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 makes sure that
79 * 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 represented 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 * the smallest multiple of the stripe value (sbi->s_stripe) which is
135 * greater than the default mb_group_prealloc.
137 * The regular allocator (using the buddy cache) supports a few tunables.
139 * /sys/fs/ext4/<partition>/mb_min_to_scan
140 * /sys/fs/ext4/<partition>/mb_max_to_scan
141 * /sys/fs/ext4/<partition>/mb_order2_req
143 * The regular allocator uses buddy scan only if the request len is power of
144 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
145 * value of s_mb_order2_reqs can be tuned via
146 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
147 * stripe size (sbi->s_stripe), we try to search for contiguous block in
148 * stripe size. This should result in better allocation on RAID setups. If
149 * not, we search in the specific group using bitmap for best extents. The
150 * tunable min_to_scan and max_to_scan control the behaviour here.
151 * min_to_scan indicate how long the mballoc __must__ look for a best
152 * extent and max_to_scan indicates how long the mballoc __can__ look for a
153 * best extent in the found extents. Searching for the blocks starts with
154 * the group specified as the goal value in allocation context via
155 * ac_g_ex. Each group is first checked based on the criteria whether it
156 * can be used for allocation. ext4_mb_good_group explains how the groups are
159 * Both the prealloc space are getting populated as above. So for the first
160 * request we will hit the buddy cache which will result in this prealloc
161 * space getting filled. The prealloc space is then later used for the
162 * subsequent request.
166 * mballoc operates on the following data:
168 * - in-core buddy (actually includes buddy and bitmap)
169 * - preallocation descriptors (PAs)
171 * there are two types of preallocations:
173 * assiged to specific inode and can be used for this inode only.
174 * it describes part of inode's space preallocated to specific
175 * physical blocks. any block from that preallocated can be used
176 * independent. the descriptor just tracks number of blocks left
177 * unused. so, before taking some block from descriptor, one must
178 * make sure corresponded logical block isn't allocated yet. this
179 * also means that freeing any block within descriptor's range
180 * must discard all preallocated blocks.
182 * assigned to specific locality group which does not translate to
183 * permanent set of inodes: inode can join and leave group. space
184 * from this type of preallocation can be used for any inode. thus
185 * it's consumed from the beginning to the end.
187 * relation between them can be expressed as:
188 * in-core buddy = on-disk bitmap + preallocation descriptors
190 * this mean blocks mballoc considers used are:
191 * - allocated blocks (persistent)
192 * - preallocated blocks (non-persistent)
194 * consistency in mballoc world means that at any time a block is either
195 * free or used in ALL structures. notice: "any time" should not be read
196 * literally -- time is discrete and delimited by locks.
198 * to keep it simple, we don't use block numbers, instead we count number of
199 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
201 * all operations can be expressed as:
202 * - init buddy: buddy = on-disk + PAs
203 * - new PA: buddy += N; PA = N
204 * - use inode PA: on-disk += N; PA -= N
205 * - discard inode PA buddy -= on-disk - PA; PA = 0
206 * - use locality group PA on-disk += N; PA -= N
207 * - discard locality group PA buddy -= PA; PA = 0
208 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
209 * is used in real operation because we can't know actual used
210 * bits from PA, only from on-disk bitmap
212 * if we follow this strict logic, then all operations above should be atomic.
213 * given some of them can block, we'd have to use something like semaphores
214 * killing performance on high-end SMP hardware. let's try to relax it using
215 * the following knowledge:
216 * 1) if buddy is referenced, it's already initialized
217 * 2) while block is used in buddy and the buddy is referenced,
218 * nobody can re-allocate that block
219 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
220 * bit set and PA claims same block, it's OK. IOW, one can set bit in
221 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
224 * so, now we're building a concurrency table:
227 * blocks for PA are allocated in the buddy, buddy must be referenced
228 * until PA is linked to allocation group to avoid concurrent buddy init
230 * we need to make sure that either on-disk bitmap or PA has uptodate data
231 * given (3) we care that PA-=N operation doesn't interfere with init
233 * the simplest way would be to have buddy initialized by the discard
234 * - use locality group PA
235 * again PA-=N must be serialized with init
236 * - discard locality group PA
237 * the simplest way would be to have buddy initialized by the discard
240 * i_data_sem serializes them
242 * discard process must wait until PA isn't used by another process
243 * - use locality group PA
244 * some mutex should serialize them
245 * - discard locality group PA
246 * discard process must wait until PA isn't used by another process
249 * i_data_sem or another mutex should serializes them
251 * discard process must wait until PA isn't used by another process
252 * - use locality group PA
253 * nothing wrong here -- they're different PAs covering different blocks
254 * - discard locality group PA
255 * discard process must wait until PA isn't used by another process
257 * now we're ready to make few consequences:
258 * - PA is referenced and while it is no discard is possible
259 * - PA is referenced until block isn't marked in on-disk bitmap
260 * - PA changes only after on-disk bitmap
261 * - discard must not compete with init. either init is done before
262 * any discard or they're serialized somehow
263 * - buddy init as sum of on-disk bitmap and PAs is done atomically
265 * a special case when we've used PA to emptiness. no need to modify buddy
266 * in this case, but we should care about concurrent init
271 * Logic in few words:
276 * mark bits in on-disk bitmap
279 * - use preallocation:
280 * find proper PA (per-inode or group)
282 * mark bits in on-disk bitmap
288 * mark bits in on-disk bitmap
291 * - discard preallocations in group:
293 * move them onto local list
294 * load on-disk bitmap
296 * remove PA from object (inode or locality group)
297 * mark free blocks in-core
299 * - discard inode's preallocations:
306 * - bitlock on a group (group)
307 * - object (inode/locality) (object)
318 * - release consumed pa:
323 * - generate in-core bitmap:
327 * - discard all for given object (inode, locality group):
332 * - discard all for given group:
339 static struct kmem_cache
*ext4_pspace_cachep
;
340 static struct kmem_cache
*ext4_ac_cachep
;
341 static struct kmem_cache
*ext4_free_ext_cachep
;
343 /* We create slab caches for groupinfo data structures based on the
344 * superblock block size. There will be one per mounted filesystem for
345 * each unique s_blocksize_bits */
346 #define NR_GRPINFO_CACHES 8
347 static struct kmem_cache
*ext4_groupinfo_caches
[NR_GRPINFO_CACHES
];
349 static const char *ext4_groupinfo_slab_names
[NR_GRPINFO_CACHES
] = {
350 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
351 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
352 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
355 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
357 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
359 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
);
361 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
363 #if BITS_PER_LONG == 64
364 *bit
+= ((unsigned long) addr
& 7UL) << 3;
365 addr
= (void *) ((unsigned long) addr
& ~7UL);
366 #elif BITS_PER_LONG == 32
367 *bit
+= ((unsigned long) addr
& 3UL) << 3;
368 addr
= (void *) ((unsigned long) addr
& ~3UL);
370 #error "how many bits you are?!"
375 static inline int mb_test_bit(int bit
, void *addr
)
378 * ext4_test_bit on architecture like powerpc
379 * needs unsigned long aligned address
381 addr
= mb_correct_addr_and_bit(&bit
, addr
);
382 return ext4_test_bit(bit
, addr
);
385 static inline void mb_set_bit(int bit
, void *addr
)
387 addr
= mb_correct_addr_and_bit(&bit
, addr
);
388 ext4_set_bit(bit
, addr
);
391 static inline void mb_clear_bit(int bit
, void *addr
)
393 addr
= mb_correct_addr_and_bit(&bit
, addr
);
394 ext4_clear_bit(bit
, addr
);
397 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
399 int fix
= 0, ret
, tmpmax
;
400 addr
= mb_correct_addr_and_bit(&fix
, addr
);
404 ret
= ext4_find_next_zero_bit(addr
, tmpmax
, start
) - fix
;
410 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
412 int fix
= 0, ret
, tmpmax
;
413 addr
= mb_correct_addr_and_bit(&fix
, addr
);
417 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
423 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
427 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
430 if (order
> e4b
->bd_blkbits
+ 1) {
435 /* at order 0 we see each particular block */
437 *max
= 1 << (e4b
->bd_blkbits
+ 3);
438 return EXT4_MB_BITMAP(e4b
);
441 bb
= EXT4_MB_BUDDY(e4b
) + EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
442 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
448 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
449 int first
, int count
)
452 struct super_block
*sb
= e4b
->bd_sb
;
454 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
456 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
457 for (i
= 0; i
< count
; i
++) {
458 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
459 ext4_fsblk_t blocknr
;
461 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
462 blocknr
+= first
+ i
;
463 ext4_grp_locked_error(sb
, e4b
->bd_group
,
464 inode
? inode
->i_ino
: 0,
466 "freeing block already freed "
470 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
474 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
478 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
480 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
481 for (i
= 0; i
< count
; i
++) {
482 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
483 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
487 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
489 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
490 unsigned char *b1
, *b2
;
492 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
493 b2
= (unsigned char *) bitmap
;
494 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
495 if (b1
[i
] != b2
[i
]) {
496 printk(KERN_ERR
"corruption in group %u "
497 "at byte %u(%u): %x in copy != %x "
498 "on disk/prealloc\n",
499 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
507 static inline void mb_free_blocks_double(struct inode
*inode
,
508 struct ext4_buddy
*e4b
, int first
, int count
)
512 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
513 int first
, int count
)
517 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
523 #ifdef AGGRESSIVE_CHECK
525 #define MB_CHECK_ASSERT(assert) \
529 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
530 function, file, line, # assert); \
535 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
536 const char *function
, int line
)
538 struct super_block
*sb
= e4b
->bd_sb
;
539 int order
= e4b
->bd_blkbits
+ 1;
546 struct ext4_group_info
*grp
;
549 struct list_head
*cur
;
554 static int mb_check_counter
;
555 if (mb_check_counter
++ % 100 != 0)
560 buddy
= mb_find_buddy(e4b
, order
, &max
);
561 MB_CHECK_ASSERT(buddy
);
562 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
563 MB_CHECK_ASSERT(buddy2
);
564 MB_CHECK_ASSERT(buddy
!= buddy2
);
565 MB_CHECK_ASSERT(max
* 2 == max2
);
568 for (i
= 0; i
< max
; i
++) {
570 if (mb_test_bit(i
, buddy
)) {
571 /* only single bit in buddy2 may be 1 */
572 if (!mb_test_bit(i
<< 1, buddy2
)) {
574 mb_test_bit((i
<<1)+1, buddy2
));
575 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
577 mb_test_bit(i
<< 1, buddy2
));
582 /* both bits in buddy2 must be 0 */
583 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
584 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
586 for (j
= 0; j
< (1 << order
); j
++) {
587 k
= (i
* (1 << order
)) + j
;
589 !mb_test_bit(k
, EXT4_MB_BITMAP(e4b
)));
593 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
598 buddy
= mb_find_buddy(e4b
, 0, &max
);
599 for (i
= 0; i
< max
; i
++) {
600 if (!mb_test_bit(i
, buddy
)) {
601 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
609 /* check used bits only */
610 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
611 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
613 MB_CHECK_ASSERT(k
< max2
);
614 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
617 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
618 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
620 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
621 list_for_each(cur
, &grp
->bb_prealloc_list
) {
622 ext4_group_t groupnr
;
623 struct ext4_prealloc_space
*pa
;
624 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
625 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
626 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
627 for (i
= 0; i
< pa
->pa_len
; i
++)
628 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
632 #undef MB_CHECK_ASSERT
633 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
634 __FILE__, __func__, __LINE__)
636 #define mb_check_buddy(e4b)
640 * Divide blocks started from @first with length @len into
641 * smaller chunks with power of 2 blocks.
642 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
643 * then increase bb_counters[] for corresponded chunk size.
645 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
646 void *buddy
, ext4_grpblk_t first
, ext4_grpblk_t len
,
647 struct ext4_group_info
*grp
)
649 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
653 unsigned short border
;
655 BUG_ON(len
> EXT4_BLOCKS_PER_GROUP(sb
));
657 border
= 2 << sb
->s_blocksize_bits
;
660 /* find how many blocks can be covered since this position */
661 max
= ffs(first
| border
) - 1;
663 /* find how many blocks of power 2 we need to mark */
670 /* mark multiblock chunks only */
671 grp
->bb_counters
[min
]++;
673 mb_clear_bit(first
>> min
,
674 buddy
+ sbi
->s_mb_offsets
[min
]);
682 * Cache the order of the largest free extent we have available in this block
686 mb_set_largest_free_order(struct super_block
*sb
, struct ext4_group_info
*grp
)
691 grp
->bb_largest_free_order
= -1; /* uninit */
693 bits
= sb
->s_blocksize_bits
+ 1;
694 for (i
= bits
; i
>= 0; i
--) {
695 if (grp
->bb_counters
[i
] > 0) {
696 grp
->bb_largest_free_order
= i
;
702 static noinline_for_stack
703 void ext4_mb_generate_buddy(struct super_block
*sb
,
704 void *buddy
, void *bitmap
, ext4_group_t group
)
706 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
707 ext4_grpblk_t max
= EXT4_BLOCKS_PER_GROUP(sb
);
712 unsigned fragments
= 0;
713 unsigned long long period
= get_cycles();
715 /* initialize buddy from bitmap which is aggregation
716 * of on-disk bitmap and preallocations */
717 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
718 grp
->bb_first_free
= i
;
722 i
= mb_find_next_bit(bitmap
, max
, i
);
726 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
728 grp
->bb_counters
[0]++;
730 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
732 grp
->bb_fragments
= fragments
;
734 if (free
!= grp
->bb_free
) {
735 ext4_grp_locked_error(sb
, group
, 0, 0,
736 "%u blocks in bitmap, %u in gd",
739 * If we intent to continue, we consider group descritor
740 * corrupt and update bb_free using bitmap value
744 mb_set_largest_free_order(sb
, grp
);
746 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
748 period
= get_cycles() - period
;
749 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
750 EXT4_SB(sb
)->s_mb_buddies_generated
++;
751 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
752 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
755 /* The buddy information is attached the buddy cache inode
756 * for convenience. The information regarding each group
757 * is loaded via ext4_mb_load_buddy. The information involve
758 * block bitmap and buddy information. The information are
759 * stored in the inode as
762 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
765 * one block each for bitmap and buddy information.
766 * So for each group we take up 2 blocks. A page can
767 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
768 * So it can have information regarding groups_per_page which
769 * is blocks_per_page/2
771 * Locking note: This routine takes the block group lock of all groups
772 * for this page; do not hold this lock when calling this routine!
775 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
777 ext4_group_t ngroups
;
783 ext4_group_t first_group
;
785 struct super_block
*sb
;
786 struct buffer_head
*bhs
;
787 struct buffer_head
**bh
;
791 struct ext4_group_info
*grinfo
;
793 mb_debug(1, "init page %lu\n", page
->index
);
795 inode
= page
->mapping
->host
;
797 ngroups
= ext4_get_groups_count(sb
);
798 blocksize
= 1 << inode
->i_blkbits
;
799 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
801 groups_per_page
= blocks_per_page
>> 1;
802 if (groups_per_page
== 0)
805 /* allocate buffer_heads to read bitmaps */
806 if (groups_per_page
> 1) {
808 i
= sizeof(struct buffer_head
*) * groups_per_page
;
809 bh
= kzalloc(i
, GFP_NOFS
);
815 first_group
= page
->index
* blocks_per_page
/ 2;
817 /* read all groups the page covers into the cache */
818 for (i
= 0; i
< groups_per_page
; i
++) {
819 struct ext4_group_desc
*desc
;
821 if (first_group
+ i
>= ngroups
)
824 grinfo
= ext4_get_group_info(sb
, first_group
+ i
);
826 * If page is uptodate then we came here after online resize
827 * which added some new uninitialized group info structs, so
828 * we must skip all initialized uptodate buddies on the page,
829 * which may be currently in use by an allocating task.
831 if (PageUptodate(page
) && !EXT4_MB_GRP_NEED_INIT(grinfo
)) {
837 desc
= ext4_get_group_desc(sb
, first_group
+ i
, NULL
);
842 bh
[i
] = sb_getblk(sb
, ext4_block_bitmap(sb
, desc
));
846 if (bitmap_uptodate(bh
[i
]))
850 if (bitmap_uptodate(bh
[i
])) {
851 unlock_buffer(bh
[i
]);
854 ext4_lock_group(sb
, first_group
+ i
);
855 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
856 ext4_init_block_bitmap(sb
, bh
[i
],
857 first_group
+ i
, desc
);
858 set_bitmap_uptodate(bh
[i
]);
859 set_buffer_uptodate(bh
[i
]);
860 ext4_unlock_group(sb
, first_group
+ i
);
861 unlock_buffer(bh
[i
]);
864 ext4_unlock_group(sb
, first_group
+ i
);
865 if (buffer_uptodate(bh
[i
])) {
867 * if not uninit if bh is uptodate,
868 * bitmap is also uptodate
870 set_bitmap_uptodate(bh
[i
]);
871 unlock_buffer(bh
[i
]);
876 * submit the buffer_head for read. We can
877 * safely mark the bitmap as uptodate now.
878 * We do it here so the bitmap uptodate bit
879 * get set with buffer lock held.
881 set_bitmap_uptodate(bh
[i
]);
882 bh
[i
]->b_end_io
= end_buffer_read_sync
;
883 submit_bh(READ
, bh
[i
]);
884 mb_debug(1, "read bitmap for group %u\n", first_group
+ i
);
887 /* wait for I/O completion */
888 for (i
= 0; i
< groups_per_page
; i
++)
890 wait_on_buffer(bh
[i
]);
893 for (i
= 0; i
< groups_per_page
; i
++)
894 if (bh
[i
] && !buffer_uptodate(bh
[i
]))
898 first_block
= page
->index
* blocks_per_page
;
899 for (i
= 0; i
< blocks_per_page
; i
++) {
902 group
= (first_block
+ i
) >> 1;
903 if (group
>= ngroups
)
906 if (!bh
[group
- first_group
])
907 /* skip initialized uptodate buddy */
911 * data carry information regarding this
912 * particular group in the format specified
916 data
= page_address(page
) + (i
* blocksize
);
917 bitmap
= bh
[group
- first_group
]->b_data
;
920 * We place the buddy block and bitmap block
923 if ((first_block
+ i
) & 1) {
924 /* this is block of buddy */
925 BUG_ON(incore
== NULL
);
926 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
927 group
, page
->index
, i
* blocksize
);
928 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
929 grinfo
= ext4_get_group_info(sb
, group
);
930 grinfo
->bb_fragments
= 0;
931 memset(grinfo
->bb_counters
, 0,
932 sizeof(*grinfo
->bb_counters
) *
933 (sb
->s_blocksize_bits
+2));
935 * incore got set to the group block bitmap below
937 ext4_lock_group(sb
, group
);
939 memset(data
, 0xff, blocksize
);
940 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
941 ext4_unlock_group(sb
, group
);
944 /* this is block of bitmap */
945 BUG_ON(incore
!= NULL
);
946 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
947 group
, page
->index
, i
* blocksize
);
948 trace_ext4_mb_bitmap_load(sb
, group
);
950 /* see comments in ext4_mb_put_pa() */
951 ext4_lock_group(sb
, group
);
952 memcpy(data
, bitmap
, blocksize
);
954 /* mark all preallocated blks used in in-core bitmap */
955 ext4_mb_generate_from_pa(sb
, data
, group
);
956 ext4_mb_generate_from_freelist(sb
, data
, group
);
957 ext4_unlock_group(sb
, group
);
959 /* set incore so that the buddy information can be
960 * generated using this
965 SetPageUptodate(page
);
969 for (i
= 0; i
< groups_per_page
; i
++)
978 * Lock the buddy and bitmap pages. This make sure other parallel init_group
979 * on the same buddy page doesn't happen whild holding the buddy page lock.
980 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
981 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
983 static int ext4_mb_get_buddy_page_lock(struct super_block
*sb
,
984 ext4_group_t group
, struct ext4_buddy
*e4b
)
986 struct inode
*inode
= EXT4_SB(sb
)->s_buddy_cache
;
987 int block
, pnum
, poff
;
991 e4b
->bd_buddy_page
= NULL
;
992 e4b
->bd_bitmap_page
= NULL
;
994 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
996 * the buddy cache inode stores the block bitmap
997 * and buddy information in consecutive blocks.
998 * So for each group we need two blocks.
1001 pnum
= block
/ blocks_per_page
;
1002 poff
= block
% blocks_per_page
;
1003 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1006 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1007 e4b
->bd_bitmap_page
= page
;
1008 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1010 if (blocks_per_page
>= 2) {
1011 /* buddy and bitmap are on the same page */
1016 pnum
= block
/ blocks_per_page
;
1017 poff
= block
% blocks_per_page
;
1018 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1021 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1022 e4b
->bd_buddy_page
= page
;
1026 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy
*e4b
)
1028 if (e4b
->bd_bitmap_page
) {
1029 unlock_page(e4b
->bd_bitmap_page
);
1030 page_cache_release(e4b
->bd_bitmap_page
);
1032 if (e4b
->bd_buddy_page
) {
1033 unlock_page(e4b
->bd_buddy_page
);
1034 page_cache_release(e4b
->bd_buddy_page
);
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
)
1047 struct ext4_group_info
*this_grp
;
1048 struct ext4_buddy e4b
;
1052 mb_debug(1, "init group %u\n", group
);
1053 this_grp
= ext4_get_group_info(sb
, group
);
1055 * This ensures that we don't reinit the buddy cache
1056 * page which map to the group from which we are already
1057 * allocating. If we are looking at the buddy cache we would
1058 * have taken a reference using ext4_mb_load_buddy and that
1059 * would have pinned buddy page to page cache.
1061 ret
= ext4_mb_get_buddy_page_lock(sb
, group
, &e4b
);
1062 if (ret
|| !EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1064 * somebody initialized the group
1065 * return without doing anything
1070 page
= e4b
.bd_bitmap_page
;
1071 ret
= ext4_mb_init_cache(page
, NULL
);
1074 if (!PageUptodate(page
)) {
1078 mark_page_accessed(page
);
1080 if (e4b
.bd_buddy_page
== NULL
) {
1082 * If both the bitmap and buddy are in
1083 * the same page we don't need to force
1089 /* init buddy cache */
1090 page
= e4b
.bd_buddy_page
;
1091 ret
= ext4_mb_init_cache(page
, e4b
.bd_bitmap
);
1094 if (!PageUptodate(page
)) {
1098 mark_page_accessed(page
);
1100 ext4_mb_put_buddy_page_lock(&e4b
);
1105 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1106 * block group lock of all groups for this page; do not hold the BG lock when
1107 * calling this routine!
1109 static noinline_for_stack
int
1110 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1111 struct ext4_buddy
*e4b
)
1113 int blocks_per_page
;
1119 struct ext4_group_info
*grp
;
1120 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1121 struct inode
*inode
= sbi
->s_buddy_cache
;
1123 mb_debug(1, "load group %u\n", group
);
1125 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1126 grp
= ext4_get_group_info(sb
, group
);
1128 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1129 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
1131 e4b
->bd_group
= group
;
1132 e4b
->bd_buddy_page
= NULL
;
1133 e4b
->bd_bitmap_page
= NULL
;
1135 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1137 * we need full data about the group
1138 * to make a good selection
1140 ret
= ext4_mb_init_group(sb
, group
);
1146 * the buddy cache inode stores the block bitmap
1147 * and buddy information in consecutive blocks.
1148 * So for each group we need two blocks.
1151 pnum
= block
/ blocks_per_page
;
1152 poff
= block
% blocks_per_page
;
1154 /* we could use find_or_create_page(), but it locks page
1155 * what we'd like to avoid in fast path ... */
1156 page
= find_get_page(inode
->i_mapping
, pnum
);
1157 if (page
== NULL
|| !PageUptodate(page
)) {
1160 * drop the page reference and try
1161 * to get the page with lock. If we
1162 * are not uptodate that implies
1163 * somebody just created the page but
1164 * is yet to initialize the same. So
1165 * wait for it to initialize.
1167 page_cache_release(page
);
1168 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1170 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1171 if (!PageUptodate(page
)) {
1172 ret
= ext4_mb_init_cache(page
, NULL
);
1177 mb_cmp_bitmaps(e4b
, page_address(page
) +
1178 (poff
* sb
->s_blocksize
));
1183 if (page
== NULL
|| !PageUptodate(page
)) {
1187 e4b
->bd_bitmap_page
= page
;
1188 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1189 mark_page_accessed(page
);
1192 pnum
= block
/ blocks_per_page
;
1193 poff
= block
% blocks_per_page
;
1195 page
= find_get_page(inode
->i_mapping
, pnum
);
1196 if (page
== NULL
|| !PageUptodate(page
)) {
1198 page_cache_release(page
);
1199 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1201 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1202 if (!PageUptodate(page
)) {
1203 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1212 if (page
== NULL
|| !PageUptodate(page
)) {
1216 e4b
->bd_buddy_page
= page
;
1217 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1218 mark_page_accessed(page
);
1220 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1221 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1227 page_cache_release(page
);
1228 if (e4b
->bd_bitmap_page
)
1229 page_cache_release(e4b
->bd_bitmap_page
);
1230 if (e4b
->bd_buddy_page
)
1231 page_cache_release(e4b
->bd_buddy_page
);
1232 e4b
->bd_buddy
= NULL
;
1233 e4b
->bd_bitmap
= NULL
;
1237 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1239 if (e4b
->bd_bitmap_page
)
1240 page_cache_release(e4b
->bd_bitmap_page
);
1241 if (e4b
->bd_buddy_page
)
1242 page_cache_release(e4b
->bd_buddy_page
);
1246 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1251 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1252 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1254 bb
= EXT4_MB_BUDDY(e4b
);
1255 while (order
<= e4b
->bd_blkbits
+ 1) {
1257 if (!mb_test_bit(block
, bb
)) {
1258 /* this block is part of buddy of order 'order' */
1261 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1267 static void mb_clear_bits(void *bm
, int cur
, int len
)
1273 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1274 /* fast path: clear whole word at once */
1275 addr
= bm
+ (cur
>> 3);
1280 mb_clear_bit(cur
, bm
);
1285 static void mb_set_bits(void *bm
, int cur
, int len
)
1291 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1292 /* fast path: set whole word at once */
1293 addr
= bm
+ (cur
>> 3);
1298 mb_set_bit(cur
, bm
);
1303 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1304 int first
, int count
)
1311 struct super_block
*sb
= e4b
->bd_sb
;
1313 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1314 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1315 mb_check_buddy(e4b
);
1316 mb_free_blocks_double(inode
, e4b
, first
, count
);
1318 e4b
->bd_info
->bb_free
+= count
;
1319 if (first
< e4b
->bd_info
->bb_first_free
)
1320 e4b
->bd_info
->bb_first_free
= first
;
1322 /* let's maintain fragments counter */
1324 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1325 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1326 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1328 e4b
->bd_info
->bb_fragments
--;
1329 else if (!block
&& !max
)
1330 e4b
->bd_info
->bb_fragments
++;
1332 /* let's maintain buddy itself */
1333 while (count
-- > 0) {
1337 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1338 ext4_fsblk_t blocknr
;
1340 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1342 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1343 inode
? inode
->i_ino
: 0,
1345 "freeing already freed block "
1348 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1349 e4b
->bd_info
->bb_counters
[order
]++;
1351 /* start of the buddy */
1352 buddy
= mb_find_buddy(e4b
, order
, &max
);
1356 if (mb_test_bit(block
, buddy
) ||
1357 mb_test_bit(block
+ 1, buddy
))
1360 /* both the buddies are free, try to coalesce them */
1361 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1367 /* for special purposes, we don't set
1368 * free bits in bitmap */
1369 mb_set_bit(block
, buddy
);
1370 mb_set_bit(block
+ 1, buddy
);
1372 e4b
->bd_info
->bb_counters
[order
]--;
1373 e4b
->bd_info
->bb_counters
[order
]--;
1377 e4b
->bd_info
->bb_counters
[order
]++;
1379 mb_clear_bit(block
, buddy2
);
1383 mb_set_largest_free_order(sb
, e4b
->bd_info
);
1384 mb_check_buddy(e4b
);
1387 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1388 int needed
, struct ext4_free_extent
*ex
)
1395 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1398 buddy
= mb_find_buddy(e4b
, order
, &max
);
1399 BUG_ON(buddy
== NULL
);
1400 BUG_ON(block
>= max
);
1401 if (mb_test_bit(block
, buddy
)) {
1408 /* FIXME dorp order completely ? */
1409 if (likely(order
== 0)) {
1410 /* find actual order */
1411 order
= mb_find_order_for_block(e4b
, block
);
1412 block
= block
>> order
;
1415 ex
->fe_len
= 1 << order
;
1416 ex
->fe_start
= block
<< order
;
1417 ex
->fe_group
= e4b
->bd_group
;
1419 /* calc difference from given start */
1420 next
= next
- ex
->fe_start
;
1422 ex
->fe_start
+= next
;
1424 while (needed
> ex
->fe_len
&&
1425 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1427 if (block
+ 1 >= max
)
1430 next
= (block
+ 1) * (1 << order
);
1431 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1434 ord
= mb_find_order_for_block(e4b
, next
);
1437 block
= next
>> order
;
1438 ex
->fe_len
+= 1 << order
;
1441 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1445 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1451 int start
= ex
->fe_start
;
1452 int len
= ex
->fe_len
;
1457 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1458 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1459 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1460 mb_check_buddy(e4b
);
1461 mb_mark_used_double(e4b
, start
, len
);
1463 e4b
->bd_info
->bb_free
-= len
;
1464 if (e4b
->bd_info
->bb_first_free
== start
)
1465 e4b
->bd_info
->bb_first_free
+= len
;
1467 /* let's maintain fragments counter */
1469 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1470 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1471 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1473 e4b
->bd_info
->bb_fragments
++;
1474 else if (!mlen
&& !max
)
1475 e4b
->bd_info
->bb_fragments
--;
1477 /* let's maintain buddy itself */
1479 ord
= mb_find_order_for_block(e4b
, start
);
1481 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1482 /* the whole chunk may be allocated at once! */
1484 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1485 BUG_ON((start
>> ord
) >= max
);
1486 mb_set_bit(start
>> ord
, buddy
);
1487 e4b
->bd_info
->bb_counters
[ord
]--;
1494 /* store for history */
1496 ret
= len
| (ord
<< 16);
1498 /* we have to split large buddy */
1500 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1501 mb_set_bit(start
>> ord
, buddy
);
1502 e4b
->bd_info
->bb_counters
[ord
]--;
1505 cur
= (start
>> ord
) & ~1U;
1506 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1507 mb_clear_bit(cur
, buddy
);
1508 mb_clear_bit(cur
+ 1, buddy
);
1509 e4b
->bd_info
->bb_counters
[ord
]++;
1510 e4b
->bd_info
->bb_counters
[ord
]++;
1512 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1514 mb_set_bits(EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1515 mb_check_buddy(e4b
);
1521 * Must be called under group lock!
1523 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1524 struct ext4_buddy
*e4b
)
1526 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1529 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1530 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1532 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1533 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1534 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1536 /* preallocation can change ac_b_ex, thus we store actually
1537 * allocated blocks for history */
1538 ac
->ac_f_ex
= ac
->ac_b_ex
;
1540 ac
->ac_status
= AC_STATUS_FOUND
;
1541 ac
->ac_tail
= ret
& 0xffff;
1542 ac
->ac_buddy
= ret
>> 16;
1545 * take the page reference. We want the page to be pinned
1546 * so that we don't get a ext4_mb_init_cache_call for this
1547 * group until we update the bitmap. That would mean we
1548 * double allocate blocks. The reference is dropped
1549 * in ext4_mb_release_context
1551 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1552 get_page(ac
->ac_bitmap_page
);
1553 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1554 get_page(ac
->ac_buddy_page
);
1555 /* store last allocated for subsequent stream allocation */
1556 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1557 spin_lock(&sbi
->s_md_lock
);
1558 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1559 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1560 spin_unlock(&sbi
->s_md_lock
);
1565 * regular allocator, for general purposes allocation
1568 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1569 struct ext4_buddy
*e4b
,
1572 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1573 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1574 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1575 struct ext4_free_extent ex
;
1578 if (ac
->ac_status
== AC_STATUS_FOUND
)
1581 * We don't want to scan for a whole year
1583 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1584 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1585 ac
->ac_status
= AC_STATUS_BREAK
;
1590 * Haven't found good chunk so far, let's continue
1592 if (bex
->fe_len
< gex
->fe_len
)
1595 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1596 && bex
->fe_group
== e4b
->bd_group
) {
1597 /* recheck chunk's availability - we don't know
1598 * when it was found (within this lock-unlock
1600 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1601 if (max
>= gex
->fe_len
) {
1602 ext4_mb_use_best_found(ac
, e4b
);
1609 * The routine checks whether found extent is good enough. If it is,
1610 * then the extent gets marked used and flag is set to the context
1611 * to stop scanning. Otherwise, the extent is compared with the
1612 * previous found extent and if new one is better, then it's stored
1613 * in the context. Later, the best found extent will be used, if
1614 * mballoc can't find good enough extent.
1616 * FIXME: real allocation policy is to be designed yet!
1618 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1619 struct ext4_free_extent
*ex
,
1620 struct ext4_buddy
*e4b
)
1622 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1623 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1625 BUG_ON(ex
->fe_len
<= 0);
1626 BUG_ON(ex
->fe_len
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1627 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1628 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1633 * The special case - take what you catch first
1635 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1637 ext4_mb_use_best_found(ac
, e4b
);
1642 * Let's check whether the chuck is good enough
1644 if (ex
->fe_len
== gex
->fe_len
) {
1646 ext4_mb_use_best_found(ac
, e4b
);
1651 * If this is first found extent, just store it in the context
1653 if (bex
->fe_len
== 0) {
1659 * If new found extent is better, store it in the context
1661 if (bex
->fe_len
< gex
->fe_len
) {
1662 /* if the request isn't satisfied, any found extent
1663 * larger than previous best one is better */
1664 if (ex
->fe_len
> bex
->fe_len
)
1666 } else if (ex
->fe_len
> gex
->fe_len
) {
1667 /* if the request is satisfied, then we try to find
1668 * an extent that still satisfy the request, but is
1669 * smaller than previous one */
1670 if (ex
->fe_len
< bex
->fe_len
)
1674 ext4_mb_check_limits(ac
, e4b
, 0);
1677 static noinline_for_stack
1678 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1679 struct ext4_buddy
*e4b
)
1681 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1682 ext4_group_t group
= ex
.fe_group
;
1686 BUG_ON(ex
.fe_len
<= 0);
1687 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1691 ext4_lock_group(ac
->ac_sb
, group
);
1692 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1696 ext4_mb_use_best_found(ac
, e4b
);
1699 ext4_unlock_group(ac
->ac_sb
, group
);
1700 ext4_mb_unload_buddy(e4b
);
1705 static noinline_for_stack
1706 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1707 struct ext4_buddy
*e4b
)
1709 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1712 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1713 struct ext4_free_extent ex
;
1715 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1718 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1722 ext4_lock_group(ac
->ac_sb
, group
);
1723 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1724 ac
->ac_g_ex
.fe_len
, &ex
);
1726 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1729 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1731 /* use do_div to get remainder (would be 64-bit modulo) */
1732 if (do_div(start
, sbi
->s_stripe
) == 0) {
1735 ext4_mb_use_best_found(ac
, e4b
);
1737 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1738 BUG_ON(ex
.fe_len
<= 0);
1739 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1740 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1743 ext4_mb_use_best_found(ac
, e4b
);
1744 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1745 /* Sometimes, caller may want to merge even small
1746 * number of blocks to an existing extent */
1747 BUG_ON(ex
.fe_len
<= 0);
1748 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1749 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1752 ext4_mb_use_best_found(ac
, e4b
);
1754 ext4_unlock_group(ac
->ac_sb
, group
);
1755 ext4_mb_unload_buddy(e4b
);
1761 * The routine scans buddy structures (not bitmap!) from given order
1762 * to max order and tries to find big enough chunk to satisfy the req
1764 static noinline_for_stack
1765 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1766 struct ext4_buddy
*e4b
)
1768 struct super_block
*sb
= ac
->ac_sb
;
1769 struct ext4_group_info
*grp
= e4b
->bd_info
;
1775 BUG_ON(ac
->ac_2order
<= 0);
1776 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1777 if (grp
->bb_counters
[i
] == 0)
1780 buddy
= mb_find_buddy(e4b
, i
, &max
);
1781 BUG_ON(buddy
== NULL
);
1783 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1788 ac
->ac_b_ex
.fe_len
= 1 << i
;
1789 ac
->ac_b_ex
.fe_start
= k
<< i
;
1790 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1792 ext4_mb_use_best_found(ac
, e4b
);
1794 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1796 if (EXT4_SB(sb
)->s_mb_stats
)
1797 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1804 * The routine scans the group and measures all found extents.
1805 * In order to optimize scanning, caller must pass number of
1806 * free blocks in the group, so the routine can know upper limit.
1808 static noinline_for_stack
1809 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1810 struct ext4_buddy
*e4b
)
1812 struct super_block
*sb
= ac
->ac_sb
;
1813 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1814 struct ext4_free_extent ex
;
1818 free
= e4b
->bd_info
->bb_free
;
1821 i
= e4b
->bd_info
->bb_first_free
;
1823 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1824 i
= mb_find_next_zero_bit(bitmap
,
1825 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1826 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1828 * IF we have corrupt bitmap, we won't find any
1829 * free blocks even though group info says we
1830 * we have free blocks
1832 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1833 "%d free blocks as per "
1834 "group info. But bitmap says 0",
1839 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1840 BUG_ON(ex
.fe_len
<= 0);
1841 if (free
< ex
.fe_len
) {
1842 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1843 "%d free blocks as per "
1844 "group info. But got %d blocks",
1847 * The number of free blocks differs. This mostly
1848 * indicate that the bitmap is corrupt. So exit
1849 * without claiming the space.
1854 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1860 ext4_mb_check_limits(ac
, e4b
, 1);
1864 * This is a special case for storages like raid5
1865 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1867 static noinline_for_stack
1868 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1869 struct ext4_buddy
*e4b
)
1871 struct super_block
*sb
= ac
->ac_sb
;
1872 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1873 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1874 struct ext4_free_extent ex
;
1875 ext4_fsblk_t first_group_block
;
1880 BUG_ON(sbi
->s_stripe
== 0);
1882 /* find first stripe-aligned block in group */
1883 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1885 a
= first_group_block
+ sbi
->s_stripe
- 1;
1886 do_div(a
, sbi
->s_stripe
);
1887 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1889 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1890 if (!mb_test_bit(i
, bitmap
)) {
1891 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1892 if (max
>= sbi
->s_stripe
) {
1895 ext4_mb_use_best_found(ac
, e4b
);
1903 /* This is now called BEFORE we load the buddy bitmap. */
1904 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1905 ext4_group_t group
, int cr
)
1907 unsigned free
, fragments
;
1908 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
1909 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1911 BUG_ON(cr
< 0 || cr
>= 4);
1913 /* We only do this if the grp has never been initialized */
1914 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1915 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
);
1920 free
= grp
->bb_free
;
1921 fragments
= grp
->bb_fragments
;
1929 BUG_ON(ac
->ac_2order
== 0);
1931 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
1934 /* Avoid using the first bg of a flexgroup for data files */
1935 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
1936 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
1937 ((group
% flex_size
) == 0))
1942 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1946 if (free
>= ac
->ac_g_ex
.fe_len
)
1958 static noinline_for_stack
int
1959 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
1961 ext4_group_t ngroups
, group
, i
;
1964 struct ext4_sb_info
*sbi
;
1965 struct super_block
*sb
;
1966 struct ext4_buddy e4b
;
1970 ngroups
= ext4_get_groups_count(sb
);
1971 /* non-extent files are limited to low blocks/groups */
1972 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
1973 ngroups
= sbi
->s_blockfile_groups
;
1975 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1977 /* first, try the goal */
1978 err
= ext4_mb_find_by_goal(ac
, &e4b
);
1979 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
1982 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
1986 * ac->ac2_order is set only if the fe_len is a power of 2
1987 * if ac2_order is set we also set criteria to 0 so that we
1988 * try exact allocation using buddy.
1990 i
= fls(ac
->ac_g_ex
.fe_len
);
1993 * We search using buddy data only if the order of the request
1994 * is greater than equal to the sbi_s_mb_order2_reqs
1995 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1997 if (i
>= sbi
->s_mb_order2_reqs
) {
1999 * This should tell if fe_len is exactly power of 2
2001 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2002 ac
->ac_2order
= i
- 1;
2005 /* if stream allocation is enabled, use global goal */
2006 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2007 /* TBD: may be hot point */
2008 spin_lock(&sbi
->s_md_lock
);
2009 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2010 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2011 spin_unlock(&sbi
->s_md_lock
);
2014 /* Let's just scan groups to find more-less suitable blocks */
2015 cr
= ac
->ac_2order
? 0 : 1;
2017 * cr == 0 try to get exact allocation,
2018 * cr == 3 try to get anything
2021 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2022 ac
->ac_criteria
= cr
;
2024 * searching for the right group start
2025 * from the goal value specified
2027 group
= ac
->ac_g_ex
.fe_group
;
2029 for (i
= 0; i
< ngroups
; group
++, i
++) {
2030 if (group
== ngroups
)
2033 /* This now checks without needing the buddy page */
2034 if (!ext4_mb_good_group(ac
, group
, cr
))
2037 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2041 ext4_lock_group(sb
, group
);
2044 * We need to check again after locking the
2047 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2048 ext4_unlock_group(sb
, group
);
2049 ext4_mb_unload_buddy(&e4b
);
2053 ac
->ac_groups_scanned
++;
2055 ext4_mb_simple_scan_group(ac
, &e4b
);
2056 else if (cr
== 1 && sbi
->s_stripe
&&
2057 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2058 ext4_mb_scan_aligned(ac
, &e4b
);
2060 ext4_mb_complex_scan_group(ac
, &e4b
);
2062 ext4_unlock_group(sb
, group
);
2063 ext4_mb_unload_buddy(&e4b
);
2065 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2070 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2071 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2073 * We've been searching too long. Let's try to allocate
2074 * the best chunk we've found so far
2077 ext4_mb_try_best_found(ac
, &e4b
);
2078 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2080 * Someone more lucky has already allocated it.
2081 * The only thing we can do is just take first
2083 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2085 ac
->ac_b_ex
.fe_group
= 0;
2086 ac
->ac_b_ex
.fe_start
= 0;
2087 ac
->ac_b_ex
.fe_len
= 0;
2088 ac
->ac_status
= AC_STATUS_CONTINUE
;
2089 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2091 atomic_inc(&sbi
->s_mb_lost_chunks
);
2099 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2101 struct super_block
*sb
= seq
->private;
2104 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2107 return (void *) ((unsigned long) group
);
2110 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2112 struct super_block
*sb
= seq
->private;
2116 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2119 return (void *) ((unsigned long) group
);
2122 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2124 struct super_block
*sb
= seq
->private;
2125 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2128 struct ext4_buddy e4b
;
2130 struct ext4_group_info info
;
2131 ext4_grpblk_t counters
[16];
2136 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2137 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2138 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2139 "group", "free", "frags", "first",
2140 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2141 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2143 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2144 sizeof(struct ext4_group_info
);
2145 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2147 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2150 ext4_lock_group(sb
, group
);
2151 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2152 ext4_unlock_group(sb
, group
);
2153 ext4_mb_unload_buddy(&e4b
);
2155 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2156 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2157 for (i
= 0; i
<= 13; i
++)
2158 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2159 sg
.info
.bb_counters
[i
] : 0);
2160 seq_printf(seq
, " ]\n");
2165 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2169 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2170 .start
= ext4_mb_seq_groups_start
,
2171 .next
= ext4_mb_seq_groups_next
,
2172 .stop
= ext4_mb_seq_groups_stop
,
2173 .show
= ext4_mb_seq_groups_show
,
2176 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2178 struct super_block
*sb
= PDE(inode
)->data
;
2181 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2183 struct seq_file
*m
= file
->private_data
;
2190 static const struct file_operations ext4_mb_seq_groups_fops
= {
2191 .owner
= THIS_MODULE
,
2192 .open
= ext4_mb_seq_groups_open
,
2194 .llseek
= seq_lseek
,
2195 .release
= seq_release
,
2198 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2200 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2201 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2207 /* Create and initialize ext4_group_info data for the given group. */
2208 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2209 struct ext4_group_desc
*desc
)
2213 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2214 struct ext4_group_info
**meta_group_info
;
2215 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2218 * First check if this group is the first of a reserved block.
2219 * If it's true, we have to allocate a new table of pointers
2220 * to ext4_group_info structures
2222 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2223 metalen
= sizeof(*meta_group_info
) <<
2224 EXT4_DESC_PER_BLOCK_BITS(sb
);
2225 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2226 if (meta_group_info
== NULL
) {
2227 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2229 goto exit_meta_group_info
;
2231 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2236 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2237 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2239 meta_group_info
[i
] = kmem_cache_alloc(cachep
, GFP_KERNEL
);
2240 if (meta_group_info
[i
] == NULL
) {
2241 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2242 goto exit_group_info
;
2244 memset(meta_group_info
[i
], 0, kmem_cache_size(cachep
));
2245 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2246 &(meta_group_info
[i
]->bb_state
));
2249 * initialize bb_free to be able to skip
2250 * empty groups without initialization
2252 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2253 meta_group_info
[i
]->bb_free
=
2254 ext4_free_blocks_after_init(sb
, group
, desc
);
2256 meta_group_info
[i
]->bb_free
=
2257 ext4_free_blks_count(sb
, desc
);
2260 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2261 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2262 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2263 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2267 struct buffer_head
*bh
;
2268 meta_group_info
[i
]->bb_bitmap
=
2269 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2270 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2271 bh
= ext4_read_block_bitmap(sb
, group
);
2273 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2282 /* If a meta_group_info table has been allocated, release it now */
2283 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2284 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2285 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] = NULL
;
2287 exit_meta_group_info
:
2289 } /* ext4_mb_add_groupinfo */
2291 static int ext4_mb_init_backend(struct super_block
*sb
)
2293 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2295 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2296 struct ext4_super_block
*es
= sbi
->s_es
;
2297 int num_meta_group_infos
;
2298 int num_meta_group_infos_max
;
2300 struct ext4_group_desc
*desc
;
2301 struct kmem_cache
*cachep
;
2303 /* This is the number of blocks used by GDT */
2304 num_meta_group_infos
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) -
2305 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2308 * This is the total number of blocks used by GDT including
2309 * the number of reserved blocks for GDT.
2310 * The s_group_info array is allocated with this value
2311 * to allow a clean online resize without a complex
2312 * manipulation of pointer.
2313 * The drawback is the unused memory when no resize
2314 * occurs but it's very low in terms of pages
2315 * (see comments below)
2316 * Need to handle this properly when META_BG resizing is allowed
2318 num_meta_group_infos_max
= num_meta_group_infos
+
2319 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2322 * array_size is the size of s_group_info array. We round it
2323 * to the next power of two because this approximation is done
2324 * internally by kmalloc so we can have some more memory
2325 * for free here (e.g. may be used for META_BG resize).
2328 while (array_size
< sizeof(*sbi
->s_group_info
) *
2329 num_meta_group_infos_max
)
2330 array_size
= array_size
<< 1;
2331 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2332 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2333 * So a two level scheme suffices for now. */
2334 sbi
->s_group_info
= kzalloc(array_size
, GFP_KERNEL
);
2335 if (sbi
->s_group_info
== NULL
) {
2336 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2339 sbi
->s_buddy_cache
= new_inode(sb
);
2340 if (sbi
->s_buddy_cache
== NULL
) {
2341 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2344 sbi
->s_buddy_cache
->i_ino
= get_next_ino();
2345 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2346 for (i
= 0; i
< ngroups
; i
++) {
2347 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2350 "EXT4-fs: can't read descriptor %u\n", i
);
2353 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2360 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2362 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2363 i
= num_meta_group_infos
;
2365 kfree(sbi
->s_group_info
[i
]);
2366 iput(sbi
->s_buddy_cache
);
2368 kfree(sbi
->s_group_info
);
2372 static void ext4_groupinfo_destroy_slabs(void)
2376 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2377 if (ext4_groupinfo_caches
[i
])
2378 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2379 ext4_groupinfo_caches
[i
] = NULL
;
2383 static int ext4_groupinfo_create_slab(size_t size
)
2385 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2387 int blocksize_bits
= order_base_2(size
);
2388 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2389 struct kmem_cache
*cachep
;
2391 if (cache_index
>= NR_GRPINFO_CACHES
)
2394 if (unlikely(cache_index
< 0))
2397 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2398 if (ext4_groupinfo_caches
[cache_index
]) {
2399 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2400 return 0; /* Already created */
2403 slab_size
= offsetof(struct ext4_group_info
,
2404 bb_counters
[blocksize_bits
+ 2]);
2406 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2407 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2410 ext4_groupinfo_caches
[cache_index
] = cachep
;
2412 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2414 printk(KERN_EMERG
"EXT4: no memory for groupinfo slab cache\n");
2421 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2423 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2429 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2431 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2432 if (sbi
->s_mb_offsets
== NULL
) {
2437 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2438 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2439 if (sbi
->s_mb_maxs
== NULL
) {
2444 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2448 /* order 0 is regular bitmap */
2449 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2450 sbi
->s_mb_offsets
[0] = 0;
2454 max
= sb
->s_blocksize
<< 2;
2456 sbi
->s_mb_offsets
[i
] = offset
;
2457 sbi
->s_mb_maxs
[i
] = max
;
2458 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2461 } while (i
<= sb
->s_blocksize_bits
+ 1);
2463 /* init file for buddy data */
2464 ret
= ext4_mb_init_backend(sb
);
2469 spin_lock_init(&sbi
->s_md_lock
);
2470 spin_lock_init(&sbi
->s_bal_lock
);
2472 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2473 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2474 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2475 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2476 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2477 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2479 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2480 * to the lowest multiple of s_stripe which is bigger than
2481 * the s_mb_group_prealloc as determined above. We want
2482 * the preallocation size to be an exact multiple of the
2483 * RAID stripe size so that preallocations don't fragment
2486 if (sbi
->s_stripe
> 1) {
2487 sbi
->s_mb_group_prealloc
= roundup(
2488 sbi
->s_mb_group_prealloc
, sbi
->s_stripe
);
2491 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2492 if (sbi
->s_locality_groups
== NULL
) {
2496 for_each_possible_cpu(i
) {
2497 struct ext4_locality_group
*lg
;
2498 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2499 mutex_init(&lg
->lg_mutex
);
2500 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2501 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2502 spin_lock_init(&lg
->lg_prealloc_lock
);
2506 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2507 &ext4_mb_seq_groups_fops
, sb
);
2510 sbi
->s_journal
->j_commit_callback
= release_blocks_on_commit
;
2513 kfree(sbi
->s_mb_offsets
);
2514 kfree(sbi
->s_mb_maxs
);
2519 /* need to called with the ext4 group lock held */
2520 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2522 struct ext4_prealloc_space
*pa
;
2523 struct list_head
*cur
, *tmp
;
2526 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2527 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2528 list_del(&pa
->pa_group_list
);
2530 kmem_cache_free(ext4_pspace_cachep
, pa
);
2533 mb_debug(1, "mballoc: %u PAs left\n", count
);
2537 int ext4_mb_release(struct super_block
*sb
)
2539 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2541 int num_meta_group_infos
;
2542 struct ext4_group_info
*grinfo
;
2543 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2544 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2546 if (sbi
->s_group_info
) {
2547 for (i
= 0; i
< ngroups
; i
++) {
2548 grinfo
= ext4_get_group_info(sb
, i
);
2550 kfree(grinfo
->bb_bitmap
);
2552 ext4_lock_group(sb
, i
);
2553 ext4_mb_cleanup_pa(grinfo
);
2554 ext4_unlock_group(sb
, i
);
2555 kmem_cache_free(cachep
, grinfo
);
2557 num_meta_group_infos
= (ngroups
+
2558 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2559 EXT4_DESC_PER_BLOCK_BITS(sb
);
2560 for (i
= 0; i
< num_meta_group_infos
; i
++)
2561 kfree(sbi
->s_group_info
[i
]);
2562 kfree(sbi
->s_group_info
);
2564 kfree(sbi
->s_mb_offsets
);
2565 kfree(sbi
->s_mb_maxs
);
2566 if (sbi
->s_buddy_cache
)
2567 iput(sbi
->s_buddy_cache
);
2568 if (sbi
->s_mb_stats
) {
2570 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2571 atomic_read(&sbi
->s_bal_allocated
),
2572 atomic_read(&sbi
->s_bal_reqs
),
2573 atomic_read(&sbi
->s_bal_success
));
2575 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2576 "%u 2^N hits, %u breaks, %u lost\n",
2577 atomic_read(&sbi
->s_bal_ex_scanned
),
2578 atomic_read(&sbi
->s_bal_goals
),
2579 atomic_read(&sbi
->s_bal_2orders
),
2580 atomic_read(&sbi
->s_bal_breaks
),
2581 atomic_read(&sbi
->s_mb_lost_chunks
));
2583 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2584 sbi
->s_mb_buddies_generated
++,
2585 sbi
->s_mb_generation_time
);
2587 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2588 atomic_read(&sbi
->s_mb_preallocated
),
2589 atomic_read(&sbi
->s_mb_discarded
));
2592 free_percpu(sbi
->s_locality_groups
);
2594 remove_proc_entry("mb_groups", sbi
->s_proc
);
2599 static inline int ext4_issue_discard(struct super_block
*sb
,
2600 ext4_group_t block_group
, ext4_grpblk_t block
, int count
)
2602 ext4_fsblk_t discard_block
;
2604 discard_block
= block
+ ext4_group_first_block_no(sb
, block_group
);
2605 trace_ext4_discard_blocks(sb
,
2606 (unsigned long long) discard_block
, count
);
2607 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, 0);
2611 * This function is called by the jbd2 layer once the commit has finished,
2612 * so we know we can free the blocks that were released with that commit.
2614 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
)
2616 struct super_block
*sb
= journal
->j_private
;
2617 struct ext4_buddy e4b
;
2618 struct ext4_group_info
*db
;
2619 int err
, count
= 0, count2
= 0;
2620 struct ext4_free_data
*entry
;
2621 struct list_head
*l
, *ltmp
;
2623 list_for_each_safe(l
, ltmp
, &txn
->t_private_list
) {
2624 entry
= list_entry(l
, struct ext4_free_data
, list
);
2626 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2627 entry
->count
, entry
->group
, entry
);
2629 if (test_opt(sb
, DISCARD
))
2630 ext4_issue_discard(sb
, entry
->group
,
2631 entry
->start_blk
, entry
->count
);
2633 err
= ext4_mb_load_buddy(sb
, entry
->group
, &e4b
);
2634 /* we expect to find existing buddy because it's pinned */
2638 /* there are blocks to put in buddy to make them really free */
2639 count
+= entry
->count
;
2641 ext4_lock_group(sb
, entry
->group
);
2642 /* Take it out of per group rb tree */
2643 rb_erase(&entry
->node
, &(db
->bb_free_root
));
2644 mb_free_blocks(NULL
, &e4b
, entry
->start_blk
, entry
->count
);
2647 * Clear the trimmed flag for the group so that the next
2648 * ext4_trim_fs can trim it.
2649 * If the volume is mounted with -o discard, online discard
2650 * is supported and the free blocks will be trimmed online.
2652 if (!test_opt(sb
, DISCARD
))
2653 EXT4_MB_GRP_CLEAR_TRIMMED(db
);
2655 if (!db
->bb_free_root
.rb_node
) {
2656 /* No more items in the per group rb tree
2657 * balance refcounts from ext4_mb_free_metadata()
2659 page_cache_release(e4b
.bd_buddy_page
);
2660 page_cache_release(e4b
.bd_bitmap_page
);
2662 ext4_unlock_group(sb
, entry
->group
);
2663 kmem_cache_free(ext4_free_ext_cachep
, entry
);
2664 ext4_mb_unload_buddy(&e4b
);
2667 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2670 #ifdef CONFIG_EXT4_DEBUG
2671 u8 mb_enable_debug __read_mostly
;
2673 static struct dentry
*debugfs_dir
;
2674 static struct dentry
*debugfs_debug
;
2676 static void __init
ext4_create_debugfs_entry(void)
2678 debugfs_dir
= debugfs_create_dir("ext4", NULL
);
2680 debugfs_debug
= debugfs_create_u8("mballoc-debug",
2686 static void ext4_remove_debugfs_entry(void)
2688 debugfs_remove(debugfs_debug
);
2689 debugfs_remove(debugfs_dir
);
2694 static void __init
ext4_create_debugfs_entry(void)
2698 static void ext4_remove_debugfs_entry(void)
2704 int __init
ext4_init_mballoc(void)
2706 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2707 SLAB_RECLAIM_ACCOUNT
);
2708 if (ext4_pspace_cachep
== NULL
)
2711 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2712 SLAB_RECLAIM_ACCOUNT
);
2713 if (ext4_ac_cachep
== NULL
) {
2714 kmem_cache_destroy(ext4_pspace_cachep
);
2718 ext4_free_ext_cachep
= KMEM_CACHE(ext4_free_data
,
2719 SLAB_RECLAIM_ACCOUNT
);
2720 if (ext4_free_ext_cachep
== NULL
) {
2721 kmem_cache_destroy(ext4_pspace_cachep
);
2722 kmem_cache_destroy(ext4_ac_cachep
);
2725 ext4_create_debugfs_entry();
2729 void ext4_exit_mballoc(void)
2732 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2733 * before destroying the slab cache.
2736 kmem_cache_destroy(ext4_pspace_cachep
);
2737 kmem_cache_destroy(ext4_ac_cachep
);
2738 kmem_cache_destroy(ext4_free_ext_cachep
);
2739 ext4_groupinfo_destroy_slabs();
2740 ext4_remove_debugfs_entry();
2745 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2746 * Returns 0 if success or error code
2748 static noinline_for_stack
int
2749 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2750 handle_t
*handle
, unsigned int reserv_blks
)
2752 struct buffer_head
*bitmap_bh
= NULL
;
2753 struct ext4_group_desc
*gdp
;
2754 struct buffer_head
*gdp_bh
;
2755 struct ext4_sb_info
*sbi
;
2756 struct super_block
*sb
;
2760 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2761 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2767 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2771 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2776 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2780 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2781 ext4_free_blks_count(sb
, gdp
));
2783 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2787 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
2789 len
= ac
->ac_b_ex
.fe_len
;
2790 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2791 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2792 "fs metadata\n", block
, block
+len
);
2793 /* File system mounted not to panic on error
2794 * Fix the bitmap and repeat the block allocation
2795 * We leak some of the blocks here.
2797 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2798 mb_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2799 ac
->ac_b_ex
.fe_len
);
2800 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2801 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2807 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2808 #ifdef AGGRESSIVE_CHECK
2811 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2812 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2813 bitmap_bh
->b_data
));
2817 mb_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,ac
->ac_b_ex
.fe_len
);
2818 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2819 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2820 ext4_free_blks_set(sb
, gdp
,
2821 ext4_free_blocks_after_init(sb
,
2822 ac
->ac_b_ex
.fe_group
, gdp
));
2824 len
= ext4_free_blks_count(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
2825 ext4_free_blks_set(sb
, gdp
, len
);
2826 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
2828 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2829 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
2831 * Now reduce the dirty block count also. Should not go negative
2833 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2834 /* release all the reserved blocks if non delalloc */
2835 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
, reserv_blks
);
2837 if (sbi
->s_log_groups_per_flex
) {
2838 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2839 ac
->ac_b_ex
.fe_group
);
2840 atomic_sub(ac
->ac_b_ex
.fe_len
,
2841 &sbi
->s_flex_groups
[flex_group
].free_blocks
);
2844 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2847 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
2850 ext4_mark_super_dirty(sb
);
2856 * here we normalize request for locality group
2857 * Group request are normalized to s_mb_group_prealloc, which goes to
2858 * s_strip if we set the same via mount option.
2859 * s_mb_group_prealloc can be configured via
2860 * /sys/fs/ext4/<partition>/mb_group_prealloc
2862 * XXX: should we try to preallocate more than the group has now?
2864 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2866 struct super_block
*sb
= ac
->ac_sb
;
2867 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2870 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2871 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2872 current
->pid
, ac
->ac_g_ex
.fe_len
);
2876 * Normalization means making request better in terms of
2877 * size and alignment
2879 static noinline_for_stack
void
2880 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
2881 struct ext4_allocation_request
*ar
)
2885 loff_t size
, orig_size
, start_off
;
2887 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
2888 struct ext4_prealloc_space
*pa
;
2890 /* do normalize only data requests, metadata requests
2891 do not need preallocation */
2892 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
2895 /* sometime caller may want exact blocks */
2896 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2899 /* caller may indicate that preallocation isn't
2900 * required (it's a tail, for example) */
2901 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
2904 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
2905 ext4_mb_normalize_group_request(ac
);
2909 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
2911 /* first, let's learn actual file size
2912 * given current request is allocated */
2913 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
2914 size
= size
<< bsbits
;
2915 if (size
< i_size_read(ac
->ac_inode
))
2916 size
= i_size_read(ac
->ac_inode
);
2919 /* max size of free chunks */
2922 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2923 (req <= (size) || max <= (chunk_size))
2925 /* first, try to predict filesize */
2926 /* XXX: should this table be tunable? */
2928 if (size
<= 16 * 1024) {
2930 } else if (size
<= 32 * 1024) {
2932 } else if (size
<= 64 * 1024) {
2934 } else if (size
<= 128 * 1024) {
2936 } else if (size
<= 256 * 1024) {
2938 } else if (size
<= 512 * 1024) {
2940 } else if (size
<= 1024 * 1024) {
2942 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
2943 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2944 (21 - bsbits
)) << 21;
2945 size
= 2 * 1024 * 1024;
2946 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
2947 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2948 (22 - bsbits
)) << 22;
2949 size
= 4 * 1024 * 1024;
2950 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
2951 (8<<20)>>bsbits
, max
, 8 * 1024)) {
2952 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2953 (23 - bsbits
)) << 23;
2954 size
= 8 * 1024 * 1024;
2956 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
2957 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
2959 size
= size
>> bsbits
;
2960 start
= start_off
>> bsbits
;
2962 /* don't cover already allocated blocks in selected range */
2963 if (ar
->pleft
&& start
<= ar
->lleft
) {
2964 size
-= ar
->lleft
+ 1 - start
;
2965 start
= ar
->lleft
+ 1;
2967 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
2968 size
-= start
+ size
- ar
->lright
;
2972 /* check we don't cross already preallocated blocks */
2974 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
2979 spin_lock(&pa
->pa_lock
);
2980 if (pa
->pa_deleted
) {
2981 spin_unlock(&pa
->pa_lock
);
2985 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
2987 /* PA must not overlap original request */
2988 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
2989 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
2991 /* skip PAs this normalized request doesn't overlap with */
2992 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
2993 spin_unlock(&pa
->pa_lock
);
2996 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
2998 /* adjust start or end to be adjacent to this pa */
2999 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3000 BUG_ON(pa_end
< start
);
3002 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3003 BUG_ON(pa
->pa_lstart
> end
);
3004 end
= pa
->pa_lstart
;
3006 spin_unlock(&pa
->pa_lock
);
3011 /* XXX: extra loop to check we really don't overlap preallocations */
3013 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3015 spin_lock(&pa
->pa_lock
);
3016 if (pa
->pa_deleted
== 0) {
3017 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3018 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3020 spin_unlock(&pa
->pa_lock
);
3024 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3025 start
> ac
->ac_o_ex
.fe_logical
) {
3026 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3027 (unsigned long) start
, (unsigned long) size
,
3028 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3030 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3031 start
> ac
->ac_o_ex
.fe_logical
);
3032 BUG_ON(size
<= 0 || size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3034 /* now prepare goal request */
3036 /* XXX: is it better to align blocks WRT to logical
3037 * placement or satisfy big request as is */
3038 ac
->ac_g_ex
.fe_logical
= start
;
3039 ac
->ac_g_ex
.fe_len
= size
;
3041 /* define goal start in order to merge */
3042 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3043 /* merge to the right */
3044 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3045 &ac
->ac_f_ex
.fe_group
,
3046 &ac
->ac_f_ex
.fe_start
);
3047 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3049 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3050 /* merge to the left */
3051 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3052 &ac
->ac_f_ex
.fe_group
,
3053 &ac
->ac_f_ex
.fe_start
);
3054 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3057 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3058 (unsigned) orig_size
, (unsigned) start
);
3061 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3063 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3065 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3066 atomic_inc(&sbi
->s_bal_reqs
);
3067 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3068 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3069 atomic_inc(&sbi
->s_bal_success
);
3070 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3071 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3072 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3073 atomic_inc(&sbi
->s_bal_goals
);
3074 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3075 atomic_inc(&sbi
->s_bal_breaks
);
3078 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3079 trace_ext4_mballoc_alloc(ac
);
3081 trace_ext4_mballoc_prealloc(ac
);
3085 * Called on failure; free up any blocks from the inode PA for this
3086 * context. We don't need this for MB_GROUP_PA because we only change
3087 * pa_free in ext4_mb_release_context(), but on failure, we've already
3088 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3090 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3092 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3095 if (pa
&& pa
->pa_type
== MB_INODE_PA
) {
3096 len
= ac
->ac_b_ex
.fe_len
;
3103 * use blocks preallocated to inode
3105 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3106 struct ext4_prealloc_space
*pa
)
3112 /* found preallocated blocks, use them */
3113 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3114 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3116 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3117 &ac
->ac_b_ex
.fe_start
);
3118 ac
->ac_b_ex
.fe_len
= len
;
3119 ac
->ac_status
= AC_STATUS_FOUND
;
3122 BUG_ON(start
< pa
->pa_pstart
);
3123 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3124 BUG_ON(pa
->pa_free
< len
);
3127 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3131 * use blocks preallocated to locality group
3133 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3134 struct ext4_prealloc_space
*pa
)
3136 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3138 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3139 &ac
->ac_b_ex
.fe_group
,
3140 &ac
->ac_b_ex
.fe_start
);
3141 ac
->ac_b_ex
.fe_len
= len
;
3142 ac
->ac_status
= AC_STATUS_FOUND
;
3145 /* we don't correct pa_pstart or pa_plen here to avoid
3146 * possible race when the group is being loaded concurrently
3147 * instead we correct pa later, after blocks are marked
3148 * in on-disk bitmap -- see ext4_mb_release_context()
3149 * Other CPUs are prevented from allocating from this pa by lg_mutex
3151 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3155 * Return the prealloc space that have minimal distance
3156 * from the goal block. @cpa is the prealloc
3157 * space that is having currently known minimal distance
3158 * from the goal block.
3160 static struct ext4_prealloc_space
*
3161 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3162 struct ext4_prealloc_space
*pa
,
3163 struct ext4_prealloc_space
*cpa
)
3165 ext4_fsblk_t cur_distance
, new_distance
;
3168 atomic_inc(&pa
->pa_count
);
3171 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3172 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3174 if (cur_distance
<= new_distance
)
3177 /* drop the previous reference */
3178 atomic_dec(&cpa
->pa_count
);
3179 atomic_inc(&pa
->pa_count
);
3184 * search goal blocks in preallocated space
3186 static noinline_for_stack
int
3187 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3190 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3191 struct ext4_locality_group
*lg
;
3192 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3193 ext4_fsblk_t goal_block
;
3195 /* only data can be preallocated */
3196 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3199 /* first, try per-file preallocation */
3201 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3203 /* all fields in this condition don't change,
3204 * so we can skip locking for them */
3205 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3206 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3209 /* non-extent files can't have physical blocks past 2^32 */
3210 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3211 pa
->pa_pstart
+ pa
->pa_len
> EXT4_MAX_BLOCK_FILE_PHYS
)
3214 /* found preallocated blocks, use them */
3215 spin_lock(&pa
->pa_lock
);
3216 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3217 atomic_inc(&pa
->pa_count
);
3218 ext4_mb_use_inode_pa(ac
, pa
);
3219 spin_unlock(&pa
->pa_lock
);
3220 ac
->ac_criteria
= 10;
3224 spin_unlock(&pa
->pa_lock
);
3228 /* can we use group allocation? */
3229 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3232 /* inode may have no locality group for some reason */
3236 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3237 if (order
> PREALLOC_TB_SIZE
- 1)
3238 /* The max size of hash table is PREALLOC_TB_SIZE */
3239 order
= PREALLOC_TB_SIZE
- 1;
3241 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3243 * search for the prealloc space that is having
3244 * minimal distance from the goal block.
3246 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3248 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3250 spin_lock(&pa
->pa_lock
);
3251 if (pa
->pa_deleted
== 0 &&
3252 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3254 cpa
= ext4_mb_check_group_pa(goal_block
,
3257 spin_unlock(&pa
->pa_lock
);
3262 ext4_mb_use_group_pa(ac
, cpa
);
3263 ac
->ac_criteria
= 20;
3270 * the function goes through all block freed in the group
3271 * but not yet committed and marks them used in in-core bitmap.
3272 * buddy must be generated from this bitmap
3273 * Need to be called with the ext4 group lock held
3275 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3279 struct ext4_group_info
*grp
;
3280 struct ext4_free_data
*entry
;
3282 grp
= ext4_get_group_info(sb
, group
);
3283 n
= rb_first(&(grp
->bb_free_root
));
3286 entry
= rb_entry(n
, struct ext4_free_data
, node
);
3287 mb_set_bits(bitmap
, entry
->start_blk
, entry
->count
);
3294 * the function goes through all preallocation in this group and marks them
3295 * used in in-core bitmap. buddy must be generated from this bitmap
3296 * Need to be called with ext4 group lock held
3298 static noinline_for_stack
3299 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3302 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3303 struct ext4_prealloc_space
*pa
;
3304 struct list_head
*cur
;
3305 ext4_group_t groupnr
;
3306 ext4_grpblk_t start
;
3307 int preallocated
= 0;
3311 /* all form of preallocation discards first load group,
3312 * so the only competing code is preallocation use.
3313 * we don't need any locking here
3314 * notice we do NOT ignore preallocations with pa_deleted
3315 * otherwise we could leave used blocks available for
3316 * allocation in buddy when concurrent ext4_mb_put_pa()
3317 * is dropping preallocation
3319 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3320 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3321 spin_lock(&pa
->pa_lock
);
3322 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3325 spin_unlock(&pa
->pa_lock
);
3326 if (unlikely(len
== 0))
3328 BUG_ON(groupnr
!= group
);
3329 mb_set_bits(bitmap
, start
, len
);
3330 preallocated
+= len
;
3333 mb_debug(1, "prellocated %u for group %u\n", preallocated
, group
);
3336 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3338 struct ext4_prealloc_space
*pa
;
3339 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3340 kmem_cache_free(ext4_pspace_cachep
, pa
);
3344 * drops a reference to preallocated space descriptor
3345 * if this was the last reference and the space is consumed
3347 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3348 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3351 ext4_fsblk_t grp_blk
;
3353 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3356 /* in this short window concurrent discard can set pa_deleted */
3357 spin_lock(&pa
->pa_lock
);
3358 if (pa
->pa_deleted
== 1) {
3359 spin_unlock(&pa
->pa_lock
);
3364 spin_unlock(&pa
->pa_lock
);
3366 grp_blk
= pa
->pa_pstart
;
3368 * If doing group-based preallocation, pa_pstart may be in the
3369 * next group when pa is used up
3371 if (pa
->pa_type
== MB_GROUP_PA
)
3374 ext4_get_group_no_and_offset(sb
, grp_blk
, &grp
, NULL
);
3379 * P1 (buddy init) P2 (regular allocation)
3380 * find block B in PA
3381 * copy on-disk bitmap to buddy
3382 * mark B in on-disk bitmap
3383 * drop PA from group
3384 * mark all PAs in buddy
3386 * thus, P1 initializes buddy with B available. to prevent this
3387 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3390 ext4_lock_group(sb
, grp
);
3391 list_del(&pa
->pa_group_list
);
3392 ext4_unlock_group(sb
, grp
);
3394 spin_lock(pa
->pa_obj_lock
);
3395 list_del_rcu(&pa
->pa_inode_list
);
3396 spin_unlock(pa
->pa_obj_lock
);
3398 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3402 * creates new preallocated space for given inode
3404 static noinline_for_stack
int
3405 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3407 struct super_block
*sb
= ac
->ac_sb
;
3408 struct ext4_prealloc_space
*pa
;
3409 struct ext4_group_info
*grp
;
3410 struct ext4_inode_info
*ei
;
3412 /* preallocate only when found space is larger then requested */
3413 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3414 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3415 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3417 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3421 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3427 /* we can't allocate as much as normalizer wants.
3428 * so, found space must get proper lstart
3429 * to cover original request */
3430 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3431 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3433 /* we're limited by original request in that
3434 * logical block must be covered any way
3435 * winl is window we can move our chunk within */
3436 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3438 /* also, we should cover whole original request */
3439 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3441 /* the smallest one defines real window */
3442 win
= min(winl
, wins
);
3444 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3445 if (offs
&& offs
< win
)
3448 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3449 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3450 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3453 /* preallocation can change ac_b_ex, thus we store actually
3454 * allocated blocks for history */
3455 ac
->ac_f_ex
= ac
->ac_b_ex
;
3457 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3458 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3459 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3460 pa
->pa_free
= pa
->pa_len
;
3461 atomic_set(&pa
->pa_count
, 1);
3462 spin_lock_init(&pa
->pa_lock
);
3463 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3464 INIT_LIST_HEAD(&pa
->pa_group_list
);
3466 pa
->pa_type
= MB_INODE_PA
;
3468 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3469 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3470 trace_ext4_mb_new_inode_pa(ac
, pa
);
3472 ext4_mb_use_inode_pa(ac
, pa
);
3473 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3475 ei
= EXT4_I(ac
->ac_inode
);
3476 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3478 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3479 pa
->pa_inode
= ac
->ac_inode
;
3481 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3482 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3483 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3485 spin_lock(pa
->pa_obj_lock
);
3486 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3487 spin_unlock(pa
->pa_obj_lock
);
3493 * creates new preallocated space for locality group inodes belongs to
3495 static noinline_for_stack
int
3496 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3498 struct super_block
*sb
= ac
->ac_sb
;
3499 struct ext4_locality_group
*lg
;
3500 struct ext4_prealloc_space
*pa
;
3501 struct ext4_group_info
*grp
;
3503 /* preallocate only when found space is larger then requested */
3504 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3505 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3506 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3508 BUG_ON(ext4_pspace_cachep
== NULL
);
3509 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3513 /* preallocation can change ac_b_ex, thus we store actually
3514 * allocated blocks for history */
3515 ac
->ac_f_ex
= ac
->ac_b_ex
;
3517 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3518 pa
->pa_lstart
= pa
->pa_pstart
;
3519 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3520 pa
->pa_free
= pa
->pa_len
;
3521 atomic_set(&pa
->pa_count
, 1);
3522 spin_lock_init(&pa
->pa_lock
);
3523 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3524 INIT_LIST_HEAD(&pa
->pa_group_list
);
3526 pa
->pa_type
= MB_GROUP_PA
;
3528 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3529 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3530 trace_ext4_mb_new_group_pa(ac
, pa
);
3532 ext4_mb_use_group_pa(ac
, pa
);
3533 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3535 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3539 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3540 pa
->pa_inode
= NULL
;
3542 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3543 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3544 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3547 * We will later add the new pa to the right bucket
3548 * after updating the pa_free in ext4_mb_release_context
3553 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3557 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3558 err
= ext4_mb_new_group_pa(ac
);
3560 err
= ext4_mb_new_inode_pa(ac
);
3565 * finds all unused blocks in on-disk bitmap, frees them in
3566 * in-core bitmap and buddy.
3567 * @pa must be unlinked from inode and group lists, so that
3568 * nobody else can find/use it.
3569 * the caller MUST hold group/inode locks.
3570 * TODO: optimize the case when there are no in-core structures yet
3572 static noinline_for_stack
int
3573 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3574 struct ext4_prealloc_space
*pa
)
3576 struct super_block
*sb
= e4b
->bd_sb
;
3577 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3582 unsigned long long grp_blk_start
;
3586 BUG_ON(pa
->pa_deleted
== 0);
3587 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3588 grp_blk_start
= pa
->pa_pstart
- bit
;
3589 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3590 end
= bit
+ pa
->pa_len
;
3593 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3596 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3597 mb_debug(1, " free preallocated %u/%u in group %u\n",
3598 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3599 (unsigned) next
- bit
, (unsigned) group
);
3602 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3603 trace_ext4_mb_release_inode_pa(pa
, grp_blk_start
+ bit
,
3605 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3608 if (free
!= pa
->pa_free
) {
3609 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3610 pa
, (unsigned long) pa
->pa_lstart
,
3611 (unsigned long) pa
->pa_pstart
,
3612 (unsigned long) pa
->pa_len
);
3613 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3616 * pa is already deleted so we use the value obtained
3617 * from the bitmap and continue.
3620 atomic_add(free
, &sbi
->s_mb_discarded
);
3625 static noinline_for_stack
int
3626 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3627 struct ext4_prealloc_space
*pa
)
3629 struct super_block
*sb
= e4b
->bd_sb
;
3633 trace_ext4_mb_release_group_pa(pa
);
3634 BUG_ON(pa
->pa_deleted
== 0);
3635 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3636 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3637 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3638 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3639 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3645 * releases all preallocations in given group
3647 * first, we need to decide discard policy:
3648 * - when do we discard
3650 * - how many do we discard
3651 * 1) how many requested
3653 static noinline_for_stack
int
3654 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3655 ext4_group_t group
, int needed
)
3657 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3658 struct buffer_head
*bitmap_bh
= NULL
;
3659 struct ext4_prealloc_space
*pa
, *tmp
;
3660 struct list_head list
;
3661 struct ext4_buddy e4b
;
3666 mb_debug(1, "discard preallocation for group %u\n", group
);
3668 if (list_empty(&grp
->bb_prealloc_list
))
3671 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3672 if (bitmap_bh
== NULL
) {
3673 ext4_error(sb
, "Error reading block bitmap for %u", group
);
3677 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3679 ext4_error(sb
, "Error loading buddy information for %u", group
);
3685 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3687 INIT_LIST_HEAD(&list
);
3689 ext4_lock_group(sb
, group
);
3690 list_for_each_entry_safe(pa
, tmp
,
3691 &grp
->bb_prealloc_list
, pa_group_list
) {
3692 spin_lock(&pa
->pa_lock
);
3693 if (atomic_read(&pa
->pa_count
)) {
3694 spin_unlock(&pa
->pa_lock
);
3698 if (pa
->pa_deleted
) {
3699 spin_unlock(&pa
->pa_lock
);
3703 /* seems this one can be freed ... */
3706 /* we can trust pa_free ... */
3707 free
+= pa
->pa_free
;
3709 spin_unlock(&pa
->pa_lock
);
3711 list_del(&pa
->pa_group_list
);
3712 list_add(&pa
->u
.pa_tmp_list
, &list
);
3715 /* if we still need more blocks and some PAs were used, try again */
3716 if (free
< needed
&& busy
) {
3718 ext4_unlock_group(sb
, group
);
3720 * Yield the CPU here so that we don't get soft lockup
3721 * in non preempt case.
3727 /* found anything to free? */
3728 if (list_empty(&list
)) {
3733 /* now free all selected PAs */
3734 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3736 /* remove from object (inode or locality group) */
3737 spin_lock(pa
->pa_obj_lock
);
3738 list_del_rcu(&pa
->pa_inode_list
);
3739 spin_unlock(pa
->pa_obj_lock
);
3741 if (pa
->pa_type
== MB_GROUP_PA
)
3742 ext4_mb_release_group_pa(&e4b
, pa
);
3744 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3746 list_del(&pa
->u
.pa_tmp_list
);
3747 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3751 ext4_unlock_group(sb
, group
);
3752 ext4_mb_unload_buddy(&e4b
);
3758 * releases all non-used preallocated blocks for given inode
3760 * It's important to discard preallocations under i_data_sem
3761 * We don't want another block to be served from the prealloc
3762 * space when we are discarding the inode prealloc space.
3764 * FIXME!! Make sure it is valid at all the call sites
3766 void ext4_discard_preallocations(struct inode
*inode
)
3768 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3769 struct super_block
*sb
= inode
->i_sb
;
3770 struct buffer_head
*bitmap_bh
= NULL
;
3771 struct ext4_prealloc_space
*pa
, *tmp
;
3772 ext4_group_t group
= 0;
3773 struct list_head list
;
3774 struct ext4_buddy e4b
;
3777 if (!S_ISREG(inode
->i_mode
)) {
3778 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3782 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
3783 trace_ext4_discard_preallocations(inode
);
3785 INIT_LIST_HEAD(&list
);
3788 /* first, collect all pa's in the inode */
3789 spin_lock(&ei
->i_prealloc_lock
);
3790 while (!list_empty(&ei
->i_prealloc_list
)) {
3791 pa
= list_entry(ei
->i_prealloc_list
.next
,
3792 struct ext4_prealloc_space
, pa_inode_list
);
3793 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3794 spin_lock(&pa
->pa_lock
);
3795 if (atomic_read(&pa
->pa_count
)) {
3796 /* this shouldn't happen often - nobody should
3797 * use preallocation while we're discarding it */
3798 spin_unlock(&pa
->pa_lock
);
3799 spin_unlock(&ei
->i_prealloc_lock
);
3800 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
3802 schedule_timeout_uninterruptible(HZ
);
3806 if (pa
->pa_deleted
== 0) {
3808 spin_unlock(&pa
->pa_lock
);
3809 list_del_rcu(&pa
->pa_inode_list
);
3810 list_add(&pa
->u
.pa_tmp_list
, &list
);
3814 /* someone is deleting pa right now */
3815 spin_unlock(&pa
->pa_lock
);
3816 spin_unlock(&ei
->i_prealloc_lock
);
3818 /* we have to wait here because pa_deleted
3819 * doesn't mean pa is already unlinked from
3820 * the list. as we might be called from
3821 * ->clear_inode() the inode will get freed
3822 * and concurrent thread which is unlinking
3823 * pa from inode's list may access already
3824 * freed memory, bad-bad-bad */
3826 /* XXX: if this happens too often, we can
3827 * add a flag to force wait only in case
3828 * of ->clear_inode(), but not in case of
3829 * regular truncate */
3830 schedule_timeout_uninterruptible(HZ
);
3833 spin_unlock(&ei
->i_prealloc_lock
);
3835 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3836 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
3837 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
3839 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3841 ext4_error(sb
, "Error loading buddy information for %u",
3846 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3847 if (bitmap_bh
== NULL
) {
3848 ext4_error(sb
, "Error reading block bitmap for %u",
3850 ext4_mb_unload_buddy(&e4b
);
3854 ext4_lock_group(sb
, group
);
3855 list_del(&pa
->pa_group_list
);
3856 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3857 ext4_unlock_group(sb
, group
);
3859 ext4_mb_unload_buddy(&e4b
);
3862 list_del(&pa
->u
.pa_tmp_list
);
3863 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3867 #ifdef CONFIG_EXT4_DEBUG
3868 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3870 struct super_block
*sb
= ac
->ac_sb
;
3871 ext4_group_t ngroups
, i
;
3873 if (!mb_enable_debug
||
3874 (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
))
3877 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
3878 " Allocation context details:\n");
3879 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
3880 ac
->ac_status
, ac
->ac_flags
);
3881 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3882 "best %lu/%lu/%lu@%lu cr %d\n",
3883 (unsigned long)ac
->ac_o_ex
.fe_group
,
3884 (unsigned long)ac
->ac_o_ex
.fe_start
,
3885 (unsigned long)ac
->ac_o_ex
.fe_len
,
3886 (unsigned long)ac
->ac_o_ex
.fe_logical
,
3887 (unsigned long)ac
->ac_g_ex
.fe_group
,
3888 (unsigned long)ac
->ac_g_ex
.fe_start
,
3889 (unsigned long)ac
->ac_g_ex
.fe_len
,
3890 (unsigned long)ac
->ac_g_ex
.fe_logical
,
3891 (unsigned long)ac
->ac_b_ex
.fe_group
,
3892 (unsigned long)ac
->ac_b_ex
.fe_start
,
3893 (unsigned long)ac
->ac_b_ex
.fe_len
,
3894 (unsigned long)ac
->ac_b_ex
.fe_logical
,
3895 (int)ac
->ac_criteria
);
3896 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
3898 printk(KERN_ERR
"EXT4-fs: groups: \n");
3899 ngroups
= ext4_get_groups_count(sb
);
3900 for (i
= 0; i
< ngroups
; i
++) {
3901 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
3902 struct ext4_prealloc_space
*pa
;
3903 ext4_grpblk_t start
;
3904 struct list_head
*cur
;
3905 ext4_lock_group(sb
, i
);
3906 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3907 pa
= list_entry(cur
, struct ext4_prealloc_space
,
3909 spin_lock(&pa
->pa_lock
);
3910 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3912 spin_unlock(&pa
->pa_lock
);
3913 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
3916 ext4_unlock_group(sb
, i
);
3918 if (grp
->bb_free
== 0)
3920 printk(KERN_ERR
"%u: %d/%d \n",
3921 i
, grp
->bb_free
, grp
->bb_fragments
);
3923 printk(KERN_ERR
"\n");
3926 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3933 * We use locality group preallocation for small size file. The size of the
3934 * file is determined by the current size or the resulting size after
3935 * allocation which ever is larger
3937 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3939 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
3941 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3942 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3945 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3948 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3951 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3952 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
3955 if ((size
== isize
) &&
3956 !ext4_fs_is_busy(sbi
) &&
3957 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
3958 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
3962 /* don't use group allocation for large files */
3963 size
= max(size
, isize
);
3964 if (size
> sbi
->s_mb_stream_request
) {
3965 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
3969 BUG_ON(ac
->ac_lg
!= NULL
);
3971 * locality group prealloc space are per cpu. The reason for having
3972 * per cpu locality group is to reduce the contention between block
3973 * request from multiple CPUs.
3975 ac
->ac_lg
= __this_cpu_ptr(sbi
->s_locality_groups
);
3977 /* we're going to use group allocation */
3978 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
3980 /* serialize all allocations in the group */
3981 mutex_lock(&ac
->ac_lg
->lg_mutex
);
3984 static noinline_for_stack
int
3985 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
3986 struct ext4_allocation_request
*ar
)
3988 struct super_block
*sb
= ar
->inode
->i_sb
;
3989 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3990 struct ext4_super_block
*es
= sbi
->s_es
;
3994 ext4_grpblk_t block
;
3996 /* we can't allocate > group size */
3999 /* just a dirty hack to filter too big requests */
4000 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4001 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4003 /* start searching from the goal */
4005 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4006 goal
>= ext4_blocks_count(es
))
4007 goal
= le32_to_cpu(es
->s_first_data_block
);
4008 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4010 /* set up allocation goals */
4011 memset(ac
, 0, sizeof(struct ext4_allocation_context
));
4012 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4013 ac
->ac_status
= AC_STATUS_CONTINUE
;
4015 ac
->ac_inode
= ar
->inode
;
4016 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4017 ac
->ac_o_ex
.fe_group
= group
;
4018 ac
->ac_o_ex
.fe_start
= block
;
4019 ac
->ac_o_ex
.fe_len
= len
;
4020 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4021 ac
->ac_g_ex
.fe_group
= group
;
4022 ac
->ac_g_ex
.fe_start
= block
;
4023 ac
->ac_g_ex
.fe_len
= len
;
4024 ac
->ac_flags
= ar
->flags
;
4026 /* we have to define context: we'll we work with a file or
4027 * locality group. this is a policy, actually */
4028 ext4_mb_group_or_file(ac
);
4030 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4031 "left: %u/%u, right %u/%u to %swritable\n",
4032 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4033 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4034 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4035 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4036 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4041 static noinline_for_stack
void
4042 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4043 struct ext4_locality_group
*lg
,
4044 int order
, int total_entries
)
4046 ext4_group_t group
= 0;
4047 struct ext4_buddy e4b
;
4048 struct list_head discard_list
;
4049 struct ext4_prealloc_space
*pa
, *tmp
;
4051 mb_debug(1, "discard locality group preallocation\n");
4053 INIT_LIST_HEAD(&discard_list
);
4055 spin_lock(&lg
->lg_prealloc_lock
);
4056 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4058 spin_lock(&pa
->pa_lock
);
4059 if (atomic_read(&pa
->pa_count
)) {
4061 * This is the pa that we just used
4062 * for block allocation. So don't
4065 spin_unlock(&pa
->pa_lock
);
4068 if (pa
->pa_deleted
) {
4069 spin_unlock(&pa
->pa_lock
);
4072 /* only lg prealloc space */
4073 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4075 /* seems this one can be freed ... */
4077 spin_unlock(&pa
->pa_lock
);
4079 list_del_rcu(&pa
->pa_inode_list
);
4080 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4083 if (total_entries
<= 5) {
4085 * we want to keep only 5 entries
4086 * allowing it to grow to 8. This
4087 * mak sure we don't call discard
4088 * soon for this list.
4093 spin_unlock(&lg
->lg_prealloc_lock
);
4095 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4097 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4098 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4099 ext4_error(sb
, "Error loading buddy information for %u",
4103 ext4_lock_group(sb
, group
);
4104 list_del(&pa
->pa_group_list
);
4105 ext4_mb_release_group_pa(&e4b
, pa
);
4106 ext4_unlock_group(sb
, group
);
4108 ext4_mb_unload_buddy(&e4b
);
4109 list_del(&pa
->u
.pa_tmp_list
);
4110 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4115 * We have incremented pa_count. So it cannot be freed at this
4116 * point. Also we hold lg_mutex. So no parallel allocation is
4117 * possible from this lg. That means pa_free cannot be updated.
4119 * A parallel ext4_mb_discard_group_preallocations is possible.
4120 * which can cause the lg_prealloc_list to be updated.
4123 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4125 int order
, added
= 0, lg_prealloc_count
= 1;
4126 struct super_block
*sb
= ac
->ac_sb
;
4127 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4128 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4130 order
= fls(pa
->pa_free
) - 1;
4131 if (order
> PREALLOC_TB_SIZE
- 1)
4132 /* The max size of hash table is PREALLOC_TB_SIZE */
4133 order
= PREALLOC_TB_SIZE
- 1;
4134 /* Add the prealloc space to lg */
4136 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4138 spin_lock(&tmp_pa
->pa_lock
);
4139 if (tmp_pa
->pa_deleted
) {
4140 spin_unlock(&tmp_pa
->pa_lock
);
4143 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4144 /* Add to the tail of the previous entry */
4145 list_add_tail_rcu(&pa
->pa_inode_list
,
4146 &tmp_pa
->pa_inode_list
);
4149 * we want to count the total
4150 * number of entries in the list
4153 spin_unlock(&tmp_pa
->pa_lock
);
4154 lg_prealloc_count
++;
4157 list_add_tail_rcu(&pa
->pa_inode_list
,
4158 &lg
->lg_prealloc_list
[order
]);
4161 /* Now trim the list to be not more than 8 elements */
4162 if (lg_prealloc_count
> 8) {
4163 ext4_mb_discard_lg_preallocations(sb
, lg
,
4164 order
, lg_prealloc_count
);
4171 * release all resource we used in allocation
4173 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4175 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4177 if (pa
->pa_type
== MB_GROUP_PA
) {
4178 /* see comment in ext4_mb_use_group_pa() */
4179 spin_lock(&pa
->pa_lock
);
4180 pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4181 pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4182 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4183 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4184 spin_unlock(&pa
->pa_lock
);
4189 * We want to add the pa to the right bucket.
4190 * Remove it from the list and while adding
4191 * make sure the list to which we are adding
4194 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4195 spin_lock(pa
->pa_obj_lock
);
4196 list_del_rcu(&pa
->pa_inode_list
);
4197 spin_unlock(pa
->pa_obj_lock
);
4198 ext4_mb_add_n_trim(ac
);
4200 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4202 if (ac
->ac_bitmap_page
)
4203 page_cache_release(ac
->ac_bitmap_page
);
4204 if (ac
->ac_buddy_page
)
4205 page_cache_release(ac
->ac_buddy_page
);
4206 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4207 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4208 ext4_mb_collect_stats(ac
);
4212 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4214 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4218 trace_ext4_mb_discard_preallocations(sb
, needed
);
4219 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4220 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4229 * Main entry point into mballoc to allocate blocks
4230 * it tries to use preallocation first, then falls back
4231 * to usual allocation
4233 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4234 struct ext4_allocation_request
*ar
, int *errp
)
4237 struct ext4_allocation_context
*ac
= NULL
;
4238 struct ext4_sb_info
*sbi
;
4239 struct super_block
*sb
;
4240 ext4_fsblk_t block
= 0;
4241 unsigned int inquota
= 0;
4242 unsigned int reserv_blks
= 0;
4244 sb
= ar
->inode
->i_sb
;
4247 trace_ext4_request_blocks(ar
);
4250 * For delayed allocation, we could skip the ENOSPC and
4251 * EDQUOT check, as blocks and quotas have been already
4252 * reserved when data being copied into pagecache.
4254 if (ext4_test_inode_state(ar
->inode
, EXT4_STATE_DELALLOC_RESERVED
))
4255 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4257 /* Without delayed allocation we need to verify
4258 * there is enough free blocks to do block allocation
4259 * and verify allocation doesn't exceed the quota limits.
4262 ext4_claim_free_blocks(sbi
, ar
->len
, ar
->flags
)) {
4264 /* let others to free the space */
4266 ar
->len
= ar
->len
>> 1;
4272 reserv_blks
= ar
->len
;
4273 if (ar
->flags
& EXT4_MB_USE_ROOT_BLOCKS
) {
4274 dquot_alloc_block_nofail(ar
->inode
, ar
->len
);
4277 dquot_alloc_block(ar
->inode
, ar
->len
)) {
4279 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4290 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4297 *errp
= ext4_mb_initialize_context(ac
, ar
);
4303 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4304 if (!ext4_mb_use_preallocated(ac
)) {
4305 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4306 ext4_mb_normalize_request(ac
, ar
);
4308 /* allocate space in core */
4309 *errp
= ext4_mb_regular_allocator(ac
);
4313 /* as we've just preallocated more space than
4314 * user requested orinally, we store allocated
4315 * space in a special descriptor */
4316 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4317 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4318 ext4_mb_new_preallocation(ac
);
4320 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4321 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_blks
);
4322 if (*errp
== -EAGAIN
) {
4324 * drop the reference that we took
4325 * in ext4_mb_use_best_found
4327 ext4_mb_release_context(ac
);
4328 ac
->ac_b_ex
.fe_group
= 0;
4329 ac
->ac_b_ex
.fe_start
= 0;
4330 ac
->ac_b_ex
.fe_len
= 0;
4331 ac
->ac_status
= AC_STATUS_CONTINUE
;
4335 ext4_discard_allocated_blocks(ac
);
4337 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4338 ar
->len
= ac
->ac_b_ex
.fe_len
;
4341 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4348 ac
->ac_b_ex
.fe_len
= 0;
4350 ext4_mb_show_ac(ac
);
4352 ext4_mb_release_context(ac
);
4355 kmem_cache_free(ext4_ac_cachep
, ac
);
4356 if (inquota
&& ar
->len
< inquota
)
4357 dquot_free_block(ar
->inode
, inquota
- ar
->len
);
4359 if (!ext4_test_inode_state(ar
->inode
,
4360 EXT4_STATE_DELALLOC_RESERVED
))
4361 /* release all the reserved blocks if non delalloc */
4362 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
,
4366 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4372 * We can merge two free data extents only if the physical blocks
4373 * are contiguous, AND the extents were freed by the same transaction,
4374 * AND the blocks are associated with the same group.
4376 static int can_merge(struct ext4_free_data
*entry1
,
4377 struct ext4_free_data
*entry2
)
4379 if ((entry1
->t_tid
== entry2
->t_tid
) &&
4380 (entry1
->group
== entry2
->group
) &&
4381 ((entry1
->start_blk
+ entry1
->count
) == entry2
->start_blk
))
4386 static noinline_for_stack
int
4387 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4388 struct ext4_free_data
*new_entry
)
4390 ext4_group_t group
= e4b
->bd_group
;
4391 ext4_grpblk_t block
;
4392 struct ext4_free_data
*entry
;
4393 struct ext4_group_info
*db
= e4b
->bd_info
;
4394 struct super_block
*sb
= e4b
->bd_sb
;
4395 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4396 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4397 struct rb_node
*parent
= NULL
, *new_node
;
4399 BUG_ON(!ext4_handle_valid(handle
));
4400 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4401 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4403 new_node
= &new_entry
->node
;
4404 block
= new_entry
->start_blk
;
4407 /* first free block exent. We need to
4408 protect buddy cache from being freed,
4409 * otherwise we'll refresh it from
4410 * on-disk bitmap and lose not-yet-available
4412 page_cache_get(e4b
->bd_buddy_page
);
4413 page_cache_get(e4b
->bd_bitmap_page
);
4417 entry
= rb_entry(parent
, struct ext4_free_data
, node
);
4418 if (block
< entry
->start_blk
)
4420 else if (block
>= (entry
->start_blk
+ entry
->count
))
4421 n
= &(*n
)->rb_right
;
4423 ext4_grp_locked_error(sb
, group
, 0,
4424 ext4_group_first_block_no(sb
, group
) + block
,
4425 "Block already on to-be-freed list");
4430 rb_link_node(new_node
, parent
, n
);
4431 rb_insert_color(new_node
, &db
->bb_free_root
);
4433 /* Now try to see the extent can be merged to left and right */
4434 node
= rb_prev(new_node
);
4436 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4437 if (can_merge(entry
, new_entry
)) {
4438 new_entry
->start_blk
= entry
->start_blk
;
4439 new_entry
->count
+= entry
->count
;
4440 rb_erase(node
, &(db
->bb_free_root
));
4441 spin_lock(&sbi
->s_md_lock
);
4442 list_del(&entry
->list
);
4443 spin_unlock(&sbi
->s_md_lock
);
4444 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4448 node
= rb_next(new_node
);
4450 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4451 if (can_merge(new_entry
, entry
)) {
4452 new_entry
->count
+= entry
->count
;
4453 rb_erase(node
, &(db
->bb_free_root
));
4454 spin_lock(&sbi
->s_md_lock
);
4455 list_del(&entry
->list
);
4456 spin_unlock(&sbi
->s_md_lock
);
4457 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4460 /* Add the extent to transaction's private list */
4461 spin_lock(&sbi
->s_md_lock
);
4462 list_add(&new_entry
->list
, &handle
->h_transaction
->t_private_list
);
4463 spin_unlock(&sbi
->s_md_lock
);
4468 * ext4_free_blocks() -- Free given blocks and update quota
4469 * @handle: handle for this transaction
4471 * @block: start physical block to free
4472 * @count: number of blocks to count
4473 * @flags: flags used by ext4_free_blocks
4475 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4476 struct buffer_head
*bh
, ext4_fsblk_t block
,
4477 unsigned long count
, int flags
)
4479 struct buffer_head
*bitmap_bh
= NULL
;
4480 struct super_block
*sb
= inode
->i_sb
;
4481 struct ext4_group_desc
*gdp
;
4482 unsigned long freed
= 0;
4483 unsigned int overflow
;
4485 struct buffer_head
*gd_bh
;
4486 ext4_group_t block_group
;
4487 struct ext4_sb_info
*sbi
;
4488 struct ext4_buddy e4b
;
4494 BUG_ON(block
!= bh
->b_blocknr
);
4496 block
= bh
->b_blocknr
;
4500 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4501 !ext4_data_block_valid(sbi
, block
, count
)) {
4502 ext4_error(sb
, "Freeing blocks not in datazone - "
4503 "block = %llu, count = %lu", block
, count
);
4507 ext4_debug("freeing block %llu\n", block
);
4508 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4510 if (flags
& EXT4_FREE_BLOCKS_FORGET
) {
4511 struct buffer_head
*tbh
= bh
;
4514 BUG_ON(bh
&& (count
> 1));
4516 for (i
= 0; i
< count
; i
++) {
4518 tbh
= sb_find_get_block(inode
->i_sb
,
4522 ext4_forget(handle
, flags
& EXT4_FREE_BLOCKS_METADATA
,
4523 inode
, tbh
, block
+ i
);
4528 * We need to make sure we don't reuse the freed block until
4529 * after the transaction is committed, which we can do by
4530 * treating the block as metadata, below. We make an
4531 * exception if the inode is to be written in writeback mode
4532 * since writeback mode has weak data consistency guarantees.
4534 if (!ext4_should_writeback_data(inode
))
4535 flags
|= EXT4_FREE_BLOCKS_METADATA
;
4539 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4542 * Check to see if we are freeing blocks across a group
4545 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4546 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4549 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4554 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4560 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4561 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4562 in_range(block
, ext4_inode_table(sb
, gdp
),
4563 EXT4_SB(sb
)->s_itb_per_group
) ||
4564 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4565 EXT4_SB(sb
)->s_itb_per_group
)) {
4567 ext4_error(sb
, "Freeing blocks in system zone - "
4568 "Block = %llu, count = %lu", block
, count
);
4569 /* err = 0. ext4_std_error should be a no op */
4573 BUFFER_TRACE(bitmap_bh
, "getting write access");
4574 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4579 * We are about to modify some metadata. Call the journal APIs
4580 * to unshare ->b_data if a currently-committing transaction is
4583 BUFFER_TRACE(gd_bh
, "get_write_access");
4584 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4587 #ifdef AGGRESSIVE_CHECK
4590 for (i
= 0; i
< count
; i
++)
4591 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4594 trace_ext4_mballoc_free(sb
, inode
, block_group
, bit
, count
);
4596 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4600 if ((flags
& EXT4_FREE_BLOCKS_METADATA
) && ext4_handle_valid(handle
)) {
4601 struct ext4_free_data
*new_entry
;
4603 * blocks being freed are metadata. these blocks shouldn't
4604 * be used until this transaction is committed
4606 new_entry
= kmem_cache_alloc(ext4_free_ext_cachep
, GFP_NOFS
);
4611 new_entry
->start_blk
= bit
;
4612 new_entry
->group
= block_group
;
4613 new_entry
->count
= count
;
4614 new_entry
->t_tid
= handle
->h_transaction
->t_tid
;
4616 ext4_lock_group(sb
, block_group
);
4617 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4618 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4620 /* need to update group_info->bb_free and bitmap
4621 * with group lock held. generate_buddy look at
4622 * them with group lock_held
4624 ext4_lock_group(sb
, block_group
);
4625 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4626 mb_free_blocks(inode
, &e4b
, bit
, count
);
4629 ret
= ext4_free_blks_count(sb
, gdp
) + count
;
4630 ext4_free_blks_set(sb
, gdp
, ret
);
4631 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4632 ext4_unlock_group(sb
, block_group
);
4633 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4635 if (sbi
->s_log_groups_per_flex
) {
4636 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4637 atomic_add(count
, &sbi
->s_flex_groups
[flex_group
].free_blocks
);
4640 ext4_mb_unload_buddy(&e4b
);
4644 /* We dirtied the bitmap block */
4645 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4646 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4648 /* And the group descriptor block */
4649 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4650 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4654 if (overflow
&& !err
) {
4660 ext4_mark_super_dirty(sb
);
4662 if (freed
&& !(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4663 dquot_free_block(inode
, freed
);
4665 ext4_std_error(sb
, err
);
4670 * ext4_add_groupblocks() -- Add given blocks to an existing group
4671 * @handle: handle to this transaction
4673 * @block: start physcial block to add to the block group
4674 * @count: number of blocks to free
4676 * This marks the blocks as free in the bitmap and buddy.
4678 void ext4_add_groupblocks(handle_t
*handle
, struct super_block
*sb
,
4679 ext4_fsblk_t block
, unsigned long count
)
4681 struct buffer_head
*bitmap_bh
= NULL
;
4682 struct buffer_head
*gd_bh
;
4683 ext4_group_t block_group
;
4686 struct ext4_group_desc
*desc
;
4687 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4688 struct ext4_buddy e4b
;
4689 int err
= 0, ret
, blk_free_count
;
4690 ext4_grpblk_t blocks_freed
;
4692 ext4_debug("Adding block(s) %llu-%llu\n", block
, block
+ count
- 1);
4694 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4696 * Check to see if we are freeing blocks across a group
4699 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
))
4702 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4705 desc
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4709 if (in_range(ext4_block_bitmap(sb
, desc
), block
, count
) ||
4710 in_range(ext4_inode_bitmap(sb
, desc
), block
, count
) ||
4711 in_range(block
, ext4_inode_table(sb
, desc
), sbi
->s_itb_per_group
) ||
4712 in_range(block
+ count
- 1, ext4_inode_table(sb
, desc
),
4713 sbi
->s_itb_per_group
)) {
4714 ext4_error(sb
, "Adding blocks in system zones - "
4715 "Block = %llu, count = %lu",
4720 BUFFER_TRACE(bitmap_bh
, "getting write access");
4721 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4726 * We are about to modify some metadata. Call the journal APIs
4727 * to unshare ->b_data if a currently-committing transaction is
4730 BUFFER_TRACE(gd_bh
, "get_write_access");
4731 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4735 for (i
= 0, blocks_freed
= 0; i
< count
; i
++) {
4736 BUFFER_TRACE(bitmap_bh
, "clear bit");
4737 if (!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
)) {
4738 ext4_error(sb
, "bit already cleared for block %llu",
4739 (ext4_fsblk_t
)(block
+ i
));
4740 BUFFER_TRACE(bitmap_bh
, "bit already cleared");
4746 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4751 * need to update group_info->bb_free and bitmap
4752 * with group lock held. generate_buddy look at
4753 * them with group lock_held
4755 ext4_lock_group(sb
, block_group
);
4756 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4757 mb_free_blocks(NULL
, &e4b
, bit
, count
);
4758 blk_free_count
= blocks_freed
+ ext4_free_blks_count(sb
, desc
);
4759 ext4_free_blks_set(sb
, desc
, blk_free_count
);
4760 desc
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, desc
);
4761 ext4_unlock_group(sb
, block_group
);
4762 percpu_counter_add(&sbi
->s_freeblocks_counter
, blocks_freed
);
4764 if (sbi
->s_log_groups_per_flex
) {
4765 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4766 atomic_add(blocks_freed
,
4767 &sbi
->s_flex_groups
[flex_group
].free_blocks
);
4770 ext4_mb_unload_buddy(&e4b
);
4772 /* We dirtied the bitmap block */
4773 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4774 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4776 /* And the group descriptor block */
4777 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4778 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4784 ext4_std_error(sb
, err
);
4789 * ext4_trim_extent -- function to TRIM one single free extent in the group
4790 * @sb: super block for the file system
4791 * @start: starting block of the free extent in the alloc. group
4792 * @count: number of blocks to TRIM
4793 * @group: alloc. group we are working with
4794 * @e4b: ext4 buddy for the group
4796 * Trim "count" blocks starting at "start" in the "group". To assure that no
4797 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4798 * be called with under the group lock.
4800 static void ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
4801 ext4_group_t group
, struct ext4_buddy
*e4b
)
4803 struct ext4_free_extent ex
;
4805 trace_ext4_trim_extent(sb
, group
, start
, count
);
4807 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
4809 ex
.fe_start
= start
;
4810 ex
.fe_group
= group
;
4814 * Mark blocks used, so no one can reuse them while
4817 mb_mark_used(e4b
, &ex
);
4818 ext4_unlock_group(sb
, group
);
4819 ext4_issue_discard(sb
, group
, start
, count
);
4820 ext4_lock_group(sb
, group
);
4821 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
4825 * ext4_trim_all_free -- function to trim all free space in alloc. group
4826 * @sb: super block for file system
4827 * @group: group to be trimmed
4828 * @start: first group block to examine
4829 * @max: last group block to examine
4830 * @minblocks: minimum extent block count
4832 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4833 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4837 * ext4_trim_all_free walks through group's block bitmap searching for free
4838 * extents. When the free extent is found, mark it as used in group buddy
4839 * bitmap. Then issue a TRIM command on this extent and free the extent in
4840 * the group buddy bitmap. This is done until whole group is scanned.
4842 static ext4_grpblk_t
4843 ext4_trim_all_free(struct super_block
*sb
, ext4_group_t group
,
4844 ext4_grpblk_t start
, ext4_grpblk_t max
,
4845 ext4_grpblk_t minblocks
)
4848 ext4_grpblk_t next
, count
= 0, free_count
= 0;
4849 struct ext4_buddy e4b
;
4852 trace_ext4_trim_all_free(sb
, group
, start
, max
);
4854 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
4856 ext4_error(sb
, "Error in loading buddy "
4857 "information for %u", group
);
4860 bitmap
= e4b
.bd_bitmap
;
4862 ext4_lock_group(sb
, group
);
4863 if (EXT4_MB_GRP_WAS_TRIMMED(e4b
.bd_info
) &&
4864 minblocks
>= atomic_read(&EXT4_SB(sb
)->s_last_trim_minblks
))
4867 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
4868 e4b
.bd_info
->bb_first_free
: start
;
4870 while (start
< max
) {
4871 start
= mb_find_next_zero_bit(bitmap
, max
, start
);
4874 next
= mb_find_next_bit(bitmap
, max
, start
);
4876 if ((next
- start
) >= minblocks
) {
4877 ext4_trim_extent(sb
, start
,
4878 next
- start
, group
, &e4b
);
4879 count
+= next
- start
;
4881 free_count
+= next
- start
;
4884 if (fatal_signal_pending(current
)) {
4885 count
= -ERESTARTSYS
;
4889 if (need_resched()) {
4890 ext4_unlock_group(sb
, group
);
4892 ext4_lock_group(sb
, group
);
4895 if ((e4b
.bd_info
->bb_free
- free_count
) < minblocks
)
4900 EXT4_MB_GRP_SET_TRIMMED(e4b
.bd_info
);
4902 ext4_unlock_group(sb
, group
);
4903 ext4_mb_unload_buddy(&e4b
);
4905 ext4_debug("trimmed %d blocks in the group %d\n",
4912 * ext4_trim_fs() -- trim ioctl handle function
4913 * @sb: superblock for filesystem
4914 * @range: fstrim_range structure
4916 * start: First Byte to trim
4917 * len: number of Bytes to trim from start
4918 * minlen: minimum extent length in Bytes
4919 * ext4_trim_fs goes through all allocation groups containing Bytes from
4920 * start to start+len. For each such a group ext4_trim_all_free function
4921 * is invoked to trim all free space.
4923 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
4925 struct ext4_group_info
*grp
;
4926 ext4_group_t first_group
, last_group
;
4927 ext4_group_t group
, ngroups
= ext4_get_groups_count(sb
);
4928 ext4_grpblk_t cnt
= 0, first_block
, last_block
;
4929 uint64_t start
, len
, minlen
, trimmed
= 0;
4930 ext4_fsblk_t first_data_blk
=
4931 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
4934 start
= range
->start
>> sb
->s_blocksize_bits
;
4935 len
= range
->len
>> sb
->s_blocksize_bits
;
4936 minlen
= range
->minlen
>> sb
->s_blocksize_bits
;
4938 if (unlikely(minlen
> EXT4_BLOCKS_PER_GROUP(sb
)))
4940 if (start
+ len
<= first_data_blk
)
4942 if (start
< first_data_blk
) {
4943 len
-= first_data_blk
- start
;
4944 start
= first_data_blk
;
4947 /* Determine first and last group to examine based on start and len */
4948 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
4949 &first_group
, &first_block
);
4950 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) (start
+ len
),
4951 &last_group
, &last_block
);
4952 last_group
= (last_group
> ngroups
- 1) ? ngroups
- 1 : last_group
;
4953 last_block
= EXT4_BLOCKS_PER_GROUP(sb
);
4955 if (first_group
> last_group
)
4958 for (group
= first_group
; group
<= last_group
; group
++) {
4959 grp
= ext4_get_group_info(sb
, group
);
4960 /* We only do this if the grp has never been initialized */
4961 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
4962 ret
= ext4_mb_init_group(sb
, group
);
4968 * For all the groups except the last one, last block will
4969 * always be EXT4_BLOCKS_PER_GROUP(sb), so we only need to
4970 * change it for the last group in which case start +
4971 * len < EXT4_BLOCKS_PER_GROUP(sb).
4973 if (first_block
+ len
< EXT4_BLOCKS_PER_GROUP(sb
))
4974 last_block
= first_block
+ len
;
4975 len
-= last_block
- first_block
;
4977 if (grp
->bb_free
>= minlen
) {
4978 cnt
= ext4_trim_all_free(sb
, group
, first_block
,
4979 last_block
, minlen
);
4988 range
->len
= trimmed
* sb
->s_blocksize
;
4991 atomic_set(&EXT4_SB(sb
)->s_last_trim_minblks
, minlen
);