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
24 #include "ext4_jbd2.h"
26 #include <linux/debugfs.h>
27 #include <linux/slab.h>
28 #include <trace/events/ext4.h>
32 * - test ext4_ext_search_left() and ext4_ext_search_right()
33 * - search for metadata in few groups
36 * - normalization should take into account whether file is still open
37 * - discard preallocations if no free space left (policy?)
38 * - don't normalize tails
40 * - reservation for superuser
43 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
44 * - track min/max extents in each group for better group selection
45 * - mb_mark_used() may allocate chunk right after splitting buddy
46 * - tree of groups sorted by number of free blocks
51 * The allocation request involve request for multiple number of blocks
52 * near to the goal(block) value specified.
54 * During initialization phase of the allocator we decide to use the
55 * group preallocation or inode preallocation depending on the size of
56 * the file. The size of the file could be the resulting file size we
57 * would have after allocation, or the current file size, which ever
58 * is larger. If the size is less than sbi->s_mb_stream_request we
59 * select to use the group preallocation. The default value of
60 * s_mb_stream_request is 16 blocks. This can also be tuned via
61 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
62 * terms of number of blocks.
64 * The main motivation for having small file use group preallocation is to
65 * ensure that we have small files closer together on the disk.
67 * First stage the allocator looks at the inode prealloc list,
68 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
69 * spaces for this particular inode. The inode prealloc space is
72 * pa_lstart -> the logical start block for this prealloc space
73 * pa_pstart -> the physical start block for this prealloc space
74 * pa_len -> length for this prealloc space (in clusters)
75 * pa_free -> free space available in this prealloc space (in clusters)
77 * The inode preallocation space is used looking at the _logical_ start
78 * block. If only the logical file block falls within the range of prealloc
79 * space we will consume the particular prealloc space. This makes sure that
80 * we have contiguous physical blocks representing the file blocks
82 * The important thing to be noted in case of inode prealloc space is that
83 * we don't modify the values associated to inode prealloc space except
86 * If we are not able to find blocks in the inode prealloc space and if we
87 * have the group allocation flag set then we look at the locality group
88 * prealloc space. These are per CPU prealloc list represented as
90 * ext4_sb_info.s_locality_groups[smp_processor_id()]
92 * The reason for having a per cpu locality group is to reduce the contention
93 * between CPUs. It is possible to get scheduled at this point.
95 * The locality group prealloc space is used looking at whether we have
96 * enough free space (pa_free) within the prealloc space.
98 * If we can't allocate blocks via inode prealloc or/and locality group
99 * prealloc then we look at the buddy cache. The buddy cache is represented
100 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
101 * mapped to the buddy and bitmap information regarding different
102 * groups. The buddy information is attached to buddy cache inode so that
103 * we can access them through the page cache. The information regarding
104 * each group is loaded via ext4_mb_load_buddy. The information involve
105 * block bitmap and buddy information. The information are stored in the
109 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
112 * one block each for bitmap and buddy information. So for each group we
113 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
114 * blocksize) blocks. So it can have information regarding groups_per_page
115 * which is blocks_per_page/2
117 * The buddy cache inode is not stored on disk. The inode is thrown
118 * away when the filesystem is unmounted.
120 * We look for count number of blocks in the buddy cache. If we were able
121 * to locate that many free blocks we return with additional information
122 * regarding rest of the contiguous physical block available
124 * Before allocating blocks via buddy cache we normalize the request
125 * blocks. This ensure we ask for more blocks that we needed. The extra
126 * blocks that we get after allocation is added to the respective prealloc
127 * list. In case of inode preallocation we follow a list of heuristics
128 * based on file size. This can be found in ext4_mb_normalize_request. If
129 * we are doing a group prealloc we try to normalize the request to
130 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
131 * dependent on the cluster size; for non-bigalloc file systems, it is
132 * 512 blocks. This can be tuned via
133 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
134 * terms of number of blocks. If we have mounted the file system with -O
135 * stripe=<value> option the group prealloc request is normalized to the
136 * the smallest multiple of the stripe value (sbi->s_stripe) which is
137 * greater than the default mb_group_prealloc.
139 * The regular allocator (using the buddy cache) supports a few tunables.
141 * /sys/fs/ext4/<partition>/mb_min_to_scan
142 * /sys/fs/ext4/<partition>/mb_max_to_scan
143 * /sys/fs/ext4/<partition>/mb_order2_req
145 * The regular allocator uses buddy scan only if the request len is power of
146 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
147 * value of s_mb_order2_reqs can be tuned via
148 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
149 * stripe size (sbi->s_stripe), we try to search for contiguous block in
150 * stripe size. This should result in better allocation on RAID setups. If
151 * not, we search in the specific group using bitmap for best extents. The
152 * tunable min_to_scan and max_to_scan control the behaviour here.
153 * min_to_scan indicate how long the mballoc __must__ look for a best
154 * extent and max_to_scan indicates how long the mballoc __can__ look for a
155 * best extent in the found extents. Searching for the blocks starts with
156 * the group specified as the goal value in allocation context via
157 * ac_g_ex. Each group is first checked based on the criteria whether it
158 * can be used for allocation. ext4_mb_good_group explains how the groups are
161 * Both the prealloc space are getting populated as above. So for the first
162 * request we will hit the buddy cache which will result in this prealloc
163 * space getting filled. The prealloc space is then later used for the
164 * subsequent request.
168 * mballoc operates on the following data:
170 * - in-core buddy (actually includes buddy and bitmap)
171 * - preallocation descriptors (PAs)
173 * there are two types of preallocations:
175 * assiged to specific inode and can be used for this inode only.
176 * it describes part of inode's space preallocated to specific
177 * physical blocks. any block from that preallocated can be used
178 * independent. the descriptor just tracks number of blocks left
179 * unused. so, before taking some block from descriptor, one must
180 * make sure corresponded logical block isn't allocated yet. this
181 * also means that freeing any block within descriptor's range
182 * must discard all preallocated blocks.
184 * assigned to specific locality group which does not translate to
185 * permanent set of inodes: inode can join and leave group. space
186 * from this type of preallocation can be used for any inode. thus
187 * it's consumed from the beginning to the end.
189 * relation between them can be expressed as:
190 * in-core buddy = on-disk bitmap + preallocation descriptors
192 * this mean blocks mballoc considers used are:
193 * - allocated blocks (persistent)
194 * - preallocated blocks (non-persistent)
196 * consistency in mballoc world means that at any time a block is either
197 * free or used in ALL structures. notice: "any time" should not be read
198 * literally -- time is discrete and delimited by locks.
200 * to keep it simple, we don't use block numbers, instead we count number of
201 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
203 * all operations can be expressed as:
204 * - init buddy: buddy = on-disk + PAs
205 * - new PA: buddy += N; PA = N
206 * - use inode PA: on-disk += N; PA -= N
207 * - discard inode PA buddy -= on-disk - PA; PA = 0
208 * - use locality group PA on-disk += N; PA -= N
209 * - discard locality group PA buddy -= PA; PA = 0
210 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
211 * is used in real operation because we can't know actual used
212 * bits from PA, only from on-disk bitmap
214 * if we follow this strict logic, then all operations above should be atomic.
215 * given some of them can block, we'd have to use something like semaphores
216 * killing performance on high-end SMP hardware. let's try to relax it using
217 * the following knowledge:
218 * 1) if buddy is referenced, it's already initialized
219 * 2) while block is used in buddy and the buddy is referenced,
220 * nobody can re-allocate that block
221 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
222 * bit set and PA claims same block, it's OK. IOW, one can set bit in
223 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
226 * so, now we're building a concurrency table:
229 * blocks for PA are allocated in the buddy, buddy must be referenced
230 * until PA is linked to allocation group to avoid concurrent buddy init
232 * we need to make sure that either on-disk bitmap or PA has uptodate data
233 * given (3) we care that PA-=N operation doesn't interfere with init
235 * the simplest way would be to have buddy initialized by the discard
236 * - use locality group PA
237 * again PA-=N must be serialized with init
238 * - discard locality group PA
239 * the simplest way would be to have buddy initialized by the discard
242 * i_data_sem serializes them
244 * discard process must wait until PA isn't used by another process
245 * - use locality group PA
246 * some mutex should serialize them
247 * - discard locality group PA
248 * discard process must wait until PA isn't used by another process
251 * i_data_sem or another mutex should serializes them
253 * discard process must wait until PA isn't used by another process
254 * - use locality group PA
255 * nothing wrong here -- they're different PAs covering different blocks
256 * - discard locality group PA
257 * discard process must wait until PA isn't used by another process
259 * now we're ready to make few consequences:
260 * - PA is referenced and while it is no discard is possible
261 * - PA is referenced until block isn't marked in on-disk bitmap
262 * - PA changes only after on-disk bitmap
263 * - discard must not compete with init. either init is done before
264 * any discard or they're serialized somehow
265 * - buddy init as sum of on-disk bitmap and PAs is done atomically
267 * a special case when we've used PA to emptiness. no need to modify buddy
268 * in this case, but we should care about concurrent init
273 * Logic in few words:
278 * mark bits in on-disk bitmap
281 * - use preallocation:
282 * find proper PA (per-inode or group)
284 * mark bits in on-disk bitmap
290 * mark bits in on-disk bitmap
293 * - discard preallocations in group:
295 * move them onto local list
296 * load on-disk bitmap
298 * remove PA from object (inode or locality group)
299 * mark free blocks in-core
301 * - discard inode's preallocations:
308 * - bitlock on a group (group)
309 * - object (inode/locality) (object)
320 * - release consumed pa:
325 * - generate in-core bitmap:
329 * - discard all for given object (inode, locality group):
334 * - discard all for given group:
341 static struct kmem_cache
*ext4_pspace_cachep
;
342 static struct kmem_cache
*ext4_ac_cachep
;
343 static struct kmem_cache
*ext4_free_data_cachep
;
345 /* We create slab caches for groupinfo data structures based on the
346 * superblock block size. There will be one per mounted filesystem for
347 * each unique s_blocksize_bits */
348 #define NR_GRPINFO_CACHES 8
349 static struct kmem_cache
*ext4_groupinfo_caches
[NR_GRPINFO_CACHES
];
351 static const char *ext4_groupinfo_slab_names
[NR_GRPINFO_CACHES
] = {
352 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
353 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
354 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
357 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
359 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
361 static void ext4_free_data_callback(struct super_block
*sb
,
362 struct ext4_journal_cb_entry
*jce
, int rc
);
364 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
366 #if BITS_PER_LONG == 64
367 *bit
+= ((unsigned long) addr
& 7UL) << 3;
368 addr
= (void *) ((unsigned long) addr
& ~7UL);
369 #elif BITS_PER_LONG == 32
370 *bit
+= ((unsigned long) addr
& 3UL) << 3;
371 addr
= (void *) ((unsigned long) addr
& ~3UL);
373 #error "how many bits you are?!"
378 static inline int mb_test_bit(int bit
, void *addr
)
381 * ext4_test_bit on architecture like powerpc
382 * needs unsigned long aligned address
384 addr
= mb_correct_addr_and_bit(&bit
, addr
);
385 return ext4_test_bit(bit
, addr
);
388 static inline void mb_set_bit(int bit
, void *addr
)
390 addr
= mb_correct_addr_and_bit(&bit
, addr
);
391 ext4_set_bit(bit
, addr
);
394 static inline void mb_clear_bit(int bit
, void *addr
)
396 addr
= mb_correct_addr_and_bit(&bit
, addr
);
397 ext4_clear_bit(bit
, addr
);
400 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
402 int fix
= 0, ret
, tmpmax
;
403 addr
= mb_correct_addr_and_bit(&fix
, addr
);
407 ret
= ext4_find_next_zero_bit(addr
, tmpmax
, start
) - fix
;
413 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
415 int fix
= 0, ret
, tmpmax
;
416 addr
= mb_correct_addr_and_bit(&fix
, addr
);
420 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
426 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
430 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
433 if (order
> e4b
->bd_blkbits
+ 1) {
438 /* at order 0 we see each particular block */
440 *max
= 1 << (e4b
->bd_blkbits
+ 3);
441 return e4b
->bd_bitmap
;
444 bb
= e4b
->bd_buddy
+ EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
445 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
451 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
452 int first
, int count
)
455 struct super_block
*sb
= e4b
->bd_sb
;
457 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
459 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
460 for (i
= 0; i
< count
; i
++) {
461 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
462 ext4_fsblk_t blocknr
;
464 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
465 blocknr
+= EXT4_C2B(EXT4_SB(sb
), first
+ i
);
466 ext4_grp_locked_error(sb
, e4b
->bd_group
,
467 inode
? inode
->i_ino
: 0,
469 "freeing block already freed "
473 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
477 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
481 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
483 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
484 for (i
= 0; i
< count
; i
++) {
485 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
486 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
490 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
492 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
493 unsigned char *b1
, *b2
;
495 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
496 b2
= (unsigned char *) bitmap
;
497 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
498 if (b1
[i
] != b2
[i
]) {
499 ext4_msg(e4b
->bd_sb
, KERN_ERR
,
500 "corruption in group %u "
501 "at byte %u(%u): %x in copy != %x "
503 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
511 static inline void mb_free_blocks_double(struct inode
*inode
,
512 struct ext4_buddy
*e4b
, int first
, int count
)
516 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
517 int first
, int count
)
521 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
527 #ifdef AGGRESSIVE_CHECK
529 #define MB_CHECK_ASSERT(assert) \
533 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
534 function, file, line, # assert); \
539 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
540 const char *function
, int line
)
542 struct super_block
*sb
= e4b
->bd_sb
;
543 int order
= e4b
->bd_blkbits
+ 1;
550 struct ext4_group_info
*grp
;
553 struct list_head
*cur
;
558 static int mb_check_counter
;
559 if (mb_check_counter
++ % 100 != 0)
564 buddy
= mb_find_buddy(e4b
, order
, &max
);
565 MB_CHECK_ASSERT(buddy
);
566 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
567 MB_CHECK_ASSERT(buddy2
);
568 MB_CHECK_ASSERT(buddy
!= buddy2
);
569 MB_CHECK_ASSERT(max
* 2 == max2
);
572 for (i
= 0; i
< max
; i
++) {
574 if (mb_test_bit(i
, buddy
)) {
575 /* only single bit in buddy2 may be 1 */
576 if (!mb_test_bit(i
<< 1, buddy2
)) {
578 mb_test_bit((i
<<1)+1, buddy2
));
579 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
581 mb_test_bit(i
<< 1, buddy2
));
586 /* both bits in buddy2 must be 1 */
587 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
588 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
590 for (j
= 0; j
< (1 << order
); j
++) {
591 k
= (i
* (1 << order
)) + j
;
593 !mb_test_bit(k
, e4b
->bd_bitmap
));
597 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
602 buddy
= mb_find_buddy(e4b
, 0, &max
);
603 for (i
= 0; i
< max
; i
++) {
604 if (!mb_test_bit(i
, buddy
)) {
605 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
613 /* check used bits only */
614 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
615 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
617 MB_CHECK_ASSERT(k
< max2
);
618 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
621 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
622 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
624 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
625 list_for_each(cur
, &grp
->bb_prealloc_list
) {
626 ext4_group_t groupnr
;
627 struct ext4_prealloc_space
*pa
;
628 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
629 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
630 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
631 for (i
= 0; i
< pa
->pa_len
; i
++)
632 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
636 #undef MB_CHECK_ASSERT
637 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
638 __FILE__, __func__, __LINE__)
640 #define mb_check_buddy(e4b)
644 * Divide blocks started from @first with length @len into
645 * smaller chunks with power of 2 blocks.
646 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
647 * then increase bb_counters[] for corresponded chunk size.
649 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
650 void *buddy
, ext4_grpblk_t first
, ext4_grpblk_t len
,
651 struct ext4_group_info
*grp
)
653 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
657 unsigned short border
;
659 BUG_ON(len
> EXT4_CLUSTERS_PER_GROUP(sb
));
661 border
= 2 << sb
->s_blocksize_bits
;
664 /* find how many blocks can be covered since this position */
665 max
= ffs(first
| border
) - 1;
667 /* find how many blocks of power 2 we need to mark */
674 /* mark multiblock chunks only */
675 grp
->bb_counters
[min
]++;
677 mb_clear_bit(first
>> min
,
678 buddy
+ sbi
->s_mb_offsets
[min
]);
686 * Cache the order of the largest free extent we have available in this block
690 mb_set_largest_free_order(struct super_block
*sb
, struct ext4_group_info
*grp
)
695 grp
->bb_largest_free_order
= -1; /* uninit */
697 bits
= sb
->s_blocksize_bits
+ 1;
698 for (i
= bits
; i
>= 0; i
--) {
699 if (grp
->bb_counters
[i
] > 0) {
700 grp
->bb_largest_free_order
= i
;
706 static noinline_for_stack
707 void ext4_mb_generate_buddy(struct super_block
*sb
,
708 void *buddy
, void *bitmap
, ext4_group_t group
)
710 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
711 ext4_grpblk_t max
= EXT4_CLUSTERS_PER_GROUP(sb
);
716 unsigned fragments
= 0;
717 unsigned long long period
= get_cycles();
719 /* initialize buddy from bitmap which is aggregation
720 * of on-disk bitmap and preallocations */
721 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
722 grp
->bb_first_free
= i
;
726 i
= mb_find_next_bit(bitmap
, max
, i
);
730 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
732 grp
->bb_counters
[0]++;
734 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
736 grp
->bb_fragments
= fragments
;
738 if (free
!= grp
->bb_free
) {
739 ext4_grp_locked_error(sb
, group
, 0, 0,
740 "%u clusters in bitmap, %u in gd",
743 * If we intent to continue, we consider group descritor
744 * corrupt and update bb_free using bitmap value
748 mb_set_largest_free_order(sb
, grp
);
750 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
752 period
= get_cycles() - period
;
753 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
754 EXT4_SB(sb
)->s_mb_buddies_generated
++;
755 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
756 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
759 /* The buddy information is attached the buddy cache inode
760 * for convenience. The information regarding each group
761 * is loaded via ext4_mb_load_buddy. The information involve
762 * block bitmap and buddy information. The information are
763 * stored in the inode as
766 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
769 * one block each for bitmap and buddy information.
770 * So for each group we take up 2 blocks. A page can
771 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
772 * So it can have information regarding groups_per_page which
773 * is blocks_per_page/2
775 * Locking note: This routine takes the block group lock of all groups
776 * for this page; do not hold this lock when calling this routine!
779 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
781 ext4_group_t ngroups
;
787 ext4_group_t first_group
, group
;
789 struct super_block
*sb
;
790 struct buffer_head
*bhs
;
791 struct buffer_head
**bh
= NULL
;
795 struct ext4_group_info
*grinfo
;
797 mb_debug(1, "init page %lu\n", page
->index
);
799 inode
= page
->mapping
->host
;
801 ngroups
= ext4_get_groups_count(sb
);
802 blocksize
= 1 << inode
->i_blkbits
;
803 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
805 groups_per_page
= blocks_per_page
>> 1;
806 if (groups_per_page
== 0)
809 /* allocate buffer_heads to read bitmaps */
810 if (groups_per_page
> 1) {
811 i
= sizeof(struct buffer_head
*) * groups_per_page
;
812 bh
= kzalloc(i
, GFP_NOFS
);
820 first_group
= page
->index
* blocks_per_page
/ 2;
822 /* read all groups the page covers into the cache */
823 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
824 if (group
>= ngroups
)
827 grinfo
= ext4_get_group_info(sb
, group
);
829 * If page is uptodate then we came here after online resize
830 * which added some new uninitialized group info structs, so
831 * we must skip all initialized uptodate buddies on the page,
832 * which may be currently in use by an allocating task.
834 if (PageUptodate(page
) && !EXT4_MB_GRP_NEED_INIT(grinfo
)) {
838 if (!(bh
[i
] = ext4_read_block_bitmap_nowait(sb
, group
))) {
842 mb_debug(1, "read bitmap for group %u\n", group
);
845 /* wait for I/O completion */
846 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
847 if (bh
[i
] && ext4_wait_block_bitmap(sb
, group
, bh
[i
])) {
853 first_block
= page
->index
* blocks_per_page
;
854 for (i
= 0; i
< blocks_per_page
; i
++) {
857 group
= (first_block
+ i
) >> 1;
858 if (group
>= ngroups
)
861 if (!bh
[group
- first_group
])
862 /* skip initialized uptodate buddy */
866 * data carry information regarding this
867 * particular group in the format specified
871 data
= page_address(page
) + (i
* blocksize
);
872 bitmap
= bh
[group
- first_group
]->b_data
;
875 * We place the buddy block and bitmap block
878 if ((first_block
+ i
) & 1) {
879 /* this is block of buddy */
880 BUG_ON(incore
== NULL
);
881 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
882 group
, page
->index
, i
* blocksize
);
883 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
884 grinfo
= ext4_get_group_info(sb
, group
);
885 grinfo
->bb_fragments
= 0;
886 memset(grinfo
->bb_counters
, 0,
887 sizeof(*grinfo
->bb_counters
) *
888 (sb
->s_blocksize_bits
+2));
890 * incore got set to the group block bitmap below
892 ext4_lock_group(sb
, group
);
894 memset(data
, 0xff, blocksize
);
895 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
896 ext4_unlock_group(sb
, group
);
899 /* this is block of bitmap */
900 BUG_ON(incore
!= NULL
);
901 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
902 group
, page
->index
, i
* blocksize
);
903 trace_ext4_mb_bitmap_load(sb
, group
);
905 /* see comments in ext4_mb_put_pa() */
906 ext4_lock_group(sb
, group
);
907 memcpy(data
, bitmap
, blocksize
);
909 /* mark all preallocated blks used in in-core bitmap */
910 ext4_mb_generate_from_pa(sb
, data
, group
);
911 ext4_mb_generate_from_freelist(sb
, data
, group
);
912 ext4_unlock_group(sb
, group
);
914 /* set incore so that the buddy information can be
915 * generated using this
920 SetPageUptodate(page
);
924 for (i
= 0; i
< groups_per_page
; i
++)
933 * Lock the buddy and bitmap pages. This make sure other parallel init_group
934 * on the same buddy page doesn't happen whild holding the buddy page lock.
935 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
936 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
938 static int ext4_mb_get_buddy_page_lock(struct super_block
*sb
,
939 ext4_group_t group
, struct ext4_buddy
*e4b
)
941 struct inode
*inode
= EXT4_SB(sb
)->s_buddy_cache
;
942 int block
, pnum
, poff
;
946 e4b
->bd_buddy_page
= NULL
;
947 e4b
->bd_bitmap_page
= NULL
;
949 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
951 * the buddy cache inode stores the block bitmap
952 * and buddy information in consecutive blocks.
953 * So for each group we need two blocks.
956 pnum
= block
/ blocks_per_page
;
957 poff
= block
% blocks_per_page
;
958 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
961 BUG_ON(page
->mapping
!= inode
->i_mapping
);
962 e4b
->bd_bitmap_page
= page
;
963 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
965 if (blocks_per_page
>= 2) {
966 /* buddy and bitmap are on the same page */
971 pnum
= block
/ blocks_per_page
;
972 poff
= block
% blocks_per_page
;
973 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
976 BUG_ON(page
->mapping
!= inode
->i_mapping
);
977 e4b
->bd_buddy_page
= page
;
981 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy
*e4b
)
983 if (e4b
->bd_bitmap_page
) {
984 unlock_page(e4b
->bd_bitmap_page
);
985 page_cache_release(e4b
->bd_bitmap_page
);
987 if (e4b
->bd_buddy_page
) {
988 unlock_page(e4b
->bd_buddy_page
);
989 page_cache_release(e4b
->bd_buddy_page
);
994 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
995 * block group lock of all groups for this page; do not hold the BG lock when
996 * calling this routine!
998 static noinline_for_stack
999 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
)
1002 struct ext4_group_info
*this_grp
;
1003 struct ext4_buddy e4b
;
1007 mb_debug(1, "init group %u\n", group
);
1008 this_grp
= ext4_get_group_info(sb
, group
);
1010 * This ensures that we don't reinit the buddy cache
1011 * page which map to the group from which we are already
1012 * allocating. If we are looking at the buddy cache we would
1013 * have taken a reference using ext4_mb_load_buddy and that
1014 * would have pinned buddy page to page cache.
1016 ret
= ext4_mb_get_buddy_page_lock(sb
, group
, &e4b
);
1017 if (ret
|| !EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1019 * somebody initialized the group
1020 * return without doing anything
1025 page
= e4b
.bd_bitmap_page
;
1026 ret
= ext4_mb_init_cache(page
, NULL
);
1029 if (!PageUptodate(page
)) {
1033 mark_page_accessed(page
);
1035 if (e4b
.bd_buddy_page
== NULL
) {
1037 * If both the bitmap and buddy are in
1038 * the same page we don't need to force
1044 /* init buddy cache */
1045 page
= e4b
.bd_buddy_page
;
1046 ret
= ext4_mb_init_cache(page
, e4b
.bd_bitmap
);
1049 if (!PageUptodate(page
)) {
1053 mark_page_accessed(page
);
1055 ext4_mb_put_buddy_page_lock(&e4b
);
1060 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1061 * block group lock of all groups for this page; do not hold the BG lock when
1062 * calling this routine!
1064 static noinline_for_stack
int
1065 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1066 struct ext4_buddy
*e4b
)
1068 int blocks_per_page
;
1074 struct ext4_group_info
*grp
;
1075 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1076 struct inode
*inode
= sbi
->s_buddy_cache
;
1078 mb_debug(1, "load group %u\n", group
);
1080 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1081 grp
= ext4_get_group_info(sb
, group
);
1083 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1086 e4b
->bd_group
= group
;
1087 e4b
->bd_buddy_page
= NULL
;
1088 e4b
->bd_bitmap_page
= NULL
;
1090 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1092 * we need full data about the group
1093 * to make a good selection
1095 ret
= ext4_mb_init_group(sb
, group
);
1101 * the buddy cache inode stores the block bitmap
1102 * and buddy information in consecutive blocks.
1103 * So for each group we need two blocks.
1106 pnum
= block
/ blocks_per_page
;
1107 poff
= block
% blocks_per_page
;
1109 /* we could use find_or_create_page(), but it locks page
1110 * what we'd like to avoid in fast path ... */
1111 page
= find_get_page(inode
->i_mapping
, pnum
);
1112 if (page
== NULL
|| !PageUptodate(page
)) {
1115 * drop the page reference and try
1116 * to get the page with lock. If we
1117 * are not uptodate that implies
1118 * somebody just created the page but
1119 * is yet to initialize the same. So
1120 * wait for it to initialize.
1122 page_cache_release(page
);
1123 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1125 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1126 if (!PageUptodate(page
)) {
1127 ret
= ext4_mb_init_cache(page
, NULL
);
1132 mb_cmp_bitmaps(e4b
, page_address(page
) +
1133 (poff
* sb
->s_blocksize
));
1138 if (page
== NULL
|| !PageUptodate(page
)) {
1142 e4b
->bd_bitmap_page
= page
;
1143 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1144 mark_page_accessed(page
);
1147 pnum
= block
/ blocks_per_page
;
1148 poff
= block
% blocks_per_page
;
1150 page
= find_get_page(inode
->i_mapping
, pnum
);
1151 if (page
== NULL
|| !PageUptodate(page
)) {
1153 page_cache_release(page
);
1154 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1156 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1157 if (!PageUptodate(page
)) {
1158 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1167 if (page
== NULL
|| !PageUptodate(page
)) {
1171 e4b
->bd_buddy_page
= page
;
1172 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1173 mark_page_accessed(page
);
1175 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1176 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1182 page_cache_release(page
);
1183 if (e4b
->bd_bitmap_page
)
1184 page_cache_release(e4b
->bd_bitmap_page
);
1185 if (e4b
->bd_buddy_page
)
1186 page_cache_release(e4b
->bd_buddy_page
);
1187 e4b
->bd_buddy
= NULL
;
1188 e4b
->bd_bitmap
= NULL
;
1192 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1194 if (e4b
->bd_bitmap_page
)
1195 page_cache_release(e4b
->bd_bitmap_page
);
1196 if (e4b
->bd_buddy_page
)
1197 page_cache_release(e4b
->bd_buddy_page
);
1201 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1206 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
1207 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1210 while (order
<= e4b
->bd_blkbits
+ 1) {
1212 if (!mb_test_bit(block
, bb
)) {
1213 /* this block is part of buddy of order 'order' */
1216 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1222 static void mb_clear_bits(void *bm
, int cur
, int len
)
1228 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1229 /* fast path: clear whole word at once */
1230 addr
= bm
+ (cur
>> 3);
1235 mb_clear_bit(cur
, bm
);
1240 void ext4_set_bits(void *bm
, int cur
, int len
)
1246 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1247 /* fast path: set whole word at once */
1248 addr
= bm
+ (cur
>> 3);
1253 mb_set_bit(cur
, bm
);
1258 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1259 int first
, int count
)
1266 struct super_block
*sb
= e4b
->bd_sb
;
1268 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1269 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1270 mb_check_buddy(e4b
);
1271 mb_free_blocks_double(inode
, e4b
, first
, count
);
1273 e4b
->bd_info
->bb_free
+= count
;
1274 if (first
< e4b
->bd_info
->bb_first_free
)
1275 e4b
->bd_info
->bb_first_free
= first
;
1277 /* let's maintain fragments counter */
1279 block
= !mb_test_bit(first
- 1, e4b
->bd_bitmap
);
1280 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1281 max
= !mb_test_bit(first
+ count
, e4b
->bd_bitmap
);
1283 e4b
->bd_info
->bb_fragments
--;
1284 else if (!block
&& !max
)
1285 e4b
->bd_info
->bb_fragments
++;
1287 /* let's maintain buddy itself */
1288 while (count
-- > 0) {
1292 if (!mb_test_bit(block
, e4b
->bd_bitmap
)) {
1293 ext4_fsblk_t blocknr
;
1295 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1296 blocknr
+= EXT4_C2B(EXT4_SB(sb
), block
);
1297 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1298 inode
? inode
->i_ino
: 0,
1300 "freeing already freed block "
1303 mb_clear_bit(block
, e4b
->bd_bitmap
);
1304 e4b
->bd_info
->bb_counters
[order
]++;
1306 /* start of the buddy */
1307 buddy
= mb_find_buddy(e4b
, order
, &max
);
1311 if (mb_test_bit(block
, buddy
) ||
1312 mb_test_bit(block
+ 1, buddy
))
1315 /* both the buddies are free, try to coalesce them */
1316 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1322 /* for special purposes, we don't set
1323 * free bits in bitmap */
1324 mb_set_bit(block
, buddy
);
1325 mb_set_bit(block
+ 1, buddy
);
1327 e4b
->bd_info
->bb_counters
[order
]--;
1328 e4b
->bd_info
->bb_counters
[order
]--;
1332 e4b
->bd_info
->bb_counters
[order
]++;
1334 mb_clear_bit(block
, buddy2
);
1338 mb_set_largest_free_order(sb
, e4b
->bd_info
);
1339 mb_check_buddy(e4b
);
1342 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1343 int needed
, struct ext4_free_extent
*ex
)
1349 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1352 buddy
= mb_find_buddy(e4b
, order
, &max
);
1353 BUG_ON(buddy
== NULL
);
1354 BUG_ON(block
>= max
);
1355 if (mb_test_bit(block
, buddy
)) {
1362 /* FIXME dorp order completely ? */
1363 if (likely(order
== 0)) {
1364 /* find actual order */
1365 order
= mb_find_order_for_block(e4b
, block
);
1366 block
= block
>> order
;
1369 ex
->fe_len
= 1 << order
;
1370 ex
->fe_start
= block
<< order
;
1371 ex
->fe_group
= e4b
->bd_group
;
1373 /* calc difference from given start */
1374 next
= next
- ex
->fe_start
;
1376 ex
->fe_start
+= next
;
1378 while (needed
> ex
->fe_len
&&
1379 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1381 if (block
+ 1 >= max
)
1384 next
= (block
+ 1) * (1 << order
);
1385 if (mb_test_bit(next
, e4b
->bd_bitmap
))
1388 order
= mb_find_order_for_block(e4b
, next
);
1390 block
= next
>> order
;
1391 ex
->fe_len
+= 1 << order
;
1394 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1398 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1404 int start
= ex
->fe_start
;
1405 int len
= ex
->fe_len
;
1410 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1411 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1412 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1413 mb_check_buddy(e4b
);
1414 mb_mark_used_double(e4b
, start
, len
);
1416 e4b
->bd_info
->bb_free
-= len
;
1417 if (e4b
->bd_info
->bb_first_free
== start
)
1418 e4b
->bd_info
->bb_first_free
+= len
;
1420 /* let's maintain fragments counter */
1422 mlen
= !mb_test_bit(start
- 1, e4b
->bd_bitmap
);
1423 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1424 max
= !mb_test_bit(start
+ len
, e4b
->bd_bitmap
);
1426 e4b
->bd_info
->bb_fragments
++;
1427 else if (!mlen
&& !max
)
1428 e4b
->bd_info
->bb_fragments
--;
1430 /* let's maintain buddy itself */
1432 ord
= mb_find_order_for_block(e4b
, start
);
1434 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1435 /* the whole chunk may be allocated at once! */
1437 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1438 BUG_ON((start
>> ord
) >= max
);
1439 mb_set_bit(start
>> ord
, buddy
);
1440 e4b
->bd_info
->bb_counters
[ord
]--;
1447 /* store for history */
1449 ret
= len
| (ord
<< 16);
1451 /* we have to split large buddy */
1453 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1454 mb_set_bit(start
>> ord
, buddy
);
1455 e4b
->bd_info
->bb_counters
[ord
]--;
1458 cur
= (start
>> ord
) & ~1U;
1459 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1460 mb_clear_bit(cur
, buddy
);
1461 mb_clear_bit(cur
+ 1, buddy
);
1462 e4b
->bd_info
->bb_counters
[ord
]++;
1463 e4b
->bd_info
->bb_counters
[ord
]++;
1465 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1467 ext4_set_bits(e4b
->bd_bitmap
, ex
->fe_start
, len0
);
1468 mb_check_buddy(e4b
);
1474 * Must be called under group lock!
1476 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1477 struct ext4_buddy
*e4b
)
1479 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1482 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1483 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1485 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1486 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1487 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1489 /* preallocation can change ac_b_ex, thus we store actually
1490 * allocated blocks for history */
1491 ac
->ac_f_ex
= ac
->ac_b_ex
;
1493 ac
->ac_status
= AC_STATUS_FOUND
;
1494 ac
->ac_tail
= ret
& 0xffff;
1495 ac
->ac_buddy
= ret
>> 16;
1498 * take the page reference. We want the page to be pinned
1499 * so that we don't get a ext4_mb_init_cache_call for this
1500 * group until we update the bitmap. That would mean we
1501 * double allocate blocks. The reference is dropped
1502 * in ext4_mb_release_context
1504 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1505 get_page(ac
->ac_bitmap_page
);
1506 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1507 get_page(ac
->ac_buddy_page
);
1508 /* store last allocated for subsequent stream allocation */
1509 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1510 spin_lock(&sbi
->s_md_lock
);
1511 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1512 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1513 spin_unlock(&sbi
->s_md_lock
);
1518 * regular allocator, for general purposes allocation
1521 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1522 struct ext4_buddy
*e4b
,
1525 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1526 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1527 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1528 struct ext4_free_extent ex
;
1531 if (ac
->ac_status
== AC_STATUS_FOUND
)
1534 * We don't want to scan for a whole year
1536 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1537 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1538 ac
->ac_status
= AC_STATUS_BREAK
;
1543 * Haven't found good chunk so far, let's continue
1545 if (bex
->fe_len
< gex
->fe_len
)
1548 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1549 && bex
->fe_group
== e4b
->bd_group
) {
1550 /* recheck chunk's availability - we don't know
1551 * when it was found (within this lock-unlock
1553 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1554 if (max
>= gex
->fe_len
) {
1555 ext4_mb_use_best_found(ac
, e4b
);
1562 * The routine checks whether found extent is good enough. If it is,
1563 * then the extent gets marked used and flag is set to the context
1564 * to stop scanning. Otherwise, the extent is compared with the
1565 * previous found extent and if new one is better, then it's stored
1566 * in the context. Later, the best found extent will be used, if
1567 * mballoc can't find good enough extent.
1569 * FIXME: real allocation policy is to be designed yet!
1571 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1572 struct ext4_free_extent
*ex
,
1573 struct ext4_buddy
*e4b
)
1575 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1576 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1578 BUG_ON(ex
->fe_len
<= 0);
1579 BUG_ON(ex
->fe_len
> EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1580 BUG_ON(ex
->fe_start
>= EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1581 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1586 * The special case - take what you catch first
1588 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1590 ext4_mb_use_best_found(ac
, e4b
);
1595 * Let's check whether the chuck is good enough
1597 if (ex
->fe_len
== gex
->fe_len
) {
1599 ext4_mb_use_best_found(ac
, e4b
);
1604 * If this is first found extent, just store it in the context
1606 if (bex
->fe_len
== 0) {
1612 * If new found extent is better, store it in the context
1614 if (bex
->fe_len
< gex
->fe_len
) {
1615 /* if the request isn't satisfied, any found extent
1616 * larger than previous best one is better */
1617 if (ex
->fe_len
> bex
->fe_len
)
1619 } else if (ex
->fe_len
> gex
->fe_len
) {
1620 /* if the request is satisfied, then we try to find
1621 * an extent that still satisfy the request, but is
1622 * smaller than previous one */
1623 if (ex
->fe_len
< bex
->fe_len
)
1627 ext4_mb_check_limits(ac
, e4b
, 0);
1630 static noinline_for_stack
1631 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1632 struct ext4_buddy
*e4b
)
1634 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1635 ext4_group_t group
= ex
.fe_group
;
1639 BUG_ON(ex
.fe_len
<= 0);
1640 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1644 ext4_lock_group(ac
->ac_sb
, group
);
1645 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1649 ext4_mb_use_best_found(ac
, e4b
);
1652 ext4_unlock_group(ac
->ac_sb
, group
);
1653 ext4_mb_unload_buddy(e4b
);
1658 static noinline_for_stack
1659 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1660 struct ext4_buddy
*e4b
)
1662 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1665 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1666 struct ext4_free_extent ex
;
1668 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1671 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1675 ext4_lock_group(ac
->ac_sb
, group
);
1676 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1677 ac
->ac_g_ex
.fe_len
, &ex
);
1679 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1682 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1684 /* use do_div to get remainder (would be 64-bit modulo) */
1685 if (do_div(start
, sbi
->s_stripe
) == 0) {
1688 ext4_mb_use_best_found(ac
, e4b
);
1690 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1691 BUG_ON(ex
.fe_len
<= 0);
1692 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1693 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1696 ext4_mb_use_best_found(ac
, e4b
);
1697 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1698 /* Sometimes, caller may want to merge even small
1699 * number of blocks to an existing extent */
1700 BUG_ON(ex
.fe_len
<= 0);
1701 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1702 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1705 ext4_mb_use_best_found(ac
, e4b
);
1707 ext4_unlock_group(ac
->ac_sb
, group
);
1708 ext4_mb_unload_buddy(e4b
);
1714 * The routine scans buddy structures (not bitmap!) from given order
1715 * to max order and tries to find big enough chunk to satisfy the req
1717 static noinline_for_stack
1718 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1719 struct ext4_buddy
*e4b
)
1721 struct super_block
*sb
= ac
->ac_sb
;
1722 struct ext4_group_info
*grp
= e4b
->bd_info
;
1728 BUG_ON(ac
->ac_2order
<= 0);
1729 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1730 if (grp
->bb_counters
[i
] == 0)
1733 buddy
= mb_find_buddy(e4b
, i
, &max
);
1734 BUG_ON(buddy
== NULL
);
1736 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1741 ac
->ac_b_ex
.fe_len
= 1 << i
;
1742 ac
->ac_b_ex
.fe_start
= k
<< i
;
1743 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1745 ext4_mb_use_best_found(ac
, e4b
);
1747 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1749 if (EXT4_SB(sb
)->s_mb_stats
)
1750 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1757 * The routine scans the group and measures all found extents.
1758 * In order to optimize scanning, caller must pass number of
1759 * free blocks in the group, so the routine can know upper limit.
1761 static noinline_for_stack
1762 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1763 struct ext4_buddy
*e4b
)
1765 struct super_block
*sb
= ac
->ac_sb
;
1766 void *bitmap
= e4b
->bd_bitmap
;
1767 struct ext4_free_extent ex
;
1771 free
= e4b
->bd_info
->bb_free
;
1774 i
= e4b
->bd_info
->bb_first_free
;
1776 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1777 i
= mb_find_next_zero_bit(bitmap
,
1778 EXT4_CLUSTERS_PER_GROUP(sb
), i
);
1779 if (i
>= EXT4_CLUSTERS_PER_GROUP(sb
)) {
1781 * IF we have corrupt bitmap, we won't find any
1782 * free blocks even though group info says we
1783 * we have free blocks
1785 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1786 "%d free clusters as per "
1787 "group info. But bitmap says 0",
1792 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1793 BUG_ON(ex
.fe_len
<= 0);
1794 if (free
< ex
.fe_len
) {
1795 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1796 "%d free clusters as per "
1797 "group info. But got %d blocks",
1800 * The number of free blocks differs. This mostly
1801 * indicate that the bitmap is corrupt. So exit
1802 * without claiming the space.
1807 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1813 ext4_mb_check_limits(ac
, e4b
, 1);
1817 * This is a special case for storages like raid5
1818 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1820 static noinline_for_stack
1821 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1822 struct ext4_buddy
*e4b
)
1824 struct super_block
*sb
= ac
->ac_sb
;
1825 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1826 void *bitmap
= e4b
->bd_bitmap
;
1827 struct ext4_free_extent ex
;
1828 ext4_fsblk_t first_group_block
;
1833 BUG_ON(sbi
->s_stripe
== 0);
1835 /* find first stripe-aligned block in group */
1836 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1838 a
= first_group_block
+ sbi
->s_stripe
- 1;
1839 do_div(a
, sbi
->s_stripe
);
1840 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1842 while (i
< EXT4_CLUSTERS_PER_GROUP(sb
)) {
1843 if (!mb_test_bit(i
, bitmap
)) {
1844 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1845 if (max
>= sbi
->s_stripe
) {
1848 ext4_mb_use_best_found(ac
, e4b
);
1856 /* This is now called BEFORE we load the buddy bitmap. */
1857 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1858 ext4_group_t group
, int cr
)
1860 unsigned free
, fragments
;
1861 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
1862 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1864 BUG_ON(cr
< 0 || cr
>= 4);
1866 /* We only do this if the grp has never been initialized */
1867 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1868 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
);
1873 free
= grp
->bb_free
;
1874 fragments
= grp
->bb_fragments
;
1882 BUG_ON(ac
->ac_2order
== 0);
1884 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
1887 /* Avoid using the first bg of a flexgroup for data files */
1888 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
1889 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
1890 ((group
% flex_size
) == 0))
1895 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1899 if (free
>= ac
->ac_g_ex
.fe_len
)
1911 static noinline_for_stack
int
1912 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
1914 ext4_group_t ngroups
, group
, i
;
1917 struct ext4_sb_info
*sbi
;
1918 struct super_block
*sb
;
1919 struct ext4_buddy e4b
;
1923 ngroups
= ext4_get_groups_count(sb
);
1924 /* non-extent files are limited to low blocks/groups */
1925 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
1926 ngroups
= sbi
->s_blockfile_groups
;
1928 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1930 /* first, try the goal */
1931 err
= ext4_mb_find_by_goal(ac
, &e4b
);
1932 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
1935 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
1939 * ac->ac2_order is set only if the fe_len is a power of 2
1940 * if ac2_order is set we also set criteria to 0 so that we
1941 * try exact allocation using buddy.
1943 i
= fls(ac
->ac_g_ex
.fe_len
);
1946 * We search using buddy data only if the order of the request
1947 * is greater than equal to the sbi_s_mb_order2_reqs
1948 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1950 if (i
>= sbi
->s_mb_order2_reqs
) {
1952 * This should tell if fe_len is exactly power of 2
1954 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
1955 ac
->ac_2order
= i
- 1;
1958 /* if stream allocation is enabled, use global goal */
1959 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1960 /* TBD: may be hot point */
1961 spin_lock(&sbi
->s_md_lock
);
1962 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
1963 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
1964 spin_unlock(&sbi
->s_md_lock
);
1967 /* Let's just scan groups to find more-less suitable blocks */
1968 cr
= ac
->ac_2order
? 0 : 1;
1970 * cr == 0 try to get exact allocation,
1971 * cr == 3 try to get anything
1974 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
1975 ac
->ac_criteria
= cr
;
1977 * searching for the right group start
1978 * from the goal value specified
1980 group
= ac
->ac_g_ex
.fe_group
;
1982 for (i
= 0; i
< ngroups
; group
++, i
++) {
1983 if (group
== ngroups
)
1986 /* This now checks without needing the buddy page */
1987 if (!ext4_mb_good_group(ac
, group
, cr
))
1990 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
1994 ext4_lock_group(sb
, group
);
1997 * We need to check again after locking the
2000 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2001 ext4_unlock_group(sb
, group
);
2002 ext4_mb_unload_buddy(&e4b
);
2006 ac
->ac_groups_scanned
++;
2008 ext4_mb_simple_scan_group(ac
, &e4b
);
2009 else if (cr
== 1 && sbi
->s_stripe
&&
2010 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2011 ext4_mb_scan_aligned(ac
, &e4b
);
2013 ext4_mb_complex_scan_group(ac
, &e4b
);
2015 ext4_unlock_group(sb
, group
);
2016 ext4_mb_unload_buddy(&e4b
);
2018 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2023 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2024 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2026 * We've been searching too long. Let's try to allocate
2027 * the best chunk we've found so far
2030 ext4_mb_try_best_found(ac
, &e4b
);
2031 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2033 * Someone more lucky has already allocated it.
2034 * The only thing we can do is just take first
2036 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2038 ac
->ac_b_ex
.fe_group
= 0;
2039 ac
->ac_b_ex
.fe_start
= 0;
2040 ac
->ac_b_ex
.fe_len
= 0;
2041 ac
->ac_status
= AC_STATUS_CONTINUE
;
2042 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2044 atomic_inc(&sbi
->s_mb_lost_chunks
);
2052 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2054 struct super_block
*sb
= seq
->private;
2057 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2060 return (void *) ((unsigned long) group
);
2063 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2065 struct super_block
*sb
= seq
->private;
2069 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2072 return (void *) ((unsigned long) group
);
2075 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2077 struct super_block
*sb
= seq
->private;
2078 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2081 struct ext4_buddy e4b
;
2083 struct ext4_group_info info
;
2084 ext4_grpblk_t counters
[16];
2089 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2090 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2091 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2092 "group", "free", "frags", "first",
2093 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2094 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2096 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2097 sizeof(struct ext4_group_info
);
2098 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2100 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2103 ext4_lock_group(sb
, group
);
2104 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2105 ext4_unlock_group(sb
, group
);
2106 ext4_mb_unload_buddy(&e4b
);
2108 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2109 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2110 for (i
= 0; i
<= 13; i
++)
2111 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2112 sg
.info
.bb_counters
[i
] : 0);
2113 seq_printf(seq
, " ]\n");
2118 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2122 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2123 .start
= ext4_mb_seq_groups_start
,
2124 .next
= ext4_mb_seq_groups_next
,
2125 .stop
= ext4_mb_seq_groups_stop
,
2126 .show
= ext4_mb_seq_groups_show
,
2129 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2131 struct super_block
*sb
= PDE(inode
)->data
;
2134 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2136 struct seq_file
*m
= file
->private_data
;
2143 static const struct file_operations ext4_mb_seq_groups_fops
= {
2144 .owner
= THIS_MODULE
,
2145 .open
= ext4_mb_seq_groups_open
,
2147 .llseek
= seq_lseek
,
2148 .release
= seq_release
,
2151 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2153 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2154 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2160 /* Create and initialize ext4_group_info data for the given group. */
2161 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2162 struct ext4_group_desc
*desc
)
2166 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2167 struct ext4_group_info
**meta_group_info
;
2168 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2171 * First check if this group is the first of a reserved block.
2172 * If it's true, we have to allocate a new table of pointers
2173 * to ext4_group_info structures
2175 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2176 metalen
= sizeof(*meta_group_info
) <<
2177 EXT4_DESC_PER_BLOCK_BITS(sb
);
2178 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2179 if (meta_group_info
== NULL
) {
2180 ext4_msg(sb
, KERN_ERR
, "can't allocate mem "
2181 "for a buddy group");
2182 goto exit_meta_group_info
;
2184 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2189 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2190 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2192 meta_group_info
[i
] = kmem_cache_alloc(cachep
, GFP_KERNEL
);
2193 if (meta_group_info
[i
] == NULL
) {
2194 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy mem");
2195 goto exit_group_info
;
2197 memset(meta_group_info
[i
], 0, kmem_cache_size(cachep
));
2198 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2199 &(meta_group_info
[i
]->bb_state
));
2202 * initialize bb_free to be able to skip
2203 * empty groups without initialization
2205 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2206 meta_group_info
[i
]->bb_free
=
2207 ext4_free_clusters_after_init(sb
, group
, desc
);
2209 meta_group_info
[i
]->bb_free
=
2210 ext4_free_group_clusters(sb
, desc
);
2213 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2214 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2215 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2216 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2220 struct buffer_head
*bh
;
2221 meta_group_info
[i
]->bb_bitmap
=
2222 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2223 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2224 bh
= ext4_read_block_bitmap(sb
, group
);
2226 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2235 /* If a meta_group_info table has been allocated, release it now */
2236 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2237 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2238 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] = NULL
;
2240 exit_meta_group_info
:
2242 } /* ext4_mb_add_groupinfo */
2244 static int ext4_mb_init_backend(struct super_block
*sb
)
2246 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2248 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2249 struct ext4_super_block
*es
= sbi
->s_es
;
2250 int num_meta_group_infos
;
2251 int num_meta_group_infos_max
;
2253 struct ext4_group_desc
*desc
;
2254 struct kmem_cache
*cachep
;
2256 /* This is the number of blocks used by GDT */
2257 num_meta_group_infos
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) -
2258 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2261 * This is the total number of blocks used by GDT including
2262 * the number of reserved blocks for GDT.
2263 * The s_group_info array is allocated with this value
2264 * to allow a clean online resize without a complex
2265 * manipulation of pointer.
2266 * The drawback is the unused memory when no resize
2267 * occurs but it's very low in terms of pages
2268 * (see comments below)
2269 * Need to handle this properly when META_BG resizing is allowed
2271 num_meta_group_infos_max
= num_meta_group_infos
+
2272 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2275 * array_size is the size of s_group_info array. We round it
2276 * to the next power of two because this approximation is done
2277 * internally by kmalloc so we can have some more memory
2278 * for free here (e.g. may be used for META_BG resize).
2281 while (array_size
< sizeof(*sbi
->s_group_info
) *
2282 num_meta_group_infos_max
)
2283 array_size
= array_size
<< 1;
2284 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2285 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2286 * So a two level scheme suffices for now. */
2287 sbi
->s_group_info
= ext4_kvzalloc(array_size
, GFP_KERNEL
);
2288 if (sbi
->s_group_info
== NULL
) {
2289 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy meta group");
2292 sbi
->s_buddy_cache
= new_inode(sb
);
2293 if (sbi
->s_buddy_cache
== NULL
) {
2294 ext4_msg(sb
, KERN_ERR
, "can't get new inode");
2297 /* To avoid potentially colliding with an valid on-disk inode number,
2298 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2299 * not in the inode hash, so it should never be found by iget(), but
2300 * this will avoid confusion if it ever shows up during debugging. */
2301 sbi
->s_buddy_cache
->i_ino
= EXT4_BAD_INO
;
2302 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2303 for (i
= 0; i
< ngroups
; i
++) {
2304 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2306 ext4_msg(sb
, KERN_ERR
, "can't read descriptor %u", i
);
2309 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2316 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2318 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2319 i
= num_meta_group_infos
;
2321 kfree(sbi
->s_group_info
[i
]);
2322 iput(sbi
->s_buddy_cache
);
2324 ext4_kvfree(sbi
->s_group_info
);
2328 static void ext4_groupinfo_destroy_slabs(void)
2332 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2333 if (ext4_groupinfo_caches
[i
])
2334 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2335 ext4_groupinfo_caches
[i
] = NULL
;
2339 static int ext4_groupinfo_create_slab(size_t size
)
2341 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2343 int blocksize_bits
= order_base_2(size
);
2344 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2345 struct kmem_cache
*cachep
;
2347 if (cache_index
>= NR_GRPINFO_CACHES
)
2350 if (unlikely(cache_index
< 0))
2353 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2354 if (ext4_groupinfo_caches
[cache_index
]) {
2355 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2356 return 0; /* Already created */
2359 slab_size
= offsetof(struct ext4_group_info
,
2360 bb_counters
[blocksize_bits
+ 2]);
2362 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2363 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2366 ext4_groupinfo_caches
[cache_index
] = cachep
;
2368 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2371 "EXT4-fs: no memory for groupinfo slab cache\n");
2378 int ext4_mb_init(struct super_block
*sb
)
2380 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2386 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2388 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2389 if (sbi
->s_mb_offsets
== NULL
) {
2394 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2395 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2396 if (sbi
->s_mb_maxs
== NULL
) {
2401 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2405 /* order 0 is regular bitmap */
2406 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2407 sbi
->s_mb_offsets
[0] = 0;
2411 max
= sb
->s_blocksize
<< 2;
2413 sbi
->s_mb_offsets
[i
] = offset
;
2414 sbi
->s_mb_maxs
[i
] = max
;
2415 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2418 } while (i
<= sb
->s_blocksize_bits
+ 1);
2420 spin_lock_init(&sbi
->s_md_lock
);
2421 spin_lock_init(&sbi
->s_bal_lock
);
2423 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2424 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2425 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2426 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2427 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2429 * The default group preallocation is 512, which for 4k block
2430 * sizes translates to 2 megabytes. However for bigalloc file
2431 * systems, this is probably too big (i.e, if the cluster size
2432 * is 1 megabyte, then group preallocation size becomes half a
2433 * gigabyte!). As a default, we will keep a two megabyte
2434 * group pralloc size for cluster sizes up to 64k, and after
2435 * that, we will force a minimum group preallocation size of
2436 * 32 clusters. This translates to 8 megs when the cluster
2437 * size is 256k, and 32 megs when the cluster size is 1 meg,
2438 * which seems reasonable as a default.
2440 sbi
->s_mb_group_prealloc
= max(MB_DEFAULT_GROUP_PREALLOC
>>
2441 sbi
->s_cluster_bits
, 32);
2443 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2444 * to the lowest multiple of s_stripe which is bigger than
2445 * the s_mb_group_prealloc as determined above. We want
2446 * the preallocation size to be an exact multiple of the
2447 * RAID stripe size so that preallocations don't fragment
2450 if (sbi
->s_stripe
> 1) {
2451 sbi
->s_mb_group_prealloc
= roundup(
2452 sbi
->s_mb_group_prealloc
, sbi
->s_stripe
);
2455 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2456 if (sbi
->s_locality_groups
== NULL
) {
2458 goto out_free_groupinfo_slab
;
2460 for_each_possible_cpu(i
) {
2461 struct ext4_locality_group
*lg
;
2462 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2463 mutex_init(&lg
->lg_mutex
);
2464 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2465 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2466 spin_lock_init(&lg
->lg_prealloc_lock
);
2469 /* init file for buddy data */
2470 ret
= ext4_mb_init_backend(sb
);
2472 goto out_free_locality_groups
;
2475 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2476 &ext4_mb_seq_groups_fops
, sb
);
2480 out_free_locality_groups
:
2481 free_percpu(sbi
->s_locality_groups
);
2482 sbi
->s_locality_groups
= NULL
;
2483 out_free_groupinfo_slab
:
2484 ext4_groupinfo_destroy_slabs();
2486 kfree(sbi
->s_mb_offsets
);
2487 sbi
->s_mb_offsets
= NULL
;
2488 kfree(sbi
->s_mb_maxs
);
2489 sbi
->s_mb_maxs
= NULL
;
2493 /* need to called with the ext4 group lock held */
2494 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2496 struct ext4_prealloc_space
*pa
;
2497 struct list_head
*cur
, *tmp
;
2500 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2501 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2502 list_del(&pa
->pa_group_list
);
2504 kmem_cache_free(ext4_pspace_cachep
, pa
);
2507 mb_debug(1, "mballoc: %u PAs left\n", count
);
2511 int ext4_mb_release(struct super_block
*sb
)
2513 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2515 int num_meta_group_infos
;
2516 struct ext4_group_info
*grinfo
;
2517 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2518 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2521 remove_proc_entry("mb_groups", sbi
->s_proc
);
2523 if (sbi
->s_group_info
) {
2524 for (i
= 0; i
< ngroups
; i
++) {
2525 grinfo
= ext4_get_group_info(sb
, i
);
2527 kfree(grinfo
->bb_bitmap
);
2529 ext4_lock_group(sb
, i
);
2530 ext4_mb_cleanup_pa(grinfo
);
2531 ext4_unlock_group(sb
, i
);
2532 kmem_cache_free(cachep
, grinfo
);
2534 num_meta_group_infos
= (ngroups
+
2535 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2536 EXT4_DESC_PER_BLOCK_BITS(sb
);
2537 for (i
= 0; i
< num_meta_group_infos
; i
++)
2538 kfree(sbi
->s_group_info
[i
]);
2539 ext4_kvfree(sbi
->s_group_info
);
2541 kfree(sbi
->s_mb_offsets
);
2542 kfree(sbi
->s_mb_maxs
);
2543 if (sbi
->s_buddy_cache
)
2544 iput(sbi
->s_buddy_cache
);
2545 if (sbi
->s_mb_stats
) {
2546 ext4_msg(sb
, KERN_INFO
,
2547 "mballoc: %u blocks %u reqs (%u success)",
2548 atomic_read(&sbi
->s_bal_allocated
),
2549 atomic_read(&sbi
->s_bal_reqs
),
2550 atomic_read(&sbi
->s_bal_success
));
2551 ext4_msg(sb
, KERN_INFO
,
2552 "mballoc: %u extents scanned, %u goal hits, "
2553 "%u 2^N hits, %u breaks, %u lost",
2554 atomic_read(&sbi
->s_bal_ex_scanned
),
2555 atomic_read(&sbi
->s_bal_goals
),
2556 atomic_read(&sbi
->s_bal_2orders
),
2557 atomic_read(&sbi
->s_bal_breaks
),
2558 atomic_read(&sbi
->s_mb_lost_chunks
));
2559 ext4_msg(sb
, KERN_INFO
,
2560 "mballoc: %lu generated and it took %Lu",
2561 sbi
->s_mb_buddies_generated
,
2562 sbi
->s_mb_generation_time
);
2563 ext4_msg(sb
, KERN_INFO
,
2564 "mballoc: %u preallocated, %u discarded",
2565 atomic_read(&sbi
->s_mb_preallocated
),
2566 atomic_read(&sbi
->s_mb_discarded
));
2569 free_percpu(sbi
->s_locality_groups
);
2574 static inline int ext4_issue_discard(struct super_block
*sb
,
2575 ext4_group_t block_group
, ext4_grpblk_t cluster
, int count
)
2577 ext4_fsblk_t discard_block
;
2579 discard_block
= (EXT4_C2B(EXT4_SB(sb
), cluster
) +
2580 ext4_group_first_block_no(sb
, block_group
));
2581 count
= EXT4_C2B(EXT4_SB(sb
), count
);
2582 trace_ext4_discard_blocks(sb
,
2583 (unsigned long long) discard_block
, count
);
2584 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, 0);
2588 * This function is called by the jbd2 layer once the commit has finished,
2589 * so we know we can free the blocks that were released with that commit.
2591 static void ext4_free_data_callback(struct super_block
*sb
,
2592 struct ext4_journal_cb_entry
*jce
,
2595 struct ext4_free_data
*entry
= (struct ext4_free_data
*)jce
;
2596 struct ext4_buddy e4b
;
2597 struct ext4_group_info
*db
;
2598 int err
, count
= 0, count2
= 0;
2600 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2601 entry
->efd_count
, entry
->efd_group
, entry
);
2603 if (test_opt(sb
, DISCARD
))
2604 ext4_issue_discard(sb
, entry
->efd_group
,
2605 entry
->efd_start_cluster
, entry
->efd_count
);
2607 err
= ext4_mb_load_buddy(sb
, entry
->efd_group
, &e4b
);
2608 /* we expect to find existing buddy because it's pinned */
2613 /* there are blocks to put in buddy to make them really free */
2614 count
+= entry
->efd_count
;
2616 ext4_lock_group(sb
, entry
->efd_group
);
2617 /* Take it out of per group rb tree */
2618 rb_erase(&entry
->efd_node
, &(db
->bb_free_root
));
2619 mb_free_blocks(NULL
, &e4b
, entry
->efd_start_cluster
, entry
->efd_count
);
2622 * Clear the trimmed flag for the group so that the next
2623 * ext4_trim_fs can trim it.
2624 * If the volume is mounted with -o discard, online discard
2625 * is supported and the free blocks will be trimmed online.
2627 if (!test_opt(sb
, DISCARD
))
2628 EXT4_MB_GRP_CLEAR_TRIMMED(db
);
2630 if (!db
->bb_free_root
.rb_node
) {
2631 /* No more items in the per group rb tree
2632 * balance refcounts from ext4_mb_free_metadata()
2634 page_cache_release(e4b
.bd_buddy_page
);
2635 page_cache_release(e4b
.bd_bitmap_page
);
2637 ext4_unlock_group(sb
, entry
->efd_group
);
2638 kmem_cache_free(ext4_free_data_cachep
, entry
);
2639 ext4_mb_unload_buddy(&e4b
);
2641 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2644 #ifdef CONFIG_EXT4_DEBUG
2645 u8 mb_enable_debug __read_mostly
;
2647 static struct dentry
*debugfs_dir
;
2648 static struct dentry
*debugfs_debug
;
2650 static void __init
ext4_create_debugfs_entry(void)
2652 debugfs_dir
= debugfs_create_dir("ext4", NULL
);
2654 debugfs_debug
= debugfs_create_u8("mballoc-debug",
2660 static void ext4_remove_debugfs_entry(void)
2662 debugfs_remove(debugfs_debug
);
2663 debugfs_remove(debugfs_dir
);
2668 static void __init
ext4_create_debugfs_entry(void)
2672 static void ext4_remove_debugfs_entry(void)
2678 int __init
ext4_init_mballoc(void)
2680 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2681 SLAB_RECLAIM_ACCOUNT
);
2682 if (ext4_pspace_cachep
== NULL
)
2685 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2686 SLAB_RECLAIM_ACCOUNT
);
2687 if (ext4_ac_cachep
== NULL
) {
2688 kmem_cache_destroy(ext4_pspace_cachep
);
2692 ext4_free_data_cachep
= KMEM_CACHE(ext4_free_data
,
2693 SLAB_RECLAIM_ACCOUNT
);
2694 if (ext4_free_data_cachep
== NULL
) {
2695 kmem_cache_destroy(ext4_pspace_cachep
);
2696 kmem_cache_destroy(ext4_ac_cachep
);
2699 ext4_create_debugfs_entry();
2703 void ext4_exit_mballoc(void)
2706 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2707 * before destroying the slab cache.
2710 kmem_cache_destroy(ext4_pspace_cachep
);
2711 kmem_cache_destroy(ext4_ac_cachep
);
2712 kmem_cache_destroy(ext4_free_data_cachep
);
2713 ext4_groupinfo_destroy_slabs();
2714 ext4_remove_debugfs_entry();
2719 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2720 * Returns 0 if success or error code
2722 static noinline_for_stack
int
2723 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2724 handle_t
*handle
, unsigned int reserv_clstrs
)
2726 struct buffer_head
*bitmap_bh
= NULL
;
2727 struct ext4_group_desc
*gdp
;
2728 struct buffer_head
*gdp_bh
;
2729 struct ext4_sb_info
*sbi
;
2730 struct super_block
*sb
;
2734 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2735 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2741 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2745 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2750 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2754 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2755 ext4_free_group_clusters(sb
, gdp
));
2757 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2761 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
2763 len
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
2764 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2765 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2766 "fs metadata", block
, block
+len
);
2767 /* File system mounted not to panic on error
2768 * Fix the bitmap and repeat the block allocation
2769 * We leak some of the blocks here.
2771 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2772 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2773 ac
->ac_b_ex
.fe_len
);
2774 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2775 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2781 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2782 #ifdef AGGRESSIVE_CHECK
2785 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2786 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2787 bitmap_bh
->b_data
));
2791 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2792 ac
->ac_b_ex
.fe_len
);
2793 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2794 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2795 ext4_free_group_clusters_set(sb
, gdp
,
2796 ext4_free_clusters_after_init(sb
,
2797 ac
->ac_b_ex
.fe_group
, gdp
));
2799 len
= ext4_free_group_clusters(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
2800 ext4_free_group_clusters_set(sb
, gdp
, len
);
2801 ext4_block_bitmap_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
, bitmap_bh
,
2802 EXT4_BLOCKS_PER_GROUP(sb
) / 8);
2803 ext4_group_desc_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
);
2805 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2806 percpu_counter_sub(&sbi
->s_freeclusters_counter
, ac
->ac_b_ex
.fe_len
);
2808 * Now reduce the dirty block count also. Should not go negative
2810 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2811 /* release all the reserved blocks if non delalloc */
2812 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
2815 if (sbi
->s_log_groups_per_flex
) {
2816 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2817 ac
->ac_b_ex
.fe_group
);
2818 atomic_sub(ac
->ac_b_ex
.fe_len
,
2819 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
2822 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2825 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
2828 ext4_mark_super_dirty(sb
);
2834 * here we normalize request for locality group
2835 * Group request are normalized to s_mb_group_prealloc, which goes to
2836 * s_strip if we set the same via mount option.
2837 * s_mb_group_prealloc can be configured via
2838 * /sys/fs/ext4/<partition>/mb_group_prealloc
2840 * XXX: should we try to preallocate more than the group has now?
2842 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2844 struct super_block
*sb
= ac
->ac_sb
;
2845 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2848 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2849 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2850 current
->pid
, ac
->ac_g_ex
.fe_len
);
2854 * Normalization means making request better in terms of
2855 * size and alignment
2857 static noinline_for_stack
void
2858 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
2859 struct ext4_allocation_request
*ar
)
2861 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2864 loff_t size
, start_off
;
2865 loff_t orig_size __maybe_unused
;
2867 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
2868 struct ext4_prealloc_space
*pa
;
2870 /* do normalize only data requests, metadata requests
2871 do not need preallocation */
2872 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
2875 /* sometime caller may want exact blocks */
2876 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2879 /* caller may indicate that preallocation isn't
2880 * required (it's a tail, for example) */
2881 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
2884 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
2885 ext4_mb_normalize_group_request(ac
);
2889 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
2891 /* first, let's learn actual file size
2892 * given current request is allocated */
2893 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
2894 size
= size
<< bsbits
;
2895 if (size
< i_size_read(ac
->ac_inode
))
2896 size
= i_size_read(ac
->ac_inode
);
2899 /* max size of free chunks */
2902 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2903 (req <= (size) || max <= (chunk_size))
2905 /* first, try to predict filesize */
2906 /* XXX: should this table be tunable? */
2908 if (size
<= 16 * 1024) {
2910 } else if (size
<= 32 * 1024) {
2912 } else if (size
<= 64 * 1024) {
2914 } else if (size
<= 128 * 1024) {
2916 } else if (size
<= 256 * 1024) {
2918 } else if (size
<= 512 * 1024) {
2920 } else if (size
<= 1024 * 1024) {
2922 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
2923 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2924 (21 - bsbits
)) << 21;
2925 size
= 2 * 1024 * 1024;
2926 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
2927 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2928 (22 - bsbits
)) << 22;
2929 size
= 4 * 1024 * 1024;
2930 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
2931 (8<<20)>>bsbits
, max
, 8 * 1024)) {
2932 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
2933 (23 - bsbits
)) << 23;
2934 size
= 8 * 1024 * 1024;
2936 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
2937 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
2939 size
= size
>> bsbits
;
2940 start
= start_off
>> bsbits
;
2942 /* don't cover already allocated blocks in selected range */
2943 if (ar
->pleft
&& start
<= ar
->lleft
) {
2944 size
-= ar
->lleft
+ 1 - start
;
2945 start
= ar
->lleft
+ 1;
2947 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
2948 size
-= start
+ size
- ar
->lright
;
2952 /* check we don't cross already preallocated blocks */
2954 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
2959 spin_lock(&pa
->pa_lock
);
2960 if (pa
->pa_deleted
) {
2961 spin_unlock(&pa
->pa_lock
);
2965 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
2968 /* PA must not overlap original request */
2969 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
2970 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
2972 /* skip PAs this normalized request doesn't overlap with */
2973 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
2974 spin_unlock(&pa
->pa_lock
);
2977 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
2979 /* adjust start or end to be adjacent to this pa */
2980 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
2981 BUG_ON(pa_end
< start
);
2983 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
2984 BUG_ON(pa
->pa_lstart
> end
);
2985 end
= pa
->pa_lstart
;
2987 spin_unlock(&pa
->pa_lock
);
2992 /* XXX: extra loop to check we really don't overlap preallocations */
2994 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
2997 spin_lock(&pa
->pa_lock
);
2998 if (pa
->pa_deleted
== 0) {
2999 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3001 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3003 spin_unlock(&pa
->pa_lock
);
3007 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3008 start
> ac
->ac_o_ex
.fe_logical
) {
3009 ext4_msg(ac
->ac_sb
, KERN_ERR
,
3010 "start %lu, size %lu, fe_logical %lu",
3011 (unsigned long) start
, (unsigned long) size
,
3012 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3014 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3015 start
> ac
->ac_o_ex
.fe_logical
);
3016 BUG_ON(size
<= 0 || size
> EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
3018 /* now prepare goal request */
3020 /* XXX: is it better to align blocks WRT to logical
3021 * placement or satisfy big request as is */
3022 ac
->ac_g_ex
.fe_logical
= start
;
3023 ac
->ac_g_ex
.fe_len
= EXT4_NUM_B2C(sbi
, size
);
3025 /* define goal start in order to merge */
3026 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3027 /* merge to the right */
3028 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3029 &ac
->ac_f_ex
.fe_group
,
3030 &ac
->ac_f_ex
.fe_start
);
3031 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3033 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3034 /* merge to the left */
3035 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3036 &ac
->ac_f_ex
.fe_group
,
3037 &ac
->ac_f_ex
.fe_start
);
3038 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3041 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3042 (unsigned) orig_size
, (unsigned) start
);
3045 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3047 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3049 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3050 atomic_inc(&sbi
->s_bal_reqs
);
3051 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3052 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3053 atomic_inc(&sbi
->s_bal_success
);
3054 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3055 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3056 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3057 atomic_inc(&sbi
->s_bal_goals
);
3058 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3059 atomic_inc(&sbi
->s_bal_breaks
);
3062 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3063 trace_ext4_mballoc_alloc(ac
);
3065 trace_ext4_mballoc_prealloc(ac
);
3069 * Called on failure; free up any blocks from the inode PA for this
3070 * context. We don't need this for MB_GROUP_PA because we only change
3071 * pa_free in ext4_mb_release_context(), but on failure, we've already
3072 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3074 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3076 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3078 if (pa
&& pa
->pa_type
== MB_INODE_PA
)
3079 pa
->pa_free
+= ac
->ac_b_ex
.fe_len
;
3083 * use blocks preallocated to inode
3085 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3086 struct ext4_prealloc_space
*pa
)
3088 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3093 /* found preallocated blocks, use them */
3094 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3095 end
= min(pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
),
3096 start
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
));
3097 len
= EXT4_NUM_B2C(sbi
, end
- start
);
3098 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3099 &ac
->ac_b_ex
.fe_start
);
3100 ac
->ac_b_ex
.fe_len
= len
;
3101 ac
->ac_status
= AC_STATUS_FOUND
;
3104 BUG_ON(start
< pa
->pa_pstart
);
3105 BUG_ON(end
> pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
));
3106 BUG_ON(pa
->pa_free
< len
);
3109 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3113 * use blocks preallocated to locality group
3115 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3116 struct ext4_prealloc_space
*pa
)
3118 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3120 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3121 &ac
->ac_b_ex
.fe_group
,
3122 &ac
->ac_b_ex
.fe_start
);
3123 ac
->ac_b_ex
.fe_len
= len
;
3124 ac
->ac_status
= AC_STATUS_FOUND
;
3127 /* we don't correct pa_pstart or pa_plen here to avoid
3128 * possible race when the group is being loaded concurrently
3129 * instead we correct pa later, after blocks are marked
3130 * in on-disk bitmap -- see ext4_mb_release_context()
3131 * Other CPUs are prevented from allocating from this pa by lg_mutex
3133 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3137 * Return the prealloc space that have minimal distance
3138 * from the goal block. @cpa is the prealloc
3139 * space that is having currently known minimal distance
3140 * from the goal block.
3142 static struct ext4_prealloc_space
*
3143 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3144 struct ext4_prealloc_space
*pa
,
3145 struct ext4_prealloc_space
*cpa
)
3147 ext4_fsblk_t cur_distance
, new_distance
;
3150 atomic_inc(&pa
->pa_count
);
3153 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3154 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3156 if (cur_distance
<= new_distance
)
3159 /* drop the previous reference */
3160 atomic_dec(&cpa
->pa_count
);
3161 atomic_inc(&pa
->pa_count
);
3166 * search goal blocks in preallocated space
3168 static noinline_for_stack
int
3169 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3171 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3173 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3174 struct ext4_locality_group
*lg
;
3175 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3176 ext4_fsblk_t goal_block
;
3178 /* only data can be preallocated */
3179 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3182 /* first, try per-file preallocation */
3184 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3186 /* all fields in this condition don't change,
3187 * so we can skip locking for them */
3188 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3189 ac
->ac_o_ex
.fe_logical
>= (pa
->pa_lstart
+
3190 EXT4_C2B(sbi
, pa
->pa_len
)))
3193 /* non-extent files can't have physical blocks past 2^32 */
3194 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3195 (pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
) >
3196 EXT4_MAX_BLOCK_FILE_PHYS
))
3199 /* found preallocated blocks, use them */
3200 spin_lock(&pa
->pa_lock
);
3201 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3202 atomic_inc(&pa
->pa_count
);
3203 ext4_mb_use_inode_pa(ac
, pa
);
3204 spin_unlock(&pa
->pa_lock
);
3205 ac
->ac_criteria
= 10;
3209 spin_unlock(&pa
->pa_lock
);
3213 /* can we use group allocation? */
3214 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3217 /* inode may have no locality group for some reason */
3221 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3222 if (order
> PREALLOC_TB_SIZE
- 1)
3223 /* The max size of hash table is PREALLOC_TB_SIZE */
3224 order
= PREALLOC_TB_SIZE
- 1;
3226 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3228 * search for the prealloc space that is having
3229 * minimal distance from the goal block.
3231 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3233 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3235 spin_lock(&pa
->pa_lock
);
3236 if (pa
->pa_deleted
== 0 &&
3237 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3239 cpa
= ext4_mb_check_group_pa(goal_block
,
3242 spin_unlock(&pa
->pa_lock
);
3247 ext4_mb_use_group_pa(ac
, cpa
);
3248 ac
->ac_criteria
= 20;
3255 * the function goes through all block freed in the group
3256 * but not yet committed and marks them used in in-core bitmap.
3257 * buddy must be generated from this bitmap
3258 * Need to be called with the ext4 group lock held
3260 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3264 struct ext4_group_info
*grp
;
3265 struct ext4_free_data
*entry
;
3267 grp
= ext4_get_group_info(sb
, group
);
3268 n
= rb_first(&(grp
->bb_free_root
));
3271 entry
= rb_entry(n
, struct ext4_free_data
, efd_node
);
3272 ext4_set_bits(bitmap
, entry
->efd_start_cluster
, entry
->efd_count
);
3279 * the function goes through all preallocation in this group and marks them
3280 * used in in-core bitmap. buddy must be generated from this bitmap
3281 * Need to be called with ext4 group lock held
3283 static noinline_for_stack
3284 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3287 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3288 struct ext4_prealloc_space
*pa
;
3289 struct list_head
*cur
;
3290 ext4_group_t groupnr
;
3291 ext4_grpblk_t start
;
3292 int preallocated
= 0;
3295 /* all form of preallocation discards first load group,
3296 * so the only competing code is preallocation use.
3297 * we don't need any locking here
3298 * notice we do NOT ignore preallocations with pa_deleted
3299 * otherwise we could leave used blocks available for
3300 * allocation in buddy when concurrent ext4_mb_put_pa()
3301 * is dropping preallocation
3303 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3304 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3305 spin_lock(&pa
->pa_lock
);
3306 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3309 spin_unlock(&pa
->pa_lock
);
3310 if (unlikely(len
== 0))
3312 BUG_ON(groupnr
!= group
);
3313 ext4_set_bits(bitmap
, start
, len
);
3314 preallocated
+= len
;
3316 mb_debug(1, "prellocated %u for group %u\n", preallocated
, group
);
3319 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3321 struct ext4_prealloc_space
*pa
;
3322 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3323 kmem_cache_free(ext4_pspace_cachep
, pa
);
3327 * drops a reference to preallocated space descriptor
3328 * if this was the last reference and the space is consumed
3330 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3331 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3334 ext4_fsblk_t grp_blk
;
3336 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3339 /* in this short window concurrent discard can set pa_deleted */
3340 spin_lock(&pa
->pa_lock
);
3341 if (pa
->pa_deleted
== 1) {
3342 spin_unlock(&pa
->pa_lock
);
3347 spin_unlock(&pa
->pa_lock
);
3349 grp_blk
= pa
->pa_pstart
;
3351 * If doing group-based preallocation, pa_pstart may be in the
3352 * next group when pa is used up
3354 if (pa
->pa_type
== MB_GROUP_PA
)
3357 ext4_get_group_no_and_offset(sb
, grp_blk
, &grp
, NULL
);
3362 * P1 (buddy init) P2 (regular allocation)
3363 * find block B in PA
3364 * copy on-disk bitmap to buddy
3365 * mark B in on-disk bitmap
3366 * drop PA from group
3367 * mark all PAs in buddy
3369 * thus, P1 initializes buddy with B available. to prevent this
3370 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3373 ext4_lock_group(sb
, grp
);
3374 list_del(&pa
->pa_group_list
);
3375 ext4_unlock_group(sb
, grp
);
3377 spin_lock(pa
->pa_obj_lock
);
3378 list_del_rcu(&pa
->pa_inode_list
);
3379 spin_unlock(pa
->pa_obj_lock
);
3381 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3385 * creates new preallocated space for given inode
3387 static noinline_for_stack
int
3388 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3390 struct super_block
*sb
= ac
->ac_sb
;
3391 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3392 struct ext4_prealloc_space
*pa
;
3393 struct ext4_group_info
*grp
;
3394 struct ext4_inode_info
*ei
;
3396 /* preallocate only when found space is larger then requested */
3397 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3398 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3399 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3401 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3405 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3411 /* we can't allocate as much as normalizer wants.
3412 * so, found space must get proper lstart
3413 * to cover original request */
3414 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3415 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3417 /* we're limited by original request in that
3418 * logical block must be covered any way
3419 * winl is window we can move our chunk within */
3420 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3422 /* also, we should cover whole original request */
3423 wins
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
);
3425 /* the smallest one defines real window */
3426 win
= min(winl
, wins
);
3428 offs
= ac
->ac_o_ex
.fe_logical
%
3429 EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
3430 if (offs
&& offs
< win
)
3433 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
-
3435 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3436 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3439 /* preallocation can change ac_b_ex, thus we store actually
3440 * allocated blocks for history */
3441 ac
->ac_f_ex
= ac
->ac_b_ex
;
3443 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3444 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3445 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3446 pa
->pa_free
= pa
->pa_len
;
3447 atomic_set(&pa
->pa_count
, 1);
3448 spin_lock_init(&pa
->pa_lock
);
3449 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3450 INIT_LIST_HEAD(&pa
->pa_group_list
);
3452 pa
->pa_type
= MB_INODE_PA
;
3454 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3455 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3456 trace_ext4_mb_new_inode_pa(ac
, pa
);
3458 ext4_mb_use_inode_pa(ac
, pa
);
3459 atomic_add(pa
->pa_free
, &sbi
->s_mb_preallocated
);
3461 ei
= EXT4_I(ac
->ac_inode
);
3462 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3464 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3465 pa
->pa_inode
= ac
->ac_inode
;
3467 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3468 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3469 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3471 spin_lock(pa
->pa_obj_lock
);
3472 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3473 spin_unlock(pa
->pa_obj_lock
);
3479 * creates new preallocated space for locality group inodes belongs to
3481 static noinline_for_stack
int
3482 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3484 struct super_block
*sb
= ac
->ac_sb
;
3485 struct ext4_locality_group
*lg
;
3486 struct ext4_prealloc_space
*pa
;
3487 struct ext4_group_info
*grp
;
3489 /* preallocate only when found space is larger then requested */
3490 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3491 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3492 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3494 BUG_ON(ext4_pspace_cachep
== NULL
);
3495 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3499 /* preallocation can change ac_b_ex, thus we store actually
3500 * allocated blocks for history */
3501 ac
->ac_f_ex
= ac
->ac_b_ex
;
3503 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3504 pa
->pa_lstart
= pa
->pa_pstart
;
3505 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3506 pa
->pa_free
= pa
->pa_len
;
3507 atomic_set(&pa
->pa_count
, 1);
3508 spin_lock_init(&pa
->pa_lock
);
3509 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3510 INIT_LIST_HEAD(&pa
->pa_group_list
);
3512 pa
->pa_type
= MB_GROUP_PA
;
3514 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3515 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3516 trace_ext4_mb_new_group_pa(ac
, pa
);
3518 ext4_mb_use_group_pa(ac
, pa
);
3519 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3521 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3525 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3526 pa
->pa_inode
= NULL
;
3528 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3529 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3530 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3533 * We will later add the new pa to the right bucket
3534 * after updating the pa_free in ext4_mb_release_context
3539 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3543 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3544 err
= ext4_mb_new_group_pa(ac
);
3546 err
= ext4_mb_new_inode_pa(ac
);
3551 * finds all unused blocks in on-disk bitmap, frees them in
3552 * in-core bitmap and buddy.
3553 * @pa must be unlinked from inode and group lists, so that
3554 * nobody else can find/use it.
3555 * the caller MUST hold group/inode locks.
3556 * TODO: optimize the case when there are no in-core structures yet
3558 static noinline_for_stack
int
3559 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3560 struct ext4_prealloc_space
*pa
)
3562 struct super_block
*sb
= e4b
->bd_sb
;
3563 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3568 unsigned long long grp_blk_start
;
3572 BUG_ON(pa
->pa_deleted
== 0);
3573 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3574 grp_blk_start
= pa
->pa_pstart
- EXT4_C2B(sbi
, bit
);
3575 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3576 end
= bit
+ pa
->pa_len
;
3579 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3582 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3583 mb_debug(1, " free preallocated %u/%u in group %u\n",
3584 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3585 (unsigned) next
- bit
, (unsigned) group
);
3588 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3589 trace_ext4_mb_release_inode_pa(pa
, (grp_blk_start
+
3590 EXT4_C2B(sbi
, bit
)),
3592 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3595 if (free
!= pa
->pa_free
) {
3596 ext4_msg(e4b
->bd_sb
, KERN_CRIT
,
3597 "pa %p: logic %lu, phys. %lu, len %lu",
3598 pa
, (unsigned long) pa
->pa_lstart
,
3599 (unsigned long) pa
->pa_pstart
,
3600 (unsigned long) pa
->pa_len
);
3601 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3604 * pa is already deleted so we use the value obtained
3605 * from the bitmap and continue.
3608 atomic_add(free
, &sbi
->s_mb_discarded
);
3613 static noinline_for_stack
int
3614 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3615 struct ext4_prealloc_space
*pa
)
3617 struct super_block
*sb
= e4b
->bd_sb
;
3621 trace_ext4_mb_release_group_pa(sb
, pa
);
3622 BUG_ON(pa
->pa_deleted
== 0);
3623 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3624 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3625 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3626 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3627 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3633 * releases all preallocations in given group
3635 * first, we need to decide discard policy:
3636 * - when do we discard
3638 * - how many do we discard
3639 * 1) how many requested
3641 static noinline_for_stack
int
3642 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3643 ext4_group_t group
, int needed
)
3645 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3646 struct buffer_head
*bitmap_bh
= NULL
;
3647 struct ext4_prealloc_space
*pa
, *tmp
;
3648 struct list_head list
;
3649 struct ext4_buddy e4b
;
3654 mb_debug(1, "discard preallocation for group %u\n", group
);
3656 if (list_empty(&grp
->bb_prealloc_list
))
3659 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3660 if (bitmap_bh
== NULL
) {
3661 ext4_error(sb
, "Error reading block bitmap for %u", group
);
3665 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3667 ext4_error(sb
, "Error loading buddy information for %u", group
);
3673 needed
= EXT4_CLUSTERS_PER_GROUP(sb
) + 1;
3675 INIT_LIST_HEAD(&list
);
3677 ext4_lock_group(sb
, group
);
3678 list_for_each_entry_safe(pa
, tmp
,
3679 &grp
->bb_prealloc_list
, pa_group_list
) {
3680 spin_lock(&pa
->pa_lock
);
3681 if (atomic_read(&pa
->pa_count
)) {
3682 spin_unlock(&pa
->pa_lock
);
3686 if (pa
->pa_deleted
) {
3687 spin_unlock(&pa
->pa_lock
);
3691 /* seems this one can be freed ... */
3694 /* we can trust pa_free ... */
3695 free
+= pa
->pa_free
;
3697 spin_unlock(&pa
->pa_lock
);
3699 list_del(&pa
->pa_group_list
);
3700 list_add(&pa
->u
.pa_tmp_list
, &list
);
3703 /* if we still need more blocks and some PAs were used, try again */
3704 if (free
< needed
&& busy
) {
3706 ext4_unlock_group(sb
, group
);
3708 * Yield the CPU here so that we don't get soft lockup
3709 * in non preempt case.
3715 /* found anything to free? */
3716 if (list_empty(&list
)) {
3721 /* now free all selected PAs */
3722 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3724 /* remove from object (inode or locality group) */
3725 spin_lock(pa
->pa_obj_lock
);
3726 list_del_rcu(&pa
->pa_inode_list
);
3727 spin_unlock(pa
->pa_obj_lock
);
3729 if (pa
->pa_type
== MB_GROUP_PA
)
3730 ext4_mb_release_group_pa(&e4b
, pa
);
3732 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3734 list_del(&pa
->u
.pa_tmp_list
);
3735 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3739 ext4_unlock_group(sb
, group
);
3740 ext4_mb_unload_buddy(&e4b
);
3746 * releases all non-used preallocated blocks for given inode
3748 * It's important to discard preallocations under i_data_sem
3749 * We don't want another block to be served from the prealloc
3750 * space when we are discarding the inode prealloc space.
3752 * FIXME!! Make sure it is valid at all the call sites
3754 void ext4_discard_preallocations(struct inode
*inode
)
3756 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3757 struct super_block
*sb
= inode
->i_sb
;
3758 struct buffer_head
*bitmap_bh
= NULL
;
3759 struct ext4_prealloc_space
*pa
, *tmp
;
3760 ext4_group_t group
= 0;
3761 struct list_head list
;
3762 struct ext4_buddy e4b
;
3765 if (!S_ISREG(inode
->i_mode
)) {
3766 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3770 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
3771 trace_ext4_discard_preallocations(inode
);
3773 INIT_LIST_HEAD(&list
);
3776 /* first, collect all pa's in the inode */
3777 spin_lock(&ei
->i_prealloc_lock
);
3778 while (!list_empty(&ei
->i_prealloc_list
)) {
3779 pa
= list_entry(ei
->i_prealloc_list
.next
,
3780 struct ext4_prealloc_space
, pa_inode_list
);
3781 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3782 spin_lock(&pa
->pa_lock
);
3783 if (atomic_read(&pa
->pa_count
)) {
3784 /* this shouldn't happen often - nobody should
3785 * use preallocation while we're discarding it */
3786 spin_unlock(&pa
->pa_lock
);
3787 spin_unlock(&ei
->i_prealloc_lock
);
3788 ext4_msg(sb
, KERN_ERR
,
3789 "uh-oh! used pa while discarding");
3791 schedule_timeout_uninterruptible(HZ
);
3795 if (pa
->pa_deleted
== 0) {
3797 spin_unlock(&pa
->pa_lock
);
3798 list_del_rcu(&pa
->pa_inode_list
);
3799 list_add(&pa
->u
.pa_tmp_list
, &list
);
3803 /* someone is deleting pa right now */
3804 spin_unlock(&pa
->pa_lock
);
3805 spin_unlock(&ei
->i_prealloc_lock
);
3807 /* we have to wait here because pa_deleted
3808 * doesn't mean pa is already unlinked from
3809 * the list. as we might be called from
3810 * ->clear_inode() the inode will get freed
3811 * and concurrent thread which is unlinking
3812 * pa from inode's list may access already
3813 * freed memory, bad-bad-bad */
3815 /* XXX: if this happens too often, we can
3816 * add a flag to force wait only in case
3817 * of ->clear_inode(), but not in case of
3818 * regular truncate */
3819 schedule_timeout_uninterruptible(HZ
);
3822 spin_unlock(&ei
->i_prealloc_lock
);
3824 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3825 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
3826 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
3828 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3830 ext4_error(sb
, "Error loading buddy information for %u",
3835 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3836 if (bitmap_bh
== NULL
) {
3837 ext4_error(sb
, "Error reading block bitmap for %u",
3839 ext4_mb_unload_buddy(&e4b
);
3843 ext4_lock_group(sb
, group
);
3844 list_del(&pa
->pa_group_list
);
3845 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3846 ext4_unlock_group(sb
, group
);
3848 ext4_mb_unload_buddy(&e4b
);
3851 list_del(&pa
->u
.pa_tmp_list
);
3852 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3856 #ifdef CONFIG_EXT4_DEBUG
3857 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3859 struct super_block
*sb
= ac
->ac_sb
;
3860 ext4_group_t ngroups
, i
;
3862 if (!mb_enable_debug
||
3863 (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
))
3866 ext4_msg(ac
->ac_sb
, KERN_ERR
, "Can't allocate:"
3867 " Allocation context details:");
3868 ext4_msg(ac
->ac_sb
, KERN_ERR
, "status %d flags %d",
3869 ac
->ac_status
, ac
->ac_flags
);
3870 ext4_msg(ac
->ac_sb
, KERN_ERR
, "orig %lu/%lu/%lu@%lu, "
3871 "goal %lu/%lu/%lu@%lu, "
3872 "best %lu/%lu/%lu@%lu cr %d",
3873 (unsigned long)ac
->ac_o_ex
.fe_group
,
3874 (unsigned long)ac
->ac_o_ex
.fe_start
,
3875 (unsigned long)ac
->ac_o_ex
.fe_len
,
3876 (unsigned long)ac
->ac_o_ex
.fe_logical
,
3877 (unsigned long)ac
->ac_g_ex
.fe_group
,
3878 (unsigned long)ac
->ac_g_ex
.fe_start
,
3879 (unsigned long)ac
->ac_g_ex
.fe_len
,
3880 (unsigned long)ac
->ac_g_ex
.fe_logical
,
3881 (unsigned long)ac
->ac_b_ex
.fe_group
,
3882 (unsigned long)ac
->ac_b_ex
.fe_start
,
3883 (unsigned long)ac
->ac_b_ex
.fe_len
,
3884 (unsigned long)ac
->ac_b_ex
.fe_logical
,
3885 (int)ac
->ac_criteria
);
3886 ext4_msg(ac
->ac_sb
, KERN_ERR
, "%lu scanned, %d found",
3887 ac
->ac_ex_scanned
, ac
->ac_found
);
3888 ext4_msg(ac
->ac_sb
, KERN_ERR
, "groups: ");
3889 ngroups
= ext4_get_groups_count(sb
);
3890 for (i
= 0; i
< ngroups
; i
++) {
3891 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
3892 struct ext4_prealloc_space
*pa
;
3893 ext4_grpblk_t start
;
3894 struct list_head
*cur
;
3895 ext4_lock_group(sb
, i
);
3896 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3897 pa
= list_entry(cur
, struct ext4_prealloc_space
,
3899 spin_lock(&pa
->pa_lock
);
3900 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3902 spin_unlock(&pa
->pa_lock
);
3903 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
3906 ext4_unlock_group(sb
, i
);
3908 if (grp
->bb_free
== 0)
3910 printk(KERN_ERR
"%u: %d/%d \n",
3911 i
, grp
->bb_free
, grp
->bb_fragments
);
3913 printk(KERN_ERR
"\n");
3916 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3923 * We use locality group preallocation for small size file. The size of the
3924 * file is determined by the current size or the resulting size after
3925 * allocation which ever is larger
3927 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3929 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
3931 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3932 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3935 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3938 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3941 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
3942 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
3945 if ((size
== isize
) &&
3946 !ext4_fs_is_busy(sbi
) &&
3947 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
3948 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
3952 if (sbi
->s_mb_group_prealloc
<= 0) {
3953 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
3957 /* don't use group allocation for large files */
3958 size
= max(size
, isize
);
3959 if (size
> sbi
->s_mb_stream_request
) {
3960 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
3964 BUG_ON(ac
->ac_lg
!= NULL
);
3966 * locality group prealloc space are per cpu. The reason for having
3967 * per cpu locality group is to reduce the contention between block
3968 * request from multiple CPUs.
3970 ac
->ac_lg
= __this_cpu_ptr(sbi
->s_locality_groups
);
3972 /* we're going to use group allocation */
3973 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
3975 /* serialize all allocations in the group */
3976 mutex_lock(&ac
->ac_lg
->lg_mutex
);
3979 static noinline_for_stack
int
3980 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
3981 struct ext4_allocation_request
*ar
)
3983 struct super_block
*sb
= ar
->inode
->i_sb
;
3984 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3985 struct ext4_super_block
*es
= sbi
->s_es
;
3989 ext4_grpblk_t block
;
3991 /* we can't allocate > group size */
3994 /* just a dirty hack to filter too big requests */
3995 if (len
>= EXT4_CLUSTERS_PER_GROUP(sb
) - 10)
3996 len
= EXT4_CLUSTERS_PER_GROUP(sb
) - 10;
3998 /* start searching from the goal */
4000 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4001 goal
>= ext4_blocks_count(es
))
4002 goal
= le32_to_cpu(es
->s_first_data_block
);
4003 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4005 /* set up allocation goals */
4006 memset(ac
, 0, sizeof(struct ext4_allocation_context
));
4007 ac
->ac_b_ex
.fe_logical
= ar
->logical
& ~(sbi
->s_cluster_ratio
- 1);
4008 ac
->ac_status
= AC_STATUS_CONTINUE
;
4010 ac
->ac_inode
= ar
->inode
;
4011 ac
->ac_o_ex
.fe_logical
= ac
->ac_b_ex
.fe_logical
;
4012 ac
->ac_o_ex
.fe_group
= group
;
4013 ac
->ac_o_ex
.fe_start
= block
;
4014 ac
->ac_o_ex
.fe_len
= len
;
4015 ac
->ac_g_ex
= ac
->ac_o_ex
;
4016 ac
->ac_flags
= ar
->flags
;
4018 /* we have to define context: we'll we work with a file or
4019 * locality group. this is a policy, actually */
4020 ext4_mb_group_or_file(ac
);
4022 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4023 "left: %u/%u, right %u/%u to %swritable\n",
4024 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4025 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4026 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4027 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4028 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4033 static noinline_for_stack
void
4034 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4035 struct ext4_locality_group
*lg
,
4036 int order
, int total_entries
)
4038 ext4_group_t group
= 0;
4039 struct ext4_buddy e4b
;
4040 struct list_head discard_list
;
4041 struct ext4_prealloc_space
*pa
, *tmp
;
4043 mb_debug(1, "discard locality group preallocation\n");
4045 INIT_LIST_HEAD(&discard_list
);
4047 spin_lock(&lg
->lg_prealloc_lock
);
4048 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4050 spin_lock(&pa
->pa_lock
);
4051 if (atomic_read(&pa
->pa_count
)) {
4053 * This is the pa that we just used
4054 * for block allocation. So don't
4057 spin_unlock(&pa
->pa_lock
);
4060 if (pa
->pa_deleted
) {
4061 spin_unlock(&pa
->pa_lock
);
4064 /* only lg prealloc space */
4065 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4067 /* seems this one can be freed ... */
4069 spin_unlock(&pa
->pa_lock
);
4071 list_del_rcu(&pa
->pa_inode_list
);
4072 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4075 if (total_entries
<= 5) {
4077 * we want to keep only 5 entries
4078 * allowing it to grow to 8. This
4079 * mak sure we don't call discard
4080 * soon for this list.
4085 spin_unlock(&lg
->lg_prealloc_lock
);
4087 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4089 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4090 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4091 ext4_error(sb
, "Error loading buddy information for %u",
4095 ext4_lock_group(sb
, group
);
4096 list_del(&pa
->pa_group_list
);
4097 ext4_mb_release_group_pa(&e4b
, pa
);
4098 ext4_unlock_group(sb
, group
);
4100 ext4_mb_unload_buddy(&e4b
);
4101 list_del(&pa
->u
.pa_tmp_list
);
4102 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4107 * We have incremented pa_count. So it cannot be freed at this
4108 * point. Also we hold lg_mutex. So no parallel allocation is
4109 * possible from this lg. That means pa_free cannot be updated.
4111 * A parallel ext4_mb_discard_group_preallocations is possible.
4112 * which can cause the lg_prealloc_list to be updated.
4115 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4117 int order
, added
= 0, lg_prealloc_count
= 1;
4118 struct super_block
*sb
= ac
->ac_sb
;
4119 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4120 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4122 order
= fls(pa
->pa_free
) - 1;
4123 if (order
> PREALLOC_TB_SIZE
- 1)
4124 /* The max size of hash table is PREALLOC_TB_SIZE */
4125 order
= PREALLOC_TB_SIZE
- 1;
4126 /* Add the prealloc space to lg */
4128 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4130 spin_lock(&tmp_pa
->pa_lock
);
4131 if (tmp_pa
->pa_deleted
) {
4132 spin_unlock(&tmp_pa
->pa_lock
);
4135 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4136 /* Add to the tail of the previous entry */
4137 list_add_tail_rcu(&pa
->pa_inode_list
,
4138 &tmp_pa
->pa_inode_list
);
4141 * we want to count the total
4142 * number of entries in the list
4145 spin_unlock(&tmp_pa
->pa_lock
);
4146 lg_prealloc_count
++;
4149 list_add_tail_rcu(&pa
->pa_inode_list
,
4150 &lg
->lg_prealloc_list
[order
]);
4153 /* Now trim the list to be not more than 8 elements */
4154 if (lg_prealloc_count
> 8) {
4155 ext4_mb_discard_lg_preallocations(sb
, lg
,
4156 order
, lg_prealloc_count
);
4163 * release all resource we used in allocation
4165 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4167 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4168 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4170 if (pa
->pa_type
== MB_GROUP_PA
) {
4171 /* see comment in ext4_mb_use_group_pa() */
4172 spin_lock(&pa
->pa_lock
);
4173 pa
->pa_pstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4174 pa
->pa_lstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4175 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4176 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4177 spin_unlock(&pa
->pa_lock
);
4182 * We want to add the pa to the right bucket.
4183 * Remove it from the list and while adding
4184 * make sure the list to which we are adding
4187 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4188 spin_lock(pa
->pa_obj_lock
);
4189 list_del_rcu(&pa
->pa_inode_list
);
4190 spin_unlock(pa
->pa_obj_lock
);
4191 ext4_mb_add_n_trim(ac
);
4193 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4195 if (ac
->ac_bitmap_page
)
4196 page_cache_release(ac
->ac_bitmap_page
);
4197 if (ac
->ac_buddy_page
)
4198 page_cache_release(ac
->ac_buddy_page
);
4199 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4200 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4201 ext4_mb_collect_stats(ac
);
4205 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4207 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4211 trace_ext4_mb_discard_preallocations(sb
, needed
);
4212 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4213 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4222 * Main entry point into mballoc to allocate blocks
4223 * it tries to use preallocation first, then falls back
4224 * to usual allocation
4226 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4227 struct ext4_allocation_request
*ar
, int *errp
)
4230 struct ext4_allocation_context
*ac
= NULL
;
4231 struct ext4_sb_info
*sbi
;
4232 struct super_block
*sb
;
4233 ext4_fsblk_t block
= 0;
4234 unsigned int inquota
= 0;
4235 unsigned int reserv_clstrs
= 0;
4237 sb
= ar
->inode
->i_sb
;
4240 trace_ext4_request_blocks(ar
);
4242 /* Allow to use superuser reservation for quota file */
4243 if (IS_NOQUOTA(ar
->inode
))
4244 ar
->flags
|= EXT4_MB_USE_ROOT_BLOCKS
;
4247 * For delayed allocation, we could skip the ENOSPC and
4248 * EDQUOT check, as blocks and quotas have been already
4249 * reserved when data being copied into pagecache.
4251 if (ext4_test_inode_state(ar
->inode
, EXT4_STATE_DELALLOC_RESERVED
))
4252 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4254 /* Without delayed allocation we need to verify
4255 * there is enough free blocks to do block allocation
4256 * and verify allocation doesn't exceed the quota limits.
4259 ext4_claim_free_clusters(sbi
, ar
->len
, ar
->flags
)) {
4261 /* let others to free the space */
4263 ar
->len
= ar
->len
>> 1;
4269 reserv_clstrs
= ar
->len
;
4270 if (ar
->flags
& EXT4_MB_USE_ROOT_BLOCKS
) {
4271 dquot_alloc_block_nofail(ar
->inode
,
4272 EXT4_C2B(sbi
, ar
->len
));
4275 dquot_alloc_block(ar
->inode
,
4276 EXT4_C2B(sbi
, ar
->len
))) {
4278 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4289 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4296 *errp
= ext4_mb_initialize_context(ac
, ar
);
4302 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4303 if (!ext4_mb_use_preallocated(ac
)) {
4304 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4305 ext4_mb_normalize_request(ac
, ar
);
4307 /* allocate space in core */
4308 *errp
= ext4_mb_regular_allocator(ac
);
4312 /* as we've just preallocated more space than
4313 * user requested orinally, we store allocated
4314 * space in a special descriptor */
4315 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4316 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4317 ext4_mb_new_preallocation(ac
);
4319 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4320 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_clstrs
);
4321 if (*errp
== -EAGAIN
) {
4323 * drop the reference that we took
4324 * in ext4_mb_use_best_found
4326 ext4_mb_release_context(ac
);
4327 ac
->ac_b_ex
.fe_group
= 0;
4328 ac
->ac_b_ex
.fe_start
= 0;
4329 ac
->ac_b_ex
.fe_len
= 0;
4330 ac
->ac_status
= AC_STATUS_CONTINUE
;
4334 ext4_discard_allocated_blocks(ac
);
4336 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4337 ar
->len
= ac
->ac_b_ex
.fe_len
;
4340 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4347 ac
->ac_b_ex
.fe_len
= 0;
4349 ext4_mb_show_ac(ac
);
4351 ext4_mb_release_context(ac
);
4354 kmem_cache_free(ext4_ac_cachep
, ac
);
4355 if (inquota
&& ar
->len
< inquota
)
4356 dquot_free_block(ar
->inode
, EXT4_C2B(sbi
, inquota
- ar
->len
));
4358 if (!ext4_test_inode_state(ar
->inode
,
4359 EXT4_STATE_DELALLOC_RESERVED
))
4360 /* release all the reserved blocks if non delalloc */
4361 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
4365 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4371 * We can merge two free data extents only if the physical blocks
4372 * are contiguous, AND the extents were freed by the same transaction,
4373 * AND the blocks are associated with the same group.
4375 static int can_merge(struct ext4_free_data
*entry1
,
4376 struct ext4_free_data
*entry2
)
4378 if ((entry1
->efd_tid
== entry2
->efd_tid
) &&
4379 (entry1
->efd_group
== entry2
->efd_group
) &&
4380 ((entry1
->efd_start_cluster
+ entry1
->efd_count
) == entry2
->efd_start_cluster
))
4385 static noinline_for_stack
int
4386 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4387 struct ext4_free_data
*new_entry
)
4389 ext4_group_t group
= e4b
->bd_group
;
4390 ext4_grpblk_t cluster
;
4391 struct ext4_free_data
*entry
;
4392 struct ext4_group_info
*db
= e4b
->bd_info
;
4393 struct super_block
*sb
= e4b
->bd_sb
;
4394 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4395 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4396 struct rb_node
*parent
= NULL
, *new_node
;
4398 BUG_ON(!ext4_handle_valid(handle
));
4399 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4400 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4402 new_node
= &new_entry
->efd_node
;
4403 cluster
= new_entry
->efd_start_cluster
;
4406 /* first free block exent. We need to
4407 protect buddy cache from being freed,
4408 * otherwise we'll refresh it from
4409 * on-disk bitmap and lose not-yet-available
4411 page_cache_get(e4b
->bd_buddy_page
);
4412 page_cache_get(e4b
->bd_bitmap_page
);
4416 entry
= rb_entry(parent
, struct ext4_free_data
, efd_node
);
4417 if (cluster
< entry
->efd_start_cluster
)
4419 else if (cluster
>= (entry
->efd_start_cluster
+ entry
->efd_count
))
4420 n
= &(*n
)->rb_right
;
4422 ext4_grp_locked_error(sb
, group
, 0,
4423 ext4_group_first_block_no(sb
, group
) +
4424 EXT4_C2B(sbi
, cluster
),
4425 "Block already on to-be-freed list");
4430 rb_link_node(new_node
, parent
, n
);
4431 rb_insert_color(new_node
, &db
->bb_free_root
);
4433 /* Now try to see the extent can be merged to left and right */
4434 node
= rb_prev(new_node
);
4436 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4437 if (can_merge(entry
, new_entry
)) {
4438 new_entry
->efd_start_cluster
= entry
->efd_start_cluster
;
4439 new_entry
->efd_count
+= entry
->efd_count
;
4440 rb_erase(node
, &(db
->bb_free_root
));
4441 ext4_journal_callback_del(handle
, &entry
->efd_jce
);
4442 kmem_cache_free(ext4_free_data_cachep
, entry
);
4446 node
= rb_next(new_node
);
4448 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4449 if (can_merge(new_entry
, entry
)) {
4450 new_entry
->efd_count
+= entry
->efd_count
;
4451 rb_erase(node
, &(db
->bb_free_root
));
4452 ext4_journal_callback_del(handle
, &entry
->efd_jce
);
4453 kmem_cache_free(ext4_free_data_cachep
, entry
);
4456 /* Add the extent to transaction's private list */
4457 ext4_journal_callback_add(handle
, ext4_free_data_callback
,
4458 &new_entry
->efd_jce
);
4463 * ext4_free_blocks() -- Free given blocks and update quota
4464 * @handle: handle for this transaction
4466 * @block: start physical block to free
4467 * @count: number of blocks to count
4468 * @flags: flags used by ext4_free_blocks
4470 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4471 struct buffer_head
*bh
, ext4_fsblk_t block
,
4472 unsigned long count
, int flags
)
4474 struct buffer_head
*bitmap_bh
= NULL
;
4475 struct super_block
*sb
= inode
->i_sb
;
4476 struct ext4_group_desc
*gdp
;
4477 unsigned long freed
= 0;
4478 unsigned int overflow
;
4480 struct buffer_head
*gd_bh
;
4481 ext4_group_t block_group
;
4482 struct ext4_sb_info
*sbi
;
4483 struct ext4_buddy e4b
;
4484 unsigned int count_clusters
;
4490 BUG_ON(block
!= bh
->b_blocknr
);
4492 block
= bh
->b_blocknr
;
4496 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4497 !ext4_data_block_valid(sbi
, block
, count
)) {
4498 ext4_error(sb
, "Freeing blocks not in datazone - "
4499 "block = %llu, count = %lu", block
, count
);
4503 ext4_debug("freeing block %llu\n", block
);
4504 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4506 if (flags
& EXT4_FREE_BLOCKS_FORGET
) {
4507 struct buffer_head
*tbh
= bh
;
4510 BUG_ON(bh
&& (count
> 1));
4512 for (i
= 0; i
< count
; i
++) {
4514 tbh
= sb_find_get_block(inode
->i_sb
,
4518 ext4_forget(handle
, flags
& EXT4_FREE_BLOCKS_METADATA
,
4519 inode
, tbh
, block
+ i
);
4524 * We need to make sure we don't reuse the freed block until
4525 * after the transaction is committed, which we can do by
4526 * treating the block as metadata, below. We make an
4527 * exception if the inode is to be written in writeback mode
4528 * since writeback mode has weak data consistency guarantees.
4530 if (!ext4_should_writeback_data(inode
))
4531 flags
|= EXT4_FREE_BLOCKS_METADATA
;
4534 * If the extent to be freed does not begin on a cluster
4535 * boundary, we need to deal with partial clusters at the
4536 * beginning and end of the extent. Normally we will free
4537 * blocks at the beginning or the end unless we are explicitly
4538 * requested to avoid doing so.
4540 overflow
= block
& (sbi
->s_cluster_ratio
- 1);
4542 if (flags
& EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER
) {
4543 overflow
= sbi
->s_cluster_ratio
- overflow
;
4545 if (count
> overflow
)
4554 overflow
= count
& (sbi
->s_cluster_ratio
- 1);
4556 if (flags
& EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER
) {
4557 if (count
> overflow
)
4562 count
+= sbi
->s_cluster_ratio
- overflow
;
4567 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4570 * Check to see if we are freeing blocks across a group
4573 if (EXT4_C2B(sbi
, bit
) + count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4574 overflow
= EXT4_C2B(sbi
, bit
) + count
-
4575 EXT4_BLOCKS_PER_GROUP(sb
);
4578 count_clusters
= EXT4_B2C(sbi
, count
);
4579 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4584 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4590 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4591 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4592 in_range(block
, ext4_inode_table(sb
, gdp
),
4593 EXT4_SB(sb
)->s_itb_per_group
) ||
4594 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4595 EXT4_SB(sb
)->s_itb_per_group
)) {
4597 ext4_error(sb
, "Freeing blocks in system zone - "
4598 "Block = %llu, count = %lu", block
, count
);
4599 /* err = 0. ext4_std_error should be a no op */
4603 BUFFER_TRACE(bitmap_bh
, "getting write access");
4604 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4609 * We are about to modify some metadata. Call the journal APIs
4610 * to unshare ->b_data if a currently-committing transaction is
4613 BUFFER_TRACE(gd_bh
, "get_write_access");
4614 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4617 #ifdef AGGRESSIVE_CHECK
4620 for (i
= 0; i
< count_clusters
; i
++)
4621 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4624 trace_ext4_mballoc_free(sb
, inode
, block_group
, bit
, count_clusters
);
4626 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4630 if ((flags
& EXT4_FREE_BLOCKS_METADATA
) && ext4_handle_valid(handle
)) {
4631 struct ext4_free_data
*new_entry
;
4633 * blocks being freed are metadata. these blocks shouldn't
4634 * be used until this transaction is committed
4636 new_entry
= kmem_cache_alloc(ext4_free_data_cachep
, GFP_NOFS
);
4638 ext4_mb_unload_buddy(&e4b
);
4642 new_entry
->efd_start_cluster
= bit
;
4643 new_entry
->efd_group
= block_group
;
4644 new_entry
->efd_count
= count_clusters
;
4645 new_entry
->efd_tid
= handle
->h_transaction
->t_tid
;
4647 ext4_lock_group(sb
, block_group
);
4648 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4649 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4651 /* need to update group_info->bb_free and bitmap
4652 * with group lock held. generate_buddy look at
4653 * them with group lock_held
4655 ext4_lock_group(sb
, block_group
);
4656 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4657 mb_free_blocks(inode
, &e4b
, bit
, count_clusters
);
4660 ret
= ext4_free_group_clusters(sb
, gdp
) + count_clusters
;
4661 ext4_free_group_clusters_set(sb
, gdp
, ret
);
4662 ext4_block_bitmap_csum_set(sb
, block_group
, gdp
, bitmap_bh
,
4663 EXT4_BLOCKS_PER_GROUP(sb
) / 8);
4664 ext4_group_desc_csum_set(sb
, block_group
, gdp
);
4665 ext4_unlock_group(sb
, block_group
);
4666 percpu_counter_add(&sbi
->s_freeclusters_counter
, count_clusters
);
4668 if (sbi
->s_log_groups_per_flex
) {
4669 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4670 atomic_add(count_clusters
,
4671 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
4674 ext4_mb_unload_buddy(&e4b
);
4678 if (!(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4679 dquot_free_block(inode
, EXT4_C2B(sbi
, count_clusters
));
4681 /* We dirtied the bitmap block */
4682 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4683 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4685 /* And the group descriptor block */
4686 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4687 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4691 if (overflow
&& !err
) {
4697 ext4_mark_super_dirty(sb
);
4700 ext4_std_error(sb
, err
);
4705 * ext4_group_add_blocks() -- Add given blocks to an existing group
4706 * @handle: handle to this transaction
4708 * @block: start physcial block to add to the block group
4709 * @count: number of blocks to free
4711 * This marks the blocks as free in the bitmap and buddy.
4713 int ext4_group_add_blocks(handle_t
*handle
, struct super_block
*sb
,
4714 ext4_fsblk_t block
, unsigned long count
)
4716 struct buffer_head
*bitmap_bh
= NULL
;
4717 struct buffer_head
*gd_bh
;
4718 ext4_group_t block_group
;
4721 struct ext4_group_desc
*desc
;
4722 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4723 struct ext4_buddy e4b
;
4724 int err
= 0, ret
, blk_free_count
;
4725 ext4_grpblk_t blocks_freed
;
4727 ext4_debug("Adding block(s) %llu-%llu\n", block
, block
+ count
- 1);
4732 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4734 * Check to see if we are freeing blocks across a group
4737 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4738 ext4_warning(sb
, "too much blocks added to group %u\n",
4744 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4750 desc
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4756 if (in_range(ext4_block_bitmap(sb
, desc
), block
, count
) ||
4757 in_range(ext4_inode_bitmap(sb
, desc
), block
, count
) ||
4758 in_range(block
, ext4_inode_table(sb
, desc
), sbi
->s_itb_per_group
) ||
4759 in_range(block
+ count
- 1, ext4_inode_table(sb
, desc
),
4760 sbi
->s_itb_per_group
)) {
4761 ext4_error(sb
, "Adding blocks in system zones - "
4762 "Block = %llu, count = %lu",
4768 BUFFER_TRACE(bitmap_bh
, "getting write access");
4769 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4774 * We are about to modify some metadata. Call the journal APIs
4775 * to unshare ->b_data if a currently-committing transaction is
4778 BUFFER_TRACE(gd_bh
, "get_write_access");
4779 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4783 for (i
= 0, blocks_freed
= 0; i
< count
; i
++) {
4784 BUFFER_TRACE(bitmap_bh
, "clear bit");
4785 if (!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
)) {
4786 ext4_error(sb
, "bit already cleared for block %llu",
4787 (ext4_fsblk_t
)(block
+ i
));
4788 BUFFER_TRACE(bitmap_bh
, "bit already cleared");
4794 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4799 * need to update group_info->bb_free and bitmap
4800 * with group lock held. generate_buddy look at
4801 * them with group lock_held
4803 ext4_lock_group(sb
, block_group
);
4804 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4805 mb_free_blocks(NULL
, &e4b
, bit
, count
);
4806 blk_free_count
= blocks_freed
+ ext4_free_group_clusters(sb
, desc
);
4807 ext4_free_group_clusters_set(sb
, desc
, blk_free_count
);
4808 ext4_block_bitmap_csum_set(sb
, block_group
, desc
, bitmap_bh
,
4809 EXT4_BLOCKS_PER_GROUP(sb
) / 8);
4810 ext4_group_desc_csum_set(sb
, block_group
, desc
);
4811 ext4_unlock_group(sb
, block_group
);
4812 percpu_counter_add(&sbi
->s_freeclusters_counter
,
4813 EXT4_B2C(sbi
, blocks_freed
));
4815 if (sbi
->s_log_groups_per_flex
) {
4816 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4817 atomic_add(EXT4_B2C(sbi
, blocks_freed
),
4818 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
4821 ext4_mb_unload_buddy(&e4b
);
4823 /* We dirtied the bitmap block */
4824 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4825 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4827 /* And the group descriptor block */
4828 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4829 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4835 ext4_std_error(sb
, err
);
4840 * ext4_trim_extent -- function to TRIM one single free extent in the group
4841 * @sb: super block for the file system
4842 * @start: starting block of the free extent in the alloc. group
4843 * @count: number of blocks to TRIM
4844 * @group: alloc. group we are working with
4845 * @e4b: ext4 buddy for the group
4847 * Trim "count" blocks starting at "start" in the "group". To assure that no
4848 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4849 * be called with under the group lock.
4851 static void ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
4852 ext4_group_t group
, struct ext4_buddy
*e4b
)
4854 struct ext4_free_extent ex
;
4856 trace_ext4_trim_extent(sb
, group
, start
, count
);
4858 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
4860 ex
.fe_start
= start
;
4861 ex
.fe_group
= group
;
4865 * Mark blocks used, so no one can reuse them while
4868 mb_mark_used(e4b
, &ex
);
4869 ext4_unlock_group(sb
, group
);
4870 ext4_issue_discard(sb
, group
, start
, count
);
4871 ext4_lock_group(sb
, group
);
4872 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
4876 * ext4_trim_all_free -- function to trim all free space in alloc. group
4877 * @sb: super block for file system
4878 * @group: group to be trimmed
4879 * @start: first group block to examine
4880 * @max: last group block to examine
4881 * @minblocks: minimum extent block count
4883 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4884 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4888 * ext4_trim_all_free walks through group's block bitmap searching for free
4889 * extents. When the free extent is found, mark it as used in group buddy
4890 * bitmap. Then issue a TRIM command on this extent and free the extent in
4891 * the group buddy bitmap. This is done until whole group is scanned.
4893 static ext4_grpblk_t
4894 ext4_trim_all_free(struct super_block
*sb
, ext4_group_t group
,
4895 ext4_grpblk_t start
, ext4_grpblk_t max
,
4896 ext4_grpblk_t minblocks
)
4899 ext4_grpblk_t next
, count
= 0, free_count
= 0;
4900 struct ext4_buddy e4b
;
4903 trace_ext4_trim_all_free(sb
, group
, start
, max
);
4905 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
4907 ext4_error(sb
, "Error in loading buddy "
4908 "information for %u", group
);
4911 bitmap
= e4b
.bd_bitmap
;
4913 ext4_lock_group(sb
, group
);
4914 if (EXT4_MB_GRP_WAS_TRIMMED(e4b
.bd_info
) &&
4915 minblocks
>= atomic_read(&EXT4_SB(sb
)->s_last_trim_minblks
))
4918 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
4919 e4b
.bd_info
->bb_first_free
: start
;
4921 while (start
<= max
) {
4922 start
= mb_find_next_zero_bit(bitmap
, max
+ 1, start
);
4925 next
= mb_find_next_bit(bitmap
, max
+ 1, start
);
4927 if ((next
- start
) >= minblocks
) {
4928 ext4_trim_extent(sb
, start
,
4929 next
- start
, group
, &e4b
);
4930 count
+= next
- start
;
4932 free_count
+= next
- start
;
4935 if (fatal_signal_pending(current
)) {
4936 count
= -ERESTARTSYS
;
4940 if (need_resched()) {
4941 ext4_unlock_group(sb
, group
);
4943 ext4_lock_group(sb
, group
);
4946 if ((e4b
.bd_info
->bb_free
- free_count
) < minblocks
)
4951 EXT4_MB_GRP_SET_TRIMMED(e4b
.bd_info
);
4953 ext4_unlock_group(sb
, group
);
4954 ext4_mb_unload_buddy(&e4b
);
4956 ext4_debug("trimmed %d blocks in the group %d\n",
4963 * ext4_trim_fs() -- trim ioctl handle function
4964 * @sb: superblock for filesystem
4965 * @range: fstrim_range structure
4967 * start: First Byte to trim
4968 * len: number of Bytes to trim from start
4969 * minlen: minimum extent length in Bytes
4970 * ext4_trim_fs goes through all allocation groups containing Bytes from
4971 * start to start+len. For each such a group ext4_trim_all_free function
4972 * is invoked to trim all free space.
4974 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
4976 struct ext4_group_info
*grp
;
4977 ext4_group_t group
, first_group
, last_group
;
4978 ext4_grpblk_t cnt
= 0, first_cluster
, last_cluster
;
4979 uint64_t start
, end
, minlen
, trimmed
= 0;
4980 ext4_fsblk_t first_data_blk
=
4981 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
4982 ext4_fsblk_t max_blks
= ext4_blocks_count(EXT4_SB(sb
)->s_es
);
4985 start
= range
->start
>> sb
->s_blocksize_bits
;
4986 end
= start
+ (range
->len
>> sb
->s_blocksize_bits
) - 1;
4987 minlen
= range
->minlen
>> sb
->s_blocksize_bits
;
4989 if (unlikely(minlen
> EXT4_CLUSTERS_PER_GROUP(sb
)) ||
4990 unlikely(start
>= max_blks
))
4992 if (end
>= max_blks
)
4994 if (end
<= first_data_blk
)
4996 if (start
< first_data_blk
)
4997 start
= first_data_blk
;
4999 /* Determine first and last group to examine based on start and end */
5000 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
5001 &first_group
, &first_cluster
);
5002 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) end
,
5003 &last_group
, &last_cluster
);
5005 /* end now represents the last cluster to discard in this group */
5006 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5008 for (group
= first_group
; group
<= last_group
; group
++) {
5009 grp
= ext4_get_group_info(sb
, group
);
5010 /* We only do this if the grp has never been initialized */
5011 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
5012 ret
= ext4_mb_init_group(sb
, group
);
5018 * For all the groups except the last one, last cluster will
5019 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5020 * change it for the last group, note that last_cluster is
5021 * already computed earlier by ext4_get_group_no_and_offset()
5023 if (group
== last_group
)
5026 if (grp
->bb_free
>= minlen
) {
5027 cnt
= ext4_trim_all_free(sb
, group
, first_cluster
,
5037 * For every group except the first one, we are sure
5038 * that the first cluster to discard will be cluster #0.
5044 atomic_set(&EXT4_SB(sb
)->s_last_trim_minblks
, minlen
);
5047 range
->len
= trimmed
* sb
->s_blocksize
;