2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
24 #include "ext4_jbd2.h"
26 #include <linux/log2.h>
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <trace/events/ext4.h>
31 #ifdef CONFIG_EXT4_DEBUG
32 ushort ext4_mballoc_debug __read_mostly
;
34 module_param_named(mballoc_debug
, ext4_mballoc_debug
, ushort
, 0644);
35 MODULE_PARM_DESC(mballoc_debug
, "Debugging level for ext4's mballoc");
40 * - test ext4_ext_search_left() and ext4_ext_search_right()
41 * - search for metadata in few groups
44 * - normalization should take into account whether file is still open
45 * - discard preallocations if no free space left (policy?)
46 * - don't normalize tails
48 * - reservation for superuser
51 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
52 * - track min/max extents in each group for better group selection
53 * - mb_mark_used() may allocate chunk right after splitting buddy
54 * - tree of groups sorted by number of free blocks
59 * The allocation request involve request for multiple number of blocks
60 * near to the goal(block) value specified.
62 * During initialization phase of the allocator we decide to use the
63 * group preallocation or inode preallocation depending on the size of
64 * the file. The size of the file could be the resulting file size we
65 * would have after allocation, or the current file size, which ever
66 * is larger. If the size is less than sbi->s_mb_stream_request we
67 * select to use the group preallocation. The default value of
68 * s_mb_stream_request is 16 blocks. This can also be tuned via
69 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
70 * terms of number of blocks.
72 * The main motivation for having small file use group preallocation is to
73 * ensure that we have small files closer together on the disk.
75 * First stage the allocator looks at the inode prealloc list,
76 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
77 * spaces for this particular inode. The inode prealloc space is
80 * pa_lstart -> the logical start block for this prealloc space
81 * pa_pstart -> the physical start block for this prealloc space
82 * pa_len -> length for this prealloc space (in clusters)
83 * pa_free -> free space available in this prealloc space (in clusters)
85 * The inode preallocation space is used looking at the _logical_ start
86 * block. If only the logical file block falls within the range of prealloc
87 * space we will consume the particular prealloc space. This makes sure that
88 * we have contiguous physical blocks representing the file blocks
90 * The important thing to be noted in case of inode prealloc space is that
91 * we don't modify the values associated to inode prealloc space except
94 * If we are not able to find blocks in the inode prealloc space and if we
95 * have the group allocation flag set then we look at the locality group
96 * prealloc space. These are per CPU prealloc list represented as
98 * ext4_sb_info.s_locality_groups[smp_processor_id()]
100 * The reason for having a per cpu locality group is to reduce the contention
101 * between CPUs. It is possible to get scheduled at this point.
103 * The locality group prealloc space is used looking at whether we have
104 * enough free space (pa_free) within the prealloc space.
106 * If we can't allocate blocks via inode prealloc or/and locality group
107 * prealloc then we look at the buddy cache. The buddy cache is represented
108 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
109 * mapped to the buddy and bitmap information regarding different
110 * groups. The buddy information is attached to buddy cache inode so that
111 * we can access them through the page cache. The information regarding
112 * each group is loaded via ext4_mb_load_buddy. The information involve
113 * block bitmap and buddy information. The information are stored in the
117 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
120 * one block each for bitmap and buddy information. So for each group we
121 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
122 * blocksize) blocks. So it can have information regarding groups_per_page
123 * which is blocks_per_page/2
125 * The buddy cache inode is not stored on disk. The inode is thrown
126 * away when the filesystem is unmounted.
128 * We look for count number of blocks in the buddy cache. If we were able
129 * to locate that many free blocks we return with additional information
130 * regarding rest of the contiguous physical block available
132 * Before allocating blocks via buddy cache we normalize the request
133 * blocks. This ensure we ask for more blocks that we needed. The extra
134 * blocks that we get after allocation is added to the respective prealloc
135 * list. In case of inode preallocation we follow a list of heuristics
136 * based on file size. This can be found in ext4_mb_normalize_request. If
137 * we are doing a group prealloc we try to normalize the request to
138 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
139 * dependent on the cluster size; for non-bigalloc file systems, it is
140 * 512 blocks. This can be tuned via
141 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
142 * terms of number of blocks. If we have mounted the file system with -O
143 * stripe=<value> option the group prealloc request is normalized to the
144 * the smallest multiple of the stripe value (sbi->s_stripe) which is
145 * greater than the default mb_group_prealloc.
147 * The regular allocator (using the buddy cache) supports a few tunables.
149 * /sys/fs/ext4/<partition>/mb_min_to_scan
150 * /sys/fs/ext4/<partition>/mb_max_to_scan
151 * /sys/fs/ext4/<partition>/mb_order2_req
153 * The regular allocator uses buddy scan only if the request len is power of
154 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
155 * value of s_mb_order2_reqs can be tuned via
156 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
157 * stripe size (sbi->s_stripe), we try to search for contiguous block in
158 * stripe size. This should result in better allocation on RAID setups. If
159 * not, we search in the specific group using bitmap for best extents. The
160 * tunable min_to_scan and max_to_scan control the behaviour here.
161 * min_to_scan indicate how long the mballoc __must__ look for a best
162 * extent and max_to_scan indicates how long the mballoc __can__ look for a
163 * best extent in the found extents. Searching for the blocks starts with
164 * the group specified as the goal value in allocation context via
165 * ac_g_ex. Each group is first checked based on the criteria whether it
166 * can be used for allocation. ext4_mb_good_group explains how the groups are
169 * Both the prealloc space are getting populated as above. So for the first
170 * request we will hit the buddy cache which will result in this prealloc
171 * space getting filled. The prealloc space is then later used for the
172 * subsequent request.
176 * mballoc operates on the following data:
178 * - in-core buddy (actually includes buddy and bitmap)
179 * - preallocation descriptors (PAs)
181 * there are two types of preallocations:
183 * assiged to specific inode and can be used for this inode only.
184 * it describes part of inode's space preallocated to specific
185 * physical blocks. any block from that preallocated can be used
186 * independent. the descriptor just tracks number of blocks left
187 * unused. so, before taking some block from descriptor, one must
188 * make sure corresponded logical block isn't allocated yet. this
189 * also means that freeing any block within descriptor's range
190 * must discard all preallocated blocks.
192 * assigned to specific locality group which does not translate to
193 * permanent set of inodes: inode can join and leave group. space
194 * from this type of preallocation can be used for any inode. thus
195 * it's consumed from the beginning to the end.
197 * relation between them can be expressed as:
198 * in-core buddy = on-disk bitmap + preallocation descriptors
200 * this mean blocks mballoc considers used are:
201 * - allocated blocks (persistent)
202 * - preallocated blocks (non-persistent)
204 * consistency in mballoc world means that at any time a block is either
205 * free or used in ALL structures. notice: "any time" should not be read
206 * literally -- time is discrete and delimited by locks.
208 * to keep it simple, we don't use block numbers, instead we count number of
209 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
211 * all operations can be expressed as:
212 * - init buddy: buddy = on-disk + PAs
213 * - new PA: buddy += N; PA = N
214 * - use inode PA: on-disk += N; PA -= N
215 * - discard inode PA buddy -= on-disk - PA; PA = 0
216 * - use locality group PA on-disk += N; PA -= N
217 * - discard locality group PA buddy -= PA; PA = 0
218 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
219 * is used in real operation because we can't know actual used
220 * bits from PA, only from on-disk bitmap
222 * if we follow this strict logic, then all operations above should be atomic.
223 * given some of them can block, we'd have to use something like semaphores
224 * killing performance on high-end SMP hardware. let's try to relax it using
225 * the following knowledge:
226 * 1) if buddy is referenced, it's already initialized
227 * 2) while block is used in buddy and the buddy is referenced,
228 * nobody can re-allocate that block
229 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
230 * bit set and PA claims same block, it's OK. IOW, one can set bit in
231 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
234 * so, now we're building a concurrency table:
237 * blocks for PA are allocated in the buddy, buddy must be referenced
238 * until PA is linked to allocation group to avoid concurrent buddy init
240 * we need to make sure that either on-disk bitmap or PA has uptodate data
241 * given (3) we care that PA-=N operation doesn't interfere with init
243 * the simplest way would be to have buddy initialized by the discard
244 * - use locality group PA
245 * again PA-=N must be serialized with init
246 * - discard locality group PA
247 * the simplest way would be to have buddy initialized by the discard
250 * i_data_sem serializes them
252 * discard process must wait until PA isn't used by another process
253 * - use locality group PA
254 * some mutex should serialize them
255 * - discard locality group PA
256 * discard process must wait until PA isn't used by another process
259 * i_data_sem or another mutex should serializes them
261 * discard process must wait until PA isn't used by another process
262 * - use locality group PA
263 * nothing wrong here -- they're different PAs covering different blocks
264 * - discard locality group PA
265 * discard process must wait until PA isn't used by another process
267 * now we're ready to make few consequences:
268 * - PA is referenced and while it is no discard is possible
269 * - PA is referenced until block isn't marked in on-disk bitmap
270 * - PA changes only after on-disk bitmap
271 * - discard must not compete with init. either init is done before
272 * any discard or they're serialized somehow
273 * - buddy init as sum of on-disk bitmap and PAs is done atomically
275 * a special case when we've used PA to emptiness. no need to modify buddy
276 * in this case, but we should care about concurrent init
281 * Logic in few words:
286 * mark bits in on-disk bitmap
289 * - use preallocation:
290 * find proper PA (per-inode or group)
292 * mark bits in on-disk bitmap
298 * mark bits in on-disk bitmap
301 * - discard preallocations in group:
303 * move them onto local list
304 * load on-disk bitmap
306 * remove PA from object (inode or locality group)
307 * mark free blocks in-core
309 * - discard inode's preallocations:
316 * - bitlock on a group (group)
317 * - object (inode/locality) (object)
328 * - release consumed pa:
333 * - generate in-core bitmap:
337 * - discard all for given object (inode, locality group):
342 * - discard all for given group:
349 static struct kmem_cache
*ext4_pspace_cachep
;
350 static struct kmem_cache
*ext4_ac_cachep
;
351 static struct kmem_cache
*ext4_free_data_cachep
;
353 /* We create slab caches for groupinfo data structures based on the
354 * superblock block size. There will be one per mounted filesystem for
355 * each unique s_blocksize_bits */
356 #define NR_GRPINFO_CACHES 8
357 static struct kmem_cache
*ext4_groupinfo_caches
[NR_GRPINFO_CACHES
];
359 static const char *ext4_groupinfo_slab_names
[NR_GRPINFO_CACHES
] = {
360 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
361 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
362 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
365 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
367 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
369 static void ext4_free_data_callback(struct super_block
*sb
,
370 struct ext4_journal_cb_entry
*jce
, int rc
);
372 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
374 #if BITS_PER_LONG == 64
375 *bit
+= ((unsigned long) addr
& 7UL) << 3;
376 addr
= (void *) ((unsigned long) addr
& ~7UL);
377 #elif BITS_PER_LONG == 32
378 *bit
+= ((unsigned long) addr
& 3UL) << 3;
379 addr
= (void *) ((unsigned long) addr
& ~3UL);
381 #error "how many bits you are?!"
386 static inline int mb_test_bit(int bit
, void *addr
)
389 * ext4_test_bit on architecture like powerpc
390 * needs unsigned long aligned address
392 addr
= mb_correct_addr_and_bit(&bit
, addr
);
393 return ext4_test_bit(bit
, addr
);
396 static inline void mb_set_bit(int bit
, void *addr
)
398 addr
= mb_correct_addr_and_bit(&bit
, addr
);
399 ext4_set_bit(bit
, addr
);
402 static inline void mb_clear_bit(int bit
, void *addr
)
404 addr
= mb_correct_addr_and_bit(&bit
, addr
);
405 ext4_clear_bit(bit
, addr
);
408 static inline int mb_test_and_clear_bit(int bit
, void *addr
)
410 addr
= mb_correct_addr_and_bit(&bit
, addr
);
411 return ext4_test_and_clear_bit(bit
, addr
);
414 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
416 int fix
= 0, ret
, tmpmax
;
417 addr
= mb_correct_addr_and_bit(&fix
, addr
);
421 ret
= ext4_find_next_zero_bit(addr
, tmpmax
, start
) - fix
;
427 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
429 int fix
= 0, ret
, tmpmax
;
430 addr
= mb_correct_addr_and_bit(&fix
, addr
);
434 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
440 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
444 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
447 if (order
> e4b
->bd_blkbits
+ 1) {
452 /* at order 0 we see each particular block */
454 *max
= 1 << (e4b
->bd_blkbits
+ 3);
455 return e4b
->bd_bitmap
;
458 bb
= e4b
->bd_buddy
+ EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
459 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
465 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
466 int first
, int count
)
469 struct super_block
*sb
= e4b
->bd_sb
;
471 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
473 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
474 for (i
= 0; i
< count
; i
++) {
475 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
476 ext4_fsblk_t blocknr
;
478 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
479 blocknr
+= EXT4_C2B(EXT4_SB(sb
), first
+ i
);
480 ext4_grp_locked_error(sb
, e4b
->bd_group
,
481 inode
? inode
->i_ino
: 0,
483 "freeing block already freed "
487 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
491 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
495 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
497 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
498 for (i
= 0; i
< count
; i
++) {
499 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
500 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
504 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
506 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
507 unsigned char *b1
, *b2
;
509 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
510 b2
= (unsigned char *) bitmap
;
511 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
512 if (b1
[i
] != b2
[i
]) {
513 ext4_msg(e4b
->bd_sb
, KERN_ERR
,
514 "corruption in group %u "
515 "at byte %u(%u): %x in copy != %x "
517 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
525 static inline void mb_free_blocks_double(struct inode
*inode
,
526 struct ext4_buddy
*e4b
, int first
, int count
)
530 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
531 int first
, int count
)
535 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
541 #ifdef AGGRESSIVE_CHECK
543 #define MB_CHECK_ASSERT(assert) \
547 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
548 function, file, line, # assert); \
553 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
554 const char *function
, int line
)
556 struct super_block
*sb
= e4b
->bd_sb
;
557 int order
= e4b
->bd_blkbits
+ 1;
564 struct ext4_group_info
*grp
;
567 struct list_head
*cur
;
572 static int mb_check_counter
;
573 if (mb_check_counter
++ % 100 != 0)
578 buddy
= mb_find_buddy(e4b
, order
, &max
);
579 MB_CHECK_ASSERT(buddy
);
580 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
581 MB_CHECK_ASSERT(buddy2
);
582 MB_CHECK_ASSERT(buddy
!= buddy2
);
583 MB_CHECK_ASSERT(max
* 2 == max2
);
586 for (i
= 0; i
< max
; i
++) {
588 if (mb_test_bit(i
, buddy
)) {
589 /* only single bit in buddy2 may be 1 */
590 if (!mb_test_bit(i
<< 1, buddy2
)) {
592 mb_test_bit((i
<<1)+1, buddy2
));
593 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
595 mb_test_bit(i
<< 1, buddy2
));
600 /* both bits in buddy2 must be 1 */
601 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
602 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
604 for (j
= 0; j
< (1 << order
); j
++) {
605 k
= (i
* (1 << order
)) + j
;
607 !mb_test_bit(k
, e4b
->bd_bitmap
));
611 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
616 buddy
= mb_find_buddy(e4b
, 0, &max
);
617 for (i
= 0; i
< max
; i
++) {
618 if (!mb_test_bit(i
, buddy
)) {
619 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
627 /* check used bits only */
628 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
629 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
631 MB_CHECK_ASSERT(k
< max2
);
632 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
635 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
636 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
638 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
639 list_for_each(cur
, &grp
->bb_prealloc_list
) {
640 ext4_group_t groupnr
;
641 struct ext4_prealloc_space
*pa
;
642 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
643 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
644 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
645 for (i
= 0; i
< pa
->pa_len
; i
++)
646 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
650 #undef MB_CHECK_ASSERT
651 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
652 __FILE__, __func__, __LINE__)
654 #define mb_check_buddy(e4b)
658 * Divide blocks started from @first with length @len into
659 * smaller chunks with power of 2 blocks.
660 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
661 * then increase bb_counters[] for corresponded chunk size.
663 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
664 void *buddy
, ext4_grpblk_t first
, ext4_grpblk_t len
,
665 struct ext4_group_info
*grp
)
667 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
671 unsigned short border
;
673 BUG_ON(len
> EXT4_CLUSTERS_PER_GROUP(sb
));
675 border
= 2 << sb
->s_blocksize_bits
;
678 /* find how many blocks can be covered since this position */
679 max
= ffs(first
| border
) - 1;
681 /* find how many blocks of power 2 we need to mark */
688 /* mark multiblock chunks only */
689 grp
->bb_counters
[min
]++;
691 mb_clear_bit(first
>> min
,
692 buddy
+ sbi
->s_mb_offsets
[min
]);
700 * Cache the order of the largest free extent we have available in this block
704 mb_set_largest_free_order(struct super_block
*sb
, struct ext4_group_info
*grp
)
709 grp
->bb_largest_free_order
= -1; /* uninit */
711 bits
= sb
->s_blocksize_bits
+ 1;
712 for (i
= bits
; i
>= 0; i
--) {
713 if (grp
->bb_counters
[i
] > 0) {
714 grp
->bb_largest_free_order
= i
;
720 static noinline_for_stack
721 void ext4_mb_generate_buddy(struct super_block
*sb
,
722 void *buddy
, void *bitmap
, ext4_group_t group
)
724 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
725 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
726 ext4_grpblk_t max
= EXT4_CLUSTERS_PER_GROUP(sb
);
731 unsigned fragments
= 0;
732 unsigned long long period
= get_cycles();
734 /* initialize buddy from bitmap which is aggregation
735 * of on-disk bitmap and preallocations */
736 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
737 grp
->bb_first_free
= i
;
741 i
= mb_find_next_bit(bitmap
, max
, i
);
745 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
747 grp
->bb_counters
[0]++;
749 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
751 grp
->bb_fragments
= fragments
;
753 if (free
!= grp
->bb_free
) {
754 ext4_grp_locked_error(sb
, group
, 0, 0,
755 "block bitmap and bg descriptor "
756 "inconsistent: %u vs %u free clusters",
759 * If we intend to continue, we consider group descriptor
760 * corrupt and update bb_free using bitmap value
763 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp
))
764 percpu_counter_sub(&sbi
->s_freeclusters_counter
,
766 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT
, &grp
->bb_state
);
768 mb_set_largest_free_order(sb
, grp
);
770 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
772 period
= get_cycles() - period
;
773 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
774 EXT4_SB(sb
)->s_mb_buddies_generated
++;
775 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
776 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
779 static void mb_regenerate_buddy(struct ext4_buddy
*e4b
)
785 while ((buddy
= mb_find_buddy(e4b
, order
++, &count
))) {
786 ext4_set_bits(buddy
, 0, count
);
788 e4b
->bd_info
->bb_fragments
= 0;
789 memset(e4b
->bd_info
->bb_counters
, 0,
790 sizeof(*e4b
->bd_info
->bb_counters
) *
791 (e4b
->bd_sb
->s_blocksize_bits
+ 2));
793 ext4_mb_generate_buddy(e4b
->bd_sb
, e4b
->bd_buddy
,
794 e4b
->bd_bitmap
, e4b
->bd_group
);
797 /* The buddy information is attached the buddy cache inode
798 * for convenience. The information regarding each group
799 * is loaded via ext4_mb_load_buddy. The information involve
800 * block bitmap and buddy information. The information are
801 * stored in the inode as
804 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
807 * one block each for bitmap and buddy information.
808 * So for each group we take up 2 blocks. A page can
809 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
810 * So it can have information regarding groups_per_page which
811 * is blocks_per_page/2
813 * Locking note: This routine takes the block group lock of all groups
814 * for this page; do not hold this lock when calling this routine!
817 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
819 ext4_group_t ngroups
;
825 ext4_group_t first_group
, group
;
827 struct super_block
*sb
;
828 struct buffer_head
*bhs
;
829 struct buffer_head
**bh
= NULL
;
833 struct ext4_group_info
*grinfo
;
835 mb_debug(1, "init page %lu\n", page
->index
);
837 inode
= page
->mapping
->host
;
839 ngroups
= ext4_get_groups_count(sb
);
840 blocksize
= 1 << inode
->i_blkbits
;
841 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
843 groups_per_page
= blocks_per_page
>> 1;
844 if (groups_per_page
== 0)
847 /* allocate buffer_heads to read bitmaps */
848 if (groups_per_page
> 1) {
849 i
= sizeof(struct buffer_head
*) * groups_per_page
;
850 bh
= kzalloc(i
, GFP_NOFS
);
858 first_group
= page
->index
* blocks_per_page
/ 2;
860 /* read all groups the page covers into the cache */
861 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
862 if (group
>= ngroups
)
865 grinfo
= ext4_get_group_info(sb
, group
);
867 * If page is uptodate then we came here after online resize
868 * which added some new uninitialized group info structs, so
869 * we must skip all initialized uptodate buddies on the page,
870 * which may be currently in use by an allocating task.
872 if (PageUptodate(page
) && !EXT4_MB_GRP_NEED_INIT(grinfo
)) {
876 if (!(bh
[i
] = ext4_read_block_bitmap_nowait(sb
, group
))) {
880 mb_debug(1, "read bitmap for group %u\n", group
);
883 /* wait for I/O completion */
884 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
885 if (bh
[i
] && ext4_wait_block_bitmap(sb
, group
, bh
[i
]))
889 first_block
= page
->index
* blocks_per_page
;
890 for (i
= 0; i
< blocks_per_page
; i
++) {
891 group
= (first_block
+ i
) >> 1;
892 if (group
>= ngroups
)
895 if (!bh
[group
- first_group
])
896 /* skip initialized uptodate buddy */
899 if (!buffer_verified(bh
[group
- first_group
]))
900 /* Skip faulty bitmaps */
905 * data carry information regarding this
906 * particular group in the format specified
910 data
= page_address(page
) + (i
* blocksize
);
911 bitmap
= bh
[group
- first_group
]->b_data
;
914 * We place the buddy block and bitmap block
917 if ((first_block
+ i
) & 1) {
918 /* this is block of buddy */
919 BUG_ON(incore
== NULL
);
920 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
921 group
, page
->index
, i
* blocksize
);
922 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
923 grinfo
= ext4_get_group_info(sb
, group
);
924 grinfo
->bb_fragments
= 0;
925 memset(grinfo
->bb_counters
, 0,
926 sizeof(*grinfo
->bb_counters
) *
927 (sb
->s_blocksize_bits
+2));
929 * incore got set to the group block bitmap below
931 ext4_lock_group(sb
, group
);
933 memset(data
, 0xff, blocksize
);
934 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
935 ext4_unlock_group(sb
, group
);
938 /* this is block of bitmap */
939 BUG_ON(incore
!= NULL
);
940 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
941 group
, page
->index
, i
* blocksize
);
942 trace_ext4_mb_bitmap_load(sb
, group
);
944 /* see comments in ext4_mb_put_pa() */
945 ext4_lock_group(sb
, group
);
946 memcpy(data
, bitmap
, blocksize
);
948 /* mark all preallocated blks used in in-core bitmap */
949 ext4_mb_generate_from_pa(sb
, data
, group
);
950 ext4_mb_generate_from_freelist(sb
, data
, group
);
951 ext4_unlock_group(sb
, group
);
953 /* set incore so that the buddy information can be
954 * generated using this
959 SetPageUptodate(page
);
963 for (i
= 0; i
< groups_per_page
; i
++)
972 * Lock the buddy and bitmap pages. This make sure other parallel init_group
973 * on the same buddy page doesn't happen whild holding the buddy page lock.
974 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
975 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
977 static int ext4_mb_get_buddy_page_lock(struct super_block
*sb
,
978 ext4_group_t group
, struct ext4_buddy
*e4b
)
980 struct inode
*inode
= EXT4_SB(sb
)->s_buddy_cache
;
981 int block
, pnum
, poff
;
985 e4b
->bd_buddy_page
= NULL
;
986 e4b
->bd_bitmap_page
= NULL
;
988 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
990 * the buddy cache inode stores the block bitmap
991 * and buddy information in consecutive blocks.
992 * So for each group we need two blocks.
995 pnum
= block
/ blocks_per_page
;
996 poff
= block
% blocks_per_page
;
997 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1000 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1001 e4b
->bd_bitmap_page
= page
;
1002 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1004 if (blocks_per_page
>= 2) {
1005 /* buddy and bitmap are on the same page */
1010 pnum
= block
/ blocks_per_page
;
1011 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1014 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1015 e4b
->bd_buddy_page
= page
;
1019 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy
*e4b
)
1021 if (e4b
->bd_bitmap_page
) {
1022 unlock_page(e4b
->bd_bitmap_page
);
1023 page_cache_release(e4b
->bd_bitmap_page
);
1025 if (e4b
->bd_buddy_page
) {
1026 unlock_page(e4b
->bd_buddy_page
);
1027 page_cache_release(e4b
->bd_buddy_page
);
1032 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1033 * block group lock of all groups for this page; do not hold the BG lock when
1034 * calling this routine!
1036 static noinline_for_stack
1037 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
)
1040 struct ext4_group_info
*this_grp
;
1041 struct ext4_buddy e4b
;
1046 mb_debug(1, "init group %u\n", group
);
1047 this_grp
= ext4_get_group_info(sb
, group
);
1049 * This ensures that we don't reinit the buddy cache
1050 * page which map to the group from which we are already
1051 * allocating. If we are looking at the buddy cache we would
1052 * have taken a reference using ext4_mb_load_buddy and that
1053 * would have pinned buddy page to page cache.
1054 * The call to ext4_mb_get_buddy_page_lock will mark the
1057 ret
= ext4_mb_get_buddy_page_lock(sb
, group
, &e4b
);
1058 if (ret
|| !EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1060 * somebody initialized the group
1061 * return without doing anything
1066 page
= e4b
.bd_bitmap_page
;
1067 ret
= ext4_mb_init_cache(page
, NULL
);
1070 if (!PageUptodate(page
)) {
1075 if (e4b
.bd_buddy_page
== NULL
) {
1077 * If both the bitmap and buddy are in
1078 * the same page we don't need to force
1084 /* init buddy cache */
1085 page
= e4b
.bd_buddy_page
;
1086 ret
= ext4_mb_init_cache(page
, e4b
.bd_bitmap
);
1089 if (!PageUptodate(page
)) {
1094 ext4_mb_put_buddy_page_lock(&e4b
);
1099 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1100 * block group lock of all groups for this page; do not hold the BG lock when
1101 * calling this routine!
1103 static noinline_for_stack
int
1104 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1105 struct ext4_buddy
*e4b
)
1107 int blocks_per_page
;
1113 struct ext4_group_info
*grp
;
1114 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1115 struct inode
*inode
= sbi
->s_buddy_cache
;
1118 mb_debug(1, "load group %u\n", group
);
1120 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1121 grp
= ext4_get_group_info(sb
, group
);
1123 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1126 e4b
->bd_group
= group
;
1127 e4b
->bd_buddy_page
= NULL
;
1128 e4b
->bd_bitmap_page
= NULL
;
1130 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1132 * we need full data about the group
1133 * to make a good selection
1135 ret
= ext4_mb_init_group(sb
, group
);
1141 * the buddy cache inode stores the block bitmap
1142 * and buddy information in consecutive blocks.
1143 * So for each group we need two blocks.
1146 pnum
= block
/ blocks_per_page
;
1147 poff
= block
% blocks_per_page
;
1149 /* we could use find_or_create_page(), but it locks page
1150 * what we'd like to avoid in fast path ... */
1151 page
= find_get_page_flags(inode
->i_mapping
, pnum
, FGP_ACCESSED
);
1152 if (page
== NULL
|| !PageUptodate(page
)) {
1155 * drop the page reference and try
1156 * to get the page with lock. If we
1157 * are not uptodate that implies
1158 * somebody just created the page but
1159 * is yet to initialize the same. So
1160 * wait for it to initialize.
1162 page_cache_release(page
);
1163 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1165 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1166 if (!PageUptodate(page
)) {
1167 ret
= ext4_mb_init_cache(page
, NULL
);
1172 mb_cmp_bitmaps(e4b
, page_address(page
) +
1173 (poff
* sb
->s_blocksize
));
1182 if (!PageUptodate(page
)) {
1187 /* Pages marked accessed already */
1188 e4b
->bd_bitmap_page
= page
;
1189 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1192 pnum
= block
/ blocks_per_page
;
1193 poff
= block
% blocks_per_page
;
1195 page
= find_get_page_flags(inode
->i_mapping
, pnum
, FGP_ACCESSED
);
1196 if (page
== NULL
|| !PageUptodate(page
)) {
1198 page_cache_release(page
);
1199 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1201 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1202 if (!PageUptodate(page
)) {
1203 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1216 if (!PageUptodate(page
)) {
1221 /* Pages marked accessed already */
1222 e4b
->bd_buddy_page
= page
;
1223 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1225 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1226 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1232 page_cache_release(page
);
1233 if (e4b
->bd_bitmap_page
)
1234 page_cache_release(e4b
->bd_bitmap_page
);
1235 if (e4b
->bd_buddy_page
)
1236 page_cache_release(e4b
->bd_buddy_page
);
1237 e4b
->bd_buddy
= NULL
;
1238 e4b
->bd_bitmap
= NULL
;
1242 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1244 if (e4b
->bd_bitmap_page
)
1245 page_cache_release(e4b
->bd_bitmap_page
);
1246 if (e4b
->bd_buddy_page
)
1247 page_cache_release(e4b
->bd_buddy_page
);
1251 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1256 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
1257 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1260 while (order
<= e4b
->bd_blkbits
+ 1) {
1262 if (!mb_test_bit(block
, bb
)) {
1263 /* this block is part of buddy of order 'order' */
1266 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1272 static void mb_clear_bits(void *bm
, int cur
, int len
)
1278 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1279 /* fast path: clear whole word at once */
1280 addr
= bm
+ (cur
>> 3);
1285 mb_clear_bit(cur
, bm
);
1290 /* clear bits in given range
1291 * will return first found zero bit if any, -1 otherwise
1293 static int mb_test_and_clear_bits(void *bm
, int cur
, int len
)
1300 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1301 /* fast path: clear whole word at once */
1302 addr
= bm
+ (cur
>> 3);
1303 if (*addr
!= (__u32
)(-1) && zero_bit
== -1)
1304 zero_bit
= cur
+ mb_find_next_zero_bit(addr
, 32, 0);
1309 if (!mb_test_and_clear_bit(cur
, bm
) && zero_bit
== -1)
1317 void ext4_set_bits(void *bm
, int cur
, int len
)
1323 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1324 /* fast path: set whole word at once */
1325 addr
= bm
+ (cur
>> 3);
1330 mb_set_bit(cur
, bm
);
1336 * _________________________________________________________________ */
1338 static inline int mb_buddy_adjust_border(int* bit
, void* bitmap
, int side
)
1340 if (mb_test_bit(*bit
+ side
, bitmap
)) {
1341 mb_clear_bit(*bit
, bitmap
);
1347 mb_set_bit(*bit
, bitmap
);
1352 static void mb_buddy_mark_free(struct ext4_buddy
*e4b
, int first
, int last
)
1356 void *buddy
= mb_find_buddy(e4b
, order
, &max
);
1361 /* Bits in range [first; last] are known to be set since
1362 * corresponding blocks were allocated. Bits in range
1363 * (first; last) will stay set because they form buddies on
1364 * upper layer. We just deal with borders if they don't
1365 * align with upper layer and then go up.
1366 * Releasing entire group is all about clearing
1367 * single bit of highest order buddy.
1371 * ---------------------------------
1373 * ---------------------------------
1374 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1375 * ---------------------------------
1377 * \_____________________/
1379 * Neither [1] nor [6] is aligned to above layer.
1380 * Left neighbour [0] is free, so mark it busy,
1381 * decrease bb_counters and extend range to
1383 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1384 * mark [6] free, increase bb_counters and shrink range to
1386 * Then shift range to [0; 2], go up and do the same.
1391 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&first
, buddy
, -1);
1393 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&last
, buddy
, 1);
1398 if (first
== last
|| !(buddy2
= mb_find_buddy(e4b
, order
, &max
))) {
1399 mb_clear_bits(buddy
, first
, last
- first
+ 1);
1400 e4b
->bd_info
->bb_counters
[order
- 1] += last
- first
+ 1;
1409 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1410 int first
, int count
)
1412 int left_is_free
= 0;
1413 int right_is_free
= 0;
1415 int last
= first
+ count
- 1;
1416 struct super_block
*sb
= e4b
->bd_sb
;
1418 if (WARN_ON(count
== 0))
1420 BUG_ON(last
>= (sb
->s_blocksize
<< 3));
1421 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1422 /* Don't bother if the block group is corrupt. */
1423 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
)))
1426 mb_check_buddy(e4b
);
1427 mb_free_blocks_double(inode
, e4b
, first
, count
);
1429 e4b
->bd_info
->bb_free
+= count
;
1430 if (first
< e4b
->bd_info
->bb_first_free
)
1431 e4b
->bd_info
->bb_first_free
= first
;
1433 /* access memory sequentially: check left neighbour,
1434 * clear range and then check right neighbour
1437 left_is_free
= !mb_test_bit(first
- 1, e4b
->bd_bitmap
);
1438 block
= mb_test_and_clear_bits(e4b
->bd_bitmap
, first
, count
);
1439 if (last
+ 1 < EXT4_SB(sb
)->s_mb_maxs
[0])
1440 right_is_free
= !mb_test_bit(last
+ 1, e4b
->bd_bitmap
);
1442 if (unlikely(block
!= -1)) {
1443 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1444 ext4_fsblk_t blocknr
;
1446 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1447 blocknr
+= EXT4_C2B(EXT4_SB(sb
), block
);
1448 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1449 inode
? inode
->i_ino
: 0,
1451 "freeing already freed block "
1452 "(bit %u); block bitmap corrupt.",
1454 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
))
1455 percpu_counter_sub(&sbi
->s_freeclusters_counter
,
1456 e4b
->bd_info
->bb_free
);
1457 /* Mark the block group as corrupt. */
1458 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT
,
1459 &e4b
->bd_info
->bb_state
);
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 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1546 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1552 int start
= ex
->fe_start
;
1553 int len
= ex
->fe_len
;
1558 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1559 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1560 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1561 mb_check_buddy(e4b
);
1562 mb_mark_used_double(e4b
, start
, len
);
1564 e4b
->bd_info
->bb_free
-= len
;
1565 if (e4b
->bd_info
->bb_first_free
== start
)
1566 e4b
->bd_info
->bb_first_free
+= len
;
1568 /* let's maintain fragments counter */
1570 mlen
= !mb_test_bit(start
- 1, e4b
->bd_bitmap
);
1571 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1572 max
= !mb_test_bit(start
+ len
, e4b
->bd_bitmap
);
1574 e4b
->bd_info
->bb_fragments
++;
1575 else if (!mlen
&& !max
)
1576 e4b
->bd_info
->bb_fragments
--;
1578 /* let's maintain buddy itself */
1580 ord
= mb_find_order_for_block(e4b
, start
);
1582 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1583 /* the whole chunk may be allocated at once! */
1585 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1586 BUG_ON((start
>> ord
) >= max
);
1587 mb_set_bit(start
>> ord
, buddy
);
1588 e4b
->bd_info
->bb_counters
[ord
]--;
1595 /* store for history */
1597 ret
= len
| (ord
<< 16);
1599 /* we have to split large buddy */
1601 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1602 mb_set_bit(start
>> ord
, buddy
);
1603 e4b
->bd_info
->bb_counters
[ord
]--;
1606 cur
= (start
>> ord
) & ~1U;
1607 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1608 mb_clear_bit(cur
, buddy
);
1609 mb_clear_bit(cur
+ 1, buddy
);
1610 e4b
->bd_info
->bb_counters
[ord
]++;
1611 e4b
->bd_info
->bb_counters
[ord
]++;
1613 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1615 ext4_set_bits(e4b
->bd_bitmap
, ex
->fe_start
, len0
);
1616 mb_check_buddy(e4b
);
1622 * Must be called under group lock!
1624 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1625 struct ext4_buddy
*e4b
)
1627 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1630 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1631 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1633 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1634 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1635 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1637 /* preallocation can change ac_b_ex, thus we store actually
1638 * allocated blocks for history */
1639 ac
->ac_f_ex
= ac
->ac_b_ex
;
1641 ac
->ac_status
= AC_STATUS_FOUND
;
1642 ac
->ac_tail
= ret
& 0xffff;
1643 ac
->ac_buddy
= ret
>> 16;
1646 * take the page reference. We want the page to be pinned
1647 * so that we don't get a ext4_mb_init_cache_call for this
1648 * group until we update the bitmap. That would mean we
1649 * double allocate blocks. The reference is dropped
1650 * in ext4_mb_release_context
1652 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1653 get_page(ac
->ac_bitmap_page
);
1654 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1655 get_page(ac
->ac_buddy_page
);
1656 /* store last allocated for subsequent stream allocation */
1657 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1658 spin_lock(&sbi
->s_md_lock
);
1659 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1660 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1661 spin_unlock(&sbi
->s_md_lock
);
1666 * regular allocator, for general purposes allocation
1669 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1670 struct ext4_buddy
*e4b
,
1673 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1674 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1675 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1676 struct ext4_free_extent ex
;
1679 if (ac
->ac_status
== AC_STATUS_FOUND
)
1682 * We don't want to scan for a whole year
1684 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1685 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1686 ac
->ac_status
= AC_STATUS_BREAK
;
1691 * Haven't found good chunk so far, let's continue
1693 if (bex
->fe_len
< gex
->fe_len
)
1696 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1697 && bex
->fe_group
== e4b
->bd_group
) {
1698 /* recheck chunk's availability - we don't know
1699 * when it was found (within this lock-unlock
1701 max
= mb_find_extent(e4b
, bex
->fe_start
, gex
->fe_len
, &ex
);
1702 if (max
>= gex
->fe_len
) {
1703 ext4_mb_use_best_found(ac
, e4b
);
1710 * The routine checks whether found extent is good enough. If it is,
1711 * then the extent gets marked used and flag is set to the context
1712 * to stop scanning. Otherwise, the extent is compared with the
1713 * previous found extent and if new one is better, then it's stored
1714 * in the context. Later, the best found extent will be used, if
1715 * mballoc can't find good enough extent.
1717 * FIXME: real allocation policy is to be designed yet!
1719 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1720 struct ext4_free_extent
*ex
,
1721 struct ext4_buddy
*e4b
)
1723 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1724 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1726 BUG_ON(ex
->fe_len
<= 0);
1727 BUG_ON(ex
->fe_len
> EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1728 BUG_ON(ex
->fe_start
>= EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1729 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1734 * The special case - take what you catch first
1736 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1738 ext4_mb_use_best_found(ac
, e4b
);
1743 * Let's check whether the chuck is good enough
1745 if (ex
->fe_len
== gex
->fe_len
) {
1747 ext4_mb_use_best_found(ac
, e4b
);
1752 * If this is first found extent, just store it in the context
1754 if (bex
->fe_len
== 0) {
1760 * If new found extent is better, store it in the context
1762 if (bex
->fe_len
< gex
->fe_len
) {
1763 /* if the request isn't satisfied, any found extent
1764 * larger than previous best one is better */
1765 if (ex
->fe_len
> bex
->fe_len
)
1767 } else if (ex
->fe_len
> gex
->fe_len
) {
1768 /* if the request is satisfied, then we try to find
1769 * an extent that still satisfy the request, but is
1770 * smaller than previous one */
1771 if (ex
->fe_len
< bex
->fe_len
)
1775 ext4_mb_check_limits(ac
, e4b
, 0);
1778 static noinline_for_stack
1779 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1780 struct ext4_buddy
*e4b
)
1782 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1783 ext4_group_t group
= ex
.fe_group
;
1787 BUG_ON(ex
.fe_len
<= 0);
1788 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1792 ext4_lock_group(ac
->ac_sb
, group
);
1793 max
= mb_find_extent(e4b
, ex
.fe_start
, ex
.fe_len
, &ex
);
1797 ext4_mb_use_best_found(ac
, e4b
);
1800 ext4_unlock_group(ac
->ac_sb
, group
);
1801 ext4_mb_unload_buddy(e4b
);
1806 static noinline_for_stack
1807 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1808 struct ext4_buddy
*e4b
)
1810 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1813 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1814 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1815 struct ext4_free_extent ex
;
1817 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1819 if (grp
->bb_free
== 0)
1822 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1826 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
))) {
1827 ext4_mb_unload_buddy(e4b
);
1831 ext4_lock_group(ac
->ac_sb
, group
);
1832 max
= mb_find_extent(e4b
, ac
->ac_g_ex
.fe_start
,
1833 ac
->ac_g_ex
.fe_len
, &ex
);
1834 ex
.fe_logical
= 0xDEADFA11; /* debug value */
1836 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1839 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1841 /* use do_div to get remainder (would be 64-bit modulo) */
1842 if (do_div(start
, sbi
->s_stripe
) == 0) {
1845 ext4_mb_use_best_found(ac
, e4b
);
1847 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1848 BUG_ON(ex
.fe_len
<= 0);
1849 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1850 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1853 ext4_mb_use_best_found(ac
, e4b
);
1854 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1855 /* Sometimes, caller may want to merge even small
1856 * number of blocks to an existing extent */
1857 BUG_ON(ex
.fe_len
<= 0);
1858 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1859 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1862 ext4_mb_use_best_found(ac
, e4b
);
1864 ext4_unlock_group(ac
->ac_sb
, group
);
1865 ext4_mb_unload_buddy(e4b
);
1871 * The routine scans buddy structures (not bitmap!) from given order
1872 * to max order and tries to find big enough chunk to satisfy the req
1874 static noinline_for_stack
1875 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1876 struct ext4_buddy
*e4b
)
1878 struct super_block
*sb
= ac
->ac_sb
;
1879 struct ext4_group_info
*grp
= e4b
->bd_info
;
1885 BUG_ON(ac
->ac_2order
<= 0);
1886 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1887 if (grp
->bb_counters
[i
] == 0)
1890 buddy
= mb_find_buddy(e4b
, i
, &max
);
1891 BUG_ON(buddy
== NULL
);
1893 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1898 ac
->ac_b_ex
.fe_len
= 1 << i
;
1899 ac
->ac_b_ex
.fe_start
= k
<< i
;
1900 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1902 ext4_mb_use_best_found(ac
, e4b
);
1904 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1906 if (EXT4_SB(sb
)->s_mb_stats
)
1907 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1914 * The routine scans the group and measures all found extents.
1915 * In order to optimize scanning, caller must pass number of
1916 * free blocks in the group, so the routine can know upper limit.
1918 static noinline_for_stack
1919 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1920 struct ext4_buddy
*e4b
)
1922 struct super_block
*sb
= ac
->ac_sb
;
1923 void *bitmap
= e4b
->bd_bitmap
;
1924 struct ext4_free_extent ex
;
1928 free
= e4b
->bd_info
->bb_free
;
1931 i
= e4b
->bd_info
->bb_first_free
;
1933 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1934 i
= mb_find_next_zero_bit(bitmap
,
1935 EXT4_CLUSTERS_PER_GROUP(sb
), i
);
1936 if (i
>= EXT4_CLUSTERS_PER_GROUP(sb
)) {
1938 * IF we have corrupt bitmap, we won't find any
1939 * free blocks even though group info says we
1940 * we have free blocks
1942 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1943 "%d free clusters as per "
1944 "group info. But bitmap says 0",
1949 mb_find_extent(e4b
, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1950 BUG_ON(ex
.fe_len
<= 0);
1951 if (free
< ex
.fe_len
) {
1952 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1953 "%d free clusters as per "
1954 "group info. But got %d blocks",
1957 * The number of free blocks differs. This mostly
1958 * indicate that the bitmap is corrupt. So exit
1959 * without claiming the space.
1963 ex
.fe_logical
= 0xDEADC0DE; /* debug value */
1964 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1970 ext4_mb_check_limits(ac
, e4b
, 1);
1974 * This is a special case for storages like raid5
1975 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1977 static noinline_for_stack
1978 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1979 struct ext4_buddy
*e4b
)
1981 struct super_block
*sb
= ac
->ac_sb
;
1982 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1983 void *bitmap
= e4b
->bd_bitmap
;
1984 struct ext4_free_extent ex
;
1985 ext4_fsblk_t first_group_block
;
1990 BUG_ON(sbi
->s_stripe
== 0);
1992 /* find first stripe-aligned block in group */
1993 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1995 a
= first_group_block
+ sbi
->s_stripe
- 1;
1996 do_div(a
, sbi
->s_stripe
);
1997 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1999 while (i
< EXT4_CLUSTERS_PER_GROUP(sb
)) {
2000 if (!mb_test_bit(i
, bitmap
)) {
2001 max
= mb_find_extent(e4b
, i
, sbi
->s_stripe
, &ex
);
2002 if (max
>= sbi
->s_stripe
) {
2004 ex
.fe_logical
= 0xDEADF00D; /* debug value */
2006 ext4_mb_use_best_found(ac
, e4b
);
2014 /* This is now called BEFORE we load the buddy bitmap. */
2015 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
2016 ext4_group_t group
, int cr
)
2018 unsigned free
, fragments
;
2019 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
2020 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
2022 BUG_ON(cr
< 0 || cr
>= 4);
2024 free
= grp
->bb_free
;
2027 if (cr
<= 2 && free
< ac
->ac_g_ex
.fe_len
)
2030 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp
)))
2033 /* We only do this if the grp has never been initialized */
2034 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
2035 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
);
2040 fragments
= grp
->bb_fragments
;
2046 BUG_ON(ac
->ac_2order
== 0);
2048 /* Avoid using the first bg of a flexgroup for data files */
2049 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
2050 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
2051 ((group
% flex_size
) == 0))
2054 if ((ac
->ac_2order
> ac
->ac_sb
->s_blocksize_bits
+1) ||
2055 (free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2058 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
2063 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2067 if (free
>= ac
->ac_g_ex
.fe_len
)
2079 static noinline_for_stack
int
2080 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
2082 ext4_group_t ngroups
, group
, i
;
2085 struct ext4_sb_info
*sbi
;
2086 struct super_block
*sb
;
2087 struct ext4_buddy e4b
;
2091 ngroups
= ext4_get_groups_count(sb
);
2092 /* non-extent files are limited to low blocks/groups */
2093 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
2094 ngroups
= sbi
->s_blockfile_groups
;
2096 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
2098 /* first, try the goal */
2099 err
= ext4_mb_find_by_goal(ac
, &e4b
);
2100 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
2103 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2107 * ac->ac2_order is set only if the fe_len is a power of 2
2108 * if ac2_order is set we also set criteria to 0 so that we
2109 * try exact allocation using buddy.
2111 i
= fls(ac
->ac_g_ex
.fe_len
);
2114 * We search using buddy data only if the order of the request
2115 * is greater than equal to the sbi_s_mb_order2_reqs
2116 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2118 if (i
>= sbi
->s_mb_order2_reqs
) {
2120 * This should tell if fe_len is exactly power of 2
2122 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2123 ac
->ac_2order
= i
- 1;
2126 /* if stream allocation is enabled, use global goal */
2127 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2128 /* TBD: may be hot point */
2129 spin_lock(&sbi
->s_md_lock
);
2130 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2131 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2132 spin_unlock(&sbi
->s_md_lock
);
2135 /* Let's just scan groups to find more-less suitable blocks */
2136 cr
= ac
->ac_2order
? 0 : 1;
2138 * cr == 0 try to get exact allocation,
2139 * cr == 3 try to get anything
2142 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2143 ac
->ac_criteria
= cr
;
2145 * searching for the right group start
2146 * from the goal value specified
2148 group
= ac
->ac_g_ex
.fe_group
;
2150 for (i
= 0; i
< ngroups
; group
++, i
++) {
2153 * Artificially restricted ngroups for non-extent
2154 * files makes group > ngroups possible on first loop.
2156 if (group
>= ngroups
)
2159 /* This now checks without needing the buddy page */
2160 if (!ext4_mb_good_group(ac
, group
, cr
))
2163 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2167 ext4_lock_group(sb
, group
);
2170 * We need to check again after locking the
2173 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2174 ext4_unlock_group(sb
, group
);
2175 ext4_mb_unload_buddy(&e4b
);
2179 ac
->ac_groups_scanned
++;
2180 if (cr
== 0 && ac
->ac_2order
< sb
->s_blocksize_bits
+2)
2181 ext4_mb_simple_scan_group(ac
, &e4b
);
2182 else if (cr
== 1 && sbi
->s_stripe
&&
2183 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2184 ext4_mb_scan_aligned(ac
, &e4b
);
2186 ext4_mb_complex_scan_group(ac
, &e4b
);
2188 ext4_unlock_group(sb
, group
);
2189 ext4_mb_unload_buddy(&e4b
);
2191 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2196 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2197 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2199 * We've been searching too long. Let's try to allocate
2200 * the best chunk we've found so far
2203 ext4_mb_try_best_found(ac
, &e4b
);
2204 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2206 * Someone more lucky has already allocated it.
2207 * The only thing we can do is just take first
2209 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2211 ac
->ac_b_ex
.fe_group
= 0;
2212 ac
->ac_b_ex
.fe_start
= 0;
2213 ac
->ac_b_ex
.fe_len
= 0;
2214 ac
->ac_status
= AC_STATUS_CONTINUE
;
2215 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2217 atomic_inc(&sbi
->s_mb_lost_chunks
);
2225 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2227 struct super_block
*sb
= seq
->private;
2230 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2233 return (void *) ((unsigned long) group
);
2236 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2238 struct super_block
*sb
= seq
->private;
2242 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2245 return (void *) ((unsigned long) group
);
2248 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2250 struct super_block
*sb
= seq
->private;
2251 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2253 int err
, buddy_loaded
= 0;
2254 struct ext4_buddy e4b
;
2255 struct ext4_group_info
*grinfo
;
2257 struct ext4_group_info info
;
2258 ext4_grpblk_t counters
[16];
2263 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2264 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2265 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2266 "group", "free", "frags", "first",
2267 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2268 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2270 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2271 sizeof(struct ext4_group_info
);
2272 grinfo
= ext4_get_group_info(sb
, group
);
2273 /* Load the group info in memory only if not already loaded. */
2274 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo
))) {
2275 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2277 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2283 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2286 ext4_mb_unload_buddy(&e4b
);
2288 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2289 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2290 for (i
= 0; i
<= 13; i
++)
2291 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2292 sg
.info
.bb_counters
[i
] : 0);
2293 seq_printf(seq
, " ]\n");
2298 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2302 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2303 .start
= ext4_mb_seq_groups_start
,
2304 .next
= ext4_mb_seq_groups_next
,
2305 .stop
= ext4_mb_seq_groups_stop
,
2306 .show
= ext4_mb_seq_groups_show
,
2309 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2311 struct super_block
*sb
= PDE_DATA(inode
);
2314 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2316 struct seq_file
*m
= file
->private_data
;
2323 static const struct file_operations ext4_mb_seq_groups_fops
= {
2324 .owner
= THIS_MODULE
,
2325 .open
= ext4_mb_seq_groups_open
,
2327 .llseek
= seq_lseek
,
2328 .release
= seq_release
,
2331 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2333 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2334 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2341 * Allocate the top-level s_group_info array for the specified number
2344 int ext4_mb_alloc_groupinfo(struct super_block
*sb
, ext4_group_t ngroups
)
2346 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2348 struct ext4_group_info
***new_groupinfo
;
2350 size
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2351 EXT4_DESC_PER_BLOCK_BITS(sb
);
2352 if (size
<= sbi
->s_group_info_size
)
2355 size
= roundup_pow_of_two(sizeof(*sbi
->s_group_info
) * size
);
2356 new_groupinfo
= ext4_kvzalloc(size
, GFP_KERNEL
);
2357 if (!new_groupinfo
) {
2358 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy meta group");
2361 if (sbi
->s_group_info
) {
2362 memcpy(new_groupinfo
, sbi
->s_group_info
,
2363 sbi
->s_group_info_size
* sizeof(*sbi
->s_group_info
));
2364 kvfree(sbi
->s_group_info
);
2366 sbi
->s_group_info
= new_groupinfo
;
2367 sbi
->s_group_info_size
= size
/ sizeof(*sbi
->s_group_info
);
2368 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2369 sbi
->s_group_info_size
);
2373 /* Create and initialize ext4_group_info data for the given group. */
2374 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2375 struct ext4_group_desc
*desc
)
2379 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2380 struct ext4_group_info
**meta_group_info
;
2381 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2384 * First check if this group is the first of a reserved block.
2385 * If it's true, we have to allocate a new table of pointers
2386 * to ext4_group_info structures
2388 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2389 metalen
= sizeof(*meta_group_info
) <<
2390 EXT4_DESC_PER_BLOCK_BITS(sb
);
2391 meta_group_info
= kmalloc(metalen
, GFP_NOFS
);
2392 if (meta_group_info
== NULL
) {
2393 ext4_msg(sb
, KERN_ERR
, "can't allocate mem "
2394 "for a buddy group");
2395 goto exit_meta_group_info
;
2397 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2402 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2403 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2405 meta_group_info
[i
] = kmem_cache_zalloc(cachep
, GFP_NOFS
);
2406 if (meta_group_info
[i
] == NULL
) {
2407 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy mem");
2408 goto exit_group_info
;
2410 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2411 &(meta_group_info
[i
]->bb_state
));
2414 * initialize bb_free to be able to skip
2415 * empty groups without initialization
2417 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2418 meta_group_info
[i
]->bb_free
=
2419 ext4_free_clusters_after_init(sb
, group
, desc
);
2421 meta_group_info
[i
]->bb_free
=
2422 ext4_free_group_clusters(sb
, desc
);
2425 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2426 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2427 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2428 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2432 struct buffer_head
*bh
;
2433 meta_group_info
[i
]->bb_bitmap
=
2434 kmalloc(sb
->s_blocksize
, GFP_NOFS
);
2435 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2436 bh
= ext4_read_block_bitmap(sb
, group
);
2438 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2447 /* If a meta_group_info table has been allocated, release it now */
2448 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2449 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2450 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] = NULL
;
2452 exit_meta_group_info
:
2454 } /* ext4_mb_add_groupinfo */
2456 static int ext4_mb_init_backend(struct super_block
*sb
)
2458 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2460 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2462 struct ext4_group_desc
*desc
;
2463 struct kmem_cache
*cachep
;
2465 err
= ext4_mb_alloc_groupinfo(sb
, ngroups
);
2469 sbi
->s_buddy_cache
= new_inode(sb
);
2470 if (sbi
->s_buddy_cache
== NULL
) {
2471 ext4_msg(sb
, KERN_ERR
, "can't get new inode");
2474 /* To avoid potentially colliding with an valid on-disk inode number,
2475 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2476 * not in the inode hash, so it should never be found by iget(), but
2477 * this will avoid confusion if it ever shows up during debugging. */
2478 sbi
->s_buddy_cache
->i_ino
= EXT4_BAD_INO
;
2479 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2480 for (i
= 0; i
< ngroups
; i
++) {
2481 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2483 ext4_msg(sb
, KERN_ERR
, "can't read descriptor %u", i
);
2486 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2493 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2495 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2496 i
= sbi
->s_group_info_size
;
2498 kfree(sbi
->s_group_info
[i
]);
2499 iput(sbi
->s_buddy_cache
);
2501 kvfree(sbi
->s_group_info
);
2505 static void ext4_groupinfo_destroy_slabs(void)
2509 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2510 if (ext4_groupinfo_caches
[i
])
2511 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2512 ext4_groupinfo_caches
[i
] = NULL
;
2516 static int ext4_groupinfo_create_slab(size_t size
)
2518 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2520 int blocksize_bits
= order_base_2(size
);
2521 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2522 struct kmem_cache
*cachep
;
2524 if (cache_index
>= NR_GRPINFO_CACHES
)
2527 if (unlikely(cache_index
< 0))
2530 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2531 if (ext4_groupinfo_caches
[cache_index
]) {
2532 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2533 return 0; /* Already created */
2536 slab_size
= offsetof(struct ext4_group_info
,
2537 bb_counters
[blocksize_bits
+ 2]);
2539 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2540 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2543 ext4_groupinfo_caches
[cache_index
] = cachep
;
2545 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2548 "EXT4-fs: no memory for groupinfo slab cache\n");
2555 int ext4_mb_init(struct super_block
*sb
)
2557 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2563 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2565 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2566 if (sbi
->s_mb_offsets
== NULL
) {
2571 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2572 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2573 if (sbi
->s_mb_maxs
== NULL
) {
2578 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2582 /* order 0 is regular bitmap */
2583 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2584 sbi
->s_mb_offsets
[0] = 0;
2588 max
= sb
->s_blocksize
<< 2;
2590 sbi
->s_mb_offsets
[i
] = offset
;
2591 sbi
->s_mb_maxs
[i
] = max
;
2592 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2595 } while (i
<= sb
->s_blocksize_bits
+ 1);
2597 spin_lock_init(&sbi
->s_md_lock
);
2598 spin_lock_init(&sbi
->s_bal_lock
);
2600 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2601 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2602 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2603 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2604 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2606 * The default group preallocation is 512, which for 4k block
2607 * sizes translates to 2 megabytes. However for bigalloc file
2608 * systems, this is probably too big (i.e, if the cluster size
2609 * is 1 megabyte, then group preallocation size becomes half a
2610 * gigabyte!). As a default, we will keep a two megabyte
2611 * group pralloc size for cluster sizes up to 64k, and after
2612 * that, we will force a minimum group preallocation size of
2613 * 32 clusters. This translates to 8 megs when the cluster
2614 * size is 256k, and 32 megs when the cluster size is 1 meg,
2615 * which seems reasonable as a default.
2617 sbi
->s_mb_group_prealloc
= max(MB_DEFAULT_GROUP_PREALLOC
>>
2618 sbi
->s_cluster_bits
, 32);
2620 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2621 * to the lowest multiple of s_stripe which is bigger than
2622 * the s_mb_group_prealloc as determined above. We want
2623 * the preallocation size to be an exact multiple of the
2624 * RAID stripe size so that preallocations don't fragment
2627 if (sbi
->s_stripe
> 1) {
2628 sbi
->s_mb_group_prealloc
= roundup(
2629 sbi
->s_mb_group_prealloc
, sbi
->s_stripe
);
2632 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2633 if (sbi
->s_locality_groups
== NULL
) {
2637 for_each_possible_cpu(i
) {
2638 struct ext4_locality_group
*lg
;
2639 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2640 mutex_init(&lg
->lg_mutex
);
2641 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2642 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2643 spin_lock_init(&lg
->lg_prealloc_lock
);
2646 /* init file for buddy data */
2647 ret
= ext4_mb_init_backend(sb
);
2649 goto out_free_locality_groups
;
2652 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2653 &ext4_mb_seq_groups_fops
, sb
);
2657 out_free_locality_groups
:
2658 free_percpu(sbi
->s_locality_groups
);
2659 sbi
->s_locality_groups
= NULL
;
2661 kfree(sbi
->s_mb_offsets
);
2662 sbi
->s_mb_offsets
= NULL
;
2663 kfree(sbi
->s_mb_maxs
);
2664 sbi
->s_mb_maxs
= NULL
;
2668 /* need to called with the ext4 group lock held */
2669 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2671 struct ext4_prealloc_space
*pa
;
2672 struct list_head
*cur
, *tmp
;
2675 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2676 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2677 list_del(&pa
->pa_group_list
);
2679 kmem_cache_free(ext4_pspace_cachep
, pa
);
2682 mb_debug(1, "mballoc: %u PAs left\n", count
);
2686 int ext4_mb_release(struct super_block
*sb
)
2688 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2690 int num_meta_group_infos
;
2691 struct ext4_group_info
*grinfo
;
2692 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2693 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2696 remove_proc_entry("mb_groups", sbi
->s_proc
);
2698 if (sbi
->s_group_info
) {
2699 for (i
= 0; i
< ngroups
; i
++) {
2700 grinfo
= ext4_get_group_info(sb
, i
);
2702 kfree(grinfo
->bb_bitmap
);
2704 ext4_lock_group(sb
, i
);
2705 ext4_mb_cleanup_pa(grinfo
);
2706 ext4_unlock_group(sb
, i
);
2707 kmem_cache_free(cachep
, grinfo
);
2709 num_meta_group_infos
= (ngroups
+
2710 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2711 EXT4_DESC_PER_BLOCK_BITS(sb
);
2712 for (i
= 0; i
< num_meta_group_infos
; i
++)
2713 kfree(sbi
->s_group_info
[i
]);
2714 kvfree(sbi
->s_group_info
);
2716 kfree(sbi
->s_mb_offsets
);
2717 kfree(sbi
->s_mb_maxs
);
2718 iput(sbi
->s_buddy_cache
);
2719 if (sbi
->s_mb_stats
) {
2720 ext4_msg(sb
, KERN_INFO
,
2721 "mballoc: %u blocks %u reqs (%u success)",
2722 atomic_read(&sbi
->s_bal_allocated
),
2723 atomic_read(&sbi
->s_bal_reqs
),
2724 atomic_read(&sbi
->s_bal_success
));
2725 ext4_msg(sb
, KERN_INFO
,
2726 "mballoc: %u extents scanned, %u goal hits, "
2727 "%u 2^N hits, %u breaks, %u lost",
2728 atomic_read(&sbi
->s_bal_ex_scanned
),
2729 atomic_read(&sbi
->s_bal_goals
),
2730 atomic_read(&sbi
->s_bal_2orders
),
2731 atomic_read(&sbi
->s_bal_breaks
),
2732 atomic_read(&sbi
->s_mb_lost_chunks
));
2733 ext4_msg(sb
, KERN_INFO
,
2734 "mballoc: %lu generated and it took %Lu",
2735 sbi
->s_mb_buddies_generated
,
2736 sbi
->s_mb_generation_time
);
2737 ext4_msg(sb
, KERN_INFO
,
2738 "mballoc: %u preallocated, %u discarded",
2739 atomic_read(&sbi
->s_mb_preallocated
),
2740 atomic_read(&sbi
->s_mb_discarded
));
2743 free_percpu(sbi
->s_locality_groups
);
2748 static inline int ext4_issue_discard(struct super_block
*sb
,
2749 ext4_group_t block_group
, ext4_grpblk_t cluster
, int count
)
2751 ext4_fsblk_t discard_block
;
2753 discard_block
= (EXT4_C2B(EXT4_SB(sb
), cluster
) +
2754 ext4_group_first_block_no(sb
, block_group
));
2755 count
= EXT4_C2B(EXT4_SB(sb
), count
);
2756 trace_ext4_discard_blocks(sb
,
2757 (unsigned long long) discard_block
, count
);
2758 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, 0);
2762 * This function is called by the jbd2 layer once the commit has finished,
2763 * so we know we can free the blocks that were released with that commit.
2765 static void ext4_free_data_callback(struct super_block
*sb
,
2766 struct ext4_journal_cb_entry
*jce
,
2769 struct ext4_free_data
*entry
= (struct ext4_free_data
*)jce
;
2770 struct ext4_buddy e4b
;
2771 struct ext4_group_info
*db
;
2772 int err
, count
= 0, count2
= 0;
2774 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2775 entry
->efd_count
, entry
->efd_group
, entry
);
2777 if (test_opt(sb
, DISCARD
)) {
2778 err
= ext4_issue_discard(sb
, entry
->efd_group
,
2779 entry
->efd_start_cluster
,
2781 if (err
&& err
!= -EOPNOTSUPP
)
2782 ext4_msg(sb
, KERN_WARNING
, "discard request in"
2783 " group:%d block:%d count:%d failed"
2784 " with %d", entry
->efd_group
,
2785 entry
->efd_start_cluster
,
2786 entry
->efd_count
, err
);
2789 err
= ext4_mb_load_buddy(sb
, entry
->efd_group
, &e4b
);
2790 /* we expect to find existing buddy because it's pinned */
2795 /* there are blocks to put in buddy to make them really free */
2796 count
+= entry
->efd_count
;
2798 ext4_lock_group(sb
, entry
->efd_group
);
2799 /* Take it out of per group rb tree */
2800 rb_erase(&entry
->efd_node
, &(db
->bb_free_root
));
2801 mb_free_blocks(NULL
, &e4b
, entry
->efd_start_cluster
, entry
->efd_count
);
2804 * Clear the trimmed flag for the group so that the next
2805 * ext4_trim_fs can trim it.
2806 * If the volume is mounted with -o discard, online discard
2807 * is supported and the free blocks will be trimmed online.
2809 if (!test_opt(sb
, DISCARD
))
2810 EXT4_MB_GRP_CLEAR_TRIMMED(db
);
2812 if (!db
->bb_free_root
.rb_node
) {
2813 /* No more items in the per group rb tree
2814 * balance refcounts from ext4_mb_free_metadata()
2816 page_cache_release(e4b
.bd_buddy_page
);
2817 page_cache_release(e4b
.bd_bitmap_page
);
2819 ext4_unlock_group(sb
, entry
->efd_group
);
2820 kmem_cache_free(ext4_free_data_cachep
, entry
);
2821 ext4_mb_unload_buddy(&e4b
);
2823 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2826 int __init
ext4_init_mballoc(void)
2828 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2829 SLAB_RECLAIM_ACCOUNT
);
2830 if (ext4_pspace_cachep
== NULL
)
2833 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2834 SLAB_RECLAIM_ACCOUNT
);
2835 if (ext4_ac_cachep
== NULL
) {
2836 kmem_cache_destroy(ext4_pspace_cachep
);
2840 ext4_free_data_cachep
= KMEM_CACHE(ext4_free_data
,
2841 SLAB_RECLAIM_ACCOUNT
);
2842 if (ext4_free_data_cachep
== NULL
) {
2843 kmem_cache_destroy(ext4_pspace_cachep
);
2844 kmem_cache_destroy(ext4_ac_cachep
);
2850 void ext4_exit_mballoc(void)
2853 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2854 * before destroying the slab cache.
2857 kmem_cache_destroy(ext4_pspace_cachep
);
2858 kmem_cache_destroy(ext4_ac_cachep
);
2859 kmem_cache_destroy(ext4_free_data_cachep
);
2860 ext4_groupinfo_destroy_slabs();
2865 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2866 * Returns 0 if success or error code
2868 static noinline_for_stack
int
2869 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2870 handle_t
*handle
, unsigned int reserv_clstrs
)
2872 struct buffer_head
*bitmap_bh
= NULL
;
2873 struct ext4_group_desc
*gdp
;
2874 struct buffer_head
*gdp_bh
;
2875 struct ext4_sb_info
*sbi
;
2876 struct super_block
*sb
;
2880 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2881 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2887 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2891 BUFFER_TRACE(bitmap_bh
, "getting write access");
2892 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2897 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2901 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2902 ext4_free_group_clusters(sb
, gdp
));
2904 BUFFER_TRACE(gdp_bh
, "get_write_access");
2905 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2909 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
2911 len
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
2912 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2913 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2914 "fs metadata", block
, block
+len
);
2915 /* File system mounted not to panic on error
2916 * Fix the bitmap and repeat the block allocation
2917 * We leak some of the blocks here.
2919 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2920 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2921 ac
->ac_b_ex
.fe_len
);
2922 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2923 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2929 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2930 #ifdef AGGRESSIVE_CHECK
2933 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2934 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2935 bitmap_bh
->b_data
));
2939 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2940 ac
->ac_b_ex
.fe_len
);
2941 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2942 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2943 ext4_free_group_clusters_set(sb
, gdp
,
2944 ext4_free_clusters_after_init(sb
,
2945 ac
->ac_b_ex
.fe_group
, gdp
));
2947 len
= ext4_free_group_clusters(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
2948 ext4_free_group_clusters_set(sb
, gdp
, len
);
2949 ext4_block_bitmap_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
, bitmap_bh
);
2950 ext4_group_desc_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
);
2952 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2953 percpu_counter_sub(&sbi
->s_freeclusters_counter
, ac
->ac_b_ex
.fe_len
);
2955 * Now reduce the dirty block count also. Should not go negative
2957 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2958 /* release all the reserved blocks if non delalloc */
2959 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
2962 if (sbi
->s_log_groups_per_flex
) {
2963 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2964 ac
->ac_b_ex
.fe_group
);
2965 atomic64_sub(ac
->ac_b_ex
.fe_len
,
2966 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
2969 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2972 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
2980 * here we normalize request for locality group
2981 * Group request are normalized to s_mb_group_prealloc, which goes to
2982 * s_strip if we set the same via mount option.
2983 * s_mb_group_prealloc can be configured via
2984 * /sys/fs/ext4/<partition>/mb_group_prealloc
2986 * XXX: should we try to preallocate more than the group has now?
2988 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2990 struct super_block
*sb
= ac
->ac_sb
;
2991 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2994 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2995 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2996 current
->pid
, ac
->ac_g_ex
.fe_len
);
3000 * Normalization means making request better in terms of
3001 * size and alignment
3003 static noinline_for_stack
void
3004 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
3005 struct ext4_allocation_request
*ar
)
3007 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3010 loff_t size
, start_off
;
3011 loff_t orig_size __maybe_unused
;
3013 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3014 struct ext4_prealloc_space
*pa
;
3016 /* do normalize only data requests, metadata requests
3017 do not need preallocation */
3018 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3021 /* sometime caller may want exact blocks */
3022 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3025 /* caller may indicate that preallocation isn't
3026 * required (it's a tail, for example) */
3027 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3030 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3031 ext4_mb_normalize_group_request(ac
);
3035 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3037 /* first, let's learn actual file size
3038 * given current request is allocated */
3039 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
3040 size
= size
<< bsbits
;
3041 if (size
< i_size_read(ac
->ac_inode
))
3042 size
= i_size_read(ac
->ac_inode
);
3045 /* max size of free chunks */
3048 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3049 (req <= (size) || max <= (chunk_size))
3051 /* first, try to predict filesize */
3052 /* XXX: should this table be tunable? */
3054 if (size
<= 16 * 1024) {
3056 } else if (size
<= 32 * 1024) {
3058 } else if (size
<= 64 * 1024) {
3060 } else if (size
<= 128 * 1024) {
3062 } else if (size
<= 256 * 1024) {
3064 } else if (size
<= 512 * 1024) {
3066 } else if (size
<= 1024 * 1024) {
3068 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3069 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3070 (21 - bsbits
)) << 21;
3071 size
= 2 * 1024 * 1024;
3072 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3073 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3074 (22 - bsbits
)) << 22;
3075 size
= 4 * 1024 * 1024;
3076 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3077 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3078 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3079 (23 - bsbits
)) << 23;
3080 size
= 8 * 1024 * 1024;
3082 start_off
= (loff_t
) ac
->ac_o_ex
.fe_logical
<< bsbits
;
3083 size
= (loff_t
) EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3084 ac
->ac_o_ex
.fe_len
) << bsbits
;
3086 size
= size
>> bsbits
;
3087 start
= start_off
>> bsbits
;
3089 /* don't cover already allocated blocks in selected range */
3090 if (ar
->pleft
&& start
<= ar
->lleft
) {
3091 size
-= ar
->lleft
+ 1 - start
;
3092 start
= ar
->lleft
+ 1;
3094 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3095 size
-= start
+ size
- ar
->lright
;
3099 /* check we don't cross already preallocated blocks */
3101 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3106 spin_lock(&pa
->pa_lock
);
3107 if (pa
->pa_deleted
) {
3108 spin_unlock(&pa
->pa_lock
);
3112 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3115 /* PA must not overlap original request */
3116 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3117 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3119 /* skip PAs this normalized request doesn't overlap with */
3120 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
3121 spin_unlock(&pa
->pa_lock
);
3124 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3126 /* adjust start or end to be adjacent to this pa */
3127 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3128 BUG_ON(pa_end
< start
);
3130 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3131 BUG_ON(pa
->pa_lstart
> end
);
3132 end
= pa
->pa_lstart
;
3134 spin_unlock(&pa
->pa_lock
);
3139 /* XXX: extra loop to check we really don't overlap preallocations */
3141 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3144 spin_lock(&pa
->pa_lock
);
3145 if (pa
->pa_deleted
== 0) {
3146 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3148 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3150 spin_unlock(&pa
->pa_lock
);
3154 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3155 start
> ac
->ac_o_ex
.fe_logical
) {
3156 ext4_msg(ac
->ac_sb
, KERN_ERR
,
3157 "start %lu, size %lu, fe_logical %lu",
3158 (unsigned long) start
, (unsigned long) size
,
3159 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3162 BUG_ON(size
<= 0 || size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3164 /* now prepare goal request */
3166 /* XXX: is it better to align blocks WRT to logical
3167 * placement or satisfy big request as is */
3168 ac
->ac_g_ex
.fe_logical
= start
;
3169 ac
->ac_g_ex
.fe_len
= EXT4_NUM_B2C(sbi
, size
);
3171 /* define goal start in order to merge */
3172 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3173 /* merge to the right */
3174 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3175 &ac
->ac_f_ex
.fe_group
,
3176 &ac
->ac_f_ex
.fe_start
);
3177 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3179 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3180 /* merge to the left */
3181 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3182 &ac
->ac_f_ex
.fe_group
,
3183 &ac
->ac_f_ex
.fe_start
);
3184 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3187 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3188 (unsigned) orig_size
, (unsigned) start
);
3191 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3193 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3195 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3196 atomic_inc(&sbi
->s_bal_reqs
);
3197 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3198 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3199 atomic_inc(&sbi
->s_bal_success
);
3200 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3201 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3202 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3203 atomic_inc(&sbi
->s_bal_goals
);
3204 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3205 atomic_inc(&sbi
->s_bal_breaks
);
3208 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3209 trace_ext4_mballoc_alloc(ac
);
3211 trace_ext4_mballoc_prealloc(ac
);
3215 * Called on failure; free up any blocks from the inode PA for this
3216 * context. We don't need this for MB_GROUP_PA because we only change
3217 * pa_free in ext4_mb_release_context(), but on failure, we've already
3218 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3220 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3222 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3223 struct ext4_buddy e4b
;
3227 if (ac
->ac_f_ex
.fe_len
== 0)
3229 err
= ext4_mb_load_buddy(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
, &e4b
);
3232 * This should never happen since we pin the
3233 * pages in the ext4_allocation_context so
3234 * ext4_mb_load_buddy() should never fail.
3236 WARN(1, "mb_load_buddy failed (%d)", err
);
3239 ext4_lock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3240 mb_free_blocks(ac
->ac_inode
, &e4b
, ac
->ac_f_ex
.fe_start
,
3241 ac
->ac_f_ex
.fe_len
);
3242 ext4_unlock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3243 ext4_mb_unload_buddy(&e4b
);
3246 if (pa
->pa_type
== MB_INODE_PA
)
3247 pa
->pa_free
+= ac
->ac_b_ex
.fe_len
;
3251 * use blocks preallocated to inode
3253 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3254 struct ext4_prealloc_space
*pa
)
3256 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3261 /* found preallocated blocks, use them */
3262 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3263 end
= min(pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
),
3264 start
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
));
3265 len
= EXT4_NUM_B2C(sbi
, end
- start
);
3266 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3267 &ac
->ac_b_ex
.fe_start
);
3268 ac
->ac_b_ex
.fe_len
= len
;
3269 ac
->ac_status
= AC_STATUS_FOUND
;
3272 BUG_ON(start
< pa
->pa_pstart
);
3273 BUG_ON(end
> pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
));
3274 BUG_ON(pa
->pa_free
< len
);
3277 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3281 * use blocks preallocated to locality group
3283 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3284 struct ext4_prealloc_space
*pa
)
3286 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3288 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3289 &ac
->ac_b_ex
.fe_group
,
3290 &ac
->ac_b_ex
.fe_start
);
3291 ac
->ac_b_ex
.fe_len
= len
;
3292 ac
->ac_status
= AC_STATUS_FOUND
;
3295 /* we don't correct pa_pstart or pa_plen here to avoid
3296 * possible race when the group is being loaded concurrently
3297 * instead we correct pa later, after blocks are marked
3298 * in on-disk bitmap -- see ext4_mb_release_context()
3299 * Other CPUs are prevented from allocating from this pa by lg_mutex
3301 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3305 * Return the prealloc space that have minimal distance
3306 * from the goal block. @cpa is the prealloc
3307 * space that is having currently known minimal distance
3308 * from the goal block.
3310 static struct ext4_prealloc_space
*
3311 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3312 struct ext4_prealloc_space
*pa
,
3313 struct ext4_prealloc_space
*cpa
)
3315 ext4_fsblk_t cur_distance
, new_distance
;
3318 atomic_inc(&pa
->pa_count
);
3321 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3322 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3324 if (cur_distance
<= new_distance
)
3327 /* drop the previous reference */
3328 atomic_dec(&cpa
->pa_count
);
3329 atomic_inc(&pa
->pa_count
);
3334 * search goal blocks in preallocated space
3336 static noinline_for_stack
int
3337 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3339 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3341 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3342 struct ext4_locality_group
*lg
;
3343 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3344 ext4_fsblk_t goal_block
;
3346 /* only data can be preallocated */
3347 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3350 /* first, try per-file preallocation */
3352 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3354 /* all fields in this condition don't change,
3355 * so we can skip locking for them */
3356 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3357 ac
->ac_o_ex
.fe_logical
>= (pa
->pa_lstart
+
3358 EXT4_C2B(sbi
, pa
->pa_len
)))
3361 /* non-extent files can't have physical blocks past 2^32 */
3362 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3363 (pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
) >
3364 EXT4_MAX_BLOCK_FILE_PHYS
))
3367 /* found preallocated blocks, use them */
3368 spin_lock(&pa
->pa_lock
);
3369 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3370 atomic_inc(&pa
->pa_count
);
3371 ext4_mb_use_inode_pa(ac
, pa
);
3372 spin_unlock(&pa
->pa_lock
);
3373 ac
->ac_criteria
= 10;
3377 spin_unlock(&pa
->pa_lock
);
3381 /* can we use group allocation? */
3382 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3385 /* inode may have no locality group for some reason */
3389 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3390 if (order
> PREALLOC_TB_SIZE
- 1)
3391 /* The max size of hash table is PREALLOC_TB_SIZE */
3392 order
= PREALLOC_TB_SIZE
- 1;
3394 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3396 * search for the prealloc space that is having
3397 * minimal distance from the goal block.
3399 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3401 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3403 spin_lock(&pa
->pa_lock
);
3404 if (pa
->pa_deleted
== 0 &&
3405 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3407 cpa
= ext4_mb_check_group_pa(goal_block
,
3410 spin_unlock(&pa
->pa_lock
);
3415 ext4_mb_use_group_pa(ac
, cpa
);
3416 ac
->ac_criteria
= 20;
3423 * the function goes through all block freed in the group
3424 * but not yet committed and marks them used in in-core bitmap.
3425 * buddy must be generated from this bitmap
3426 * Need to be called with the ext4 group lock held
3428 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3432 struct ext4_group_info
*grp
;
3433 struct ext4_free_data
*entry
;
3435 grp
= ext4_get_group_info(sb
, group
);
3436 n
= rb_first(&(grp
->bb_free_root
));
3439 entry
= rb_entry(n
, struct ext4_free_data
, efd_node
);
3440 ext4_set_bits(bitmap
, entry
->efd_start_cluster
, entry
->efd_count
);
3447 * the function goes through all preallocation in this group and marks them
3448 * used in in-core bitmap. buddy must be generated from this bitmap
3449 * Need to be called with ext4 group lock held
3451 static noinline_for_stack
3452 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3455 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3456 struct ext4_prealloc_space
*pa
;
3457 struct list_head
*cur
;
3458 ext4_group_t groupnr
;
3459 ext4_grpblk_t start
;
3460 int preallocated
= 0;
3463 /* all form of preallocation discards first load group,
3464 * so the only competing code is preallocation use.
3465 * we don't need any locking here
3466 * notice we do NOT ignore preallocations with pa_deleted
3467 * otherwise we could leave used blocks available for
3468 * allocation in buddy when concurrent ext4_mb_put_pa()
3469 * is dropping preallocation
3471 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3472 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3473 spin_lock(&pa
->pa_lock
);
3474 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3477 spin_unlock(&pa
->pa_lock
);
3478 if (unlikely(len
== 0))
3480 BUG_ON(groupnr
!= group
);
3481 ext4_set_bits(bitmap
, start
, len
);
3482 preallocated
+= len
;
3484 mb_debug(1, "prellocated %u for group %u\n", preallocated
, group
);
3487 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3489 struct ext4_prealloc_space
*pa
;
3490 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3492 BUG_ON(atomic_read(&pa
->pa_count
));
3493 BUG_ON(pa
->pa_deleted
== 0);
3494 kmem_cache_free(ext4_pspace_cachep
, pa
);
3498 * drops a reference to preallocated space descriptor
3499 * if this was the last reference and the space is consumed
3501 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3502 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3505 ext4_fsblk_t grp_blk
;
3507 /* in this short window concurrent discard can set pa_deleted */
3508 spin_lock(&pa
->pa_lock
);
3509 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0) {
3510 spin_unlock(&pa
->pa_lock
);
3514 if (pa
->pa_deleted
== 1) {
3515 spin_unlock(&pa
->pa_lock
);
3520 spin_unlock(&pa
->pa_lock
);
3522 grp_blk
= pa
->pa_pstart
;
3524 * If doing group-based preallocation, pa_pstart may be in the
3525 * next group when pa is used up
3527 if (pa
->pa_type
== MB_GROUP_PA
)
3530 grp
= ext4_get_group_number(sb
, grp_blk
);
3535 * P1 (buddy init) P2 (regular allocation)
3536 * find block B in PA
3537 * copy on-disk bitmap to buddy
3538 * mark B in on-disk bitmap
3539 * drop PA from group
3540 * mark all PAs in buddy
3542 * thus, P1 initializes buddy with B available. to prevent this
3543 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3546 ext4_lock_group(sb
, grp
);
3547 list_del(&pa
->pa_group_list
);
3548 ext4_unlock_group(sb
, grp
);
3550 spin_lock(pa
->pa_obj_lock
);
3551 list_del_rcu(&pa
->pa_inode_list
);
3552 spin_unlock(pa
->pa_obj_lock
);
3554 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3558 * creates new preallocated space for given inode
3560 static noinline_for_stack
int
3561 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3563 struct super_block
*sb
= ac
->ac_sb
;
3564 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3565 struct ext4_prealloc_space
*pa
;
3566 struct ext4_group_info
*grp
;
3567 struct ext4_inode_info
*ei
;
3569 /* preallocate only when found space is larger then requested */
3570 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3571 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3572 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3574 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3578 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3584 /* we can't allocate as much as normalizer wants.
3585 * so, found space must get proper lstart
3586 * to cover original request */
3587 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3588 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3590 /* we're limited by original request in that
3591 * logical block must be covered any way
3592 * winl is window we can move our chunk within */
3593 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3595 /* also, we should cover whole original request */
3596 wins
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
);
3598 /* the smallest one defines real window */
3599 win
= min(winl
, wins
);
3601 offs
= ac
->ac_o_ex
.fe_logical
%
3602 EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
3603 if (offs
&& offs
< win
)
3606 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
-
3607 EXT4_NUM_B2C(sbi
, win
);
3608 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3609 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3612 /* preallocation can change ac_b_ex, thus we store actually
3613 * allocated blocks for history */
3614 ac
->ac_f_ex
= ac
->ac_b_ex
;
3616 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3617 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3618 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3619 pa
->pa_free
= pa
->pa_len
;
3620 atomic_set(&pa
->pa_count
, 1);
3621 spin_lock_init(&pa
->pa_lock
);
3622 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3623 INIT_LIST_HEAD(&pa
->pa_group_list
);
3625 pa
->pa_type
= MB_INODE_PA
;
3627 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3628 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3629 trace_ext4_mb_new_inode_pa(ac
, pa
);
3631 ext4_mb_use_inode_pa(ac
, pa
);
3632 atomic_add(pa
->pa_free
, &sbi
->s_mb_preallocated
);
3634 ei
= EXT4_I(ac
->ac_inode
);
3635 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3637 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3638 pa
->pa_inode
= ac
->ac_inode
;
3640 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3641 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3642 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3644 spin_lock(pa
->pa_obj_lock
);
3645 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3646 spin_unlock(pa
->pa_obj_lock
);
3652 * creates new preallocated space for locality group inodes belongs to
3654 static noinline_for_stack
int
3655 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3657 struct super_block
*sb
= ac
->ac_sb
;
3658 struct ext4_locality_group
*lg
;
3659 struct ext4_prealloc_space
*pa
;
3660 struct ext4_group_info
*grp
;
3662 /* preallocate only when found space is larger then requested */
3663 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3664 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3665 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3667 BUG_ON(ext4_pspace_cachep
== NULL
);
3668 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3672 /* preallocation can change ac_b_ex, thus we store actually
3673 * allocated blocks for history */
3674 ac
->ac_f_ex
= ac
->ac_b_ex
;
3676 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3677 pa
->pa_lstart
= pa
->pa_pstart
;
3678 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3679 pa
->pa_free
= pa
->pa_len
;
3680 atomic_set(&pa
->pa_count
, 1);
3681 spin_lock_init(&pa
->pa_lock
);
3682 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3683 INIT_LIST_HEAD(&pa
->pa_group_list
);
3685 pa
->pa_type
= MB_GROUP_PA
;
3687 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3688 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3689 trace_ext4_mb_new_group_pa(ac
, pa
);
3691 ext4_mb_use_group_pa(ac
, pa
);
3692 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3694 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3698 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3699 pa
->pa_inode
= NULL
;
3701 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3702 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3703 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3706 * We will later add the new pa to the right bucket
3707 * after updating the pa_free in ext4_mb_release_context
3712 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3716 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3717 err
= ext4_mb_new_group_pa(ac
);
3719 err
= ext4_mb_new_inode_pa(ac
);
3724 * finds all unused blocks in on-disk bitmap, frees them in
3725 * in-core bitmap and buddy.
3726 * @pa must be unlinked from inode and group lists, so that
3727 * nobody else can find/use it.
3728 * the caller MUST hold group/inode locks.
3729 * TODO: optimize the case when there are no in-core structures yet
3731 static noinline_for_stack
int
3732 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3733 struct ext4_prealloc_space
*pa
)
3735 struct super_block
*sb
= e4b
->bd_sb
;
3736 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3741 unsigned long long grp_blk_start
;
3745 BUG_ON(pa
->pa_deleted
== 0);
3746 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3747 grp_blk_start
= pa
->pa_pstart
- EXT4_C2B(sbi
, bit
);
3748 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3749 end
= bit
+ pa
->pa_len
;
3752 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3755 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3756 mb_debug(1, " free preallocated %u/%u in group %u\n",
3757 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3758 (unsigned) next
- bit
, (unsigned) group
);
3761 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3762 trace_ext4_mb_release_inode_pa(pa
, (grp_blk_start
+
3763 EXT4_C2B(sbi
, bit
)),
3765 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3768 if (free
!= pa
->pa_free
) {
3769 ext4_msg(e4b
->bd_sb
, KERN_CRIT
,
3770 "pa %p: logic %lu, phys. %lu, len %lu",
3771 pa
, (unsigned long) pa
->pa_lstart
,
3772 (unsigned long) pa
->pa_pstart
,
3773 (unsigned long) pa
->pa_len
);
3774 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3777 * pa is already deleted so we use the value obtained
3778 * from the bitmap and continue.
3781 atomic_add(free
, &sbi
->s_mb_discarded
);
3786 static noinline_for_stack
int
3787 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3788 struct ext4_prealloc_space
*pa
)
3790 struct super_block
*sb
= e4b
->bd_sb
;
3794 trace_ext4_mb_release_group_pa(sb
, pa
);
3795 BUG_ON(pa
->pa_deleted
== 0);
3796 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3797 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3798 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3799 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3800 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3806 * releases all preallocations in given group
3808 * first, we need to decide discard policy:
3809 * - when do we discard
3811 * - how many do we discard
3812 * 1) how many requested
3814 static noinline_for_stack
int
3815 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3816 ext4_group_t group
, int needed
)
3818 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3819 struct buffer_head
*bitmap_bh
= NULL
;
3820 struct ext4_prealloc_space
*pa
, *tmp
;
3821 struct list_head list
;
3822 struct ext4_buddy e4b
;
3827 mb_debug(1, "discard preallocation for group %u\n", group
);
3829 if (list_empty(&grp
->bb_prealloc_list
))
3832 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3833 if (bitmap_bh
== NULL
) {
3834 ext4_error(sb
, "Error reading block bitmap for %u", group
);
3838 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3840 ext4_error(sb
, "Error loading buddy information for %u", group
);
3846 needed
= EXT4_CLUSTERS_PER_GROUP(sb
) + 1;
3848 INIT_LIST_HEAD(&list
);
3850 ext4_lock_group(sb
, group
);
3851 list_for_each_entry_safe(pa
, tmp
,
3852 &grp
->bb_prealloc_list
, pa_group_list
) {
3853 spin_lock(&pa
->pa_lock
);
3854 if (atomic_read(&pa
->pa_count
)) {
3855 spin_unlock(&pa
->pa_lock
);
3859 if (pa
->pa_deleted
) {
3860 spin_unlock(&pa
->pa_lock
);
3864 /* seems this one can be freed ... */
3867 /* we can trust pa_free ... */
3868 free
+= pa
->pa_free
;
3870 spin_unlock(&pa
->pa_lock
);
3872 list_del(&pa
->pa_group_list
);
3873 list_add(&pa
->u
.pa_tmp_list
, &list
);
3876 /* if we still need more blocks and some PAs were used, try again */
3877 if (free
< needed
&& busy
) {
3879 ext4_unlock_group(sb
, group
);
3884 /* found anything to free? */
3885 if (list_empty(&list
)) {
3890 /* now free all selected PAs */
3891 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3893 /* remove from object (inode or locality group) */
3894 spin_lock(pa
->pa_obj_lock
);
3895 list_del_rcu(&pa
->pa_inode_list
);
3896 spin_unlock(pa
->pa_obj_lock
);
3898 if (pa
->pa_type
== MB_GROUP_PA
)
3899 ext4_mb_release_group_pa(&e4b
, pa
);
3901 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3903 list_del(&pa
->u
.pa_tmp_list
);
3904 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3908 ext4_unlock_group(sb
, group
);
3909 ext4_mb_unload_buddy(&e4b
);
3915 * releases all non-used preallocated blocks for given inode
3917 * It's important to discard preallocations under i_data_sem
3918 * We don't want another block to be served from the prealloc
3919 * space when we are discarding the inode prealloc space.
3921 * FIXME!! Make sure it is valid at all the call sites
3923 void ext4_discard_preallocations(struct inode
*inode
)
3925 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3926 struct super_block
*sb
= inode
->i_sb
;
3927 struct buffer_head
*bitmap_bh
= NULL
;
3928 struct ext4_prealloc_space
*pa
, *tmp
;
3929 ext4_group_t group
= 0;
3930 struct list_head list
;
3931 struct ext4_buddy e4b
;
3934 if (!S_ISREG(inode
->i_mode
)) {
3935 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3939 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
3940 trace_ext4_discard_preallocations(inode
);
3942 INIT_LIST_HEAD(&list
);
3945 /* first, collect all pa's in the inode */
3946 spin_lock(&ei
->i_prealloc_lock
);
3947 while (!list_empty(&ei
->i_prealloc_list
)) {
3948 pa
= list_entry(ei
->i_prealloc_list
.next
,
3949 struct ext4_prealloc_space
, pa_inode_list
);
3950 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3951 spin_lock(&pa
->pa_lock
);
3952 if (atomic_read(&pa
->pa_count
)) {
3953 /* this shouldn't happen often - nobody should
3954 * use preallocation while we're discarding it */
3955 spin_unlock(&pa
->pa_lock
);
3956 spin_unlock(&ei
->i_prealloc_lock
);
3957 ext4_msg(sb
, KERN_ERR
,
3958 "uh-oh! used pa while discarding");
3960 schedule_timeout_uninterruptible(HZ
);
3964 if (pa
->pa_deleted
== 0) {
3966 spin_unlock(&pa
->pa_lock
);
3967 list_del_rcu(&pa
->pa_inode_list
);
3968 list_add(&pa
->u
.pa_tmp_list
, &list
);
3972 /* someone is deleting pa right now */
3973 spin_unlock(&pa
->pa_lock
);
3974 spin_unlock(&ei
->i_prealloc_lock
);
3976 /* we have to wait here because pa_deleted
3977 * doesn't mean pa is already unlinked from
3978 * the list. as we might be called from
3979 * ->clear_inode() the inode will get freed
3980 * and concurrent thread which is unlinking
3981 * pa from inode's list may access already
3982 * freed memory, bad-bad-bad */
3984 /* XXX: if this happens too often, we can
3985 * add a flag to force wait only in case
3986 * of ->clear_inode(), but not in case of
3987 * regular truncate */
3988 schedule_timeout_uninterruptible(HZ
);
3991 spin_unlock(&ei
->i_prealloc_lock
);
3993 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3994 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
3995 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
3997 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3999 ext4_error(sb
, "Error loading buddy information for %u",
4004 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
4005 if (bitmap_bh
== NULL
) {
4006 ext4_error(sb
, "Error reading block bitmap for %u",
4008 ext4_mb_unload_buddy(&e4b
);
4012 ext4_lock_group(sb
, group
);
4013 list_del(&pa
->pa_group_list
);
4014 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
4015 ext4_unlock_group(sb
, group
);
4017 ext4_mb_unload_buddy(&e4b
);
4020 list_del(&pa
->u
.pa_tmp_list
);
4021 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4025 #ifdef CONFIG_EXT4_DEBUG
4026 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4028 struct super_block
*sb
= ac
->ac_sb
;
4029 ext4_group_t ngroups
, i
;
4031 if (!ext4_mballoc_debug
||
4032 (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
))
4035 ext4_msg(ac
->ac_sb
, KERN_ERR
, "Can't allocate:"
4036 " Allocation context details:");
4037 ext4_msg(ac
->ac_sb
, KERN_ERR
, "status %d flags %d",
4038 ac
->ac_status
, ac
->ac_flags
);
4039 ext4_msg(ac
->ac_sb
, KERN_ERR
, "orig %lu/%lu/%lu@%lu, "
4040 "goal %lu/%lu/%lu@%lu, "
4041 "best %lu/%lu/%lu@%lu cr %d",
4042 (unsigned long)ac
->ac_o_ex
.fe_group
,
4043 (unsigned long)ac
->ac_o_ex
.fe_start
,
4044 (unsigned long)ac
->ac_o_ex
.fe_len
,
4045 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4046 (unsigned long)ac
->ac_g_ex
.fe_group
,
4047 (unsigned long)ac
->ac_g_ex
.fe_start
,
4048 (unsigned long)ac
->ac_g_ex
.fe_len
,
4049 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4050 (unsigned long)ac
->ac_b_ex
.fe_group
,
4051 (unsigned long)ac
->ac_b_ex
.fe_start
,
4052 (unsigned long)ac
->ac_b_ex
.fe_len
,
4053 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4054 (int)ac
->ac_criteria
);
4055 ext4_msg(ac
->ac_sb
, KERN_ERR
, "%d found", ac
->ac_found
);
4056 ext4_msg(ac
->ac_sb
, KERN_ERR
, "groups: ");
4057 ngroups
= ext4_get_groups_count(sb
);
4058 for (i
= 0; i
< ngroups
; i
++) {
4059 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4060 struct ext4_prealloc_space
*pa
;
4061 ext4_grpblk_t start
;
4062 struct list_head
*cur
;
4063 ext4_lock_group(sb
, i
);
4064 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4065 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4067 spin_lock(&pa
->pa_lock
);
4068 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4070 spin_unlock(&pa
->pa_lock
);
4071 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
4074 ext4_unlock_group(sb
, i
);
4076 if (grp
->bb_free
== 0)
4078 printk(KERN_ERR
"%u: %d/%d \n",
4079 i
, grp
->bb_free
, grp
->bb_fragments
);
4081 printk(KERN_ERR
"\n");
4084 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4091 * We use locality group preallocation for small size file. The size of the
4092 * file is determined by the current size or the resulting size after
4093 * allocation which ever is larger
4095 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4097 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4099 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4100 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4103 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4106 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4109 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
4110 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
4113 if ((size
== isize
) &&
4114 !ext4_fs_is_busy(sbi
) &&
4115 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
4116 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
4120 if (sbi
->s_mb_group_prealloc
<= 0) {
4121 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4125 /* don't use group allocation for large files */
4126 size
= max(size
, isize
);
4127 if (size
> sbi
->s_mb_stream_request
) {
4128 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4132 BUG_ON(ac
->ac_lg
!= NULL
);
4134 * locality group prealloc space are per cpu. The reason for having
4135 * per cpu locality group is to reduce the contention between block
4136 * request from multiple CPUs.
4138 ac
->ac_lg
= raw_cpu_ptr(sbi
->s_locality_groups
);
4140 /* we're going to use group allocation */
4141 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4143 /* serialize all allocations in the group */
4144 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4147 static noinline_for_stack
int
4148 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4149 struct ext4_allocation_request
*ar
)
4151 struct super_block
*sb
= ar
->inode
->i_sb
;
4152 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4153 struct ext4_super_block
*es
= sbi
->s_es
;
4157 ext4_grpblk_t block
;
4159 /* we can't allocate > group size */
4162 /* just a dirty hack to filter too big requests */
4163 if (len
>= EXT4_CLUSTERS_PER_GROUP(sb
))
4164 len
= EXT4_CLUSTERS_PER_GROUP(sb
);
4166 /* start searching from the goal */
4168 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4169 goal
>= ext4_blocks_count(es
))
4170 goal
= le32_to_cpu(es
->s_first_data_block
);
4171 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4173 /* set up allocation goals */
4174 ac
->ac_b_ex
.fe_logical
= EXT4_LBLK_CMASK(sbi
, ar
->logical
);
4175 ac
->ac_status
= AC_STATUS_CONTINUE
;
4177 ac
->ac_inode
= ar
->inode
;
4178 ac
->ac_o_ex
.fe_logical
= ac
->ac_b_ex
.fe_logical
;
4179 ac
->ac_o_ex
.fe_group
= group
;
4180 ac
->ac_o_ex
.fe_start
= block
;
4181 ac
->ac_o_ex
.fe_len
= len
;
4182 ac
->ac_g_ex
= ac
->ac_o_ex
;
4183 ac
->ac_flags
= ar
->flags
;
4185 /* we have to define context: we'll we work with a file or
4186 * locality group. this is a policy, actually */
4187 ext4_mb_group_or_file(ac
);
4189 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4190 "left: %u/%u, right %u/%u to %swritable\n",
4191 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4192 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4193 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4194 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4195 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4200 static noinline_for_stack
void
4201 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4202 struct ext4_locality_group
*lg
,
4203 int order
, int total_entries
)
4205 ext4_group_t group
= 0;
4206 struct ext4_buddy e4b
;
4207 struct list_head discard_list
;
4208 struct ext4_prealloc_space
*pa
, *tmp
;
4210 mb_debug(1, "discard locality group preallocation\n");
4212 INIT_LIST_HEAD(&discard_list
);
4214 spin_lock(&lg
->lg_prealloc_lock
);
4215 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4217 spin_lock(&pa
->pa_lock
);
4218 if (atomic_read(&pa
->pa_count
)) {
4220 * This is the pa that we just used
4221 * for block allocation. So don't
4224 spin_unlock(&pa
->pa_lock
);
4227 if (pa
->pa_deleted
) {
4228 spin_unlock(&pa
->pa_lock
);
4231 /* only lg prealloc space */
4232 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4234 /* seems this one can be freed ... */
4236 spin_unlock(&pa
->pa_lock
);
4238 list_del_rcu(&pa
->pa_inode_list
);
4239 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4242 if (total_entries
<= 5) {
4244 * we want to keep only 5 entries
4245 * allowing it to grow to 8. This
4246 * mak sure we don't call discard
4247 * soon for this list.
4252 spin_unlock(&lg
->lg_prealloc_lock
);
4254 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4256 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4257 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4258 ext4_error(sb
, "Error loading buddy information for %u",
4262 ext4_lock_group(sb
, group
);
4263 list_del(&pa
->pa_group_list
);
4264 ext4_mb_release_group_pa(&e4b
, pa
);
4265 ext4_unlock_group(sb
, group
);
4267 ext4_mb_unload_buddy(&e4b
);
4268 list_del(&pa
->u
.pa_tmp_list
);
4269 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4274 * We have incremented pa_count. So it cannot be freed at this
4275 * point. Also we hold lg_mutex. So no parallel allocation is
4276 * possible from this lg. That means pa_free cannot be updated.
4278 * A parallel ext4_mb_discard_group_preallocations is possible.
4279 * which can cause the lg_prealloc_list to be updated.
4282 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4284 int order
, added
= 0, lg_prealloc_count
= 1;
4285 struct super_block
*sb
= ac
->ac_sb
;
4286 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4287 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4289 order
= fls(pa
->pa_free
) - 1;
4290 if (order
> PREALLOC_TB_SIZE
- 1)
4291 /* The max size of hash table is PREALLOC_TB_SIZE */
4292 order
= PREALLOC_TB_SIZE
- 1;
4293 /* Add the prealloc space to lg */
4294 spin_lock(&lg
->lg_prealloc_lock
);
4295 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4297 spin_lock(&tmp_pa
->pa_lock
);
4298 if (tmp_pa
->pa_deleted
) {
4299 spin_unlock(&tmp_pa
->pa_lock
);
4302 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4303 /* Add to the tail of the previous entry */
4304 list_add_tail_rcu(&pa
->pa_inode_list
,
4305 &tmp_pa
->pa_inode_list
);
4308 * we want to count the total
4309 * number of entries in the list
4312 spin_unlock(&tmp_pa
->pa_lock
);
4313 lg_prealloc_count
++;
4316 list_add_tail_rcu(&pa
->pa_inode_list
,
4317 &lg
->lg_prealloc_list
[order
]);
4318 spin_unlock(&lg
->lg_prealloc_lock
);
4320 /* Now trim the list to be not more than 8 elements */
4321 if (lg_prealloc_count
> 8) {
4322 ext4_mb_discard_lg_preallocations(sb
, lg
,
4323 order
, lg_prealloc_count
);
4330 * release all resource we used in allocation
4332 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4334 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4335 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4337 if (pa
->pa_type
== MB_GROUP_PA
) {
4338 /* see comment in ext4_mb_use_group_pa() */
4339 spin_lock(&pa
->pa_lock
);
4340 pa
->pa_pstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4341 pa
->pa_lstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4342 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4343 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4344 spin_unlock(&pa
->pa_lock
);
4349 * We want to add the pa to the right bucket.
4350 * Remove it from the list and while adding
4351 * make sure the list to which we are adding
4354 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4355 spin_lock(pa
->pa_obj_lock
);
4356 list_del_rcu(&pa
->pa_inode_list
);
4357 spin_unlock(pa
->pa_obj_lock
);
4358 ext4_mb_add_n_trim(ac
);
4360 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4362 if (ac
->ac_bitmap_page
)
4363 page_cache_release(ac
->ac_bitmap_page
);
4364 if (ac
->ac_buddy_page
)
4365 page_cache_release(ac
->ac_buddy_page
);
4366 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4367 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4368 ext4_mb_collect_stats(ac
);
4372 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4374 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4378 trace_ext4_mb_discard_preallocations(sb
, needed
);
4379 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4380 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4389 * Main entry point into mballoc to allocate blocks
4390 * it tries to use preallocation first, then falls back
4391 * to usual allocation
4393 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4394 struct ext4_allocation_request
*ar
, int *errp
)
4397 struct ext4_allocation_context
*ac
= NULL
;
4398 struct ext4_sb_info
*sbi
;
4399 struct super_block
*sb
;
4400 ext4_fsblk_t block
= 0;
4401 unsigned int inquota
= 0;
4402 unsigned int reserv_clstrs
= 0;
4405 sb
= ar
->inode
->i_sb
;
4408 trace_ext4_request_blocks(ar
);
4410 /* Allow to use superuser reservation for quota file */
4411 if (IS_NOQUOTA(ar
->inode
))
4412 ar
->flags
|= EXT4_MB_USE_ROOT_BLOCKS
;
4414 if ((ar
->flags
& EXT4_MB_DELALLOC_RESERVED
) == 0) {
4415 /* Without delayed allocation we need to verify
4416 * there is enough free blocks to do block allocation
4417 * and verify allocation doesn't exceed the quota limits.
4420 ext4_claim_free_clusters(sbi
, ar
->len
, ar
->flags
)) {
4422 /* let others to free the space */
4424 ar
->len
= ar
->len
>> 1;
4430 reserv_clstrs
= ar
->len
;
4431 if (ar
->flags
& EXT4_MB_USE_ROOT_BLOCKS
) {
4432 dquot_alloc_block_nofail(ar
->inode
,
4433 EXT4_C2B(sbi
, ar
->len
));
4436 dquot_alloc_block(ar
->inode
,
4437 EXT4_C2B(sbi
, ar
->len
))) {
4439 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4450 ac
= kmem_cache_zalloc(ext4_ac_cachep
, GFP_NOFS
);
4457 *errp
= ext4_mb_initialize_context(ac
, ar
);
4463 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4464 if (!ext4_mb_use_preallocated(ac
)) {
4465 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4466 ext4_mb_normalize_request(ac
, ar
);
4468 /* allocate space in core */
4469 *errp
= ext4_mb_regular_allocator(ac
);
4471 goto discard_and_exit
;
4473 /* as we've just preallocated more space than
4474 * user requested originally, we store allocated
4475 * space in a special descriptor */
4476 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4477 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4478 *errp
= ext4_mb_new_preallocation(ac
);
4481 ext4_discard_allocated_blocks(ac
);
4485 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4486 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_clstrs
);
4487 if (*errp
== -EAGAIN
) {
4489 * drop the reference that we took
4490 * in ext4_mb_use_best_found
4492 ext4_mb_release_context(ac
);
4493 ac
->ac_b_ex
.fe_group
= 0;
4494 ac
->ac_b_ex
.fe_start
= 0;
4495 ac
->ac_b_ex
.fe_len
= 0;
4496 ac
->ac_status
= AC_STATUS_CONTINUE
;
4499 ext4_discard_allocated_blocks(ac
);
4502 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4503 ar
->len
= ac
->ac_b_ex
.fe_len
;
4506 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4514 ac
->ac_b_ex
.fe_len
= 0;
4516 ext4_mb_show_ac(ac
);
4518 ext4_mb_release_context(ac
);
4521 kmem_cache_free(ext4_ac_cachep
, ac
);
4522 if (inquota
&& ar
->len
< inquota
)
4523 dquot_free_block(ar
->inode
, EXT4_C2B(sbi
, inquota
- ar
->len
));
4525 if ((ar
->flags
& EXT4_MB_DELALLOC_RESERVED
) == 0)
4526 /* release all the reserved blocks if non delalloc */
4527 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
4531 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4537 * We can merge two free data extents only if the physical blocks
4538 * are contiguous, AND the extents were freed by the same transaction,
4539 * AND the blocks are associated with the same group.
4541 static int can_merge(struct ext4_free_data
*entry1
,
4542 struct ext4_free_data
*entry2
)
4544 if ((entry1
->efd_tid
== entry2
->efd_tid
) &&
4545 (entry1
->efd_group
== entry2
->efd_group
) &&
4546 ((entry1
->efd_start_cluster
+ entry1
->efd_count
) == entry2
->efd_start_cluster
))
4551 static noinline_for_stack
int
4552 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4553 struct ext4_free_data
*new_entry
)
4555 ext4_group_t group
= e4b
->bd_group
;
4556 ext4_grpblk_t cluster
;
4557 struct ext4_free_data
*entry
;
4558 struct ext4_group_info
*db
= e4b
->bd_info
;
4559 struct super_block
*sb
= e4b
->bd_sb
;
4560 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4561 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4562 struct rb_node
*parent
= NULL
, *new_node
;
4564 BUG_ON(!ext4_handle_valid(handle
));
4565 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4566 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4568 new_node
= &new_entry
->efd_node
;
4569 cluster
= new_entry
->efd_start_cluster
;
4572 /* first free block exent. We need to
4573 protect buddy cache from being freed,
4574 * otherwise we'll refresh it from
4575 * on-disk bitmap and lose not-yet-available
4577 page_cache_get(e4b
->bd_buddy_page
);
4578 page_cache_get(e4b
->bd_bitmap_page
);
4582 entry
= rb_entry(parent
, struct ext4_free_data
, efd_node
);
4583 if (cluster
< entry
->efd_start_cluster
)
4585 else if (cluster
>= (entry
->efd_start_cluster
+ entry
->efd_count
))
4586 n
= &(*n
)->rb_right
;
4588 ext4_grp_locked_error(sb
, group
, 0,
4589 ext4_group_first_block_no(sb
, group
) +
4590 EXT4_C2B(sbi
, cluster
),
4591 "Block already on to-be-freed list");
4596 rb_link_node(new_node
, parent
, n
);
4597 rb_insert_color(new_node
, &db
->bb_free_root
);
4599 /* Now try to see the extent can be merged to left and right */
4600 node
= rb_prev(new_node
);
4602 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4603 if (can_merge(entry
, new_entry
) &&
4604 ext4_journal_callback_try_del(handle
, &entry
->efd_jce
)) {
4605 new_entry
->efd_start_cluster
= entry
->efd_start_cluster
;
4606 new_entry
->efd_count
+= entry
->efd_count
;
4607 rb_erase(node
, &(db
->bb_free_root
));
4608 kmem_cache_free(ext4_free_data_cachep
, entry
);
4612 node
= rb_next(new_node
);
4614 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4615 if (can_merge(new_entry
, entry
) &&
4616 ext4_journal_callback_try_del(handle
, &entry
->efd_jce
)) {
4617 new_entry
->efd_count
+= entry
->efd_count
;
4618 rb_erase(node
, &(db
->bb_free_root
));
4619 kmem_cache_free(ext4_free_data_cachep
, entry
);
4622 /* Add the extent to transaction's private list */
4623 ext4_journal_callback_add(handle
, ext4_free_data_callback
,
4624 &new_entry
->efd_jce
);
4629 * ext4_free_blocks() -- Free given blocks and update quota
4630 * @handle: handle for this transaction
4632 * @block: start physical block to free
4633 * @count: number of blocks to count
4634 * @flags: flags used by ext4_free_blocks
4636 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4637 struct buffer_head
*bh
, ext4_fsblk_t block
,
4638 unsigned long count
, int flags
)
4640 struct buffer_head
*bitmap_bh
= NULL
;
4641 struct super_block
*sb
= inode
->i_sb
;
4642 struct ext4_group_desc
*gdp
;
4643 unsigned int overflow
;
4645 struct buffer_head
*gd_bh
;
4646 ext4_group_t block_group
;
4647 struct ext4_sb_info
*sbi
;
4648 struct ext4_buddy e4b
;
4649 unsigned int count_clusters
;
4656 BUG_ON(block
!= bh
->b_blocknr
);
4658 block
= bh
->b_blocknr
;
4662 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4663 !ext4_data_block_valid(sbi
, block
, count
)) {
4664 ext4_error(sb
, "Freeing blocks not in datazone - "
4665 "block = %llu, count = %lu", block
, count
);
4669 ext4_debug("freeing block %llu\n", block
);
4670 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4672 if (flags
& EXT4_FREE_BLOCKS_FORGET
) {
4673 struct buffer_head
*tbh
= bh
;
4676 BUG_ON(bh
&& (count
> 1));
4678 for (i
= 0; i
< count
; i
++) {
4681 tbh
= sb_find_get_block(inode
->i_sb
,
4685 ext4_forget(handle
, flags
& EXT4_FREE_BLOCKS_METADATA
,
4686 inode
, tbh
, block
+ i
);
4691 * We need to make sure we don't reuse the freed block until
4692 * after the transaction is committed, which we can do by
4693 * treating the block as metadata, below. We make an
4694 * exception if the inode is to be written in writeback mode
4695 * since writeback mode has weak data consistency guarantees.
4697 if (!ext4_should_writeback_data(inode
))
4698 flags
|= EXT4_FREE_BLOCKS_METADATA
;
4701 * If the extent to be freed does not begin on a cluster
4702 * boundary, we need to deal with partial clusters at the
4703 * beginning and end of the extent. Normally we will free
4704 * blocks at the beginning or the end unless we are explicitly
4705 * requested to avoid doing so.
4707 overflow
= EXT4_PBLK_COFF(sbi
, block
);
4709 if (flags
& EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER
) {
4710 overflow
= sbi
->s_cluster_ratio
- overflow
;
4712 if (count
> overflow
)
4721 overflow
= EXT4_LBLK_COFF(sbi
, count
);
4723 if (flags
& EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER
) {
4724 if (count
> overflow
)
4729 count
+= sbi
->s_cluster_ratio
- overflow
;
4734 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4736 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4737 ext4_get_group_info(sb
, block_group
))))
4741 * Check to see if we are freeing blocks across a group
4744 if (EXT4_C2B(sbi
, bit
) + count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4745 overflow
= EXT4_C2B(sbi
, bit
) + count
-
4746 EXT4_BLOCKS_PER_GROUP(sb
);
4749 count_clusters
= EXT4_NUM_B2C(sbi
, count
);
4750 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4755 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4761 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4762 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4763 in_range(block
, ext4_inode_table(sb
, gdp
),
4764 EXT4_SB(sb
)->s_itb_per_group
) ||
4765 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4766 EXT4_SB(sb
)->s_itb_per_group
)) {
4768 ext4_error(sb
, "Freeing blocks in system zone - "
4769 "Block = %llu, count = %lu", block
, count
);
4770 /* err = 0. ext4_std_error should be a no op */
4774 BUFFER_TRACE(bitmap_bh
, "getting write access");
4775 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4780 * We are about to modify some metadata. Call the journal APIs
4781 * to unshare ->b_data if a currently-committing transaction is
4784 BUFFER_TRACE(gd_bh
, "get_write_access");
4785 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4788 #ifdef AGGRESSIVE_CHECK
4791 for (i
= 0; i
< count_clusters
; i
++)
4792 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4795 trace_ext4_mballoc_free(sb
, inode
, block_group
, bit
, count_clusters
);
4797 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4801 if ((flags
& EXT4_FREE_BLOCKS_METADATA
) && ext4_handle_valid(handle
)) {
4802 struct ext4_free_data
*new_entry
;
4804 * blocks being freed are metadata. these blocks shouldn't
4805 * be used until this transaction is committed
4808 new_entry
= kmem_cache_alloc(ext4_free_data_cachep
, GFP_NOFS
);
4811 * We use a retry loop because
4812 * ext4_free_blocks() is not allowed to fail.
4815 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
4818 new_entry
->efd_start_cluster
= bit
;
4819 new_entry
->efd_group
= block_group
;
4820 new_entry
->efd_count
= count_clusters
;
4821 new_entry
->efd_tid
= handle
->h_transaction
->t_tid
;
4823 ext4_lock_group(sb
, block_group
);
4824 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4825 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4827 /* need to update group_info->bb_free and bitmap
4828 * with group lock held. generate_buddy look at
4829 * them with group lock_held
4831 if (test_opt(sb
, DISCARD
)) {
4832 err
= ext4_issue_discard(sb
, block_group
, bit
, count
);
4833 if (err
&& err
!= -EOPNOTSUPP
)
4834 ext4_msg(sb
, KERN_WARNING
, "discard request in"
4835 " group:%d block:%d count:%lu failed"
4836 " with %d", block_group
, bit
, count
,
4839 EXT4_MB_GRP_CLEAR_TRIMMED(e4b
.bd_info
);
4841 ext4_lock_group(sb
, block_group
);
4842 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4843 mb_free_blocks(inode
, &e4b
, bit
, count_clusters
);
4846 ret
= ext4_free_group_clusters(sb
, gdp
) + count_clusters
;
4847 ext4_free_group_clusters_set(sb
, gdp
, ret
);
4848 ext4_block_bitmap_csum_set(sb
, block_group
, gdp
, bitmap_bh
);
4849 ext4_group_desc_csum_set(sb
, block_group
, gdp
);
4850 ext4_unlock_group(sb
, block_group
);
4852 if (sbi
->s_log_groups_per_flex
) {
4853 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4854 atomic64_add(count_clusters
,
4855 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
4858 if (!(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4859 dquot_free_block(inode
, EXT4_C2B(sbi
, count_clusters
));
4860 percpu_counter_add(&sbi
->s_freeclusters_counter
, count_clusters
);
4862 ext4_mb_unload_buddy(&e4b
);
4864 /* We dirtied the bitmap block */
4865 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4866 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4868 /* And the group descriptor block */
4869 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4870 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4874 if (overflow
&& !err
) {
4882 ext4_std_error(sb
, err
);
4887 * ext4_group_add_blocks() -- Add given blocks to an existing group
4888 * @handle: handle to this transaction
4890 * @block: start physical block to add to the block group
4891 * @count: number of blocks to free
4893 * This marks the blocks as free in the bitmap and buddy.
4895 int ext4_group_add_blocks(handle_t
*handle
, struct super_block
*sb
,
4896 ext4_fsblk_t block
, unsigned long count
)
4898 struct buffer_head
*bitmap_bh
= NULL
;
4899 struct buffer_head
*gd_bh
;
4900 ext4_group_t block_group
;
4903 struct ext4_group_desc
*desc
;
4904 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4905 struct ext4_buddy e4b
;
4906 int err
= 0, ret
, blk_free_count
;
4907 ext4_grpblk_t blocks_freed
;
4909 ext4_debug("Adding block(s) %llu-%llu\n", block
, block
+ count
- 1);
4914 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4916 * Check to see if we are freeing blocks across a group
4919 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4920 ext4_warning(sb
, "too much blocks added to group %u\n",
4926 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4932 desc
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4938 if (in_range(ext4_block_bitmap(sb
, desc
), block
, count
) ||
4939 in_range(ext4_inode_bitmap(sb
, desc
), block
, count
) ||
4940 in_range(block
, ext4_inode_table(sb
, desc
), sbi
->s_itb_per_group
) ||
4941 in_range(block
+ count
- 1, ext4_inode_table(sb
, desc
),
4942 sbi
->s_itb_per_group
)) {
4943 ext4_error(sb
, "Adding blocks in system zones - "
4944 "Block = %llu, count = %lu",
4950 BUFFER_TRACE(bitmap_bh
, "getting write access");
4951 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4956 * We are about to modify some metadata. Call the journal APIs
4957 * to unshare ->b_data if a currently-committing transaction is
4960 BUFFER_TRACE(gd_bh
, "get_write_access");
4961 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4965 for (i
= 0, blocks_freed
= 0; i
< count
; i
++) {
4966 BUFFER_TRACE(bitmap_bh
, "clear bit");
4967 if (!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
)) {
4968 ext4_error(sb
, "bit already cleared for block %llu",
4969 (ext4_fsblk_t
)(block
+ i
));
4970 BUFFER_TRACE(bitmap_bh
, "bit already cleared");
4976 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4981 * need to update group_info->bb_free and bitmap
4982 * with group lock held. generate_buddy look at
4983 * them with group lock_held
4985 ext4_lock_group(sb
, block_group
);
4986 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4987 mb_free_blocks(NULL
, &e4b
, bit
, count
);
4988 blk_free_count
= blocks_freed
+ ext4_free_group_clusters(sb
, desc
);
4989 ext4_free_group_clusters_set(sb
, desc
, blk_free_count
);
4990 ext4_block_bitmap_csum_set(sb
, block_group
, desc
, bitmap_bh
);
4991 ext4_group_desc_csum_set(sb
, block_group
, desc
);
4992 ext4_unlock_group(sb
, block_group
);
4993 percpu_counter_add(&sbi
->s_freeclusters_counter
,
4994 EXT4_NUM_B2C(sbi
, blocks_freed
));
4996 if (sbi
->s_log_groups_per_flex
) {
4997 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4998 atomic64_add(EXT4_NUM_B2C(sbi
, blocks_freed
),
4999 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
5002 ext4_mb_unload_buddy(&e4b
);
5004 /* We dirtied the bitmap block */
5005 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
5006 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
5008 /* And the group descriptor block */
5009 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
5010 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
5016 ext4_std_error(sb
, err
);
5021 * ext4_trim_extent -- function to TRIM one single free extent in the group
5022 * @sb: super block for the file system
5023 * @start: starting block of the free extent in the alloc. group
5024 * @count: number of blocks to TRIM
5025 * @group: alloc. group we are working with
5026 * @e4b: ext4 buddy for the group
5028 * Trim "count" blocks starting at "start" in the "group". To assure that no
5029 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5030 * be called with under the group lock.
5032 static int ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
5033 ext4_group_t group
, struct ext4_buddy
*e4b
)
5037 struct ext4_free_extent ex
;
5040 trace_ext4_trim_extent(sb
, group
, start
, count
);
5042 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
5044 ex
.fe_start
= start
;
5045 ex
.fe_group
= group
;
5049 * Mark blocks used, so no one can reuse them while
5052 mb_mark_used(e4b
, &ex
);
5053 ext4_unlock_group(sb
, group
);
5054 ret
= ext4_issue_discard(sb
, group
, start
, count
);
5055 ext4_lock_group(sb
, group
);
5056 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
5061 * ext4_trim_all_free -- function to trim all free space in alloc. group
5062 * @sb: super block for file system
5063 * @group: group to be trimmed
5064 * @start: first group block to examine
5065 * @max: last group block to examine
5066 * @minblocks: minimum extent block count
5068 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5069 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5073 * ext4_trim_all_free walks through group's block bitmap searching for free
5074 * extents. When the free extent is found, mark it as used in group buddy
5075 * bitmap. Then issue a TRIM command on this extent and free the extent in
5076 * the group buddy bitmap. This is done until whole group is scanned.
5078 static ext4_grpblk_t
5079 ext4_trim_all_free(struct super_block
*sb
, ext4_group_t group
,
5080 ext4_grpblk_t start
, ext4_grpblk_t max
,
5081 ext4_grpblk_t minblocks
)
5084 ext4_grpblk_t next
, count
= 0, free_count
= 0;
5085 struct ext4_buddy e4b
;
5088 trace_ext4_trim_all_free(sb
, group
, start
, max
);
5090 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
5092 ext4_error(sb
, "Error in loading buddy "
5093 "information for %u", group
);
5096 bitmap
= e4b
.bd_bitmap
;
5098 ext4_lock_group(sb
, group
);
5099 if (EXT4_MB_GRP_WAS_TRIMMED(e4b
.bd_info
) &&
5100 minblocks
>= atomic_read(&EXT4_SB(sb
)->s_last_trim_minblks
))
5103 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
5104 e4b
.bd_info
->bb_first_free
: start
;
5106 while (start
<= max
) {
5107 start
= mb_find_next_zero_bit(bitmap
, max
+ 1, start
);
5110 next
= mb_find_next_bit(bitmap
, max
+ 1, start
);
5112 if ((next
- start
) >= minblocks
) {
5113 ret
= ext4_trim_extent(sb
, start
,
5114 next
- start
, group
, &e4b
);
5115 if (ret
&& ret
!= -EOPNOTSUPP
)
5118 count
+= next
- start
;
5120 free_count
+= next
- start
;
5123 if (fatal_signal_pending(current
)) {
5124 count
= -ERESTARTSYS
;
5128 if (need_resched()) {
5129 ext4_unlock_group(sb
, group
);
5131 ext4_lock_group(sb
, group
);
5134 if ((e4b
.bd_info
->bb_free
- free_count
) < minblocks
)
5140 EXT4_MB_GRP_SET_TRIMMED(e4b
.bd_info
);
5143 ext4_unlock_group(sb
, group
);
5144 ext4_mb_unload_buddy(&e4b
);
5146 ext4_debug("trimmed %d blocks in the group %d\n",
5153 * ext4_trim_fs() -- trim ioctl handle function
5154 * @sb: superblock for filesystem
5155 * @range: fstrim_range structure
5157 * start: First Byte to trim
5158 * len: number of Bytes to trim from start
5159 * minlen: minimum extent length in Bytes
5160 * ext4_trim_fs goes through all allocation groups containing Bytes from
5161 * start to start+len. For each such a group ext4_trim_all_free function
5162 * is invoked to trim all free space.
5164 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
5166 struct ext4_group_info
*grp
;
5167 ext4_group_t group
, first_group
, last_group
;
5168 ext4_grpblk_t cnt
= 0, first_cluster
, last_cluster
;
5169 uint64_t start
, end
, minlen
, trimmed
= 0;
5170 ext4_fsblk_t first_data_blk
=
5171 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
5172 ext4_fsblk_t max_blks
= ext4_blocks_count(EXT4_SB(sb
)->s_es
);
5175 start
= range
->start
>> sb
->s_blocksize_bits
;
5176 end
= start
+ (range
->len
>> sb
->s_blocksize_bits
) - 1;
5177 minlen
= EXT4_NUM_B2C(EXT4_SB(sb
),
5178 range
->minlen
>> sb
->s_blocksize_bits
);
5180 if (minlen
> EXT4_CLUSTERS_PER_GROUP(sb
) ||
5181 start
>= max_blks
||
5182 range
->len
< sb
->s_blocksize
)
5184 if (end
>= max_blks
)
5186 if (end
<= first_data_blk
)
5188 if (start
< first_data_blk
)
5189 start
= first_data_blk
;
5191 /* Determine first and last group to examine based on start and end */
5192 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
5193 &first_group
, &first_cluster
);
5194 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) end
,
5195 &last_group
, &last_cluster
);
5197 /* end now represents the last cluster to discard in this group */
5198 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5200 for (group
= first_group
; group
<= last_group
; group
++) {
5201 grp
= ext4_get_group_info(sb
, group
);
5202 /* We only do this if the grp has never been initialized */
5203 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
5204 ret
= ext4_mb_init_group(sb
, group
);
5210 * For all the groups except the last one, last cluster will
5211 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5212 * change it for the last group, note that last_cluster is
5213 * already computed earlier by ext4_get_group_no_and_offset()
5215 if (group
== last_group
)
5218 if (grp
->bb_free
>= minlen
) {
5219 cnt
= ext4_trim_all_free(sb
, group
, first_cluster
,
5229 * For every group except the first one, we are sure
5230 * that the first cluster to discard will be cluster #0.
5236 atomic_set(&EXT4_SB(sb
)->s_last_trim_minblks
, minlen
);
5239 range
->len
= EXT4_C2B(EXT4_SB(sb
), trimmed
) << sb
->s_blocksize_bits
;