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
);
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",
757 * If we intent to continue, we consider group descritor
758 * corrupt and update bb_free using bitmap value
762 mb_set_largest_free_order(sb
, grp
);
764 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
766 period
= get_cycles() - period
;
767 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
768 EXT4_SB(sb
)->s_mb_buddies_generated
++;
769 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
770 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
773 static void mb_regenerate_buddy(struct ext4_buddy
*e4b
)
779 while ((buddy
= mb_find_buddy(e4b
, order
++, &count
))) {
780 ext4_set_bits(buddy
, 0, count
);
782 e4b
->bd_info
->bb_fragments
= 0;
783 memset(e4b
->bd_info
->bb_counters
, 0,
784 sizeof(*e4b
->bd_info
->bb_counters
) *
785 (e4b
->bd_sb
->s_blocksize_bits
+ 2));
787 ext4_mb_generate_buddy(e4b
->bd_sb
, e4b
->bd_buddy
,
788 e4b
->bd_bitmap
, e4b
->bd_group
);
791 /* The buddy information is attached the buddy cache inode
792 * for convenience. The information regarding each group
793 * is loaded via ext4_mb_load_buddy. The information involve
794 * block bitmap and buddy information. The information are
795 * stored in the inode as
798 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
801 * one block each for bitmap and buddy information.
802 * So for each group we take up 2 blocks. A page can
803 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
804 * So it can have information regarding groups_per_page which
805 * is blocks_per_page/2
807 * Locking note: This routine takes the block group lock of all groups
808 * for this page; do not hold this lock when calling this routine!
811 static int ext4_mb_init_cache(struct page
*page
, char *incore
, gfp_t gfp
)
813 ext4_group_t ngroups
;
819 ext4_group_t first_group
, group
;
821 struct super_block
*sb
;
822 struct buffer_head
*bhs
;
823 struct buffer_head
**bh
= NULL
;
827 struct ext4_group_info
*grinfo
;
829 mb_debug(1, "init page %lu\n", page
->index
);
831 inode
= page
->mapping
->host
;
833 ngroups
= ext4_get_groups_count(sb
);
834 blocksize
= 1 << inode
->i_blkbits
;
835 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
837 groups_per_page
= blocks_per_page
>> 1;
838 if (groups_per_page
== 0)
841 /* allocate buffer_heads to read bitmaps */
842 if (groups_per_page
> 1) {
843 i
= sizeof(struct buffer_head
*) * groups_per_page
;
844 bh
= kzalloc(i
, gfp
);
852 first_group
= page
->index
* blocks_per_page
/ 2;
854 /* read all groups the page covers into the cache */
855 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
856 if (group
>= ngroups
)
859 grinfo
= ext4_get_group_info(sb
, group
);
861 * If page is uptodate then we came here after online resize
862 * which added some new uninitialized group info structs, so
863 * we must skip all initialized uptodate buddies on the page,
864 * which may be currently in use by an allocating task.
866 if (PageUptodate(page
) && !EXT4_MB_GRP_NEED_INIT(grinfo
)) {
870 if (!(bh
[i
] = ext4_read_block_bitmap_nowait(sb
, group
))) {
874 mb_debug(1, "read bitmap for group %u\n", group
);
877 /* wait for I/O completion */
878 for (i
= 0, group
= first_group
; i
< groups_per_page
; i
++, group
++) {
879 if (bh
[i
] && ext4_wait_block_bitmap(sb
, group
, bh
[i
])) {
885 first_block
= page
->index
* blocks_per_page
;
886 for (i
= 0; i
< blocks_per_page
; i
++) {
887 group
= (first_block
+ i
) >> 1;
888 if (group
>= ngroups
)
891 if (!bh
[group
- first_group
])
892 /* skip initialized uptodate buddy */
896 * data carry information regarding this
897 * particular group in the format specified
901 data
= page_address(page
) + (i
* blocksize
);
902 bitmap
= bh
[group
- first_group
]->b_data
;
905 * We place the buddy block and bitmap block
908 if ((first_block
+ i
) & 1) {
909 /* this is block of buddy */
910 BUG_ON(incore
== NULL
);
911 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
912 group
, page
->index
, i
* blocksize
);
913 trace_ext4_mb_buddy_bitmap_load(sb
, group
);
914 grinfo
= ext4_get_group_info(sb
, group
);
915 grinfo
->bb_fragments
= 0;
916 memset(grinfo
->bb_counters
, 0,
917 sizeof(*grinfo
->bb_counters
) *
918 (sb
->s_blocksize_bits
+2));
920 * incore got set to the group block bitmap below
922 ext4_lock_group(sb
, group
);
924 memset(data
, 0xff, blocksize
);
925 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
926 ext4_unlock_group(sb
, group
);
929 /* this is block of bitmap */
930 BUG_ON(incore
!= NULL
);
931 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
932 group
, page
->index
, i
* blocksize
);
933 trace_ext4_mb_bitmap_load(sb
, group
);
935 /* see comments in ext4_mb_put_pa() */
936 ext4_lock_group(sb
, group
);
937 memcpy(data
, bitmap
, blocksize
);
939 /* mark all preallocated blks used in in-core bitmap */
940 ext4_mb_generate_from_pa(sb
, data
, group
);
941 ext4_mb_generate_from_freelist(sb
, data
, group
);
942 ext4_unlock_group(sb
, group
);
944 /* set incore so that the buddy information can be
945 * generated using this
950 SetPageUptodate(page
);
954 for (i
= 0; i
< groups_per_page
; i
++)
963 * Lock the buddy and bitmap pages. This make sure other parallel init_group
964 * on the same buddy page doesn't happen whild holding the buddy page lock.
965 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
966 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
968 static int ext4_mb_get_buddy_page_lock(struct super_block
*sb
,
969 ext4_group_t group
, struct ext4_buddy
*e4b
, gfp_t gfp
)
971 struct inode
*inode
= EXT4_SB(sb
)->s_buddy_cache
;
972 int block
, pnum
, poff
;
976 e4b
->bd_buddy_page
= NULL
;
977 e4b
->bd_bitmap_page
= NULL
;
979 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
981 * the buddy cache inode stores the block bitmap
982 * and buddy information in consecutive blocks.
983 * So for each group we need two blocks.
986 pnum
= block
/ blocks_per_page
;
987 poff
= block
% blocks_per_page
;
988 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
991 BUG_ON(page
->mapping
!= inode
->i_mapping
);
992 e4b
->bd_bitmap_page
= page
;
993 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
995 if (blocks_per_page
>= 2) {
996 /* buddy and bitmap are on the same page */
1001 pnum
= block
/ blocks_per_page
;
1002 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1005 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1006 e4b
->bd_buddy_page
= page
;
1010 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy
*e4b
)
1012 if (e4b
->bd_bitmap_page
) {
1013 unlock_page(e4b
->bd_bitmap_page
);
1014 page_cache_release(e4b
->bd_bitmap_page
);
1016 if (e4b
->bd_buddy_page
) {
1017 unlock_page(e4b
->bd_buddy_page
);
1018 page_cache_release(e4b
->bd_buddy_page
);
1023 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1024 * block group lock of all groups for this page; do not hold the BG lock when
1025 * calling this routine!
1027 static noinline_for_stack
1028 int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
, gfp_t gfp
)
1031 struct ext4_group_info
*this_grp
;
1032 struct ext4_buddy e4b
;
1037 mb_debug(1, "init group %u\n", group
);
1038 this_grp
= ext4_get_group_info(sb
, group
);
1040 * This ensures that we don't reinit the buddy cache
1041 * page which map to the group from which we are already
1042 * allocating. If we are looking at the buddy cache we would
1043 * have taken a reference using ext4_mb_load_buddy and that
1044 * would have pinned buddy page to page cache.
1046 ret
= ext4_mb_get_buddy_page_lock(sb
, group
, &e4b
, gfp
);
1047 if (ret
|| !EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1049 * somebody initialized the group
1050 * return without doing anything
1055 page
= e4b
.bd_bitmap_page
;
1056 ret
= ext4_mb_init_cache(page
, NULL
, gfp
);
1059 if (!PageUptodate(page
)) {
1063 mark_page_accessed(page
);
1065 if (e4b
.bd_buddy_page
== NULL
) {
1067 * If both the bitmap and buddy are in
1068 * the same page we don't need to force
1074 /* init buddy cache */
1075 page
= e4b
.bd_buddy_page
;
1076 ret
= ext4_mb_init_cache(page
, e4b
.bd_bitmap
, gfp
);
1079 if (!PageUptodate(page
)) {
1083 mark_page_accessed(page
);
1085 ext4_mb_put_buddy_page_lock(&e4b
);
1090 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1091 * block group lock of all groups for this page; do not hold the BG lock when
1092 * calling this routine!
1094 static noinline_for_stack
int
1095 ext4_mb_load_buddy_gfp(struct super_block
*sb
, ext4_group_t group
,
1096 struct ext4_buddy
*e4b
, gfp_t gfp
)
1098 int blocks_per_page
;
1104 struct ext4_group_info
*grp
;
1105 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1106 struct inode
*inode
= sbi
->s_buddy_cache
;
1109 mb_debug(1, "load group %u\n", group
);
1111 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1112 grp
= ext4_get_group_info(sb
, group
);
1114 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1117 e4b
->bd_group
= group
;
1118 e4b
->bd_buddy_page
= NULL
;
1119 e4b
->bd_bitmap_page
= NULL
;
1121 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
1123 * we need full data about the group
1124 * to make a good selection
1126 ret
= ext4_mb_init_group(sb
, group
, gfp
);
1132 * the buddy cache inode stores the block bitmap
1133 * and buddy information in consecutive blocks.
1134 * So for each group we need two blocks.
1137 pnum
= block
/ blocks_per_page
;
1138 poff
= block
% blocks_per_page
;
1140 /* we could use find_or_create_page(), but it locks page
1141 * what we'd like to avoid in fast path ... */
1142 page
= find_get_page(inode
->i_mapping
, pnum
);
1143 if (page
== NULL
|| !PageUptodate(page
)) {
1146 * drop the page reference and try
1147 * to get the page with lock. If we
1148 * are not uptodate that implies
1149 * somebody just created the page but
1150 * is yet to initialize the same. So
1151 * wait for it to initialize.
1153 page_cache_release(page
);
1154 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1156 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1157 if (!PageUptodate(page
)) {
1158 ret
= ext4_mb_init_cache(page
, NULL
, gfp
);
1163 mb_cmp_bitmaps(e4b
, page_address(page
) +
1164 (poff
* sb
->s_blocksize
));
1169 if (page
== NULL
|| !PageUptodate(page
)) {
1173 e4b
->bd_bitmap_page
= page
;
1174 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1175 mark_page_accessed(page
);
1178 pnum
= block
/ blocks_per_page
;
1179 poff
= block
% blocks_per_page
;
1181 page
= find_get_page(inode
->i_mapping
, pnum
);
1182 if (page
== NULL
|| !PageUptodate(page
)) {
1184 page_cache_release(page
);
1185 page
= find_or_create_page(inode
->i_mapping
, pnum
, gfp
);
1187 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1188 if (!PageUptodate(page
)) {
1189 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
,
1199 if (page
== NULL
|| !PageUptodate(page
)) {
1203 e4b
->bd_buddy_page
= page
;
1204 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1205 mark_page_accessed(page
);
1207 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1208 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1214 page_cache_release(page
);
1215 if (e4b
->bd_bitmap_page
)
1216 page_cache_release(e4b
->bd_bitmap_page
);
1217 if (e4b
->bd_buddy_page
)
1218 page_cache_release(e4b
->bd_buddy_page
);
1219 e4b
->bd_buddy
= NULL
;
1220 e4b
->bd_bitmap
= NULL
;
1224 static int ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1225 struct ext4_buddy
*e4b
)
1227 return ext4_mb_load_buddy_gfp(sb
, group
, e4b
, GFP_NOFS
);
1230 static void ext4_mb_unload_buddy(struct ext4_buddy
*e4b
)
1232 if (e4b
->bd_bitmap_page
)
1233 page_cache_release(e4b
->bd_bitmap_page
);
1234 if (e4b
->bd_buddy_page
)
1235 page_cache_release(e4b
->bd_buddy_page
);
1239 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1242 int bb_incr
= 1 << (e4b
->bd_blkbits
- 1);
1245 BUG_ON(e4b
->bd_bitmap
== e4b
->bd_buddy
);
1246 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1249 while (order
<= e4b
->bd_blkbits
+ 1) {
1251 if (!mb_test_bit(block
, bb
)) {
1252 /* this block is part of buddy of order 'order' */
1262 static void mb_clear_bits(void *bm
, int cur
, int len
)
1268 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1269 /* fast path: clear whole word at once */
1270 addr
= bm
+ (cur
>> 3);
1275 mb_clear_bit(cur
, bm
);
1280 /* clear bits in given range
1281 * will return first found zero bit if any, -1 otherwise
1283 static int mb_test_and_clear_bits(void *bm
, int cur
, int len
)
1290 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1291 /* fast path: clear whole word at once */
1292 addr
= bm
+ (cur
>> 3);
1293 if (*addr
!= (__u32
)(-1) && zero_bit
== -1)
1294 zero_bit
= cur
+ mb_find_next_zero_bit(addr
, 32, 0);
1299 if (!mb_test_and_clear_bit(cur
, bm
) && zero_bit
== -1)
1307 void ext4_set_bits(void *bm
, int cur
, int len
)
1313 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1314 /* fast path: set whole word at once */
1315 addr
= bm
+ (cur
>> 3);
1320 mb_set_bit(cur
, bm
);
1326 * _________________________________________________________________ */
1328 static inline int mb_buddy_adjust_border(int* bit
, void* bitmap
, int side
)
1330 if (mb_test_bit(*bit
+ side
, bitmap
)) {
1331 mb_clear_bit(*bit
, bitmap
);
1337 mb_set_bit(*bit
, bitmap
);
1342 static void mb_buddy_mark_free(struct ext4_buddy
*e4b
, int first
, int last
)
1346 void *buddy
= mb_find_buddy(e4b
, order
, &max
);
1351 /* Bits in range [first; last] are known to be set since
1352 * corresponding blocks were allocated. Bits in range
1353 * (first; last) will stay set because they form buddies on
1354 * upper layer. We just deal with borders if they don't
1355 * align with upper layer and then go up.
1356 * Releasing entire group is all about clearing
1357 * single bit of highest order buddy.
1361 * ---------------------------------
1363 * ---------------------------------
1364 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1365 * ---------------------------------
1367 * \_____________________/
1369 * Neither [1] nor [6] is aligned to above layer.
1370 * Left neighbour [0] is free, so mark it busy,
1371 * decrease bb_counters and extend range to
1373 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1374 * mark [6] free, increase bb_counters and shrink range to
1376 * Then shift range to [0; 2], go up and do the same.
1381 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&first
, buddy
, -1);
1383 e4b
->bd_info
->bb_counters
[order
] += mb_buddy_adjust_border(&last
, buddy
, 1);
1388 if (first
== last
|| !(buddy2
= mb_find_buddy(e4b
, order
, &max
))) {
1389 mb_clear_bits(buddy
, first
, last
- first
+ 1);
1390 e4b
->bd_info
->bb_counters
[order
- 1] += last
- first
+ 1;
1399 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1400 int first
, int count
)
1402 int left_is_free
= 0;
1403 int right_is_free
= 0;
1405 int last
= first
+ count
- 1;
1406 struct super_block
*sb
= e4b
->bd_sb
;
1408 if (WARN_ON(count
== 0))
1410 BUG_ON(last
>= (sb
->s_blocksize
<< 3));
1411 assert_spin_locked(ext4_group_lock_ptr(sb
, e4b
->bd_group
));
1412 mb_check_buddy(e4b
);
1413 mb_free_blocks_double(inode
, e4b
, first
, count
);
1415 e4b
->bd_info
->bb_free
+= count
;
1416 if (first
< e4b
->bd_info
->bb_first_free
)
1417 e4b
->bd_info
->bb_first_free
= first
;
1419 /* access memory sequentially: check left neighbour,
1420 * clear range and then check right neighbour
1423 left_is_free
= !mb_test_bit(first
- 1, e4b
->bd_bitmap
);
1424 block
= mb_test_and_clear_bits(e4b
->bd_bitmap
, first
, count
);
1425 if (last
+ 1 < EXT4_SB(sb
)->s_mb_maxs
[0])
1426 right_is_free
= !mb_test_bit(last
+ 1, e4b
->bd_bitmap
);
1428 if (unlikely(block
!= -1)) {
1429 ext4_fsblk_t blocknr
;
1431 blocknr
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1432 blocknr
+= EXT4_C2B(EXT4_SB(sb
), block
);
1433 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1434 inode
? inode
->i_ino
: 0,
1436 "freeing already freed block "
1438 mb_regenerate_buddy(e4b
);
1442 /* let's maintain fragments counter */
1443 if (left_is_free
&& right_is_free
)
1444 e4b
->bd_info
->bb_fragments
--;
1445 else if (!left_is_free
&& !right_is_free
)
1446 e4b
->bd_info
->bb_fragments
++;
1448 /* buddy[0] == bd_bitmap is a special case, so handle
1449 * it right away and let mb_buddy_mark_free stay free of
1450 * zero order checks.
1451 * Check if neighbours are to be coaleasced,
1452 * adjust bitmap bb_counters and borders appropriately.
1455 first
+= !left_is_free
;
1456 e4b
->bd_info
->bb_counters
[0] += left_is_free
? -1 : 1;
1459 last
-= !right_is_free
;
1460 e4b
->bd_info
->bb_counters
[0] += right_is_free
? -1 : 1;
1464 mb_buddy_mark_free(e4b
, first
>> 1, last
>> 1);
1467 mb_set_largest_free_order(sb
, e4b
->bd_info
);
1468 mb_check_buddy(e4b
);
1471 static int mb_find_extent(struct ext4_buddy
*e4b
, int block
,
1472 int needed
, struct ext4_free_extent
*ex
)
1478 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1481 buddy
= mb_find_buddy(e4b
, 0, &max
);
1482 BUG_ON(buddy
== NULL
);
1483 BUG_ON(block
>= max
);
1484 if (mb_test_bit(block
, buddy
)) {
1491 /* find actual order */
1492 order
= mb_find_order_for_block(e4b
, block
);
1493 block
= block
>> order
;
1495 ex
->fe_len
= 1 << order
;
1496 ex
->fe_start
= block
<< order
;
1497 ex
->fe_group
= e4b
->bd_group
;
1499 /* calc difference from given start */
1500 next
= next
- ex
->fe_start
;
1502 ex
->fe_start
+= next
;
1504 while (needed
> ex
->fe_len
&&
1505 mb_find_buddy(e4b
, order
, &max
)) {
1507 if (block
+ 1 >= max
)
1510 next
= (block
+ 1) * (1 << order
);
1511 if (mb_test_bit(next
, e4b
->bd_bitmap
))
1514 order
= mb_find_order_for_block(e4b
, next
);
1516 block
= next
>> order
;
1517 ex
->fe_len
+= 1 << order
;
1520 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1524 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1530 int start
= ex
->fe_start
;
1531 int len
= ex
->fe_len
;
1536 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1537 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1538 assert_spin_locked(ext4_group_lock_ptr(e4b
->bd_sb
, e4b
->bd_group
));
1539 mb_check_buddy(e4b
);
1540 mb_mark_used_double(e4b
, start
, len
);
1542 e4b
->bd_info
->bb_free
-= len
;
1543 if (e4b
->bd_info
->bb_first_free
== start
)
1544 e4b
->bd_info
->bb_first_free
+= len
;
1546 /* let's maintain fragments counter */
1548 mlen
= !mb_test_bit(start
- 1, e4b
->bd_bitmap
);
1549 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1550 max
= !mb_test_bit(start
+ len
, e4b
->bd_bitmap
);
1552 e4b
->bd_info
->bb_fragments
++;
1553 else if (!mlen
&& !max
)
1554 e4b
->bd_info
->bb_fragments
--;
1556 /* let's maintain buddy itself */
1558 ord
= mb_find_order_for_block(e4b
, start
);
1560 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1561 /* the whole chunk may be allocated at once! */
1563 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1564 BUG_ON((start
>> ord
) >= max
);
1565 mb_set_bit(start
>> ord
, buddy
);
1566 e4b
->bd_info
->bb_counters
[ord
]--;
1573 /* store for history */
1575 ret
= len
| (ord
<< 16);
1577 /* we have to split large buddy */
1579 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1580 mb_set_bit(start
>> ord
, buddy
);
1581 e4b
->bd_info
->bb_counters
[ord
]--;
1584 cur
= (start
>> ord
) & ~1U;
1585 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1586 mb_clear_bit(cur
, buddy
);
1587 mb_clear_bit(cur
+ 1, buddy
);
1588 e4b
->bd_info
->bb_counters
[ord
]++;
1589 e4b
->bd_info
->bb_counters
[ord
]++;
1591 mb_set_largest_free_order(e4b
->bd_sb
, e4b
->bd_info
);
1593 ext4_set_bits(e4b
->bd_bitmap
, ex
->fe_start
, len0
);
1594 mb_check_buddy(e4b
);
1600 * Must be called under group lock!
1602 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1603 struct ext4_buddy
*e4b
)
1605 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1608 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1609 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1611 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1612 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1613 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1615 /* preallocation can change ac_b_ex, thus we store actually
1616 * allocated blocks for history */
1617 ac
->ac_f_ex
= ac
->ac_b_ex
;
1619 ac
->ac_status
= AC_STATUS_FOUND
;
1620 ac
->ac_tail
= ret
& 0xffff;
1621 ac
->ac_buddy
= ret
>> 16;
1624 * take the page reference. We want the page to be pinned
1625 * so that we don't get a ext4_mb_init_cache_call for this
1626 * group until we update the bitmap. That would mean we
1627 * double allocate blocks. The reference is dropped
1628 * in ext4_mb_release_context
1630 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1631 get_page(ac
->ac_bitmap_page
);
1632 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1633 get_page(ac
->ac_buddy_page
);
1634 /* store last allocated for subsequent stream allocation */
1635 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
1636 spin_lock(&sbi
->s_md_lock
);
1637 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1638 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1639 spin_unlock(&sbi
->s_md_lock
);
1644 * regular allocator, for general purposes allocation
1647 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1648 struct ext4_buddy
*e4b
,
1651 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1652 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1653 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1654 struct ext4_free_extent ex
;
1657 if (ac
->ac_status
== AC_STATUS_FOUND
)
1660 * We don't want to scan for a whole year
1662 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1663 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1664 ac
->ac_status
= AC_STATUS_BREAK
;
1669 * Haven't found good chunk so far, let's continue
1671 if (bex
->fe_len
< gex
->fe_len
)
1674 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1675 && bex
->fe_group
== e4b
->bd_group
) {
1676 /* recheck chunk's availability - we don't know
1677 * when it was found (within this lock-unlock
1679 max
= mb_find_extent(e4b
, bex
->fe_start
, gex
->fe_len
, &ex
);
1680 if (max
>= gex
->fe_len
) {
1681 ext4_mb_use_best_found(ac
, e4b
);
1688 * The routine checks whether found extent is good enough. If it is,
1689 * then the extent gets marked used and flag is set to the context
1690 * to stop scanning. Otherwise, the extent is compared with the
1691 * previous found extent and if new one is better, then it's stored
1692 * in the context. Later, the best found extent will be used, if
1693 * mballoc can't find good enough extent.
1695 * FIXME: real allocation policy is to be designed yet!
1697 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1698 struct ext4_free_extent
*ex
,
1699 struct ext4_buddy
*e4b
)
1701 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1702 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1704 BUG_ON(ex
->fe_len
<= 0);
1705 BUG_ON(ex
->fe_len
> EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1706 BUG_ON(ex
->fe_start
>= EXT4_CLUSTERS_PER_GROUP(ac
->ac_sb
));
1707 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1712 * The special case - take what you catch first
1714 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1716 ext4_mb_use_best_found(ac
, e4b
);
1721 * Let's check whether the chuck is good enough
1723 if (ex
->fe_len
== gex
->fe_len
) {
1725 ext4_mb_use_best_found(ac
, e4b
);
1730 * If this is first found extent, just store it in the context
1732 if (bex
->fe_len
== 0) {
1738 * If new found extent is better, store it in the context
1740 if (bex
->fe_len
< gex
->fe_len
) {
1741 /* if the request isn't satisfied, any found extent
1742 * larger than previous best one is better */
1743 if (ex
->fe_len
> bex
->fe_len
)
1745 } else if (ex
->fe_len
> gex
->fe_len
) {
1746 /* if the request is satisfied, then we try to find
1747 * an extent that still satisfy the request, but is
1748 * smaller than previous one */
1749 if (ex
->fe_len
< bex
->fe_len
)
1753 ext4_mb_check_limits(ac
, e4b
, 0);
1756 static noinline_for_stack
1757 int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1758 struct ext4_buddy
*e4b
)
1760 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1761 ext4_group_t group
= ex
.fe_group
;
1765 BUG_ON(ex
.fe_len
<= 0);
1766 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1770 ext4_lock_group(ac
->ac_sb
, group
);
1771 max
= mb_find_extent(e4b
, ex
.fe_start
, ex
.fe_len
, &ex
);
1775 ext4_mb_use_best_found(ac
, e4b
);
1778 ext4_unlock_group(ac
->ac_sb
, group
);
1779 ext4_mb_unload_buddy(e4b
);
1784 static noinline_for_stack
1785 int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1786 struct ext4_buddy
*e4b
)
1788 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1791 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1792 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1793 struct ext4_free_extent ex
;
1795 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1797 if (grp
->bb_free
== 0)
1800 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1804 ext4_lock_group(ac
->ac_sb
, group
);
1805 max
= mb_find_extent(e4b
, ac
->ac_g_ex
.fe_start
,
1806 ac
->ac_g_ex
.fe_len
, &ex
);
1808 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1811 start
= ext4_group_first_block_no(ac
->ac_sb
, e4b
->bd_group
) +
1813 /* use do_div to get remainder (would be 64-bit modulo) */
1814 if (do_div(start
, sbi
->s_stripe
) == 0) {
1817 ext4_mb_use_best_found(ac
, e4b
);
1819 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1820 BUG_ON(ex
.fe_len
<= 0);
1821 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1822 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1825 ext4_mb_use_best_found(ac
, e4b
);
1826 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1827 /* Sometimes, caller may want to merge even small
1828 * number of blocks to an existing extent */
1829 BUG_ON(ex
.fe_len
<= 0);
1830 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1831 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1834 ext4_mb_use_best_found(ac
, e4b
);
1836 ext4_unlock_group(ac
->ac_sb
, group
);
1837 ext4_mb_unload_buddy(e4b
);
1843 * The routine scans buddy structures (not bitmap!) from given order
1844 * to max order and tries to find big enough chunk to satisfy the req
1846 static noinline_for_stack
1847 void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1848 struct ext4_buddy
*e4b
)
1850 struct super_block
*sb
= ac
->ac_sb
;
1851 struct ext4_group_info
*grp
= e4b
->bd_info
;
1857 BUG_ON(ac
->ac_2order
<= 0);
1858 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1859 if (grp
->bb_counters
[i
] == 0)
1862 buddy
= mb_find_buddy(e4b
, i
, &max
);
1863 BUG_ON(buddy
== NULL
);
1865 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1870 ac
->ac_b_ex
.fe_len
= 1 << i
;
1871 ac
->ac_b_ex
.fe_start
= k
<< i
;
1872 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1874 ext4_mb_use_best_found(ac
, e4b
);
1876 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1878 if (EXT4_SB(sb
)->s_mb_stats
)
1879 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1886 * The routine scans the group and measures all found extents.
1887 * In order to optimize scanning, caller must pass number of
1888 * free blocks in the group, so the routine can know upper limit.
1890 static noinline_for_stack
1891 void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1892 struct ext4_buddy
*e4b
)
1894 struct super_block
*sb
= ac
->ac_sb
;
1895 void *bitmap
= e4b
->bd_bitmap
;
1896 struct ext4_free_extent ex
;
1900 free
= e4b
->bd_info
->bb_free
;
1903 i
= e4b
->bd_info
->bb_first_free
;
1905 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1906 i
= mb_find_next_zero_bit(bitmap
,
1907 EXT4_CLUSTERS_PER_GROUP(sb
), i
);
1908 if (i
>= EXT4_CLUSTERS_PER_GROUP(sb
)) {
1910 * IF we have corrupt bitmap, we won't find any
1911 * free blocks even though group info says we
1912 * we have free blocks
1914 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1915 "%d free clusters as per "
1916 "group info. But bitmap says 0",
1921 mb_find_extent(e4b
, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1922 BUG_ON(ex
.fe_len
<= 0);
1923 if (free
< ex
.fe_len
) {
1924 ext4_grp_locked_error(sb
, e4b
->bd_group
, 0, 0,
1925 "%d free clusters as per "
1926 "group info. But got %d blocks",
1929 * The number of free blocks differs. This mostly
1930 * indicate that the bitmap is corrupt. So exit
1931 * without claiming the space.
1936 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1942 ext4_mb_check_limits(ac
, e4b
, 1);
1946 * This is a special case for storages like raid5
1947 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1949 static noinline_for_stack
1950 void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1951 struct ext4_buddy
*e4b
)
1953 struct super_block
*sb
= ac
->ac_sb
;
1954 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1955 void *bitmap
= e4b
->bd_bitmap
;
1956 struct ext4_free_extent ex
;
1957 ext4_fsblk_t first_group_block
;
1962 BUG_ON(sbi
->s_stripe
== 0);
1964 /* find first stripe-aligned block in group */
1965 first_group_block
= ext4_group_first_block_no(sb
, e4b
->bd_group
);
1967 a
= first_group_block
+ sbi
->s_stripe
- 1;
1968 do_div(a
, sbi
->s_stripe
);
1969 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1971 while (i
< EXT4_CLUSTERS_PER_GROUP(sb
)) {
1972 if (!mb_test_bit(i
, bitmap
)) {
1973 max
= mb_find_extent(e4b
, i
, sbi
->s_stripe
, &ex
);
1974 if (max
>= sbi
->s_stripe
) {
1977 ext4_mb_use_best_found(ac
, e4b
);
1985 /* This is now called BEFORE we load the buddy bitmap. */
1986 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1987 ext4_group_t group
, int cr
)
1989 unsigned free
, fragments
;
1990 int flex_size
= ext4_flex_bg_size(EXT4_SB(ac
->ac_sb
));
1991 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1993 BUG_ON(cr
< 0 || cr
>= 4);
1995 free
= grp
->bb_free
;
1998 if (cr
<= 2 && free
< ac
->ac_g_ex
.fe_len
)
2001 /* We only do this if the grp has never been initialized */
2002 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
2003 int ret
= ext4_mb_init_group(ac
->ac_sb
, group
, GFP_NOFS
);
2008 fragments
= grp
->bb_fragments
;
2014 BUG_ON(ac
->ac_2order
== 0);
2016 /* Avoid using the first bg of a flexgroup for data files */
2017 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
) &&
2018 (flex_size
>= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
) &&
2019 ((group
% flex_size
) == 0))
2022 if ((ac
->ac_2order
> ac
->ac_sb
->s_blocksize_bits
+1) ||
2023 (free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2026 if (grp
->bb_largest_free_order
< ac
->ac_2order
)
2031 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
2035 if (free
>= ac
->ac_g_ex
.fe_len
)
2047 static noinline_for_stack
int
2048 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
2050 ext4_group_t ngroups
, group
, i
;
2053 struct ext4_sb_info
*sbi
;
2054 struct super_block
*sb
;
2055 struct ext4_buddy e4b
;
2059 ngroups
= ext4_get_groups_count(sb
);
2060 /* non-extent files are limited to low blocks/groups */
2061 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)))
2062 ngroups
= sbi
->s_blockfile_groups
;
2064 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
2066 /* first, try the goal */
2067 err
= ext4_mb_find_by_goal(ac
, &e4b
);
2068 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
2071 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2075 * ac->ac2_order is set only if the fe_len is a power of 2
2076 * if ac2_order is set we also set criteria to 0 so that we
2077 * try exact allocation using buddy.
2079 i
= fls(ac
->ac_g_ex
.fe_len
);
2082 * We search using buddy data only if the order of the request
2083 * is greater than equal to the sbi_s_mb_order2_reqs
2084 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2086 if (i
>= sbi
->s_mb_order2_reqs
) {
2088 * This should tell if fe_len is exactly power of 2
2090 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2091 ac
->ac_2order
= i
- 1;
2094 /* if stream allocation is enabled, use global goal */
2095 if (ac
->ac_flags
& EXT4_MB_STREAM_ALLOC
) {
2096 /* TBD: may be hot point */
2097 spin_lock(&sbi
->s_md_lock
);
2098 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2099 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2100 spin_unlock(&sbi
->s_md_lock
);
2103 /* Let's just scan groups to find more-less suitable blocks */
2104 cr
= ac
->ac_2order
? 0 : 1;
2106 * cr == 0 try to get exact allocation,
2107 * cr == 3 try to get anything
2110 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2111 ac
->ac_criteria
= cr
;
2113 * searching for the right group start
2114 * from the goal value specified
2116 group
= ac
->ac_g_ex
.fe_group
;
2118 for (i
= 0; i
< ngroups
; group
++, i
++) {
2120 * Artificially restricted ngroups for non-extent
2121 * files makes group > ngroups possible on first loop.
2123 if (group
>= ngroups
)
2126 /* This now checks without needing the buddy page */
2127 if (!ext4_mb_good_group(ac
, group
, cr
))
2130 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2134 ext4_lock_group(sb
, group
);
2137 * We need to check again after locking the
2140 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2141 ext4_unlock_group(sb
, group
);
2142 ext4_mb_unload_buddy(&e4b
);
2146 ac
->ac_groups_scanned
++;
2147 if (cr
== 0 && ac
->ac_2order
< sb
->s_blocksize_bits
+2)
2148 ext4_mb_simple_scan_group(ac
, &e4b
);
2149 else if (cr
== 1 && sbi
->s_stripe
&&
2150 !(ac
->ac_g_ex
.fe_len
% sbi
->s_stripe
))
2151 ext4_mb_scan_aligned(ac
, &e4b
);
2153 ext4_mb_complex_scan_group(ac
, &e4b
);
2155 ext4_unlock_group(sb
, group
);
2156 ext4_mb_unload_buddy(&e4b
);
2158 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2163 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2164 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2166 * We've been searching too long. Let's try to allocate
2167 * the best chunk we've found so far
2170 ext4_mb_try_best_found(ac
, &e4b
);
2171 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2173 * Someone more lucky has already allocated it.
2174 * The only thing we can do is just take first
2176 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2178 ac
->ac_b_ex
.fe_group
= 0;
2179 ac
->ac_b_ex
.fe_start
= 0;
2180 ac
->ac_b_ex
.fe_len
= 0;
2181 ac
->ac_status
= AC_STATUS_CONTINUE
;
2182 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2184 atomic_inc(&sbi
->s_mb_lost_chunks
);
2192 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2194 struct super_block
*sb
= seq
->private;
2197 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2200 return (void *) ((unsigned long) group
);
2203 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2205 struct super_block
*sb
= seq
->private;
2209 if (*pos
< 0 || *pos
>= ext4_get_groups_count(sb
))
2212 return (void *) ((unsigned long) group
);
2215 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2217 struct super_block
*sb
= seq
->private;
2218 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2220 int err
, buddy_loaded
= 0;
2221 struct ext4_buddy e4b
;
2222 struct ext4_group_info
*grinfo
;
2224 struct ext4_group_info info
;
2225 ext4_grpblk_t counters
[EXT4_MAX_BLOCK_LOG_SIZE
+ 2];
2230 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2231 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2232 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2233 "group", "free", "frags", "first",
2234 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2235 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2237 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2238 sizeof(struct ext4_group_info
);
2239 grinfo
= ext4_get_group_info(sb
, group
);
2240 /* Load the group info in memory only if not already loaded. */
2241 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo
))) {
2242 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2244 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2250 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2253 ext4_mb_unload_buddy(&e4b
);
2255 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2256 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2257 for (i
= 0; i
<= 13; i
++)
2258 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2259 sg
.info
.bb_counters
[i
] : 0);
2260 seq_printf(seq
, " ]\n");
2265 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2269 static const struct seq_operations ext4_mb_seq_groups_ops
= {
2270 .start
= ext4_mb_seq_groups_start
,
2271 .next
= ext4_mb_seq_groups_next
,
2272 .stop
= ext4_mb_seq_groups_stop
,
2273 .show
= ext4_mb_seq_groups_show
,
2276 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2278 struct super_block
*sb
= PDE_DATA(inode
);
2281 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2283 struct seq_file
*m
= file
->private_data
;
2290 static const struct file_operations ext4_mb_seq_groups_fops
= {
2291 .owner
= THIS_MODULE
,
2292 .open
= ext4_mb_seq_groups_open
,
2294 .llseek
= seq_lseek
,
2295 .release
= seq_release
,
2298 static struct kmem_cache
*get_groupinfo_cache(int blocksize_bits
)
2300 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2301 struct kmem_cache
*cachep
= ext4_groupinfo_caches
[cache_index
];
2308 * Allocate the top-level s_group_info array for the specified number
2311 int ext4_mb_alloc_groupinfo(struct super_block
*sb
, ext4_group_t ngroups
)
2313 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2315 struct ext4_group_info
***new_groupinfo
;
2317 size
= (ngroups
+ EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2318 EXT4_DESC_PER_BLOCK_BITS(sb
);
2319 if (size
<= sbi
->s_group_info_size
)
2322 size
= roundup_pow_of_two(sizeof(*sbi
->s_group_info
) * size
);
2323 new_groupinfo
= ext4_kvzalloc(size
, GFP_KERNEL
);
2324 if (!new_groupinfo
) {
2325 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy meta group");
2328 if (sbi
->s_group_info
) {
2329 memcpy(new_groupinfo
, sbi
->s_group_info
,
2330 sbi
->s_group_info_size
* sizeof(*sbi
->s_group_info
));
2331 ext4_kvfree(sbi
->s_group_info
);
2333 sbi
->s_group_info
= new_groupinfo
;
2334 sbi
->s_group_info_size
= size
/ sizeof(*sbi
->s_group_info
);
2335 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2336 sbi
->s_group_info_size
);
2340 /* Create and initialize ext4_group_info data for the given group. */
2341 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2342 struct ext4_group_desc
*desc
)
2346 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2347 struct ext4_group_info
**meta_group_info
;
2348 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2351 * First check if this group is the first of a reserved block.
2352 * If it's true, we have to allocate a new table of pointers
2353 * to ext4_group_info structures
2355 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2356 metalen
= sizeof(*meta_group_info
) <<
2357 EXT4_DESC_PER_BLOCK_BITS(sb
);
2358 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2359 if (meta_group_info
== NULL
) {
2360 ext4_msg(sb
, KERN_ERR
, "can't allocate mem "
2361 "for a buddy group");
2362 goto exit_meta_group_info
;
2364 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2369 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2370 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2372 meta_group_info
[i
] = kmem_cache_zalloc(cachep
, GFP_KERNEL
);
2373 if (meta_group_info
[i
] == NULL
) {
2374 ext4_msg(sb
, KERN_ERR
, "can't allocate buddy mem");
2375 goto exit_group_info
;
2377 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2378 &(meta_group_info
[i
]->bb_state
));
2381 * initialize bb_free to be able to skip
2382 * empty groups without initialization
2384 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2385 meta_group_info
[i
]->bb_free
=
2386 ext4_free_clusters_after_init(sb
, group
, desc
);
2388 meta_group_info
[i
]->bb_free
=
2389 ext4_free_group_clusters(sb
, desc
);
2392 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2393 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2394 meta_group_info
[i
]->bb_free_root
= RB_ROOT
;
2395 meta_group_info
[i
]->bb_largest_free_order
= -1; /* uninit */
2399 struct buffer_head
*bh
;
2400 meta_group_info
[i
]->bb_bitmap
=
2401 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2402 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2403 bh
= ext4_read_block_bitmap(sb
, group
);
2405 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2414 /* If a meta_group_info table has been allocated, release it now */
2415 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2416 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2417 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] = NULL
;
2419 exit_meta_group_info
:
2421 } /* ext4_mb_add_groupinfo */
2423 static int ext4_mb_init_backend(struct super_block
*sb
)
2425 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2427 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2429 struct ext4_group_desc
*desc
;
2430 struct kmem_cache
*cachep
;
2432 err
= ext4_mb_alloc_groupinfo(sb
, ngroups
);
2436 sbi
->s_buddy_cache
= new_inode(sb
);
2437 if (sbi
->s_buddy_cache
== NULL
) {
2438 ext4_msg(sb
, KERN_ERR
, "can't get new inode");
2441 /* To avoid potentially colliding with an valid on-disk inode number,
2442 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2443 * not in the inode hash, so it should never be found by iget(), but
2444 * this will avoid confusion if it ever shows up during debugging. */
2445 sbi
->s_buddy_cache
->i_ino
= EXT4_BAD_INO
;
2446 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2447 for (i
= 0; i
< ngroups
; i
++) {
2448 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2450 ext4_msg(sb
, KERN_ERR
, "can't read descriptor %u", i
);
2453 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2460 cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2462 kmem_cache_free(cachep
, ext4_get_group_info(sb
, i
));
2463 i
= sbi
->s_group_info_size
;
2465 kfree(sbi
->s_group_info
[i
]);
2466 iput(sbi
->s_buddy_cache
);
2468 ext4_kvfree(sbi
->s_group_info
);
2472 static void ext4_groupinfo_destroy_slabs(void)
2476 for (i
= 0; i
< NR_GRPINFO_CACHES
; i
++) {
2477 if (ext4_groupinfo_caches
[i
])
2478 kmem_cache_destroy(ext4_groupinfo_caches
[i
]);
2479 ext4_groupinfo_caches
[i
] = NULL
;
2483 static int ext4_groupinfo_create_slab(size_t size
)
2485 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex
);
2487 int blocksize_bits
= order_base_2(size
);
2488 int cache_index
= blocksize_bits
- EXT4_MIN_BLOCK_LOG_SIZE
;
2489 struct kmem_cache
*cachep
;
2491 if (cache_index
>= NR_GRPINFO_CACHES
)
2494 if (unlikely(cache_index
< 0))
2497 mutex_lock(&ext4_grpinfo_slab_create_mutex
);
2498 if (ext4_groupinfo_caches
[cache_index
]) {
2499 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2500 return 0; /* Already created */
2503 slab_size
= offsetof(struct ext4_group_info
,
2504 bb_counters
[blocksize_bits
+ 2]);
2506 cachep
= kmem_cache_create(ext4_groupinfo_slab_names
[cache_index
],
2507 slab_size
, 0, SLAB_RECLAIM_ACCOUNT
,
2510 ext4_groupinfo_caches
[cache_index
] = cachep
;
2512 mutex_unlock(&ext4_grpinfo_slab_create_mutex
);
2515 "EXT4-fs: no memory for groupinfo slab cache\n");
2522 int ext4_mb_init(struct super_block
*sb
)
2524 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2526 unsigned offset
, offset_incr
;
2530 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_offsets
);
2532 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2533 if (sbi
->s_mb_offsets
== NULL
) {
2538 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(*sbi
->s_mb_maxs
);
2539 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2540 if (sbi
->s_mb_maxs
== NULL
) {
2545 ret
= ext4_groupinfo_create_slab(sb
->s_blocksize
);
2549 /* order 0 is regular bitmap */
2550 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2551 sbi
->s_mb_offsets
[0] = 0;
2555 offset_incr
= 1 << (sb
->s_blocksize_bits
- 1);
2556 max
= sb
->s_blocksize
<< 2;
2558 sbi
->s_mb_offsets
[i
] = offset
;
2559 sbi
->s_mb_maxs
[i
] = max
;
2560 offset
+= offset_incr
;
2561 offset_incr
= offset_incr
>> 1;
2564 } while (i
<= sb
->s_blocksize_bits
+ 1);
2566 spin_lock_init(&sbi
->s_md_lock
);
2567 spin_lock_init(&sbi
->s_bal_lock
);
2569 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2570 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2571 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2572 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2573 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2575 * The default group preallocation is 512, which for 4k block
2576 * sizes translates to 2 megabytes. However for bigalloc file
2577 * systems, this is probably too big (i.e, if the cluster size
2578 * is 1 megabyte, then group preallocation size becomes half a
2579 * gigabyte!). As a default, we will keep a two megabyte
2580 * group pralloc size for cluster sizes up to 64k, and after
2581 * that, we will force a minimum group preallocation size of
2582 * 32 clusters. This translates to 8 megs when the cluster
2583 * size is 256k, and 32 megs when the cluster size is 1 meg,
2584 * which seems reasonable as a default.
2586 sbi
->s_mb_group_prealloc
= max(MB_DEFAULT_GROUP_PREALLOC
>>
2587 sbi
->s_cluster_bits
, 32);
2589 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2590 * to the lowest multiple of s_stripe which is bigger than
2591 * the s_mb_group_prealloc as determined above. We want
2592 * the preallocation size to be an exact multiple of the
2593 * RAID stripe size so that preallocations don't fragment
2596 if (sbi
->s_stripe
> 1) {
2597 sbi
->s_mb_group_prealloc
= roundup(
2598 sbi
->s_mb_group_prealloc
, sbi
->s_stripe
);
2601 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2602 if (sbi
->s_locality_groups
== NULL
) {
2604 goto out_free_groupinfo_slab
;
2606 for_each_possible_cpu(i
) {
2607 struct ext4_locality_group
*lg
;
2608 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2609 mutex_init(&lg
->lg_mutex
);
2610 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2611 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2612 spin_lock_init(&lg
->lg_prealloc_lock
);
2615 /* init file for buddy data */
2616 ret
= ext4_mb_init_backend(sb
);
2618 goto out_free_locality_groups
;
2621 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2622 &ext4_mb_seq_groups_fops
, sb
);
2626 out_free_locality_groups
:
2627 free_percpu(sbi
->s_locality_groups
);
2628 sbi
->s_locality_groups
= NULL
;
2629 out_free_groupinfo_slab
:
2630 ext4_groupinfo_destroy_slabs();
2632 kfree(sbi
->s_mb_offsets
);
2633 sbi
->s_mb_offsets
= NULL
;
2634 kfree(sbi
->s_mb_maxs
);
2635 sbi
->s_mb_maxs
= NULL
;
2639 /* need to called with the ext4 group lock held */
2640 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2642 struct ext4_prealloc_space
*pa
;
2643 struct list_head
*cur
, *tmp
;
2646 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2647 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2648 list_del(&pa
->pa_group_list
);
2650 kmem_cache_free(ext4_pspace_cachep
, pa
);
2653 mb_debug(1, "mballoc: %u PAs left\n", count
);
2657 int ext4_mb_release(struct super_block
*sb
)
2659 ext4_group_t ngroups
= ext4_get_groups_count(sb
);
2661 int num_meta_group_infos
;
2662 struct ext4_group_info
*grinfo
;
2663 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2664 struct kmem_cache
*cachep
= get_groupinfo_cache(sb
->s_blocksize_bits
);
2667 remove_proc_entry("mb_groups", sbi
->s_proc
);
2669 if (sbi
->s_group_info
) {
2670 for (i
= 0; i
< ngroups
; i
++) {
2671 grinfo
= ext4_get_group_info(sb
, i
);
2673 kfree(grinfo
->bb_bitmap
);
2675 ext4_lock_group(sb
, i
);
2676 ext4_mb_cleanup_pa(grinfo
);
2677 ext4_unlock_group(sb
, i
);
2678 kmem_cache_free(cachep
, grinfo
);
2680 num_meta_group_infos
= (ngroups
+
2681 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2682 EXT4_DESC_PER_BLOCK_BITS(sb
);
2683 for (i
= 0; i
< num_meta_group_infos
; i
++)
2684 kfree(sbi
->s_group_info
[i
]);
2685 ext4_kvfree(sbi
->s_group_info
);
2687 kfree(sbi
->s_mb_offsets
);
2688 kfree(sbi
->s_mb_maxs
);
2689 if (sbi
->s_buddy_cache
)
2690 iput(sbi
->s_buddy_cache
);
2691 if (sbi
->s_mb_stats
) {
2692 ext4_msg(sb
, KERN_INFO
,
2693 "mballoc: %u blocks %u reqs (%u success)",
2694 atomic_read(&sbi
->s_bal_allocated
),
2695 atomic_read(&sbi
->s_bal_reqs
),
2696 atomic_read(&sbi
->s_bal_success
));
2697 ext4_msg(sb
, KERN_INFO
,
2698 "mballoc: %u extents scanned, %u goal hits, "
2699 "%u 2^N hits, %u breaks, %u lost",
2700 atomic_read(&sbi
->s_bal_ex_scanned
),
2701 atomic_read(&sbi
->s_bal_goals
),
2702 atomic_read(&sbi
->s_bal_2orders
),
2703 atomic_read(&sbi
->s_bal_breaks
),
2704 atomic_read(&sbi
->s_mb_lost_chunks
));
2705 ext4_msg(sb
, KERN_INFO
,
2706 "mballoc: %lu generated and it took %Lu",
2707 sbi
->s_mb_buddies_generated
,
2708 sbi
->s_mb_generation_time
);
2709 ext4_msg(sb
, KERN_INFO
,
2710 "mballoc: %u preallocated, %u discarded",
2711 atomic_read(&sbi
->s_mb_preallocated
),
2712 atomic_read(&sbi
->s_mb_discarded
));
2715 free_percpu(sbi
->s_locality_groups
);
2720 static inline int ext4_issue_discard(struct super_block
*sb
,
2721 ext4_group_t block_group
, ext4_grpblk_t cluster
, int count
,
2722 unsigned long flags
)
2724 ext4_fsblk_t discard_block
;
2726 discard_block
= (EXT4_C2B(EXT4_SB(sb
), cluster
) +
2727 ext4_group_first_block_no(sb
, block_group
));
2728 count
= EXT4_C2B(EXT4_SB(sb
), count
);
2729 trace_ext4_discard_blocks(sb
,
2730 (unsigned long long) discard_block
, count
);
2731 return sb_issue_discard(sb
, discard_block
, count
, GFP_NOFS
, flags
);
2735 * This function is called by the jbd2 layer once the commit has finished,
2736 * so we know we can free the blocks that were released with that commit.
2738 static void ext4_free_data_callback(struct super_block
*sb
,
2739 struct ext4_journal_cb_entry
*jce
,
2742 struct ext4_free_data
*entry
= (struct ext4_free_data
*)jce
;
2743 struct ext4_buddy e4b
;
2744 struct ext4_group_info
*db
;
2745 int err
, count
= 0, count2
= 0;
2747 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2748 entry
->efd_count
, entry
->efd_group
, entry
);
2750 if (test_opt(sb
, DISCARD
)) {
2751 err
= ext4_issue_discard(sb
, entry
->efd_group
,
2752 entry
->efd_start_cluster
,
2753 entry
->efd_count
, 0);
2754 if (err
&& err
!= -EOPNOTSUPP
)
2755 ext4_msg(sb
, KERN_WARNING
, "discard request in"
2756 " group:%d block:%d count:%d failed"
2757 " with %d", entry
->efd_group
,
2758 entry
->efd_start_cluster
,
2759 entry
->efd_count
, err
);
2762 err
= ext4_mb_load_buddy(sb
, entry
->efd_group
, &e4b
);
2763 /* we expect to find existing buddy because it's pinned */
2768 /* there are blocks to put in buddy to make them really free */
2769 count
+= entry
->efd_count
;
2771 ext4_lock_group(sb
, entry
->efd_group
);
2772 /* Take it out of per group rb tree */
2773 rb_erase(&entry
->efd_node
, &(db
->bb_free_root
));
2774 mb_free_blocks(NULL
, &e4b
, entry
->efd_start_cluster
, entry
->efd_count
);
2777 * Clear the trimmed flag for the group so that the next
2778 * ext4_trim_fs can trim it.
2779 * If the volume is mounted with -o discard, online discard
2780 * is supported and the free blocks will be trimmed online.
2782 if (!test_opt(sb
, DISCARD
))
2783 EXT4_MB_GRP_CLEAR_TRIMMED(db
);
2785 if (!db
->bb_free_root
.rb_node
) {
2786 /* No more items in the per group rb tree
2787 * balance refcounts from ext4_mb_free_metadata()
2789 page_cache_release(e4b
.bd_buddy_page
);
2790 page_cache_release(e4b
.bd_bitmap_page
);
2792 ext4_unlock_group(sb
, entry
->efd_group
);
2793 kmem_cache_free(ext4_free_data_cachep
, entry
);
2794 ext4_mb_unload_buddy(&e4b
);
2796 mb_debug(1, "freed %u blocks in %u structures\n", count
, count2
);
2799 int __init
ext4_init_mballoc(void)
2801 ext4_pspace_cachep
= KMEM_CACHE(ext4_prealloc_space
,
2802 SLAB_RECLAIM_ACCOUNT
);
2803 if (ext4_pspace_cachep
== NULL
)
2806 ext4_ac_cachep
= KMEM_CACHE(ext4_allocation_context
,
2807 SLAB_RECLAIM_ACCOUNT
);
2808 if (ext4_ac_cachep
== NULL
) {
2809 kmem_cache_destroy(ext4_pspace_cachep
);
2813 ext4_free_data_cachep
= KMEM_CACHE(ext4_free_data
,
2814 SLAB_RECLAIM_ACCOUNT
);
2815 if (ext4_free_data_cachep
== NULL
) {
2816 kmem_cache_destroy(ext4_pspace_cachep
);
2817 kmem_cache_destroy(ext4_ac_cachep
);
2823 void ext4_exit_mballoc(void)
2826 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2827 * before destroying the slab cache.
2830 kmem_cache_destroy(ext4_pspace_cachep
);
2831 kmem_cache_destroy(ext4_ac_cachep
);
2832 kmem_cache_destroy(ext4_free_data_cachep
);
2833 ext4_groupinfo_destroy_slabs();
2838 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2839 * Returns 0 if success or error code
2841 static noinline_for_stack
int
2842 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2843 handle_t
*handle
, unsigned int reserv_clstrs
)
2845 struct buffer_head
*bitmap_bh
= NULL
;
2846 struct ext4_group_desc
*gdp
;
2847 struct buffer_head
*gdp_bh
;
2848 struct ext4_sb_info
*sbi
;
2849 struct super_block
*sb
;
2853 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2854 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2860 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2864 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2869 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2873 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2874 ext4_free_group_clusters(sb
, gdp
));
2876 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
2880 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
2882 len
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
2883 if (!ext4_data_block_valid(sbi
, block
, len
)) {
2884 ext4_error(sb
, "Allocating blocks %llu-%llu which overlap "
2885 "fs metadata", block
, block
+len
);
2886 /* File system mounted not to panic on error
2887 * Fix the bitmap and return EUCLEAN
2888 * We leak some of the blocks here.
2890 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2891 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2892 ac
->ac_b_ex
.fe_len
);
2893 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2894 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2900 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
2901 #ifdef AGGRESSIVE_CHECK
2904 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
2905 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
2906 bitmap_bh
->b_data
));
2910 ext4_set_bits(bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
2911 ac
->ac_b_ex
.fe_len
);
2912 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2913 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
2914 ext4_free_group_clusters_set(sb
, gdp
,
2915 ext4_free_clusters_after_init(sb
,
2916 ac
->ac_b_ex
.fe_group
, gdp
));
2918 len
= ext4_free_group_clusters(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
2919 ext4_free_group_clusters_set(sb
, gdp
, len
);
2920 ext4_block_bitmap_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
, bitmap_bh
);
2921 ext4_group_desc_csum_set(sb
, ac
->ac_b_ex
.fe_group
, gdp
);
2923 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
2924 percpu_counter_sub(&sbi
->s_freeclusters_counter
, ac
->ac_b_ex
.fe_len
);
2926 * Now reduce the dirty block count also. Should not go negative
2928 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
2929 /* release all the reserved blocks if non delalloc */
2930 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
2933 if (sbi
->s_log_groups_per_flex
) {
2934 ext4_group_t flex_group
= ext4_flex_group(sbi
,
2935 ac
->ac_b_ex
.fe_group
);
2936 atomic64_sub(ac
->ac_b_ex
.fe_len
,
2937 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
2940 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
2943 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
2951 * here we normalize request for locality group
2952 * Group request are normalized to s_mb_group_prealloc, which goes to
2953 * s_strip if we set the same via mount option.
2954 * s_mb_group_prealloc can be configured via
2955 * /sys/fs/ext4/<partition>/mb_group_prealloc
2957 * XXX: should we try to preallocate more than the group has now?
2959 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
2961 struct super_block
*sb
= ac
->ac_sb
;
2962 struct ext4_locality_group
*lg
= ac
->ac_lg
;
2965 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
2966 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2967 current
->pid
, ac
->ac_g_ex
.fe_len
);
2971 * Normalization means making request better in terms of
2972 * size and alignment
2974 static noinline_for_stack
void
2975 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
2976 struct ext4_allocation_request
*ar
)
2978 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2981 loff_t size
, start_off
;
2982 loff_t orig_size __maybe_unused
;
2984 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
2985 struct ext4_prealloc_space
*pa
;
2987 /* do normalize only data requests, metadata requests
2988 do not need preallocation */
2989 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
2992 /* sometime caller may want exact blocks */
2993 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
2996 /* caller may indicate that preallocation isn't
2997 * required (it's a tail, for example) */
2998 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3001 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3002 ext4_mb_normalize_group_request(ac
);
3006 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3008 /* first, let's learn actual file size
3009 * given current request is allocated */
3010 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
3011 size
= size
<< bsbits
;
3012 if (size
< i_size_read(ac
->ac_inode
))
3013 size
= i_size_read(ac
->ac_inode
);
3016 /* max size of free chunks */
3019 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3020 (req <= (size) || max <= (chunk_size))
3022 /* first, try to predict filesize */
3023 /* XXX: should this table be tunable? */
3025 if (size
<= 16 * 1024) {
3027 } else if (size
<= 32 * 1024) {
3029 } else if (size
<= 64 * 1024) {
3031 } else if (size
<= 128 * 1024) {
3033 } else if (size
<= 256 * 1024) {
3035 } else if (size
<= 512 * 1024) {
3037 } else if (size
<= 1024 * 1024) {
3039 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3040 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3041 (21 - bsbits
)) << 21;
3042 size
= 2 * 1024 * 1024;
3043 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3044 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3045 (22 - bsbits
)) << 22;
3046 size
= 4 * 1024 * 1024;
3047 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3048 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3049 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3050 (23 - bsbits
)) << 23;
3051 size
= 8 * 1024 * 1024;
3053 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3054 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3056 size
= size
>> bsbits
;
3057 start
= start_off
>> bsbits
;
3059 /* don't cover already allocated blocks in selected range */
3060 if (ar
->pleft
&& start
<= ar
->lleft
) {
3061 size
-= ar
->lleft
+ 1 - start
;
3062 start
= ar
->lleft
+ 1;
3064 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3065 size
-= start
+ size
- ar
->lright
;
3068 * Trim allocation request for filesystems with artificially small
3071 if (size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
))
3072 size
= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
);
3076 /* check we don't cross already preallocated blocks */
3078 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3083 spin_lock(&pa
->pa_lock
);
3084 if (pa
->pa_deleted
) {
3085 spin_unlock(&pa
->pa_lock
);
3089 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3092 /* PA must not overlap original request */
3093 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3094 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3096 /* skip PAs this normalized request doesn't overlap with */
3097 if (pa
->pa_lstart
>= end
|| pa_end
<= start
) {
3098 spin_unlock(&pa
->pa_lock
);
3101 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3103 /* adjust start or end to be adjacent to this pa */
3104 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3105 BUG_ON(pa_end
< start
);
3107 } else if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3108 BUG_ON(pa
->pa_lstart
> end
);
3109 end
= pa
->pa_lstart
;
3111 spin_unlock(&pa
->pa_lock
);
3116 /* XXX: extra loop to check we really don't overlap preallocations */
3118 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3121 spin_lock(&pa
->pa_lock
);
3122 if (pa
->pa_deleted
== 0) {
3123 pa_end
= pa
->pa_lstart
+ EXT4_C2B(EXT4_SB(ac
->ac_sb
),
3125 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3127 spin_unlock(&pa
->pa_lock
);
3131 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3132 start
> ac
->ac_o_ex
.fe_logical
) {
3133 ext4_msg(ac
->ac_sb
, KERN_ERR
,
3134 "start %lu, size %lu, fe_logical %lu",
3135 (unsigned long) start
, (unsigned long) size
,
3136 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3138 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3139 start
> ac
->ac_o_ex
.fe_logical
);
3140 BUG_ON(size
<= 0 || size
> EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3142 /* now prepare goal request */
3144 /* XXX: is it better to align blocks WRT to logical
3145 * placement or satisfy big request as is */
3146 ac
->ac_g_ex
.fe_logical
= start
;
3147 ac
->ac_g_ex
.fe_len
= EXT4_NUM_B2C(sbi
, size
);
3149 /* define goal start in order to merge */
3150 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3151 /* merge to the right */
3152 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3153 &ac
->ac_f_ex
.fe_group
,
3154 &ac
->ac_f_ex
.fe_start
);
3155 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3157 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3158 /* merge to the left */
3159 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3160 &ac
->ac_f_ex
.fe_group
,
3161 &ac
->ac_f_ex
.fe_start
);
3162 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3165 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3166 (unsigned) orig_size
, (unsigned) start
);
3169 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3171 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3173 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3174 atomic_inc(&sbi
->s_bal_reqs
);
3175 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3176 if (ac
->ac_b_ex
.fe_len
>= ac
->ac_o_ex
.fe_len
)
3177 atomic_inc(&sbi
->s_bal_success
);
3178 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3179 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3180 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3181 atomic_inc(&sbi
->s_bal_goals
);
3182 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3183 atomic_inc(&sbi
->s_bal_breaks
);
3186 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
)
3187 trace_ext4_mballoc_alloc(ac
);
3189 trace_ext4_mballoc_prealloc(ac
);
3193 * Called on failure; free up any blocks from the inode PA for this
3194 * context. We don't need this for MB_GROUP_PA because we only change
3195 * pa_free in ext4_mb_release_context(), but on failure, we've already
3196 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3198 static void ext4_discard_allocated_blocks(struct ext4_allocation_context
*ac
)
3200 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
3201 struct ext4_buddy e4b
;
3205 if (ac
->ac_f_ex
.fe_len
== 0)
3207 err
= ext4_mb_load_buddy(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
, &e4b
);
3210 * This should never happen since we pin the
3211 * pages in the ext4_allocation_context so
3212 * ext4_mb_load_buddy() should never fail.
3214 WARN(1, "mb_load_buddy failed (%d)", err
);
3217 ext4_lock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3218 mb_free_blocks(ac
->ac_inode
, &e4b
, ac
->ac_f_ex
.fe_start
,
3219 ac
->ac_f_ex
.fe_len
);
3220 ext4_unlock_group(ac
->ac_sb
, ac
->ac_f_ex
.fe_group
);
3221 ext4_mb_unload_buddy(&e4b
);
3224 if (pa
->pa_type
== MB_INODE_PA
)
3225 pa
->pa_free
+= ac
->ac_b_ex
.fe_len
;
3229 * use blocks preallocated to inode
3231 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3232 struct ext4_prealloc_space
*pa
)
3234 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3239 /* found preallocated blocks, use them */
3240 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3241 end
= min(pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
),
3242 start
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
));
3243 len
= EXT4_NUM_B2C(sbi
, end
- start
);
3244 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3245 &ac
->ac_b_ex
.fe_start
);
3246 ac
->ac_b_ex
.fe_len
= len
;
3247 ac
->ac_status
= AC_STATUS_FOUND
;
3250 BUG_ON(start
< pa
->pa_pstart
);
3251 BUG_ON(end
> pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
));
3252 BUG_ON(pa
->pa_free
< len
);
3255 mb_debug(1, "use %llu/%u from inode pa %p\n", start
, len
, pa
);
3259 * use blocks preallocated to locality group
3261 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3262 struct ext4_prealloc_space
*pa
)
3264 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3266 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3267 &ac
->ac_b_ex
.fe_group
,
3268 &ac
->ac_b_ex
.fe_start
);
3269 ac
->ac_b_ex
.fe_len
= len
;
3270 ac
->ac_status
= AC_STATUS_FOUND
;
3273 /* we don't correct pa_pstart or pa_plen here to avoid
3274 * possible race when the group is being loaded concurrently
3275 * instead we correct pa later, after blocks are marked
3276 * in on-disk bitmap -- see ext4_mb_release_context()
3277 * Other CPUs are prevented from allocating from this pa by lg_mutex
3279 mb_debug(1, "use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3283 * Return the prealloc space that have minimal distance
3284 * from the goal block. @cpa is the prealloc
3285 * space that is having currently known minimal distance
3286 * from the goal block.
3288 static struct ext4_prealloc_space
*
3289 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3290 struct ext4_prealloc_space
*pa
,
3291 struct ext4_prealloc_space
*cpa
)
3293 ext4_fsblk_t cur_distance
, new_distance
;
3296 atomic_inc(&pa
->pa_count
);
3299 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3300 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3302 if (cur_distance
<= new_distance
)
3305 /* drop the previous reference */
3306 atomic_dec(&cpa
->pa_count
);
3307 atomic_inc(&pa
->pa_count
);
3312 * search goal blocks in preallocated space
3314 static noinline_for_stack
int
3315 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3317 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3319 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3320 struct ext4_locality_group
*lg
;
3321 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3322 ext4_fsblk_t goal_block
;
3324 /* only data can be preallocated */
3325 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3328 /* first, try per-file preallocation */
3330 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3332 /* all fields in this condition don't change,
3333 * so we can skip locking for them */
3334 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3335 ac
->ac_o_ex
.fe_logical
>= (pa
->pa_lstart
+
3336 EXT4_C2B(sbi
, pa
->pa_len
)))
3339 /* non-extent files can't have physical blocks past 2^32 */
3340 if (!(ext4_test_inode_flag(ac
->ac_inode
, EXT4_INODE_EXTENTS
)) &&
3341 (pa
->pa_pstart
+ EXT4_C2B(sbi
, pa
->pa_len
) >
3342 EXT4_MAX_BLOCK_FILE_PHYS
))
3345 /* found preallocated blocks, use them */
3346 spin_lock(&pa
->pa_lock
);
3347 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3348 atomic_inc(&pa
->pa_count
);
3349 ext4_mb_use_inode_pa(ac
, pa
);
3350 spin_unlock(&pa
->pa_lock
);
3351 ac
->ac_criteria
= 10;
3355 spin_unlock(&pa
->pa_lock
);
3359 /* can we use group allocation? */
3360 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3363 /* inode may have no locality group for some reason */
3367 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3368 if (order
> PREALLOC_TB_SIZE
- 1)
3369 /* The max size of hash table is PREALLOC_TB_SIZE */
3370 order
= PREALLOC_TB_SIZE
- 1;
3372 goal_block
= ext4_grp_offs_to_block(ac
->ac_sb
, &ac
->ac_g_ex
);
3374 * search for the prealloc space that is having
3375 * minimal distance from the goal block.
3377 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3379 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3381 spin_lock(&pa
->pa_lock
);
3382 if (pa
->pa_deleted
== 0 &&
3383 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3385 cpa
= ext4_mb_check_group_pa(goal_block
,
3388 spin_unlock(&pa
->pa_lock
);
3393 ext4_mb_use_group_pa(ac
, cpa
);
3394 ac
->ac_criteria
= 20;
3401 * the function goes through all block freed in the group
3402 * but not yet committed and marks them used in in-core bitmap.
3403 * buddy must be generated from this bitmap
3404 * Need to be called with the ext4 group lock held
3406 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3410 struct ext4_group_info
*grp
;
3411 struct ext4_free_data
*entry
;
3413 grp
= ext4_get_group_info(sb
, group
);
3414 n
= rb_first(&(grp
->bb_free_root
));
3417 entry
= rb_entry(n
, struct ext4_free_data
, efd_node
);
3418 ext4_set_bits(bitmap
, entry
->efd_start_cluster
, entry
->efd_count
);
3425 * the function goes through all preallocation in this group and marks them
3426 * used in in-core bitmap. buddy must be generated from this bitmap
3427 * Need to be called with ext4 group lock held
3429 static noinline_for_stack
3430 void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3433 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3434 struct ext4_prealloc_space
*pa
;
3435 struct list_head
*cur
;
3436 ext4_group_t groupnr
;
3437 ext4_grpblk_t start
;
3438 int preallocated
= 0;
3441 /* all form of preallocation discards first load group,
3442 * so the only competing code is preallocation use.
3443 * we don't need any locking here
3444 * notice we do NOT ignore preallocations with pa_deleted
3445 * otherwise we could leave used blocks available for
3446 * allocation in buddy when concurrent ext4_mb_put_pa()
3447 * is dropping preallocation
3449 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3450 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3451 spin_lock(&pa
->pa_lock
);
3452 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3455 spin_unlock(&pa
->pa_lock
);
3456 if (unlikely(len
== 0))
3458 BUG_ON(groupnr
!= group
);
3459 ext4_set_bits(bitmap
, start
, len
);
3460 preallocated
+= len
;
3462 mb_debug(1, "prellocated %u for group %u\n", preallocated
, group
);
3465 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3467 struct ext4_prealloc_space
*pa
;
3468 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3470 BUG_ON(atomic_read(&pa
->pa_count
));
3471 BUG_ON(pa
->pa_deleted
== 0);
3472 kmem_cache_free(ext4_pspace_cachep
, pa
);
3476 * drops a reference to preallocated space descriptor
3477 * if this was the last reference and the space is consumed
3479 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3480 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3483 ext4_fsblk_t grp_blk
;
3485 /* in this short window concurrent discard can set pa_deleted */
3486 spin_lock(&pa
->pa_lock
);
3487 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0) {
3488 spin_unlock(&pa
->pa_lock
);
3492 if (pa
->pa_deleted
== 1) {
3493 spin_unlock(&pa
->pa_lock
);
3498 spin_unlock(&pa
->pa_lock
);
3500 grp_blk
= pa
->pa_pstart
;
3502 * If doing group-based preallocation, pa_pstart may be in the
3503 * next group when pa is used up
3505 if (pa
->pa_type
== MB_GROUP_PA
)
3508 grp
= ext4_get_group_number(sb
, grp_blk
);
3513 * P1 (buddy init) P2 (regular allocation)
3514 * find block B in PA
3515 * copy on-disk bitmap to buddy
3516 * mark B in on-disk bitmap
3517 * drop PA from group
3518 * mark all PAs in buddy
3520 * thus, P1 initializes buddy with B available. to prevent this
3521 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3524 ext4_lock_group(sb
, grp
);
3525 list_del(&pa
->pa_group_list
);
3526 ext4_unlock_group(sb
, grp
);
3528 spin_lock(pa
->pa_obj_lock
);
3529 list_del_rcu(&pa
->pa_inode_list
);
3530 spin_unlock(pa
->pa_obj_lock
);
3532 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3536 * creates new preallocated space for given inode
3538 static noinline_for_stack
int
3539 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3541 struct super_block
*sb
= ac
->ac_sb
;
3542 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3543 struct ext4_prealloc_space
*pa
;
3544 struct ext4_group_info
*grp
;
3545 struct ext4_inode_info
*ei
;
3547 /* preallocate only when found space is larger then requested */
3548 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3549 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3550 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3552 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3556 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3562 /* we can't allocate as much as normalizer wants.
3563 * so, found space must get proper lstart
3564 * to cover original request */
3565 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3566 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3568 /* we're limited by original request in that
3569 * logical block must be covered any way
3570 * winl is window we can move our chunk within */
3571 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3573 /* also, we should cover whole original request */
3574 wins
= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
);
3576 /* the smallest one defines real window */
3577 win
= min(winl
, wins
);
3579 offs
= ac
->ac_o_ex
.fe_logical
%
3580 EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
3581 if (offs
&& offs
< win
)
3584 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
-
3585 EXT4_NUM_B2C(sbi
, win
);
3586 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3587 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3590 /* preallocation can change ac_b_ex, thus we store actually
3591 * allocated blocks for history */
3592 ac
->ac_f_ex
= ac
->ac_b_ex
;
3594 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3595 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3596 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3597 pa
->pa_free
= pa
->pa_len
;
3598 atomic_set(&pa
->pa_count
, 1);
3599 spin_lock_init(&pa
->pa_lock
);
3600 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3601 INIT_LIST_HEAD(&pa
->pa_group_list
);
3603 pa
->pa_type
= MB_INODE_PA
;
3605 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa
,
3606 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3607 trace_ext4_mb_new_inode_pa(ac
, pa
);
3609 ext4_mb_use_inode_pa(ac
, pa
);
3610 atomic_add(pa
->pa_free
, &sbi
->s_mb_preallocated
);
3612 ei
= EXT4_I(ac
->ac_inode
);
3613 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3615 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3616 pa
->pa_inode
= ac
->ac_inode
;
3618 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3619 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3620 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3622 spin_lock(pa
->pa_obj_lock
);
3623 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3624 spin_unlock(pa
->pa_obj_lock
);
3630 * creates new preallocated space for locality group inodes belongs to
3632 static noinline_for_stack
int
3633 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3635 struct super_block
*sb
= ac
->ac_sb
;
3636 struct ext4_locality_group
*lg
;
3637 struct ext4_prealloc_space
*pa
;
3638 struct ext4_group_info
*grp
;
3640 /* preallocate only when found space is larger then requested */
3641 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3642 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3643 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3645 BUG_ON(ext4_pspace_cachep
== NULL
);
3646 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3650 /* preallocation can change ac_b_ex, thus we store actually
3651 * allocated blocks for history */
3652 ac
->ac_f_ex
= ac
->ac_b_ex
;
3654 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3655 pa
->pa_lstart
= pa
->pa_pstart
;
3656 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3657 pa
->pa_free
= pa
->pa_len
;
3658 atomic_set(&pa
->pa_count
, 1);
3659 spin_lock_init(&pa
->pa_lock
);
3660 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3661 INIT_LIST_HEAD(&pa
->pa_group_list
);
3663 pa
->pa_type
= MB_GROUP_PA
;
3665 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa
,
3666 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3667 trace_ext4_mb_new_group_pa(ac
, pa
);
3669 ext4_mb_use_group_pa(ac
, pa
);
3670 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3672 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3676 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3677 pa
->pa_inode
= NULL
;
3679 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3680 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3681 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3684 * We will later add the new pa to the right bucket
3685 * after updating the pa_free in ext4_mb_release_context
3690 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3694 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3695 err
= ext4_mb_new_group_pa(ac
);
3697 err
= ext4_mb_new_inode_pa(ac
);
3702 * finds all unused blocks in on-disk bitmap, frees them in
3703 * in-core bitmap and buddy.
3704 * @pa must be unlinked from inode and group lists, so that
3705 * nobody else can find/use it.
3706 * the caller MUST hold group/inode locks.
3707 * TODO: optimize the case when there are no in-core structures yet
3709 static noinline_for_stack
int
3710 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3711 struct ext4_prealloc_space
*pa
)
3713 struct super_block
*sb
= e4b
->bd_sb
;
3714 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3719 unsigned long long grp_blk_start
;
3723 BUG_ON(pa
->pa_deleted
== 0);
3724 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3725 grp_blk_start
= pa
->pa_pstart
- EXT4_C2B(sbi
, bit
);
3726 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3727 end
= bit
+ pa
->pa_len
;
3730 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3733 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3734 mb_debug(1, " free preallocated %u/%u in group %u\n",
3735 (unsigned) ext4_group_first_block_no(sb
, group
) + bit
,
3736 (unsigned) next
- bit
, (unsigned) group
);
3739 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, next
- bit
);
3740 trace_ext4_mb_release_inode_pa(pa
, (grp_blk_start
+
3741 EXT4_C2B(sbi
, bit
)),
3743 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3746 if (free
!= pa
->pa_free
) {
3747 ext4_msg(e4b
->bd_sb
, KERN_CRIT
,
3748 "pa %p: logic %lu, phys. %lu, len %lu",
3749 pa
, (unsigned long) pa
->pa_lstart
,
3750 (unsigned long) pa
->pa_pstart
,
3751 (unsigned long) pa
->pa_len
);
3752 ext4_grp_locked_error(sb
, group
, 0, 0, "free %u, pa_free %u",
3755 * pa is already deleted so we use the value obtained
3756 * from the bitmap and continue.
3759 atomic_add(free
, &sbi
->s_mb_discarded
);
3764 static noinline_for_stack
int
3765 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3766 struct ext4_prealloc_space
*pa
)
3768 struct super_block
*sb
= e4b
->bd_sb
;
3772 trace_ext4_mb_release_group_pa(sb
, pa
);
3773 BUG_ON(pa
->pa_deleted
== 0);
3774 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3775 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3776 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3777 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3778 trace_ext4_mballoc_discard(sb
, NULL
, group
, bit
, pa
->pa_len
);
3784 * releases all preallocations in given group
3786 * first, we need to decide discard policy:
3787 * - when do we discard
3789 * - how many do we discard
3790 * 1) how many requested
3792 static noinline_for_stack
int
3793 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3794 ext4_group_t group
, int needed
)
3796 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3797 struct buffer_head
*bitmap_bh
= NULL
;
3798 struct ext4_prealloc_space
*pa
, *tmp
;
3799 struct list_head list
;
3800 struct ext4_buddy e4b
;
3805 mb_debug(1, "discard preallocation for group %u\n", group
);
3807 if (list_empty(&grp
->bb_prealloc_list
))
3810 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3811 if (bitmap_bh
== NULL
) {
3812 ext4_error(sb
, "Error reading block bitmap for %u", group
);
3816 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3818 ext4_error(sb
, "Error loading buddy information for %u", group
);
3824 needed
= EXT4_CLUSTERS_PER_GROUP(sb
) + 1;
3826 INIT_LIST_HEAD(&list
);
3828 ext4_lock_group(sb
, group
);
3829 list_for_each_entry_safe(pa
, tmp
,
3830 &grp
->bb_prealloc_list
, pa_group_list
) {
3831 spin_lock(&pa
->pa_lock
);
3832 if (atomic_read(&pa
->pa_count
)) {
3833 spin_unlock(&pa
->pa_lock
);
3837 if (pa
->pa_deleted
) {
3838 spin_unlock(&pa
->pa_lock
);
3842 /* seems this one can be freed ... */
3845 /* we can trust pa_free ... */
3846 free
+= pa
->pa_free
;
3848 spin_unlock(&pa
->pa_lock
);
3850 list_del(&pa
->pa_group_list
);
3851 list_add(&pa
->u
.pa_tmp_list
, &list
);
3854 /* if we still need more blocks and some PAs were used, try again */
3855 if (free
< needed
&& busy
) {
3857 ext4_unlock_group(sb
, group
);
3862 /* found anything to free? */
3863 if (list_empty(&list
)) {
3868 /* now free all selected PAs */
3869 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3871 /* remove from object (inode or locality group) */
3872 spin_lock(pa
->pa_obj_lock
);
3873 list_del_rcu(&pa
->pa_inode_list
);
3874 spin_unlock(pa
->pa_obj_lock
);
3876 if (pa
->pa_type
== MB_GROUP_PA
)
3877 ext4_mb_release_group_pa(&e4b
, pa
);
3879 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3881 list_del(&pa
->u
.pa_tmp_list
);
3882 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3886 ext4_unlock_group(sb
, group
);
3887 ext4_mb_unload_buddy(&e4b
);
3893 * releases all non-used preallocated blocks for given inode
3895 * It's important to discard preallocations under i_data_sem
3896 * We don't want another block to be served from the prealloc
3897 * space when we are discarding the inode prealloc space.
3899 * FIXME!! Make sure it is valid at all the call sites
3901 void ext4_discard_preallocations(struct inode
*inode
)
3903 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3904 struct super_block
*sb
= inode
->i_sb
;
3905 struct buffer_head
*bitmap_bh
= NULL
;
3906 struct ext4_prealloc_space
*pa
, *tmp
;
3907 ext4_group_t group
= 0;
3908 struct list_head list
;
3909 struct ext4_buddy e4b
;
3912 if (!S_ISREG(inode
->i_mode
)) {
3913 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3917 mb_debug(1, "discard preallocation for inode %lu\n", inode
->i_ino
);
3918 trace_ext4_discard_preallocations(inode
);
3920 INIT_LIST_HEAD(&list
);
3923 /* first, collect all pa's in the inode */
3924 spin_lock(&ei
->i_prealloc_lock
);
3925 while (!list_empty(&ei
->i_prealloc_list
)) {
3926 pa
= list_entry(ei
->i_prealloc_list
.next
,
3927 struct ext4_prealloc_space
, pa_inode_list
);
3928 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3929 spin_lock(&pa
->pa_lock
);
3930 if (atomic_read(&pa
->pa_count
)) {
3931 /* this shouldn't happen often - nobody should
3932 * use preallocation while we're discarding it */
3933 spin_unlock(&pa
->pa_lock
);
3934 spin_unlock(&ei
->i_prealloc_lock
);
3935 ext4_msg(sb
, KERN_ERR
,
3936 "uh-oh! used pa while discarding");
3938 schedule_timeout_uninterruptible(HZ
);
3942 if (pa
->pa_deleted
== 0) {
3944 spin_unlock(&pa
->pa_lock
);
3945 list_del_rcu(&pa
->pa_inode_list
);
3946 list_add(&pa
->u
.pa_tmp_list
, &list
);
3950 /* someone is deleting pa right now */
3951 spin_unlock(&pa
->pa_lock
);
3952 spin_unlock(&ei
->i_prealloc_lock
);
3954 /* we have to wait here because pa_deleted
3955 * doesn't mean pa is already unlinked from
3956 * the list. as we might be called from
3957 * ->clear_inode() the inode will get freed
3958 * and concurrent thread which is unlinking
3959 * pa from inode's list may access already
3960 * freed memory, bad-bad-bad */
3962 /* XXX: if this happens too often, we can
3963 * add a flag to force wait only in case
3964 * of ->clear_inode(), but not in case of
3965 * regular truncate */
3966 schedule_timeout_uninterruptible(HZ
);
3969 spin_unlock(&ei
->i_prealloc_lock
);
3971 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3972 BUG_ON(pa
->pa_type
!= MB_INODE_PA
);
3973 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
3975 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3977 ext4_error(sb
, "Error loading buddy information for %u",
3982 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3983 if (bitmap_bh
== NULL
) {
3984 ext4_error(sb
, "Error reading block bitmap for %u",
3986 ext4_mb_unload_buddy(&e4b
);
3990 ext4_lock_group(sb
, group
);
3991 list_del(&pa
->pa_group_list
);
3992 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3993 ext4_unlock_group(sb
, group
);
3995 ext4_mb_unload_buddy(&e4b
);
3998 list_del(&pa
->u
.pa_tmp_list
);
3999 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4003 #ifdef CONFIG_EXT4_DEBUG
4004 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4006 struct super_block
*sb
= ac
->ac_sb
;
4007 ext4_group_t ngroups
, i
;
4009 if (!ext4_mballoc_debug
||
4010 (EXT4_SB(sb
)->s_mount_flags
& EXT4_MF_FS_ABORTED
))
4013 ext4_msg(ac
->ac_sb
, KERN_ERR
, "Can't allocate:"
4014 " Allocation context details:");
4015 ext4_msg(ac
->ac_sb
, KERN_ERR
, "status %d flags %d",
4016 ac
->ac_status
, ac
->ac_flags
);
4017 ext4_msg(ac
->ac_sb
, KERN_ERR
, "orig %lu/%lu/%lu@%lu, "
4018 "goal %lu/%lu/%lu@%lu, "
4019 "best %lu/%lu/%lu@%lu cr %d",
4020 (unsigned long)ac
->ac_o_ex
.fe_group
,
4021 (unsigned long)ac
->ac_o_ex
.fe_start
,
4022 (unsigned long)ac
->ac_o_ex
.fe_len
,
4023 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4024 (unsigned long)ac
->ac_g_ex
.fe_group
,
4025 (unsigned long)ac
->ac_g_ex
.fe_start
,
4026 (unsigned long)ac
->ac_g_ex
.fe_len
,
4027 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4028 (unsigned long)ac
->ac_b_ex
.fe_group
,
4029 (unsigned long)ac
->ac_b_ex
.fe_start
,
4030 (unsigned long)ac
->ac_b_ex
.fe_len
,
4031 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4032 (int)ac
->ac_criteria
);
4033 ext4_msg(ac
->ac_sb
, KERN_ERR
, "%lu scanned, %d found",
4034 ac
->ac_ex_scanned
, ac
->ac_found
);
4035 ext4_msg(ac
->ac_sb
, KERN_ERR
, "groups: ");
4036 ngroups
= ext4_get_groups_count(sb
);
4037 for (i
= 0; i
< ngroups
; i
++) {
4038 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4039 struct ext4_prealloc_space
*pa
;
4040 ext4_grpblk_t start
;
4041 struct list_head
*cur
;
4042 ext4_lock_group(sb
, i
);
4043 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4044 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4046 spin_lock(&pa
->pa_lock
);
4047 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4049 spin_unlock(&pa
->pa_lock
);
4050 printk(KERN_ERR
"PA:%u:%d:%u \n", i
,
4053 ext4_unlock_group(sb
, i
);
4055 if (grp
->bb_free
== 0)
4057 printk(KERN_ERR
"%u: %d/%d \n",
4058 i
, grp
->bb_free
, grp
->bb_fragments
);
4060 printk(KERN_ERR
"\n");
4063 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4070 * We use locality group preallocation for small size file. The size of the
4071 * file is determined by the current size or the resulting size after
4072 * allocation which ever is larger
4074 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4076 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4078 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4079 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4082 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4085 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4088 size
= ac
->ac_o_ex
.fe_logical
+ EXT4_C2B(sbi
, ac
->ac_o_ex
.fe_len
);
4089 isize
= (i_size_read(ac
->ac_inode
) + ac
->ac_sb
->s_blocksize
- 1)
4092 if ((size
== isize
) &&
4093 !ext4_fs_is_busy(sbi
) &&
4094 (atomic_read(&ac
->ac_inode
->i_writecount
) == 0)) {
4095 ac
->ac_flags
|= EXT4_MB_HINT_NOPREALLOC
;
4099 if (sbi
->s_mb_group_prealloc
<= 0) {
4100 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4104 /* don't use group allocation for large files */
4105 size
= max(size
, isize
);
4106 if (size
> sbi
->s_mb_stream_request
) {
4107 ac
->ac_flags
|= EXT4_MB_STREAM_ALLOC
;
4111 BUG_ON(ac
->ac_lg
!= NULL
);
4113 * locality group prealloc space are per cpu. The reason for having
4114 * per cpu locality group is to reduce the contention between block
4115 * request from multiple CPUs.
4117 ac
->ac_lg
= __this_cpu_ptr(sbi
->s_locality_groups
);
4119 /* we're going to use group allocation */
4120 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4122 /* serialize all allocations in the group */
4123 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4126 static noinline_for_stack
int
4127 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4128 struct ext4_allocation_request
*ar
)
4130 struct super_block
*sb
= ar
->inode
->i_sb
;
4131 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4132 struct ext4_super_block
*es
= sbi
->s_es
;
4136 ext4_grpblk_t block
;
4138 /* we can't allocate > group size */
4141 /* just a dirty hack to filter too big requests */
4142 if (len
>= EXT4_CLUSTERS_PER_GROUP(sb
))
4143 len
= EXT4_CLUSTERS_PER_GROUP(sb
);
4145 /* start searching from the goal */
4147 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4148 goal
>= ext4_blocks_count(es
))
4149 goal
= le32_to_cpu(es
->s_first_data_block
);
4150 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4152 /* set up allocation goals */
4153 ac
->ac_b_ex
.fe_logical
= EXT4_LBLK_CMASK(sbi
, ar
->logical
);
4154 ac
->ac_status
= AC_STATUS_CONTINUE
;
4156 ac
->ac_inode
= ar
->inode
;
4157 ac
->ac_o_ex
.fe_logical
= ac
->ac_b_ex
.fe_logical
;
4158 ac
->ac_o_ex
.fe_group
= group
;
4159 ac
->ac_o_ex
.fe_start
= block
;
4160 ac
->ac_o_ex
.fe_len
= len
;
4161 ac
->ac_g_ex
= ac
->ac_o_ex
;
4162 ac
->ac_flags
= ar
->flags
;
4164 /* we have to define context: we'll we work with a file or
4165 * locality group. this is a policy, actually */
4166 ext4_mb_group_or_file(ac
);
4168 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4169 "left: %u/%u, right %u/%u to %swritable\n",
4170 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4171 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4172 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4173 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4174 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4179 static noinline_for_stack
void
4180 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4181 struct ext4_locality_group
*lg
,
4182 int order
, int total_entries
)
4184 ext4_group_t group
= 0;
4185 struct ext4_buddy e4b
;
4186 struct list_head discard_list
;
4187 struct ext4_prealloc_space
*pa
, *tmp
;
4189 mb_debug(1, "discard locality group preallocation\n");
4191 INIT_LIST_HEAD(&discard_list
);
4193 spin_lock(&lg
->lg_prealloc_lock
);
4194 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4196 spin_lock(&pa
->pa_lock
);
4197 if (atomic_read(&pa
->pa_count
)) {
4199 * This is the pa that we just used
4200 * for block allocation. So don't
4203 spin_unlock(&pa
->pa_lock
);
4206 if (pa
->pa_deleted
) {
4207 spin_unlock(&pa
->pa_lock
);
4210 /* only lg prealloc space */
4211 BUG_ON(pa
->pa_type
!= MB_GROUP_PA
);
4213 /* seems this one can be freed ... */
4215 spin_unlock(&pa
->pa_lock
);
4217 list_del_rcu(&pa
->pa_inode_list
);
4218 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4221 if (total_entries
<= 5) {
4223 * we want to keep only 5 entries
4224 * allowing it to grow to 8. This
4225 * mak sure we don't call discard
4226 * soon for this list.
4231 spin_unlock(&lg
->lg_prealloc_lock
);
4233 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4235 group
= ext4_get_group_number(sb
, pa
->pa_pstart
);
4236 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4237 ext4_error(sb
, "Error loading buddy information for %u",
4241 ext4_lock_group(sb
, group
);
4242 list_del(&pa
->pa_group_list
);
4243 ext4_mb_release_group_pa(&e4b
, pa
);
4244 ext4_unlock_group(sb
, group
);
4246 ext4_mb_unload_buddy(&e4b
);
4247 list_del(&pa
->u
.pa_tmp_list
);
4248 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4253 * We have incremented pa_count. So it cannot be freed at this
4254 * point. Also we hold lg_mutex. So no parallel allocation is
4255 * possible from this lg. That means pa_free cannot be updated.
4257 * A parallel ext4_mb_discard_group_preallocations is possible.
4258 * which can cause the lg_prealloc_list to be updated.
4261 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4263 int order
, added
= 0, lg_prealloc_count
= 1;
4264 struct super_block
*sb
= ac
->ac_sb
;
4265 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4266 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4268 order
= fls(pa
->pa_free
) - 1;
4269 if (order
> PREALLOC_TB_SIZE
- 1)
4270 /* The max size of hash table is PREALLOC_TB_SIZE */
4271 order
= PREALLOC_TB_SIZE
- 1;
4272 /* Add the prealloc space to lg */
4273 spin_lock(&lg
->lg_prealloc_lock
);
4274 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4276 spin_lock(&tmp_pa
->pa_lock
);
4277 if (tmp_pa
->pa_deleted
) {
4278 spin_unlock(&tmp_pa
->pa_lock
);
4281 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4282 /* Add to the tail of the previous entry */
4283 list_add_tail_rcu(&pa
->pa_inode_list
,
4284 &tmp_pa
->pa_inode_list
);
4287 * we want to count the total
4288 * number of entries in the list
4291 spin_unlock(&tmp_pa
->pa_lock
);
4292 lg_prealloc_count
++;
4295 list_add_tail_rcu(&pa
->pa_inode_list
,
4296 &lg
->lg_prealloc_list
[order
]);
4297 spin_unlock(&lg
->lg_prealloc_lock
);
4299 /* Now trim the list to be not more than 8 elements */
4300 if (lg_prealloc_count
> 8) {
4301 ext4_mb_discard_lg_preallocations(sb
, lg
,
4302 order
, lg_prealloc_count
);
4309 * release all resource we used in allocation
4311 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4313 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4314 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4316 if (pa
->pa_type
== MB_GROUP_PA
) {
4317 /* see comment in ext4_mb_use_group_pa() */
4318 spin_lock(&pa
->pa_lock
);
4319 pa
->pa_pstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4320 pa
->pa_lstart
+= EXT4_C2B(sbi
, ac
->ac_b_ex
.fe_len
);
4321 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4322 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4323 spin_unlock(&pa
->pa_lock
);
4328 * We want to add the pa to the right bucket.
4329 * Remove it from the list and while adding
4330 * make sure the list to which we are adding
4333 if ((pa
->pa_type
== MB_GROUP_PA
) && likely(pa
->pa_free
)) {
4334 spin_lock(pa
->pa_obj_lock
);
4335 list_del_rcu(&pa
->pa_inode_list
);
4336 spin_unlock(pa
->pa_obj_lock
);
4337 ext4_mb_add_n_trim(ac
);
4339 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4341 if (ac
->ac_bitmap_page
)
4342 page_cache_release(ac
->ac_bitmap_page
);
4343 if (ac
->ac_buddy_page
)
4344 page_cache_release(ac
->ac_buddy_page
);
4345 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4346 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4347 ext4_mb_collect_stats(ac
);
4351 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4353 ext4_group_t i
, ngroups
= ext4_get_groups_count(sb
);
4357 trace_ext4_mb_discard_preallocations(sb
, needed
);
4358 for (i
= 0; i
< ngroups
&& needed
> 0; i
++) {
4359 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4368 * Main entry point into mballoc to allocate blocks
4369 * it tries to use preallocation first, then falls back
4370 * to usual allocation
4372 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4373 struct ext4_allocation_request
*ar
, int *errp
)
4376 struct ext4_allocation_context
*ac
= NULL
;
4377 struct ext4_sb_info
*sbi
;
4378 struct super_block
*sb
;
4379 ext4_fsblk_t block
= 0;
4380 unsigned int inquota
= 0;
4381 unsigned int reserv_clstrs
= 0;
4384 sb
= ar
->inode
->i_sb
;
4387 trace_ext4_request_blocks(ar
);
4389 /* Allow to use superuser reservation for quota file */
4390 if (IS_NOQUOTA(ar
->inode
))
4391 ar
->flags
|= EXT4_MB_USE_ROOT_BLOCKS
;
4394 * For delayed allocation, we could skip the ENOSPC and
4395 * EDQUOT check, as blocks and quotas have been already
4396 * reserved when data being copied into pagecache.
4398 if (ext4_test_inode_state(ar
->inode
, EXT4_STATE_DELALLOC_RESERVED
))
4399 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4401 /* Without delayed allocation we need to verify
4402 * there is enough free blocks to do block allocation
4403 * and verify allocation doesn't exceed the quota limits.
4406 ext4_claim_free_clusters(sbi
, ar
->len
, ar
->flags
)) {
4408 /* let others to free the space */
4410 ar
->len
= ar
->len
>> 1;
4416 reserv_clstrs
= ar
->len
;
4417 if (ar
->flags
& EXT4_MB_USE_ROOT_BLOCKS
) {
4418 dquot_alloc_block_nofail(ar
->inode
,
4419 EXT4_C2B(sbi
, ar
->len
));
4422 dquot_alloc_block(ar
->inode
,
4423 EXT4_C2B(sbi
, ar
->len
))) {
4425 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4436 ac
= kmem_cache_zalloc(ext4_ac_cachep
, GFP_NOFS
);
4443 *errp
= ext4_mb_initialize_context(ac
, ar
);
4449 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4450 if (!ext4_mb_use_preallocated(ac
)) {
4451 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4452 ext4_mb_normalize_request(ac
, ar
);
4454 /* allocate space in core */
4455 *errp
= ext4_mb_regular_allocator(ac
);
4457 ext4_discard_allocated_blocks(ac
);
4461 /* as we've just preallocated more space than
4462 * user requested orinally, we store allocated
4463 * space in a special descriptor */
4464 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4465 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4466 ext4_mb_new_preallocation(ac
);
4468 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4469 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_clstrs
);
4471 ext4_discard_allocated_blocks(ac
);
4474 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4475 ar
->len
= ac
->ac_b_ex
.fe_len
;
4478 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4486 ac
->ac_b_ex
.fe_len
= 0;
4488 ext4_mb_show_ac(ac
);
4490 ext4_mb_release_context(ac
);
4493 kmem_cache_free(ext4_ac_cachep
, ac
);
4494 if (inquota
&& ar
->len
< inquota
)
4495 dquot_free_block(ar
->inode
, EXT4_C2B(sbi
, inquota
- ar
->len
));
4497 if (!ext4_test_inode_state(ar
->inode
,
4498 EXT4_STATE_DELALLOC_RESERVED
))
4499 /* release all the reserved blocks if non delalloc */
4500 percpu_counter_sub(&sbi
->s_dirtyclusters_counter
,
4504 trace_ext4_allocate_blocks(ar
, (unsigned long long)block
);
4510 * We can merge two free data extents only if the physical blocks
4511 * are contiguous, AND the extents were freed by the same transaction,
4512 * AND the blocks are associated with the same group.
4514 static int can_merge(struct ext4_free_data
*entry1
,
4515 struct ext4_free_data
*entry2
)
4517 if ((entry1
->efd_tid
== entry2
->efd_tid
) &&
4518 (entry1
->efd_group
== entry2
->efd_group
) &&
4519 ((entry1
->efd_start_cluster
+ entry1
->efd_count
) == entry2
->efd_start_cluster
))
4524 static noinline_for_stack
int
4525 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4526 struct ext4_free_data
*new_entry
)
4528 ext4_group_t group
= e4b
->bd_group
;
4529 ext4_grpblk_t cluster
;
4530 struct ext4_free_data
*entry
;
4531 struct ext4_group_info
*db
= e4b
->bd_info
;
4532 struct super_block
*sb
= e4b
->bd_sb
;
4533 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4534 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4535 struct rb_node
*parent
= NULL
, *new_node
;
4537 BUG_ON(!ext4_handle_valid(handle
));
4538 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4539 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4541 new_node
= &new_entry
->efd_node
;
4542 cluster
= new_entry
->efd_start_cluster
;
4545 /* first free block exent. We need to
4546 protect buddy cache from being freed,
4547 * otherwise we'll refresh it from
4548 * on-disk bitmap and lose not-yet-available
4550 page_cache_get(e4b
->bd_buddy_page
);
4551 page_cache_get(e4b
->bd_bitmap_page
);
4555 entry
= rb_entry(parent
, struct ext4_free_data
, efd_node
);
4556 if (cluster
< entry
->efd_start_cluster
)
4558 else if (cluster
>= (entry
->efd_start_cluster
+ entry
->efd_count
))
4559 n
= &(*n
)->rb_right
;
4561 ext4_grp_locked_error(sb
, group
, 0,
4562 ext4_group_first_block_no(sb
, group
) +
4563 EXT4_C2B(sbi
, cluster
),
4564 "Block already on to-be-freed list");
4569 rb_link_node(new_node
, parent
, n
);
4570 rb_insert_color(new_node
, &db
->bb_free_root
);
4572 /* Now try to see the extent can be merged to left and right */
4573 node
= rb_prev(new_node
);
4575 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4576 if (can_merge(entry
, new_entry
) &&
4577 ext4_journal_callback_try_del(handle
, &entry
->efd_jce
)) {
4578 new_entry
->efd_start_cluster
= entry
->efd_start_cluster
;
4579 new_entry
->efd_count
+= entry
->efd_count
;
4580 rb_erase(node
, &(db
->bb_free_root
));
4581 kmem_cache_free(ext4_free_data_cachep
, entry
);
4585 node
= rb_next(new_node
);
4587 entry
= rb_entry(node
, struct ext4_free_data
, efd_node
);
4588 if (can_merge(new_entry
, entry
) &&
4589 ext4_journal_callback_try_del(handle
, &entry
->efd_jce
)) {
4590 new_entry
->efd_count
+= entry
->efd_count
;
4591 rb_erase(node
, &(db
->bb_free_root
));
4592 kmem_cache_free(ext4_free_data_cachep
, entry
);
4595 /* Add the extent to transaction's private list */
4596 ext4_journal_callback_add(handle
, ext4_free_data_callback
,
4597 &new_entry
->efd_jce
);
4602 * ext4_free_blocks() -- Free given blocks and update quota
4603 * @handle: handle for this transaction
4605 * @block: start physical block to free
4606 * @count: number of blocks to count
4607 * @flags: flags used by ext4_free_blocks
4609 void ext4_free_blocks(handle_t
*handle
, struct inode
*inode
,
4610 struct buffer_head
*bh
, ext4_fsblk_t block
,
4611 unsigned long count
, int flags
)
4613 struct buffer_head
*bitmap_bh
= NULL
;
4614 struct super_block
*sb
= inode
->i_sb
;
4615 struct ext4_group_desc
*gdp
;
4616 unsigned int overflow
;
4618 struct buffer_head
*gd_bh
;
4619 ext4_group_t block_group
;
4620 struct ext4_sb_info
*sbi
;
4621 struct ext4_inode_info
*ei
= EXT4_I(inode
);
4622 struct ext4_buddy e4b
;
4623 unsigned int count_clusters
;
4630 BUG_ON(block
!= bh
->b_blocknr
);
4632 block
= bh
->b_blocknr
;
4636 if (!(flags
& EXT4_FREE_BLOCKS_VALIDATED
) &&
4637 !ext4_data_block_valid(sbi
, block
, count
)) {
4638 ext4_error(sb
, "Freeing blocks not in datazone - "
4639 "block = %llu, count = %lu", block
, count
);
4643 ext4_debug("freeing block %llu\n", block
);
4644 trace_ext4_free_blocks(inode
, block
, count
, flags
);
4646 if (flags
& EXT4_FREE_BLOCKS_FORGET
) {
4647 struct buffer_head
*tbh
= bh
;
4650 BUG_ON(bh
&& (count
> 1));
4652 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
);
4710 * Check to see if we are freeing blocks across a group
4713 if (EXT4_C2B(sbi
, bit
) + count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4714 overflow
= EXT4_C2B(sbi
, bit
) + count
-
4715 EXT4_BLOCKS_PER_GROUP(sb
);
4718 count_clusters
= EXT4_NUM_B2C(sbi
, count
);
4719 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4724 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4730 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4731 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4732 in_range(block
, ext4_inode_table(sb
, gdp
),
4733 EXT4_SB(sb
)->s_itb_per_group
) ||
4734 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4735 EXT4_SB(sb
)->s_itb_per_group
)) {
4737 ext4_error(sb
, "Freeing blocks in system zone - "
4738 "Block = %llu, count = %lu", block
, count
);
4739 /* err = 0. ext4_std_error should be a no op */
4743 BUFFER_TRACE(bitmap_bh
, "getting write access");
4744 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4749 * We are about to modify some metadata. Call the journal APIs
4750 * to unshare ->b_data if a currently-committing transaction is
4753 BUFFER_TRACE(gd_bh
, "get_write_access");
4754 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4757 #ifdef AGGRESSIVE_CHECK
4760 for (i
= 0; i
< count_clusters
; i
++)
4761 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4764 trace_ext4_mballoc_free(sb
, inode
, block_group
, bit
, count_clusters
);
4766 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4767 err
= ext4_mb_load_buddy_gfp(sb
, block_group
, &e4b
,
4768 GFP_NOFS
|__GFP_NOFAIL
);
4772 if ((flags
& EXT4_FREE_BLOCKS_METADATA
) && ext4_handle_valid(handle
)) {
4773 struct ext4_free_data
*new_entry
;
4775 * blocks being freed are metadata. these blocks shouldn't
4776 * be used until this transaction is committed
4778 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4781 new_entry
= kmem_cache_alloc(ext4_free_data_cachep
,
4782 GFP_NOFS
|__GFP_NOFAIL
);
4783 new_entry
->efd_start_cluster
= bit
;
4784 new_entry
->efd_group
= block_group
;
4785 new_entry
->efd_count
= count_clusters
;
4786 new_entry
->efd_tid
= handle
->h_transaction
->t_tid
;
4788 ext4_lock_group(sb
, block_group
);
4789 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4790 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4792 /* need to update group_info->bb_free and bitmap
4793 * with group lock held. generate_buddy look at
4794 * them with group lock_held
4796 if (test_opt(sb
, DISCARD
)) {
4797 err
= ext4_issue_discard(sb
, block_group
, bit
, count
,
4799 if (err
&& err
!= -EOPNOTSUPP
)
4800 ext4_msg(sb
, KERN_WARNING
, "discard request in"
4801 " group:%d block:%d count:%lu failed"
4802 " with %d", block_group
, bit
, count
,
4805 EXT4_MB_GRP_CLEAR_TRIMMED(e4b
.bd_info
);
4807 ext4_lock_group(sb
, block_group
);
4808 mb_clear_bits(bitmap_bh
->b_data
, bit
, count_clusters
);
4809 mb_free_blocks(inode
, &e4b
, bit
, count_clusters
);
4812 ret
= ext4_free_group_clusters(sb
, gdp
) + count_clusters
;
4813 ext4_free_group_clusters_set(sb
, gdp
, ret
);
4814 ext4_block_bitmap_csum_set(sb
, block_group
, gdp
, bitmap_bh
);
4815 ext4_group_desc_csum_set(sb
, block_group
, gdp
);
4816 ext4_unlock_group(sb
, block_group
);
4818 if (sbi
->s_log_groups_per_flex
) {
4819 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4820 atomic64_add(count_clusters
,
4821 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
4824 if (flags
& EXT4_FREE_BLOCKS_RESERVE
&& ei
->i_reserved_data_blocks
) {
4825 percpu_counter_add(&sbi
->s_dirtyclusters_counter
,
4827 spin_lock(&ei
->i_block_reservation_lock
);
4828 if (flags
& EXT4_FREE_BLOCKS_METADATA
)
4829 ei
->i_reserved_meta_blocks
+= count_clusters
;
4831 ei
->i_reserved_data_blocks
+= count_clusters
;
4832 spin_unlock(&ei
->i_block_reservation_lock
);
4833 if (!(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4834 dquot_reclaim_block(inode
,
4835 EXT4_C2B(sbi
, count_clusters
));
4836 } else if (!(flags
& EXT4_FREE_BLOCKS_NO_QUOT_UPDATE
))
4837 dquot_free_block(inode
, EXT4_C2B(sbi
, count_clusters
));
4838 percpu_counter_add(&sbi
->s_freeclusters_counter
, count_clusters
);
4840 ext4_mb_unload_buddy(&e4b
);
4842 /* We dirtied the bitmap block */
4843 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4844 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4846 /* And the group descriptor block */
4847 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4848 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4852 if (overflow
&& !err
) {
4860 ext4_std_error(sb
, err
);
4865 * ext4_group_add_blocks() -- Add given blocks to an existing group
4866 * @handle: handle to this transaction
4868 * @block: start physical block to add to the block group
4869 * @count: number of blocks to free
4871 * This marks the blocks as free in the bitmap and buddy.
4873 int ext4_group_add_blocks(handle_t
*handle
, struct super_block
*sb
,
4874 ext4_fsblk_t block
, unsigned long count
)
4876 struct buffer_head
*bitmap_bh
= NULL
;
4877 struct buffer_head
*gd_bh
;
4878 ext4_group_t block_group
;
4881 struct ext4_group_desc
*desc
;
4882 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4883 struct ext4_buddy e4b
;
4884 int err
= 0, ret
, blk_free_count
;
4885 ext4_grpblk_t blocks_freed
;
4887 ext4_debug("Adding block(s) %llu-%llu\n", block
, block
+ count
- 1);
4892 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4894 * Check to see if we are freeing blocks across a group
4897 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4898 ext4_warning(sb
, "too much blocks added to group %u\n",
4904 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4910 desc
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4916 if (in_range(ext4_block_bitmap(sb
, desc
), block
, count
) ||
4917 in_range(ext4_inode_bitmap(sb
, desc
), block
, count
) ||
4918 in_range(block
, ext4_inode_table(sb
, desc
), sbi
->s_itb_per_group
) ||
4919 in_range(block
+ count
- 1, ext4_inode_table(sb
, desc
),
4920 sbi
->s_itb_per_group
)) {
4921 ext4_error(sb
, "Adding blocks in system zones - "
4922 "Block = %llu, count = %lu",
4928 BUFFER_TRACE(bitmap_bh
, "getting write access");
4929 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4934 * We are about to modify some metadata. Call the journal APIs
4935 * to unshare ->b_data if a currently-committing transaction is
4938 BUFFER_TRACE(gd_bh
, "get_write_access");
4939 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4943 for (i
= 0, blocks_freed
= 0; i
< count
; i
++) {
4944 BUFFER_TRACE(bitmap_bh
, "clear bit");
4945 if (!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
)) {
4946 ext4_error(sb
, "bit already cleared for block %llu",
4947 (ext4_fsblk_t
)(block
+ i
));
4948 BUFFER_TRACE(bitmap_bh
, "bit already cleared");
4954 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4959 * need to update group_info->bb_free and bitmap
4960 * with group lock held. generate_buddy look at
4961 * them with group lock_held
4963 ext4_lock_group(sb
, block_group
);
4964 mb_clear_bits(bitmap_bh
->b_data
, bit
, count
);
4965 mb_free_blocks(NULL
, &e4b
, bit
, count
);
4966 blk_free_count
= blocks_freed
+ ext4_free_group_clusters(sb
, desc
);
4967 ext4_free_group_clusters_set(sb
, desc
, blk_free_count
);
4968 ext4_block_bitmap_csum_set(sb
, block_group
, desc
, bitmap_bh
);
4969 ext4_group_desc_csum_set(sb
, block_group
, desc
);
4970 ext4_unlock_group(sb
, block_group
);
4971 percpu_counter_add(&sbi
->s_freeclusters_counter
,
4972 EXT4_NUM_B2C(sbi
, blocks_freed
));
4974 if (sbi
->s_log_groups_per_flex
) {
4975 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4976 atomic64_add(EXT4_NUM_B2C(sbi
, blocks_freed
),
4977 &sbi
->s_flex_groups
[flex_group
].free_clusters
);
4980 ext4_mb_unload_buddy(&e4b
);
4982 /* We dirtied the bitmap block */
4983 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4984 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4986 /* And the group descriptor block */
4987 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4988 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4994 ext4_std_error(sb
, err
);
4999 * ext4_trim_extent -- function to TRIM one single free extent in the group
5000 * @sb: super block for the file system
5001 * @start: starting block of the free extent in the alloc. group
5002 * @count: number of blocks to TRIM
5003 * @group: alloc. group we are working with
5004 * @e4b: ext4 buddy for the group
5005 * @blkdev_flags: flags for the block device
5007 * Trim "count" blocks starting at "start" in the "group". To assure that no
5008 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5009 * be called with under the group lock.
5011 static int ext4_trim_extent(struct super_block
*sb
, int start
, int count
,
5012 ext4_group_t group
, struct ext4_buddy
*e4b
,
5013 unsigned long blkdev_flags
)
5015 struct ext4_free_extent ex
;
5018 trace_ext4_trim_extent(sb
, group
, start
, count
);
5020 assert_spin_locked(ext4_group_lock_ptr(sb
, group
));
5022 ex
.fe_start
= start
;
5023 ex
.fe_group
= group
;
5027 * Mark blocks used, so no one can reuse them while
5030 mb_mark_used(e4b
, &ex
);
5031 ext4_unlock_group(sb
, group
);
5032 ret
= ext4_issue_discard(sb
, group
, start
, count
, blkdev_flags
);
5033 ext4_lock_group(sb
, group
);
5034 mb_free_blocks(NULL
, e4b
, start
, ex
.fe_len
);
5039 * ext4_trim_all_free -- function to trim all free space in alloc. group
5040 * @sb: super block for file system
5041 * @group: group to be trimmed
5042 * @start: first group block to examine
5043 * @max: last group block to examine
5044 * @minblocks: minimum extent block count
5045 * @blkdev_flags: flags for the block device
5047 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5048 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5052 * ext4_trim_all_free walks through group's block bitmap searching for free
5053 * extents. When the free extent is found, mark it as used in group buddy
5054 * bitmap. Then issue a TRIM command on this extent and free the extent in
5055 * the group buddy bitmap. This is done until whole group is scanned.
5057 static ext4_grpblk_t
5058 ext4_trim_all_free(struct super_block
*sb
, ext4_group_t group
,
5059 ext4_grpblk_t start
, ext4_grpblk_t max
,
5060 ext4_grpblk_t minblocks
, unsigned long blkdev_flags
)
5063 ext4_grpblk_t next
, count
= 0, free_count
= 0;
5064 struct ext4_buddy e4b
;
5067 trace_ext4_trim_all_free(sb
, group
, start
, max
);
5069 ret
= ext4_mb_load_buddy(sb
, group
, &e4b
);
5071 ext4_error(sb
, "Error in loading buddy "
5072 "information for %u", group
);
5075 bitmap
= e4b
.bd_bitmap
;
5077 ext4_lock_group(sb
, group
);
5078 if (EXT4_MB_GRP_WAS_TRIMMED(e4b
.bd_info
) &&
5079 minblocks
>= atomic_read(&EXT4_SB(sb
)->s_last_trim_minblks
))
5082 start
= (e4b
.bd_info
->bb_first_free
> start
) ?
5083 e4b
.bd_info
->bb_first_free
: start
;
5085 while (start
<= max
) {
5086 start
= mb_find_next_zero_bit(bitmap
, max
+ 1, start
);
5089 next
= mb_find_next_bit(bitmap
, max
+ 1, start
);
5091 if ((next
- start
) >= minblocks
) {
5092 ret
= ext4_trim_extent(sb
, start
,
5093 next
- start
, group
, &e4b
,
5095 if (ret
&& ret
!= -EOPNOTSUPP
)
5098 count
+= next
- start
;
5100 free_count
+= next
- start
;
5103 if (fatal_signal_pending(current
)) {
5104 count
= -ERESTARTSYS
;
5108 if (need_resched()) {
5109 ext4_unlock_group(sb
, group
);
5111 ext4_lock_group(sb
, group
);
5114 if ((e4b
.bd_info
->bb_free
- free_count
) < minblocks
)
5120 EXT4_MB_GRP_SET_TRIMMED(e4b
.bd_info
);
5123 ext4_unlock_group(sb
, group
);
5124 ext4_mb_unload_buddy(&e4b
);
5126 ext4_debug("trimmed %d blocks in the group %d\n",
5133 * ext4_trim_fs() -- trim ioctl handle function
5134 * @sb: superblock for filesystem
5135 * @range: fstrim_range structure
5136 * @blkdev_flags: flags for the block device
5138 * start: First Byte to trim
5139 * len: number of Bytes to trim from start
5140 * minlen: minimum extent length in Bytes
5141 * ext4_trim_fs goes through all allocation groups containing Bytes from
5142 * start to start+len. For each such a group ext4_trim_all_free function
5143 * is invoked to trim all free space.
5145 int ext4_trim_fs(struct super_block
*sb
, struct fstrim_range
*range
,
5146 unsigned long blkdev_flags
)
5148 struct ext4_group_info
*grp
;
5149 ext4_group_t group
, first_group
, last_group
;
5150 ext4_grpblk_t cnt
= 0, first_cluster
, last_cluster
;
5151 uint64_t start
, end
, minlen
, trimmed
= 0;
5152 ext4_fsblk_t first_data_blk
=
5153 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
5154 ext4_fsblk_t max_blks
= ext4_blocks_count(EXT4_SB(sb
)->s_es
);
5157 start
= range
->start
>> sb
->s_blocksize_bits
;
5158 end
= start
+ (range
->len
>> sb
->s_blocksize_bits
) - 1;
5159 minlen
= EXT4_NUM_B2C(EXT4_SB(sb
),
5160 range
->minlen
>> sb
->s_blocksize_bits
);
5162 if (minlen
> EXT4_CLUSTERS_PER_GROUP(sb
) ||
5163 start
>= max_blks
||
5164 range
->len
< sb
->s_blocksize
)
5166 if (end
>= max_blks
)
5168 if (end
<= first_data_blk
)
5170 if (start
< first_data_blk
)
5171 start
= first_data_blk
;
5173 /* Determine first and last group to examine based on start and end */
5174 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) start
,
5175 &first_group
, &first_cluster
);
5176 ext4_get_group_no_and_offset(sb
, (ext4_fsblk_t
) end
,
5177 &last_group
, &last_cluster
);
5179 /* end now represents the last cluster to discard in this group */
5180 end
= EXT4_CLUSTERS_PER_GROUP(sb
) - 1;
5182 for (group
= first_group
; group
<= last_group
; group
++) {
5183 grp
= ext4_get_group_info(sb
, group
);
5184 /* We only do this if the grp has never been initialized */
5185 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp
))) {
5186 ret
= ext4_mb_init_group(sb
, group
, GFP_NOFS
);
5192 * For all the groups except the last one, last cluster will
5193 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5194 * change it for the last group, note that last_cluster is
5195 * already computed earlier by ext4_get_group_no_and_offset()
5197 if (group
== last_group
)
5200 if (grp
->bb_free
>= minlen
) {
5201 cnt
= ext4_trim_all_free(sb
, group
, first_cluster
,
5202 end
, minlen
, blkdev_flags
);
5211 * For every group except the first one, we are sure
5212 * that the first cluster to discard will be cluster #0.
5218 atomic_set(&EXT4_SB(sb
)->s_last_trim_minblks
, minlen
);
5221 range
->len
= EXT4_C2B(EXT4_SB(sb
), trimmed
) << sb
->s_blocksize_bits
;