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 License
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/log2.h>
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <linux/backing-dev.h>
30 #include <trace/events/ext4.h>
32 #ifdef CONFIG_EXT4_DEBUG
33 ushort ext4_mballoc_debug __read_mostly
;
35 module_param_named(mballoc_debug
, ext4_mballoc_debug
, ushort
, 0644);
36 MODULE_PARM_DESC(mballoc_debug
, "Debugging level for ext4's mballoc");
41 * - test ext4_ext_search_left() and ext4_ext_search_right()
42 * - search for metadata in few groups
45 * - normalization should take into account whether file is still open
46 * - discard preallocations if no free space left (policy?)
47 * - don't normalize tails
49 * - reservation for superuser
52 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
53 * - track min/max extents in each group for better group selection
54 * - mb_mark_used() may allocate chunk right after splitting buddy
55 * - tree of groups sorted by number of free blocks
60 * The allocation request involve request for multiple number of blocks
61 * near to the goal(block) value specified.
63 * During initialization phase of the allocator we decide to use the
64 * group preallocation or inode preallocation depending on the size of
65 * the file. The size of the file could be the resulting file size we
66 * would have after allocation, or the current file size, which ever
67 * is larger. If the size is less than sbi->s_mb_stream_request we
68 * select to use the group preallocation. The default value of
69 * s_mb_stream_request is 16 blocks. This can also be tuned via
70 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
71 * terms of number of blocks.
73 * The main motivation for having small file use group preallocation is to
74 * ensure that we have small files closer together on the disk.
76 * First stage the allocator looks at the inode prealloc list,
77 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
78 * spaces for this particular inode. The inode prealloc space is
81 * pa_lstart -> the logical start block for this prealloc space
82 * pa_pstart -> the physical start block for this prealloc space
83 * pa_len -> length for this prealloc space (in clusters)
84 * pa_free -> free space available in this prealloc space (in clusters)
86 * The inode preallocation space is used looking at the _logical_ start
87 * block. If only the logical file block falls within the range of prealloc
88 * space we will consume the particular prealloc space. This makes sure that
89 * we have contiguous physical blocks representing the file blocks
91 * The important thing to be noted in case of inode prealloc space is that
92 * we don't modify the values associated to inode prealloc space except
95 * If we are not able to find blocks in the inode prealloc space and if we
96 * have the group allocation flag set then we look at the locality group
97 * prealloc space. These are per CPU prealloc list represented as
99 * ext4_sb_info.s_locality_groups[smp_processor_id()]
101 * The reason for having a per cpu locality group is to reduce the contention
102 * between CPUs. It is possible to get scheduled at this point.
104 * The locality group prealloc space is used looking at whether we have
105 * enough free space (pa_free) within the prealloc space.
107 * If we can't allocate blocks via inode prealloc or/and locality group
108 * prealloc then we look at the buddy cache. The buddy cache is represented
109 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
110 * mapped to the buddy and bitmap information regarding different
111 * groups. The buddy information is attached to buddy cache inode so that
112 * we can access them through the page cache. The information regarding
113 * each group is loaded via ext4_mb_load_buddy. The information involve
114 * block bitmap and buddy information. The information are stored in the
118 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
121 * one block each for bitmap and buddy information. So for each group we
122 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
123 * blocksize) blocks. So it can have information regarding groups_per_page
124 * which is blocks_per_page/2
126 * The buddy cache inode is not stored on disk. The inode is thrown
127 * away when the filesystem is unmounted.
129 * We look for count number of blocks in the buddy cache. If we were able
130 * to locate that many free blocks we return with additional information
131 * regarding rest of the contiguous physical block available
133 * Before allocating blocks via buddy cache we normalize the request
134 * blocks. This ensure we ask for more blocks that we needed. The extra
135 * blocks that we get after allocation is added to the respective prealloc
136 * list. In case of inode preallocation we follow a list of heuristics
137 * based on file size. This can be found in ext4_mb_normalize_request. If
138 * we are doing a group prealloc we try to normalize the request to
139 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
140 * dependent on the cluster size; for non-bigalloc file systems, it is
141 * 512 blocks. This can be tuned via
142 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
143 * terms of number of blocks. If we have mounted the file system with -O
144 * stripe=<value> option the group prealloc request is normalized to the
145 * the smallest multiple of the stripe value (sbi->s_stripe) which is
146 * greater than the default mb_group_prealloc.
148 * The regular allocator (using the buddy cache) supports a few tunables.
150 * /sys/fs/ext4/<partition>/mb_min_to_scan
151 * /sys/fs/ext4/<partition>/mb_max_to_scan
152 * /sys/fs/ext4/<partition>/mb_order2_req
154 * The regular allocator uses buddy scan only if the request len is power of
155 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
156 * value of s_mb_order2_reqs can be tuned via
157 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
158 * stripe size (sbi->s_stripe), we try to search for contiguous block in
159 * stripe size. This should result in better allocation on RAID setups. If
160 * not, we search in the specific group using bitmap for best extents. The
161 * tunable min_to_scan and max_to_scan control the behaviour here.
162 * min_to_scan indicate how long the mballoc __must__ look for a best
163 * extent and max_to_scan indicates how long the mballoc __can__ look for a
164 * best extent in the found extents. Searching for the blocks starts with
165 * the group specified as the goal value in allocation context via
166 * ac_g_ex. Each group is first checked based on the criteria whether it
167 * can be used for allocation. ext4_mb_good_group explains how the groups are
170 * Both the prealloc space are getting populated as above. So for the first
171 * request we will hit the buddy cache which will result in this prealloc
172 * space getting filled. The prealloc space is then later used for the
173 * subsequent request.
177 * mballoc operates on the following data:
179 * - in-core buddy (actually includes buddy and bitmap)
180 * - preallocation descriptors (PAs)
182 * there are two types of preallocations:
184 * assiged to specific inode and can be used for this inode only.
185 * it describes part of inode's space preallocated to specific
186 * physical blocks. any block from that preallocated can be used
187 * independent. the descriptor just tracks number of blocks left
188 * unused. so, before taking some block from descriptor, one must
189 * make sure corresponded logical block isn't allocated yet. this
190 * also means that freeing any block within descriptor's range
191 * must discard all preallocated blocks.
193 * assigned to specific locality group which does not translate to
194 * permanent set of inodes: inode can join and leave group. space
195 * from this type of preallocation can be used for any inode. thus
196 * it's consumed from the beginning to the end.
198 * relation between them can be expressed as:
199 * in-core buddy = on-disk bitmap + preallocation descriptors
201 * this mean blocks mballoc considers used are:
202 * - allocated blocks (persistent)
203 * - preallocated blocks (non-persistent)
205 * consistency in mballoc world means that at any time a block is either
206 * free or used in ALL structures. notice: "any time" should not be read
207 * literally -- time is discrete and delimited by locks.
209 * to keep it simple, we don't use block numbers, instead we count number of
210 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
212 * all operations can be expressed as:
213 * - init buddy: buddy = on-disk + PAs
214 * - new PA: buddy += N; PA = N
215 * - use inode PA: on-disk += N; PA -= N
216 * - discard inode PA buddy -= on-disk - PA; PA = 0
217 * - use locality group PA on-disk += N; PA -= N
218 * - discard locality group PA buddy -= PA; PA = 0
219 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
220 * is used in real operation because we can't know actual used
221 * bits from PA, only from on-disk bitmap
223 * if we follow this strict logic, then all operations above should be atomic.
224 * given some of them can block, we'd have to use something like semaphores
225 * killing performance on high-end SMP hardware. let's try to relax it using
226 * the following knowledge:
227 * 1) if buddy is referenced, it's already initialized
228 * 2) while block is used in buddy and the buddy is referenced,
229 * nobody can re-allocate that block
230 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
231 * bit set and PA claims same block, it's OK. IOW, one can set bit in
232 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
235 * so, now we're building a concurrency table:
238 * blocks for PA are allocated in the buddy, buddy must be referenced
239 * until PA is linked to allocation group to avoid concurrent buddy init
241 * we need to make sure that either on-disk bitmap or PA has uptodate data
242 * given (3) we care that PA-=N operation doesn't interfere with init
244 * the simplest way would be to have buddy initialized by the discard
245 * - use locality group PA
246 * again PA-=N must be serialized with init
247 * - discard locality group PA
248 * the simplest way would be to have buddy initialized by the discard
251 * i_data_sem serializes them
253 * discard process must wait until PA isn't used by another process
254 * - use locality group PA
255 * some mutex should serialize them
256 * - discard locality group PA
257 * discard process must wait until PA isn't used by another process
260 * i_data_sem or another mutex should serializes them
262 * discard process must wait until PA isn't used by another process
263 * - use locality group PA
264 * nothing wrong here -- they're different PAs covering different blocks
265 * - discard locality group PA
266 * discard process must wait until PA isn't used by another process
268 * now we're ready to make few consequences:
269 * - PA is referenced and while it is no discard is possible
270 * - PA is referenced until block isn't marked in on-disk bitmap
271 * - PA changes only after on-disk bitmap
272 * - discard must not compete with init. either init is done before
273 * any discard or they're serialized somehow
274 * - buddy init as sum of on-disk bitmap and PAs is done atomically
276 * a special case when we've used PA to emptiness. no need to modify buddy
277 * in this case, but we should care about concurrent init
282 * Logic in few words:
287 * mark bits in on-disk bitmap
290 * - use preallocation:
291 * find proper PA (per-inode or group)
293 * mark bits in on-disk bitmap
299 * mark bits in on-disk bitmap
302 * - discard preallocations in group:
304 * move them onto local list
305 * load on-disk bitmap
307 * remove PA from object (inode or locality group)
308 * mark free blocks in-core
310 * - discard inode's preallocations:
317 * - bitlock on a group (group)
318 * - object (inode/locality) (object)
329 * - release consumed pa:
334 * - generate in-core bitmap:
338 * - discard all for given object (inode, locality group):
343 * - discard all for given group:
350 static struct kmem_cache
*ext4_pspace_cachep
;
351 static struct kmem_cache
*ext4_ac_cachep
;
352 static struct kmem_cache
*ext4_free_data_cachep
;
354 /* We create slab caches for groupinfo data structures based on the
355 * superblock block size. There will be one per mounted filesystem for
356 * each unique s_blocksize_bits */
357 #define NR_GRPINFO_CACHES 8
358 static struct kmem_cache
*ext4_groupinfo_caches
[NR_GRPINFO_CACHES
];
360 static const char * const ext4_groupinfo_slab_names
[NR_GRPINFO_CACHES
] = {
361 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
362 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
363 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
366 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
368 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
371 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
373 #if BITS_PER_LONG == 64
374 *bit
+= ((unsigned long) addr
& 7UL) << 3;
375 addr
= (void *) ((unsigned long) addr
& ~7UL);
376 #elif BITS_PER_LONG == 32
377 *bit
+= ((unsigned long) addr
& 3UL) << 3;
378 addr
= (void *) ((unsigned long) addr
& ~3UL);
380 #error "how many bits you are?!"
385 static inline int mb_test_bit(int bit
, void *addr
)
388 * ext4_test_bit on architecture like powerpc
389 * needs unsigned long aligned address
391 addr
= mb_correct_addr_and_bit(&bit
, addr
);
392 return ext4_test_bit(bit
, addr
);
395 static inline void mb_set_bit(int bit
, void *addr
)
397 addr
= mb_correct_addr_and_bit(&bit
, addr
);
398 ext4_set_bit(bit
, addr
);
401 static inline void mb_clear_bit(int bit
, void *addr
)
403 addr
= mb_correct_addr_and_bit(&bit
, addr
);
404 ext4_clear_bit(bit
, addr
);
407 static inline int mb_test_and_clear_bit(int bit
, void *addr
)
409 addr
= mb_correct_addr_and_bit(&bit
, addr
);
410 return ext4_test_and_clear_bit(bit
, addr
);
413 static inline int mb_find_next_zero_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_zero_bit(addr
, tmpmax
, start
) - fix
;
426 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
428 int fix
= 0, ret
, tmpmax
;
429 addr
= mb_correct_addr_and_bit(&fix
, addr
);
433 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
439 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
443 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
446 if (order
> e4b
->bd_blkbits
+ 1) {
451 /* at order 0 we see each particular block */
453 *max
= 1 << (e4b
->bd_blkbits
+ 3);
454 return e4b
->bd_bitmap
;
457 bb
= e4b
->bd_buddy
+ EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
458 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
464 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
465 int first
, int count
)
468 struct super_block
*sb
= e4b
->bd_sb
;
470 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
472 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
473 for (i
= 0; i
< count
; i
++) {
474 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
475 ext4_fsblk_t blocknr
;
477 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
478 blocknr
+= EXT4_C2B(EXT4_SB(sb
), first
+ i
);
479 ext4_grp_locked_error(sb
, e4b
->bd_group
,
480 inode
? inode
->i_ino
: 0,
482 "freeing block already freed "
486 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
490 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
494 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
496 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
497 for (i
= 0; i
< count
; i
++) {
498 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
499 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
503 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
505 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
506 unsigned char *b1
, *b2
;
508 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
509 b2
= (unsigned char *) bitmap
;
510 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
511 if (b1
[i
] != b2
[i
]) {
512 ext4_msg(e4b
->bd_sb
, KERN_ERR
,
513 "corruption in group %u "
514 "at byte %u(%u): %x in copy != %x "
516 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
524 static inline void mb_free_blocks_double(struct inode
*inode
,
525 struct ext4_buddy
*e4b
, int first
, int count
)
529 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
530 int first
, int count
)
534 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
540 #ifdef AGGRESSIVE_CHECK
542 #define MB_CHECK_ASSERT(assert) \
546 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
547 function, file, line, # assert); \
552 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
553 const char *function
, int line
)
555 struct super_block
*sb
= e4b
->bd_sb
;
556 int order
= e4b
->bd_blkbits
+ 1;
563 struct ext4_group_info
*grp
;
566 struct list_head
*cur
;
571 static int mb_check_counter
;
572 if (mb_check_counter
++ % 100 != 0)
577 buddy
= mb_find_buddy(e4b
, order
, &max
);
578 MB_CHECK_ASSERT(buddy
);
579 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
580 MB_CHECK_ASSERT(buddy2
);
581 MB_CHECK_ASSERT(buddy
!= buddy2
);
582 MB_CHECK_ASSERT(max
* 2 == max2
);
585 for (i
= 0; i
< max
; i
++) {
587 if (mb_test_bit(i
, buddy
)) {
588 /* only single bit in buddy2 may be 1 */
589 if (!mb_test_bit(i
<< 1, buddy2
)) {
591 mb_test_bit((i
<<1)+1, buddy2
));
592 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
594 mb_test_bit(i
<< 1, buddy2
));
599 /* both bits in buddy2 must be 1 */
600 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
601 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
603 for (j
= 0; j
< (1 << order
); j
++) {
604 k
= (i
* (1 << order
)) + j
;
606 !mb_test_bit(k
, e4b
->bd_bitmap
));
610 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
615 buddy
= mb_find_buddy(e4b
, 0, &max
);
616 for (i
= 0; i
< max
; i
++) {
617 if (!mb_test_bit(i
, buddy
)) {
618 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
626 /* check used bits only */
627 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
628 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
630 MB_CHECK_ASSERT(k
< max2
);
631 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
634 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
635 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
637 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
638 list_for_each(cur
, &grp
->bb_prealloc_list
) {
639 ext4_group_t groupnr
;
640 struct ext4_prealloc_space
*pa
;
641 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
642 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
643 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
644 for (i
= 0; i
< pa
->pa_len
; i
++)
645 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
649 #undef MB_CHECK_ASSERT
650 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
651 __FILE__, __func__, __LINE__)
653 #define mb_check_buddy(e4b)
657 * Divide blocks started from @first with length @len into
658 * smaller chunks with power of 2 blocks.
659 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
660 * then increase bb_counters[] for corresponded chunk size.
662 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
663 void *buddy
, ext4_grpblk_t first
, ext4_grpblk_t len
,
664 struct ext4_group_info
*grp
)
666 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
672 BUG_ON(len
> EXT4_CLUSTERS_PER_GROUP(sb
));
674 border
= 2 << sb
->s_blocksize_bits
;
677 /* find how many blocks can be covered since this position */
678 max
= ffs(first
| border
) - 1;
680 /* find how many blocks of power 2 we need to mark */
687 /* mark multiblock chunks only */
688 grp
->bb_counters
[min
]++;
690 mb_clear_bit(first
>> min
,
691 buddy
+ sbi
->s_mb_offsets
[min
]);
699 * Cache the order of the largest free extent we have available in this block
703 mb_set_largest_free_order(struct super_block
*sb
, struct ext4_group_info
*grp
)
708 grp
->bb_largest_free_order
= -1; /* uninit */
710 bits
= sb
->s_blocksize_bits
+ 1;
711 for (i
= bits
; i
>= 0; i
--) {
712 if (grp
->bb_counters
[i
] > 0) {
713 grp
->bb_largest_free_order
= i
;
719 static noinline_for_stack
720 void ext4_mb_generate_buddy(struct super_block
*sb
,
721 void *buddy
, void *bitmap
, ext4_group_t group
)
723 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
724 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
725 ext4_grpblk_t max
= EXT4_CLUSTERS_PER_GROUP(sb
);
730 unsigned fragments
= 0;
731 unsigned long long period
= get_cycles();
733 /* initialize buddy from bitmap which is aggregation
734 * of on-disk bitmap and preallocations */
735 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
736 grp
->bb_first_free
= i
;
740 i
= mb_find_next_bit(bitmap
, max
, i
);
744 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
746 grp
->bb_counters
[0]++;
748 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
750 grp
->bb_fragments
= fragments
;
752 if (free
!= grp
->bb_free
) {
753 ext4_grp_locked_error(sb
, group
, 0, 0,
754 "block bitmap and bg descriptor "
755 "inconsistent: %u vs %u free clusters",
758 * If we intend to continue, we consider group descriptor
759 * corrupt and update bb_free using bitmap value
762 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp
))
763 percpu_counter_sub(&sbi
->s_freeclusters_counter
,
765 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT
, &grp
->bb_state
);
767 mb_set_largest_free_order(sb
, grp
);
769 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
771 period
= get_cycles() - period
;
772 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
773 EXT4_SB(sb
)->s_mb_buddies_generated
++;
774 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
775 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
778 static void mb_regenerate_buddy(struct ext4_buddy
*e4b
)
784 while ((buddy
= mb_find_buddy(e4b
, order
++, &count
))) {
785 ext4_set_bits(buddy
, 0, count
);
787 e4b
->bd_info
->bb_fragments
= 0;
788 memset(e4b
->bd_info
->bb_counters
, 0,
789 sizeof(*e4b
->bd_info
->bb_counters
) *
790 (e4b
->bd_sb
->s_blocksize_bits
+ 2));
792 ext4_mb_generate_buddy(e4b
->bd_sb
, e4b
->bd_buddy
,
793 e4b
->bd_bitmap
, e4b
->bd_group
);
796 /* The buddy information is attached the buddy cache inode
797 * for convenience. The information regarding each group
798 * is loaded via ext4_mb_load_buddy. The information involve
799 * block bitmap and buddy information. The information are
800 * stored in the inode as
803 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
806 * one block each for bitmap and buddy information.
807 * So for each group we take up 2 blocks. A page can
808 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
809 * So it can have information regarding groups_per_page which
810 * is blocks_per_page/2
812 * Locking note: This routine takes the block group lock of all groups
813 * for this page; do not hold this lock when calling this routine!
816 static int ext4_mb_init_cache(struct page
*page
, char *incore
, gfp_t gfp
)
818 ext4_group_t ngroups
;
824 ext4_group_t first_group
, group
;
826 struct super_block
*sb
;
827 struct buffer_head
*bhs
;
828 struct buffer_head
**bh
= NULL
;
832 struct ext4_group_info
*grinfo
;
834 mb_debug(1, "init page %lu\n", page
->index
);
836 inode
= page
->mapping
->host
;
838 ngroups
= ext4_get_groups_count(sb
);
839 blocksize
= i_blocksize(inode
);
840 blocks_per_page
= PAGE_SIZE
/ blocksize
;
842 groups_per_page
= blocks_per_page
>> 1;
843 if (groups_per_page
== 0)
846 /* allocate buffer_heads to read bitmaps */
847 if (groups_per_page
> 1) {
848 i
= sizeof(struct buffer_head
*) * groups_per_page
;
849 bh
= kzalloc(i
, gfp
);
857 first_group
= page
->index
* blocks_per_page
/ 2;
859 /* read all groups the page covers into the cache */
860 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
861 if (group
>= ngroups
)
864 grinfo
= ext4_get_group_info(sb
, group
);
866 * If page is uptodate then we came here after online resize
867 * which added some new uninitialized group info structs, so
868 * we must skip all initialized uptodate buddies on the page,
869 * which may be currently in use by an allocating task.
871 if (PageUptodate(page
) && !EXT4_MB_GRP_NEED_INIT(grinfo
)) {
875 bh
[i
] = ext4_read_block_bitmap_nowait(sb
, group
);
877 err
= PTR_ERR(bh
[i
]);
881 mb_debug(1, "read bitmap for group %u\n", group
);
884 /* wait for I/O completion */
885 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
890 err2
= ext4_wait_block_bitmap(sb
, group
, bh
[i
]);
895 first_block
= page
->index
* blocks_per_page
;
896 for (i
= 0; i
< blocks_per_page
; i
++) {
897 group
= (first_block
+ i
) >> 1;
898 if (group
>= ngroups
)
901 if (!bh
[group
- first_group
])
902 /* skip initialized uptodate buddy */
905 if (!buffer_verified(bh
[group
- first_group
]))
906 /* Skip faulty bitmaps */
911 * data carry information regarding this
912 * particular group in the format specified
916 data
= page_address(page
) + (i
* blocksize
);
917 bitmap
= bh
[group
- first_group
]->b_data
;
920 * We place the buddy block and bitmap block
923 if ((first_block
+ i
) & 1) {
924 /* this is block of buddy */
925 BUG_ON(incore
== NULL
);
926 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
927 group
, page
->index
, i
* blocksize
);
928 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
929 grinfo
= ext4_get_group_info(sb
, group
);
930 grinfo
->bb_fragments
= 0;
931 memset(grinfo
->bb_counters
, 0,
932 sizeof(*grinfo
->bb_counters
) *
933 (sb
->s_blocksize_bits
+2));
935 * incore got set to the group block bitmap below
937 ext4_lock_group(sb
, group
);
939 memset(data
, 0xff, blocksize
);
940 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
941 ext4_unlock_group(sb
, group
);
944 /* this is block of bitmap */
945 BUG_ON(incore
!= NULL
);
946 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
947 group
, page
->index
, i
* blocksize
);
948 trace_ext4_mb_bitmap_load(sb
, group
);
950 /* see comments in ext4_mb_put_pa() */
951 ext4_lock_group(sb
, group
);
952 memcpy(data
, bitmap
, blocksize
);
954 /* mark all preallocated blks used in in-core bitmap */
955 ext4_mb_generate_from_pa(sb
, data
, group
);
956 ext4_mb_generate_from_freelist(sb
, data
, group
);
957 ext4_unlock_group(sb
, group
);
959 /* set incore so that the buddy information can be
960 * generated using this
965 SetPageUptodate(page
);
969 for (i
= 0; i
< groups_per_page
; i
++)
978 * Lock the buddy and bitmap pages. This make sure other parallel init_group
979 * on the same buddy page doesn't happen whild holding the buddy page lock.
980 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
981 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
983 static int ext4_mb_get_buddy_page_lock(struct super_block
*sb
,
984 ext4_group_t group
, struct ext4_buddy
*e4b
, gfp_t gfp
)
986 struct inode
*inode
= EXT4_SB(sb
)->s_buddy_cache
;
987 int block
, pnum
, poff
;
991 e4b
->bd_buddy_page
= NULL
;
992 e4b
->bd_bitmap_page
= NULL
;
994 blocks_per_page
= PAGE_SIZE
/ sb
->s_blocksize
;
996 * the buddy cache inode stores the block bitmap
997 * and buddy information in consecutive blocks.
998 * So for each group we need two blocks.
1001 pnum
= block
/ blocks_per_page
;
1002 poff
= block
% blocks_per_page
;
1003 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1006 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1007 e4b
->bd_bitmap_page
= page
;
1008 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1010 if (blocks_per_page
>= 2) {
1011 /* buddy and bitmap are on the same page */
1016 pnum
= block
/ blocks_per_page
;
1017 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1020 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1021 e4b
->bd_buddy_page
= page
;
1025 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy
*e4b
)
1027 if (e4b
->bd_bitmap_page
) {
1028 unlock_page(e4b
->bd_bitmap_page
);
1029 put_page(e4b
->bd_bitmap_page
);
1031 if (e4b
->bd_buddy_page
) {
1032 unlock_page(e4b
->bd_buddy_page
);
1033 put_page(e4b
->bd_buddy_page
);
1038 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1039 * block group lock of all groups for this page; do not hold the BG lock when
1040 * calling this routine!
1042 static noinline_for_stack
1043 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
, gfp_t gfp
)
1046 struct ext4_group_info
*this_grp
;
1047 struct ext4_buddy e4b
;
1052 mb_debug(1, "init group %u\n", group
);
1053 this_grp
= ext4_get_group_info(sb
, group
);
1055 * This ensures that we don't reinit the buddy cache
1056 * page which map to the group from which we are already
1057 * allocating. If we are looking at the buddy cache we would
1058 * have taken a reference using ext4_mb_load_buddy and that
1059 * would have pinned buddy page to page cache.
1060 * The call to ext4_mb_get_buddy_page_lock will mark the
1063 ret
= ext4_mb_get_buddy_page_lock(sb
, group
, &e4b
, gfp
);
1064 if (ret
|| !EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1066 * somebody initialized the group
1067 * return without doing anything
1072 page
= e4b
.bd_bitmap_page
;
1073 ret
= ext4_mb_init_cache(page
, NULL
, gfp
);
1076 if (!PageUptodate(page
)) {
1081 if (e4b
.bd_buddy_page
== NULL
) {
1083 * If both the bitmap and buddy are in
1084 * the same page we don't need to force
1090 /* init buddy cache */
1091 page
= e4b
.bd_buddy_page
;
1092 ret
= ext4_mb_init_cache(page
, e4b
.bd_bitmap
, gfp
);
1095 if (!PageUptodate(page
)) {
1100 ext4_mb_put_buddy_page_lock(&e4b
);
1105 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1106 * block group lock of all groups for this page; do not hold the BG lock when
1107 * calling this routine!
1109 static noinline_for_stack
int
1110 ext4_mb_load_buddy_gfp(struct super_block
*sb
, ext4_group_t group
,
1111 struct ext4_buddy
*e4b
, gfp_t gfp
)
1113 int blocks_per_page
;
1119 struct ext4_group_info
*grp
;
1120 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1121 struct inode
*inode
= sbi
->s_buddy_cache
;
1124 mb_debug(1, "load group %u\n", group
);
1126 blocks_per_page
= PAGE_SIZE
/ sb
->s_blocksize
;
1127 grp
= ext4_get_group_info(sb
, group
);
1129 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1132 e4b
->bd_group
= group
;
1133 e4b
->bd_buddy_page
= NULL
;
1134 e4b
->bd_bitmap_page
= NULL
;
1136 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1138 * we need full data about the group
1139 * to make a good selection
1141 ret
= ext4_mb_init_group(sb
, group
, gfp
);
1147 * the buddy cache inode stores the block bitmap
1148 * and buddy information in consecutive blocks.
1149 * So for each group we need two blocks.
1152 pnum
= block
/ blocks_per_page
;
1153 poff
= block
% blocks_per_page
;
1155 /* we could use find_or_create_page(), but it locks page
1156 * what we'd like to avoid in fast path ... */
1157 page
= find_get_page_flags(inode
->i_mapping
, pnum
, FGP_ACCESSED
);
1158 if (page
== NULL
|| !PageUptodate(page
)) {
1161 * drop the page reference and try
1162 * to get the page with lock. If we
1163 * are not uptodate that implies
1164 * somebody just created the page but
1165 * is yet to initialize the same. So
1166 * wait for it to initialize.
1169 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1171 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1172 if (!PageUptodate(page
)) {
1173 ret
= ext4_mb_init_cache(page
, NULL
, gfp
);
1178 mb_cmp_bitmaps(e4b
, page_address(page
) +
1179 (poff
* sb
->s_blocksize
));
1188 if (!PageUptodate(page
)) {
1193 /* Pages marked accessed already */
1194 e4b
->bd_bitmap_page
= page
;
1195 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1198 pnum
= block
/ blocks_per_page
;
1199 poff
= block
% blocks_per_page
;
1201 page
= find_get_page_flags(inode
->i_mapping
, pnum
, FGP_ACCESSED
);
1202 if (page
== NULL
|| !PageUptodate(page
)) {
1205 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1207 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1208 if (!PageUptodate(page
)) {
1209 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
,
1223 if (!PageUptodate(page
)) {
1228 /* Pages marked accessed already */
1229 e4b
->bd_buddy_page
= page
;
1230 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1232 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1233 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1240 if (e4b
->bd_bitmap_page
)
1241 put_page(e4b
->bd_bitmap_page
);
1242 if (e4b
->bd_buddy_page
)
1243 put_page(e4b
->bd_buddy_page
);
1244 e4b
->bd_buddy
= NULL
;
1245 e4b
->bd_bitmap
= NULL
;
1249 static int ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1250 struct ext4_buddy
*e4b
)
1252 return ext4_mb_load_buddy_gfp(sb
, group
, e4b
, GFP_NOFS
);
1255 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1257 if (e4b
->bd_bitmap_page
)
1258 put_page(e4b
->bd_bitmap_page
);
1259 if (e4b
->bd_buddy_page
)
1260 put_page(e4b
->bd_buddy_page
);
1264 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1267 int bb_incr
= 1 << (e4b
->bd_blkbits
- 1);
1270 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
1271 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1274 while (order
<= e4b
->bd_blkbits
+ 1) {
1276 if (!mb_test_bit(block
, bb
)) {
1277 /* this block is part of buddy of order 'order' */
1287 static void mb_clear_bits(void *bm
, int cur
, int len
)
1293 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1294 /* fast path: clear whole word at once */
1295 addr
= bm
+ (cur
>> 3);
1300 mb_clear_bit(cur
, bm
);
1305 /* clear bits in given range
1306 * will return first found zero bit if any, -1 otherwise
1308 static int mb_test_and_clear_bits(void *bm
, int cur
, int len
)
1315 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1316 /* fast path: clear whole word at once */
1317 addr
= bm
+ (cur
>> 3);
1318 if (*addr
!= (__u32
)(-1) && zero_bit
== -1)
1319 zero_bit
= cur
+ mb_find_next_zero_bit(addr
, 32, 0);
1324 if (!mb_test_and_clear_bit(cur
, bm
) && zero_bit
== -1)
1332 void ext4_set_bits(void *bm
, int cur
, int len
)
1338 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1339 /* fast path: set whole word at once */
1340 addr
= bm
+ (cur
>> 3);
1345 mb_set_bit(cur
, bm
);
1351 * _________________________________________________________________ */
1353 static inline int mb_buddy_adjust_border(int* bit
, void* bitmap
, int side
)
1355 if (mb_test_bit(*bit
+ side
, bitmap
)) {
1356 mb_clear_bit(*bit
, bitmap
);
1362 mb_set_bit(*bit
, bitmap
);
1367 static void mb_buddy_mark_free(struct ext4_buddy
*e4b
, int first
, int last
)
1371 void *buddy
= mb_find_buddy(e4b
, order
, &max
);
1376 /* Bits in range [first; last] are known to be set since
1377 * corresponding blocks were allocated. Bits in range
1378 * (first; last) will stay set because they form buddies on
1379 * upper layer. We just deal with borders if they don't
1380 * align with upper layer and then go up.
1381 * Releasing entire group is all about clearing
1382 * single bit of highest order buddy.
1386 * ---------------------------------
1388 * ---------------------------------
1389 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1390 * ---------------------------------
1392 * \_____________________/
1394 * Neither [1] nor [6] is aligned to above layer.
1395 * Left neighbour [0] is free, so mark it busy,
1396 * decrease bb_counters and extend range to
1398 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1399 * mark [6] free, increase bb_counters and shrink range to
1401 * Then shift range to [0; 2], go up and do the same.
1406 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&first
, buddy
, -1);
1408 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&last
, buddy
, 1);
1413 if (first
== last
|| !(buddy2
= mb_find_buddy(e4b
, order
, &max
))) {
1414 mb_clear_bits(buddy
, first
, last
- first
+ 1);
1415 e4b
->bd_info
->bb_counters
[order
- 1] += last
- first
+ 1;
1424 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1425 int first
, int count
)
1427 int left_is_free
= 0;
1428 int right_is_free
= 0;
1430 int last
= first
+ count
- 1;
1431 struct super_block
*sb
= e4b
->bd_sb
;
1433 if (WARN_ON(count
== 0))
1435 BUG_ON(last
>= (sb
->s_blocksize
<< 3));
1436 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1437 /* Don't bother if the block group is corrupt. */
1438 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
)))
1441 mb_check_buddy(e4b
);
1442 mb_free_blocks_double(inode
, e4b
, first
, count
);
1444 e4b
->bd_info
->bb_free
+= count
;
1445 if (first
< e4b
->bd_info
->bb_first_free
)
1446 e4b
->bd_info
->bb_first_free
= first
;
1448 /* access memory sequentially: check left neighbour,
1449 * clear range and then check right neighbour
1452 left_is_free
= !mb_test_bit(first
- 1, e4b
->bd_bitmap
);
1453 block
= mb_test_and_clear_bits(e4b
->bd_bitmap
, first
, count
);
1454 if (last
+ 1 < EXT4_SB(sb
)->s_mb_maxs
[0])
1455 right_is_free
= !mb_test_bit(last
+ 1, e4b
->bd_bitmap
);
1457 if (unlikely(block
!= -1)) {
1458 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1459 ext4_fsblk_t blocknr
;
1461 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1462 blocknr
+= EXT4_C2B(EXT4_SB(sb
), block
);
1463 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1464 inode
? inode
->i_ino
: 0,
1466 "freeing already freed block "
1467 "(bit %u); block bitmap corrupt.",
1469 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
))
1470 percpu_counter_sub(&sbi
->s_freeclusters_counter
,
1471 e4b
->bd_info
->bb_free
);
1472 /* Mark the block group as corrupt. */
1473 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT
,
1474 &e4b
->bd_info
->bb_state
);
1475 mb_regenerate_buddy(e4b
);
1479 /* let's maintain fragments counter */
1480 if (left_is_free
&& right_is_free
)
1481 e4b
->bd_info
->bb_fragments
--;
1482 else if (!left_is_free
&& !right_is_free
)
1483 e4b
->bd_info
->bb_fragments
++;
1485 /* buddy[0] == bd_bitmap is a special case, so handle
1486 * it right away and let mb_buddy_mark_free stay free of
1487 * zero order checks.
1488 * Check if neighbours are to be coaleasced,
1489 * adjust bitmap bb_counters and borders appropriately.
1492 first
+= !left_is_free
;
1493 e4b
->bd_info
->bb_counters
[0] += left_is_free
? -1 : 1;
1496 last
-= !right_is_free
;
1497 e4b
->bd_info
->bb_counters
[0] += right_is_free
? -1 : 1;
1501 mb_buddy_mark_free(e4b
, first
>> 1, last
>> 1);
1504 mb_set_largest_free_order(sb
, e4b
->bd_info
);
1505 mb_check_buddy(e4b
);
1508 static int mb_find_extent(struct ext4_buddy
*e4b
, int block
,
1509 int needed
, struct ext4_free_extent
*ex
)
1515 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1518 buddy
= mb_find_buddy(e4b
, 0, &max
);
1519 BUG_ON(buddy
== NULL
);
1520 BUG_ON(block
>= max
);
1521 if (mb_test_bit(block
, buddy
)) {
1528 /* find actual order */
1529 order
= mb_find_order_for_block(e4b
, block
);
1530 block
= block
>> order
;
1532 ex
->fe_len
= 1 << order
;
1533 ex
->fe_start
= block
<< order
;
1534 ex
->fe_group
= e4b
->bd_group
;
1536 /* calc difference from given start */
1537 next
= next
- ex
->fe_start
;
1539 ex
->fe_start
+= next
;
1541 while (needed
> ex
->fe_len
&&
1542 mb_find_buddy(e4b
, order
, &max
)) {
1544 if (block
+ 1 >= max
)
1547 next
= (block
+ 1) * (1 << order
);
1548 if (mb_test_bit(next
, e4b
->bd_bitmap
))
1551 order
= mb_find_order_for_block(e4b
, next
);
1553 block
= next
>> order
;
1554 ex
->fe_len
+= 1 << order
;
1557 if (ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3))) {
1558 /* Should never happen! (but apparently sometimes does?!?) */
1560 ext4_error(e4b
->bd_sb
, "corruption or bug in mb_find_extent "
1561 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1562 block
, order
, needed
, ex
->fe_group
, ex
->fe_start
,
1563 ex
->fe_len
, ex
->fe_logical
);
1571 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1577 int start
= ex
->fe_start
;
1578 int len
= ex
->fe_len
;
1583 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1584 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1585 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1586 mb_check_buddy(e4b
);
1587 mb_mark_used_double(e4b
, start
, len
);
1589 e4b
->bd_info
->bb_free
-= len
;
1590 if (e4b
->bd_info
->bb_first_free
== start
)
1591 e4b
->bd_info
->bb_first_free
+= len
;
1593 /* let's maintain fragments counter */
1595 mlen
= !mb_test_bit(start
- 1, e4b
->bd_bitmap
);
1596 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1597 max
= !mb_test_bit(start
+ len
, e4b
->bd_bitmap
);
1599 e4b
->bd_info
->bb_fragments
++;
1600 else if (!mlen
&& !max
)
1601 e4b
->bd_info
->bb_fragments
--;
1603 /* let's maintain buddy itself */
1605 ord
= mb_find_order_for_block(e4b
, start
);
1607 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1608 /* the whole chunk may be allocated at once! */
1610 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1611 BUG_ON((start
>> ord
) >= max
);
1612 mb_set_bit(start
>> ord
, buddy
);
1613 e4b
->bd_info
->bb_counters
[ord
]--;
1620 /* store for history */
1622 ret
= len
| (ord
<< 16);
1624 /* we have to split large buddy */
1626 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1627 mb_set_bit(start
>> ord
, buddy
);
1628 e4b
->bd_info
->bb_counters
[ord
]--;
1631 cur
= (start
>> ord
) & ~1U;
1632 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1633 mb_clear_bit(cur
, buddy
);
1634 mb_clear_bit(cur
+ 1, buddy
);
1635 e4b
->bd_info
->bb_counters
[ord
]++;
1636 e4b
->bd_info
->bb_counters
[ord
]++;
1638 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1640 ext4_set_bits(e4b
->bd_bitmap
, ex
->fe_start
, len0
);
1641 mb_check_buddy(e4b
);
1647 * Must be called under group lock!
1649 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1650 struct ext4_buddy
*e4b
)
1652 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1655 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1656 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1658 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1659 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1660 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1662 /* preallocation can change ac_b_ex, thus we store actually
1663 * allocated blocks for history */
1664 ac
->ac_f_ex
= ac
->ac_b_ex
;
1666 ac
->ac_status
= AC_STATUS_FOUND
;
1667 ac
->ac_tail
= ret
& 0xffff;
1668 ac
->ac_buddy
= ret
>> 16;
1671 * take the page reference. We want the page to be pinned
1672 * so that we don't get a ext4_mb_init_cache_call for this
1673 * group until we update the bitmap. That would mean we
1674 * double allocate blocks. The reference is dropped
1675 * in ext4_mb_release_context
1677 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1678 get_page(ac
->ac_bitmap_page
);
1679 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1680 get_page(ac
->ac_buddy_page
);
1681 /* store last allocated for subsequent stream allocation */
1682 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1683 spin_lock(&sbi
->s_md_lock
);
1684 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1685 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1686 spin_unlock(&sbi
->s_md_lock
);
1691 * regular allocator, for general purposes allocation
1694 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1695 struct ext4_buddy
*e4b
,
1698 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1699 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1700 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1701 struct ext4_free_extent ex
;
1704 if (ac
->ac_status
== AC_STATUS_FOUND
)
1707 * We don't want to scan for a whole year
1709 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1710 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1711 ac
->ac_status
= AC_STATUS_BREAK
;
1716 * Haven't found good chunk so far, let's continue
1718 if (bex
->fe_len
< gex
->fe_len
)
1721 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1722 && bex
->fe_group
== e4b
->bd_group
) {
1723 /* recheck chunk's availability - we don't know
1724 * when it was found (within this lock-unlock
1726 max
= mb_find_extent(e4b
, bex
->fe_start
, gex
->fe_len
, &ex
);
1727 if (max
>= gex
->fe_len
) {
1728 ext4_mb_use_best_found(ac
, e4b
);
1735 * The routine checks whether found extent is good enough. If it is,
1736 * then the extent gets marked used and flag is set to the context
1737 * to stop scanning. Otherwise, the extent is compared with the
1738 * previous found extent and if new one is better, then it's stored
1739 * in the context. Later, the best found extent will be used, if
1740 * mballoc can't find good enough extent.
1742 * FIXME: real allocation policy is to be designed yet!
1744 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1745 struct ext4_free_extent
*ex
,
1746 struct ext4_buddy
*e4b
)
1748 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1749 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1751 BUG_ON(ex
->fe_len
<= 0);
1752 BUG_ON(ex
->fe_len
> EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1753 BUG_ON(ex
->fe_start
>= EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1754 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1759 * The special case - take what you catch first
1761 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1763 ext4_mb_use_best_found(ac
, e4b
);
1768 * Let's check whether the chuck is good enough
1770 if (ex
->fe_len
== gex
->fe_len
) {
1772 ext4_mb_use_best_found(ac
, e4b
);
1777 * If this is first found extent, just store it in the context
1779 if (bex
->fe_len
== 0) {
1785 * If new found extent is better, store it in the context
1787 if (bex
->fe_len
< gex
->fe_len
) {
1788 /* if the request isn't satisfied, any found extent
1789 * larger than previous best one is better */
1790 if (ex
->fe_len
> bex
->fe_len
)
1792 } else if (ex
->fe_len
> gex
->fe_len
) {
1793 /* if the request is satisfied, then we try to find
1794 * an extent that still satisfy the request, but is
1795 * smaller than previous one */
1796 if (ex
->fe_len
< bex
->fe_len
)
1800 ext4_mb_check_limits(ac
, e4b
, 0);
1803 static noinline_for_stack
1804 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1805 struct ext4_buddy
*e4b
)
1807 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1808 ext4_group_t group
= ex
.fe_group
;
1812 BUG_ON(ex
.fe_len
<= 0);
1813 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1817 ext4_lock_group(ac
->ac_sb
, group
);
1818 max
= mb_find_extent(e4b
, ex
.fe_start
, ex
.fe_len
, &ex
);
1822 ext4_mb_use_best_found(ac
, e4b
);
1825 ext4_unlock_group(ac
->ac_sb
, group
);
1826 ext4_mb_unload_buddy(e4b
);
1831 static noinline_for_stack
1832 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1833 struct ext4_buddy
*e4b
)
1835 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1838 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1839 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1840 struct ext4_free_extent ex
;
1842 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1844 if (grp
->bb_free
== 0)
1847 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1851 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
))) {
1852 ext4_mb_unload_buddy(e4b
);
1856 ext4_lock_group(ac
->ac_sb
, group
);
1857 max
= mb_find_extent(e4b
, ac
->ac_g_ex
.fe_start
,
1858 ac
->ac_g_ex
.fe_len
, &ex
);
1859 ex
.fe_logical
= 0xDEADFA11; /* debug value */
1861 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1864 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1866 /* use do_div to get remainder (would be 64-bit modulo) */
1867 if (do_div(start
, sbi
->s_stripe
) == 0) {
1870 ext4_mb_use_best_found(ac
, e4b
);
1872 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1873 BUG_ON(ex
.fe_len
<= 0);
1874 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1875 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1878 ext4_mb_use_best_found(ac
, e4b
);
1879 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1880 /* Sometimes, caller may want to merge even small
1881 * number of blocks to an existing extent */
1882 BUG_ON(ex
.fe_len
<= 0);
1883 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1884 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1887 ext4_mb_use_best_found(ac
, e4b
);
1889 ext4_unlock_group(ac
->ac_sb
, group
);
1890 ext4_mb_unload_buddy(e4b
);
1896 * The routine scans buddy structures (not bitmap!) from given order
1897 * to max order and tries to find big enough chunk to satisfy the req
1899 static noinline_for_stack
1900 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1901 struct ext4_buddy
*e4b
)
1903 struct super_block
*sb
= ac
->ac_sb
;
1904 struct ext4_group_info
*grp
= e4b
->bd_info
;
1910 BUG_ON(ac
->ac_2order
<= 0);
1911 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1912 if (grp
->bb_counters
[i
] == 0)
1915 buddy
= mb_find_buddy(e4b
, i
, &max
);
1916 BUG_ON(buddy
== NULL
);
1918 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1923 ac
->ac_b_ex
.fe_len
= 1 << i
;
1924 ac
->ac_b_ex
.fe_start
= k
<< i
;
1925 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1927 ext4_mb_use_best_found(ac
, e4b
);
1929 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1931 if (EXT4_SB(sb
)->s_mb_stats
)
1932 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1939 * The routine scans the group and measures all found extents.
1940 * In order to optimize scanning, caller must pass number of
1941 * free blocks in the group, so the routine can know upper limit.
1943 static noinline_for_stack
1944 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1945 struct ext4_buddy
*e4b
)
1947 struct super_block
*sb
= ac
->ac_sb
;
1948 void *bitmap
= e4b
->bd_bitmap
;
1949 struct ext4_free_extent ex
;
1953 free
= e4b
->bd_info
->bb_free
;
1956 i
= e4b
->bd_info
->bb_first_free
;
1958 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1959 i
= mb_find_next_zero_bit(bitmap
,
1960 EXT4_CLUSTERS_PER_GROUP(sb
), i
);
1961 if (i
>= EXT4_CLUSTERS_PER_GROUP(sb
)) {
1963 * IF we have corrupt bitmap, we won't find any
1964 * free blocks even though group info says we
1965 * we have free blocks
1967 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1968 "%d free clusters as per "
1969 "group info. But bitmap says 0",
1974 mb_find_extent(e4b
, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1975 BUG_ON(ex
.fe_len
<= 0);
1976 if (free
< ex
.fe_len
) {
1977 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1978 "%d free clusters as per "
1979 "group info. But got %d blocks",
1982 * The number of free blocks differs. This mostly
1983 * indicate that the bitmap is corrupt. So exit
1984 * without claiming the space.
1988 ex
.fe_logical
= 0xDEADC0DE; /* debug value */
1989 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1995 ext4_mb_check_limits(ac
, e4b
, 1);
1999 * This is a special case for storages like raid5
2000 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2002 static noinline_for_stack
2003 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
2004 struct ext4_buddy
*e4b
)
2006 struct super_block
*sb
= ac
->ac_sb
;
2007 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2008 void *bitmap
= e4b
->bd_bitmap
;
2009 struct ext4_free_extent ex
;
2010 ext4_fsblk_t first_group_block
;
2015 BUG_ON(sbi
->s_stripe
== 0);
2017 /* find first stripe-aligned block in group */
2018 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
2020 a
= first_group_block
+ sbi
->s_stripe
- 1;
2021 do_div(a
, sbi
->s_stripe
);
2022 i
= (a
* sbi
->s_stripe
) - first_group_block
;
2024 while (i
< EXT4_CLUSTERS_PER_GROUP(sb
)) {
2025 if (!mb_test_bit(i
, bitmap
)) {
2026 max
= mb_find_extent(e4b
, i
, sbi
->s_stripe
, &ex
);
2027 if (max
>= sbi
->s_stripe
) {
2029 ex
.fe_logical
= 0xDEADF00D; /* debug value */
2031 ext4_mb_use_best_found(ac
, e4b
);
2040 * This is now called BEFORE we load the buddy bitmap.
2041 * Returns either 1 or 0 indicating that the group is either suitable
2042 * for the allocation or not. In addition it can also return negative
2043 * error code when something goes wrong.
2045 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
2046 ext4_group_t group
, int cr
)
2048 unsigned free
, fragments
;
2049 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
2050 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
2052 BUG_ON(cr
< 0 || cr
>= 4);
2054 free
= grp
->bb_free
;
2057 if (cr
<= 2 && free
< ac
->ac_g_ex
.fe_len
)
2060 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp
)))
2063 /* We only do this if the grp has never been initialized */
2064 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
2065 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
, GFP_NOFS
);
2070 fragments
= grp
->bb_fragments
;
2076 BUG_ON(ac
->ac_2order
== 0);
2078 /* Avoid using the first bg of a flexgroup for data files */
2079 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
2080 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
2081 ((group
% flex_size
) == 0))
2084 if ((ac
->ac_2order
> ac
->ac_sb
->s_blocksize_bits
+1) ||
2085 (free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2088 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
2093 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2097 if (free
>= ac
->ac_g_ex
.fe_len
)
2109 static noinline_for_stack
int
2110 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
2112 ext4_group_t ngroups
, group
, i
;
2114 int err
= 0, first_err
= 0;
2115 struct ext4_sb_info
*sbi
;
2116 struct super_block
*sb
;
2117 struct ext4_buddy e4b
;
2121 ngroups
= ext4_get_groups_count(sb
);
2122 /* non-extent files are limited to low blocks/groups */
2123 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
2124 ngroups
= sbi
->s_blockfile_groups
;
2126 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
2128 /* first, try the goal */
2129 err
= ext4_mb_find_by_goal(ac
, &e4b
);
2130 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
2133 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2137 * ac->ac2_order is set only if the fe_len is a power of 2
2138 * if ac2_order is set we also set criteria to 0 so that we
2139 * try exact allocation using buddy.
2141 i
= fls(ac
->ac_g_ex
.fe_len
);
2144 * We search using buddy data only if the order of the request
2145 * is greater than equal to the sbi_s_mb_order2_reqs
2146 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2147 * We also support searching for power-of-two requests only for
2148 * requests upto maximum buddy size we have constructed.
2150 if (i
>= sbi
->s_mb_order2_reqs
&& i
<= sb
->s_blocksize_bits
+ 2) {
2152 * This should tell if fe_len is exactly power of 2
2154 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2155 ac
->ac_2order
= i
- 1;
2158 /* if stream allocation is enabled, use global goal */
2159 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2160 /* TBD: may be hot point */
2161 spin_lock(&sbi
->s_md_lock
);
2162 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2163 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2164 spin_unlock(&sbi
->s_md_lock
);
2167 /* Let's just scan groups to find more-less suitable blocks */
2168 cr
= ac
->ac_2order
? 0 : 1;
2170 * cr == 0 try to get exact allocation,
2171 * cr == 3 try to get anything
2174 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2175 ac
->ac_criteria
= cr
;
2177 * searching for the right group start
2178 * from the goal value specified
2180 group
= ac
->ac_g_ex
.fe_group
;
2182 for (i
= 0; i
< ngroups
; group
++, i
++) {
2186 * Artificially restricted ngroups for non-extent
2187 * files makes group > ngroups possible on first loop.
2189 if (group
>= ngroups
)
2192 /* This now checks without needing the buddy page */
2193 ret
= ext4_mb_good_group(ac
, group
, cr
);
2200 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2204 ext4_lock_group(sb
, group
);
2207 * We need to check again after locking the
2210 ret
= ext4_mb_good_group(ac
, group
, cr
);
2212 ext4_unlock_group(sb
, group
);
2213 ext4_mb_unload_buddy(&e4b
);
2219 ac
->ac_groups_scanned
++;
2221 ext4_mb_simple_scan_group(ac
, &e4b
);
2222 else if (cr
== 1 && sbi
->s_stripe
&&
2223 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2224 ext4_mb_scan_aligned(ac
, &e4b
);
2226 ext4_mb_complex_scan_group(ac
, &e4b
);
2228 ext4_unlock_group(sb
, group
);
2229 ext4_mb_unload_buddy(&e4b
);
2231 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2236 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2237 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2239 * We've been searching too long. Let's try to allocate
2240 * the best chunk we've found so far
2243 ext4_mb_try_best_found(ac
, &e4b
);
2244 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2246 * Someone more lucky has already allocated it.
2247 * The only thing we can do is just take first
2249 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2251 ac
->ac_b_ex
.fe_group
= 0;
2252 ac
->ac_b_ex
.fe_start
= 0;
2253 ac
->ac_b_ex
.fe_len
= 0;
2254 ac
->ac_status
= AC_STATUS_CONTINUE
;
2255 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2257 atomic_inc(&sbi
->s_mb_lost_chunks
);
2262 if (!err
&& ac
->ac_status
!= AC_STATUS_FOUND
&& first_err
)
2267 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2269 struct super_block
*sb
= seq
->private;
2272 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2275 return (void *) ((unsigned long) group
);
2278 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2280 struct super_block
*sb
= seq
->private;
2284 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2287 return (void *) ((unsigned long) group
);
2290 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2292 struct super_block
*sb
= seq
->private;
2293 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2295 int err
, buddy_loaded
= 0;
2296 struct ext4_buddy e4b
;
2297 struct ext4_group_info
*grinfo
;
2299 struct ext4_group_info info
;
2300 ext4_grpblk_t counters
[EXT4_MAX_BLOCK_LOG_SIZE
+ 2];
2305 seq_puts(seq
, "#group: free frags first ["
2306 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2307 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2309 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2310 sizeof(struct ext4_group_info
);
2311 grinfo
= ext4_get_group_info(sb
, group
);
2312 /* Load the group info in memory only if not already loaded. */
2313 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo
))) {
2314 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2316 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2322 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2325 ext4_mb_unload_buddy(&e4b
);
2327 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2328 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2329 for (i
= 0; i
<= 13; i
++)
2330 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2331 sg
.info
.bb_counters
[i
] : 0);
2332 seq_printf(seq
, " ]\n");
2337 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2341 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2342 .start
= ext4_mb_seq_groups_start
,
2343 .next
= ext4_mb_seq_groups_next
,
2344 .stop
= ext4_mb_seq_groups_stop
,
2345 .show
= ext4_mb_seq_groups_show
,
2348 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2350 struct super_block
*sb
= PDE_DATA(inode
);
2353 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2355 struct seq_file
*m
= file
->private_data
;
2362 const struct file_operations ext4_seq_mb_groups_fops
= {
2363 .open
= ext4_mb_seq_groups_open
,
2365 .llseek
= seq_lseek
,
2366 .release
= seq_release
,
2369 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2371 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2372 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2379 * Allocate the top-level s_group_info array for the specified number
2382 int ext4_mb_alloc_groupinfo(struct super_block
*sb
, ext4_group_t ngroups
)
2384 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2386 struct ext4_group_info
***new_groupinfo
;
2388 size
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2389 EXT4_DESC_PER_BLOCK_BITS(sb
);
2390 if (size
<= sbi
->s_group_info_size
)
2393 size
= roundup_pow_of_two(sizeof(*sbi
->s_group_info
) * size
);
2394 new_groupinfo
= kvzalloc(size
, GFP_KERNEL
);
2395 if (!new_groupinfo
) {
2396 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy meta group");
2399 if (sbi
->s_group_info
) {
2400 memcpy(new_groupinfo
, sbi
->s_group_info
,
2401 sbi
->s_group_info_size
* sizeof(*sbi
->s_group_info
));
2402 kvfree(sbi
->s_group_info
);
2404 sbi
->s_group_info
= new_groupinfo
;
2405 sbi
->s_group_info_size
= size
/ sizeof(*sbi
->s_group_info
);
2406 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2407 sbi
->s_group_info_size
);
2411 /* Create and initialize ext4_group_info data for the given group. */
2412 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2413 struct ext4_group_desc
*desc
)
2417 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2418 struct ext4_group_info
**meta_group_info
;
2419 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2422 * First check if this group is the first of a reserved block.
2423 * If it's true, we have to allocate a new table of pointers
2424 * to ext4_group_info structures
2426 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2427 metalen
= sizeof(*meta_group_info
) <<
2428 EXT4_DESC_PER_BLOCK_BITS(sb
);
2429 meta_group_info
= kmalloc(metalen
, GFP_NOFS
);
2430 if (meta_group_info
== NULL
) {
2431 ext4_msg(sb
, KERN_ERR
, "can't allocate mem "
2432 "for a buddy group");
2433 goto exit_meta_group_info
;
2435 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2440 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2441 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2443 meta_group_info
[i
] = kmem_cache_zalloc(cachep
, GFP_NOFS
);
2444 if (meta_group_info
[i
] == NULL
) {
2445 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy mem");
2446 goto exit_group_info
;
2448 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2449 &(meta_group_info
[i
]->bb_state
));
2452 * initialize bb_free to be able to skip
2453 * empty groups without initialization
2455 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2456 meta_group_info
[i
]->bb_free
=
2457 ext4_free_clusters_after_init(sb
, group
, desc
);
2459 meta_group_info
[i
]->bb_free
=
2460 ext4_free_group_clusters(sb
, desc
);
2463 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2464 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2465 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2466 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2470 struct buffer_head
*bh
;
2471 meta_group_info
[i
]->bb_bitmap
=
2472 kmalloc(sb
->s_blocksize
, GFP_NOFS
);
2473 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2474 bh
= ext4_read_block_bitmap(sb
, group
);
2475 BUG_ON(IS_ERR_OR_NULL(bh
));
2476 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2485 /* If a meta_group_info table has been allocated, release it now */
2486 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2487 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2488 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] = NULL
;
2490 exit_meta_group_info
:
2492 } /* ext4_mb_add_groupinfo */
2494 static int ext4_mb_init_backend(struct super_block
*sb
)
2496 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2498 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2500 struct ext4_group_desc
*desc
;
2501 struct kmem_cache
*cachep
;
2503 err
= ext4_mb_alloc_groupinfo(sb
, ngroups
);
2507 sbi
->s_buddy_cache
= new_inode(sb
);
2508 if (sbi
->s_buddy_cache
== NULL
) {
2509 ext4_msg(sb
, KERN_ERR
, "can't get new inode");
2512 /* To avoid potentially colliding with an valid on-disk inode number,
2513 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2514 * not in the inode hash, so it should never be found by iget(), but
2515 * this will avoid confusion if it ever shows up during debugging. */
2516 sbi
->s_buddy_cache
->i_ino
= EXT4_BAD_INO
;
2517 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2518 for (i
= 0; i
< ngroups
; i
++) {
2519 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2521 ext4_msg(sb
, KERN_ERR
, "can't read descriptor %u", i
);
2524 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2531 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2533 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2534 i
= sbi
->s_group_info_size
;
2536 kfree(sbi
->s_group_info
[i
]);
2537 iput(sbi
->s_buddy_cache
);
2539 kvfree(sbi
->s_group_info
);
2543 static void ext4_groupinfo_destroy_slabs(void)
2547 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2548 if (ext4_groupinfo_caches
[i
])
2549 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2550 ext4_groupinfo_caches
[i
] = NULL
;
2554 static int ext4_groupinfo_create_slab(size_t size
)
2556 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2558 int blocksize_bits
= order_base_2(size
);
2559 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2560 struct kmem_cache
*cachep
;
2562 if (cache_index
>= NR_GRPINFO_CACHES
)
2565 if (unlikely(cache_index
< 0))
2568 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2569 if (ext4_groupinfo_caches
[cache_index
]) {
2570 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2571 return 0; /* Already created */
2574 slab_size
= offsetof(struct ext4_group_info
,
2575 bb_counters
[blocksize_bits
+ 2]);
2577 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2578 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2581 ext4_groupinfo_caches
[cache_index
] = cachep
;
2583 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2586 "EXT4-fs: no memory for groupinfo slab cache\n");
2593 int ext4_mb_init(struct super_block
*sb
)
2595 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2597 unsigned offset
, offset_incr
;
2601 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2603 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2604 if (sbi
->s_mb_offsets
== NULL
) {
2609 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2610 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2611 if (sbi
->s_mb_maxs
== NULL
) {
2616 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2620 /* order 0 is regular bitmap */
2621 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2622 sbi
->s_mb_offsets
[0] = 0;
2626 offset_incr
= 1 << (sb
->s_blocksize_bits
- 1);
2627 max
= sb
->s_blocksize
<< 2;
2629 sbi
->s_mb_offsets
[i
] = offset
;
2630 sbi
->s_mb_maxs
[i
] = max
;
2631 offset
+= offset_incr
;
2632 offset_incr
= offset_incr
>> 1;
2635 } while (i
<= sb
->s_blocksize_bits
+ 1);
2637 spin_lock_init(&sbi
->s_md_lock
);
2638 spin_lock_init(&sbi
->s_bal_lock
);
2639 sbi
->s_mb_free_pending
= 0;
2640 INIT_LIST_HEAD(&sbi
->s_freed_data_list
);
2642 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2643 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2644 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2645 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2646 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2648 * The default group preallocation is 512, which for 4k block
2649 * sizes translates to 2 megabytes. However for bigalloc file
2650 * systems, this is probably too big (i.e, if the cluster size
2651 * is 1 megabyte, then group preallocation size becomes half a
2652 * gigabyte!). As a default, we will keep a two megabyte
2653 * group pralloc size for cluster sizes up to 64k, and after
2654 * that, we will force a minimum group preallocation size of
2655 * 32 clusters. This translates to 8 megs when the cluster
2656 * size is 256k, and 32 megs when the cluster size is 1 meg,
2657 * which seems reasonable as a default.
2659 sbi
->s_mb_group_prealloc
= max(MB_DEFAULT_GROUP_PREALLOC
>>
2660 sbi
->s_cluster_bits
, 32);
2662 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2663 * to the lowest multiple of s_stripe which is bigger than
2664 * the s_mb_group_prealloc as determined above. We want
2665 * the preallocation size to be an exact multiple of the
2666 * RAID stripe size so that preallocations don't fragment
2669 if (sbi
->s_stripe
> 1) {
2670 sbi
->s_mb_group_prealloc
= roundup(
2671 sbi
->s_mb_group_prealloc
, sbi
->s_stripe
);
2674 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2675 if (sbi
->s_locality_groups
== NULL
) {
2679 for_each_possible_cpu(i
) {
2680 struct ext4_locality_group
*lg
;
2681 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2682 mutex_init(&lg
->lg_mutex
);
2683 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2684 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2685 spin_lock_init(&lg
->lg_prealloc_lock
);
2688 /* init file for buddy data */
2689 ret
= ext4_mb_init_backend(sb
);
2691 goto out_free_locality_groups
;
2695 out_free_locality_groups
:
2696 free_percpu(sbi
->s_locality_groups
);
2697 sbi
->s_locality_groups
= NULL
;
2699 kfree(sbi
->s_mb_offsets
);
2700 sbi
->s_mb_offsets
= NULL
;
2701 kfree(sbi
->s_mb_maxs
);
2702 sbi
->s_mb_maxs
= NULL
;
2706 /* need to called with the ext4 group lock held */
2707 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2709 struct ext4_prealloc_space
*pa
;
2710 struct list_head
*cur
, *tmp
;
2713 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2714 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2715 list_del(&pa
->pa_group_list
);
2717 kmem_cache_free(ext4_pspace_cachep
, pa
);
2720 mb_debug(1, "mballoc: %u PAs left\n", count
);
2724 int ext4_mb_release(struct super_block
*sb
)
2726 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2728 int num_meta_group_infos
;
2729 struct ext4_group_info
*grinfo
;
2730 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2731 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2733 if (sbi
->s_group_info
) {
2734 for (i
= 0; i
< ngroups
; i
++) {
2735 grinfo
= ext4_get_group_info(sb
, i
);
2737 kfree(grinfo
->bb_bitmap
);
2739 ext4_lock_group(sb
, i
);
2740 ext4_mb_cleanup_pa(grinfo
);
2741 ext4_unlock_group(sb
, i
);
2742 kmem_cache_free(cachep
, grinfo
);
2744 num_meta_group_infos
= (ngroups
+
2745 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2746 EXT4_DESC_PER_BLOCK_BITS(sb
);
2747 for (i
= 0; i
< num_meta_group_infos
; i
++)
2748 kfree(sbi
->s_group_info
[i
]);
2749 kvfree(sbi
->s_group_info
);
2751 kfree(sbi
->s_mb_offsets
);
2752 kfree(sbi
->s_mb_maxs
);
2753 iput(sbi
->s_buddy_cache
);
2754 if (sbi
->s_mb_stats
) {
2755 ext4_msg(sb
, KERN_INFO
,
2756 "mballoc: %u blocks %u reqs (%u success)",
2757 atomic_read(&sbi
->s_bal_allocated
),
2758 atomic_read(&sbi
->s_bal_reqs
),
2759 atomic_read(&sbi
->s_bal_success
));
2760 ext4_msg(sb
, KERN_INFO
,
2761 "mballoc: %u extents scanned, %u goal hits, "
2762 "%u 2^N hits, %u breaks, %u lost",
2763 atomic_read(&sbi
->s_bal_ex_scanned
),
2764 atomic_read(&sbi
->s_bal_goals
),
2765 atomic_read(&sbi
->s_bal_2orders
),
2766 atomic_read(&sbi
->s_bal_breaks
),
2767 atomic_read(&sbi
->s_mb_lost_chunks
));
2768 ext4_msg(sb
, KERN_INFO
,
2769 "mballoc: %lu generated and it took %Lu",
2770 sbi
->s_mb_buddies_generated
,
2771 sbi
->s_mb_generation_time
);
2772 ext4_msg(sb
, KERN_INFO
,
2773 "mballoc: %u preallocated, %u discarded",
2774 atomic_read(&sbi
->s_mb_preallocated
),
2775 atomic_read(&sbi
->s_mb_discarded
));
2778 free_percpu(sbi
->s_locality_groups
);
2783 static inline int ext4_issue_discard(struct super_block
*sb
,
2784 ext4_group_t block_group
, ext4_grpblk_t cluster
, int count
,
2787 ext4_fsblk_t discard_block
;
2789 discard_block
= (EXT4_C2B(EXT4_SB(sb
), cluster
) +
2790 ext4_group_first_block_no(sb
, block_group
));
2791 count
= EXT4_C2B(EXT4_SB(sb
), count
);
2792 trace_ext4_discard_blocks(sb
,
2793 (unsigned long long) discard_block
, count
);
2795 return __blkdev_issue_discard(sb
->s_bdev
,
2796 (sector_t
)discard_block
<< (sb
->s_blocksize_bits
- 9),
2797 (sector_t
)count
<< (sb
->s_blocksize_bits
- 9),
2800 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, 0);
2803 static void ext4_free_data_in_buddy(struct super_block
*sb
,
2804 struct ext4_free_data
*entry
)
2806 struct ext4_buddy e4b
;
2807 struct ext4_group_info
*db
;
2808 int err
, count
= 0, count2
= 0;
2810 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2811 entry
->efd_count
, entry
->efd_group
, entry
);
2813 err
= ext4_mb_load_buddy(sb
, entry
->efd_group
, &e4b
);
2814 /* we expect to find existing buddy because it's pinned */
2817 spin_lock(&EXT4_SB(sb
)->s_md_lock
);
2818 EXT4_SB(sb
)->s_mb_free_pending
-= entry
->efd_count
;
2819 spin_unlock(&EXT4_SB(sb
)->s_md_lock
);
2822 /* there are blocks to put in buddy to make them really free */
2823 count
+= entry
->efd_count
;
2825 ext4_lock_group(sb
, entry
->efd_group
);
2826 /* Take it out of per group rb tree */
2827 rb_erase(&entry
->efd_node
, &(db
->bb_free_root
));
2828 mb_free_blocks(NULL
, &e4b
, entry
->efd_start_cluster
, entry
->efd_count
);
2831 * Clear the trimmed flag for the group so that the next
2832 * ext4_trim_fs can trim it.
2833 * If the volume is mounted with -o discard, online discard
2834 * is supported and the free blocks will be trimmed online.
2836 if (!test_opt(sb
, DISCARD
))
2837 EXT4_MB_GRP_CLEAR_TRIMMED(db
);
2839 if (!db
->bb_free_root
.rb_node
) {
2840 /* No more items in the per group rb tree
2841 * balance refcounts from ext4_mb_free_metadata()
2843 put_page(e4b
.bd_buddy_page
);
2844 put_page(e4b
.bd_bitmap_page
);
2846 ext4_unlock_group(sb
, entry
->efd_group
);
2847 kmem_cache_free(ext4_free_data_cachep
, entry
);
2848 ext4_mb_unload_buddy(&e4b
);
2850 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2854 * This function is called by the jbd2 layer once the commit has finished,
2855 * so we know we can free the blocks that were released with that commit.
2857 void ext4_process_freed_data(struct super_block
*sb
, tid_t commit_tid
)
2859 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2860 struct ext4_free_data
*entry
, *tmp
;
2861 struct bio
*discard_bio
= NULL
;
2862 struct list_head freed_data_list
;
2863 struct list_head
*cut_pos
= NULL
;
2866 INIT_LIST_HEAD(&freed_data_list
);
2868 spin_lock(&sbi
->s_md_lock
);
2869 list_for_each_entry(entry
, &sbi
->s_freed_data_list
, efd_list
) {
2870 if (entry
->efd_tid
!= commit_tid
)
2872 cut_pos
= &entry
->efd_list
;
2875 list_cut_position(&freed_data_list
, &sbi
->s_freed_data_list
,
2877 spin_unlock(&sbi
->s_md_lock
);
2879 if (test_opt(sb
, DISCARD
)) {
2880 list_for_each_entry(entry
, &freed_data_list
, efd_list
) {
2881 err
= ext4_issue_discard(sb
, entry
->efd_group
,
2882 entry
->efd_start_cluster
,
2885 if (err
&& err
!= -EOPNOTSUPP
) {
2886 ext4_msg(sb
, KERN_WARNING
, "discard request in"
2887 " group:%d block:%d count:%d failed"
2888 " with %d", entry
->efd_group
,
2889 entry
->efd_start_cluster
,
2890 entry
->efd_count
, err
);
2891 } else if (err
== -EOPNOTSUPP
)
2896 submit_bio_wait(discard_bio
);
2899 list_for_each_entry_safe(entry
, tmp
, &freed_data_list
, efd_list
)
2900 ext4_free_data_in_buddy(sb
, entry
);
2903 int __init
ext4_init_mballoc(void)
2905 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2906 SLAB_RECLAIM_ACCOUNT
);
2907 if (ext4_pspace_cachep
== NULL
)
2910 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2911 SLAB_RECLAIM_ACCOUNT
);
2912 if (ext4_ac_cachep
== NULL
) {
2913 kmem_cache_destroy(ext4_pspace_cachep
);
2917 ext4_free_data_cachep
= KMEM_CACHE(ext4_free_data
,
2918 SLAB_RECLAIM_ACCOUNT
);
2919 if (ext4_free_data_cachep
== NULL
) {
2920 kmem_cache_destroy(ext4_pspace_cachep
);
2921 kmem_cache_destroy(ext4_ac_cachep
);
2927 void ext4_exit_mballoc(void)
2930 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2931 * before destroying the slab cache.
2934 kmem_cache_destroy(ext4_pspace_cachep
);
2935 kmem_cache_destroy(ext4_ac_cachep
);
2936 kmem_cache_destroy(ext4_free_data_cachep
);
2937 ext4_groupinfo_destroy_slabs();
2942 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2943 * Returns 0 if success or error code
2945 static noinline_for_stack
int
2946 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2947 handle_t
*handle
, unsigned int reserv_clstrs
)
2949 struct buffer_head
*bitmap_bh
= NULL
;
2950 struct ext4_group_desc
*gdp
;
2951 struct buffer_head
*gdp_bh
;
2952 struct ext4_sb_info
*sbi
;
2953 struct super_block
*sb
;
2957 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2958 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2963 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2964 if (IS_ERR(bitmap_bh
)) {
2965 err
= PTR_ERR(bitmap_bh
);
2970 BUFFER_TRACE(bitmap_bh
, "getting write access");
2971 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2976 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2980 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2981 ext4_free_group_clusters(sb
, gdp
));
2983 BUFFER_TRACE(gdp_bh
, "get_write_access");
2984 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2988 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
2990 len
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
2991 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2992 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2993 "fs metadata", block
, block
+len
);
2994 /* File system mounted not to panic on error
2995 * Fix the bitmap and return EFSCORRUPTED
2996 * We leak some of the blocks here.
2998 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2999 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
3000 ac
->ac_b_ex
.fe_len
);
3001 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3002 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3004 err
= -EFSCORRUPTED
;
3008 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3009 #ifdef AGGRESSIVE_CHECK
3012 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
3013 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
3014 bitmap_bh
->b_data
));
3018 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
3019 ac
->ac_b_ex
.fe_len
);
3020 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
3021 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
3022 ext4_free_group_clusters_set(sb
, gdp
,
3023 ext4_free_clusters_after_init(sb
,
3024 ac
->ac_b_ex
.fe_group
, gdp
));
3026 len
= ext4_free_group_clusters(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
3027 ext4_free_group_clusters_set(sb
, gdp
, len
);
3028 ext4_block_bitmap_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
, bitmap_bh
);
3029 ext4_group_desc_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
);
3031 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3032 percpu_counter_sub(&sbi
->s_freeclusters_counter
, ac
->ac_b_ex
.fe_len
);
3034 * Now reduce the dirty block count also. Should not go negative
3036 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
3037 /* release all the reserved blocks if non delalloc */
3038 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
3041 if (sbi
->s_log_groups_per_flex
) {
3042 ext4_group_t flex_group
= ext4_flex_group(sbi
,
3043 ac
->ac_b_ex
.fe_group
);
3044 atomic64_sub(ac
->ac_b_ex
.fe_len
,
3045 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
3048 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3051 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
3059 * here we normalize request for locality group
3060 * Group request are normalized to s_mb_group_prealloc, which goes to
3061 * s_strip if we set the same via mount option.
3062 * s_mb_group_prealloc can be configured via
3063 * /sys/fs/ext4/<partition>/mb_group_prealloc
3065 * XXX: should we try to preallocate more than the group has now?
3067 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
3069 struct super_block
*sb
= ac
->ac_sb
;
3070 struct ext4_locality_group
*lg
= ac
->ac_lg
;
3073 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
3074 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3075 current
->pid
, ac
->ac_g_ex
.fe_len
);
3079 * Normalization means making request better in terms of
3080 * size and alignment
3082 static noinline_for_stack
void
3083 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
3084 struct ext4_allocation_request
*ar
)
3086 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3089 loff_t size
, start_off
;
3090 loff_t orig_size __maybe_unused
;
3092 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3093 struct ext4_prealloc_space
*pa
;
3095 /* do normalize only data requests, metadata requests
3096 do not need preallocation */
3097 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3100 /* sometime caller may want exact blocks */
3101 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3104 /* caller may indicate that preallocation isn't
3105 * required (it's a tail, for example) */
3106 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3109 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3110 ext4_mb_normalize_group_request(ac
);
3114 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3116 /* first, let's learn actual file size
3117 * given current request is allocated */
3118 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
3119 size
= size
<< bsbits
;
3120 if (size
< i_size_read(ac
->ac_inode
))
3121 size
= i_size_read(ac
->ac_inode
);
3124 /* max size of free chunks */
3127 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3128 (req <= (size) || max <= (chunk_size))
3130 /* first, try to predict filesize */
3131 /* XXX: should this table be tunable? */
3133 if (size
<= 16 * 1024) {
3135 } else if (size
<= 32 * 1024) {
3137 } else if (size
<= 64 * 1024) {
3139 } else if (size
<= 128 * 1024) {
3141 } else if (size
<= 256 * 1024) {
3143 } else if (size
<= 512 * 1024) {
3145 } else if (size
<= 1024 * 1024) {
3147 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3148 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3149 (21 - bsbits
)) << 21;
3150 size
= 2 * 1024 * 1024;
3151 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3152 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3153 (22 - bsbits
)) << 22;
3154 size
= 4 * 1024 * 1024;
3155 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3156 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3157 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3158 (23 - bsbits
)) << 23;
3159 size
= 8 * 1024 * 1024;
3161 start_off
= (loff_t
) ac
->ac_o_ex
.fe_logical
<< bsbits
;
3162 size
= (loff_t
) EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3163 ac
->ac_o_ex
.fe_len
) << bsbits
;
3165 size
= size
>> bsbits
;
3166 start
= start_off
>> bsbits
;
3168 /* don't cover already allocated blocks in selected range */
3169 if (ar
->pleft
&& start
<= ar
->lleft
) {
3170 size
-= ar
->lleft
+ 1 - start
;
3171 start
= ar
->lleft
+ 1;
3173 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3174 size
-= start
+ size
- ar
->lright
;
3177 * Trim allocation request for filesystems with artificially small
3180 if (size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
))
3181 size
= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
);
3185 /* check we don't cross already preallocated blocks */
3187 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3192 spin_lock(&pa
->pa_lock
);
3193 if (pa
->pa_deleted
) {
3194 spin_unlock(&pa
->pa_lock
);
3198 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3201 /* PA must not overlap original request */
3202 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3203 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3205 /* skip PAs this normalized request doesn't overlap with */
3206 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
3207 spin_unlock(&pa
->pa_lock
);
3210 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3212 /* adjust start or end to be adjacent to this pa */
3213 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3214 BUG_ON(pa_end
< start
);
3216 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3217 BUG_ON(pa
->pa_lstart
> end
);
3218 end
= pa
->pa_lstart
;
3220 spin_unlock(&pa
->pa_lock
);
3225 /* XXX: extra loop to check we really don't overlap preallocations */
3227 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3230 spin_lock(&pa
->pa_lock
);
3231 if (pa
->pa_deleted
== 0) {
3232 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3234 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3236 spin_unlock(&pa
->pa_lock
);
3240 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3241 start
> ac
->ac_o_ex
.fe_logical
) {
3242 ext4_msg(ac
->ac_sb
, KERN_ERR
,
3243 "start %lu, size %lu, fe_logical %lu",
3244 (unsigned long) start
, (unsigned long) size
,
3245 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3248 BUG_ON(size
<= 0 || size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3250 /* now prepare goal request */
3252 /* XXX: is it better to align blocks WRT to logical
3253 * placement or satisfy big request as is */
3254 ac
->ac_g_ex
.fe_logical
= start
;
3255 ac
->ac_g_ex
.fe_len
= EXT4_NUM_B2C(sbi
, size
);
3257 /* define goal start in order to merge */
3258 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3259 /* merge to the right */
3260 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3261 &ac
->ac_f_ex
.fe_group
,
3262 &ac
->ac_f_ex
.fe_start
);
3263 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3265 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3266 /* merge to the left */
3267 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3268 &ac
->ac_f_ex
.fe_group
,
3269 &ac
->ac_f_ex
.fe_start
);
3270 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3273 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3274 (unsigned) orig_size
, (unsigned) start
);
3277 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3279 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3281 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3282 atomic_inc(&sbi
->s_bal_reqs
);
3283 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3284 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3285 atomic_inc(&sbi
->s_bal_success
);
3286 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3287 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3288 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3289 atomic_inc(&sbi
->s_bal_goals
);
3290 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3291 atomic_inc(&sbi
->s_bal_breaks
);
3294 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3295 trace_ext4_mballoc_alloc(ac
);
3297 trace_ext4_mballoc_prealloc(ac
);
3301 * Called on failure; free up any blocks from the inode PA for this
3302 * context. We don't need this for MB_GROUP_PA because we only change
3303 * pa_free in ext4_mb_release_context(), but on failure, we've already
3304 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3306 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3308 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3309 struct ext4_buddy e4b
;
3313 if (ac
->ac_f_ex
.fe_len
== 0)
3315 err
= ext4_mb_load_buddy(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
, &e4b
);
3318 * This should never happen since we pin the
3319 * pages in the ext4_allocation_context so
3320 * ext4_mb_load_buddy() should never fail.
3322 WARN(1, "mb_load_buddy failed (%d)", err
);
3325 ext4_lock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3326 mb_free_blocks(ac
->ac_inode
, &e4b
, ac
->ac_f_ex
.fe_start
,
3327 ac
->ac_f_ex
.fe_len
);
3328 ext4_unlock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3329 ext4_mb_unload_buddy(&e4b
);
3332 if (pa
->pa_type
== MB_INODE_PA
)
3333 pa
->pa_free
+= ac
->ac_b_ex
.fe_len
;
3337 * use blocks preallocated to inode
3339 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3340 struct ext4_prealloc_space
*pa
)
3342 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3347 /* found preallocated blocks, use them */
3348 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3349 end
= min(pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
),
3350 start
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
));
3351 len
= EXT4_NUM_B2C(sbi
, end
- start
);
3352 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3353 &ac
->ac_b_ex
.fe_start
);
3354 ac
->ac_b_ex
.fe_len
= len
;
3355 ac
->ac_status
= AC_STATUS_FOUND
;
3358 BUG_ON(start
< pa
->pa_pstart
);
3359 BUG_ON(end
> pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
));
3360 BUG_ON(pa
->pa_free
< len
);
3363 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3367 * use blocks preallocated to locality group
3369 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3370 struct ext4_prealloc_space
*pa
)
3372 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3374 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3375 &ac
->ac_b_ex
.fe_group
,
3376 &ac
->ac_b_ex
.fe_start
);
3377 ac
->ac_b_ex
.fe_len
= len
;
3378 ac
->ac_status
= AC_STATUS_FOUND
;
3381 /* we don't correct pa_pstart or pa_plen here to avoid
3382 * possible race when the group is being loaded concurrently
3383 * instead we correct pa later, after blocks are marked
3384 * in on-disk bitmap -- see ext4_mb_release_context()
3385 * Other CPUs are prevented from allocating from this pa by lg_mutex
3387 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3391 * Return the prealloc space that have minimal distance
3392 * from the goal block. @cpa is the prealloc
3393 * space that is having currently known minimal distance
3394 * from the goal block.
3396 static struct ext4_prealloc_space
*
3397 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3398 struct ext4_prealloc_space
*pa
,
3399 struct ext4_prealloc_space
*cpa
)
3401 ext4_fsblk_t cur_distance
, new_distance
;
3404 atomic_inc(&pa
->pa_count
);
3407 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3408 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3410 if (cur_distance
<= new_distance
)
3413 /* drop the previous reference */
3414 atomic_dec(&cpa
->pa_count
);
3415 atomic_inc(&pa
->pa_count
);
3420 * search goal blocks in preallocated space
3422 static noinline_for_stack
int
3423 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3425 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3427 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3428 struct ext4_locality_group
*lg
;
3429 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3430 ext4_fsblk_t goal_block
;
3432 /* only data can be preallocated */
3433 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3436 /* first, try per-file preallocation */
3438 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3440 /* all fields in this condition don't change,
3441 * so we can skip locking for them */
3442 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3443 ac
->ac_o_ex
.fe_logical
>= (pa
->pa_lstart
+
3444 EXT4_C2B(sbi
, pa
->pa_len
)))
3447 /* non-extent files can't have physical blocks past 2^32 */
3448 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3449 (pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
) >
3450 EXT4_MAX_BLOCK_FILE_PHYS
))
3453 /* found preallocated blocks, use them */
3454 spin_lock(&pa
->pa_lock
);
3455 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3456 atomic_inc(&pa
->pa_count
);
3457 ext4_mb_use_inode_pa(ac
, pa
);
3458 spin_unlock(&pa
->pa_lock
);
3459 ac
->ac_criteria
= 10;
3463 spin_unlock(&pa
->pa_lock
);
3467 /* can we use group allocation? */
3468 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3471 /* inode may have no locality group for some reason */
3475 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3476 if (order
> PREALLOC_TB_SIZE
- 1)
3477 /* The max size of hash table is PREALLOC_TB_SIZE */
3478 order
= PREALLOC_TB_SIZE
- 1;
3480 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3482 * search for the prealloc space that is having
3483 * minimal distance from the goal block.
3485 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3487 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3489 spin_lock(&pa
->pa_lock
);
3490 if (pa
->pa_deleted
== 0 &&
3491 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3493 cpa
= ext4_mb_check_group_pa(goal_block
,
3496 spin_unlock(&pa
->pa_lock
);
3501 ext4_mb_use_group_pa(ac
, cpa
);
3502 ac
->ac_criteria
= 20;
3509 * the function goes through all block freed in the group
3510 * but not yet committed and marks them used in in-core bitmap.
3511 * buddy must be generated from this bitmap
3512 * Need to be called with the ext4 group lock held
3514 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3518 struct ext4_group_info
*grp
;
3519 struct ext4_free_data
*entry
;
3521 grp
= ext4_get_group_info(sb
, group
);
3522 n
= rb_first(&(grp
->bb_free_root
));
3525 entry
= rb_entry(n
, struct ext4_free_data
, efd_node
);
3526 ext4_set_bits(bitmap
, entry
->efd_start_cluster
, entry
->efd_count
);
3533 * the function goes through all preallocation in this group and marks them
3534 * used in in-core bitmap. buddy must be generated from this bitmap
3535 * Need to be called with ext4 group lock held
3537 static noinline_for_stack
3538 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3541 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3542 struct ext4_prealloc_space
*pa
;
3543 struct list_head
*cur
;
3544 ext4_group_t groupnr
;
3545 ext4_grpblk_t start
;
3546 int preallocated
= 0;
3549 /* all form of preallocation discards first load group,
3550 * so the only competing code is preallocation use.
3551 * we don't need any locking here
3552 * notice we do NOT ignore preallocations with pa_deleted
3553 * otherwise we could leave used blocks available for
3554 * allocation in buddy when concurrent ext4_mb_put_pa()
3555 * is dropping preallocation
3557 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3558 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3559 spin_lock(&pa
->pa_lock
);
3560 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3563 spin_unlock(&pa
->pa_lock
);
3564 if (unlikely(len
== 0))
3566 BUG_ON(groupnr
!= group
);
3567 ext4_set_bits(bitmap
, start
, len
);
3568 preallocated
+= len
;
3570 mb_debug(1, "preallocated %u for group %u\n", preallocated
, group
);
3573 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3575 struct ext4_prealloc_space
*pa
;
3576 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3578 BUG_ON(atomic_read(&pa
->pa_count
));
3579 BUG_ON(pa
->pa_deleted
== 0);
3580 kmem_cache_free(ext4_pspace_cachep
, pa
);
3584 * drops a reference to preallocated space descriptor
3585 * if this was the last reference and the space is consumed
3587 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3588 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3591 ext4_fsblk_t grp_blk
;
3593 /* in this short window concurrent discard can set pa_deleted */
3594 spin_lock(&pa
->pa_lock
);
3595 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0) {
3596 spin_unlock(&pa
->pa_lock
);
3600 if (pa
->pa_deleted
== 1) {
3601 spin_unlock(&pa
->pa_lock
);
3606 spin_unlock(&pa
->pa_lock
);
3608 grp_blk
= pa
->pa_pstart
;
3610 * If doing group-based preallocation, pa_pstart may be in the
3611 * next group when pa is used up
3613 if (pa
->pa_type
== MB_GROUP_PA
)
3616 grp
= ext4_get_group_number(sb
, grp_blk
);
3621 * P1 (buddy init) P2 (regular allocation)
3622 * find block B in PA
3623 * copy on-disk bitmap to buddy
3624 * mark B in on-disk bitmap
3625 * drop PA from group
3626 * mark all PAs in buddy
3628 * thus, P1 initializes buddy with B available. to prevent this
3629 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3632 ext4_lock_group(sb
, grp
);
3633 list_del(&pa
->pa_group_list
);
3634 ext4_unlock_group(sb
, grp
);
3636 spin_lock(pa
->pa_obj_lock
);
3637 list_del_rcu(&pa
->pa_inode_list
);
3638 spin_unlock(pa
->pa_obj_lock
);
3640 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3644 * creates new preallocated space for given inode
3646 static noinline_for_stack
int
3647 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3649 struct super_block
*sb
= ac
->ac_sb
;
3650 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3651 struct ext4_prealloc_space
*pa
;
3652 struct ext4_group_info
*grp
;
3653 struct ext4_inode_info
*ei
;
3655 /* preallocate only when found space is larger then requested */
3656 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3657 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3658 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3660 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3664 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3670 /* we can't allocate as much as normalizer wants.
3671 * so, found space must get proper lstart
3672 * to cover original request */
3673 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3674 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3676 /* we're limited by original request in that
3677 * logical block must be covered any way
3678 * winl is window we can move our chunk within */
3679 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3681 /* also, we should cover whole original request */
3682 wins
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
);
3684 /* the smallest one defines real window */
3685 win
= min(winl
, wins
);
3687 offs
= ac
->ac_o_ex
.fe_logical
%
3688 EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
3689 if (offs
&& offs
< win
)
3692 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
-
3693 EXT4_NUM_B2C(sbi
, win
);
3694 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3695 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3698 /* preallocation can change ac_b_ex, thus we store actually
3699 * allocated blocks for history */
3700 ac
->ac_f_ex
= ac
->ac_b_ex
;
3702 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3703 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3704 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3705 pa
->pa_free
= pa
->pa_len
;
3706 atomic_set(&pa
->pa_count
, 1);
3707 spin_lock_init(&pa
->pa_lock
);
3708 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3709 INIT_LIST_HEAD(&pa
->pa_group_list
);
3711 pa
->pa_type
= MB_INODE_PA
;
3713 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3714 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3715 trace_ext4_mb_new_inode_pa(ac
, pa
);
3717 ext4_mb_use_inode_pa(ac
, pa
);
3718 atomic_add(pa
->pa_free
, &sbi
->s_mb_preallocated
);
3720 ei
= EXT4_I(ac
->ac_inode
);
3721 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3723 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3724 pa
->pa_inode
= ac
->ac_inode
;
3726 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3727 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3728 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3730 spin_lock(pa
->pa_obj_lock
);
3731 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3732 spin_unlock(pa
->pa_obj_lock
);
3738 * creates new preallocated space for locality group inodes belongs to
3740 static noinline_for_stack
int
3741 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3743 struct super_block
*sb
= ac
->ac_sb
;
3744 struct ext4_locality_group
*lg
;
3745 struct ext4_prealloc_space
*pa
;
3746 struct ext4_group_info
*grp
;
3748 /* preallocate only when found space is larger then requested */
3749 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3750 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3751 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3753 BUG_ON(ext4_pspace_cachep
== NULL
);
3754 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3758 /* preallocation can change ac_b_ex, thus we store actually
3759 * allocated blocks for history */
3760 ac
->ac_f_ex
= ac
->ac_b_ex
;
3762 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3763 pa
->pa_lstart
= pa
->pa_pstart
;
3764 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3765 pa
->pa_free
= pa
->pa_len
;
3766 atomic_set(&pa
->pa_count
, 1);
3767 spin_lock_init(&pa
->pa_lock
);
3768 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3769 INIT_LIST_HEAD(&pa
->pa_group_list
);
3771 pa
->pa_type
= MB_GROUP_PA
;
3773 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3774 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3775 trace_ext4_mb_new_group_pa(ac
, pa
);
3777 ext4_mb_use_group_pa(ac
, pa
);
3778 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3780 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3784 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3785 pa
->pa_inode
= NULL
;
3787 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3788 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3789 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3792 * We will later add the new pa to the right bucket
3793 * after updating the pa_free in ext4_mb_release_context
3798 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3802 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3803 err
= ext4_mb_new_group_pa(ac
);
3805 err
= ext4_mb_new_inode_pa(ac
);
3810 * finds all unused blocks in on-disk bitmap, frees them in
3811 * in-core bitmap and buddy.
3812 * @pa must be unlinked from inode and group lists, so that
3813 * nobody else can find/use it.
3814 * the caller MUST hold group/inode locks.
3815 * TODO: optimize the case when there are no in-core structures yet
3817 static noinline_for_stack
int
3818 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3819 struct ext4_prealloc_space
*pa
)
3821 struct super_block
*sb
= e4b
->bd_sb
;
3822 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3827 unsigned long long grp_blk_start
;
3831 BUG_ON(pa
->pa_deleted
== 0);
3832 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3833 grp_blk_start
= pa
->pa_pstart
- EXT4_C2B(sbi
, bit
);
3834 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3835 end
= bit
+ pa
->pa_len
;
3838 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3841 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3842 mb_debug(1, " free preallocated %u/%u in group %u\n",
3843 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3844 (unsigned) next
- bit
, (unsigned) group
);
3847 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3848 trace_ext4_mb_release_inode_pa(pa
, (grp_blk_start
+
3849 EXT4_C2B(sbi
, bit
)),
3851 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3854 if (free
!= pa
->pa_free
) {
3855 ext4_msg(e4b
->bd_sb
, KERN_CRIT
,
3856 "pa %p: logic %lu, phys. %lu, len %lu",
3857 pa
, (unsigned long) pa
->pa_lstart
,
3858 (unsigned long) pa
->pa_pstart
,
3859 (unsigned long) pa
->pa_len
);
3860 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3863 * pa is already deleted so we use the value obtained
3864 * from the bitmap and continue.
3867 atomic_add(free
, &sbi
->s_mb_discarded
);
3872 static noinline_for_stack
int
3873 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3874 struct ext4_prealloc_space
*pa
)
3876 struct super_block
*sb
= e4b
->bd_sb
;
3880 trace_ext4_mb_release_group_pa(sb
, pa
);
3881 BUG_ON(pa
->pa_deleted
== 0);
3882 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3883 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3884 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3885 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3886 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3892 * releases all preallocations in given group
3894 * first, we need to decide discard policy:
3895 * - when do we discard
3897 * - how many do we discard
3898 * 1) how many requested
3900 static noinline_for_stack
int
3901 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3902 ext4_group_t group
, int needed
)
3904 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3905 struct buffer_head
*bitmap_bh
= NULL
;
3906 struct ext4_prealloc_space
*pa
, *tmp
;
3907 struct list_head list
;
3908 struct ext4_buddy e4b
;
3913 mb_debug(1, "discard preallocation for group %u\n", group
);
3915 if (list_empty(&grp
->bb_prealloc_list
))
3918 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3919 if (IS_ERR(bitmap_bh
)) {
3920 err
= PTR_ERR(bitmap_bh
);
3921 ext4_error(sb
, "Error %d reading block bitmap for %u",
3926 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3928 ext4_warning(sb
, "Error %d loading buddy information for %u",
3935 needed
= EXT4_CLUSTERS_PER_GROUP(sb
) + 1;
3937 INIT_LIST_HEAD(&list
);
3939 ext4_lock_group(sb
, group
);
3940 list_for_each_entry_safe(pa
, tmp
,
3941 &grp
->bb_prealloc_list
, pa_group_list
) {
3942 spin_lock(&pa
->pa_lock
);
3943 if (atomic_read(&pa
->pa_count
)) {
3944 spin_unlock(&pa
->pa_lock
);
3948 if (pa
->pa_deleted
) {
3949 spin_unlock(&pa
->pa_lock
);
3953 /* seems this one can be freed ... */
3956 /* we can trust pa_free ... */
3957 free
+= pa
->pa_free
;
3959 spin_unlock(&pa
->pa_lock
);
3961 list_del(&pa
->pa_group_list
);
3962 list_add(&pa
->u
.pa_tmp_list
, &list
);
3965 /* if we still need more blocks and some PAs were used, try again */
3966 if (free
< needed
&& busy
) {
3968 ext4_unlock_group(sb
, group
);
3973 /* found anything to free? */
3974 if (list_empty(&list
)) {
3979 /* now free all selected PAs */
3980 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3982 /* remove from object (inode or locality group) */
3983 spin_lock(pa
->pa_obj_lock
);
3984 list_del_rcu(&pa
->pa_inode_list
);
3985 spin_unlock(pa
->pa_obj_lock
);
3987 if (pa
->pa_type
== MB_GROUP_PA
)
3988 ext4_mb_release_group_pa(&e4b
, pa
);
3990 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3992 list_del(&pa
->u
.pa_tmp_list
);
3993 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3997 ext4_unlock_group(sb
, group
);
3998 ext4_mb_unload_buddy(&e4b
);
4004 * releases all non-used preallocated blocks for given inode
4006 * It's important to discard preallocations under i_data_sem
4007 * We don't want another block to be served from the prealloc
4008 * space when we are discarding the inode prealloc space.
4010 * FIXME!! Make sure it is valid at all the call sites
4012 void ext4_discard_preallocations(struct inode
*inode
)
4014 struct ext4_inode_info
*ei
= EXT4_I(inode
);
4015 struct super_block
*sb
= inode
->i_sb
;
4016 struct buffer_head
*bitmap_bh
= NULL
;
4017 struct ext4_prealloc_space
*pa
, *tmp
;
4018 ext4_group_t group
= 0;
4019 struct list_head list
;
4020 struct ext4_buddy e4b
;
4023 if (!S_ISREG(inode
->i_mode
)) {
4024 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4028 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
4029 trace_ext4_discard_preallocations(inode
);
4031 INIT_LIST_HEAD(&list
);
4034 /* first, collect all pa's in the inode */
4035 spin_lock(&ei
->i_prealloc_lock
);
4036 while (!list_empty(&ei
->i_prealloc_list
)) {
4037 pa
= list_entry(ei
->i_prealloc_list
.next
,
4038 struct ext4_prealloc_space
, pa_inode_list
);
4039 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
4040 spin_lock(&pa
->pa_lock
);
4041 if (atomic_read(&pa
->pa_count
)) {
4042 /* this shouldn't happen often - nobody should
4043 * use preallocation while we're discarding it */
4044 spin_unlock(&pa
->pa_lock
);
4045 spin_unlock(&ei
->i_prealloc_lock
);
4046 ext4_msg(sb
, KERN_ERR
,
4047 "uh-oh! used pa while discarding");
4049 schedule_timeout_uninterruptible(HZ
);
4053 if (pa
->pa_deleted
== 0) {
4055 spin_unlock(&pa
->pa_lock
);
4056 list_del_rcu(&pa
->pa_inode_list
);
4057 list_add(&pa
->u
.pa_tmp_list
, &list
);
4061 /* someone is deleting pa right now */
4062 spin_unlock(&pa
->pa_lock
);
4063 spin_unlock(&ei
->i_prealloc_lock
);
4065 /* we have to wait here because pa_deleted
4066 * doesn't mean pa is already unlinked from
4067 * the list. as we might be called from
4068 * ->clear_inode() the inode will get freed
4069 * and concurrent thread which is unlinking
4070 * pa from inode's list may access already
4071 * freed memory, bad-bad-bad */
4073 /* XXX: if this happens too often, we can
4074 * add a flag to force wait only in case
4075 * of ->clear_inode(), but not in case of
4076 * regular truncate */
4077 schedule_timeout_uninterruptible(HZ
);
4080 spin_unlock(&ei
->i_prealloc_lock
);
4082 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
4083 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
4084 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4086 err
= ext4_mb_load_buddy_gfp(sb
, group
, &e4b
,
4087 GFP_NOFS
|__GFP_NOFAIL
);
4089 ext4_error(sb
, "Error %d loading buddy information for %u",
4094 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
4095 if (IS_ERR(bitmap_bh
)) {
4096 err
= PTR_ERR(bitmap_bh
);
4097 ext4_error(sb
, "Error %d reading block bitmap for %u",
4099 ext4_mb_unload_buddy(&e4b
);
4103 ext4_lock_group(sb
, group
);
4104 list_del(&pa
->pa_group_list
);
4105 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
4106 ext4_unlock_group(sb
, group
);
4108 ext4_mb_unload_buddy(&e4b
);
4111 list_del(&pa
->u
.pa_tmp_list
);
4112 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4116 #ifdef CONFIG_EXT4_DEBUG
4117 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4119 struct super_block
*sb
= ac
->ac_sb
;
4120 ext4_group_t ngroups
, i
;
4122 if (!ext4_mballoc_debug
||
4123 (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
))
4126 ext4_msg(ac
->ac_sb
, KERN_ERR
, "Can't allocate:"
4127 " Allocation context details:");
4128 ext4_msg(ac
->ac_sb
, KERN_ERR
, "status %d flags %d",
4129 ac
->ac_status
, ac
->ac_flags
);
4130 ext4_msg(ac
->ac_sb
, KERN_ERR
, "orig %lu/%lu/%lu@%lu, "
4131 "goal %lu/%lu/%lu@%lu, "
4132 "best %lu/%lu/%lu@%lu cr %d",
4133 (unsigned long)ac
->ac_o_ex
.fe_group
,
4134 (unsigned long)ac
->ac_o_ex
.fe_start
,
4135 (unsigned long)ac
->ac_o_ex
.fe_len
,
4136 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4137 (unsigned long)ac
->ac_g_ex
.fe_group
,
4138 (unsigned long)ac
->ac_g_ex
.fe_start
,
4139 (unsigned long)ac
->ac_g_ex
.fe_len
,
4140 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4141 (unsigned long)ac
->ac_b_ex
.fe_group
,
4142 (unsigned long)ac
->ac_b_ex
.fe_start
,
4143 (unsigned long)ac
->ac_b_ex
.fe_len
,
4144 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4145 (int)ac
->ac_criteria
);
4146 ext4_msg(ac
->ac_sb
, KERN_ERR
, "%d found", ac
->ac_found
);
4147 ext4_msg(ac
->ac_sb
, KERN_ERR
, "groups: ");
4148 ngroups
= ext4_get_groups_count(sb
);
4149 for (i
= 0; i
< ngroups
; i
++) {
4150 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4151 struct ext4_prealloc_space
*pa
;
4152 ext4_grpblk_t start
;
4153 struct list_head
*cur
;
4154 ext4_lock_group(sb
, i
);
4155 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4156 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4158 spin_lock(&pa
->pa_lock
);
4159 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4161 spin_unlock(&pa
->pa_lock
);
4162 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
4165 ext4_unlock_group(sb
, i
);
4167 if (grp
->bb_free
== 0)
4169 printk(KERN_ERR
"%u: %d/%d \n",
4170 i
, grp
->bb_free
, grp
->bb_fragments
);
4172 printk(KERN_ERR
"\n");
4175 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4182 * We use locality group preallocation for small size file. The size of the
4183 * file is determined by the current size or the resulting size after
4184 * allocation which ever is larger
4186 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4188 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4190 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4191 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4194 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4197 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4200 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
4201 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
4204 if ((size
== isize
) &&
4205 !ext4_fs_is_busy(sbi
) &&
4206 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
4207 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
4211 if (sbi
->s_mb_group_prealloc
<= 0) {
4212 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4216 /* don't use group allocation for large files */
4217 size
= max(size
, isize
);
4218 if (size
> sbi
->s_mb_stream_request
) {
4219 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4223 BUG_ON(ac
->ac_lg
!= NULL
);
4225 * locality group prealloc space are per cpu. The reason for having
4226 * per cpu locality group is to reduce the contention between block
4227 * request from multiple CPUs.
4229 ac
->ac_lg
= raw_cpu_ptr(sbi
->s_locality_groups
);
4231 /* we're going to use group allocation */
4232 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4234 /* serialize all allocations in the group */
4235 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4238 static noinline_for_stack
int
4239 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4240 struct ext4_allocation_request
*ar
)
4242 struct super_block
*sb
= ar
->inode
->i_sb
;
4243 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4244 struct ext4_super_block
*es
= sbi
->s_es
;
4248 ext4_grpblk_t block
;
4250 /* we can't allocate > group size */
4253 /* just a dirty hack to filter too big requests */
4254 if (len
>= EXT4_CLUSTERS_PER_GROUP(sb
))
4255 len
= EXT4_CLUSTERS_PER_GROUP(sb
);
4257 /* start searching from the goal */
4259 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4260 goal
>= ext4_blocks_count(es
))
4261 goal
= le32_to_cpu(es
->s_first_data_block
);
4262 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4264 /* set up allocation goals */
4265 ac
->ac_b_ex
.fe_logical
= EXT4_LBLK_CMASK(sbi
, ar
->logical
);
4266 ac
->ac_status
= AC_STATUS_CONTINUE
;
4268 ac
->ac_inode
= ar
->inode
;
4269 ac
->ac_o_ex
.fe_logical
= ac
->ac_b_ex
.fe_logical
;
4270 ac
->ac_o_ex
.fe_group
= group
;
4271 ac
->ac_o_ex
.fe_start
= block
;
4272 ac
->ac_o_ex
.fe_len
= len
;
4273 ac
->ac_g_ex
= ac
->ac_o_ex
;
4274 ac
->ac_flags
= ar
->flags
;
4276 /* we have to define context: we'll we work with a file or
4277 * locality group. this is a policy, actually */
4278 ext4_mb_group_or_file(ac
);
4280 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4281 "left: %u/%u, right %u/%u to %swritable\n",
4282 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4283 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4284 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4285 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4286 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4291 static noinline_for_stack
void
4292 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4293 struct ext4_locality_group
*lg
,
4294 int order
, int total_entries
)
4296 ext4_group_t group
= 0;
4297 struct ext4_buddy e4b
;
4298 struct list_head discard_list
;
4299 struct ext4_prealloc_space
*pa
, *tmp
;
4301 mb_debug(1, "discard locality group preallocation\n");
4303 INIT_LIST_HEAD(&discard_list
);
4305 spin_lock(&lg
->lg_prealloc_lock
);
4306 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4308 spin_lock(&pa
->pa_lock
);
4309 if (atomic_read(&pa
->pa_count
)) {
4311 * This is the pa that we just used
4312 * for block allocation. So don't
4315 spin_unlock(&pa
->pa_lock
);
4318 if (pa
->pa_deleted
) {
4319 spin_unlock(&pa
->pa_lock
);
4322 /* only lg prealloc space */
4323 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4325 /* seems this one can be freed ... */
4327 spin_unlock(&pa
->pa_lock
);
4329 list_del_rcu(&pa
->pa_inode_list
);
4330 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4333 if (total_entries
<= 5) {
4335 * we want to keep only 5 entries
4336 * allowing it to grow to 8. This
4337 * mak sure we don't call discard
4338 * soon for this list.
4343 spin_unlock(&lg
->lg_prealloc_lock
);
4345 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4348 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4349 err
= ext4_mb_load_buddy_gfp(sb
, group
, &e4b
,
4350 GFP_NOFS
|__GFP_NOFAIL
);
4352 ext4_error(sb
, "Error %d loading buddy information for %u",
4356 ext4_lock_group(sb
, group
);
4357 list_del(&pa
->pa_group_list
);
4358 ext4_mb_release_group_pa(&e4b
, pa
);
4359 ext4_unlock_group(sb
, group
);
4361 ext4_mb_unload_buddy(&e4b
);
4362 list_del(&pa
->u
.pa_tmp_list
);
4363 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4368 * We have incremented pa_count. So it cannot be freed at this
4369 * point. Also we hold lg_mutex. So no parallel allocation is
4370 * possible from this lg. That means pa_free cannot be updated.
4372 * A parallel ext4_mb_discard_group_preallocations is possible.
4373 * which can cause the lg_prealloc_list to be updated.
4376 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4378 int order
, added
= 0, lg_prealloc_count
= 1;
4379 struct super_block
*sb
= ac
->ac_sb
;
4380 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4381 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4383 order
= fls(pa
->pa_free
) - 1;
4384 if (order
> PREALLOC_TB_SIZE
- 1)
4385 /* The max size of hash table is PREALLOC_TB_SIZE */
4386 order
= PREALLOC_TB_SIZE
- 1;
4387 /* Add the prealloc space to lg */
4388 spin_lock(&lg
->lg_prealloc_lock
);
4389 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4391 spin_lock(&tmp_pa
->pa_lock
);
4392 if (tmp_pa
->pa_deleted
) {
4393 spin_unlock(&tmp_pa
->pa_lock
);
4396 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4397 /* Add to the tail of the previous entry */
4398 list_add_tail_rcu(&pa
->pa_inode_list
,
4399 &tmp_pa
->pa_inode_list
);
4402 * we want to count the total
4403 * number of entries in the list
4406 spin_unlock(&tmp_pa
->pa_lock
);
4407 lg_prealloc_count
++;
4410 list_add_tail_rcu(&pa
->pa_inode_list
,
4411 &lg
->lg_prealloc_list
[order
]);
4412 spin_unlock(&lg
->lg_prealloc_lock
);
4414 /* Now trim the list to be not more than 8 elements */
4415 if (lg_prealloc_count
> 8) {
4416 ext4_mb_discard_lg_preallocations(sb
, lg
,
4417 order
, lg_prealloc_count
);
4424 * release all resource we used in allocation
4426 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4428 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4429 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4431 if (pa
->pa_type
== MB_GROUP_PA
) {
4432 /* see comment in ext4_mb_use_group_pa() */
4433 spin_lock(&pa
->pa_lock
);
4434 pa
->pa_pstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4435 pa
->pa_lstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4436 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4437 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4438 spin_unlock(&pa
->pa_lock
);
4443 * We want to add the pa to the right bucket.
4444 * Remove it from the list and while adding
4445 * make sure the list to which we are adding
4448 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4449 spin_lock(pa
->pa_obj_lock
);
4450 list_del_rcu(&pa
->pa_inode_list
);
4451 spin_unlock(pa
->pa_obj_lock
);
4452 ext4_mb_add_n_trim(ac
);
4454 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4456 if (ac
->ac_bitmap_page
)
4457 put_page(ac
->ac_bitmap_page
);
4458 if (ac
->ac_buddy_page
)
4459 put_page(ac
->ac_buddy_page
);
4460 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4461 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4462 ext4_mb_collect_stats(ac
);
4466 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4468 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4472 trace_ext4_mb_discard_preallocations(sb
, needed
);
4473 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4474 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4483 * Main entry point into mballoc to allocate blocks
4484 * it tries to use preallocation first, then falls back
4485 * to usual allocation
4487 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4488 struct ext4_allocation_request
*ar
, int *errp
)
4491 struct ext4_allocation_context
*ac
= NULL
;
4492 struct ext4_sb_info
*sbi
;
4493 struct super_block
*sb
;
4494 ext4_fsblk_t block
= 0;
4495 unsigned int inquota
= 0;
4496 unsigned int reserv_clstrs
= 0;
4499 sb
= ar
->inode
->i_sb
;
4502 trace_ext4_request_blocks(ar
);
4504 /* Allow to use superuser reservation for quota file */
4505 if (ext4_is_quota_file(ar
->inode
))
4506 ar
->flags
|= EXT4_MB_USE_ROOT_BLOCKS
;
4508 if ((ar
->flags
& EXT4_MB_DELALLOC_RESERVED
) == 0) {
4509 /* Without delayed allocation we need to verify
4510 * there is enough free blocks to do block allocation
4511 * and verify allocation doesn't exceed the quota limits.
4514 ext4_claim_free_clusters(sbi
, ar
->len
, ar
->flags
)) {
4516 /* let others to free the space */
4518 ar
->len
= ar
->len
>> 1;
4524 reserv_clstrs
= ar
->len
;
4525 if (ar
->flags
& EXT4_MB_USE_ROOT_BLOCKS
) {
4526 dquot_alloc_block_nofail(ar
->inode
,
4527 EXT4_C2B(sbi
, ar
->len
));
4530 dquot_alloc_block(ar
->inode
,
4531 EXT4_C2B(sbi
, ar
->len
))) {
4533 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4544 ac
= kmem_cache_zalloc(ext4_ac_cachep
, GFP_NOFS
);
4551 *errp
= ext4_mb_initialize_context(ac
, ar
);
4557 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4558 if (!ext4_mb_use_preallocated(ac
)) {
4559 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4560 ext4_mb_normalize_request(ac
, ar
);
4562 /* allocate space in core */
4563 *errp
= ext4_mb_regular_allocator(ac
);
4565 goto discard_and_exit
;
4567 /* as we've just preallocated more space than
4568 * user requested originally, we store allocated
4569 * space in a special descriptor */
4570 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4571 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4572 *errp
= ext4_mb_new_preallocation(ac
);
4575 ext4_discard_allocated_blocks(ac
);
4579 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4580 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_clstrs
);
4582 ext4_discard_allocated_blocks(ac
);
4585 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4586 ar
->len
= ac
->ac_b_ex
.fe_len
;
4589 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4597 ac
->ac_b_ex
.fe_len
= 0;
4599 ext4_mb_show_ac(ac
);
4601 ext4_mb_release_context(ac
);
4604 kmem_cache_free(ext4_ac_cachep
, ac
);
4605 if (inquota
&& ar
->len
< inquota
)
4606 dquot_free_block(ar
->inode
, EXT4_C2B(sbi
, inquota
- ar
->len
));
4608 if ((ar
->flags
& EXT4_MB_DELALLOC_RESERVED
) == 0)
4609 /* release all the reserved blocks if non delalloc */
4610 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
4614 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4620 * We can merge two free data extents only if the physical blocks
4621 * are contiguous, AND the extents were freed by the same transaction,
4622 * AND the blocks are associated with the same group.
4624 static void ext4_try_merge_freed_extent(struct ext4_sb_info
*sbi
,
4625 struct ext4_free_data
*entry
,
4626 struct ext4_free_data
*new_entry
,
4627 struct rb_root
*entry_rb_root
)
4629 if ((entry
->efd_tid
!= new_entry
->efd_tid
) ||
4630 (entry
->efd_group
!= new_entry
->efd_group
))
4632 if (entry
->efd_start_cluster
+ entry
->efd_count
==
4633 new_entry
->efd_start_cluster
) {
4634 new_entry
->efd_start_cluster
= entry
->efd_start_cluster
;
4635 new_entry
->efd_count
+= entry
->efd_count
;
4636 } else if (new_entry
->efd_start_cluster
+ new_entry
->efd_count
==
4637 entry
->efd_start_cluster
) {
4638 new_entry
->efd_count
+= entry
->efd_count
;
4641 spin_lock(&sbi
->s_md_lock
);
4642 list_del(&entry
->efd_list
);
4643 spin_unlock(&sbi
->s_md_lock
);
4644 rb_erase(&entry
->efd_node
, entry_rb_root
);
4645 kmem_cache_free(ext4_free_data_cachep
, entry
);
4648 static noinline_for_stack
int
4649 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4650 struct ext4_free_data
*new_entry
)
4652 ext4_group_t group
= e4b
->bd_group
;
4653 ext4_grpblk_t cluster
;
4654 ext4_grpblk_t clusters
= new_entry
->efd_count
;
4655 struct ext4_free_data
*entry
;
4656 struct ext4_group_info
*db
= e4b
->bd_info
;
4657 struct super_block
*sb
= e4b
->bd_sb
;
4658 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4659 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4660 struct rb_node
*parent
= NULL
, *new_node
;
4662 BUG_ON(!ext4_handle_valid(handle
));
4663 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4664 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4666 new_node
= &new_entry
->efd_node
;
4667 cluster
= new_entry
->efd_start_cluster
;
4670 /* first free block exent. We need to
4671 protect buddy cache from being freed,
4672 * otherwise we'll refresh it from
4673 * on-disk bitmap and lose not-yet-available
4675 get_page(e4b
->bd_buddy_page
);
4676 get_page(e4b
->bd_bitmap_page
);
4680 entry
= rb_entry(parent
, struct ext4_free_data
, efd_node
);
4681 if (cluster
< entry
->efd_start_cluster
)
4683 else if (cluster
>= (entry
->efd_start_cluster
+ entry
->efd_count
))
4684 n
= &(*n
)->rb_right
;
4686 ext4_grp_locked_error(sb
, group
, 0,
4687 ext4_group_first_block_no(sb
, group
) +
4688 EXT4_C2B(sbi
, cluster
),
4689 "Block already on to-be-freed list");
4694 rb_link_node(new_node
, parent
, n
);
4695 rb_insert_color(new_node
, &db
->bb_free_root
);
4697 /* Now try to see the extent can be merged to left and right */
4698 node
= rb_prev(new_node
);
4700 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4701 ext4_try_merge_freed_extent(sbi
, entry
, new_entry
,
4702 &(db
->bb_free_root
));
4705 node
= rb_next(new_node
);
4707 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4708 ext4_try_merge_freed_extent(sbi
, entry
, new_entry
,
4709 &(db
->bb_free_root
));
4712 spin_lock(&sbi
->s_md_lock
);
4713 list_add_tail(&new_entry
->efd_list
, &sbi
->s_freed_data_list
);
4714 sbi
->s_mb_free_pending
+= clusters
;
4715 spin_unlock(&sbi
->s_md_lock
);
4720 * ext4_free_blocks() -- Free given blocks and update quota
4721 * @handle: handle for this transaction
4723 * @block: start physical block to free
4724 * @count: number of blocks to count
4725 * @flags: flags used by ext4_free_blocks
4727 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4728 struct buffer_head
*bh
, ext4_fsblk_t block
,
4729 unsigned long count
, int flags
)
4731 struct buffer_head
*bitmap_bh
= NULL
;
4732 struct super_block
*sb
= inode
->i_sb
;
4733 struct ext4_group_desc
*gdp
;
4734 unsigned int overflow
;
4736 struct buffer_head
*gd_bh
;
4737 ext4_group_t block_group
;
4738 struct ext4_sb_info
*sbi
;
4739 struct ext4_buddy e4b
;
4740 unsigned int count_clusters
;
4747 BUG_ON(block
!= bh
->b_blocknr
);
4749 block
= bh
->b_blocknr
;
4753 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4754 !ext4_data_block_valid(sbi
, block
, count
)) {
4755 ext4_error(sb
, "Freeing blocks not in datazone - "
4756 "block = %llu, count = %lu", block
, count
);
4760 ext4_debug("freeing block %llu\n", block
);
4761 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4763 if (bh
&& (flags
& EXT4_FREE_BLOCKS_FORGET
)) {
4766 ext4_forget(handle
, flags
& EXT4_FREE_BLOCKS_METADATA
,
4771 * If the extent to be freed does not begin on a cluster
4772 * boundary, we need to deal with partial clusters at the
4773 * beginning and end of the extent. Normally we will free
4774 * blocks at the beginning or the end unless we are explicitly
4775 * requested to avoid doing so.
4777 overflow
= EXT4_PBLK_COFF(sbi
, block
);
4779 if (flags
& EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER
) {
4780 overflow
= sbi
->s_cluster_ratio
- overflow
;
4782 if (count
> overflow
)
4791 overflow
= EXT4_LBLK_COFF(sbi
, count
);
4793 if (flags
& EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER
) {
4794 if (count
> overflow
)
4799 count
+= sbi
->s_cluster_ratio
- overflow
;
4802 if (!bh
&& (flags
& EXT4_FREE_BLOCKS_FORGET
)) {
4804 int is_metadata
= flags
& EXT4_FREE_BLOCKS_METADATA
;
4806 for (i
= 0; i
< count
; i
++) {
4809 bh
= sb_find_get_block(inode
->i_sb
, block
+ i
);
4810 ext4_forget(handle
, is_metadata
, inode
, bh
, block
+ i
);
4816 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4818 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4819 ext4_get_group_info(sb
, block_group
))))
4823 * Check to see if we are freeing blocks across a group
4826 if (EXT4_C2B(sbi
, bit
) + count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4827 overflow
= EXT4_C2B(sbi
, bit
) + count
-
4828 EXT4_BLOCKS_PER_GROUP(sb
);
4831 count_clusters
= EXT4_NUM_B2C(sbi
, count
);
4832 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4833 if (IS_ERR(bitmap_bh
)) {
4834 err
= PTR_ERR(bitmap_bh
);
4838 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4844 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4845 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4846 in_range(block
, ext4_inode_table(sb
, gdp
),
4847 EXT4_SB(sb
)->s_itb_per_group
) ||
4848 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4849 EXT4_SB(sb
)->s_itb_per_group
)) {
4851 ext4_error(sb
, "Freeing blocks in system zone - "
4852 "Block = %llu, count = %lu", block
, count
);
4853 /* err = 0. ext4_std_error should be a no op */
4857 BUFFER_TRACE(bitmap_bh
, "getting write access");
4858 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4863 * We are about to modify some metadata. Call the journal APIs
4864 * to unshare ->b_data if a currently-committing transaction is
4867 BUFFER_TRACE(gd_bh
, "get_write_access");
4868 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4871 #ifdef AGGRESSIVE_CHECK
4874 for (i
= 0; i
< count_clusters
; i
++)
4875 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4878 trace_ext4_mballoc_free(sb
, inode
, block_group
, bit
, count_clusters
);
4880 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4881 err
= ext4_mb_load_buddy_gfp(sb
, block_group
, &e4b
,
4882 GFP_NOFS
|__GFP_NOFAIL
);
4887 * We need to make sure we don't reuse the freed block until after the
4888 * transaction is committed. We make an exception if the inode is to be
4889 * written in writeback mode since writeback mode has weak data
4890 * consistency guarantees.
4892 if (ext4_handle_valid(handle
) &&
4893 ((flags
& EXT4_FREE_BLOCKS_METADATA
) ||
4894 !ext4_should_writeback_data(inode
))) {
4895 struct ext4_free_data
*new_entry
;
4897 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4900 new_entry
= kmem_cache_alloc(ext4_free_data_cachep
,
4901 GFP_NOFS
|__GFP_NOFAIL
);
4902 new_entry
->efd_start_cluster
= bit
;
4903 new_entry
->efd_group
= block_group
;
4904 new_entry
->efd_count
= count_clusters
;
4905 new_entry
->efd_tid
= handle
->h_transaction
->t_tid
;
4907 ext4_lock_group(sb
, block_group
);
4908 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4909 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4911 /* need to update group_info->bb_free and bitmap
4912 * with group lock held. generate_buddy look at
4913 * them with group lock_held
4915 if (test_opt(sb
, DISCARD
)) {
4916 err
= ext4_issue_discard(sb
, block_group
, bit
, count
,
4918 if (err
&& err
!= -EOPNOTSUPP
)
4919 ext4_msg(sb
, KERN_WARNING
, "discard request in"
4920 " group:%d block:%d count:%lu failed"
4921 " with %d", block_group
, bit
, count
,
4924 EXT4_MB_GRP_CLEAR_TRIMMED(e4b
.bd_info
);
4926 ext4_lock_group(sb
, block_group
);
4927 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4928 mb_free_blocks(inode
, &e4b
, bit
, count_clusters
);
4931 ret
= ext4_free_group_clusters(sb
, gdp
) + count_clusters
;
4932 ext4_free_group_clusters_set(sb
, gdp
, ret
);
4933 ext4_block_bitmap_csum_set(sb
, block_group
, gdp
, bitmap_bh
);
4934 ext4_group_desc_csum_set(sb
, block_group
, gdp
);
4935 ext4_unlock_group(sb
, block_group
);
4937 if (sbi
->s_log_groups_per_flex
) {
4938 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4939 atomic64_add(count_clusters
,
4940 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
4943 if (!(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4944 dquot_free_block(inode
, EXT4_C2B(sbi
, count_clusters
));
4945 percpu_counter_add(&sbi
->s_freeclusters_counter
, count_clusters
);
4947 ext4_mb_unload_buddy(&e4b
);
4949 /* We dirtied the bitmap block */
4950 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4951 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4953 /* And the group descriptor block */
4954 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4955 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4959 if (overflow
&& !err
) {
4967 ext4_std_error(sb
, err
);
4972 * ext4_group_add_blocks() -- Add given blocks to an existing group
4973 * @handle: handle to this transaction
4975 * @block: start physical block to add to the block group
4976 * @count: number of blocks to free
4978 * This marks the blocks as free in the bitmap and buddy.
4980 int ext4_group_add_blocks(handle_t
*handle
, struct super_block
*sb
,
4981 ext4_fsblk_t block
, unsigned long count
)
4983 struct buffer_head
*bitmap_bh
= NULL
;
4984 struct buffer_head
*gd_bh
;
4985 ext4_group_t block_group
;
4988 struct ext4_group_desc
*desc
;
4989 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4990 struct ext4_buddy e4b
;
4991 int err
= 0, ret
, blk_free_count
;
4992 ext4_grpblk_t blocks_freed
;
4994 ext4_debug("Adding block(s) %llu-%llu\n", block
, block
+ count
- 1);
4999 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
5001 * Check to see if we are freeing blocks across a group
5004 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
5005 ext4_warning(sb
, "too much blocks added to group %u",
5011 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
5012 if (IS_ERR(bitmap_bh
)) {
5013 err
= PTR_ERR(bitmap_bh
);
5018 desc
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
5024 if (in_range(ext4_block_bitmap(sb
, desc
), block
, count
) ||
5025 in_range(ext4_inode_bitmap(sb
, desc
), block
, count
) ||
5026 in_range(block
, ext4_inode_table(sb
, desc
), sbi
->s_itb_per_group
) ||
5027 in_range(block
+ count
- 1, ext4_inode_table(sb
, desc
),
5028 sbi
->s_itb_per_group
)) {
5029 ext4_error(sb
, "Adding blocks in system zones - "
5030 "Block = %llu, count = %lu",
5036 BUFFER_TRACE(bitmap_bh
, "getting write access");
5037 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
5042 * We are about to modify some metadata. Call the journal APIs
5043 * to unshare ->b_data if a currently-committing transaction is
5046 BUFFER_TRACE(gd_bh
, "get_write_access");
5047 err
= ext4_journal_get_write_access(handle
, gd_bh
);
5051 for (i
= 0, blocks_freed
= 0; i
< count
; i
++) {
5052 BUFFER_TRACE(bitmap_bh
, "clear bit");
5053 if (!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
)) {
5054 ext4_error(sb
, "bit already cleared for block %llu",
5055 (ext4_fsblk_t
)(block
+ i
));
5056 BUFFER_TRACE(bitmap_bh
, "bit already cleared");
5062 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
5067 * need to update group_info->bb_free and bitmap
5068 * with group lock held. generate_buddy look at
5069 * them with group lock_held
5071 ext4_lock_group(sb
, block_group
);
5072 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
5073 mb_free_blocks(NULL
, &e4b
, bit
, count
);
5074 blk_free_count
= blocks_freed
+ ext4_free_group_clusters(sb
, desc
);
5075 ext4_free_group_clusters_set(sb
, desc
, blk_free_count
);
5076 ext4_block_bitmap_csum_set(sb
, block_group
, desc
, bitmap_bh
);
5077 ext4_group_desc_csum_set(sb
, block_group
, desc
);
5078 ext4_unlock_group(sb
, block_group
);
5079 percpu_counter_add(&sbi
->s_freeclusters_counter
,
5080 EXT4_NUM_B2C(sbi
, blocks_freed
));
5082 if (sbi
->s_log_groups_per_flex
) {
5083 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
5084 atomic64_add(EXT4_NUM_B2C(sbi
, blocks_freed
),
5085 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
5088 ext4_mb_unload_buddy(&e4b
);
5090 /* We dirtied the bitmap block */
5091 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
5092 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
5094 /* And the group descriptor block */
5095 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
5096 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
5102 ext4_std_error(sb
, err
);
5107 * ext4_trim_extent -- function to TRIM one single free extent in the group
5108 * @sb: super block for the file system
5109 * @start: starting block of the free extent in the alloc. group
5110 * @count: number of blocks to TRIM
5111 * @group: alloc. group we are working with
5112 * @e4b: ext4 buddy for the group
5114 * Trim "count" blocks starting at "start" in the "group". To assure that no
5115 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5116 * be called with under the group lock.
5118 static int ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
5119 ext4_group_t group
, struct ext4_buddy
*e4b
)
5123 struct ext4_free_extent ex
;
5126 trace_ext4_trim_extent(sb
, group
, start
, count
);
5128 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
5130 ex
.fe_start
= start
;
5131 ex
.fe_group
= group
;
5135 * Mark blocks used, so no one can reuse them while
5138 mb_mark_used(e4b
, &ex
);
5139 ext4_unlock_group(sb
, group
);
5140 ret
= ext4_issue_discard(sb
, group
, start
, count
, NULL
);
5141 ext4_lock_group(sb
, group
);
5142 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
5147 * ext4_trim_all_free -- function to trim all free space in alloc. group
5148 * @sb: super block for file system
5149 * @group: group to be trimmed
5150 * @start: first group block to examine
5151 * @max: last group block to examine
5152 * @minblocks: minimum extent block count
5154 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5155 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5159 * ext4_trim_all_free walks through group's block bitmap searching for free
5160 * extents. When the free extent is found, mark it as used in group buddy
5161 * bitmap. Then issue a TRIM command on this extent and free the extent in
5162 * the group buddy bitmap. This is done until whole group is scanned.
5164 static ext4_grpblk_t
5165 ext4_trim_all_free(struct super_block
*sb
, ext4_group_t group
,
5166 ext4_grpblk_t start
, ext4_grpblk_t max
,
5167 ext4_grpblk_t minblocks
)
5170 ext4_grpblk_t next
, count
= 0, free_count
= 0;
5171 struct ext4_buddy e4b
;
5174 trace_ext4_trim_all_free(sb
, group
, start
, max
);
5176 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
5178 ext4_warning(sb
, "Error %d loading buddy information for %u",
5182 bitmap
= e4b
.bd_bitmap
;
5184 ext4_lock_group(sb
, group
);
5185 if (EXT4_MB_GRP_WAS_TRIMMED(e4b
.bd_info
) &&
5186 minblocks
>= atomic_read(&EXT4_SB(sb
)->s_last_trim_minblks
))
5189 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
5190 e4b
.bd_info
->bb_first_free
: start
;
5192 while (start
<= max
) {
5193 start
= mb_find_next_zero_bit(bitmap
, max
+ 1, start
);
5196 next
= mb_find_next_bit(bitmap
, max
+ 1, start
);
5198 if ((next
- start
) >= minblocks
) {
5199 ret
= ext4_trim_extent(sb
, start
,
5200 next
- start
, group
, &e4b
);
5201 if (ret
&& ret
!= -EOPNOTSUPP
)
5204 count
+= next
- start
;
5206 free_count
+= next
- start
;
5209 if (fatal_signal_pending(current
)) {
5210 count
= -ERESTARTSYS
;
5214 if (need_resched()) {
5215 ext4_unlock_group(sb
, group
);
5217 ext4_lock_group(sb
, group
);
5220 if ((e4b
.bd_info
->bb_free
- free_count
) < minblocks
)
5226 EXT4_MB_GRP_SET_TRIMMED(e4b
.bd_info
);
5229 ext4_unlock_group(sb
, group
);
5230 ext4_mb_unload_buddy(&e4b
);
5232 ext4_debug("trimmed %d blocks in the group %d\n",
5239 * ext4_trim_fs() -- trim ioctl handle function
5240 * @sb: superblock for filesystem
5241 * @range: fstrim_range structure
5243 * start: First Byte to trim
5244 * len: number of Bytes to trim from start
5245 * minlen: minimum extent length in Bytes
5246 * ext4_trim_fs goes through all allocation groups containing Bytes from
5247 * start to start+len. For each such a group ext4_trim_all_free function
5248 * is invoked to trim all free space.
5250 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
5252 struct ext4_group_info
*grp
;
5253 ext4_group_t group
, first_group
, last_group
;
5254 ext4_grpblk_t cnt
= 0, first_cluster
, last_cluster
;
5255 uint64_t start
, end
, minlen
, trimmed
= 0;
5256 ext4_fsblk_t first_data_blk
=
5257 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
5258 ext4_fsblk_t max_blks
= ext4_blocks_count(EXT4_SB(sb
)->s_es
);
5261 start
= range
->start
>> sb
->s_blocksize_bits
;
5262 end
= start
+ (range
->len
>> sb
->s_blocksize_bits
) - 1;
5263 minlen
= EXT4_NUM_B2C(EXT4_SB(sb
),
5264 range
->minlen
>> sb
->s_blocksize_bits
);
5266 if (minlen
> EXT4_CLUSTERS_PER_GROUP(sb
) ||
5267 start
>= max_blks
||
5268 range
->len
< sb
->s_blocksize
)
5270 if (end
>= max_blks
)
5272 if (end
<= first_data_blk
)
5274 if (start
< first_data_blk
)
5275 start
= first_data_blk
;
5277 /* Determine first and last group to examine based on start and end */
5278 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
5279 &first_group
, &first_cluster
);
5280 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) end
,
5281 &last_group
, &last_cluster
);
5283 /* end now represents the last cluster to discard in this group */
5284 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5286 for (group
= first_group
; group
<= last_group
; group
++) {
5287 grp
= ext4_get_group_info(sb
, group
);
5288 /* We only do this if the grp has never been initialized */
5289 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
5290 ret
= ext4_mb_init_group(sb
, group
, GFP_NOFS
);
5296 * For all the groups except the last one, last cluster will
5297 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5298 * change it for the last group, note that last_cluster is
5299 * already computed earlier by ext4_get_group_no_and_offset()
5301 if (group
== last_group
)
5304 if (grp
->bb_free
>= minlen
) {
5305 cnt
= ext4_trim_all_free(sb
, group
, first_cluster
,
5315 * For every group except the first one, we are sure
5316 * that the first cluster to discard will be cluster #0.
5322 atomic_set(&EXT4_SB(sb
)->s_last_trim_minblks
, minlen
);
5325 range
->len
= EXT4_C2B(EXT4_SB(sb
), trimmed
) << sb
->s_blocksize_bits
;
5329 /* Iterate all the free extents in the group. */
5331 ext4_mballoc_query_range(
5332 struct super_block
*sb
,
5334 ext4_grpblk_t start
,
5336 ext4_mballoc_query_range_fn formatter
,
5341 struct ext4_buddy e4b
;
5344 error
= ext4_mb_load_buddy(sb
, group
, &e4b
);
5347 bitmap
= e4b
.bd_bitmap
;
5349 ext4_lock_group(sb
, group
);
5351 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
5352 e4b
.bd_info
->bb_first_free
: start
;
5353 if (end
>= EXT4_CLUSTERS_PER_GROUP(sb
))
5354 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5356 while (start
<= end
) {
5357 start
= mb_find_next_zero_bit(bitmap
, end
+ 1, start
);
5360 next
= mb_find_next_bit(bitmap
, end
+ 1, start
);
5362 ext4_unlock_group(sb
, group
);
5363 error
= formatter(sb
, group
, start
, next
- start
, priv
);
5366 ext4_lock_group(sb
, group
);
5371 ext4_unlock_group(sb
, group
);
5373 ext4_mb_unload_buddy(&e4b
);