2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
24 #include "ext4_jbd2.h"
26 #include <linux/log2.h>
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <linux/nospec.h>
30 #include <linux/backing-dev.h>
31 #include <trace/events/ext4.h>
33 #ifdef CONFIG_EXT4_DEBUG
34 ushort ext4_mballoc_debug __read_mostly
;
36 module_param_named(mballoc_debug
, ext4_mballoc_debug
, ushort
, 0644);
37 MODULE_PARM_DESC(mballoc_debug
, "Debugging level for ext4's mballoc");
42 * - test ext4_ext_search_left() and ext4_ext_search_right()
43 * - search for metadata in few groups
46 * - normalization should take into account whether file is still open
47 * - discard preallocations if no free space left (policy?)
48 * - don't normalize tails
50 * - reservation for superuser
53 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
54 * - track min/max extents in each group for better group selection
55 * - mb_mark_used() may allocate chunk right after splitting buddy
56 * - tree of groups sorted by number of free blocks
61 * The allocation request involve request for multiple number of blocks
62 * near to the goal(block) value specified.
64 * During initialization phase of the allocator we decide to use the
65 * group preallocation or inode preallocation depending on the size of
66 * the file. The size of the file could be the resulting file size we
67 * would have after allocation, or the current file size, which ever
68 * is larger. If the size is less than sbi->s_mb_stream_request we
69 * select to use the group preallocation. The default value of
70 * s_mb_stream_request is 16 blocks. This can also be tuned via
71 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
72 * terms of number of blocks.
74 * The main motivation for having small file use group preallocation is to
75 * ensure that we have small files closer together on the disk.
77 * First stage the allocator looks at the inode prealloc list,
78 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
79 * spaces for this particular inode. The inode prealloc space is
82 * pa_lstart -> the logical start block for this prealloc space
83 * pa_pstart -> the physical start block for this prealloc space
84 * pa_len -> length for this prealloc space (in clusters)
85 * pa_free -> free space available in this prealloc space (in clusters)
87 * The inode preallocation space is used looking at the _logical_ start
88 * block. If only the logical file block falls within the range of prealloc
89 * space we will consume the particular prealloc space. This makes sure that
90 * we have contiguous physical blocks representing the file blocks
92 * The important thing to be noted in case of inode prealloc space is that
93 * we don't modify the values associated to inode prealloc space except
96 * If we are not able to find blocks in the inode prealloc space and if we
97 * have the group allocation flag set then we look at the locality group
98 * prealloc space. These are per CPU prealloc list represented as
100 * ext4_sb_info.s_locality_groups[smp_processor_id()]
102 * The reason for having a per cpu locality group is to reduce the contention
103 * between CPUs. It is possible to get scheduled at this point.
105 * The locality group prealloc space is used looking at whether we have
106 * enough free space (pa_free) within the prealloc space.
108 * If we can't allocate blocks via inode prealloc or/and locality group
109 * prealloc then we look at the buddy cache. The buddy cache is represented
110 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
111 * mapped to the buddy and bitmap information regarding different
112 * groups. The buddy information is attached to buddy cache inode so that
113 * we can access them through the page cache. The information regarding
114 * each group is loaded via ext4_mb_load_buddy. The information involve
115 * block bitmap and buddy information. The information are stored in the
119 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
122 * one block each for bitmap and buddy information. So for each group we
123 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
124 * blocksize) blocks. So it can have information regarding groups_per_page
125 * which is blocks_per_page/2
127 * The buddy cache inode is not stored on disk. The inode is thrown
128 * away when the filesystem is unmounted.
130 * We look for count number of blocks in the buddy cache. If we were able
131 * to locate that many free blocks we return with additional information
132 * regarding rest of the contiguous physical block available
134 * Before allocating blocks via buddy cache we normalize the request
135 * blocks. This ensure we ask for more blocks that we needed. The extra
136 * blocks that we get after allocation is added to the respective prealloc
137 * list. In case of inode preallocation we follow a list of heuristics
138 * based on file size. This can be found in ext4_mb_normalize_request. If
139 * we are doing a group prealloc we try to normalize the request to
140 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
141 * dependent on the cluster size; for non-bigalloc file systems, it is
142 * 512 blocks. This can be tuned via
143 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
144 * terms of number of blocks. If we have mounted the file system with -O
145 * stripe=<value> option the group prealloc request is normalized to the
146 * the smallest multiple of the stripe value (sbi->s_stripe) which is
147 * greater than the default mb_group_prealloc.
149 * The regular allocator (using the buddy cache) supports a few tunables.
151 * /sys/fs/ext4/<partition>/mb_min_to_scan
152 * /sys/fs/ext4/<partition>/mb_max_to_scan
153 * /sys/fs/ext4/<partition>/mb_order2_req
155 * The regular allocator uses buddy scan only if the request len is power of
156 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
157 * value of s_mb_order2_reqs can be tuned via
158 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
159 * stripe size (sbi->s_stripe), we try to search for contiguous block in
160 * stripe size. This should result in better allocation on RAID setups. If
161 * not, we search in the specific group using bitmap for best extents. The
162 * tunable min_to_scan and max_to_scan control the behaviour here.
163 * min_to_scan indicate how long the mballoc __must__ look for a best
164 * extent and max_to_scan indicates how long the mballoc __can__ look for a
165 * best extent in the found extents. Searching for the blocks starts with
166 * the group specified as the goal value in allocation context via
167 * ac_g_ex. Each group is first checked based on the criteria whether it
168 * can be used for allocation. ext4_mb_good_group explains how the groups are
171 * Both the prealloc space are getting populated as above. So for the first
172 * request we will hit the buddy cache which will result in this prealloc
173 * space getting filled. The prealloc space is then later used for the
174 * subsequent request.
178 * mballoc operates on the following data:
180 * - in-core buddy (actually includes buddy and bitmap)
181 * - preallocation descriptors (PAs)
183 * there are two types of preallocations:
185 * assiged to specific inode and can be used for this inode only.
186 * it describes part of inode's space preallocated to specific
187 * physical blocks. any block from that preallocated can be used
188 * independent. the descriptor just tracks number of blocks left
189 * unused. so, before taking some block from descriptor, one must
190 * make sure corresponded logical block isn't allocated yet. this
191 * also means that freeing any block within descriptor's range
192 * must discard all preallocated blocks.
194 * assigned to specific locality group which does not translate to
195 * permanent set of inodes: inode can join and leave group. space
196 * from this type of preallocation can be used for any inode. thus
197 * it's consumed from the beginning to the end.
199 * relation between them can be expressed as:
200 * in-core buddy = on-disk bitmap + preallocation descriptors
202 * this mean blocks mballoc considers used are:
203 * - allocated blocks (persistent)
204 * - preallocated blocks (non-persistent)
206 * consistency in mballoc world means that at any time a block is either
207 * free or used in ALL structures. notice: "any time" should not be read
208 * literally -- time is discrete and delimited by locks.
210 * to keep it simple, we don't use block numbers, instead we count number of
211 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
213 * all operations can be expressed as:
214 * - init buddy: buddy = on-disk + PAs
215 * - new PA: buddy += N; PA = N
216 * - use inode PA: on-disk += N; PA -= N
217 * - discard inode PA buddy -= on-disk - PA; PA = 0
218 * - use locality group PA on-disk += N; PA -= N
219 * - discard locality group PA buddy -= PA; PA = 0
220 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
221 * is used in real operation because we can't know actual used
222 * bits from PA, only from on-disk bitmap
224 * if we follow this strict logic, then all operations above should be atomic.
225 * given some of them can block, we'd have to use something like semaphores
226 * killing performance on high-end SMP hardware. let's try to relax it using
227 * the following knowledge:
228 * 1) if buddy is referenced, it's already initialized
229 * 2) while block is used in buddy and the buddy is referenced,
230 * nobody can re-allocate that block
231 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
232 * bit set and PA claims same block, it's OK. IOW, one can set bit in
233 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
236 * so, now we're building a concurrency table:
239 * blocks for PA are allocated in the buddy, buddy must be referenced
240 * until PA is linked to allocation group to avoid concurrent buddy init
242 * we need to make sure that either on-disk bitmap or PA has uptodate data
243 * given (3) we care that PA-=N operation doesn't interfere with init
245 * the simplest way would be to have buddy initialized by the discard
246 * - use locality group PA
247 * again PA-=N must be serialized with init
248 * - discard locality group PA
249 * the simplest way would be to have buddy initialized by the discard
252 * i_data_sem serializes them
254 * discard process must wait until PA isn't used by another process
255 * - use locality group PA
256 * some mutex should serialize them
257 * - discard locality group PA
258 * discard process must wait until PA isn't used by another process
261 * i_data_sem or another mutex should serializes them
263 * discard process must wait until PA isn't used by another process
264 * - use locality group PA
265 * nothing wrong here -- they're different PAs covering different blocks
266 * - discard locality group PA
267 * discard process must wait until PA isn't used by another process
269 * now we're ready to make few consequences:
270 * - PA is referenced and while it is no discard is possible
271 * - PA is referenced until block isn't marked in on-disk bitmap
272 * - PA changes only after on-disk bitmap
273 * - discard must not compete with init. either init is done before
274 * any discard or they're serialized somehow
275 * - buddy init as sum of on-disk bitmap and PAs is done atomically
277 * a special case when we've used PA to emptiness. no need to modify buddy
278 * in this case, but we should care about concurrent init
283 * Logic in few words:
288 * mark bits in on-disk bitmap
291 * - use preallocation:
292 * find proper PA (per-inode or group)
294 * mark bits in on-disk bitmap
300 * mark bits in on-disk bitmap
303 * - discard preallocations in group:
305 * move them onto local list
306 * load on-disk bitmap
308 * remove PA from object (inode or locality group)
309 * mark free blocks in-core
311 * - discard inode's preallocations:
318 * - bitlock on a group (group)
319 * - object (inode/locality) (object)
330 * - release consumed pa:
335 * - generate in-core bitmap:
339 * - discard all for given object (inode, locality group):
344 * - discard all for given group:
351 static struct kmem_cache
*ext4_pspace_cachep
;
352 static struct kmem_cache
*ext4_ac_cachep
;
353 static struct kmem_cache
*ext4_free_data_cachep
;
355 /* We create slab caches for groupinfo data structures based on the
356 * superblock block size. There will be one per mounted filesystem for
357 * each unique s_blocksize_bits */
358 #define NR_GRPINFO_CACHES 8
359 static struct kmem_cache
*ext4_groupinfo_caches
[NR_GRPINFO_CACHES
];
361 static const char * const ext4_groupinfo_slab_names
[NR_GRPINFO_CACHES
] = {
362 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
363 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
364 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
367 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
369 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
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
);
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_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
, gfp_t gfp
)
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
= i_blocksize(inode
);
841 blocks_per_page
= PAGE_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
);
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 bh
[i
] = ext4_read_block_bitmap_nowait(sb
, group
);
878 err
= PTR_ERR(bh
[i
]);
882 mb_debug(1, "read bitmap for group %u\n", group
);
885 /* wait for I/O completion */
886 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
891 err2
= ext4_wait_block_bitmap(sb
, group
, bh
[i
]);
896 first_block
= page
->index
* blocks_per_page
;
897 for (i
= 0; i
< blocks_per_page
; i
++) {
898 group
= (first_block
+ i
) >> 1;
899 if (group
>= ngroups
)
902 if (!bh
[group
- first_group
])
903 /* skip initialized uptodate buddy */
906 if (!buffer_verified(bh
[group
- first_group
]))
907 /* Skip faulty bitmaps */
912 * data carry information regarding this
913 * particular group in the format specified
917 data
= page_address(page
) + (i
* blocksize
);
918 bitmap
= bh
[group
- first_group
]->b_data
;
921 * We place the buddy block and bitmap block
924 if ((first_block
+ i
) & 1) {
925 /* this is block of buddy */
926 BUG_ON(incore
== NULL
);
927 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
928 group
, page
->index
, i
* blocksize
);
929 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
930 grinfo
= ext4_get_group_info(sb
, group
);
931 grinfo
->bb_fragments
= 0;
932 memset(grinfo
->bb_counters
, 0,
933 sizeof(*grinfo
->bb_counters
) *
934 (sb
->s_blocksize_bits
+2));
936 * incore got set to the group block bitmap below
938 ext4_lock_group(sb
, group
);
940 memset(data
, 0xff, blocksize
);
941 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
942 ext4_unlock_group(sb
, group
);
945 /* this is block of bitmap */
946 BUG_ON(incore
!= NULL
);
947 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
948 group
, page
->index
, i
* blocksize
);
949 trace_ext4_mb_bitmap_load(sb
, group
);
951 /* see comments in ext4_mb_put_pa() */
952 ext4_lock_group(sb
, group
);
953 memcpy(data
, bitmap
, blocksize
);
955 /* mark all preallocated blks used in in-core bitmap */
956 ext4_mb_generate_from_pa(sb
, data
, group
);
957 ext4_mb_generate_from_freelist(sb
, data
, group
);
958 ext4_unlock_group(sb
, group
);
960 /* set incore so that the buddy information can be
961 * generated using this
966 SetPageUptodate(page
);
970 for (i
= 0; i
< groups_per_page
; i
++)
979 * Lock the buddy and bitmap pages. This make sure other parallel init_group
980 * on the same buddy page doesn't happen whild holding the buddy page lock.
981 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
982 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
984 static int ext4_mb_get_buddy_page_lock(struct super_block
*sb
,
985 ext4_group_t group
, struct ext4_buddy
*e4b
, gfp_t gfp
)
987 struct inode
*inode
= EXT4_SB(sb
)->s_buddy_cache
;
988 int block
, pnum
, poff
;
992 e4b
->bd_buddy_page
= NULL
;
993 e4b
->bd_bitmap_page
= NULL
;
995 blocks_per_page
= PAGE_SIZE
/ sb
->s_blocksize
;
997 * the buddy cache inode stores the block bitmap
998 * and buddy information in consecutive blocks.
999 * So for each group we need two blocks.
1002 pnum
= block
/ blocks_per_page
;
1003 poff
= block
% blocks_per_page
;
1004 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1007 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1008 e4b
->bd_bitmap_page
= page
;
1009 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1011 if (blocks_per_page
>= 2) {
1012 /* buddy and bitmap are on the same page */
1017 pnum
= block
/ blocks_per_page
;
1018 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1021 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1022 e4b
->bd_buddy_page
= page
;
1026 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy
*e4b
)
1028 if (e4b
->bd_bitmap_page
) {
1029 unlock_page(e4b
->bd_bitmap_page
);
1030 put_page(e4b
->bd_bitmap_page
);
1032 if (e4b
->bd_buddy_page
) {
1033 unlock_page(e4b
->bd_buddy_page
);
1034 put_page(e4b
->bd_buddy_page
);
1039 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1040 * block group lock of all groups for this page; do not hold the BG lock when
1041 * calling this routine!
1043 static noinline_for_stack
1044 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
, gfp_t gfp
)
1047 struct ext4_group_info
*this_grp
;
1048 struct ext4_buddy e4b
;
1053 mb_debug(1, "init group %u\n", group
);
1054 this_grp
= ext4_get_group_info(sb
, group
);
1056 * This ensures that we don't reinit the buddy cache
1057 * page which map to the group from which we are already
1058 * allocating. If we are looking at the buddy cache we would
1059 * have taken a reference using ext4_mb_load_buddy and that
1060 * would have pinned buddy page to page cache.
1061 * The call to ext4_mb_get_buddy_page_lock will mark the
1064 ret
= ext4_mb_get_buddy_page_lock(sb
, group
, &e4b
, gfp
);
1065 if (ret
|| !EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1067 * somebody initialized the group
1068 * return without doing anything
1073 page
= e4b
.bd_bitmap_page
;
1074 ret
= ext4_mb_init_cache(page
, NULL
, gfp
);
1077 if (!PageUptodate(page
)) {
1082 if (e4b
.bd_buddy_page
== NULL
) {
1084 * If both the bitmap and buddy are in
1085 * the same page we don't need to force
1091 /* init buddy cache */
1092 page
= e4b
.bd_buddy_page
;
1093 ret
= ext4_mb_init_cache(page
, e4b
.bd_bitmap
, gfp
);
1096 if (!PageUptodate(page
)) {
1101 ext4_mb_put_buddy_page_lock(&e4b
);
1106 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1107 * block group lock of all groups for this page; do not hold the BG lock when
1108 * calling this routine!
1110 static noinline_for_stack
int
1111 ext4_mb_load_buddy_gfp(struct super_block
*sb
, ext4_group_t group
,
1112 struct ext4_buddy
*e4b
, gfp_t gfp
)
1114 int blocks_per_page
;
1120 struct ext4_group_info
*grp
;
1121 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1122 struct inode
*inode
= sbi
->s_buddy_cache
;
1125 mb_debug(1, "load group %u\n", group
);
1127 blocks_per_page
= PAGE_SIZE
/ sb
->s_blocksize
;
1128 grp
= ext4_get_group_info(sb
, group
);
1130 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1133 e4b
->bd_group
= group
;
1134 e4b
->bd_buddy_page
= NULL
;
1135 e4b
->bd_bitmap_page
= NULL
;
1137 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1139 * we need full data about the group
1140 * to make a good selection
1142 ret
= ext4_mb_init_group(sb
, group
, gfp
);
1148 * the buddy cache inode stores the block bitmap
1149 * and buddy information in consecutive blocks.
1150 * So for each group we need two blocks.
1153 pnum
= block
/ blocks_per_page
;
1154 poff
= block
% blocks_per_page
;
1156 /* we could use find_or_create_page(), but it locks page
1157 * what we'd like to avoid in fast path ... */
1158 page
= find_get_page_flags(inode
->i_mapping
, pnum
, FGP_ACCESSED
);
1159 if (page
== NULL
|| !PageUptodate(page
)) {
1162 * drop the page reference and try
1163 * to get the page with lock. If we
1164 * are not uptodate that implies
1165 * somebody just created the page but
1166 * is yet to initialize the same. So
1167 * wait for it to initialize.
1170 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1172 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1173 if (!PageUptodate(page
)) {
1174 ret
= ext4_mb_init_cache(page
, NULL
, gfp
);
1179 mb_cmp_bitmaps(e4b
, page_address(page
) +
1180 (poff
* sb
->s_blocksize
));
1189 if (!PageUptodate(page
)) {
1194 /* Pages marked accessed already */
1195 e4b
->bd_bitmap_page
= page
;
1196 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1199 pnum
= block
/ blocks_per_page
;
1200 poff
= block
% blocks_per_page
;
1202 page
= find_get_page_flags(inode
->i_mapping
, pnum
, FGP_ACCESSED
);
1203 if (page
== NULL
|| !PageUptodate(page
)) {
1206 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1208 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1209 if (!PageUptodate(page
)) {
1210 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
,
1224 if (!PageUptodate(page
)) {
1229 /* Pages marked accessed already */
1230 e4b
->bd_buddy_page
= page
;
1231 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1233 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1234 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1241 if (e4b
->bd_bitmap_page
)
1242 put_page(e4b
->bd_bitmap_page
);
1243 if (e4b
->bd_buddy_page
)
1244 put_page(e4b
->bd_buddy_page
);
1245 e4b
->bd_buddy
= NULL
;
1246 e4b
->bd_bitmap
= NULL
;
1250 static int ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1251 struct ext4_buddy
*e4b
)
1253 return ext4_mb_load_buddy_gfp(sb
, group
, e4b
, GFP_NOFS
);
1256 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1258 if (e4b
->bd_bitmap_page
)
1259 put_page(e4b
->bd_bitmap_page
);
1260 if (e4b
->bd_buddy_page
)
1261 put_page(e4b
->bd_buddy_page
);
1265 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1268 int bb_incr
= 1 << (e4b
->bd_blkbits
- 1);
1271 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
1272 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1275 while (order
<= e4b
->bd_blkbits
+ 1) {
1277 if (!mb_test_bit(block
, bb
)) {
1278 /* this block is part of buddy of order 'order' */
1288 static void mb_clear_bits(void *bm
, int cur
, int len
)
1294 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1295 /* fast path: clear whole word at once */
1296 addr
= bm
+ (cur
>> 3);
1301 mb_clear_bit(cur
, bm
);
1306 /* clear bits in given range
1307 * will return first found zero bit if any, -1 otherwise
1309 static int mb_test_and_clear_bits(void *bm
, int cur
, int len
)
1316 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1317 /* fast path: clear whole word at once */
1318 addr
= bm
+ (cur
>> 3);
1319 if (*addr
!= (__u32
)(-1) && zero_bit
== -1)
1320 zero_bit
= cur
+ mb_find_next_zero_bit(addr
, 32, 0);
1325 if (!mb_test_and_clear_bit(cur
, bm
) && zero_bit
== -1)
1333 void ext4_set_bits(void *bm
, int cur
, int len
)
1339 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1340 /* fast path: set whole word at once */
1341 addr
= bm
+ (cur
>> 3);
1346 mb_set_bit(cur
, bm
);
1352 * _________________________________________________________________ */
1354 static inline int mb_buddy_adjust_border(int* bit
, void* bitmap
, int side
)
1356 if (mb_test_bit(*bit
+ side
, bitmap
)) {
1357 mb_clear_bit(*bit
, bitmap
);
1363 mb_set_bit(*bit
, bitmap
);
1368 static void mb_buddy_mark_free(struct ext4_buddy
*e4b
, int first
, int last
)
1372 void *buddy
= mb_find_buddy(e4b
, order
, &max
);
1377 /* Bits in range [first; last] are known to be set since
1378 * corresponding blocks were allocated. Bits in range
1379 * (first; last) will stay set because they form buddies on
1380 * upper layer. We just deal with borders if they don't
1381 * align with upper layer and then go up.
1382 * Releasing entire group is all about clearing
1383 * single bit of highest order buddy.
1387 * ---------------------------------
1389 * ---------------------------------
1390 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1391 * ---------------------------------
1393 * \_____________________/
1395 * Neither [1] nor [6] is aligned to above layer.
1396 * Left neighbour [0] is free, so mark it busy,
1397 * decrease bb_counters and extend range to
1399 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1400 * mark [6] free, increase bb_counters and shrink range to
1402 * Then shift range to [0; 2], go up and do the same.
1407 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&first
, buddy
, -1);
1409 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&last
, buddy
, 1);
1414 if (first
== last
|| !(buddy2
= mb_find_buddy(e4b
, order
, &max
))) {
1415 mb_clear_bits(buddy
, first
, last
- first
+ 1);
1416 e4b
->bd_info
->bb_counters
[order
- 1] += last
- first
+ 1;
1425 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1426 int first
, int count
)
1428 int left_is_free
= 0;
1429 int right_is_free
= 0;
1431 int last
= first
+ count
- 1;
1432 struct super_block
*sb
= e4b
->bd_sb
;
1434 if (WARN_ON(count
== 0))
1436 BUG_ON(last
>= (sb
->s_blocksize
<< 3));
1437 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1438 /* Don't bother if the block group is corrupt. */
1439 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
)))
1442 mb_check_buddy(e4b
);
1443 mb_free_blocks_double(inode
, e4b
, first
, count
);
1445 e4b
->bd_info
->bb_free
+= count
;
1446 if (first
< e4b
->bd_info
->bb_first_free
)
1447 e4b
->bd_info
->bb_first_free
= first
;
1449 /* access memory sequentially: check left neighbour,
1450 * clear range and then check right neighbour
1453 left_is_free
= !mb_test_bit(first
- 1, e4b
->bd_bitmap
);
1454 block
= mb_test_and_clear_bits(e4b
->bd_bitmap
, first
, count
);
1455 if (last
+ 1 < EXT4_SB(sb
)->s_mb_maxs
[0])
1456 right_is_free
= !mb_test_bit(last
+ 1, e4b
->bd_bitmap
);
1458 if (unlikely(block
!= -1)) {
1459 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1460 ext4_fsblk_t blocknr
;
1462 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1463 blocknr
+= EXT4_C2B(EXT4_SB(sb
), block
);
1464 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1465 inode
? inode
->i_ino
: 0,
1467 "freeing already freed block "
1468 "(bit %u); block bitmap corrupt.",
1470 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
))
1471 percpu_counter_sub(&sbi
->s_freeclusters_counter
,
1472 e4b
->bd_info
->bb_free
);
1473 /* Mark the block group as corrupt. */
1474 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT
,
1475 &e4b
->bd_info
->bb_state
);
1476 mb_regenerate_buddy(e4b
);
1480 /* let's maintain fragments counter */
1481 if (left_is_free
&& right_is_free
)
1482 e4b
->bd_info
->bb_fragments
--;
1483 else if (!left_is_free
&& !right_is_free
)
1484 e4b
->bd_info
->bb_fragments
++;
1486 /* buddy[0] == bd_bitmap is a special case, so handle
1487 * it right away and let mb_buddy_mark_free stay free of
1488 * zero order checks.
1489 * Check if neighbours are to be coaleasced,
1490 * adjust bitmap bb_counters and borders appropriately.
1493 first
+= !left_is_free
;
1494 e4b
->bd_info
->bb_counters
[0] += left_is_free
? -1 : 1;
1497 last
-= !right_is_free
;
1498 e4b
->bd_info
->bb_counters
[0] += right_is_free
? -1 : 1;
1502 mb_buddy_mark_free(e4b
, first
>> 1, last
>> 1);
1505 mb_set_largest_free_order(sb
, e4b
->bd_info
);
1506 mb_check_buddy(e4b
);
1509 static int mb_find_extent(struct ext4_buddy
*e4b
, int block
,
1510 int needed
, struct ext4_free_extent
*ex
)
1516 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1519 buddy
= mb_find_buddy(e4b
, 0, &max
);
1520 BUG_ON(buddy
== NULL
);
1521 BUG_ON(block
>= max
);
1522 if (mb_test_bit(block
, buddy
)) {
1529 /* find actual order */
1530 order
= mb_find_order_for_block(e4b
, block
);
1531 block
= block
>> order
;
1533 ex
->fe_len
= 1 << order
;
1534 ex
->fe_start
= block
<< order
;
1535 ex
->fe_group
= e4b
->bd_group
;
1537 /* calc difference from given start */
1538 next
= next
- ex
->fe_start
;
1540 ex
->fe_start
+= next
;
1542 while (needed
> ex
->fe_len
&&
1543 mb_find_buddy(e4b
, order
, &max
)) {
1545 if (block
+ 1 >= max
)
1548 next
= (block
+ 1) * (1 << order
);
1549 if (mb_test_bit(next
, e4b
->bd_bitmap
))
1552 order
= mb_find_order_for_block(e4b
, next
);
1554 block
= next
>> order
;
1555 ex
->fe_len
+= 1 << order
;
1558 if (ex
->fe_start
+ ex
->fe_len
> EXT4_CLUSTERS_PER_GROUP(e4b
->bd_sb
)) {
1559 /* Should never happen! (but apparently sometimes does?!?) */
1561 ext4_error(e4b
->bd_sb
, "corruption or bug in mb_find_extent "
1562 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1563 block
, order
, needed
, ex
->fe_group
, ex
->fe_start
,
1564 ex
->fe_len
, ex
->fe_logical
);
1572 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1578 int start
= ex
->fe_start
;
1579 int len
= ex
->fe_len
;
1584 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1585 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1586 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1587 mb_check_buddy(e4b
);
1588 mb_mark_used_double(e4b
, start
, len
);
1590 e4b
->bd_info
->bb_free
-= len
;
1591 if (e4b
->bd_info
->bb_first_free
== start
)
1592 e4b
->bd_info
->bb_first_free
+= len
;
1594 /* let's maintain fragments counter */
1596 mlen
= !mb_test_bit(start
- 1, e4b
->bd_bitmap
);
1597 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1598 max
= !mb_test_bit(start
+ len
, e4b
->bd_bitmap
);
1600 e4b
->bd_info
->bb_fragments
++;
1601 else if (!mlen
&& !max
)
1602 e4b
->bd_info
->bb_fragments
--;
1604 /* let's maintain buddy itself */
1606 ord
= mb_find_order_for_block(e4b
, start
);
1608 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1609 /* the whole chunk may be allocated at once! */
1611 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1612 BUG_ON((start
>> ord
) >= max
);
1613 mb_set_bit(start
>> ord
, buddy
);
1614 e4b
->bd_info
->bb_counters
[ord
]--;
1621 /* store for history */
1623 ret
= len
| (ord
<< 16);
1625 /* we have to split large buddy */
1627 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1628 mb_set_bit(start
>> ord
, buddy
);
1629 e4b
->bd_info
->bb_counters
[ord
]--;
1632 cur
= (start
>> ord
) & ~1U;
1633 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1634 mb_clear_bit(cur
, buddy
);
1635 mb_clear_bit(cur
+ 1, buddy
);
1636 e4b
->bd_info
->bb_counters
[ord
]++;
1637 e4b
->bd_info
->bb_counters
[ord
]++;
1639 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1641 ext4_set_bits(e4b
->bd_bitmap
, ex
->fe_start
, len0
);
1642 mb_check_buddy(e4b
);
1648 * Must be called under group lock!
1650 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1651 struct ext4_buddy
*e4b
)
1653 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1656 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1657 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1659 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1660 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1661 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1663 /* preallocation can change ac_b_ex, thus we store actually
1664 * allocated blocks for history */
1665 ac
->ac_f_ex
= ac
->ac_b_ex
;
1667 ac
->ac_status
= AC_STATUS_FOUND
;
1668 ac
->ac_tail
= ret
& 0xffff;
1669 ac
->ac_buddy
= ret
>> 16;
1672 * take the page reference. We want the page to be pinned
1673 * so that we don't get a ext4_mb_init_cache_call for this
1674 * group until we update the bitmap. That would mean we
1675 * double allocate blocks. The reference is dropped
1676 * in ext4_mb_release_context
1678 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1679 get_page(ac
->ac_bitmap_page
);
1680 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1681 get_page(ac
->ac_buddy_page
);
1682 /* store last allocated for subsequent stream allocation */
1683 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1684 spin_lock(&sbi
->s_md_lock
);
1685 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1686 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1687 spin_unlock(&sbi
->s_md_lock
);
1692 * regular allocator, for general purposes allocation
1695 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1696 struct ext4_buddy
*e4b
,
1699 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1700 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1701 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1702 struct ext4_free_extent ex
;
1705 if (ac
->ac_status
== AC_STATUS_FOUND
)
1708 * We don't want to scan for a whole year
1710 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1711 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1712 ac
->ac_status
= AC_STATUS_BREAK
;
1717 * Haven't found good chunk so far, let's continue
1719 if (bex
->fe_len
< gex
->fe_len
)
1722 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1723 && bex
->fe_group
== e4b
->bd_group
) {
1724 /* recheck chunk's availability - we don't know
1725 * when it was found (within this lock-unlock
1727 max
= mb_find_extent(e4b
, bex
->fe_start
, gex
->fe_len
, &ex
);
1728 if (max
>= gex
->fe_len
) {
1729 ext4_mb_use_best_found(ac
, e4b
);
1736 * The routine checks whether found extent is good enough. If it is,
1737 * then the extent gets marked used and flag is set to the context
1738 * to stop scanning. Otherwise, the extent is compared with the
1739 * previous found extent and if new one is better, then it's stored
1740 * in the context. Later, the best found extent will be used, if
1741 * mballoc can't find good enough extent.
1743 * FIXME: real allocation policy is to be designed yet!
1745 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1746 struct ext4_free_extent
*ex
,
1747 struct ext4_buddy
*e4b
)
1749 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1750 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1752 BUG_ON(ex
->fe_len
<= 0);
1753 BUG_ON(ex
->fe_len
> EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1754 BUG_ON(ex
->fe_start
>= EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1755 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1760 * The special case - take what you catch first
1762 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1764 ext4_mb_use_best_found(ac
, e4b
);
1769 * Let's check whether the chuck is good enough
1771 if (ex
->fe_len
== gex
->fe_len
) {
1773 ext4_mb_use_best_found(ac
, e4b
);
1778 * If this is first found extent, just store it in the context
1780 if (bex
->fe_len
== 0) {
1786 * If new found extent is better, store it in the context
1788 if (bex
->fe_len
< gex
->fe_len
) {
1789 /* if the request isn't satisfied, any found extent
1790 * larger than previous best one is better */
1791 if (ex
->fe_len
> bex
->fe_len
)
1793 } else if (ex
->fe_len
> gex
->fe_len
) {
1794 /* if the request is satisfied, then we try to find
1795 * an extent that still satisfy the request, but is
1796 * smaller than previous one */
1797 if (ex
->fe_len
< bex
->fe_len
)
1801 ext4_mb_check_limits(ac
, e4b
, 0);
1804 static noinline_for_stack
1805 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1806 struct ext4_buddy
*e4b
)
1808 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1809 ext4_group_t group
= ex
.fe_group
;
1813 BUG_ON(ex
.fe_len
<= 0);
1814 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1818 ext4_lock_group(ac
->ac_sb
, group
);
1819 max
= mb_find_extent(e4b
, ex
.fe_start
, ex
.fe_len
, &ex
);
1823 ext4_mb_use_best_found(ac
, e4b
);
1826 ext4_unlock_group(ac
->ac_sb
, group
);
1827 ext4_mb_unload_buddy(e4b
);
1832 static noinline_for_stack
1833 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1834 struct ext4_buddy
*e4b
)
1836 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1839 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1840 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1841 struct ext4_free_extent ex
;
1843 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1845 if (grp
->bb_free
== 0)
1848 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1852 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
))) {
1853 ext4_mb_unload_buddy(e4b
);
1857 ext4_lock_group(ac
->ac_sb
, group
);
1858 max
= mb_find_extent(e4b
, ac
->ac_g_ex
.fe_start
,
1859 ac
->ac_g_ex
.fe_len
, &ex
);
1860 ex
.fe_logical
= 0xDEADFA11; /* debug value */
1862 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1865 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1867 /* use do_div to get remainder (would be 64-bit modulo) */
1868 if (do_div(start
, sbi
->s_stripe
) == 0) {
1871 ext4_mb_use_best_found(ac
, e4b
);
1873 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1874 BUG_ON(ex
.fe_len
<= 0);
1875 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1876 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1879 ext4_mb_use_best_found(ac
, e4b
);
1880 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1881 /* Sometimes, caller may want to merge even small
1882 * number of blocks to an existing extent */
1883 BUG_ON(ex
.fe_len
<= 0);
1884 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1885 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1888 ext4_mb_use_best_found(ac
, e4b
);
1890 ext4_unlock_group(ac
->ac_sb
, group
);
1891 ext4_mb_unload_buddy(e4b
);
1897 * The routine scans buddy structures (not bitmap!) from given order
1898 * to max order and tries to find big enough chunk to satisfy the req
1900 static noinline_for_stack
1901 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1902 struct ext4_buddy
*e4b
)
1904 struct super_block
*sb
= ac
->ac_sb
;
1905 struct ext4_group_info
*grp
= e4b
->bd_info
;
1911 BUG_ON(ac
->ac_2order
<= 0);
1912 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1913 if (grp
->bb_counters
[i
] == 0)
1916 buddy
= mb_find_buddy(e4b
, i
, &max
);
1917 BUG_ON(buddy
== NULL
);
1919 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1924 ac
->ac_b_ex
.fe_len
= 1 << i
;
1925 ac
->ac_b_ex
.fe_start
= k
<< i
;
1926 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1928 ext4_mb_use_best_found(ac
, e4b
);
1930 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1932 if (EXT4_SB(sb
)->s_mb_stats
)
1933 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1940 * The routine scans the group and measures all found extents.
1941 * In order to optimize scanning, caller must pass number of
1942 * free blocks in the group, so the routine can know upper limit.
1944 static noinline_for_stack
1945 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1946 struct ext4_buddy
*e4b
)
1948 struct super_block
*sb
= ac
->ac_sb
;
1949 void *bitmap
= e4b
->bd_bitmap
;
1950 struct ext4_free_extent ex
;
1954 free
= e4b
->bd_info
->bb_free
;
1955 if (WARN_ON(free
<= 0))
1958 i
= e4b
->bd_info
->bb_first_free
;
1960 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1961 i
= mb_find_next_zero_bit(bitmap
,
1962 EXT4_CLUSTERS_PER_GROUP(sb
), i
);
1963 if (i
>= EXT4_CLUSTERS_PER_GROUP(sb
)) {
1965 * IF we have corrupt bitmap, we won't find any
1966 * free blocks even though group info says we
1967 * we have free blocks
1969 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1970 "%d free clusters as per "
1971 "group info. But bitmap says 0",
1976 mb_find_extent(e4b
, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1977 if (WARN_ON(ex
.fe_len
<= 0))
1979 if (free
< ex
.fe_len
) {
1980 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1981 "%d free clusters as per "
1982 "group info. But got %d blocks",
1985 * The number of free blocks differs. This mostly
1986 * indicate that the bitmap is corrupt. So exit
1987 * without claiming the space.
1991 ex
.fe_logical
= 0xDEADC0DE; /* debug value */
1992 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1998 ext4_mb_check_limits(ac
, e4b
, 1);
2002 * This is a special case for storages like raid5
2003 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2005 static noinline_for_stack
2006 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
2007 struct ext4_buddy
*e4b
)
2009 struct super_block
*sb
= ac
->ac_sb
;
2010 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2011 void *bitmap
= e4b
->bd_bitmap
;
2012 struct ext4_free_extent ex
;
2013 ext4_fsblk_t first_group_block
;
2018 BUG_ON(sbi
->s_stripe
== 0);
2020 /* find first stripe-aligned block in group */
2021 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
2023 a
= first_group_block
+ sbi
->s_stripe
- 1;
2024 do_div(a
, sbi
->s_stripe
);
2025 i
= (a
* sbi
->s_stripe
) - first_group_block
;
2027 while (i
< EXT4_CLUSTERS_PER_GROUP(sb
)) {
2028 if (!mb_test_bit(i
, bitmap
)) {
2029 max
= mb_find_extent(e4b
, i
, sbi
->s_stripe
, &ex
);
2030 if (max
>= sbi
->s_stripe
) {
2032 ex
.fe_logical
= 0xDEADF00D; /* debug value */
2034 ext4_mb_use_best_found(ac
, e4b
);
2043 * This is now called BEFORE we load the buddy bitmap.
2044 * Returns either 1 or 0 indicating that the group is either suitable
2045 * for the allocation or not. In addition it can also return negative
2046 * error code when something goes wrong.
2048 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
2049 ext4_group_t group
, int cr
)
2051 unsigned free
, fragments
;
2052 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
2053 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
2055 BUG_ON(cr
< 0 || cr
>= 4);
2057 free
= grp
->bb_free
;
2060 if (cr
<= 2 && free
< ac
->ac_g_ex
.fe_len
)
2063 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp
)))
2066 /* We only do this if the grp has never been initialized */
2067 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
2068 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
, GFP_NOFS
);
2073 fragments
= grp
->bb_fragments
;
2079 BUG_ON(ac
->ac_2order
== 0);
2081 /* Avoid using the first bg of a flexgroup for data files */
2082 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
2083 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
2084 ((group
% flex_size
) == 0))
2087 if ((ac
->ac_2order
> ac
->ac_sb
->s_blocksize_bits
+1) ||
2088 (free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2091 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
2096 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2100 if (free
>= ac
->ac_g_ex
.fe_len
)
2112 static noinline_for_stack
int
2113 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
2115 ext4_group_t ngroups
, group
, i
;
2117 int err
= 0, first_err
= 0;
2118 struct ext4_sb_info
*sbi
;
2119 struct super_block
*sb
;
2120 struct ext4_buddy e4b
;
2124 ngroups
= ext4_get_groups_count(sb
);
2125 /* non-extent files are limited to low blocks/groups */
2126 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
2127 ngroups
= sbi
->s_blockfile_groups
;
2129 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
2131 /* first, try the goal */
2132 err
= ext4_mb_find_by_goal(ac
, &e4b
);
2133 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
2136 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2140 * ac->ac2_order is set only if the fe_len is a power of 2
2141 * if ac2_order is set we also set criteria to 0 so that we
2142 * try exact allocation using buddy.
2144 i
= fls(ac
->ac_g_ex
.fe_len
);
2147 * We search using buddy data only if the order of the request
2148 * is greater than equal to the sbi_s_mb_order2_reqs
2149 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2150 * We also support searching for power-of-two requests only for
2151 * requests upto maximum buddy size we have constructed.
2153 if (i
>= sbi
->s_mb_order2_reqs
&& i
<= sb
->s_blocksize_bits
+ 2) {
2155 * This should tell if fe_len is exactly power of 2
2157 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2158 ac
->ac_2order
= array_index_nospec(i
- 1,
2159 sb
->s_blocksize_bits
+ 2);
2162 /* if stream allocation is enabled, use global goal */
2163 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2164 /* TBD: may be hot point */
2165 spin_lock(&sbi
->s_md_lock
);
2166 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2167 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2168 spin_unlock(&sbi
->s_md_lock
);
2171 /* Let's just scan groups to find more-less suitable blocks */
2172 cr
= ac
->ac_2order
? 0 : 1;
2174 * cr == 0 try to get exact allocation,
2175 * cr == 3 try to get anything
2178 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2179 ac
->ac_criteria
= cr
;
2181 * searching for the right group start
2182 * from the goal value specified
2184 group
= ac
->ac_g_ex
.fe_group
;
2186 for (i
= 0; i
< ngroups
; group
++, i
++) {
2190 * Artificially restricted ngroups for non-extent
2191 * files makes group > ngroups possible on first loop.
2193 if (group
>= ngroups
)
2196 /* This now checks without needing the buddy page */
2197 ret
= ext4_mb_good_group(ac
, group
, cr
);
2204 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2208 ext4_lock_group(sb
, group
);
2211 * We need to check again after locking the
2214 ret
= ext4_mb_good_group(ac
, group
, cr
);
2216 ext4_unlock_group(sb
, group
);
2217 ext4_mb_unload_buddy(&e4b
);
2223 ac
->ac_groups_scanned
++;
2225 ext4_mb_simple_scan_group(ac
, &e4b
);
2226 else if (cr
== 1 && sbi
->s_stripe
&&
2227 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2228 ext4_mb_scan_aligned(ac
, &e4b
);
2230 ext4_mb_complex_scan_group(ac
, &e4b
);
2232 ext4_unlock_group(sb
, group
);
2233 ext4_mb_unload_buddy(&e4b
);
2235 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2240 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2241 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2243 * We've been searching too long. Let's try to allocate
2244 * the best chunk we've found so far
2247 ext4_mb_try_best_found(ac
, &e4b
);
2248 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2250 * Someone more lucky has already allocated it.
2251 * The only thing we can do is just take first
2253 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2255 ac
->ac_b_ex
.fe_group
= 0;
2256 ac
->ac_b_ex
.fe_start
= 0;
2257 ac
->ac_b_ex
.fe_len
= 0;
2258 ac
->ac_status
= AC_STATUS_CONTINUE
;
2259 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2261 atomic_inc(&sbi
->s_mb_lost_chunks
);
2266 if (!err
&& ac
->ac_status
!= AC_STATUS_FOUND
&& first_err
)
2271 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2273 struct super_block
*sb
= seq
->private;
2276 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2279 return (void *) ((unsigned long) group
);
2282 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2284 struct super_block
*sb
= seq
->private;
2288 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2291 return (void *) ((unsigned long) group
);
2294 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2296 struct super_block
*sb
= seq
->private;
2297 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2299 int err
, buddy_loaded
= 0;
2300 struct ext4_buddy e4b
;
2301 struct ext4_group_info
*grinfo
;
2302 unsigned char blocksize_bits
= min_t(unsigned char,
2303 sb
->s_blocksize_bits
,
2304 EXT4_MAX_BLOCK_LOG_SIZE
);
2306 struct ext4_group_info info
;
2307 ext4_grpblk_t counters
[EXT4_MAX_BLOCK_LOG_SIZE
+ 2];
2312 seq_puts(seq
, "#group: free frags first ["
2313 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2314 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2316 i
= (blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2317 sizeof(struct ext4_group_info
);
2319 grinfo
= ext4_get_group_info(sb
, group
);
2320 /* Load the group info in memory only if not already loaded. */
2321 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo
))) {
2322 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2324 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2330 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2333 ext4_mb_unload_buddy(&e4b
);
2335 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2336 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2337 for (i
= 0; i
<= 13; i
++)
2338 seq_printf(seq
, " %-5u", i
<= blocksize_bits
+ 1 ?
2339 sg
.info
.bb_counters
[i
] : 0);
2340 seq_printf(seq
, " ]\n");
2345 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2349 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2350 .start
= ext4_mb_seq_groups_start
,
2351 .next
= ext4_mb_seq_groups_next
,
2352 .stop
= ext4_mb_seq_groups_stop
,
2353 .show
= ext4_mb_seq_groups_show
,
2356 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2358 struct super_block
*sb
= PDE_DATA(inode
);
2361 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2363 struct seq_file
*m
= file
->private_data
;
2370 const struct file_operations ext4_seq_mb_groups_fops
= {
2371 .open
= ext4_mb_seq_groups_open
,
2373 .llseek
= seq_lseek
,
2374 .release
= seq_release
,
2377 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2379 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2380 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2387 * Allocate the top-level s_group_info array for the specified number
2390 int ext4_mb_alloc_groupinfo(struct super_block
*sb
, ext4_group_t ngroups
)
2392 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2394 struct ext4_group_info
***old_groupinfo
, ***new_groupinfo
;
2396 size
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2397 EXT4_DESC_PER_BLOCK_BITS(sb
);
2398 if (size
<= sbi
->s_group_info_size
)
2401 size
= roundup_pow_of_two(sizeof(*sbi
->s_group_info
) * size
);
2402 new_groupinfo
= kvzalloc(size
, GFP_KERNEL
);
2403 if (!new_groupinfo
) {
2404 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy meta group");
2408 old_groupinfo
= rcu_dereference(sbi
->s_group_info
);
2410 memcpy(new_groupinfo
, old_groupinfo
,
2411 sbi
->s_group_info_size
* sizeof(*sbi
->s_group_info
));
2413 rcu_assign_pointer(sbi
->s_group_info
, new_groupinfo
);
2414 sbi
->s_group_info_size
= size
/ sizeof(*sbi
->s_group_info
);
2416 ext4_kvfree_array_rcu(old_groupinfo
);
2417 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2418 sbi
->s_group_info_size
);
2422 /* Create and initialize ext4_group_info data for the given group. */
2423 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2424 struct ext4_group_desc
*desc
)
2428 int idx
= group
>> EXT4_DESC_PER_BLOCK_BITS(sb
);
2429 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2430 struct ext4_group_info
**meta_group_info
;
2431 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2434 * First check if this group is the first of a reserved block.
2435 * If it's true, we have to allocate a new table of pointers
2436 * to ext4_group_info structures
2438 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2439 metalen
= sizeof(*meta_group_info
) <<
2440 EXT4_DESC_PER_BLOCK_BITS(sb
);
2441 meta_group_info
= kmalloc(metalen
, GFP_NOFS
);
2442 if (meta_group_info
== NULL
) {
2443 ext4_msg(sb
, KERN_ERR
, "can't allocate mem "
2444 "for a buddy group");
2445 goto exit_meta_group_info
;
2448 rcu_dereference(sbi
->s_group_info
)[idx
] = meta_group_info
;
2452 meta_group_info
= sbi_array_rcu_deref(sbi
, s_group_info
, idx
);
2453 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2455 meta_group_info
[i
] = kmem_cache_zalloc(cachep
, GFP_NOFS
);
2456 if (meta_group_info
[i
] == NULL
) {
2457 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy mem");
2458 goto exit_group_info
;
2460 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2461 &(meta_group_info
[i
]->bb_state
));
2464 * initialize bb_free to be able to skip
2465 * empty groups without initialization
2467 if (ext4_has_group_desc_csum(sb
) &&
2468 (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))) {
2469 meta_group_info
[i
]->bb_free
=
2470 ext4_free_clusters_after_init(sb
, group
, desc
);
2472 meta_group_info
[i
]->bb_free
=
2473 ext4_free_group_clusters(sb
, desc
);
2476 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2477 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2478 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2479 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2483 struct buffer_head
*bh
;
2484 meta_group_info
[i
]->bb_bitmap
=
2485 kmalloc(sb
->s_blocksize
, GFP_NOFS
);
2486 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2487 bh
= ext4_read_block_bitmap(sb
, group
);
2488 BUG_ON(IS_ERR_OR_NULL(bh
));
2489 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2498 /* If a meta_group_info table has been allocated, release it now */
2499 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2500 struct ext4_group_info
***group_info
;
2503 group_info
= rcu_dereference(sbi
->s_group_info
);
2504 kfree(group_info
[idx
]);
2505 group_info
[idx
] = NULL
;
2508 exit_meta_group_info
:
2510 } /* ext4_mb_add_groupinfo */
2512 static int ext4_mb_init_backend(struct super_block
*sb
)
2514 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2516 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2518 struct ext4_group_desc
*desc
;
2519 struct ext4_group_info
***group_info
;
2520 struct kmem_cache
*cachep
;
2522 err
= ext4_mb_alloc_groupinfo(sb
, ngroups
);
2526 sbi
->s_buddy_cache
= new_inode(sb
);
2527 if (sbi
->s_buddy_cache
== NULL
) {
2528 ext4_msg(sb
, KERN_ERR
, "can't get new inode");
2531 /* To avoid potentially colliding with an valid on-disk inode number,
2532 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2533 * not in the inode hash, so it should never be found by iget(), but
2534 * this will avoid confusion if it ever shows up during debugging. */
2535 sbi
->s_buddy_cache
->i_ino
= EXT4_BAD_INO
;
2536 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2537 for (i
= 0; i
< ngroups
; i
++) {
2538 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2540 ext4_msg(sb
, KERN_ERR
, "can't read descriptor %u", i
);
2543 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2550 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2552 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2553 i
= sbi
->s_group_info_size
;
2555 group_info
= rcu_dereference(sbi
->s_group_info
);
2557 kfree(group_info
[i
]);
2559 iput(sbi
->s_buddy_cache
);
2562 kvfree(rcu_dereference(sbi
->s_group_info
));
2567 static void ext4_groupinfo_destroy_slabs(void)
2571 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2572 if (ext4_groupinfo_caches
[i
])
2573 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2574 ext4_groupinfo_caches
[i
] = NULL
;
2578 static int ext4_groupinfo_create_slab(size_t size
)
2580 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2582 int blocksize_bits
= order_base_2(size
);
2583 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2584 struct kmem_cache
*cachep
;
2586 if (cache_index
>= NR_GRPINFO_CACHES
)
2589 if (unlikely(cache_index
< 0))
2592 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2593 if (ext4_groupinfo_caches
[cache_index
]) {
2594 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2595 return 0; /* Already created */
2598 slab_size
= offsetof(struct ext4_group_info
,
2599 bb_counters
[blocksize_bits
+ 2]);
2601 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2602 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2605 ext4_groupinfo_caches
[cache_index
] = cachep
;
2607 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2610 "EXT4-fs: no memory for groupinfo slab cache\n");
2617 int ext4_mb_init(struct super_block
*sb
)
2619 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2621 unsigned offset
, offset_incr
;
2625 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2627 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2628 if (sbi
->s_mb_offsets
== NULL
) {
2633 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2634 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2635 if (sbi
->s_mb_maxs
== NULL
) {
2640 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2644 /* order 0 is regular bitmap */
2645 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2646 sbi
->s_mb_offsets
[0] = 0;
2650 offset_incr
= 1 << (sb
->s_blocksize_bits
- 1);
2651 max
= sb
->s_blocksize
<< 2;
2653 sbi
->s_mb_offsets
[i
] = offset
;
2654 sbi
->s_mb_maxs
[i
] = max
;
2655 offset
+= offset_incr
;
2656 offset_incr
= offset_incr
>> 1;
2659 } while (i
<= sb
->s_blocksize_bits
+ 1);
2661 spin_lock_init(&sbi
->s_md_lock
);
2662 spin_lock_init(&sbi
->s_bal_lock
);
2663 sbi
->s_mb_free_pending
= 0;
2664 INIT_LIST_HEAD(&sbi
->s_freed_data_list
);
2666 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2667 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2668 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2669 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2670 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2672 * The default group preallocation is 512, which for 4k block
2673 * sizes translates to 2 megabytes. However for bigalloc file
2674 * systems, this is probably too big (i.e, if the cluster size
2675 * is 1 megabyte, then group preallocation size becomes half a
2676 * gigabyte!). As a default, we will keep a two megabyte
2677 * group pralloc size for cluster sizes up to 64k, and after
2678 * that, we will force a minimum group preallocation size of
2679 * 32 clusters. This translates to 8 megs when the cluster
2680 * size is 256k, and 32 megs when the cluster size is 1 meg,
2681 * which seems reasonable as a default.
2683 sbi
->s_mb_group_prealloc
= max(MB_DEFAULT_GROUP_PREALLOC
>>
2684 sbi
->s_cluster_bits
, 32);
2686 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2687 * to the lowest multiple of s_stripe which is bigger than
2688 * the s_mb_group_prealloc as determined above. We want
2689 * the preallocation size to be an exact multiple of the
2690 * RAID stripe size so that preallocations don't fragment
2693 if (sbi
->s_stripe
> 1) {
2694 sbi
->s_mb_group_prealloc
= roundup(
2695 sbi
->s_mb_group_prealloc
, sbi
->s_stripe
);
2698 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2699 if (sbi
->s_locality_groups
== NULL
) {
2703 for_each_possible_cpu(i
) {
2704 struct ext4_locality_group
*lg
;
2705 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2706 mutex_init(&lg
->lg_mutex
);
2707 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2708 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2709 spin_lock_init(&lg
->lg_prealloc_lock
);
2712 /* init file for buddy data */
2713 ret
= ext4_mb_init_backend(sb
);
2715 goto out_free_locality_groups
;
2719 out_free_locality_groups
:
2720 free_percpu(sbi
->s_locality_groups
);
2721 sbi
->s_locality_groups
= NULL
;
2723 kfree(sbi
->s_mb_offsets
);
2724 sbi
->s_mb_offsets
= NULL
;
2725 kfree(sbi
->s_mb_maxs
);
2726 sbi
->s_mb_maxs
= NULL
;
2730 /* need to called with the ext4 group lock held */
2731 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2733 struct ext4_prealloc_space
*pa
;
2734 struct list_head
*cur
, *tmp
;
2737 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2738 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2739 list_del(&pa
->pa_group_list
);
2741 kmem_cache_free(ext4_pspace_cachep
, pa
);
2744 mb_debug(1, "mballoc: %u PAs left\n", count
);
2748 int ext4_mb_release(struct super_block
*sb
)
2750 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2752 int num_meta_group_infos
;
2753 struct ext4_group_info
*grinfo
, ***group_info
;
2754 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2755 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2757 if (sbi
->s_group_info
) {
2758 for (i
= 0; i
< ngroups
; i
++) {
2759 grinfo
= ext4_get_group_info(sb
, i
);
2761 kfree(grinfo
->bb_bitmap
);
2763 ext4_lock_group(sb
, i
);
2764 ext4_mb_cleanup_pa(grinfo
);
2765 ext4_unlock_group(sb
, i
);
2766 kmem_cache_free(cachep
, grinfo
);
2768 num_meta_group_infos
= (ngroups
+
2769 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2770 EXT4_DESC_PER_BLOCK_BITS(sb
);
2772 group_info
= rcu_dereference(sbi
->s_group_info
);
2773 for (i
= 0; i
< num_meta_group_infos
; i
++)
2774 kfree(group_info
[i
]);
2778 kfree(sbi
->s_mb_offsets
);
2779 kfree(sbi
->s_mb_maxs
);
2780 iput(sbi
->s_buddy_cache
);
2781 if (sbi
->s_mb_stats
) {
2782 ext4_msg(sb
, KERN_INFO
,
2783 "mballoc: %u blocks %u reqs (%u success)",
2784 atomic_read(&sbi
->s_bal_allocated
),
2785 atomic_read(&sbi
->s_bal_reqs
),
2786 atomic_read(&sbi
->s_bal_success
));
2787 ext4_msg(sb
, KERN_INFO
,
2788 "mballoc: %u extents scanned, %u goal hits, "
2789 "%u 2^N hits, %u breaks, %u lost",
2790 atomic_read(&sbi
->s_bal_ex_scanned
),
2791 atomic_read(&sbi
->s_bal_goals
),
2792 atomic_read(&sbi
->s_bal_2orders
),
2793 atomic_read(&sbi
->s_bal_breaks
),
2794 atomic_read(&sbi
->s_mb_lost_chunks
));
2795 ext4_msg(sb
, KERN_INFO
,
2796 "mballoc: %lu generated and it took %Lu",
2797 sbi
->s_mb_buddies_generated
,
2798 sbi
->s_mb_generation_time
);
2799 ext4_msg(sb
, KERN_INFO
,
2800 "mballoc: %u preallocated, %u discarded",
2801 atomic_read(&sbi
->s_mb_preallocated
),
2802 atomic_read(&sbi
->s_mb_discarded
));
2805 free_percpu(sbi
->s_locality_groups
);
2810 static inline int ext4_issue_discard(struct super_block
*sb
,
2811 ext4_group_t block_group
, ext4_grpblk_t cluster
, int count
,
2814 ext4_fsblk_t discard_block
;
2816 discard_block
= (EXT4_C2B(EXT4_SB(sb
), cluster
) +
2817 ext4_group_first_block_no(sb
, block_group
));
2818 count
= EXT4_C2B(EXT4_SB(sb
), count
);
2819 trace_ext4_discard_blocks(sb
,
2820 (unsigned long long) discard_block
, count
);
2822 return __blkdev_issue_discard(sb
->s_bdev
,
2823 (sector_t
)discard_block
<< (sb
->s_blocksize_bits
- 9),
2824 (sector_t
)count
<< (sb
->s_blocksize_bits
- 9),
2827 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, 0);
2830 static void ext4_free_data_in_buddy(struct super_block
*sb
,
2831 struct ext4_free_data
*entry
)
2833 struct ext4_buddy e4b
;
2834 struct ext4_group_info
*db
;
2835 int err
, count
= 0, count2
= 0;
2837 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2838 entry
->efd_count
, entry
->efd_group
, entry
);
2840 err
= ext4_mb_load_buddy(sb
, entry
->efd_group
, &e4b
);
2841 /* we expect to find existing buddy because it's pinned */
2844 spin_lock(&EXT4_SB(sb
)->s_md_lock
);
2845 EXT4_SB(sb
)->s_mb_free_pending
-= entry
->efd_count
;
2846 spin_unlock(&EXT4_SB(sb
)->s_md_lock
);
2849 /* there are blocks to put in buddy to make them really free */
2850 count
+= entry
->efd_count
;
2852 ext4_lock_group(sb
, entry
->efd_group
);
2853 /* Take it out of per group rb tree */
2854 rb_erase(&entry
->efd_node
, &(db
->bb_free_root
));
2855 mb_free_blocks(NULL
, &e4b
, entry
->efd_start_cluster
, entry
->efd_count
);
2858 * Clear the trimmed flag for the group so that the next
2859 * ext4_trim_fs can trim it.
2860 * If the volume is mounted with -o discard, online discard
2861 * is supported and the free blocks will be trimmed online.
2863 if (!test_opt(sb
, DISCARD
))
2864 EXT4_MB_GRP_CLEAR_TRIMMED(db
);
2866 if (!db
->bb_free_root
.rb_node
) {
2867 /* No more items in the per group rb tree
2868 * balance refcounts from ext4_mb_free_metadata()
2870 put_page(e4b
.bd_buddy_page
);
2871 put_page(e4b
.bd_bitmap_page
);
2873 ext4_unlock_group(sb
, entry
->efd_group
);
2874 kmem_cache_free(ext4_free_data_cachep
, entry
);
2875 ext4_mb_unload_buddy(&e4b
);
2877 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2881 * This function is called by the jbd2 layer once the commit has finished,
2882 * so we know we can free the blocks that were released with that commit.
2884 void ext4_process_freed_data(struct super_block
*sb
, tid_t commit_tid
)
2886 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2887 struct ext4_free_data
*entry
, *tmp
;
2888 struct bio
*discard_bio
= NULL
;
2889 struct list_head freed_data_list
;
2890 struct list_head
*cut_pos
= NULL
;
2893 INIT_LIST_HEAD(&freed_data_list
);
2895 spin_lock(&sbi
->s_md_lock
);
2896 list_for_each_entry(entry
, &sbi
->s_freed_data_list
, efd_list
) {
2897 if (entry
->efd_tid
!= commit_tid
)
2899 cut_pos
= &entry
->efd_list
;
2902 list_cut_position(&freed_data_list
, &sbi
->s_freed_data_list
,
2904 spin_unlock(&sbi
->s_md_lock
);
2906 if (test_opt(sb
, DISCARD
)) {
2907 list_for_each_entry(entry
, &freed_data_list
, efd_list
) {
2908 err
= ext4_issue_discard(sb
, entry
->efd_group
,
2909 entry
->efd_start_cluster
,
2912 if (err
&& err
!= -EOPNOTSUPP
) {
2913 ext4_msg(sb
, KERN_WARNING
, "discard request in"
2914 " group:%d block:%d count:%d failed"
2915 " with %d", entry
->efd_group
,
2916 entry
->efd_start_cluster
,
2917 entry
->efd_count
, err
);
2918 } else if (err
== -EOPNOTSUPP
)
2923 submit_bio_wait(discard_bio
);
2924 bio_put(discard_bio
);
2928 list_for_each_entry_safe(entry
, tmp
, &freed_data_list
, efd_list
)
2929 ext4_free_data_in_buddy(sb
, entry
);
2932 int __init
ext4_init_mballoc(void)
2934 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2935 SLAB_RECLAIM_ACCOUNT
);
2936 if (ext4_pspace_cachep
== NULL
)
2939 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2940 SLAB_RECLAIM_ACCOUNT
);
2941 if (ext4_ac_cachep
== NULL
) {
2942 kmem_cache_destroy(ext4_pspace_cachep
);
2946 ext4_free_data_cachep
= KMEM_CACHE(ext4_free_data
,
2947 SLAB_RECLAIM_ACCOUNT
);
2948 if (ext4_free_data_cachep
== NULL
) {
2949 kmem_cache_destroy(ext4_pspace_cachep
);
2950 kmem_cache_destroy(ext4_ac_cachep
);
2956 void ext4_exit_mballoc(void)
2959 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2960 * before destroying the slab cache.
2963 kmem_cache_destroy(ext4_pspace_cachep
);
2964 kmem_cache_destroy(ext4_ac_cachep
);
2965 kmem_cache_destroy(ext4_free_data_cachep
);
2966 ext4_groupinfo_destroy_slabs();
2971 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2972 * Returns 0 if success or error code
2974 static noinline_for_stack
int
2975 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2976 handle_t
*handle
, unsigned int reserv_clstrs
)
2978 struct buffer_head
*bitmap_bh
= NULL
;
2979 struct ext4_group_desc
*gdp
;
2980 struct buffer_head
*gdp_bh
;
2981 struct ext4_sb_info
*sbi
;
2982 struct super_block
*sb
;
2986 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2987 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2992 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2993 if (IS_ERR(bitmap_bh
)) {
2994 err
= PTR_ERR(bitmap_bh
);
2999 BUFFER_TRACE(bitmap_bh
, "getting write access");
3000 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
3005 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
3009 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
3010 ext4_free_group_clusters(sb
, gdp
));
3012 BUFFER_TRACE(gdp_bh
, "get_write_access");
3013 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
3017 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3019 len
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
3020 if (!ext4_data_block_valid(sbi
, block
, len
)) {
3021 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
3022 "fs metadata", block
, block
+len
);
3023 /* File system mounted not to panic on error
3024 * Fix the bitmap and return EFSCORRUPTED
3025 * We leak some of the blocks here.
3027 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3028 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
3029 ac
->ac_b_ex
.fe_len
);
3030 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3031 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3033 err
= -EFSCORRUPTED
;
3037 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3038 #ifdef AGGRESSIVE_CHECK
3041 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
3042 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
3043 bitmap_bh
->b_data
));
3047 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
3048 ac
->ac_b_ex
.fe_len
);
3049 if (ext4_has_group_desc_csum(sb
) &&
3050 (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))) {
3051 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
3052 ext4_free_group_clusters_set(sb
, gdp
,
3053 ext4_free_clusters_after_init(sb
,
3054 ac
->ac_b_ex
.fe_group
, gdp
));
3056 len
= ext4_free_group_clusters(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
3057 ext4_free_group_clusters_set(sb
, gdp
, len
);
3058 ext4_block_bitmap_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
, bitmap_bh
);
3059 ext4_group_desc_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
);
3061 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3062 percpu_counter_sub(&sbi
->s_freeclusters_counter
, ac
->ac_b_ex
.fe_len
);
3064 * Now reduce the dirty block count also. Should not go negative
3066 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
3067 /* release all the reserved blocks if non delalloc */
3068 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
3071 if (sbi
->s_log_groups_per_flex
) {
3072 ext4_group_t flex_group
= ext4_flex_group(sbi
,
3073 ac
->ac_b_ex
.fe_group
);
3074 atomic64_sub(ac
->ac_b_ex
.fe_len
,
3075 &sbi_array_rcu_deref(sbi
, s_flex_groups
,
3076 flex_group
)->free_clusters
);
3079 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3082 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
3090 * here we normalize request for locality group
3091 * Group request are normalized to s_mb_group_prealloc, which goes to
3092 * s_strip if we set the same via mount option.
3093 * s_mb_group_prealloc can be configured via
3094 * /sys/fs/ext4/<partition>/mb_group_prealloc
3096 * XXX: should we try to preallocate more than the group has now?
3098 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
3100 struct super_block
*sb
= ac
->ac_sb
;
3101 struct ext4_locality_group
*lg
= ac
->ac_lg
;
3104 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
3105 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3106 current
->pid
, ac
->ac_g_ex
.fe_len
);
3110 * Normalization means making request better in terms of
3111 * size and alignment
3113 static noinline_for_stack
void
3114 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
3115 struct ext4_allocation_request
*ar
)
3117 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3120 loff_t size
, start_off
;
3121 loff_t orig_size __maybe_unused
;
3123 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3124 struct ext4_prealloc_space
*pa
;
3126 /* do normalize only data requests, metadata requests
3127 do not need preallocation */
3128 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3131 /* sometime caller may want exact blocks */
3132 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3135 /* caller may indicate that preallocation isn't
3136 * required (it's a tail, for example) */
3137 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3140 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3141 ext4_mb_normalize_group_request(ac
);
3145 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3147 /* first, let's learn actual file size
3148 * given current request is allocated */
3149 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
3150 size
= size
<< bsbits
;
3151 if (size
< i_size_read(ac
->ac_inode
))
3152 size
= i_size_read(ac
->ac_inode
);
3155 /* max size of free chunks */
3158 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3159 (req <= (size) || max <= (chunk_size))
3161 /* first, try to predict filesize */
3162 /* XXX: should this table be tunable? */
3164 if (size
<= 16 * 1024) {
3166 } else if (size
<= 32 * 1024) {
3168 } else if (size
<= 64 * 1024) {
3170 } else if (size
<= 128 * 1024) {
3172 } else if (size
<= 256 * 1024) {
3174 } else if (size
<= 512 * 1024) {
3176 } else if (size
<= 1024 * 1024) {
3178 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3179 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3180 (21 - bsbits
)) << 21;
3181 size
= 2 * 1024 * 1024;
3182 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3183 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3184 (22 - bsbits
)) << 22;
3185 size
= 4 * 1024 * 1024;
3186 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3187 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3188 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3189 (23 - bsbits
)) << 23;
3190 size
= 8 * 1024 * 1024;
3192 start_off
= (loff_t
) ac
->ac_o_ex
.fe_logical
<< bsbits
;
3193 size
= (loff_t
) EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3194 ac
->ac_o_ex
.fe_len
) << bsbits
;
3196 size
= size
>> bsbits
;
3197 start
= start_off
>> bsbits
;
3199 /* don't cover already allocated blocks in selected range */
3200 if (ar
->pleft
&& start
<= ar
->lleft
) {
3201 size
-= ar
->lleft
+ 1 - start
;
3202 start
= ar
->lleft
+ 1;
3204 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3205 size
-= start
+ size
- ar
->lright
;
3208 * Trim allocation request for filesystems with artificially small
3211 if (size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
))
3212 size
= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
);
3216 /* check we don't cross already preallocated blocks */
3218 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3223 spin_lock(&pa
->pa_lock
);
3224 if (pa
->pa_deleted
) {
3225 spin_unlock(&pa
->pa_lock
);
3229 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3232 /* PA must not overlap original request */
3233 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3234 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3236 /* skip PAs this normalized request doesn't overlap with */
3237 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
3238 spin_unlock(&pa
->pa_lock
);
3241 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3243 /* adjust start or end to be adjacent to this pa */
3244 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3245 BUG_ON(pa_end
< start
);
3247 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3248 BUG_ON(pa
->pa_lstart
> end
);
3249 end
= pa
->pa_lstart
;
3251 spin_unlock(&pa
->pa_lock
);
3256 /* XXX: extra loop to check we really don't overlap preallocations */
3258 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3261 spin_lock(&pa
->pa_lock
);
3262 if (pa
->pa_deleted
== 0) {
3263 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3265 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3267 spin_unlock(&pa
->pa_lock
);
3271 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3272 start
> ac
->ac_o_ex
.fe_logical
) {
3273 ext4_msg(ac
->ac_sb
, KERN_ERR
,
3274 "start %lu, size %lu, fe_logical %lu",
3275 (unsigned long) start
, (unsigned long) size
,
3276 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3279 BUG_ON(size
<= 0 || size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3281 /* now prepare goal request */
3283 /* XXX: is it better to align blocks WRT to logical
3284 * placement or satisfy big request as is */
3285 ac
->ac_g_ex
.fe_logical
= start
;
3286 ac
->ac_g_ex
.fe_len
= EXT4_NUM_B2C(sbi
, size
);
3288 /* define goal start in order to merge */
3289 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3290 /* merge to the right */
3291 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3292 &ac
->ac_f_ex
.fe_group
,
3293 &ac
->ac_f_ex
.fe_start
);
3294 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3296 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3297 /* merge to the left */
3298 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3299 &ac
->ac_f_ex
.fe_group
,
3300 &ac
->ac_f_ex
.fe_start
);
3301 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3304 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3305 (unsigned) orig_size
, (unsigned) start
);
3308 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3310 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3312 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3313 atomic_inc(&sbi
->s_bal_reqs
);
3314 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3315 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3316 atomic_inc(&sbi
->s_bal_success
);
3317 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3318 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3319 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3320 atomic_inc(&sbi
->s_bal_goals
);
3321 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3322 atomic_inc(&sbi
->s_bal_breaks
);
3325 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3326 trace_ext4_mballoc_alloc(ac
);
3328 trace_ext4_mballoc_prealloc(ac
);
3332 * Called on failure; free up any blocks from the inode PA for this
3333 * context. We don't need this for MB_GROUP_PA because we only change
3334 * pa_free in ext4_mb_release_context(), but on failure, we've already
3335 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3337 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3339 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3340 struct ext4_buddy e4b
;
3344 if (ac
->ac_f_ex
.fe_len
== 0)
3346 err
= ext4_mb_load_buddy(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
, &e4b
);
3349 * This should never happen since we pin the
3350 * pages in the ext4_allocation_context so
3351 * ext4_mb_load_buddy() should never fail.
3353 WARN(1, "mb_load_buddy failed (%d)", err
);
3356 ext4_lock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3357 mb_free_blocks(ac
->ac_inode
, &e4b
, ac
->ac_f_ex
.fe_start
,
3358 ac
->ac_f_ex
.fe_len
);
3359 ext4_unlock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3360 ext4_mb_unload_buddy(&e4b
);
3363 if (pa
->pa_type
== MB_INODE_PA
)
3364 pa
->pa_free
+= ac
->ac_b_ex
.fe_len
;
3368 * use blocks preallocated to inode
3370 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3371 struct ext4_prealloc_space
*pa
)
3373 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3378 /* found preallocated blocks, use them */
3379 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3380 end
= min(pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
),
3381 start
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
));
3382 len
= EXT4_NUM_B2C(sbi
, end
- start
);
3383 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3384 &ac
->ac_b_ex
.fe_start
);
3385 ac
->ac_b_ex
.fe_len
= len
;
3386 ac
->ac_status
= AC_STATUS_FOUND
;
3389 BUG_ON(start
< pa
->pa_pstart
);
3390 BUG_ON(end
> pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
));
3391 BUG_ON(pa
->pa_free
< len
);
3394 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3398 * use blocks preallocated to locality group
3400 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3401 struct ext4_prealloc_space
*pa
)
3403 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3405 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3406 &ac
->ac_b_ex
.fe_group
,
3407 &ac
->ac_b_ex
.fe_start
);
3408 ac
->ac_b_ex
.fe_len
= len
;
3409 ac
->ac_status
= AC_STATUS_FOUND
;
3412 /* we don't correct pa_pstart or pa_plen here to avoid
3413 * possible race when the group is being loaded concurrently
3414 * instead we correct pa later, after blocks are marked
3415 * in on-disk bitmap -- see ext4_mb_release_context()
3416 * Other CPUs are prevented from allocating from this pa by lg_mutex
3418 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3422 * Return the prealloc space that have minimal distance
3423 * from the goal block. @cpa is the prealloc
3424 * space that is having currently known minimal distance
3425 * from the goal block.
3427 static struct ext4_prealloc_space
*
3428 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3429 struct ext4_prealloc_space
*pa
,
3430 struct ext4_prealloc_space
*cpa
)
3432 ext4_fsblk_t cur_distance
, new_distance
;
3435 atomic_inc(&pa
->pa_count
);
3438 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3439 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3441 if (cur_distance
<= new_distance
)
3444 /* drop the previous reference */
3445 atomic_dec(&cpa
->pa_count
);
3446 atomic_inc(&pa
->pa_count
);
3451 * search goal blocks in preallocated space
3453 static noinline_for_stack
int
3454 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3456 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3458 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3459 struct ext4_locality_group
*lg
;
3460 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3461 ext4_fsblk_t goal_block
;
3463 /* only data can be preallocated */
3464 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3467 /* first, try per-file preallocation */
3469 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3471 /* all fields in this condition don't change,
3472 * so we can skip locking for them */
3473 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3474 ac
->ac_o_ex
.fe_logical
>= (pa
->pa_lstart
+
3475 EXT4_C2B(sbi
, pa
->pa_len
)))
3478 /* non-extent files can't have physical blocks past 2^32 */
3479 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3480 (pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
) >
3481 EXT4_MAX_BLOCK_FILE_PHYS
))
3484 /* found preallocated blocks, use them */
3485 spin_lock(&pa
->pa_lock
);
3486 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3487 atomic_inc(&pa
->pa_count
);
3488 ext4_mb_use_inode_pa(ac
, pa
);
3489 spin_unlock(&pa
->pa_lock
);
3490 ac
->ac_criteria
= 10;
3494 spin_unlock(&pa
->pa_lock
);
3498 /* can we use group allocation? */
3499 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3502 /* inode may have no locality group for some reason */
3506 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3507 if (order
> PREALLOC_TB_SIZE
- 1)
3508 /* The max size of hash table is PREALLOC_TB_SIZE */
3509 order
= PREALLOC_TB_SIZE
- 1;
3511 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3513 * search for the prealloc space that is having
3514 * minimal distance from the goal block.
3516 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3518 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3520 spin_lock(&pa
->pa_lock
);
3521 if (pa
->pa_deleted
== 0 &&
3522 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3524 cpa
= ext4_mb_check_group_pa(goal_block
,
3527 spin_unlock(&pa
->pa_lock
);
3532 ext4_mb_use_group_pa(ac
, cpa
);
3533 ac
->ac_criteria
= 20;
3540 * the function goes through all block freed in the group
3541 * but not yet committed and marks them used in in-core bitmap.
3542 * buddy must be generated from this bitmap
3543 * Need to be called with the ext4 group lock held
3545 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3549 struct ext4_group_info
*grp
;
3550 struct ext4_free_data
*entry
;
3552 grp
= ext4_get_group_info(sb
, group
);
3553 n
= rb_first(&(grp
->bb_free_root
));
3556 entry
= rb_entry(n
, struct ext4_free_data
, efd_node
);
3557 ext4_set_bits(bitmap
, entry
->efd_start_cluster
, entry
->efd_count
);
3564 * the function goes through all preallocation in this group and marks them
3565 * used in in-core bitmap. buddy must be generated from this bitmap
3566 * Need to be called with ext4 group lock held
3568 static noinline_for_stack
3569 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3572 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3573 struct ext4_prealloc_space
*pa
;
3574 struct list_head
*cur
;
3575 ext4_group_t groupnr
;
3576 ext4_grpblk_t start
;
3577 int preallocated
= 0;
3580 /* all form of preallocation discards first load group,
3581 * so the only competing code is preallocation use.
3582 * we don't need any locking here
3583 * notice we do NOT ignore preallocations with pa_deleted
3584 * otherwise we could leave used blocks available for
3585 * allocation in buddy when concurrent ext4_mb_put_pa()
3586 * is dropping preallocation
3588 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3589 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3590 spin_lock(&pa
->pa_lock
);
3591 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3594 spin_unlock(&pa
->pa_lock
);
3595 if (unlikely(len
== 0))
3597 BUG_ON(groupnr
!= group
);
3598 ext4_set_bits(bitmap
, start
, len
);
3599 preallocated
+= len
;
3601 mb_debug(1, "preallocated %u for group %u\n", preallocated
, group
);
3604 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3606 struct ext4_prealloc_space
*pa
;
3607 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3609 BUG_ON(atomic_read(&pa
->pa_count
));
3610 BUG_ON(pa
->pa_deleted
== 0);
3611 kmem_cache_free(ext4_pspace_cachep
, pa
);
3615 * drops a reference to preallocated space descriptor
3616 * if this was the last reference and the space is consumed
3618 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3619 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3622 ext4_fsblk_t grp_blk
;
3624 /* in this short window concurrent discard can set pa_deleted */
3625 spin_lock(&pa
->pa_lock
);
3626 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0) {
3627 spin_unlock(&pa
->pa_lock
);
3631 if (pa
->pa_deleted
== 1) {
3632 spin_unlock(&pa
->pa_lock
);
3637 spin_unlock(&pa
->pa_lock
);
3639 grp_blk
= pa
->pa_pstart
;
3641 * If doing group-based preallocation, pa_pstart may be in the
3642 * next group when pa is used up
3644 if (pa
->pa_type
== MB_GROUP_PA
)
3647 grp
= ext4_get_group_number(sb
, grp_blk
);
3652 * P1 (buddy init) P2 (regular allocation)
3653 * find block B in PA
3654 * copy on-disk bitmap to buddy
3655 * mark B in on-disk bitmap
3656 * drop PA from group
3657 * mark all PAs in buddy
3659 * thus, P1 initializes buddy with B available. to prevent this
3660 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3663 ext4_lock_group(sb
, grp
);
3664 list_del(&pa
->pa_group_list
);
3665 ext4_unlock_group(sb
, grp
);
3667 spin_lock(pa
->pa_obj_lock
);
3668 list_del_rcu(&pa
->pa_inode_list
);
3669 spin_unlock(pa
->pa_obj_lock
);
3671 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3675 * creates new preallocated space for given inode
3677 static noinline_for_stack
int
3678 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3680 struct super_block
*sb
= ac
->ac_sb
;
3681 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3682 struct ext4_prealloc_space
*pa
;
3683 struct ext4_group_info
*grp
;
3684 struct ext4_inode_info
*ei
;
3686 /* preallocate only when found space is larger then requested */
3687 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3688 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3689 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3691 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3695 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3701 /* we can't allocate as much as normalizer wants.
3702 * so, found space must get proper lstart
3703 * to cover original request */
3704 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3705 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3707 /* we're limited by original request in that
3708 * logical block must be covered any way
3709 * winl is window we can move our chunk within */
3710 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3712 /* also, we should cover whole original request */
3713 wins
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
);
3715 /* the smallest one defines real window */
3716 win
= min(winl
, wins
);
3718 offs
= ac
->ac_o_ex
.fe_logical
%
3719 EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
3720 if (offs
&& offs
< win
)
3723 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
-
3724 EXT4_NUM_B2C(sbi
, win
);
3725 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3726 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3729 /* preallocation can change ac_b_ex, thus we store actually
3730 * allocated blocks for history */
3731 ac
->ac_f_ex
= ac
->ac_b_ex
;
3733 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3734 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3735 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3736 pa
->pa_free
= pa
->pa_len
;
3737 atomic_set(&pa
->pa_count
, 1);
3738 spin_lock_init(&pa
->pa_lock
);
3739 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3740 INIT_LIST_HEAD(&pa
->pa_group_list
);
3742 pa
->pa_type
= MB_INODE_PA
;
3744 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3745 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3746 trace_ext4_mb_new_inode_pa(ac
, pa
);
3748 ext4_mb_use_inode_pa(ac
, pa
);
3749 atomic_add(pa
->pa_free
, &sbi
->s_mb_preallocated
);
3751 ei
= EXT4_I(ac
->ac_inode
);
3752 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3754 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3755 pa
->pa_inode
= ac
->ac_inode
;
3757 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3758 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3759 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3761 spin_lock(pa
->pa_obj_lock
);
3762 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3763 spin_unlock(pa
->pa_obj_lock
);
3769 * creates new preallocated space for locality group inodes belongs to
3771 static noinline_for_stack
int
3772 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3774 struct super_block
*sb
= ac
->ac_sb
;
3775 struct ext4_locality_group
*lg
;
3776 struct ext4_prealloc_space
*pa
;
3777 struct ext4_group_info
*grp
;
3779 /* preallocate only when found space is larger then requested */
3780 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3781 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3782 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3784 BUG_ON(ext4_pspace_cachep
== NULL
);
3785 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3789 /* preallocation can change ac_b_ex, thus we store actually
3790 * allocated blocks for history */
3791 ac
->ac_f_ex
= ac
->ac_b_ex
;
3793 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3794 pa
->pa_lstart
= pa
->pa_pstart
;
3795 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3796 pa
->pa_free
= pa
->pa_len
;
3797 atomic_set(&pa
->pa_count
, 1);
3798 spin_lock_init(&pa
->pa_lock
);
3799 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3800 INIT_LIST_HEAD(&pa
->pa_group_list
);
3802 pa
->pa_type
= MB_GROUP_PA
;
3804 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3805 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3806 trace_ext4_mb_new_group_pa(ac
, pa
);
3808 ext4_mb_use_group_pa(ac
, pa
);
3809 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3811 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3815 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3816 pa
->pa_inode
= NULL
;
3818 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3819 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3820 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3823 * We will later add the new pa to the right bucket
3824 * after updating the pa_free in ext4_mb_release_context
3829 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3833 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3834 err
= ext4_mb_new_group_pa(ac
);
3836 err
= ext4_mb_new_inode_pa(ac
);
3841 * finds all unused blocks in on-disk bitmap, frees them in
3842 * in-core bitmap and buddy.
3843 * @pa must be unlinked from inode and group lists, so that
3844 * nobody else can find/use it.
3845 * the caller MUST hold group/inode locks.
3846 * TODO: optimize the case when there are no in-core structures yet
3848 static noinline_for_stack
int
3849 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3850 struct ext4_prealloc_space
*pa
)
3852 struct super_block
*sb
= e4b
->bd_sb
;
3853 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3858 unsigned long long grp_blk_start
;
3862 BUG_ON(pa
->pa_deleted
== 0);
3863 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3864 grp_blk_start
= pa
->pa_pstart
- EXT4_C2B(sbi
, bit
);
3865 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3866 end
= bit
+ pa
->pa_len
;
3869 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3872 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3873 mb_debug(1, " free preallocated %u/%u in group %u\n",
3874 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3875 (unsigned) next
- bit
, (unsigned) group
);
3878 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3879 trace_ext4_mb_release_inode_pa(pa
, (grp_blk_start
+
3880 EXT4_C2B(sbi
, bit
)),
3882 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3885 if (free
!= pa
->pa_free
) {
3886 ext4_msg(e4b
->bd_sb
, KERN_CRIT
,
3887 "pa %p: logic %lu, phys. %lu, len %lu",
3888 pa
, (unsigned long) pa
->pa_lstart
,
3889 (unsigned long) pa
->pa_pstart
,
3890 (unsigned long) pa
->pa_len
);
3891 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3894 * pa is already deleted so we use the value obtained
3895 * from the bitmap and continue.
3898 atomic_add(free
, &sbi
->s_mb_discarded
);
3903 static noinline_for_stack
int
3904 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3905 struct ext4_prealloc_space
*pa
)
3907 struct super_block
*sb
= e4b
->bd_sb
;
3911 trace_ext4_mb_release_group_pa(sb
, pa
);
3912 BUG_ON(pa
->pa_deleted
== 0);
3913 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3914 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3915 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3916 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3917 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3923 * releases all preallocations in given group
3925 * first, we need to decide discard policy:
3926 * - when do we discard
3928 * - how many do we discard
3929 * 1) how many requested
3931 static noinline_for_stack
int
3932 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3933 ext4_group_t group
, int needed
)
3935 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3936 struct buffer_head
*bitmap_bh
= NULL
;
3937 struct ext4_prealloc_space
*pa
, *tmp
;
3938 struct list_head list
;
3939 struct ext4_buddy e4b
;
3944 mb_debug(1, "discard preallocation for group %u\n", group
);
3946 if (list_empty(&grp
->bb_prealloc_list
))
3949 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3950 if (IS_ERR(bitmap_bh
)) {
3951 err
= PTR_ERR(bitmap_bh
);
3952 ext4_error(sb
, "Error %d reading block bitmap for %u",
3957 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3959 ext4_warning(sb
, "Error %d loading buddy information for %u",
3966 needed
= EXT4_CLUSTERS_PER_GROUP(sb
) + 1;
3968 INIT_LIST_HEAD(&list
);
3970 ext4_lock_group(sb
, group
);
3971 list_for_each_entry_safe(pa
, tmp
,
3972 &grp
->bb_prealloc_list
, pa_group_list
) {
3973 spin_lock(&pa
->pa_lock
);
3974 if (atomic_read(&pa
->pa_count
)) {
3975 spin_unlock(&pa
->pa_lock
);
3979 if (pa
->pa_deleted
) {
3980 spin_unlock(&pa
->pa_lock
);
3984 /* seems this one can be freed ... */
3987 /* we can trust pa_free ... */
3988 free
+= pa
->pa_free
;
3990 spin_unlock(&pa
->pa_lock
);
3992 list_del(&pa
->pa_group_list
);
3993 list_add(&pa
->u
.pa_tmp_list
, &list
);
3996 /* if we still need more blocks and some PAs were used, try again */
3997 if (free
< needed
&& busy
) {
3999 ext4_unlock_group(sb
, group
);
4004 /* found anything to free? */
4005 if (list_empty(&list
)) {
4010 /* now free all selected PAs */
4011 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
4013 /* remove from object (inode or locality group) */
4014 spin_lock(pa
->pa_obj_lock
);
4015 list_del_rcu(&pa
->pa_inode_list
);
4016 spin_unlock(pa
->pa_obj_lock
);
4018 if (pa
->pa_type
== MB_GROUP_PA
)
4019 ext4_mb_release_group_pa(&e4b
, pa
);
4021 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
4023 list_del(&pa
->u
.pa_tmp_list
);
4024 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4028 ext4_unlock_group(sb
, group
);
4029 ext4_mb_unload_buddy(&e4b
);
4035 * releases all non-used preallocated blocks for given inode
4037 * It's important to discard preallocations under i_data_sem
4038 * We don't want another block to be served from the prealloc
4039 * space when we are discarding the inode prealloc space.
4041 * FIXME!! Make sure it is valid at all the call sites
4043 void ext4_discard_preallocations(struct inode
*inode
)
4045 struct ext4_inode_info
*ei
= EXT4_I(inode
);
4046 struct super_block
*sb
= inode
->i_sb
;
4047 struct buffer_head
*bitmap_bh
= NULL
;
4048 struct ext4_prealloc_space
*pa
, *tmp
;
4049 ext4_group_t group
= 0;
4050 struct list_head list
;
4051 struct ext4_buddy e4b
;
4054 if (!S_ISREG(inode
->i_mode
)) {
4055 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4059 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
4060 trace_ext4_discard_preallocations(inode
);
4062 INIT_LIST_HEAD(&list
);
4065 /* first, collect all pa's in the inode */
4066 spin_lock(&ei
->i_prealloc_lock
);
4067 while (!list_empty(&ei
->i_prealloc_list
)) {
4068 pa
= list_entry(ei
->i_prealloc_list
.next
,
4069 struct ext4_prealloc_space
, pa_inode_list
);
4070 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
4071 spin_lock(&pa
->pa_lock
);
4072 if (atomic_read(&pa
->pa_count
)) {
4073 /* this shouldn't happen often - nobody should
4074 * use preallocation while we're discarding it */
4075 spin_unlock(&pa
->pa_lock
);
4076 spin_unlock(&ei
->i_prealloc_lock
);
4077 ext4_msg(sb
, KERN_ERR
,
4078 "uh-oh! used pa while discarding");
4080 schedule_timeout_uninterruptible(HZ
);
4084 if (pa
->pa_deleted
== 0) {
4086 spin_unlock(&pa
->pa_lock
);
4087 list_del_rcu(&pa
->pa_inode_list
);
4088 list_add(&pa
->u
.pa_tmp_list
, &list
);
4092 /* someone is deleting pa right now */
4093 spin_unlock(&pa
->pa_lock
);
4094 spin_unlock(&ei
->i_prealloc_lock
);
4096 /* we have to wait here because pa_deleted
4097 * doesn't mean pa is already unlinked from
4098 * the list. as we might be called from
4099 * ->clear_inode() the inode will get freed
4100 * and concurrent thread which is unlinking
4101 * pa from inode's list may access already
4102 * freed memory, bad-bad-bad */
4104 /* XXX: if this happens too often, we can
4105 * add a flag to force wait only in case
4106 * of ->clear_inode(), but not in case of
4107 * regular truncate */
4108 schedule_timeout_uninterruptible(HZ
);
4111 spin_unlock(&ei
->i_prealloc_lock
);
4113 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
4114 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
4115 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4117 err
= ext4_mb_load_buddy_gfp(sb
, group
, &e4b
,
4118 GFP_NOFS
|__GFP_NOFAIL
);
4120 ext4_error(sb
, "Error %d loading buddy information for %u",
4125 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
4126 if (IS_ERR(bitmap_bh
)) {
4127 err
= PTR_ERR(bitmap_bh
);
4128 ext4_error(sb
, "Error %d reading block bitmap for %u",
4130 ext4_mb_unload_buddy(&e4b
);
4134 ext4_lock_group(sb
, group
);
4135 list_del(&pa
->pa_group_list
);
4136 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
4137 ext4_unlock_group(sb
, group
);
4139 ext4_mb_unload_buddy(&e4b
);
4142 list_del(&pa
->u
.pa_tmp_list
);
4143 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4147 #ifdef CONFIG_EXT4_DEBUG
4148 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4150 struct super_block
*sb
= ac
->ac_sb
;
4151 ext4_group_t ngroups
, i
;
4153 if (!ext4_mballoc_debug
||
4154 (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
))
4157 ext4_msg(ac
->ac_sb
, KERN_ERR
, "Can't allocate:"
4158 " Allocation context details:");
4159 ext4_msg(ac
->ac_sb
, KERN_ERR
, "status %d flags %d",
4160 ac
->ac_status
, ac
->ac_flags
);
4161 ext4_msg(ac
->ac_sb
, KERN_ERR
, "orig %lu/%lu/%lu@%lu, "
4162 "goal %lu/%lu/%lu@%lu, "
4163 "best %lu/%lu/%lu@%lu cr %d",
4164 (unsigned long)ac
->ac_o_ex
.fe_group
,
4165 (unsigned long)ac
->ac_o_ex
.fe_start
,
4166 (unsigned long)ac
->ac_o_ex
.fe_len
,
4167 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4168 (unsigned long)ac
->ac_g_ex
.fe_group
,
4169 (unsigned long)ac
->ac_g_ex
.fe_start
,
4170 (unsigned long)ac
->ac_g_ex
.fe_len
,
4171 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4172 (unsigned long)ac
->ac_b_ex
.fe_group
,
4173 (unsigned long)ac
->ac_b_ex
.fe_start
,
4174 (unsigned long)ac
->ac_b_ex
.fe_len
,
4175 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4176 (int)ac
->ac_criteria
);
4177 ext4_msg(ac
->ac_sb
, KERN_ERR
, "%d found", ac
->ac_found
);
4178 ext4_msg(ac
->ac_sb
, KERN_ERR
, "groups: ");
4179 ngroups
= ext4_get_groups_count(sb
);
4180 for (i
= 0; i
< ngroups
; i
++) {
4181 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4182 struct ext4_prealloc_space
*pa
;
4183 ext4_grpblk_t start
;
4184 struct list_head
*cur
;
4185 ext4_lock_group(sb
, i
);
4186 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4187 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4189 spin_lock(&pa
->pa_lock
);
4190 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4192 spin_unlock(&pa
->pa_lock
);
4193 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
4196 ext4_unlock_group(sb
, i
);
4198 if (grp
->bb_free
== 0)
4200 printk(KERN_ERR
"%u: %d/%d \n",
4201 i
, grp
->bb_free
, grp
->bb_fragments
);
4203 printk(KERN_ERR
"\n");
4206 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4213 * We use locality group preallocation for small size file. The size of the
4214 * file is determined by the current size or the resulting size after
4215 * allocation which ever is larger
4217 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4219 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4221 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4222 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4225 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4228 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4231 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
4232 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
4235 if ((size
== isize
) &&
4236 !ext4_fs_is_busy(sbi
) &&
4237 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
4238 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
4242 if (sbi
->s_mb_group_prealloc
<= 0) {
4243 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4247 /* don't use group allocation for large files */
4248 size
= max(size
, isize
);
4249 if (size
> sbi
->s_mb_stream_request
) {
4250 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4254 BUG_ON(ac
->ac_lg
!= NULL
);
4256 * locality group prealloc space are per cpu. The reason for having
4257 * per cpu locality group is to reduce the contention between block
4258 * request from multiple CPUs.
4260 ac
->ac_lg
= raw_cpu_ptr(sbi
->s_locality_groups
);
4262 /* we're going to use group allocation */
4263 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4265 /* serialize all allocations in the group */
4266 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4269 static noinline_for_stack
int
4270 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4271 struct ext4_allocation_request
*ar
)
4273 struct super_block
*sb
= ar
->inode
->i_sb
;
4274 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4275 struct ext4_super_block
*es
= sbi
->s_es
;
4279 ext4_grpblk_t block
;
4281 /* we can't allocate > group size */
4284 /* just a dirty hack to filter too big requests */
4285 if (len
>= EXT4_CLUSTERS_PER_GROUP(sb
))
4286 len
= EXT4_CLUSTERS_PER_GROUP(sb
);
4288 /* start searching from the goal */
4290 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4291 goal
>= ext4_blocks_count(es
))
4292 goal
= le32_to_cpu(es
->s_first_data_block
);
4293 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4295 /* set up allocation goals */
4296 ac
->ac_b_ex
.fe_logical
= EXT4_LBLK_CMASK(sbi
, ar
->logical
);
4297 ac
->ac_status
= AC_STATUS_CONTINUE
;
4299 ac
->ac_inode
= ar
->inode
;
4300 ac
->ac_o_ex
.fe_logical
= ac
->ac_b_ex
.fe_logical
;
4301 ac
->ac_o_ex
.fe_group
= group
;
4302 ac
->ac_o_ex
.fe_start
= block
;
4303 ac
->ac_o_ex
.fe_len
= len
;
4304 ac
->ac_g_ex
= ac
->ac_o_ex
;
4305 ac
->ac_flags
= ar
->flags
;
4307 /* we have to define context: we'll we work with a file or
4308 * locality group. this is a policy, actually */
4309 ext4_mb_group_or_file(ac
);
4311 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4312 "left: %u/%u, right %u/%u to %swritable\n",
4313 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4314 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4315 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4316 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4317 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4322 static noinline_for_stack
void
4323 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4324 struct ext4_locality_group
*lg
,
4325 int order
, int total_entries
)
4327 ext4_group_t group
= 0;
4328 struct ext4_buddy e4b
;
4329 struct list_head discard_list
;
4330 struct ext4_prealloc_space
*pa
, *tmp
;
4332 mb_debug(1, "discard locality group preallocation\n");
4334 INIT_LIST_HEAD(&discard_list
);
4336 spin_lock(&lg
->lg_prealloc_lock
);
4337 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4339 spin_lock(&pa
->pa_lock
);
4340 if (atomic_read(&pa
->pa_count
)) {
4342 * This is the pa that we just used
4343 * for block allocation. So don't
4346 spin_unlock(&pa
->pa_lock
);
4349 if (pa
->pa_deleted
) {
4350 spin_unlock(&pa
->pa_lock
);
4353 /* only lg prealloc space */
4354 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4356 /* seems this one can be freed ... */
4358 spin_unlock(&pa
->pa_lock
);
4360 list_del_rcu(&pa
->pa_inode_list
);
4361 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4364 if (total_entries
<= 5) {
4366 * we want to keep only 5 entries
4367 * allowing it to grow to 8. This
4368 * mak sure we don't call discard
4369 * soon for this list.
4374 spin_unlock(&lg
->lg_prealloc_lock
);
4376 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4379 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4380 err
= ext4_mb_load_buddy_gfp(sb
, group
, &e4b
,
4381 GFP_NOFS
|__GFP_NOFAIL
);
4383 ext4_error(sb
, "Error %d loading buddy information for %u",
4387 ext4_lock_group(sb
, group
);
4388 list_del(&pa
->pa_group_list
);
4389 ext4_mb_release_group_pa(&e4b
, pa
);
4390 ext4_unlock_group(sb
, group
);
4392 ext4_mb_unload_buddy(&e4b
);
4393 list_del(&pa
->u
.pa_tmp_list
);
4394 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4399 * We have incremented pa_count. So it cannot be freed at this
4400 * point. Also we hold lg_mutex. So no parallel allocation is
4401 * possible from this lg. That means pa_free cannot be updated.
4403 * A parallel ext4_mb_discard_group_preallocations is possible.
4404 * which can cause the lg_prealloc_list to be updated.
4407 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4409 int order
, added
= 0, lg_prealloc_count
= 1;
4410 struct super_block
*sb
= ac
->ac_sb
;
4411 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4412 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4414 order
= fls(pa
->pa_free
) - 1;
4415 if (order
> PREALLOC_TB_SIZE
- 1)
4416 /* The max size of hash table is PREALLOC_TB_SIZE */
4417 order
= PREALLOC_TB_SIZE
- 1;
4418 /* Add the prealloc space to lg */
4419 spin_lock(&lg
->lg_prealloc_lock
);
4420 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4422 spin_lock(&tmp_pa
->pa_lock
);
4423 if (tmp_pa
->pa_deleted
) {
4424 spin_unlock(&tmp_pa
->pa_lock
);
4427 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4428 /* Add to the tail of the previous entry */
4429 list_add_tail_rcu(&pa
->pa_inode_list
,
4430 &tmp_pa
->pa_inode_list
);
4433 * we want to count the total
4434 * number of entries in the list
4437 spin_unlock(&tmp_pa
->pa_lock
);
4438 lg_prealloc_count
++;
4441 list_add_tail_rcu(&pa
->pa_inode_list
,
4442 &lg
->lg_prealloc_list
[order
]);
4443 spin_unlock(&lg
->lg_prealloc_lock
);
4445 /* Now trim the list to be not more than 8 elements */
4446 if (lg_prealloc_count
> 8) {
4447 ext4_mb_discard_lg_preallocations(sb
, lg
,
4448 order
, lg_prealloc_count
);
4455 * release all resource we used in allocation
4457 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4459 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4460 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4462 if (pa
->pa_type
== MB_GROUP_PA
) {
4463 /* see comment in ext4_mb_use_group_pa() */
4464 spin_lock(&pa
->pa_lock
);
4465 pa
->pa_pstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4466 pa
->pa_lstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4467 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4468 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4469 spin_unlock(&pa
->pa_lock
);
4474 * We want to add the pa to the right bucket.
4475 * Remove it from the list and while adding
4476 * make sure the list to which we are adding
4479 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4480 spin_lock(pa
->pa_obj_lock
);
4481 list_del_rcu(&pa
->pa_inode_list
);
4482 spin_unlock(pa
->pa_obj_lock
);
4483 ext4_mb_add_n_trim(ac
);
4485 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4487 if (ac
->ac_bitmap_page
)
4488 put_page(ac
->ac_bitmap_page
);
4489 if (ac
->ac_buddy_page
)
4490 put_page(ac
->ac_buddy_page
);
4491 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4492 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4493 ext4_mb_collect_stats(ac
);
4497 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4499 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4503 trace_ext4_mb_discard_preallocations(sb
, needed
);
4504 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4505 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4514 * Main entry point into mballoc to allocate blocks
4515 * it tries to use preallocation first, then falls back
4516 * to usual allocation
4518 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4519 struct ext4_allocation_request
*ar
, int *errp
)
4522 struct ext4_allocation_context
*ac
= NULL
;
4523 struct ext4_sb_info
*sbi
;
4524 struct super_block
*sb
;
4525 ext4_fsblk_t block
= 0;
4526 unsigned int inquota
= 0;
4527 unsigned int reserv_clstrs
= 0;
4530 sb
= ar
->inode
->i_sb
;
4533 trace_ext4_request_blocks(ar
);
4535 /* Allow to use superuser reservation for quota file */
4536 if (ext4_is_quota_file(ar
->inode
))
4537 ar
->flags
|= EXT4_MB_USE_ROOT_BLOCKS
;
4539 if ((ar
->flags
& EXT4_MB_DELALLOC_RESERVED
) == 0) {
4540 /* Without delayed allocation we need to verify
4541 * there is enough free blocks to do block allocation
4542 * and verify allocation doesn't exceed the quota limits.
4545 ext4_claim_free_clusters(sbi
, ar
->len
, ar
->flags
)) {
4547 /* let others to free the space */
4549 ar
->len
= ar
->len
>> 1;
4555 reserv_clstrs
= ar
->len
;
4556 if (ar
->flags
& EXT4_MB_USE_ROOT_BLOCKS
) {
4557 dquot_alloc_block_nofail(ar
->inode
,
4558 EXT4_C2B(sbi
, ar
->len
));
4561 dquot_alloc_block(ar
->inode
,
4562 EXT4_C2B(sbi
, ar
->len
))) {
4564 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4575 ac
= kmem_cache_zalloc(ext4_ac_cachep
, GFP_NOFS
);
4582 *errp
= ext4_mb_initialize_context(ac
, ar
);
4588 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4589 if (!ext4_mb_use_preallocated(ac
)) {
4590 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4591 ext4_mb_normalize_request(ac
, ar
);
4593 /* allocate space in core */
4594 *errp
= ext4_mb_regular_allocator(ac
);
4596 goto discard_and_exit
;
4598 /* as we've just preallocated more space than
4599 * user requested originally, we store allocated
4600 * space in a special descriptor */
4601 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4602 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4603 *errp
= ext4_mb_new_preallocation(ac
);
4606 ext4_discard_allocated_blocks(ac
);
4610 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4611 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_clstrs
);
4613 ext4_discard_allocated_blocks(ac
);
4616 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4617 ar
->len
= ac
->ac_b_ex
.fe_len
;
4620 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4628 ac
->ac_b_ex
.fe_len
= 0;
4630 ext4_mb_show_ac(ac
);
4632 ext4_mb_release_context(ac
);
4635 kmem_cache_free(ext4_ac_cachep
, ac
);
4636 if (inquota
&& ar
->len
< inquota
)
4637 dquot_free_block(ar
->inode
, EXT4_C2B(sbi
, inquota
- ar
->len
));
4639 if ((ar
->flags
& EXT4_MB_DELALLOC_RESERVED
) == 0)
4640 /* release all the reserved blocks if non delalloc */
4641 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
4645 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4651 * We can merge two free data extents only if the physical blocks
4652 * are contiguous, AND the extents were freed by the same transaction,
4653 * AND the blocks are associated with the same group.
4655 static void ext4_try_merge_freed_extent(struct ext4_sb_info
*sbi
,
4656 struct ext4_free_data
*entry
,
4657 struct ext4_free_data
*new_entry
,
4658 struct rb_root
*entry_rb_root
)
4660 if ((entry
->efd_tid
!= new_entry
->efd_tid
) ||
4661 (entry
->efd_group
!= new_entry
->efd_group
))
4663 if (entry
->efd_start_cluster
+ entry
->efd_count
==
4664 new_entry
->efd_start_cluster
) {
4665 new_entry
->efd_start_cluster
= entry
->efd_start_cluster
;
4666 new_entry
->efd_count
+= entry
->efd_count
;
4667 } else if (new_entry
->efd_start_cluster
+ new_entry
->efd_count
==
4668 entry
->efd_start_cluster
) {
4669 new_entry
->efd_count
+= entry
->efd_count
;
4672 spin_lock(&sbi
->s_md_lock
);
4673 list_del(&entry
->efd_list
);
4674 spin_unlock(&sbi
->s_md_lock
);
4675 rb_erase(&entry
->efd_node
, entry_rb_root
);
4676 kmem_cache_free(ext4_free_data_cachep
, entry
);
4679 static noinline_for_stack
int
4680 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4681 struct ext4_free_data
*new_entry
)
4683 ext4_group_t group
= e4b
->bd_group
;
4684 ext4_grpblk_t cluster
;
4685 ext4_grpblk_t clusters
= new_entry
->efd_count
;
4686 struct ext4_free_data
*entry
;
4687 struct ext4_group_info
*db
= e4b
->bd_info
;
4688 struct super_block
*sb
= e4b
->bd_sb
;
4689 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4690 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4691 struct rb_node
*parent
= NULL
, *new_node
;
4693 BUG_ON(!ext4_handle_valid(handle
));
4694 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4695 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4697 new_node
= &new_entry
->efd_node
;
4698 cluster
= new_entry
->efd_start_cluster
;
4701 /* first free block exent. We need to
4702 protect buddy cache from being freed,
4703 * otherwise we'll refresh it from
4704 * on-disk bitmap and lose not-yet-available
4706 get_page(e4b
->bd_buddy_page
);
4707 get_page(e4b
->bd_bitmap_page
);
4711 entry
= rb_entry(parent
, struct ext4_free_data
, efd_node
);
4712 if (cluster
< entry
->efd_start_cluster
)
4714 else if (cluster
>= (entry
->efd_start_cluster
+ entry
->efd_count
))
4715 n
= &(*n
)->rb_right
;
4717 ext4_grp_locked_error(sb
, group
, 0,
4718 ext4_group_first_block_no(sb
, group
) +
4719 EXT4_C2B(sbi
, cluster
),
4720 "Block already on to-be-freed list");
4725 rb_link_node(new_node
, parent
, n
);
4726 rb_insert_color(new_node
, &db
->bb_free_root
);
4728 /* Now try to see the extent can be merged to left and right */
4729 node
= rb_prev(new_node
);
4731 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4732 ext4_try_merge_freed_extent(sbi
, entry
, new_entry
,
4733 &(db
->bb_free_root
));
4736 node
= rb_next(new_node
);
4738 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4739 ext4_try_merge_freed_extent(sbi
, entry
, new_entry
,
4740 &(db
->bb_free_root
));
4743 spin_lock(&sbi
->s_md_lock
);
4744 list_add_tail(&new_entry
->efd_list
, &sbi
->s_freed_data_list
);
4745 sbi
->s_mb_free_pending
+= clusters
;
4746 spin_unlock(&sbi
->s_md_lock
);
4751 * ext4_free_blocks() -- Free given blocks and update quota
4752 * @handle: handle for this transaction
4754 * @block: start physical block to free
4755 * @count: number of blocks to count
4756 * @flags: flags used by ext4_free_blocks
4758 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4759 struct buffer_head
*bh
, ext4_fsblk_t block
,
4760 unsigned long count
, int flags
)
4762 struct buffer_head
*bitmap_bh
= NULL
;
4763 struct super_block
*sb
= inode
->i_sb
;
4764 struct ext4_group_desc
*gdp
;
4765 unsigned int overflow
;
4767 struct buffer_head
*gd_bh
;
4768 ext4_group_t block_group
;
4769 struct ext4_sb_info
*sbi
;
4770 struct ext4_buddy e4b
;
4771 unsigned int count_clusters
;
4778 BUG_ON(block
!= bh
->b_blocknr
);
4780 block
= bh
->b_blocknr
;
4784 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4785 !ext4_data_block_valid(sbi
, block
, count
)) {
4786 ext4_error(sb
, "Freeing blocks not in datazone - "
4787 "block = %llu, count = %lu", block
, count
);
4791 ext4_debug("freeing block %llu\n", block
);
4792 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4794 if (bh
&& (flags
& EXT4_FREE_BLOCKS_FORGET
)) {
4797 ext4_forget(handle
, flags
& EXT4_FREE_BLOCKS_METADATA
,
4802 * If the extent to be freed does not begin on a cluster
4803 * boundary, we need to deal with partial clusters at the
4804 * beginning and end of the extent. Normally we will free
4805 * blocks at the beginning or the end unless we are explicitly
4806 * requested to avoid doing so.
4808 overflow
= EXT4_PBLK_COFF(sbi
, block
);
4810 if (flags
& EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER
) {
4811 overflow
= sbi
->s_cluster_ratio
- overflow
;
4813 if (count
> overflow
)
4822 overflow
= EXT4_LBLK_COFF(sbi
, count
);
4824 if (flags
& EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER
) {
4825 if (count
> overflow
)
4830 count
+= sbi
->s_cluster_ratio
- overflow
;
4833 if (!bh
&& (flags
& EXT4_FREE_BLOCKS_FORGET
)) {
4835 int is_metadata
= flags
& EXT4_FREE_BLOCKS_METADATA
;
4837 for (i
= 0; i
< count
; i
++) {
4840 bh
= sb_find_get_block(inode
->i_sb
, block
+ i
);
4841 ext4_forget(handle
, is_metadata
, inode
, bh
, block
+ i
);
4847 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4849 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4850 ext4_get_group_info(sb
, block_group
))))
4854 * Check to see if we are freeing blocks across a group
4857 if (EXT4_C2B(sbi
, bit
) + count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4858 overflow
= EXT4_C2B(sbi
, bit
) + count
-
4859 EXT4_BLOCKS_PER_GROUP(sb
);
4862 count_clusters
= EXT4_NUM_B2C(sbi
, count
);
4863 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4864 if (IS_ERR(bitmap_bh
)) {
4865 err
= PTR_ERR(bitmap_bh
);
4869 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4875 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4876 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4877 in_range(block
, ext4_inode_table(sb
, gdp
),
4878 EXT4_SB(sb
)->s_itb_per_group
) ||
4879 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4880 EXT4_SB(sb
)->s_itb_per_group
)) {
4882 ext4_error(sb
, "Freeing blocks in system zone - "
4883 "Block = %llu, count = %lu", block
, count
);
4884 /* err = 0. ext4_std_error should be a no op */
4888 BUFFER_TRACE(bitmap_bh
, "getting write access");
4889 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4894 * We are about to modify some metadata. Call the journal APIs
4895 * to unshare ->b_data if a currently-committing transaction is
4898 BUFFER_TRACE(gd_bh
, "get_write_access");
4899 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4902 #ifdef AGGRESSIVE_CHECK
4905 for (i
= 0; i
< count_clusters
; i
++)
4906 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4909 trace_ext4_mballoc_free(sb
, inode
, block_group
, bit
, count_clusters
);
4911 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4912 err
= ext4_mb_load_buddy_gfp(sb
, block_group
, &e4b
,
4913 GFP_NOFS
|__GFP_NOFAIL
);
4918 * We need to make sure we don't reuse the freed block until after the
4919 * transaction is committed. We make an exception if the inode is to be
4920 * written in writeback mode since writeback mode has weak data
4921 * consistency guarantees.
4923 if (ext4_handle_valid(handle
) &&
4924 ((flags
& EXT4_FREE_BLOCKS_METADATA
) ||
4925 !ext4_should_writeback_data(inode
))) {
4926 struct ext4_free_data
*new_entry
;
4928 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4931 new_entry
= kmem_cache_alloc(ext4_free_data_cachep
,
4932 GFP_NOFS
|__GFP_NOFAIL
);
4933 new_entry
->efd_start_cluster
= bit
;
4934 new_entry
->efd_group
= block_group
;
4935 new_entry
->efd_count
= count_clusters
;
4936 new_entry
->efd_tid
= handle
->h_transaction
->t_tid
;
4938 ext4_lock_group(sb
, block_group
);
4939 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4940 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4942 /* need to update group_info->bb_free and bitmap
4943 * with group lock held. generate_buddy look at
4944 * them with group lock_held
4946 if (test_opt(sb
, DISCARD
)) {
4947 err
= ext4_issue_discard(sb
, block_group
, bit
, count
,
4949 if (err
&& err
!= -EOPNOTSUPP
)
4950 ext4_msg(sb
, KERN_WARNING
, "discard request in"
4951 " group:%d block:%d count:%lu failed"
4952 " with %d", block_group
, bit
, count
,
4955 EXT4_MB_GRP_CLEAR_TRIMMED(e4b
.bd_info
);
4957 ext4_lock_group(sb
, block_group
);
4958 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4959 mb_free_blocks(inode
, &e4b
, bit
, count_clusters
);
4962 ret
= ext4_free_group_clusters(sb
, gdp
) + count_clusters
;
4963 ext4_free_group_clusters_set(sb
, gdp
, ret
);
4964 ext4_block_bitmap_csum_set(sb
, block_group
, gdp
, bitmap_bh
);
4965 ext4_group_desc_csum_set(sb
, block_group
, gdp
);
4966 ext4_unlock_group(sb
, block_group
);
4968 if (sbi
->s_log_groups_per_flex
) {
4969 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4970 atomic64_add(count_clusters
,
4971 &sbi_array_rcu_deref(sbi
, s_flex_groups
,
4972 flex_group
)->free_clusters
);
4975 if (!(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4976 dquot_free_block(inode
, EXT4_C2B(sbi
, count_clusters
));
4977 percpu_counter_add(&sbi
->s_freeclusters_counter
, count_clusters
);
4979 ext4_mb_unload_buddy(&e4b
);
4981 /* We dirtied the bitmap block */
4982 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4983 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4985 /* And the group descriptor block */
4986 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4987 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4991 if (overflow
&& !err
) {
4999 ext4_std_error(sb
, err
);
5004 * ext4_group_add_blocks() -- Add given blocks to an existing group
5005 * @handle: handle to this transaction
5007 * @block: start physical block to add to the block group
5008 * @count: number of blocks to free
5010 * This marks the blocks as free in the bitmap and buddy.
5012 int ext4_group_add_blocks(handle_t
*handle
, struct super_block
*sb
,
5013 ext4_fsblk_t block
, unsigned long count
)
5015 struct buffer_head
*bitmap_bh
= NULL
;
5016 struct buffer_head
*gd_bh
;
5017 ext4_group_t block_group
;
5020 struct ext4_group_desc
*desc
;
5021 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
5022 struct ext4_buddy e4b
;
5023 int err
= 0, ret
, blk_free_count
;
5024 ext4_grpblk_t blocks_freed
;
5026 ext4_debug("Adding block(s) %llu-%llu\n", block
, block
+ count
- 1);
5031 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
5033 * Check to see if we are freeing blocks across a group
5036 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
5037 ext4_warning(sb
, "too much blocks added to group %u",
5043 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
5044 if (IS_ERR(bitmap_bh
)) {
5045 err
= PTR_ERR(bitmap_bh
);
5050 desc
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
5056 if (in_range(ext4_block_bitmap(sb
, desc
), block
, count
) ||
5057 in_range(ext4_inode_bitmap(sb
, desc
), block
, count
) ||
5058 in_range(block
, ext4_inode_table(sb
, desc
), sbi
->s_itb_per_group
) ||
5059 in_range(block
+ count
- 1, ext4_inode_table(sb
, desc
),
5060 sbi
->s_itb_per_group
)) {
5061 ext4_error(sb
, "Adding blocks in system zones - "
5062 "Block = %llu, count = %lu",
5068 BUFFER_TRACE(bitmap_bh
, "getting write access");
5069 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
5074 * We are about to modify some metadata. Call the journal APIs
5075 * to unshare ->b_data if a currently-committing transaction is
5078 BUFFER_TRACE(gd_bh
, "get_write_access");
5079 err
= ext4_journal_get_write_access(handle
, gd_bh
);
5083 for (i
= 0, blocks_freed
= 0; i
< count
; i
++) {
5084 BUFFER_TRACE(bitmap_bh
, "clear bit");
5085 if (!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
)) {
5086 ext4_error(sb
, "bit already cleared for block %llu",
5087 (ext4_fsblk_t
)(block
+ i
));
5088 BUFFER_TRACE(bitmap_bh
, "bit already cleared");
5094 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
5099 * need to update group_info->bb_free and bitmap
5100 * with group lock held. generate_buddy look at
5101 * them with group lock_held
5103 ext4_lock_group(sb
, block_group
);
5104 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
5105 mb_free_blocks(NULL
, &e4b
, bit
, count
);
5106 blk_free_count
= blocks_freed
+ ext4_free_group_clusters(sb
, desc
);
5107 ext4_free_group_clusters_set(sb
, desc
, blk_free_count
);
5108 ext4_block_bitmap_csum_set(sb
, block_group
, desc
, bitmap_bh
);
5109 ext4_group_desc_csum_set(sb
, block_group
, desc
);
5110 ext4_unlock_group(sb
, block_group
);
5111 percpu_counter_add(&sbi
->s_freeclusters_counter
,
5112 EXT4_NUM_B2C(sbi
, blocks_freed
));
5114 if (sbi
->s_log_groups_per_flex
) {
5115 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
5116 atomic64_add(EXT4_NUM_B2C(sbi
, blocks_freed
),
5117 &sbi_array_rcu_deref(sbi
, s_flex_groups
,
5118 flex_group
)->free_clusters
);
5121 ext4_mb_unload_buddy(&e4b
);
5123 /* We dirtied the bitmap block */
5124 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
5125 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
5127 /* And the group descriptor block */
5128 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
5129 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
5135 ext4_std_error(sb
, err
);
5140 * ext4_trim_extent -- function to TRIM one single free extent in the group
5141 * @sb: super block for the file system
5142 * @start: starting block of the free extent in the alloc. group
5143 * @count: number of blocks to TRIM
5144 * @group: alloc. group we are working with
5145 * @e4b: ext4 buddy for the group
5147 * Trim "count" blocks starting at "start" in the "group". To assure that no
5148 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5149 * be called with under the group lock.
5151 static int ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
5152 ext4_group_t group
, struct ext4_buddy
*e4b
)
5156 struct ext4_free_extent ex
;
5159 trace_ext4_trim_extent(sb
, group
, start
, count
);
5161 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
5163 ex
.fe_start
= start
;
5164 ex
.fe_group
= group
;
5168 * Mark blocks used, so no one can reuse them while
5171 mb_mark_used(e4b
, &ex
);
5172 ext4_unlock_group(sb
, group
);
5173 ret
= ext4_issue_discard(sb
, group
, start
, count
, NULL
);
5174 ext4_lock_group(sb
, group
);
5175 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
5180 * ext4_trim_all_free -- function to trim all free space in alloc. group
5181 * @sb: super block for file system
5182 * @group: group to be trimmed
5183 * @start: first group block to examine
5184 * @max: last group block to examine
5185 * @minblocks: minimum extent block count
5187 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5188 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5192 * ext4_trim_all_free walks through group's block bitmap searching for free
5193 * extents. When the free extent is found, mark it as used in group buddy
5194 * bitmap. Then issue a TRIM command on this extent and free the extent in
5195 * the group buddy bitmap. This is done until whole group is scanned.
5197 static ext4_grpblk_t
5198 ext4_trim_all_free(struct super_block
*sb
, ext4_group_t group
,
5199 ext4_grpblk_t start
, ext4_grpblk_t max
,
5200 ext4_grpblk_t minblocks
)
5203 ext4_grpblk_t next
, count
= 0, free_count
= 0;
5204 struct ext4_buddy e4b
;
5207 trace_ext4_trim_all_free(sb
, group
, start
, max
);
5209 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
5211 ext4_warning(sb
, "Error %d loading buddy information for %u",
5215 bitmap
= e4b
.bd_bitmap
;
5217 ext4_lock_group(sb
, group
);
5218 if (EXT4_MB_GRP_WAS_TRIMMED(e4b
.bd_info
) &&
5219 minblocks
>= atomic_read(&EXT4_SB(sb
)->s_last_trim_minblks
))
5222 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
5223 e4b
.bd_info
->bb_first_free
: start
;
5225 while (start
<= max
) {
5226 start
= mb_find_next_zero_bit(bitmap
, max
+ 1, start
);
5229 next
= mb_find_next_bit(bitmap
, max
+ 1, start
);
5231 if ((next
- start
) >= minblocks
) {
5232 ret
= ext4_trim_extent(sb
, start
,
5233 next
- start
, group
, &e4b
);
5234 if (ret
&& ret
!= -EOPNOTSUPP
)
5237 count
+= next
- start
;
5239 free_count
+= next
- start
;
5242 if (fatal_signal_pending(current
)) {
5243 count
= -ERESTARTSYS
;
5247 if (need_resched()) {
5248 ext4_unlock_group(sb
, group
);
5250 ext4_lock_group(sb
, group
);
5253 if ((e4b
.bd_info
->bb_free
- free_count
) < minblocks
)
5259 EXT4_MB_GRP_SET_TRIMMED(e4b
.bd_info
);
5262 ext4_unlock_group(sb
, group
);
5263 ext4_mb_unload_buddy(&e4b
);
5265 ext4_debug("trimmed %d blocks in the group %d\n",
5272 * ext4_trim_fs() -- trim ioctl handle function
5273 * @sb: superblock for filesystem
5274 * @range: fstrim_range structure
5276 * start: First Byte to trim
5277 * len: number of Bytes to trim from start
5278 * minlen: minimum extent length in Bytes
5279 * ext4_trim_fs goes through all allocation groups containing Bytes from
5280 * start to start+len. For each such a group ext4_trim_all_free function
5281 * is invoked to trim all free space.
5283 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
5285 struct ext4_group_info
*grp
;
5286 ext4_group_t group
, first_group
, last_group
;
5287 ext4_grpblk_t cnt
= 0, first_cluster
, last_cluster
;
5288 uint64_t start
, end
, minlen
, trimmed
= 0;
5289 ext4_fsblk_t first_data_blk
=
5290 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
5291 ext4_fsblk_t max_blks
= ext4_blocks_count(EXT4_SB(sb
)->s_es
);
5294 start
= range
->start
>> sb
->s_blocksize_bits
;
5295 end
= start
+ (range
->len
>> sb
->s_blocksize_bits
) - 1;
5296 minlen
= EXT4_NUM_B2C(EXT4_SB(sb
),
5297 range
->minlen
>> sb
->s_blocksize_bits
);
5299 if (minlen
> EXT4_CLUSTERS_PER_GROUP(sb
) ||
5300 start
>= max_blks
||
5301 range
->len
< sb
->s_blocksize
)
5303 if (end
>= max_blks
)
5305 if (end
<= first_data_blk
)
5307 if (start
< first_data_blk
)
5308 start
= first_data_blk
;
5310 /* Determine first and last group to examine based on start and end */
5311 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
5312 &first_group
, &first_cluster
);
5313 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) end
,
5314 &last_group
, &last_cluster
);
5316 /* end now represents the last cluster to discard in this group */
5317 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5319 for (group
= first_group
; group
<= last_group
; group
++) {
5320 grp
= ext4_get_group_info(sb
, group
);
5321 /* We only do this if the grp has never been initialized */
5322 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
5323 ret
= ext4_mb_init_group(sb
, group
, GFP_NOFS
);
5329 * For all the groups except the last one, last cluster will
5330 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5331 * change it for the last group, note that last_cluster is
5332 * already computed earlier by ext4_get_group_no_and_offset()
5334 if (group
== last_group
)
5337 if (grp
->bb_free
>= minlen
) {
5338 cnt
= ext4_trim_all_free(sb
, group
, first_cluster
,
5348 * For every group except the first one, we are sure
5349 * that the first cluster to discard will be cluster #0.
5355 atomic_set(&EXT4_SB(sb
)->s_last_trim_minblks
, minlen
);
5358 range
->len
= EXT4_C2B(EXT4_SB(sb
), trimmed
) << sb
->s_blocksize_bits
;
5362 /* Iterate all the free extents in the group. */
5364 ext4_mballoc_query_range(
5365 struct super_block
*sb
,
5367 ext4_grpblk_t start
,
5369 ext4_mballoc_query_range_fn formatter
,
5374 struct ext4_buddy e4b
;
5377 error
= ext4_mb_load_buddy(sb
, group
, &e4b
);
5380 bitmap
= e4b
.bd_bitmap
;
5382 ext4_lock_group(sb
, group
);
5384 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
5385 e4b
.bd_info
->bb_first_free
: start
;
5386 if (end
>= EXT4_CLUSTERS_PER_GROUP(sb
))
5387 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5389 while (start
<= end
) {
5390 start
= mb_find_next_zero_bit(bitmap
, end
+ 1, start
);
5393 next
= mb_find_next_bit(bitmap
, end
+ 1, start
);
5395 ext4_unlock_group(sb
, group
);
5396 error
= formatter(sb
, group
, start
, next
- start
, priv
);
5399 ext4_lock_group(sb
, group
);
5404 ext4_unlock_group(sb
, group
);
5406 ext4_mb_unload_buddy(&e4b
);