
->fe_start
= block
<< order
;
1473 ex
->fe_group
= e4b
->bd_group
;
1475 /* calc difference from given start */
1476 next
= next
- ex
->fe_start
;
1478 ex
->fe_start
+= next
;
1480 while (needed
> ex
->fe_len
&&
1481 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1483 if (block
+ 1 >= max
)
1486 next
= (block
+ 1) * (1 << order
);
1487 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1490 ord
= mb_find_order_for_block(e4b
, next
);
1493 block
= next
>> order
;
1494 ex
->fe_len
+= 1 << order
;
1497 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1501 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1507 int start
= ex
->fe_start
;
1508 int len
= ex
->fe_len
;
1513 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1514 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1515 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1516 mb_check_buddy(e4b
);
1517 mb_mark_used_double(e4b
, start
, len
);
1519 e4b
->bd_info
->bb_free
-= len
;
1520 if (e4b
->bd_info
->bb_first_free
== start
)
1521 e4b
->bd_info
->bb_first_free
+= len
;
1523 /* let's maintain fragments counter */
1525 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1526 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1527 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1529 e4b
->bd_info
->bb_fragments
++;
1530 else if (!mlen
&& !max
)
1531 e4b
->bd_info
->bb_fragments
--;
1533 /* let's maintain buddy itself */
1535 ord
= mb_find_order_for_block(e4b
, start
);
1537 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1538 /* the whole chunk may be allocated at once! */
1540 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1541 BUG_ON((start
>> ord
) >= max
);
1542 mb_set_bit(start
>> ord
, buddy
);
1543 e4b
->bd_info
->bb_counters
[ord
]--;
1550 /* store for history */
1552 ret
= len
| (ord
<< 16);
1554 /* we have to split large buddy */
1556 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1557 mb_set_bit(start
>> ord
, buddy
);
1558 e4b
->bd_info
->bb_counters
[ord
]--;
1561 cur
= (start
>> ord
) & ~1U;
1562 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1563 mb_clear_bit(cur
, buddy
);
1564 mb_clear_bit(cur
+ 1, buddy
);
1565 e4b
->bd_info
->bb_counters
[ord
]++;
1566 e4b
->bd_info
->bb_counters
[ord
]++;
1568 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1570 mb_set_bits(EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1571 mb_check_buddy(e4b
);
1577 * Must be called under group lock!
1579 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1580 struct ext4_buddy
*e4b
)
1582 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1585 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1586 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1588 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1589 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1590 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1592 /* preallocation can change ac_b_ex, thus we store actually
1593 * allocated blocks for history */
1594 ac
->ac_f_ex
= ac
->ac_b_ex
;
1596 ac
->ac_status
= AC_STATUS_FOUND
;
1597 ac
->ac_tail
= ret
& 0xffff;
1598 ac
->ac_buddy
= ret
>> 16;
1601 * take the page reference. We want the page to be pinned
1602 * so that we don't get a ext4_mb_init_cache_call for this
1603 * group until we update the bitmap. That would mean we
1604 * double allocate blocks. The reference is dropped
1605 * in ext4_mb_release_context
1607 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1608 get_page(ac
->ac_bitmap_page
);
1609 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1610 get_page(ac
->ac_buddy_page
);
1611 /* on allocation we use ac to track the held semaphore */
1612 ac
->alloc_semp
= e4b
->alloc_semp
;
1613 e4b
->alloc_semp
= NULL
;
1614 /* store last allocated for subsequent stream allocation */
1615 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1616 spin_lock(&sbi
->s_md_lock
);
1617 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1618 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1619 spin_unlock(&sbi
->s_md_lock
);
1624 * regular allocator, for general purposes allocation
1627 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1628 struct ext4_buddy
*e4b
,
1631 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1632 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1633 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1634 struct ext4_free_extent ex
;
1637 if (ac
->ac_status
== AC_STATUS_FOUND
)
1640 * We don't want to scan for a whole year
1642 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1643 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1644 ac
->ac_status
= AC_STATUS_BREAK
;
1649 * Haven't found good chunk so far, let's continue
1651 if (bex
->fe_len
< gex
->fe_len
)
1654 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1655 && bex
->fe_group
== e4b
->bd_group
) {
1656 /* recheck chunk's availability - we don't know
1657 * when it was found (within this lock-unlock
1659 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1660 if (max
>= gex
->fe_len
) {
1661 ext4_mb_use_best_found(ac
, e4b
);
1668 * The routine checks whether found extent is good enough. If it is,
1669 * then the extent gets marked used and flag is set to the context
1670 * to stop scanning. Otherwise, the extent is compared with the
1671 * previous found extent and if new one is better, then it's stored
1672 * in the context. Later, the best found extent will be used, if
1673 * mballoc can't find good enough extent.
1675 * FIXME: real allocation policy is to be designed yet!
1677 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1678 struct ext4_free_extent
*ex
,
1679 struct ext4_buddy
*e4b
)
1681 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1682 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1684 BUG_ON(ex
->fe_len
<= 0);
1685 BUG_ON(ex
->fe_len
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1686 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1687 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1692 * The special case - take what you catch first
1694 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1696 ext4_mb_use_best_found(ac
, e4b
);
1701 * Let's check whether the chuck is good enough
1703 if (ex
->fe_len
== gex
->fe_len
) {
1705 ext4_mb_use_best_found(ac
, e4b
);
1710 * If this is first found extent, just store it in the context
1712 if (bex
->fe_len
== 0) {
1718 * If new found extent is better, store it in the context
1720 if (bex
->fe_len
< gex
->fe_len
) {
1721 /* if the request isn't satisfied, any found extent
1722 * larger than previous best one is better */
1723 if (ex
->fe_len
> bex
->fe_len
)
1725 } else if (ex
->fe_len
> gex
->fe_len
) {
1726 /* if the request is satisfied, then we try to find
1727 * an extent that still satisfy the request, but is
1728 * smaller than previous one */
1729 if (ex
->fe_len
< bex
->fe_len
)
1733 ext4_mb_check_limits(ac
, e4b
, 0);
1736 static noinline_for_stack
1737 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1738 struct ext4_buddy
*e4b
)
1740 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1741 ext4_group_t group
= ex
.fe_group
;
1745 BUG_ON(ex
.fe_len
<= 0);
1746 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1750 ext4_lock_group(ac
->ac_sb
, group
);
1751 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1755 ext4_mb_use_best_found(ac
, e4b
);
1758 ext4_unlock_group(ac
->ac_sb
, group
);
1759 ext4_mb_unload_buddy(e4b
);
1764 static noinline_for_stack
1765 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1766 struct ext4_buddy
*e4b
)
1768 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1771 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1772 struct ext4_free_extent ex
;
1774 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1777 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1781 ext4_lock_group(ac
->ac_sb
, group
);
1782 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1783 ac
->ac_g_ex
.fe_len
, &ex
);
1785 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1788 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1790 /* use do_div to get remainder (would be 64-bit modulo) */
1791 if (do_div(start
, sbi
->s_stripe
) == 0) {
1794 ext4_mb_use_best_found(ac
, e4b
);
1796 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1797 BUG_ON(ex
.fe_len
<= 0);
1798 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1799 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1802 ext4_mb_use_best_found(ac
, e4b
);
1803 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1804 /* Sometimes, caller may want to merge even small
1805 * number of blocks to an existing extent */
1806 BUG_ON(ex
.fe_len
<= 0);
1807 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1808 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1811 ext4_mb_use_best_found(ac
, e4b
);
1813 ext4_unlock_group(ac
->ac_sb
, group
);
1814 ext4_mb_unload_buddy(e4b
);
1820 * The routine scans buddy structures (not bitmap!) from given order
1821 * to max order and tries to find big enough chunk to satisfy the req
1823 static noinline_for_stack
1824 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1825 struct ext4_buddy
*e4b
)
1827 struct super_block
*sb
= ac
->ac_sb
;
1828 struct ext4_group_info
*grp
= e4b
->bd_info
;
1834 BUG_ON(ac
->ac_2order
<= 0);
1835 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1836 if (grp
->bb_counters
[i
] == 0)
1839 buddy
= mb_find_buddy(e4b
, i
, &max
);
1840 BUG_ON(buddy
== NULL
);
1842 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1847 ac
->ac_b_ex
.fe_len
= 1 << i
;
1848 ac
->ac_b_ex
.fe_start
= k
<< i
;
1849 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1851 ext4_mb_use_best_found(ac
, e4b
);
1853 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1855 if (EXT4_SB(sb
)->s_mb_stats
)
1856 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1863 * The routine scans the group and measures all found extents.
1864 * In order to optimize scanning, caller must pass number of
1865 * free blocks in the group, so the routine can know upper limit.
1867 static noinline_for_stack
1868 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1869 struct ext4_buddy
*e4b
)
1871 struct super_block
*sb
= ac
->ac_sb
;
1872 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1873 struct ext4_free_extent ex
;
1877 free
= e4b
->bd_info
->bb_free
;
1880 i
= e4b
->bd_info
->bb_first_free
;
1882 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1883 i
= mb_find_next_zero_bit(bitmap
,
1884 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1885 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1887 * IF we have corrupt bitmap, we won't find any
1888 * free blocks even though group info says we
1889 * we have free blocks
1891 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1892 "%d free blocks as per "
1893 "group info. But bitmap says 0",
1898 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1899 BUG_ON(ex
.fe_len
<= 0);
1900 if (free
< ex
.fe_len
) {
1901 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1902 "%d free blocks as per "
1903 "group info. But got %d blocks",
1906 * The number of free blocks differs. This mostly
1907 * indicate that the bitmap is corrupt. So exit
1908 * without claiming the space.
1913 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1919 ext4_mb_check_limits(ac
, e4b
, 1);
1923 * This is a special case for storages like raid5
1924 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1926 static noinline_for_stack
1927 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1928 struct ext4_buddy
*e4b
)
1930 struct super_block
*sb
= ac
->ac_sb
;
1931 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1932 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1933 struct ext4_free_extent ex
;
1934 ext4_fsblk_t first_group_block
;
1939 BUG_ON(sbi
->s_stripe
== 0);
1941 /* find first stripe-aligned block in group */
1942 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1944 a
= first_group_block
+ sbi
->s_stripe
- 1;
1945 do_div(a
, sbi
->s_stripe
);
1946 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1948 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1949 if (!mb_test_bit(i
, bitmap
)) {
1950 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1951 if (max
>= sbi
->s_stripe
) {
1954 ext4_mb_use_best_found(ac
, e4b
);
1962 /* This is now called BEFORE we load the buddy bitmap. */
1963 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1964 ext4_group_t group
, int cr
)
1966 unsigned free
, fragments
;
1967 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
1968 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1970 BUG_ON(cr
< 0 || cr
>= 4);
1972 /* We only do this if the grp has never been initialized */
1973 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1974 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
);
1979 free
= grp
->bb_free
;
1980 fragments
= grp
->bb_fragments
;
1988 BUG_ON(ac
->ac_2order
== 0);
1990 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
1993 /* Avoid using the first bg of a flexgroup for data files */
1994 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
1995 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
1996 ((group
% flex_size
) == 0))
2001 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2005 if (free
>= ac
->ac_g_ex
.fe_len
)
2017 static noinline_for_stack
int
2018 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
2020 ext4_group_t ngroups
, group
, i
;
2023 struct ext4_sb_info
*sbi
;
2024 struct super_block
*sb
;
2025 struct ext4_buddy e4b
;
2029 ngroups
= ext4_get_groups_count(sb
);
2030 /* non-extent files are limited to low blocks/groups */
2031 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
2032 ngroups
= sbi
->s_blockfile_groups
;
2034 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
2036 /* first, try the goal */
2037 err
= ext4_mb_find_by_goal(ac
, &e4b
);
2038 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
2041 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2045 * ac->ac2_order is set only if the fe_len is a power of 2
2046 * if ac2_order is set we also set criteria to 0 so that we
2047 * try exact allocation using buddy.
2049 i
= fls(ac
->ac_g_ex
.fe_len
);
2052 * We search using buddy data only if the order of the request
2053 * is greater than equal to the sbi_s_mb_order2_reqs
2054 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2056 if (i
>= sbi
->s_mb_order2_reqs
) {
2058 * This should tell if fe_len is exactly power of 2
2060 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2061 ac
->ac_2order
= i
- 1;
2064 /* if stream allocation is enabled, use global goal */
2065 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2066 /* TBD: may be hot point */
2067 spin_lock(&sbi
->s_md_lock
);
2068 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2069 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2070 spin_unlock(&sbi
->s_md_lock
);
2073 /* Let's just scan groups to find more-less suitable blocks */
2074 cr
= ac
->ac_2order
? 0 : 1;
2076 * cr == 0 try to get exact allocation,
2077 * cr == 3 try to get anything
2080 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2081 ac
->ac_criteria
= cr
;
2083 * searching for the right group start
2084 * from the goal value specified
2086 group
= ac
->ac_g_ex
.fe_group
;
2088 for (i
= 0; i
< ngroups
; group
++, i
++) {
2089 if (group
== ngroups
)
2092 /* This now checks without needing the buddy page */
2093 if (!ext4_mb_good_group(ac
, group
, cr
))
2096 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2100 ext4_lock_group(sb
, group
);
2103 * We need to check again after locking the
2106 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2107 ext4_unlock_group(sb
, group
);
2108 ext4_mb_unload_buddy(&e4b
);
2112 ac
->ac_groups_scanned
++;
2114 ext4_mb_simple_scan_group(ac
, &e4b
);
2115 else if (cr
== 1 && sbi
->s_stripe
&&
2116 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2117 ext4_mb_scan_aligned(ac
, &e4b
);
2119 ext4_mb_complex_scan_group(ac
, &e4b
);
2121 ext4_unlock_group(sb
, group
);
2122 ext4_mb_unload_buddy(&e4b
);
2124 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2129 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2130 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2132 * We've been searching too long. Let's try to allocate
2133 * the best chunk we've found so far
2136 ext4_mb_try_best_found(ac
, &e4b
);
2137 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2139 * Someone more lucky has already allocated it.
2140 * The only thing we can do is just take first
2142 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2144 ac
->ac_b_ex
.fe_group
= 0;
2145 ac
->ac_b_ex
.fe_start
= 0;
2146 ac
->ac_b_ex
.fe_len
= 0;
2147 ac
->ac_status
= AC_STATUS_CONTINUE
;
2148 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2150 atomic_inc(&sbi
->s_mb_lost_chunks
);
2158 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2160 struct super_block
*sb
= seq
->private;
2163 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2166 return (void *) ((unsigned long) group
);
2169 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2171 struct super_block
*sb
= seq
->private;
2175 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2178 return (void *) ((unsigned long) group
);
2181 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2183 struct super_block
*sb
= seq
->private;
2184 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2187 struct ext4_buddy e4b
;
2189 struct ext4_group_info info
;
2190 ext4_grpblk_t counters
[16];
2195 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2196 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2197 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2198 "group", "free", "frags", "first",
2199 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2200 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2202 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2203 sizeof(struct ext4_group_info
);
2204 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2206 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2209 ext4_lock_group(sb
, group
);
2210 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2211 ext4_unlock_group(sb
, group
);
2212 ext4_mb_unload_buddy(&e4b
);
2214 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2215 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2216 for (i
= 0; i
<= 13; i
++)
2217 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2218 sg
.info
.bb_counters
[i
] : 0);
2219 seq_printf(seq
, " ]\n");
2224 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2228 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2229 .start
= ext4_mb_seq_groups_start
,
2230 .next
= ext4_mb_seq_groups_next
,
2231 .stop
= ext4_mb_seq_groups_stop
,
2232 .show
= ext4_mb_seq_groups_show
,
2235 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2237 struct super_block
*sb
= PDE(inode
)->data
;
2240 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2242 struct seq_file
*m
= file
->private_data
;
2249 static const struct file_operations ext4_mb_seq_groups_fops
= {
2250 .owner
= THIS_MODULE
,
2251 .open
= ext4_mb_seq_groups_open
,
2253 .llseek
= seq_lseek
,
2254 .release
= seq_release
,
2257 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2259 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2260 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2266 /* Create and initialize ext4_group_info data for the given group. */
2267 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2268 struct ext4_group_desc
*desc
)
2272 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2273 struct ext4_group_info
**meta_group_info
;
2274 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2277 * First check if this group is the first of a reserved block.
2278 * If it's true, we have to allocate a new table of pointers
2279 * to ext4_group_info structures
2281 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2282 metalen
= sizeof(*meta_group_info
) <<
2283 EXT4_DESC_PER_BLOCK_BITS(sb
);
2284 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2285 if (meta_group_info
== NULL
) {
2286 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2288 goto exit_meta_group_info
;
2290 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2295 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2296 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2298 meta_group_info
[i
] = kmem_cache_alloc(cachep
, GFP_KERNEL
);
2299 if (meta_group_info
[i
] == NULL
) {
2300 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2301 goto exit_group_info
;
2303 memset(meta_group_info
[i
], 0, kmem_cache_size(cachep
));
2304 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2305 &(meta_group_info
[i
]->bb_state
));
2308 * initialize bb_free to be able to skip
2309 * empty groups without initialization
2311 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2312 meta_group_info
[i
]->bb_free
=
2313 ext4_free_blocks_after_init(sb
, group
, desc
);
2315 meta_group_info
[i
]->bb_free
=
2316 ext4_free_blks_count(sb
, desc
);
2319 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2320 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2321 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2322 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2326 struct buffer_head
*bh
;
2327 meta_group_info
[i
]->bb_bitmap
=
2328 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2329 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2330 bh
= ext4_read_block_bitmap(sb
, group
);
2332 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2341 /* If a meta_group_info table has been allocated, release it now */
2342 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0)
2343 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2344 exit_meta_group_info
:
2346 } /* ext4_mb_add_groupinfo */
2348 static int ext4_mb_init_backend(struct super_block
*sb
)
2350 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2352 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2353 struct ext4_super_block
*es
= sbi
->s_es
;
2354 int num_meta_group_infos
;
2355 int num_meta_group_infos_max
;
2357 struct ext4_group_desc
*desc
;
2358 struct kmem_cache
*cachep
;
2360 /* This is the number of blocks used by GDT */
2361 num_meta_group_infos
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) -
2362 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2365 * This is the total number of blocks used by GDT including
2366 * the number of reserved blocks for GDT.
2367 * The s_group_info array is allocated with this value
2368 * to allow a clean online resize without a complex
2369 * manipulation of pointer.
2370 * The drawback is the unused memory when no resize
2371 * occurs but it's very low in terms of pages
2372 * (see comments below)
2373 * Need to handle this properly when META_BG resizing is allowed
2375 num_meta_group_infos_max
= num_meta_group_infos
+
2376 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2379 * array_size is the size of s_group_info array. We round it
2380 * to the next power of two because this approximation is done
2381 * internally by kmalloc so we can have some more memory
2382 * for free here (e.g. may be used for META_BG resize).
2385 while (array_size
< sizeof(*sbi
->s_group_info
) *
2386 num_meta_group_infos_max
)
2387 array_size
= array_size
<< 1;
2388 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2389 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2390 * So a two level scheme suffices for now. */
2391 sbi
->s_group_info
= kzalloc(array_size
, GFP_KERNEL
);
2392 if (sbi
->s_group_info
== NULL
) {
2393 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2396 sbi
->s_buddy_cache
= new_inode(sb
);
2397 if (sbi
->s_buddy_cache
== NULL
) {
2398 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2401 sbi
->s_buddy_cache
->i_ino
= get_next_ino();
2402 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2403 for (i
= 0; i
< ngroups
; i
++) {
2404 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2407 "EXT4-fs: can't read descriptor %u\n", i
);
2410 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2417 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2419 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2420 i
= num_meta_group_infos
;
2422 kfree(sbi
->s_group_info
[i
]);
2423 iput(sbi
->s_buddy_cache
);
2425 kfree(sbi
->s_group_info
);
2429 static void ext4_groupinfo_destroy_slabs(void)
2433 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2434 if (ext4_groupinfo_caches
[i
])
2435 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2436 ext4_groupinfo_caches
[i
] = NULL
;
2440 static int ext4_groupinfo_create_slab(size_t size
)
2442 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2444 int blocksize_bits
= order_base_2(size
);
2445 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2446 struct kmem_cache
*cachep
;
2448 if (cache_index
>= NR_GRPINFO_CACHES
)
2451 if (unlikely(cache_index
< 0))
2454 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2455 if (ext4_groupinfo_caches
[cache_index
]) {
2456 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2457 return 0; /* Already created */
2460 slab_size
= offsetof(struct ext4_group_info
,
2461 bb_counters
[blocksize_bits
+ 2]);
2463 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2464 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2467 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2469 printk(KERN_EMERG
"EXT4: no memory for groupinfo slab cache\n");
2473 ext4_groupinfo_caches
[cache_index
] = cachep
;
2478 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2480 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2486 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2488 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2489 if (sbi
->s_mb_offsets
== NULL
) {
2494 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2495 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2496 if (sbi
->s_mb_maxs
== NULL
) {
2501 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2505 /* order 0 is regular bitmap */
2506 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2507 sbi
->s_mb_offsets
[0] = 0;
2511 max
= sb
->s_blocksize
<< 2;
2513 sbi
->s_mb_offsets
[i
] = offset
;
2514 sbi
->s_mb_maxs
[i
] = max
;
2515 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2518 } while (i
<= sb
->s_blocksize_bits
+ 1);
2520 /* init file for buddy data */
2521 ret
= ext4_mb_init_backend(sb
);
2526 spin_lock_init(&sbi
->s_md_lock
);
2527 spin_lock_init(&sbi
->s_bal_lock
);
2529 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2530 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2531 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2532 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2533 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2534 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2536 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2537 if (sbi
->s_locality_groups
== NULL
) {
2541 for_each_possible_cpu(i
) {
2542 struct ext4_locality_group
*lg
;
2543 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2544 mutex_init(&lg
->lg_mutex
);
2545 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2546 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2547 spin_lock_init(&lg
->lg_prealloc_lock
);
2551 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2552 &ext4_mb_seq_groups_fops
, sb
);
2555 sbi
->s_journal
->j_commit_callback
= release_blocks_on_commit
;
2558 kfree(sbi
->s_mb_offsets
);
2559 kfree(sbi
->s_mb_maxs
);
2564 /* need to called with the ext4 group lock held */
2565 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2567 struct ext4_prealloc_space
*pa
;
2568 struct list_head
*cur
, *tmp
;
2571 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2572 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2573 list_del(&pa
->pa_group_list
);
2575 kmem_cache_free(ext4_pspace_cachep
, pa
);
2578 mb_debug(1, "mballoc: %u PAs left\n", count
);
2582 int ext4_mb_release(struct super_block
*sb
)
2584 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2586 int num_meta_group_infos
;
2587 struct ext4_group_info
*grinfo
;
2588 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2589 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2591 if (sbi
->s_group_info
) {
2592 for (i
= 0; i
< ngroups
; i
++) {
2593 grinfo
= ext4_get_group_info(sb
, i
);
2595 kfree(grinfo
->bb_bitmap
);
2597 ext4_lock_group(sb
, i
);
2598 ext4_mb_cleanup_pa(grinfo
);
2599 ext4_unlock_group(sb
, i
);
2600 kmem_cache_free(cachep
, grinfo
);
2602 num_meta_group_infos
= (ngroups
+
2603 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2604 EXT4_DESC_PER_BLOCK_BITS(sb
);
2605 for (i
= 0; i
< num_meta_group_infos
; i
++)
2606 kfree(sbi
->s_group_info
[i
]);
2607 kfree(sbi
->s_group_info
);
2609 kfree(sbi
->s_mb_offsets
);
2610 kfree(sbi
->s_mb_maxs
);
2611 if (sbi
->s_buddy_cache
)
2612 iput(sbi
->s_buddy_cache
);
2613 if (sbi
->s_mb_stats
) {
2615 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2616 atomic_read(&sbi
->s_bal_allocated
),
2617 atomic_read(&sbi
->s_bal_reqs
),
2618 atomic_read(&sbi
->s_bal_success
));
2620 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2621 "%u 2^N hits, %u breaks, %u lost\n",
2622 atomic_read(&sbi
->s_bal_ex_scanned
),
2623 atomic_read(&sbi
->s_bal_goals
),
2624 atomic_read(&sbi
->s_bal_2orders
),
2625 atomic_read(&sbi
->s_bal_breaks
),
2626 atomic_read(&sbi
->s_mb_lost_chunks
));
2628 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2629 sbi
->s_mb_buddies_generated
++,
2630 sbi
->s_mb_generation_time
);
2632 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2633 atomic_read(&sbi
->s_mb_preallocated
),
2634 atomic_read(&sbi
->s_mb_discarded
));
2637 free_percpu(sbi
->s_locality_groups
);
2639 remove_proc_entry("mb_groups", sbi
->s_proc
);
2644 static inline int ext4_issue_discard(struct super_block
*sb
,
2645 ext4_group_t block_group
, ext4_grpblk_t block
, int count
)
2647 ext4_fsblk_t discard_block
;
2649 discard_block
= block
+ ext4_group_first_block_no(sb
, block_group
);
2650 trace_ext4_discard_blocks(sb
,
2651 (unsigned long long) discard_block
, count
);
2652 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, 0);
2656 * This function is called by the jbd2 layer once the commit has finished,
2657 * so we know we can free the blocks that were released with that commit.
2659 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
)
2661 struct super_block
*sb
= journal
->j_private
;
2662 struct ext4_buddy e4b
;
2663 struct ext4_group_info
*db
;
2664 int err
, ret
, count
= 0, count2
= 0;
2665 struct ext4_free_data
*entry
;
2666 struct list_head
*l
, *ltmp
;
2668 list_for_each_safe(l
, ltmp
, &txn
->t_private_list
) {
2669 entry
= list_entry(l
, struct ext4_free_data
, list
);
2671 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2672 entry
->count
, entry
->group
, entry
);
2674 if (test_opt(sb
, DISCARD
)) {
2675 ret
= ext4_issue_discard(sb
, entry
->group
,
2676 entry
->start_blk
, entry
->count
);
2677 if (unlikely(ret
== -EOPNOTSUPP
)) {
2678 ext4_warning(sb
, "discard not supported, "
2680 clear_opt(sb
, DISCARD
);
2684 err
= ext4_mb_load_buddy(sb
, entry
->group
, &e4b
);
2685 /* we expect to find existing buddy because it's pinned */
2689 /* there are blocks to put in buddy to make them really free */
2690 count
+= entry
->count
;
2692 ext4_lock_group(sb
, entry
->group
);
2693 /* Take it out of per group rb tree */
2694 rb_erase(&entry
->node
, &(db
->bb_free_root
));
2695 mb_free_blocks(NULL
, &e4b
, entry
->start_blk
, entry
->count
);
2697 if (!db
->bb_free_root
.rb_node
) {
2698 /* No more items in the per group rb tree
2699 * balance refcounts from ext4_mb_free_metadata()
2701 page_cache_release(e4b
.bd_buddy_page
);
2702 page_cache_release(e4b
.bd_bitmap_page
);
2704 ext4_unlock_group(sb
, entry
->group
);
2705 kmem_cache_free(ext4_free_ext_cachep
, entry
);
2706 ext4_mb_unload_buddy(&e4b
);
2709 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2712 #ifdef CONFIG_EXT4_DEBUG
2713 u8 mb_enable_debug __read_mostly
;
2715 static struct dentry
*debugfs_dir
;
2716 static struct dentry
*debugfs_debug
;
2718 static void __init
ext4_create_debugfs_entry(void)
2720 debugfs_dir
= debugfs_create_dir("ext4", NULL
);
2722 debugfs_debug
= debugfs_create_u8("mballoc-debug",
2728 static void ext4_remove_debugfs_entry(void)
2730 debugfs_remove(debugfs_debug
);
2731 debugfs_remove(debugfs_dir
);
2736 static void __init
ext4_create_debugfs_entry(void)
2740 static void ext4_remove_debugfs_entry(void)
2746 int __init
ext4_init_mballoc(void)
2748 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2749 SLAB_RECLAIM_ACCOUNT
);
2750 if (ext4_pspace_cachep
== NULL
)
2753 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2754 SLAB_RECLAIM_ACCOUNT
);
2755 if (ext4_ac_cachep
== NULL
) {
2756 kmem_cache_destroy(ext4_pspace_cachep
);
2760 ext4_free_ext_cachep
= KMEM_CACHE(ext4_free_data
,
2761 SLAB_RECLAIM_ACCOUNT
);
2762 if (ext4_free_ext_cachep
== NULL
) {
2763 kmem_cache_destroy(ext4_pspace_cachep
);
2764 kmem_cache_destroy(ext4_ac_cachep
);
2767 ext4_create_debugfs_entry();
2771 void ext4_exit_mballoc(void)
2774 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2775 * before destroying the slab cache.
2778 kmem_cache_destroy(ext4_pspace_cachep
);
2779 kmem_cache_destroy(ext4_ac_cachep
);
2780 kmem_cache_destroy(ext4_free_ext_cachep
);
2781 ext4_groupinfo_destroy_slabs();
2782 ext4_remove_debugfs_entry();
2787 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2788 * Returns 0 if success or error code
2790 static noinline_for_stack
int
2791 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2792 handle_t
*handle
, unsigned int reserv_blks
)
2794 struct buffer_head
*bitmap_bh
= NULL
;
2795 struct ext4_group_desc
*gdp
;
2796 struct buffer_head
*gdp_bh
;
2797 struct ext4_sb_info
*sbi
;
2798 struct super_block
*sb
;
2802 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2803 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2809 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2813 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2818 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2822 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2823 ext4_free_blks_count(sb
, gdp
));
2825 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2829 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
2831 len
= ac
->ac_b_ex
.fe_len
;
2832 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2833 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2834 "fs metadata\n", block
, block
+len
);
2835 /* File system mounted not to panic on error
2836 * Fix the bitmap and repeat the block allocation
2837 * We leak some of the blocks here.
2839 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2840 mb_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2841 ac
->ac_b_ex
.fe_len
);
2842 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2843 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2849 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2850 #ifdef AGGRESSIVE_CHECK
2853 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2854 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2855 bitmap_bh
->b_data
));
2859 mb_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,ac
->ac_b_ex
.fe_len
);
2860 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2861 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2862 ext4_free_blks_set(sb
, gdp
,
2863 ext4_free_blocks_after_init(sb
,
2864 ac
->ac_b_ex
.fe_group
, gdp
));
2866 len
= ext4_free_blks_count(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
2867 ext4_free_blks_set(sb
, gdp
, len
);
2868 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
2870 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2871 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
2873 * Now reduce the dirty block count also. Should not go negative
2875 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2876 /* release all the reserved blocks if non delalloc */
2877 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
, reserv_blks
);
2879 if (sbi
->s_log_groups_per_flex
) {
2880 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2881 ac
->ac_b_ex
.fe_group
);
2882 atomic_sub(ac
->ac_b_ex
.fe_len
,
2883 &sbi
->s_flex_groups
[flex_group
].free_blocks
);
2886 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2889 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
2892 ext4_mark_super_dirty(sb
);
2898 * here we normalize request for locality group
2899 * Group request are normalized to s_strip size if we set the same via mount
2900 * option. If not we set it to s_mb_group_prealloc which can be configured via
2901 * /sys/fs/ext4/<partition>/mb_group_prealloc
2903 * XXX: should we try to preallocate more than the group has now?
2905 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2907 struct super_block
*sb
= ac
->ac_sb
;
2908 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2911 if (EXT4_SB(sb
)->s_stripe
)
2912 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
2914 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2915 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2916 current
->pid
, ac
->ac_g_ex
.fe_len
);
2920 * Normalization means making request better in terms of
2921 * size and alignment
2923 static noinline_for_stack
void
2924 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
2925 struct ext4_allocation_request
*ar
)
2929 loff_t size
, orig_size
, start_off
;
2931 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
2932 struct ext4_prealloc_space
*pa
;
2934 /* do normalize only data requests, metadata requests
2935 do not need preallocation */
2936 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
2939 /* sometime caller may want exact blocks */
2940 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2943 /* caller may indicate that preallocation isn't
2944 * required (it's a tail, for example) */
2945 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
2948 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
2949 ext4_mb_normalize_group_request(ac
);
2953 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
2955 /* first, let's learn actual file size
2956 * given current request is allocated */
2957 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
2958 size
= size
<< bsbits
;
2959 if (size
< i_size_read(ac
->ac_inode
))
2960 size
= i_size_read(ac
->ac_inode
);
2963 /* max size of free chunks */
2966 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2967 (req <= (size) || max <= (chunk_size))
2969 /* first, try to predict filesize */
2970 /* XXX: should this table be tunable? */
2972 if (size
<= 16 * 1024) {
2974 } else if (size
<= 32 * 1024) {
2976 } else if (size
<= 64 * 1024) {
2978 } else if (size
<= 128 * 1024) {
2980 } else if (size
<= 256 * 1024) {
2982 } else if (size
<= 512 * 1024) {
2984 } else if (size
<= 1024 * 1024) {
2986 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
2987 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2988 (21 - bsbits
)) << 21;
2989 size
= 2 * 1024 * 1024;
2990 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
2991 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2992 (22 - bsbits
)) << 22;
2993 size
= 4 * 1024 * 1024;
2994 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
2995 (8<<20)>>bsbits
, max
, 8 * 1024)) {
2996 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2997 (23 - bsbits
)) << 23;
2998 size
= 8 * 1024 * 1024;
3000 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3001 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3003 size
= size
>> bsbits
;
3004 start
= start_off
>> bsbits
;
3006 /* don't cover already allocated blocks in selected range */
3007 if (ar
->pleft
&& start
<= ar
->lleft
) {
3008 size
-= ar
->lleft
+ 1 - start
;
3009 start
= ar
->lleft
+ 1;
3011 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3012 size
-= start
+ size
- ar
->lright
;
3016 /* check we don't cross already preallocated blocks */
3018 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3023 spin_lock(&pa
->pa_lock
);
3024 if (pa
->pa_deleted
) {
3025 spin_unlock(&pa
->pa_lock
);
3029 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3031 /* PA must not overlap original request */
3032 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3033 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3035 /* skip PAs this normalized request doesn't overlap with */
3036 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
3037 spin_unlock(&pa
->pa_lock
);
3040 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3042 /* adjust start or end to be adjacent to this pa */
3043 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3044 BUG_ON(pa_end
< start
);
3046 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3047 BUG_ON(pa
->pa_lstart
> end
);
3048 end
= pa
->pa_lstart
;
3050 spin_unlock(&pa
->pa_lock
);
3055 /* XXX: extra loop to check we really don't overlap preallocations */
3057 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3059 spin_lock(&pa
->pa_lock
);
3060 if (pa
->pa_deleted
== 0) {
3061 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3062 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3064 spin_unlock(&pa
->pa_lock
);
3068 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3069 start
> ac
->ac_o_ex
.fe_logical
) {
3070 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3071 (unsigned long) start
, (unsigned long) size
,
3072 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3074 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3075 start
> ac
->ac_o_ex
.fe_logical
);
3076 BUG_ON(size
<= 0 || size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3078 /* now prepare goal request */
3080 /* XXX: is it better to align blocks WRT to logical
3081 * placement or satisfy big request as is */
3082 ac
->ac_g_ex
.fe_logical
= start
;
3083 ac
->ac_g_ex
.fe_len
= size
;
3085 /* define goal start in order to merge */
3086 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3087 /* merge to the right */
3088 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3089 &ac
->ac_f_ex
.fe_group
,
3090 &ac
->ac_f_ex
.fe_start
);
3091 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3093 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3094 /* merge to the left */
3095 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3096 &ac
->ac_f_ex
.fe_group
,
3097 &ac
->ac_f_ex
.fe_start
);
3098 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3101 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3102 (unsigned) orig_size
, (unsigned) start
);
3105 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3107 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3109 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3110 atomic_inc(&sbi
->s_bal_reqs
);
3111 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3112 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3113 atomic_inc(&sbi
->s_bal_success
);
3114 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3115 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3116 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3117 atomic_inc(&sbi
->s_bal_goals
);
3118 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3119 atomic_inc(&sbi
->s_bal_breaks
);
3122 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3123 trace_ext4_mballoc_alloc(ac
);
3125 trace_ext4_mballoc_prealloc(ac
);
3129 * Called on failure; free up any blocks from the inode PA for this
3130 * context. We don't need this for MB_GROUP_PA because we only change
3131 * pa_free in ext4_mb_release_context(), but on failure, we've already
3132 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3134 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3136 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3139 if (pa
&& pa
->pa_type
== MB_INODE_PA
) {
3140 len
= ac
->ac_b_ex
.fe_len
;
3147 * use blocks preallocated to inode
3149 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3150 struct ext4_prealloc_space
*pa
)
3156 /* found preallocated blocks, use them */
3157 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3158 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3160 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3161 &ac
->ac_b_ex
.fe_start
);
3162 ac
->ac_b_ex
.fe_len
= len
;
3163 ac
->ac_status
= AC_STATUS_FOUND
;
3166 BUG_ON(start
< pa
->pa_pstart
);
3167 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3168 BUG_ON(pa
->pa_free
< len
);
3171 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3175 * use blocks preallocated to locality group
3177 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3178 struct ext4_prealloc_space
*pa
)
3180 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3182 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3183 &ac
->ac_b_ex
.fe_group
,
3184 &ac
->ac_b_ex
.fe_start
);
3185 ac
->ac_b_ex
.fe_len
= len
;
3186 ac
->ac_status
= AC_STATUS_FOUND
;
3189 /* we don't correct pa_pstart or pa_plen here to avoid
3190 * possible race when the group is being loaded concurrently
3191 * instead we correct pa later, after blocks are marked
3192 * in on-disk bitmap -- see ext4_mb_release_context()
3193 * Other CPUs are prevented from allocating from this pa by lg_mutex
3195 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3199 * Return the prealloc space that have minimal distance
3200 * from the goal block. @cpa is the prealloc
3201 * space that is having currently known minimal distance
3202 * from the goal block.
3204 static struct ext4_prealloc_space
*
3205 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3206 struct ext4_prealloc_space
*pa
,
3207 struct ext4_prealloc_space
*cpa
)
3209 ext4_fsblk_t cur_distance
, new_distance
;
3212 atomic_inc(&pa
->pa_count
);
3215 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3216 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3218 if (cur_distance
<= new_distance
)
3221 /* drop the previous reference */
3222 atomic_dec(&cpa
->pa_count
);
3223 atomic_inc(&pa
->pa_count
);
3228 * search goal blocks in preallocated space
3230 static noinline_for_stack
int
3231 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3234 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3235 struct ext4_locality_group
*lg
;
3236 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3237 ext4_fsblk_t goal_block
;
3239 /* only data can be preallocated */
3240 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3243 /* first, try per-file preallocation */
3245 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3247 /* all fields in this condition don't change,
3248 * so we can skip locking for them */
3249 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3250 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3253 /* non-extent files can't have physical blocks past 2^32 */
3254 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3255 pa
->pa_pstart
+ pa
->pa_len
> EXT4_MAX_BLOCK_FILE_PHYS
)
3258 /* found preallocated blocks, use them */
3259 spin_lock(&pa
->pa_lock
);
3260 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3261 atomic_inc(&pa
->pa_count
);
3262 ext4_mb_use_inode_pa(ac
, pa
);
3263 spin_unlock(&pa
->pa_lock
);
3264 ac
->ac_criteria
= 10;
3268 spin_unlock(&pa
->pa_lock
);
3272 /* can we use group allocation? */
3273 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3276 /* inode may have no locality group for some reason */
3280 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3281 if (order
> PREALLOC_TB_SIZE
- 1)
3282 /* The max size of hash table is PREALLOC_TB_SIZE */
3283 order
= PREALLOC_TB_SIZE
- 1;
3285 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3287 * search for the prealloc space that is having
3288 * minimal distance from the goal block.
3290 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3292 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3294 spin_lock(&pa
->pa_lock
);
3295 if (pa
->pa_deleted
== 0 &&
3296 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3298 cpa
= ext4_mb_check_group_pa(goal_block
,
3301 spin_unlock(&pa
->pa_lock
);
3306 ext4_mb_use_group_pa(ac
, cpa
);
3307 ac
->ac_criteria
= 20;
3314 * the function goes through all block freed in the group
3315 * but not yet committed and marks them used in in-core bitmap.
3316 * buddy must be generated from this bitmap
3317 * Need to be called with the ext4 group lock held
3319 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3323 struct ext4_group_info
*grp
;
3324 struct ext4_free_data
*entry
;
3326 grp
= ext4_get_group_info(sb
, group
);
3327 n
= rb_first(&(grp
->bb_free_root
));
3330 entry
= rb_entry(n
, struct ext4_free_data
, node
);
3331 mb_set_bits(bitmap
, entry
->start_blk
, entry
->count
);
3338 * the function goes through all preallocation in this group and marks them
3339 * used in in-core bitmap. buddy must be generated from this bitmap
3340 * Need to be called with ext4 group lock held
3342 static noinline_for_stack
3343 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3346 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3347 struct ext4_prealloc_space
*pa
;
3348 struct list_head
*cur
;
3349 ext4_group_t groupnr
;
3350 ext4_grpblk_t start
;
3351 int preallocated
= 0;
3355 /* all form of preallocation discards first load group,
3356 * so the only competing code is preallocation use.
3357 * we don't need any locking here
3358 * notice we do NOT ignore preallocations with pa_deleted
3359 * otherwise we could leave used blocks available for
3360 * allocation in buddy when concurrent ext4_mb_put_pa()
3361 * is dropping preallocation
3363 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3364 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3365 spin_lock(&pa
->pa_lock
);
3366 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3369 spin_unlock(&pa
->pa_lock
);
3370 if (unlikely(len
== 0))
3372 BUG_ON(groupnr
!= group
);
3373 mb_set_bits(bitmap
, start
, len
);
3374 preallocated
+= len
;
3377 mb_debug(1, "prellocated %u for group %u\n", preallocated
, group
);
3380 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3382 struct ext4_prealloc_space
*pa
;
3383 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3384 kmem_cache_free(ext4_pspace_cachep
, pa
);
3388 * drops a reference to preallocated space descriptor
3389 * if this was the last reference and the space is consumed
3391 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3392 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3395 ext4_fsblk_t grp_blk
;
3397 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3400 /* in this short window concurrent discard can set pa_deleted */
3401 spin_lock(&pa
->pa_lock
);
3402 if (pa
->pa_deleted
== 1) {
3403 spin_unlock(&pa
->pa_lock
);
3408 spin_unlock(&pa
->pa_lock
);
3410 grp_blk
= pa
->pa_pstart
;
3412 * If doing group-based preallocation, pa_pstart may be in the
3413 * next group when pa is used up
3415 if (pa
->pa_type
== MB_GROUP_PA
)
3418 ext4_get_group_no_and_offset(sb
, grp_blk
, &grp
, NULL
);
3423 * P1 (buddy init) P2 (regular allocation)
3424 * find block B in PA
3425 * copy on-disk bitmap to buddy
3426 * mark B in on-disk bitmap
3427 * drop PA from group
3428 * mark all PAs in buddy
3430 * thus, P1 initializes buddy with B available. to prevent this
3431 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3434 ext4_lock_group(sb
, grp
);
3435 list_del(&pa
->pa_group_list
);
3436 ext4_unlock_group(sb
, grp
);
3438 spin_lock(pa
->pa_obj_lock
);
3439 list_del_rcu(&pa
->pa_inode_list
);
3440 spin_unlock(pa
->pa_obj_lock
);
3442 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3446 * creates new preallocated space for given inode
3448 static noinline_for_stack
int
3449 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3451 struct super_block
*sb
= ac
->ac_sb
;
3452 struct ext4_prealloc_space
*pa
;
3453 struct ext4_group_info
*grp
;
3454 struct ext4_inode_info
*ei
;
3456 /* preallocate only when found space is larger then requested */
3457 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3458 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3459 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3461 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3465 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3471 /* we can't allocate as much as normalizer wants.
3472 * so, found space must get proper lstart
3473 * to cover original request */
3474 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3475 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3477 /* we're limited by original request in that
3478 * logical block must be covered any way
3479 * winl is window we can move our chunk within */
3480 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3482 /* also, we should cover whole original request */
3483 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3485 /* the smallest one defines real window */
3486 win
= min(winl
, wins
);
3488 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3489 if (offs
&& offs
< win
)
3492 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3493 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3494 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3497 /* preallocation can change ac_b_ex, thus we store actually
3498 * allocated blocks for history */
3499 ac
->ac_f_ex
= ac
->ac_b_ex
;
3501 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3502 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3503 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3504 pa
->pa_free
= pa
->pa_len
;
3505 atomic_set(&pa
->pa_count
, 1);
3506 spin_lock_init(&pa
->pa_lock
);
3507 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3508 INIT_LIST_HEAD(&pa
->pa_group_list
);
3510 pa
->pa_type
= MB_INODE_PA
;
3512 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3513 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3514 trace_ext4_mb_new_inode_pa(ac
, pa
);
3516 ext4_mb_use_inode_pa(ac
, pa
);
3517 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3519 ei
= EXT4_I(ac
->ac_inode
);
3520 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3522 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3523 pa
->pa_inode
= ac
->ac_inode
;
3525 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3526 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3527 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3529 spin_lock(pa
->pa_obj_lock
);
3530 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3531 spin_unlock(pa
->pa_obj_lock
);
3537 * creates new preallocated space for locality group inodes belongs to
3539 static noinline_for_stack
int
3540 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3542 struct super_block
*sb
= ac
->ac_sb
;
3543 struct ext4_locality_group
*lg
;
3544 struct ext4_prealloc_space
*pa
;
3545 struct ext4_group_info
*grp
;
3547 /* preallocate only when found space is larger then requested */
3548 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3549 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3550 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3552 BUG_ON(ext4_pspace_cachep
== NULL
);
3553 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3557 /* preallocation can change ac_b_ex, thus we store actually
3558 * allocated blocks for history */
3559 ac
->ac_f_ex
= ac
->ac_b_ex
;
3561 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3562 pa
->pa_lstart
= pa
->pa_pstart
;
3563 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3564 pa
->pa_free
= pa
->pa_len
;
3565 atomic_set(&pa
->pa_count
, 1);
3566 spin_lock_init(&pa
->pa_lock
);
3567 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3568 INIT_LIST_HEAD(&pa
->pa_group_list
);
3570 pa
->pa_type
= MB_GROUP_PA
;
3572 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3573 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3574 trace_ext4_mb_new_group_pa(ac
, pa
);
3576 ext4_mb_use_group_pa(ac
, pa
);
3577 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3579 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3583 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3584 pa
->pa_inode
= NULL
;
3586 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3587 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3588 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3591 * We will later add the new pa to the right bucket
3592 * after updating the pa_free in ext4_mb_release_context
3597 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3601 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3602 err
= ext4_mb_new_group_pa(ac
);
3604 err
= ext4_mb_new_inode_pa(ac
);
3609 * finds all unused blocks in on-disk bitmap, frees them in
3610 * in-core bitmap and buddy.
3611 * @pa must be unlinked from inode and group lists, so that
3612 * nobody else can find/use it.
3613 * the caller MUST hold group/inode locks.
3614 * TODO: optimize the case when there are no in-core structures yet
3616 static noinline_for_stack
int
3617 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3618 struct ext4_prealloc_space
*pa
)
3620 struct super_block
*sb
= e4b
->bd_sb
;
3621 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3626 unsigned long long grp_blk_start
;
3630 BUG_ON(pa
->pa_deleted
== 0);
3631 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3632 grp_blk_start
= pa
->pa_pstart
- bit
;
3633 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3634 end
= bit
+ pa
->pa_len
;
3637 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3640 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3641 mb_debug(1, " free preallocated %u/%u in group %u\n",
3642 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3643 (unsigned) next
- bit
, (unsigned) group
);
3646 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3647 trace_ext4_mb_release_inode_pa(sb
, pa
->pa_inode
, pa
,
3648 grp_blk_start
+ bit
, next
- bit
);
3649 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3652 if (free
!= pa
->pa_free
) {
3653 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3654 pa
, (unsigned long) pa
->pa_lstart
,
3655 (unsigned long) pa
->pa_pstart
,
3656 (unsigned long) pa
->pa_len
);
3657 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3660 * pa is already deleted so we use the value obtained
3661 * from the bitmap and continue.
3664 atomic_add(free
, &sbi
->s_mb_discarded
);
3669 static noinline_for_stack
int
3670 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3671 struct ext4_prealloc_space
*pa
)
3673 struct super_block
*sb
= e4b
->bd_sb
;
3677 trace_ext4_mb_release_group_pa(sb
, pa
);
3678 BUG_ON(pa
->pa_deleted
== 0);
3679 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3680 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3681 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3682 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3683 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3689 * releases all preallocations in given group
3691 * first, we need to decide discard policy:
3692 * - when do we discard
3694 * - how many do we discard
3695 * 1) how many requested
3697 static noinline_for_stack
int
3698 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3699 ext4_group_t group
, int needed
)
3701 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3702 struct buffer_head
*bitmap_bh
= NULL
;
3703 struct ext4_prealloc_space
*pa
, *tmp
;
3704 struct list_head list
;
3705 struct ext4_buddy e4b
;
3710 mb_debug(1, "discard preallocation for group %u\n", group
);
3712 if (list_empty(&grp
->bb_prealloc_list
))
3715 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3716 if (bitmap_bh
== NULL
) {
3717 ext4_error(sb
, "Error reading block bitmap for %u", group
);
3721 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3723 ext4_error(sb
, "Error loading buddy information for %u", group
);
3729 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3731 INIT_LIST_HEAD(&list
);
3733 ext4_lock_group(sb
, group
);
3734 list_for_each_entry_safe(pa
, tmp
,
3735 &grp
->bb_prealloc_list
, pa_group_list
) {
3736 spin_lock(&pa
->pa_lock
);
3737 if (atomic_read(&pa
->pa_count
)) {
3738 spin_unlock(&pa
->pa_lock
);
3742 if (pa
->pa_deleted
) {
3743 spin_unlock(&pa
->pa_lock
);
3747 /* seems this one can be freed ... */
3750 /* we can trust pa_free ... */
3751 free
+= pa
->pa_free
;
3753 spin_unlock(&pa
->pa_lock
);
3755 list_del(&pa
->pa_group_list
);
3756 list_add(&pa
->u
.pa_tmp_list
, &list
);
3759 /* if we still need more blocks and some PAs were used, try again */
3760 if (free
< needed
&& busy
) {
3762 ext4_unlock_group(sb
, group
);
3764 * Yield the CPU here so that we don't get soft lockup
3765 * in non preempt case.
3771 /* found anything to free? */
3772 if (list_empty(&list
)) {
3777 /* now free all selected PAs */
3778 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3780 /* remove from object (inode or locality group) */
3781 spin_lock(pa
->pa_obj_lock
);
3782 list_del_rcu(&pa
->pa_inode_list
);
3783 spin_unlock(pa
->pa_obj_lock
);
3785 if (pa
->pa_type
== MB_GROUP_PA
)
3786 ext4_mb_release_group_pa(&e4b
, pa
);
3788 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3790 list_del(&pa
->u
.pa_tmp_list
);
3791 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3795 ext4_unlock_group(sb
, group
);
3796 ext4_mb_unload_buddy(&e4b
);
3802 * releases all non-used preallocated blocks for given inode
3804 * It's important to discard preallocations under i_data_sem
3805 * We don't want another block to be served from the prealloc
3806 * space when we are discarding the inode prealloc space.
3808 * FIXME!! Make sure it is valid at all the call sites
3810 void ext4_discard_preallocations(struct inode
*inode
)
3812 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3813 struct super_block
*sb
= inode
->i_sb
;
3814 struct buffer_head
*bitmap_bh
= NULL
;
3815 struct ext4_prealloc_space
*pa
, *tmp
;
3816 ext4_group_t group
= 0;
3817 struct list_head list
;
3818 struct ext4_buddy e4b
;
3821 if (!S_ISREG(inode
->i_mode
)) {
3822 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3826 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
3827 trace_ext4_discard_preallocations(inode
);
3829 INIT_LIST_HEAD(&list
);
3832 /* first, collect all pa's in the inode */
3833 spin_lock(&ei
->i_prealloc_lock
);
3834 while (!list_empty(&ei
->i_prealloc_list
)) {
3835 pa
= list_entry(ei
->i_prealloc_list
.next
,
3836 struct ext4_prealloc_space
, pa_inode_list
);
3837 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3838 spin_lock(&pa
->pa_lock
);
3839 if (atomic_read(&pa
->pa_count
)) {
3840 /* this shouldn't happen often - nobody should
3841 * use preallocation while we're discarding it */
3842 spin_unlock(&pa
->pa_lock
);
3843 spin_unlock(&ei
->i_prealloc_lock
);
3844 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
3846 schedule_timeout_uninterruptible(HZ
);
3850 if (pa
->pa_deleted
== 0) {
3852 spin_unlock(&pa
->pa_lock
);
3853 list_del_rcu(&pa
->pa_inode_list
);
3854 list_add(&pa
->u
.pa_tmp_list
, &list
);
3858 /* someone is deleting pa right now */
3859 spin_unlock(&pa
->pa_lock
);
3860 spin_unlock(&ei
->i_prealloc_lock
);
3862 /* we have to wait here because pa_deleted
3863 * doesn't mean pa is already unlinked from
3864 * the list. as we might be called from
3865 * ->clear_inode() the inode will get freed
3866 * and concurrent thread which is unlinking
3867 * pa from inode's list may access already
3868 * freed memory, bad-bad-bad */
3870 /* XXX: if this happens too often, we can
3871 * add a flag to force wait only in case
3872 * of ->clear_inode(), but not in case of
3873 * regular truncate */
3874 schedule_timeout_uninterruptible(HZ
);
3877 spin_unlock(&ei
->i_prealloc_lock
);
3879 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3880 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
3881 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
3883 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3885 ext4_error(sb
, "Error loading buddy information for %u",
3890 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3891 if (bitmap_bh
== NULL
) {
3892 ext4_error(sb
, "Error reading block bitmap for %u",
3894 ext4_mb_unload_buddy(&e4b
);
3898 ext4_lock_group(sb
, group
);
3899 list_del(&pa
->pa_group_list
);
3900 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3901 ext4_unlock_group(sb
, group
);
3903 ext4_mb_unload_buddy(&e4b
);
3906 list_del(&pa
->u
.pa_tmp_list
);
3907 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3911 #ifdef CONFIG_EXT4_DEBUG
3912 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3914 struct super_block
*sb
= ac
->ac_sb
;
3915 ext4_group_t ngroups
, i
;
3917 if (!mb_enable_debug
||
3918 (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
))
3921 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
3922 " Allocation context details:\n");
3923 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
3924 ac
->ac_status
, ac
->ac_flags
);
3925 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3926 "best %lu/%lu/%lu@%lu cr %d\n",
3927 (unsigned long)ac
->ac_o_ex
.fe_group
,
3928 (unsigned long)ac
->ac_o_ex
.fe_start
,
3929 (unsigned long)ac
->ac_o_ex
.fe_len
,
3930 (unsigned long)ac
->ac_o_ex
.fe_logical
,
3931 (unsigned long)ac
->ac_g_ex
.fe_group
,
3932 (unsigned long)ac
->ac_g_ex
.fe_start
,
3933 (unsigned long)ac
->ac_g_ex
.fe_len
,
3934 (unsigned long)ac
->ac_g_ex
.fe_logical
,
3935 (unsigned long)ac
->ac_b_ex
.fe_group
,
3936 (unsigned long)ac
->ac_b_ex
.fe_start
,
3937 (unsigned long)ac
->ac_b_ex
.fe_len
,
3938 (unsigned long)ac
->ac_b_ex
.fe_logical
,
3939 (int)ac
->ac_criteria
);
3940 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
3942 printk(KERN_ERR
"EXT4-fs: groups: \n");
3943 ngroups
= ext4_get_groups_count(sb
);
3944 for (i
= 0; i
< ngroups
; i
++) {
3945 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
3946 struct ext4_prealloc_space
*pa
;
3947 ext4_grpblk_t start
;
3948 struct list_head
*cur
;
3949 ext4_lock_group(sb
, i
);
3950 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3951 pa
= list_entry(cur
, struct ext4_prealloc_space
,
3953 spin_lock(&pa
->pa_lock
);
3954 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3956 spin_unlock(&pa
->pa_lock
);
3957 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
3960 ext4_unlock_group(sb
, i
);
3962 if (grp
->bb_free
== 0)
3964 printk(KERN_ERR
"%u: %d/%d \n",
3965 i
, grp
->bb_free
, grp
->bb_fragments
);
3967 printk(KERN_ERR
"\n");
3970 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3977 * We use locality group preallocation for small size file. The size of the
3978 * file is determined by the current size or the resulting size after
3979 * allocation which ever is larger
3981 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3983 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
3985 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3986 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3989 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3992 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3995 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3996 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
3999 if ((size
== isize
) &&
4000 !ext4_fs_is_busy(sbi
) &&
4001 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
4002 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
4006 /* don't use group allocation for large files */
4007 size
= max(size
, isize
);
4008 if (size
> sbi
->s_mb_stream_request
) {
4009 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4013 BUG_ON(ac
->ac_lg
!= NULL
);
4015 * locality group prealloc space are per cpu. The reason for having
4016 * per cpu locality group is to reduce the contention between block
4017 * request from multiple CPUs.
4019 ac
->ac_lg
= __this_cpu_ptr(sbi
->s_locality_groups
);
4021 /* we're going to use group allocation */
4022 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4024 /* serialize all allocations in the group */
4025 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4028 static noinline_for_stack
int
4029 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4030 struct ext4_allocation_request
*ar
)
4032 struct super_block
*sb
= ar
->inode
->i_sb
;
4033 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4034 struct ext4_super_block
*es
= sbi
->s_es
;
4038 ext4_grpblk_t block
;
4040 /* we can't allocate > group size */
4043 /* just a dirty hack to filter too big requests */
4044 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4045 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4047 /* start searching from the goal */
4049 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4050 goal
>= ext4_blocks_count(es
))
4051 goal
= le32_to_cpu(es
->s_first_data_block
);
4052 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4054 /* set up allocation goals */
4055 memset(ac
, 0, sizeof(struct ext4_allocation_context
));
4056 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4057 ac
->ac_status
= AC_STATUS_CONTINUE
;
4059 ac
->ac_inode
= ar
->inode
;
4060 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4061 ac
->ac_o_ex
.fe_group
= group
;
4062 ac
->ac_o_ex
.fe_start
= block
;
4063 ac
->ac_o_ex
.fe_len
= len
;
4064 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4065 ac
->ac_g_ex
.fe_group
= group
;
4066 ac
->ac_g_ex
.fe_start
= block
;
4067 ac
->ac_g_ex
.fe_len
= len
;
4068 ac
->ac_flags
= ar
->flags
;
4070 /* we have to define context: we'll we work with a file or
4071 * locality group. this is a policy, actually */
4072 ext4_mb_group_or_file(ac
);
4074 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4075 "left: %u/%u, right %u/%u to %swritable\n",
4076 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4077 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4078 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4079 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4080 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4085 static noinline_for_stack
void
4086 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4087 struct ext4_locality_group
*lg
,
4088 int order
, int total_entries
)
4090 ext4_group_t group
= 0;
4091 struct ext4_buddy e4b
;
4092 struct list_head discard_list
;
4093 struct ext4_prealloc_space
*pa
, *tmp
;
4095 mb_debug(1, "discard locality group preallocation\n");
4097 INIT_LIST_HEAD(&discard_list
);
4099 spin_lock(&lg
->lg_prealloc_lock
);
4100 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4102 spin_lock(&pa
->pa_lock
);
4103 if (atomic_read(&pa
->pa_count
)) {
4105 * This is the pa that we just used
4106 * for block allocation. So don't
4109 spin_unlock(&pa
->pa_lock
);
4112 if (pa
->pa_deleted
) {
4113 spin_unlock(&pa
->pa_lock
);
4116 /* only lg prealloc space */
4117 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4119 /* seems this one can be freed ... */
4121 spin_unlock(&pa
->pa_lock
);
4123 list_del_rcu(&pa
->pa_inode_list
);
4124 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4127 if (total_entries
<= 5) {
4129 * we want to keep only 5 entries
4130 * allowing it to grow to 8. This
4131 * mak sure we don't call discard
4132 * soon for this list.
4137 spin_unlock(&lg
->lg_prealloc_lock
);
4139 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4141 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4142 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4143 ext4_error(sb
, "Error loading buddy information for %u",
4147 ext4_lock_group(sb
, group
);
4148 list_del(&pa
->pa_group_list
);
4149 ext4_mb_release_group_pa(&e4b
, pa
);
4150 ext4_unlock_group(sb
, group
);
4152 ext4_mb_unload_buddy(&e4b
);
4153 list_del(&pa
->u
.pa_tmp_list
);
4154 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4159 * We have incremented pa_count. So it cannot be freed at this
4160 * point. Also we hold lg_mutex. So no parallel allocation is
4161 * possible from this lg. That means pa_free cannot be updated.
4163 * A parallel ext4_mb_discard_group_preallocations is possible.
4164 * which can cause the lg_prealloc_list to be updated.
4167 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4169 int order
, added
= 0, lg_prealloc_count
= 1;
4170 struct super_block
*sb
= ac
->ac_sb
;
4171 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4172 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4174 order
= fls(pa
->pa_free
) - 1;
4175 if (order
> PREALLOC_TB_SIZE
- 1)
4176 /* The max size of hash table is PREALLOC_TB_SIZE */
4177 order
= PREALLOC_TB_SIZE
- 1;
4178 /* Add the prealloc space to lg */
4180 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4182 spin_lock(&tmp_pa
->pa_lock
);
4183 if (tmp_pa
->pa_deleted
) {
4184 spin_unlock(&tmp_pa
->pa_lock
);
4187 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4188 /* Add to the tail of the previous entry */
4189 list_add_tail_rcu(&pa
->pa_inode_list
,
4190 &tmp_pa
->pa_inode_list
);
4193 * we want to count the total
4194 * number of entries in the list
4197 spin_unlock(&tmp_pa
->pa_lock
);
4198 lg_prealloc_count
++;
4201 list_add_tail_rcu(&pa
->pa_inode_list
,
4202 &lg
->lg_prealloc_list
[order
]);
4205 /* Now trim the list to be not more than 8 elements */
4206 if (lg_prealloc_count
> 8) {
4207 ext4_mb_discard_lg_preallocations(sb
, lg
,
4208 order
, lg_prealloc_count
);
4215 * release all resource we used in allocation
4217 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4219 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4221 if (pa
->pa_type
== MB_GROUP_PA
) {
4222 /* see comment in ext4_mb_use_group_pa() */
4223 spin_lock(&pa
->pa_lock
);
4224 pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4225 pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4226 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4227 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4228 spin_unlock(&pa
->pa_lock
);
4232 up_read(ac
->alloc_semp
);
4235 * We want to add the pa to the right bucket.
4236 * Remove it from the list and while adding
4237 * make sure the list to which we are adding
4238 * doesn't grow big. We need to release
4239 * alloc_semp before calling ext4_mb_add_n_trim()
4241 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4242 spin_lock(pa
->pa_obj_lock
);
4243 list_del_rcu(&pa
->pa_inode_list
);
4244 spin_unlock(pa
->pa_obj_lock
);
4245 ext4_mb_add_n_trim(ac
);
4247 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4249 if (ac
->ac_bitmap_page
)
4250 page_cache_release(ac
->ac_bitmap_page
);
4251 if (ac
->ac_buddy_page
)
4252 page_cache_release(ac
->ac_buddy_page
);
4253 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4254 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4255 ext4_mb_collect_stats(ac
);
4259 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4261 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4265 trace_ext4_mb_discard_preallocations(sb
, needed
);
4266 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4267 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4276 * Main entry point into mballoc to allocate blocks
4277 * it tries to use preallocation first, then falls back
4278 * to usual allocation
4280 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4281 struct ext4_allocation_request
*ar
, int *errp
)
4284 struct ext4_allocation_context
*ac
= NULL
;
4285 struct ext4_sb_info
*sbi
;
4286 struct super_block
*sb
;
4287 ext4_fsblk_t block
= 0;
4288 unsigned int inquota
= 0;
4289 unsigned int reserv_blks
= 0;
4291 sb
= ar
->inode
->i_sb
;
4294 trace_ext4_request_blocks(ar
);
4297 * For delayed allocation, we could skip the ENOSPC and
4298 * EDQUOT check, as blocks and quotas have been already
4299 * reserved when data being copied into pagecache.
4301 if (ext4_test_inode_state(ar
->inode
, EXT4_STATE_DELALLOC_RESERVED
))
4302 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4304 /* Without delayed allocation we need to verify
4305 * there is enough free blocks to do block allocation
4306 * and verify allocation doesn't exceed the quota limits.
4308 while (ar
->len
&& ext4_claim_free_blocks(sbi
, ar
->len
)) {
4309 /* let others to free the space */
4311 ar
->len
= ar
->len
>> 1;
4317 reserv_blks
= ar
->len
;
4318 while (ar
->len
&& dquot_alloc_block(ar
->inode
, ar
->len
)) {
4319 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4329 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4336 *errp
= ext4_mb_initialize_context(ac
, ar
);
4342 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4343 if (!ext4_mb_use_preallocated(ac
)) {
4344 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4345 ext4_mb_normalize_request(ac
, ar
);
4347 /* allocate space in core */
4348 *errp
= ext4_mb_regular_allocator(ac
);
4352 /* as we've just preallocated more space than
4353 * user requested orinally, we store allocated
4354 * space in a special descriptor */
4355 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4356 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4357 ext4_mb_new_preallocation(ac
);
4359 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4360 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_blks
);
4361 if (*errp
== -EAGAIN
) {
4363 * drop the reference that we took
4364 * in ext4_mb_use_best_found
4366 ext4_mb_release_context(ac
);
4367 ac
->ac_b_ex
.fe_group
= 0;
4368 ac
->ac_b_ex
.fe_start
= 0;
4369 ac
->ac_b_ex
.fe_len
= 0;
4370 ac
->ac_status
= AC_STATUS_CONTINUE
;
4374 ext4_discard_allocated_blocks(ac
);
4376 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4377 ar
->len
= ac
->ac_b_ex
.fe_len
;
4380 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4387 ac
->ac_b_ex
.fe_len
= 0;
4389 ext4_mb_show_ac(ac
);
4391 ext4_mb_release_context(ac
);
4394 kmem_cache_free(ext4_ac_cachep
, ac
);
4395 if (inquota
&& ar
->len
< inquota
)
4396 dquot_free_block(ar
->inode
, inquota
- ar
->len
);
4398 if (!ext4_test_inode_state(ar
->inode
,
4399 EXT4_STATE_DELALLOC_RESERVED
))
4400 /* release all the reserved blocks if non delalloc */
4401 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
,
4405 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4411 * We can merge two free data extents only if the physical blocks
4412 * are contiguous, AND the extents were freed by the same transaction,
4413 * AND the blocks are associated with the same group.
4415 static int can_merge(struct ext4_free_data
*entry1
,
4416 struct ext4_free_data
*entry2
)
4418 if ((entry1
->t_tid
== entry2
->t_tid
) &&
4419 (entry1
->group
== entry2
->group
) &&
4420 ((entry1
->start_blk
+ entry1
->count
) == entry2
->start_blk
))
4425 static noinline_for_stack
int
4426 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4427 struct ext4_free_data
*new_entry
)
4429 ext4_group_t group
= e4b
->bd_group
;
4430 ext4_grpblk_t block
;
4431 struct ext4_free_data
*entry
;
4432 struct ext4_group_info
*db
= e4b
->bd_info
;
4433 struct super_block
*sb
= e4b
->bd_sb
;
4434 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4435 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4436 struct rb_node
*parent
= NULL
, *new_node
;
4438 BUG_ON(!ext4_handle_valid(handle
));
4439 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4440 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4442 new_node
= &new_entry
->node
;
4443 block
= new_entry
->start_blk
;
4446 /* first free block exent. We need to
4447 protect buddy cache from being freed,
4448 * otherwise we'll refresh it from
4449 * on-disk bitmap and lose not-yet-available
4451 page_cache_get(e4b
->bd_buddy_page
);
4452 page_cache_get(e4b
->bd_bitmap_page
);
4456 entry
= rb_entry(parent
, struct ext4_free_data
, node
);
4457 if (block
< entry
->start_blk
)
4459 else if (block
>= (entry
->start_blk
+ entry
->count
))
4460 n
= &(*n
)->rb_right
;
4462 ext4_grp_locked_error(sb
, group
, 0,
4463 ext4_group_first_block_no(sb
, group
) + block
,
4464 "Block already on to-be-freed list");
4469 rb_link_node(new_node
, parent
, n
);
4470 rb_insert_color(new_node
, &db
->bb_free_root
);
4472 /* Now try to see the extent can be merged to left and right */
4473 node
= rb_prev(new_node
);
4475 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4476 if (can_merge(entry
, new_entry
)) {
4477 new_entry
->start_blk
= entry
->start_blk
;
4478 new_entry
->count
+= entry
->count
;
4479 rb_erase(node
, &(db
->bb_free_root
));
4480 spin_lock(&sbi
->s_md_lock
);
4481 list_del(&entry
->list
);
4482 spin_unlock(&sbi
->s_md_lock
);
4483 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4487 node
= rb_next(new_node
);
4489 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4490 if (can_merge(new_entry
, entry
)) {
4491 new_entry
->count
+= entry
->count
;
4492 rb_erase(node
, &(db
->bb_free_root
));
4493 spin_lock(&sbi
->s_md_lock
);
4494 list_del(&entry
->list
);
4495 spin_unlock(&sbi
->s_md_lock
);
4496 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4499 /* Add the extent to transaction's private list */
4500 spin_lock(&sbi
->s_md_lock
);
4501 list_add(&new_entry
->list
, &handle
->h_transaction
->t_private_list
);
4502 spin_unlock(&sbi
->s_md_lock
);
4507 * ext4_free_blocks() -- Free given blocks and update quota
4508 * @handle: handle for this transaction
4510 * @block: start physical block to free
4511 * @count: number of blocks to count
4512 * @metadata: Are these metadata blocks
4514 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4515 struct buffer_head
*bh
, ext4_fsblk_t block
,
4516 unsigned long count
, int flags
)
4518 struct buffer_head
*bitmap_bh
= NULL
;
4519 struct super_block
*sb
= inode
->i_sb
;
4520 struct ext4_group_desc
*gdp
;
4521 unsigned long freed
= 0;
4522 unsigned int overflow
;
4524 struct buffer_head
*gd_bh
;
4525 ext4_group_t block_group
;
4526 struct ext4_sb_info
*sbi
;
4527 struct ext4_buddy e4b
;
4533 BUG_ON(block
!= bh
->b_blocknr
);
4535 block
= bh
->b_blocknr
;
4539 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4540 !ext4_data_block_valid(sbi
, block
, count
)) {
4541 ext4_error(sb
, "Freeing blocks not in datazone - "
4542 "block = %llu, count = %lu", block
, count
);
4546 ext4_debug("freeing block %llu\n", block
);
4547 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4549 if (flags
& EXT4_FREE_BLOCKS_FORGET
) {
4550 struct buffer_head
*tbh
= bh
;
4553 BUG_ON(bh
&& (count
> 1));
4555 for (i
= 0; i
< count
; i
++) {
4557 tbh
= sb_find_get_block(inode
->i_sb
,
4561 ext4_forget(handle
, flags
& EXT4_FREE_BLOCKS_METADATA
,
4562 inode
, tbh
, block
+ i
);
4567 * We need to make sure we don't reuse the freed block until
4568 * after the transaction is committed, which we can do by
4569 * treating the block as metadata, below. We make an
4570 * exception if the inode is to be written in writeback mode
4571 * since writeback mode has weak data consistency guarantees.
4573 if (!ext4_should_writeback_data(inode
))
4574 flags
|= EXT4_FREE_BLOCKS_METADATA
;
4578 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4581 * Check to see if we are freeing blocks across a group
4584 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4585 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4588 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4593 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4599 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4600 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4601 in_range(block
, ext4_inode_table(sb
, gdp
),
4602 EXT4_SB(sb
)->s_itb_per_group
) ||
4603 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4604 EXT4_SB(sb
)->s_itb_per_group
)) {
4606 ext4_error(sb
, "Freeing blocks in system zone - "
4607 "Block = %llu, count = %lu", block
, count
);
4608 /* err = 0. ext4_std_error should be a no op */
4612 BUFFER_TRACE(bitmap_bh
, "getting write access");
4613 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4618 * We are about to modify some metadata. Call the journal APIs
4619 * to unshare ->b_data if a currently-committing transaction is
4622 BUFFER_TRACE(gd_bh
, "get_write_access");
4623 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4626 #ifdef AGGRESSIVE_CHECK
4629 for (i
= 0; i
< count
; i
++)
4630 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4633 trace_ext4_mballoc_free(sb
, inode
, block_group
, bit
, count
);
4635 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4639 if ((flags
& EXT4_FREE_BLOCKS_METADATA
) && ext4_handle_valid(handle
)) {
4640 struct ext4_free_data
*new_entry
;
4642 * blocks being freed are metadata. these blocks shouldn't
4643 * be used until this transaction is committed
4645 new_entry
= kmem_cache_alloc(ext4_free_ext_cachep
, GFP_NOFS
);
4650 new_entry
->start_blk
= bit
;
4651 new_entry
->group
= block_group
;
4652 new_entry
->count
= count
;
4653 new_entry
->t_tid
= handle
->h_transaction
->t_tid
;
4655 ext4_lock_group(sb
, block_group
);
4656 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4657 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4659 /* need to update group_info->bb_free and bitmap
4660 * with group lock held. generate_buddy look at
4661 * them with group lock_held
4663 ext4_lock_group(sb
, block_group
);
4664 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4665 mb_free_blocks(inode
, &e4b
, bit
, count
);
4668 ret
= ext4_free_blks_count(sb
, gdp
) + count
;
4669 ext4_free_blks_set(sb
, gdp
, ret
);
4670 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4671 ext4_unlock_group(sb
, block_group
);
4672 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4674 if (sbi
->s_log_groups_per_flex
) {
4675 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4676 atomic_add(count
, &sbi
->s_flex_groups
[flex_group
].free_blocks
);
4679 ext4_mb_unload_buddy(&e4b
);
4683 /* We dirtied the bitmap block */
4684 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4685 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4687 /* And the group descriptor block */
4688 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4689 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4693 if (overflow
&& !err
) {
4699 ext4_mark_super_dirty(sb
);
4702 dquot_free_block(inode
, freed
);
4704 ext4_std_error(sb
, err
);
4709 * ext4_trim_extent -- function to TRIM one single free extent in the group
4710 * @sb: super block for the file system
4711 * @start: starting block of the free extent in the alloc. group
4712 * @count: number of blocks to TRIM
4713 * @group: alloc. group we are working with
4714 * @e4b: ext4 buddy for the group
4716 * Trim "count" blocks starting at "start" in the "group". To assure that no
4717 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4718 * be called with under the group lock.
4720 static int ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
4721 ext4_group_t group
, struct ext4_buddy
*e4b
)
4723 struct ext4_free_extent ex
;
4726 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
4728 ex
.fe_start
= start
;
4729 ex
.fe_group
= group
;
4733 * Mark blocks used, so no one can reuse them while
4736 mb_mark_used(e4b
, &ex
);
4737 ext4_unlock_group(sb
, group
);
4739 ret
= ext4_issue_discard(sb
, group
, start
, count
);
4741 ext4_lock_group(sb
, group
);
4742 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
4747 * ext4_trim_all_free -- function to trim all free space in alloc. group
4748 * @sb: super block for file system
4750 * @start: first group block to examine
4751 * @max: last group block to examine
4752 * @minblocks: minimum extent block count
4754 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4755 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4759 * ext4_trim_all_free walks through group's block bitmap searching for free
4760 * extents. When the free extent is found, mark it as used in group buddy
4761 * bitmap. Then issue a TRIM command on this extent and free the extent in
4762 * the group buddy bitmap. This is done until whole group is scanned.
4764 static ext4_grpblk_t
4765 ext4_trim_all_free(struct super_block
*sb
, struct ext4_buddy
*e4b
,
4766 ext4_grpblk_t start
, ext4_grpblk_t max
, ext4_grpblk_t minblocks
)
4769 ext4_grpblk_t next
, count
= 0;
4773 BUG_ON(e4b
== NULL
);
4775 bitmap
= e4b
->bd_bitmap
;
4776 group
= e4b
->bd_group
;
4777 start
= (e4b
->bd_info
->bb_first_free
> start
) ?
4778 e4b
->bd_info
->bb_first_free
: start
;
4779 ext4_lock_group(sb
, group
);
4781 while (start
< max
) {
4782 start
= mb_find_next_zero_bit(bitmap
, max
, start
);
4785 next
= mb_find_next_bit(bitmap
, max
, start
);
4787 if ((next
- start
) >= minblocks
) {
4788 ret
= ext4_trim_extent(sb
, start
,
4789 next
- start
, group
, e4b
);
4792 count
+= next
- start
;
4796 if (fatal_signal_pending(current
)) {
4797 count
= -ERESTARTSYS
;
4801 if (need_resched()) {
4802 ext4_unlock_group(sb
, group
);
4804 ext4_lock_group(sb
, group
);
4807 if ((e4b
->bd_info
->bb_free
- count
) < minblocks
)
4810 ext4_unlock_group(sb
, group
);
4812 ext4_debug("trimmed %d blocks in the group %d\n",
4822 * ext4_trim_fs() -- trim ioctl handle function
4823 * @sb: superblock for filesystem
4824 * @range: fstrim_range structure
4826 * start: First Byte to trim
4827 * len: number of Bytes to trim from start
4828 * minlen: minimum extent length in Bytes
4829 * ext4_trim_fs goes through all allocation groups containing Bytes from
4830 * start to start+len. For each such a group ext4_trim_all_free function
4831 * is invoked to trim all free space.
4833 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
4835 struct ext4_buddy e4b
;
4836 ext4_group_t first_group
, last_group
;
4837 ext4_group_t group
, ngroups
= ext4_get_groups_count(sb
);
4838 ext4_grpblk_t cnt
= 0, first_block
, last_block
;
4839 uint64_t start
, len
, minlen
, trimmed
;
4840 ext4_fsblk_t first_data_blk
=
4841 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
4844 start
= range
->start
>> sb
->s_blocksize_bits
;
4845 len
= range
->len
>> sb
->s_blocksize_bits
;
4846 minlen
= range
->minlen
>> sb
->s_blocksize_bits
;
4849 if (unlikely(minlen
> EXT4_BLOCKS_PER_GROUP(sb
)))
4851 if (start
< first_data_blk
) {
4852 len
-= first_data_blk
- start
;
4853 start
= first_data_blk
;
4856 /* Determine first and last group to examine based on start and len */
4857 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
4858 &first_group
, &first_block
);
4859 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) (start
+ len
),
4860 &last_group
, &last_block
);
4861 last_group
= (last_group
> ngroups
- 1) ? ngroups
- 1 : last_group
;
4862 last_block
= EXT4_BLOCKS_PER_GROUP(sb
);
4864 if (first_group
> last_group
)
4867 for (group
= first_group
; group
<= last_group
; group
++) {
4868 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
4870 ext4_error(sb
, "Error in loading buddy "
4871 "information for %u", group
);
4876 * For all the groups except the last one, last block will
4877 * always be EXT4_BLOCKS_PER_GROUP(sb), so we only need to
4878 * change it for the last group in which case start +
4879 * len < EXT4_BLOCKS_PER_GROUP(sb).
4881 if (first_block
+ len
< EXT4_BLOCKS_PER_GROUP(sb
))
4882 last_block
= first_block
+ len
;
4883 len
-= last_block
- first_block
;
4885 if (e4b
.bd_info
->bb_free
>= minlen
) {
4886 cnt
= ext4_trim_all_free(sb
, &e4b
, first_block
,
4887 last_block
, minlen
);
4890 ext4_mb_unload_buddy(&e4b
);
4894 ext4_mb_unload_buddy(&e4b
);
4898 range
->len
= trimmed
* sb
->s_blocksize
;