2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
24 #include "ext4_jbd2.h"
26 #include <linux/log2.h>
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <linux/backing-dev.h>
30 #include <trace/events/ext4.h>
32 #ifdef CONFIG_EXT4_DEBUG
33 ushort ext4_mballoc_debug __read_mostly
;
35 module_param_named(mballoc_debug
, ext4_mballoc_debug
, ushort
, 0644);
36 MODULE_PARM_DESC(mballoc_debug
, "Debugging level for ext4's mballoc");
41 * - test ext4_ext_search_left() and ext4_ext_search_right()
42 * - search for metadata in few groups
45 * - normalization should take into account whether file is still open
46 * - discard preallocations if no free space left (policy?)
47 * - don't normalize tails
49 * - reservation for superuser
52 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
53 * - track min/max extents in each group for better group selection
54 * - mb_mark_used() may allocate chunk right after splitting buddy
55 * - tree of groups sorted by number of free blocks
60 * The allocation request involve request for multiple number of blocks
61 * near to the goal(block) value specified.
63 * During initialization phase of the allocator we decide to use the
64 * group preallocation or inode preallocation depending on the size of
65 * the file. The size of the file could be the resulting file size we
66 * would have after allocation, or the current file size, which ever
67 * is larger. If the size is less than sbi->s_mb_stream_request we
68 * select to use the group preallocation. The default value of
69 * s_mb_stream_request is 16 blocks. This can also be tuned via
70 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
71 * terms of number of blocks.
73 * The main motivation for having small file use group preallocation is to
74 * ensure that we have small files closer together on the disk.
76 * First stage the allocator looks at the inode prealloc list,
77 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
78 * spaces for this particular inode. The inode prealloc space is
81 * pa_lstart -> the logical start block for this prealloc space
82 * pa_pstart -> the physical start block for this prealloc space
83 * pa_len -> length for this prealloc space (in clusters)
84 * pa_free -> free space available in this prealloc space (in clusters)
86 * The inode preallocation space is used looking at the _logical_ start
87 * block. If only the logical file block falls within the range of prealloc
88 * space we will consume the particular prealloc space. This makes sure that
89 * we have contiguous physical blocks representing the file blocks
91 * The important thing to be noted in case of inode prealloc space is that
92 * we don't modify the values associated to inode prealloc space except
95 * If we are not able to find blocks in the inode prealloc space and if we
96 * have the group allocation flag set then we look at the locality group
97 * prealloc space. These are per CPU prealloc list represented as
99 * ext4_sb_info.s_locality_groups[smp_processor_id()]
101 * The reason for having a per cpu locality group is to reduce the contention
102 * between CPUs. It is possible to get scheduled at this point.
104 * The locality group prealloc space is used looking at whether we have
105 * enough free space (pa_free) within the prealloc space.
107 * If we can't allocate blocks via inode prealloc or/and locality group
108 * prealloc then we look at the buddy cache. The buddy cache is represented
109 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
110 * mapped to the buddy and bitmap information regarding different
111 * groups. The buddy information is attached to buddy cache inode so that
112 * we can access them through the page cache. The information regarding
113 * each group is loaded via ext4_mb_load_buddy. The information involve
114 * block bitmap and buddy information. The information are stored in the
118 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
121 * one block each for bitmap and buddy information. So for each group we
122 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
123 * blocksize) blocks. So it can have information regarding groups_per_page
124 * which is blocks_per_page/2
126 * The buddy cache inode is not stored on disk. The inode is thrown
127 * away when the filesystem is unmounted.
129 * We look for count number of blocks in the buddy cache. If we were able
130 * to locate that many free blocks we return with additional information
131 * regarding rest of the contiguous physical block available
133 * Before allocating blocks via buddy cache we normalize the request
134 * blocks. This ensure we ask for more blocks that we needed. The extra
135 * blocks that we get after allocation is added to the respective prealloc
136 * list. In case of inode preallocation we follow a list of heuristics
137 * based on file size. This can be found in ext4_mb_normalize_request. If
138 * we are doing a group prealloc we try to normalize the request to
139 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
140 * dependent on the cluster size; for non-bigalloc file systems, it is
141 * 512 blocks. This can be tuned via
142 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
143 * terms of number of blocks. If we have mounted the file system with -O
144 * stripe=<value> option the group prealloc request is normalized to the
145 * the smallest multiple of the stripe value (sbi->s_stripe) which is
146 * greater than the default mb_group_prealloc.
148 * The regular allocator (using the buddy cache) supports a few tunables.
150 * /sys/fs/ext4/<partition>/mb_min_to_scan
151 * /sys/fs/ext4/<partition>/mb_max_to_scan
152 * /sys/fs/ext4/<partition>/mb_order2_req
154 * The regular allocator uses buddy scan only if the request len is power of
155 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
156 * value of s_mb_order2_reqs can be tuned via
157 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
158 * stripe size (sbi->s_stripe), we try to search for contiguous block in
159 * stripe size. This should result in better allocation on RAID setups. If
160 * not, we search in the specific group using bitmap for best extents. The
161 * tunable min_to_scan and max_to_scan control the behaviour here.
162 * min_to_scan indicate how long the mballoc __must__ look for a best
163 * extent and max_to_scan indicates how long the mballoc __can__ look for a
164 * best extent in the found extents. Searching for the blocks starts with
165 * the group specified as the goal value in allocation context via
166 * ac_g_ex. Each group is first checked based on the criteria whether it
167 * can be used for allocation. ext4_mb_good_group explains how the groups are
170 * Both the prealloc space are getting populated as above. So for the first
171 * request we will hit the buddy cache which will result in this prealloc
172 * space getting filled. The prealloc space is then later used for the
173 * subsequent request.
177 * mballoc operates on the following data:
179 * - in-core buddy (actually includes buddy and bitmap)
180 * - preallocation descriptors (PAs)
182 * there are two types of preallocations:
184 * assiged to specific inode and can be used for this inode only.
185 * it describes part of inode's space preallocated to specific
186 * physical blocks. any block from that preallocated can be used
187 * independent. the descriptor just tracks number of blocks left
188 * unused. so, before taking some block from descriptor, one must
189 * make sure corresponded logical block isn't allocated yet. this
190 * also means that freeing any block within descriptor's range
191 * must discard all preallocated blocks.
193 * assigned to specific locality group which does not translate to
194 * permanent set of inodes: inode can join and leave group. space
195 * from this type of preallocation can be used for any inode. thus
196 * it's consumed from the beginning to the end.
198 * relation between them can be expressed as:
199 * in-core buddy = on-disk bitmap + preallocation descriptors
201 * this mean blocks mballoc considers used are:
202 * - allocated blocks (persistent)
203 * - preallocated blocks (non-persistent)
205 * consistency in mballoc world means that at any time a block is either
206 * free or used in ALL structures. notice: "any time" should not be read
207 * literally -- time is discrete and delimited by locks.
209 * to keep it simple, we don't use block numbers, instead we count number of
210 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
212 * all operations can be expressed as:
213 * - init buddy: buddy = on-disk + PAs
214 * - new PA: buddy += N; PA = N
215 * - use inode PA: on-disk += N; PA -= N
216 * - discard inode PA buddy -= on-disk - PA; PA = 0
217 * - use locality group PA on-disk += N; PA -= N
218 * - discard locality group PA buddy -= PA; PA = 0
219 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
220 * is used in real operation because we can't know actual used
221 * bits from PA, only from on-disk bitmap
223 * if we follow this strict logic, then all operations above should be atomic.
224 * given some of them can block, we'd have to use something like semaphores
225 * killing performance on high-end SMP hardware. let's try to relax it using
226 * the following knowledge:
227 * 1) if buddy is referenced, it's already initialized
228 * 2) while block is used in buddy and the buddy is referenced,
229 * nobody can re-allocate that block
230 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
231 * bit set and PA claims same block, it's OK. IOW, one can set bit in
232 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
235 * so, now we're building a concurrency table:
238 * blocks for PA are allocated in the buddy, buddy must be referenced
239 * until PA is linked to allocation group to avoid concurrent buddy init
241 * we need to make sure that either on-disk bitmap or PA has uptodate data
242 * given (3) we care that PA-=N operation doesn't interfere with init
244 * the simplest way would be to have buddy initialized by the discard
245 * - use locality group PA
246 * again PA-=N must be serialized with init
247 * - discard locality group PA
248 * the simplest way would be to have buddy initialized by the discard
251 * i_data_sem serializes them
253 * discard process must wait until PA isn't used by another process
254 * - use locality group PA
255 * some mutex should serialize them
256 * - discard locality group PA
257 * discard process must wait until PA isn't used by another process
260 * i_data_sem or another mutex should serializes them
262 * discard process must wait until PA isn't used by another process
263 * - use locality group PA
264 * nothing wrong here -- they're different PAs covering different blocks
265 * - discard locality group PA
266 * discard process must wait until PA isn't used by another process
268 * now we're ready to make few consequences:
269 * - PA is referenced and while it is no discard is possible
270 * - PA is referenced until block isn't marked in on-disk bitmap
271 * - PA changes only after on-disk bitmap
272 * - discard must not compete with init. either init is done before
273 * any discard or they're serialized somehow
274 * - buddy init as sum of on-disk bitmap and PAs is done atomically
276 * a special case when we've used PA to emptiness. no need to modify buddy
277 * in this case, but we should care about concurrent init
282 * Logic in few words:
287 * mark bits in on-disk bitmap
290 * - use preallocation:
291 * find proper PA (per-inode or group)
293 * mark bits in on-disk bitmap
299 * mark bits in on-disk bitmap
302 * - discard preallocations in group:
304 * move them onto local list
305 * load on-disk bitmap
307 * remove PA from object (inode or locality group)
308 * mark free blocks in-core
310 * - discard inode's preallocations:
317 * - bitlock on a group (group)
318 * - object (inode/locality) (object)
329 * - release consumed pa:
334 * - generate in-core bitmap:
338 * - discard all for given object (inode, locality group):
343 * - discard all for given group:
350 static struct kmem_cache
*ext4_pspace_cachep
;
351 static struct kmem_cache
*ext4_ac_cachep
;
352 static struct kmem_cache
*ext4_free_data_cachep
;
354 /* We create slab caches for groupinfo data structures based on the
355 * superblock block size. There will be one per mounted filesystem for
356 * each unique s_blocksize_bits */
357 #define NR_GRPINFO_CACHES 8
358 static struct kmem_cache
*ext4_groupinfo_caches
[NR_GRPINFO_CACHES
];
360 static const char *ext4_groupinfo_slab_names
[NR_GRPINFO_CACHES
] = {
361 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
362 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
363 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
366 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
368 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
370 static void ext4_free_data_callback(struct super_block
*sb
,
371 struct ext4_journal_cb_entry
*jce
, int rc
);
373 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
375 #if BITS_PER_LONG == 64
376 *bit
+= ((unsigned long) addr
& 7UL) << 3;
377 addr
= (void *) ((unsigned long) addr
& ~7UL);
378 #elif BITS_PER_LONG == 32
379 *bit
+= ((unsigned long) addr
& 3UL) << 3;
380 addr
= (void *) ((unsigned long) addr
& ~3UL);
382 #error "how many bits you are?!"
387 static inline int mb_test_bit(int bit
, void *addr
)
390 * ext4_test_bit on architecture like powerpc
391 * needs unsigned long aligned address
393 addr
= mb_correct_addr_and_bit(&bit
, addr
);
394 return ext4_test_bit(bit
, addr
);
397 static inline void mb_set_bit(int bit
, void *addr
)
399 addr
= mb_correct_addr_and_bit(&bit
, addr
);
400 ext4_set_bit(bit
, addr
);
403 static inline void mb_clear_bit(int bit
, void *addr
)
405 addr
= mb_correct_addr_and_bit(&bit
, addr
);
406 ext4_clear_bit(bit
, addr
);
409 static inline int mb_test_and_clear_bit(int bit
, void *addr
)
411 addr
= mb_correct_addr_and_bit(&bit
, addr
);
412 return ext4_test_and_clear_bit(bit
, addr
);
415 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
417 int fix
= 0, ret
, tmpmax
;
418 addr
= mb_correct_addr_and_bit(&fix
, addr
);
422 ret
= ext4_find_next_zero_bit(addr
, tmpmax
, start
) - fix
;
428 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
430 int fix
= 0, ret
, tmpmax
;
431 addr
= mb_correct_addr_and_bit(&fix
, addr
);
435 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
441 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
445 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
448 if (order
> e4b
->bd_blkbits
+ 1) {
453 /* at order 0 we see each particular block */
455 *max
= 1 << (e4b
->bd_blkbits
+ 3);
456 return e4b
->bd_bitmap
;
459 bb
= e4b
->bd_buddy
+ EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
460 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
466 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
467 int first
, int count
)
470 struct super_block
*sb
= e4b
->bd_sb
;
472 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
474 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
475 for (i
= 0; i
< count
; i
++) {
476 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
477 ext4_fsblk_t blocknr
;
479 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
480 blocknr
+= EXT4_C2B(EXT4_SB(sb
), first
+ i
);
481 ext4_grp_locked_error(sb
, e4b
->bd_group
,
482 inode
? inode
->i_ino
: 0,
484 "freeing block already freed "
488 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
492 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
496 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
498 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
499 for (i
= 0; i
< count
; i
++) {
500 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
501 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
505 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
507 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
508 unsigned char *b1
, *b2
;
510 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
511 b2
= (unsigned char *) bitmap
;
512 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
513 if (b1
[i
] != b2
[i
]) {
514 ext4_msg(e4b
->bd_sb
, KERN_ERR
,
515 "corruption in group %u "
516 "at byte %u(%u): %x in copy != %x "
518 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
526 static inline void mb_free_blocks_double(struct inode
*inode
,
527 struct ext4_buddy
*e4b
, int first
, int count
)
531 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
532 int first
, int count
)
536 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
542 #ifdef AGGRESSIVE_CHECK
544 #define MB_CHECK_ASSERT(assert) \
548 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
549 function, file, line, # assert); \
554 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
555 const char *function
, int line
)
557 struct super_block
*sb
= e4b
->bd_sb
;
558 int order
= e4b
->bd_blkbits
+ 1;
565 struct ext4_group_info
*grp
;
568 struct list_head
*cur
;
573 static int mb_check_counter
;
574 if (mb_check_counter
++ % 100 != 0)
579 buddy
= mb_find_buddy(e4b
, order
, &max
);
580 MB_CHECK_ASSERT(buddy
);
581 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
582 MB_CHECK_ASSERT(buddy2
);
583 MB_CHECK_ASSERT(buddy
!= buddy2
);
584 MB_CHECK_ASSERT(max
* 2 == max2
);
587 for (i
= 0; i
< max
; i
++) {
589 if (mb_test_bit(i
, buddy
)) {
590 /* only single bit in buddy2 may be 1 */
591 if (!mb_test_bit(i
<< 1, buddy2
)) {
593 mb_test_bit((i
<<1)+1, buddy2
));
594 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
596 mb_test_bit(i
<< 1, buddy2
));
601 /* both bits in buddy2 must be 1 */
602 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
603 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
605 for (j
= 0; j
< (1 << order
); j
++) {
606 k
= (i
* (1 << order
)) + j
;
608 !mb_test_bit(k
, e4b
->bd_bitmap
));
612 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
617 buddy
= mb_find_buddy(e4b
, 0, &max
);
618 for (i
= 0; i
< max
; i
++) {
619 if (!mb_test_bit(i
, buddy
)) {
620 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
628 /* check used bits only */
629 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
630 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
632 MB_CHECK_ASSERT(k
< max2
);
633 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
636 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
637 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
639 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
640 list_for_each(cur
, &grp
->bb_prealloc_list
) {
641 ext4_group_t groupnr
;
642 struct ext4_prealloc_space
*pa
;
643 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
644 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
645 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
646 for (i
= 0; i
< pa
->pa_len
; i
++)
647 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
651 #undef MB_CHECK_ASSERT
652 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
653 __FILE__, __func__, __LINE__)
655 #define mb_check_buddy(e4b)
659 * Divide blocks started from @first with length @len into
660 * smaller chunks with power of 2 blocks.
661 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
662 * then increase bb_counters[] for corresponded chunk size.
664 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
665 void *buddy
, ext4_grpblk_t first
, ext4_grpblk_t len
,
666 struct ext4_group_info
*grp
)
668 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
672 unsigned short border
;
674 BUG_ON(len
> EXT4_CLUSTERS_PER_GROUP(sb
));
676 border
= 2 << sb
->s_blocksize_bits
;
679 /* find how many blocks can be covered since this position */
680 max
= ffs(first
| border
) - 1;
682 /* find how many blocks of power 2 we need to mark */
689 /* mark multiblock chunks only */
690 grp
->bb_counters
[min
]++;
692 mb_clear_bit(first
>> min
,
693 buddy
+ sbi
->s_mb_offsets
[min
]);
701 * Cache the order of the largest free extent we have available in this block
705 mb_set_largest_free_order(struct super_block
*sb
, struct ext4_group_info
*grp
)
710 grp
->bb_largest_free_order
= -1; /* uninit */
712 bits
= sb
->s_blocksize_bits
+ 1;
713 for (i
= bits
; i
>= 0; i
--) {
714 if (grp
->bb_counters
[i
] > 0) {
715 grp
->bb_largest_free_order
= i
;
721 static noinline_for_stack
722 void ext4_mb_generate_buddy(struct super_block
*sb
,
723 void *buddy
, void *bitmap
, ext4_group_t group
)
725 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
726 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
727 ext4_grpblk_t max
= EXT4_CLUSTERS_PER_GROUP(sb
);
732 unsigned fragments
= 0;
733 unsigned long long period
= get_cycles();
735 /* initialize buddy from bitmap which is aggregation
736 * of on-disk bitmap and preallocations */
737 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
738 grp
->bb_first_free
= i
;
742 i
= mb_find_next_bit(bitmap
, max
, i
);
746 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
748 grp
->bb_counters
[0]++;
750 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
752 grp
->bb_fragments
= fragments
;
754 if (free
!= grp
->bb_free
) {
755 ext4_grp_locked_error(sb
, group
, 0, 0,
756 "block bitmap and bg descriptor "
757 "inconsistent: %u vs %u free clusters",
760 * If we intend to continue, we consider group descriptor
761 * corrupt and update bb_free using bitmap value
764 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp
))
765 percpu_counter_sub(&sbi
->s_freeclusters_counter
,
767 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT
, &grp
->bb_state
);
769 mb_set_largest_free_order(sb
, grp
);
771 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
773 period
= get_cycles() - period
;
774 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
775 EXT4_SB(sb
)->s_mb_buddies_generated
++;
776 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
777 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
780 static void mb_regenerate_buddy(struct ext4_buddy
*e4b
)
786 while ((buddy
= mb_find_buddy(e4b
, order
++, &count
))) {
787 ext4_set_bits(buddy
, 0, count
);
789 e4b
->bd_info
->bb_fragments
= 0;
790 memset(e4b
->bd_info
->bb_counters
, 0,
791 sizeof(*e4b
->bd_info
->bb_counters
) *
792 (e4b
->bd_sb
->s_blocksize_bits
+ 2));
794 ext4_mb_generate_buddy(e4b
->bd_sb
, e4b
->bd_buddy
,
795 e4b
->bd_bitmap
, e4b
->bd_group
);
798 /* The buddy information is attached the buddy cache inode
799 * for convenience. The information regarding each group
800 * is loaded via ext4_mb_load_buddy. The information involve
801 * block bitmap and buddy information. The information are
802 * stored in the inode as
805 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
808 * one block each for bitmap and buddy information.
809 * So for each group we take up 2 blocks. A page can
810 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
811 * So it can have information regarding groups_per_page which
812 * is blocks_per_page/2
814 * Locking note: This routine takes the block group lock of all groups
815 * for this page; do not hold this lock when calling this routine!
818 static int ext4_mb_init_cache(struct page
*page
, char *incore
, gfp_t gfp
)
820 ext4_group_t ngroups
;
826 ext4_group_t first_group
, group
;
828 struct super_block
*sb
;
829 struct buffer_head
*bhs
;
830 struct buffer_head
**bh
= NULL
;
834 struct ext4_group_info
*grinfo
;
836 mb_debug(1, "init page %lu\n", page
->index
);
838 inode
= page
->mapping
->host
;
840 ngroups
= ext4_get_groups_count(sb
);
841 blocksize
= 1 << inode
->i_blkbits
;
842 blocks_per_page
= PAGE_SIZE
/ blocksize
;
844 groups_per_page
= blocks_per_page
>> 1;
845 if (groups_per_page
== 0)
848 /* allocate buffer_heads to read bitmaps */
849 if (groups_per_page
> 1) {
850 i
= sizeof(struct buffer_head
*) * groups_per_page
;
851 bh
= kzalloc(i
, gfp
);
859 first_group
= page
->index
* blocks_per_page
/ 2;
861 /* read all groups the page covers into the cache */
862 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
863 if (group
>= ngroups
)
866 grinfo
= ext4_get_group_info(sb
, group
);
868 * If page is uptodate then we came here after online resize
869 * which added some new uninitialized group info structs, so
870 * we must skip all initialized uptodate buddies on the page,
871 * which may be currently in use by an allocating task.
873 if (PageUptodate(page
) && !EXT4_MB_GRP_NEED_INIT(grinfo
)) {
877 bh
[i
] = ext4_read_block_bitmap_nowait(sb
, group
);
879 err
= PTR_ERR(bh
[i
]);
883 mb_debug(1, "read bitmap for group %u\n", group
);
886 /* wait for I/O completion */
887 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
892 err2
= ext4_wait_block_bitmap(sb
, group
, bh
[i
]);
897 first_block
= page
->index
* blocks_per_page
;
898 for (i
= 0; i
< blocks_per_page
; i
++) {
899 group
= (first_block
+ i
) >> 1;
900 if (group
>= ngroups
)
903 if (!bh
[group
- first_group
])
904 /* skip initialized uptodate buddy */
907 if (!buffer_verified(bh
[group
- first_group
]))
908 /* Skip faulty bitmaps */
913 * data carry information regarding this
914 * particular group in the format specified
918 data
= page_address(page
) + (i
* blocksize
);
919 bitmap
= bh
[group
- first_group
]->b_data
;
922 * We place the buddy block and bitmap block
925 if ((first_block
+ i
) & 1) {
926 /* this is block of buddy */
927 BUG_ON(incore
== NULL
);
928 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
929 group
, page
->index
, i
* blocksize
);
930 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
931 grinfo
= ext4_get_group_info(sb
, group
);
932 grinfo
->bb_fragments
= 0;
933 memset(grinfo
->bb_counters
, 0,
934 sizeof(*grinfo
->bb_counters
) *
935 (sb
->s_blocksize_bits
+2));
937 * incore got set to the group block bitmap below
939 ext4_lock_group(sb
, group
);
941 memset(data
, 0xff, blocksize
);
942 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
943 ext4_unlock_group(sb
, group
);
946 /* this is block of bitmap */
947 BUG_ON(incore
!= NULL
);
948 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
949 group
, page
->index
, i
* blocksize
);
950 trace_ext4_mb_bitmap_load(sb
, group
);
952 /* see comments in ext4_mb_put_pa() */
953 ext4_lock_group(sb
, group
);
954 memcpy(data
, bitmap
, blocksize
);
956 /* mark all preallocated blks used in in-core bitmap */
957 ext4_mb_generate_from_pa(sb
, data
, group
);
958 ext4_mb_generate_from_freelist(sb
, data
, group
);
959 ext4_unlock_group(sb
, group
);
961 /* set incore so that the buddy information can be
962 * generated using this
967 SetPageUptodate(page
);
971 for (i
= 0; i
< groups_per_page
; i
++)
980 * Lock the buddy and bitmap pages. This make sure other parallel init_group
981 * on the same buddy page doesn't happen whild holding the buddy page lock.
982 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
983 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
985 static int ext4_mb_get_buddy_page_lock(struct super_block
*sb
,
986 ext4_group_t group
, struct ext4_buddy
*e4b
, gfp_t gfp
)
988 struct inode
*inode
= EXT4_SB(sb
)->s_buddy_cache
;
989 int block
, pnum
, poff
;
993 e4b
->bd_buddy_page
= NULL
;
994 e4b
->bd_bitmap_page
= NULL
;
996 blocks_per_page
= PAGE_SIZE
/ sb
->s_blocksize
;
998 * the buddy cache inode stores the block bitmap
999 * and buddy information in consecutive blocks.
1000 * So for each group we need two blocks.
1003 pnum
= block
/ blocks_per_page
;
1004 poff
= block
% blocks_per_page
;
1005 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1008 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1009 e4b
->bd_bitmap_page
= page
;
1010 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1012 if (blocks_per_page
>= 2) {
1013 /* buddy and bitmap are on the same page */
1018 pnum
= block
/ blocks_per_page
;
1019 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1022 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1023 e4b
->bd_buddy_page
= page
;
1027 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy
*e4b
)
1029 if (e4b
->bd_bitmap_page
) {
1030 unlock_page(e4b
->bd_bitmap_page
);
1031 put_page(e4b
->bd_bitmap_page
);
1033 if (e4b
->bd_buddy_page
) {
1034 unlock_page(e4b
->bd_buddy_page
);
1035 put_page(e4b
->bd_buddy_page
);
1040 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1041 * block group lock of all groups for this page; do not hold the BG lock when
1042 * calling this routine!
1044 static noinline_for_stack
1045 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
, gfp_t gfp
)
1048 struct ext4_group_info
*this_grp
;
1049 struct ext4_buddy e4b
;
1054 mb_debug(1, "init group %u\n", group
);
1055 this_grp
= ext4_get_group_info(sb
, group
);
1057 * This ensures that we don't reinit the buddy cache
1058 * page which map to the group from which we are already
1059 * allocating. If we are looking at the buddy cache we would
1060 * have taken a reference using ext4_mb_load_buddy and that
1061 * would have pinned buddy page to page cache.
1062 * The call to ext4_mb_get_buddy_page_lock will mark the
1065 ret
= ext4_mb_get_buddy_page_lock(sb
, group
, &e4b
, gfp
);
1066 if (ret
|| !EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1068 * somebody initialized the group
1069 * return without doing anything
1074 page
= e4b
.bd_bitmap_page
;
1075 ret
= ext4_mb_init_cache(page
, NULL
, gfp
);
1078 if (!PageUptodate(page
)) {
1083 if (e4b
.bd_buddy_page
== NULL
) {
1085 * If both the bitmap and buddy are in
1086 * the same page we don't need to force
1092 /* init buddy cache */
1093 page
= e4b
.bd_buddy_page
;
1094 ret
= ext4_mb_init_cache(page
, e4b
.bd_bitmap
, gfp
);
1097 if (!PageUptodate(page
)) {
1102 ext4_mb_put_buddy_page_lock(&e4b
);
1107 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1108 * block group lock of all groups for this page; do not hold the BG lock when
1109 * calling this routine!
1111 static noinline_for_stack
int
1112 ext4_mb_load_buddy_gfp(struct super_block
*sb
, ext4_group_t group
,
1113 struct ext4_buddy
*e4b
, gfp_t gfp
)
1115 int blocks_per_page
;
1121 struct ext4_group_info
*grp
;
1122 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1123 struct inode
*inode
= sbi
->s_buddy_cache
;
1126 mb_debug(1, "load group %u\n", group
);
1128 blocks_per_page
= PAGE_SIZE
/ sb
->s_blocksize
;
1129 grp
= ext4_get_group_info(sb
, group
);
1131 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1134 e4b
->bd_group
= group
;
1135 e4b
->bd_buddy_page
= NULL
;
1136 e4b
->bd_bitmap_page
= NULL
;
1138 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1140 * we need full data about the group
1141 * to make a good selection
1143 ret
= ext4_mb_init_group(sb
, group
, gfp
);
1149 * the buddy cache inode stores the block bitmap
1150 * and buddy information in consecutive blocks.
1151 * So for each group we need two blocks.
1154 pnum
= block
/ blocks_per_page
;
1155 poff
= block
% blocks_per_page
;
1157 /* we could use find_or_create_page(), but it locks page
1158 * what we'd like to avoid in fast path ... */
1159 page
= find_get_page_flags(inode
->i_mapping
, pnum
, FGP_ACCESSED
);
1160 if (page
== NULL
|| !PageUptodate(page
)) {
1163 * drop the page reference and try
1164 * to get the page with lock. If we
1165 * are not uptodate that implies
1166 * somebody just created the page but
1167 * is yet to initialize the same. So
1168 * wait for it to initialize.
1171 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1173 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1174 if (!PageUptodate(page
)) {
1175 ret
= ext4_mb_init_cache(page
, NULL
, gfp
);
1180 mb_cmp_bitmaps(e4b
, page_address(page
) +
1181 (poff
* sb
->s_blocksize
));
1190 if (!PageUptodate(page
)) {
1195 /* Pages marked accessed already */
1196 e4b
->bd_bitmap_page
= page
;
1197 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1200 pnum
= block
/ blocks_per_page
;
1201 poff
= block
% blocks_per_page
;
1203 page
= find_get_page_flags(inode
->i_mapping
, pnum
, FGP_ACCESSED
);
1204 if (page
== NULL
|| !PageUptodate(page
)) {
1207 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1209 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1210 if (!PageUptodate(page
)) {
1211 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
,
1225 if (!PageUptodate(page
)) {
1230 /* Pages marked accessed already */
1231 e4b
->bd_buddy_page
= page
;
1232 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1234 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1235 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1242 if (e4b
->bd_bitmap_page
)
1243 put_page(e4b
->bd_bitmap_page
);
1244 if (e4b
->bd_buddy_page
)
1245 put_page(e4b
->bd_buddy_page
);
1246 e4b
->bd_buddy
= NULL
;
1247 e4b
->bd_bitmap
= NULL
;
1251 static int ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1252 struct ext4_buddy
*e4b
)
1254 return ext4_mb_load_buddy_gfp(sb
, group
, e4b
, GFP_NOFS
);
1257 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1259 if (e4b
->bd_bitmap_page
)
1260 put_page(e4b
->bd_bitmap_page
);
1261 if (e4b
->bd_buddy_page
)
1262 put_page(e4b
->bd_buddy_page
);
1266 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1269 int bb_incr
= 1 << (e4b
->bd_blkbits
- 1);
1272 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
1273 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1276 while (order
<= e4b
->bd_blkbits
+ 1) {
1278 if (!mb_test_bit(block
, bb
)) {
1279 /* this block is part of buddy of order 'order' */
1289 static void mb_clear_bits(void *bm
, int cur
, int len
)
1295 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1296 /* fast path: clear whole word at once */
1297 addr
= bm
+ (cur
>> 3);
1302 mb_clear_bit(cur
, bm
);
1307 /* clear bits in given range
1308 * will return first found zero bit if any, -1 otherwise
1310 static int mb_test_and_clear_bits(void *bm
, int cur
, int len
)
1317 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1318 /* fast path: clear whole word at once */
1319 addr
= bm
+ (cur
>> 3);
1320 if (*addr
!= (__u32
)(-1) && zero_bit
== -1)
1321 zero_bit
= cur
+ mb_find_next_zero_bit(addr
, 32, 0);
1326 if (!mb_test_and_clear_bit(cur
, bm
) && zero_bit
== -1)
1334 void ext4_set_bits(void *bm
, int cur
, int len
)
1340 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1341 /* fast path: set whole word at once */
1342 addr
= bm
+ (cur
>> 3);
1347 mb_set_bit(cur
, bm
);
1353 * _________________________________________________________________ */
1355 static inline int mb_buddy_adjust_border(int* bit
, void* bitmap
, int side
)
1357 if (mb_test_bit(*bit
+ side
, bitmap
)) {
1358 mb_clear_bit(*bit
, bitmap
);
1364 mb_set_bit(*bit
, bitmap
);
1369 static void mb_buddy_mark_free(struct ext4_buddy
*e4b
, int first
, int last
)
1373 void *buddy
= mb_find_buddy(e4b
, order
, &max
);
1378 /* Bits in range [first; last] are known to be set since
1379 * corresponding blocks were allocated. Bits in range
1380 * (first; last) will stay set because they form buddies on
1381 * upper layer. We just deal with borders if they don't
1382 * align with upper layer and then go up.
1383 * Releasing entire group is all about clearing
1384 * single bit of highest order buddy.
1388 * ---------------------------------
1390 * ---------------------------------
1391 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1392 * ---------------------------------
1394 * \_____________________/
1396 * Neither [1] nor [6] is aligned to above layer.
1397 * Left neighbour [0] is free, so mark it busy,
1398 * decrease bb_counters and extend range to
1400 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1401 * mark [6] free, increase bb_counters and shrink range to
1403 * Then shift range to [0; 2], go up and do the same.
1408 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&first
, buddy
, -1);
1410 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&last
, buddy
, 1);
1415 if (first
== last
|| !(buddy2
= mb_find_buddy(e4b
, order
, &max
))) {
1416 mb_clear_bits(buddy
, first
, last
- first
+ 1);
1417 e4b
->bd_info
->bb_counters
[order
- 1] += last
- first
+ 1;
1426 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1427 int first
, int count
)
1429 int left_is_free
= 0;
1430 int right_is_free
= 0;
1432 int last
= first
+ count
- 1;
1433 struct super_block
*sb
= e4b
->bd_sb
;
1435 if (WARN_ON(count
== 0))
1437 BUG_ON(last
>= (sb
->s_blocksize
<< 3));
1438 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1439 /* Don't bother if the block group is corrupt. */
1440 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
)))
1443 mb_check_buddy(e4b
);
1444 mb_free_blocks_double(inode
, e4b
, first
, count
);
1446 e4b
->bd_info
->bb_free
+= count
;
1447 if (first
< e4b
->bd_info
->bb_first_free
)
1448 e4b
->bd_info
->bb_first_free
= first
;
1450 /* access memory sequentially: check left neighbour,
1451 * clear range and then check right neighbour
1454 left_is_free
= !mb_test_bit(first
- 1, e4b
->bd_bitmap
);
1455 block
= mb_test_and_clear_bits(e4b
->bd_bitmap
, first
, count
);
1456 if (last
+ 1 < EXT4_SB(sb
)->s_mb_maxs
[0])
1457 right_is_free
= !mb_test_bit(last
+ 1, e4b
->bd_bitmap
);
1459 if (unlikely(block
!= -1)) {
1460 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1461 ext4_fsblk_t blocknr
;
1463 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1464 blocknr
+= EXT4_C2B(EXT4_SB(sb
), block
);
1465 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1466 inode
? inode
->i_ino
: 0,
1468 "freeing already freed block "
1469 "(bit %u); block bitmap corrupt.",
1471 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
))
1472 percpu_counter_sub(&sbi
->s_freeclusters_counter
,
1473 e4b
->bd_info
->bb_free
);
1474 /* Mark the block group as corrupt. */
1475 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT
,
1476 &e4b
->bd_info
->bb_state
);
1477 mb_regenerate_buddy(e4b
);
1481 /* let's maintain fragments counter */
1482 if (left_is_free
&& right_is_free
)
1483 e4b
->bd_info
->bb_fragments
--;
1484 else if (!left_is_free
&& !right_is_free
)
1485 e4b
->bd_info
->bb_fragments
++;
1487 /* buddy[0] == bd_bitmap is a special case, so handle
1488 * it right away and let mb_buddy_mark_free stay free of
1489 * zero order checks.
1490 * Check if neighbours are to be coaleasced,
1491 * adjust bitmap bb_counters and borders appropriately.
1494 first
+= !left_is_free
;
1495 e4b
->bd_info
->bb_counters
[0] += left_is_free
? -1 : 1;
1498 last
-= !right_is_free
;
1499 e4b
->bd_info
->bb_counters
[0] += right_is_free
? -1 : 1;
1503 mb_buddy_mark_free(e4b
, first
>> 1, last
>> 1);
1506 mb_set_largest_free_order(sb
, e4b
->bd_info
);
1507 mb_check_buddy(e4b
);
1510 static int mb_find_extent(struct ext4_buddy
*e4b
, int block
,
1511 int needed
, struct ext4_free_extent
*ex
)
1517 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1520 buddy
= mb_find_buddy(e4b
, 0, &max
);
1521 BUG_ON(buddy
== NULL
);
1522 BUG_ON(block
>= max
);
1523 if (mb_test_bit(block
, buddy
)) {
1530 /* find actual order */
1531 order
= mb_find_order_for_block(e4b
, block
);
1532 block
= block
>> order
;
1534 ex
->fe_len
= 1 << order
;
1535 ex
->fe_start
= block
<< order
;
1536 ex
->fe_group
= e4b
->bd_group
;
1538 /* calc difference from given start */
1539 next
= next
- ex
->fe_start
;
1541 ex
->fe_start
+= next
;
1543 while (needed
> ex
->fe_len
&&
1544 mb_find_buddy(e4b
, order
, &max
)) {
1546 if (block
+ 1 >= max
)
1549 next
= (block
+ 1) * (1 << order
);
1550 if (mb_test_bit(next
, e4b
->bd_bitmap
))
1553 order
= mb_find_order_for_block(e4b
, next
);
1555 block
= next
>> order
;
1556 ex
->fe_len
+= 1 << order
;
1559 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1563 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1569 int start
= ex
->fe_start
;
1570 int len
= ex
->fe_len
;
1575 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1576 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1577 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1578 mb_check_buddy(e4b
);
1579 mb_mark_used_double(e4b
, start
, len
);
1581 e4b
->bd_info
->bb_free
-= len
;
1582 if (e4b
->bd_info
->bb_first_free
== start
)
1583 e4b
->bd_info
->bb_first_free
+= len
;
1585 /* let's maintain fragments counter */
1587 mlen
= !mb_test_bit(start
- 1, e4b
->bd_bitmap
);
1588 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1589 max
= !mb_test_bit(start
+ len
, e4b
->bd_bitmap
);
1591 e4b
->bd_info
->bb_fragments
++;
1592 else if (!mlen
&& !max
)
1593 e4b
->bd_info
->bb_fragments
--;
1595 /* let's maintain buddy itself */
1597 ord
= mb_find_order_for_block(e4b
, start
);
1599 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1600 /* the whole chunk may be allocated at once! */
1602 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1603 BUG_ON((start
>> ord
) >= max
);
1604 mb_set_bit(start
>> ord
, buddy
);
1605 e4b
->bd_info
->bb_counters
[ord
]--;
1612 /* store for history */
1614 ret
= len
| (ord
<< 16);
1616 /* we have to split large buddy */
1618 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1619 mb_set_bit(start
>> ord
, buddy
);
1620 e4b
->bd_info
->bb_counters
[ord
]--;
1623 cur
= (start
>> ord
) & ~1U;
1624 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1625 mb_clear_bit(cur
, buddy
);
1626 mb_clear_bit(cur
+ 1, buddy
);
1627 e4b
->bd_info
->bb_counters
[ord
]++;
1628 e4b
->bd_info
->bb_counters
[ord
]++;
1630 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1632 ext4_set_bits(e4b
->bd_bitmap
, ex
->fe_start
, len0
);
1633 mb_check_buddy(e4b
);
1639 * Must be called under group lock!
1641 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1642 struct ext4_buddy
*e4b
)
1644 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1647 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1648 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1650 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1651 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1652 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1654 /* preallocation can change ac_b_ex, thus we store actually
1655 * allocated blocks for history */
1656 ac
->ac_f_ex
= ac
->ac_b_ex
;
1658 ac
->ac_status
= AC_STATUS_FOUND
;
1659 ac
->ac_tail
= ret
& 0xffff;
1660 ac
->ac_buddy
= ret
>> 16;
1663 * take the page reference. We want the page to be pinned
1664 * so that we don't get a ext4_mb_init_cache_call for this
1665 * group until we update the bitmap. That would mean we
1666 * double allocate blocks. The reference is dropped
1667 * in ext4_mb_release_context
1669 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1670 get_page(ac
->ac_bitmap_page
);
1671 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1672 get_page(ac
->ac_buddy_page
);
1673 /* store last allocated for subsequent stream allocation */
1674 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1675 spin_lock(&sbi
->s_md_lock
);
1676 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1677 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1678 spin_unlock(&sbi
->s_md_lock
);
1683 * regular allocator, for general purposes allocation
1686 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1687 struct ext4_buddy
*e4b
,
1690 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1691 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1692 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1693 struct ext4_free_extent ex
;
1696 if (ac
->ac_status
== AC_STATUS_FOUND
)
1699 * We don't want to scan for a whole year
1701 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1702 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1703 ac
->ac_status
= AC_STATUS_BREAK
;
1708 * Haven't found good chunk so far, let's continue
1710 if (bex
->fe_len
< gex
->fe_len
)
1713 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1714 && bex
->fe_group
== e4b
->bd_group
) {
1715 /* recheck chunk's availability - we don't know
1716 * when it was found (within this lock-unlock
1718 max
= mb_find_extent(e4b
, bex
->fe_start
, gex
->fe_len
, &ex
);
1719 if (max
>= gex
->fe_len
) {
1720 ext4_mb_use_best_found(ac
, e4b
);
1727 * The routine checks whether found extent is good enough. If it is,
1728 * then the extent gets marked used and flag is set to the context
1729 * to stop scanning. Otherwise, the extent is compared with the
1730 * previous found extent and if new one is better, then it's stored
1731 * in the context. Later, the best found extent will be used, if
1732 * mballoc can't find good enough extent.
1734 * FIXME: real allocation policy is to be designed yet!
1736 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1737 struct ext4_free_extent
*ex
,
1738 struct ext4_buddy
*e4b
)
1740 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1741 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1743 BUG_ON(ex
->fe_len
<= 0);
1744 BUG_ON(ex
->fe_len
> EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1745 BUG_ON(ex
->fe_start
>= EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1746 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1751 * The special case - take what you catch first
1753 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1755 ext4_mb_use_best_found(ac
, e4b
);
1760 * Let's check whether the chuck is good enough
1762 if (ex
->fe_len
== gex
->fe_len
) {
1764 ext4_mb_use_best_found(ac
, e4b
);
1769 * If this is first found extent, just store it in the context
1771 if (bex
->fe_len
== 0) {
1777 * If new found extent is better, store it in the context
1779 if (bex
->fe_len
< gex
->fe_len
) {
1780 /* if the request isn't satisfied, any found extent
1781 * larger than previous best one is better */
1782 if (ex
->fe_len
> bex
->fe_len
)
1784 } else if (ex
->fe_len
> gex
->fe_len
) {
1785 /* if the request is satisfied, then we try to find
1786 * an extent that still satisfy the request, but is
1787 * smaller than previous one */
1788 if (ex
->fe_len
< bex
->fe_len
)
1792 ext4_mb_check_limits(ac
, e4b
, 0);
1795 static noinline_for_stack
1796 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1797 struct ext4_buddy
*e4b
)
1799 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1800 ext4_group_t group
= ex
.fe_group
;
1804 BUG_ON(ex
.fe_len
<= 0);
1805 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1809 ext4_lock_group(ac
->ac_sb
, group
);
1810 max
= mb_find_extent(e4b
, ex
.fe_start
, ex
.fe_len
, &ex
);
1814 ext4_mb_use_best_found(ac
, e4b
);
1817 ext4_unlock_group(ac
->ac_sb
, group
);
1818 ext4_mb_unload_buddy(e4b
);
1823 static noinline_for_stack
1824 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1825 struct ext4_buddy
*e4b
)
1827 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1830 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1831 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1832 struct ext4_free_extent ex
;
1834 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1836 if (grp
->bb_free
== 0)
1839 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1843 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
))) {
1844 ext4_mb_unload_buddy(e4b
);
1848 ext4_lock_group(ac
->ac_sb
, group
);
1849 max
= mb_find_extent(e4b
, ac
->ac_g_ex
.fe_start
,
1850 ac
->ac_g_ex
.fe_len
, &ex
);
1851 ex
.fe_logical
= 0xDEADFA11; /* debug value */
1853 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1856 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1858 /* use do_div to get remainder (would be 64-bit modulo) */
1859 if (do_div(start
, sbi
->s_stripe
) == 0) {
1862 ext4_mb_use_best_found(ac
, e4b
);
1864 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1865 BUG_ON(ex
.fe_len
<= 0);
1866 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1867 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1870 ext4_mb_use_best_found(ac
, e4b
);
1871 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1872 /* Sometimes, caller may want to merge even small
1873 * number of blocks to an existing extent */
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
);
1881 ext4_unlock_group(ac
->ac_sb
, group
);
1882 ext4_mb_unload_buddy(e4b
);
1888 * The routine scans buddy structures (not bitmap!) from given order
1889 * to max order and tries to find big enough chunk to satisfy the req
1891 static noinline_for_stack
1892 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1893 struct ext4_buddy
*e4b
)
1895 struct super_block
*sb
= ac
->ac_sb
;
1896 struct ext4_group_info
*grp
= e4b
->bd_info
;
1902 BUG_ON(ac
->ac_2order
<= 0);
1903 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1904 if (grp
->bb_counters
[i
] == 0)
1907 buddy
= mb_find_buddy(e4b
, i
, &max
);
1908 BUG_ON(buddy
== NULL
);
1910 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1915 ac
->ac_b_ex
.fe_len
= 1 << i
;
1916 ac
->ac_b_ex
.fe_start
= k
<< i
;
1917 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1919 ext4_mb_use_best_found(ac
, e4b
);
1921 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1923 if (EXT4_SB(sb
)->s_mb_stats
)
1924 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1931 * The routine scans the group and measures all found extents.
1932 * In order to optimize scanning, caller must pass number of
1933 * free blocks in the group, so the routine can know upper limit.
1935 static noinline_for_stack
1936 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1937 struct ext4_buddy
*e4b
)
1939 struct super_block
*sb
= ac
->ac_sb
;
1940 void *bitmap
= e4b
->bd_bitmap
;
1941 struct ext4_free_extent ex
;
1945 free
= e4b
->bd_info
->bb_free
;
1948 i
= e4b
->bd_info
->bb_first_free
;
1950 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1951 i
= mb_find_next_zero_bit(bitmap
,
1952 EXT4_CLUSTERS_PER_GROUP(sb
), i
);
1953 if (i
>= EXT4_CLUSTERS_PER_GROUP(sb
)) {
1955 * IF we have corrupt bitmap, we won't find any
1956 * free blocks even though group info says we
1957 * we have free blocks
1959 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1960 "%d free clusters as per "
1961 "group info. But bitmap says 0",
1966 mb_find_extent(e4b
, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1967 BUG_ON(ex
.fe_len
<= 0);
1968 if (free
< ex
.fe_len
) {
1969 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1970 "%d free clusters as per "
1971 "group info. But got %d blocks",
1974 * The number of free blocks differs. This mostly
1975 * indicate that the bitmap is corrupt. So exit
1976 * without claiming the space.
1980 ex
.fe_logical
= 0xDEADC0DE; /* debug value */
1981 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1987 ext4_mb_check_limits(ac
, e4b
, 1);
1991 * This is a special case for storages like raid5
1992 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1994 static noinline_for_stack
1995 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1996 struct ext4_buddy
*e4b
)
1998 struct super_block
*sb
= ac
->ac_sb
;
1999 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2000 void *bitmap
= e4b
->bd_bitmap
;
2001 struct ext4_free_extent ex
;
2002 ext4_fsblk_t first_group_block
;
2007 BUG_ON(sbi
->s_stripe
== 0);
2009 /* find first stripe-aligned block in group */
2010 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
2012 a
= first_group_block
+ sbi
->s_stripe
- 1;
2013 do_div(a
, sbi
->s_stripe
);
2014 i
= (a
* sbi
->s_stripe
) - first_group_block
;
2016 while (i
< EXT4_CLUSTERS_PER_GROUP(sb
)) {
2017 if (!mb_test_bit(i
, bitmap
)) {
2018 max
= mb_find_extent(e4b
, i
, sbi
->s_stripe
, &ex
);
2019 if (max
>= sbi
->s_stripe
) {
2021 ex
.fe_logical
= 0xDEADF00D; /* debug value */
2023 ext4_mb_use_best_found(ac
, e4b
);
2032 * This is now called BEFORE we load the buddy bitmap.
2033 * Returns either 1 or 0 indicating that the group is either suitable
2034 * for the allocation or not. In addition it can also return negative
2035 * error code when something goes wrong.
2037 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
2038 ext4_group_t group
, int cr
)
2040 unsigned free
, fragments
;
2041 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
2042 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
2044 BUG_ON(cr
< 0 || cr
>= 4);
2046 free
= grp
->bb_free
;
2049 if (cr
<= 2 && free
< ac
->ac_g_ex
.fe_len
)
2052 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp
)))
2055 /* We only do this if the grp has never been initialized */
2056 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
2057 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
, GFP_NOFS
);
2062 fragments
= grp
->bb_fragments
;
2068 BUG_ON(ac
->ac_2order
== 0);
2070 /* Avoid using the first bg of a flexgroup for data files */
2071 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
2072 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
2073 ((group
% flex_size
) == 0))
2076 if ((ac
->ac_2order
> ac
->ac_sb
->s_blocksize_bits
+1) ||
2077 (free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2080 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
2085 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2089 if (free
>= ac
->ac_g_ex
.fe_len
)
2101 static noinline_for_stack
int
2102 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
2104 ext4_group_t ngroups
, group
, i
;
2106 int err
= 0, first_err
= 0;
2107 struct ext4_sb_info
*sbi
;
2108 struct super_block
*sb
;
2109 struct ext4_buddy e4b
;
2113 ngroups
= ext4_get_groups_count(sb
);
2114 /* non-extent files are limited to low blocks/groups */
2115 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
2116 ngroups
= sbi
->s_blockfile_groups
;
2118 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
2120 /* first, try the goal */
2121 err
= ext4_mb_find_by_goal(ac
, &e4b
);
2122 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
2125 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2129 * ac->ac2_order is set only if the fe_len is a power of 2
2130 * if ac2_order is set we also set criteria to 0 so that we
2131 * try exact allocation using buddy.
2133 i
= fls(ac
->ac_g_ex
.fe_len
);
2136 * We search using buddy data only if the order of the request
2137 * is greater than equal to the sbi_s_mb_order2_reqs
2138 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2140 if (i
>= sbi
->s_mb_order2_reqs
) {
2142 * This should tell if fe_len is exactly power of 2
2144 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2145 ac
->ac_2order
= i
- 1;
2148 /* if stream allocation is enabled, use global goal */
2149 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2150 /* TBD: may be hot point */
2151 spin_lock(&sbi
->s_md_lock
);
2152 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2153 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2154 spin_unlock(&sbi
->s_md_lock
);
2157 /* Let's just scan groups to find more-less suitable blocks */
2158 cr
= ac
->ac_2order
? 0 : 1;
2160 * cr == 0 try to get exact allocation,
2161 * cr == 3 try to get anything
2164 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2165 ac
->ac_criteria
= cr
;
2167 * searching for the right group start
2168 * from the goal value specified
2170 group
= ac
->ac_g_ex
.fe_group
;
2172 for (i
= 0; i
< ngroups
; group
++, i
++) {
2176 * Artificially restricted ngroups for non-extent
2177 * files makes group > ngroups possible on first loop.
2179 if (group
>= ngroups
)
2182 /* This now checks without needing the buddy page */
2183 ret
= ext4_mb_good_group(ac
, group
, cr
);
2190 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2194 ext4_lock_group(sb
, group
);
2197 * We need to check again after locking the
2200 ret
= ext4_mb_good_group(ac
, group
, cr
);
2202 ext4_unlock_group(sb
, group
);
2203 ext4_mb_unload_buddy(&e4b
);
2209 ac
->ac_groups_scanned
++;
2210 if (cr
== 0 && ac
->ac_2order
< sb
->s_blocksize_bits
+2)
2211 ext4_mb_simple_scan_group(ac
, &e4b
);
2212 else if (cr
== 1 && sbi
->s_stripe
&&
2213 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2214 ext4_mb_scan_aligned(ac
, &e4b
);
2216 ext4_mb_complex_scan_group(ac
, &e4b
);
2218 ext4_unlock_group(sb
, group
);
2219 ext4_mb_unload_buddy(&e4b
);
2221 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2226 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2227 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2229 * We've been searching too long. Let's try to allocate
2230 * the best chunk we've found so far
2233 ext4_mb_try_best_found(ac
, &e4b
);
2234 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2236 * Someone more lucky has already allocated it.
2237 * The only thing we can do is just take first
2239 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2241 ac
->ac_b_ex
.fe_group
= 0;
2242 ac
->ac_b_ex
.fe_start
= 0;
2243 ac
->ac_b_ex
.fe_len
= 0;
2244 ac
->ac_status
= AC_STATUS_CONTINUE
;
2245 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2247 atomic_inc(&sbi
->s_mb_lost_chunks
);
2252 if (!err
&& ac
->ac_status
!= AC_STATUS_FOUND
&& first_err
)
2257 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2259 struct super_block
*sb
= seq
->private;
2262 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2265 return (void *) ((unsigned long) group
);
2268 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2270 struct super_block
*sb
= seq
->private;
2274 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2277 return (void *) ((unsigned long) group
);
2280 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2282 struct super_block
*sb
= seq
->private;
2283 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2285 int err
, buddy_loaded
= 0;
2286 struct ext4_buddy e4b
;
2287 struct ext4_group_info
*grinfo
;
2289 struct ext4_group_info info
;
2290 ext4_grpblk_t counters
[16];
2295 seq_puts(seq
, "#group: free frags first ["
2296 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2297 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2299 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2300 sizeof(struct ext4_group_info
);
2301 grinfo
= ext4_get_group_info(sb
, group
);
2302 /* Load the group info in memory only if not already loaded. */
2303 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo
))) {
2304 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2306 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2312 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2315 ext4_mb_unload_buddy(&e4b
);
2317 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2318 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2319 for (i
= 0; i
<= 13; i
++)
2320 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2321 sg
.info
.bb_counters
[i
] : 0);
2322 seq_printf(seq
, " ]\n");
2327 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2331 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2332 .start
= ext4_mb_seq_groups_start
,
2333 .next
= ext4_mb_seq_groups_next
,
2334 .stop
= ext4_mb_seq_groups_stop
,
2335 .show
= ext4_mb_seq_groups_show
,
2338 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2340 struct super_block
*sb
= PDE_DATA(inode
);
2343 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2345 struct seq_file
*m
= file
->private_data
;
2352 const struct file_operations ext4_seq_mb_groups_fops
= {
2353 .owner
= THIS_MODULE
,
2354 .open
= ext4_mb_seq_groups_open
,
2356 .llseek
= seq_lseek
,
2357 .release
= seq_release
,
2360 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2362 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2363 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2370 * Allocate the top-level s_group_info array for the specified number
2373 int ext4_mb_alloc_groupinfo(struct super_block
*sb
, ext4_group_t ngroups
)
2375 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2377 struct ext4_group_info
***new_groupinfo
;
2379 size
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2380 EXT4_DESC_PER_BLOCK_BITS(sb
);
2381 if (size
<= sbi
->s_group_info_size
)
2384 size
= roundup_pow_of_two(sizeof(*sbi
->s_group_info
) * size
);
2385 new_groupinfo
= ext4_kvzalloc(size
, GFP_KERNEL
);
2386 if (!new_groupinfo
) {
2387 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy meta group");
2390 if (sbi
->s_group_info
) {
2391 memcpy(new_groupinfo
, sbi
->s_group_info
,
2392 sbi
->s_group_info_size
* sizeof(*sbi
->s_group_info
));
2393 kvfree(sbi
->s_group_info
);
2395 sbi
->s_group_info
= new_groupinfo
;
2396 sbi
->s_group_info_size
= size
/ sizeof(*sbi
->s_group_info
);
2397 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2398 sbi
->s_group_info_size
);
2402 /* Create and initialize ext4_group_info data for the given group. */
2403 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2404 struct ext4_group_desc
*desc
)
2408 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2409 struct ext4_group_info
**meta_group_info
;
2410 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2413 * First check if this group is the first of a reserved block.
2414 * If it's true, we have to allocate a new table of pointers
2415 * to ext4_group_info structures
2417 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2418 metalen
= sizeof(*meta_group_info
) <<
2419 EXT4_DESC_PER_BLOCK_BITS(sb
);
2420 meta_group_info
= kmalloc(metalen
, GFP_NOFS
);
2421 if (meta_group_info
== NULL
) {
2422 ext4_msg(sb
, KERN_ERR
, "can't allocate mem "
2423 "for a buddy group");
2424 goto exit_meta_group_info
;
2426 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2431 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2432 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2434 meta_group_info
[i
] = kmem_cache_zalloc(cachep
, GFP_NOFS
);
2435 if (meta_group_info
[i
] == NULL
) {
2436 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy mem");
2437 goto exit_group_info
;
2439 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2440 &(meta_group_info
[i
]->bb_state
));
2443 * initialize bb_free to be able to skip
2444 * empty groups without initialization
2446 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2447 meta_group_info
[i
]->bb_free
=
2448 ext4_free_clusters_after_init(sb
, group
, desc
);
2450 meta_group_info
[i
]->bb_free
=
2451 ext4_free_group_clusters(sb
, desc
);
2454 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2455 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2456 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2457 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2461 struct buffer_head
*bh
;
2462 meta_group_info
[i
]->bb_bitmap
=
2463 kmalloc(sb
->s_blocksize
, GFP_NOFS
);
2464 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2465 bh
= ext4_read_block_bitmap(sb
, group
);
2466 BUG_ON(IS_ERR_OR_NULL(bh
));
2467 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2476 /* If a meta_group_info table has been allocated, release it now */
2477 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2478 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2479 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] = NULL
;
2481 exit_meta_group_info
:
2483 } /* ext4_mb_add_groupinfo */
2485 static int ext4_mb_init_backend(struct super_block
*sb
)
2487 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2489 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2491 struct ext4_group_desc
*desc
;
2492 struct kmem_cache
*cachep
;
2494 err
= ext4_mb_alloc_groupinfo(sb
, ngroups
);
2498 sbi
->s_buddy_cache
= new_inode(sb
);
2499 if (sbi
->s_buddy_cache
== NULL
) {
2500 ext4_msg(sb
, KERN_ERR
, "can't get new inode");
2503 /* To avoid potentially colliding with an valid on-disk inode number,
2504 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2505 * not in the inode hash, so it should never be found by iget(), but
2506 * this will avoid confusion if it ever shows up during debugging. */
2507 sbi
->s_buddy_cache
->i_ino
= EXT4_BAD_INO
;
2508 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2509 for (i
= 0; i
< ngroups
; i
++) {
2510 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2512 ext4_msg(sb
, KERN_ERR
, "can't read descriptor %u", i
);
2515 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2522 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2524 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2525 i
= sbi
->s_group_info_size
;
2527 kfree(sbi
->s_group_info
[i
]);
2528 iput(sbi
->s_buddy_cache
);
2530 kvfree(sbi
->s_group_info
);
2534 static void ext4_groupinfo_destroy_slabs(void)
2538 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2539 if (ext4_groupinfo_caches
[i
])
2540 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2541 ext4_groupinfo_caches
[i
] = NULL
;
2545 static int ext4_groupinfo_create_slab(size_t size
)
2547 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2549 int blocksize_bits
= order_base_2(size
);
2550 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2551 struct kmem_cache
*cachep
;
2553 if (cache_index
>= NR_GRPINFO_CACHES
)
2556 if (unlikely(cache_index
< 0))
2559 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2560 if (ext4_groupinfo_caches
[cache_index
]) {
2561 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2562 return 0; /* Already created */
2565 slab_size
= offsetof(struct ext4_group_info
,
2566 bb_counters
[blocksize_bits
+ 2]);
2568 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2569 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2572 ext4_groupinfo_caches
[cache_index
] = cachep
;
2574 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2577 "EXT4-fs: no memory for groupinfo slab cache\n");
2584 int ext4_mb_init(struct super_block
*sb
)
2586 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2588 unsigned offset
, offset_incr
;
2592 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2594 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2595 if (sbi
->s_mb_offsets
== NULL
) {
2600 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2601 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2602 if (sbi
->s_mb_maxs
== NULL
) {
2607 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2611 /* order 0 is regular bitmap */
2612 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2613 sbi
->s_mb_offsets
[0] = 0;
2617 offset_incr
= 1 << (sb
->s_blocksize_bits
- 1);
2618 max
= sb
->s_blocksize
<< 2;
2620 sbi
->s_mb_offsets
[i
] = offset
;
2621 sbi
->s_mb_maxs
[i
] = max
;
2622 offset
+= offset_incr
;
2623 offset_incr
= offset_incr
>> 1;
2626 } while (i
<= sb
->s_blocksize_bits
+ 1);
2628 spin_lock_init(&sbi
->s_md_lock
);
2629 spin_lock_init(&sbi
->s_bal_lock
);
2630 sbi
->s_mb_free_pending
= 0;
2632 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2633 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2634 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2635 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2636 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2638 * The default group preallocation is 512, which for 4k block
2639 * sizes translates to 2 megabytes. However for bigalloc file
2640 * systems, this is probably too big (i.e, if the cluster size
2641 * is 1 megabyte, then group preallocation size becomes half a
2642 * gigabyte!). As a default, we will keep a two megabyte
2643 * group pralloc size for cluster sizes up to 64k, and after
2644 * that, we will force a minimum group preallocation size of
2645 * 32 clusters. This translates to 8 megs when the cluster
2646 * size is 256k, and 32 megs when the cluster size is 1 meg,
2647 * which seems reasonable as a default.
2649 sbi
->s_mb_group_prealloc
= max(MB_DEFAULT_GROUP_PREALLOC
>>
2650 sbi
->s_cluster_bits
, 32);
2652 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2653 * to the lowest multiple of s_stripe which is bigger than
2654 * the s_mb_group_prealloc as determined above. We want
2655 * the preallocation size to be an exact multiple of the
2656 * RAID stripe size so that preallocations don't fragment
2659 if (sbi
->s_stripe
> 1) {
2660 sbi
->s_mb_group_prealloc
= roundup(
2661 sbi
->s_mb_group_prealloc
, sbi
->s_stripe
);
2664 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2665 if (sbi
->s_locality_groups
== NULL
) {
2669 for_each_possible_cpu(i
) {
2670 struct ext4_locality_group
*lg
;
2671 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2672 mutex_init(&lg
->lg_mutex
);
2673 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2674 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2675 spin_lock_init(&lg
->lg_prealloc_lock
);
2678 /* init file for buddy data */
2679 ret
= ext4_mb_init_backend(sb
);
2681 goto out_free_locality_groups
;
2685 out_free_locality_groups
:
2686 free_percpu(sbi
->s_locality_groups
);
2687 sbi
->s_locality_groups
= NULL
;
2689 kfree(sbi
->s_mb_offsets
);
2690 sbi
->s_mb_offsets
= NULL
;
2691 kfree(sbi
->s_mb_maxs
);
2692 sbi
->s_mb_maxs
= NULL
;
2696 /* need to called with the ext4 group lock held */
2697 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2699 struct ext4_prealloc_space
*pa
;
2700 struct list_head
*cur
, *tmp
;
2703 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2704 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2705 list_del(&pa
->pa_group_list
);
2707 kmem_cache_free(ext4_pspace_cachep
, pa
);
2710 mb_debug(1, "mballoc: %u PAs left\n", count
);
2714 int ext4_mb_release(struct super_block
*sb
)
2716 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2718 int num_meta_group_infos
;
2719 struct ext4_group_info
*grinfo
;
2720 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2721 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2723 if (sbi
->s_group_info
) {
2724 for (i
= 0; i
< ngroups
; i
++) {
2725 grinfo
= ext4_get_group_info(sb
, i
);
2727 kfree(grinfo
->bb_bitmap
);
2729 ext4_lock_group(sb
, i
);
2730 ext4_mb_cleanup_pa(grinfo
);
2731 ext4_unlock_group(sb
, i
);
2732 kmem_cache_free(cachep
, grinfo
);
2734 num_meta_group_infos
= (ngroups
+
2735 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2736 EXT4_DESC_PER_BLOCK_BITS(sb
);
2737 for (i
= 0; i
< num_meta_group_infos
; i
++)
2738 kfree(sbi
->s_group_info
[i
]);
2739 kvfree(sbi
->s_group_info
);
2741 kfree(sbi
->s_mb_offsets
);
2742 kfree(sbi
->s_mb_maxs
);
2743 iput(sbi
->s_buddy_cache
);
2744 if (sbi
->s_mb_stats
) {
2745 ext4_msg(sb
, KERN_INFO
,
2746 "mballoc: %u blocks %u reqs (%u success)",
2747 atomic_read(&sbi
->s_bal_allocated
),
2748 atomic_read(&sbi
->s_bal_reqs
),
2749 atomic_read(&sbi
->s_bal_success
));
2750 ext4_msg(sb
, KERN_INFO
,
2751 "mballoc: %u extents scanned, %u goal hits, "
2752 "%u 2^N hits, %u breaks, %u lost",
2753 atomic_read(&sbi
->s_bal_ex_scanned
),
2754 atomic_read(&sbi
->s_bal_goals
),
2755 atomic_read(&sbi
->s_bal_2orders
),
2756 atomic_read(&sbi
->s_bal_breaks
),
2757 atomic_read(&sbi
->s_mb_lost_chunks
));
2758 ext4_msg(sb
, KERN_INFO
,
2759 "mballoc: %lu generated and it took %Lu",
2760 sbi
->s_mb_buddies_generated
,
2761 sbi
->s_mb_generation_time
);
2762 ext4_msg(sb
, KERN_INFO
,
2763 "mballoc: %u preallocated, %u discarded",
2764 atomic_read(&sbi
->s_mb_preallocated
),
2765 atomic_read(&sbi
->s_mb_discarded
));
2768 free_percpu(sbi
->s_locality_groups
);
2773 static inline int ext4_issue_discard(struct super_block
*sb
,
2774 ext4_group_t block_group
, ext4_grpblk_t cluster
, int count
)
2776 ext4_fsblk_t discard_block
;
2778 discard_block
= (EXT4_C2B(EXT4_SB(sb
), cluster
) +
2779 ext4_group_first_block_no(sb
, block_group
));
2780 count
= EXT4_C2B(EXT4_SB(sb
), count
);
2781 trace_ext4_discard_blocks(sb
,
2782 (unsigned long long) discard_block
, count
);
2783 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, 0);
2787 * This function is called by the jbd2 layer once the commit has finished,
2788 * so we know we can free the blocks that were released with that commit.
2790 static void ext4_free_data_callback(struct super_block
*sb
,
2791 struct ext4_journal_cb_entry
*jce
,
2794 struct ext4_free_data
*entry
= (struct ext4_free_data
*)jce
;
2795 struct ext4_buddy e4b
;
2796 struct ext4_group_info
*db
;
2797 int err
, count
= 0, count2
= 0;
2799 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2800 entry
->efd_count
, entry
->efd_group
, entry
);
2802 if (test_opt(sb
, DISCARD
)) {
2803 err
= ext4_issue_discard(sb
, entry
->efd_group
,
2804 entry
->efd_start_cluster
,
2806 if (err
&& err
!= -EOPNOTSUPP
)
2807 ext4_msg(sb
, KERN_WARNING
, "discard request in"
2808 " group:%d block:%d count:%d failed"
2809 " with %d", entry
->efd_group
,
2810 entry
->efd_start_cluster
,
2811 entry
->efd_count
, err
);
2814 err
= ext4_mb_load_buddy(sb
, entry
->efd_group
, &e4b
);
2815 /* we expect to find existing buddy because it's pinned */
2818 spin_lock(&EXT4_SB(sb
)->s_md_lock
);
2819 EXT4_SB(sb
)->s_mb_free_pending
-= entry
->efd_count
;
2820 spin_unlock(&EXT4_SB(sb
)->s_md_lock
);
2823 /* there are blocks to put in buddy to make them really free */
2824 count
+= entry
->efd_count
;
2826 ext4_lock_group(sb
, entry
->efd_group
);
2827 /* Take it out of per group rb tree */
2828 rb_erase(&entry
->efd_node
, &(db
->bb_free_root
));
2829 mb_free_blocks(NULL
, &e4b
, entry
->efd_start_cluster
, entry
->efd_count
);
2832 * Clear the trimmed flag for the group so that the next
2833 * ext4_trim_fs can trim it.
2834 * If the volume is mounted with -o discard, online discard
2835 * is supported and the free blocks will be trimmed online.
2837 if (!test_opt(sb
, DISCARD
))
2838 EXT4_MB_GRP_CLEAR_TRIMMED(db
);
2840 if (!db
->bb_free_root
.rb_node
) {
2841 /* No more items in the per group rb tree
2842 * balance refcounts from ext4_mb_free_metadata()
2844 put_page(e4b
.bd_buddy_page
);
2845 put_page(e4b
.bd_bitmap_page
);
2847 ext4_unlock_group(sb
, entry
->efd_group
);
2848 kmem_cache_free(ext4_free_data_cachep
, entry
);
2849 ext4_mb_unload_buddy(&e4b
);
2851 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2854 int __init
ext4_init_mballoc(void)
2856 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2857 SLAB_RECLAIM_ACCOUNT
);
2858 if (ext4_pspace_cachep
== NULL
)
2861 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2862 SLAB_RECLAIM_ACCOUNT
);
2863 if (ext4_ac_cachep
== NULL
) {
2864 kmem_cache_destroy(ext4_pspace_cachep
);
2868 ext4_free_data_cachep
= KMEM_CACHE(ext4_free_data
,
2869 SLAB_RECLAIM_ACCOUNT
);
2870 if (ext4_free_data_cachep
== NULL
) {
2871 kmem_cache_destroy(ext4_pspace_cachep
);
2872 kmem_cache_destroy(ext4_ac_cachep
);
2878 void ext4_exit_mballoc(void)
2881 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2882 * before destroying the slab cache.
2885 kmem_cache_destroy(ext4_pspace_cachep
);
2886 kmem_cache_destroy(ext4_ac_cachep
);
2887 kmem_cache_destroy(ext4_free_data_cachep
);
2888 ext4_groupinfo_destroy_slabs();
2893 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2894 * Returns 0 if success or error code
2896 static noinline_for_stack
int
2897 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2898 handle_t
*handle
, unsigned int reserv_clstrs
)
2900 struct buffer_head
*bitmap_bh
= NULL
;
2901 struct ext4_group_desc
*gdp
;
2902 struct buffer_head
*gdp_bh
;
2903 struct ext4_sb_info
*sbi
;
2904 struct super_block
*sb
;
2908 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2909 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2914 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2915 if (IS_ERR(bitmap_bh
)) {
2916 err
= PTR_ERR(bitmap_bh
);
2921 BUFFER_TRACE(bitmap_bh
, "getting write access");
2922 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2927 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2931 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2932 ext4_free_group_clusters(sb
, gdp
));
2934 BUFFER_TRACE(gdp_bh
, "get_write_access");
2935 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2939 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
2941 len
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
2942 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2943 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2944 "fs metadata", block
, block
+len
);
2945 /* File system mounted not to panic on error
2946 * Fix the bitmap and return EFSCORRUPTED
2947 * We leak some of the blocks here.
2949 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2950 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2951 ac
->ac_b_ex
.fe_len
);
2952 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2953 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2955 err
= -EFSCORRUPTED
;
2959 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2960 #ifdef AGGRESSIVE_CHECK
2963 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2964 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2965 bitmap_bh
->b_data
));
2969 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2970 ac
->ac_b_ex
.fe_len
);
2971 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2972 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2973 ext4_free_group_clusters_set(sb
, gdp
,
2974 ext4_free_clusters_after_init(sb
,
2975 ac
->ac_b_ex
.fe_group
, gdp
));
2977 len
= ext4_free_group_clusters(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
2978 ext4_free_group_clusters_set(sb
, gdp
, len
);
2979 ext4_block_bitmap_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
, bitmap_bh
);
2980 ext4_group_desc_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
);
2982 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2983 percpu_counter_sub(&sbi
->s_freeclusters_counter
, ac
->ac_b_ex
.fe_len
);
2985 * Now reduce the dirty block count also. Should not go negative
2987 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2988 /* release all the reserved blocks if non delalloc */
2989 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
2992 if (sbi
->s_log_groups_per_flex
) {
2993 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2994 ac
->ac_b_ex
.fe_group
);
2995 atomic64_sub(ac
->ac_b_ex
.fe_len
,
2996 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
2999 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3002 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
3010 * here we normalize request for locality group
3011 * Group request are normalized to s_mb_group_prealloc, which goes to
3012 * s_strip if we set the same via mount option.
3013 * s_mb_group_prealloc can be configured via
3014 * /sys/fs/ext4/<partition>/mb_group_prealloc
3016 * XXX: should we try to preallocate more than the group has now?
3018 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
3020 struct super_block
*sb
= ac
->ac_sb
;
3021 struct ext4_locality_group
*lg
= ac
->ac_lg
;
3024 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
3025 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3026 current
->pid
, ac
->ac_g_ex
.fe_len
);
3030 * Normalization means making request better in terms of
3031 * size and alignment
3033 static noinline_for_stack
void
3034 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
3035 struct ext4_allocation_request
*ar
)
3037 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3040 loff_t size
, start_off
;
3041 loff_t orig_size __maybe_unused
;
3043 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3044 struct ext4_prealloc_space
*pa
;
3046 /* do normalize only data requests, metadata requests
3047 do not need preallocation */
3048 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3051 /* sometime caller may want exact blocks */
3052 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3055 /* caller may indicate that preallocation isn't
3056 * required (it's a tail, for example) */
3057 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3060 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3061 ext4_mb_normalize_group_request(ac
);
3065 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3067 /* first, let's learn actual file size
3068 * given current request is allocated */
3069 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
3070 size
= size
<< bsbits
;
3071 if (size
< i_size_read(ac
->ac_inode
))
3072 size
= i_size_read(ac
->ac_inode
);
3075 /* max size of free chunks */
3078 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3079 (req <= (size) || max <= (chunk_size))
3081 /* first, try to predict filesize */
3082 /* XXX: should this table be tunable? */
3084 if (size
<= 16 * 1024) {
3086 } else if (size
<= 32 * 1024) {
3088 } else if (size
<= 64 * 1024) {
3090 } else if (size
<= 128 * 1024) {
3092 } else if (size
<= 256 * 1024) {
3094 } else if (size
<= 512 * 1024) {
3096 } else if (size
<= 1024 * 1024) {
3098 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3099 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3100 (21 - bsbits
)) << 21;
3101 size
= 2 * 1024 * 1024;
3102 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3103 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3104 (22 - bsbits
)) << 22;
3105 size
= 4 * 1024 * 1024;
3106 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3107 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3108 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3109 (23 - bsbits
)) << 23;
3110 size
= 8 * 1024 * 1024;
3112 start_off
= (loff_t
) ac
->ac_o_ex
.fe_logical
<< bsbits
;
3113 size
= (loff_t
) EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3114 ac
->ac_o_ex
.fe_len
) << bsbits
;
3116 size
= size
>> bsbits
;
3117 start
= start_off
>> bsbits
;
3119 /* don't cover already allocated blocks in selected range */
3120 if (ar
->pleft
&& start
<= ar
->lleft
) {
3121 size
-= ar
->lleft
+ 1 - start
;
3122 start
= ar
->lleft
+ 1;
3124 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3125 size
-= start
+ size
- ar
->lright
;
3129 /* check we don't cross already preallocated blocks */
3131 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3136 spin_lock(&pa
->pa_lock
);
3137 if (pa
->pa_deleted
) {
3138 spin_unlock(&pa
->pa_lock
);
3142 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3145 /* PA must not overlap original request */
3146 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3147 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3149 /* skip PAs this normalized request doesn't overlap with */
3150 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
3151 spin_unlock(&pa
->pa_lock
);
3154 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3156 /* adjust start or end to be adjacent to this pa */
3157 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3158 BUG_ON(pa_end
< start
);
3160 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3161 BUG_ON(pa
->pa_lstart
> end
);
3162 end
= pa
->pa_lstart
;
3164 spin_unlock(&pa
->pa_lock
);
3169 /* XXX: extra loop to check we really don't overlap preallocations */
3171 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3174 spin_lock(&pa
->pa_lock
);
3175 if (pa
->pa_deleted
== 0) {
3176 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3178 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3180 spin_unlock(&pa
->pa_lock
);
3184 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3185 start
> ac
->ac_o_ex
.fe_logical
) {
3186 ext4_msg(ac
->ac_sb
, KERN_ERR
,
3187 "start %lu, size %lu, fe_logical %lu",
3188 (unsigned long) start
, (unsigned long) size
,
3189 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3192 BUG_ON(size
<= 0 || size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3194 /* now prepare goal request */
3196 /* XXX: is it better to align blocks WRT to logical
3197 * placement or satisfy big request as is */
3198 ac
->ac_g_ex
.fe_logical
= start
;
3199 ac
->ac_g_ex
.fe_len
= EXT4_NUM_B2C(sbi
, size
);
3201 /* define goal start in order to merge */
3202 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3203 /* merge to the right */
3204 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3205 &ac
->ac_f_ex
.fe_group
,
3206 &ac
->ac_f_ex
.fe_start
);
3207 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3209 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3210 /* merge to the left */
3211 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3212 &ac
->ac_f_ex
.fe_group
,
3213 &ac
->ac_f_ex
.fe_start
);
3214 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3217 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3218 (unsigned) orig_size
, (unsigned) start
);
3221 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3223 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3225 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3226 atomic_inc(&sbi
->s_bal_reqs
);
3227 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3228 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3229 atomic_inc(&sbi
->s_bal_success
);
3230 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3231 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3232 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3233 atomic_inc(&sbi
->s_bal_goals
);
3234 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3235 atomic_inc(&sbi
->s_bal_breaks
);
3238 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3239 trace_ext4_mballoc_alloc(ac
);
3241 trace_ext4_mballoc_prealloc(ac
);
3245 * Called on failure; free up any blocks from the inode PA for this
3246 * context. We don't need this for MB_GROUP_PA because we only change
3247 * pa_free in ext4_mb_release_context(), but on failure, we've already
3248 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3250 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3252 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3253 struct ext4_buddy e4b
;
3257 if (ac
->ac_f_ex
.fe_len
== 0)
3259 err
= ext4_mb_load_buddy(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
, &e4b
);
3262 * This should never happen since we pin the
3263 * pages in the ext4_allocation_context so
3264 * ext4_mb_load_buddy() should never fail.
3266 WARN(1, "mb_load_buddy failed (%d)", err
);
3269 ext4_lock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3270 mb_free_blocks(ac
->ac_inode
, &e4b
, ac
->ac_f_ex
.fe_start
,
3271 ac
->ac_f_ex
.fe_len
);
3272 ext4_unlock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3273 ext4_mb_unload_buddy(&e4b
);
3276 if (pa
->pa_type
== MB_INODE_PA
)
3277 pa
->pa_free
+= ac
->ac_b_ex
.fe_len
;
3281 * use blocks preallocated to inode
3283 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3284 struct ext4_prealloc_space
*pa
)
3286 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3291 /* found preallocated blocks, use them */
3292 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3293 end
= min(pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
),
3294 start
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
));
3295 len
= EXT4_NUM_B2C(sbi
, end
- start
);
3296 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3297 &ac
->ac_b_ex
.fe_start
);
3298 ac
->ac_b_ex
.fe_len
= len
;
3299 ac
->ac_status
= AC_STATUS_FOUND
;
3302 BUG_ON(start
< pa
->pa_pstart
);
3303 BUG_ON(end
> pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
));
3304 BUG_ON(pa
->pa_free
< len
);
3307 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3311 * use blocks preallocated to locality group
3313 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3314 struct ext4_prealloc_space
*pa
)
3316 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3318 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3319 &ac
->ac_b_ex
.fe_group
,
3320 &ac
->ac_b_ex
.fe_start
);
3321 ac
->ac_b_ex
.fe_len
= len
;
3322 ac
->ac_status
= AC_STATUS_FOUND
;
3325 /* we don't correct pa_pstart or pa_plen here to avoid
3326 * possible race when the group is being loaded concurrently
3327 * instead we correct pa later, after blocks are marked
3328 * in on-disk bitmap -- see ext4_mb_release_context()
3329 * Other CPUs are prevented from allocating from this pa by lg_mutex
3331 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3335 * Return the prealloc space that have minimal distance
3336 * from the goal block. @cpa is the prealloc
3337 * space that is having currently known minimal distance
3338 * from the goal block.
3340 static struct ext4_prealloc_space
*
3341 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3342 struct ext4_prealloc_space
*pa
,
3343 struct ext4_prealloc_space
*cpa
)
3345 ext4_fsblk_t cur_distance
, new_distance
;
3348 atomic_inc(&pa
->pa_count
);
3351 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3352 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3354 if (cur_distance
<= new_distance
)
3357 /* drop the previous reference */
3358 atomic_dec(&cpa
->pa_count
);
3359 atomic_inc(&pa
->pa_count
);
3364 * search goal blocks in preallocated space
3366 static noinline_for_stack
int
3367 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3369 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3371 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3372 struct ext4_locality_group
*lg
;
3373 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3374 ext4_fsblk_t goal_block
;
3376 /* only data can be preallocated */
3377 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3380 /* first, try per-file preallocation */
3382 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3384 /* all fields in this condition don't change,
3385 * so we can skip locking for them */
3386 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3387 ac
->ac_o_ex
.fe_logical
>= (pa
->pa_lstart
+
3388 EXT4_C2B(sbi
, pa
->pa_len
)))
3391 /* non-extent files can't have physical blocks past 2^32 */
3392 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3393 (pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
) >
3394 EXT4_MAX_BLOCK_FILE_PHYS
))
3397 /* found preallocated blocks, use them */
3398 spin_lock(&pa
->pa_lock
);
3399 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3400 atomic_inc(&pa
->pa_count
);
3401 ext4_mb_use_inode_pa(ac
, pa
);
3402 spin_unlock(&pa
->pa_lock
);
3403 ac
->ac_criteria
= 10;
3407 spin_unlock(&pa
->pa_lock
);
3411 /* can we use group allocation? */
3412 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3415 /* inode may have no locality group for some reason */
3419 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3420 if (order
> PREALLOC_TB_SIZE
- 1)
3421 /* The max size of hash table is PREALLOC_TB_SIZE */
3422 order
= PREALLOC_TB_SIZE
- 1;
3424 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3426 * search for the prealloc space that is having
3427 * minimal distance from the goal block.
3429 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3431 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3433 spin_lock(&pa
->pa_lock
);
3434 if (pa
->pa_deleted
== 0 &&
3435 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3437 cpa
= ext4_mb_check_group_pa(goal_block
,
3440 spin_unlock(&pa
->pa_lock
);
3445 ext4_mb_use_group_pa(ac
, cpa
);
3446 ac
->ac_criteria
= 20;
3453 * the function goes through all block freed in the group
3454 * but not yet committed and marks them used in in-core bitmap.
3455 * buddy must be generated from this bitmap
3456 * Need to be called with the ext4 group lock held
3458 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3462 struct ext4_group_info
*grp
;
3463 struct ext4_free_data
*entry
;
3465 grp
= ext4_get_group_info(sb
, group
);
3466 n
= rb_first(&(grp
->bb_free_root
));
3469 entry
= rb_entry(n
, struct ext4_free_data
, efd_node
);
3470 ext4_set_bits(bitmap
, entry
->efd_start_cluster
, entry
->efd_count
);
3477 * the function goes through all preallocation in this group and marks them
3478 * used in in-core bitmap. buddy must be generated from this bitmap
3479 * Need to be called with ext4 group lock held
3481 static noinline_for_stack
3482 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3485 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3486 struct ext4_prealloc_space
*pa
;
3487 struct list_head
*cur
;
3488 ext4_group_t groupnr
;
3489 ext4_grpblk_t start
;
3490 int preallocated
= 0;
3493 /* all form of preallocation discards first load group,
3494 * so the only competing code is preallocation use.
3495 * we don't need any locking here
3496 * notice we do NOT ignore preallocations with pa_deleted
3497 * otherwise we could leave used blocks available for
3498 * allocation in buddy when concurrent ext4_mb_put_pa()
3499 * is dropping preallocation
3501 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3502 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3503 spin_lock(&pa
->pa_lock
);
3504 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3507 spin_unlock(&pa
->pa_lock
);
3508 if (unlikely(len
== 0))
3510 BUG_ON(groupnr
!= group
);
3511 ext4_set_bits(bitmap
, start
, len
);
3512 preallocated
+= len
;
3514 mb_debug(1, "prellocated %u for group %u\n", preallocated
, group
);
3517 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3519 struct ext4_prealloc_space
*pa
;
3520 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3522 BUG_ON(atomic_read(&pa
->pa_count
));
3523 BUG_ON(pa
->pa_deleted
== 0);
3524 kmem_cache_free(ext4_pspace_cachep
, pa
);
3528 * drops a reference to preallocated space descriptor
3529 * if this was the last reference and the space is consumed
3531 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3532 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3535 ext4_fsblk_t grp_blk
;
3537 /* in this short window concurrent discard can set pa_deleted */
3538 spin_lock(&pa
->pa_lock
);
3539 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0) {
3540 spin_unlock(&pa
->pa_lock
);
3544 if (pa
->pa_deleted
== 1) {
3545 spin_unlock(&pa
->pa_lock
);
3550 spin_unlock(&pa
->pa_lock
);
3552 grp_blk
= pa
->pa_pstart
;
3554 * If doing group-based preallocation, pa_pstart may be in the
3555 * next group when pa is used up
3557 if (pa
->pa_type
== MB_GROUP_PA
)
3560 grp
= ext4_get_group_number(sb
, grp_blk
);
3565 * P1 (buddy init) P2 (regular allocation)
3566 * find block B in PA
3567 * copy on-disk bitmap to buddy
3568 * mark B in on-disk bitmap
3569 * drop PA from group
3570 * mark all PAs in buddy
3572 * thus, P1 initializes buddy with B available. to prevent this
3573 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3576 ext4_lock_group(sb
, grp
);
3577 list_del(&pa
->pa_group_list
);
3578 ext4_unlock_group(sb
, grp
);
3580 spin_lock(pa
->pa_obj_lock
);
3581 list_del_rcu(&pa
->pa_inode_list
);
3582 spin_unlock(pa
->pa_obj_lock
);
3584 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3588 * creates new preallocated space for given inode
3590 static noinline_for_stack
int
3591 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3593 struct super_block
*sb
= ac
->ac_sb
;
3594 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3595 struct ext4_prealloc_space
*pa
;
3596 struct ext4_group_info
*grp
;
3597 struct ext4_inode_info
*ei
;
3599 /* preallocate only when found space is larger then requested */
3600 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3601 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3602 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3604 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3608 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3614 /* we can't allocate as much as normalizer wants.
3615 * so, found space must get proper lstart
3616 * to cover original request */
3617 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3618 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3620 /* we're limited by original request in that
3621 * logical block must be covered any way
3622 * winl is window we can move our chunk within */
3623 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3625 /* also, we should cover whole original request */
3626 wins
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
);
3628 /* the smallest one defines real window */
3629 win
= min(winl
, wins
);
3631 offs
= ac
->ac_o_ex
.fe_logical
%
3632 EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
3633 if (offs
&& offs
< win
)
3636 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
-
3637 EXT4_NUM_B2C(sbi
, win
);
3638 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3639 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3642 /* preallocation can change ac_b_ex, thus we store actually
3643 * allocated blocks for history */
3644 ac
->ac_f_ex
= ac
->ac_b_ex
;
3646 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3647 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3648 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3649 pa
->pa_free
= pa
->pa_len
;
3650 atomic_set(&pa
->pa_count
, 1);
3651 spin_lock_init(&pa
->pa_lock
);
3652 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3653 INIT_LIST_HEAD(&pa
->pa_group_list
);
3655 pa
->pa_type
= MB_INODE_PA
;
3657 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3658 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3659 trace_ext4_mb_new_inode_pa(ac
, pa
);
3661 ext4_mb_use_inode_pa(ac
, pa
);
3662 atomic_add(pa
->pa_free
, &sbi
->s_mb_preallocated
);
3664 ei
= EXT4_I(ac
->ac_inode
);
3665 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3667 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3668 pa
->pa_inode
= ac
->ac_inode
;
3670 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3671 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3672 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3674 spin_lock(pa
->pa_obj_lock
);
3675 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3676 spin_unlock(pa
->pa_obj_lock
);
3682 * creates new preallocated space for locality group inodes belongs to
3684 static noinline_for_stack
int
3685 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3687 struct super_block
*sb
= ac
->ac_sb
;
3688 struct ext4_locality_group
*lg
;
3689 struct ext4_prealloc_space
*pa
;
3690 struct ext4_group_info
*grp
;
3692 /* preallocate only when found space is larger then requested */
3693 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3694 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3695 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3697 BUG_ON(ext4_pspace_cachep
== NULL
);
3698 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3702 /* preallocation can change ac_b_ex, thus we store actually
3703 * allocated blocks for history */
3704 ac
->ac_f_ex
= ac
->ac_b_ex
;
3706 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3707 pa
->pa_lstart
= pa
->pa_pstart
;
3708 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3709 pa
->pa_free
= pa
->pa_len
;
3710 atomic_set(&pa
->pa_count
, 1);
3711 spin_lock_init(&pa
->pa_lock
);
3712 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3713 INIT_LIST_HEAD(&pa
->pa_group_list
);
3715 pa
->pa_type
= MB_GROUP_PA
;
3717 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3718 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3719 trace_ext4_mb_new_group_pa(ac
, pa
);
3721 ext4_mb_use_group_pa(ac
, pa
);
3722 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3724 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3728 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3729 pa
->pa_inode
= NULL
;
3731 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3732 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3733 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3736 * We will later add the new pa to the right bucket
3737 * after updating the pa_free in ext4_mb_release_context
3742 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3746 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3747 err
= ext4_mb_new_group_pa(ac
);
3749 err
= ext4_mb_new_inode_pa(ac
);
3754 * finds all unused blocks in on-disk bitmap, frees them in
3755 * in-core bitmap and buddy.
3756 * @pa must be unlinked from inode and group lists, so that
3757 * nobody else can find/use it.
3758 * the caller MUST hold group/inode locks.
3759 * TODO: optimize the case when there are no in-core structures yet
3761 static noinline_for_stack
int
3762 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3763 struct ext4_prealloc_space
*pa
)
3765 struct super_block
*sb
= e4b
->bd_sb
;
3766 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3771 unsigned long long grp_blk_start
;
3775 BUG_ON(pa
->pa_deleted
== 0);
3776 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3777 grp_blk_start
= pa
->pa_pstart
- EXT4_C2B(sbi
, bit
);
3778 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3779 end
= bit
+ pa
->pa_len
;
3782 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3785 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3786 mb_debug(1, " free preallocated %u/%u in group %u\n",
3787 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3788 (unsigned) next
- bit
, (unsigned) group
);
3791 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3792 trace_ext4_mb_release_inode_pa(pa
, (grp_blk_start
+
3793 EXT4_C2B(sbi
, bit
)),
3795 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3798 if (free
!= pa
->pa_free
) {
3799 ext4_msg(e4b
->bd_sb
, KERN_CRIT
,
3800 "pa %p: logic %lu, phys. %lu, len %lu",
3801 pa
, (unsigned long) pa
->pa_lstart
,
3802 (unsigned long) pa
->pa_pstart
,
3803 (unsigned long) pa
->pa_len
);
3804 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3807 * pa is already deleted so we use the value obtained
3808 * from the bitmap and continue.
3811 atomic_add(free
, &sbi
->s_mb_discarded
);
3816 static noinline_for_stack
int
3817 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3818 struct ext4_prealloc_space
*pa
)
3820 struct super_block
*sb
= e4b
->bd_sb
;
3824 trace_ext4_mb_release_group_pa(sb
, pa
);
3825 BUG_ON(pa
->pa_deleted
== 0);
3826 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3827 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3828 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3829 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3830 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3836 * releases all preallocations in given group
3838 * first, we need to decide discard policy:
3839 * - when do we discard
3841 * - how many do we discard
3842 * 1) how many requested
3844 static noinline_for_stack
int
3845 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3846 ext4_group_t group
, int needed
)
3848 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3849 struct buffer_head
*bitmap_bh
= NULL
;
3850 struct ext4_prealloc_space
*pa
, *tmp
;
3851 struct list_head list
;
3852 struct ext4_buddy e4b
;
3857 mb_debug(1, "discard preallocation for group %u\n", group
);
3859 if (list_empty(&grp
->bb_prealloc_list
))
3862 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3863 if (IS_ERR(bitmap_bh
)) {
3864 err
= PTR_ERR(bitmap_bh
);
3865 ext4_error(sb
, "Error %d reading block bitmap for %u",
3870 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3872 ext4_error(sb
, "Error loading buddy information for %u", group
);
3878 needed
= EXT4_CLUSTERS_PER_GROUP(sb
) + 1;
3880 INIT_LIST_HEAD(&list
);
3882 ext4_lock_group(sb
, group
);
3883 list_for_each_entry_safe(pa
, tmp
,
3884 &grp
->bb_prealloc_list
, pa_group_list
) {
3885 spin_lock(&pa
->pa_lock
);
3886 if (atomic_read(&pa
->pa_count
)) {
3887 spin_unlock(&pa
->pa_lock
);
3891 if (pa
->pa_deleted
) {
3892 spin_unlock(&pa
->pa_lock
);
3896 /* seems this one can be freed ... */
3899 /* we can trust pa_free ... */
3900 free
+= pa
->pa_free
;
3902 spin_unlock(&pa
->pa_lock
);
3904 list_del(&pa
->pa_group_list
);
3905 list_add(&pa
->u
.pa_tmp_list
, &list
);
3908 /* if we still need more blocks and some PAs were used, try again */
3909 if (free
< needed
&& busy
) {
3911 ext4_unlock_group(sb
, group
);
3916 /* found anything to free? */
3917 if (list_empty(&list
)) {
3922 /* now free all selected PAs */
3923 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3925 /* remove from object (inode or locality group) */
3926 spin_lock(pa
->pa_obj_lock
);
3927 list_del_rcu(&pa
->pa_inode_list
);
3928 spin_unlock(pa
->pa_obj_lock
);
3930 if (pa
->pa_type
== MB_GROUP_PA
)
3931 ext4_mb_release_group_pa(&e4b
, pa
);
3933 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3935 list_del(&pa
->u
.pa_tmp_list
);
3936 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3940 ext4_unlock_group(sb
, group
);
3941 ext4_mb_unload_buddy(&e4b
);
3947 * releases all non-used preallocated blocks for given inode
3949 * It's important to discard preallocations under i_data_sem
3950 * We don't want another block to be served from the prealloc
3951 * space when we are discarding the inode prealloc space.
3953 * FIXME!! Make sure it is valid at all the call sites
3955 void ext4_discard_preallocations(struct inode
*inode
)
3957 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3958 struct super_block
*sb
= inode
->i_sb
;
3959 struct buffer_head
*bitmap_bh
= NULL
;
3960 struct ext4_prealloc_space
*pa
, *tmp
;
3961 ext4_group_t group
= 0;
3962 struct list_head list
;
3963 struct ext4_buddy e4b
;
3966 if (!S_ISREG(inode
->i_mode
)) {
3967 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3971 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
3972 trace_ext4_discard_preallocations(inode
);
3974 INIT_LIST_HEAD(&list
);
3977 /* first, collect all pa's in the inode */
3978 spin_lock(&ei
->i_prealloc_lock
);
3979 while (!list_empty(&ei
->i_prealloc_list
)) {
3980 pa
= list_entry(ei
->i_prealloc_list
.next
,
3981 struct ext4_prealloc_space
, pa_inode_list
);
3982 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3983 spin_lock(&pa
->pa_lock
);
3984 if (atomic_read(&pa
->pa_count
)) {
3985 /* this shouldn't happen often - nobody should
3986 * use preallocation while we're discarding it */
3987 spin_unlock(&pa
->pa_lock
);
3988 spin_unlock(&ei
->i_prealloc_lock
);
3989 ext4_msg(sb
, KERN_ERR
,
3990 "uh-oh! used pa while discarding");
3992 schedule_timeout_uninterruptible(HZ
);
3996 if (pa
->pa_deleted
== 0) {
3998 spin_unlock(&pa
->pa_lock
);
3999 list_del_rcu(&pa
->pa_inode_list
);
4000 list_add(&pa
->u
.pa_tmp_list
, &list
);
4004 /* someone is deleting pa right now */
4005 spin_unlock(&pa
->pa_lock
);
4006 spin_unlock(&ei
->i_prealloc_lock
);
4008 /* we have to wait here because pa_deleted
4009 * doesn't mean pa is already unlinked from
4010 * the list. as we might be called from
4011 * ->clear_inode() the inode will get freed
4012 * and concurrent thread which is unlinking
4013 * pa from inode's list may access already
4014 * freed memory, bad-bad-bad */
4016 /* XXX: if this happens too often, we can
4017 * add a flag to force wait only in case
4018 * of ->clear_inode(), but not in case of
4019 * regular truncate */
4020 schedule_timeout_uninterruptible(HZ
);
4023 spin_unlock(&ei
->i_prealloc_lock
);
4025 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
4026 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
4027 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4029 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
4031 ext4_error(sb
, "Error loading buddy information for %u",
4036 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
4037 if (IS_ERR(bitmap_bh
)) {
4038 err
= PTR_ERR(bitmap_bh
);
4039 ext4_error(sb
, "Error %d reading block bitmap for %u",
4041 ext4_mb_unload_buddy(&e4b
);
4045 ext4_lock_group(sb
, group
);
4046 list_del(&pa
->pa_group_list
);
4047 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
4048 ext4_unlock_group(sb
, group
);
4050 ext4_mb_unload_buddy(&e4b
);
4053 list_del(&pa
->u
.pa_tmp_list
);
4054 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4058 #ifdef CONFIG_EXT4_DEBUG
4059 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4061 struct super_block
*sb
= ac
->ac_sb
;
4062 ext4_group_t ngroups
, i
;
4064 if (!ext4_mballoc_debug
||
4065 (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
))
4068 ext4_msg(ac
->ac_sb
, KERN_ERR
, "Can't allocate:"
4069 " Allocation context details:");
4070 ext4_msg(ac
->ac_sb
, KERN_ERR
, "status %d flags %d",
4071 ac
->ac_status
, ac
->ac_flags
);
4072 ext4_msg(ac
->ac_sb
, KERN_ERR
, "orig %lu/%lu/%lu@%lu, "
4073 "goal %lu/%lu/%lu@%lu, "
4074 "best %lu/%lu/%lu@%lu cr %d",
4075 (unsigned long)ac
->ac_o_ex
.fe_group
,
4076 (unsigned long)ac
->ac_o_ex
.fe_start
,
4077 (unsigned long)ac
->ac_o_ex
.fe_len
,
4078 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4079 (unsigned long)ac
->ac_g_ex
.fe_group
,
4080 (unsigned long)ac
->ac_g_ex
.fe_start
,
4081 (unsigned long)ac
->ac_g_ex
.fe_len
,
4082 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4083 (unsigned long)ac
->ac_b_ex
.fe_group
,
4084 (unsigned long)ac
->ac_b_ex
.fe_start
,
4085 (unsigned long)ac
->ac_b_ex
.fe_len
,
4086 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4087 (int)ac
->ac_criteria
);
4088 ext4_msg(ac
->ac_sb
, KERN_ERR
, "%d found", ac
->ac_found
);
4089 ext4_msg(ac
->ac_sb
, KERN_ERR
, "groups: ");
4090 ngroups
= ext4_get_groups_count(sb
);
4091 for (i
= 0; i
< ngroups
; i
++) {
4092 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4093 struct ext4_prealloc_space
*pa
;
4094 ext4_grpblk_t start
;
4095 struct list_head
*cur
;
4096 ext4_lock_group(sb
, i
);
4097 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4098 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4100 spin_lock(&pa
->pa_lock
);
4101 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4103 spin_unlock(&pa
->pa_lock
);
4104 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
4107 ext4_unlock_group(sb
, i
);
4109 if (grp
->bb_free
== 0)
4111 printk(KERN_ERR
"%u: %d/%d \n",
4112 i
, grp
->bb_free
, grp
->bb_fragments
);
4114 printk(KERN_ERR
"\n");
4117 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4124 * We use locality group preallocation for small size file. The size of the
4125 * file is determined by the current size or the resulting size after
4126 * allocation which ever is larger
4128 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4130 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4132 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4133 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4136 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4139 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4142 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
4143 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
4146 if ((size
== isize
) &&
4147 !ext4_fs_is_busy(sbi
) &&
4148 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
4149 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
4153 if (sbi
->s_mb_group_prealloc
<= 0) {
4154 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4158 /* don't use group allocation for large files */
4159 size
= max(size
, isize
);
4160 if (size
> sbi
->s_mb_stream_request
) {
4161 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4165 BUG_ON(ac
->ac_lg
!= NULL
);
4167 * locality group prealloc space are per cpu. The reason for having
4168 * per cpu locality group is to reduce the contention between block
4169 * request from multiple CPUs.
4171 ac
->ac_lg
= raw_cpu_ptr(sbi
->s_locality_groups
);
4173 /* we're going to use group allocation */
4174 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4176 /* serialize all allocations in the group */
4177 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4180 static noinline_for_stack
int
4181 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4182 struct ext4_allocation_request
*ar
)
4184 struct super_block
*sb
= ar
->inode
->i_sb
;
4185 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4186 struct ext4_super_block
*es
= sbi
->s_es
;
4190 ext4_grpblk_t block
;
4192 /* we can't allocate > group size */
4195 /* just a dirty hack to filter too big requests */
4196 if (len
>= EXT4_CLUSTERS_PER_GROUP(sb
))
4197 len
= EXT4_CLUSTERS_PER_GROUP(sb
);
4199 /* start searching from the goal */
4201 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4202 goal
>= ext4_blocks_count(es
))
4203 goal
= le32_to_cpu(es
->s_first_data_block
);
4204 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4206 /* set up allocation goals */
4207 ac
->ac_b_ex
.fe_logical
= EXT4_LBLK_CMASK(sbi
, ar
->logical
);
4208 ac
->ac_status
= AC_STATUS_CONTINUE
;
4210 ac
->ac_inode
= ar
->inode
;
4211 ac
->ac_o_ex
.fe_logical
= ac
->ac_b_ex
.fe_logical
;
4212 ac
->ac_o_ex
.fe_group
= group
;
4213 ac
->ac_o_ex
.fe_start
= block
;
4214 ac
->ac_o_ex
.fe_len
= len
;
4215 ac
->ac_g_ex
= ac
->ac_o_ex
;
4216 ac
->ac_flags
= ar
->flags
;
4218 /* we have to define context: we'll we work with a file or
4219 * locality group. this is a policy, actually */
4220 ext4_mb_group_or_file(ac
);
4222 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4223 "left: %u/%u, right %u/%u to %swritable\n",
4224 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4225 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4226 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4227 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4228 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4233 static noinline_for_stack
void
4234 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4235 struct ext4_locality_group
*lg
,
4236 int order
, int total_entries
)
4238 ext4_group_t group
= 0;
4239 struct ext4_buddy e4b
;
4240 struct list_head discard_list
;
4241 struct ext4_prealloc_space
*pa
, *tmp
;
4243 mb_debug(1, "discard locality group preallocation\n");
4245 INIT_LIST_HEAD(&discard_list
);
4247 spin_lock(&lg
->lg_prealloc_lock
);
4248 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4250 spin_lock(&pa
->pa_lock
);
4251 if (atomic_read(&pa
->pa_count
)) {
4253 * This is the pa that we just used
4254 * for block allocation. So don't
4257 spin_unlock(&pa
->pa_lock
);
4260 if (pa
->pa_deleted
) {
4261 spin_unlock(&pa
->pa_lock
);
4264 /* only lg prealloc space */
4265 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4267 /* seems this one can be freed ... */
4269 spin_unlock(&pa
->pa_lock
);
4271 list_del_rcu(&pa
->pa_inode_list
);
4272 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4275 if (total_entries
<= 5) {
4277 * we want to keep only 5 entries
4278 * allowing it to grow to 8. This
4279 * mak sure we don't call discard
4280 * soon for this list.
4285 spin_unlock(&lg
->lg_prealloc_lock
);
4287 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4289 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4290 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4291 ext4_error(sb
, "Error loading buddy information for %u",
4295 ext4_lock_group(sb
, group
);
4296 list_del(&pa
->pa_group_list
);
4297 ext4_mb_release_group_pa(&e4b
, pa
);
4298 ext4_unlock_group(sb
, group
);
4300 ext4_mb_unload_buddy(&e4b
);
4301 list_del(&pa
->u
.pa_tmp_list
);
4302 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4307 * We have incremented pa_count. So it cannot be freed at this
4308 * point. Also we hold lg_mutex. So no parallel allocation is
4309 * possible from this lg. That means pa_free cannot be updated.
4311 * A parallel ext4_mb_discard_group_preallocations is possible.
4312 * which can cause the lg_prealloc_list to be updated.
4315 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4317 int order
, added
= 0, lg_prealloc_count
= 1;
4318 struct super_block
*sb
= ac
->ac_sb
;
4319 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4320 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4322 order
= fls(pa
->pa_free
) - 1;
4323 if (order
> PREALLOC_TB_SIZE
- 1)
4324 /* The max size of hash table is PREALLOC_TB_SIZE */
4325 order
= PREALLOC_TB_SIZE
- 1;
4326 /* Add the prealloc space to lg */
4327 spin_lock(&lg
->lg_prealloc_lock
);
4328 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4330 spin_lock(&tmp_pa
->pa_lock
);
4331 if (tmp_pa
->pa_deleted
) {
4332 spin_unlock(&tmp_pa
->pa_lock
);
4335 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4336 /* Add to the tail of the previous entry */
4337 list_add_tail_rcu(&pa
->pa_inode_list
,
4338 &tmp_pa
->pa_inode_list
);
4341 * we want to count the total
4342 * number of entries in the list
4345 spin_unlock(&tmp_pa
->pa_lock
);
4346 lg_prealloc_count
++;
4349 list_add_tail_rcu(&pa
->pa_inode_list
,
4350 &lg
->lg_prealloc_list
[order
]);
4351 spin_unlock(&lg
->lg_prealloc_lock
);
4353 /* Now trim the list to be not more than 8 elements */
4354 if (lg_prealloc_count
> 8) {
4355 ext4_mb_discard_lg_preallocations(sb
, lg
,
4356 order
, lg_prealloc_count
);
4363 * release all resource we used in allocation
4365 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4367 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4368 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4370 if (pa
->pa_type
== MB_GROUP_PA
) {
4371 /* see comment in ext4_mb_use_group_pa() */
4372 spin_lock(&pa
->pa_lock
);
4373 pa
->pa_pstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4374 pa
->pa_lstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4375 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4376 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4377 spin_unlock(&pa
->pa_lock
);
4382 * We want to add the pa to the right bucket.
4383 * Remove it from the list and while adding
4384 * make sure the list to which we are adding
4387 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4388 spin_lock(pa
->pa_obj_lock
);
4389 list_del_rcu(&pa
->pa_inode_list
);
4390 spin_unlock(pa
->pa_obj_lock
);
4391 ext4_mb_add_n_trim(ac
);
4393 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4395 if (ac
->ac_bitmap_page
)
4396 put_page(ac
->ac_bitmap_page
);
4397 if (ac
->ac_buddy_page
)
4398 put_page(ac
->ac_buddy_page
);
4399 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4400 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4401 ext4_mb_collect_stats(ac
);
4405 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4407 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4411 trace_ext4_mb_discard_preallocations(sb
, needed
);
4412 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4413 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4422 * Main entry point into mballoc to allocate blocks
4423 * it tries to use preallocation first, then falls back
4424 * to usual allocation
4426 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4427 struct ext4_allocation_request
*ar
, int *errp
)
4430 struct ext4_allocation_context
*ac
= NULL
;
4431 struct ext4_sb_info
*sbi
;
4432 struct super_block
*sb
;
4433 ext4_fsblk_t block
= 0;
4434 unsigned int inquota
= 0;
4435 unsigned int reserv_clstrs
= 0;
4438 sb
= ar
->inode
->i_sb
;
4441 trace_ext4_request_blocks(ar
);
4443 /* Allow to use superuser reservation for quota file */
4444 if (IS_NOQUOTA(ar
->inode
))
4445 ar
->flags
|= EXT4_MB_USE_ROOT_BLOCKS
;
4447 if ((ar
->flags
& EXT4_MB_DELALLOC_RESERVED
) == 0) {
4448 /* Without delayed allocation we need to verify
4449 * there is enough free blocks to do block allocation
4450 * and verify allocation doesn't exceed the quota limits.
4453 ext4_claim_free_clusters(sbi
, ar
->len
, ar
->flags
)) {
4455 /* let others to free the space */
4457 ar
->len
= ar
->len
>> 1;
4463 reserv_clstrs
= ar
->len
;
4464 if (ar
->flags
& EXT4_MB_USE_ROOT_BLOCKS
) {
4465 dquot_alloc_block_nofail(ar
->inode
,
4466 EXT4_C2B(sbi
, ar
->len
));
4469 dquot_alloc_block(ar
->inode
,
4470 EXT4_C2B(sbi
, ar
->len
))) {
4472 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4483 ac
= kmem_cache_zalloc(ext4_ac_cachep
, GFP_NOFS
);
4490 *errp
= ext4_mb_initialize_context(ac
, ar
);
4496 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4497 if (!ext4_mb_use_preallocated(ac
)) {
4498 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4499 ext4_mb_normalize_request(ac
, ar
);
4501 /* allocate space in core */
4502 *errp
= ext4_mb_regular_allocator(ac
);
4504 goto discard_and_exit
;
4506 /* as we've just preallocated more space than
4507 * user requested originally, we store allocated
4508 * space in a special descriptor */
4509 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4510 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4511 *errp
= ext4_mb_new_preallocation(ac
);
4514 ext4_discard_allocated_blocks(ac
);
4518 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4519 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_clstrs
);
4521 ext4_discard_allocated_blocks(ac
);
4524 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4525 ar
->len
= ac
->ac_b_ex
.fe_len
;
4528 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4536 ac
->ac_b_ex
.fe_len
= 0;
4538 ext4_mb_show_ac(ac
);
4540 ext4_mb_release_context(ac
);
4543 kmem_cache_free(ext4_ac_cachep
, ac
);
4544 if (inquota
&& ar
->len
< inquota
)
4545 dquot_free_block(ar
->inode
, EXT4_C2B(sbi
, inquota
- ar
->len
));
4547 if ((ar
->flags
& EXT4_MB_DELALLOC_RESERVED
) == 0)
4548 /* release all the reserved blocks if non delalloc */
4549 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
4553 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4559 * We can merge two free data extents only if the physical blocks
4560 * are contiguous, AND the extents were freed by the same transaction,
4561 * AND the blocks are associated with the same group.
4563 static int can_merge(struct ext4_free_data
*entry1
,
4564 struct ext4_free_data
*entry2
)
4566 if ((entry1
->efd_tid
== entry2
->efd_tid
) &&
4567 (entry1
->efd_group
== entry2
->efd_group
) &&
4568 ((entry1
->efd_start_cluster
+ entry1
->efd_count
) == entry2
->efd_start_cluster
))
4573 static noinline_for_stack
int
4574 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4575 struct ext4_free_data
*new_entry
)
4577 ext4_group_t group
= e4b
->bd_group
;
4578 ext4_grpblk_t cluster
;
4579 ext4_grpblk_t clusters
= new_entry
->efd_count
;
4580 struct ext4_free_data
*entry
;
4581 struct ext4_group_info
*db
= e4b
->bd_info
;
4582 struct super_block
*sb
= e4b
->bd_sb
;
4583 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4584 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4585 struct rb_node
*parent
= NULL
, *new_node
;
4587 BUG_ON(!ext4_handle_valid(handle
));
4588 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4589 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4591 new_node
= &new_entry
->efd_node
;
4592 cluster
= new_entry
->efd_start_cluster
;
4595 /* first free block exent. We need to
4596 protect buddy cache from being freed,
4597 * otherwise we'll refresh it from
4598 * on-disk bitmap and lose not-yet-available
4600 get_page(e4b
->bd_buddy_page
);
4601 get_page(e4b
->bd_bitmap_page
);
4605 entry
= rb_entry(parent
, struct ext4_free_data
, efd_node
);
4606 if (cluster
< entry
->efd_start_cluster
)
4608 else if (cluster
>= (entry
->efd_start_cluster
+ entry
->efd_count
))
4609 n
= &(*n
)->rb_right
;
4611 ext4_grp_locked_error(sb
, group
, 0,
4612 ext4_group_first_block_no(sb
, group
) +
4613 EXT4_C2B(sbi
, cluster
),
4614 "Block already on to-be-freed list");
4619 rb_link_node(new_node
, parent
, n
);
4620 rb_insert_color(new_node
, &db
->bb_free_root
);
4622 /* Now try to see the extent can be merged to left and right */
4623 node
= rb_prev(new_node
);
4625 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4626 if (can_merge(entry
, new_entry
) &&
4627 ext4_journal_callback_try_del(handle
, &entry
->efd_jce
)) {
4628 new_entry
->efd_start_cluster
= entry
->efd_start_cluster
;
4629 new_entry
->efd_count
+= entry
->efd_count
;
4630 rb_erase(node
, &(db
->bb_free_root
));
4631 kmem_cache_free(ext4_free_data_cachep
, entry
);
4635 node
= rb_next(new_node
);
4637 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4638 if (can_merge(new_entry
, entry
) &&
4639 ext4_journal_callback_try_del(handle
, &entry
->efd_jce
)) {
4640 new_entry
->efd_count
+= entry
->efd_count
;
4641 rb_erase(node
, &(db
->bb_free_root
));
4642 kmem_cache_free(ext4_free_data_cachep
, entry
);
4645 /* Add the extent to transaction's private list */
4646 new_entry
->efd_jce
.jce_func
= ext4_free_data_callback
;
4647 spin_lock(&sbi
->s_md_lock
);
4648 _ext4_journal_callback_add(handle
, &new_entry
->efd_jce
);
4649 sbi
->s_mb_free_pending
+= clusters
;
4650 spin_unlock(&sbi
->s_md_lock
);
4655 * ext4_free_blocks() -- Free given blocks and update quota
4656 * @handle: handle for this transaction
4658 * @block: start physical block to free
4659 * @count: number of blocks to count
4660 * @flags: flags used by ext4_free_blocks
4662 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4663 struct buffer_head
*bh
, ext4_fsblk_t block
,
4664 unsigned long count
, int flags
)
4666 struct buffer_head
*bitmap_bh
= NULL
;
4667 struct super_block
*sb
= inode
->i_sb
;
4668 struct ext4_group_desc
*gdp
;
4669 unsigned int overflow
;
4671 struct buffer_head
*gd_bh
;
4672 ext4_group_t block_group
;
4673 struct ext4_sb_info
*sbi
;
4674 struct ext4_buddy e4b
;
4675 unsigned int count_clusters
;
4682 BUG_ON(block
!= bh
->b_blocknr
);
4684 block
= bh
->b_blocknr
;
4688 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4689 !ext4_data_block_valid(sbi
, block
, count
)) {
4690 ext4_error(sb
, "Freeing blocks not in datazone - "
4691 "block = %llu, count = %lu", block
, count
);
4695 ext4_debug("freeing block %llu\n", block
);
4696 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4698 if (bh
&& (flags
& EXT4_FREE_BLOCKS_FORGET
)) {
4701 ext4_forget(handle
, flags
& EXT4_FREE_BLOCKS_METADATA
,
4706 * If the extent to be freed does not begin on a cluster
4707 * boundary, we need to deal with partial clusters at the
4708 * beginning and end of the extent. Normally we will free
4709 * blocks at the beginning or the end unless we are explicitly
4710 * requested to avoid doing so.
4712 overflow
= EXT4_PBLK_COFF(sbi
, block
);
4714 if (flags
& EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER
) {
4715 overflow
= sbi
->s_cluster_ratio
- overflow
;
4717 if (count
> overflow
)
4726 overflow
= EXT4_LBLK_COFF(sbi
, count
);
4728 if (flags
& EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER
) {
4729 if (count
> overflow
)
4734 count
+= sbi
->s_cluster_ratio
- overflow
;
4737 if (!bh
&& (flags
& EXT4_FREE_BLOCKS_FORGET
)) {
4739 int is_metadata
= flags
& EXT4_FREE_BLOCKS_METADATA
;
4741 for (i
= 0; i
< count
; i
++) {
4744 bh
= sb_find_get_block(inode
->i_sb
, block
+ i
);
4745 ext4_forget(handle
, is_metadata
, inode
, bh
, block
+ i
);
4751 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4753 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4754 ext4_get_group_info(sb
, block_group
))))
4758 * Check to see if we are freeing blocks across a group
4761 if (EXT4_C2B(sbi
, bit
) + count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4762 overflow
= EXT4_C2B(sbi
, bit
) + count
-
4763 EXT4_BLOCKS_PER_GROUP(sb
);
4766 count_clusters
= EXT4_NUM_B2C(sbi
, count
);
4767 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4768 if (IS_ERR(bitmap_bh
)) {
4769 err
= PTR_ERR(bitmap_bh
);
4773 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4779 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4780 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4781 in_range(block
, ext4_inode_table(sb
, gdp
),
4782 EXT4_SB(sb
)->s_itb_per_group
) ||
4783 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4784 EXT4_SB(sb
)->s_itb_per_group
)) {
4786 ext4_error(sb
, "Freeing blocks in system zone - "
4787 "Block = %llu, count = %lu", block
, count
);
4788 /* err = 0. ext4_std_error should be a no op */
4792 BUFFER_TRACE(bitmap_bh
, "getting write access");
4793 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4798 * We are about to modify some metadata. Call the journal APIs
4799 * to unshare ->b_data if a currently-committing transaction is
4802 BUFFER_TRACE(gd_bh
, "get_write_access");
4803 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4806 #ifdef AGGRESSIVE_CHECK
4809 for (i
= 0; i
< count_clusters
; i
++)
4810 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4813 trace_ext4_mballoc_free(sb
, inode
, block_group
, bit
, count_clusters
);
4815 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4816 err
= ext4_mb_load_buddy_gfp(sb
, block_group
, &e4b
,
4817 GFP_NOFS
|__GFP_NOFAIL
);
4822 * We need to make sure we don't reuse the freed block until after the
4823 * transaction is committed. We make an exception if the inode is to be
4824 * written in writeback mode since writeback mode has weak data
4825 * consistency guarantees.
4827 if (ext4_handle_valid(handle
) &&
4828 ((flags
& EXT4_FREE_BLOCKS_METADATA
) ||
4829 !ext4_should_writeback_data(inode
))) {
4830 struct ext4_free_data
*new_entry
;
4832 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4835 new_entry
= kmem_cache_alloc(ext4_free_data_cachep
,
4836 GFP_NOFS
|__GFP_NOFAIL
);
4837 new_entry
->efd_start_cluster
= bit
;
4838 new_entry
->efd_group
= block_group
;
4839 new_entry
->efd_count
= count_clusters
;
4840 new_entry
->efd_tid
= handle
->h_transaction
->t_tid
;
4842 ext4_lock_group(sb
, block_group
);
4843 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4844 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4846 /* need to update group_info->bb_free and bitmap
4847 * with group lock held. generate_buddy look at
4848 * them with group lock_held
4850 if (test_opt(sb
, DISCARD
)) {
4851 err
= ext4_issue_discard(sb
, block_group
, bit
, count
);
4852 if (err
&& err
!= -EOPNOTSUPP
)
4853 ext4_msg(sb
, KERN_WARNING
, "discard request in"
4854 " group:%d block:%d count:%lu failed"
4855 " with %d", block_group
, bit
, count
,
4858 EXT4_MB_GRP_CLEAR_TRIMMED(e4b
.bd_info
);
4860 ext4_lock_group(sb
, block_group
);
4861 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4862 mb_free_blocks(inode
, &e4b
, bit
, count_clusters
);
4865 ret
= ext4_free_group_clusters(sb
, gdp
) + count_clusters
;
4866 ext4_free_group_clusters_set(sb
, gdp
, ret
);
4867 ext4_block_bitmap_csum_set(sb
, block_group
, gdp
, bitmap_bh
);
4868 ext4_group_desc_csum_set(sb
, block_group
, gdp
);
4869 ext4_unlock_group(sb
, block_group
);
4871 if (sbi
->s_log_groups_per_flex
) {
4872 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4873 atomic64_add(count_clusters
,
4874 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
4877 if (!(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4878 dquot_free_block(inode
, EXT4_C2B(sbi
, count_clusters
));
4879 percpu_counter_add(&sbi
->s_freeclusters_counter
, count_clusters
);
4881 ext4_mb_unload_buddy(&e4b
);
4883 /* We dirtied the bitmap block */
4884 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4885 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4887 /* And the group descriptor block */
4888 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4889 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4893 if (overflow
&& !err
) {
4901 ext4_std_error(sb
, err
);
4906 * ext4_group_add_blocks() -- Add given blocks to an existing group
4907 * @handle: handle to this transaction
4909 * @block: start physical block to add to the block group
4910 * @count: number of blocks to free
4912 * This marks the blocks as free in the bitmap and buddy.
4914 int ext4_group_add_blocks(handle_t
*handle
, struct super_block
*sb
,
4915 ext4_fsblk_t block
, unsigned long count
)
4917 struct buffer_head
*bitmap_bh
= NULL
;
4918 struct buffer_head
*gd_bh
;
4919 ext4_group_t block_group
;
4922 struct ext4_group_desc
*desc
;
4923 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4924 struct ext4_buddy e4b
;
4925 int err
= 0, ret
, blk_free_count
;
4926 ext4_grpblk_t blocks_freed
;
4928 ext4_debug("Adding block(s) %llu-%llu\n", block
, block
+ count
- 1);
4933 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4935 * Check to see if we are freeing blocks across a group
4938 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4939 ext4_warning(sb
, "too much blocks added to group %u",
4945 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4946 if (IS_ERR(bitmap_bh
)) {
4947 err
= PTR_ERR(bitmap_bh
);
4952 desc
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4958 if (in_range(ext4_block_bitmap(sb
, desc
), block
, count
) ||
4959 in_range(ext4_inode_bitmap(sb
, desc
), block
, count
) ||
4960 in_range(block
, ext4_inode_table(sb
, desc
), sbi
->s_itb_per_group
) ||
4961 in_range(block
+ count
- 1, ext4_inode_table(sb
, desc
),
4962 sbi
->s_itb_per_group
)) {
4963 ext4_error(sb
, "Adding blocks in system zones - "
4964 "Block = %llu, count = %lu",
4970 BUFFER_TRACE(bitmap_bh
, "getting write access");
4971 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4976 * We are about to modify some metadata. Call the journal APIs
4977 * to unshare ->b_data if a currently-committing transaction is
4980 BUFFER_TRACE(gd_bh
, "get_write_access");
4981 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4985 for (i
= 0, blocks_freed
= 0; i
< count
; i
++) {
4986 BUFFER_TRACE(bitmap_bh
, "clear bit");
4987 if (!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
)) {
4988 ext4_error(sb
, "bit already cleared for block %llu",
4989 (ext4_fsblk_t
)(block
+ i
));
4990 BUFFER_TRACE(bitmap_bh
, "bit already cleared");
4996 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
5001 * need to update group_info->bb_free and bitmap
5002 * with group lock held. generate_buddy look at
5003 * them with group lock_held
5005 ext4_lock_group(sb
, block_group
);
5006 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
5007 mb_free_blocks(NULL
, &e4b
, bit
, count
);
5008 blk_free_count
= blocks_freed
+ ext4_free_group_clusters(sb
, desc
);
5009 ext4_free_group_clusters_set(sb
, desc
, blk_free_count
);
5010 ext4_block_bitmap_csum_set(sb
, block_group
, desc
, bitmap_bh
);
5011 ext4_group_desc_csum_set(sb
, block_group
, desc
);
5012 ext4_unlock_group(sb
, block_group
);
5013 percpu_counter_add(&sbi
->s_freeclusters_counter
,
5014 EXT4_NUM_B2C(sbi
, blocks_freed
));
5016 if (sbi
->s_log_groups_per_flex
) {
5017 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
5018 atomic64_add(EXT4_NUM_B2C(sbi
, blocks_freed
),
5019 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
5022 ext4_mb_unload_buddy(&e4b
);
5024 /* We dirtied the bitmap block */
5025 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
5026 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
5028 /* And the group descriptor block */
5029 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
5030 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
5036 ext4_std_error(sb
, err
);
5041 * ext4_trim_extent -- function to TRIM one single free extent in the group
5042 * @sb: super block for the file system
5043 * @start: starting block of the free extent in the alloc. group
5044 * @count: number of blocks to TRIM
5045 * @group: alloc. group we are working with
5046 * @e4b: ext4 buddy for the group
5048 * Trim "count" blocks starting at "start" in the "group". To assure that no
5049 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5050 * be called with under the group lock.
5052 static int ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
5053 ext4_group_t group
, struct ext4_buddy
*e4b
)
5057 struct ext4_free_extent ex
;
5060 trace_ext4_trim_extent(sb
, group
, start
, count
);
5062 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
5064 ex
.fe_start
= start
;
5065 ex
.fe_group
= group
;
5069 * Mark blocks used, so no one can reuse them while
5072 mb_mark_used(e4b
, &ex
);
5073 ext4_unlock_group(sb
, group
);
5074 ret
= ext4_issue_discard(sb
, group
, start
, count
);
5075 ext4_lock_group(sb
, group
);
5076 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
5081 * ext4_trim_all_free -- function to trim all free space in alloc. group
5082 * @sb: super block for file system
5083 * @group: group to be trimmed
5084 * @start: first group block to examine
5085 * @max: last group block to examine
5086 * @minblocks: minimum extent block count
5088 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5089 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5093 * ext4_trim_all_free walks through group's block bitmap searching for free
5094 * extents. When the free extent is found, mark it as used in group buddy
5095 * bitmap. Then issue a TRIM command on this extent and free the extent in
5096 * the group buddy bitmap. This is done until whole group is scanned.
5098 static ext4_grpblk_t
5099 ext4_trim_all_free(struct super_block
*sb
, ext4_group_t group
,
5100 ext4_grpblk_t start
, ext4_grpblk_t max
,
5101 ext4_grpblk_t minblocks
)
5104 ext4_grpblk_t next
, count
= 0, free_count
= 0;
5105 struct ext4_buddy e4b
;
5108 trace_ext4_trim_all_free(sb
, group
, start
, max
);
5110 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
5112 ext4_error(sb
, "Error in loading buddy "
5113 "information for %u", group
);
5116 bitmap
= e4b
.bd_bitmap
;
5118 ext4_lock_group(sb
, group
);
5119 if (EXT4_MB_GRP_WAS_TRIMMED(e4b
.bd_info
) &&
5120 minblocks
>= atomic_read(&EXT4_SB(sb
)->s_last_trim_minblks
))
5123 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
5124 e4b
.bd_info
->bb_first_free
: start
;
5126 while (start
<= max
) {
5127 start
= mb_find_next_zero_bit(bitmap
, max
+ 1, start
);
5130 next
= mb_find_next_bit(bitmap
, max
+ 1, start
);
5132 if ((next
- start
) >= minblocks
) {
5133 ret
= ext4_trim_extent(sb
, start
,
5134 next
- start
, group
, &e4b
);
5135 if (ret
&& ret
!= -EOPNOTSUPP
)
5138 count
+= next
- start
;
5140 free_count
+= next
- start
;
5143 if (fatal_signal_pending(current
)) {
5144 count
= -ERESTARTSYS
;
5148 if (need_resched()) {
5149 ext4_unlock_group(sb
, group
);
5151 ext4_lock_group(sb
, group
);
5154 if ((e4b
.bd_info
->bb_free
- free_count
) < minblocks
)
5160 EXT4_MB_GRP_SET_TRIMMED(e4b
.bd_info
);
5163 ext4_unlock_group(sb
, group
);
5164 ext4_mb_unload_buddy(&e4b
);
5166 ext4_debug("trimmed %d blocks in the group %d\n",
5173 * ext4_trim_fs() -- trim ioctl handle function
5174 * @sb: superblock for filesystem
5175 * @range: fstrim_range structure
5177 * start: First Byte to trim
5178 * len: number of Bytes to trim from start
5179 * minlen: minimum extent length in Bytes
5180 * ext4_trim_fs goes through all allocation groups containing Bytes from
5181 * start to start+len. For each such a group ext4_trim_all_free function
5182 * is invoked to trim all free space.
5184 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
5186 struct ext4_group_info
*grp
;
5187 ext4_group_t group
, first_group
, last_group
;
5188 ext4_grpblk_t cnt
= 0, first_cluster
, last_cluster
;
5189 uint64_t start
, end
, minlen
, trimmed
= 0;
5190 ext4_fsblk_t first_data_blk
=
5191 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
5192 ext4_fsblk_t max_blks
= ext4_blocks_count(EXT4_SB(sb
)->s_es
);
5195 start
= range
->start
>> sb
->s_blocksize_bits
;
5196 end
= start
+ (range
->len
>> sb
->s_blocksize_bits
) - 1;
5197 minlen
= EXT4_NUM_B2C(EXT4_SB(sb
),
5198 range
->minlen
>> sb
->s_blocksize_bits
);
5200 if (minlen
> EXT4_CLUSTERS_PER_GROUP(sb
) ||
5201 start
>= max_blks
||
5202 range
->len
< sb
->s_blocksize
)
5204 if (end
>= max_blks
)
5206 if (end
<= first_data_blk
)
5208 if (start
< first_data_blk
)
5209 start
= first_data_blk
;
5211 /* Determine first and last group to examine based on start and end */
5212 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
5213 &first_group
, &first_cluster
);
5214 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) end
,
5215 &last_group
, &last_cluster
);
5217 /* end now represents the last cluster to discard in this group */
5218 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5220 for (group
= first_group
; group
<= last_group
; group
++) {
5221 grp
= ext4_get_group_info(sb
, group
);
5222 /* We only do this if the grp has never been initialized */
5223 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
5224 ret
= ext4_mb_init_group(sb
, group
, GFP_NOFS
);
5230 * For all the groups except the last one, last cluster will
5231 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5232 * change it for the last group, note that last_cluster is
5233 * already computed earlier by ext4_get_group_no_and_offset()
5235 if (group
== last_group
)
5238 if (grp
->bb_free
>= minlen
) {
5239 cnt
= ext4_trim_all_free(sb
, group
, first_cluster
,
5249 * For every group except the first one, we are sure
5250 * that the first cluster to discard will be cluster #0.
5256 atomic_set(&EXT4_SB(sb
)->s_last_trim_minblks
, minlen
);
5259 range
->len
= EXT4_C2B(EXT4_SB(sb
), trimmed
) << sb
->s_blocksize_bits
;