1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
4 * Written by Alex Tomas <alex@clusterfs.com>
9 * mballoc.c contains the multiblocks allocation routines
12 #include "ext4_jbd2.h"
14 #include <linux/log2.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/nospec.h>
18 #include <linux/backing-dev.h>
19 #include <trace/events/ext4.h>
21 #ifdef CONFIG_EXT4_DEBUG
22 ushort ext4_mballoc_debug __read_mostly
;
24 module_param_named(mballoc_debug
, ext4_mballoc_debug
, ushort
, 0644);
25 MODULE_PARM_DESC(mballoc_debug
, "Debugging level for ext4's mballoc");
30 * - test ext4_ext_search_left() and ext4_ext_search_right()
31 * - search for metadata in few groups
34 * - normalization should take into account whether file is still open
35 * - discard preallocations if no free space left (policy?)
36 * - don't normalize tails
38 * - reservation for superuser
41 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
42 * - track min/max extents in each group for better group selection
43 * - mb_mark_used() may allocate chunk right after splitting buddy
44 * - tree of groups sorted by number of free blocks
49 * The allocation request involve request for multiple number of blocks
50 * near to the goal(block) value specified.
52 * During initialization phase of the allocator we decide to use the
53 * group preallocation or inode preallocation depending on the size of
54 * the file. The size of the file could be the resulting file size we
55 * would have after allocation, or the current file size, which ever
56 * is larger. If the size is less than sbi->s_mb_stream_request we
57 * select to use the group preallocation. The default value of
58 * s_mb_stream_request is 16 blocks. This can also be tuned via
59 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
60 * terms of number of blocks.
62 * The main motivation for having small file use group preallocation is to
63 * ensure that we have small files closer together on the disk.
65 * First stage the allocator looks at the inode prealloc list,
66 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
67 * spaces for this particular inode. The inode prealloc space is
70 * pa_lstart -> the logical start block for this prealloc space
71 * pa_pstart -> the physical start block for this prealloc space
72 * pa_len -> length for this prealloc space (in clusters)
73 * pa_free -> free space available in this prealloc space (in clusters)
75 * The inode preallocation space is used looking at the _logical_ start
76 * block. If only the logical file block falls within the range of prealloc
77 * space we will consume the particular prealloc space. This makes sure that
78 * we have contiguous physical blocks representing the file blocks
80 * The important thing to be noted in case of inode prealloc space is that
81 * we don't modify the values associated to inode prealloc space except
84 * If we are not able to find blocks in the inode prealloc space and if we
85 * have the group allocation flag set then we look at the locality group
86 * prealloc space. These are per CPU prealloc list represented as
88 * ext4_sb_info.s_locality_groups[smp_processor_id()]
90 * The reason for having a per cpu locality group is to reduce the contention
91 * between CPUs. It is possible to get scheduled at this point.
93 * The locality group prealloc space is used looking at whether we have
94 * enough free space (pa_free) within the prealloc space.
96 * If we can't allocate blocks via inode prealloc or/and locality group
97 * prealloc then we look at the buddy cache. The buddy cache is represented
98 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
99 * mapped to the buddy and bitmap information regarding different
100 * groups. The buddy information is attached to buddy cache inode so that
101 * we can access them through the page cache. The information regarding
102 * each group is loaded via ext4_mb_load_buddy. The information involve
103 * block bitmap and buddy information. The information are stored in the
107 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
110 * one block each for bitmap and buddy information. So for each group we
111 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
112 * blocksize) blocks. So it can have information regarding groups_per_page
113 * which is blocks_per_page/2
115 * The buddy cache inode is not stored on disk. The inode is thrown
116 * away when the filesystem is unmounted.
118 * We look for count number of blocks in the buddy cache. If we were able
119 * to locate that many free blocks we return with additional information
120 * regarding rest of the contiguous physical block available
122 * Before allocating blocks via buddy cache we normalize the request
123 * blocks. This ensure we ask for more blocks that we needed. The extra
124 * blocks that we get after allocation is added to the respective prealloc
125 * list. In case of inode preallocation we follow a list of heuristics
126 * based on file size. This can be found in ext4_mb_normalize_request. If
127 * we are doing a group prealloc we try to normalize the request to
128 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
129 * dependent on the cluster size; for non-bigalloc file systems, it is
130 * 512 blocks. This can be tuned via
131 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
132 * terms of number of blocks. If we have mounted the file system with -O
133 * stripe=<value> option the group prealloc request is normalized to the
134 * the smallest multiple of the stripe value (sbi->s_stripe) which is
135 * greater than the default mb_group_prealloc.
137 * The regular allocator (using the buddy cache) supports a few tunables.
139 * /sys/fs/ext4/<partition>/mb_min_to_scan
140 * /sys/fs/ext4/<partition>/mb_max_to_scan
141 * /sys/fs/ext4/<partition>/mb_order2_req
143 * The regular allocator uses buddy scan only if the request len is power of
144 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
145 * value of s_mb_order2_reqs can be tuned via
146 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
147 * stripe size (sbi->s_stripe), we try to search for contiguous block in
148 * stripe size. This should result in better allocation on RAID setups. If
149 * not, we search in the specific group using bitmap for best extents. The
150 * tunable min_to_scan and max_to_scan control the behaviour here.
151 * min_to_scan indicate how long the mballoc __must__ look for a best
152 * extent and max_to_scan indicates how long the mballoc __can__ look for a
153 * best extent in the found extents. Searching for the blocks starts with
154 * the group specified as the goal value in allocation context via
155 * ac_g_ex. Each group is first checked based on the criteria whether it
156 * can be used for allocation. ext4_mb_good_group explains how the groups are
159 * Both the prealloc space are getting populated as above. So for the first
160 * request we will hit the buddy cache which will result in this prealloc
161 * space getting filled. The prealloc space is then later used for the
162 * subsequent request.
166 * mballoc operates on the following data:
168 * - in-core buddy (actually includes buddy and bitmap)
169 * - preallocation descriptors (PAs)
171 * there are two types of preallocations:
173 * assiged to specific inode and can be used for this inode only.
174 * it describes part of inode's space preallocated to specific
175 * physical blocks. any block from that preallocated can be used
176 * independent. the descriptor just tracks number of blocks left
177 * unused. so, before taking some block from descriptor, one must
178 * make sure corresponded logical block isn't allocated yet. this
179 * also means that freeing any block within descriptor's range
180 * must discard all preallocated blocks.
182 * assigned to specific locality group which does not translate to
183 * permanent set of inodes: inode can join and leave group. space
184 * from this type of preallocation can be used for any inode. thus
185 * it's consumed from the beginning to the end.
187 * relation between them can be expressed as:
188 * in-core buddy = on-disk bitmap + preallocation descriptors
190 * this mean blocks mballoc considers used are:
191 * - allocated blocks (persistent)
192 * - preallocated blocks (non-persistent)
194 * consistency in mballoc world means that at any time a block is either
195 * free or used in ALL structures. notice: "any time" should not be read
196 * literally -- time is discrete and delimited by locks.
198 * to keep it simple, we don't use block numbers, instead we count number of
199 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
201 * all operations can be expressed as:
202 * - init buddy: buddy = on-disk + PAs
203 * - new PA: buddy += N; PA = N
204 * - use inode PA: on-disk += N; PA -= N
205 * - discard inode PA buddy -= on-disk - PA; PA = 0
206 * - use locality group PA on-disk += N; PA -= N
207 * - discard locality group PA buddy -= PA; PA = 0
208 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
209 * is used in real operation because we can't know actual used
210 * bits from PA, only from on-disk bitmap
212 * if we follow this strict logic, then all operations above should be atomic.
213 * given some of them can block, we'd have to use something like semaphores
214 * killing performance on high-end SMP hardware. let's try to relax it using
215 * the following knowledge:
216 * 1) if buddy is referenced, it's already initialized
217 * 2) while block is used in buddy and the buddy is referenced,
218 * nobody can re-allocate that block
219 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
220 * bit set and PA claims same block, it's OK. IOW, one can set bit in
221 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
224 * so, now we're building a concurrency table:
227 * blocks for PA are allocated in the buddy, buddy must be referenced
228 * until PA is linked to allocation group to avoid concurrent buddy init
230 * we need to make sure that either on-disk bitmap or PA has uptodate data
231 * given (3) we care that PA-=N operation doesn't interfere with init
233 * the simplest way would be to have buddy initialized by the discard
234 * - use locality group PA
235 * again PA-=N must be serialized with init
236 * - discard locality group PA
237 * the simplest way would be to have buddy initialized by the discard
240 * i_data_sem serializes them
242 * discard process must wait until PA isn't used by another process
243 * - use locality group PA
244 * some mutex should serialize them
245 * - discard locality group PA
246 * discard process must wait until PA isn't used by another process
249 * i_data_sem or another mutex should serializes them
251 * discard process must wait until PA isn't used by another process
252 * - use locality group PA
253 * nothing wrong here -- they're different PAs covering different blocks
254 * - discard locality group PA
255 * discard process must wait until PA isn't used by another process
257 * now we're ready to make few consequences:
258 * - PA is referenced and while it is no discard is possible
259 * - PA is referenced until block isn't marked in on-disk bitmap
260 * - PA changes only after on-disk bitmap
261 * - discard must not compete with init. either init is done before
262 * any discard or they're serialized somehow
263 * - buddy init as sum of on-disk bitmap and PAs is done atomically
265 * a special case when we've used PA to emptiness. no need to modify buddy
266 * in this case, but we should care about concurrent init
271 * Logic in few words:
276 * mark bits in on-disk bitmap
279 * - use preallocation:
280 * find proper PA (per-inode or group)
282 * mark bits in on-disk bitmap
288 * mark bits in on-disk bitmap
291 * - discard preallocations in group:
293 * move them onto local list
294 * load on-disk bitmap
296 * remove PA from object (inode or locality group)
297 * mark free blocks in-core
299 * - discard inode's preallocations:
306 * - bitlock on a group (group)
307 * - object (inode/locality) (object)
318 * - release consumed pa:
323 * - generate in-core bitmap:
327 * - discard all for given object (inode, locality group):
332 * - discard all for given group:
339 static struct kmem_cache
*ext4_pspace_cachep
;
340 static struct kmem_cache
*ext4_ac_cachep
;
341 static struct kmem_cache
*ext4_free_data_cachep
;
343 /* We create slab caches for groupinfo data structures based on the
344 * superblock block size. There will be one per mounted filesystem for
345 * each unique s_blocksize_bits */
346 #define NR_GRPINFO_CACHES 8
347 static struct kmem_cache
*ext4_groupinfo_caches
[NR_GRPINFO_CACHES
];
349 static const char * const ext4_groupinfo_slab_names
[NR_GRPINFO_CACHES
] = {
350 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
351 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
352 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
355 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
357 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
360 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
362 #if BITS_PER_LONG == 64
363 *bit
+= ((unsigned long) addr
& 7UL) << 3;
364 addr
= (void *) ((unsigned long) addr
& ~7UL);
365 #elif BITS_PER_LONG == 32
366 *bit
+= ((unsigned long) addr
& 3UL) << 3;
367 addr
= (void *) ((unsigned long) addr
& ~3UL);
369 #error "how many bits you are?!"
374 static inline int mb_test_bit(int bit
, void *addr
)
377 * ext4_test_bit on architecture like powerpc
378 * needs unsigned long aligned address
380 addr
= mb_correct_addr_and_bit(&bit
, addr
);
381 return ext4_test_bit(bit
, addr
);
384 static inline void mb_set_bit(int bit
, void *addr
)
386 addr
= mb_correct_addr_and_bit(&bit
, addr
);
387 ext4_set_bit(bit
, addr
);
390 static inline void mb_clear_bit(int bit
, void *addr
)
392 addr
= mb_correct_addr_and_bit(&bit
, addr
);
393 ext4_clear_bit(bit
, addr
);
396 static inline int mb_test_and_clear_bit(int bit
, void *addr
)
398 addr
= mb_correct_addr_and_bit(&bit
, addr
);
399 return ext4_test_and_clear_bit(bit
, addr
);
402 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
404 int fix
= 0, ret
, tmpmax
;
405 addr
= mb_correct_addr_and_bit(&fix
, addr
);
409 ret
= ext4_find_next_zero_bit(addr
, tmpmax
, start
) - fix
;
415 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
417 int fix
= 0, ret
, tmpmax
;
418 addr
= mb_correct_addr_and_bit(&fix
, addr
);
422 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
428 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
432 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
435 if (order
> e4b
->bd_blkbits
+ 1) {
440 /* at order 0 we see each particular block */
442 *max
= 1 << (e4b
->bd_blkbits
+ 3);
443 return e4b
->bd_bitmap
;
446 bb
= e4b
->bd_buddy
+ EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
447 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
453 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
454 int first
, int count
)
457 struct super_block
*sb
= e4b
->bd_sb
;
459 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
461 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
462 for (i
= 0; i
< count
; i
++) {
463 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
464 ext4_fsblk_t blocknr
;
466 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
467 blocknr
+= EXT4_C2B(EXT4_SB(sb
), first
+ i
);
468 ext4_grp_locked_error(sb
, e4b
->bd_group
,
469 inode
? inode
->i_ino
: 0,
471 "freeing block already freed "
474 ext4_mark_group_bitmap_corrupted(sb
, e4b
->bd_group
,
475 EXT4_GROUP_INFO_BBITMAP_CORRUPT
);
477 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
481 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
485 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
487 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
488 for (i
= 0; i
< count
; i
++) {
489 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
490 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
494 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
496 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
497 unsigned char *b1
, *b2
;
499 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
500 b2
= (unsigned char *) bitmap
;
501 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
502 if (b1
[i
] != b2
[i
]) {
503 ext4_msg(e4b
->bd_sb
, KERN_ERR
,
504 "corruption in group %u "
505 "at byte %u(%u): %x in copy != %x "
507 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
515 static inline void mb_free_blocks_double(struct inode
*inode
,
516 struct ext4_buddy
*e4b
, int first
, int count
)
520 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
521 int first
, int count
)
525 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
531 #ifdef AGGRESSIVE_CHECK
533 #define MB_CHECK_ASSERT(assert) \
537 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
538 function, file, line, # assert); \
543 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
544 const char *function
, int line
)
546 struct super_block
*sb
= e4b
->bd_sb
;
547 int order
= e4b
->bd_blkbits
+ 1;
554 struct ext4_group_info
*grp
;
557 struct list_head
*cur
;
562 static int mb_check_counter
;
563 if (mb_check_counter
++ % 100 != 0)
568 buddy
= mb_find_buddy(e4b
, order
, &max
);
569 MB_CHECK_ASSERT(buddy
);
570 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
571 MB_CHECK_ASSERT(buddy2
);
572 MB_CHECK_ASSERT(buddy
!= buddy2
);
573 MB_CHECK_ASSERT(max
* 2 == max2
);
576 for (i
= 0; i
< max
; i
++) {
578 if (mb_test_bit(i
, buddy
)) {
579 /* only single bit in buddy2 may be 1 */
580 if (!mb_test_bit(i
<< 1, buddy2
)) {
582 mb_test_bit((i
<<1)+1, buddy2
));
583 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
585 mb_test_bit(i
<< 1, buddy2
));
590 /* both bits in buddy2 must be 1 */
591 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
592 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
594 for (j
= 0; j
< (1 << order
); j
++) {
595 k
= (i
* (1 << order
)) + j
;
597 !mb_test_bit(k
, e4b
->bd_bitmap
));
601 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
606 buddy
= mb_find_buddy(e4b
, 0, &max
);
607 for (i
= 0; i
< max
; i
++) {
608 if (!mb_test_bit(i
, buddy
)) {
609 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
617 /* check used bits only */
618 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
619 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
621 MB_CHECK_ASSERT(k
< max2
);
622 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
625 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
626 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
628 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
629 list_for_each(cur
, &grp
->bb_prealloc_list
) {
630 ext4_group_t groupnr
;
631 struct ext4_prealloc_space
*pa
;
632 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
633 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
634 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
635 for (i
= 0; i
< pa
->pa_len
; i
++)
636 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
640 #undef MB_CHECK_ASSERT
641 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
642 __FILE__, __func__, __LINE__)
644 #define mb_check_buddy(e4b)
648 * Divide blocks started from @first with length @len into
649 * smaller chunks with power of 2 blocks.
650 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
651 * then increase bb_counters[] for corresponded chunk size.
653 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
654 void *buddy
, ext4_grpblk_t first
, ext4_grpblk_t len
,
655 struct ext4_group_info
*grp
)
657 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
663 BUG_ON(len
> EXT4_CLUSTERS_PER_GROUP(sb
));
665 border
= 2 << sb
->s_blocksize_bits
;
668 /* find how many blocks can be covered since this position */
669 max
= ffs(first
| border
) - 1;
671 /* find how many blocks of power 2 we need to mark */
678 /* mark multiblock chunks only */
679 grp
->bb_counters
[min
]++;
681 mb_clear_bit(first
>> min
,
682 buddy
+ sbi
->s_mb_offsets
[min
]);
690 * Cache the order of the largest free extent we have available in this block
694 mb_set_largest_free_order(struct super_block
*sb
, struct ext4_group_info
*grp
)
699 grp
->bb_largest_free_order
= -1; /* uninit */
701 bits
= sb
->s_blocksize_bits
+ 1;
702 for (i
= bits
; i
>= 0; i
--) {
703 if (grp
->bb_counters
[i
] > 0) {
704 grp
->bb_largest_free_order
= i
;
710 static noinline_for_stack
711 void ext4_mb_generate_buddy(struct super_block
*sb
,
712 void *buddy
, void *bitmap
, ext4_group_t group
)
714 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
715 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
716 ext4_grpblk_t max
= EXT4_CLUSTERS_PER_GROUP(sb
);
721 unsigned fragments
= 0;
722 unsigned long long period
= get_cycles();
724 /* initialize buddy from bitmap which is aggregation
725 * of on-disk bitmap and preallocations */
726 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
727 grp
->bb_first_free
= i
;
731 i
= mb_find_next_bit(bitmap
, max
, i
);
735 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
737 grp
->bb_counters
[0]++;
739 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
741 grp
->bb_fragments
= fragments
;
743 if (free
!= grp
->bb_free
) {
744 ext4_grp_locked_error(sb
, group
, 0, 0,
745 "block bitmap and bg descriptor "
746 "inconsistent: %u vs %u free clusters",
749 * If we intend to continue, we consider group descriptor
750 * corrupt and update bb_free using bitmap value
753 ext4_mark_group_bitmap_corrupted(sb
, group
,
754 EXT4_GROUP_INFO_BBITMAP_CORRUPT
);
756 mb_set_largest_free_order(sb
, grp
);
758 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
760 period
= get_cycles() - period
;
761 spin_lock(&sbi
->s_bal_lock
);
762 sbi
->s_mb_buddies_generated
++;
763 sbi
->s_mb_generation_time
+= period
;
764 spin_unlock(&sbi
->s_bal_lock
);
767 static void mb_regenerate_buddy(struct ext4_buddy
*e4b
)
773 while ((buddy
= mb_find_buddy(e4b
, order
++, &count
))) {
774 ext4_set_bits(buddy
, 0, count
);
776 e4b
->bd_info
->bb_fragments
= 0;
777 memset(e4b
->bd_info
->bb_counters
, 0,
778 sizeof(*e4b
->bd_info
->bb_counters
) *
779 (e4b
->bd_sb
->s_blocksize_bits
+ 2));
781 ext4_mb_generate_buddy(e4b
->bd_sb
, e4b
->bd_buddy
,
782 e4b
->bd_bitmap
, e4b
->bd_group
);
785 /* The buddy information is attached the buddy cache inode
786 * for convenience. The information regarding each group
787 * is loaded via ext4_mb_load_buddy. The information involve
788 * block bitmap and buddy information. The information are
789 * stored in the inode as
792 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
795 * one block each for bitmap and buddy information.
796 * So for each group we take up 2 blocks. A page can
797 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
798 * So it can have information regarding groups_per_page which
799 * is blocks_per_page/2
801 * Locking note: This routine takes the block group lock of all groups
802 * for this page; do not hold this lock when calling this routine!
805 static int ext4_mb_init_cache(struct page
*page
, char *incore
, gfp_t gfp
)
807 ext4_group_t ngroups
;
813 ext4_group_t first_group
, group
;
815 struct super_block
*sb
;
816 struct buffer_head
*bhs
;
817 struct buffer_head
**bh
= NULL
;
821 struct ext4_group_info
*grinfo
;
823 mb_debug(1, "init page %lu\n", page
->index
);
825 inode
= page
->mapping
->host
;
827 ngroups
= ext4_get_groups_count(sb
);
828 blocksize
= i_blocksize(inode
);
829 blocks_per_page
= PAGE_SIZE
/ blocksize
;
831 groups_per_page
= blocks_per_page
>> 1;
832 if (groups_per_page
== 0)
835 /* allocate buffer_heads to read bitmaps */
836 if (groups_per_page
> 1) {
837 i
= sizeof(struct buffer_head
*) * groups_per_page
;
838 bh
= kzalloc(i
, gfp
);
846 first_group
= page
->index
* blocks_per_page
/ 2;
848 /* read all groups the page covers into the cache */
849 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
850 if (group
>= ngroups
)
853 grinfo
= ext4_get_group_info(sb
, group
);
855 * If page is uptodate then we came here after online resize
856 * which added some new uninitialized group info structs, so
857 * we must skip all initialized uptodate buddies on the page,
858 * which may be currently in use by an allocating task.
860 if (PageUptodate(page
) && !EXT4_MB_GRP_NEED_INIT(grinfo
)) {
864 bh
[i
] = ext4_read_block_bitmap_nowait(sb
, group
);
866 err
= PTR_ERR(bh
[i
]);
870 mb_debug(1, "read bitmap for group %u\n", group
);
873 /* wait for I/O completion */
874 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
879 err2
= ext4_wait_block_bitmap(sb
, group
, bh
[i
]);
884 first_block
= page
->index
* blocks_per_page
;
885 for (i
= 0; i
< blocks_per_page
; i
++) {
886 group
= (first_block
+ i
) >> 1;
887 if (group
>= ngroups
)
890 if (!bh
[group
- first_group
])
891 /* skip initialized uptodate buddy */
894 if (!buffer_verified(bh
[group
- first_group
]))
895 /* Skip faulty bitmaps */
900 * data carry information regarding this
901 * particular group in the format specified
905 data
= page_address(page
) + (i
* blocksize
);
906 bitmap
= bh
[group
- first_group
]->b_data
;
909 * We place the buddy block and bitmap block
912 if ((first_block
+ i
) & 1) {
913 /* this is block of buddy */
914 BUG_ON(incore
== NULL
);
915 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
916 group
, page
->index
, i
* blocksize
);
917 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
918 grinfo
= ext4_get_group_info(sb
, group
);
919 grinfo
->bb_fragments
= 0;
920 memset(grinfo
->bb_counters
, 0,
921 sizeof(*grinfo
->bb_counters
) *
922 (sb
->s_blocksize_bits
+2));
924 * incore got set to the group block bitmap below
926 ext4_lock_group(sb
, group
);
928 memset(data
, 0xff, blocksize
);
929 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
930 ext4_unlock_group(sb
, group
);
933 /* this is block of bitmap */
934 BUG_ON(incore
!= NULL
);
935 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
936 group
, page
->index
, i
* blocksize
);
937 trace_ext4_mb_bitmap_load(sb
, group
);
939 /* see comments in ext4_mb_put_pa() */
940 ext4_lock_group(sb
, group
);
941 memcpy(data
, bitmap
, blocksize
);
943 /* mark all preallocated blks used in in-core bitmap */
944 ext4_mb_generate_from_pa(sb
, data
, group
);
945 ext4_mb_generate_from_freelist(sb
, data
, group
);
946 ext4_unlock_group(sb
, group
);
948 /* set incore so that the buddy information can be
949 * generated using this
954 SetPageUptodate(page
);
958 for (i
= 0; i
< groups_per_page
; i
++)
967 * Lock the buddy and bitmap pages. This make sure other parallel init_group
968 * on the same buddy page doesn't happen whild holding the buddy page lock.
969 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
970 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
972 static int ext4_mb_get_buddy_page_lock(struct super_block
*sb
,
973 ext4_group_t group
, struct ext4_buddy
*e4b
, gfp_t gfp
)
975 struct inode
*inode
= EXT4_SB(sb
)->s_buddy_cache
;
976 int block
, pnum
, poff
;
980 e4b
->bd_buddy_page
= NULL
;
981 e4b
->bd_bitmap_page
= NULL
;
983 blocks_per_page
= PAGE_SIZE
/ sb
->s_blocksize
;
985 * the buddy cache inode stores the block bitmap
986 * and buddy information in consecutive blocks.
987 * So for each group we need two blocks.
990 pnum
= block
/ blocks_per_page
;
991 poff
= block
% blocks_per_page
;
992 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
995 BUG_ON(page
->mapping
!= inode
->i_mapping
);
996 e4b
->bd_bitmap_page
= page
;
997 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
999 if (blocks_per_page
>= 2) {
1000 /* buddy and bitmap are on the same page */
1005 pnum
= block
/ blocks_per_page
;
1006 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1009 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1010 e4b
->bd_buddy_page
= page
;
1014 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy
*e4b
)
1016 if (e4b
->bd_bitmap_page
) {
1017 unlock_page(e4b
->bd_bitmap_page
);
1018 put_page(e4b
->bd_bitmap_page
);
1020 if (e4b
->bd_buddy_page
) {
1021 unlock_page(e4b
->bd_buddy_page
);
1022 put_page(e4b
->bd_buddy_page
);
1027 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1028 * block group lock of all groups for this page; do not hold the BG lock when
1029 * calling this routine!
1031 static noinline_for_stack
1032 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
, gfp_t gfp
)
1035 struct ext4_group_info
*this_grp
;
1036 struct ext4_buddy e4b
;
1041 mb_debug(1, "init group %u\n", group
);
1042 this_grp
= ext4_get_group_info(sb
, group
);
1044 * This ensures that we don't reinit the buddy cache
1045 * page which map to the group from which we are already
1046 * allocating. If we are looking at the buddy cache we would
1047 * have taken a reference using ext4_mb_load_buddy and that
1048 * would have pinned buddy page to page cache.
1049 * The call to ext4_mb_get_buddy_page_lock will mark the
1052 ret
= ext4_mb_get_buddy_page_lock(sb
, group
, &e4b
, gfp
);
1053 if (ret
|| !EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1055 * somebody initialized the group
1056 * return without doing anything
1061 page
= e4b
.bd_bitmap_page
;
1062 ret
= ext4_mb_init_cache(page
, NULL
, gfp
);
1065 if (!PageUptodate(page
)) {
1070 if (e4b
.bd_buddy_page
== NULL
) {
1072 * If both the bitmap and buddy are in
1073 * the same page we don't need to force
1079 /* init buddy cache */
1080 page
= e4b
.bd_buddy_page
;
1081 ret
= ext4_mb_init_cache(page
, e4b
.bd_bitmap
, gfp
);
1084 if (!PageUptodate(page
)) {
1089 ext4_mb_put_buddy_page_lock(&e4b
);
1094 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1095 * block group lock of all groups for this page; do not hold the BG lock when
1096 * calling this routine!
1098 static noinline_for_stack
int
1099 ext4_mb_load_buddy_gfp(struct super_block
*sb
, ext4_group_t group
,
1100 struct ext4_buddy
*e4b
, gfp_t gfp
)
1102 int blocks_per_page
;
1108 struct ext4_group_info
*grp
;
1109 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1110 struct inode
*inode
= sbi
->s_buddy_cache
;
1113 mb_debug(1, "load group %u\n", group
);
1115 blocks_per_page
= PAGE_SIZE
/ sb
->s_blocksize
;
1116 grp
= ext4_get_group_info(sb
, group
);
1118 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1121 e4b
->bd_group
= group
;
1122 e4b
->bd_buddy_page
= NULL
;
1123 e4b
->bd_bitmap_page
= NULL
;
1125 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1127 * we need full data about the group
1128 * to make a good selection
1130 ret
= ext4_mb_init_group(sb
, group
, gfp
);
1136 * the buddy cache inode stores the block bitmap
1137 * and buddy information in consecutive blocks.
1138 * So for each group we need two blocks.
1141 pnum
= block
/ blocks_per_page
;
1142 poff
= block
% blocks_per_page
;
1144 /* we could use find_or_create_page(), but it locks page
1145 * what we'd like to avoid in fast path ... */
1146 page
= find_get_page_flags(inode
->i_mapping
, pnum
, FGP_ACCESSED
);
1147 if (page
== NULL
|| !PageUptodate(page
)) {
1150 * drop the page reference and try
1151 * to get the page with lock. If we
1152 * are not uptodate that implies
1153 * somebody just created the page but
1154 * is yet to initialize the same. So
1155 * wait for it to initialize.
1158 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1160 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1161 if (!PageUptodate(page
)) {
1162 ret
= ext4_mb_init_cache(page
, NULL
, gfp
);
1167 mb_cmp_bitmaps(e4b
, page_address(page
) +
1168 (poff
* sb
->s_blocksize
));
1177 if (!PageUptodate(page
)) {
1182 /* Pages marked accessed already */
1183 e4b
->bd_bitmap_page
= page
;
1184 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1187 pnum
= block
/ blocks_per_page
;
1188 poff
= block
% blocks_per_page
;
1190 page
= find_get_page_flags(inode
->i_mapping
, pnum
, FGP_ACCESSED
);
1191 if (page
== NULL
|| !PageUptodate(page
)) {
1194 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1196 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1197 if (!PageUptodate(page
)) {
1198 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
,
1212 if (!PageUptodate(page
)) {
1217 /* Pages marked accessed already */
1218 e4b
->bd_buddy_page
= page
;
1219 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1221 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1222 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1229 if (e4b
->bd_bitmap_page
)
1230 put_page(e4b
->bd_bitmap_page
);
1231 if (e4b
->bd_buddy_page
)
1232 put_page(e4b
->bd_buddy_page
);
1233 e4b
->bd_buddy
= NULL
;
1234 e4b
->bd_bitmap
= NULL
;
1238 static int ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1239 struct ext4_buddy
*e4b
)
1241 return ext4_mb_load_buddy_gfp(sb
, group
, e4b
, GFP_NOFS
);
1244 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1246 if (e4b
->bd_bitmap_page
)
1247 put_page(e4b
->bd_bitmap_page
);
1248 if (e4b
->bd_buddy_page
)
1249 put_page(e4b
->bd_buddy_page
);
1253 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1256 int bb_incr
= 1 << (e4b
->bd_blkbits
- 1);
1259 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
1260 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1263 while (order
<= e4b
->bd_blkbits
+ 1) {
1265 if (!mb_test_bit(block
, bb
)) {
1266 /* this block is part of buddy of order 'order' */
1276 static void mb_clear_bits(void *bm
, int cur
, int len
)
1282 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1283 /* fast path: clear whole word at once */
1284 addr
= bm
+ (cur
>> 3);
1289 mb_clear_bit(cur
, bm
);
1294 /* clear bits in given range
1295 * will return first found zero bit if any, -1 otherwise
1297 static int mb_test_and_clear_bits(void *bm
, int cur
, int len
)
1304 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1305 /* fast path: clear whole word at once */
1306 addr
= bm
+ (cur
>> 3);
1307 if (*addr
!= (__u32
)(-1) && zero_bit
== -1)
1308 zero_bit
= cur
+ mb_find_next_zero_bit(addr
, 32, 0);
1313 if (!mb_test_and_clear_bit(cur
, bm
) && zero_bit
== -1)
1321 void ext4_set_bits(void *bm
, int cur
, int len
)
1327 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1328 /* fast path: set whole word at once */
1329 addr
= bm
+ (cur
>> 3);
1334 mb_set_bit(cur
, bm
);
1340 * _________________________________________________________________ */
1342 static inline int mb_buddy_adjust_border(int* bit
, void* bitmap
, int side
)
1344 if (mb_test_bit(*bit
+ side
, bitmap
)) {
1345 mb_clear_bit(*bit
, bitmap
);
1351 mb_set_bit(*bit
, bitmap
);
1356 static void mb_buddy_mark_free(struct ext4_buddy
*e4b
, int first
, int last
)
1360 void *buddy
= mb_find_buddy(e4b
, order
, &max
);
1365 /* Bits in range [first; last] are known to be set since
1366 * corresponding blocks were allocated. Bits in range
1367 * (first; last) will stay set because they form buddies on
1368 * upper layer. We just deal with borders if they don't
1369 * align with upper layer and then go up.
1370 * Releasing entire group is all about clearing
1371 * single bit of highest order buddy.
1375 * ---------------------------------
1377 * ---------------------------------
1378 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1379 * ---------------------------------
1381 * \_____________________/
1383 * Neither [1] nor [6] is aligned to above layer.
1384 * Left neighbour [0] is free, so mark it busy,
1385 * decrease bb_counters and extend range to
1387 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1388 * mark [6] free, increase bb_counters and shrink range to
1390 * Then shift range to [0; 2], go up and do the same.
1395 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&first
, buddy
, -1);
1397 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&last
, buddy
, 1);
1402 if (first
== last
|| !(buddy2
= mb_find_buddy(e4b
, order
, &max
))) {
1403 mb_clear_bits(buddy
, first
, last
- first
+ 1);
1404 e4b
->bd_info
->bb_counters
[order
- 1] += last
- first
+ 1;
1413 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1414 int first
, int count
)
1416 int left_is_free
= 0;
1417 int right_is_free
= 0;
1419 int last
= first
+ count
- 1;
1420 struct super_block
*sb
= e4b
->bd_sb
;
1422 if (WARN_ON(count
== 0))
1424 BUG_ON(last
>= (sb
->s_blocksize
<< 3));
1425 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1426 /* Don't bother if the block group is corrupt. */
1427 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
)))
1430 mb_check_buddy(e4b
);
1431 mb_free_blocks_double(inode
, e4b
, first
, count
);
1433 e4b
->bd_info
->bb_free
+= count
;
1434 if (first
< e4b
->bd_info
->bb_first_free
)
1435 e4b
->bd_info
->bb_first_free
= first
;
1437 /* access memory sequentially: check left neighbour,
1438 * clear range and then check right neighbour
1441 left_is_free
= !mb_test_bit(first
- 1, e4b
->bd_bitmap
);
1442 block
= mb_test_and_clear_bits(e4b
->bd_bitmap
, first
, count
);
1443 if (last
+ 1 < EXT4_SB(sb
)->s_mb_maxs
[0])
1444 right_is_free
= !mb_test_bit(last
+ 1, e4b
->bd_bitmap
);
1446 if (unlikely(block
!= -1)) {
1447 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1448 ext4_fsblk_t blocknr
;
1450 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1451 blocknr
+= EXT4_C2B(sbi
, block
);
1452 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1453 inode
? inode
->i_ino
: 0,
1455 "freeing already freed block "
1456 "(bit %u); block bitmap corrupt.",
1458 ext4_mark_group_bitmap_corrupted(sb
, e4b
->bd_group
,
1459 EXT4_GROUP_INFO_BBITMAP_CORRUPT
);
1460 mb_regenerate_buddy(e4b
);
1464 /* let's maintain fragments counter */
1465 if (left_is_free
&& right_is_free
)
1466 e4b
->bd_info
->bb_fragments
--;
1467 else if (!left_is_free
&& !right_is_free
)
1468 e4b
->bd_info
->bb_fragments
++;
1470 /* buddy[0] == bd_bitmap is a special case, so handle
1471 * it right away and let mb_buddy_mark_free stay free of
1472 * zero order checks.
1473 * Check if neighbours are to be coaleasced,
1474 * adjust bitmap bb_counters and borders appropriately.
1477 first
+= !left_is_free
;
1478 e4b
->bd_info
->bb_counters
[0] += left_is_free
? -1 : 1;
1481 last
-= !right_is_free
;
1482 e4b
->bd_info
->bb_counters
[0] += right_is_free
? -1 : 1;
1486 mb_buddy_mark_free(e4b
, first
>> 1, last
>> 1);
1489 mb_set_largest_free_order(sb
, e4b
->bd_info
);
1490 mb_check_buddy(e4b
);
1493 static int mb_find_extent(struct ext4_buddy
*e4b
, int block
,
1494 int needed
, struct ext4_free_extent
*ex
)
1500 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1503 buddy
= mb_find_buddy(e4b
, 0, &max
);
1504 BUG_ON(buddy
== NULL
);
1505 BUG_ON(block
>= max
);
1506 if (mb_test_bit(block
, buddy
)) {
1513 /* find actual order */
1514 order
= mb_find_order_for_block(e4b
, block
);
1515 block
= block
>> order
;
1517 ex
->fe_len
= 1 << order
;
1518 ex
->fe_start
= block
<< order
;
1519 ex
->fe_group
= e4b
->bd_group
;
1521 /* calc difference from given start */
1522 next
= next
- ex
->fe_start
;
1524 ex
->fe_start
+= next
;
1526 while (needed
> ex
->fe_len
&&
1527 mb_find_buddy(e4b
, order
, &max
)) {
1529 if (block
+ 1 >= max
)
1532 next
= (block
+ 1) * (1 << order
);
1533 if (mb_test_bit(next
, e4b
->bd_bitmap
))
1536 order
= mb_find_order_for_block(e4b
, next
);
1538 block
= next
>> order
;
1539 ex
->fe_len
+= 1 << order
;
1542 if (ex
->fe_start
+ ex
->fe_len
> EXT4_CLUSTERS_PER_GROUP(e4b
->bd_sb
)) {
1543 /* Should never happen! (but apparently sometimes does?!?) */
1545 ext4_error(e4b
->bd_sb
, "corruption or bug in mb_find_extent "
1546 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1547 block
, order
, needed
, ex
->fe_group
, ex
->fe_start
,
1548 ex
->fe_len
, ex
->fe_logical
);
1556 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1562 int start
= ex
->fe_start
;
1563 int len
= ex
->fe_len
;
1568 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1569 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1570 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1571 mb_check_buddy(e4b
);
1572 mb_mark_used_double(e4b
, start
, len
);
1574 e4b
->bd_info
->bb_free
-= len
;
1575 if (e4b
->bd_info
->bb_first_free
== start
)
1576 e4b
->bd_info
->bb_first_free
+= len
;
1578 /* let's maintain fragments counter */
1580 mlen
= !mb_test_bit(start
- 1, e4b
->bd_bitmap
);
1581 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1582 max
= !mb_test_bit(start
+ len
, e4b
->bd_bitmap
);
1584 e4b
->bd_info
->bb_fragments
++;
1585 else if (!mlen
&& !max
)
1586 e4b
->bd_info
->bb_fragments
--;
1588 /* let's maintain buddy itself */
1590 ord
= mb_find_order_for_block(e4b
, start
);
1592 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1593 /* the whole chunk may be allocated at once! */
1595 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1596 BUG_ON((start
>> ord
) >= max
);
1597 mb_set_bit(start
>> ord
, buddy
);
1598 e4b
->bd_info
->bb_counters
[ord
]--;
1605 /* store for history */
1607 ret
= len
| (ord
<< 16);
1609 /* we have to split large buddy */
1611 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1612 mb_set_bit(start
>> ord
, buddy
);
1613 e4b
->bd_info
->bb_counters
[ord
]--;
1616 cur
= (start
>> ord
) & ~1U;
1617 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1618 mb_clear_bit(cur
, buddy
);
1619 mb_clear_bit(cur
+ 1, buddy
);
1620 e4b
->bd_info
->bb_counters
[ord
]++;
1621 e4b
->bd_info
->bb_counters
[ord
]++;
1623 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1625 ext4_set_bits(e4b
->bd_bitmap
, ex
->fe_start
, len0
);
1626 mb_check_buddy(e4b
);
1632 * Must be called under group lock!
1634 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1635 struct ext4_buddy
*e4b
)
1637 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1640 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1641 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1643 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1644 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1645 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1647 /* preallocation can change ac_b_ex, thus we store actually
1648 * allocated blocks for history */
1649 ac
->ac_f_ex
= ac
->ac_b_ex
;
1651 ac
->ac_status
= AC_STATUS_FOUND
;
1652 ac
->ac_tail
= ret
& 0xffff;
1653 ac
->ac_buddy
= ret
>> 16;
1656 * take the page reference. We want the page to be pinned
1657 * so that we don't get a ext4_mb_init_cache_call for this
1658 * group until we update the bitmap. That would mean we
1659 * double allocate blocks. The reference is dropped
1660 * in ext4_mb_release_context
1662 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1663 get_page(ac
->ac_bitmap_page
);
1664 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1665 get_page(ac
->ac_buddy_page
);
1666 /* store last allocated for subsequent stream allocation */
1667 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1668 spin_lock(&sbi
->s_md_lock
);
1669 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1670 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1671 spin_unlock(&sbi
->s_md_lock
);
1676 * regular allocator, for general purposes allocation
1679 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1680 struct ext4_buddy
*e4b
,
1683 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1684 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1685 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1686 struct ext4_free_extent ex
;
1689 if (ac
->ac_status
== AC_STATUS_FOUND
)
1692 * We don't want to scan for a whole year
1694 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1695 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1696 ac
->ac_status
= AC_STATUS_BREAK
;
1701 * Haven't found good chunk so far, let's continue
1703 if (bex
->fe_len
< gex
->fe_len
)
1706 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1707 && bex
->fe_group
== e4b
->bd_group
) {
1708 /* recheck chunk's availability - we don't know
1709 * when it was found (within this lock-unlock
1711 max
= mb_find_extent(e4b
, bex
->fe_start
, gex
->fe_len
, &ex
);
1712 if (max
>= gex
->fe_len
) {
1713 ext4_mb_use_best_found(ac
, e4b
);
1720 * The routine checks whether found extent is good enough. If it is,
1721 * then the extent gets marked used and flag is set to the context
1722 * to stop scanning. Otherwise, the extent is compared with the
1723 * previous found extent and if new one is better, then it's stored
1724 * in the context. Later, the best found extent will be used, if
1725 * mballoc can't find good enough extent.
1727 * FIXME: real allocation policy is to be designed yet!
1729 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1730 struct ext4_free_extent
*ex
,
1731 struct ext4_buddy
*e4b
)
1733 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1734 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1736 BUG_ON(ex
->fe_len
<= 0);
1737 BUG_ON(ex
->fe_len
> EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1738 BUG_ON(ex
->fe_start
>= EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1739 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1744 * The special case - take what you catch first
1746 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1748 ext4_mb_use_best_found(ac
, e4b
);
1753 * Let's check whether the chuck is good enough
1755 if (ex
->fe_len
== gex
->fe_len
) {
1757 ext4_mb_use_best_found(ac
, e4b
);
1762 * If this is first found extent, just store it in the context
1764 if (bex
->fe_len
== 0) {
1770 * If new found extent is better, store it in the context
1772 if (bex
->fe_len
< gex
->fe_len
) {
1773 /* if the request isn't satisfied, any found extent
1774 * larger than previous best one is better */
1775 if (ex
->fe_len
> bex
->fe_len
)
1777 } else if (ex
->fe_len
> gex
->fe_len
) {
1778 /* if the request is satisfied, then we try to find
1779 * an extent that still satisfy the request, but is
1780 * smaller than previous one */
1781 if (ex
->fe_len
< bex
->fe_len
)
1785 ext4_mb_check_limits(ac
, e4b
, 0);
1788 static noinline_for_stack
1789 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1790 struct ext4_buddy
*e4b
)
1792 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1793 ext4_group_t group
= ex
.fe_group
;
1797 BUG_ON(ex
.fe_len
<= 0);
1798 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1802 ext4_lock_group(ac
->ac_sb
, group
);
1803 max
= mb_find_extent(e4b
, ex
.fe_start
, ex
.fe_len
, &ex
);
1807 ext4_mb_use_best_found(ac
, e4b
);
1810 ext4_unlock_group(ac
->ac_sb
, group
);
1811 ext4_mb_unload_buddy(e4b
);
1816 static noinline_for_stack
1817 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1818 struct ext4_buddy
*e4b
)
1820 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1823 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1824 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1825 struct ext4_free_extent ex
;
1827 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1829 if (grp
->bb_free
== 0)
1832 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1836 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
))) {
1837 ext4_mb_unload_buddy(e4b
);
1841 ext4_lock_group(ac
->ac_sb
, group
);
1842 max
= mb_find_extent(e4b
, ac
->ac_g_ex
.fe_start
,
1843 ac
->ac_g_ex
.fe_len
, &ex
);
1844 ex
.fe_logical
= 0xDEADFA11; /* debug value */
1846 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1849 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1851 /* use do_div to get remainder (would be 64-bit modulo) */
1852 if (do_div(start
, sbi
->s_stripe
) == 0) {
1855 ext4_mb_use_best_found(ac
, e4b
);
1857 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1858 BUG_ON(ex
.fe_len
<= 0);
1859 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1860 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1863 ext4_mb_use_best_found(ac
, e4b
);
1864 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1865 /* Sometimes, caller may want to merge even small
1866 * number of blocks to an existing extent */
1867 BUG_ON(ex
.fe_len
<= 0);
1868 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1869 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1872 ext4_mb_use_best_found(ac
, e4b
);
1874 ext4_unlock_group(ac
->ac_sb
, group
);
1875 ext4_mb_unload_buddy(e4b
);
1881 * The routine scans buddy structures (not bitmap!) from given order
1882 * to max order and tries to find big enough chunk to satisfy the req
1884 static noinline_for_stack
1885 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1886 struct ext4_buddy
*e4b
)
1888 struct super_block
*sb
= ac
->ac_sb
;
1889 struct ext4_group_info
*grp
= e4b
->bd_info
;
1895 BUG_ON(ac
->ac_2order
<= 0);
1896 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1897 if (grp
->bb_counters
[i
] == 0)
1900 buddy
= mb_find_buddy(e4b
, i
, &max
);
1901 BUG_ON(buddy
== NULL
);
1903 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1908 ac
->ac_b_ex
.fe_len
= 1 << i
;
1909 ac
->ac_b_ex
.fe_start
= k
<< i
;
1910 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1912 ext4_mb_use_best_found(ac
, e4b
);
1914 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1916 if (EXT4_SB(sb
)->s_mb_stats
)
1917 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1924 * The routine scans the group and measures all found extents.
1925 * In order to optimize scanning, caller must pass number of
1926 * free blocks in the group, so the routine can know upper limit.
1928 static noinline_for_stack
1929 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1930 struct ext4_buddy
*e4b
)
1932 struct super_block
*sb
= ac
->ac_sb
;
1933 void *bitmap
= e4b
->bd_bitmap
;
1934 struct ext4_free_extent ex
;
1938 free
= e4b
->bd_info
->bb_free
;
1941 i
= e4b
->bd_info
->bb_first_free
;
1943 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1944 i
= mb_find_next_zero_bit(bitmap
,
1945 EXT4_CLUSTERS_PER_GROUP(sb
), i
);
1946 if (i
>= EXT4_CLUSTERS_PER_GROUP(sb
)) {
1948 * IF we have corrupt bitmap, we won't find any
1949 * free blocks even though group info says we
1950 * we have free blocks
1952 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1953 "%d free clusters as per "
1954 "group info. But bitmap says 0",
1956 ext4_mark_group_bitmap_corrupted(sb
, e4b
->bd_group
,
1957 EXT4_GROUP_INFO_BBITMAP_CORRUPT
);
1961 mb_find_extent(e4b
, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1962 BUG_ON(ex
.fe_len
<= 0);
1963 if (free
< ex
.fe_len
) {
1964 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1965 "%d free clusters as per "
1966 "group info. But got %d blocks",
1968 ext4_mark_group_bitmap_corrupted(sb
, e4b
->bd_group
,
1969 EXT4_GROUP_INFO_BBITMAP_CORRUPT
);
1971 * The number of free blocks differs. This mostly
1972 * indicate that the bitmap is corrupt. So exit
1973 * without claiming the space.
1977 ex
.fe_logical
= 0xDEADC0DE; /* debug value */
1978 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1984 ext4_mb_check_limits(ac
, e4b
, 1);
1988 * This is a special case for storages like raid5
1989 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1991 static noinline_for_stack
1992 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1993 struct ext4_buddy
*e4b
)
1995 struct super_block
*sb
= ac
->ac_sb
;
1996 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1997 void *bitmap
= e4b
->bd_bitmap
;
1998 struct ext4_free_extent ex
;
1999 ext4_fsblk_t first_group_block
;
2004 BUG_ON(sbi
->s_stripe
== 0);
2006 /* find first stripe-aligned block in group */
2007 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
2009 a
= first_group_block
+ sbi
->s_stripe
- 1;
2010 do_div(a
, sbi
->s_stripe
);
2011 i
= (a
* sbi
->s_stripe
) - first_group_block
;
2013 while (i
< EXT4_CLUSTERS_PER_GROUP(sb
)) {
2014 if (!mb_test_bit(i
, bitmap
)) {
2015 max
= mb_find_extent(e4b
, i
, sbi
->s_stripe
, &ex
);
2016 if (max
>= sbi
->s_stripe
) {
2018 ex
.fe_logical
= 0xDEADF00D; /* debug value */
2020 ext4_mb_use_best_found(ac
, e4b
);
2029 * This is now called BEFORE we load the buddy bitmap.
2030 * Returns either 1 or 0 indicating that the group is either suitable
2031 * for the allocation or not. In addition it can also return negative
2032 * error code when something goes wrong.
2034 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
2035 ext4_group_t group
, int cr
)
2037 unsigned free
, fragments
;
2038 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
2039 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
2041 BUG_ON(cr
< 0 || cr
>= 4);
2043 free
= grp
->bb_free
;
2046 if (cr
<= 2 && free
< ac
->ac_g_ex
.fe_len
)
2049 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp
)))
2052 /* We only do this if the grp has never been initialized */
2053 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
2054 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
, GFP_NOFS
);
2059 fragments
= grp
->bb_fragments
;
2065 BUG_ON(ac
->ac_2order
== 0);
2067 /* Avoid using the first bg of a flexgroup for data files */
2068 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
2069 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
2070 ((group
% flex_size
) == 0))
2073 if ((ac
->ac_2order
> ac
->ac_sb
->s_blocksize_bits
+1) ||
2074 (free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2077 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
2082 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2086 if (free
>= ac
->ac_g_ex
.fe_len
)
2098 static noinline_for_stack
int
2099 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
2101 ext4_group_t ngroups
, group
, i
;
2103 int err
= 0, first_err
= 0;
2104 struct ext4_sb_info
*sbi
;
2105 struct super_block
*sb
;
2106 struct ext4_buddy e4b
;
2110 ngroups
= ext4_get_groups_count(sb
);
2111 /* non-extent files are limited to low blocks/groups */
2112 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
2113 ngroups
= sbi
->s_blockfile_groups
;
2115 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
2117 /* first, try the goal */
2118 err
= ext4_mb_find_by_goal(ac
, &e4b
);
2119 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
2122 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2126 * ac->ac2_order is set only if the fe_len is a power of 2
2127 * if ac2_order is set we also set criteria to 0 so that we
2128 * try exact allocation using buddy.
2130 i
= fls(ac
->ac_g_ex
.fe_len
);
2133 * We search using buddy data only if the order of the request
2134 * is greater than equal to the sbi_s_mb_order2_reqs
2135 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2136 * We also support searching for power-of-two requests only for
2137 * requests upto maximum buddy size we have constructed.
2139 if (i
>= sbi
->s_mb_order2_reqs
&& i
<= sb
->s_blocksize_bits
+ 2) {
2141 * This should tell if fe_len is exactly power of 2
2143 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2144 ac
->ac_2order
= array_index_nospec(i
- 1,
2145 sb
->s_blocksize_bits
+ 2);
2148 /* if stream allocation is enabled, use global goal */
2149 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2150 /* TBD: may be hot point */
2151 spin_lock(&sbi
->s_md_lock
);
2152 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2153 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2154 spin_unlock(&sbi
->s_md_lock
);
2157 /* Let's just scan groups to find more-less suitable blocks */
2158 cr
= ac
->ac_2order
? 0 : 1;
2160 * cr == 0 try to get exact allocation,
2161 * cr == 3 try to get anything
2164 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2165 ac
->ac_criteria
= cr
;
2167 * searching for the right group start
2168 * from the goal value specified
2170 group
= ac
->ac_g_ex
.fe_group
;
2172 for (i
= 0; i
< ngroups
; group
++, i
++) {
2176 * Artificially restricted ngroups for non-extent
2177 * files makes group > ngroups possible on first loop.
2179 if (group
>= ngroups
)
2182 /* This now checks without needing the buddy page */
2183 ret
= ext4_mb_good_group(ac
, group
, cr
);
2190 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2194 ext4_lock_group(sb
, group
);
2197 * We need to check again after locking the
2200 ret
= ext4_mb_good_group(ac
, group
, cr
);
2202 ext4_unlock_group(sb
, group
);
2203 ext4_mb_unload_buddy(&e4b
);
2209 ac
->ac_groups_scanned
++;
2211 ext4_mb_simple_scan_group(ac
, &e4b
);
2212 else if (cr
== 1 && sbi
->s_stripe
&&
2213 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2214 ext4_mb_scan_aligned(ac
, &e4b
);
2216 ext4_mb_complex_scan_group(ac
, &e4b
);
2218 ext4_unlock_group(sb
, group
);
2219 ext4_mb_unload_buddy(&e4b
);
2221 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2226 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2227 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2229 * We've been searching too long. Let's try to allocate
2230 * the best chunk we've found so far
2233 ext4_mb_try_best_found(ac
, &e4b
);
2234 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2236 * Someone more lucky has already allocated it.
2237 * The only thing we can do is just take first
2239 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2241 ac
->ac_b_ex
.fe_group
= 0;
2242 ac
->ac_b_ex
.fe_start
= 0;
2243 ac
->ac_b_ex
.fe_len
= 0;
2244 ac
->ac_status
= AC_STATUS_CONTINUE
;
2245 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2247 atomic_inc(&sbi
->s_mb_lost_chunks
);
2252 if (!err
&& ac
->ac_status
!= AC_STATUS_FOUND
&& first_err
)
2257 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2259 struct super_block
*sb
= PDE_DATA(file_inode(seq
->file
));
2262 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2265 return (void *) ((unsigned long) group
);
2268 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2270 struct super_block
*sb
= PDE_DATA(file_inode(seq
->file
));
2274 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2277 return (void *) ((unsigned long) group
);
2280 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2282 struct super_block
*sb
= PDE_DATA(file_inode(seq
->file
));
2283 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2285 int err
, buddy_loaded
= 0;
2286 struct ext4_buddy e4b
;
2287 struct ext4_group_info
*grinfo
;
2288 unsigned char blocksize_bits
= min_t(unsigned char,
2289 sb
->s_blocksize_bits
,
2290 EXT4_MAX_BLOCK_LOG_SIZE
);
2292 struct ext4_group_info info
;
2293 ext4_grpblk_t counters
[EXT4_MAX_BLOCK_LOG_SIZE
+ 2];
2298 seq_puts(seq
, "#group: free frags first ["
2299 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2300 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2302 i
= (blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2303 sizeof(struct ext4_group_info
);
2305 grinfo
= ext4_get_group_info(sb
, group
);
2306 /* Load the group info in memory only if not already loaded. */
2307 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo
))) {
2308 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2310 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2316 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2319 ext4_mb_unload_buddy(&e4b
);
2321 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2322 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2323 for (i
= 0; i
<= 13; i
++)
2324 seq_printf(seq
, " %-5u", i
<= blocksize_bits
+ 1 ?
2325 sg
.info
.bb_counters
[i
] : 0);
2326 seq_printf(seq
, " ]\n");
2331 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2335 const struct seq_operations ext4_mb_seq_groups_ops
= {
2336 .start
= ext4_mb_seq_groups_start
,
2337 .next
= ext4_mb_seq_groups_next
,
2338 .stop
= ext4_mb_seq_groups_stop
,
2339 .show
= ext4_mb_seq_groups_show
,
2342 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2344 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2345 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2352 * Allocate the top-level s_group_info array for the specified number
2355 int ext4_mb_alloc_groupinfo(struct super_block
*sb
, ext4_group_t ngroups
)
2357 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2359 struct ext4_group_info
***old_groupinfo
, ***new_groupinfo
;
2361 size
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2362 EXT4_DESC_PER_BLOCK_BITS(sb
);
2363 if (size
<= sbi
->s_group_info_size
)
2366 size
= roundup_pow_of_two(sizeof(*sbi
->s_group_info
) * size
);
2367 new_groupinfo
= kvzalloc(size
, GFP_KERNEL
);
2368 if (!new_groupinfo
) {
2369 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy meta group");
2373 old_groupinfo
= rcu_dereference(sbi
->s_group_info
);
2375 memcpy(new_groupinfo
, old_groupinfo
,
2376 sbi
->s_group_info_size
* sizeof(*sbi
->s_group_info
));
2378 rcu_assign_pointer(sbi
->s_group_info
, new_groupinfo
);
2379 sbi
->s_group_info_size
= size
/ sizeof(*sbi
->s_group_info
);
2381 ext4_kvfree_array_rcu(old_groupinfo
);
2382 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2383 sbi
->s_group_info_size
);
2387 /* Create and initialize ext4_group_info data for the given group. */
2388 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2389 struct ext4_group_desc
*desc
)
2393 int idx
= group
>> EXT4_DESC_PER_BLOCK_BITS(sb
);
2394 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2395 struct ext4_group_info
**meta_group_info
;
2396 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2399 * First check if this group is the first of a reserved block.
2400 * If it's true, we have to allocate a new table of pointers
2401 * to ext4_group_info structures
2403 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2404 metalen
= sizeof(*meta_group_info
) <<
2405 EXT4_DESC_PER_BLOCK_BITS(sb
);
2406 meta_group_info
= kmalloc(metalen
, GFP_NOFS
);
2407 if (meta_group_info
== NULL
) {
2408 ext4_msg(sb
, KERN_ERR
, "can't allocate mem "
2409 "for a buddy group");
2410 goto exit_meta_group_info
;
2413 rcu_dereference(sbi
->s_group_info
)[idx
] = meta_group_info
;
2417 meta_group_info
= sbi_array_rcu_deref(sbi
, s_group_info
, idx
);
2418 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2420 meta_group_info
[i
] = kmem_cache_zalloc(cachep
, GFP_NOFS
);
2421 if (meta_group_info
[i
] == NULL
) {
2422 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy mem");
2423 goto exit_group_info
;
2425 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2426 &(meta_group_info
[i
]->bb_state
));
2429 * initialize bb_free to be able to skip
2430 * empty groups without initialization
2432 if (ext4_has_group_desc_csum(sb
) &&
2433 (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))) {
2434 meta_group_info
[i
]->bb_free
=
2435 ext4_free_clusters_after_init(sb
, group
, desc
);
2437 meta_group_info
[i
]->bb_free
=
2438 ext4_free_group_clusters(sb
, desc
);
2441 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2442 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2443 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2444 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2448 struct buffer_head
*bh
;
2449 meta_group_info
[i
]->bb_bitmap
=
2450 kmalloc(sb
->s_blocksize
, GFP_NOFS
);
2451 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2452 bh
= ext4_read_block_bitmap(sb
, group
);
2453 BUG_ON(IS_ERR_OR_NULL(bh
));
2454 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2463 /* If a meta_group_info table has been allocated, release it now */
2464 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2465 struct ext4_group_info
***group_info
;
2468 group_info
= rcu_dereference(sbi
->s_group_info
);
2469 kfree(group_info
[idx
]);
2470 group_info
[idx
] = NULL
;
2473 exit_meta_group_info
:
2475 } /* ext4_mb_add_groupinfo */
2477 static int ext4_mb_init_backend(struct super_block
*sb
)
2479 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2481 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2483 struct ext4_group_desc
*desc
;
2484 struct ext4_group_info
***group_info
;
2485 struct kmem_cache
*cachep
;
2487 err
= ext4_mb_alloc_groupinfo(sb
, ngroups
);
2491 sbi
->s_buddy_cache
= new_inode(sb
);
2492 if (sbi
->s_buddy_cache
== NULL
) {
2493 ext4_msg(sb
, KERN_ERR
, "can't get new inode");
2496 /* To avoid potentially colliding with an valid on-disk inode number,
2497 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2498 * not in the inode hash, so it should never be found by iget(), but
2499 * this will avoid confusion if it ever shows up during debugging. */
2500 sbi
->s_buddy_cache
->i_ino
= EXT4_BAD_INO
;
2501 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2502 for (i
= 0; i
< ngroups
; i
++) {
2504 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2506 ext4_msg(sb
, KERN_ERR
, "can't read descriptor %u", i
);
2509 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2516 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2518 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2519 i
= sbi
->s_group_info_size
;
2521 group_info
= rcu_dereference(sbi
->s_group_info
);
2523 kfree(group_info
[i
]);
2525 iput(sbi
->s_buddy_cache
);
2528 kvfree(rcu_dereference(sbi
->s_group_info
));
2533 static void ext4_groupinfo_destroy_slabs(void)
2537 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2538 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2539 ext4_groupinfo_caches
[i
] = NULL
;
2543 static int ext4_groupinfo_create_slab(size_t size
)
2545 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2547 int blocksize_bits
= order_base_2(size
);
2548 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2549 struct kmem_cache
*cachep
;
2551 if (cache_index
>= NR_GRPINFO_CACHES
)
2554 if (unlikely(cache_index
< 0))
2557 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2558 if (ext4_groupinfo_caches
[cache_index
]) {
2559 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2560 return 0; /* Already created */
2563 slab_size
= offsetof(struct ext4_group_info
,
2564 bb_counters
[blocksize_bits
+ 2]);
2566 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2567 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2570 ext4_groupinfo_caches
[cache_index
] = cachep
;
2572 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2575 "EXT4-fs: no memory for groupinfo slab cache\n");
2582 int ext4_mb_init(struct super_block
*sb
)
2584 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2586 unsigned offset
, offset_incr
;
2590 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2592 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2593 if (sbi
->s_mb_offsets
== NULL
) {
2598 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2599 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2600 if (sbi
->s_mb_maxs
== NULL
) {
2605 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2609 /* order 0 is regular bitmap */
2610 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2611 sbi
->s_mb_offsets
[0] = 0;
2615 offset_incr
= 1 << (sb
->s_blocksize_bits
- 1);
2616 max
= sb
->s_blocksize
<< 2;
2618 sbi
->s_mb_offsets
[i
] = offset
;
2619 sbi
->s_mb_maxs
[i
] = max
;
2620 offset
+= offset_incr
;
2621 offset_incr
= offset_incr
>> 1;
2624 } while (i
<= sb
->s_blocksize_bits
+ 1);
2626 spin_lock_init(&sbi
->s_md_lock
);
2627 spin_lock_init(&sbi
->s_bal_lock
);
2628 sbi
->s_mb_free_pending
= 0;
2629 INIT_LIST_HEAD(&sbi
->s_freed_data_list
);
2631 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2632 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2633 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2634 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2635 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2637 * The default group preallocation is 512, which for 4k block
2638 * sizes translates to 2 megabytes. However for bigalloc file
2639 * systems, this is probably too big (i.e, if the cluster size
2640 * is 1 megabyte, then group preallocation size becomes half a
2641 * gigabyte!). As a default, we will keep a two megabyte
2642 * group pralloc size for cluster sizes up to 64k, and after
2643 * that, we will force a minimum group preallocation size of
2644 * 32 clusters. This translates to 8 megs when the cluster
2645 * size is 256k, and 32 megs when the cluster size is 1 meg,
2646 * which seems reasonable as a default.
2648 sbi
->s_mb_group_prealloc
= max(MB_DEFAULT_GROUP_PREALLOC
>>
2649 sbi
->s_cluster_bits
, 32);
2651 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2652 * to the lowest multiple of s_stripe which is bigger than
2653 * the s_mb_group_prealloc as determined above. We want
2654 * the preallocation size to be an exact multiple of the
2655 * RAID stripe size so that preallocations don't fragment
2658 if (sbi
->s_stripe
> 1) {
2659 sbi
->s_mb_group_prealloc
= roundup(
2660 sbi
->s_mb_group_prealloc
, sbi
->s_stripe
);
2663 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2664 if (sbi
->s_locality_groups
== NULL
) {
2668 for_each_possible_cpu(i
) {
2669 struct ext4_locality_group
*lg
;
2670 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2671 mutex_init(&lg
->lg_mutex
);
2672 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2673 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2674 spin_lock_init(&lg
->lg_prealloc_lock
);
2677 /* init file for buddy data */
2678 ret
= ext4_mb_init_backend(sb
);
2680 goto out_free_locality_groups
;
2684 out_free_locality_groups
:
2685 free_percpu(sbi
->s_locality_groups
);
2686 sbi
->s_locality_groups
= NULL
;
2688 kfree(sbi
->s_mb_offsets
);
2689 sbi
->s_mb_offsets
= NULL
;
2690 kfree(sbi
->s_mb_maxs
);
2691 sbi
->s_mb_maxs
= NULL
;
2695 /* need to called with the ext4 group lock held */
2696 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2698 struct ext4_prealloc_space
*pa
;
2699 struct list_head
*cur
, *tmp
;
2702 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2703 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2704 list_del(&pa
->pa_group_list
);
2706 kmem_cache_free(ext4_pspace_cachep
, pa
);
2709 mb_debug(1, "mballoc: %u PAs left\n", count
);
2713 int ext4_mb_release(struct super_block
*sb
)
2715 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2717 int num_meta_group_infos
;
2718 struct ext4_group_info
*grinfo
, ***group_info
;
2719 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2720 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2722 if (sbi
->s_group_info
) {
2723 for (i
= 0; i
< ngroups
; i
++) {
2725 grinfo
= ext4_get_group_info(sb
, i
);
2727 kfree(grinfo
->bb_bitmap
);
2729 ext4_lock_group(sb
, i
);
2730 ext4_mb_cleanup_pa(grinfo
);
2731 ext4_unlock_group(sb
, i
);
2732 kmem_cache_free(cachep
, grinfo
);
2734 num_meta_group_infos
= (ngroups
+
2735 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2736 EXT4_DESC_PER_BLOCK_BITS(sb
);
2738 group_info
= rcu_dereference(sbi
->s_group_info
);
2739 for (i
= 0; i
< num_meta_group_infos
; i
++)
2740 kfree(group_info
[i
]);
2744 kfree(sbi
->s_mb_offsets
);
2745 kfree(sbi
->s_mb_maxs
);
2746 iput(sbi
->s_buddy_cache
);
2747 if (sbi
->s_mb_stats
) {
2748 ext4_msg(sb
, KERN_INFO
,
2749 "mballoc: %u blocks %u reqs (%u success)",
2750 atomic_read(&sbi
->s_bal_allocated
),
2751 atomic_read(&sbi
->s_bal_reqs
),
2752 atomic_read(&sbi
->s_bal_success
));
2753 ext4_msg(sb
, KERN_INFO
,
2754 "mballoc: %u extents scanned, %u goal hits, "
2755 "%u 2^N hits, %u breaks, %u lost",
2756 atomic_read(&sbi
->s_bal_ex_scanned
),
2757 atomic_read(&sbi
->s_bal_goals
),
2758 atomic_read(&sbi
->s_bal_2orders
),
2759 atomic_read(&sbi
->s_bal_breaks
),
2760 atomic_read(&sbi
->s_mb_lost_chunks
));
2761 ext4_msg(sb
, KERN_INFO
,
2762 "mballoc: %lu generated and it took %Lu",
2763 sbi
->s_mb_buddies_generated
,
2764 sbi
->s_mb_generation_time
);
2765 ext4_msg(sb
, KERN_INFO
,
2766 "mballoc: %u preallocated, %u discarded",
2767 atomic_read(&sbi
->s_mb_preallocated
),
2768 atomic_read(&sbi
->s_mb_discarded
));
2771 free_percpu(sbi
->s_locality_groups
);
2776 static inline int ext4_issue_discard(struct super_block
*sb
,
2777 ext4_group_t block_group
, ext4_grpblk_t cluster
, int count
,
2780 ext4_fsblk_t discard_block
;
2782 discard_block
= (EXT4_C2B(EXT4_SB(sb
), cluster
) +
2783 ext4_group_first_block_no(sb
, block_group
));
2784 count
= EXT4_C2B(EXT4_SB(sb
), count
);
2785 trace_ext4_discard_blocks(sb
,
2786 (unsigned long long) discard_block
, count
);
2788 return __blkdev_issue_discard(sb
->s_bdev
,
2789 (sector_t
)discard_block
<< (sb
->s_blocksize_bits
- 9),
2790 (sector_t
)count
<< (sb
->s_blocksize_bits
- 9),
2793 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, 0);
2796 static void ext4_free_data_in_buddy(struct super_block
*sb
,
2797 struct ext4_free_data
*entry
)
2799 struct ext4_buddy e4b
;
2800 struct ext4_group_info
*db
;
2801 int err
, count
= 0, count2
= 0;
2803 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2804 entry
->efd_count
, entry
->efd_group
, entry
);
2806 err
= ext4_mb_load_buddy(sb
, entry
->efd_group
, &e4b
);
2807 /* we expect to find existing buddy because it's pinned */
2810 spin_lock(&EXT4_SB(sb
)->s_md_lock
);
2811 EXT4_SB(sb
)->s_mb_free_pending
-= entry
->efd_count
;
2812 spin_unlock(&EXT4_SB(sb
)->s_md_lock
);
2815 /* there are blocks to put in buddy to make them really free */
2816 count
+= entry
->efd_count
;
2818 ext4_lock_group(sb
, entry
->efd_group
);
2819 /* Take it out of per group rb tree */
2820 rb_erase(&entry
->efd_node
, &(db
->bb_free_root
));
2821 mb_free_blocks(NULL
, &e4b
, entry
->efd_start_cluster
, entry
->efd_count
);
2824 * Clear the trimmed flag for the group so that the next
2825 * ext4_trim_fs can trim it.
2826 * If the volume is mounted with -o discard, online discard
2827 * is supported and the free blocks will be trimmed online.
2829 if (!test_opt(sb
, DISCARD
))
2830 EXT4_MB_GRP_CLEAR_TRIMMED(db
);
2832 if (!db
->bb_free_root
.rb_node
) {
2833 /* No more items in the per group rb tree
2834 * balance refcounts from ext4_mb_free_metadata()
2836 put_page(e4b
.bd_buddy_page
);
2837 put_page(e4b
.bd_bitmap_page
);
2839 ext4_unlock_group(sb
, entry
->efd_group
);
2840 kmem_cache_free(ext4_free_data_cachep
, entry
);
2841 ext4_mb_unload_buddy(&e4b
);
2843 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2847 * This function is called by the jbd2 layer once the commit has finished,
2848 * so we know we can free the blocks that were released with that commit.
2850 void ext4_process_freed_data(struct super_block
*sb
, tid_t commit_tid
)
2852 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2853 struct ext4_free_data
*entry
, *tmp
;
2854 struct bio
*discard_bio
= NULL
;
2855 struct list_head freed_data_list
;
2856 struct list_head
*cut_pos
= NULL
;
2859 INIT_LIST_HEAD(&freed_data_list
);
2861 spin_lock(&sbi
->s_md_lock
);
2862 list_for_each_entry(entry
, &sbi
->s_freed_data_list
, efd_list
) {
2863 if (entry
->efd_tid
!= commit_tid
)
2865 cut_pos
= &entry
->efd_list
;
2868 list_cut_position(&freed_data_list
, &sbi
->s_freed_data_list
,
2870 spin_unlock(&sbi
->s_md_lock
);
2872 if (test_opt(sb
, DISCARD
)) {
2873 list_for_each_entry(entry
, &freed_data_list
, efd_list
) {
2874 err
= ext4_issue_discard(sb
, entry
->efd_group
,
2875 entry
->efd_start_cluster
,
2878 if (err
&& err
!= -EOPNOTSUPP
) {
2879 ext4_msg(sb
, KERN_WARNING
, "discard request in"
2880 " group:%d block:%d count:%d failed"
2881 " with %d", entry
->efd_group
,
2882 entry
->efd_start_cluster
,
2883 entry
->efd_count
, err
);
2884 } else if (err
== -EOPNOTSUPP
)
2889 submit_bio_wait(discard_bio
);
2890 bio_put(discard_bio
);
2894 list_for_each_entry_safe(entry
, tmp
, &freed_data_list
, efd_list
)
2895 ext4_free_data_in_buddy(sb
, entry
);
2898 int __init
ext4_init_mballoc(void)
2900 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2901 SLAB_RECLAIM_ACCOUNT
);
2902 if (ext4_pspace_cachep
== NULL
)
2905 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2906 SLAB_RECLAIM_ACCOUNT
);
2907 if (ext4_ac_cachep
== NULL
) {
2908 kmem_cache_destroy(ext4_pspace_cachep
);
2912 ext4_free_data_cachep
= KMEM_CACHE(ext4_free_data
,
2913 SLAB_RECLAIM_ACCOUNT
);
2914 if (ext4_free_data_cachep
== NULL
) {
2915 kmem_cache_destroy(ext4_pspace_cachep
);
2916 kmem_cache_destroy(ext4_ac_cachep
);
2922 void ext4_exit_mballoc(void)
2925 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2926 * before destroying the slab cache.
2929 kmem_cache_destroy(ext4_pspace_cachep
);
2930 kmem_cache_destroy(ext4_ac_cachep
);
2931 kmem_cache_destroy(ext4_free_data_cachep
);
2932 ext4_groupinfo_destroy_slabs();
2937 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2938 * Returns 0 if success or error code
2940 static noinline_for_stack
int
2941 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2942 handle_t
*handle
, unsigned int reserv_clstrs
)
2944 struct buffer_head
*bitmap_bh
= NULL
;
2945 struct ext4_group_desc
*gdp
;
2946 struct buffer_head
*gdp_bh
;
2947 struct ext4_sb_info
*sbi
;
2948 struct super_block
*sb
;
2952 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2953 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2958 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2959 if (IS_ERR(bitmap_bh
)) {
2960 err
= PTR_ERR(bitmap_bh
);
2965 BUFFER_TRACE(bitmap_bh
, "getting write access");
2966 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2971 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2975 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2976 ext4_free_group_clusters(sb
, gdp
));
2978 BUFFER_TRACE(gdp_bh
, "get_write_access");
2979 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2983 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
2985 len
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
2986 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2987 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2988 "fs metadata", block
, block
+len
);
2989 /* File system mounted not to panic on error
2990 * Fix the bitmap and return EFSCORRUPTED
2991 * We leak some of the blocks here.
2993 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2994 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2995 ac
->ac_b_ex
.fe_len
);
2996 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2997 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2999 err
= -EFSCORRUPTED
;
3003 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3004 #ifdef AGGRESSIVE_CHECK
3007 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
3008 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
3009 bitmap_bh
->b_data
));
3013 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
3014 ac
->ac_b_ex
.fe_len
);
3015 if (ext4_has_group_desc_csum(sb
) &&
3016 (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))) {
3017 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
3018 ext4_free_group_clusters_set(sb
, gdp
,
3019 ext4_free_clusters_after_init(sb
,
3020 ac
->ac_b_ex
.fe_group
, gdp
));
3022 len
= ext4_free_group_clusters(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
3023 ext4_free_group_clusters_set(sb
, gdp
, len
);
3024 ext4_block_bitmap_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
, bitmap_bh
);
3025 ext4_group_desc_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
);
3027 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3028 percpu_counter_sub(&sbi
->s_freeclusters_counter
, ac
->ac_b_ex
.fe_len
);
3030 * Now reduce the dirty block count also. Should not go negative
3032 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
3033 /* release all the reserved blocks if non delalloc */
3034 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
3037 if (sbi
->s_log_groups_per_flex
) {
3038 ext4_group_t flex_group
= ext4_flex_group(sbi
,
3039 ac
->ac_b_ex
.fe_group
);
3040 atomic64_sub(ac
->ac_b_ex
.fe_len
,
3041 &sbi_array_rcu_deref(sbi
, s_flex_groups
,
3042 flex_group
)->free_clusters
);
3045 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3048 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
3056 * here we normalize request for locality group
3057 * Group request are normalized to s_mb_group_prealloc, which goes to
3058 * s_strip if we set the same via mount option.
3059 * s_mb_group_prealloc can be configured via
3060 * /sys/fs/ext4/<partition>/mb_group_prealloc
3062 * XXX: should we try to preallocate more than the group has now?
3064 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
3066 struct super_block
*sb
= ac
->ac_sb
;
3067 struct ext4_locality_group
*lg
= ac
->ac_lg
;
3070 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
3071 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3072 current
->pid
, ac
->ac_g_ex
.fe_len
);
3076 * Normalization means making request better in terms of
3077 * size and alignment
3079 static noinline_for_stack
void
3080 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
3081 struct ext4_allocation_request
*ar
)
3083 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3086 loff_t size
, start_off
;
3087 loff_t orig_size __maybe_unused
;
3089 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3090 struct ext4_prealloc_space
*pa
;
3092 /* do normalize only data requests, metadata requests
3093 do not need preallocation */
3094 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3097 /* sometime caller may want exact blocks */
3098 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3101 /* caller may indicate that preallocation isn't
3102 * required (it's a tail, for example) */
3103 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3106 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3107 ext4_mb_normalize_group_request(ac
);
3111 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3113 /* first, let's learn actual file size
3114 * given current request is allocated */
3115 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
3116 size
= size
<< bsbits
;
3117 if (size
< i_size_read(ac
->ac_inode
))
3118 size
= i_size_read(ac
->ac_inode
);
3121 /* max size of free chunks */
3124 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3125 (req <= (size) || max <= (chunk_size))
3127 /* first, try to predict filesize */
3128 /* XXX: should this table be tunable? */
3130 if (size
<= 16 * 1024) {
3132 } else if (size
<= 32 * 1024) {
3134 } else if (size
<= 64 * 1024) {
3136 } else if (size
<= 128 * 1024) {
3138 } else if (size
<= 256 * 1024) {
3140 } else if (size
<= 512 * 1024) {
3142 } else if (size
<= 1024 * 1024) {
3144 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3145 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3146 (21 - bsbits
)) << 21;
3147 size
= 2 * 1024 * 1024;
3148 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3149 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3150 (22 - bsbits
)) << 22;
3151 size
= 4 * 1024 * 1024;
3152 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3153 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3154 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3155 (23 - bsbits
)) << 23;
3156 size
= 8 * 1024 * 1024;
3158 start_off
= (loff_t
) ac
->ac_o_ex
.fe_logical
<< bsbits
;
3159 size
= (loff_t
) EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3160 ac
->ac_o_ex
.fe_len
) << bsbits
;
3162 size
= size
>> bsbits
;
3163 start
= start_off
>> bsbits
;
3165 /* don't cover already allocated blocks in selected range */
3166 if (ar
->pleft
&& start
<= ar
->lleft
) {
3167 size
-= ar
->lleft
+ 1 - start
;
3168 start
= ar
->lleft
+ 1;
3170 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3171 size
-= start
+ size
- ar
->lright
;
3174 * Trim allocation request for filesystems with artificially small
3177 if (size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
))
3178 size
= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
);
3182 /* check we don't cross already preallocated blocks */
3184 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3189 spin_lock(&pa
->pa_lock
);
3190 if (pa
->pa_deleted
) {
3191 spin_unlock(&pa
->pa_lock
);
3195 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3198 /* PA must not overlap original request */
3199 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3200 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3202 /* skip PAs this normalized request doesn't overlap with */
3203 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
3204 spin_unlock(&pa
->pa_lock
);
3207 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3209 /* adjust start or end to be adjacent to this pa */
3210 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3211 BUG_ON(pa_end
< start
);
3213 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3214 BUG_ON(pa
->pa_lstart
> end
);
3215 end
= pa
->pa_lstart
;
3217 spin_unlock(&pa
->pa_lock
);
3222 /* XXX: extra loop to check we really don't overlap preallocations */
3224 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3227 spin_lock(&pa
->pa_lock
);
3228 if (pa
->pa_deleted
== 0) {
3229 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3231 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3233 spin_unlock(&pa
->pa_lock
);
3237 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3238 start
> ac
->ac_o_ex
.fe_logical
) {
3239 ext4_msg(ac
->ac_sb
, KERN_ERR
,
3240 "start %lu, size %lu, fe_logical %lu",
3241 (unsigned long) start
, (unsigned long) size
,
3242 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3245 BUG_ON(size
<= 0 || size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3247 /* now prepare goal request */
3249 /* XXX: is it better to align blocks WRT to logical
3250 * placement or satisfy big request as is */
3251 ac
->ac_g_ex
.fe_logical
= start
;
3252 ac
->ac_g_ex
.fe_len
= EXT4_NUM_B2C(sbi
, size
);
3254 /* define goal start in order to merge */
3255 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3256 /* merge to the right */
3257 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3258 &ac
->ac_f_ex
.fe_group
,
3259 &ac
->ac_f_ex
.fe_start
);
3260 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3262 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3263 /* merge to the left */
3264 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3265 &ac
->ac_f_ex
.fe_group
,
3266 &ac
->ac_f_ex
.fe_start
);
3267 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3270 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3271 (unsigned) orig_size
, (unsigned) start
);
3274 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3276 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3278 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3279 atomic_inc(&sbi
->s_bal_reqs
);
3280 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3281 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3282 atomic_inc(&sbi
->s_bal_success
);
3283 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3284 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3285 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3286 atomic_inc(&sbi
->s_bal_goals
);
3287 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3288 atomic_inc(&sbi
->s_bal_breaks
);
3291 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3292 trace_ext4_mballoc_alloc(ac
);
3294 trace_ext4_mballoc_prealloc(ac
);
3298 * Called on failure; free up any blocks from the inode PA for this
3299 * context. We don't need this for MB_GROUP_PA because we only change
3300 * pa_free in ext4_mb_release_context(), but on failure, we've already
3301 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3303 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3305 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3306 struct ext4_buddy e4b
;
3310 if (ac
->ac_f_ex
.fe_len
== 0)
3312 err
= ext4_mb_load_buddy(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
, &e4b
);
3315 * This should never happen since we pin the
3316 * pages in the ext4_allocation_context so
3317 * ext4_mb_load_buddy() should never fail.
3319 WARN(1, "mb_load_buddy failed (%d)", err
);
3322 ext4_lock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3323 mb_free_blocks(ac
->ac_inode
, &e4b
, ac
->ac_f_ex
.fe_start
,
3324 ac
->ac_f_ex
.fe_len
);
3325 ext4_unlock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3326 ext4_mb_unload_buddy(&e4b
);
3329 if (pa
->pa_type
== MB_INODE_PA
)
3330 pa
->pa_free
+= ac
->ac_b_ex
.fe_len
;
3334 * use blocks preallocated to inode
3336 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3337 struct ext4_prealloc_space
*pa
)
3339 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3344 /* found preallocated blocks, use them */
3345 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3346 end
= min(pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
),
3347 start
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
));
3348 len
= EXT4_NUM_B2C(sbi
, end
- start
);
3349 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3350 &ac
->ac_b_ex
.fe_start
);
3351 ac
->ac_b_ex
.fe_len
= len
;
3352 ac
->ac_status
= AC_STATUS_FOUND
;
3355 BUG_ON(start
< pa
->pa_pstart
);
3356 BUG_ON(end
> pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
));
3357 BUG_ON(pa
->pa_free
< len
);
3360 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3364 * use blocks preallocated to locality group
3366 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3367 struct ext4_prealloc_space
*pa
)
3369 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3371 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3372 &ac
->ac_b_ex
.fe_group
,
3373 &ac
->ac_b_ex
.fe_start
);
3374 ac
->ac_b_ex
.fe_len
= len
;
3375 ac
->ac_status
= AC_STATUS_FOUND
;
3378 /* we don't correct pa_pstart or pa_plen here to avoid
3379 * possible race when the group is being loaded concurrently
3380 * instead we correct pa later, after blocks are marked
3381 * in on-disk bitmap -- see ext4_mb_release_context()
3382 * Other CPUs are prevented from allocating from this pa by lg_mutex
3384 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3388 * Return the prealloc space that have minimal distance
3389 * from the goal block. @cpa is the prealloc
3390 * space that is having currently known minimal distance
3391 * from the goal block.
3393 static struct ext4_prealloc_space
*
3394 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3395 struct ext4_prealloc_space
*pa
,
3396 struct ext4_prealloc_space
*cpa
)
3398 ext4_fsblk_t cur_distance
, new_distance
;
3401 atomic_inc(&pa
->pa_count
);
3404 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3405 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3407 if (cur_distance
<= new_distance
)
3410 /* drop the previous reference */
3411 atomic_dec(&cpa
->pa_count
);
3412 atomic_inc(&pa
->pa_count
);
3417 * search goal blocks in preallocated space
3419 static noinline_for_stack
int
3420 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3422 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3424 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3425 struct ext4_locality_group
*lg
;
3426 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3427 ext4_fsblk_t goal_block
;
3429 /* only data can be preallocated */
3430 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3433 /* first, try per-file preallocation */
3435 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3437 /* all fields in this condition don't change,
3438 * so we can skip locking for them */
3439 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3440 ac
->ac_o_ex
.fe_logical
>= (pa
->pa_lstart
+
3441 EXT4_C2B(sbi
, pa
->pa_len
)))
3444 /* non-extent files can't have physical blocks past 2^32 */
3445 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3446 (pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
) >
3447 EXT4_MAX_BLOCK_FILE_PHYS
))
3450 /* found preallocated blocks, use them */
3451 spin_lock(&pa
->pa_lock
);
3452 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3453 atomic_inc(&pa
->pa_count
);
3454 ext4_mb_use_inode_pa(ac
, pa
);
3455 spin_unlock(&pa
->pa_lock
);
3456 ac
->ac_criteria
= 10;
3460 spin_unlock(&pa
->pa_lock
);
3464 /* can we use group allocation? */
3465 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3468 /* inode may have no locality group for some reason */
3472 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3473 if (order
> PREALLOC_TB_SIZE
- 1)
3474 /* The max size of hash table is PREALLOC_TB_SIZE */
3475 order
= PREALLOC_TB_SIZE
- 1;
3477 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3479 * search for the prealloc space that is having
3480 * minimal distance from the goal block.
3482 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3484 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3486 spin_lock(&pa
->pa_lock
);
3487 if (pa
->pa_deleted
== 0 &&
3488 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3490 cpa
= ext4_mb_check_group_pa(goal_block
,
3493 spin_unlock(&pa
->pa_lock
);
3498 ext4_mb_use_group_pa(ac
, cpa
);
3499 ac
->ac_criteria
= 20;
3506 * the function goes through all block freed in the group
3507 * but not yet committed and marks them used in in-core bitmap.
3508 * buddy must be generated from this bitmap
3509 * Need to be called with the ext4 group lock held
3511 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3515 struct ext4_group_info
*grp
;
3516 struct ext4_free_data
*entry
;
3518 grp
= ext4_get_group_info(sb
, group
);
3519 n
= rb_first(&(grp
->bb_free_root
));
3522 entry
= rb_entry(n
, struct ext4_free_data
, efd_node
);
3523 ext4_set_bits(bitmap
, entry
->efd_start_cluster
, entry
->efd_count
);
3530 * the function goes through all preallocation in this group and marks them
3531 * used in in-core bitmap. buddy must be generated from this bitmap
3532 * Need to be called with ext4 group lock held
3534 static noinline_for_stack
3535 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3538 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3539 struct ext4_prealloc_space
*pa
;
3540 struct list_head
*cur
;
3541 ext4_group_t groupnr
;
3542 ext4_grpblk_t start
;
3543 int preallocated
= 0;
3546 /* all form of preallocation discards first load group,
3547 * so the only competing code is preallocation use.
3548 * we don't need any locking here
3549 * notice we do NOT ignore preallocations with pa_deleted
3550 * otherwise we could leave used blocks available for
3551 * allocation in buddy when concurrent ext4_mb_put_pa()
3552 * is dropping preallocation
3554 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3555 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3556 spin_lock(&pa
->pa_lock
);
3557 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3560 spin_unlock(&pa
->pa_lock
);
3561 if (unlikely(len
== 0))
3563 BUG_ON(groupnr
!= group
);
3564 ext4_set_bits(bitmap
, start
, len
);
3565 preallocated
+= len
;
3567 mb_debug(1, "preallocated %u for group %u\n", preallocated
, group
);
3570 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3572 struct ext4_prealloc_space
*pa
;
3573 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3575 BUG_ON(atomic_read(&pa
->pa_count
));
3576 BUG_ON(pa
->pa_deleted
== 0);
3577 kmem_cache_free(ext4_pspace_cachep
, pa
);
3581 * drops a reference to preallocated space descriptor
3582 * if this was the last reference and the space is consumed
3584 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3585 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3588 ext4_fsblk_t grp_blk
;
3590 /* in this short window concurrent discard can set pa_deleted */
3591 spin_lock(&pa
->pa_lock
);
3592 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0) {
3593 spin_unlock(&pa
->pa_lock
);
3597 if (pa
->pa_deleted
== 1) {
3598 spin_unlock(&pa
->pa_lock
);
3603 spin_unlock(&pa
->pa_lock
);
3605 grp_blk
= pa
->pa_pstart
;
3607 * If doing group-based preallocation, pa_pstart may be in the
3608 * next group when pa is used up
3610 if (pa
->pa_type
== MB_GROUP_PA
)
3613 grp
= ext4_get_group_number(sb
, grp_blk
);
3618 * P1 (buddy init) P2 (regular allocation)
3619 * find block B in PA
3620 * copy on-disk bitmap to buddy
3621 * mark B in on-disk bitmap
3622 * drop PA from group
3623 * mark all PAs in buddy
3625 * thus, P1 initializes buddy with B available. to prevent this
3626 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3629 ext4_lock_group(sb
, grp
);
3630 list_del(&pa
->pa_group_list
);
3631 ext4_unlock_group(sb
, grp
);
3633 spin_lock(pa
->pa_obj_lock
);
3634 list_del_rcu(&pa
->pa_inode_list
);
3635 spin_unlock(pa
->pa_obj_lock
);
3637 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3641 * creates new preallocated space for given inode
3643 static noinline_for_stack
int
3644 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3646 struct super_block
*sb
= ac
->ac_sb
;
3647 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3648 struct ext4_prealloc_space
*pa
;
3649 struct ext4_group_info
*grp
;
3650 struct ext4_inode_info
*ei
;
3652 /* preallocate only when found space is larger then requested */
3653 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3654 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3655 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3657 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3661 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3667 /* we can't allocate as much as normalizer wants.
3668 * so, found space must get proper lstart
3669 * to cover original request */
3670 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3671 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3673 /* we're limited by original request in that
3674 * logical block must be covered any way
3675 * winl is window we can move our chunk within */
3676 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3678 /* also, we should cover whole original request */
3679 wins
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
);
3681 /* the smallest one defines real window */
3682 win
= min(winl
, wins
);
3684 offs
= ac
->ac_o_ex
.fe_logical
%
3685 EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
3686 if (offs
&& offs
< win
)
3689 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
-
3690 EXT4_NUM_B2C(sbi
, win
);
3691 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3692 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3695 /* preallocation can change ac_b_ex, thus we store actually
3696 * allocated blocks for history */
3697 ac
->ac_f_ex
= ac
->ac_b_ex
;
3699 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3700 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3701 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3702 pa
->pa_free
= pa
->pa_len
;
3703 atomic_set(&pa
->pa_count
, 1);
3704 spin_lock_init(&pa
->pa_lock
);
3705 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3706 INIT_LIST_HEAD(&pa
->pa_group_list
);
3708 pa
->pa_type
= MB_INODE_PA
;
3710 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3711 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3712 trace_ext4_mb_new_inode_pa(ac
, pa
);
3714 ext4_mb_use_inode_pa(ac
, pa
);
3715 atomic_add(pa
->pa_free
, &sbi
->s_mb_preallocated
);
3717 ei
= EXT4_I(ac
->ac_inode
);
3718 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3720 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3721 pa
->pa_inode
= ac
->ac_inode
;
3723 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3724 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3725 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3727 spin_lock(pa
->pa_obj_lock
);
3728 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3729 spin_unlock(pa
->pa_obj_lock
);
3735 * creates new preallocated space for locality group inodes belongs to
3737 static noinline_for_stack
int
3738 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3740 struct super_block
*sb
= ac
->ac_sb
;
3741 struct ext4_locality_group
*lg
;
3742 struct ext4_prealloc_space
*pa
;
3743 struct ext4_group_info
*grp
;
3745 /* preallocate only when found space is larger then requested */
3746 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3747 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3748 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3750 BUG_ON(ext4_pspace_cachep
== NULL
);
3751 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3755 /* preallocation can change ac_b_ex, thus we store actually
3756 * allocated blocks for history */
3757 ac
->ac_f_ex
= ac
->ac_b_ex
;
3759 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3760 pa
->pa_lstart
= pa
->pa_pstart
;
3761 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3762 pa
->pa_free
= pa
->pa_len
;
3763 atomic_set(&pa
->pa_count
, 1);
3764 spin_lock_init(&pa
->pa_lock
);
3765 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3766 INIT_LIST_HEAD(&pa
->pa_group_list
);
3768 pa
->pa_type
= MB_GROUP_PA
;
3770 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3771 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3772 trace_ext4_mb_new_group_pa(ac
, pa
);
3774 ext4_mb_use_group_pa(ac
, pa
);
3775 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3777 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3781 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3782 pa
->pa_inode
= NULL
;
3784 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3785 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3786 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3789 * We will later add the new pa to the right bucket
3790 * after updating the pa_free in ext4_mb_release_context
3795 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3799 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3800 err
= ext4_mb_new_group_pa(ac
);
3802 err
= ext4_mb_new_inode_pa(ac
);
3807 * finds all unused blocks in on-disk bitmap, frees them in
3808 * in-core bitmap and buddy.
3809 * @pa must be unlinked from inode and group lists, so that
3810 * nobody else can find/use it.
3811 * the caller MUST hold group/inode locks.
3812 * TODO: optimize the case when there are no in-core structures yet
3814 static noinline_for_stack
int
3815 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3816 struct ext4_prealloc_space
*pa
)
3818 struct super_block
*sb
= e4b
->bd_sb
;
3819 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3824 unsigned long long grp_blk_start
;
3827 BUG_ON(pa
->pa_deleted
== 0);
3828 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3829 grp_blk_start
= pa
->pa_pstart
- EXT4_C2B(sbi
, bit
);
3830 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3831 end
= bit
+ pa
->pa_len
;
3834 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3837 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3838 mb_debug(1, " free preallocated %u/%u in group %u\n",
3839 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3840 (unsigned) next
- bit
, (unsigned) group
);
3843 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3844 trace_ext4_mb_release_inode_pa(pa
, (grp_blk_start
+
3845 EXT4_C2B(sbi
, bit
)),
3847 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3850 if (free
!= pa
->pa_free
) {
3851 ext4_msg(e4b
->bd_sb
, KERN_CRIT
,
3852 "pa %p: logic %lu, phys. %lu, len %lu",
3853 pa
, (unsigned long) pa
->pa_lstart
,
3854 (unsigned long) pa
->pa_pstart
,
3855 (unsigned long) pa
->pa_len
);
3856 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3859 * pa is already deleted so we use the value obtained
3860 * from the bitmap and continue.
3863 atomic_add(free
, &sbi
->s_mb_discarded
);
3868 static noinline_for_stack
int
3869 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3870 struct ext4_prealloc_space
*pa
)
3872 struct super_block
*sb
= e4b
->bd_sb
;
3876 trace_ext4_mb_release_group_pa(sb
, pa
);
3877 BUG_ON(pa
->pa_deleted
== 0);
3878 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3879 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3880 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3881 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3882 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3888 * releases all preallocations in given group
3890 * first, we need to decide discard policy:
3891 * - when do we discard
3893 * - how many do we discard
3894 * 1) how many requested
3896 static noinline_for_stack
int
3897 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3898 ext4_group_t group
, int needed
)
3900 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3901 struct buffer_head
*bitmap_bh
= NULL
;
3902 struct ext4_prealloc_space
*pa
, *tmp
;
3903 struct list_head list
;
3904 struct ext4_buddy e4b
;
3909 mb_debug(1, "discard preallocation for group %u\n", group
);
3911 if (list_empty(&grp
->bb_prealloc_list
))
3914 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3915 if (IS_ERR(bitmap_bh
)) {
3916 err
= PTR_ERR(bitmap_bh
);
3917 ext4_set_errno(sb
, -err
);
3918 ext4_error(sb
, "Error %d reading block bitmap for %u",
3923 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3925 ext4_warning(sb
, "Error %d loading buddy information for %u",
3932 needed
= EXT4_CLUSTERS_PER_GROUP(sb
) + 1;
3934 INIT_LIST_HEAD(&list
);
3936 ext4_lock_group(sb
, group
);
3937 list_for_each_entry_safe(pa
, tmp
,
3938 &grp
->bb_prealloc_list
, pa_group_list
) {
3939 spin_lock(&pa
->pa_lock
);
3940 if (atomic_read(&pa
->pa_count
)) {
3941 spin_unlock(&pa
->pa_lock
);
3945 if (pa
->pa_deleted
) {
3946 spin_unlock(&pa
->pa_lock
);
3950 /* seems this one can be freed ... */
3953 /* we can trust pa_free ... */
3954 free
+= pa
->pa_free
;
3956 spin_unlock(&pa
->pa_lock
);
3958 list_del(&pa
->pa_group_list
);
3959 list_add(&pa
->u
.pa_tmp_list
, &list
);
3962 /* if we still need more blocks and some PAs were used, try again */
3963 if (free
< needed
&& busy
) {
3965 ext4_unlock_group(sb
, group
);
3970 /* found anything to free? */
3971 if (list_empty(&list
)) {
3976 /* now free all selected PAs */
3977 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3979 /* remove from object (inode or locality group) */
3980 spin_lock(pa
->pa_obj_lock
);
3981 list_del_rcu(&pa
->pa_inode_list
);
3982 spin_unlock(pa
->pa_obj_lock
);
3984 if (pa
->pa_type
== MB_GROUP_PA
)
3985 ext4_mb_release_group_pa(&e4b
, pa
);
3987 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3989 list_del(&pa
->u
.pa_tmp_list
);
3990 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3994 ext4_unlock_group(sb
, group
);
3995 ext4_mb_unload_buddy(&e4b
);
4001 * releases all non-used preallocated blocks for given inode
4003 * It's important to discard preallocations under i_data_sem
4004 * We don't want another block to be served from the prealloc
4005 * space when we are discarding the inode prealloc space.
4007 * FIXME!! Make sure it is valid at all the call sites
4009 void ext4_discard_preallocations(struct inode
*inode
)
4011 struct ext4_inode_info
*ei
= EXT4_I(inode
);
4012 struct super_block
*sb
= inode
->i_sb
;
4013 struct buffer_head
*bitmap_bh
= NULL
;
4014 struct ext4_prealloc_space
*pa
, *tmp
;
4015 ext4_group_t group
= 0;
4016 struct list_head list
;
4017 struct ext4_buddy e4b
;
4020 if (!S_ISREG(inode
->i_mode
)) {
4021 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4025 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
4026 trace_ext4_discard_preallocations(inode
);
4028 INIT_LIST_HEAD(&list
);
4031 /* first, collect all pa's in the inode */
4032 spin_lock(&ei
->i_prealloc_lock
);
4033 while (!list_empty(&ei
->i_prealloc_list
)) {
4034 pa
= list_entry(ei
->i_prealloc_list
.next
,
4035 struct ext4_prealloc_space
, pa_inode_list
);
4036 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
4037 spin_lock(&pa
->pa_lock
);
4038 if (atomic_read(&pa
->pa_count
)) {
4039 /* this shouldn't happen often - nobody should
4040 * use preallocation while we're discarding it */
4041 spin_unlock(&pa
->pa_lock
);
4042 spin_unlock(&ei
->i_prealloc_lock
);
4043 ext4_msg(sb
, KERN_ERR
,
4044 "uh-oh! used pa while discarding");
4046 schedule_timeout_uninterruptible(HZ
);
4050 if (pa
->pa_deleted
== 0) {
4052 spin_unlock(&pa
->pa_lock
);
4053 list_del_rcu(&pa
->pa_inode_list
);
4054 list_add(&pa
->u
.pa_tmp_list
, &list
);
4058 /* someone is deleting pa right now */
4059 spin_unlock(&pa
->pa_lock
);
4060 spin_unlock(&ei
->i_prealloc_lock
);
4062 /* we have to wait here because pa_deleted
4063 * doesn't mean pa is already unlinked from
4064 * the list. as we might be called from
4065 * ->clear_inode() the inode will get freed
4066 * and concurrent thread which is unlinking
4067 * pa from inode's list may access already
4068 * freed memory, bad-bad-bad */
4070 /* XXX: if this happens too often, we can
4071 * add a flag to force wait only in case
4072 * of ->clear_inode(), but not in case of
4073 * regular truncate */
4074 schedule_timeout_uninterruptible(HZ
);
4077 spin_unlock(&ei
->i_prealloc_lock
);
4079 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
4080 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
4081 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4083 err
= ext4_mb_load_buddy_gfp(sb
, group
, &e4b
,
4084 GFP_NOFS
|__GFP_NOFAIL
);
4086 ext4_set_errno(sb
, -err
);
4087 ext4_error(sb
, "Error %d loading buddy information for %u",
4092 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
4093 if (IS_ERR(bitmap_bh
)) {
4094 err
= PTR_ERR(bitmap_bh
);
4095 ext4_set_errno(sb
, -err
);
4096 ext4_error(sb
, "Error %d reading block bitmap for %u",
4098 ext4_mb_unload_buddy(&e4b
);
4102 ext4_lock_group(sb
, group
);
4103 list_del(&pa
->pa_group_list
);
4104 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
4105 ext4_unlock_group(sb
, group
);
4107 ext4_mb_unload_buddy(&e4b
);
4110 list_del(&pa
->u
.pa_tmp_list
);
4111 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4115 #ifdef CONFIG_EXT4_DEBUG
4116 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4118 struct super_block
*sb
= ac
->ac_sb
;
4119 ext4_group_t ngroups
, i
;
4121 if (!ext4_mballoc_debug
||
4122 (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
))
4125 ext4_msg(ac
->ac_sb
, KERN_ERR
, "Can't allocate:"
4126 " Allocation context details:");
4127 ext4_msg(ac
->ac_sb
, KERN_ERR
, "status %d flags %d",
4128 ac
->ac_status
, ac
->ac_flags
);
4129 ext4_msg(ac
->ac_sb
, KERN_ERR
, "orig %lu/%lu/%lu@%lu, "
4130 "goal %lu/%lu/%lu@%lu, "
4131 "best %lu/%lu/%lu@%lu cr %d",
4132 (unsigned long)ac
->ac_o_ex
.fe_group
,
4133 (unsigned long)ac
->ac_o_ex
.fe_start
,
4134 (unsigned long)ac
->ac_o_ex
.fe_len
,
4135 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4136 (unsigned long)ac
->ac_g_ex
.fe_group
,
4137 (unsigned long)ac
->ac_g_ex
.fe_start
,
4138 (unsigned long)ac
->ac_g_ex
.fe_len
,
4139 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4140 (unsigned long)ac
->ac_b_ex
.fe_group
,
4141 (unsigned long)ac
->ac_b_ex
.fe_start
,
4142 (unsigned long)ac
->ac_b_ex
.fe_len
,
4143 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4144 (int)ac
->ac_criteria
);
4145 ext4_msg(ac
->ac_sb
, KERN_ERR
, "%d found", ac
->ac_found
);
4146 ext4_msg(ac
->ac_sb
, KERN_ERR
, "groups: ");
4147 ngroups
= ext4_get_groups_count(sb
);
4148 for (i
= 0; i
< ngroups
; i
++) {
4149 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4150 struct ext4_prealloc_space
*pa
;
4151 ext4_grpblk_t start
;
4152 struct list_head
*cur
;
4153 ext4_lock_group(sb
, i
);
4154 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4155 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4157 spin_lock(&pa
->pa_lock
);
4158 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4160 spin_unlock(&pa
->pa_lock
);
4161 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
4164 ext4_unlock_group(sb
, i
);
4166 if (grp
->bb_free
== 0)
4168 printk(KERN_ERR
"%u: %d/%d \n",
4169 i
, grp
->bb_free
, grp
->bb_fragments
);
4171 printk(KERN_ERR
"\n");
4174 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4181 * We use locality group preallocation for small size file. The size of the
4182 * file is determined by the current size or the resulting size after
4183 * allocation which ever is larger
4185 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4187 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4189 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4190 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4193 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4196 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4199 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
4200 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
4203 if ((size
== isize
) && !ext4_fs_is_busy(sbi
) &&
4204 !inode_is_open_for_write(ac
->ac_inode
)) {
4205 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
4209 if (sbi
->s_mb_group_prealloc
<= 0) {
4210 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4214 /* don't use group allocation for large files */
4215 size
= max(size
, isize
);
4216 if (size
> sbi
->s_mb_stream_request
) {
4217 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4221 BUG_ON(ac
->ac_lg
!= NULL
);
4223 * locality group prealloc space are per cpu. The reason for having
4224 * per cpu locality group is to reduce the contention between block
4225 * request from multiple CPUs.
4227 ac
->ac_lg
= raw_cpu_ptr(sbi
->s_locality_groups
);
4229 /* we're going to use group allocation */
4230 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4232 /* serialize all allocations in the group */
4233 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4236 static noinline_for_stack
int
4237 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4238 struct ext4_allocation_request
*ar
)
4240 struct super_block
*sb
= ar
->inode
->i_sb
;
4241 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4242 struct ext4_super_block
*es
= sbi
->s_es
;
4246 ext4_grpblk_t block
;
4248 /* we can't allocate > group size */
4251 /* just a dirty hack to filter too big requests */
4252 if (len
>= EXT4_CLUSTERS_PER_GROUP(sb
))
4253 len
= EXT4_CLUSTERS_PER_GROUP(sb
);
4255 /* start searching from the goal */
4257 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4258 goal
>= ext4_blocks_count(es
))
4259 goal
= le32_to_cpu(es
->s_first_data_block
);
4260 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4262 /* set up allocation goals */
4263 ac
->ac_b_ex
.fe_logical
= EXT4_LBLK_CMASK(sbi
, ar
->logical
);
4264 ac
->ac_status
= AC_STATUS_CONTINUE
;
4266 ac
->ac_inode
= ar
->inode
;
4267 ac
->ac_o_ex
.fe_logical
= ac
->ac_b_ex
.fe_logical
;
4268 ac
->ac_o_ex
.fe_group
= group
;
4269 ac
->ac_o_ex
.fe_start
= block
;
4270 ac
->ac_o_ex
.fe_len
= len
;
4271 ac
->ac_g_ex
= ac
->ac_o_ex
;
4272 ac
->ac_flags
= ar
->flags
;
4274 /* we have to define context: we'll we work with a file or
4275 * locality group. this is a policy, actually */
4276 ext4_mb_group_or_file(ac
);
4278 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4279 "left: %u/%u, right %u/%u to %swritable\n",
4280 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4281 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4282 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4283 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4284 inode_is_open_for_write(ar
->inode
) ? "" : "non-");
4289 static noinline_for_stack
void
4290 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4291 struct ext4_locality_group
*lg
,
4292 int order
, int total_entries
)
4294 ext4_group_t group
= 0;
4295 struct ext4_buddy e4b
;
4296 struct list_head discard_list
;
4297 struct ext4_prealloc_space
*pa
, *tmp
;
4299 mb_debug(1, "discard locality group preallocation\n");
4301 INIT_LIST_HEAD(&discard_list
);
4303 spin_lock(&lg
->lg_prealloc_lock
);
4304 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4306 spin_lock(&pa
->pa_lock
);
4307 if (atomic_read(&pa
->pa_count
)) {
4309 * This is the pa that we just used
4310 * for block allocation. So don't
4313 spin_unlock(&pa
->pa_lock
);
4316 if (pa
->pa_deleted
) {
4317 spin_unlock(&pa
->pa_lock
);
4320 /* only lg prealloc space */
4321 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4323 /* seems this one can be freed ... */
4325 spin_unlock(&pa
->pa_lock
);
4327 list_del_rcu(&pa
->pa_inode_list
);
4328 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4331 if (total_entries
<= 5) {
4333 * we want to keep only 5 entries
4334 * allowing it to grow to 8. This
4335 * mak sure we don't call discard
4336 * soon for this list.
4341 spin_unlock(&lg
->lg_prealloc_lock
);
4343 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4346 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4347 err
= ext4_mb_load_buddy_gfp(sb
, group
, &e4b
,
4348 GFP_NOFS
|__GFP_NOFAIL
);
4350 ext4_set_errno(sb
, -err
);
4351 ext4_error(sb
, "Error %d loading buddy information for %u",
4355 ext4_lock_group(sb
, group
);
4356 list_del(&pa
->pa_group_list
);
4357 ext4_mb_release_group_pa(&e4b
, pa
);
4358 ext4_unlock_group(sb
, group
);
4360 ext4_mb_unload_buddy(&e4b
);
4361 list_del(&pa
->u
.pa_tmp_list
);
4362 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4367 * We have incremented pa_count. So it cannot be freed at this
4368 * point. Also we hold lg_mutex. So no parallel allocation is
4369 * possible from this lg. That means pa_free cannot be updated.
4371 * A parallel ext4_mb_discard_group_preallocations is possible.
4372 * which can cause the lg_prealloc_list to be updated.
4375 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4377 int order
, added
= 0, lg_prealloc_count
= 1;
4378 struct super_block
*sb
= ac
->ac_sb
;
4379 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4380 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4382 order
= fls(pa
->pa_free
) - 1;
4383 if (order
> PREALLOC_TB_SIZE
- 1)
4384 /* The max size of hash table is PREALLOC_TB_SIZE */
4385 order
= PREALLOC_TB_SIZE
- 1;
4386 /* Add the prealloc space to lg */
4387 spin_lock(&lg
->lg_prealloc_lock
);
4388 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4390 spin_lock(&tmp_pa
->pa_lock
);
4391 if (tmp_pa
->pa_deleted
) {
4392 spin_unlock(&tmp_pa
->pa_lock
);
4395 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4396 /* Add to the tail of the previous entry */
4397 list_add_tail_rcu(&pa
->pa_inode_list
,
4398 &tmp_pa
->pa_inode_list
);
4401 * we want to count the total
4402 * number of entries in the list
4405 spin_unlock(&tmp_pa
->pa_lock
);
4406 lg_prealloc_count
++;
4409 list_add_tail_rcu(&pa
->pa_inode_list
,
4410 &lg
->lg_prealloc_list
[order
]);
4411 spin_unlock(&lg
->lg_prealloc_lock
);
4413 /* Now trim the list to be not more than 8 elements */
4414 if (lg_prealloc_count
> 8) {
4415 ext4_mb_discard_lg_preallocations(sb
, lg
,
4416 order
, lg_prealloc_count
);
4423 * release all resource we used in allocation
4425 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4427 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4428 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4430 if (pa
->pa_type
== MB_GROUP_PA
) {
4431 /* see comment in ext4_mb_use_group_pa() */
4432 spin_lock(&pa
->pa_lock
);
4433 pa
->pa_pstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4434 pa
->pa_lstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4435 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4436 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4437 spin_unlock(&pa
->pa_lock
);
4442 * We want to add the pa to the right bucket.
4443 * Remove it from the list and while adding
4444 * make sure the list to which we are adding
4447 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4448 spin_lock(pa
->pa_obj_lock
);
4449 list_del_rcu(&pa
->pa_inode_list
);
4450 spin_unlock(pa
->pa_obj_lock
);
4451 ext4_mb_add_n_trim(ac
);
4453 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4455 if (ac
->ac_bitmap_page
)
4456 put_page(ac
->ac_bitmap_page
);
4457 if (ac
->ac_buddy_page
)
4458 put_page(ac
->ac_buddy_page
);
4459 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4460 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4461 ext4_mb_collect_stats(ac
);
4465 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4467 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4471 trace_ext4_mb_discard_preallocations(sb
, needed
);
4472 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4473 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4482 * Main entry point into mballoc to allocate blocks
4483 * it tries to use preallocation first, then falls back
4484 * to usual allocation
4486 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4487 struct ext4_allocation_request
*ar
, int *errp
)
4490 struct ext4_allocation_context
*ac
= NULL
;
4491 struct ext4_sb_info
*sbi
;
4492 struct super_block
*sb
;
4493 ext4_fsblk_t block
= 0;
4494 unsigned int inquota
= 0;
4495 unsigned int reserv_clstrs
= 0;
4498 sb
= ar
->inode
->i_sb
;
4501 trace_ext4_request_blocks(ar
);
4503 /* Allow to use superuser reservation for quota file */
4504 if (ext4_is_quota_file(ar
->inode
))
4505 ar
->flags
|= EXT4_MB_USE_ROOT_BLOCKS
;
4507 if ((ar
->flags
& EXT4_MB_DELALLOC_RESERVED
) == 0) {
4508 /* Without delayed allocation we need to verify
4509 * there is enough free blocks to do block allocation
4510 * and verify allocation doesn't exceed the quota limits.
4513 ext4_claim_free_clusters(sbi
, ar
->len
, ar
->flags
)) {
4515 /* let others to free the space */
4517 ar
->len
= ar
->len
>> 1;
4523 reserv_clstrs
= ar
->len
;
4524 if (ar
->flags
& EXT4_MB_USE_ROOT_BLOCKS
) {
4525 dquot_alloc_block_nofail(ar
->inode
,
4526 EXT4_C2B(sbi
, ar
->len
));
4529 dquot_alloc_block(ar
->inode
,
4530 EXT4_C2B(sbi
, ar
->len
))) {
4532 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4543 ac
= kmem_cache_zalloc(ext4_ac_cachep
, GFP_NOFS
);
4550 *errp
= ext4_mb_initialize_context(ac
, ar
);
4556 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4557 if (!ext4_mb_use_preallocated(ac
)) {
4558 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4559 ext4_mb_normalize_request(ac
, ar
);
4561 /* allocate space in core */
4562 *errp
= ext4_mb_regular_allocator(ac
);
4564 goto discard_and_exit
;
4566 /* as we've just preallocated more space than
4567 * user requested originally, we store allocated
4568 * space in a special descriptor */
4569 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4570 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4571 *errp
= ext4_mb_new_preallocation(ac
);
4574 ext4_discard_allocated_blocks(ac
);
4578 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4579 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_clstrs
);
4581 ext4_discard_allocated_blocks(ac
);
4584 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4585 ar
->len
= ac
->ac_b_ex
.fe_len
;
4588 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4596 ac
->ac_b_ex
.fe_len
= 0;
4598 ext4_mb_show_ac(ac
);
4600 ext4_mb_release_context(ac
);
4603 kmem_cache_free(ext4_ac_cachep
, ac
);
4604 if (inquota
&& ar
->len
< inquota
)
4605 dquot_free_block(ar
->inode
, EXT4_C2B(sbi
, inquota
- ar
->len
));
4607 if ((ar
->flags
& EXT4_MB_DELALLOC_RESERVED
) == 0)
4608 /* release all the reserved blocks if non delalloc */
4609 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
4613 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4619 * We can merge two free data extents only if the physical blocks
4620 * are contiguous, AND the extents were freed by the same transaction,
4621 * AND the blocks are associated with the same group.
4623 static void ext4_try_merge_freed_extent(struct ext4_sb_info
*sbi
,
4624 struct ext4_free_data
*entry
,
4625 struct ext4_free_data
*new_entry
,
4626 struct rb_root
*entry_rb_root
)
4628 if ((entry
->efd_tid
!= new_entry
->efd_tid
) ||
4629 (entry
->efd_group
!= new_entry
->efd_group
))
4631 if (entry
->efd_start_cluster
+ entry
->efd_count
==
4632 new_entry
->efd_start_cluster
) {
4633 new_entry
->efd_start_cluster
= entry
->efd_start_cluster
;
4634 new_entry
->efd_count
+= entry
->efd_count
;
4635 } else if (new_entry
->efd_start_cluster
+ new_entry
->efd_count
==
4636 entry
->efd_start_cluster
) {
4637 new_entry
->efd_count
+= entry
->efd_count
;
4640 spin_lock(&sbi
->s_md_lock
);
4641 list_del(&entry
->efd_list
);
4642 spin_unlock(&sbi
->s_md_lock
);
4643 rb_erase(&entry
->efd_node
, entry_rb_root
);
4644 kmem_cache_free(ext4_free_data_cachep
, entry
);
4647 static noinline_for_stack
int
4648 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4649 struct ext4_free_data
*new_entry
)
4651 ext4_group_t group
= e4b
->bd_group
;
4652 ext4_grpblk_t cluster
;
4653 ext4_grpblk_t clusters
= new_entry
->efd_count
;
4654 struct ext4_free_data
*entry
;
4655 struct ext4_group_info
*db
= e4b
->bd_info
;
4656 struct super_block
*sb
= e4b
->bd_sb
;
4657 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4658 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4659 struct rb_node
*parent
= NULL
, *new_node
;
4661 BUG_ON(!ext4_handle_valid(handle
));
4662 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4663 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4665 new_node
= &new_entry
->efd_node
;
4666 cluster
= new_entry
->efd_start_cluster
;
4669 /* first free block exent. We need to
4670 protect buddy cache from being freed,
4671 * otherwise we'll refresh it from
4672 * on-disk bitmap and lose not-yet-available
4674 get_page(e4b
->bd_buddy_page
);
4675 get_page(e4b
->bd_bitmap_page
);
4679 entry
= rb_entry(parent
, struct ext4_free_data
, efd_node
);
4680 if (cluster
< entry
->efd_start_cluster
)
4682 else if (cluster
>= (entry
->efd_start_cluster
+ entry
->efd_count
))
4683 n
= &(*n
)->rb_right
;
4685 ext4_grp_locked_error(sb
, group
, 0,
4686 ext4_group_first_block_no(sb
, group
) +
4687 EXT4_C2B(sbi
, cluster
),
4688 "Block already on to-be-freed list");
4693 rb_link_node(new_node
, parent
, n
);
4694 rb_insert_color(new_node
, &db
->bb_free_root
);
4696 /* Now try to see the extent can be merged to left and right */
4697 node
= rb_prev(new_node
);
4699 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4700 ext4_try_merge_freed_extent(sbi
, entry
, new_entry
,
4701 &(db
->bb_free_root
));
4704 node
= rb_next(new_node
);
4706 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4707 ext4_try_merge_freed_extent(sbi
, entry
, new_entry
,
4708 &(db
->bb_free_root
));
4711 spin_lock(&sbi
->s_md_lock
);
4712 list_add_tail(&new_entry
->efd_list
, &sbi
->s_freed_data_list
);
4713 sbi
->s_mb_free_pending
+= clusters
;
4714 spin_unlock(&sbi
->s_md_lock
);
4719 * ext4_free_blocks() -- Free given blocks and update quota
4720 * @handle: handle for this transaction
4722 * @bh: optional buffer of the block to be freed
4723 * @block: starting physical block to be freed
4724 * @count: number of blocks to be freed
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 sbi
->s_itb_per_group
) ||
4848 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4849 sbi
->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_array_rcu_deref(sbi
, s_flex_groups
,
4941 flex_group
)->free_clusters
);
4945 * on a bigalloc file system, defer the s_freeclusters_counter
4946 * update to the caller (ext4_remove_space and friends) so they
4947 * can determine if a cluster freed here should be rereserved
4949 if (!(flags
& EXT4_FREE_BLOCKS_RERESERVE_CLUSTER
)) {
4950 if (!(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4951 dquot_free_block(inode
, EXT4_C2B(sbi
, count_clusters
));
4952 percpu_counter_add(&sbi
->s_freeclusters_counter
,
4956 ext4_mb_unload_buddy(&e4b
);
4958 /* We dirtied the bitmap block */
4959 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4960 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4962 /* And the group descriptor block */
4963 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4964 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4968 if (overflow
&& !err
) {
4976 ext4_std_error(sb
, err
);
4981 * ext4_group_add_blocks() -- Add given blocks to an existing group
4982 * @handle: handle to this transaction
4984 * @block: start physical block to add to the block group
4985 * @count: number of blocks to free
4987 * This marks the blocks as free in the bitmap and buddy.
4989 int ext4_group_add_blocks(handle_t
*handle
, struct super_block
*sb
,
4990 ext4_fsblk_t block
, unsigned long count
)
4992 struct buffer_head
*bitmap_bh
= NULL
;
4993 struct buffer_head
*gd_bh
;
4994 ext4_group_t block_group
;
4997 struct ext4_group_desc
*desc
;
4998 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4999 struct ext4_buddy e4b
;
5000 int err
= 0, ret
, free_clusters_count
;
5001 ext4_grpblk_t clusters_freed
;
5002 ext4_fsblk_t first_cluster
= EXT4_B2C(sbi
, block
);
5003 ext4_fsblk_t last_cluster
= EXT4_B2C(sbi
, block
+ count
- 1);
5004 unsigned long cluster_count
= last_cluster
- first_cluster
+ 1;
5006 ext4_debug("Adding block(s) %llu-%llu\n", block
, block
+ count
- 1);
5011 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
5013 * Check to see if we are freeing blocks across a group
5016 if (bit
+ cluster_count
> EXT4_CLUSTERS_PER_GROUP(sb
)) {
5017 ext4_warning(sb
, "too many blocks added to group %u",
5023 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
5024 if (IS_ERR(bitmap_bh
)) {
5025 err
= PTR_ERR(bitmap_bh
);
5030 desc
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
5036 if (in_range(ext4_block_bitmap(sb
, desc
), block
, count
) ||
5037 in_range(ext4_inode_bitmap(sb
, desc
), block
, count
) ||
5038 in_range(block
, ext4_inode_table(sb
, desc
), sbi
->s_itb_per_group
) ||
5039 in_range(block
+ count
- 1, ext4_inode_table(sb
, desc
),
5040 sbi
->s_itb_per_group
)) {
5041 ext4_error(sb
, "Adding blocks in system zones - "
5042 "Block = %llu, count = %lu",
5048 BUFFER_TRACE(bitmap_bh
, "getting write access");
5049 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
5054 * We are about to modify some metadata. Call the journal APIs
5055 * to unshare ->b_data if a currently-committing transaction is
5058 BUFFER_TRACE(gd_bh
, "get_write_access");
5059 err
= ext4_journal_get_write_access(handle
, gd_bh
);
5063 for (i
= 0, clusters_freed
= 0; i
< cluster_count
; i
++) {
5064 BUFFER_TRACE(bitmap_bh
, "clear bit");
5065 if (!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
)) {
5066 ext4_error(sb
, "bit already cleared for block %llu",
5067 (ext4_fsblk_t
)(block
+ i
));
5068 BUFFER_TRACE(bitmap_bh
, "bit already cleared");
5074 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
5079 * need to update group_info->bb_free and bitmap
5080 * with group lock held. generate_buddy look at
5081 * them with group lock_held
5083 ext4_lock_group(sb
, block_group
);
5084 mb_clear_bits(bitmap_bh
->b_data
, bit
, cluster_count
);
5085 mb_free_blocks(NULL
, &e4b
, bit
, cluster_count
);
5086 free_clusters_count
= clusters_freed
+
5087 ext4_free_group_clusters(sb
, desc
);
5088 ext4_free_group_clusters_set(sb
, desc
, free_clusters_count
);
5089 ext4_block_bitmap_csum_set(sb
, block_group
, desc
, bitmap_bh
);
5090 ext4_group_desc_csum_set(sb
, block_group
, desc
);
5091 ext4_unlock_group(sb
, block_group
);
5092 percpu_counter_add(&sbi
->s_freeclusters_counter
,
5095 if (sbi
->s_log_groups_per_flex
) {
5096 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
5097 atomic64_add(clusters_freed
,
5098 &sbi_array_rcu_deref(sbi
, s_flex_groups
,
5099 flex_group
)->free_clusters
);
5102 ext4_mb_unload_buddy(&e4b
);
5104 /* We dirtied the bitmap block */
5105 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
5106 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
5108 /* And the group descriptor block */
5109 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
5110 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
5116 ext4_std_error(sb
, err
);
5121 * ext4_trim_extent -- function to TRIM one single free extent in the group
5122 * @sb: super block for the file system
5123 * @start: starting block of the free extent in the alloc. group
5124 * @count: number of blocks to TRIM
5125 * @group: alloc. group we are working with
5126 * @e4b: ext4 buddy for the group
5128 * Trim "count" blocks starting at "start" in the "group". To assure that no
5129 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5130 * be called with under the group lock.
5132 static int ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
5133 ext4_group_t group
, struct ext4_buddy
*e4b
)
5137 struct ext4_free_extent ex
;
5140 trace_ext4_trim_extent(sb
, group
, start
, count
);
5142 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
5144 ex
.fe_start
= start
;
5145 ex
.fe_group
= group
;
5149 * Mark blocks used, so no one can reuse them while
5152 mb_mark_used(e4b
, &ex
);
5153 ext4_unlock_group(sb
, group
);
5154 ret
= ext4_issue_discard(sb
, group
, start
, count
, NULL
);
5155 ext4_lock_group(sb
, group
);
5156 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
5161 * ext4_trim_all_free -- function to trim all free space in alloc. group
5162 * @sb: super block for file system
5163 * @group: group to be trimmed
5164 * @start: first group block to examine
5165 * @max: last group block to examine
5166 * @minblocks: minimum extent block count
5168 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5169 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5173 * ext4_trim_all_free walks through group's block bitmap searching for free
5174 * extents. When the free extent is found, mark it as used in group buddy
5175 * bitmap. Then issue a TRIM command on this extent and free the extent in
5176 * the group buddy bitmap. This is done until whole group is scanned.
5178 static ext4_grpblk_t
5179 ext4_trim_all_free(struct super_block
*sb
, ext4_group_t group
,
5180 ext4_grpblk_t start
, ext4_grpblk_t max
,
5181 ext4_grpblk_t minblocks
)
5184 ext4_grpblk_t next
, count
= 0, free_count
= 0;
5185 struct ext4_buddy e4b
;
5188 trace_ext4_trim_all_free(sb
, group
, start
, max
);
5190 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
5192 ext4_warning(sb
, "Error %d loading buddy information for %u",
5196 bitmap
= e4b
.bd_bitmap
;
5198 ext4_lock_group(sb
, group
);
5199 if (EXT4_MB_GRP_WAS_TRIMMED(e4b
.bd_info
) &&
5200 minblocks
>= atomic_read(&EXT4_SB(sb
)->s_last_trim_minblks
))
5203 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
5204 e4b
.bd_info
->bb_first_free
: start
;
5206 while (start
<= max
) {
5207 start
= mb_find_next_zero_bit(bitmap
, max
+ 1, start
);
5210 next
= mb_find_next_bit(bitmap
, max
+ 1, start
);
5212 if ((next
- start
) >= minblocks
) {
5213 ret
= ext4_trim_extent(sb
, start
,
5214 next
- start
, group
, &e4b
);
5215 if (ret
&& ret
!= -EOPNOTSUPP
)
5218 count
+= next
- start
;
5220 free_count
+= next
- start
;
5223 if (fatal_signal_pending(current
)) {
5224 count
= -ERESTARTSYS
;
5228 if (need_resched()) {
5229 ext4_unlock_group(sb
, group
);
5231 ext4_lock_group(sb
, group
);
5234 if ((e4b
.bd_info
->bb_free
- free_count
) < minblocks
)
5240 EXT4_MB_GRP_SET_TRIMMED(e4b
.bd_info
);
5243 ext4_unlock_group(sb
, group
);
5244 ext4_mb_unload_buddy(&e4b
);
5246 ext4_debug("trimmed %d blocks in the group %d\n",
5253 * ext4_trim_fs() -- trim ioctl handle function
5254 * @sb: superblock for filesystem
5255 * @range: fstrim_range structure
5257 * start: First Byte to trim
5258 * len: number of Bytes to trim from start
5259 * minlen: minimum extent length in Bytes
5260 * ext4_trim_fs goes through all allocation groups containing Bytes from
5261 * start to start+len. For each such a group ext4_trim_all_free function
5262 * is invoked to trim all free space.
5264 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
5266 struct ext4_group_info
*grp
;
5267 ext4_group_t group
, first_group
, last_group
;
5268 ext4_grpblk_t cnt
= 0, first_cluster
, last_cluster
;
5269 uint64_t start
, end
, minlen
, trimmed
= 0;
5270 ext4_fsblk_t first_data_blk
=
5271 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
5272 ext4_fsblk_t max_blks
= ext4_blocks_count(EXT4_SB(sb
)->s_es
);
5275 start
= range
->start
>> sb
->s_blocksize_bits
;
5276 end
= start
+ (range
->len
>> sb
->s_blocksize_bits
) - 1;
5277 minlen
= EXT4_NUM_B2C(EXT4_SB(sb
),
5278 range
->minlen
>> sb
->s_blocksize_bits
);
5280 if (minlen
> EXT4_CLUSTERS_PER_GROUP(sb
) ||
5281 start
>= max_blks
||
5282 range
->len
< sb
->s_blocksize
)
5284 if (end
>= max_blks
)
5286 if (end
<= first_data_blk
)
5288 if (start
< first_data_blk
)
5289 start
= first_data_blk
;
5291 /* Determine first and last group to examine based on start and end */
5292 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
5293 &first_group
, &first_cluster
);
5294 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) end
,
5295 &last_group
, &last_cluster
);
5297 /* end now represents the last cluster to discard in this group */
5298 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5300 for (group
= first_group
; group
<= last_group
; group
++) {
5301 grp
= ext4_get_group_info(sb
, group
);
5302 /* We only do this if the grp has never been initialized */
5303 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
5304 ret
= ext4_mb_init_group(sb
, group
, GFP_NOFS
);
5310 * For all the groups except the last one, last cluster will
5311 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5312 * change it for the last group, note that last_cluster is
5313 * already computed earlier by ext4_get_group_no_and_offset()
5315 if (group
== last_group
)
5318 if (grp
->bb_free
>= minlen
) {
5319 cnt
= ext4_trim_all_free(sb
, group
, first_cluster
,
5329 * For every group except the first one, we are sure
5330 * that the first cluster to discard will be cluster #0.
5336 atomic_set(&EXT4_SB(sb
)->s_last_trim_minblks
, minlen
);
5339 range
->len
= EXT4_C2B(EXT4_SB(sb
), trimmed
) << sb
->s_blocksize_bits
;
5343 /* Iterate all the free extents in the group. */
5345 ext4_mballoc_query_range(
5346 struct super_block
*sb
,
5348 ext4_grpblk_t start
,
5350 ext4_mballoc_query_range_fn formatter
,
5355 struct ext4_buddy e4b
;
5358 error
= ext4_mb_load_buddy(sb
, group
, &e4b
);
5361 bitmap
= e4b
.bd_bitmap
;
5363 ext4_lock_group(sb
, group
);
5365 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
5366 e4b
.bd_info
->bb_first_free
: start
;
5367 if (end
>= EXT4_CLUSTERS_PER_GROUP(sb
))
5368 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5370 while (start
<= end
) {
5371 start
= mb_find_next_zero_bit(bitmap
, end
+ 1, start
);
5374 next
= mb_find_next_bit(bitmap
, end
+ 1, start
);
5376 ext4_unlock_group(sb
, group
);
5377 error
= formatter(sb
, group
, start
, next
- start
, priv
);
5380 ext4_lock_group(sb
, group
);
5385 ext4_unlock_group(sb
, group
);
5387 ext4_mb_unload_buddy(&e4b
);