2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
24 #include "ext4_jbd2.h"
26 #include <linux/log2.h>
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <trace/events/ext4.h>
31 #ifdef CONFIG_EXT4_DEBUG
32 ushort ext4_mballoc_debug __read_mostly
;
34 module_param_named(mballoc_debug
, ext4_mballoc_debug
, ushort
, 0644);
35 MODULE_PARM_DESC(mballoc_debug
, "Debugging level for ext4's mballoc");
40 * - test ext4_ext_search_left() and ext4_ext_search_right()
41 * - search for metadata in few groups
44 * - normalization should take into account whether file is still open
45 * - discard preallocations if no free space left (policy?)
46 * - don't normalize tails
48 * - reservation for superuser
51 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
52 * - track min/max extents in each group for better group selection
53 * - mb_mark_used() may allocate chunk right after splitting buddy
54 * - tree of groups sorted by number of free blocks
59 * The allocation request involve request for multiple number of blocks
60 * near to the goal(block) value specified.
62 * During initialization phase of the allocator we decide to use the
63 * group preallocation or inode preallocation depending on the size of
64 * the file. The size of the file could be the resulting file size we
65 * would have after allocation, or the current file size, which ever
66 * is larger. If the size is less than sbi->s_mb_stream_request we
67 * select to use the group preallocation. The default value of
68 * s_mb_stream_request is 16 blocks. This can also be tuned via
69 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
70 * terms of number of blocks.
72 * The main motivation for having small file use group preallocation is to
73 * ensure that we have small files closer together on the disk.
75 * First stage the allocator looks at the inode prealloc list,
76 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
77 * spaces for this particular inode. The inode prealloc space is
80 * pa_lstart -> the logical start block for this prealloc space
81 * pa_pstart -> the physical start block for this prealloc space
82 * pa_len -> length for this prealloc space (in clusters)
83 * pa_free -> free space available in this prealloc space (in clusters)
85 * The inode preallocation space is used looking at the _logical_ start
86 * block. If only the logical file block falls within the range of prealloc
87 * space we will consume the particular prealloc space. This makes sure that
88 * we have contiguous physical blocks representing the file blocks
90 * The important thing to be noted in case of inode prealloc space is that
91 * we don't modify the values associated to inode prealloc space except
94 * If we are not able to find blocks in the inode prealloc space and if we
95 * have the group allocation flag set then we look at the locality group
96 * prealloc space. These are per CPU prealloc list represented as
98 * ext4_sb_info.s_locality_groups[smp_processor_id()]
100 * The reason for having a per cpu locality group is to reduce the contention
101 * between CPUs. It is possible to get scheduled at this point.
103 * The locality group prealloc space is used looking at whether we have
104 * enough free space (pa_free) within the prealloc space.
106 * If we can't allocate blocks via inode prealloc or/and locality group
107 * prealloc then we look at the buddy cache. The buddy cache is represented
108 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
109 * mapped to the buddy and bitmap information regarding different
110 * groups. The buddy information is attached to buddy cache inode so that
111 * we can access them through the page cache. The information regarding
112 * each group is loaded via ext4_mb_load_buddy. The information involve
113 * block bitmap and buddy information. The information are stored in the
117 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
120 * one block each for bitmap and buddy information. So for each group we
121 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
122 * blocksize) blocks. So it can have information regarding groups_per_page
123 * which is blocks_per_page/2
125 * The buddy cache inode is not stored on disk. The inode is thrown
126 * away when the filesystem is unmounted.
128 * We look for count number of blocks in the buddy cache. If we were able
129 * to locate that many free blocks we return with additional information
130 * regarding rest of the contiguous physical block available
132 * Before allocating blocks via buddy cache we normalize the request
133 * blocks. This ensure we ask for more blocks that we needed. The extra
134 * blocks that we get after allocation is added to the respective prealloc
135 * list. In case of inode preallocation we follow a list of heuristics
136 * based on file size. This can be found in ext4_mb_normalize_request. If
137 * we are doing a group prealloc we try to normalize the request to
138 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
139 * dependent on the cluster size; for non-bigalloc file systems, it is
140 * 512 blocks. This can be tuned via
141 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
142 * terms of number of blocks. If we have mounted the file system with -O
143 * stripe=<value> option the group prealloc request is normalized to the
144 * the smallest multiple of the stripe value (sbi->s_stripe) which is
145 * greater than the default mb_group_prealloc.
147 * The regular allocator (using the buddy cache) supports a few tunables.
149 * /sys/fs/ext4/<partition>/mb_min_to_scan
150 * /sys/fs/ext4/<partition>/mb_max_to_scan
151 * /sys/fs/ext4/<partition>/mb_order2_req
153 * The regular allocator uses buddy scan only if the request len is power of
154 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
155 * value of s_mb_order2_reqs can be tuned via
156 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
157 * stripe size (sbi->s_stripe), we try to search for contiguous block in
158 * stripe size. This should result in better allocation on RAID setups. If
159 * not, we search in the specific group using bitmap for best extents. The
160 * tunable min_to_scan and max_to_scan control the behaviour here.
161 * min_to_scan indicate how long the mballoc __must__ look for a best
162 * extent and max_to_scan indicates how long the mballoc __can__ look for a
163 * best extent in the found extents. Searching for the blocks starts with
164 * the group specified as the goal value in allocation context via
165 * ac_g_ex. Each group is first checked based on the criteria whether it
166 * can be used for allocation. ext4_mb_good_group explains how the groups are
169 * Both the prealloc space are getting populated as above. So for the first
170 * request we will hit the buddy cache which will result in this prealloc
171 * space getting filled. The prealloc space is then later used for the
172 * subsequent request.
176 * mballoc operates on the following data:
178 * - in-core buddy (actually includes buddy and bitmap)
179 * - preallocation descriptors (PAs)
181 * there are two types of preallocations:
183 * assiged to specific inode and can be used for this inode only.
184 * it describes part of inode's space preallocated to specific
185 * physical blocks. any block from that preallocated can be used
186 * independent. the descriptor just tracks number of blocks left
187 * unused. so, before taking some block from descriptor, one must
188 * make sure corresponded logical block isn't allocated yet. this
189 * also means that freeing any block within descriptor's range
190 * must discard all preallocated blocks.
192 * assigned to specific locality group which does not translate to
193 * permanent set of inodes: inode can join and leave group. space
194 * from this type of preallocation can be used for any inode. thus
195 * it's consumed from the beginning to the end.
197 * relation between them can be expressed as:
198 * in-core buddy = on-disk bitmap + preallocation descriptors
200 * this mean blocks mballoc considers used are:
201 * - allocated blocks (persistent)
202 * - preallocated blocks (non-persistent)
204 * consistency in mballoc world means that at any time a block is either
205 * free or used in ALL structures. notice: "any time" should not be read
206 * literally -- time is discrete and delimited by locks.
208 * to keep it simple, we don't use block numbers, instead we count number of
209 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
211 * all operations can be expressed as:
212 * - init buddy: buddy = on-disk + PAs
213 * - new PA: buddy += N; PA = N
214 * - use inode PA: on-disk += N; PA -= N
215 * - discard inode PA buddy -= on-disk - PA; PA = 0
216 * - use locality group PA on-disk += N; PA -= N
217 * - discard locality group PA buddy -= PA; PA = 0
218 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
219 * is used in real operation because we can't know actual used
220 * bits from PA, only from on-disk bitmap
222 * if we follow this strict logic, then all operations above should be atomic.
223 * given some of them can block, we'd have to use something like semaphores
224 * killing performance on high-end SMP hardware. let's try to relax it using
225 * the following knowledge:
226 * 1) if buddy is referenced, it's already initialized
227 * 2) while block is used in buddy and the buddy is referenced,
228 * nobody can re-allocate that block
229 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
230 * bit set and PA claims same block, it's OK. IOW, one can set bit in
231 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
234 * so, now we're building a concurrency table:
237 * blocks for PA are allocated in the buddy, buddy must be referenced
238 * until PA is linked to allocation group to avoid concurrent buddy init
240 * we need to make sure that either on-disk bitmap or PA has uptodate data
241 * given (3) we care that PA-=N operation doesn't interfere with init
243 * the simplest way would be to have buddy initialized by the discard
244 * - use locality group PA
245 * again PA-=N must be serialized with init
246 * - discard locality group PA
247 * the simplest way would be to have buddy initialized by the discard
250 * i_data_sem serializes them
252 * discard process must wait until PA isn't used by another process
253 * - use locality group PA
254 * some mutex should serialize them
255 * - discard locality group PA
256 * discard process must wait until PA isn't used by another process
259 * i_data_sem or another mutex should serializes them
261 * discard process must wait until PA isn't used by another process
262 * - use locality group PA
263 * nothing wrong here -- they're different PAs covering different blocks
264 * - discard locality group PA
265 * discard process must wait until PA isn't used by another process
267 * now we're ready to make few consequences:
268 * - PA is referenced and while it is no discard is possible
269 * - PA is referenced until block isn't marked in on-disk bitmap
270 * - PA changes only after on-disk bitmap
271 * - discard must not compete with init. either init is done before
272 * any discard or they're serialized somehow
273 * - buddy init as sum of on-disk bitmap and PAs is done atomically
275 * a special case when we've used PA to emptiness. no need to modify buddy
276 * in this case, but we should care about concurrent init
281 * Logic in few words:
286 * mark bits in on-disk bitmap
289 * - use preallocation:
290 * find proper PA (per-inode or group)
292 * mark bits in on-disk bitmap
298 * mark bits in on-disk bitmap
301 * - discard preallocations in group:
303 * move them onto local list
304 * load on-disk bitmap
306 * remove PA from object (inode or locality group)
307 * mark free blocks in-core
309 * - discard inode's preallocations:
316 * - bitlock on a group (group)
317 * - object (inode/locality) (object)
328 * - release consumed pa:
333 * - generate in-core bitmap:
337 * - discard all for given object (inode, locality group):
342 * - discard all for given group:
349 static struct kmem_cache
*ext4_pspace_cachep
;
350 static struct kmem_cache
*ext4_ac_cachep
;
351 static struct kmem_cache
*ext4_free_data_cachep
;
353 /* We create slab caches for groupinfo data structures based on the
354 * superblock block size. There will be one per mounted filesystem for
355 * each unique s_blocksize_bits */
356 #define NR_GRPINFO_CACHES 8
357 static struct kmem_cache
*ext4_groupinfo_caches
[NR_GRPINFO_CACHES
];
359 static const char *ext4_groupinfo_slab_names
[NR_GRPINFO_CACHES
] = {
360 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
361 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
362 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
365 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
367 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
369 static void ext4_free_data_callback(struct super_block
*sb
,
370 struct ext4_journal_cb_entry
*jce
, int rc
);
372 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
374 #if BITS_PER_LONG == 64
375 *bit
+= ((unsigned long) addr
& 7UL) << 3;
376 addr
= (void *) ((unsigned long) addr
& ~7UL);
377 #elif BITS_PER_LONG == 32
378 *bit
+= ((unsigned long) addr
& 3UL) << 3;
379 addr
= (void *) ((unsigned long) addr
& ~3UL);
381 #error "how many bits you are?!"
386 static inline int mb_test_bit(int bit
, void *addr
)
389 * ext4_test_bit on architecture like powerpc
390 * needs unsigned long aligned address
392 addr
= mb_correct_addr_and_bit(&bit
, addr
);
393 return ext4_test_bit(bit
, addr
);
396 static inline void mb_set_bit(int bit
, void *addr
)
398 addr
= mb_correct_addr_and_bit(&bit
, addr
);
399 ext4_set_bit(bit
, addr
);
402 static inline void mb_clear_bit(int bit
, void *addr
)
404 addr
= mb_correct_addr_and_bit(&bit
, addr
);
405 ext4_clear_bit(bit
, addr
);
408 static inline int mb_test_and_clear_bit(int bit
, void *addr
)
410 addr
= mb_correct_addr_and_bit(&bit
, addr
);
411 return ext4_test_and_clear_bit(bit
, addr
);
414 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
416 int fix
= 0, ret
, tmpmax
;
417 addr
= mb_correct_addr_and_bit(&fix
, addr
);
421 ret
= ext4_find_next_zero_bit(addr
, tmpmax
, start
) - fix
;
427 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
429 int fix
= 0, ret
, tmpmax
;
430 addr
= mb_correct_addr_and_bit(&fix
, addr
);
434 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
440 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
444 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
447 if (order
> e4b
->bd_blkbits
+ 1) {
452 /* at order 0 we see each particular block */
454 *max
= 1 << (e4b
->bd_blkbits
+ 3);
455 return e4b
->bd_bitmap
;
458 bb
= e4b
->bd_buddy
+ EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
459 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
465 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
466 int first
, int count
)
469 struct super_block
*sb
= e4b
->bd_sb
;
471 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
473 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
474 for (i
= 0; i
< count
; i
++) {
475 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
476 ext4_fsblk_t blocknr
;
478 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
479 blocknr
+= EXT4_C2B(EXT4_SB(sb
), first
+ i
);
480 ext4_grp_locked_error(sb
, e4b
->bd_group
,
481 inode
? inode
->i_ino
: 0,
483 "freeing block already freed "
487 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
491 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
495 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
497 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
498 for (i
= 0; i
< count
; i
++) {
499 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
500 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
504 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
506 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
507 unsigned char *b1
, *b2
;
509 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
510 b2
= (unsigned char *) bitmap
;
511 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
512 if (b1
[i
] != b2
[i
]) {
513 ext4_msg(e4b
->bd_sb
, KERN_ERR
,
514 "corruption in group %u "
515 "at byte %u(%u): %x in copy != %x "
517 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
525 static inline void mb_free_blocks_double(struct inode
*inode
,
526 struct ext4_buddy
*e4b
, int first
, int count
)
530 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
531 int first
, int count
)
535 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
541 #ifdef AGGRESSIVE_CHECK
543 #define MB_CHECK_ASSERT(assert) \
547 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
548 function, file, line, # assert); \
553 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
554 const char *function
, int line
)
556 struct super_block
*sb
= e4b
->bd_sb
;
557 int order
= e4b
->bd_blkbits
+ 1;
564 struct ext4_group_info
*grp
;
567 struct list_head
*cur
;
572 static int mb_check_counter
;
573 if (mb_check_counter
++ % 100 != 0)
578 buddy
= mb_find_buddy(e4b
, order
, &max
);
579 MB_CHECK_ASSERT(buddy
);
580 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
581 MB_CHECK_ASSERT(buddy2
);
582 MB_CHECK_ASSERT(buddy
!= buddy2
);
583 MB_CHECK_ASSERT(max
* 2 == max2
);
586 for (i
= 0; i
< max
; i
++) {
588 if (mb_test_bit(i
, buddy
)) {
589 /* only single bit in buddy2 may be 1 */
590 if (!mb_test_bit(i
<< 1, buddy2
)) {
592 mb_test_bit((i
<<1)+1, buddy2
));
593 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
595 mb_test_bit(i
<< 1, buddy2
));
600 /* both bits in buddy2 must be 1 */
601 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
602 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
604 for (j
= 0; j
< (1 << order
); j
++) {
605 k
= (i
* (1 << order
)) + j
;
607 !mb_test_bit(k
, e4b
->bd_bitmap
));
611 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
616 buddy
= mb_find_buddy(e4b
, 0, &max
);
617 for (i
= 0; i
< max
; i
++) {
618 if (!mb_test_bit(i
, buddy
)) {
619 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
627 /* check used bits only */
628 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
629 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
631 MB_CHECK_ASSERT(k
< max2
);
632 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
635 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
636 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
638 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
639 list_for_each(cur
, &grp
->bb_prealloc_list
) {
640 ext4_group_t groupnr
;
641 struct ext4_prealloc_space
*pa
;
642 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
643 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
644 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
645 for (i
= 0; i
< pa
->pa_len
; i
++)
646 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
650 #undef MB_CHECK_ASSERT
651 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
652 __FILE__, __func__, __LINE__)
654 #define mb_check_buddy(e4b)
658 * Divide blocks started from @first with length @len into
659 * smaller chunks with power of 2 blocks.
660 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
661 * then increase bb_counters[] for corresponded chunk size.
663 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
664 void *buddy
, ext4_grpblk_t first
, ext4_grpblk_t len
,
665 struct ext4_group_info
*grp
)
667 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
671 unsigned short border
;
673 BUG_ON(len
> EXT4_CLUSTERS_PER_GROUP(sb
));
675 border
= 2 << sb
->s_blocksize_bits
;
678 /* find how many blocks can be covered since this position */
679 max
= ffs(first
| border
) - 1;
681 /* find how many blocks of power 2 we need to mark */
688 /* mark multiblock chunks only */
689 grp
->bb_counters
[min
]++;
691 mb_clear_bit(first
>> min
,
692 buddy
+ sbi
->s_mb_offsets
[min
]);
700 * Cache the order of the largest free extent we have available in this block
704 mb_set_largest_free_order(struct super_block
*sb
, struct ext4_group_info
*grp
)
709 grp
->bb_largest_free_order
= -1; /* uninit */
711 bits
= sb
->s_blocksize_bits
+ 1;
712 for (i
= bits
; i
>= 0; i
--) {
713 if (grp
->bb_counters
[i
] > 0) {
714 grp
->bb_largest_free_order
= i
;
720 static noinline_for_stack
721 void ext4_mb_generate_buddy(struct super_block
*sb
,
722 void *buddy
, void *bitmap
, ext4_group_t group
)
724 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
725 ext4_grpblk_t max
= EXT4_CLUSTERS_PER_GROUP(sb
);
730 unsigned fragments
= 0;
731 unsigned long long period
= get_cycles();
733 /* initialize buddy from bitmap which is aggregation
734 * of on-disk bitmap and preallocations */
735 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
736 grp
->bb_first_free
= i
;
740 i
= mb_find_next_bit(bitmap
, max
, i
);
744 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
746 grp
->bb_counters
[0]++;
748 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
750 grp
->bb_fragments
= fragments
;
752 if (free
!= grp
->bb_free
) {
753 ext4_grp_locked_error(sb
, group
, 0, 0,
754 "%u clusters in bitmap, %u in gd; "
755 "block bitmap corrupt.",
758 * If we intend to continue, we consider group descriptor
759 * corrupt and update bb_free using bitmap value
762 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT
, &grp
->bb_state
);
764 mb_set_largest_free_order(sb
, grp
);
766 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
768 period
= get_cycles() - period
;
769 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
770 EXT4_SB(sb
)->s_mb_buddies_generated
++;
771 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
772 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
775 static void mb_regenerate_buddy(struct ext4_buddy
*e4b
)
781 while ((buddy
= mb_find_buddy(e4b
, order
++, &count
))) {
782 ext4_set_bits(buddy
, 0, count
);
784 e4b
->bd_info
->bb_fragments
= 0;
785 memset(e4b
->bd_info
->bb_counters
, 0,
786 sizeof(*e4b
->bd_info
->bb_counters
) *
787 (e4b
->bd_sb
->s_blocksize_bits
+ 2));
789 ext4_mb_generate_buddy(e4b
->bd_sb
, e4b
->bd_buddy
,
790 e4b
->bd_bitmap
, e4b
->bd_group
);
793 /* The buddy information is attached the buddy cache inode
794 * for convenience. The information regarding each group
795 * is loaded via ext4_mb_load_buddy. The information involve
796 * block bitmap and buddy information. The information are
797 * stored in the inode as
800 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
803 * one block each for bitmap and buddy information.
804 * So for each group we take up 2 blocks. A page can
805 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
806 * So it can have information regarding groups_per_page which
807 * is blocks_per_page/2
809 * Locking note: This routine takes the block group lock of all groups
810 * for this page; do not hold this lock when calling this routine!
813 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
815 ext4_group_t ngroups
;
821 ext4_group_t first_group
, group
;
823 struct super_block
*sb
;
824 struct buffer_head
*bhs
;
825 struct buffer_head
**bh
= NULL
;
829 struct ext4_group_info
*grinfo
;
831 mb_debug(1, "init page %lu\n", page
->index
);
833 inode
= page
->mapping
->host
;
835 ngroups
= ext4_get_groups_count(sb
);
836 blocksize
= 1 << inode
->i_blkbits
;
837 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
839 groups_per_page
= blocks_per_page
>> 1;
840 if (groups_per_page
== 0)
843 /* allocate buffer_heads to read bitmaps */
844 if (groups_per_page
> 1) {
845 i
= sizeof(struct buffer_head
*) * groups_per_page
;
846 bh
= kzalloc(i
, GFP_NOFS
);
854 first_group
= page
->index
* blocks_per_page
/ 2;
856 /* read all groups the page covers into the cache */
857 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
858 if (group
>= ngroups
)
861 grinfo
= ext4_get_group_info(sb
, group
);
863 * If page is uptodate then we came here after online resize
864 * which added some new uninitialized group info structs, so
865 * we must skip all initialized uptodate buddies on the page,
866 * which may be currently in use by an allocating task.
868 if (PageUptodate(page
) && !EXT4_MB_GRP_NEED_INIT(grinfo
)) {
872 if (!(bh
[i
] = ext4_read_block_bitmap_nowait(sb
, group
))) {
876 mb_debug(1, "read bitmap for group %u\n", group
);
879 /* wait for I/O completion */
880 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
881 if (bh
[i
] && ext4_wait_block_bitmap(sb
, group
, bh
[i
])) {
887 first_block
= page
->index
* blocks_per_page
;
888 for (i
= 0; i
< blocks_per_page
; i
++) {
889 group
= (first_block
+ i
) >> 1;
890 if (group
>= ngroups
)
893 if (!bh
[group
- first_group
])
894 /* skip initialized uptodate buddy */
898 * data carry information regarding this
899 * particular group in the format specified
903 data
= page_address(page
) + (i
* blocksize
);
904 bitmap
= bh
[group
- first_group
]->b_data
;
907 * We place the buddy block and bitmap block
910 if ((first_block
+ i
) & 1) {
911 /* this is block of buddy */
912 BUG_ON(incore
== NULL
);
913 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
914 group
, page
->index
, i
* blocksize
);
915 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
916 grinfo
= ext4_get_group_info(sb
, group
);
917 grinfo
->bb_fragments
= 0;
918 memset(grinfo
->bb_counters
, 0,
919 sizeof(*grinfo
->bb_counters
) *
920 (sb
->s_blocksize_bits
+2));
922 * incore got set to the group block bitmap below
924 ext4_lock_group(sb
, group
);
926 memset(data
, 0xff, blocksize
);
927 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
928 ext4_unlock_group(sb
, group
);
931 /* this is block of bitmap */
932 BUG_ON(incore
!= NULL
);
933 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
934 group
, page
->index
, i
* blocksize
);
935 trace_ext4_mb_bitmap_load(sb
, group
);
937 /* see comments in ext4_mb_put_pa() */
938 ext4_lock_group(sb
, group
);
939 memcpy(data
, bitmap
, blocksize
);
941 /* mark all preallocated blks used in in-core bitmap */
942 ext4_mb_generate_from_pa(sb
, data
, group
);
943 ext4_mb_generate_from_freelist(sb
, data
, group
);
944 ext4_unlock_group(sb
, group
);
946 /* set incore so that the buddy information can be
947 * generated using this
952 SetPageUptodate(page
);
956 for (i
= 0; i
< groups_per_page
; i
++)
965 * Lock the buddy and bitmap pages. This make sure other parallel init_group
966 * on the same buddy page doesn't happen whild holding the buddy page lock.
967 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
968 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
970 static int ext4_mb_get_buddy_page_lock(struct super_block
*sb
,
971 ext4_group_t group
, struct ext4_buddy
*e4b
)
973 struct inode
*inode
= EXT4_SB(sb
)->s_buddy_cache
;
974 int block
, pnum
, poff
;
978 e4b
->bd_buddy_page
= NULL
;
979 e4b
->bd_bitmap_page
= NULL
;
981 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
983 * the buddy cache inode stores the block bitmap
984 * and buddy information in consecutive blocks.
985 * So for each group we need two blocks.
988 pnum
= block
/ blocks_per_page
;
989 poff
= block
% blocks_per_page
;
990 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
993 BUG_ON(page
->mapping
!= inode
->i_mapping
);
994 e4b
->bd_bitmap_page
= page
;
995 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
997 if (blocks_per_page
>= 2) {
998 /* buddy and bitmap are on the same page */
1003 pnum
= block
/ blocks_per_page
;
1004 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1007 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1008 e4b
->bd_buddy_page
= page
;
1012 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy
*e4b
)
1014 if (e4b
->bd_bitmap_page
) {
1015 unlock_page(e4b
->bd_bitmap_page
);
1016 page_cache_release(e4b
->bd_bitmap_page
);
1018 if (e4b
->bd_buddy_page
) {
1019 unlock_page(e4b
->bd_buddy_page
);
1020 page_cache_release(e4b
->bd_buddy_page
);
1025 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1026 * block group lock of all groups for this page; do not hold the BG lock when
1027 * calling this routine!
1029 static noinline_for_stack
1030 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
)
1033 struct ext4_group_info
*this_grp
;
1034 struct ext4_buddy e4b
;
1039 mb_debug(1, "init group %u\n", group
);
1040 this_grp
= ext4_get_group_info(sb
, group
);
1042 * This ensures that we don't reinit the buddy cache
1043 * page which map to the group from which we are already
1044 * allocating. If we are looking at the buddy cache we would
1045 * have taken a reference using ext4_mb_load_buddy and that
1046 * would have pinned buddy page to page cache.
1048 ret
= ext4_mb_get_buddy_page_lock(sb
, group
, &e4b
);
1049 if (ret
|| !EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1051 * somebody initialized the group
1052 * return without doing anything
1057 page
= e4b
.bd_bitmap_page
;
1058 ret
= ext4_mb_init_cache(page
, NULL
);
1061 if (!PageUptodate(page
)) {
1065 mark_page_accessed(page
);
1067 if (e4b
.bd_buddy_page
== NULL
) {
1069 * If both the bitmap and buddy are in
1070 * the same page we don't need to force
1076 /* init buddy cache */
1077 page
= e4b
.bd_buddy_page
;
1078 ret
= ext4_mb_init_cache(page
, e4b
.bd_bitmap
);
1081 if (!PageUptodate(page
)) {
1085 mark_page_accessed(page
);
1087 ext4_mb_put_buddy_page_lock(&e4b
);
1092 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1093 * block group lock of all groups for this page; do not hold the BG lock when
1094 * calling this routine!
1096 static noinline_for_stack
int
1097 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1098 struct ext4_buddy
*e4b
)
1100 int blocks_per_page
;
1106 struct ext4_group_info
*grp
;
1107 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1108 struct inode
*inode
= sbi
->s_buddy_cache
;
1111 mb_debug(1, "load group %u\n", group
);
1113 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1114 grp
= ext4_get_group_info(sb
, group
);
1116 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1119 e4b
->bd_group
= group
;
1120 e4b
->bd_buddy_page
= NULL
;
1121 e4b
->bd_bitmap_page
= NULL
;
1123 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1125 * we need full data about the group
1126 * to make a good selection
1128 ret
= ext4_mb_init_group(sb
, group
);
1134 * the buddy cache inode stores the block bitmap
1135 * and buddy information in consecutive blocks.
1136 * So for each group we need two blocks.
1139 pnum
= block
/ blocks_per_page
;
1140 poff
= block
% blocks_per_page
;
1142 /* we could use find_or_create_page(), but it locks page
1143 * what we'd like to avoid in fast path ... */
1144 page
= find_get_page(inode
->i_mapping
, pnum
);
1145 if (page
== NULL
|| !PageUptodate(page
)) {
1148 * drop the page reference and try
1149 * to get the page with lock. If we
1150 * are not uptodate that implies
1151 * somebody just created the page but
1152 * is yet to initialize the same. So
1153 * wait for it to initialize.
1155 page_cache_release(page
);
1156 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1158 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1159 if (!PageUptodate(page
)) {
1160 ret
= ext4_mb_init_cache(page
, NULL
);
1165 mb_cmp_bitmaps(e4b
, page_address(page
) +
1166 (poff
* sb
->s_blocksize
));
1175 if (!PageUptodate(page
)) {
1179 e4b
->bd_bitmap_page
= page
;
1180 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1181 mark_page_accessed(page
);
1184 pnum
= block
/ blocks_per_page
;
1185 poff
= block
% blocks_per_page
;
1187 page
= find_get_page(inode
->i_mapping
, pnum
);
1188 if (page
== NULL
|| !PageUptodate(page
)) {
1190 page_cache_release(page
);
1191 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1193 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1194 if (!PageUptodate(page
)) {
1195 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1208 if (!PageUptodate(page
)) {
1212 e4b
->bd_buddy_page
= page
;
1213 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1214 mark_page_accessed(page
);
1216 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1217 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1223 page_cache_release(page
);
1224 if (e4b
->bd_bitmap_page
)
1225 page_cache_release(e4b
->bd_bitmap_page
);
1226 if (e4b
->bd_buddy_page
)
1227 page_cache_release(e4b
->bd_buddy_page
);
1228 e4b
->bd_buddy
= NULL
;
1229 e4b
->bd_bitmap
= NULL
;
1233 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1235 if (e4b
->bd_bitmap_page
)
1236 page_cache_release(e4b
->bd_bitmap_page
);
1237 if (e4b
->bd_buddy_page
)
1238 page_cache_release(e4b
->bd_buddy_page
);
1242 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1247 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
1248 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1251 while (order
<= e4b
->bd_blkbits
+ 1) {
1253 if (!mb_test_bit(block
, bb
)) {
1254 /* this block is part of buddy of order 'order' */
1257 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1263 static void mb_clear_bits(void *bm
, int cur
, int len
)
1269 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1270 /* fast path: clear whole word at once */
1271 addr
= bm
+ (cur
>> 3);
1276 mb_clear_bit(cur
, bm
);
1281 /* clear bits in given range
1282 * will return first found zero bit if any, -1 otherwise
1284 static int mb_test_and_clear_bits(void *bm
, int cur
, int len
)
1291 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1292 /* fast path: clear whole word at once */
1293 addr
= bm
+ (cur
>> 3);
1294 if (*addr
!= (__u32
)(-1) && zero_bit
== -1)
1295 zero_bit
= cur
+ mb_find_next_zero_bit(addr
, 32, 0);
1300 if (!mb_test_and_clear_bit(cur
, bm
) && zero_bit
== -1)
1308 void ext4_set_bits(void *bm
, int cur
, int len
)
1314 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1315 /* fast path: set whole word at once */
1316 addr
= bm
+ (cur
>> 3);
1321 mb_set_bit(cur
, bm
);
1327 * _________________________________________________________________ */
1329 static inline int mb_buddy_adjust_border(int* bit
, void* bitmap
, int side
)
1331 if (mb_test_bit(*bit
+ side
, bitmap
)) {
1332 mb_clear_bit(*bit
, bitmap
);
1338 mb_set_bit(*bit
, bitmap
);
1343 static void mb_buddy_mark_free(struct ext4_buddy
*e4b
, int first
, int last
)
1347 void *buddy
= mb_find_buddy(e4b
, order
, &max
);
1352 /* Bits in range [first; last] are known to be set since
1353 * corresponding blocks were allocated. Bits in range
1354 * (first; last) will stay set because they form buddies on
1355 * upper layer. We just deal with borders if they don't
1356 * align with upper layer and then go up.
1357 * Releasing entire group is all about clearing
1358 * single bit of highest order buddy.
1362 * ---------------------------------
1364 * ---------------------------------
1365 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1366 * ---------------------------------
1368 * \_____________________/
1370 * Neither [1] nor [6] is aligned to above layer.
1371 * Left neighbour [0] is free, so mark it busy,
1372 * decrease bb_counters and extend range to
1374 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1375 * mark [6] free, increase bb_counters and shrink range to
1377 * Then shift range to [0; 2], go up and do the same.
1382 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&first
, buddy
, -1);
1384 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&last
, buddy
, 1);
1389 if (first
== last
|| !(buddy2
= mb_find_buddy(e4b
, order
, &max
))) {
1390 mb_clear_bits(buddy
, first
, last
- first
+ 1);
1391 e4b
->bd_info
->bb_counters
[order
- 1] += last
- first
+ 1;
1400 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1401 int first
, int count
)
1403 int left_is_free
= 0;
1404 int right_is_free
= 0;
1406 int last
= first
+ count
- 1;
1407 struct super_block
*sb
= e4b
->bd_sb
;
1409 BUG_ON(last
>= (sb
->s_blocksize
<< 3));
1410 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1411 /* Don't bother if the block group is corrupt. */
1412 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
)))
1415 mb_check_buddy(e4b
);
1416 mb_free_blocks_double(inode
, e4b
, first
, count
);
1418 e4b
->bd_info
->bb_free
+= count
;
1419 if (first
< e4b
->bd_info
->bb_first_free
)
1420 e4b
->bd_info
->bb_first_free
= first
;
1422 /* access memory sequentially: check left neighbour,
1423 * clear range and then check right neighbour
1426 left_is_free
= !mb_test_bit(first
- 1, e4b
->bd_bitmap
);
1427 block
= mb_test_and_clear_bits(e4b
->bd_bitmap
, first
, count
);
1428 if (last
+ 1 < EXT4_SB(sb
)->s_mb_maxs
[0])
1429 right_is_free
= !mb_test_bit(last
+ 1, e4b
->bd_bitmap
);
1431 if (unlikely(block
!= -1)) {
1432 ext4_fsblk_t blocknr
;
1434 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1435 blocknr
+= EXT4_C2B(EXT4_SB(sb
), block
);
1436 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1437 inode
? inode
->i_ino
: 0,
1439 "freeing already freed block "
1440 "(bit %u); block bitmap corrupt.",
1442 /* Mark the block group as corrupt. */
1443 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT
,
1444 &e4b
->bd_info
->bb_state
);
1445 mb_regenerate_buddy(e4b
);
1449 /* let's maintain fragments counter */
1450 if (left_is_free
&& right_is_free
)
1451 e4b
->bd_info
->bb_fragments
--;
1452 else if (!left_is_free
&& !right_is_free
)
1453 e4b
->bd_info
->bb_fragments
++;
1455 /* buddy[0] == bd_bitmap is a special case, so handle
1456 * it right away and let mb_buddy_mark_free stay free of
1457 * zero order checks.
1458 * Check if neighbours are to be coaleasced,
1459 * adjust bitmap bb_counters and borders appropriately.
1462 first
+= !left_is_free
;
1463 e4b
->bd_info
->bb_counters
[0] += left_is_free
? -1 : 1;
1466 last
-= !right_is_free
;
1467 e4b
->bd_info
->bb_counters
[0] += right_is_free
? -1 : 1;
1471 mb_buddy_mark_free(e4b
, first
>> 1, last
>> 1);
1474 mb_set_largest_free_order(sb
, e4b
->bd_info
);
1475 mb_check_buddy(e4b
);
1478 static int mb_find_extent(struct ext4_buddy
*e4b
, int block
,
1479 int needed
, struct ext4_free_extent
*ex
)
1485 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1488 buddy
= mb_find_buddy(e4b
, 0, &max
);
1489 BUG_ON(buddy
== NULL
);
1490 BUG_ON(block
>= max
);
1491 if (mb_test_bit(block
, buddy
)) {
1498 /* find actual order */
1499 order
= mb_find_order_for_block(e4b
, block
);
1500 block
= block
>> order
;
1502 ex
->fe_len
= 1 << order
;
1503 ex
->fe_start
= block
<< order
;
1504 ex
->fe_group
= e4b
->bd_group
;
1506 /* calc difference from given start */
1507 next
= next
- ex
->fe_start
;
1509 ex
->fe_start
+= next
;
1511 while (needed
> ex
->fe_len
&&
1512 mb_find_buddy(e4b
, order
, &max
)) {
1514 if (block
+ 1 >= max
)
1517 next
= (block
+ 1) * (1 << order
);
1518 if (mb_test_bit(next
, e4b
->bd_bitmap
))
1521 order
= mb_find_order_for_block(e4b
, next
);
1523 block
= next
>> order
;
1524 ex
->fe_len
+= 1 << order
;
1527 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1531 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1537 int start
= ex
->fe_start
;
1538 int len
= ex
->fe_len
;
1543 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1544 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1545 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1546 mb_check_buddy(e4b
);
1547 mb_mark_used_double(e4b
, start
, len
);
1549 e4b
->bd_info
->bb_free
-= len
;
1550 if (e4b
->bd_info
->bb_first_free
== start
)
1551 e4b
->bd_info
->bb_first_free
+= len
;
1553 /* let's maintain fragments counter */
1555 mlen
= !mb_test_bit(start
- 1, e4b
->bd_bitmap
);
1556 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1557 max
= !mb_test_bit(start
+ len
, e4b
->bd_bitmap
);
1559 e4b
->bd_info
->bb_fragments
++;
1560 else if (!mlen
&& !max
)
1561 e4b
->bd_info
->bb_fragments
--;
1563 /* let's maintain buddy itself */
1565 ord
= mb_find_order_for_block(e4b
, start
);
1567 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1568 /* the whole chunk may be allocated at once! */
1570 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1571 BUG_ON((start
>> ord
) >= max
);
1572 mb_set_bit(start
>> ord
, buddy
);
1573 e4b
->bd_info
->bb_counters
[ord
]--;
1580 /* store for history */
1582 ret
= len
| (ord
<< 16);
1584 /* we have to split large buddy */
1586 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1587 mb_set_bit(start
>> ord
, buddy
);
1588 e4b
->bd_info
->bb_counters
[ord
]--;
1591 cur
= (start
>> ord
) & ~1U;
1592 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1593 mb_clear_bit(cur
, buddy
);
1594 mb_clear_bit(cur
+ 1, buddy
);
1595 e4b
->bd_info
->bb_counters
[ord
]++;
1596 e4b
->bd_info
->bb_counters
[ord
]++;
1598 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1600 ext4_set_bits(e4b
->bd_bitmap
, ex
->fe_start
, len0
);
1601 mb_check_buddy(e4b
);
1607 * Must be called under group lock!
1609 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1610 struct ext4_buddy
*e4b
)
1612 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1615 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1616 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1618 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1619 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1620 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1622 /* preallocation can change ac_b_ex, thus we store actually
1623 * allocated blocks for history */
1624 ac
->ac_f_ex
= ac
->ac_b_ex
;
1626 ac
->ac_status
= AC_STATUS_FOUND
;
1627 ac
->ac_tail
= ret
& 0xffff;
1628 ac
->ac_buddy
= ret
>> 16;
1631 * take the page reference. We want the page to be pinned
1632 * so that we don't get a ext4_mb_init_cache_call for this
1633 * group until we update the bitmap. That would mean we
1634 * double allocate blocks. The reference is dropped
1635 * in ext4_mb_release_context
1637 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1638 get_page(ac
->ac_bitmap_page
);
1639 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1640 get_page(ac
->ac_buddy_page
);
1641 /* store last allocated for subsequent stream allocation */
1642 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1643 spin_lock(&sbi
->s_md_lock
);
1644 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1645 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1646 spin_unlock(&sbi
->s_md_lock
);
1651 * regular allocator, for general purposes allocation
1654 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1655 struct ext4_buddy
*e4b
,
1658 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1659 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1660 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1661 struct ext4_free_extent ex
;
1664 if (ac
->ac_status
== AC_STATUS_FOUND
)
1667 * We don't want to scan for a whole year
1669 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1670 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1671 ac
->ac_status
= AC_STATUS_BREAK
;
1676 * Haven't found good chunk so far, let's continue
1678 if (bex
->fe_len
< gex
->fe_len
)
1681 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1682 && bex
->fe_group
== e4b
->bd_group
) {
1683 /* recheck chunk's availability - we don't know
1684 * when it was found (within this lock-unlock
1686 max
= mb_find_extent(e4b
, bex
->fe_start
, gex
->fe_len
, &ex
);
1687 if (max
>= gex
->fe_len
) {
1688 ext4_mb_use_best_found(ac
, e4b
);
1695 * The routine checks whether found extent is good enough. If it is,
1696 * then the extent gets marked used and flag is set to the context
1697 * to stop scanning. Otherwise, the extent is compared with the
1698 * previous found extent and if new one is better, then it's stored
1699 * in the context. Later, the best found extent will be used, if
1700 * mballoc can't find good enough extent.
1702 * FIXME: real allocation policy is to be designed yet!
1704 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1705 struct ext4_free_extent
*ex
,
1706 struct ext4_buddy
*e4b
)
1708 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1709 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1711 BUG_ON(ex
->fe_len
<= 0);
1712 BUG_ON(ex
->fe_len
> EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1713 BUG_ON(ex
->fe_start
>= EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1714 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1719 * The special case - take what you catch first
1721 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1723 ext4_mb_use_best_found(ac
, e4b
);
1728 * Let's check whether the chuck is good enough
1730 if (ex
->fe_len
== gex
->fe_len
) {
1732 ext4_mb_use_best_found(ac
, e4b
);
1737 * If this is first found extent, just store it in the context
1739 if (bex
->fe_len
== 0) {
1745 * If new found extent is better, store it in the context
1747 if (bex
->fe_len
< gex
->fe_len
) {
1748 /* if the request isn't satisfied, any found extent
1749 * larger than previous best one is better */
1750 if (ex
->fe_len
> bex
->fe_len
)
1752 } else if (ex
->fe_len
> gex
->fe_len
) {
1753 /* if the request is satisfied, then we try to find
1754 * an extent that still satisfy the request, but is
1755 * smaller than previous one */
1756 if (ex
->fe_len
< bex
->fe_len
)
1760 ext4_mb_check_limits(ac
, e4b
, 0);
1763 static noinline_for_stack
1764 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1765 struct ext4_buddy
*e4b
)
1767 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1768 ext4_group_t group
= ex
.fe_group
;
1772 BUG_ON(ex
.fe_len
<= 0);
1773 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1777 ext4_lock_group(ac
->ac_sb
, group
);
1778 max
= mb_find_extent(e4b
, ex
.fe_start
, ex
.fe_len
, &ex
);
1782 ext4_mb_use_best_found(ac
, e4b
);
1785 ext4_unlock_group(ac
->ac_sb
, group
);
1786 ext4_mb_unload_buddy(e4b
);
1791 static noinline_for_stack
1792 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1793 struct ext4_buddy
*e4b
)
1795 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1798 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1799 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1800 struct ext4_free_extent ex
;
1802 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1804 if (grp
->bb_free
== 0)
1807 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1811 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b
->bd_info
))) {
1812 ext4_mb_unload_buddy(e4b
);
1816 ext4_lock_group(ac
->ac_sb
, group
);
1817 max
= mb_find_extent(e4b
, ac
->ac_g_ex
.fe_start
,
1818 ac
->ac_g_ex
.fe_len
, &ex
);
1819 ex
.fe_logical
= 0xDEADFA11; /* debug value */
1821 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1824 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1826 /* use do_div to get remainder (would be 64-bit modulo) */
1827 if (do_div(start
, sbi
->s_stripe
) == 0) {
1830 ext4_mb_use_best_found(ac
, e4b
);
1832 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1833 BUG_ON(ex
.fe_len
<= 0);
1834 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1835 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1838 ext4_mb_use_best_found(ac
, e4b
);
1839 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1840 /* Sometimes, caller may want to merge even small
1841 * number of blocks to an existing extent */
1842 BUG_ON(ex
.fe_len
<= 0);
1843 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1844 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1847 ext4_mb_use_best_found(ac
, e4b
);
1849 ext4_unlock_group(ac
->ac_sb
, group
);
1850 ext4_mb_unload_buddy(e4b
);
1856 * The routine scans buddy structures (not bitmap!) from given order
1857 * to max order and tries to find big enough chunk to satisfy the req
1859 static noinline_for_stack
1860 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1861 struct ext4_buddy
*e4b
)
1863 struct super_block
*sb
= ac
->ac_sb
;
1864 struct ext4_group_info
*grp
= e4b
->bd_info
;
1870 BUG_ON(ac
->ac_2order
<= 0);
1871 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1872 if (grp
->bb_counters
[i
] == 0)
1875 buddy
= mb_find_buddy(e4b
, i
, &max
);
1876 BUG_ON(buddy
== NULL
);
1878 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1883 ac
->ac_b_ex
.fe_len
= 1 << i
;
1884 ac
->ac_b_ex
.fe_start
= k
<< i
;
1885 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1887 ext4_mb_use_best_found(ac
, e4b
);
1889 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1891 if (EXT4_SB(sb
)->s_mb_stats
)
1892 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1899 * The routine scans the group and measures all found extents.
1900 * In order to optimize scanning, caller must pass number of
1901 * free blocks in the group, so the routine can know upper limit.
1903 static noinline_for_stack
1904 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1905 struct ext4_buddy
*e4b
)
1907 struct super_block
*sb
= ac
->ac_sb
;
1908 void *bitmap
= e4b
->bd_bitmap
;
1909 struct ext4_free_extent ex
;
1913 free
= e4b
->bd_info
->bb_free
;
1916 i
= e4b
->bd_info
->bb_first_free
;
1918 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1919 i
= mb_find_next_zero_bit(bitmap
,
1920 EXT4_CLUSTERS_PER_GROUP(sb
), i
);
1921 if (i
>= EXT4_CLUSTERS_PER_GROUP(sb
)) {
1923 * IF we have corrupt bitmap, we won't find any
1924 * free blocks even though group info says we
1925 * we have free blocks
1927 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1928 "%d free clusters as per "
1929 "group info. But bitmap says 0",
1934 mb_find_extent(e4b
, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1935 BUG_ON(ex
.fe_len
<= 0);
1936 if (free
< ex
.fe_len
) {
1937 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1938 "%d free clusters as per "
1939 "group info. But got %d blocks",
1942 * The number of free blocks differs. This mostly
1943 * indicate that the bitmap is corrupt. So exit
1944 * without claiming the space.
1948 ex
.fe_logical
= 0xDEADC0DE; /* debug value */
1949 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1955 ext4_mb_check_limits(ac
, e4b
, 1);
1959 * This is a special case for storages like raid5
1960 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1962 static noinline_for_stack
1963 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1964 struct ext4_buddy
*e4b
)
1966 struct super_block
*sb
= ac
->ac_sb
;
1967 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1968 void *bitmap
= e4b
->bd_bitmap
;
1969 struct ext4_free_extent ex
;
1970 ext4_fsblk_t first_group_block
;
1975 BUG_ON(sbi
->s_stripe
== 0);
1977 /* find first stripe-aligned block in group */
1978 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1980 a
= first_group_block
+ sbi
->s_stripe
- 1;
1981 do_div(a
, sbi
->s_stripe
);
1982 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1984 while (i
< EXT4_CLUSTERS_PER_GROUP(sb
)) {
1985 if (!mb_test_bit(i
, bitmap
)) {
1986 max
= mb_find_extent(e4b
, i
, sbi
->s_stripe
, &ex
);
1987 if (max
>= sbi
->s_stripe
) {
1989 ex
.fe_logical
= 0xDEADF00D; /* debug value */
1991 ext4_mb_use_best_found(ac
, e4b
);
1999 /* This is now called BEFORE we load the buddy bitmap. */
2000 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
2001 ext4_group_t group
, int cr
)
2003 unsigned free
, fragments
;
2004 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
2005 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
2007 BUG_ON(cr
< 0 || cr
>= 4);
2009 free
= grp
->bb_free
;
2012 if (cr
<= 2 && free
< ac
->ac_g_ex
.fe_len
)
2015 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp
)))
2018 /* We only do this if the grp has never been initialized */
2019 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
2020 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
);
2025 fragments
= grp
->bb_fragments
;
2031 BUG_ON(ac
->ac_2order
== 0);
2033 /* Avoid using the first bg of a flexgroup for data files */
2034 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
2035 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
2036 ((group
% flex_size
) == 0))
2039 if ((ac
->ac_2order
> ac
->ac_sb
->s_blocksize_bits
+1) ||
2040 (free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2043 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
2048 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2052 if (free
>= ac
->ac_g_ex
.fe_len
)
2064 static noinline_for_stack
int
2065 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
2067 ext4_group_t ngroups
, group
, i
;
2070 struct ext4_sb_info
*sbi
;
2071 struct super_block
*sb
;
2072 struct ext4_buddy e4b
;
2076 ngroups
= ext4_get_groups_count(sb
);
2077 /* non-extent files are limited to low blocks/groups */
2078 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
2079 ngroups
= sbi
->s_blockfile_groups
;
2081 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
2083 /* first, try the goal */
2084 err
= ext4_mb_find_by_goal(ac
, &e4b
);
2085 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
2088 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2092 * ac->ac2_order is set only if the fe_len is a power of 2
2093 * if ac2_order is set we also set criteria to 0 so that we
2094 * try exact allocation using buddy.
2096 i
= fls(ac
->ac_g_ex
.fe_len
);
2099 * We search using buddy data only if the order of the request
2100 * is greater than equal to the sbi_s_mb_order2_reqs
2101 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2103 if (i
>= sbi
->s_mb_order2_reqs
) {
2105 * This should tell if fe_len is exactly power of 2
2107 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2108 ac
->ac_2order
= i
- 1;
2111 /* if stream allocation is enabled, use global goal */
2112 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2113 /* TBD: may be hot point */
2114 spin_lock(&sbi
->s_md_lock
);
2115 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2116 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2117 spin_unlock(&sbi
->s_md_lock
);
2120 /* Let's just scan groups to find more-less suitable blocks */
2121 cr
= ac
->ac_2order
? 0 : 1;
2123 * cr == 0 try to get exact allocation,
2124 * cr == 3 try to get anything
2127 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2128 ac
->ac_criteria
= cr
;
2130 * searching for the right group start
2131 * from the goal value specified
2133 group
= ac
->ac_g_ex
.fe_group
;
2135 for (i
= 0; i
< ngroups
; group
++, i
++) {
2138 * Artificially restricted ngroups for non-extent
2139 * files makes group > ngroups possible on first loop.
2141 if (group
>= ngroups
)
2144 /* This now checks without needing the buddy page */
2145 if (!ext4_mb_good_group(ac
, group
, cr
))
2148 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2152 ext4_lock_group(sb
, group
);
2155 * We need to check again after locking the
2158 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2159 ext4_unlock_group(sb
, group
);
2160 ext4_mb_unload_buddy(&e4b
);
2164 ac
->ac_groups_scanned
++;
2165 if (cr
== 0 && ac
->ac_2order
< sb
->s_blocksize_bits
+2)
2166 ext4_mb_simple_scan_group(ac
, &e4b
);
2167 else if (cr
== 1 && sbi
->s_stripe
&&
2168 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2169 ext4_mb_scan_aligned(ac
, &e4b
);
2171 ext4_mb_complex_scan_group(ac
, &e4b
);
2173 ext4_unlock_group(sb
, group
);
2174 ext4_mb_unload_buddy(&e4b
);
2176 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2181 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2182 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2184 * We've been searching too long. Let's try to allocate
2185 * the best chunk we've found so far
2188 ext4_mb_try_best_found(ac
, &e4b
);
2189 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2191 * Someone more lucky has already allocated it.
2192 * The only thing we can do is just take first
2194 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2196 ac
->ac_b_ex
.fe_group
= 0;
2197 ac
->ac_b_ex
.fe_start
= 0;
2198 ac
->ac_b_ex
.fe_len
= 0;
2199 ac
->ac_status
= AC_STATUS_CONTINUE
;
2200 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2202 atomic_inc(&sbi
->s_mb_lost_chunks
);
2210 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2212 struct super_block
*sb
= seq
->private;
2215 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2218 return (void *) ((unsigned long) group
);
2221 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2223 struct super_block
*sb
= seq
->private;
2227 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2230 return (void *) ((unsigned long) group
);
2233 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2235 struct super_block
*sb
= seq
->private;
2236 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2238 int err
, buddy_loaded
= 0;
2239 struct ext4_buddy e4b
;
2240 struct ext4_group_info
*grinfo
;
2242 struct ext4_group_info info
;
2243 ext4_grpblk_t counters
[16];
2248 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2249 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2250 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2251 "group", "free", "frags", "first",
2252 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2253 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2255 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2256 sizeof(struct ext4_group_info
);
2257 grinfo
= ext4_get_group_info(sb
, group
);
2258 /* Load the group info in memory only if not already loaded. */
2259 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo
))) {
2260 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2262 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2268 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2271 ext4_mb_unload_buddy(&e4b
);
2273 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2274 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2275 for (i
= 0; i
<= 13; i
++)
2276 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2277 sg
.info
.bb_counters
[i
] : 0);
2278 seq_printf(seq
, " ]\n");
2283 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2287 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2288 .start
= ext4_mb_seq_groups_start
,
2289 .next
= ext4_mb_seq_groups_next
,
2290 .stop
= ext4_mb_seq_groups_stop
,
2291 .show
= ext4_mb_seq_groups_show
,
2294 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2296 struct super_block
*sb
= PDE_DATA(inode
);
2299 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2301 struct seq_file
*m
= file
->private_data
;
2308 static const struct file_operations ext4_mb_seq_groups_fops
= {
2309 .owner
= THIS_MODULE
,
2310 .open
= ext4_mb_seq_groups_open
,
2312 .llseek
= seq_lseek
,
2313 .release
= seq_release
,
2316 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2318 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2319 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2326 * Allocate the top-level s_group_info array for the specified number
2329 int ext4_mb_alloc_groupinfo(struct super_block
*sb
, ext4_group_t ngroups
)
2331 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2333 struct ext4_group_info
***new_groupinfo
;
2335 size
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2336 EXT4_DESC_PER_BLOCK_BITS(sb
);
2337 if (size
<= sbi
->s_group_info_size
)
2340 size
= roundup_pow_of_two(sizeof(*sbi
->s_group_info
) * size
);
2341 new_groupinfo
= ext4_kvzalloc(size
, GFP_KERNEL
);
2342 if (!new_groupinfo
) {
2343 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy meta group");
2346 if (sbi
->s_group_info
) {
2347 memcpy(new_groupinfo
, sbi
->s_group_info
,
2348 sbi
->s_group_info_size
* sizeof(*sbi
->s_group_info
));
2349 ext4_kvfree(sbi
->s_group_info
);
2351 sbi
->s_group_info
= new_groupinfo
;
2352 sbi
->s_group_info_size
= size
/ sizeof(*sbi
->s_group_info
);
2353 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2354 sbi
->s_group_info_size
);
2358 /* Create and initialize ext4_group_info data for the given group. */
2359 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2360 struct ext4_group_desc
*desc
)
2364 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2365 struct ext4_group_info
**meta_group_info
;
2366 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2369 * First check if this group is the first of a reserved block.
2370 * If it's true, we have to allocate a new table of pointers
2371 * to ext4_group_info structures
2373 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2374 metalen
= sizeof(*meta_group_info
) <<
2375 EXT4_DESC_PER_BLOCK_BITS(sb
);
2376 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2377 if (meta_group_info
== NULL
) {
2378 ext4_msg(sb
, KERN_ERR
, "can't allocate mem "
2379 "for a buddy group");
2380 goto exit_meta_group_info
;
2382 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2387 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2388 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2390 meta_group_info
[i
] = kmem_cache_zalloc(cachep
, GFP_KERNEL
);
2391 if (meta_group_info
[i
] == NULL
) {
2392 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy mem");
2393 goto exit_group_info
;
2395 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2396 &(meta_group_info
[i
]->bb_state
));
2399 * initialize bb_free to be able to skip
2400 * empty groups without initialization
2402 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2403 meta_group_info
[i
]->bb_free
=
2404 ext4_free_clusters_after_init(sb
, group
, desc
);
2406 meta_group_info
[i
]->bb_free
=
2407 ext4_free_group_clusters(sb
, desc
);
2410 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2411 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2412 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2413 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2417 struct buffer_head
*bh
;
2418 meta_group_info
[i
]->bb_bitmap
=
2419 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2420 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2421 bh
= ext4_read_block_bitmap(sb
, group
);
2423 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2432 /* If a meta_group_info table has been allocated, release it now */
2433 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2434 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2435 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] = NULL
;
2437 exit_meta_group_info
:
2439 } /* ext4_mb_add_groupinfo */
2441 static int ext4_mb_init_backend(struct super_block
*sb
)
2443 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2445 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2447 struct ext4_group_desc
*desc
;
2448 struct kmem_cache
*cachep
;
2450 err
= ext4_mb_alloc_groupinfo(sb
, ngroups
);
2454 sbi
->s_buddy_cache
= new_inode(sb
);
2455 if (sbi
->s_buddy_cache
== NULL
) {
2456 ext4_msg(sb
, KERN_ERR
, "can't get new inode");
2459 /* To avoid potentially colliding with an valid on-disk inode number,
2460 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2461 * not in the inode hash, so it should never be found by iget(), but
2462 * this will avoid confusion if it ever shows up during debugging. */
2463 sbi
->s_buddy_cache
->i_ino
= EXT4_BAD_INO
;
2464 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2465 for (i
= 0; i
< ngroups
; i
++) {
2466 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2468 ext4_msg(sb
, KERN_ERR
, "can't read descriptor %u", i
);
2471 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2478 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2480 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2481 i
= sbi
->s_group_info_size
;
2483 kfree(sbi
->s_group_info
[i
]);
2484 iput(sbi
->s_buddy_cache
);
2486 ext4_kvfree(sbi
->s_group_info
);
2490 static void ext4_groupinfo_destroy_slabs(void)
2494 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2495 if (ext4_groupinfo_caches
[i
])
2496 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2497 ext4_groupinfo_caches
[i
] = NULL
;
2501 static int ext4_groupinfo_create_slab(size_t size
)
2503 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2505 int blocksize_bits
= order_base_2(size
);
2506 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2507 struct kmem_cache
*cachep
;
2509 if (cache_index
>= NR_GRPINFO_CACHES
)
2512 if (unlikely(cache_index
< 0))
2515 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2516 if (ext4_groupinfo_caches
[cache_index
]) {
2517 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2518 return 0; /* Already created */
2521 slab_size
= offsetof(struct ext4_group_info
,
2522 bb_counters
[blocksize_bits
+ 2]);
2524 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2525 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2528 ext4_groupinfo_caches
[cache_index
] = cachep
;
2530 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2533 "EXT4-fs: no memory for groupinfo slab cache\n");
2540 int ext4_mb_init(struct super_block
*sb
)
2542 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2548 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2550 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2551 if (sbi
->s_mb_offsets
== NULL
) {
2556 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2557 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2558 if (sbi
->s_mb_maxs
== NULL
) {
2563 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2567 /* order 0 is regular bitmap */
2568 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2569 sbi
->s_mb_offsets
[0] = 0;
2573 max
= sb
->s_blocksize
<< 2;
2575 sbi
->s_mb_offsets
[i
] = offset
;
2576 sbi
->s_mb_maxs
[i
] = max
;
2577 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2580 } while (i
<= sb
->s_blocksize_bits
+ 1);
2582 spin_lock_init(&sbi
->s_md_lock
);
2583 spin_lock_init(&sbi
->s_bal_lock
);
2585 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2586 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2587 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2588 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2589 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2591 * The default group preallocation is 512, which for 4k block
2592 * sizes translates to 2 megabytes. However for bigalloc file
2593 * systems, this is probably too big (i.e, if the cluster size
2594 * is 1 megabyte, then group preallocation size becomes half a
2595 * gigabyte!). As a default, we will keep a two megabyte
2596 * group pralloc size for cluster sizes up to 64k, and after
2597 * that, we will force a minimum group preallocation size of
2598 * 32 clusters. This translates to 8 megs when the cluster
2599 * size is 256k, and 32 megs when the cluster size is 1 meg,
2600 * which seems reasonable as a default.
2602 sbi
->s_mb_group_prealloc
= max(MB_DEFAULT_GROUP_PREALLOC
>>
2603 sbi
->s_cluster_bits
, 32);
2605 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2606 * to the lowest multiple of s_stripe which is bigger than
2607 * the s_mb_group_prealloc as determined above. We want
2608 * the preallocation size to be an exact multiple of the
2609 * RAID stripe size so that preallocations don't fragment
2612 if (sbi
->s_stripe
> 1) {
2613 sbi
->s_mb_group_prealloc
= roundup(
2614 sbi
->s_mb_group_prealloc
, sbi
->s_stripe
);
2617 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2618 if (sbi
->s_locality_groups
== NULL
) {
2620 goto out_free_groupinfo_slab
;
2622 for_each_possible_cpu(i
) {
2623 struct ext4_locality_group
*lg
;
2624 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2625 mutex_init(&lg
->lg_mutex
);
2626 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2627 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2628 spin_lock_init(&lg
->lg_prealloc_lock
);
2631 /* init file for buddy data */
2632 ret
= ext4_mb_init_backend(sb
);
2634 goto out_free_locality_groups
;
2637 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2638 &ext4_mb_seq_groups_fops
, sb
);
2642 out_free_locality_groups
:
2643 free_percpu(sbi
->s_locality_groups
);
2644 sbi
->s_locality_groups
= NULL
;
2645 out_free_groupinfo_slab
:
2646 ext4_groupinfo_destroy_slabs();
2648 kfree(sbi
->s_mb_offsets
);
2649 sbi
->s_mb_offsets
= NULL
;
2650 kfree(sbi
->s_mb_maxs
);
2651 sbi
->s_mb_maxs
= NULL
;
2655 /* need to called with the ext4 group lock held */
2656 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2658 struct ext4_prealloc_space
*pa
;
2659 struct list_head
*cur
, *tmp
;
2662 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2663 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2664 list_del(&pa
->pa_group_list
);
2666 kmem_cache_free(ext4_pspace_cachep
, pa
);
2669 mb_debug(1, "mballoc: %u PAs left\n", count
);
2673 int ext4_mb_release(struct super_block
*sb
)
2675 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2677 int num_meta_group_infos
;
2678 struct ext4_group_info
*grinfo
;
2679 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2680 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2683 remove_proc_entry("mb_groups", sbi
->s_proc
);
2685 if (sbi
->s_group_info
) {
2686 for (i
= 0; i
< ngroups
; i
++) {
2687 grinfo
= ext4_get_group_info(sb
, i
);
2689 kfree(grinfo
->bb_bitmap
);
2691 ext4_lock_group(sb
, i
);
2692 ext4_mb_cleanup_pa(grinfo
);
2693 ext4_unlock_group(sb
, i
);
2694 kmem_cache_free(cachep
, grinfo
);
2696 num_meta_group_infos
= (ngroups
+
2697 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2698 EXT4_DESC_PER_BLOCK_BITS(sb
);
2699 for (i
= 0; i
< num_meta_group_infos
; i
++)
2700 kfree(sbi
->s_group_info
[i
]);
2701 ext4_kvfree(sbi
->s_group_info
);
2703 kfree(sbi
->s_mb_offsets
);
2704 kfree(sbi
->s_mb_maxs
);
2705 if (sbi
->s_buddy_cache
)
2706 iput(sbi
->s_buddy_cache
);
2707 if (sbi
->s_mb_stats
) {
2708 ext4_msg(sb
, KERN_INFO
,
2709 "mballoc: %u blocks %u reqs (%u success)",
2710 atomic_read(&sbi
->s_bal_allocated
),
2711 atomic_read(&sbi
->s_bal_reqs
),
2712 atomic_read(&sbi
->s_bal_success
));
2713 ext4_msg(sb
, KERN_INFO
,
2714 "mballoc: %u extents scanned, %u goal hits, "
2715 "%u 2^N hits, %u breaks, %u lost",
2716 atomic_read(&sbi
->s_bal_ex_scanned
),
2717 atomic_read(&sbi
->s_bal_goals
),
2718 atomic_read(&sbi
->s_bal_2orders
),
2719 atomic_read(&sbi
->s_bal_breaks
),
2720 atomic_read(&sbi
->s_mb_lost_chunks
));
2721 ext4_msg(sb
, KERN_INFO
,
2722 "mballoc: %lu generated and it took %Lu",
2723 sbi
->s_mb_buddies_generated
,
2724 sbi
->s_mb_generation_time
);
2725 ext4_msg(sb
, KERN_INFO
,
2726 "mballoc: %u preallocated, %u discarded",
2727 atomic_read(&sbi
->s_mb_preallocated
),
2728 atomic_read(&sbi
->s_mb_discarded
));
2731 free_percpu(sbi
->s_locality_groups
);
2736 static inline int ext4_issue_discard(struct super_block
*sb
,
2737 ext4_group_t block_group
, ext4_grpblk_t cluster
, int count
)
2739 ext4_fsblk_t discard_block
;
2741 discard_block
= (EXT4_C2B(EXT4_SB(sb
), cluster
) +
2742 ext4_group_first_block_no(sb
, block_group
));
2743 count
= EXT4_C2B(EXT4_SB(sb
), count
);
2744 trace_ext4_discard_blocks(sb
,
2745 (unsigned long long) discard_block
, count
);
2746 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, 0);
2750 * This function is called by the jbd2 layer once the commit has finished,
2751 * so we know we can free the blocks that were released with that commit.
2753 static void ext4_free_data_callback(struct super_block
*sb
,
2754 struct ext4_journal_cb_entry
*jce
,
2757 struct ext4_free_data
*entry
= (struct ext4_free_data
*)jce
;
2758 struct ext4_buddy e4b
;
2759 struct ext4_group_info
*db
;
2760 int err
, count
= 0, count2
= 0;
2762 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2763 entry
->efd_count
, entry
->efd_group
, entry
);
2765 if (test_opt(sb
, DISCARD
)) {
2766 err
= ext4_issue_discard(sb
, entry
->efd_group
,
2767 entry
->efd_start_cluster
,
2769 if (err
&& err
!= -EOPNOTSUPP
)
2770 ext4_msg(sb
, KERN_WARNING
, "discard request in"
2771 " group:%d block:%d count:%d failed"
2772 " with %d", entry
->efd_group
,
2773 entry
->efd_start_cluster
,
2774 entry
->efd_count
, err
);
2777 err
= ext4_mb_load_buddy(sb
, entry
->efd_group
, &e4b
);
2778 /* we expect to find existing buddy because it's pinned */
2783 /* there are blocks to put in buddy to make them really free */
2784 count
+= entry
->efd_count
;
2786 ext4_lock_group(sb
, entry
->efd_group
);
2787 /* Take it out of per group rb tree */
2788 rb_erase(&entry
->efd_node
, &(db
->bb_free_root
));
2789 mb_free_blocks(NULL
, &e4b
, entry
->efd_start_cluster
, entry
->efd_count
);
2792 * Clear the trimmed flag for the group so that the next
2793 * ext4_trim_fs can trim it.
2794 * If the volume is mounted with -o discard, online discard
2795 * is supported and the free blocks will be trimmed online.
2797 if (!test_opt(sb
, DISCARD
))
2798 EXT4_MB_GRP_CLEAR_TRIMMED(db
);
2800 if (!db
->bb_free_root
.rb_node
) {
2801 /* No more items in the per group rb tree
2802 * balance refcounts from ext4_mb_free_metadata()
2804 page_cache_release(e4b
.bd_buddy_page
);
2805 page_cache_release(e4b
.bd_bitmap_page
);
2807 ext4_unlock_group(sb
, entry
->efd_group
);
2808 kmem_cache_free(ext4_free_data_cachep
, entry
);
2809 ext4_mb_unload_buddy(&e4b
);
2811 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2814 int __init
ext4_init_mballoc(void)
2816 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2817 SLAB_RECLAIM_ACCOUNT
);
2818 if (ext4_pspace_cachep
== NULL
)
2821 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2822 SLAB_RECLAIM_ACCOUNT
);
2823 if (ext4_ac_cachep
== NULL
) {
2824 kmem_cache_destroy(ext4_pspace_cachep
);
2828 ext4_free_data_cachep
= KMEM_CACHE(ext4_free_data
,
2829 SLAB_RECLAIM_ACCOUNT
);
2830 if (ext4_free_data_cachep
== NULL
) {
2831 kmem_cache_destroy(ext4_pspace_cachep
);
2832 kmem_cache_destroy(ext4_ac_cachep
);
2838 void ext4_exit_mballoc(void)
2841 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2842 * before destroying the slab cache.
2845 kmem_cache_destroy(ext4_pspace_cachep
);
2846 kmem_cache_destroy(ext4_ac_cachep
);
2847 kmem_cache_destroy(ext4_free_data_cachep
);
2848 ext4_groupinfo_destroy_slabs();
2853 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2854 * Returns 0 if success or error code
2856 static noinline_for_stack
int
2857 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2858 handle_t
*handle
, unsigned int reserv_clstrs
)
2860 struct buffer_head
*bitmap_bh
= NULL
;
2861 struct ext4_group_desc
*gdp
;
2862 struct buffer_head
*gdp_bh
;
2863 struct ext4_sb_info
*sbi
;
2864 struct super_block
*sb
;
2868 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2869 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2875 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2879 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2884 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2888 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2889 ext4_free_group_clusters(sb
, gdp
));
2891 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2895 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
2897 len
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
2898 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2899 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2900 "fs metadata", block
, block
+len
);
2901 /* File system mounted not to panic on error
2902 * Fix the bitmap and repeat the block allocation
2903 * We leak some of the blocks here.
2905 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2906 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2907 ac
->ac_b_ex
.fe_len
);
2908 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2909 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2915 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2916 #ifdef AGGRESSIVE_CHECK
2919 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2920 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2921 bitmap_bh
->b_data
));
2925 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2926 ac
->ac_b_ex
.fe_len
);
2927 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2928 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2929 ext4_free_group_clusters_set(sb
, gdp
,
2930 ext4_free_clusters_after_init(sb
,
2931 ac
->ac_b_ex
.fe_group
, gdp
));
2933 len
= ext4_free_group_clusters(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
2934 ext4_free_group_clusters_set(sb
, gdp
, len
);
2935 ext4_block_bitmap_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
, bitmap_bh
);
2936 ext4_group_desc_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
);
2938 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2939 percpu_counter_sub(&sbi
->s_freeclusters_counter
, ac
->ac_b_ex
.fe_len
);
2941 * Now reduce the dirty block count also. Should not go negative
2943 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2944 /* release all the reserved blocks if non delalloc */
2945 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
2948 if (sbi
->s_log_groups_per_flex
) {
2949 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2950 ac
->ac_b_ex
.fe_group
);
2951 atomic64_sub(ac
->ac_b_ex
.fe_len
,
2952 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
2955 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2958 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
2966 * here we normalize request for locality group
2967 * Group request are normalized to s_mb_group_prealloc, which goes to
2968 * s_strip if we set the same via mount option.
2969 * s_mb_group_prealloc can be configured via
2970 * /sys/fs/ext4/<partition>/mb_group_prealloc
2972 * XXX: should we try to preallocate more than the group has now?
2974 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2976 struct super_block
*sb
= ac
->ac_sb
;
2977 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2980 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2981 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2982 current
->pid
, ac
->ac_g_ex
.fe_len
);
2986 * Normalization means making request better in terms of
2987 * size and alignment
2989 static noinline_for_stack
void
2990 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
2991 struct ext4_allocation_request
*ar
)
2993 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2996 loff_t size
, start_off
;
2997 loff_t orig_size __maybe_unused
;
2999 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3000 struct ext4_prealloc_space
*pa
;
3002 /* do normalize only data requests, metadata requests
3003 do not need preallocation */
3004 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3007 /* sometime caller may want exact blocks */
3008 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3011 /* caller may indicate that preallocation isn't
3012 * required (it's a tail, for example) */
3013 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3016 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3017 ext4_mb_normalize_group_request(ac
);
3021 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3023 /* first, let's learn actual file size
3024 * given current request is allocated */
3025 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
3026 size
= size
<< bsbits
;
3027 if (size
< i_size_read(ac
->ac_inode
))
3028 size
= i_size_read(ac
->ac_inode
);
3031 /* max size of free chunks */
3034 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3035 (req <= (size) || max <= (chunk_size))
3037 /* first, try to predict filesize */
3038 /* XXX: should this table be tunable? */
3040 if (size
<= 16 * 1024) {
3042 } else if (size
<= 32 * 1024) {
3044 } else if (size
<= 64 * 1024) {
3046 } else if (size
<= 128 * 1024) {
3048 } else if (size
<= 256 * 1024) {
3050 } else if (size
<= 512 * 1024) {
3052 } else if (size
<= 1024 * 1024) {
3054 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3055 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3056 (21 - bsbits
)) << 21;
3057 size
= 2 * 1024 * 1024;
3058 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3059 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3060 (22 - bsbits
)) << 22;
3061 size
= 4 * 1024 * 1024;
3062 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3063 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3064 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3065 (23 - bsbits
)) << 23;
3066 size
= 8 * 1024 * 1024;
3068 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3069 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3071 size
= size
>> bsbits
;
3072 start
= start_off
>> bsbits
;
3074 /* don't cover already allocated blocks in selected range */
3075 if (ar
->pleft
&& start
<= ar
->lleft
) {
3076 size
-= ar
->lleft
+ 1 - start
;
3077 start
= ar
->lleft
+ 1;
3079 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3080 size
-= start
+ size
- ar
->lright
;
3084 /* check we don't cross already preallocated blocks */
3086 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3091 spin_lock(&pa
->pa_lock
);
3092 if (pa
->pa_deleted
) {
3093 spin_unlock(&pa
->pa_lock
);
3097 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3100 /* PA must not overlap original request */
3101 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3102 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3104 /* skip PAs this normalized request doesn't overlap with */
3105 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
3106 spin_unlock(&pa
->pa_lock
);
3109 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3111 /* adjust start or end to be adjacent to this pa */
3112 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3113 BUG_ON(pa_end
< start
);
3115 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3116 BUG_ON(pa
->pa_lstart
> end
);
3117 end
= pa
->pa_lstart
;
3119 spin_unlock(&pa
->pa_lock
);
3124 /* XXX: extra loop to check we really don't overlap preallocations */
3126 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3129 spin_lock(&pa
->pa_lock
);
3130 if (pa
->pa_deleted
== 0) {
3131 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3133 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3135 spin_unlock(&pa
->pa_lock
);
3139 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3140 start
> ac
->ac_o_ex
.fe_logical
) {
3141 ext4_msg(ac
->ac_sb
, KERN_ERR
,
3142 "start %lu, size %lu, fe_logical %lu",
3143 (unsigned long) start
, (unsigned long) size
,
3144 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3146 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3147 start
> ac
->ac_o_ex
.fe_logical
);
3148 BUG_ON(size
<= 0 || size
> EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
3150 /* now prepare goal request */
3152 /* XXX: is it better to align blocks WRT to logical
3153 * placement or satisfy big request as is */
3154 ac
->ac_g_ex
.fe_logical
= start
;
3155 ac
->ac_g_ex
.fe_len
= EXT4_NUM_B2C(sbi
, size
);
3157 /* define goal start in order to merge */
3158 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3159 /* merge to the right */
3160 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3161 &ac
->ac_f_ex
.fe_group
,
3162 &ac
->ac_f_ex
.fe_start
);
3163 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3165 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3166 /* merge to the left */
3167 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3168 &ac
->ac_f_ex
.fe_group
,
3169 &ac
->ac_f_ex
.fe_start
);
3170 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3173 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3174 (unsigned) orig_size
, (unsigned) start
);
3177 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3179 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3181 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3182 atomic_inc(&sbi
->s_bal_reqs
);
3183 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3184 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3185 atomic_inc(&sbi
->s_bal_success
);
3186 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3187 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3188 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3189 atomic_inc(&sbi
->s_bal_goals
);
3190 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3191 atomic_inc(&sbi
->s_bal_breaks
);
3194 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3195 trace_ext4_mballoc_alloc(ac
);
3197 trace_ext4_mballoc_prealloc(ac
);
3201 * Called on failure; free up any blocks from the inode PA for this
3202 * context. We don't need this for MB_GROUP_PA because we only change
3203 * pa_free in ext4_mb_release_context(), but on failure, we've already
3204 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3206 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3208 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3210 if (pa
&& pa
->pa_type
== MB_INODE_PA
)
3211 pa
->pa_free
+= ac
->ac_b_ex
.fe_len
;
3215 * use blocks preallocated to inode
3217 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3218 struct ext4_prealloc_space
*pa
)
3220 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3225 /* found preallocated blocks, use them */
3226 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3227 end
= min(pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
),
3228 start
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
));
3229 len
= EXT4_NUM_B2C(sbi
, end
- start
);
3230 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3231 &ac
->ac_b_ex
.fe_start
);
3232 ac
->ac_b_ex
.fe_len
= len
;
3233 ac
->ac_status
= AC_STATUS_FOUND
;
3236 BUG_ON(start
< pa
->pa_pstart
);
3237 BUG_ON(end
> pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
));
3238 BUG_ON(pa
->pa_free
< len
);
3241 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3245 * use blocks preallocated to locality group
3247 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3248 struct ext4_prealloc_space
*pa
)
3250 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3252 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3253 &ac
->ac_b_ex
.fe_group
,
3254 &ac
->ac_b_ex
.fe_start
);
3255 ac
->ac_b_ex
.fe_len
= len
;
3256 ac
->ac_status
= AC_STATUS_FOUND
;
3259 /* we don't correct pa_pstart or pa_plen here to avoid
3260 * possible race when the group is being loaded concurrently
3261 * instead we correct pa later, after blocks are marked
3262 * in on-disk bitmap -- see ext4_mb_release_context()
3263 * Other CPUs are prevented from allocating from this pa by lg_mutex
3265 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3269 * Return the prealloc space that have minimal distance
3270 * from the goal block. @cpa is the prealloc
3271 * space that is having currently known minimal distance
3272 * from the goal block.
3274 static struct ext4_prealloc_space
*
3275 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3276 struct ext4_prealloc_space
*pa
,
3277 struct ext4_prealloc_space
*cpa
)
3279 ext4_fsblk_t cur_distance
, new_distance
;
3282 atomic_inc(&pa
->pa_count
);
3285 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3286 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3288 if (cur_distance
<= new_distance
)
3291 /* drop the previous reference */
3292 atomic_dec(&cpa
->pa_count
);
3293 atomic_inc(&pa
->pa_count
);
3298 * search goal blocks in preallocated space
3300 static noinline_for_stack
int
3301 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3303 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3305 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3306 struct ext4_locality_group
*lg
;
3307 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3308 ext4_fsblk_t goal_block
;
3310 /* only data can be preallocated */
3311 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3314 /* first, try per-file preallocation */
3316 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3318 /* all fields in this condition don't change,
3319 * so we can skip locking for them */
3320 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3321 ac
->ac_o_ex
.fe_logical
>= (pa
->pa_lstart
+
3322 EXT4_C2B(sbi
, pa
->pa_len
)))
3325 /* non-extent files can't have physical blocks past 2^32 */
3326 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3327 (pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
) >
3328 EXT4_MAX_BLOCK_FILE_PHYS
))
3331 /* found preallocated blocks, use them */
3332 spin_lock(&pa
->pa_lock
);
3333 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3334 atomic_inc(&pa
->pa_count
);
3335 ext4_mb_use_inode_pa(ac
, pa
);
3336 spin_unlock(&pa
->pa_lock
);
3337 ac
->ac_criteria
= 10;
3341 spin_unlock(&pa
->pa_lock
);
3345 /* can we use group allocation? */
3346 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3349 /* inode may have no locality group for some reason */
3353 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3354 if (order
> PREALLOC_TB_SIZE
- 1)
3355 /* The max size of hash table is PREALLOC_TB_SIZE */
3356 order
= PREALLOC_TB_SIZE
- 1;
3358 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3360 * search for the prealloc space that is having
3361 * minimal distance from the goal block.
3363 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3365 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3367 spin_lock(&pa
->pa_lock
);
3368 if (pa
->pa_deleted
== 0 &&
3369 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3371 cpa
= ext4_mb_check_group_pa(goal_block
,
3374 spin_unlock(&pa
->pa_lock
);
3379 ext4_mb_use_group_pa(ac
, cpa
);
3380 ac
->ac_criteria
= 20;
3387 * the function goes through all block freed in the group
3388 * but not yet committed and marks them used in in-core bitmap.
3389 * buddy must be generated from this bitmap
3390 * Need to be called with the ext4 group lock held
3392 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3396 struct ext4_group_info
*grp
;
3397 struct ext4_free_data
*entry
;
3399 grp
= ext4_get_group_info(sb
, group
);
3400 n
= rb_first(&(grp
->bb_free_root
));
3403 entry
= rb_entry(n
, struct ext4_free_data
, efd_node
);
3404 ext4_set_bits(bitmap
, entry
->efd_start_cluster
, entry
->efd_count
);
3411 * the function goes through all preallocation in this group and marks them
3412 * used in in-core bitmap. buddy must be generated from this bitmap
3413 * Need to be called with ext4 group lock held
3415 static noinline_for_stack
3416 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3419 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3420 struct ext4_prealloc_space
*pa
;
3421 struct list_head
*cur
;
3422 ext4_group_t groupnr
;
3423 ext4_grpblk_t start
;
3424 int preallocated
= 0;
3427 /* all form of preallocation discards first load group,
3428 * so the only competing code is preallocation use.
3429 * we don't need any locking here
3430 * notice we do NOT ignore preallocations with pa_deleted
3431 * otherwise we could leave used blocks available for
3432 * allocation in buddy when concurrent ext4_mb_put_pa()
3433 * is dropping preallocation
3435 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3436 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3437 spin_lock(&pa
->pa_lock
);
3438 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3441 spin_unlock(&pa
->pa_lock
);
3442 if (unlikely(len
== 0))
3444 BUG_ON(groupnr
!= group
);
3445 ext4_set_bits(bitmap
, start
, len
);
3446 preallocated
+= len
;
3448 mb_debug(1, "prellocated %u for group %u\n", preallocated
, group
);
3451 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3453 struct ext4_prealloc_space
*pa
;
3454 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3456 BUG_ON(atomic_read(&pa
->pa_count
));
3457 BUG_ON(pa
->pa_deleted
== 0);
3458 kmem_cache_free(ext4_pspace_cachep
, pa
);
3462 * drops a reference to preallocated space descriptor
3463 * if this was the last reference and the space is consumed
3465 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3466 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3469 ext4_fsblk_t grp_blk
;
3471 /* in this short window concurrent discard can set pa_deleted */
3472 spin_lock(&pa
->pa_lock
);
3473 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0) {
3474 spin_unlock(&pa
->pa_lock
);
3478 if (pa
->pa_deleted
== 1) {
3479 spin_unlock(&pa
->pa_lock
);
3484 spin_unlock(&pa
->pa_lock
);
3486 grp_blk
= pa
->pa_pstart
;
3488 * If doing group-based preallocation, pa_pstart may be in the
3489 * next group when pa is used up
3491 if (pa
->pa_type
== MB_GROUP_PA
)
3494 grp
= ext4_get_group_number(sb
, grp_blk
);
3499 * P1 (buddy init) P2 (regular allocation)
3500 * find block B in PA
3501 * copy on-disk bitmap to buddy
3502 * mark B in on-disk bitmap
3503 * drop PA from group
3504 * mark all PAs in buddy
3506 * thus, P1 initializes buddy with B available. to prevent this
3507 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3510 ext4_lock_group(sb
, grp
);
3511 list_del(&pa
->pa_group_list
);
3512 ext4_unlock_group(sb
, grp
);
3514 spin_lock(pa
->pa_obj_lock
);
3515 list_del_rcu(&pa
->pa_inode_list
);
3516 spin_unlock(pa
->pa_obj_lock
);
3518 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3522 * creates new preallocated space for given inode
3524 static noinline_for_stack
int
3525 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3527 struct super_block
*sb
= ac
->ac_sb
;
3528 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3529 struct ext4_prealloc_space
*pa
;
3530 struct ext4_group_info
*grp
;
3531 struct ext4_inode_info
*ei
;
3533 /* preallocate only when found space is larger then requested */
3534 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3535 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3536 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3538 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3542 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3548 /* we can't allocate as much as normalizer wants.
3549 * so, found space must get proper lstart
3550 * to cover original request */
3551 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3552 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3554 /* we're limited by original request in that
3555 * logical block must be covered any way
3556 * winl is window we can move our chunk within */
3557 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3559 /* also, we should cover whole original request */
3560 wins
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
);
3562 /* the smallest one defines real window */
3563 win
= min(winl
, wins
);
3565 offs
= ac
->ac_o_ex
.fe_logical
%
3566 EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
3567 if (offs
&& offs
< win
)
3570 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
-
3571 EXT4_NUM_B2C(sbi
, win
);
3572 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3573 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3576 /* preallocation can change ac_b_ex, thus we store actually
3577 * allocated blocks for history */
3578 ac
->ac_f_ex
= ac
->ac_b_ex
;
3580 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3581 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3582 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3583 pa
->pa_free
= pa
->pa_len
;
3584 atomic_set(&pa
->pa_count
, 1);
3585 spin_lock_init(&pa
->pa_lock
);
3586 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3587 INIT_LIST_HEAD(&pa
->pa_group_list
);
3589 pa
->pa_type
= MB_INODE_PA
;
3591 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3592 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3593 trace_ext4_mb_new_inode_pa(ac
, pa
);
3595 ext4_mb_use_inode_pa(ac
, pa
);
3596 atomic_add(pa
->pa_free
, &sbi
->s_mb_preallocated
);
3598 ei
= EXT4_I(ac
->ac_inode
);
3599 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3601 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3602 pa
->pa_inode
= ac
->ac_inode
;
3604 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3605 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3606 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3608 spin_lock(pa
->pa_obj_lock
);
3609 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3610 spin_unlock(pa
->pa_obj_lock
);
3616 * creates new preallocated space for locality group inodes belongs to
3618 static noinline_for_stack
int
3619 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3621 struct super_block
*sb
= ac
->ac_sb
;
3622 struct ext4_locality_group
*lg
;
3623 struct ext4_prealloc_space
*pa
;
3624 struct ext4_group_info
*grp
;
3626 /* preallocate only when found space is larger then requested */
3627 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3628 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3629 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3631 BUG_ON(ext4_pspace_cachep
== NULL
);
3632 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3636 /* preallocation can change ac_b_ex, thus we store actually
3637 * allocated blocks for history */
3638 ac
->ac_f_ex
= ac
->ac_b_ex
;
3640 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3641 pa
->pa_lstart
= pa
->pa_pstart
;
3642 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3643 pa
->pa_free
= pa
->pa_len
;
3644 atomic_set(&pa
->pa_count
, 1);
3645 spin_lock_init(&pa
->pa_lock
);
3646 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3647 INIT_LIST_HEAD(&pa
->pa_group_list
);
3649 pa
->pa_type
= MB_GROUP_PA
;
3651 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3652 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3653 trace_ext4_mb_new_group_pa(ac
, pa
);
3655 ext4_mb_use_group_pa(ac
, pa
);
3656 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3658 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3662 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3663 pa
->pa_inode
= NULL
;
3665 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3666 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3667 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3670 * We will later add the new pa to the right bucket
3671 * after updating the pa_free in ext4_mb_release_context
3676 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3680 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3681 err
= ext4_mb_new_group_pa(ac
);
3683 err
= ext4_mb_new_inode_pa(ac
);
3688 * finds all unused blocks in on-disk bitmap, frees them in
3689 * in-core bitmap and buddy.
3690 * @pa must be unlinked from inode and group lists, so that
3691 * nobody else can find/use it.
3692 * the caller MUST hold group/inode locks.
3693 * TODO: optimize the case when there are no in-core structures yet
3695 static noinline_for_stack
int
3696 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3697 struct ext4_prealloc_space
*pa
)
3699 struct super_block
*sb
= e4b
->bd_sb
;
3700 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3705 unsigned long long grp_blk_start
;
3709 BUG_ON(pa
->pa_deleted
== 0);
3710 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3711 grp_blk_start
= pa
->pa_pstart
- EXT4_C2B(sbi
, bit
);
3712 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3713 end
= bit
+ pa
->pa_len
;
3716 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3719 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3720 mb_debug(1, " free preallocated %u/%u in group %u\n",
3721 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3722 (unsigned) next
- bit
, (unsigned) group
);
3725 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3726 trace_ext4_mb_release_inode_pa(pa
, (grp_blk_start
+
3727 EXT4_C2B(sbi
, bit
)),
3729 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3732 if (free
!= pa
->pa_free
) {
3733 ext4_msg(e4b
->bd_sb
, KERN_CRIT
,
3734 "pa %p: logic %lu, phys. %lu, len %lu",
3735 pa
, (unsigned long) pa
->pa_lstart
,
3736 (unsigned long) pa
->pa_pstart
,
3737 (unsigned long) pa
->pa_len
);
3738 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3741 * pa is already deleted so we use the value obtained
3742 * from the bitmap and continue.
3745 atomic_add(free
, &sbi
->s_mb_discarded
);
3750 static noinline_for_stack
int
3751 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3752 struct ext4_prealloc_space
*pa
)
3754 struct super_block
*sb
= e4b
->bd_sb
;
3758 trace_ext4_mb_release_group_pa(sb
, pa
);
3759 BUG_ON(pa
->pa_deleted
== 0);
3760 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3761 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3762 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3763 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3764 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3770 * releases all preallocations in given group
3772 * first, we need to decide discard policy:
3773 * - when do we discard
3775 * - how many do we discard
3776 * 1) how many requested
3778 static noinline_for_stack
int
3779 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3780 ext4_group_t group
, int needed
)
3782 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3783 struct buffer_head
*bitmap_bh
= NULL
;
3784 struct ext4_prealloc_space
*pa
, *tmp
;
3785 struct list_head list
;
3786 struct ext4_buddy e4b
;
3791 mb_debug(1, "discard preallocation for group %u\n", group
);
3793 if (list_empty(&grp
->bb_prealloc_list
))
3796 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3797 if (bitmap_bh
== NULL
) {
3798 ext4_error(sb
, "Error reading block bitmap for %u", group
);
3802 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3804 ext4_error(sb
, "Error loading buddy information for %u", group
);
3810 needed
= EXT4_CLUSTERS_PER_GROUP(sb
) + 1;
3812 INIT_LIST_HEAD(&list
);
3814 ext4_lock_group(sb
, group
);
3815 list_for_each_entry_safe(pa
, tmp
,
3816 &grp
->bb_prealloc_list
, pa_group_list
) {
3817 spin_lock(&pa
->pa_lock
);
3818 if (atomic_read(&pa
->pa_count
)) {
3819 spin_unlock(&pa
->pa_lock
);
3823 if (pa
->pa_deleted
) {
3824 spin_unlock(&pa
->pa_lock
);
3828 /* seems this one can be freed ... */
3831 /* we can trust pa_free ... */
3832 free
+= pa
->pa_free
;
3834 spin_unlock(&pa
->pa_lock
);
3836 list_del(&pa
->pa_group_list
);
3837 list_add(&pa
->u
.pa_tmp_list
, &list
);
3840 /* if we still need more blocks and some PAs were used, try again */
3841 if (free
< needed
&& busy
) {
3843 ext4_unlock_group(sb
, group
);
3848 /* found anything to free? */
3849 if (list_empty(&list
)) {
3854 /* now free all selected PAs */
3855 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3857 /* remove from object (inode or locality group) */
3858 spin_lock(pa
->pa_obj_lock
);
3859 list_del_rcu(&pa
->pa_inode_list
);
3860 spin_unlock(pa
->pa_obj_lock
);
3862 if (pa
->pa_type
== MB_GROUP_PA
)
3863 ext4_mb_release_group_pa(&e4b
, pa
);
3865 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3867 list_del(&pa
->u
.pa_tmp_list
);
3868 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3872 ext4_unlock_group(sb
, group
);
3873 ext4_mb_unload_buddy(&e4b
);
3879 * releases all non-used preallocated blocks for given inode
3881 * It's important to discard preallocations under i_data_sem
3882 * We don't want another block to be served from the prealloc
3883 * space when we are discarding the inode prealloc space.
3885 * FIXME!! Make sure it is valid at all the call sites
3887 void ext4_discard_preallocations(struct inode
*inode
)
3889 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3890 struct super_block
*sb
= inode
->i_sb
;
3891 struct buffer_head
*bitmap_bh
= NULL
;
3892 struct ext4_prealloc_space
*pa
, *tmp
;
3893 ext4_group_t group
= 0;
3894 struct list_head list
;
3895 struct ext4_buddy e4b
;
3898 if (!S_ISREG(inode
->i_mode
)) {
3899 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3903 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
3904 trace_ext4_discard_preallocations(inode
);
3906 INIT_LIST_HEAD(&list
);
3909 /* first, collect all pa's in the inode */
3910 spin_lock(&ei
->i_prealloc_lock
);
3911 while (!list_empty(&ei
->i_prealloc_list
)) {
3912 pa
= list_entry(ei
->i_prealloc_list
.next
,
3913 struct ext4_prealloc_space
, pa_inode_list
);
3914 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3915 spin_lock(&pa
->pa_lock
);
3916 if (atomic_read(&pa
->pa_count
)) {
3917 /* this shouldn't happen often - nobody should
3918 * use preallocation while we're discarding it */
3919 spin_unlock(&pa
->pa_lock
);
3920 spin_unlock(&ei
->i_prealloc_lock
);
3921 ext4_msg(sb
, KERN_ERR
,
3922 "uh-oh! used pa while discarding");
3924 schedule_timeout_uninterruptible(HZ
);
3928 if (pa
->pa_deleted
== 0) {
3930 spin_unlock(&pa
->pa_lock
);
3931 list_del_rcu(&pa
->pa_inode_list
);
3932 list_add(&pa
->u
.pa_tmp_list
, &list
);
3936 /* someone is deleting pa right now */
3937 spin_unlock(&pa
->pa_lock
);
3938 spin_unlock(&ei
->i_prealloc_lock
);
3940 /* we have to wait here because pa_deleted
3941 * doesn't mean pa is already unlinked from
3942 * the list. as we might be called from
3943 * ->clear_inode() the inode will get freed
3944 * and concurrent thread which is unlinking
3945 * pa from inode's list may access already
3946 * freed memory, bad-bad-bad */
3948 /* XXX: if this happens too often, we can
3949 * add a flag to force wait only in case
3950 * of ->clear_inode(), but not in case of
3951 * regular truncate */
3952 schedule_timeout_uninterruptible(HZ
);
3955 spin_unlock(&ei
->i_prealloc_lock
);
3957 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3958 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
3959 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
3961 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3963 ext4_error(sb
, "Error loading buddy information for %u",
3968 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3969 if (bitmap_bh
== NULL
) {
3970 ext4_error(sb
, "Error reading block bitmap for %u",
3972 ext4_mb_unload_buddy(&e4b
);
3976 ext4_lock_group(sb
, group
);
3977 list_del(&pa
->pa_group_list
);
3978 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3979 ext4_unlock_group(sb
, group
);
3981 ext4_mb_unload_buddy(&e4b
);
3984 list_del(&pa
->u
.pa_tmp_list
);
3985 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3989 #ifdef CONFIG_EXT4_DEBUG
3990 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
3992 struct super_block
*sb
= ac
->ac_sb
;
3993 ext4_group_t ngroups
, i
;
3995 if (!ext4_mballoc_debug
||
3996 (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
))
3999 ext4_msg(ac
->ac_sb
, KERN_ERR
, "Can't allocate:"
4000 " Allocation context details:");
4001 ext4_msg(ac
->ac_sb
, KERN_ERR
, "status %d flags %d",
4002 ac
->ac_status
, ac
->ac_flags
);
4003 ext4_msg(ac
->ac_sb
, KERN_ERR
, "orig %lu/%lu/%lu@%lu, "
4004 "goal %lu/%lu/%lu@%lu, "
4005 "best %lu/%lu/%lu@%lu cr %d",
4006 (unsigned long)ac
->ac_o_ex
.fe_group
,
4007 (unsigned long)ac
->ac_o_ex
.fe_start
,
4008 (unsigned long)ac
->ac_o_ex
.fe_len
,
4009 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4010 (unsigned long)ac
->ac_g_ex
.fe_group
,
4011 (unsigned long)ac
->ac_g_ex
.fe_start
,
4012 (unsigned long)ac
->ac_g_ex
.fe_len
,
4013 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4014 (unsigned long)ac
->ac_b_ex
.fe_group
,
4015 (unsigned long)ac
->ac_b_ex
.fe_start
,
4016 (unsigned long)ac
->ac_b_ex
.fe_len
,
4017 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4018 (int)ac
->ac_criteria
);
4019 ext4_msg(ac
->ac_sb
, KERN_ERR
, "%d found", ac
->ac_found
);
4020 ext4_msg(ac
->ac_sb
, KERN_ERR
, "groups: ");
4021 ngroups
= ext4_get_groups_count(sb
);
4022 for (i
= 0; i
< ngroups
; i
++) {
4023 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4024 struct ext4_prealloc_space
*pa
;
4025 ext4_grpblk_t start
;
4026 struct list_head
*cur
;
4027 ext4_lock_group(sb
, i
);
4028 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4029 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4031 spin_lock(&pa
->pa_lock
);
4032 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4034 spin_unlock(&pa
->pa_lock
);
4035 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
4038 ext4_unlock_group(sb
, i
);
4040 if (grp
->bb_free
== 0)
4042 printk(KERN_ERR
"%u: %d/%d \n",
4043 i
, grp
->bb_free
, grp
->bb_fragments
);
4045 printk(KERN_ERR
"\n");
4048 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4055 * We use locality group preallocation for small size file. The size of the
4056 * file is determined by the current size or the resulting size after
4057 * allocation which ever is larger
4059 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4061 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4063 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4064 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4067 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4070 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4073 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
4074 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
4077 if ((size
== isize
) &&
4078 !ext4_fs_is_busy(sbi
) &&
4079 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
4080 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
4084 if (sbi
->s_mb_group_prealloc
<= 0) {
4085 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4089 /* don't use group allocation for large files */
4090 size
= max(size
, isize
);
4091 if (size
> sbi
->s_mb_stream_request
) {
4092 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4096 BUG_ON(ac
->ac_lg
!= NULL
);
4098 * locality group prealloc space are per cpu. The reason for having
4099 * per cpu locality group is to reduce the contention between block
4100 * request from multiple CPUs.
4102 ac
->ac_lg
= __this_cpu_ptr(sbi
->s_locality_groups
);
4104 /* we're going to use group allocation */
4105 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4107 /* serialize all allocations in the group */
4108 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4111 static noinline_for_stack
int
4112 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4113 struct ext4_allocation_request
*ar
)
4115 struct super_block
*sb
= ar
->inode
->i_sb
;
4116 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4117 struct ext4_super_block
*es
= sbi
->s_es
;
4121 ext4_grpblk_t block
;
4123 /* we can't allocate > group size */
4126 /* just a dirty hack to filter too big requests */
4127 if (len
>= EXT4_CLUSTERS_PER_GROUP(sb
))
4128 len
= EXT4_CLUSTERS_PER_GROUP(sb
);
4130 /* start searching from the goal */
4132 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4133 goal
>= ext4_blocks_count(es
))
4134 goal
= le32_to_cpu(es
->s_first_data_block
);
4135 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4137 /* set up allocation goals */
4138 ac
->ac_b_ex
.fe_logical
= EXT4_LBLK_CMASK(sbi
, ar
->logical
);
4139 ac
->ac_status
= AC_STATUS_CONTINUE
;
4141 ac
->ac_inode
= ar
->inode
;
4142 ac
->ac_o_ex
.fe_logical
= ac
->ac_b_ex
.fe_logical
;
4143 ac
->ac_o_ex
.fe_group
= group
;
4144 ac
->ac_o_ex
.fe_start
= block
;
4145 ac
->ac_o_ex
.fe_len
= len
;
4146 ac
->ac_g_ex
= ac
->ac_o_ex
;
4147 ac
->ac_flags
= ar
->flags
;
4149 /* we have to define context: we'll we work with a file or
4150 * locality group. this is a policy, actually */
4151 ext4_mb_group_or_file(ac
);
4153 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4154 "left: %u/%u, right %u/%u to %swritable\n",
4155 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4156 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4157 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4158 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4159 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4164 static noinline_for_stack
void
4165 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4166 struct ext4_locality_group
*lg
,
4167 int order
, int total_entries
)
4169 ext4_group_t group
= 0;
4170 struct ext4_buddy e4b
;
4171 struct list_head discard_list
;
4172 struct ext4_prealloc_space
*pa
, *tmp
;
4174 mb_debug(1, "discard locality group preallocation\n");
4176 INIT_LIST_HEAD(&discard_list
);
4178 spin_lock(&lg
->lg_prealloc_lock
);
4179 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4181 spin_lock(&pa
->pa_lock
);
4182 if (atomic_read(&pa
->pa_count
)) {
4184 * This is the pa that we just used
4185 * for block allocation. So don't
4188 spin_unlock(&pa
->pa_lock
);
4191 if (pa
->pa_deleted
) {
4192 spin_unlock(&pa
->pa_lock
);
4195 /* only lg prealloc space */
4196 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4198 /* seems this one can be freed ... */
4200 spin_unlock(&pa
->pa_lock
);
4202 list_del_rcu(&pa
->pa_inode_list
);
4203 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4206 if (total_entries
<= 5) {
4208 * we want to keep only 5 entries
4209 * allowing it to grow to 8. This
4210 * mak sure we don't call discard
4211 * soon for this list.
4216 spin_unlock(&lg
->lg_prealloc_lock
);
4218 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4220 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4221 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4222 ext4_error(sb
, "Error loading buddy information for %u",
4226 ext4_lock_group(sb
, group
);
4227 list_del(&pa
->pa_group_list
);
4228 ext4_mb_release_group_pa(&e4b
, pa
);
4229 ext4_unlock_group(sb
, group
);
4231 ext4_mb_unload_buddy(&e4b
);
4232 list_del(&pa
->u
.pa_tmp_list
);
4233 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4238 * We have incremented pa_count. So it cannot be freed at this
4239 * point. Also we hold lg_mutex. So no parallel allocation is
4240 * possible from this lg. That means pa_free cannot be updated.
4242 * A parallel ext4_mb_discard_group_preallocations is possible.
4243 * which can cause the lg_prealloc_list to be updated.
4246 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4248 int order
, added
= 0, lg_prealloc_count
= 1;
4249 struct super_block
*sb
= ac
->ac_sb
;
4250 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4251 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4253 order
= fls(pa
->pa_free
) - 1;
4254 if (order
> PREALLOC_TB_SIZE
- 1)
4255 /* The max size of hash table is PREALLOC_TB_SIZE */
4256 order
= PREALLOC_TB_SIZE
- 1;
4257 /* Add the prealloc space to lg */
4258 spin_lock(&lg
->lg_prealloc_lock
);
4259 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4261 spin_lock(&tmp_pa
->pa_lock
);
4262 if (tmp_pa
->pa_deleted
) {
4263 spin_unlock(&tmp_pa
->pa_lock
);
4266 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4267 /* Add to the tail of the previous entry */
4268 list_add_tail_rcu(&pa
->pa_inode_list
,
4269 &tmp_pa
->pa_inode_list
);
4272 * we want to count the total
4273 * number of entries in the list
4276 spin_unlock(&tmp_pa
->pa_lock
);
4277 lg_prealloc_count
++;
4280 list_add_tail_rcu(&pa
->pa_inode_list
,
4281 &lg
->lg_prealloc_list
[order
]);
4282 spin_unlock(&lg
->lg_prealloc_lock
);
4284 /* Now trim the list to be not more than 8 elements */
4285 if (lg_prealloc_count
> 8) {
4286 ext4_mb_discard_lg_preallocations(sb
, lg
,
4287 order
, lg_prealloc_count
);
4294 * release all resource we used in allocation
4296 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4298 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4299 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4301 if (pa
->pa_type
== MB_GROUP_PA
) {
4302 /* see comment in ext4_mb_use_group_pa() */
4303 spin_lock(&pa
->pa_lock
);
4304 pa
->pa_pstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4305 pa
->pa_lstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4306 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4307 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4308 spin_unlock(&pa
->pa_lock
);
4313 * We want to add the pa to the right bucket.
4314 * Remove it from the list and while adding
4315 * make sure the list to which we are adding
4318 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4319 spin_lock(pa
->pa_obj_lock
);
4320 list_del_rcu(&pa
->pa_inode_list
);
4321 spin_unlock(pa
->pa_obj_lock
);
4322 ext4_mb_add_n_trim(ac
);
4324 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4326 if (ac
->ac_bitmap_page
)
4327 page_cache_release(ac
->ac_bitmap_page
);
4328 if (ac
->ac_buddy_page
)
4329 page_cache_release(ac
->ac_buddy_page
);
4330 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4331 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4332 ext4_mb_collect_stats(ac
);
4336 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4338 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4342 trace_ext4_mb_discard_preallocations(sb
, needed
);
4343 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4344 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4353 * Main entry point into mballoc to allocate blocks
4354 * it tries to use preallocation first, then falls back
4355 * to usual allocation
4357 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4358 struct ext4_allocation_request
*ar
, int *errp
)
4361 struct ext4_allocation_context
*ac
= NULL
;
4362 struct ext4_sb_info
*sbi
;
4363 struct super_block
*sb
;
4364 ext4_fsblk_t block
= 0;
4365 unsigned int inquota
= 0;
4366 unsigned int reserv_clstrs
= 0;
4369 sb
= ar
->inode
->i_sb
;
4372 trace_ext4_request_blocks(ar
);
4374 /* Allow to use superuser reservation for quota file */
4375 if (IS_NOQUOTA(ar
->inode
))
4376 ar
->flags
|= EXT4_MB_USE_ROOT_BLOCKS
;
4379 * For delayed allocation, we could skip the ENOSPC and
4380 * EDQUOT check, as blocks and quotas have been already
4381 * reserved when data being copied into pagecache.
4383 if (ext4_test_inode_state(ar
->inode
, EXT4_STATE_DELALLOC_RESERVED
))
4384 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4386 /* Without delayed allocation we need to verify
4387 * there is enough free blocks to do block allocation
4388 * and verify allocation doesn't exceed the quota limits.
4391 ext4_claim_free_clusters(sbi
, ar
->len
, ar
->flags
)) {
4393 /* let others to free the space */
4395 ar
->len
= ar
->len
>> 1;
4401 reserv_clstrs
= ar
->len
;
4402 if (ar
->flags
& EXT4_MB_USE_ROOT_BLOCKS
) {
4403 dquot_alloc_block_nofail(ar
->inode
,
4404 EXT4_C2B(sbi
, ar
->len
));
4407 dquot_alloc_block(ar
->inode
,
4408 EXT4_C2B(sbi
, ar
->len
))) {
4410 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4421 ac
= kmem_cache_zalloc(ext4_ac_cachep
, GFP_NOFS
);
4428 *errp
= ext4_mb_initialize_context(ac
, ar
);
4434 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4435 if (!ext4_mb_use_preallocated(ac
)) {
4436 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4437 ext4_mb_normalize_request(ac
, ar
);
4439 /* allocate space in core */
4440 *errp
= ext4_mb_regular_allocator(ac
);
4442 goto discard_and_exit
;
4444 /* as we've just preallocated more space than
4445 * user requested originally, we store allocated
4446 * space in a special descriptor */
4447 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4448 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4449 *errp
= ext4_mb_new_preallocation(ac
);
4452 ext4_discard_allocated_blocks(ac
);
4456 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4457 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_clstrs
);
4458 if (*errp
== -EAGAIN
) {
4460 * drop the reference that we took
4461 * in ext4_mb_use_best_found
4463 ext4_mb_release_context(ac
);
4464 ac
->ac_b_ex
.fe_group
= 0;
4465 ac
->ac_b_ex
.fe_start
= 0;
4466 ac
->ac_b_ex
.fe_len
= 0;
4467 ac
->ac_status
= AC_STATUS_CONTINUE
;
4470 ext4_discard_allocated_blocks(ac
);
4473 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4474 ar
->len
= ac
->ac_b_ex
.fe_len
;
4477 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4485 ac
->ac_b_ex
.fe_len
= 0;
4487 ext4_mb_show_ac(ac
);
4489 ext4_mb_release_context(ac
);
4492 kmem_cache_free(ext4_ac_cachep
, ac
);
4493 if (inquota
&& ar
->len
< inquota
)
4494 dquot_free_block(ar
->inode
, EXT4_C2B(sbi
, inquota
- ar
->len
));
4496 if (!ext4_test_inode_state(ar
->inode
,
4497 EXT4_STATE_DELALLOC_RESERVED
))
4498 /* release all the reserved blocks if non delalloc */
4499 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
4503 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4509 * We can merge two free data extents only if the physical blocks
4510 * are contiguous, AND the extents were freed by the same transaction,
4511 * AND the blocks are associated with the same group.
4513 static int can_merge(struct ext4_free_data
*entry1
,
4514 struct ext4_free_data
*entry2
)
4516 if ((entry1
->efd_tid
== entry2
->efd_tid
) &&
4517 (entry1
->efd_group
== entry2
->efd_group
) &&
4518 ((entry1
->efd_start_cluster
+ entry1
->efd_count
) == entry2
->efd_start_cluster
))
4523 static noinline_for_stack
int
4524 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4525 struct ext4_free_data
*new_entry
)
4527 ext4_group_t group
= e4b
->bd_group
;
4528 ext4_grpblk_t cluster
;
4529 struct ext4_free_data
*entry
;
4530 struct ext4_group_info
*db
= e4b
->bd_info
;
4531 struct super_block
*sb
= e4b
->bd_sb
;
4532 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4533 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4534 struct rb_node
*parent
= NULL
, *new_node
;
4536 BUG_ON(!ext4_handle_valid(handle
));
4537 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4538 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4540 new_node
= &new_entry
->efd_node
;
4541 cluster
= new_entry
->efd_start_cluster
;
4544 /* first free block exent. We need to
4545 protect buddy cache from being freed,
4546 * otherwise we'll refresh it from
4547 * on-disk bitmap and lose not-yet-available
4549 page_cache_get(e4b
->bd_buddy_page
);
4550 page_cache_get(e4b
->bd_bitmap_page
);
4554 entry
= rb_entry(parent
, struct ext4_free_data
, efd_node
);
4555 if (cluster
< entry
->efd_start_cluster
)
4557 else if (cluster
>= (entry
->efd_start_cluster
+ entry
->efd_count
))
4558 n
= &(*n
)->rb_right
;
4560 ext4_grp_locked_error(sb
, group
, 0,
4561 ext4_group_first_block_no(sb
, group
) +
4562 EXT4_C2B(sbi
, cluster
),
4563 "Block already on to-be-freed list");
4568 rb_link_node(new_node
, parent
, n
);
4569 rb_insert_color(new_node
, &db
->bb_free_root
);
4571 /* Now try to see the extent can be merged to left and right */
4572 node
= rb_prev(new_node
);
4574 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4575 if (can_merge(entry
, new_entry
) &&
4576 ext4_journal_callback_try_del(handle
, &entry
->efd_jce
)) {
4577 new_entry
->efd_start_cluster
= entry
->efd_start_cluster
;
4578 new_entry
->efd_count
+= entry
->efd_count
;
4579 rb_erase(node
, &(db
->bb_free_root
));
4580 kmem_cache_free(ext4_free_data_cachep
, entry
);
4584 node
= rb_next(new_node
);
4586 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4587 if (can_merge(new_entry
, entry
) &&
4588 ext4_journal_callback_try_del(handle
, &entry
->efd_jce
)) {
4589 new_entry
->efd_count
+= entry
->efd_count
;
4590 rb_erase(node
, &(db
->bb_free_root
));
4591 kmem_cache_free(ext4_free_data_cachep
, entry
);
4594 /* Add the extent to transaction's private list */
4595 ext4_journal_callback_add(handle
, ext4_free_data_callback
,
4596 &new_entry
->efd_jce
);
4601 * ext4_free_blocks() -- Free given blocks and update quota
4602 * @handle: handle for this transaction
4604 * @block: start physical block to free
4605 * @count: number of blocks to count
4606 * @flags: flags used by ext4_free_blocks
4608 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4609 struct buffer_head
*bh
, ext4_fsblk_t block
,
4610 unsigned long count
, int flags
)
4612 struct buffer_head
*bitmap_bh
= NULL
;
4613 struct super_block
*sb
= inode
->i_sb
;
4614 struct ext4_group_desc
*gdp
;
4615 unsigned int overflow
;
4617 struct buffer_head
*gd_bh
;
4618 ext4_group_t block_group
;
4619 struct ext4_sb_info
*sbi
;
4620 struct ext4_inode_info
*ei
= EXT4_I(inode
);
4621 struct ext4_buddy e4b
;
4622 unsigned int count_clusters
;
4629 BUG_ON(block
!= bh
->b_blocknr
);
4631 block
= bh
->b_blocknr
;
4635 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4636 !ext4_data_block_valid(sbi
, block
, count
)) {
4637 ext4_error(sb
, "Freeing blocks not in datazone - "
4638 "block = %llu, count = %lu", block
, count
);
4642 ext4_debug("freeing block %llu\n", block
);
4643 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4645 if (flags
& EXT4_FREE_BLOCKS_FORGET
) {
4646 struct buffer_head
*tbh
= bh
;
4649 BUG_ON(bh
&& (count
> 1));
4651 for (i
= 0; i
< count
; i
++) {
4654 tbh
= sb_find_get_block(inode
->i_sb
,
4658 ext4_forget(handle
, flags
& EXT4_FREE_BLOCKS_METADATA
,
4659 inode
, tbh
, block
+ i
);
4664 * We need to make sure we don't reuse the freed block until
4665 * after the transaction is committed, which we can do by
4666 * treating the block as metadata, below. We make an
4667 * exception if the inode is to be written in writeback mode
4668 * since writeback mode has weak data consistency guarantees.
4670 if (!ext4_should_writeback_data(inode
))
4671 flags
|= EXT4_FREE_BLOCKS_METADATA
;
4674 * If the extent to be freed does not begin on a cluster
4675 * boundary, we need to deal with partial clusters at the
4676 * beginning and end of the extent. Normally we will free
4677 * blocks at the beginning or the end unless we are explicitly
4678 * requested to avoid doing so.
4680 overflow
= EXT4_PBLK_COFF(sbi
, block
);
4682 if (flags
& EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER
) {
4683 overflow
= sbi
->s_cluster_ratio
- overflow
;
4685 if (count
> overflow
)
4694 overflow
= EXT4_LBLK_COFF(sbi
, count
);
4696 if (flags
& EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER
) {
4697 if (count
> overflow
)
4702 count
+= sbi
->s_cluster_ratio
- overflow
;
4707 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4709 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4710 ext4_get_group_info(sb
, block_group
))))
4714 * Check to see if we are freeing blocks across a group
4717 if (EXT4_C2B(sbi
, bit
) + count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4718 overflow
= EXT4_C2B(sbi
, bit
) + count
-
4719 EXT4_BLOCKS_PER_GROUP(sb
);
4722 count_clusters
= EXT4_NUM_B2C(sbi
, count
);
4723 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4728 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4734 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4735 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4736 in_range(block
, ext4_inode_table(sb
, gdp
),
4737 EXT4_SB(sb
)->s_itb_per_group
) ||
4738 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4739 EXT4_SB(sb
)->s_itb_per_group
)) {
4741 ext4_error(sb
, "Freeing blocks in system zone - "
4742 "Block = %llu, count = %lu", block
, count
);
4743 /* err = 0. ext4_std_error should be a no op */
4747 BUFFER_TRACE(bitmap_bh
, "getting write access");
4748 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4753 * We are about to modify some metadata. Call the journal APIs
4754 * to unshare ->b_data if a currently-committing transaction is
4757 BUFFER_TRACE(gd_bh
, "get_write_access");
4758 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4761 #ifdef AGGRESSIVE_CHECK
4764 for (i
= 0; i
< count_clusters
; i
++)
4765 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4768 trace_ext4_mballoc_free(sb
, inode
, block_group
, bit
, count_clusters
);
4770 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4774 if ((flags
& EXT4_FREE_BLOCKS_METADATA
) && ext4_handle_valid(handle
)) {
4775 struct ext4_free_data
*new_entry
;
4777 * blocks being freed are metadata. these blocks shouldn't
4778 * be used until this transaction is committed
4781 new_entry
= kmem_cache_alloc(ext4_free_data_cachep
, GFP_NOFS
);
4784 * We use a retry loop because
4785 * ext4_free_blocks() is not allowed to fail.
4788 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
4791 new_entry
->efd_start_cluster
= bit
;
4792 new_entry
->efd_group
= block_group
;
4793 new_entry
->efd_count
= count_clusters
;
4794 new_entry
->efd_tid
= handle
->h_transaction
->t_tid
;
4796 ext4_lock_group(sb
, block_group
);
4797 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4798 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4800 /* need to update group_info->bb_free and bitmap
4801 * with group lock held. generate_buddy look at
4802 * them with group lock_held
4804 if (test_opt(sb
, DISCARD
)) {
4805 err
= ext4_issue_discard(sb
, block_group
, bit
, count
);
4806 if (err
&& err
!= -EOPNOTSUPP
)
4807 ext4_msg(sb
, KERN_WARNING
, "discard request in"
4808 " group:%d block:%d count:%lu failed"
4809 " with %d", block_group
, bit
, count
,
4812 EXT4_MB_GRP_CLEAR_TRIMMED(e4b
.bd_info
);
4814 ext4_lock_group(sb
, block_group
);
4815 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4816 mb_free_blocks(inode
, &e4b
, bit
, count_clusters
);
4819 ret
= ext4_free_group_clusters(sb
, gdp
) + count_clusters
;
4820 ext4_free_group_clusters_set(sb
, gdp
, ret
);
4821 ext4_block_bitmap_csum_set(sb
, block_group
, gdp
, bitmap_bh
);
4822 ext4_group_desc_csum_set(sb
, block_group
, gdp
);
4823 ext4_unlock_group(sb
, block_group
);
4825 if (sbi
->s_log_groups_per_flex
) {
4826 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4827 atomic64_add(count_clusters
,
4828 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
4831 if (flags
& EXT4_FREE_BLOCKS_RESERVE
&& ei
->i_reserved_data_blocks
) {
4832 percpu_counter_add(&sbi
->s_dirtyclusters_counter
,
4834 spin_lock(&ei
->i_block_reservation_lock
);
4835 if (flags
& EXT4_FREE_BLOCKS_METADATA
)
4836 ei
->i_reserved_meta_blocks
+= count_clusters
;
4838 ei
->i_reserved_data_blocks
+= count_clusters
;
4839 spin_unlock(&ei
->i_block_reservation_lock
);
4840 if (!(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4841 dquot_reclaim_block(inode
,
4842 EXT4_C2B(sbi
, count_clusters
));
4843 } else if (!(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4844 dquot_free_block(inode
, EXT4_C2B(sbi
, count_clusters
));
4845 percpu_counter_add(&sbi
->s_freeclusters_counter
, count_clusters
);
4847 ext4_mb_unload_buddy(&e4b
);
4849 /* We dirtied the bitmap block */
4850 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4851 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4853 /* And the group descriptor block */
4854 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4855 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4859 if (overflow
&& !err
) {
4867 ext4_std_error(sb
, err
);
4872 * ext4_group_add_blocks() -- Add given blocks to an existing group
4873 * @handle: handle to this transaction
4875 * @block: start physical block to add to the block group
4876 * @count: number of blocks to free
4878 * This marks the blocks as free in the bitmap and buddy.
4880 int ext4_group_add_blocks(handle_t
*handle
, struct super_block
*sb
,
4881 ext4_fsblk_t block
, unsigned long count
)
4883 struct buffer_head
*bitmap_bh
= NULL
;
4884 struct buffer_head
*gd_bh
;
4885 ext4_group_t block_group
;
4888 struct ext4_group_desc
*desc
;
4889 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4890 struct ext4_buddy e4b
;
4891 int err
= 0, ret
, blk_free_count
;
4892 ext4_grpblk_t blocks_freed
;
4894 ext4_debug("Adding block(s) %llu-%llu\n", block
, block
+ count
- 1);
4899 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4901 * Check to see if we are freeing blocks across a group
4904 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4905 ext4_warning(sb
, "too much blocks added to group %u\n",
4911 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4917 desc
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4923 if (in_range(ext4_block_bitmap(sb
, desc
), block
, count
) ||
4924 in_range(ext4_inode_bitmap(sb
, desc
), block
, count
) ||
4925 in_range(block
, ext4_inode_table(sb
, desc
), sbi
->s_itb_per_group
) ||
4926 in_range(block
+ count
- 1, ext4_inode_table(sb
, desc
),
4927 sbi
->s_itb_per_group
)) {
4928 ext4_error(sb
, "Adding blocks in system zones - "
4929 "Block = %llu, count = %lu",
4935 BUFFER_TRACE(bitmap_bh
, "getting write access");
4936 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4941 * We are about to modify some metadata. Call the journal APIs
4942 * to unshare ->b_data if a currently-committing transaction is
4945 BUFFER_TRACE(gd_bh
, "get_write_access");
4946 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4950 for (i
= 0, blocks_freed
= 0; i
< count
; i
++) {
4951 BUFFER_TRACE(bitmap_bh
, "clear bit");
4952 if (!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
)) {
4953 ext4_error(sb
, "bit already cleared for block %llu",
4954 (ext4_fsblk_t
)(block
+ i
));
4955 BUFFER_TRACE(bitmap_bh
, "bit already cleared");
4961 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4966 * need to update group_info->bb_free and bitmap
4967 * with group lock held. generate_buddy look at
4968 * them with group lock_held
4970 ext4_lock_group(sb
, block_group
);
4971 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4972 mb_free_blocks(NULL
, &e4b
, bit
, count
);
4973 blk_free_count
= blocks_freed
+ ext4_free_group_clusters(sb
, desc
);
4974 ext4_free_group_clusters_set(sb
, desc
, blk_free_count
);
4975 ext4_block_bitmap_csum_set(sb
, block_group
, desc
, bitmap_bh
);
4976 ext4_group_desc_csum_set(sb
, block_group
, desc
);
4977 ext4_unlock_group(sb
, block_group
);
4978 percpu_counter_add(&sbi
->s_freeclusters_counter
,
4979 EXT4_NUM_B2C(sbi
, blocks_freed
));
4981 if (sbi
->s_log_groups_per_flex
) {
4982 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4983 atomic64_add(EXT4_NUM_B2C(sbi
, blocks_freed
),
4984 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
4987 ext4_mb_unload_buddy(&e4b
);
4989 /* We dirtied the bitmap block */
4990 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4991 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4993 /* And the group descriptor block */
4994 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4995 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
5001 ext4_std_error(sb
, err
);
5006 * ext4_trim_extent -- function to TRIM one single free extent in the group
5007 * @sb: super block for the file system
5008 * @start: starting block of the free extent in the alloc. group
5009 * @count: number of blocks to TRIM
5010 * @group: alloc. group we are working with
5011 * @e4b: ext4 buddy for the group
5013 * Trim "count" blocks starting at "start" in the "group". To assure that no
5014 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5015 * be called with under the group lock.
5017 static int ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
5018 ext4_group_t group
, struct ext4_buddy
*e4b
)
5022 struct ext4_free_extent ex
;
5025 trace_ext4_trim_extent(sb
, group
, start
, count
);
5027 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
5029 ex
.fe_start
= start
;
5030 ex
.fe_group
= group
;
5034 * Mark blocks used, so no one can reuse them while
5037 mb_mark_used(e4b
, &ex
);
5038 ext4_unlock_group(sb
, group
);
5039 ret
= ext4_issue_discard(sb
, group
, start
, count
);
5040 ext4_lock_group(sb
, group
);
5041 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
5046 * ext4_trim_all_free -- function to trim all free space in alloc. group
5047 * @sb: super block for file system
5048 * @group: group to be trimmed
5049 * @start: first group block to examine
5050 * @max: last group block to examine
5051 * @minblocks: minimum extent block count
5053 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5054 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5058 * ext4_trim_all_free walks through group's block bitmap searching for free
5059 * extents. When the free extent is found, mark it as used in group buddy
5060 * bitmap. Then issue a TRIM command on this extent and free the extent in
5061 * the group buddy bitmap. This is done until whole group is scanned.
5063 static ext4_grpblk_t
5064 ext4_trim_all_free(struct super_block
*sb
, ext4_group_t group
,
5065 ext4_grpblk_t start
, ext4_grpblk_t max
,
5066 ext4_grpblk_t minblocks
)
5069 ext4_grpblk_t next
, count
= 0, free_count
= 0;
5070 struct ext4_buddy e4b
;
5073 trace_ext4_trim_all_free(sb
, group
, start
, max
);
5075 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
5077 ext4_error(sb
, "Error in loading buddy "
5078 "information for %u", group
);
5081 bitmap
= e4b
.bd_bitmap
;
5083 ext4_lock_group(sb
, group
);
5084 if (EXT4_MB_GRP_WAS_TRIMMED(e4b
.bd_info
) &&
5085 minblocks
>= atomic_read(&EXT4_SB(sb
)->s_last_trim_minblks
))
5088 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
5089 e4b
.bd_info
->bb_first_free
: start
;
5091 while (start
<= max
) {
5092 start
= mb_find_next_zero_bit(bitmap
, max
+ 1, start
);
5095 next
= mb_find_next_bit(bitmap
, max
+ 1, start
);
5097 if ((next
- start
) >= minblocks
) {
5098 ret
= ext4_trim_extent(sb
, start
,
5099 next
- start
, group
, &e4b
);
5100 if (ret
&& ret
!= -EOPNOTSUPP
)
5103 count
+= next
- start
;
5105 free_count
+= next
- start
;
5108 if (fatal_signal_pending(current
)) {
5109 count
= -ERESTARTSYS
;
5113 if (need_resched()) {
5114 ext4_unlock_group(sb
, group
);
5116 ext4_lock_group(sb
, group
);
5119 if ((e4b
.bd_info
->bb_free
- free_count
) < minblocks
)
5125 EXT4_MB_GRP_SET_TRIMMED(e4b
.bd_info
);
5128 ext4_unlock_group(sb
, group
);
5129 ext4_mb_unload_buddy(&e4b
);
5131 ext4_debug("trimmed %d blocks in the group %d\n",
5138 * ext4_trim_fs() -- trim ioctl handle function
5139 * @sb: superblock for filesystem
5140 * @range: fstrim_range structure
5142 * start: First Byte to trim
5143 * len: number of Bytes to trim from start
5144 * minlen: minimum extent length in Bytes
5145 * ext4_trim_fs goes through all allocation groups containing Bytes from
5146 * start to start+len. For each such a group ext4_trim_all_free function
5147 * is invoked to trim all free space.
5149 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
)
5151 struct ext4_group_info
*grp
;
5152 ext4_group_t group
, first_group
, last_group
;
5153 ext4_grpblk_t cnt
= 0, first_cluster
, last_cluster
;
5154 uint64_t start
, end
, minlen
, trimmed
= 0;
5155 ext4_fsblk_t first_data_blk
=
5156 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
5157 ext4_fsblk_t max_blks
= ext4_blocks_count(EXT4_SB(sb
)->s_es
);
5160 start
= range
->start
>> sb
->s_blocksize_bits
;
5161 end
= start
+ (range
->len
>> sb
->s_blocksize_bits
) - 1;
5162 minlen
= EXT4_NUM_B2C(EXT4_SB(sb
),
5163 range
->minlen
>> sb
->s_blocksize_bits
);
5165 if (minlen
> EXT4_CLUSTERS_PER_GROUP(sb
) ||
5166 start
>= max_blks
||
5167 range
->len
< sb
->s_blocksize
)
5169 if (end
>= max_blks
)
5171 if (end
<= first_data_blk
)
5173 if (start
< first_data_blk
)
5174 start
= first_data_blk
;
5176 /* Determine first and last group to examine based on start and end */
5177 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
5178 &first_group
, &first_cluster
);
5179 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) end
,
5180 &last_group
, &last_cluster
);
5182 /* end now represents the last cluster to discard in this group */
5183 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5185 for (group
= first_group
; group
<= last_group
; group
++) {
5186 grp
= ext4_get_group_info(sb
, group
);
5187 /* We only do this if the grp has never been initialized */
5188 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
5189 ret
= ext4_mb_init_group(sb
, group
);
5195 * For all the groups except the last one, last cluster will
5196 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5197 * change it for the last group, note that last_cluster is
5198 * already computed earlier by ext4_get_group_no_and_offset()
5200 if (group
== last_group
)
5203 if (grp
->bb_free
>= minlen
) {
5204 cnt
= ext4_trim_all_free(sb
, group
, first_cluster
,
5214 * For every group except the first one, we are sure
5215 * that the first cluster to discard will be cluster #0.
5221 atomic_set(&EXT4_SB(sb
)->s_last_trim_minblks
, minlen
);
5224 range
->len
= EXT4_C2B(EXT4_SB(sb
), trimmed
) << sb
->s_blocksize_bits
;