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 <linux/time.h>
26 #include <linux/namei.h>
27 #include <linux/ext4_jbd2.h>
28 #include <linux/ext4_fs.h>
29 #include <linux/quotaops.h>
30 #include <linux/buffer_head.h>
31 #include <linux/module.h>
32 #include <linux/swap.h>
33 #include <linux/proc_fs.h>
34 #include <linux/pagemap.h>
35 #include <linux/seq_file.h>
36 #include <linux/version.h>
41 * - test ext4_ext_search_left() and ext4_ext_search_right()
42 * - search for metadata in few groups
45 * - normalization should take into account whether file is still open
46 * - discard preallocations if no free space left (policy?)
47 * - don't normalize tails
49 * - reservation for superuser
52 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
53 * - track min/max extents in each group for better group selection
54 * - mb_mark_used() may allocate chunk right after splitting buddy
55 * - tree of groups sorted by number of free blocks
60 * The allocation request involve request for multiple number of blocks
61 * near to the goal(block) value specified.
63 * During initialization phase of the allocator we decide to use the group
64 * preallocation or inode preallocation depending on the size file. The
65 * size of the file could be the resulting file size we would have after
66 * allocation or the current file size which ever is larger. If the size is
67 * less that sbi->s_mb_stream_request we select the group
68 * preallocation. The default value of s_mb_stream_request is 16
69 * blocks. This can also be tuned via
70 * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms
71 * of number of blocks.
73 * The main motivation for having small file use group preallocation is to
74 * ensure that we have small file closer in the disk.
76 * First stage the allocator looks at the inode prealloc list
77 * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for
78 * this particular inode. The inode prealloc space is represented as:
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 -> lenght for this prealloc space
83 * pa_free -> free space available in this prealloc space
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 make sure that
88 * that the 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 repreasented 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) withing 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 buddy][ group 0 bitmap] [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. Default value of s_mb_group_prealloc is set to
139 * 512 blocks. This can be tuned via
140 * /proc/fs/ext4/<partition/group_prealloc. The value is represented in
141 * terms of number of blocks. If we have mounted the file system with -O
142 * stripe=<value> option the group prealloc request is normalized to the
143 * stripe value (sbi->s_stripe)
145 * The regular allocator(using the buddy cache) support few tunables.
147 * /proc/fs/ext4/<partition>/min_to_scan
148 * /proc/fs/ext4/<partition>/max_to_scan
149 * /proc/fs/ext4/<partition>/order2_req
151 * The regular allocator use buddy scan only if the request len is power of
152 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
153 * value of s_mb_order2_reqs can be tuned via
154 * /proc/fs/ext4/<partition>/order2_req. If the request len is equal to
155 * stripe size (sbi->s_stripe), we try to search for contigous block in
156 * stripe size. This should result in better allocation on RAID setup. If
157 * not we search in the specific group using bitmap for best extents. The
158 * tunable min_to_scan and max_to_scan controll the behaviour here.
159 * min_to_scan indicate how long the mballoc __must__ look for a best
160 * extent and max_to_scanindicate how long the mballoc __can__ look for a
161 * best extent in the found extents. Searching for the blocks starts with
162 * the group specified as the goal value in allocation context via
163 * ac_g_ex. Each group is first checked based on the criteria whether it
164 * can used for allocation. ext4_mb_good_group explains how the groups are
167 * Both the prealloc space are getting populated as above. So for the first
168 * request we will hit the buddy cache which will result in this prealloc
169 * space getting filled. The prealloc space is then later used for the
170 * subsequent request.
174 * mballoc operates on the following data:
176 * - in-core buddy (actually includes buddy and bitmap)
177 * - preallocation descriptors (PAs)
179 * there are two types of preallocations:
181 * assiged to specific inode and can be used for this inode only.
182 * it describes part of inode's space preallocated to specific
183 * physical blocks. any block from that preallocated can be used
184 * independent. the descriptor just tracks number of blocks left
185 * unused. so, before taking some block from descriptor, one must
186 * make sure corresponded logical block isn't allocated yet. this
187 * also means that freeing any block within descriptor's range
188 * must discard all preallocated blocks.
190 * assigned to specific locality group which does not translate to
191 * permanent set of inodes: inode can join and leave group. space
192 * from this type of preallocation can be used for any inode. thus
193 * it's consumed from the beginning to the end.
195 * relation between them can be expressed as:
196 * in-core buddy = on-disk bitmap + preallocation descriptors
198 * this mean blocks mballoc considers used are:
199 * - allocated blocks (persistent)
200 * - preallocated blocks (non-persistent)
202 * consistency in mballoc world means that at any time a block is either
203 * free or used in ALL structures. notice: "any time" should not be read
204 * literally -- time is discrete and delimited by locks.
206 * to keep it simple, we don't use block numbers, instead we count number of
207 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
209 * all operations can be expressed as:
210 * - init buddy: buddy = on-disk + PAs
211 * - new PA: buddy += N; PA = N
212 * - use inode PA: on-disk += N; PA -= N
213 * - discard inode PA buddy -= on-disk - PA; PA = 0
214 * - use locality group PA on-disk += N; PA -= N
215 * - discard locality group PA buddy -= PA; PA = 0
216 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
217 * is used in real operation because we can't know actual used
218 * bits from PA, only from on-disk bitmap
220 * if we follow this strict logic, then all operations above should be atomic.
221 * given some of them can block, we'd have to use something like semaphores
222 * killing performance on high-end SMP hardware. let's try to relax it using
223 * the following knowledge:
224 * 1) if buddy is referenced, it's already initialized
225 * 2) while block is used in buddy and the buddy is referenced,
226 * nobody can re-allocate that block
227 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
228 * bit set and PA claims same block, it's OK. IOW, one can set bit in
229 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
232 * so, now we're building a concurrency table:
235 * blocks for PA are allocated in the buddy, buddy must be referenced
236 * until PA is linked to allocation group to avoid concurrent buddy init
238 * we need to make sure that either on-disk bitmap or PA has uptodate data
239 * given (3) we care that PA-=N operation doesn't interfere with init
241 * the simplest way would be to have buddy initialized by the discard
242 * - use locality group PA
243 * again PA-=N must be serialized with init
244 * - discard locality group PA
245 * the simplest way would be to have buddy initialized by the discard
248 * i_data_sem serializes them
250 * discard process must wait until PA isn't used by another process
251 * - use locality group PA
252 * some mutex should serialize them
253 * - discard locality group PA
254 * discard process must wait until PA isn't used by another process
257 * i_data_sem or another mutex should serializes them
259 * discard process must wait until PA isn't used by another process
260 * - use locality group PA
261 * nothing wrong here -- they're different PAs covering different blocks
262 * - discard locality group PA
263 * discard process must wait until PA isn't used by another process
265 * now we're ready to make few consequences:
266 * - PA is referenced and while it is no discard is possible
267 * - PA is referenced until block isn't marked in on-disk bitmap
268 * - PA changes only after on-disk bitmap
269 * - discard must not compete with init. either init is done before
270 * any discard or they're serialized somehow
271 * - buddy init as sum of on-disk bitmap and PAs is done atomically
273 * a special case when we've used PA to emptiness. no need to modify buddy
274 * in this case, but we should care about concurrent init
279 * Logic in few words:
284 * mark bits in on-disk bitmap
287 * - use preallocation:
288 * find proper PA (per-inode or group)
290 * mark bits in on-disk bitmap
296 * mark bits in on-disk bitmap
299 * - discard preallocations in group:
301 * move them onto local list
302 * load on-disk bitmap
304 * remove PA from object (inode or locality group)
305 * mark free blocks in-core
307 * - discard inode's preallocations:
314 * - bitlock on a group (group)
315 * - object (inode/locality) (object)
326 * - release consumed pa:
331 * - generate in-core bitmap:
335 * - discard all for given object (inode, locality group):
340 * - discard all for given group:
349 * with AGGRESSIVE_CHECK allocator runs consistency checks over
350 * structures. these checks slow things down a lot
352 #define AGGRESSIVE_CHECK__
355 * with DOUBLE_CHECK defined mballoc creates persistent in-core
356 * bitmaps, maintains and uses them to check for double allocations
358 #define DOUBLE_CHECK__
364 #define mb_debug(fmt, a...) printk(fmt, ##a)
366 #define mb_debug(fmt, a...)
370 * with EXT4_MB_HISTORY mballoc stores last N allocations in memory
371 * and you can monitor it in /proc/fs/ext4/<dev>/mb_history
373 #define EXT4_MB_HISTORY
374 #define EXT4_MB_HISTORY_ALLOC 1 /* allocation */
375 #define EXT4_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */
376 #define EXT4_MB_HISTORY_DISCARD 4 /* preallocation discarded */
377 #define EXT4_MB_HISTORY_FREE 8 /* free */
379 #define EXT4_MB_HISTORY_DEFAULT (EXT4_MB_HISTORY_ALLOC | \
380 EXT4_MB_HISTORY_PREALLOC)
383 * How long mballoc can look for a best extent (in found extents)
385 #define MB_DEFAULT_MAX_TO_SCAN 200
388 * How long mballoc must look for a best extent
390 #define MB_DEFAULT_MIN_TO_SCAN 10
393 * How many groups mballoc will scan looking for the best chunk
395 #define MB_DEFAULT_MAX_GROUPS_TO_SCAN 5
398 * with 'ext4_mb_stats' allocator will collect stats that will be
399 * shown at umount. The collecting costs though!
401 #define MB_DEFAULT_STATS 1
404 * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served
405 * by the stream allocator, which purpose is to pack requests
406 * as close each to other as possible to produce smooth I/O traffic
407 * We use locality group prealloc space for stream request.
408 * We can tune the same via /proc/fs/ext4/<parition>/stream_req
410 #define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */
413 * for which requests use 2^N search using buddies
415 #define MB_DEFAULT_ORDER2_REQS 2
418 * default group prealloc size 512 blocks
420 #define MB_DEFAULT_GROUP_PREALLOC 512
422 static struct kmem_cache
*ext4_pspace_cachep
;
423 static struct kmem_cache
*ext4_ac_cachep
;
425 #ifdef EXT4_BB_MAX_BLOCKS
426 #undef EXT4_BB_MAX_BLOCKS
428 #define EXT4_BB_MAX_BLOCKS 30
430 struct ext4_free_metadata
{
433 ext4_grpblk_t blocks
[EXT4_BB_MAX_BLOCKS
];
434 struct list_head list
;
437 struct ext4_group_info
{
438 unsigned long bb_state
;
439 unsigned long bb_tid
;
440 struct ext4_free_metadata
*bb_md_cur
;
441 unsigned short bb_first_free
;
442 unsigned short bb_free
;
443 unsigned short bb_fragments
;
444 struct list_head bb_prealloc_list
;
448 unsigned short bb_counters
[];
451 #define EXT4_GROUP_INFO_NEED_INIT_BIT 0
452 #define EXT4_GROUP_INFO_LOCKED_BIT 1
454 #define EXT4_MB_GRP_NEED_INIT(grp) \
455 (test_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &((grp)->bb_state)))
458 struct ext4_prealloc_space
{
459 struct list_head pa_inode_list
;
460 struct list_head pa_group_list
;
462 struct list_head pa_tmp_list
;
463 struct rcu_head pa_rcu
;
468 ext4_fsblk_t pa_pstart
; /* phys. block */
469 ext4_lblk_t pa_lstart
; /* log. block */
470 unsigned short pa_len
; /* len of preallocated chunk */
471 unsigned short pa_free
; /* how many blocks are free */
472 unsigned short pa_linear
; /* consumed in one direction
473 * strictly, for grp prealloc */
474 spinlock_t
*pa_obj_lock
;
475 struct inode
*pa_inode
; /* hack, for history only */
479 struct ext4_free_extent
{
480 ext4_lblk_t fe_logical
;
481 ext4_grpblk_t fe_start
;
482 ext4_group_t fe_group
;
488 * we try to group all related changes together
489 * so that writeback can flush/allocate them together as well
491 struct ext4_locality_group
{
493 struct mutex lg_mutex
; /* to serialize allocates */
494 struct list_head lg_prealloc_list
;/* list of preallocations */
495 spinlock_t lg_prealloc_lock
;
498 struct ext4_allocation_context
{
499 struct inode
*ac_inode
;
500 struct super_block
*ac_sb
;
502 /* original request */
503 struct ext4_free_extent ac_o_ex
;
505 /* goal request (after normalization) */
506 struct ext4_free_extent ac_g_ex
;
508 /* the best found extent */
509 struct ext4_free_extent ac_b_ex
;
511 /* copy of the bext found extent taken before preallocation efforts */
512 struct ext4_free_extent ac_f_ex
;
514 /* number of iterations done. we have to track to limit searching */
515 unsigned long ac_ex_scanned
;
516 __u16 ac_groups_scanned
;
520 __u16 ac_flags
; /* allocation hints */
524 __u8 ac_2order
; /* if request is to allocate 2^N blocks and
525 * N > 0, the field stores N, otherwise 0 */
526 __u8 ac_op
; /* operation, for history only */
527 struct page
*ac_bitmap_page
;
528 struct page
*ac_buddy_page
;
529 struct ext4_prealloc_space
*ac_pa
;
530 struct ext4_locality_group
*ac_lg
;
533 #define AC_STATUS_CONTINUE 1
534 #define AC_STATUS_FOUND 2
535 #define AC_STATUS_BREAK 3
537 struct ext4_mb_history
{
538 struct ext4_free_extent orig
; /* orig allocation */
539 struct ext4_free_extent goal
; /* goal allocation */
540 struct ext4_free_extent result
; /* result allocation */
543 __u16 found
; /* how many extents have been found */
544 __u16 groups
; /* how many groups have been scanned */
545 __u16 tail
; /* what tail broke some buddy */
546 __u16 buddy
; /* buddy the tail ^^^ broke */
548 __u8 cr
:3; /* which phase the result extent was found at */
554 struct page
*bd_buddy_page
;
556 struct page
*bd_bitmap_page
;
558 struct ext4_group_info
*bd_info
;
559 struct super_block
*bd_sb
;
561 ext4_group_t bd_group
;
563 #define EXT4_MB_BITMAP(e4b) ((e4b)->bd_bitmap)
564 #define EXT4_MB_BUDDY(e4b) ((e4b)->bd_buddy)
566 #ifndef EXT4_MB_HISTORY
567 static inline void ext4_mb_store_history(struct ext4_allocation_context
*ac
)
572 static void ext4_mb_store_history(struct ext4_allocation_context
*ac
);
575 #define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
577 static struct proc_dir_entry
*proc_root_ext4
;
578 struct buffer_head
*read_block_bitmap(struct super_block
*, ext4_group_t
);
579 ext4_fsblk_t
ext4_new_blocks_old(handle_t
*handle
, struct inode
*inode
,
580 ext4_fsblk_t goal
, unsigned long *count
, int *errp
);
582 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
584 static void ext4_mb_poll_new_transaction(struct super_block
*, handle_t
*);
585 static void ext4_mb_free_committed_blocks(struct super_block
*);
586 static void ext4_mb_return_to_preallocation(struct inode
*inode
,
587 struct ext4_buddy
*e4b
, sector_t block
,
589 static void ext4_mb_put_pa(struct ext4_allocation_context
*,
590 struct super_block
*, struct ext4_prealloc_space
*pa
);
591 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
);
592 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
);
595 static inline void ext4_lock_group(struct super_block
*sb
, ext4_group_t group
)
597 struct ext4_group_info
*grinfo
= ext4_get_group_info(sb
, group
);
599 bit_spin_lock(EXT4_GROUP_INFO_LOCKED_BIT
, &(grinfo
->bb_state
));
602 static inline void ext4_unlock_group(struct super_block
*sb
,
605 struct ext4_group_info
*grinfo
= ext4_get_group_info(sb
, group
);
607 bit_spin_unlock(EXT4_GROUP_INFO_LOCKED_BIT
, &(grinfo
->bb_state
));
610 static inline int ext4_is_group_locked(struct super_block
*sb
,
613 struct ext4_group_info
*grinfo
= ext4_get_group_info(sb
, group
);
615 return bit_spin_is_locked(EXT4_GROUP_INFO_LOCKED_BIT
,
616 &(grinfo
->bb_state
));
619 static ext4_fsblk_t
ext4_grp_offs_to_block(struct super_block
*sb
,
620 struct ext4_free_extent
*fex
)
624 block
= (ext4_fsblk_t
) fex
->fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
626 + le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
630 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
632 #if BITS_PER_LONG == 64
633 *bit
+= ((unsigned long) addr
& 7UL) << 3;
634 addr
= (void *) ((unsigned long) addr
& ~7UL);
635 #elif BITS_PER_LONG == 32
636 *bit
+= ((unsigned long) addr
& 3UL) << 3;
637 addr
= (void *) ((unsigned long) addr
& ~3UL);
639 #error "how many bits you are?!"
644 static inline int mb_test_bit(int bit
, void *addr
)
647 * ext4_test_bit on architecture like powerpc
648 * needs unsigned long aligned address
650 addr
= mb_correct_addr_and_bit(&bit
, addr
);
651 return ext4_test_bit(bit
, addr
);
654 static inline void mb_set_bit(int bit
, void *addr
)
656 addr
= mb_correct_addr_and_bit(&bit
, addr
);
657 ext4_set_bit(bit
, addr
);
660 static inline void mb_set_bit_atomic(spinlock_t
*lock
, int bit
, void *addr
)
662 addr
= mb_correct_addr_and_bit(&bit
, addr
);
663 ext4_set_bit_atomic(lock
, bit
, addr
);
666 static inline void mb_clear_bit(int bit
, void *addr
)
668 addr
= mb_correct_addr_and_bit(&bit
, addr
);
669 ext4_clear_bit(bit
, addr
);
672 static inline void mb_clear_bit_atomic(spinlock_t
*lock
, int bit
, void *addr
)
674 addr
= mb_correct_addr_and_bit(&bit
, addr
);
675 ext4_clear_bit_atomic(lock
, bit
, addr
);
678 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
681 addr
= mb_correct_addr_and_bit(&fix
, addr
);
685 return ext4_find_next_zero_bit(addr
, max
, start
) - fix
;
688 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
691 addr
= mb_correct_addr_and_bit(&fix
, addr
);
695 return ext4_find_next_bit(addr
, max
, start
) - fix
;
698 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
702 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
705 if (order
> e4b
->bd_blkbits
+ 1) {
710 /* at order 0 we see each particular block */
711 *max
= 1 << (e4b
->bd_blkbits
+ 3);
713 return EXT4_MB_BITMAP(e4b
);
715 bb
= EXT4_MB_BUDDY(e4b
) + EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
716 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
722 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
723 int first
, int count
)
726 struct super_block
*sb
= e4b
->bd_sb
;
728 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
730 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
731 for (i
= 0; i
< count
; i
++) {
732 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
733 ext4_fsblk_t blocknr
;
734 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
735 blocknr
+= first
+ i
;
737 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
739 ext4_error(sb
, __FUNCTION__
, "double-free of inode"
740 " %lu's block %llu(bit %u in group %lu)\n",
741 inode
? inode
->i_ino
: 0, blocknr
,
742 first
+ i
, e4b
->bd_group
);
744 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
748 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
752 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
754 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
755 for (i
= 0; i
< count
; i
++) {
756 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
757 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
761 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
763 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
764 unsigned char *b1
, *b2
;
766 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
767 b2
= (unsigned char *) bitmap
;
768 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
769 if (b1
[i
] != b2
[i
]) {
770 printk("corruption in group %lu at byte %u(%u):"
771 " %x in copy != %x on disk/prealloc\n",
772 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
780 static inline void mb_free_blocks_double(struct inode
*inode
,
781 struct ext4_buddy
*e4b
, int first
, int count
)
785 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
786 int first
, int count
)
790 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
796 #ifdef AGGRESSIVE_CHECK
798 #define MB_CHECK_ASSERT(assert) \
802 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
803 function, file, line, # assert); \
808 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
809 const char *function
, int line
)
811 struct super_block
*sb
= e4b
->bd_sb
;
812 int order
= e4b
->bd_blkbits
+ 1;
819 struct ext4_group_info
*grp
;
822 struct list_head
*cur
;
826 if (!test_opt(sb
, MBALLOC
))
830 static int mb_check_counter
;
831 if (mb_check_counter
++ % 100 != 0)
836 buddy
= mb_find_buddy(e4b
, order
, &max
);
837 MB_CHECK_ASSERT(buddy
);
838 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
839 MB_CHECK_ASSERT(buddy2
);
840 MB_CHECK_ASSERT(buddy
!= buddy2
);
841 MB_CHECK_ASSERT(max
* 2 == max2
);
844 for (i
= 0; i
< max
; i
++) {
846 if (mb_test_bit(i
, buddy
)) {
847 /* only single bit in buddy2 may be 1 */
848 if (!mb_test_bit(i
<< 1, buddy2
)) {
850 mb_test_bit((i
<<1)+1, buddy2
));
851 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
853 mb_test_bit(i
<< 1, buddy2
));
858 /* both bits in buddy2 must be 0 */
859 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
860 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
862 for (j
= 0; j
< (1 << order
); j
++) {
863 k
= (i
* (1 << order
)) + j
;
865 !mb_test_bit(k
, EXT4_MB_BITMAP(e4b
)));
869 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
874 buddy
= mb_find_buddy(e4b
, 0, &max
);
875 for (i
= 0; i
< max
; i
++) {
876 if (!mb_test_bit(i
, buddy
)) {
877 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
885 /* check used bits only */
886 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
887 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
889 MB_CHECK_ASSERT(k
< max2
);
890 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
893 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
894 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
896 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
897 buddy
= mb_find_buddy(e4b
, 0, &max
);
898 list_for_each(cur
, &grp
->bb_prealloc_list
) {
899 ext4_group_t groupnr
;
900 struct ext4_prealloc_space
*pa
;
901 pa
= list_entry(cur
, struct ext4_prealloc_space
, group_list
);
902 ext4_get_group_no_and_offset(sb
, pa
->pstart
, &groupnr
, &k
);
903 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
904 for (i
= 0; i
< pa
->len
; i
++)
905 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
909 #undef MB_CHECK_ASSERT
910 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
911 __FILE__, __FUNCTION__, __LINE__)
913 #define mb_check_buddy(e4b)
916 /* FIXME!! need more doc */
917 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
918 void *buddy
, unsigned first
, int len
,
919 struct ext4_group_info
*grp
)
921 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
924 unsigned short chunk
;
925 unsigned short border
;
927 BUG_ON(len
> EXT4_BLOCKS_PER_GROUP(sb
));
929 border
= 2 << sb
->s_blocksize_bits
;
932 /* find how many blocks can be covered since this position */
933 max
= ffs(first
| border
) - 1;
935 /* find how many blocks of power 2 we need to mark */
942 /* mark multiblock chunks only */
943 grp
->bb_counters
[min
]++;
945 mb_clear_bit(first
>> min
,
946 buddy
+ sbi
->s_mb_offsets
[min
]);
953 static void ext4_mb_generate_buddy(struct super_block
*sb
,
954 void *buddy
, void *bitmap
, ext4_group_t group
)
956 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
957 unsigned short max
= EXT4_BLOCKS_PER_GROUP(sb
);
958 unsigned short i
= 0;
959 unsigned short first
;
962 unsigned fragments
= 0;
963 unsigned long long period
= get_cycles();
965 /* initialize buddy from bitmap which is aggregation
966 * of on-disk bitmap and preallocations */
967 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
968 grp
->bb_first_free
= i
;
972 i
= mb_find_next_bit(bitmap
, max
, i
);
976 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
978 grp
->bb_counters
[0]++;
980 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
982 grp
->bb_fragments
= fragments
;
984 if (free
!= grp
->bb_free
) {
985 ext4_error(sb
, __FUNCTION__
,
986 "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
987 group
, free
, grp
->bb_free
);
989 * If we intent to continue, we consider group descritor
990 * corrupt and update bb_free using bitmap value
995 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
997 period
= get_cycles() - period
;
998 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
999 EXT4_SB(sb
)->s_mb_buddies_generated
++;
1000 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
1001 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
1004 /* The buddy information is attached the buddy cache inode
1005 * for convenience. The information regarding each group
1006 * is loaded via ext4_mb_load_buddy. The information involve
1007 * block bitmap and buddy information. The information are
1008 * stored in the inode as
1011 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
1014 * one block each for bitmap and buddy information.
1015 * So for each group we take up 2 blocks. A page can
1016 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
1017 * So it can have information regarding groups_per_page which
1018 * is blocks_per_page/2
1021 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
1024 int blocks_per_page
;
1025 int groups_per_page
;
1028 ext4_group_t first_group
;
1030 struct super_block
*sb
;
1031 struct buffer_head
*bhs
;
1032 struct buffer_head
**bh
;
1033 struct inode
*inode
;
1037 mb_debug("init page %lu\n", page
->index
);
1039 inode
= page
->mapping
->host
;
1041 blocksize
= 1 << inode
->i_blkbits
;
1042 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
1044 groups_per_page
= blocks_per_page
>> 1;
1045 if (groups_per_page
== 0)
1046 groups_per_page
= 1;
1048 /* allocate buffer_heads to read bitmaps */
1049 if (groups_per_page
> 1) {
1051 i
= sizeof(struct buffer_head
*) * groups_per_page
;
1052 bh
= kzalloc(i
, GFP_NOFS
);
1058 first_group
= page
->index
* blocks_per_page
/ 2;
1060 /* read all groups the page covers into the cache */
1061 for (i
= 0; i
< groups_per_page
; i
++) {
1062 struct ext4_group_desc
*desc
;
1064 if (first_group
+ i
>= EXT4_SB(sb
)->s_groups_count
)
1068 desc
= ext4_get_group_desc(sb
, first_group
+ i
, NULL
);
1073 bh
[i
] = sb_getblk(sb
, ext4_block_bitmap(sb
, desc
));
1077 if (bh_uptodate_or_lock(bh
[i
]))
1080 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
1081 ext4_init_block_bitmap(sb
, bh
[i
],
1082 first_group
+ i
, desc
);
1083 set_buffer_uptodate(bh
[i
]);
1084 unlock_buffer(bh
[i
]);
1088 bh
[i
]->b_end_io
= end_buffer_read_sync
;
1089 submit_bh(READ
, bh
[i
]);
1090 mb_debug("read bitmap for group %lu\n", first_group
+ i
);
1093 /* wait for I/O completion */
1094 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
1095 wait_on_buffer(bh
[i
]);
1098 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
1099 if (!buffer_uptodate(bh
[i
]))
1102 first_block
= page
->index
* blocks_per_page
;
1103 for (i
= 0; i
< blocks_per_page
; i
++) {
1105 struct ext4_group_info
*grinfo
;
1107 group
= (first_block
+ i
) >> 1;
1108 if (group
>= EXT4_SB(sb
)->s_groups_count
)
1112 * data carry information regarding this
1113 * particular group in the format specified
1117 data
= page_address(page
) + (i
* blocksize
);
1118 bitmap
= bh
[group
- first_group
]->b_data
;
1121 * We place the buddy block and bitmap block
1124 if ((first_block
+ i
) & 1) {
1125 /* this is block of buddy */
1126 BUG_ON(incore
== NULL
);
1127 mb_debug("put buddy for group %u in page %lu/%x\n",
1128 group
, page
->index
, i
* blocksize
);
1129 memset(data
, 0xff, blocksize
);
1130 grinfo
= ext4_get_group_info(sb
, group
);
1131 grinfo
->bb_fragments
= 0;
1132 memset(grinfo
->bb_counters
, 0,
1133 sizeof(unsigned short)*(sb
->s_blocksize_bits
+2));
1135 * incore got set to the group block bitmap below
1137 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
1140 /* this is block of bitmap */
1141 BUG_ON(incore
!= NULL
);
1142 mb_debug("put bitmap for group %u in page %lu/%x\n",
1143 group
, page
->index
, i
* blocksize
);
1145 /* see comments in ext4_mb_put_pa() */
1146 ext4_lock_group(sb
, group
);
1147 memcpy(data
, bitmap
, blocksize
);
1149 /* mark all preallocated blks used in in-core bitmap */
1150 ext4_mb_generate_from_pa(sb
, data
, group
);
1151 ext4_unlock_group(sb
, group
);
1153 /* set incore so that the buddy information can be
1154 * generated using this
1159 SetPageUptodate(page
);
1163 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
1171 static int ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1172 struct ext4_buddy
*e4b
)
1174 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1175 struct inode
*inode
= sbi
->s_buddy_cache
;
1176 int blocks_per_page
;
1182 mb_debug("load group %lu\n", group
);
1184 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1186 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1187 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
1189 e4b
->bd_group
= group
;
1190 e4b
->bd_buddy_page
= NULL
;
1191 e4b
->bd_bitmap_page
= NULL
;
1194 * the buddy cache inode stores the block bitmap
1195 * and buddy information in consecutive blocks.
1196 * So for each group we need two blocks.
1199 pnum
= block
/ blocks_per_page
;
1200 poff
= block
% blocks_per_page
;
1202 /* we could use find_or_create_page(), but it locks page
1203 * what we'd like to avoid in fast path ... */
1204 page
= find_get_page(inode
->i_mapping
, pnum
);
1205 if (page
== NULL
|| !PageUptodate(page
)) {
1207 page_cache_release(page
);
1208 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1210 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1211 if (!PageUptodate(page
)) {
1212 ext4_mb_init_cache(page
, NULL
);
1213 mb_cmp_bitmaps(e4b
, page_address(page
) +
1214 (poff
* sb
->s_blocksize
));
1219 if (page
== NULL
|| !PageUptodate(page
))
1221 e4b
->bd_bitmap_page
= page
;
1222 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1223 mark_page_accessed(page
);
1226 pnum
= block
/ blocks_per_page
;
1227 poff
= block
% blocks_per_page
;
1229 page
= find_get_page(inode
->i_mapping
, pnum
);
1230 if (page
== NULL
|| !PageUptodate(page
)) {
1232 page_cache_release(page
);
1233 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1235 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1236 if (!PageUptodate(page
))
1237 ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1242 if (page
== NULL
|| !PageUptodate(page
))
1244 e4b
->bd_buddy_page
= page
;
1245 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1246 mark_page_accessed(page
);
1248 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1249 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1254 if (e4b
->bd_bitmap_page
)
1255 page_cache_release(e4b
->bd_bitmap_page
);
1256 if (e4b
->bd_buddy_page
)
1257 page_cache_release(e4b
->bd_buddy_page
);
1258 e4b
->bd_buddy
= NULL
;
1259 e4b
->bd_bitmap
= NULL
;
1263 static void ext4_mb_release_desc(struct ext4_buddy
*e4b
)
1265 if (e4b
->bd_bitmap_page
)
1266 page_cache_release(e4b
->bd_bitmap_page
);
1267 if (e4b
->bd_buddy_page
)
1268 page_cache_release(e4b
->bd_buddy_page
);
1272 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1277 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1278 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1280 bb
= EXT4_MB_BUDDY(e4b
);
1281 while (order
<= e4b
->bd_blkbits
+ 1) {
1283 if (!mb_test_bit(block
, bb
)) {
1284 /* this block is part of buddy of order 'order' */
1287 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1293 static void mb_clear_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1299 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1300 /* fast path: clear whole word at once */
1301 addr
= bm
+ (cur
>> 3);
1306 mb_clear_bit_atomic(lock
, cur
, bm
);
1311 static void mb_set_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1317 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1318 /* fast path: set whole word at once */
1319 addr
= bm
+ (cur
>> 3);
1324 mb_set_bit_atomic(lock
, cur
, bm
);
1329 static int mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1330 int first
, int count
)
1337 struct super_block
*sb
= e4b
->bd_sb
;
1339 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1340 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
1341 mb_check_buddy(e4b
);
1342 mb_free_blocks_double(inode
, e4b
, first
, count
);
1344 e4b
->bd_info
->bb_free
+= count
;
1345 if (first
< e4b
->bd_info
->bb_first_free
)
1346 e4b
->bd_info
->bb_first_free
= first
;
1348 /* let's maintain fragments counter */
1350 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1351 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1352 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1354 e4b
->bd_info
->bb_fragments
--;
1355 else if (!block
&& !max
)
1356 e4b
->bd_info
->bb_fragments
++;
1358 /* let's maintain buddy itself */
1359 while (count
-- > 0) {
1363 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1364 ext4_fsblk_t blocknr
;
1365 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
1368 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
1370 ext4_error(sb
, __FUNCTION__
, "double-free of inode"
1371 " %lu's block %llu(bit %u in group %lu)\n",
1372 inode
? inode
->i_ino
: 0, blocknr
, block
,
1375 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1376 e4b
->bd_info
->bb_counters
[order
]++;
1378 /* start of the buddy */
1379 buddy
= mb_find_buddy(e4b
, order
, &max
);
1383 if (mb_test_bit(block
, buddy
) ||
1384 mb_test_bit(block
+ 1, buddy
))
1387 /* both the buddies are free, try to coalesce them */
1388 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1394 /* for special purposes, we don't set
1395 * free bits in bitmap */
1396 mb_set_bit(block
, buddy
);
1397 mb_set_bit(block
+ 1, buddy
);
1399 e4b
->bd_info
->bb_counters
[order
]--;
1400 e4b
->bd_info
->bb_counters
[order
]--;
1404 e4b
->bd_info
->bb_counters
[order
]++;
1406 mb_clear_bit(block
, buddy2
);
1410 mb_check_buddy(e4b
);
1415 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1416 int needed
, struct ext4_free_extent
*ex
)
1423 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1426 buddy
= mb_find_buddy(e4b
, order
, &max
);
1427 BUG_ON(buddy
== NULL
);
1428 BUG_ON(block
>= max
);
1429 if (mb_test_bit(block
, buddy
)) {
1436 /* FIXME dorp order completely ? */
1437 if (likely(order
== 0)) {
1438 /* find actual order */
1439 order
= mb_find_order_for_block(e4b
, block
);
1440 block
= block
>> order
;
1443 ex
->fe_len
= 1 << order
;
1444 ex
->fe_start
= block
<< order
;
1445 ex
->fe_group
= e4b
->bd_group
;
1447 /* calc difference from given start */
1448 next
= next
- ex
->fe_start
;
1450 ex
->fe_start
+= next
;
1452 while (needed
> ex
->fe_len
&&
1453 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1455 if (block
+ 1 >= max
)
1458 next
= (block
+ 1) * (1 << order
);
1459 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1462 ord
= mb_find_order_for_block(e4b
, next
);
1465 block
= next
>> order
;
1466 ex
->fe_len
+= 1 << order
;
1469 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1473 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1479 int start
= ex
->fe_start
;
1480 int len
= ex
->fe_len
;
1485 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1486 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1487 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1488 mb_check_buddy(e4b
);
1489 mb_mark_used_double(e4b
, start
, len
);
1491 e4b
->bd_info
->bb_free
-= len
;
1492 if (e4b
->bd_info
->bb_first_free
== start
)
1493 e4b
->bd_info
->bb_first_free
+= len
;
1495 /* let's maintain fragments counter */
1497 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1498 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1499 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1501 e4b
->bd_info
->bb_fragments
++;
1502 else if (!mlen
&& !max
)
1503 e4b
->bd_info
->bb_fragments
--;
1505 /* let's maintain buddy itself */
1507 ord
= mb_find_order_for_block(e4b
, start
);
1509 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1510 /* the whole chunk may be allocated at once! */
1512 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1513 BUG_ON((start
>> ord
) >= max
);
1514 mb_set_bit(start
>> ord
, buddy
);
1515 e4b
->bd_info
->bb_counters
[ord
]--;
1522 /* store for history */
1524 ret
= len
| (ord
<< 16);
1526 /* we have to split large buddy */
1528 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1529 mb_set_bit(start
>> ord
, buddy
);
1530 e4b
->bd_info
->bb_counters
[ord
]--;
1533 cur
= (start
>> ord
) & ~1U;
1534 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1535 mb_clear_bit(cur
, buddy
);
1536 mb_clear_bit(cur
+ 1, buddy
);
1537 e4b
->bd_info
->bb_counters
[ord
]++;
1538 e4b
->bd_info
->bb_counters
[ord
]++;
1541 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b
->bd_sb
), ex
->fe_group
),
1542 EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1543 mb_check_buddy(e4b
);
1549 * Must be called under group lock!
1551 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1552 struct ext4_buddy
*e4b
)
1554 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1557 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1558 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1560 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1561 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1562 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1564 /* preallocation can change ac_b_ex, thus we store actually
1565 * allocated blocks for history */
1566 ac
->ac_f_ex
= ac
->ac_b_ex
;
1568 ac
->ac_status
= AC_STATUS_FOUND
;
1569 ac
->ac_tail
= ret
& 0xffff;
1570 ac
->ac_buddy
= ret
>> 16;
1572 /* XXXXXXX: SUCH A HORRIBLE **CK */
1574 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1575 get_page(ac
->ac_bitmap_page
);
1576 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1577 get_page(ac
->ac_buddy_page
);
1579 /* store last allocated for subsequent stream allocation */
1580 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1581 spin_lock(&sbi
->s_md_lock
);
1582 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1583 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1584 spin_unlock(&sbi
->s_md_lock
);
1589 * regular allocator, for general purposes allocation
1592 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1593 struct ext4_buddy
*e4b
,
1596 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1597 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1598 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1599 struct ext4_free_extent ex
;
1603 * We don't want to scan for a whole year
1605 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1606 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1607 ac
->ac_status
= AC_STATUS_BREAK
;
1612 * Haven't found good chunk so far, let's continue
1614 if (bex
->fe_len
< gex
->fe_len
)
1617 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1618 && bex
->fe_group
== e4b
->bd_group
) {
1619 /* recheck chunk's availability - we don't know
1620 * when it was found (within this lock-unlock
1622 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1623 if (max
>= gex
->fe_len
) {
1624 ext4_mb_use_best_found(ac
, e4b
);
1631 * The routine checks whether found extent is good enough. If it is,
1632 * then the extent gets marked used and flag is set to the context
1633 * to stop scanning. Otherwise, the extent is compared with the
1634 * previous found extent and if new one is better, then it's stored
1635 * in the context. Later, the best found extent will be used, if
1636 * mballoc can't find good enough extent.
1638 * FIXME: real allocation policy is to be designed yet!
1640 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1641 struct ext4_free_extent
*ex
,
1642 struct ext4_buddy
*e4b
)
1644 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1645 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1647 BUG_ON(ex
->fe_len
<= 0);
1648 BUG_ON(ex
->fe_len
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1649 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1650 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1655 * The special case - take what you catch first
1657 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1659 ext4_mb_use_best_found(ac
, e4b
);
1664 * Let's check whether the chuck is good enough
1666 if (ex
->fe_len
== gex
->fe_len
) {
1668 ext4_mb_use_best_found(ac
, e4b
);
1673 * If this is first found extent, just store it in the context
1675 if (bex
->fe_len
== 0) {
1681 * If new found extent is better, store it in the context
1683 if (bex
->fe_len
< gex
->fe_len
) {
1684 /* if the request isn't satisfied, any found extent
1685 * larger than previous best one is better */
1686 if (ex
->fe_len
> bex
->fe_len
)
1688 } else if (ex
->fe_len
> gex
->fe_len
) {
1689 /* if the request is satisfied, then we try to find
1690 * an extent that still satisfy the request, but is
1691 * smaller than previous one */
1692 if (ex
->fe_len
< bex
->fe_len
)
1696 ext4_mb_check_limits(ac
, e4b
, 0);
1699 static int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1700 struct ext4_buddy
*e4b
)
1702 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1703 ext4_group_t group
= ex
.fe_group
;
1707 BUG_ON(ex
.fe_len
<= 0);
1708 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1712 ext4_lock_group(ac
->ac_sb
, group
);
1713 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1717 ext4_mb_use_best_found(ac
, e4b
);
1720 ext4_unlock_group(ac
->ac_sb
, group
);
1721 ext4_mb_release_desc(e4b
);
1726 static int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1727 struct ext4_buddy
*e4b
)
1729 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1732 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1733 struct ext4_super_block
*es
= sbi
->s_es
;
1734 struct ext4_free_extent ex
;
1736 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1739 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1743 ext4_lock_group(ac
->ac_sb
, group
);
1744 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1745 ac
->ac_g_ex
.fe_len
, &ex
);
1747 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1750 start
= (e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
)) +
1751 ex
.fe_start
+ le32_to_cpu(es
->s_first_data_block
);
1752 /* use do_div to get remainder (would be 64-bit modulo) */
1753 if (do_div(start
, sbi
->s_stripe
) == 0) {
1756 ext4_mb_use_best_found(ac
, e4b
);
1758 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1759 BUG_ON(ex
.fe_len
<= 0);
1760 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1761 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1764 ext4_mb_use_best_found(ac
, e4b
);
1765 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1766 /* Sometimes, caller may want to merge even small
1767 * number of blocks to an existing extent */
1768 BUG_ON(ex
.fe_len
<= 0);
1769 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1770 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1773 ext4_mb_use_best_found(ac
, e4b
);
1775 ext4_unlock_group(ac
->ac_sb
, group
);
1776 ext4_mb_release_desc(e4b
);
1782 * The routine scans buddy structures (not bitmap!) from given order
1783 * to max order and tries to find big enough chunk to satisfy the req
1785 static void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1786 struct ext4_buddy
*e4b
)
1788 struct super_block
*sb
= ac
->ac_sb
;
1789 struct ext4_group_info
*grp
= e4b
->bd_info
;
1795 BUG_ON(ac
->ac_2order
<= 0);
1796 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1797 if (grp
->bb_counters
[i
] == 0)
1800 buddy
= mb_find_buddy(e4b
, i
, &max
);
1801 BUG_ON(buddy
== NULL
);
1803 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1808 ac
->ac_b_ex
.fe_len
= 1 << i
;
1809 ac
->ac_b_ex
.fe_start
= k
<< i
;
1810 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1812 ext4_mb_use_best_found(ac
, e4b
);
1814 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1816 if (EXT4_SB(sb
)->s_mb_stats
)
1817 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1824 * The routine scans the group and measures all found extents.
1825 * In order to optimize scanning, caller must pass number of
1826 * free blocks in the group, so the routine can know upper limit.
1828 static void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1829 struct ext4_buddy
*e4b
)
1831 struct super_block
*sb
= ac
->ac_sb
;
1832 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1833 struct ext4_free_extent ex
;
1837 free
= e4b
->bd_info
->bb_free
;
1840 i
= e4b
->bd_info
->bb_first_free
;
1842 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1843 i
= mb_find_next_zero_bit(bitmap
,
1844 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1845 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1847 * IF we have corrupt bitmap, we won't find any
1848 * free blocks even though group info says we
1849 * we have free blocks
1851 ext4_error(sb
, __FUNCTION__
, "%d free blocks as per "
1852 "group info. But bitmap says 0\n",
1857 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1858 BUG_ON(ex
.fe_len
<= 0);
1859 if (free
< ex
.fe_len
) {
1860 ext4_error(sb
, __FUNCTION__
, "%d free blocks as per "
1861 "group info. But got %d blocks\n",
1864 * The number of free blocks differs. This mostly
1865 * indicate that the bitmap is corrupt. So exit
1866 * without claiming the space.
1871 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1877 ext4_mb_check_limits(ac
, e4b
, 1);
1881 * This is a special case for storages like raid5
1882 * we try to find stripe-aligned chunks for stripe-size requests
1883 * XXX should do so at least for multiples of stripe size as well
1885 static void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1886 struct ext4_buddy
*e4b
)
1888 struct super_block
*sb
= ac
->ac_sb
;
1889 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1890 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1891 struct ext4_free_extent ex
;
1892 ext4_fsblk_t first_group_block
;
1897 BUG_ON(sbi
->s_stripe
== 0);
1899 /* find first stripe-aligned block in group */
1900 first_group_block
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
)
1901 + le32_to_cpu(sbi
->s_es
->s_first_data_block
);
1902 a
= first_group_block
+ sbi
->s_stripe
- 1;
1903 do_div(a
, sbi
->s_stripe
);
1904 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1906 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1907 if (!mb_test_bit(i
, bitmap
)) {
1908 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1909 if (max
>= sbi
->s_stripe
) {
1912 ext4_mb_use_best_found(ac
, e4b
);
1920 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1921 ext4_group_t group
, int cr
)
1923 unsigned free
, fragments
;
1925 struct ext4_group_desc
*desc
;
1926 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1928 BUG_ON(cr
< 0 || cr
>= 4);
1929 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp
));
1931 free
= grp
->bb_free
;
1932 fragments
= grp
->bb_fragments
;
1940 BUG_ON(ac
->ac_2order
== 0);
1941 /* If this group is uninitialized, skip it initially */
1942 desc
= ext4_get_group_desc(ac
->ac_sb
, group
, NULL
);
1943 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))
1946 bits
= ac
->ac_sb
->s_blocksize_bits
+ 1;
1947 for (i
= ac
->ac_2order
; i
<= bits
; i
++)
1948 if (grp
->bb_counters
[i
] > 0)
1952 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1956 if (free
>= ac
->ac_g_ex
.fe_len
)
1968 static int ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
1975 struct ext4_sb_info
*sbi
;
1976 struct super_block
*sb
;
1977 struct ext4_buddy e4b
;
1982 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1984 /* first, try the goal */
1985 err
= ext4_mb_find_by_goal(ac
, &e4b
);
1986 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
1989 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
1993 * ac->ac2_order is set only if the fe_len is a power of 2
1994 * if ac2_order is set we also set criteria to 0 so that we
1995 * try exact allocation using buddy.
1997 i
= fls(ac
->ac_g_ex
.fe_len
);
2000 * We search using buddy data only if the order of the request
2001 * is greater than equal to the sbi_s_mb_order2_reqs
2002 * You can tune it via /proc/fs/ext4/<partition>/order2_req
2004 if (i
>= sbi
->s_mb_order2_reqs
) {
2006 * This should tell if fe_len is exactly power of 2
2008 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
2009 ac
->ac_2order
= i
- 1;
2012 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
2013 /* if stream allocation is enabled, use global goal */
2014 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
2015 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
2019 if (size
< sbi
->s_mb_stream_request
&&
2020 (ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
2021 /* TBD: may be hot point */
2022 spin_lock(&sbi
->s_md_lock
);
2023 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
2024 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
2025 spin_unlock(&sbi
->s_md_lock
);
2028 /* searching for the right group start from the goal value specified */
2029 group
= ac
->ac_g_ex
.fe_group
;
2031 /* Let's just scan groups to find more-less suitable blocks */
2032 cr
= ac
->ac_2order
? 0 : 1;
2034 * cr == 0 try to get exact allocation,
2035 * cr == 3 try to get anything
2038 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2039 ac
->ac_criteria
= cr
;
2040 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; group
++, i
++) {
2041 struct ext4_group_info
*grp
;
2042 struct ext4_group_desc
*desc
;
2044 if (group
== EXT4_SB(sb
)->s_groups_count
)
2047 /* quick check to skip empty groups */
2048 grp
= ext4_get_group_info(ac
->ac_sb
, group
);
2049 if (grp
->bb_free
== 0)
2053 * if the group is already init we check whether it is
2054 * a good group and if not we don't load the buddy
2056 if (EXT4_MB_GRP_NEED_INIT(grp
)) {
2058 * we need full data about the group
2059 * to make a good selection
2061 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2064 ext4_mb_release_desc(&e4b
);
2068 * If the particular group doesn't satisfy our
2069 * criteria we continue with the next group
2071 if (!ext4_mb_good_group(ac
, group
, cr
))
2074 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2078 ext4_lock_group(sb
, group
);
2079 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2080 /* someone did allocation from this group */
2081 ext4_unlock_group(sb
, group
);
2082 ext4_mb_release_desc(&e4b
);
2086 ac
->ac_groups_scanned
++;
2087 desc
= ext4_get_group_desc(sb
, group
, NULL
);
2088 if (cr
== 0 || (desc
->bg_flags
&
2089 cpu_to_le16(EXT4_BG_BLOCK_UNINIT
) &&
2090 ac
->ac_2order
!= 0))
2091 ext4_mb_simple_scan_group(ac
, &e4b
);
2093 ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
)
2094 ext4_mb_scan_aligned(ac
, &e4b
);
2096 ext4_mb_complex_scan_group(ac
, &e4b
);
2098 ext4_unlock_group(sb
, group
);
2099 ext4_mb_release_desc(&e4b
);
2101 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2106 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2107 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2109 * We've been searching too long. Let's try to allocate
2110 * the best chunk we've found so far
2113 ext4_mb_try_best_found(ac
, &e4b
);
2114 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2116 * Someone more lucky has already allocated it.
2117 * The only thing we can do is just take first
2119 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2121 ac
->ac_b_ex
.fe_group
= 0;
2122 ac
->ac_b_ex
.fe_start
= 0;
2123 ac
->ac_b_ex
.fe_len
= 0;
2124 ac
->ac_status
= AC_STATUS_CONTINUE
;
2125 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2127 atomic_inc(&sbi
->s_mb_lost_chunks
);
2135 #ifdef EXT4_MB_HISTORY
2136 struct ext4_mb_proc_session
{
2137 struct ext4_mb_history
*history
;
2138 struct super_block
*sb
;
2143 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session
*s
,
2144 struct ext4_mb_history
*hs
,
2147 if (hs
== s
->history
+ s
->max
)
2149 if (!first
&& hs
== s
->history
+ s
->start
)
2151 while (hs
->orig
.fe_len
== 0) {
2153 if (hs
== s
->history
+ s
->max
)
2155 if (hs
== s
->history
+ s
->start
)
2161 static void *ext4_mb_seq_history_start(struct seq_file
*seq
, loff_t
*pos
)
2163 struct ext4_mb_proc_session
*s
= seq
->private;
2164 struct ext4_mb_history
*hs
;
2168 return SEQ_START_TOKEN
;
2169 hs
= ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
2172 while (--l
&& (hs
= ext4_mb_history_skip_empty(s
, ++hs
, 0)) != NULL
);
2176 static void *ext4_mb_seq_history_next(struct seq_file
*seq
, void *v
,
2179 struct ext4_mb_proc_session
*s
= seq
->private;
2180 struct ext4_mb_history
*hs
= v
;
2183 if (v
== SEQ_START_TOKEN
)
2184 return ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
2186 return ext4_mb_history_skip_empty(s
, ++hs
, 0);
2189 static int ext4_mb_seq_history_show(struct seq_file
*seq
, void *v
)
2191 char buf
[25], buf2
[25], buf3
[25], *fmt
;
2192 struct ext4_mb_history
*hs
= v
;
2194 if (v
== SEQ_START_TOKEN
) {
2195 seq_printf(seq
, "%-5s %-8s %-23s %-23s %-23s %-5s "
2196 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
2197 "pid", "inode", "original", "goal", "result", "found",
2198 "grps", "cr", "flags", "merge", "tail", "broken");
2202 if (hs
->op
== EXT4_MB_HISTORY_ALLOC
) {
2203 fmt
= "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2204 "%-5u %-5s %-5u %-6u\n";
2205 sprintf(buf2
, "%lu/%d/%u@%u", hs
->result
.fe_group
,
2206 hs
->result
.fe_start
, hs
->result
.fe_len
,
2207 hs
->result
.fe_logical
);
2208 sprintf(buf
, "%lu/%d/%u@%u", hs
->orig
.fe_group
,
2209 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
2210 hs
->orig
.fe_logical
);
2211 sprintf(buf3
, "%lu/%d/%u@%u", hs
->goal
.fe_group
,
2212 hs
->goal
.fe_start
, hs
->goal
.fe_len
,
2213 hs
->goal
.fe_logical
);
2214 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, buf3
, buf2
,
2215 hs
->found
, hs
->groups
, hs
->cr
, hs
->flags
,
2216 hs
->merged
? "M" : "", hs
->tail
,
2217 hs
->buddy
? 1 << hs
->buddy
: 0);
2218 } else if (hs
->op
== EXT4_MB_HISTORY_PREALLOC
) {
2219 fmt
= "%-5u %-8u %-23s %-23s %-23s\n";
2220 sprintf(buf2
, "%lu/%d/%u@%u", hs
->result
.fe_group
,
2221 hs
->result
.fe_start
, hs
->result
.fe_len
,
2222 hs
->result
.fe_logical
);
2223 sprintf(buf
, "%lu/%d/%u@%u", hs
->orig
.fe_group
,
2224 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
2225 hs
->orig
.fe_logical
);
2226 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, "", buf2
);
2227 } else if (hs
->op
== EXT4_MB_HISTORY_DISCARD
) {
2228 sprintf(buf2
, "%lu/%d/%u", hs
->result
.fe_group
,
2229 hs
->result
.fe_start
, hs
->result
.fe_len
);
2230 seq_printf(seq
, "%-5u %-8u %-23s discard\n",
2231 hs
->pid
, hs
->ino
, buf2
);
2232 } else if (hs
->op
== EXT4_MB_HISTORY_FREE
) {
2233 sprintf(buf2
, "%lu/%d/%u", hs
->result
.fe_group
,
2234 hs
->result
.fe_start
, hs
->result
.fe_len
);
2235 seq_printf(seq
, "%-5u %-8u %-23s free\n",
2236 hs
->pid
, hs
->ino
, buf2
);
2241 static void ext4_mb_seq_history_stop(struct seq_file
*seq
, void *v
)
2245 static struct seq_operations ext4_mb_seq_history_ops
= {
2246 .start
= ext4_mb_seq_history_start
,
2247 .next
= ext4_mb_seq_history_next
,
2248 .stop
= ext4_mb_seq_history_stop
,
2249 .show
= ext4_mb_seq_history_show
,
2252 static int ext4_mb_seq_history_open(struct inode
*inode
, struct file
*file
)
2254 struct super_block
*sb
= PDE(inode
)->data
;
2255 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2256 struct ext4_mb_proc_session
*s
;
2260 s
= kmalloc(sizeof(*s
), GFP_KERNEL
);
2264 size
= sizeof(struct ext4_mb_history
) * sbi
->s_mb_history_max
;
2265 s
->history
= kmalloc(size
, GFP_KERNEL
);
2266 if (s
->history
== NULL
) {
2271 spin_lock(&sbi
->s_mb_history_lock
);
2272 memcpy(s
->history
, sbi
->s_mb_history
, size
);
2273 s
->max
= sbi
->s_mb_history_max
;
2274 s
->start
= sbi
->s_mb_history_cur
% s
->max
;
2275 spin_unlock(&sbi
->s_mb_history_lock
);
2277 rc
= seq_open(file
, &ext4_mb_seq_history_ops
);
2279 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2289 static int ext4_mb_seq_history_release(struct inode
*inode
, struct file
*file
)
2291 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2292 struct ext4_mb_proc_session
*s
= seq
->private;
2295 return seq_release(inode
, file
);
2298 static ssize_t
ext4_mb_seq_history_write(struct file
*file
,
2299 const char __user
*buffer
,
2300 size_t count
, loff_t
*ppos
)
2302 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2303 struct ext4_mb_proc_session
*s
= seq
->private;
2304 struct super_block
*sb
= s
->sb
;
2308 if (count
>= sizeof(str
)) {
2309 printk(KERN_ERR
"EXT4-fs: %s string too long, max %u bytes\n",
2310 "mb_history", (int)sizeof(str
));
2314 if (copy_from_user(str
, buffer
, count
))
2317 value
= simple_strtol(str
, NULL
, 0);
2320 EXT4_SB(sb
)->s_mb_history_filter
= value
;
2325 static struct file_operations ext4_mb_seq_history_fops
= {
2326 .owner
= THIS_MODULE
,
2327 .open
= ext4_mb_seq_history_open
,
2329 .write
= ext4_mb_seq_history_write
,
2330 .llseek
= seq_lseek
,
2331 .release
= ext4_mb_seq_history_release
,
2334 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2336 struct super_block
*sb
= seq
->private;
2337 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2340 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2344 return (void *) group
;
2347 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2349 struct super_block
*sb
= seq
->private;
2350 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2354 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2357 return (void *) group
;;
2360 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2362 struct super_block
*sb
= seq
->private;
2363 long group
= (long) v
;
2366 struct ext4_buddy e4b
;
2368 struct ext4_group_info info
;
2369 unsigned short counters
[16];
2374 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2375 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2376 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2377 "group", "free", "frags", "first",
2378 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2379 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2381 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2382 sizeof(struct ext4_group_info
);
2383 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2385 seq_printf(seq
, "#%-5lu: I/O error\n", group
);
2388 ext4_lock_group(sb
, group
);
2389 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2390 ext4_unlock_group(sb
, group
);
2391 ext4_mb_release_desc(&e4b
);
2393 seq_printf(seq
, "#%-5lu: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2394 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2395 for (i
= 0; i
<= 13; i
++)
2396 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2397 sg
.info
.bb_counters
[i
] : 0);
2398 seq_printf(seq
, " ]\n");
2403 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2407 static struct seq_operations ext4_mb_seq_groups_ops
= {
2408 .start
= ext4_mb_seq_groups_start
,
2409 .next
= ext4_mb_seq_groups_next
,
2410 .stop
= ext4_mb_seq_groups_stop
,
2411 .show
= ext4_mb_seq_groups_show
,
2414 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2416 struct super_block
*sb
= PDE(inode
)->data
;
2419 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2421 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2428 static struct file_operations ext4_mb_seq_groups_fops
= {
2429 .owner
= THIS_MODULE
,
2430 .open
= ext4_mb_seq_groups_open
,
2432 .llseek
= seq_lseek
,
2433 .release
= seq_release
,
2436 static void ext4_mb_history_release(struct super_block
*sb
)
2438 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2440 remove_proc_entry("mb_groups", sbi
->s_mb_proc
);
2441 remove_proc_entry("mb_history", sbi
->s_mb_proc
);
2443 kfree(sbi
->s_mb_history
);
2446 static void ext4_mb_history_init(struct super_block
*sb
)
2448 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2451 if (sbi
->s_mb_proc
!= NULL
) {
2452 proc_create_data("mb_history", S_IRUGO
, sbi
->s_mb_proc
,
2453 &ext4_mb_seq_history_fops
, sb
);
2454 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_mb_proc
,
2455 &ext4_mb_seq_groups_fops
, sb
);
2458 sbi
->s_mb_history_max
= 1000;
2459 sbi
->s_mb_history_cur
= 0;
2460 spin_lock_init(&sbi
->s_mb_history_lock
);
2461 i
= sbi
->s_mb_history_max
* sizeof(struct ext4_mb_history
);
2462 sbi
->s_mb_history
= kmalloc(i
, GFP_KERNEL
);
2463 if (likely(sbi
->s_mb_history
!= NULL
))
2464 memset(sbi
->s_mb_history
, 0, i
);
2465 /* if we can't allocate history, then we simple won't use it */
2468 static void ext4_mb_store_history(struct ext4_allocation_context
*ac
)
2470 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2471 struct ext4_mb_history h
;
2473 if (unlikely(sbi
->s_mb_history
== NULL
))
2476 if (!(ac
->ac_op
& sbi
->s_mb_history_filter
))
2480 h
.pid
= current
->pid
;
2481 h
.ino
= ac
->ac_inode
? ac
->ac_inode
->i_ino
: 0;
2482 h
.orig
= ac
->ac_o_ex
;
2483 h
.result
= ac
->ac_b_ex
;
2484 h
.flags
= ac
->ac_flags
;
2485 h
.found
= ac
->ac_found
;
2486 h
.groups
= ac
->ac_groups_scanned
;
2487 h
.cr
= ac
->ac_criteria
;
2488 h
.tail
= ac
->ac_tail
;
2489 h
.buddy
= ac
->ac_buddy
;
2491 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
) {
2492 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
2493 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
2495 h
.goal
= ac
->ac_g_ex
;
2496 h
.result
= ac
->ac_f_ex
;
2499 spin_lock(&sbi
->s_mb_history_lock
);
2500 memcpy(sbi
->s_mb_history
+ sbi
->s_mb_history_cur
, &h
, sizeof(h
));
2501 if (++sbi
->s_mb_history_cur
>= sbi
->s_mb_history_max
)
2502 sbi
->s_mb_history_cur
= 0;
2503 spin_unlock(&sbi
->s_mb_history_lock
);
2507 #define ext4_mb_history_release(sb)
2508 #define ext4_mb_history_init(sb)
2511 static int ext4_mb_init_backend(struct super_block
*sb
)
2514 int j
, len
, metalen
;
2515 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2516 int num_meta_group_infos
=
2517 (sbi
->s_groups_count
+ EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2518 EXT4_DESC_PER_BLOCK_BITS(sb
);
2519 struct ext4_group_info
**meta_group_info
;
2521 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2522 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2523 * So a two level scheme suffices for now. */
2524 sbi
->s_group_info
= kmalloc(sizeof(*sbi
->s_group_info
) *
2525 num_meta_group_infos
, GFP_KERNEL
);
2526 if (sbi
->s_group_info
== NULL
) {
2527 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2530 sbi
->s_buddy_cache
= new_inode(sb
);
2531 if (sbi
->s_buddy_cache
== NULL
) {
2532 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2535 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2537 metalen
= sizeof(*meta_group_info
) << EXT4_DESC_PER_BLOCK_BITS(sb
);
2538 for (i
= 0; i
< num_meta_group_infos
; i
++) {
2539 if ((i
+ 1) == num_meta_group_infos
)
2540 metalen
= sizeof(*meta_group_info
) *
2541 (sbi
->s_groups_count
-
2542 (i
<< EXT4_DESC_PER_BLOCK_BITS(sb
)));
2543 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2544 if (meta_group_info
== NULL
) {
2545 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2549 sbi
->s_group_info
[i
] = meta_group_info
;
2553 * calculate needed size. if change bb_counters size,
2554 * don't forget about ext4_mb_generate_buddy()
2556 len
= sizeof(struct ext4_group_info
);
2557 len
+= sizeof(unsigned short) * (sb
->s_blocksize_bits
+ 2);
2558 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2559 struct ext4_group_desc
*desc
;
2562 sbi
->s_group_info
[i
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2563 j
= i
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2565 meta_group_info
[j
] = kzalloc(len
, GFP_KERNEL
);
2566 if (meta_group_info
[j
] == NULL
) {
2567 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2571 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2574 "EXT4-fs: can't read descriptor %lu\n", i
);
2577 memset(meta_group_info
[j
], 0, len
);
2578 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2579 &(meta_group_info
[j
]->bb_state
));
2582 * initialize bb_free to be able to skip
2583 * empty groups without initialization
2585 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2586 meta_group_info
[j
]->bb_free
=
2587 ext4_free_blocks_after_init(sb
, i
, desc
);
2589 meta_group_info
[j
]->bb_free
=
2590 le16_to_cpu(desc
->bg_free_blocks_count
);
2593 INIT_LIST_HEAD(&meta_group_info
[j
]->bb_prealloc_list
);
2597 struct buffer_head
*bh
;
2598 meta_group_info
[j
]->bb_bitmap
=
2599 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2600 BUG_ON(meta_group_info
[j
]->bb_bitmap
== NULL
);
2601 bh
= read_block_bitmap(sb
, i
);
2603 memcpy(meta_group_info
[j
]->bb_bitmap
, bh
->b_data
,
2615 kfree(ext4_get_group_info(sb
, i
));
2618 i
= num_meta_group_infos
;
2621 kfree(sbi
->s_group_info
[i
]);
2622 iput(sbi
->s_buddy_cache
);
2624 kfree(sbi
->s_group_info
);
2628 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2630 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2635 if (!test_opt(sb
, MBALLOC
))
2638 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned short);
2640 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2641 if (sbi
->s_mb_offsets
== NULL
) {
2642 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2645 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2646 if (sbi
->s_mb_maxs
== NULL
) {
2647 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2648 kfree(sbi
->s_mb_maxs
);
2652 /* order 0 is regular bitmap */
2653 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2654 sbi
->s_mb_offsets
[0] = 0;
2658 max
= sb
->s_blocksize
<< 2;
2660 sbi
->s_mb_offsets
[i
] = offset
;
2661 sbi
->s_mb_maxs
[i
] = max
;
2662 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2665 } while (i
<= sb
->s_blocksize_bits
+ 1);
2667 /* init file for buddy data */
2668 i
= ext4_mb_init_backend(sb
);
2670 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2671 kfree(sbi
->s_mb_offsets
);
2672 kfree(sbi
->s_mb_maxs
);
2676 spin_lock_init(&sbi
->s_md_lock
);
2677 INIT_LIST_HEAD(&sbi
->s_active_transaction
);
2678 INIT_LIST_HEAD(&sbi
->s_closed_transaction
);
2679 INIT_LIST_HEAD(&sbi
->s_committed_transaction
);
2680 spin_lock_init(&sbi
->s_bal_lock
);
2682 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2683 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2684 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2685 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2686 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2687 sbi
->s_mb_history_filter
= EXT4_MB_HISTORY_DEFAULT
;
2688 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2690 i
= sizeof(struct ext4_locality_group
) * NR_CPUS
;
2691 sbi
->s_locality_groups
= kmalloc(i
, GFP_KERNEL
);
2692 if (sbi
->s_locality_groups
== NULL
) {
2693 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2694 kfree(sbi
->s_mb_offsets
);
2695 kfree(sbi
->s_mb_maxs
);
2698 for (i
= 0; i
< NR_CPUS
; i
++) {
2699 struct ext4_locality_group
*lg
;
2700 lg
= &sbi
->s_locality_groups
[i
];
2701 mutex_init(&lg
->lg_mutex
);
2702 INIT_LIST_HEAD(&lg
->lg_prealloc_list
);
2703 spin_lock_init(&lg
->lg_prealloc_lock
);
2706 ext4_mb_init_per_dev_proc(sb
);
2707 ext4_mb_history_init(sb
);
2709 printk("EXT4-fs: mballoc enabled\n");
2713 /* need to called with ext4 group lock (ext4_lock_group) */
2714 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2716 struct ext4_prealloc_space
*pa
;
2717 struct list_head
*cur
, *tmp
;
2720 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2721 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2722 list_del(&pa
->pa_group_list
);
2727 mb_debug("mballoc: %u PAs left\n", count
);
2731 int ext4_mb_release(struct super_block
*sb
)
2734 int num_meta_group_infos
;
2735 struct ext4_group_info
*grinfo
;
2736 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2738 if (!test_opt(sb
, MBALLOC
))
2741 /* release freed, non-committed blocks */
2742 spin_lock(&sbi
->s_md_lock
);
2743 list_splice_init(&sbi
->s_closed_transaction
,
2744 &sbi
->s_committed_transaction
);
2745 list_splice_init(&sbi
->s_active_transaction
,
2746 &sbi
->s_committed_transaction
);
2747 spin_unlock(&sbi
->s_md_lock
);
2748 ext4_mb_free_committed_blocks(sb
);
2750 if (sbi
->s_group_info
) {
2751 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2752 grinfo
= ext4_get_group_info(sb
, i
);
2754 kfree(grinfo
->bb_bitmap
);
2756 ext4_lock_group(sb
, i
);
2757 ext4_mb_cleanup_pa(grinfo
);
2758 ext4_unlock_group(sb
, i
);
2761 num_meta_group_infos
= (sbi
->s_groups_count
+
2762 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2763 EXT4_DESC_PER_BLOCK_BITS(sb
);
2764 for (i
= 0; i
< num_meta_group_infos
; i
++)
2765 kfree(sbi
->s_group_info
[i
]);
2766 kfree(sbi
->s_group_info
);
2768 kfree(sbi
->s_mb_offsets
);
2769 kfree(sbi
->s_mb_maxs
);
2770 if (sbi
->s_buddy_cache
)
2771 iput(sbi
->s_buddy_cache
);
2772 if (sbi
->s_mb_stats
) {
2774 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2775 atomic_read(&sbi
->s_bal_allocated
),
2776 atomic_read(&sbi
->s_bal_reqs
),
2777 atomic_read(&sbi
->s_bal_success
));
2779 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2780 "%u 2^N hits, %u breaks, %u lost\n",
2781 atomic_read(&sbi
->s_bal_ex_scanned
),
2782 atomic_read(&sbi
->s_bal_goals
),
2783 atomic_read(&sbi
->s_bal_2orders
),
2784 atomic_read(&sbi
->s_bal_breaks
),
2785 atomic_read(&sbi
->s_mb_lost_chunks
));
2787 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2788 sbi
->s_mb_buddies_generated
++,
2789 sbi
->s_mb_generation_time
);
2791 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2792 atomic_read(&sbi
->s_mb_preallocated
),
2793 atomic_read(&sbi
->s_mb_discarded
));
2796 kfree(sbi
->s_locality_groups
);
2798 ext4_mb_history_release(sb
);
2799 ext4_mb_destroy_per_dev_proc(sb
);
2804 static void ext4_mb_free_committed_blocks(struct super_block
*sb
)
2806 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2811 struct ext4_free_metadata
*md
;
2812 struct ext4_buddy e4b
;
2814 if (list_empty(&sbi
->s_committed_transaction
))
2817 /* there is committed blocks to be freed yet */
2819 /* get next array of blocks */
2821 spin_lock(&sbi
->s_md_lock
);
2822 if (!list_empty(&sbi
->s_committed_transaction
)) {
2823 md
= list_entry(sbi
->s_committed_transaction
.next
,
2824 struct ext4_free_metadata
, list
);
2825 list_del(&md
->list
);
2827 spin_unlock(&sbi
->s_md_lock
);
2832 mb_debug("gonna free %u blocks in group %lu (0x%p):",
2833 md
->num
, md
->group
, md
);
2835 err
= ext4_mb_load_buddy(sb
, md
->group
, &e4b
);
2836 /* we expect to find existing buddy because it's pinned */
2839 /* there are blocks to put in buddy to make them really free */
2842 ext4_lock_group(sb
, md
->group
);
2843 for (i
= 0; i
< md
->num
; i
++) {
2844 mb_debug(" %u", md
->blocks
[i
]);
2845 err
= mb_free_blocks(NULL
, &e4b
, md
->blocks
[i
], 1);
2849 ext4_unlock_group(sb
, md
->group
);
2851 /* balance refcounts from ext4_mb_free_metadata() */
2852 page_cache_release(e4b
.bd_buddy_page
);
2853 page_cache_release(e4b
.bd_bitmap_page
);
2856 ext4_mb_release_desc(&e4b
);
2860 mb_debug("freed %u blocks in %u structures\n", count
, count2
);
2863 #define EXT4_MB_STATS_NAME "stats"
2864 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2865 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2866 #define EXT4_MB_ORDER2_REQ "order2_req"
2867 #define EXT4_MB_STREAM_REQ "stream_req"
2868 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2872 #define MB_PROC_VALUE_READ(name) \
2873 static int ext4_mb_read_##name(char *page, char **start, \
2874 off_t off, int count, int *eof, void *data) \
2876 struct ext4_sb_info *sbi = data; \
2881 len = sprintf(page, "%ld\n", sbi->s_mb_##name); \
2886 #define MB_PROC_VALUE_WRITE(name) \
2887 static int ext4_mb_write_##name(struct file *file, \
2888 const char __user *buf, unsigned long cnt, void *data) \
2890 struct ext4_sb_info *sbi = data; \
2893 if (cnt >= sizeof(str)) \
2895 if (copy_from_user(str, buf, cnt)) \
2897 value = simple_strtol(str, NULL, 0); \
2900 sbi->s_mb_##name = value; \
2904 MB_PROC_VALUE_READ(stats
);
2905 MB_PROC_VALUE_WRITE(stats
);
2906 MB_PROC_VALUE_READ(max_to_scan
);
2907 MB_PROC_VALUE_WRITE(max_to_scan
);
2908 MB_PROC_VALUE_READ(min_to_scan
);
2909 MB_PROC_VALUE_WRITE(min_to_scan
);
2910 MB_PROC_VALUE_READ(order2_reqs
);
2911 MB_PROC_VALUE_WRITE(order2_reqs
);
2912 MB_PROC_VALUE_READ(stream_request
);
2913 MB_PROC_VALUE_WRITE(stream_request
);
2914 MB_PROC_VALUE_READ(group_prealloc
);
2915 MB_PROC_VALUE_WRITE(group_prealloc
);
2917 #define MB_PROC_HANDLER(name, var) \
2919 proc = create_proc_entry(name, mode, sbi->s_mb_proc); \
2920 if (proc == NULL) { \
2921 printk(KERN_ERR "EXT4-fs: can't to create %s\n", name); \
2925 proc->read_proc = ext4_mb_read_##var ; \
2926 proc->write_proc = ext4_mb_write_##var; \
2929 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
)
2931 mode_t mode
= S_IFREG
| S_IRUGO
| S_IWUSR
;
2932 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2933 struct proc_dir_entry
*proc
;
2936 snprintf(devname
, sizeof(devname
) - 1, "%s",
2937 bdevname(sb
->s_bdev
, devname
));
2938 sbi
->s_mb_proc
= proc_mkdir(devname
, proc_root_ext4
);
2940 MB_PROC_HANDLER(EXT4_MB_STATS_NAME
, stats
);
2941 MB_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME
, max_to_scan
);
2942 MB_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME
, min_to_scan
);
2943 MB_PROC_HANDLER(EXT4_MB_ORDER2_REQ
, order2_reqs
);
2944 MB_PROC_HANDLER(EXT4_MB_STREAM_REQ
, stream_request
);
2945 MB_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC
, group_prealloc
);
2950 printk(KERN_ERR
"EXT4-fs: Unable to create %s\n", devname
);
2951 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_mb_proc
);
2952 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_mb_proc
);
2953 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_mb_proc
);
2954 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_mb_proc
);
2955 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_mb_proc
);
2956 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_mb_proc
);
2957 remove_proc_entry(devname
, proc_root_ext4
);
2958 sbi
->s_mb_proc
= NULL
;
2963 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
)
2965 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2968 if (sbi
->s_mb_proc
== NULL
)
2971 snprintf(devname
, sizeof(devname
) - 1, "%s",
2972 bdevname(sb
->s_bdev
, devname
));
2973 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_mb_proc
);
2974 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_mb_proc
);
2975 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_mb_proc
);
2976 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_mb_proc
);
2977 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_mb_proc
);
2978 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_mb_proc
);
2979 remove_proc_entry(devname
, proc_root_ext4
);
2984 int __init
init_ext4_mballoc(void)
2986 ext4_pspace_cachep
=
2987 kmem_cache_create("ext4_prealloc_space",
2988 sizeof(struct ext4_prealloc_space
),
2989 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2990 if (ext4_pspace_cachep
== NULL
)
2994 kmem_cache_create("ext4_alloc_context",
2995 sizeof(struct ext4_allocation_context
),
2996 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2997 if (ext4_ac_cachep
== NULL
) {
2998 kmem_cache_destroy(ext4_pspace_cachep
);
3001 #ifdef CONFIG_PROC_FS
3002 proc_root_ext4
= proc_mkdir("fs/ext4", NULL
);
3003 if (proc_root_ext4
== NULL
)
3004 printk(KERN_ERR
"EXT4-fs: Unable to create fs/ext4\n");
3009 void exit_ext4_mballoc(void)
3011 /* XXX: synchronize_rcu(); */
3012 kmem_cache_destroy(ext4_pspace_cachep
);
3013 kmem_cache_destroy(ext4_ac_cachep
);
3014 #ifdef CONFIG_PROC_FS
3015 remove_proc_entry("fs/ext4", NULL
);
3021 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
3022 * Returns 0 if success or error code
3024 static int ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
3027 struct buffer_head
*bitmap_bh
= NULL
;
3028 struct ext4_super_block
*es
;
3029 struct ext4_group_desc
*gdp
;
3030 struct buffer_head
*gdp_bh
;
3031 struct ext4_sb_info
*sbi
;
3032 struct super_block
*sb
;
3036 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3037 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
3043 ext4_debug("using block group %lu(%d)\n", ac
->ac_b_ex
.fe_group
,
3044 gdp
->bg_free_blocks_count
);
3047 bitmap_bh
= read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
3051 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
3056 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
3060 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
3064 block
= ac
->ac_b_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
3065 + ac
->ac_b_ex
.fe_start
3066 + le32_to_cpu(es
->s_first_data_block
);
3068 if (block
== ext4_block_bitmap(sb
, gdp
) ||
3069 block
== ext4_inode_bitmap(sb
, gdp
) ||
3070 in_range(block
, ext4_inode_table(sb
, gdp
),
3071 EXT4_SB(sb
)->s_itb_per_group
)) {
3073 ext4_error(sb
, __FUNCTION__
,
3074 "Allocating block in system zone - block = %llu",
3077 #ifdef AGGRESSIVE_CHECK
3080 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
3081 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
3082 bitmap_bh
->b_data
));
3086 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
), bitmap_bh
->b_data
,
3087 ac
->ac_b_ex
.fe_start
, ac
->ac_b_ex
.fe_len
);
3089 spin_lock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
3090 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
3091 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
3092 gdp
->bg_free_blocks_count
=
3093 cpu_to_le16(ext4_free_blocks_after_init(sb
,
3094 ac
->ac_b_ex
.fe_group
,
3097 gdp
->bg_free_blocks_count
=
3098 cpu_to_le16(le16_to_cpu(gdp
->bg_free_blocks_count
)
3099 - ac
->ac_b_ex
.fe_len
);
3100 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
3101 spin_unlock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
3102 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
3104 err
= ext4_journal_dirty_metadata(handle
, bitmap_bh
);
3107 err
= ext4_journal_dirty_metadata(handle
, gdp_bh
);
3116 * here we normalize request for locality group
3117 * Group request are normalized to s_strip size if we set the same via mount
3118 * option. If not we set it to s_mb_group_prealloc which can be configured via
3119 * /proc/fs/ext4/<partition>/group_prealloc
3121 * XXX: should we try to preallocate more than the group has now?
3123 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
3125 struct super_block
*sb
= ac
->ac_sb
;
3126 struct ext4_locality_group
*lg
= ac
->ac_lg
;
3129 if (EXT4_SB(sb
)->s_stripe
)
3130 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
3132 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
3133 mb_debug("#%u: goal %lu blocks for locality group\n",
3134 current
->pid
, ac
->ac_g_ex
.fe_len
);
3138 * Normalization means making request better in terms of
3139 * size and alignment
3141 static void ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
3142 struct ext4_allocation_request
*ar
)
3146 struct list_head
*cur
;
3147 loff_t size
, orig_size
, start_off
;
3148 ext4_lblk_t start
, orig_start
;
3149 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3151 /* do normalize only data requests, metadata requests
3152 do not need preallocation */
3153 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3156 /* sometime caller may want exact blocks */
3157 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3160 /* caller may indicate that preallocation isn't
3161 * required (it's a tail, for example) */
3162 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3165 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3166 ext4_mb_normalize_group_request(ac
);
3170 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3172 /* first, let's learn actual file size
3173 * given current request is allocated */
3174 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3175 size
= size
<< bsbits
;
3176 if (size
< i_size_read(ac
->ac_inode
))
3177 size
= i_size_read(ac
->ac_inode
);
3179 /* max available blocks in a free group */
3180 max
= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
) - 1 - 1 -
3181 EXT4_SB(ac
->ac_sb
)->s_itb_per_group
;
3183 #define NRL_CHECK_SIZE(req, size, max,bits) \
3184 (req <= (size) || max <= ((size) >> bits))
3186 /* first, try to predict filesize */
3187 /* XXX: should this table be tunable? */
3189 if (size
<= 16 * 1024) {
3191 } else if (size
<= 32 * 1024) {
3193 } else if (size
<= 64 * 1024) {
3195 } else if (size
<= 128 * 1024) {
3197 } else if (size
<= 256 * 1024) {
3199 } else if (size
<= 512 * 1024) {
3201 } else if (size
<= 1024 * 1024) {
3203 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, bsbits
)) {
3204 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3205 (20 - bsbits
)) << 20;
3207 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, bsbits
)) {
3208 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3209 (22 - bsbits
)) << 22;
3210 size
= 4 * 1024 * 1024;
3211 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3212 (8<<20)>>bsbits
, max
, bsbits
)) {
3213 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3214 (23 - bsbits
)) << 23;
3215 size
= 8 * 1024 * 1024;
3217 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3218 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3220 orig_size
= size
= size
>> bsbits
;
3221 orig_start
= start
= start_off
>> bsbits
;
3223 /* don't cover already allocated blocks in selected range */
3224 if (ar
->pleft
&& start
<= ar
->lleft
) {
3225 size
-= ar
->lleft
+ 1 - start
;
3226 start
= ar
->lleft
+ 1;
3228 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3229 size
-= start
+ size
- ar
->lright
;
3233 /* check we don't cross already preallocated blocks */
3235 list_for_each_rcu(cur
, &ei
->i_prealloc_list
) {
3236 struct ext4_prealloc_space
*pa
;
3237 unsigned long pa_end
;
3239 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_inode_list
);
3243 spin_lock(&pa
->pa_lock
);
3244 if (pa
->pa_deleted
) {
3245 spin_unlock(&pa
->pa_lock
);
3249 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3251 /* PA must not overlap original request */
3252 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3253 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3255 /* skip PA normalized request doesn't overlap with */
3256 if (pa
->pa_lstart
>= end
) {
3257 spin_unlock(&pa
->pa_lock
);
3260 if (pa_end
<= start
) {
3261 spin_unlock(&pa
->pa_lock
);
3264 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3266 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3267 BUG_ON(pa_end
< start
);
3271 if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3272 BUG_ON(pa
->pa_lstart
> end
);
3273 end
= pa
->pa_lstart
;
3275 spin_unlock(&pa
->pa_lock
);
3280 /* XXX: extra loop to check we really don't overlap preallocations */
3282 list_for_each_rcu(cur
, &ei
->i_prealloc_list
) {
3283 struct ext4_prealloc_space
*pa
;
3284 unsigned long pa_end
;
3285 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_inode_list
);
3286 spin_lock(&pa
->pa_lock
);
3287 if (pa
->pa_deleted
== 0) {
3288 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3289 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3291 spin_unlock(&pa
->pa_lock
);
3295 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3296 start
> ac
->ac_o_ex
.fe_logical
) {
3297 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3298 (unsigned long) start
, (unsigned long) size
,
3299 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3301 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3302 start
> ac
->ac_o_ex
.fe_logical
);
3303 BUG_ON(size
<= 0 || size
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3305 /* now prepare goal request */
3307 /* XXX: is it better to align blocks WRT to logical
3308 * placement or satisfy big request as is */
3309 ac
->ac_g_ex
.fe_logical
= start
;
3310 ac
->ac_g_ex
.fe_len
= size
;
3312 /* define goal start in order to merge */
3313 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3314 /* merge to the right */
3315 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3316 &ac
->ac_f_ex
.fe_group
,
3317 &ac
->ac_f_ex
.fe_start
);
3318 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3320 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3321 /* merge to the left */
3322 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3323 &ac
->ac_f_ex
.fe_group
,
3324 &ac
->ac_f_ex
.fe_start
);
3325 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3328 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3329 (unsigned) orig_size
, (unsigned) start
);
3332 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3334 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3336 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3337 atomic_inc(&sbi
->s_bal_reqs
);
3338 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3339 if (ac
->ac_o_ex
.fe_len
>= ac
->ac_g_ex
.fe_len
)
3340 atomic_inc(&sbi
->s_bal_success
);
3341 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3342 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3343 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3344 atomic_inc(&sbi
->s_bal_goals
);
3345 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3346 atomic_inc(&sbi
->s_bal_breaks
);
3349 ext4_mb_store_history(ac
);
3353 * use blocks preallocated to inode
3355 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3356 struct ext4_prealloc_space
*pa
)
3362 /* found preallocated blocks, use them */
3363 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3364 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3366 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3367 &ac
->ac_b_ex
.fe_start
);
3368 ac
->ac_b_ex
.fe_len
= len
;
3369 ac
->ac_status
= AC_STATUS_FOUND
;
3372 BUG_ON(start
< pa
->pa_pstart
);
3373 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3374 BUG_ON(pa
->pa_free
< len
);
3377 mb_debug("use %llu/%lu from inode pa %p\n", start
, len
, pa
);
3381 * use blocks preallocated to locality group
3383 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3384 struct ext4_prealloc_space
*pa
)
3386 unsigned len
= ac
->ac_o_ex
.fe_len
;
3388 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3389 &ac
->ac_b_ex
.fe_group
,
3390 &ac
->ac_b_ex
.fe_start
);
3391 ac
->ac_b_ex
.fe_len
= len
;
3392 ac
->ac_status
= AC_STATUS_FOUND
;
3395 /* we don't correct pa_pstart or pa_plen here to avoid
3396 * possible race when the group is being loaded concurrently
3397 * instead we correct pa later, after blocks are marked
3398 * in on-disk bitmap -- see ext4_mb_release_context()
3399 * Other CPUs are prevented from allocating from this pa by lg_mutex
3401 mb_debug("use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3405 * search goal blocks in preallocated space
3407 static int ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3409 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3410 struct ext4_locality_group
*lg
;
3411 struct ext4_prealloc_space
*pa
;
3412 struct list_head
*cur
;
3414 /* only data can be preallocated */
3415 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3418 /* first, try per-file preallocation */
3420 list_for_each_rcu(cur
, &ei
->i_prealloc_list
) {
3421 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_inode_list
);
3423 /* all fields in this condition don't change,
3424 * so we can skip locking for them */
3425 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3426 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3429 /* found preallocated blocks, use them */
3430 spin_lock(&pa
->pa_lock
);
3431 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3432 atomic_inc(&pa
->pa_count
);
3433 ext4_mb_use_inode_pa(ac
, pa
);
3434 spin_unlock(&pa
->pa_lock
);
3435 ac
->ac_criteria
= 10;
3439 spin_unlock(&pa
->pa_lock
);
3443 /* can we use group allocation? */
3444 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3447 /* inode may have no locality group for some reason */
3453 list_for_each_rcu(cur
, &lg
->lg_prealloc_list
) {
3454 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_inode_list
);
3455 spin_lock(&pa
->pa_lock
);
3456 if (pa
->pa_deleted
== 0 && pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3457 atomic_inc(&pa
->pa_count
);
3458 ext4_mb_use_group_pa(ac
, pa
);
3459 spin_unlock(&pa
->pa_lock
);
3460 ac
->ac_criteria
= 20;
3464 spin_unlock(&pa
->pa_lock
);
3472 * the function goes through all preallocation in this group and marks them
3473 * used in in-core bitmap. buddy must be generated from this bitmap
3474 * Need to be called with ext4 group lock (ext4_lock_group)
3476 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3479 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3480 struct ext4_prealloc_space
*pa
;
3481 struct list_head
*cur
;
3482 ext4_group_t groupnr
;
3483 ext4_grpblk_t start
;
3484 int preallocated
= 0;
3488 /* all form of preallocation discards first load group,
3489 * so the only competing code is preallocation use.
3490 * we don't need any locking here
3491 * notice we do NOT ignore preallocations with pa_deleted
3492 * otherwise we could leave used blocks available for
3493 * allocation in buddy when concurrent ext4_mb_put_pa()
3494 * is dropping preallocation
3496 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3497 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3498 spin_lock(&pa
->pa_lock
);
3499 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3502 spin_unlock(&pa
->pa_lock
);
3503 if (unlikely(len
== 0))
3505 BUG_ON(groupnr
!= group
);
3506 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3507 bitmap
, start
, len
);
3508 preallocated
+= len
;
3511 mb_debug("prellocated %u for group %lu\n", preallocated
, group
);
3514 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3516 struct ext4_prealloc_space
*pa
;
3517 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3518 kmem_cache_free(ext4_pspace_cachep
, pa
);
3522 * drops a reference to preallocated space descriptor
3523 * if this was the last reference and the space is consumed
3525 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3526 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3530 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3533 /* in this short window concurrent discard can set pa_deleted */
3534 spin_lock(&pa
->pa_lock
);
3535 if (pa
->pa_deleted
== 1) {
3536 spin_unlock(&pa
->pa_lock
);
3541 spin_unlock(&pa
->pa_lock
);
3543 /* -1 is to protect from crossing allocation group */
3544 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
- 1, &grp
, NULL
);
3549 * P1 (buddy init) P2 (regular allocation)
3550 * find block B in PA
3551 * copy on-disk bitmap to buddy
3552 * mark B in on-disk bitmap
3553 * drop PA from group
3554 * mark all PAs in buddy
3556 * thus, P1 initializes buddy with B available. to prevent this
3557 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3560 ext4_lock_group(sb
, grp
);
3561 list_del(&pa
->pa_group_list
);
3562 ext4_unlock_group(sb
, grp
);
3564 spin_lock(pa
->pa_obj_lock
);
3565 list_del_rcu(&pa
->pa_inode_list
);
3566 spin_unlock(pa
->pa_obj_lock
);
3568 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3572 * creates new preallocated space for given inode
3574 static int ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3576 struct super_block
*sb
= ac
->ac_sb
;
3577 struct ext4_prealloc_space
*pa
;
3578 struct ext4_group_info
*grp
;
3579 struct ext4_inode_info
*ei
;
3581 /* preallocate only when found space is larger then requested */
3582 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3583 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3584 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3586 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3590 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3596 /* we can't allocate as much as normalizer wants.
3597 * so, found space must get proper lstart
3598 * to cover original request */
3599 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3600 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3602 /* we're limited by original request in that
3603 * logical block must be covered any way
3604 * winl is window we can move our chunk within */
3605 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3607 /* also, we should cover whole original request */
3608 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3610 /* the smallest one defines real window */
3611 win
= min(winl
, wins
);
3613 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3614 if (offs
&& offs
< win
)
3617 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3618 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3619 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3622 /* preallocation can change ac_b_ex, thus we store actually
3623 * allocated blocks for history */
3624 ac
->ac_f_ex
= ac
->ac_b_ex
;
3626 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3627 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3628 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3629 pa
->pa_free
= pa
->pa_len
;
3630 atomic_set(&pa
->pa_count
, 1);
3631 spin_lock_init(&pa
->pa_lock
);
3635 mb_debug("new inode pa %p: %llu/%u for %u\n", pa
,
3636 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3638 ext4_mb_use_inode_pa(ac
, pa
);
3639 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3641 ei
= EXT4_I(ac
->ac_inode
);
3642 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3644 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3645 pa
->pa_inode
= ac
->ac_inode
;
3647 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3648 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3649 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3651 spin_lock(pa
->pa_obj_lock
);
3652 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3653 spin_unlock(pa
->pa_obj_lock
);
3659 * creates new preallocated space for locality group inodes belongs to
3661 static int ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3663 struct super_block
*sb
= ac
->ac_sb
;
3664 struct ext4_locality_group
*lg
;
3665 struct ext4_prealloc_space
*pa
;
3666 struct ext4_group_info
*grp
;
3668 /* preallocate only when found space is larger then requested */
3669 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3670 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3671 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3673 BUG_ON(ext4_pspace_cachep
== NULL
);
3674 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3678 /* preallocation can change ac_b_ex, thus we store actually
3679 * allocated blocks for history */
3680 ac
->ac_f_ex
= ac
->ac_b_ex
;
3682 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3683 pa
->pa_lstart
= pa
->pa_pstart
;
3684 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3685 pa
->pa_free
= pa
->pa_len
;
3686 atomic_set(&pa
->pa_count
, 1);
3687 spin_lock_init(&pa
->pa_lock
);
3691 mb_debug("new group pa %p: %llu/%u for %u\n", pa
,
3692 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3694 ext4_mb_use_group_pa(ac
, pa
);
3695 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3697 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3701 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3702 pa
->pa_inode
= NULL
;
3704 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3705 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3706 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3708 spin_lock(pa
->pa_obj_lock
);
3709 list_add_tail_rcu(&pa
->pa_inode_list
, &lg
->lg_prealloc_list
);
3710 spin_unlock(pa
->pa_obj_lock
);
3715 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3719 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3720 err
= ext4_mb_new_group_pa(ac
);
3722 err
= ext4_mb_new_inode_pa(ac
);
3727 * finds all unused blocks in on-disk bitmap, frees them in
3728 * in-core bitmap and buddy.
3729 * @pa must be unlinked from inode and group lists, so that
3730 * nobody else can find/use it.
3731 * the caller MUST hold group/inode locks.
3732 * TODO: optimize the case when there are no in-core structures yet
3734 static int ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
,
3735 struct buffer_head
*bitmap_bh
,
3736 struct ext4_prealloc_space
*pa
)
3738 struct ext4_allocation_context
*ac
;
3739 struct super_block
*sb
= e4b
->bd_sb
;
3740 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3749 BUG_ON(pa
->pa_deleted
== 0);
3750 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3751 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3752 end
= bit
+ pa
->pa_len
;
3754 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3758 ac
->ac_inode
= pa
->pa_inode
;
3759 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3763 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3766 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3769 start
= group
* EXT4_BLOCKS_PER_GROUP(sb
) + bit
+
3770 le32_to_cpu(sbi
->s_es
->s_first_data_block
);
3771 mb_debug(" free preallocated %u/%u in group %u\n",
3772 (unsigned) start
, (unsigned) next
- bit
,
3777 ac
->ac_b_ex
.fe_group
= group
;
3778 ac
->ac_b_ex
.fe_start
= bit
;
3779 ac
->ac_b_ex
.fe_len
= next
- bit
;
3780 ac
->ac_b_ex
.fe_logical
= 0;
3781 ext4_mb_store_history(ac
);
3784 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3787 if (free
!= pa
->pa_free
) {
3788 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3789 pa
, (unsigned long) pa
->pa_lstart
,
3790 (unsigned long) pa
->pa_pstart
,
3791 (unsigned long) pa
->pa_len
);
3792 ext4_error(sb
, __FUNCTION__
, "free %u, pa_free %u\n",
3795 * pa is already deleted so we use the value obtained
3796 * from the bitmap and continue.
3799 atomic_add(free
, &sbi
->s_mb_discarded
);
3801 kmem_cache_free(ext4_ac_cachep
, ac
);
3806 static int ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3807 struct ext4_prealloc_space
*pa
)
3809 struct ext4_allocation_context
*ac
;
3810 struct super_block
*sb
= e4b
->bd_sb
;
3814 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3817 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3819 BUG_ON(pa
->pa_deleted
== 0);
3820 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3821 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3822 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3823 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3827 ac
->ac_inode
= NULL
;
3828 ac
->ac_b_ex
.fe_group
= group
;
3829 ac
->ac_b_ex
.fe_start
= bit
;
3830 ac
->ac_b_ex
.fe_len
= pa
->pa_len
;
3831 ac
->ac_b_ex
.fe_logical
= 0;
3832 ext4_mb_store_history(ac
);
3833 kmem_cache_free(ext4_ac_cachep
, ac
);
3840 * releases all preallocations in given group
3842 * first, we need to decide discard policy:
3843 * - when do we discard
3845 * - how many do we discard
3846 * 1) how many requested
3848 static int ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3849 ext4_group_t group
, int needed
)
3851 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3852 struct buffer_head
*bitmap_bh
= NULL
;
3853 struct ext4_prealloc_space
*pa
, *tmp
;
3854 struct list_head list
;
3855 struct ext4_buddy e4b
;
3860 mb_debug("discard preallocation for group %lu\n", group
);
3862 if (list_empty(&grp
->bb_prealloc_list
))
3865 bitmap_bh
= read_block_bitmap(sb
, group
);
3866 if (bitmap_bh
== NULL
) {
3867 /* error handling here */
3868 ext4_mb_release_desc(&e4b
);
3869 BUG_ON(bitmap_bh
== NULL
);
3872 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3873 BUG_ON(err
!= 0); /* error handling here */
3876 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3878 grp
= ext4_get_group_info(sb
, group
);
3879 INIT_LIST_HEAD(&list
);
3882 ext4_lock_group(sb
, group
);
3883 list_for_each_entry_safe(pa
, tmp
,
3884 &grp
->bb_prealloc_list
, pa_group_list
) {
3885 spin_lock(&pa
->pa_lock
);
3886 if (atomic_read(&pa
->pa_count
)) {
3887 spin_unlock(&pa
->pa_lock
);
3891 if (pa
->pa_deleted
) {
3892 spin_unlock(&pa
->pa_lock
);
3896 /* seems this one can be freed ... */
3899 /* we can trust pa_free ... */
3900 free
+= pa
->pa_free
;
3902 spin_unlock(&pa
->pa_lock
);
3904 list_del(&pa
->pa_group_list
);
3905 list_add(&pa
->u
.pa_tmp_list
, &list
);
3908 /* if we still need more blocks and some PAs were used, try again */
3909 if (free
< needed
&& busy
) {
3911 ext4_unlock_group(sb
, group
);
3913 * Yield the CPU here so that we don't get soft lockup
3914 * in non preempt case.
3920 /* found anything to free? */
3921 if (list_empty(&list
)) {
3926 /* now free all selected PAs */
3927 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3929 /* remove from object (inode or locality group) */
3930 spin_lock(pa
->pa_obj_lock
);
3931 list_del_rcu(&pa
->pa_inode_list
);
3932 spin_unlock(pa
->pa_obj_lock
);
3935 ext4_mb_release_group_pa(&e4b
, pa
);
3937 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
3939 list_del(&pa
->u
.pa_tmp_list
);
3940 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3944 ext4_unlock_group(sb
, group
);
3945 ext4_mb_release_desc(&e4b
);
3951 * releases all non-used preallocated blocks for given inode
3953 * It's important to discard preallocations under i_data_sem
3954 * We don't want another block to be served from the prealloc
3955 * space when we are discarding the inode prealloc space.
3957 * FIXME!! Make sure it is valid at all the call sites
3959 void ext4_mb_discard_inode_preallocations(struct inode
*inode
)
3961 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3962 struct super_block
*sb
= inode
->i_sb
;
3963 struct buffer_head
*bitmap_bh
= NULL
;
3964 struct ext4_prealloc_space
*pa
, *tmp
;
3965 ext4_group_t group
= 0;
3966 struct list_head list
;
3967 struct ext4_buddy e4b
;
3970 if (!test_opt(sb
, MBALLOC
) || !S_ISREG(inode
->i_mode
)) {
3971 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3975 mb_debug("discard preallocation for inode %lu\n", inode
->i_ino
);
3977 INIT_LIST_HEAD(&list
);
3980 /* first, collect all pa's in the inode */
3981 spin_lock(&ei
->i_prealloc_lock
);
3982 while (!list_empty(&ei
->i_prealloc_list
)) {
3983 pa
= list_entry(ei
->i_prealloc_list
.next
,
3984 struct ext4_prealloc_space
, pa_inode_list
);
3985 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3986 spin_lock(&pa
->pa_lock
);
3987 if (atomic_read(&pa
->pa_count
)) {
3988 /* this shouldn't happen often - nobody should
3989 * use preallocation while we're discarding it */
3990 spin_unlock(&pa
->pa_lock
);
3991 spin_unlock(&ei
->i_prealloc_lock
);
3992 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
3994 schedule_timeout_uninterruptible(HZ
);
3998 if (pa
->pa_deleted
== 0) {
4000 spin_unlock(&pa
->pa_lock
);
4001 list_del_rcu(&pa
->pa_inode_list
);
4002 list_add(&pa
->u
.pa_tmp_list
, &list
);
4006 /* someone is deleting pa right now */
4007 spin_unlock(&pa
->pa_lock
);
4008 spin_unlock(&ei
->i_prealloc_lock
);
4010 /* we have to wait here because pa_deleted
4011 * doesn't mean pa is already unlinked from
4012 * the list. as we might be called from
4013 * ->clear_inode() the inode will get freed
4014 * and concurrent thread which is unlinking
4015 * pa from inode's list may access already
4016 * freed memory, bad-bad-bad */
4018 /* XXX: if this happens too often, we can
4019 * add a flag to force wait only in case
4020 * of ->clear_inode(), but not in case of
4021 * regular truncate */
4022 schedule_timeout_uninterruptible(HZ
);
4025 spin_unlock(&ei
->i_prealloc_lock
);
4027 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
4028 BUG_ON(pa
->pa_linear
!= 0);
4029 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4031 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
4032 BUG_ON(err
!= 0); /* error handling here */
4034 bitmap_bh
= read_block_bitmap(sb
, group
);
4035 if (bitmap_bh
== NULL
) {
4036 /* error handling here */
4037 ext4_mb_release_desc(&e4b
);
4038 BUG_ON(bitmap_bh
== NULL
);
4041 ext4_lock_group(sb
, group
);
4042 list_del(&pa
->pa_group_list
);
4043 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
);
4044 ext4_unlock_group(sb
, group
);
4046 ext4_mb_release_desc(&e4b
);
4049 list_del(&pa
->u
.pa_tmp_list
);
4050 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4055 * finds all preallocated spaces and return blocks being freed to them
4056 * if preallocated space becomes full (no block is used from the space)
4057 * then the function frees space in buddy
4058 * XXX: at the moment, truncate (which is the only way to free blocks)
4059 * discards all preallocations
4061 static void ext4_mb_return_to_preallocation(struct inode
*inode
,
4062 struct ext4_buddy
*e4b
,
4063 sector_t block
, int count
)
4065 BUG_ON(!list_empty(&EXT4_I(inode
)->i_prealloc_list
));
4068 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4070 struct super_block
*sb
= ac
->ac_sb
;
4073 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
4074 " Allocation context details:\n");
4075 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
4076 ac
->ac_status
, ac
->ac_flags
);
4077 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4078 "best %lu/%lu/%lu@%lu cr %d\n",
4079 (unsigned long)ac
->ac_o_ex
.fe_group
,
4080 (unsigned long)ac
->ac_o_ex
.fe_start
,
4081 (unsigned long)ac
->ac_o_ex
.fe_len
,
4082 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4083 (unsigned long)ac
->ac_g_ex
.fe_group
,
4084 (unsigned long)ac
->ac_g_ex
.fe_start
,
4085 (unsigned long)ac
->ac_g_ex
.fe_len
,
4086 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4087 (unsigned long)ac
->ac_b_ex
.fe_group
,
4088 (unsigned long)ac
->ac_b_ex
.fe_start
,
4089 (unsigned long)ac
->ac_b_ex
.fe_len
,
4090 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4091 (int)ac
->ac_criteria
);
4092 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
4094 printk(KERN_ERR
"EXT4-fs: groups: \n");
4095 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; i
++) {
4096 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4097 struct ext4_prealloc_space
*pa
;
4098 ext4_grpblk_t start
;
4099 struct list_head
*cur
;
4100 ext4_lock_group(sb
, i
);
4101 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4102 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4104 spin_lock(&pa
->pa_lock
);
4105 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4107 spin_unlock(&pa
->pa_lock
);
4108 printk(KERN_ERR
"PA:%lu:%d:%u \n", i
,
4111 ext4_lock_group(sb
, i
);
4113 if (grp
->bb_free
== 0)
4115 printk(KERN_ERR
"%lu: %d/%d \n",
4116 i
, grp
->bb_free
, grp
->bb_fragments
);
4118 printk(KERN_ERR
"\n");
4121 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4128 * We use locality group preallocation for small size file. The size of the
4129 * file is determined by the current size or the resulting size after
4130 * allocation which ever is larger
4132 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
4134 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4136 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4137 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4140 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4143 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
4144 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
4145 size
= max(size
, isize
);
4147 /* don't use group allocation for large files */
4148 if (size
>= sbi
->s_mb_stream_request
)
4151 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4154 BUG_ON(ac
->ac_lg
!= NULL
);
4156 * locality group prealloc space are per cpu. The reason for having
4157 * per cpu locality group is to reduce the contention between block
4158 * request from multiple CPUs.
4160 ac
->ac_lg
= &sbi
->s_locality_groups
[get_cpu()];
4163 /* we're going to use group allocation */
4164 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4166 /* serialize all allocations in the group */
4167 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4170 static int ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4171 struct ext4_allocation_request
*ar
)
4173 struct super_block
*sb
= ar
->inode
->i_sb
;
4174 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4175 struct ext4_super_block
*es
= sbi
->s_es
;
4179 ext4_grpblk_t block
;
4181 /* we can't allocate > group size */
4184 /* just a dirty hack to filter too big requests */
4185 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4186 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4188 /* start searching from the goal */
4190 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4191 goal
>= ext4_blocks_count(es
))
4192 goal
= le32_to_cpu(es
->s_first_data_block
);
4193 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4195 /* set up allocation goals */
4196 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4197 ac
->ac_b_ex
.fe_group
= 0;
4198 ac
->ac_b_ex
.fe_start
= 0;
4199 ac
->ac_b_ex
.fe_len
= 0;
4200 ac
->ac_status
= AC_STATUS_CONTINUE
;
4201 ac
->ac_groups_scanned
= 0;
4202 ac
->ac_ex_scanned
= 0;
4205 ac
->ac_inode
= ar
->inode
;
4206 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4207 ac
->ac_o_ex
.fe_group
= group
;
4208 ac
->ac_o_ex
.fe_start
= block
;
4209 ac
->ac_o_ex
.fe_len
= len
;
4210 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4211 ac
->ac_g_ex
.fe_group
= group
;
4212 ac
->ac_g_ex
.fe_start
= block
;
4213 ac
->ac_g_ex
.fe_len
= len
;
4214 ac
->ac_f_ex
.fe_len
= 0;
4215 ac
->ac_flags
= ar
->flags
;
4217 ac
->ac_criteria
= 0;
4219 ac
->ac_bitmap_page
= NULL
;
4220 ac
->ac_buddy_page
= NULL
;
4223 /* we have to define context: we'll we work with a file or
4224 * locality group. this is a policy, actually */
4225 ext4_mb_group_or_file(ac
);
4227 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4228 "left: %u/%u, right %u/%u to %swritable\n",
4229 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4230 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4231 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4232 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4233 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4239 * release all resource we used in allocation
4241 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4244 if (ac
->ac_pa
->pa_linear
) {
4245 /* see comment in ext4_mb_use_group_pa() */
4246 spin_lock(&ac
->ac_pa
->pa_lock
);
4247 ac
->ac_pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4248 ac
->ac_pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4249 ac
->ac_pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4250 ac
->ac_pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4251 spin_unlock(&ac
->ac_pa
->pa_lock
);
4253 ext4_mb_put_pa(ac
, ac
->ac_sb
, ac
->ac_pa
);
4255 if (ac
->ac_bitmap_page
)
4256 page_cache_release(ac
->ac_bitmap_page
);
4257 if (ac
->ac_buddy_page
)
4258 page_cache_release(ac
->ac_buddy_page
);
4259 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4260 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4261 ext4_mb_collect_stats(ac
);
4265 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4271 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
&& needed
> 0; i
++) {
4272 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4281 * Main entry point into mballoc to allocate blocks
4282 * it tries to use preallocation first, then falls back
4283 * to usual allocation
4285 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4286 struct ext4_allocation_request
*ar
, int *errp
)
4288 struct ext4_allocation_context
*ac
= NULL
;
4289 struct ext4_sb_info
*sbi
;
4290 struct super_block
*sb
;
4291 ext4_fsblk_t block
= 0;
4295 sb
= ar
->inode
->i_sb
;
4298 if (!test_opt(sb
, MBALLOC
)) {
4299 block
= ext4_new_blocks_old(handle
, ar
->inode
, ar
->goal
,
4304 while (ar
->len
&& DQUOT_ALLOC_BLOCK(ar
->inode
, ar
->len
)) {
4305 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4314 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4320 ext4_mb_poll_new_transaction(sb
, handle
);
4322 *errp
= ext4_mb_initialize_context(ac
, ar
);
4328 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4329 if (!ext4_mb_use_preallocated(ac
)) {
4331 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4332 ext4_mb_normalize_request(ac
, ar
);
4335 /* allocate space in core */
4336 ext4_mb_regular_allocator(ac
);
4338 /* as we've just preallocated more space than
4339 * user requested orinally, we store allocated
4340 * space in a special descriptor */
4341 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4342 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4343 ext4_mb_new_preallocation(ac
);
4346 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4347 ext4_mb_mark_diskspace_used(ac
, handle
);
4349 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4350 ar
->len
= ac
->ac_b_ex
.fe_len
;
4352 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4356 ac
->ac_b_ex
.fe_len
= 0;
4358 ext4_mb_show_ac(ac
);
4361 ext4_mb_release_context(ac
);
4364 if (ar
->len
< inquota
)
4365 DQUOT_FREE_BLOCK(ar
->inode
, inquota
- ar
->len
);
4367 kmem_cache_free(ext4_ac_cachep
, ac
);
4370 static void ext4_mb_poll_new_transaction(struct super_block
*sb
,
4373 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4375 if (sbi
->s_last_transaction
== handle
->h_transaction
->t_tid
)
4378 /* new transaction! time to close last one and free blocks for
4379 * committed transaction. we know that only transaction can be
4380 * active, so previos transaction can be being logged and we
4381 * know that transaction before previous is known to be already
4382 * logged. this means that now we may free blocks freed in all
4383 * transactions before previous one. hope I'm clear enough ... */
4385 spin_lock(&sbi
->s_md_lock
);
4386 if (sbi
->s_last_transaction
!= handle
->h_transaction
->t_tid
) {
4387 mb_debug("new transaction %lu, old %lu\n",
4388 (unsigned long) handle
->h_transaction
->t_tid
,
4389 (unsigned long) sbi
->s_last_transaction
);
4390 list_splice_init(&sbi
->s_closed_transaction
,
4391 &sbi
->s_committed_transaction
);
4392 list_splice_init(&sbi
->s_active_transaction
,
4393 &sbi
->s_closed_transaction
);
4394 sbi
->s_last_transaction
= handle
->h_transaction
->t_tid
;
4396 spin_unlock(&sbi
->s_md_lock
);
4398 ext4_mb_free_committed_blocks(sb
);
4401 static int ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4402 ext4_group_t group
, ext4_grpblk_t block
, int count
)
4404 struct ext4_group_info
*db
= e4b
->bd_info
;
4405 struct super_block
*sb
= e4b
->bd_sb
;
4406 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4407 struct ext4_free_metadata
*md
;
4410 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4411 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4413 ext4_lock_group(sb
, group
);
4414 for (i
= 0; i
< count
; i
++) {
4416 if (md
&& db
->bb_tid
!= handle
->h_transaction
->t_tid
) {
4417 db
->bb_md_cur
= NULL
;
4422 ext4_unlock_group(sb
, group
);
4423 md
= kmalloc(sizeof(*md
), GFP_NOFS
);
4429 ext4_lock_group(sb
, group
);
4430 if (db
->bb_md_cur
== NULL
) {
4431 spin_lock(&sbi
->s_md_lock
);
4432 list_add(&md
->list
, &sbi
->s_active_transaction
);
4433 spin_unlock(&sbi
->s_md_lock
);
4434 /* protect buddy cache from being freed,
4435 * otherwise we'll refresh it from
4436 * on-disk bitmap and lose not-yet-available
4438 page_cache_get(e4b
->bd_buddy_page
);
4439 page_cache_get(e4b
->bd_bitmap_page
);
4441 db
->bb_tid
= handle
->h_transaction
->t_tid
;
4442 mb_debug("new md 0x%p for group %lu\n",
4450 BUG_ON(md
->num
>= EXT4_BB_MAX_BLOCKS
);
4451 md
->blocks
[md
->num
] = block
+ i
;
4453 if (md
->num
== EXT4_BB_MAX_BLOCKS
) {
4454 /* no more space, put full container on a sb's list */
4455 db
->bb_md_cur
= NULL
;
4458 ext4_unlock_group(sb
, group
);
4463 * Main entry point into mballoc to free blocks
4465 void ext4_mb_free_blocks(handle_t
*handle
, struct inode
*inode
,
4466 unsigned long block
, unsigned long count
,
4467 int metadata
, unsigned long *freed
)
4469 struct buffer_head
*bitmap_bh
= NULL
;
4470 struct super_block
*sb
= inode
->i_sb
;
4471 struct ext4_allocation_context
*ac
= NULL
;
4472 struct ext4_group_desc
*gdp
;
4473 struct ext4_super_block
*es
;
4474 unsigned long overflow
;
4476 struct buffer_head
*gd_bh
;
4477 ext4_group_t block_group
;
4478 struct ext4_sb_info
*sbi
;
4479 struct ext4_buddy e4b
;
4485 ext4_mb_poll_new_transaction(sb
, handle
);
4488 es
= EXT4_SB(sb
)->s_es
;
4489 if (block
< le32_to_cpu(es
->s_first_data_block
) ||
4490 block
+ count
< block
||
4491 block
+ count
> ext4_blocks_count(es
)) {
4492 ext4_error(sb
, __FUNCTION__
,
4493 "Freeing blocks not in datazone - "
4494 "block = %lu, count = %lu", block
, count
);
4498 ext4_debug("freeing block %lu\n", block
);
4500 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4502 ac
->ac_op
= EXT4_MB_HISTORY_FREE
;
4503 ac
->ac_inode
= inode
;
4509 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4512 * Check to see if we are freeing blocks across a group
4515 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4516 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4519 bitmap_bh
= read_block_bitmap(sb
, block_group
);
4522 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4526 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4527 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4528 in_range(block
, ext4_inode_table(sb
, gdp
),
4529 EXT4_SB(sb
)->s_itb_per_group
) ||
4530 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4531 EXT4_SB(sb
)->s_itb_per_group
)) {
4533 ext4_error(sb
, __FUNCTION__
,
4534 "Freeing blocks in system zone - "
4535 "Block = %lu, count = %lu", block
, count
);
4538 BUFFER_TRACE(bitmap_bh
, "getting write access");
4539 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4544 * We are about to modify some metadata. Call the journal APIs
4545 * to unshare ->b_data if a currently-committing transaction is
4548 BUFFER_TRACE(gd_bh
, "get_write_access");
4549 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4553 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4557 #ifdef AGGRESSIVE_CHECK
4560 for (i
= 0; i
< count
; i
++)
4561 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4564 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4567 /* We dirtied the bitmap block */
4568 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4569 err
= ext4_journal_dirty_metadata(handle
, bitmap_bh
);
4572 ac
->ac_b_ex
.fe_group
= block_group
;
4573 ac
->ac_b_ex
.fe_start
= bit
;
4574 ac
->ac_b_ex
.fe_len
= count
;
4575 ext4_mb_store_history(ac
);
4579 /* blocks being freed are metadata. these blocks shouldn't
4580 * be used until this transaction is committed */
4581 ext4_mb_free_metadata(handle
, &e4b
, block_group
, bit
, count
);
4583 ext4_lock_group(sb
, block_group
);
4584 err
= mb_free_blocks(inode
, &e4b
, bit
, count
);
4585 ext4_mb_return_to_preallocation(inode
, &e4b
, block
, count
);
4586 ext4_unlock_group(sb
, block_group
);
4590 spin_lock(sb_bgl_lock(sbi
, block_group
));
4591 gdp
->bg_free_blocks_count
=
4592 cpu_to_le16(le16_to_cpu(gdp
->bg_free_blocks_count
) + count
);
4593 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4594 spin_unlock(sb_bgl_lock(sbi
, block_group
));
4595 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4597 ext4_mb_release_desc(&e4b
);
4601 /* And the group descriptor block */
4602 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4603 ret
= ext4_journal_dirty_metadata(handle
, gd_bh
);
4607 if (overflow
&& !err
) {
4616 ext4_std_error(sb
, err
);
4618 kmem_cache_free(ext4_ac_cachep
, ac
);