2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
25 #include <linux/debugfs.h>
26 #include <linux/slab.h>
27 #include <trace/events/ext4.h>
31 * - test ext4_ext_search_left() and ext4_ext_search_right()
32 * - search for metadata in few groups
35 * - normalization should take into account whether file is still open
36 * - discard preallocations if no free space left (policy?)
37 * - don't normalize tails
39 * - reservation for superuser
42 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
43 * - track min/max extents in each group for better group selection
44 * - mb_mark_used() may allocate chunk right after splitting buddy
45 * - tree of groups sorted by number of free blocks
50 * The allocation request involve request for multiple number of blocks
51 * near to the goal(block) value specified.
53 * During initialization phase of the allocator we decide to use the
54 * group preallocation or inode preallocation depending on the size of
55 * the file. The size of the file could be the resulting file size we
56 * would have after allocation, or the current file size, which ever
57 * is larger. If the size is less than sbi->s_mb_stream_request we
58 * select to use the group preallocation. The default value of
59 * s_mb_stream_request is 16 blocks. This can also be tuned via
60 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
61 * terms of number of blocks.
63 * The main motivation for having small file use group preallocation is to
64 * ensure that we have small files closer together on the disk.
66 * First stage the allocator looks at the inode prealloc list,
67 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
68 * spaces for this particular inode. The inode prealloc space is
71 * pa_lstart -> the logical start block for this prealloc space
72 * pa_pstart -> the physical start block for this prealloc space
73 * pa_len -> length for this prealloc space
74 * pa_free -> free space available in this prealloc space
76 * The inode preallocation space is used looking at the _logical_ start
77 * block. If only the logical file block falls within the range of prealloc
78 * space we will consume the particular prealloc space. This make sure that
79 * that the we have contiguous physical blocks representing the file blocks
81 * The important thing to be noted in case of inode prealloc space is that
82 * we don't modify the values associated to inode prealloc space except
85 * If we are not able to find blocks in the inode prealloc space and if we
86 * have the group allocation flag set then we look at the locality group
87 * prealloc space. These are per CPU prealloc list repreasented as
89 * ext4_sb_info.s_locality_groups[smp_processor_id()]
91 * The reason for having a per cpu locality group is to reduce the contention
92 * between CPUs. It is possible to get scheduled at this point.
94 * The locality group prealloc space is used looking at whether we have
95 * enough free space (pa_free) withing the prealloc space.
97 * If we can't allocate blocks via inode prealloc or/and locality group
98 * prealloc then we look at the buddy cache. The buddy cache is represented
99 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
100 * mapped to the buddy and bitmap information regarding different
101 * groups. The buddy information is attached to buddy cache inode so that
102 * we can access them through the page cache. The information regarding
103 * each group is loaded via ext4_mb_load_buddy. The information involve
104 * block bitmap and buddy information. The information are stored in the
108 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
111 * one block each for bitmap and buddy information. So for each group we
112 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
113 * blocksize) blocks. So it can have information regarding groups_per_page
114 * which is blocks_per_page/2
116 * The buddy cache inode is not stored on disk. The inode is thrown
117 * away when the filesystem is unmounted.
119 * We look for count number of blocks in the buddy cache. If we were able
120 * to locate that many free blocks we return with additional information
121 * regarding rest of the contiguous physical block available
123 * Before allocating blocks via buddy cache we normalize the request
124 * blocks. This ensure we ask for more blocks that we needed. The extra
125 * blocks that we get after allocation is added to the respective prealloc
126 * list. In case of inode preallocation we follow a list of heuristics
127 * based on file size. This can be found in ext4_mb_normalize_request. If
128 * we are doing a group prealloc we try to normalize the request to
129 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
130 * 512 blocks. This can be tuned via
131 * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
132 * terms of number of blocks. If we have mounted the file system with -O
133 * stripe=<value> option the group prealloc request is normalized to the
134 * stripe value (sbi->s_stripe)
136 * The regular allocator(using the buddy cache) supports few tunables.
138 * /sys/fs/ext4/<partition>/mb_min_to_scan
139 * /sys/fs/ext4/<partition>/mb_max_to_scan
140 * /sys/fs/ext4/<partition>/mb_order2_req
142 * The regular allocator uses buddy scan only if the request len is power of
143 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
144 * value of s_mb_order2_reqs can be tuned via
145 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
146 * stripe size (sbi->s_stripe), we try to search for contiguous block in
147 * stripe size. This should result in better allocation on RAID setups. If
148 * not, we search in the specific group using bitmap for best extents. The
149 * tunable min_to_scan and max_to_scan control the behaviour here.
150 * min_to_scan indicate how long the mballoc __must__ look for a best
151 * extent and max_to_scan indicates how long the mballoc __can__ look for a
152 * best extent in the found extents. Searching for the blocks starts with
153 * the group specified as the goal value in allocation context via
154 * ac_g_ex. Each group is first checked based on the criteria whether it
155 * can used for allocation. ext4_mb_good_group explains how the groups are
158 * Both the prealloc space are getting populated as above. So for the first
159 * request we will hit the buddy cache which will result in this prealloc
160 * space getting filled. The prealloc space is then later used for the
161 * subsequent request.
165 * mballoc operates on the following data:
167 * - in-core buddy (actually includes buddy and bitmap)
168 * - preallocation descriptors (PAs)
170 * there are two types of preallocations:
172 * assiged to specific inode and can be used for this inode only.
173 * it describes part of inode's space preallocated to specific
174 * physical blocks. any block from that preallocated can be used
175 * independent. the descriptor just tracks number of blocks left
176 * unused. so, before taking some block from descriptor, one must
177 * make sure corresponded logical block isn't allocated yet. this
178 * also means that freeing any block within descriptor's range
179 * must discard all preallocated blocks.
181 * assigned to specific locality group which does not translate to
182 * permanent set of inodes: inode can join and leave group. space
183 * from this type of preallocation can be used for any inode. thus
184 * it's consumed from the beginning to the end.
186 * relation between them can be expressed as:
187 * in-core buddy = on-disk bitmap + preallocation descriptors
189 * this mean blocks mballoc considers used are:
190 * - allocated blocks (persistent)
191 * - preallocated blocks (non-persistent)
193 * consistency in mballoc world means that at any time a block is either
194 * free or used in ALL structures. notice: "any time" should not be read
195 * literally -- time is discrete and delimited by locks.
197 * to keep it simple, we don't use block numbers, instead we count number of
198 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
200 * all operations can be expressed as:
201 * - init buddy: buddy = on-disk + PAs
202 * - new PA: buddy += N; PA = N
203 * - use inode PA: on-disk += N; PA -= N
204 * - discard inode PA buddy -= on-disk - PA; PA = 0
205 * - use locality group PA on-disk += N; PA -= N
206 * - discard locality group PA buddy -= PA; PA = 0
207 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
208 * is used in real operation because we can't know actual used
209 * bits from PA, only from on-disk bitmap
211 * if we follow this strict logic, then all operations above should be atomic.
212 * given some of them can block, we'd have to use something like semaphores
213 * killing performance on high-end SMP hardware. let's try to relax it using
214 * the following knowledge:
215 * 1) if buddy is referenced, it's already initialized
216 * 2) while block is used in buddy and the buddy is referenced,
217 * nobody can re-allocate that block
218 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
219 * bit set and PA claims same block, it's OK. IOW, one can set bit in
220 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
223 * so, now we're building a concurrency table:
226 * blocks for PA are allocated in the buddy, buddy must be referenced
227 * until PA is linked to allocation group to avoid concurrent buddy init
229 * we need to make sure that either on-disk bitmap or PA has uptodate data
230 * given (3) we care that PA-=N operation doesn't interfere with init
232 * the simplest way would be to have buddy initialized by the discard
233 * - use locality group PA
234 * again PA-=N must be serialized with init
235 * - discard locality group PA
236 * the simplest way would be to have buddy initialized by the discard
239 * i_data_sem serializes them
241 * discard process must wait until PA isn't used by another process
242 * - use locality group PA
243 * some mutex should serialize them
244 * - discard locality group PA
245 * discard process must wait until PA isn't used by another process
248 * i_data_sem or another mutex should serializes them
250 * discard process must wait until PA isn't used by another process
251 * - use locality group PA
252 * nothing wrong here -- they're different PAs covering different blocks
253 * - discard locality group PA
254 * discard process must wait until PA isn't used by another process
256 * now we're ready to make few consequences:
257 * - PA is referenced and while it is no discard is possible
258 * - PA is referenced until block isn't marked in on-disk bitmap
259 * - PA changes only after on-disk bitmap
260 * - discard must not compete with init. either init is done before
261 * any discard or they're serialized somehow
262 * - buddy init as sum of on-disk bitmap and PAs is done atomically
264 * a special case when we've used PA to emptiness. no need to modify buddy
265 * in this case, but we should care about concurrent init
270 * Logic in few words:
275 * mark bits in on-disk bitmap
278 * - use preallocation:
279 * find proper PA (per-inode or group)
281 * mark bits in on-disk bitmap
287 * mark bits in on-disk bitmap
290 * - discard preallocations in group:
292 * move them onto local list
293 * load on-disk bitmap
295 * remove PA from object (inode or locality group)
296 * mark free blocks in-core
298 * - discard inode's preallocations:
305 * - bitlock on a group (group)
306 * - object (inode/locality) (object)
317 * - release consumed pa:
322 * - generate in-core bitmap:
326 * - discard all for given object (inode, locality group):
331 * - discard all for given group:
338 static struct kmem_cache
*ext4_pspace_cachep
;
339 static struct kmem_cache
*ext4_ac_cachep
;
340 static struct kmem_cache
*ext4_free_ext_cachep
;
342 /* We create slab caches for groupinfo data structures based on the
343 * superblock block size. There will be one per mounted filesystem for
344 * each unique s_blocksize_bits */
345 #define NR_GRPINFO_CACHES 8
346 static struct kmem_cache
*ext4_groupinfo_caches
[NR_GRPINFO_CACHES
];
348 static const char *ext4_groupinfo_slab_names
[NR_GRPINFO_CACHES
] = {
349 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
350 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
351 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
354 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
356 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
358 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
);
360 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
362 #if BITS_PER_LONG == 64
363 *bit
+= ((unsigned long) addr
& 7UL) << 3;
364 addr
= (void *) ((unsigned long) addr
& ~7UL);
365 #elif BITS_PER_LONG == 32
366 *bit
+= ((unsigned long) addr
& 3UL) << 3;
367 addr
= (void *) ((unsigned long) addr
& ~3UL);
369 #error "how many bits you are?!"
374 static inline int mb_test_bit(int bit
, void *addr
)
377 * ext4_test_bit on architecture like powerpc
378 * needs unsigned long aligned address
380 addr
= mb_correct_addr_and_bit(&bit
, addr
);
381 return ext4_test_bit(bit
, addr
);
384 static inline void mb_set_bit(int bit
, void *addr
)
386 addr
= mb_correct_addr_and_bit(&bit
, addr
);
387 ext4_set_bit(bit
, addr
);
390 static inline void mb_clear_bit(int bit
, void *addr
)
392 addr
= mb_correct_addr_and_bit(&bit
, addr
);
393 ext4_clear_bit(bit
, addr
);
396 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
398 int fix
= 0, ret
, tmpmax
;
399 addr
= mb_correct_addr_and_bit(&fix
, addr
);
403 ret
= ext4_find_next_zero_bit(addr
, tmpmax
, start
) - fix
;
409 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
411 int fix
= 0, ret
, tmpmax
;
412 addr
= mb_correct_addr_and_bit(&fix
, addr
);
416 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
422 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
426 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
429 if (order
> e4b
->bd_blkbits
+ 1) {
434 /* at order 0 we see each particular block */
435 *max
= 1 << (e4b
->bd_blkbits
+ 3);
437 return EXT4_MB_BITMAP(e4b
);
439 bb
= EXT4_MB_BUDDY(e4b
) + EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
440 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
446 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
447 int first
, int count
)
450 struct super_block
*sb
= e4b
->bd_sb
;
452 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
454 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
455 for (i
= 0; i
< count
; i
++) {
456 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
457 ext4_fsblk_t blocknr
;
459 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
460 blocknr
+= first
+ i
;
461 ext4_grp_locked_error(sb
, e4b
->bd_group
,
462 inode
? inode
->i_ino
: 0,
464 "freeing block already freed "
468 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
472 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
476 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
478 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
479 for (i
= 0; i
< count
; i
++) {
480 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
481 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
485 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
487 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
488 unsigned char *b1
, *b2
;
490 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
491 b2
= (unsigned char *) bitmap
;
492 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
493 if (b1
[i
] != b2
[i
]) {
494 printk(KERN_ERR
"corruption in group %u "
495 "at byte %u(%u): %x in copy != %x "
496 "on disk/prealloc\n",
497 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
505 static inline void mb_free_blocks_double(struct inode
*inode
,
506 struct ext4_buddy
*e4b
, int first
, int count
)
510 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
511 int first
, int count
)
515 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
521 #ifdef AGGRESSIVE_CHECK
523 #define MB_CHECK_ASSERT(assert) \
527 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
528 function, file, line, # assert); \
533 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
534 const char *function
, int line
)
536 struct super_block
*sb
= e4b
->bd_sb
;
537 int order
= e4b
->bd_blkbits
+ 1;
544 struct ext4_group_info
*grp
;
547 struct list_head
*cur
;
552 static int mb_check_counter
;
553 if (mb_check_counter
++ % 100 != 0)
558 buddy
= mb_find_buddy(e4b
, order
, &max
);
559 MB_CHECK_ASSERT(buddy
);
560 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
561 MB_CHECK_ASSERT(buddy2
);
562 MB_CHECK_ASSERT(buddy
!= buddy2
);
563 MB_CHECK_ASSERT(max
* 2 == max2
);
566 for (i
= 0; i
< max
; i
++) {
568 if (mb_test_bit(i
, buddy
)) {
569 /* only single bit in buddy2 may be 1 */
570 if (!mb_test_bit(i
<< 1, buddy2
)) {
572 mb_test_bit((i
<<1)+1, buddy2
));
573 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
575 mb_test_bit(i
<< 1, buddy2
));
580 /* both bits in buddy2 must be 0 */
581 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
582 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
584 for (j
= 0; j
< (1 << order
); j
++) {
585 k
= (i
* (1 << order
)) + j
;
587 !mb_test_bit(k
, EXT4_MB_BITMAP(e4b
)));
591 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
596 buddy
= mb_find_buddy(e4b
, 0, &max
);
597 for (i
= 0; i
< max
; i
++) {
598 if (!mb_test_bit(i
, buddy
)) {
599 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
607 /* check used bits only */
608 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
609 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
611 MB_CHECK_ASSERT(k
< max2
);
612 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
615 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
616 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
618 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
619 buddy
= mb_find_buddy(e4b
, 0, &max
);
620 list_for_each(cur
, &grp
->bb_prealloc_list
) {
621 ext4_group_t groupnr
;
622 struct ext4_prealloc_space
*pa
;
623 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
624 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
625 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
626 for (i
= 0; i
< pa
->pa_len
; i
++)
627 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
631 #undef MB_CHECK_ASSERT
632 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
633 __FILE__, __func__, __LINE__)
635 #define mb_check_buddy(e4b)
638 /* FIXME!! need more doc */
639 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
640 void *buddy
, ext4_grpblk_t first
, ext4_grpblk_t len
,
641 struct ext4_group_info
*grp
)
643 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
647 unsigned short border
;
649 BUG_ON(len
> EXT4_BLOCKS_PER_GROUP(sb
));
651 border
= 2 << sb
->s_blocksize_bits
;
654 /* find how many blocks can be covered since this position */
655 max
= ffs(first
| border
) - 1;
657 /* find how many blocks of power 2 we need to mark */
664 /* mark multiblock chunks only */
665 grp
->bb_counters
[min
]++;
667 mb_clear_bit(first
>> min
,
668 buddy
+ sbi
->s_mb_offsets
[min
]);
676 * Cache the order of the largest free extent we have available in this block
680 mb_set_largest_free_order(struct super_block
*sb
, struct ext4_group_info
*grp
)
685 grp
->bb_largest_free_order
= -1; /* uninit */
687 bits
= sb
->s_blocksize_bits
+ 1;
688 for (i
= bits
; i
>= 0; i
--) {
689 if (grp
->bb_counters
[i
] > 0) {
690 grp
->bb_largest_free_order
= i
;
696 static noinline_for_stack
697 void ext4_mb_generate_buddy(struct super_block
*sb
,
698 void *buddy
, void *bitmap
, ext4_group_t group
)
700 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
701 ext4_grpblk_t max
= EXT4_BLOCKS_PER_GROUP(sb
);
706 unsigned fragments
= 0;
707 unsigned long long period
= get_cycles();
709 /* initialize buddy from bitmap which is aggregation
710 * of on-disk bitmap and preallocations */
711 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
712 grp
->bb_first_free
= i
;
716 i
= mb_find_next_bit(bitmap
, max
, i
);
720 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
722 grp
->bb_counters
[0]++;
724 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
726 grp
->bb_fragments
= fragments
;
728 if (free
!= grp
->bb_free
) {
729 ext4_grp_locked_error(sb
, group
, 0, 0,
730 "%u blocks in bitmap, %u in gd",
733 * If we intent to continue, we consider group descritor
734 * corrupt and update bb_free using bitmap value
738 mb_set_largest_free_order(sb
, grp
);
740 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
742 period
= get_cycles() - period
;
743 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
744 EXT4_SB(sb
)->s_mb_buddies_generated
++;
745 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
746 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
749 /* The buddy information is attached the buddy cache inode
750 * for convenience. The information regarding each group
751 * is loaded via ext4_mb_load_buddy. The information involve
752 * block bitmap and buddy information. The information are
753 * stored in the inode as
756 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
759 * one block each for bitmap and buddy information.
760 * So for each group we take up 2 blocks. A page can
761 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
762 * So it can have information regarding groups_per_page which
763 * is blocks_per_page/2
765 * Locking note: This routine takes the block group lock of all groups
766 * for this page; do not hold this lock when calling this routine!
769 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
771 ext4_group_t ngroups
;
777 ext4_group_t first_group
;
779 struct super_block
*sb
;
780 struct buffer_head
*bhs
;
781 struct buffer_head
**bh
;
786 mb_debug(1, "init page %lu\n", page
->index
);
788 inode
= page
->mapping
->host
;
790 ngroups
= ext4_get_groups_count(sb
);
791 blocksize
= 1 << inode
->i_blkbits
;
792 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
794 groups_per_page
= blocks_per_page
>> 1;
795 if (groups_per_page
== 0)
798 /* allocate buffer_heads to read bitmaps */
799 if (groups_per_page
> 1) {
801 i
= sizeof(struct buffer_head
*) * groups_per_page
;
802 bh
= kzalloc(i
, GFP_NOFS
);
808 first_group
= page
->index
* blocks_per_page
/ 2;
810 /* read all groups the page covers into the cache */
811 for (i
= 0; i
< groups_per_page
; i
++) {
812 struct ext4_group_desc
*desc
;
814 if (first_group
+ i
>= ngroups
)
818 desc
= ext4_get_group_desc(sb
, first_group
+ i
, NULL
);
823 bh
[i
] = sb_getblk(sb
, ext4_block_bitmap(sb
, desc
));
827 if (bitmap_uptodate(bh
[i
]))
831 if (bitmap_uptodate(bh
[i
])) {
832 unlock_buffer(bh
[i
]);
835 ext4_lock_group(sb
, first_group
+ i
);
836 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
837 ext4_init_block_bitmap(sb
, bh
[i
],
838 first_group
+ i
, desc
);
839 set_bitmap_uptodate(bh
[i
]);
840 set_buffer_uptodate(bh
[i
]);
841 ext4_unlock_group(sb
, first_group
+ i
);
842 unlock_buffer(bh
[i
]);
845 ext4_unlock_group(sb
, first_group
+ i
);
846 if (buffer_uptodate(bh
[i
])) {
848 * if not uninit if bh is uptodate,
849 * bitmap is also uptodate
851 set_bitmap_uptodate(bh
[i
]);
852 unlock_buffer(bh
[i
]);
857 * submit the buffer_head for read. We can
858 * safely mark the bitmap as uptodate now.
859 * We do it here so the bitmap uptodate bit
860 * get set with buffer lock held.
862 set_bitmap_uptodate(bh
[i
]);
863 bh
[i
]->b_end_io
= end_buffer_read_sync
;
864 submit_bh(READ
, bh
[i
]);
865 mb_debug(1, "read bitmap for group %u\n", first_group
+ i
);
868 /* wait for I/O completion */
869 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
870 wait_on_buffer(bh
[i
]);
873 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
874 if (!buffer_uptodate(bh
[i
]))
878 first_block
= page
->index
* blocks_per_page
;
880 memset(page_address(page
), 0xff, PAGE_CACHE_SIZE
);
881 for (i
= 0; i
< blocks_per_page
; i
++) {
883 struct ext4_group_info
*grinfo
;
885 group
= (first_block
+ i
) >> 1;
886 if (group
>= ngroups
)
890 * data carry information regarding this
891 * particular group in the format specified
895 data
= page_address(page
) + (i
* blocksize
);
896 bitmap
= bh
[group
- first_group
]->b_data
;
899 * We place the buddy block and bitmap block
902 if ((first_block
+ i
) & 1) {
903 /* this is block of buddy */
904 BUG_ON(incore
== NULL
);
905 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
906 group
, page
->index
, i
* blocksize
);
907 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
908 grinfo
= ext4_get_group_info(sb
, group
);
909 grinfo
->bb_fragments
= 0;
910 memset(grinfo
->bb_counters
, 0,
911 sizeof(*grinfo
->bb_counters
) *
912 (sb
->s_blocksize_bits
+2));
914 * incore got set to the group block bitmap below
916 ext4_lock_group(sb
, group
);
917 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
918 ext4_unlock_group(sb
, group
);
921 /* this is block of bitmap */
922 BUG_ON(incore
!= NULL
);
923 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
924 group
, page
->index
, i
* blocksize
);
925 trace_ext4_mb_bitmap_load(sb
, group
);
927 /* see comments in ext4_mb_put_pa() */
928 ext4_lock_group(sb
, group
);
929 memcpy(data
, bitmap
, blocksize
);
931 /* mark all preallocated blks used in in-core bitmap */
932 ext4_mb_generate_from_pa(sb
, data
, group
);
933 ext4_mb_generate_from_freelist(sb
, data
, group
);
934 ext4_unlock_group(sb
, group
);
936 /* set incore so that the buddy information can be
937 * generated using this
942 SetPageUptodate(page
);
946 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
955 * lock the group_info alloc_sem of all the groups
956 * belonging to the same buddy cache page. This
957 * make sure other parallel operation on the buddy
958 * cache doesn't happen whild holding the buddy cache
961 static int ext4_mb_get_buddy_cache_lock(struct super_block
*sb
,
968 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
969 ext4_group_t first_group
;
970 struct ext4_group_info
*grp
;
972 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
974 * the buddy cache inode stores the block bitmap
975 * and buddy information in consecutive blocks.
976 * So for each group we need two blocks.
979 pnum
= block
/ blocks_per_page
;
980 first_group
= pnum
* blocks_per_page
/ 2;
982 groups_per_page
= blocks_per_page
>> 1;
983 if (groups_per_page
== 0)
985 /* read all groups the page covers into the cache */
986 for (i
= 0; i
< groups_per_page
; i
++) {
988 if ((first_group
+ i
) >= ngroups
)
990 grp
= ext4_get_group_info(sb
, first_group
+ i
);
991 /* take all groups write allocation
992 * semaphore. This make sure there is
993 * no block allocation going on in any
996 down_write_nested(&grp
->alloc_sem
, i
);
1001 static void ext4_mb_put_buddy_cache_lock(struct super_block
*sb
,
1002 ext4_group_t group
, int locked_group
)
1006 int blocks_per_page
;
1007 ext4_group_t first_group
;
1008 struct ext4_group_info
*grp
;
1010 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1012 * the buddy cache inode stores the block bitmap
1013 * and buddy information in consecutive blocks.
1014 * So for each group we need two blocks.
1017 pnum
= block
/ blocks_per_page
;
1018 first_group
= pnum
* blocks_per_page
/ 2;
1019 /* release locks on all the groups */
1020 for (i
= 0; i
< locked_group
; i
++) {
1022 grp
= ext4_get_group_info(sb
, first_group
+ i
);
1023 /* take all groups write allocation
1024 * semaphore. This make sure there is
1025 * no block allocation going on in any
1028 up_write(&grp
->alloc_sem
);
1034 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1035 * block group lock of all groups for this page; do not hold the BG lock when
1036 * calling this routine!
1038 static noinline_for_stack
1039 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
)
1044 int blocks_per_page
;
1045 int block
, pnum
, poff
;
1046 int num_grp_locked
= 0;
1047 struct ext4_group_info
*this_grp
;
1048 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1049 struct inode
*inode
= sbi
->s_buddy_cache
;
1050 struct page
*page
= NULL
, *bitmap_page
= NULL
;
1052 mb_debug(1, "init group %u\n", group
);
1053 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1054 this_grp
= ext4_get_group_info(sb
, group
);
1056 * This ensures that we don't reinit the buddy cache
1057 * page which map to the group from which we are already
1058 * allocating. If we are looking at the buddy cache we would
1059 * have taken a reference using ext4_mb_load_buddy and that
1060 * would have taken the alloc_sem lock.
1062 num_grp_locked
= ext4_mb_get_buddy_cache_lock(sb
, group
);
1063 if (!EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1065 * somebody initialized the group
1066 * return without doing anything
1072 * the buddy cache inode stores the block bitmap
1073 * and buddy information in consecutive blocks.
1074 * So for each group we need two blocks.
1077 pnum
= block
/ blocks_per_page
;
1078 poff
= block
% blocks_per_page
;
1079 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1081 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1082 ret
= ext4_mb_init_cache(page
, NULL
);
1089 if (page
== NULL
|| !PageUptodate(page
)) {
1093 mark_page_accessed(page
);
1095 bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1097 /* init buddy cache */
1099 pnum
= block
/ blocks_per_page
;
1100 poff
= block
% blocks_per_page
;
1101 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1102 if (page
== bitmap_page
) {
1104 * If both the bitmap and buddy are in
1105 * the same page we don't need to force
1110 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1111 ret
= ext4_mb_init_cache(page
, bitmap
);
1118 if (page
== NULL
|| !PageUptodate(page
)) {
1122 mark_page_accessed(page
);
1124 ext4_mb_put_buddy_cache_lock(sb
, group
, num_grp_locked
);
1126 page_cache_release(bitmap_page
);
1128 page_cache_release(page
);
1133 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1134 * block group lock of all groups for this page; do not hold the BG lock when
1135 * calling this routine!
1137 static noinline_for_stack
int
1138 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1139 struct ext4_buddy
*e4b
)
1141 int blocks_per_page
;
1147 struct ext4_group_info
*grp
;
1148 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1149 struct inode
*inode
= sbi
->s_buddy_cache
;
1151 mb_debug(1, "load group %u\n", group
);
1153 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1154 grp
= ext4_get_group_info(sb
, group
);
1156 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1157 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
1159 e4b
->bd_group
= group
;
1160 e4b
->bd_buddy_page
= NULL
;
1161 e4b
->bd_bitmap_page
= NULL
;
1162 e4b
->alloc_semp
= &grp
->alloc_sem
;
1164 /* Take the read lock on the group alloc
1165 * sem. This would make sure a parallel
1166 * ext4_mb_init_group happening on other
1167 * groups mapped by the page is blocked
1168 * till we are done with allocation
1171 down_read(e4b
->alloc_semp
);
1173 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1174 /* we need to check for group need init flag
1175 * with alloc_semp held so that we can be sure
1176 * that new blocks didn't get added to the group
1177 * when we are loading the buddy cache
1179 up_read(e4b
->alloc_semp
);
1181 * we need full data about the group
1182 * to make a good selection
1184 ret
= ext4_mb_init_group(sb
, group
);
1187 goto repeat_load_buddy
;
1191 * the buddy cache inode stores the block bitmap
1192 * and buddy information in consecutive blocks.
1193 * So for each group we need two blocks.
1196 pnum
= block
/ blocks_per_page
;
1197 poff
= block
% blocks_per_page
;
1199 /* we could use find_or_create_page(), but it locks page
1200 * what we'd like to avoid in fast path ... */
1201 page
= find_get_page(inode
->i_mapping
, pnum
);
1202 if (page
== NULL
|| !PageUptodate(page
)) {
1205 * drop the page reference and try
1206 * to get the page with lock. If we
1207 * are not uptodate that implies
1208 * somebody just created the page but
1209 * is yet to initialize the same. So
1210 * wait for it to initialize.
1212 page_cache_release(page
);
1213 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1215 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1216 if (!PageUptodate(page
)) {
1217 ret
= ext4_mb_init_cache(page
, NULL
);
1222 mb_cmp_bitmaps(e4b
, page_address(page
) +
1223 (poff
* sb
->s_blocksize
));
1228 if (page
== NULL
|| !PageUptodate(page
)) {
1232 e4b
->bd_bitmap_page
= page
;
1233 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1234 mark_page_accessed(page
);
1237 pnum
= block
/ blocks_per_page
;
1238 poff
= block
% blocks_per_page
;
1240 page
= find_get_page(inode
->i_mapping
, pnum
);
1241 if (page
== NULL
|| !PageUptodate(page
)) {
1243 page_cache_release(page
);
1244 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1246 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1247 if (!PageUptodate(page
)) {
1248 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1257 if (page
== NULL
|| !PageUptodate(page
)) {
1261 e4b
->bd_buddy_page
= page
;
1262 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1263 mark_page_accessed(page
);
1265 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1266 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1272 page_cache_release(page
);
1273 if (e4b
->bd_bitmap_page
)
1274 page_cache_release(e4b
->bd_bitmap_page
);
1275 if (e4b
->bd_buddy_page
)
1276 page_cache_release(e4b
->bd_buddy_page
);
1277 e4b
->bd_buddy
= NULL
;
1278 e4b
->bd_bitmap
= NULL
;
1280 /* Done with the buddy cache */
1281 up_read(e4b
->alloc_semp
);
1285 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1287 if (e4b
->bd_bitmap_page
)
1288 page_cache_release(e4b
->bd_bitmap_page
);
1289 if (e4b
->bd_buddy_page
)
1290 page_cache_release(e4b
->bd_buddy_page
);
1291 /* Done with the buddy cache */
1292 if (e4b
->alloc_semp
)
1293 up_read(e4b
->alloc_semp
);
1297 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1302 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1303 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1305 bb
= EXT4_MB_BUDDY(e4b
);
1306 while (order
<= e4b
->bd_blkbits
+ 1) {
1308 if (!mb_test_bit(block
, bb
)) {
1309 /* this block is part of buddy of order 'order' */
1312 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1318 static void mb_clear_bits(void *bm
, int cur
, int len
)
1324 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1325 /* fast path: clear whole word at once */
1326 addr
= bm
+ (cur
>> 3);
1331 mb_clear_bit(cur
, bm
);
1336 static void mb_set_bits(void *bm
, int cur
, int len
)
1342 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1343 /* fast path: set whole word at once */
1344 addr
= bm
+ (cur
>> 3);
1349 mb_set_bit(cur
, bm
);
1354 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1355 int first
, int count
)
1362 struct super_block
*sb
= e4b
->bd_sb
;
1364 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1365 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1366 mb_check_buddy(e4b
);
1367 mb_free_blocks_double(inode
, e4b
, first
, count
);
1369 e4b
->bd_info
->bb_free
+= count
;
1370 if (first
< e4b
->bd_info
->bb_first_free
)
1371 e4b
->bd_info
->bb_first_free
= first
;
1373 /* let's maintain fragments counter */
1375 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1376 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1377 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1379 e4b
->bd_info
->bb_fragments
--;
1380 else if (!block
&& !max
)
1381 e4b
->bd_info
->bb_fragments
++;
1383 /* let's maintain buddy itself */
1384 while (count
-- > 0) {
1388 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1389 ext4_fsblk_t blocknr
;
1391 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1393 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1394 inode
? inode
->i_ino
: 0,
1396 "freeing already freed block "
1399 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1400 e4b
->bd_info
->bb_counters
[order
]++;
1402 /* start of the buddy */
1403 buddy
= mb_find_buddy(e4b
, order
, &max
);
1407 if (mb_test_bit(block
, buddy
) ||
1408 mb_test_bit(block
+ 1, buddy
))
1411 /* both the buddies are free, try to coalesce them */
1412 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1418 /* for special purposes, we don't set
1419 * free bits in bitmap */
1420 mb_set_bit(block
, buddy
);
1421 mb_set_bit(block
+ 1, buddy
);
1423 e4b
->bd_info
->bb_counters
[order
]--;
1424 e4b
->bd_info
->bb_counters
[order
]--;
1428 e4b
->bd_info
->bb_counters
[order
]++;
1430 mb_clear_bit(block
, buddy2
);
1434 mb_set_largest_free_order(sb
, e4b
->bd_info
);
1435 mb_check_buddy(e4b
);
1438 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1439 int needed
, struct ext4_free_extent
*ex
)
1446 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1449 buddy
= mb_find_buddy(e4b
, order
, &max
);
1450 BUG_ON(buddy
== NULL
);
1451 BUG_ON(block
>= max
);
1452 if (mb_test_bit(block
, buddy
)) {
1459 /* FIXME dorp order completely ? */
1460 if (likely(order
== 0)) {
1461 /* find actual order */
1462 order
= mb_find_order_for_block(e4b
, block
);
1463 block
= block
>> order
;
1466 ex
->fe_len
= 1 << order
;
1467 ex
->fe_start
= block
<< order
;
1468 ex
->fe_group
= e4b
->bd_group
;
1470 /* calc difference from given start */
1471 next
= next
- ex
->fe_start
;
1473 ex
->fe_start
+= next
;
1475 while (needed
> ex
->fe_len
&&
1476 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1478 if (block
+ 1 >= max
)
1481 next
= (block
+ 1) * (1 << order
);
1482 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1485 ord
= mb_find_order_for_block(e4b
, next
);
1488 block
= next
>> order
;
1489 ex
->fe_len
+= 1 << order
;
1492 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1496 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1502 int start
= ex
->fe_start
;
1503 int len
= ex
->fe_len
;
1508 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1509 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1510 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1511 mb_check_buddy(e4b
);
1512 mb_mark_used_double(e4b
, start
, len
);
1514 e4b
->bd_info
->bb_free
-= len
;
1515 if (e4b
->bd_info
->bb_first_free
== start
)
1516 e4b
->bd_info
->bb_first_free
+= len
;
1518 /* let's maintain fragments counter */
1520 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1521 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1522 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1524 e4b
->bd_info
->bb_fragments
++;
1525 else if (!mlen
&& !max
)
1526 e4b
->bd_info
->bb_fragments
--;
1528 /* let's maintain buddy itself */
1530 ord
= mb_find_order_for_block(e4b
, start
);
1532 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1533 /* the whole chunk may be allocated at once! */
1535 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1536 BUG_ON((start
>> ord
) >= max
);
1537 mb_set_bit(start
>> ord
, buddy
);
1538 e4b
->bd_info
->bb_counters
[ord
]--;
1545 /* store for history */
1547 ret
= len
| (ord
<< 16);
1549 /* we have to split large buddy */
1551 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1552 mb_set_bit(start
>> ord
, buddy
);
1553 e4b
->bd_info
->bb_counters
[ord
]--;
1556 cur
= (start
>> ord
) & ~1U;
1557 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1558 mb_clear_bit(cur
, buddy
);
1559 mb_clear_bit(cur
+ 1, buddy
);
1560 e4b
->bd_info
->bb_counters
[ord
]++;
1561 e4b
->bd_info
->bb_counters
[ord
]++;
1563 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1565 mb_set_bits(EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1566 mb_check_buddy(e4b
);
1572 * Must be called under group lock!
1574 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1575 struct ext4_buddy
*e4b
)
1577 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1580 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1581 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1583 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1584 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1585 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1587 /* preallocation can change ac_b_ex, thus we store actually
1588 * allocated blocks for history */
1589 ac
->ac_f_ex
= ac
->ac_b_ex
;
1591 ac
->ac_status
= AC_STATUS_FOUND
;
1592 ac
->ac_tail
= ret
& 0xffff;
1593 ac
->ac_buddy
= ret
>> 16;
1596 * take the page reference. We want the page to be pinned
1597 * so that we don't get a ext4_mb_init_cache_call for this
1598 * group until we update the bitmap. That would mean we
1599 * double allocate blocks. The reference is dropped
1600 * in ext4_mb_release_context
1602 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1603 get_page(ac
->ac_bitmap_page
);
1604 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1605 get_page(ac
->ac_buddy_page
);
1606 /* on allocation we use ac to track the held semaphore */
1607 ac
->alloc_semp
= e4b
->alloc_semp
;
1608 e4b
->alloc_semp
= NULL
;
1609 /* store last allocated for subsequent stream allocation */
1610 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1611 spin_lock(&sbi
->s_md_lock
);
1612 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1613 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1614 spin_unlock(&sbi
->s_md_lock
);
1619 * regular allocator, for general purposes allocation
1622 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1623 struct ext4_buddy
*e4b
,
1626 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1627 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1628 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1629 struct ext4_free_extent ex
;
1632 if (ac
->ac_status
== AC_STATUS_FOUND
)
1635 * We don't want to scan for a whole year
1637 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1638 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1639 ac
->ac_status
= AC_STATUS_BREAK
;
1644 * Haven't found good chunk so far, let's continue
1646 if (bex
->fe_len
< gex
->fe_len
)
1649 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1650 && bex
->fe_group
== e4b
->bd_group
) {
1651 /* recheck chunk's availability - we don't know
1652 * when it was found (within this lock-unlock
1654 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1655 if (max
>= gex
->fe_len
) {
1656 ext4_mb_use_best_found(ac
, e4b
);
1663 * The routine checks whether found extent is good enough. If it is,
1664 * then the extent gets marked used and flag is set to the context
1665 * to stop scanning. Otherwise, the extent is compared with the
1666 * previous found extent and if new one is better, then it's stored
1667 * in the context. Later, the best found extent will be used, if
1668 * mballoc can't find good enough extent.
1670 * FIXME: real allocation policy is to be designed yet!
1672 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1673 struct ext4_free_extent
*ex
,
1674 struct ext4_buddy
*e4b
)
1676 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1677 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1679 BUG_ON(ex
->fe_len
<= 0);
1680 BUG_ON(ex
->fe_len
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1681 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1682 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1687 * The special case - take what you catch first
1689 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1691 ext4_mb_use_best_found(ac
, e4b
);
1696 * Let's check whether the chuck is good enough
1698 if (ex
->fe_len
== gex
->fe_len
) {
1700 ext4_mb_use_best_found(ac
, e4b
);
1705 * If this is first found extent, just store it in the context
1707 if (bex
->fe_len
== 0) {
1713 * If new found extent is better, store it in the context
1715 if (bex
->fe_len
< gex
->fe_len
) {
1716 /* if the request isn't satisfied, any found extent
1717 * larger than previous best one is better */
1718 if (ex
->fe_len
> bex
->fe_len
)
1720 } else if (ex
->fe_len
> gex
->fe_len
) {
1721 /* if the request is satisfied, then we try to find
1722 * an extent that still satisfy the request, but is
1723 * smaller than previous one */
1724 if (ex
->fe_len
< bex
->fe_len
)
1728 ext4_mb_check_limits(ac
, e4b
, 0);
1731 static noinline_for_stack
1732 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1733 struct ext4_buddy
*e4b
)
1735 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1736 ext4_group_t group
= ex
.fe_group
;
1740 BUG_ON(ex
.fe_len
<= 0);
1741 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1745 ext4_lock_group(ac
->ac_sb
, group
);
1746 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1750 ext4_mb_use_best_found(ac
, e4b
);
1753 ext4_unlock_group(ac
->ac_sb
, group
);
1754 ext4_mb_unload_buddy(e4b
);
1759 static noinline_for_stack
1760 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1761 struct ext4_buddy
*e4b
)
1763 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1766 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1767 struct ext4_free_extent ex
;
1769 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1772 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1776 ext4_lock_group(ac
->ac_sb
, group
);
1777 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1778 ac
->ac_g_ex
.fe_len
, &ex
);
1780 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1783 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1785 /* use do_div to get remainder (would be 64-bit modulo) */
1786 if (do_div(start
, sbi
->s_stripe
) == 0) {
1789 ext4_mb_use_best_found(ac
, e4b
);
1791 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1792 BUG_ON(ex
.fe_len
<= 0);
1793 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1794 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1797 ext4_mb_use_best_found(ac
, e4b
);
1798 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1799 /* Sometimes, caller may want to merge even small
1800 * number of blocks to an existing extent */
1801 BUG_ON(ex
.fe_len
<= 0);
1802 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1803 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1806 ext4_mb_use_best_found(ac
, e4b
);
1808 ext4_unlock_group(ac
->ac_sb
, group
);
1809 ext4_mb_unload_buddy(e4b
);
1815 * The routine scans buddy structures (not bitmap!) from given order
1816 * to max order and tries to find big enough chunk to satisfy the req
1818 static noinline_for_stack
1819 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1820 struct ext4_buddy
*e4b
)
1822 struct super_block
*sb
= ac
->ac_sb
;
1823 struct ext4_group_info
*grp
= e4b
->bd_info
;
1829 BUG_ON(ac
->ac_2order
<= 0);
1830 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1831 if (grp
->bb_counters
[i
] == 0)
1834 buddy
= mb_find_buddy(e4b
, i
, &max
);
1835 BUG_ON(buddy
== NULL
);
1837 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1842 ac
->ac_b_ex
.fe_len
= 1 << i
;
1843 ac
->ac_b_ex
.fe_start
= k
<< i
;
1844 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1846 ext4_mb_use_best_found(ac
, e4b
);
1848 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1850 if (EXT4_SB(sb
)->s_mb_stats
)
1851 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1858 * The routine scans the group and measures all found extents.
1859 * In order to optimize scanning, caller must pass number of
1860 * free blocks in the group, so the routine can know upper limit.
1862 static noinline_for_stack
1863 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1864 struct ext4_buddy
*e4b
)
1866 struct super_block
*sb
= ac
->ac_sb
;
1867 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1868 struct ext4_free_extent ex
;
1872 free
= e4b
->bd_info
->bb_free
;
1875 i
= e4b
->bd_info
->bb_first_free
;
1877 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1878 i
= mb_find_next_zero_bit(bitmap
,
1879 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1880 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1882 * IF we have corrupt bitmap, we won't find any
1883 * free blocks even though group info says we
1884 * we have free blocks
1886 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1887 "%d free blocks as per "
1888 "group info. But bitmap says 0",
1893 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1894 BUG_ON(ex
.fe_len
<= 0);
1895 if (free
< ex
.fe_len
) {
1896 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1897 "%d free blocks as per "
1898 "group info. But got %d blocks",
1901 * The number of free blocks differs. This mostly
1902 * indicate that the bitmap is corrupt. So exit
1903 * without claiming the space.
1908 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1914 ext4_mb_check_limits(ac
, e4b
, 1);
1918 * This is a special case for storages like raid5
1919 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1921 static noinline_for_stack
1922 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1923 struct ext4_buddy
*e4b
)
1925 struct super_block
*sb
= ac
->ac_sb
;
1926 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1927 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1928 struct ext4_free_extent ex
;
1929 ext4_fsblk_t first_group_block
;
1934 BUG_ON(sbi
->s_stripe
== 0);
1936 /* find first stripe-aligned block in group */
1937 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1939 a
= first_group_block
+ sbi
->s_stripe
- 1;
1940 do_div(a
, sbi
->s_stripe
);
1941 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1943 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1944 if (!mb_test_bit(i
, bitmap
)) {
1945 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1946 if (max
>= sbi
->s_stripe
) {
1949 ext4_mb_use_best_found(ac
, e4b
);
1957 /* This is now called BEFORE we load the buddy bitmap. */
1958 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1959 ext4_group_t group
, int cr
)
1961 unsigned free
, fragments
;
1962 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
1963 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1965 BUG_ON(cr
< 0 || cr
>= 4);
1967 /* We only do this if the grp has never been initialized */
1968 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1969 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
);
1974 free
= grp
->bb_free
;
1975 fragments
= grp
->bb_fragments
;
1983 BUG_ON(ac
->ac_2order
== 0);
1985 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
1988 /* Avoid using the first bg of a flexgroup for data files */
1989 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
1990 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
1991 ((group
% flex_size
) == 0))
1996 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2000 if (free
>= ac
->ac_g_ex
.fe_len
)
2012 static noinline_for_stack
int
2013 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
2015 ext4_group_t ngroups
, group
, i
;
2018 struct ext4_sb_info
*sbi
;
2019 struct super_block
*sb
;
2020 struct ext4_buddy e4b
;
2024 ngroups
= ext4_get_groups_count(sb
);
2025 /* non-extent files are limited to low blocks/groups */
2026 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
2027 ngroups
= sbi
->s_blockfile_groups
;
2029 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
2031 /* first, try the goal */
2032 err
= ext4_mb_find_by_goal(ac
, &e4b
);
2033 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
2036 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2040 * ac->ac2_order is set only if the fe_len is a power of 2
2041 * if ac2_order is set we also set criteria to 0 so that we
2042 * try exact allocation using buddy.
2044 i
= fls(ac
->ac_g_ex
.fe_len
);
2047 * We search using buddy data only if the order of the request
2048 * is greater than equal to the sbi_s_mb_order2_reqs
2049 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2051 if (i
>= sbi
->s_mb_order2_reqs
) {
2053 * This should tell if fe_len is exactly power of 2
2055 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2056 ac
->ac_2order
= i
- 1;
2059 /* if stream allocation is enabled, use global goal */
2060 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2061 /* TBD: may be hot point */
2062 spin_lock(&sbi
->s_md_lock
);
2063 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2064 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2065 spin_unlock(&sbi
->s_md_lock
);
2068 /* Let's just scan groups to find more-less suitable blocks */
2069 cr
= ac
->ac_2order
? 0 : 1;
2071 * cr == 0 try to get exact allocation,
2072 * cr == 3 try to get anything
2075 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2076 ac
->ac_criteria
= cr
;
2078 * searching for the right group start
2079 * from the goal value specified
2081 group
= ac
->ac_g_ex
.fe_group
;
2083 for (i
= 0; i
< ngroups
; group
++, i
++) {
2084 if (group
== ngroups
)
2087 /* This now checks without needing the buddy page */
2088 if (!ext4_mb_good_group(ac
, group
, cr
))
2091 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2095 ext4_lock_group(sb
, group
);
2098 * We need to check again after locking the
2101 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2102 ext4_unlock_group(sb
, group
);
2103 ext4_mb_unload_buddy(&e4b
);
2107 ac
->ac_groups_scanned
++;
2109 ext4_mb_simple_scan_group(ac
, &e4b
);
2110 else if (cr
== 1 && sbi
->s_stripe
&&
2111 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2112 ext4_mb_scan_aligned(ac
, &e4b
);
2114 ext4_mb_complex_scan_group(ac
, &e4b
);
2116 ext4_unlock_group(sb
, group
);
2117 ext4_mb_unload_buddy(&e4b
);
2119 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2124 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2125 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2127 * We've been searching too long. Let's try to allocate
2128 * the best chunk we've found so far
2131 ext4_mb_try_best_found(ac
, &e4b
);
2132 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2134 * Someone more lucky has already allocated it.
2135 * The only thing we can do is just take first
2137 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2139 ac
->ac_b_ex
.fe_group
= 0;
2140 ac
->ac_b_ex
.fe_start
= 0;
2141 ac
->ac_b_ex
.fe_len
= 0;
2142 ac
->ac_status
= AC_STATUS_CONTINUE
;
2143 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2145 atomic_inc(&sbi
->s_mb_lost_chunks
);
2153 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2155 struct super_block
*sb
= seq
->private;
2158 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2161 return (void *) ((unsigned long) group
);
2164 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2166 struct super_block
*sb
= seq
->private;
2170 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2173 return (void *) ((unsigned long) group
);
2176 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2178 struct super_block
*sb
= seq
->private;
2179 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2182 struct ext4_buddy e4b
;
2184 struct ext4_group_info info
;
2185 ext4_grpblk_t counters
[16];
2190 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2191 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2192 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2193 "group", "free", "frags", "first",
2194 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2195 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2197 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2198 sizeof(struct ext4_group_info
);
2199 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2201 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2204 ext4_lock_group(sb
, group
);
2205 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2206 ext4_unlock_group(sb
, group
);
2207 ext4_mb_unload_buddy(&e4b
);
2209 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2210 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2211 for (i
= 0; i
<= 13; i
++)
2212 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2213 sg
.info
.bb_counters
[i
] : 0);
2214 seq_printf(seq
, " ]\n");
2219 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2223 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2224 .start
= ext4_mb_seq_groups_start
,
2225 .next
= ext4_mb_seq_groups_next
,
2226 .stop
= ext4_mb_seq_groups_stop
,
2227 .show
= ext4_mb_seq_groups_show
,
2230 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2232 struct super_block
*sb
= PDE(inode
)->data
;
2235 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2237 struct seq_file
*m
= file
->private_data
;
2244 static const struct file_operations ext4_mb_seq_groups_fops
= {
2245 .owner
= THIS_MODULE
,
2246 .open
= ext4_mb_seq_groups_open
,
2248 .llseek
= seq_lseek
,
2249 .release
= seq_release
,
2252 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2254 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2255 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2261 /* Create and initialize ext4_group_info data for the given group. */
2262 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2263 struct ext4_group_desc
*desc
)
2267 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2268 struct ext4_group_info
**meta_group_info
;
2269 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2272 * First check if this group is the first of a reserved block.
2273 * If it's true, we have to allocate a new table of pointers
2274 * to ext4_group_info structures
2276 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2277 metalen
= sizeof(*meta_group_info
) <<
2278 EXT4_DESC_PER_BLOCK_BITS(sb
);
2279 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2280 if (meta_group_info
== NULL
) {
2281 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2283 goto exit_meta_group_info
;
2285 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2290 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2291 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2293 meta_group_info
[i
] = kmem_cache_alloc(cachep
, GFP_KERNEL
);
2294 if (meta_group_info
[i
] == NULL
) {
2295 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2296 goto exit_group_info
;
2298 memset(meta_group_info
[i
], 0, kmem_cache_size(cachep
));
2299 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2300 &(meta_group_info
[i
]->bb_state
));
2303 * initialize bb_free to be able to skip
2304 * empty groups without initialization
2306 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2307 meta_group_info
[i
]->bb_free
=
2308 ext4_free_blocks_after_init(sb
, group
, desc
);
2310 meta_group_info
[i
]->bb_free
=
2311 ext4_free_blks_count(sb
, desc
);
2314 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2315 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2316 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2317 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2321 struct buffer_head
*bh
;
2322 meta_group_info
[i
]->bb_bitmap
=
2323 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2324 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2325 bh
= ext4_read_block_bitmap(sb
, group
);
2327 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2336 /* If a meta_group_info table has been allocated, release it now */
2337 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0)
2338 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2339 exit_meta_group_info
:
2341 } /* ext4_mb_add_groupinfo */
2343 static int ext4_mb_init_backend(struct super_block
*sb
)
2345 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2347 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2348 struct ext4_super_block
*es
= sbi
->s_es
;
2349 int num_meta_group_infos
;
2350 int num_meta_group_infos_max
;
2352 struct ext4_group_desc
*desc
;
2353 struct kmem_cache
*cachep
;
2355 /* This is the number of blocks used by GDT */
2356 num_meta_group_infos
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) -
2357 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2360 * This is the total number of blocks used by GDT including
2361 * the number of reserved blocks for GDT.
2362 * The s_group_info array is allocated with this value
2363 * to allow a clean online resize without a complex
2364 * manipulation of pointer.
2365 * The drawback is the unused memory when no resize
2366 * occurs but it's very low in terms of pages
2367 * (see comments below)
2368 * Need to handle this properly when META_BG resizing is allowed
2370 num_meta_group_infos_max
= num_meta_group_infos
+
2371 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2374 * array_size is the size of s_group_info array. We round it
2375 * to the next power of two because this approximation is done
2376 * internally by kmalloc so we can have some more memory
2377 * for free here (e.g. may be used for META_BG resize).
2380 while (array_size
< sizeof(*sbi
->s_group_info
) *
2381 num_meta_group_infos_max
)
2382 array_size
= array_size
<< 1;
2383 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2384 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2385 * So a two level scheme suffices for now. */
2386 sbi
->s_group_info
= kmalloc(array_size
, GFP_KERNEL
);
2387 if (sbi
->s_group_info
== NULL
) {
2388 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2391 sbi
->s_buddy_cache
= new_inode(sb
);
2392 if (sbi
->s_buddy_cache
== NULL
) {
2393 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2396 sbi
->s_buddy_cache
->i_ino
= get_next_ino();
2397 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2398 for (i
= 0; i
< ngroups
; i
++) {
2399 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2402 "EXT4-fs: can't read descriptor %u\n", i
);
2405 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2412 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2414 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2415 i
= num_meta_group_infos
;
2417 kfree(sbi
->s_group_info
[i
]);
2418 iput(sbi
->s_buddy_cache
);
2420 kfree(sbi
->s_group_info
);
2424 static void ext4_groupinfo_destroy_slabs(void)
2428 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2429 if (ext4_groupinfo_caches
[i
])
2430 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2431 ext4_groupinfo_caches
[i
] = NULL
;
2435 static int ext4_groupinfo_create_slab(size_t size
)
2437 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2439 int blocksize_bits
= order_base_2(size
);
2440 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2441 struct kmem_cache
*cachep
;
2443 if (cache_index
>= NR_GRPINFO_CACHES
)
2446 if (unlikely(cache_index
< 0))
2449 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2450 if (ext4_groupinfo_caches
[cache_index
]) {
2451 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2452 return 0; /* Already created */
2455 slab_size
= offsetof(struct ext4_group_info
,
2456 bb_counters
[blocksize_bits
+ 2]);
2458 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2459 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2462 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2464 printk(KERN_EMERG
"EXT4: no memory for groupinfo slab cache\n");
2468 ext4_groupinfo_caches
[cache_index
] = cachep
;
2473 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2475 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2481 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2483 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2484 if (sbi
->s_mb_offsets
== NULL
) {
2489 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2490 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2491 if (sbi
->s_mb_maxs
== NULL
) {
2496 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2500 /* order 0 is regular bitmap */
2501 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2502 sbi
->s_mb_offsets
[0] = 0;
2506 max
= sb
->s_blocksize
<< 2;
2508 sbi
->s_mb_offsets
[i
] = offset
;
2509 sbi
->s_mb_maxs
[i
] = max
;
2510 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2513 } while (i
<= sb
->s_blocksize_bits
+ 1);
2515 /* init file for buddy data */
2516 ret
= ext4_mb_init_backend(sb
);
2521 spin_lock_init(&sbi
->s_md_lock
);
2522 spin_lock_init(&sbi
->s_bal_lock
);
2524 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2525 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2526 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2527 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2528 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2529 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2531 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2532 if (sbi
->s_locality_groups
== NULL
) {
2536 for_each_possible_cpu(i
) {
2537 struct ext4_locality_group
*lg
;
2538 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2539 mutex_init(&lg
->lg_mutex
);
2540 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2541 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2542 spin_lock_init(&lg
->lg_prealloc_lock
);
2546 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2547 &ext4_mb_seq_groups_fops
, sb
);
2550 sbi
->s_journal
->j_commit_callback
= release_blocks_on_commit
;
2553 kfree(sbi
->s_mb_offsets
);
2554 kfree(sbi
->s_mb_maxs
);
2559 /* need to called with the ext4 group lock held */
2560 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2562 struct ext4_prealloc_space
*pa
;
2563 struct list_head
*cur
, *tmp
;
2566 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2567 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2568 list_del(&pa
->pa_group_list
);
2570 kmem_cache_free(ext4_pspace_cachep
, pa
);
2573 mb_debug(1, "mballoc: %u PAs left\n", count
);
2577 int ext4_mb_release(struct super_block
*sb
)
2579 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2581 int num_meta_group_infos
;
2582 struct ext4_group_info
*grinfo
;
2583 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2584 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2586 if (sbi
->s_group_info
) {
2587 for (i
= 0; i
< ngroups
; i
++) {
2588 grinfo
= ext4_get_group_info(sb
, i
);
2590 kfree(grinfo
->bb_bitmap
);
2592 ext4_lock_group(sb
, i
);
2593 ext4_mb_cleanup_pa(grinfo
);
2594 ext4_unlock_group(sb
, i
);
2595 kmem_cache_free(cachep
, grinfo
);
2597 num_meta_group_infos
= (ngroups
+
2598 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2599 EXT4_DESC_PER_BLOCK_BITS(sb
);
2600 for (i
= 0; i
< num_meta_group_infos
; i
++)
2601 kfree(sbi
->s_group_info
[i
]);
2602 kfree(sbi
->s_group_info
);
2604 kfree(sbi
->s_mb_offsets
);
2605 kfree(sbi
->s_mb_maxs
);
2606 if (sbi
->s_buddy_cache
)
2607 iput(sbi
->s_buddy_cache
);
2608 if (sbi
->s_mb_stats
) {
2610 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2611 atomic_read(&sbi
->s_bal_allocated
),
2612 atomic_read(&sbi
->s_bal_reqs
),
2613 atomic_read(&sbi
->s_bal_success
));
2615 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2616 "%u 2^N hits, %u breaks, %u lost\n",
2617 atomic_read(&sbi
->s_bal_ex_scanned
),
2618 atomic_read(&sbi
->s_bal_goals
),
2619 atomic_read(&sbi
->s_bal_2orders
),
2620 atomic_read(&sbi
->s_bal_breaks
),
2621 atomic_read(&sbi
->s_mb_lost_chunks
));
2623 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2624 sbi
->s_mb_buddies_generated
++,
2625 sbi
->s_mb_generation_time
);
2627 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2628 atomic_read(&sbi
->s_mb_preallocated
),
2629 atomic_read(&sbi
->s_mb_discarded
));
2632 free_percpu(sbi
->s_locality_groups
);
2634 remove_proc_entry("mb_groups", sbi
->s_proc
);
2639 static inline int ext4_issue_discard(struct super_block
*sb
,
2640 ext4_group_t block_group
, ext4_grpblk_t block
, int count
)
2642 ext4_fsblk_t discard_block
;
2644 discard_block
= block
+ ext4_group_first_block_no(sb
, block_group
);
2645 trace_ext4_discard_blocks(sb
,
2646 (unsigned long long) discard_block
, count
);
2647 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, 0);
2651 * This function is called by the jbd2 layer once the commit has finished,
2652 * so we know we can free the blocks that were released with that commit.
2654 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
)
2656 struct super_block
*sb
= journal
->j_private
;
2657 struct ext4_buddy e4b
;
2658 struct ext4_group_info
*db
;
2659 int err
, ret
, count
= 0, count2
= 0;
2660 struct ext4_free_data
*entry
;
2661 struct list_head
*l
, *ltmp
;
2663 list_for_each_safe(l
, ltmp
, &txn
->t_private_list
) {
2664 entry
= list_entry(l
, struct ext4_free_data
, list
);
2666 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2667 entry
->count
, entry
->group
, entry
);
2669 if (test_opt(sb
, DISCARD
)) {
2670 ret
= ext4_issue_discard(sb
, entry
->group
,
2671 entry
->start_blk
, entry
->count
);
2672 if (unlikely(ret
== -EOPNOTSUPP
)) {
2673 ext4_warning(sb
, "discard not supported, "
2675 clear_opt(sb
, DISCARD
);
2679 err
= ext4_mb_load_buddy(sb
, entry
->group
, &e4b
);
2680 /* we expect to find existing buddy because it's pinned */
2684 /* there are blocks to put in buddy to make them really free */
2685 count
+= entry
->count
;
2687 ext4_lock_group(sb
, entry
->group
);
2688 /* Take it out of per group rb tree */
2689 rb_erase(&entry
->node
, &(db
->bb_free_root
));
2690 mb_free_blocks(NULL
, &e4b
, entry
->start_blk
, entry
->count
);
2692 if (!db
->bb_free_root
.rb_node
) {
2693 /* No more items in the per group rb tree
2694 * balance refcounts from ext4_mb_free_metadata()
2696 page_cache_release(e4b
.bd_buddy_page
);
2697 page_cache_release(e4b
.bd_bitmap_page
);
2699 ext4_unlock_group(sb
, entry
->group
);
2700 kmem_cache_free(ext4_free_ext_cachep
, entry
);
2701 ext4_mb_unload_buddy(&e4b
);
2704 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2707 #ifdef CONFIG_EXT4_DEBUG
2708 u8 mb_enable_debug __read_mostly
;
2710 static struct dentry
*debugfs_dir
;
2711 static struct dentry
*debugfs_debug
;
2713 static void __init
ext4_create_debugfs_entry(void)
2715 debugfs_dir
= debugfs_create_dir("ext4", NULL
);
2717 debugfs_debug
= debugfs_create_u8("mballoc-debug",
2723 static void ext4_remove_debugfs_entry(void)
2725 debugfs_remove(debugfs_debug
);
2726 debugfs_remove(debugfs_dir
);
2731 static void __init
ext4_create_debugfs_entry(void)
2735 static void ext4_remove_debugfs_entry(void)
2741 int __init
ext4_init_mballoc(void)
2743 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2744 SLAB_RECLAIM_ACCOUNT
);
2745 if (ext4_pspace_cachep
== NULL
)
2748 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2749 SLAB_RECLAIM_ACCOUNT
);
2750 if (ext4_ac_cachep
== NULL
) {
2751 kmem_cache_destroy(ext4_pspace_cachep
);
2755 ext4_free_ext_cachep
= KMEM_CACHE(ext4_free_data
,
2756 SLAB_RECLAIM_ACCOUNT
);
2757 if (ext4_free_ext_cachep
== NULL
) {
2758 kmem_cache_destroy(ext4_pspace_cachep
);
2759 kmem_cache_destroy(ext4_ac_cachep
);
2762 ext4_create_debugfs_entry();
2766 void ext4_exit_mballoc(void)
2769 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2770 * before destroying the slab cache.
2773 kmem_cache_destroy(ext4_pspace_cachep
);
2774 kmem_cache_destroy(ext4_ac_cachep
);
2775 kmem_cache_destroy(ext4_free_ext_cachep
);
2776 ext4_groupinfo_destroy_slabs();
2777 ext4_remove_debugfs_entry();
2782 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2783 * Returns 0 if success or error code
2785 static noinline_for_stack
int
2786 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2787 handle_t
*handle
, unsigned int reserv_blks
)
2789 struct buffer_head
*bitmap_bh
= NULL
;
2790 struct ext4_group_desc
*gdp
;
2791 struct buffer_head
*gdp_bh
;
2792 struct ext4_sb_info
*sbi
;
2793 struct super_block
*sb
;
2797 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2798 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2804 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2808 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2813 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2817 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2818 ext4_free_blks_count(sb
, gdp
));
2820 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2824 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
2826 len
= ac
->ac_b_ex
.fe_len
;
2827 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2828 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2829 "fs metadata\n", block
, block
+len
);
2830 /* File system mounted not to panic on error
2831 * Fix the bitmap and repeat the block allocation
2832 * We leak some of the blocks here.
2834 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2835 mb_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2836 ac
->ac_b_ex
.fe_len
);
2837 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2838 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2844 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2845 #ifdef AGGRESSIVE_CHECK
2848 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2849 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2850 bitmap_bh
->b_data
));
2854 mb_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,ac
->ac_b_ex
.fe_len
);
2855 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2856 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2857 ext4_free_blks_set(sb
, gdp
,
2858 ext4_free_blocks_after_init(sb
,
2859 ac
->ac_b_ex
.fe_group
, gdp
));
2861 len
= ext4_free_blks_count(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
2862 ext4_free_blks_set(sb
, gdp
, len
);
2863 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
2865 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2866 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
2868 * Now reduce the dirty block count also. Should not go negative
2870 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2871 /* release all the reserved blocks if non delalloc */
2872 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
, reserv_blks
);
2874 if (sbi
->s_log_groups_per_flex
) {
2875 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2876 ac
->ac_b_ex
.fe_group
);
2877 atomic_sub(ac
->ac_b_ex
.fe_len
,
2878 &sbi
->s_flex_groups
[flex_group
].free_blocks
);
2881 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2884 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
2887 ext4_mark_super_dirty(sb
);
2893 * here we normalize request for locality group
2894 * Group request are normalized to s_strip size if we set the same via mount
2895 * option. If not we set it to s_mb_group_prealloc which can be configured via
2896 * /sys/fs/ext4/<partition>/mb_group_prealloc
2898 * XXX: should we try to preallocate more than the group has now?
2900 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2902 struct super_block
*sb
= ac
->ac_sb
;
2903 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2906 if (EXT4_SB(sb
)->s_stripe
)
2907 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
2909 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2910 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2911 current
->pid
, ac
->ac_g_ex
.fe_len
);
2915 * Normalization means making request better in terms of
2916 * size and alignment
2918 static noinline_for_stack
void
2919 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
2920 struct ext4_allocation_request
*ar
)
2924 loff_t size
, orig_size
, start_off
;
2926 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
2927 struct ext4_prealloc_space
*pa
;
2929 /* do normalize only data requests, metadata requests
2930 do not need preallocation */
2931 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
2934 /* sometime caller may want exact blocks */
2935 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2938 /* caller may indicate that preallocation isn't
2939 * required (it's a tail, for example) */
2940 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
2943 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
2944 ext4_mb_normalize_group_request(ac
);
2948 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
2950 /* first, let's learn actual file size
2951 * given current request is allocated */
2952 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
2953 size
= size
<< bsbits
;
2954 if (size
< i_size_read(ac
->ac_inode
))
2955 size
= i_size_read(ac
->ac_inode
);
2958 /* max size of free chunks */
2961 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2962 (req <= (size) || max <= (chunk_size))
2964 /* first, try to predict filesize */
2965 /* XXX: should this table be tunable? */
2967 if (size
<= 16 * 1024) {
2969 } else if (size
<= 32 * 1024) {
2971 } else if (size
<= 64 * 1024) {
2973 } else if (size
<= 128 * 1024) {
2975 } else if (size
<= 256 * 1024) {
2977 } else if (size
<= 512 * 1024) {
2979 } else if (size
<= 1024 * 1024) {
2981 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
2982 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2983 (21 - bsbits
)) << 21;
2984 size
= 2 * 1024 * 1024;
2985 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
2986 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2987 (22 - bsbits
)) << 22;
2988 size
= 4 * 1024 * 1024;
2989 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
2990 (8<<20)>>bsbits
, max
, 8 * 1024)) {
2991 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2992 (23 - bsbits
)) << 23;
2993 size
= 8 * 1024 * 1024;
2995 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
2996 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
2998 size
= size
>> bsbits
;
2999 start
= start_off
>> bsbits
;
3001 /* don't cover already allocated blocks in selected range */
3002 if (ar
->pleft
&& start
<= ar
->lleft
) {
3003 size
-= ar
->lleft
+ 1 - start
;
3004 start
= ar
->lleft
+ 1;
3006 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3007 size
-= start
+ size
- ar
->lright
;
3011 /* check we don't cross already preallocated blocks */
3013 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3018 spin_lock(&pa
->pa_lock
);
3019 if (pa
->pa_deleted
) {
3020 spin_unlock(&pa
->pa_lock
);
3024 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3026 /* PA must not overlap original request */
3027 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3028 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3030 /* skip PAs this normalized request doesn't overlap with */
3031 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
3032 spin_unlock(&pa
->pa_lock
);
3035 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3037 /* adjust start or end to be adjacent to this pa */
3038 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3039 BUG_ON(pa_end
< start
);
3041 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3042 BUG_ON(pa
->pa_lstart
> end
);
3043 end
= pa
->pa_lstart
;
3045 spin_unlock(&pa
->pa_lock
);
3050 /* XXX: extra loop to check we really don't overlap preallocations */
3052 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3054 spin_lock(&pa
->pa_lock
);
3055 if (pa
->pa_deleted
== 0) {
3056 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3057 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3059 spin_unlock(&pa
->pa_lock
);
3063 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3064 start
> ac
->ac_o_ex
.fe_logical
) {
3065 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3066 (unsigned long) start
, (unsigned long) size
,
3067 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3069 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3070 start
> ac
->ac_o_ex
.fe_logical
);
3071 BUG_ON(size
<= 0 || size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3073 /* now prepare goal request */
3075 /* XXX: is it better to align blocks WRT to logical
3076 * placement or satisfy big request as is */
3077 ac
->ac_g_ex
.fe_logical
= start
;
3078 ac
->ac_g_ex
.fe_len
= size
;
3080 /* define goal start in order to merge */
3081 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3082 /* merge to the right */
3083 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3084 &ac
->ac_f_ex
.fe_group
,
3085 &ac
->ac_f_ex
.fe_start
);
3086 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3088 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3089 /* merge to the left */
3090 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3091 &ac
->ac_f_ex
.fe_group
,
3092 &ac
->ac_f_ex
.fe_start
);
3093 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3096 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3097 (unsigned) orig_size
, (unsigned) start
);
3100 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3102 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3104 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3105 atomic_inc(&sbi
->s_bal_reqs
);
3106 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3107 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3108 atomic_inc(&sbi
->s_bal_success
);
3109 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3110 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3111 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3112 atomic_inc(&sbi
->s_bal_goals
);
3113 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3114 atomic_inc(&sbi
->s_bal_breaks
);
3117 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3118 trace_ext4_mballoc_alloc(ac
);
3120 trace_ext4_mballoc_prealloc(ac
);
3124 * Called on failure; free up any blocks from the inode PA for this
3125 * context. We don't need this for MB_GROUP_PA because we only change
3126 * pa_free in ext4_mb_release_context(), but on failure, we've already
3127 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3129 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3131 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3134 if (pa
&& pa
->pa_type
== MB_INODE_PA
) {
3135 len
= ac
->ac_b_ex
.fe_len
;
3142 * use blocks preallocated to inode
3144 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3145 struct ext4_prealloc_space
*pa
)
3151 /* found preallocated blocks, use them */
3152 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3153 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3155 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3156 &ac
->ac_b_ex
.fe_start
);
3157 ac
->ac_b_ex
.fe_len
= len
;
3158 ac
->ac_status
= AC_STATUS_FOUND
;
3161 BUG_ON(start
< pa
->pa_pstart
);
3162 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3163 BUG_ON(pa
->pa_free
< len
);
3166 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3170 * use blocks preallocated to locality group
3172 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3173 struct ext4_prealloc_space
*pa
)
3175 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3177 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3178 &ac
->ac_b_ex
.fe_group
,
3179 &ac
->ac_b_ex
.fe_start
);
3180 ac
->ac_b_ex
.fe_len
= len
;
3181 ac
->ac_status
= AC_STATUS_FOUND
;
3184 /* we don't correct pa_pstart or pa_plen here to avoid
3185 * possible race when the group is being loaded concurrently
3186 * instead we correct pa later, after blocks are marked
3187 * in on-disk bitmap -- see ext4_mb_release_context()
3188 * Other CPUs are prevented from allocating from this pa by lg_mutex
3190 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3194 * Return the prealloc space that have minimal distance
3195 * from the goal block. @cpa is the prealloc
3196 * space that is having currently known minimal distance
3197 * from the goal block.
3199 static struct ext4_prealloc_space
*
3200 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3201 struct ext4_prealloc_space
*pa
,
3202 struct ext4_prealloc_space
*cpa
)
3204 ext4_fsblk_t cur_distance
, new_distance
;
3207 atomic_inc(&pa
->pa_count
);
3210 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3211 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3213 if (cur_distance
< new_distance
)
3216 /* drop the previous reference */
3217 atomic_dec(&cpa
->pa_count
);
3218 atomic_inc(&pa
->pa_count
);
3223 * search goal blocks in preallocated space
3225 static noinline_for_stack
int
3226 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3229 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3230 struct ext4_locality_group
*lg
;
3231 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3232 ext4_fsblk_t goal_block
;
3234 /* only data can be preallocated */
3235 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3238 /* first, try per-file preallocation */
3240 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3242 /* all fields in this condition don't change,
3243 * so we can skip locking for them */
3244 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3245 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3248 /* non-extent files can't have physical blocks past 2^32 */
3249 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3250 pa
->pa_pstart
+ pa
->pa_len
> EXT4_MAX_BLOCK_FILE_PHYS
)
3253 /* found preallocated blocks, use them */
3254 spin_lock(&pa
->pa_lock
);
3255 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3256 atomic_inc(&pa
->pa_count
);
3257 ext4_mb_use_inode_pa(ac
, pa
);
3258 spin_unlock(&pa
->pa_lock
);
3259 ac
->ac_criteria
= 10;
3263 spin_unlock(&pa
->pa_lock
);
3267 /* can we use group allocation? */
3268 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3271 /* inode may have no locality group for some reason */
3275 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3276 if (order
> PREALLOC_TB_SIZE
- 1)
3277 /* The max size of hash table is PREALLOC_TB_SIZE */
3278 order
= PREALLOC_TB_SIZE
- 1;
3280 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3282 * search for the prealloc space that is having
3283 * minimal distance from the goal block.
3285 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3287 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3289 spin_lock(&pa
->pa_lock
);
3290 if (pa
->pa_deleted
== 0 &&
3291 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3293 cpa
= ext4_mb_check_group_pa(goal_block
,
3296 spin_unlock(&pa
->pa_lock
);
3301 ext4_mb_use_group_pa(ac
, cpa
);
3302 ac
->ac_criteria
= 20;
3309 * the function goes through all block freed in the group
3310 * but not yet committed and marks them used in in-core bitmap.
3311 * buddy must be generated from this bitmap
3312 * Need to be called with the ext4 group lock held
3314 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3318 struct ext4_group_info
*grp
;
3319 struct ext4_free_data
*entry
;
3321 grp
= ext4_get_group_info(sb
, group
);
3322 n
= rb_first(&(grp
->bb_free_root
));
3325 entry
= rb_entry(n
, struct ext4_free_data
, node
);
3326 mb_set_bits(bitmap
, entry
->start_blk
, entry
->count
);
3333 * the function goes through all preallocation in this group and marks them
3334 * used in in-core bitmap. buddy must be generated from this bitmap
3335 * Need to be called with ext4 group lock held
3337 static noinline_for_stack
3338 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3341 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3342 struct ext4_prealloc_space
*pa
;
3343 struct list_head
*cur
;
3344 ext4_group_t groupnr
;
3345 ext4_grpblk_t start
;
3346 int preallocated
= 0;
3350 /* all form of preallocation discards first load group,
3351 * so the only competing code is preallocation use.
3352 * we don't need any locking here
3353 * notice we do NOT ignore preallocations with pa_deleted
3354 * otherwise we could leave used blocks available for
3355 * allocation in buddy when concurrent ext4_mb_put_pa()
3356 * is dropping preallocation
3358 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3359 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3360 spin_lock(&pa
->pa_lock
);
3361 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3364 spin_unlock(&pa
->pa_lock
);
3365 if (unlikely(len
== 0))
3367 BUG_ON(groupnr
!= group
);
3368 mb_set_bits(bitmap
, start
, len
);
3369 preallocated
+= len
;
3372 mb_debug(1, "prellocated %u for group %u\n", preallocated
, group
);
3375 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3377 struct ext4_prealloc_space
*pa
;
3378 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3379 kmem_cache_free(ext4_pspace_cachep
, pa
);
3383 * drops a reference to preallocated space descriptor
3384 * if this was the last reference and the space is consumed
3386 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3387 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3390 ext4_fsblk_t grp_blk
;
3392 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3395 /* in this short window concurrent discard can set pa_deleted */
3396 spin_lock(&pa
->pa_lock
);
3397 if (pa
->pa_deleted
== 1) {
3398 spin_unlock(&pa
->pa_lock
);
3403 spin_unlock(&pa
->pa_lock
);
3405 grp_blk
= pa
->pa_pstart
;
3407 * If doing group-based preallocation, pa_pstart may be in the
3408 * next group when pa is used up
3410 if (pa
->pa_type
== MB_GROUP_PA
)
3413 ext4_get_group_no_and_offset(sb
, grp_blk
, &grp
, NULL
);
3418 * P1 (buddy init) P2 (regular allocation)
3419 * find block B in PA
3420 * copy on-disk bitmap to buddy
3421 * mark B in on-disk bitmap
3422 * drop PA from group
3423 * mark all PAs in buddy
3425 * thus, P1 initializes buddy with B available. to prevent this
3426 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3429 ext4_lock_group(sb
, grp
);
3430 list_del(&pa
->pa_group_list
);
3431 ext4_unlock_group(sb
, grp
);
3433 spin_lock(pa
->pa_obj_lock
);
3434 list_del_rcu(&pa
->pa_inode_list
);
3435 spin_unlock(pa
->pa_obj_lock
);
3437 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3441 * creates new preallocated space for given inode
3443 static noinline_for_stack
int
3444 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3446 struct super_block
*sb
= ac
->ac_sb
;
3447 struct ext4_prealloc_space
*pa
;
3448 struct ext4_group_info
*grp
;
3449 struct ext4_inode_info
*ei
;
3451 /* preallocate only when found space is larger then requested */
3452 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3453 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3454 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3456 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3460 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3466 /* we can't allocate as much as normalizer wants.
3467 * so, found space must get proper lstart
3468 * to cover original request */
3469 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3470 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3472 /* we're limited by original request in that
3473 * logical block must be covered any way
3474 * winl is window we can move our chunk within */
3475 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3477 /* also, we should cover whole original request */
3478 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3480 /* the smallest one defines real window */
3481 win
= min(winl
, wins
);
3483 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3484 if (offs
&& offs
< win
)
3487 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3488 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3489 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3492 /* preallocation can change ac_b_ex, thus we store actually
3493 * allocated blocks for history */
3494 ac
->ac_f_ex
= ac
->ac_b_ex
;
3496 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3497 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3498 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3499 pa
->pa_free
= pa
->pa_len
;
3500 atomic_set(&pa
->pa_count
, 1);
3501 spin_lock_init(&pa
->pa_lock
);
3502 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3503 INIT_LIST_HEAD(&pa
->pa_group_list
);
3505 pa
->pa_type
= MB_INODE_PA
;
3507 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3508 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3509 trace_ext4_mb_new_inode_pa(ac
, pa
);
3511 ext4_mb_use_inode_pa(ac
, pa
);
3512 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3514 ei
= EXT4_I(ac
->ac_inode
);
3515 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3517 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3518 pa
->pa_inode
= ac
->ac_inode
;
3520 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3521 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3522 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3524 spin_lock(pa
->pa_obj_lock
);
3525 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3526 spin_unlock(pa
->pa_obj_lock
);
3532 * creates new preallocated space for locality group inodes belongs to
3534 static noinline_for_stack
int
3535 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3537 struct super_block
*sb
= ac
->ac_sb
;
3538 struct ext4_locality_group
*lg
;
3539 struct ext4_prealloc_space
*pa
;
3540 struct ext4_group_info
*grp
;
3542 /* preallocate only when found space is larger then requested */
3543 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3544 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3545 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3547 BUG_ON(ext4_pspace_cachep
== NULL
);
3548 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3552 /* preallocation can change ac_b_ex, thus we store actually
3553 * allocated blocks for history */
3554 ac
->ac_f_ex
= ac
->ac_b_ex
;
3556 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3557 pa
->pa_lstart
= pa
->pa_pstart
;
3558 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3559 pa
->pa_free
= pa
->pa_len
;
3560 atomic_set(&pa
->pa_count
, 1);
3561 spin_lock_init(&pa
->pa_lock
);
3562 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3563 INIT_LIST_HEAD(&pa
->pa_group_list
);
3565 pa
->pa_type
= MB_GROUP_PA
;
3567 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3568 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3569 trace_ext4_mb_new_group_pa(ac
, pa
);
3571 ext4_mb_use_group_pa(ac
, pa
);
3572 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3574 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3578 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3579 pa
->pa_inode
= NULL
;
3581 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3582 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3583 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3586 * We will later add the new pa to the right bucket
3587 * after updating the pa_free in ext4_mb_release_context
3592 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3596 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3597 err
= ext4_mb_new_group_pa(ac
);
3599 err
= ext4_mb_new_inode_pa(ac
);
3604 * finds all unused blocks in on-disk bitmap, frees them in
3605 * in-core bitmap and buddy.
3606 * @pa must be unlinked from inode and group lists, so that
3607 * nobody else can find/use it.
3608 * the caller MUST hold group/inode locks.
3609 * TODO: optimize the case when there are no in-core structures yet
3611 static noinline_for_stack
int
3612 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3613 struct ext4_prealloc_space
*pa
)
3615 struct super_block
*sb
= e4b
->bd_sb
;
3616 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3621 unsigned long long grp_blk_start
;
3625 BUG_ON(pa
->pa_deleted
== 0);
3626 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3627 grp_blk_start
= pa
->pa_pstart
- bit
;
3628 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3629 end
= bit
+ pa
->pa_len
;
3632 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3635 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3636 mb_debug(1, " free preallocated %u/%u in group %u\n",
3637 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3638 (unsigned) next
- bit
, (unsigned) group
);
3641 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3642 trace_ext4_mb_release_inode_pa(sb
, pa
->pa_inode
, pa
,
3643 grp_blk_start
+ bit
, next
- bit
);
3644 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3647 if (free
!= pa
->pa_free
) {
3648 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3649 pa
, (unsigned long) pa
->pa_lstart
,
3650 (unsigned long) pa
->pa_pstart
,
3651 (unsigned long) pa
->pa_len
);
3652 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3655 * pa is already deleted so we use the value obtained
3656 * from the bitmap and continue.
3659 atomic_add(free
, &sbi
->s_mb_discarded
);
3664 static noinline_for_stack
int
3665 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3666 struct ext4_prealloc_space
*pa
)
3668 struct super_block
*sb
= e4b
->bd_sb
;
3672 trace_ext4_mb_release_group_pa(sb
, pa
);
3673 BUG_ON(pa
->pa_deleted
== 0);
3674 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3675 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3676 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3677 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3678 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3684 * releases all preallocations in given group
3686 * first, we need to decide discard policy:
3687 * - when do we discard
3689 * - how many do we discard
3690 * 1) how many requested
3692 static noinline_for_stack
int
3693 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3694 ext4_group_t group
, int needed
)
3696 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3697 struct buffer_head
*bitmap_bh
= NULL
;
3698 struct ext4_prealloc_space
*pa
, *tmp
;
3699 struct list_head list
;
3700 struct ext4_buddy e4b
;
3705 mb_debug(1, "discard preallocation for group %u\n", group
);
3707 if (list_empty(&grp
->bb_prealloc_list
))
3710 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3711 if (bitmap_bh
== NULL
) {
3712 ext4_error(sb
, "Error reading block bitmap for %u", group
);
3716 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3718 ext4_error(sb
, "Error loading buddy information for %u", group
);
3724 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3726 INIT_LIST_HEAD(&list
);
3728 ext4_lock_group(sb
, group
);
3729 list_for_each_entry_safe(pa
, tmp
,
3730 &grp
->bb_prealloc_list
, pa_group_list
) {
3731 spin_lock(&pa
->pa_lock
);
3732 if (atomic_read(&pa
->pa_count
)) {
3733 spin_unlock(&pa
->pa_lock
);
3737 if (pa
->pa_deleted
) {
3738 spin_unlock(&pa
->pa_lock
);
3742 /* seems this one can be freed ... */
3745 /* we can trust pa_free ... */
3746 free
+= pa
->pa_free
;
3748 spin_unlock(&pa
->pa_lock
);
3750 list_del(&pa
->pa_group_list
);
3751 list_add(&pa
->u
.pa_tmp_list
, &list
);
3754 /* if we still need more blocks and some PAs were used, try again */
3755 if (free
< needed
&& busy
) {
3757 ext4_unlock_group(sb
, group
);
3759 * Yield the CPU here so that we don't get soft lockup
3760 * in non preempt case.
3766 /* found anything to free? */
3767 if (list_empty(&list
)) {
3772 /* now free all selected PAs */
3773 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3775 /* remove from object (inode or locality group) */
3776 spin_lock(pa
->pa_obj_lock
);
3777 list_del_rcu(&pa
->pa_inode_list
);
3778 spin_unlock(pa
->pa_obj_lock
);
3780 if (pa
->pa_type
== MB_GROUP_PA
)
3781 ext4_mb_release_group_pa(&e4b
, pa
);
3783 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3785 list_del(&pa
->u
.pa_tmp_list
);
3786 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3790 ext4_unlock_group(sb
, group
);
3791 ext4_mb_unload_buddy(&e4b
);
3797 * releases all non-used preallocated blocks for given inode
3799 * It's important to discard preallocations under i_data_sem
3800 * We don't want another block to be served from the prealloc
3801 * space when we are discarding the inode prealloc space.
3803 * FIXME!! Make sure it is valid at all the call sites
3805 void ext4_discard_preallocations(struct inode
*inode
)
3807 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3808 struct super_block
*sb
= inode
->i_sb
;
3809 struct buffer_head
*bitmap_bh
= NULL
;
3810 struct ext4_prealloc_space
*pa
, *tmp
;
3811 ext4_group_t group
= 0;
3812 struct list_head list
;
3813 struct ext4_buddy e4b
;
3816 if (!S_ISREG(inode
->i_mode
)) {
3817 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3821 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
3822 trace_ext4_discard_preallocations(inode
);
3824 INIT_LIST_HEAD(&list
);
3827 /* first, collect all pa's in the inode */
3828 spin_lock(&ei
->i_prealloc_lock
);
3829 while (!list_empty(&ei
->i_prealloc_list
)) {
3830 pa
= list_entry(ei
->i_prealloc_list
.next
,
3831 struct ext4_prealloc_space
, pa_inode_list
);
3832 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3833 spin_lock(&pa
->pa_lock
);
3834 if (atomic_read(&pa
->pa_count
)) {
3835 /* this shouldn't happen often - nobody should
3836 * use preallocation while we're discarding it */
3837 spin_unlock(&pa
->pa_lock
);
3838 spin_unlock(&ei
->i_prealloc_lock
);
3839 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
3841 schedule_timeout_uninterruptible(HZ
);
3845 if (pa
->pa_deleted
== 0) {
3847 spin_unlock(&pa
->pa_lock
);
3848 list_del_rcu(&pa
->pa_inode_list
);
3849 list_add(&pa
->u
.pa_tmp_list
, &list
);
3853 /* someone is deleting pa right now */
3854 spin_unlock(&pa
->pa_lock
);
3855 spin_unlock(&ei
->i_prealloc_lock
);
3857 /* we have to wait here because pa_deleted
3858 * doesn't mean pa is already unlinked from
3859 * the list. as we might be called from
3860 * ->clear_inode() the inode will get freed
3861 * and concurrent thread which is unlinking
3862 * pa from inode's list may access already
3863 * freed memory, bad-bad-bad */
3865 /* XXX: if this happens too often, we can
3866 * add a flag to force wait only in case
3867 * of ->clear_inode(), but not in case of
3868 * regular truncate */
3869 schedule_timeout_uninterruptible(HZ
);
3872 spin_unlock(&ei
->i_prealloc_lock
);
3874 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3875 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
3876 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
3878 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3880 ext4_error(sb
, "Error loading buddy information for %u",
3885 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3886 if (bitmap_bh
== NULL
) {
3887 ext4_error(sb
, "Error reading block bitmap for %u",
3889 ext4_mb_unload_buddy(&e4b
);
3893 ext4_lock_group(sb
, group
);
3894 list_del(&pa
->pa_group_list
);
3895 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3896 ext4_unlock_group(sb
, group
);
3898 ext4_mb_unload_buddy(&e4b
);
3901 list_del(&pa
->u
.pa_tmp_list
);
3902 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3906 #ifdef CONFIG_EXT4_DEBUG
3907 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3909 struct super_block
*sb
= ac
->ac_sb
;
3910 ext4_group_t ngroups
, i
;
3912 if (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
)
3915 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
3916 " Allocation context details:\n");
3917 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
3918 ac
->ac_status
, ac
->ac_flags
);
3919 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3920 "best %lu/%lu/%lu@%lu cr %d\n",
3921 (unsigned long)ac
->ac_o_ex
.fe_group
,
3922 (unsigned long)ac
->ac_o_ex
.fe_start
,
3923 (unsigned long)ac
->ac_o_ex
.fe_len
,
3924 (unsigned long)ac
->ac_o_ex
.fe_logical
,
3925 (unsigned long)ac
->ac_g_ex
.fe_group
,
3926 (unsigned long)ac
->ac_g_ex
.fe_start
,
3927 (unsigned long)ac
->ac_g_ex
.fe_len
,
3928 (unsigned long)ac
->ac_g_ex
.fe_logical
,
3929 (unsigned long)ac
->ac_b_ex
.fe_group
,
3930 (unsigned long)ac
->ac_b_ex
.fe_start
,
3931 (unsigned long)ac
->ac_b_ex
.fe_len
,
3932 (unsigned long)ac
->ac_b_ex
.fe_logical
,
3933 (int)ac
->ac_criteria
);
3934 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
3936 printk(KERN_ERR
"EXT4-fs: groups: \n");
3937 ngroups
= ext4_get_groups_count(sb
);
3938 for (i
= 0; i
< ngroups
; i
++) {
3939 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
3940 struct ext4_prealloc_space
*pa
;
3941 ext4_grpblk_t start
;
3942 struct list_head
*cur
;
3943 ext4_lock_group(sb
, i
);
3944 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3945 pa
= list_entry(cur
, struct ext4_prealloc_space
,
3947 spin_lock(&pa
->pa_lock
);
3948 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3950 spin_unlock(&pa
->pa_lock
);
3951 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
3954 ext4_unlock_group(sb
, i
);
3956 if (grp
->bb_free
== 0)
3958 printk(KERN_ERR
"%u: %d/%d \n",
3959 i
, grp
->bb_free
, grp
->bb_fragments
);
3961 printk(KERN_ERR
"\n");
3964 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3971 * We use locality group preallocation for small size file. The size of the
3972 * file is determined by the current size or the resulting size after
3973 * allocation which ever is larger
3975 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3977 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
3979 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3980 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3983 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3986 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3989 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3990 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
3993 if ((size
== isize
) &&
3994 !ext4_fs_is_busy(sbi
) &&
3995 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
3996 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
4000 /* don't use group allocation for large files */
4001 size
= max(size
, isize
);
4002 if (size
> sbi
->s_mb_stream_request
) {
4003 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4007 BUG_ON(ac
->ac_lg
!= NULL
);
4009 * locality group prealloc space are per cpu. The reason for having
4010 * per cpu locality group is to reduce the contention between block
4011 * request from multiple CPUs.
4013 ac
->ac_lg
= __this_cpu_ptr(sbi
->s_locality_groups
);
4015 /* we're going to use group allocation */
4016 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4018 /* serialize all allocations in the group */
4019 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4022 static noinline_for_stack
int
4023 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4024 struct ext4_allocation_request
*ar
)
4026 struct super_block
*sb
= ar
->inode
->i_sb
;
4027 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4028 struct ext4_super_block
*es
= sbi
->s_es
;
4032 ext4_grpblk_t block
;
4034 /* we can't allocate > group size */
4037 /* just a dirty hack to filter too big requests */
4038 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4039 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4041 /* start searching from the goal */
4043 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4044 goal
>= ext4_blocks_count(es
))
4045 goal
= le32_to_cpu(es
->s_first_data_block
);
4046 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4048 /* set up allocation goals */
4049 memset(ac
, 0, sizeof(struct ext4_allocation_context
));
4050 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4051 ac
->ac_status
= AC_STATUS_CONTINUE
;
4053 ac
->ac_inode
= ar
->inode
;
4054 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4055 ac
->ac_o_ex
.fe_group
= group
;
4056 ac
->ac_o_ex
.fe_start
= block
;
4057 ac
->ac_o_ex
.fe_len
= len
;
4058 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4059 ac
->ac_g_ex
.fe_group
= group
;
4060 ac
->ac_g_ex
.fe_start
= block
;
4061 ac
->ac_g_ex
.fe_len
= len
;
4062 ac
->ac_flags
= ar
->flags
;
4064 /* we have to define context: we'll we work with a file or
4065 * locality group. this is a policy, actually */
4066 ext4_mb_group_or_file(ac
);
4068 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4069 "left: %u/%u, right %u/%u to %swritable\n",
4070 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4071 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4072 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4073 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4074 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4079 static noinline_for_stack
void
4080 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4081 struct ext4_locality_group
*lg
,
4082 int order
, int total_entries
)
4084 ext4_group_t group
= 0;
4085 struct ext4_buddy e4b
;
4086 struct list_head discard_list
;
4087 struct ext4_prealloc_space
*pa
, *tmp
;
4089 mb_debug(1, "discard locality group preallocation\n");
4091 INIT_LIST_HEAD(&discard_list
);
4093 spin_lock(&lg
->lg_prealloc_lock
);
4094 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4096 spin_lock(&pa
->pa_lock
);
4097 if (atomic_read(&pa
->pa_count
)) {
4099 * This is the pa that we just used
4100 * for block allocation. So don't
4103 spin_unlock(&pa
->pa_lock
);
4106 if (pa
->pa_deleted
) {
4107 spin_unlock(&pa
->pa_lock
);
4110 /* only lg prealloc space */
4111 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4113 /* seems this one can be freed ... */
4115 spin_unlock(&pa
->pa_lock
);
4117 list_del_rcu(&pa
->pa_inode_list
);
4118 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4121 if (total_entries
<= 5) {
4123 * we want to keep only 5 entries
4124 * allowing it to grow to 8. This
4125 * mak sure we don't call discard
4126 * soon for this list.
4131 spin_unlock(&lg
->lg_prealloc_lock
);
4133 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4135 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4136 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4137 ext4_error(sb
, "Error loading buddy information for %u",
4141 ext4_lock_group(sb
, group
);
4142 list_del(&pa
->pa_group_list
);
4143 ext4_mb_release_group_pa(&e4b
, pa
);
4144 ext4_unlock_group(sb
, group
);
4146 ext4_mb_unload_buddy(&e4b
);
4147 list_del(&pa
->u
.pa_tmp_list
);
4148 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4153 * We have incremented pa_count. So it cannot be freed at this
4154 * point. Also we hold lg_mutex. So no parallel allocation is
4155 * possible from this lg. That means pa_free cannot be updated.
4157 * A parallel ext4_mb_discard_group_preallocations is possible.
4158 * which can cause the lg_prealloc_list to be updated.
4161 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4163 int order
, added
= 0, lg_prealloc_count
= 1;
4164 struct super_block
*sb
= ac
->ac_sb
;
4165 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4166 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4168 order
= fls(pa
->pa_free
) - 1;
4169 if (order
> PREALLOC_TB_SIZE
- 1)
4170 /* The max size of hash table is PREALLOC_TB_SIZE */
4171 order
= PREALLOC_TB_SIZE
- 1;
4172 /* Add the prealloc space to lg */
4174 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4176 spin_lock(&tmp_pa
->pa_lock
);
4177 if (tmp_pa
->pa_deleted
) {
4178 spin_unlock(&tmp_pa
->pa_lock
);
4181 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4182 /* Add to the tail of the previous entry */
4183 list_add_tail_rcu(&pa
->pa_inode_list
,
4184 &tmp_pa
->pa_inode_list
);
4187 * we want to count the total
4188 * number of entries in the list
4191 spin_unlock(&tmp_pa
->pa_lock
);
4192 lg_prealloc_count
++;
4195 list_add_tail_rcu(&pa
->pa_inode_list
,
4196 &lg
->lg_prealloc_list
[order
]);
4199 /* Now trim the list to be not more than 8 elements */
4200 if (lg_prealloc_count
> 8) {
4201 ext4_mb_discard_lg_preallocations(sb
, lg
,
4202 order
, lg_prealloc_count
);
4209 * release all resource we used in allocation
4211 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4213 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4215 if (pa
->pa_type
== MB_GROUP_PA
) {
4216 /* see comment in ext4_mb_use_group_pa() */
4217 spin_lock(&pa
->pa_lock
);
4218 pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4219 pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4220 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4221 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4222 spin_unlock(&pa
->pa_lock
);
4226 up_read(ac
->alloc_semp
);
4229 * We want to add the pa to the right bucket.
4230 * Remove it from the list and while adding
4231 * make sure the list to which we are adding
4232 * doesn't grow big. We need to release
4233 * alloc_semp before calling ext4_mb_add_n_trim()
4235 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4236 spin_lock(pa
->pa_obj_lock
);
4237 list_del_rcu(&pa
->pa_inode_list
);
4238 spin_unlock(pa
->pa_obj_lock
);
4239 ext4_mb_add_n_trim(ac
);
4241 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4243 if (ac
->ac_bitmap_page
)
4244 page_cache_release(ac
->ac_bitmap_page
);
4245 if (ac
->ac_buddy_page
)
4246 page_cache_release(ac
->ac_buddy_page
);
4247 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4248 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4249 ext4_mb_collect_stats(ac
);
4253 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4255 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4259 trace_ext4_mb_discard_preallocations(sb
, needed
);
4260 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4261 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4270 * Main entry point into mballoc to allocate blocks
4271 * it tries to use preallocation first, then falls back
4272 * to usual allocation
4274 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4275 struct ext4_allocation_request
*ar
, int *errp
)
4278 struct ext4_allocation_context
*ac
= NULL
;
4279 struct ext4_sb_info
*sbi
;
4280 struct super_block
*sb
;
4281 ext4_fsblk_t block
= 0;
4282 unsigned int inquota
= 0;
4283 unsigned int reserv_blks
= 0;
4285 sb
= ar
->inode
->i_sb
;
4288 trace_ext4_request_blocks(ar
);
4291 * For delayed allocation, we could skip the ENOSPC and
4292 * EDQUOT check, as blocks and quotas have been already
4293 * reserved when data being copied into pagecache.
4295 if (ext4_test_inode_state(ar
->inode
, EXT4_STATE_DELALLOC_RESERVED
))
4296 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4298 /* Without delayed allocation we need to verify
4299 * there is enough free blocks to do block allocation
4300 * and verify allocation doesn't exceed the quota limits.
4302 while (ar
->len
&& ext4_claim_free_blocks(sbi
, ar
->len
)) {
4303 /* let others to free the space */
4305 ar
->len
= ar
->len
>> 1;
4311 reserv_blks
= ar
->len
;
4312 while (ar
->len
&& dquot_alloc_block(ar
->inode
, ar
->len
)) {
4313 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4323 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4330 *errp
= ext4_mb_initialize_context(ac
, ar
);
4336 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4337 if (!ext4_mb_use_preallocated(ac
)) {
4338 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4339 ext4_mb_normalize_request(ac
, ar
);
4341 /* allocate space in core */
4342 *errp
= ext4_mb_regular_allocator(ac
);
4346 /* as we've just preallocated more space than
4347 * user requested orinally, we store allocated
4348 * space in a special descriptor */
4349 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4350 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4351 ext4_mb_new_preallocation(ac
);
4353 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4354 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_blks
);
4355 if (*errp
== -EAGAIN
) {
4357 * drop the reference that we took
4358 * in ext4_mb_use_best_found
4360 ext4_mb_release_context(ac
);
4361 ac
->ac_b_ex
.fe_group
= 0;
4362 ac
->ac_b_ex
.fe_start
= 0;
4363 ac
->ac_b_ex
.fe_len
= 0;
4364 ac
->ac_status
= AC_STATUS_CONTINUE
;
4368 ext4_discard_allocated_blocks(ac
);
4370 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4371 ar
->len
= ac
->ac_b_ex
.fe_len
;
4374 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4381 ac
->ac_b_ex
.fe_len
= 0;
4383 ext4_mb_show_ac(ac
);
4385 ext4_mb_release_context(ac
);
4388 kmem_cache_free(ext4_ac_cachep
, ac
);
4389 if (inquota
&& ar
->len
< inquota
)
4390 dquot_free_block(ar
->inode
, inquota
- ar
->len
);
4392 if (!ext4_test_inode_state(ar
->inode
,
4393 EXT4_STATE_DELALLOC_RESERVED
))
4394 /* release all the reserved blocks if non delalloc */
4395 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
,
4399 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4405 * We can merge two free data extents only if the physical blocks
4406 * are contiguous, AND the extents were freed by the same transaction,
4407 * AND the blocks are associated with the same group.
4409 static int can_merge(struct ext4_free_data
*entry1
,
4410 struct ext4_free_data
*entry2
)
4412 if ((entry1
->t_tid
== entry2
->t_tid
) &&
4413 (entry1
->group
== entry2
->group
) &&
4414 ((entry1
->start_blk
+ entry1
->count
) == entry2
->start_blk
))
4419 static noinline_for_stack
int
4420 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4421 struct ext4_free_data
*new_entry
)
4423 ext4_group_t group
= e4b
->bd_group
;
4424 ext4_grpblk_t block
;
4425 struct ext4_free_data
*entry
;
4426 struct ext4_group_info
*db
= e4b
->bd_info
;
4427 struct super_block
*sb
= e4b
->bd_sb
;
4428 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4429 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4430 struct rb_node
*parent
= NULL
, *new_node
;
4432 BUG_ON(!ext4_handle_valid(handle
));
4433 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4434 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4436 new_node
= &new_entry
->node
;
4437 block
= new_entry
->start_blk
;
4440 /* first free block exent. We need to
4441 protect buddy cache from being freed,
4442 * otherwise we'll refresh it from
4443 * on-disk bitmap and lose not-yet-available
4445 page_cache_get(e4b
->bd_buddy_page
);
4446 page_cache_get(e4b
->bd_bitmap_page
);
4450 entry
= rb_entry(parent
, struct ext4_free_data
, node
);
4451 if (block
< entry
->start_blk
)
4453 else if (block
>= (entry
->start_blk
+ entry
->count
))
4454 n
= &(*n
)->rb_right
;
4456 ext4_grp_locked_error(sb
, group
, 0,
4457 ext4_group_first_block_no(sb
, group
) + block
,
4458 "Block already on to-be-freed list");
4463 rb_link_node(new_node
, parent
, n
);
4464 rb_insert_color(new_node
, &db
->bb_free_root
);
4466 /* Now try to see the extent can be merged to left and right */
4467 node
= rb_prev(new_node
);
4469 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4470 if (can_merge(entry
, new_entry
)) {
4471 new_entry
->start_blk
= entry
->start_blk
;
4472 new_entry
->count
+= entry
->count
;
4473 rb_erase(node
, &(db
->bb_free_root
));
4474 spin_lock(&sbi
->s_md_lock
);
4475 list_del(&entry
->list
);
4476 spin_unlock(&sbi
->s_md_lock
);
4477 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4481 node
= rb_next(new_node
);
4483 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4484 if (can_merge(new_entry
, entry
)) {
4485 new_entry
->count
+= entry
->count
;
4486 rb_erase(node
, &(db
->bb_free_root
));
4487 spin_lock(&sbi
->s_md_lock
);
4488 list_del(&entry
->list
);
4489 spin_unlock(&sbi
->s_md_lock
);
4490 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4493 /* Add the extent to transaction's private list */
4494 spin_lock(&sbi
->s_md_lock
);
4495 list_add(&new_entry
->list
, &handle
->h_transaction
->t_private_list
);
4496 spin_unlock(&sbi
->s_md_lock
);
4501 * ext4_free_blocks() -- Free given blocks and update quota
4502 * @handle: handle for this transaction
4504 * @block: start physical block to free
4505 * @count: number of blocks to count
4506 * @metadata: Are these metadata blocks
4508 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4509 struct buffer_head
*bh
, ext4_fsblk_t block
,
4510 unsigned long count
, int flags
)
4512 struct buffer_head
*bitmap_bh
= NULL
;
4513 struct super_block
*sb
= inode
->i_sb
;
4514 struct ext4_group_desc
*gdp
;
4515 unsigned long freed
= 0;
4516 unsigned int overflow
;
4518 struct buffer_head
*gd_bh
;
4519 ext4_group_t block_group
;
4520 struct ext4_sb_info
*sbi
;
4521 struct ext4_buddy e4b
;
4527 BUG_ON(block
!= bh
->b_blocknr
);
4529 block
= bh
->b_blocknr
;
4533 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4534 !ext4_data_block_valid(sbi
, block
, count
)) {
4535 ext4_error(sb
, "Freeing blocks not in datazone - "
4536 "block = %llu, count = %lu", block
, count
);
4540 ext4_debug("freeing block %llu\n", block
);
4541 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4543 if (flags
& EXT4_FREE_BLOCKS_FORGET
) {
4544 struct buffer_head
*tbh
= bh
;
4547 BUG_ON(bh
&& (count
> 1));
4549 for (i
= 0; i
< count
; i
++) {
4551 tbh
= sb_find_get_block(inode
->i_sb
,
4555 ext4_forget(handle
, flags
& EXT4_FREE_BLOCKS_METADATA
,
4556 inode
, tbh
, block
+ i
);
4561 * We need to make sure we don't reuse the freed block until
4562 * after the transaction is committed, which we can do by
4563 * treating the block as metadata, below. We make an
4564 * exception if the inode is to be written in writeback mode
4565 * since writeback mode has weak data consistency guarantees.
4567 if (!ext4_should_writeback_data(inode
))
4568 flags
|= EXT4_FREE_BLOCKS_METADATA
;
4572 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4575 * Check to see if we are freeing blocks across a group
4578 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4579 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4582 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4587 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4593 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4594 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4595 in_range(block
, ext4_inode_table(sb
, gdp
),
4596 EXT4_SB(sb
)->s_itb_per_group
) ||
4597 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4598 EXT4_SB(sb
)->s_itb_per_group
)) {
4600 ext4_error(sb
, "Freeing blocks in system zone - "
4601 "Block = %llu, count = %lu", block
, count
);
4602 /* err = 0. ext4_std_error should be a no op */
4606 BUFFER_TRACE(bitmap_bh
, "getting write access");
4607 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4612 * We are about to modify some metadata. Call the journal APIs
4613 * to unshare ->b_data if a currently-committing transaction is
4616 BUFFER_TRACE(gd_bh
, "get_write_access");
4617 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4620 #ifdef AGGRESSIVE_CHECK
4623 for (i
= 0; i
< count
; i
++)
4624 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4627 trace_ext4_mballoc_free(sb
, inode
, block_group
, bit
, count
);
4629 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4633 if ((flags
& EXT4_FREE_BLOCKS_METADATA
) && ext4_handle_valid(handle
)) {
4634 struct ext4_free_data
*new_entry
;
4636 * blocks being freed are metadata. these blocks shouldn't
4637 * be used until this transaction is committed
4639 new_entry
= kmem_cache_alloc(ext4_free_ext_cachep
, GFP_NOFS
);
4644 new_entry
->start_blk
= bit
;
4645 new_entry
->group
= block_group
;
4646 new_entry
->count
= count
;
4647 new_entry
->t_tid
= handle
->h_transaction
->t_tid
;
4649 ext4_lock_group(sb
, block_group
);
4650 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4651 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4653 /* need to update group_info->bb_free and bitmap
4654 * with group lock held. generate_buddy look at
4655 * them with group lock_held
4657 ext4_lock_group(sb
, block_group
);
4658 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4659 mb_free_blocks(inode
, &e4b
, bit
, count
);
4662 ret
= ext4_free_blks_count(sb
, gdp
) + count
;
4663 ext4_free_blks_set(sb
, gdp
, ret
);
4664 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4665 ext4_unlock_group(sb
, block_group
);
4666 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4668 if (sbi
->s_log_groups_per_flex
) {
4669 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4670 atomic_add(count
, &sbi
->s_flex_groups
[flex_group
].free_blocks
);
4673 ext4_mb_unload_buddy(&e4b
);
4677 /* We dirtied the bitmap block */
4678 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4679 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4681 /* And the group descriptor block */
4682 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4683 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4687 if (overflow
&& !err
) {
4693 ext4_mark_super_dirty(sb
);
4696 dquot_free_block(inode
, freed
);
4698 ext4_std_error(sb
, err
);
4703 * ext4_trim_extent -- function to TRIM one single free extent in the group
4704 * @sb: super block for the file system
4705 * @start: starting block of the free extent in the alloc. group
4706 * @count: number of blocks to TRIM
4707 * @group: alloc. group we are working with
4708 * @e4b: ext4 buddy for the group
4710 * Trim "count" blocks starting at "start" in the "group". To assure that no
4711 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4712 * be called with under the group lock.
4714 static int ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
4715 ext4_group_t group
, struct ext4_buddy
*e4b
)
4717 struct ext4_free_extent ex
;
4720 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
4722 ex
.fe_start
= start
;
4723 ex
.fe_group
= group
;
4727 * Mark blocks used, so no one can reuse them while
4730 mb_mark_used(e4b
, &ex
);
4731 ext4_unlock_group(sb
, group
);
4733 ret
= ext4_issue_discard(sb
, group
, start
, count
);
4735 ext4_lock_group(sb
, group
);
4736 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
4741 * ext4_trim_all_free -- function to trim all free space in alloc. group
4742 * @sb: super block for file system
4744 * @start: first group block to examine
4745 * @max: last group block to examine
4746 * @minblocks: minimum extent block count
4748 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4749 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4753 * ext4_trim_all_free walks through group's block bitmap searching for free
4754 * extents. When the free extent is found, mark it as used in group buddy
4755 * bitmap. Then issue a TRIM command on this extent and free the extent in
4756 * the group buddy bitmap. This is done until whole group is scanned.
4758 ext4_grpblk_t
ext4_trim_all_free(struct super_block
*sb
, struct ext4_buddy
*e4b
,
4759 ext4_grpblk_t start
, ext4_grpblk_t max
, ext4_grpblk_t minblocks
)
4762 ext4_grpblk_t next
, count
= 0;
4766 BUG_ON(e4b
== NULL
);
4768 bitmap
= e4b
->bd_bitmap
;
4769 group
= e4b
->bd_group
;
4770 start
= (e4b
->bd_info
->bb_first_free
> start
) ?
4771 e4b
->bd_info
->bb_first_free
: start
;
4772 ext4_lock_group(sb
, group
);
4774 while (start
< max
) {
4775 start
= mb_find_next_zero_bit(bitmap
, max
, start
);
4778 next
= mb_find_next_bit(bitmap
, max
, start
);
4780 if ((next
- start
) >= minblocks
) {
4781 ret
= ext4_trim_extent(sb
, start
,
4782 next
- start
, group
, e4b
);
4785 count
+= next
- start
;
4789 if (fatal_signal_pending(current
)) {
4790 count
= -ERESTARTSYS
;
4794 if (need_resched()) {
4795 ext4_unlock_group(sb
, group
);
4797 ext4_lock_group(sb
, group
);
4800 if ((e4b
->bd_info
->bb_free
- count
) < minblocks
)
4803 ext4_unlock_group(sb
, group
);
4805 ext4_debug("trimmed %d blocks in the group %d\n",
4815 * ext4_trim_fs() -- trim ioctl handle function
4816 * @sb: superblock for filesystem
4817 * @range: fstrim_range structure
4819 * start: First Byte to trim
4820 * len: number of Bytes to trim from start
4821 * minlen: minimum extent length in Bytes
4822 * ext4_trim_fs goes through all allocation groups containing Bytes from
4823 * start to start+len. For each such a group ext4_trim_all_free function
4824 * is invoked to trim all free space.
4826 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
4828 struct ext4_buddy e4b
;
4829 ext4_group_t first_group
, last_group
;
4830 ext4_group_t group
, ngroups
= ext4_get_groups_count(sb
);
4831 ext4_grpblk_t cnt
= 0, first_block
, last_block
;
4832 uint64_t start
, len
, minlen
, trimmed
;
4833 ext4_fsblk_t first_data_blk
=
4834 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
4837 start
= range
->start
>> sb
->s_blocksize_bits
;
4838 len
= range
->len
>> sb
->s_blocksize_bits
;
4839 minlen
= range
->minlen
>> sb
->s_blocksize_bits
;
4842 if (unlikely(minlen
> EXT4_BLOCKS_PER_GROUP(sb
)))
4844 if (start
< first_data_blk
) {
4845 len
-= first_data_blk
- start
;
4846 start
= first_data_blk
;
4849 /* Determine first and last group to examine based on start and len */
4850 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
4851 &first_group
, &first_block
);
4852 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) (start
+ len
),
4853 &last_group
, &last_block
);
4854 last_group
= (last_group
> ngroups
- 1) ? ngroups
- 1 : last_group
;
4855 last_block
= EXT4_BLOCKS_PER_GROUP(sb
);
4857 if (first_group
> last_group
)
4860 for (group
= first_group
; group
<= last_group
; group
++) {
4861 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
4863 ext4_error(sb
, "Error in loading buddy "
4864 "information for %u", group
);
4868 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
))
4869 len
-= (EXT4_BLOCKS_PER_GROUP(sb
) - first_block
);
4871 last_block
= first_block
+ len
;
4873 if (e4b
.bd_info
->bb_free
>= minlen
) {
4874 cnt
= ext4_trim_all_free(sb
, &e4b
, first_block
,
4875 last_block
, minlen
);
4878 ext4_mb_unload_buddy(&e4b
);
4882 ext4_mb_unload_buddy(&e4b
);
4886 range
->len
= trimmed
* sb
->s_blocksize
;