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/quotaops.h>
28 #include <linux/buffer_head.h>
29 #include <linux/module.h>
30 #include <linux/swap.h>
31 #include <linux/proc_fs.h>
32 #include <linux/pagemap.h>
33 #include <linux/seq_file.h>
34 #include <linux/version.h>
35 #include "ext4_jbd2.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
);
580 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
582 static void ext4_mb_poll_new_transaction(struct super_block
*, handle_t
*);
583 static void ext4_mb_free_committed_blocks(struct super_block
*);
584 static void ext4_mb_return_to_preallocation(struct inode
*inode
,
585 struct ext4_buddy
*e4b
, sector_t block
,
587 static void ext4_mb_put_pa(struct ext4_allocation_context
*,
588 struct super_block
*, struct ext4_prealloc_space
*pa
);
589 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
);
590 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
);
593 static inline void ext4_lock_group(struct super_block
*sb
, ext4_group_t group
)
595 struct ext4_group_info
*grinfo
= ext4_get_group_info(sb
, group
);
597 bit_spin_lock(EXT4_GROUP_INFO_LOCKED_BIT
, &(grinfo
->bb_state
));
600 static inline void ext4_unlock_group(struct super_block
*sb
,
603 struct ext4_group_info
*grinfo
= ext4_get_group_info(sb
, group
);
605 bit_spin_unlock(EXT4_GROUP_INFO_LOCKED_BIT
, &(grinfo
->bb_state
));
608 static inline int ext4_is_group_locked(struct super_block
*sb
,
611 struct ext4_group_info
*grinfo
= ext4_get_group_info(sb
, group
);
613 return bit_spin_is_locked(EXT4_GROUP_INFO_LOCKED_BIT
,
614 &(grinfo
->bb_state
));
617 static ext4_fsblk_t
ext4_grp_offs_to_block(struct super_block
*sb
,
618 struct ext4_free_extent
*fex
)
622 block
= (ext4_fsblk_t
) fex
->fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
624 + le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
628 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
630 #if BITS_PER_LONG == 64
631 *bit
+= ((unsigned long) addr
& 7UL) << 3;
632 addr
= (void *) ((unsigned long) addr
& ~7UL);
633 #elif BITS_PER_LONG == 32
634 *bit
+= ((unsigned long) addr
& 3UL) << 3;
635 addr
= (void *) ((unsigned long) addr
& ~3UL);
637 #error "how many bits you are?!"
642 static inline int mb_test_bit(int bit
, void *addr
)
645 * ext4_test_bit on architecture like powerpc
646 * needs unsigned long aligned address
648 addr
= mb_correct_addr_and_bit(&bit
, addr
);
649 return ext4_test_bit(bit
, addr
);
652 static inline void mb_set_bit(int bit
, void *addr
)
654 addr
= mb_correct_addr_and_bit(&bit
, addr
);
655 ext4_set_bit(bit
, addr
);
658 static inline void mb_set_bit_atomic(spinlock_t
*lock
, int bit
, void *addr
)
660 addr
= mb_correct_addr_and_bit(&bit
, addr
);
661 ext4_set_bit_atomic(lock
, bit
, addr
);
664 static inline void mb_clear_bit(int bit
, void *addr
)
666 addr
= mb_correct_addr_and_bit(&bit
, addr
);
667 ext4_clear_bit(bit
, addr
);
670 static inline void mb_clear_bit_atomic(spinlock_t
*lock
, int bit
, void *addr
)
672 addr
= mb_correct_addr_and_bit(&bit
, addr
);
673 ext4_clear_bit_atomic(lock
, bit
, addr
);
676 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
679 addr
= mb_correct_addr_and_bit(&fix
, addr
);
683 return ext4_find_next_zero_bit(addr
, max
, start
) - fix
;
686 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
689 addr
= mb_correct_addr_and_bit(&fix
, addr
);
693 return ext4_find_next_bit(addr
, max
, start
) - fix
;
696 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
700 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
703 if (order
> e4b
->bd_blkbits
+ 1) {
708 /* at order 0 we see each particular block */
709 *max
= 1 << (e4b
->bd_blkbits
+ 3);
711 return EXT4_MB_BITMAP(e4b
);
713 bb
= EXT4_MB_BUDDY(e4b
) + EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
714 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
720 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
721 int first
, int count
)
724 struct super_block
*sb
= e4b
->bd_sb
;
726 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
728 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
729 for (i
= 0; i
< count
; i
++) {
730 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
731 ext4_fsblk_t blocknr
;
732 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
733 blocknr
+= first
+ i
;
735 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
737 ext4_error(sb
, __func__
, "double-free of inode"
738 " %lu's block %llu(bit %u in group %lu)\n",
739 inode
? inode
->i_ino
: 0, blocknr
,
740 first
+ i
, e4b
->bd_group
);
742 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
746 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
750 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
752 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
753 for (i
= 0; i
< count
; i
++) {
754 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
755 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
759 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
761 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
762 unsigned char *b1
, *b2
;
764 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
765 b2
= (unsigned char *) bitmap
;
766 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
767 if (b1
[i
] != b2
[i
]) {
768 printk("corruption in group %lu at byte %u(%u):"
769 " %x in copy != %x on disk/prealloc\n",
770 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
778 static inline void mb_free_blocks_double(struct inode
*inode
,
779 struct ext4_buddy
*e4b
, int first
, int count
)
783 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
784 int first
, int count
)
788 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
794 #ifdef AGGRESSIVE_CHECK
796 #define MB_CHECK_ASSERT(assert) \
800 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
801 function, file, line, # assert); \
806 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
807 const char *function
, int line
)
809 struct super_block
*sb
= e4b
->bd_sb
;
810 int order
= e4b
->bd_blkbits
+ 1;
817 struct ext4_group_info
*grp
;
820 struct list_head
*cur
;
824 if (!test_opt(sb
, MBALLOC
))
828 static int mb_check_counter
;
829 if (mb_check_counter
++ % 100 != 0)
834 buddy
= mb_find_buddy(e4b
, order
, &max
);
835 MB_CHECK_ASSERT(buddy
);
836 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
837 MB_CHECK_ASSERT(buddy2
);
838 MB_CHECK_ASSERT(buddy
!= buddy2
);
839 MB_CHECK_ASSERT(max
* 2 == max2
);
842 for (i
= 0; i
< max
; i
++) {
844 if (mb_test_bit(i
, buddy
)) {
845 /* only single bit in buddy2 may be 1 */
846 if (!mb_test_bit(i
<< 1, buddy2
)) {
848 mb_test_bit((i
<<1)+1, buddy2
));
849 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
851 mb_test_bit(i
<< 1, buddy2
));
856 /* both bits in buddy2 must be 0 */
857 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
858 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
860 for (j
= 0; j
< (1 << order
); j
++) {
861 k
= (i
* (1 << order
)) + j
;
863 !mb_test_bit(k
, EXT4_MB_BITMAP(e4b
)));
867 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
872 buddy
= mb_find_buddy(e4b
, 0, &max
);
873 for (i
= 0; i
< max
; i
++) {
874 if (!mb_test_bit(i
, buddy
)) {
875 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
883 /* check used bits only */
884 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
885 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
887 MB_CHECK_ASSERT(k
< max2
);
888 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
891 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
892 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
894 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
895 buddy
= mb_find_buddy(e4b
, 0, &max
);
896 list_for_each(cur
, &grp
->bb_prealloc_list
) {
897 ext4_group_t groupnr
;
898 struct ext4_prealloc_space
*pa
;
899 pa
= list_entry(cur
, struct ext4_prealloc_space
, group_list
);
900 ext4_get_group_no_and_offset(sb
, pa
->pstart
, &groupnr
, &k
);
901 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
902 for (i
= 0; i
< pa
->len
; i
++)
903 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
907 #undef MB_CHECK_ASSERT
908 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
909 __FILE__, __func__, __LINE__)
911 #define mb_check_buddy(e4b)
914 /* FIXME!! need more doc */
915 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
916 void *buddy
, unsigned first
, int len
,
917 struct ext4_group_info
*grp
)
919 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
922 unsigned short chunk
;
923 unsigned short border
;
925 BUG_ON(len
> EXT4_BLOCKS_PER_GROUP(sb
));
927 border
= 2 << sb
->s_blocksize_bits
;
930 /* find how many blocks can be covered since this position */
931 max
= ffs(first
| border
) - 1;
933 /* find how many blocks of power 2 we need to mark */
940 /* mark multiblock chunks only */
941 grp
->bb_counters
[min
]++;
943 mb_clear_bit(first
>> min
,
944 buddy
+ sbi
->s_mb_offsets
[min
]);
951 static void ext4_mb_generate_buddy(struct super_block
*sb
,
952 void *buddy
, void *bitmap
, ext4_group_t group
)
954 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
955 unsigned short max
= EXT4_BLOCKS_PER_GROUP(sb
);
956 unsigned short i
= 0;
957 unsigned short first
;
960 unsigned fragments
= 0;
961 unsigned long long period
= get_cycles();
963 /* initialize buddy from bitmap which is aggregation
964 * of on-disk bitmap and preallocations */
965 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
966 grp
->bb_first_free
= i
;
970 i
= mb_find_next_bit(bitmap
, max
, i
);
974 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
976 grp
->bb_counters
[0]++;
978 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
980 grp
->bb_fragments
= fragments
;
982 if (free
!= grp
->bb_free
) {
983 ext4_error(sb
, __func__
,
984 "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
985 group
, free
, grp
->bb_free
);
987 * If we intent to continue, we consider group descritor
988 * corrupt and update bb_free using bitmap value
993 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
995 period
= get_cycles() - period
;
996 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
997 EXT4_SB(sb
)->s_mb_buddies_generated
++;
998 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
999 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
1002 /* The buddy information is attached the buddy cache inode
1003 * for convenience. The information regarding each group
1004 * is loaded via ext4_mb_load_buddy. The information involve
1005 * block bitmap and buddy information. The information are
1006 * stored in the inode as
1009 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
1012 * one block each for bitmap and buddy information.
1013 * So for each group we take up 2 blocks. A page can
1014 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
1015 * So it can have information regarding groups_per_page which
1016 * is blocks_per_page/2
1019 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
1022 int blocks_per_page
;
1023 int groups_per_page
;
1026 ext4_group_t first_group
;
1028 struct super_block
*sb
;
1029 struct buffer_head
*bhs
;
1030 struct buffer_head
**bh
;
1031 struct inode
*inode
;
1035 mb_debug("init page %lu\n", page
->index
);
1037 inode
= page
->mapping
->host
;
1039 blocksize
= 1 << inode
->i_blkbits
;
1040 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
1042 groups_per_page
= blocks_per_page
>> 1;
1043 if (groups_per_page
== 0)
1044 groups_per_page
= 1;
1046 /* allocate buffer_heads to read bitmaps */
1047 if (groups_per_page
> 1) {
1049 i
= sizeof(struct buffer_head
*) * groups_per_page
;
1050 bh
= kzalloc(i
, GFP_NOFS
);
1056 first_group
= page
->index
* blocks_per_page
/ 2;
1058 /* read all groups the page covers into the cache */
1059 for (i
= 0; i
< groups_per_page
; i
++) {
1060 struct ext4_group_desc
*desc
;
1062 if (first_group
+ i
>= EXT4_SB(sb
)->s_groups_count
)
1066 desc
= ext4_get_group_desc(sb
, first_group
+ i
, NULL
);
1071 bh
[i
] = sb_getblk(sb
, ext4_block_bitmap(sb
, desc
));
1075 if (bh_uptodate_or_lock(bh
[i
]))
1078 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
1079 ext4_init_block_bitmap(sb
, bh
[i
],
1080 first_group
+ i
, desc
);
1081 set_buffer_uptodate(bh
[i
]);
1082 unlock_buffer(bh
[i
]);
1086 bh
[i
]->b_end_io
= end_buffer_read_sync
;
1087 submit_bh(READ
, bh
[i
]);
1088 mb_debug("read bitmap for group %lu\n", first_group
+ i
);
1091 /* wait for I/O completion */
1092 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
1093 wait_on_buffer(bh
[i
]);
1096 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
1097 if (!buffer_uptodate(bh
[i
]))
1100 first_block
= page
->index
* blocks_per_page
;
1101 for (i
= 0; i
< blocks_per_page
; i
++) {
1103 struct ext4_group_info
*grinfo
;
1105 group
= (first_block
+ i
) >> 1;
1106 if (group
>= EXT4_SB(sb
)->s_groups_count
)
1110 * data carry information regarding this
1111 * particular group in the format specified
1115 data
= page_address(page
) + (i
* blocksize
);
1116 bitmap
= bh
[group
- first_group
]->b_data
;
1119 * We place the buddy block and bitmap block
1122 if ((first_block
+ i
) & 1) {
1123 /* this is block of buddy */
1124 BUG_ON(incore
== NULL
);
1125 mb_debug("put buddy for group %u in page %lu/%x\n",
1126 group
, page
->index
, i
* blocksize
);
1127 memset(data
, 0xff, blocksize
);
1128 grinfo
= ext4_get_group_info(sb
, group
);
1129 grinfo
->bb_fragments
= 0;
1130 memset(grinfo
->bb_counters
, 0,
1131 sizeof(unsigned short)*(sb
->s_blocksize_bits
+2));
1133 * incore got set to the group block bitmap below
1135 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
1138 /* this is block of bitmap */
1139 BUG_ON(incore
!= NULL
);
1140 mb_debug("put bitmap for group %u in page %lu/%x\n",
1141 group
, page
->index
, i
* blocksize
);
1143 /* see comments in ext4_mb_put_pa() */
1144 ext4_lock_group(sb
, group
);
1145 memcpy(data
, bitmap
, blocksize
);
1147 /* mark all preallocated blks used in in-core bitmap */
1148 ext4_mb_generate_from_pa(sb
, data
, group
);
1149 ext4_unlock_group(sb
, group
);
1151 /* set incore so that the buddy information can be
1152 * generated using this
1157 SetPageUptodate(page
);
1161 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
1169 static noinline_for_stack
int
1170 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
1171 struct ext4_buddy
*e4b
)
1173 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1174 struct inode
*inode
= sbi
->s_buddy_cache
;
1175 int blocks_per_page
;
1181 mb_debug("load group %lu\n", group
);
1183 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1185 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
1186 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
1188 e4b
->bd_group
= group
;
1189 e4b
->bd_buddy_page
= NULL
;
1190 e4b
->bd_bitmap_page
= NULL
;
1193 * the buddy cache inode stores the block bitmap
1194 * and buddy information in consecutive blocks.
1195 * So for each group we need two blocks.
1198 pnum
= block
/ blocks_per_page
;
1199 poff
= block
% blocks_per_page
;
1201 /* we could use find_or_create_page(), but it locks page
1202 * what we'd like to avoid in fast path ... */
1203 page
= find_get_page(inode
->i_mapping
, pnum
);
1204 if (page
== NULL
|| !PageUptodate(page
)) {
1206 page_cache_release(page
);
1207 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1209 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1210 if (!PageUptodate(page
)) {
1211 ext4_mb_init_cache(page
, NULL
);
1212 mb_cmp_bitmaps(e4b
, page_address(page
) +
1213 (poff
* sb
->s_blocksize
));
1218 if (page
== NULL
|| !PageUptodate(page
))
1220 e4b
->bd_bitmap_page
= page
;
1221 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1222 mark_page_accessed(page
);
1225 pnum
= block
/ blocks_per_page
;
1226 poff
= block
% blocks_per_page
;
1228 page
= find_get_page(inode
->i_mapping
, pnum
);
1229 if (page
== NULL
|| !PageUptodate(page
)) {
1231 page_cache_release(page
);
1232 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1234 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1235 if (!PageUptodate(page
))
1236 ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1241 if (page
== NULL
|| !PageUptodate(page
))
1243 e4b
->bd_buddy_page
= page
;
1244 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1245 mark_page_accessed(page
);
1247 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1248 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1253 if (e4b
->bd_bitmap_page
)
1254 page_cache_release(e4b
->bd_bitmap_page
);
1255 if (e4b
->bd_buddy_page
)
1256 page_cache_release(e4b
->bd_buddy_page
);
1257 e4b
->bd_buddy
= NULL
;
1258 e4b
->bd_bitmap
= NULL
;
1262 static void ext4_mb_release_desc(struct ext4_buddy
*e4b
)
1264 if (e4b
->bd_bitmap_page
)
1265 page_cache_release(e4b
->bd_bitmap_page
);
1266 if (e4b
->bd_buddy_page
)
1267 page_cache_release(e4b
->bd_buddy_page
);
1271 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1276 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1277 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1279 bb
= EXT4_MB_BUDDY(e4b
);
1280 while (order
<= e4b
->bd_blkbits
+ 1) {
1282 if (!mb_test_bit(block
, bb
)) {
1283 /* this block is part of buddy of order 'order' */
1286 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1292 static void mb_clear_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1298 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1299 /* fast path: clear whole word at once */
1300 addr
= bm
+ (cur
>> 3);
1305 mb_clear_bit_atomic(lock
, cur
, bm
);
1310 static void mb_set_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1316 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1317 /* fast path: set whole word at once */
1318 addr
= bm
+ (cur
>> 3);
1323 mb_set_bit_atomic(lock
, cur
, bm
);
1328 static int mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1329 int first
, int count
)
1336 struct super_block
*sb
= e4b
->bd_sb
;
1338 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1339 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
1340 mb_check_buddy(e4b
);
1341 mb_free_blocks_double(inode
, e4b
, first
, count
);
1343 e4b
->bd_info
->bb_free
+= count
;
1344 if (first
< e4b
->bd_info
->bb_first_free
)
1345 e4b
->bd_info
->bb_first_free
= first
;
1347 /* let's maintain fragments counter */
1349 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1350 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1351 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1353 e4b
->bd_info
->bb_fragments
--;
1354 else if (!block
&& !max
)
1355 e4b
->bd_info
->bb_fragments
++;
1357 /* let's maintain buddy itself */
1358 while (count
-- > 0) {
1362 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1363 ext4_fsblk_t blocknr
;
1364 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
1367 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
1369 ext4_error(sb
, __func__
, "double-free of inode"
1370 " %lu's block %llu(bit %u in group %lu)\n",
1371 inode
? inode
->i_ino
: 0, blocknr
, block
,
1374 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1375 e4b
->bd_info
->bb_counters
[order
]++;
1377 /* start of the buddy */
1378 buddy
= mb_find_buddy(e4b
, order
, &max
);
1382 if (mb_test_bit(block
, buddy
) ||
1383 mb_test_bit(block
+ 1, buddy
))
1386 /* both the buddies are free, try to coalesce them */
1387 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1393 /* for special purposes, we don't set
1394 * free bits in bitmap */
1395 mb_set_bit(block
, buddy
);
1396 mb_set_bit(block
+ 1, buddy
);
1398 e4b
->bd_info
->bb_counters
[order
]--;
1399 e4b
->bd_info
->bb_counters
[order
]--;
1403 e4b
->bd_info
->bb_counters
[order
]++;
1405 mb_clear_bit(block
, buddy2
);
1409 mb_check_buddy(e4b
);
1414 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1415 int needed
, struct ext4_free_extent
*ex
)
1422 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1425 buddy
= mb_find_buddy(e4b
, order
, &max
);
1426 BUG_ON(buddy
== NULL
);
1427 BUG_ON(block
>= max
);
1428 if (mb_test_bit(block
, buddy
)) {
1435 /* FIXME dorp order completely ? */
1436 if (likely(order
== 0)) {
1437 /* find actual order */
1438 order
= mb_find_order_for_block(e4b
, block
);
1439 block
= block
>> order
;
1442 ex
->fe_len
= 1 << order
;
1443 ex
->fe_start
= block
<< order
;
1444 ex
->fe_group
= e4b
->bd_group
;
1446 /* calc difference from given start */
1447 next
= next
- ex
->fe_start
;
1449 ex
->fe_start
+= next
;
1451 while (needed
> ex
->fe_len
&&
1452 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1454 if (block
+ 1 >= max
)
1457 next
= (block
+ 1) * (1 << order
);
1458 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1461 ord
= mb_find_order_for_block(e4b
, next
);
1464 block
= next
>> order
;
1465 ex
->fe_len
+= 1 << order
;
1468 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1472 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1478 int start
= ex
->fe_start
;
1479 int len
= ex
->fe_len
;
1484 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1485 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1486 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1487 mb_check_buddy(e4b
);
1488 mb_mark_used_double(e4b
, start
, len
);
1490 e4b
->bd_info
->bb_free
-= len
;
1491 if (e4b
->bd_info
->bb_first_free
== start
)
1492 e4b
->bd_info
->bb_first_free
+= len
;
1494 /* let's maintain fragments counter */
1496 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1497 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1498 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1500 e4b
->bd_info
->bb_fragments
++;
1501 else if (!mlen
&& !max
)
1502 e4b
->bd_info
->bb_fragments
--;
1504 /* let's maintain buddy itself */
1506 ord
= mb_find_order_for_block(e4b
, start
);
1508 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1509 /* the whole chunk may be allocated at once! */
1511 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1512 BUG_ON((start
>> ord
) >= max
);
1513 mb_set_bit(start
>> ord
, buddy
);
1514 e4b
->bd_info
->bb_counters
[ord
]--;
1521 /* store for history */
1523 ret
= len
| (ord
<< 16);
1525 /* we have to split large buddy */
1527 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1528 mb_set_bit(start
>> ord
, buddy
);
1529 e4b
->bd_info
->bb_counters
[ord
]--;
1532 cur
= (start
>> ord
) & ~1U;
1533 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1534 mb_clear_bit(cur
, buddy
);
1535 mb_clear_bit(cur
+ 1, buddy
);
1536 e4b
->bd_info
->bb_counters
[ord
]++;
1537 e4b
->bd_info
->bb_counters
[ord
]++;
1540 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b
->bd_sb
), ex
->fe_group
),
1541 EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1542 mb_check_buddy(e4b
);
1548 * Must be called under group lock!
1550 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1551 struct ext4_buddy
*e4b
)
1553 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1556 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1557 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1559 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1560 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1561 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1563 /* preallocation can change ac_b_ex, thus we store actually
1564 * allocated blocks for history */
1565 ac
->ac_f_ex
= ac
->ac_b_ex
;
1567 ac
->ac_status
= AC_STATUS_FOUND
;
1568 ac
->ac_tail
= ret
& 0xffff;
1569 ac
->ac_buddy
= ret
>> 16;
1571 /* XXXXXXX: SUCH A HORRIBLE **CK */
1573 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1574 get_page(ac
->ac_bitmap_page
);
1575 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1576 get_page(ac
->ac_buddy_page
);
1578 /* store last allocated for subsequent stream allocation */
1579 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1580 spin_lock(&sbi
->s_md_lock
);
1581 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1582 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1583 spin_unlock(&sbi
->s_md_lock
);
1588 * regular allocator, for general purposes allocation
1591 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1592 struct ext4_buddy
*e4b
,
1595 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1596 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1597 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1598 struct ext4_free_extent ex
;
1602 * We don't want to scan for a whole year
1604 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1605 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1606 ac
->ac_status
= AC_STATUS_BREAK
;
1611 * Haven't found good chunk so far, let's continue
1613 if (bex
->fe_len
< gex
->fe_len
)
1616 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1617 && bex
->fe_group
== e4b
->bd_group
) {
1618 /* recheck chunk's availability - we don't know
1619 * when it was found (within this lock-unlock
1621 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1622 if (max
>= gex
->fe_len
) {
1623 ext4_mb_use_best_found(ac
, e4b
);
1630 * The routine checks whether found extent is good enough. If it is,
1631 * then the extent gets marked used and flag is set to the context
1632 * to stop scanning. Otherwise, the extent is compared with the
1633 * previous found extent and if new one is better, then it's stored
1634 * in the context. Later, the best found extent will be used, if
1635 * mballoc can't find good enough extent.
1637 * FIXME: real allocation policy is to be designed yet!
1639 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1640 struct ext4_free_extent
*ex
,
1641 struct ext4_buddy
*e4b
)
1643 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1644 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1646 BUG_ON(ex
->fe_len
<= 0);
1647 BUG_ON(ex
->fe_len
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1648 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1649 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1654 * The special case - take what you catch first
1656 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1658 ext4_mb_use_best_found(ac
, e4b
);
1663 * Let's check whether the chuck is good enough
1665 if (ex
->fe_len
== gex
->fe_len
) {
1667 ext4_mb_use_best_found(ac
, e4b
);
1672 * If this is first found extent, just store it in the context
1674 if (bex
->fe_len
== 0) {
1680 * If new found extent is better, store it in the context
1682 if (bex
->fe_len
< gex
->fe_len
) {
1683 /* if the request isn't satisfied, any found extent
1684 * larger than previous best one is better */
1685 if (ex
->fe_len
> bex
->fe_len
)
1687 } else if (ex
->fe_len
> gex
->fe_len
) {
1688 /* if the request is satisfied, then we try to find
1689 * an extent that still satisfy the request, but is
1690 * smaller than previous one */
1691 if (ex
->fe_len
< bex
->fe_len
)
1695 ext4_mb_check_limits(ac
, e4b
, 0);
1698 static int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1699 struct ext4_buddy
*e4b
)
1701 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1702 ext4_group_t group
= ex
.fe_group
;
1706 BUG_ON(ex
.fe_len
<= 0);
1707 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1711 ext4_lock_group(ac
->ac_sb
, group
);
1712 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1716 ext4_mb_use_best_found(ac
, e4b
);
1719 ext4_unlock_group(ac
->ac_sb
, group
);
1720 ext4_mb_release_desc(e4b
);
1725 static int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1726 struct ext4_buddy
*e4b
)
1728 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1731 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1732 struct ext4_super_block
*es
= sbi
->s_es
;
1733 struct ext4_free_extent ex
;
1735 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1738 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1742 ext4_lock_group(ac
->ac_sb
, group
);
1743 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1744 ac
->ac_g_ex
.fe_len
, &ex
);
1746 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1749 start
= (e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
)) +
1750 ex
.fe_start
+ le32_to_cpu(es
->s_first_data_block
);
1751 /* use do_div to get remainder (would be 64-bit modulo) */
1752 if (do_div(start
, sbi
->s_stripe
) == 0) {
1755 ext4_mb_use_best_found(ac
, e4b
);
1757 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1758 BUG_ON(ex
.fe_len
<= 0);
1759 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1760 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1763 ext4_mb_use_best_found(ac
, e4b
);
1764 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1765 /* Sometimes, caller may want to merge even small
1766 * number of blocks to an existing extent */
1767 BUG_ON(ex
.fe_len
<= 0);
1768 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1769 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1772 ext4_mb_use_best_found(ac
, e4b
);
1774 ext4_unlock_group(ac
->ac_sb
, group
);
1775 ext4_mb_release_desc(e4b
);
1781 * The routine scans buddy structures (not bitmap!) from given order
1782 * to max order and tries to find big enough chunk to satisfy the req
1784 static void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1785 struct ext4_buddy
*e4b
)
1787 struct super_block
*sb
= ac
->ac_sb
;
1788 struct ext4_group_info
*grp
= e4b
->bd_info
;
1794 BUG_ON(ac
->ac_2order
<= 0);
1795 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1796 if (grp
->bb_counters
[i
] == 0)
1799 buddy
= mb_find_buddy(e4b
, i
, &max
);
1800 BUG_ON(buddy
== NULL
);
1802 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1807 ac
->ac_b_ex
.fe_len
= 1 << i
;
1808 ac
->ac_b_ex
.fe_start
= k
<< i
;
1809 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1811 ext4_mb_use_best_found(ac
, e4b
);
1813 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1815 if (EXT4_SB(sb
)->s_mb_stats
)
1816 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1823 * The routine scans the group and measures all found extents.
1824 * In order to optimize scanning, caller must pass number of
1825 * free blocks in the group, so the routine can know upper limit.
1827 static void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1828 struct ext4_buddy
*e4b
)
1830 struct super_block
*sb
= ac
->ac_sb
;
1831 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1832 struct ext4_free_extent ex
;
1836 free
= e4b
->bd_info
->bb_free
;
1839 i
= e4b
->bd_info
->bb_first_free
;
1841 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1842 i
= mb_find_next_zero_bit(bitmap
,
1843 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1844 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1846 * IF we have corrupt bitmap, we won't find any
1847 * free blocks even though group info says we
1848 * we have free blocks
1850 ext4_error(sb
, __func__
, "%d free blocks as per "
1851 "group info. But bitmap says 0\n",
1856 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1857 BUG_ON(ex
.fe_len
<= 0);
1858 if (free
< ex
.fe_len
) {
1859 ext4_error(sb
, __func__
, "%d free blocks as per "
1860 "group info. But got %d blocks\n",
1863 * The number of free blocks differs. This mostly
1864 * indicate that the bitmap is corrupt. So exit
1865 * without claiming the space.
1870 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1876 ext4_mb_check_limits(ac
, e4b
, 1);
1880 * This is a special case for storages like raid5
1881 * we try to find stripe-aligned chunks for stripe-size requests
1882 * XXX should do so at least for multiples of stripe size as well
1884 static void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1885 struct ext4_buddy
*e4b
)
1887 struct super_block
*sb
= ac
->ac_sb
;
1888 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1889 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1890 struct ext4_free_extent ex
;
1891 ext4_fsblk_t first_group_block
;
1896 BUG_ON(sbi
->s_stripe
== 0);
1898 /* find first stripe-aligned block in group */
1899 first_group_block
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
)
1900 + le32_to_cpu(sbi
->s_es
->s_first_data_block
);
1901 a
= first_group_block
+ sbi
->s_stripe
- 1;
1902 do_div(a
, sbi
->s_stripe
);
1903 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1905 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1906 if (!mb_test_bit(i
, bitmap
)) {
1907 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1908 if (max
>= sbi
->s_stripe
) {
1911 ext4_mb_use_best_found(ac
, e4b
);
1919 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1920 ext4_group_t group
, int cr
)
1922 unsigned free
, fragments
;
1924 struct ext4_group_desc
*desc
;
1925 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1927 BUG_ON(cr
< 0 || cr
>= 4);
1928 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp
));
1930 free
= grp
->bb_free
;
1931 fragments
= grp
->bb_fragments
;
1939 BUG_ON(ac
->ac_2order
== 0);
1940 /* If this group is uninitialized, skip it initially */
1941 desc
= ext4_get_group_desc(ac
->ac_sb
, group
, NULL
);
1942 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))
1945 bits
= ac
->ac_sb
->s_blocksize_bits
+ 1;
1946 for (i
= ac
->ac_2order
; i
<= bits
; i
++)
1947 if (grp
->bb_counters
[i
] > 0)
1951 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1955 if (free
>= ac
->ac_g_ex
.fe_len
)
1967 static noinline_for_stack
int
1968 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 noinline_for_stack
void
2469 ext4_mb_store_history(struct ext4_allocation_context
*ac
)
2471 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2472 struct ext4_mb_history h
;
2474 if (unlikely(sbi
->s_mb_history
== NULL
))
2477 if (!(ac
->ac_op
& sbi
->s_mb_history_filter
))
2481 h
.pid
= current
->pid
;
2482 h
.ino
= ac
->ac_inode
? ac
->ac_inode
->i_ino
: 0;
2483 h
.orig
= ac
->ac_o_ex
;
2484 h
.result
= ac
->ac_b_ex
;
2485 h
.flags
= ac
->ac_flags
;
2486 h
.found
= ac
->ac_found
;
2487 h
.groups
= ac
->ac_groups_scanned
;
2488 h
.cr
= ac
->ac_criteria
;
2489 h
.tail
= ac
->ac_tail
;
2490 h
.buddy
= ac
->ac_buddy
;
2492 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
) {
2493 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
2494 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
2496 h
.goal
= ac
->ac_g_ex
;
2497 h
.result
= ac
->ac_f_ex
;
2500 spin_lock(&sbi
->s_mb_history_lock
);
2501 memcpy(sbi
->s_mb_history
+ sbi
->s_mb_history_cur
, &h
, sizeof(h
));
2502 if (++sbi
->s_mb_history_cur
>= sbi
->s_mb_history_max
)
2503 sbi
->s_mb_history_cur
= 0;
2504 spin_unlock(&sbi
->s_mb_history_lock
);
2508 #define ext4_mb_history_release(sb)
2509 #define ext4_mb_history_init(sb)
2512 static int ext4_mb_init_backend(struct super_block
*sb
)
2515 int j
, len
, metalen
;
2516 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2517 int num_meta_group_infos
=
2518 (sbi
->s_groups_count
+ EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2519 EXT4_DESC_PER_BLOCK_BITS(sb
);
2520 struct ext4_group_info
**meta_group_info
;
2522 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2523 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2524 * So a two level scheme suffices for now. */
2525 sbi
->s_group_info
= kmalloc(sizeof(*sbi
->s_group_info
) *
2526 num_meta_group_infos
, GFP_KERNEL
);
2527 if (sbi
->s_group_info
== NULL
) {
2528 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2531 sbi
->s_buddy_cache
= new_inode(sb
);
2532 if (sbi
->s_buddy_cache
== NULL
) {
2533 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2536 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2538 metalen
= sizeof(*meta_group_info
) << EXT4_DESC_PER_BLOCK_BITS(sb
);
2539 for (i
= 0; i
< num_meta_group_infos
; i
++) {
2540 if ((i
+ 1) == num_meta_group_infos
)
2541 metalen
= sizeof(*meta_group_info
) *
2542 (sbi
->s_groups_count
-
2543 (i
<< EXT4_DESC_PER_BLOCK_BITS(sb
)));
2544 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2545 if (meta_group_info
== NULL
) {
2546 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2550 sbi
->s_group_info
[i
] = meta_group_info
;
2554 * calculate needed size. if change bb_counters size,
2555 * don't forget about ext4_mb_generate_buddy()
2557 len
= sizeof(struct ext4_group_info
);
2558 len
+= sizeof(unsigned short) * (sb
->s_blocksize_bits
+ 2);
2559 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2560 struct ext4_group_desc
*desc
;
2563 sbi
->s_group_info
[i
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2564 j
= i
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2566 meta_group_info
[j
] = kzalloc(len
, GFP_KERNEL
);
2567 if (meta_group_info
[j
] == NULL
) {
2568 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2572 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2575 "EXT4-fs: can't read descriptor %lu\n", i
);
2578 memset(meta_group_info
[j
], 0, len
);
2579 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2580 &(meta_group_info
[j
]->bb_state
));
2583 * initialize bb_free to be able to skip
2584 * empty groups without initialization
2586 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2587 meta_group_info
[j
]->bb_free
=
2588 ext4_free_blocks_after_init(sb
, i
, desc
);
2590 meta_group_info
[j
]->bb_free
=
2591 le16_to_cpu(desc
->bg_free_blocks_count
);
2594 INIT_LIST_HEAD(&meta_group_info
[j
]->bb_prealloc_list
);
2598 struct buffer_head
*bh
;
2599 meta_group_info
[j
]->bb_bitmap
=
2600 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2601 BUG_ON(meta_group_info
[j
]->bb_bitmap
== NULL
);
2602 bh
= read_block_bitmap(sb
, i
);
2604 memcpy(meta_group_info
[j
]->bb_bitmap
, bh
->b_data
,
2616 kfree(ext4_get_group_info(sb
, i
));
2619 i
= num_meta_group_infos
;
2622 kfree(sbi
->s_group_info
[i
]);
2623 iput(sbi
->s_buddy_cache
);
2625 kfree(sbi
->s_group_info
);
2629 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2631 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2636 if (!test_opt(sb
, MBALLOC
))
2639 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned short);
2641 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2642 if (sbi
->s_mb_offsets
== NULL
) {
2643 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2646 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2647 if (sbi
->s_mb_maxs
== NULL
) {
2648 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2649 kfree(sbi
->s_mb_maxs
);
2653 /* order 0 is regular bitmap */
2654 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2655 sbi
->s_mb_offsets
[0] = 0;
2659 max
= sb
->s_blocksize
<< 2;
2661 sbi
->s_mb_offsets
[i
] = offset
;
2662 sbi
->s_mb_maxs
[i
] = max
;
2663 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2666 } while (i
<= sb
->s_blocksize_bits
+ 1);
2668 /* init file for buddy data */
2669 i
= ext4_mb_init_backend(sb
);
2671 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2672 kfree(sbi
->s_mb_offsets
);
2673 kfree(sbi
->s_mb_maxs
);
2677 spin_lock_init(&sbi
->s_md_lock
);
2678 INIT_LIST_HEAD(&sbi
->s_active_transaction
);
2679 INIT_LIST_HEAD(&sbi
->s_closed_transaction
);
2680 INIT_LIST_HEAD(&sbi
->s_committed_transaction
);
2681 spin_lock_init(&sbi
->s_bal_lock
);
2683 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2684 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2685 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2686 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2687 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2688 sbi
->s_mb_history_filter
= EXT4_MB_HISTORY_DEFAULT
;
2689 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2691 i
= sizeof(struct ext4_locality_group
) * NR_CPUS
;
2692 sbi
->s_locality_groups
= kmalloc(i
, GFP_KERNEL
);
2693 if (sbi
->s_locality_groups
== NULL
) {
2694 clear_opt(sbi
->s_mount_opt
, MBALLOC
);
2695 kfree(sbi
->s_mb_offsets
);
2696 kfree(sbi
->s_mb_maxs
);
2699 for (i
= 0; i
< NR_CPUS
; i
++) {
2700 struct ext4_locality_group
*lg
;
2701 lg
= &sbi
->s_locality_groups
[i
];
2702 mutex_init(&lg
->lg_mutex
);
2703 INIT_LIST_HEAD(&lg
->lg_prealloc_list
);
2704 spin_lock_init(&lg
->lg_prealloc_lock
);
2707 ext4_mb_init_per_dev_proc(sb
);
2708 ext4_mb_history_init(sb
);
2710 printk("EXT4-fs: mballoc enabled\n");
2714 /* need to called with ext4 group lock (ext4_lock_group) */
2715 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2717 struct ext4_prealloc_space
*pa
;
2718 struct list_head
*cur
, *tmp
;
2721 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2722 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2723 list_del(&pa
->pa_group_list
);
2728 mb_debug("mballoc: %u PAs left\n", count
);
2732 int ext4_mb_release(struct super_block
*sb
)
2735 int num_meta_group_infos
;
2736 struct ext4_group_info
*grinfo
;
2737 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2739 if (!test_opt(sb
, MBALLOC
))
2742 /* release freed, non-committed blocks */
2743 spin_lock(&sbi
->s_md_lock
);
2744 list_splice_init(&sbi
->s_closed_transaction
,
2745 &sbi
->s_committed_transaction
);
2746 list_splice_init(&sbi
->s_active_transaction
,
2747 &sbi
->s_committed_transaction
);
2748 spin_unlock(&sbi
->s_md_lock
);
2749 ext4_mb_free_committed_blocks(sb
);
2751 if (sbi
->s_group_info
) {
2752 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2753 grinfo
= ext4_get_group_info(sb
, i
);
2755 kfree(grinfo
->bb_bitmap
);
2757 ext4_lock_group(sb
, i
);
2758 ext4_mb_cleanup_pa(grinfo
);
2759 ext4_unlock_group(sb
, i
);
2762 num_meta_group_infos
= (sbi
->s_groups_count
+
2763 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2764 EXT4_DESC_PER_BLOCK_BITS(sb
);
2765 for (i
= 0; i
< num_meta_group_infos
; i
++)
2766 kfree(sbi
->s_group_info
[i
]);
2767 kfree(sbi
->s_group_info
);
2769 kfree(sbi
->s_mb_offsets
);
2770 kfree(sbi
->s_mb_maxs
);
2771 if (sbi
->s_buddy_cache
)
2772 iput(sbi
->s_buddy_cache
);
2773 if (sbi
->s_mb_stats
) {
2775 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2776 atomic_read(&sbi
->s_bal_allocated
),
2777 atomic_read(&sbi
->s_bal_reqs
),
2778 atomic_read(&sbi
->s_bal_success
));
2780 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2781 "%u 2^N hits, %u breaks, %u lost\n",
2782 atomic_read(&sbi
->s_bal_ex_scanned
),
2783 atomic_read(&sbi
->s_bal_goals
),
2784 atomic_read(&sbi
->s_bal_2orders
),
2785 atomic_read(&sbi
->s_bal_breaks
),
2786 atomic_read(&sbi
->s_mb_lost_chunks
));
2788 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2789 sbi
->s_mb_buddies_generated
++,
2790 sbi
->s_mb_generation_time
);
2792 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2793 atomic_read(&sbi
->s_mb_preallocated
),
2794 atomic_read(&sbi
->s_mb_discarded
));
2797 kfree(sbi
->s_locality_groups
);
2799 ext4_mb_history_release(sb
);
2800 ext4_mb_destroy_per_dev_proc(sb
);
2805 static noinline_for_stack
void
2806 ext4_mb_free_committed_blocks(struct super_block
*sb
)
2808 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2813 struct ext4_free_metadata
*md
;
2814 struct ext4_buddy e4b
;
2816 if (list_empty(&sbi
->s_committed_transaction
))
2819 /* there is committed blocks to be freed yet */
2821 /* get next array of blocks */
2823 spin_lock(&sbi
->s_md_lock
);
2824 if (!list_empty(&sbi
->s_committed_transaction
)) {
2825 md
= list_entry(sbi
->s_committed_transaction
.next
,
2826 struct ext4_free_metadata
, list
);
2827 list_del(&md
->list
);
2829 spin_unlock(&sbi
->s_md_lock
);
2834 mb_debug("gonna free %u blocks in group %lu (0x%p):",
2835 md
->num
, md
->group
, md
);
2837 err
= ext4_mb_load_buddy(sb
, md
->group
, &e4b
);
2838 /* we expect to find existing buddy because it's pinned */
2841 /* there are blocks to put in buddy to make them really free */
2844 ext4_lock_group(sb
, md
->group
);
2845 for (i
= 0; i
< md
->num
; i
++) {
2846 mb_debug(" %u", md
->blocks
[i
]);
2847 err
= mb_free_blocks(NULL
, &e4b
, md
->blocks
[i
], 1);
2851 ext4_unlock_group(sb
, md
->group
);
2853 /* balance refcounts from ext4_mb_free_metadata() */
2854 page_cache_release(e4b
.bd_buddy_page
);
2855 page_cache_release(e4b
.bd_bitmap_page
);
2858 ext4_mb_release_desc(&e4b
);
2862 mb_debug("freed %u blocks in %u structures\n", count
, count2
);
2865 #define EXT4_MB_STATS_NAME "stats"
2866 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2867 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2868 #define EXT4_MB_ORDER2_REQ "order2_req"
2869 #define EXT4_MB_STREAM_REQ "stream_req"
2870 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2874 #define MB_PROC_VALUE_READ(name) \
2875 static int ext4_mb_read_##name(char *page, char **start, \
2876 off_t off, int count, int *eof, void *data) \
2878 struct ext4_sb_info *sbi = data; \
2883 len = sprintf(page, "%ld\n", sbi->s_mb_##name); \
2888 #define MB_PROC_VALUE_WRITE(name) \
2889 static int ext4_mb_write_##name(struct file *file, \
2890 const char __user *buf, unsigned long cnt, void *data) \
2892 struct ext4_sb_info *sbi = data; \
2895 if (cnt >= sizeof(str)) \
2897 if (copy_from_user(str, buf, cnt)) \
2899 value = simple_strtol(str, NULL, 0); \
2902 sbi->s_mb_##name = value; \
2906 MB_PROC_VALUE_READ(stats
);
2907 MB_PROC_VALUE_WRITE(stats
);
2908 MB_PROC_VALUE_READ(max_to_scan
);
2909 MB_PROC_VALUE_WRITE(max_to_scan
);
2910 MB_PROC_VALUE_READ(min_to_scan
);
2911 MB_PROC_VALUE_WRITE(min_to_scan
);
2912 MB_PROC_VALUE_READ(order2_reqs
);
2913 MB_PROC_VALUE_WRITE(order2_reqs
);
2914 MB_PROC_VALUE_READ(stream_request
);
2915 MB_PROC_VALUE_WRITE(stream_request
);
2916 MB_PROC_VALUE_READ(group_prealloc
);
2917 MB_PROC_VALUE_WRITE(group_prealloc
);
2919 #define MB_PROC_HANDLER(name, var) \
2921 proc = create_proc_entry(name, mode, sbi->s_mb_proc); \
2922 if (proc == NULL) { \
2923 printk(KERN_ERR "EXT4-fs: can't to create %s\n", name); \
2927 proc->read_proc = ext4_mb_read_##var ; \
2928 proc->write_proc = ext4_mb_write_##var; \
2931 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
)
2933 mode_t mode
= S_IFREG
| S_IRUGO
| S_IWUSR
;
2934 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2935 struct proc_dir_entry
*proc
;
2938 snprintf(devname
, sizeof(devname
) - 1, "%s",
2939 bdevname(sb
->s_bdev
, devname
));
2940 sbi
->s_mb_proc
= proc_mkdir(devname
, proc_root_ext4
);
2942 MB_PROC_HANDLER(EXT4_MB_STATS_NAME
, stats
);
2943 MB_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME
, max_to_scan
);
2944 MB_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME
, min_to_scan
);
2945 MB_PROC_HANDLER(EXT4_MB_ORDER2_REQ
, order2_reqs
);
2946 MB_PROC_HANDLER(EXT4_MB_STREAM_REQ
, stream_request
);
2947 MB_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC
, group_prealloc
);
2952 printk(KERN_ERR
"EXT4-fs: Unable to create %s\n", devname
);
2953 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_mb_proc
);
2954 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_mb_proc
);
2955 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_mb_proc
);
2956 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_mb_proc
);
2957 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_mb_proc
);
2958 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_mb_proc
);
2959 remove_proc_entry(devname
, proc_root_ext4
);
2960 sbi
->s_mb_proc
= NULL
;
2965 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
)
2967 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2970 if (sbi
->s_mb_proc
== NULL
)
2973 snprintf(devname
, sizeof(devname
) - 1, "%s",
2974 bdevname(sb
->s_bdev
, devname
));
2975 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_mb_proc
);
2976 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_mb_proc
);
2977 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_mb_proc
);
2978 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_mb_proc
);
2979 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_mb_proc
);
2980 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_mb_proc
);
2981 remove_proc_entry(devname
, proc_root_ext4
);
2986 int __init
init_ext4_mballoc(void)
2988 ext4_pspace_cachep
=
2989 kmem_cache_create("ext4_prealloc_space",
2990 sizeof(struct ext4_prealloc_space
),
2991 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2992 if (ext4_pspace_cachep
== NULL
)
2996 kmem_cache_create("ext4_alloc_context",
2997 sizeof(struct ext4_allocation_context
),
2998 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2999 if (ext4_ac_cachep
== NULL
) {
3000 kmem_cache_destroy(ext4_pspace_cachep
);
3003 #ifdef CONFIG_PROC_FS
3004 proc_root_ext4
= proc_mkdir("fs/ext4", NULL
);
3005 if (proc_root_ext4
== NULL
)
3006 printk(KERN_ERR
"EXT4-fs: Unable to create fs/ext4\n");
3011 void exit_ext4_mballoc(void)
3013 /* XXX: synchronize_rcu(); */
3014 kmem_cache_destroy(ext4_pspace_cachep
);
3015 kmem_cache_destroy(ext4_ac_cachep
);
3016 #ifdef CONFIG_PROC_FS
3017 remove_proc_entry("fs/ext4", NULL
);
3023 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
3024 * Returns 0 if success or error code
3026 static noinline_for_stack
int
3027 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
3030 struct buffer_head
*bitmap_bh
= NULL
;
3031 struct ext4_super_block
*es
;
3032 struct ext4_group_desc
*gdp
;
3033 struct buffer_head
*gdp_bh
;
3034 struct ext4_sb_info
*sbi
;
3035 struct super_block
*sb
;
3039 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3040 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
3046 ext4_debug("using block group %lu(%d)\n", ac
->ac_b_ex
.fe_group
,
3047 gdp
->bg_free_blocks_count
);
3050 bitmap_bh
= read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
3054 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
3059 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
3063 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
3067 block
= ac
->ac_b_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
3068 + ac
->ac_b_ex
.fe_start
3069 + le32_to_cpu(es
->s_first_data_block
);
3071 if (block
== ext4_block_bitmap(sb
, gdp
) ||
3072 block
== ext4_inode_bitmap(sb
, gdp
) ||
3073 in_range(block
, ext4_inode_table(sb
, gdp
),
3074 EXT4_SB(sb
)->s_itb_per_group
)) {
3076 ext4_error(sb
, __func__
,
3077 "Allocating block in system zone - block = %llu",
3080 #ifdef AGGRESSIVE_CHECK
3083 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
3084 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
3085 bitmap_bh
->b_data
));
3089 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
), bitmap_bh
->b_data
,
3090 ac
->ac_b_ex
.fe_start
, ac
->ac_b_ex
.fe_len
);
3092 spin_lock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
3093 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
3094 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
3095 gdp
->bg_free_blocks_count
=
3096 cpu_to_le16(ext4_free_blocks_after_init(sb
,
3097 ac
->ac_b_ex
.fe_group
,
3100 le16_add_cpu(&gdp
->bg_free_blocks_count
, -ac
->ac_b_ex
.fe_len
);
3101 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
3102 spin_unlock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
3103 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
3105 err
= ext4_journal_dirty_metadata(handle
, bitmap_bh
);
3108 err
= ext4_journal_dirty_metadata(handle
, gdp_bh
);
3117 * here we normalize request for locality group
3118 * Group request are normalized to s_strip size if we set the same via mount
3119 * option. If not we set it to s_mb_group_prealloc which can be configured via
3120 * /proc/fs/ext4/<partition>/group_prealloc
3122 * XXX: should we try to preallocate more than the group has now?
3124 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
3126 struct super_block
*sb
= ac
->ac_sb
;
3127 struct ext4_locality_group
*lg
= ac
->ac_lg
;
3130 if (EXT4_SB(sb
)->s_stripe
)
3131 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
3133 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
3134 mb_debug("#%u: goal %lu blocks for locality group\n",
3135 current
->pid
, ac
->ac_g_ex
.fe_len
);
3139 * Normalization means making request better in terms of
3140 * size and alignment
3142 static noinline_for_stack
void
3143 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
3144 struct ext4_allocation_request
*ar
)
3148 loff_t size
, orig_size
, start_off
;
3149 ext4_lblk_t start
, orig_start
;
3150 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3151 struct ext4_prealloc_space
*pa
;
3153 /* do normalize only data requests, metadata requests
3154 do not need preallocation */
3155 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3158 /* sometime caller may want exact blocks */
3159 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3162 /* caller may indicate that preallocation isn't
3163 * required (it's a tail, for example) */
3164 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3167 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3168 ext4_mb_normalize_group_request(ac
);
3172 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3174 /* first, let's learn actual file size
3175 * given current request is allocated */
3176 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3177 size
= size
<< bsbits
;
3178 if (size
< i_size_read(ac
->ac_inode
))
3179 size
= i_size_read(ac
->ac_inode
);
3181 /* max available blocks in a free group */
3182 max
= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
) - 1 - 1 -
3183 EXT4_SB(ac
->ac_sb
)->s_itb_per_group
;
3185 #define NRL_CHECK_SIZE(req, size, max,bits) \
3186 (req <= (size) || max <= ((size) >> bits))
3188 /* first, try to predict filesize */
3189 /* XXX: should this table be tunable? */
3191 if (size
<= 16 * 1024) {
3193 } else if (size
<= 32 * 1024) {
3195 } else if (size
<= 64 * 1024) {
3197 } else if (size
<= 128 * 1024) {
3199 } else if (size
<= 256 * 1024) {
3201 } else if (size
<= 512 * 1024) {
3203 } else if (size
<= 1024 * 1024) {
3205 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, bsbits
)) {
3206 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3207 (20 - bsbits
)) << 20;
3209 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, bsbits
)) {
3210 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3211 (22 - bsbits
)) << 22;
3212 size
= 4 * 1024 * 1024;
3213 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3214 (8<<20)>>bsbits
, max
, bsbits
)) {
3215 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3216 (23 - bsbits
)) << 23;
3217 size
= 8 * 1024 * 1024;
3219 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3220 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3222 orig_size
= size
= size
>> bsbits
;
3223 orig_start
= start
= start_off
>> bsbits
;
3225 /* don't cover already allocated blocks in selected range */
3226 if (ar
->pleft
&& start
<= ar
->lleft
) {
3227 size
-= ar
->lleft
+ 1 - start
;
3228 start
= ar
->lleft
+ 1;
3230 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3231 size
-= start
+ size
- ar
->lright
;
3235 /* check we don't cross already preallocated blocks */
3237 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3238 unsigned long pa_end
;
3242 spin_lock(&pa
->pa_lock
);
3243 if (pa
->pa_deleted
) {
3244 spin_unlock(&pa
->pa_lock
);
3248 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3250 /* PA must not overlap original request */
3251 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3252 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3254 /* skip PA normalized request doesn't overlap with */
3255 if (pa
->pa_lstart
>= end
) {
3256 spin_unlock(&pa
->pa_lock
);
3259 if (pa_end
<= start
) {
3260 spin_unlock(&pa
->pa_lock
);
3263 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3265 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3266 BUG_ON(pa_end
< start
);
3270 if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3271 BUG_ON(pa
->pa_lstart
> end
);
3272 end
= pa
->pa_lstart
;
3274 spin_unlock(&pa
->pa_lock
);
3279 /* XXX: extra loop to check we really don't overlap preallocations */
3281 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3282 unsigned long pa_end
;
3283 spin_lock(&pa
->pa_lock
);
3284 if (pa
->pa_deleted
== 0) {
3285 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3286 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3288 spin_unlock(&pa
->pa_lock
);
3292 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3293 start
> ac
->ac_o_ex
.fe_logical
) {
3294 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3295 (unsigned long) start
, (unsigned long) size
,
3296 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3298 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3299 start
> ac
->ac_o_ex
.fe_logical
);
3300 BUG_ON(size
<= 0 || size
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3302 /* now prepare goal request */
3304 /* XXX: is it better to align blocks WRT to logical
3305 * placement or satisfy big request as is */
3306 ac
->ac_g_ex
.fe_logical
= start
;
3307 ac
->ac_g_ex
.fe_len
= size
;
3309 /* define goal start in order to merge */
3310 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3311 /* merge to the right */
3312 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3313 &ac
->ac_f_ex
.fe_group
,
3314 &ac
->ac_f_ex
.fe_start
);
3315 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3317 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3318 /* merge to the left */
3319 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3320 &ac
->ac_f_ex
.fe_group
,
3321 &ac
->ac_f_ex
.fe_start
);
3322 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3325 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3326 (unsigned) orig_size
, (unsigned) start
);
3329 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3331 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3333 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3334 atomic_inc(&sbi
->s_bal_reqs
);
3335 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3336 if (ac
->ac_o_ex
.fe_len
>= ac
->ac_g_ex
.fe_len
)
3337 atomic_inc(&sbi
->s_bal_success
);
3338 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3339 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3340 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3341 atomic_inc(&sbi
->s_bal_goals
);
3342 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3343 atomic_inc(&sbi
->s_bal_breaks
);
3346 ext4_mb_store_history(ac
);
3350 * use blocks preallocated to inode
3352 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3353 struct ext4_prealloc_space
*pa
)
3359 /* found preallocated blocks, use them */
3360 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3361 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3363 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3364 &ac
->ac_b_ex
.fe_start
);
3365 ac
->ac_b_ex
.fe_len
= len
;
3366 ac
->ac_status
= AC_STATUS_FOUND
;
3369 BUG_ON(start
< pa
->pa_pstart
);
3370 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3371 BUG_ON(pa
->pa_free
< len
);
3374 mb_debug("use %llu/%lu from inode pa %p\n", start
, len
, pa
);
3378 * use blocks preallocated to locality group
3380 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3381 struct ext4_prealloc_space
*pa
)
3383 unsigned len
= ac
->ac_o_ex
.fe_len
;
3385 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3386 &ac
->ac_b_ex
.fe_group
,
3387 &ac
->ac_b_ex
.fe_start
);
3388 ac
->ac_b_ex
.fe_len
= len
;
3389 ac
->ac_status
= AC_STATUS_FOUND
;
3392 /* we don't correct pa_pstart or pa_plen here to avoid
3393 * possible race when the group is being loaded concurrently
3394 * instead we correct pa later, after blocks are marked
3395 * in on-disk bitmap -- see ext4_mb_release_context()
3396 * Other CPUs are prevented from allocating from this pa by lg_mutex
3398 mb_debug("use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3402 * search goal blocks in preallocated space
3404 static noinline_for_stack
int
3405 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3407 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3408 struct ext4_locality_group
*lg
;
3409 struct ext4_prealloc_space
*pa
;
3411 /* only data can be preallocated */
3412 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3415 /* first, try per-file preallocation */
3417 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3419 /* all fields in this condition don't change,
3420 * so we can skip locking for them */
3421 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3422 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3425 /* found preallocated blocks, use them */
3426 spin_lock(&pa
->pa_lock
);
3427 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3428 atomic_inc(&pa
->pa_count
);
3429 ext4_mb_use_inode_pa(ac
, pa
);
3430 spin_unlock(&pa
->pa_lock
);
3431 ac
->ac_criteria
= 10;
3435 spin_unlock(&pa
->pa_lock
);
3439 /* can we use group allocation? */
3440 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3443 /* inode may have no locality group for some reason */
3449 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
, pa_inode_list
) {
3450 spin_lock(&pa
->pa_lock
);
3451 if (pa
->pa_deleted
== 0 && pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3452 atomic_inc(&pa
->pa_count
);
3453 ext4_mb_use_group_pa(ac
, pa
);
3454 spin_unlock(&pa
->pa_lock
);
3455 ac
->ac_criteria
= 20;
3459 spin_unlock(&pa
->pa_lock
);
3467 * the function goes through all preallocation in this group and marks them
3468 * used in in-core bitmap. buddy must be generated from this bitmap
3469 * Need to be called with ext4 group lock (ext4_lock_group)
3471 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3474 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3475 struct ext4_prealloc_space
*pa
;
3476 struct list_head
*cur
;
3477 ext4_group_t groupnr
;
3478 ext4_grpblk_t start
;
3479 int preallocated
= 0;
3483 /* all form of preallocation discards first load group,
3484 * so the only competing code is preallocation use.
3485 * we don't need any locking here
3486 * notice we do NOT ignore preallocations with pa_deleted
3487 * otherwise we could leave used blocks available for
3488 * allocation in buddy when concurrent ext4_mb_put_pa()
3489 * is dropping preallocation
3491 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3492 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3493 spin_lock(&pa
->pa_lock
);
3494 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3497 spin_unlock(&pa
->pa_lock
);
3498 if (unlikely(len
== 0))
3500 BUG_ON(groupnr
!= group
);
3501 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3502 bitmap
, start
, len
);
3503 preallocated
+= len
;
3506 mb_debug("prellocated %u for group %lu\n", preallocated
, group
);
3509 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3511 struct ext4_prealloc_space
*pa
;
3512 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3513 kmem_cache_free(ext4_pspace_cachep
, pa
);
3517 * drops a reference to preallocated space descriptor
3518 * if this was the last reference and the space is consumed
3520 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3521 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3525 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3528 /* in this short window concurrent discard can set pa_deleted */
3529 spin_lock(&pa
->pa_lock
);
3530 if (pa
->pa_deleted
== 1) {
3531 spin_unlock(&pa
->pa_lock
);
3536 spin_unlock(&pa
->pa_lock
);
3538 /* -1 is to protect from crossing allocation group */
3539 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
- 1, &grp
, NULL
);
3544 * P1 (buddy init) P2 (regular allocation)
3545 * find block B in PA
3546 * copy on-disk bitmap to buddy
3547 * mark B in on-disk bitmap
3548 * drop PA from group
3549 * mark all PAs in buddy
3551 * thus, P1 initializes buddy with B available. to prevent this
3552 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3555 ext4_lock_group(sb
, grp
);
3556 list_del(&pa
->pa_group_list
);
3557 ext4_unlock_group(sb
, grp
);
3559 spin_lock(pa
->pa_obj_lock
);
3560 list_del_rcu(&pa
->pa_inode_list
);
3561 spin_unlock(pa
->pa_obj_lock
);
3563 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3567 * creates new preallocated space for given inode
3569 static noinline_for_stack
int
3570 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3572 struct super_block
*sb
= ac
->ac_sb
;
3573 struct ext4_prealloc_space
*pa
;
3574 struct ext4_group_info
*grp
;
3575 struct ext4_inode_info
*ei
;
3577 /* preallocate only when found space is larger then requested */
3578 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3579 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3580 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3582 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3586 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3592 /* we can't allocate as much as normalizer wants.
3593 * so, found space must get proper lstart
3594 * to cover original request */
3595 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3596 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3598 /* we're limited by original request in that
3599 * logical block must be covered any way
3600 * winl is window we can move our chunk within */
3601 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3603 /* also, we should cover whole original request */
3604 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3606 /* the smallest one defines real window */
3607 win
= min(winl
, wins
);
3609 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3610 if (offs
&& offs
< win
)
3613 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3614 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3615 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3618 /* preallocation can change ac_b_ex, thus we store actually
3619 * allocated blocks for history */
3620 ac
->ac_f_ex
= ac
->ac_b_ex
;
3622 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3623 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3624 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3625 pa
->pa_free
= pa
->pa_len
;
3626 atomic_set(&pa
->pa_count
, 1);
3627 spin_lock_init(&pa
->pa_lock
);
3631 mb_debug("new inode pa %p: %llu/%u for %u\n", pa
,
3632 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3634 ext4_mb_use_inode_pa(ac
, pa
);
3635 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3637 ei
= EXT4_I(ac
->ac_inode
);
3638 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3640 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3641 pa
->pa_inode
= ac
->ac_inode
;
3643 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3644 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3645 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3647 spin_lock(pa
->pa_obj_lock
);
3648 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3649 spin_unlock(pa
->pa_obj_lock
);
3655 * creates new preallocated space for locality group inodes belongs to
3657 static noinline_for_stack
int
3658 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3660 struct super_block
*sb
= ac
->ac_sb
;
3661 struct ext4_locality_group
*lg
;
3662 struct ext4_prealloc_space
*pa
;
3663 struct ext4_group_info
*grp
;
3665 /* preallocate only when found space is larger then requested */
3666 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3667 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3668 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3670 BUG_ON(ext4_pspace_cachep
== NULL
);
3671 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3675 /* preallocation can change ac_b_ex, thus we store actually
3676 * allocated blocks for history */
3677 ac
->ac_f_ex
= ac
->ac_b_ex
;
3679 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3680 pa
->pa_lstart
= pa
->pa_pstart
;
3681 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3682 pa
->pa_free
= pa
->pa_len
;
3683 atomic_set(&pa
->pa_count
, 1);
3684 spin_lock_init(&pa
->pa_lock
);
3688 mb_debug("new group pa %p: %llu/%u for %u\n", pa
,
3689 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3691 ext4_mb_use_group_pa(ac
, pa
);
3692 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3694 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3698 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3699 pa
->pa_inode
= NULL
;
3701 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3702 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3703 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3705 spin_lock(pa
->pa_obj_lock
);
3706 list_add_tail_rcu(&pa
->pa_inode_list
, &lg
->lg_prealloc_list
);
3707 spin_unlock(pa
->pa_obj_lock
);
3712 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3716 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3717 err
= ext4_mb_new_group_pa(ac
);
3719 err
= ext4_mb_new_inode_pa(ac
);
3724 * finds all unused blocks in on-disk bitmap, frees them in
3725 * in-core bitmap and buddy.
3726 * @pa must be unlinked from inode and group lists, so that
3727 * nobody else can find/use it.
3728 * the caller MUST hold group/inode locks.
3729 * TODO: optimize the case when there are no in-core structures yet
3731 static noinline_for_stack
int
3732 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3733 struct ext4_prealloc_space
*pa
,
3734 struct ext4_allocation_context
*ac
)
3736 struct super_block
*sb
= e4b
->bd_sb
;
3737 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3746 BUG_ON(pa
->pa_deleted
== 0);
3747 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3748 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3749 end
= bit
+ pa
->pa_len
;
3753 ac
->ac_inode
= pa
->pa_inode
;
3754 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3758 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3761 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3764 start
= group
* EXT4_BLOCKS_PER_GROUP(sb
) + bit
+
3765 le32_to_cpu(sbi
->s_es
->s_first_data_block
);
3766 mb_debug(" free preallocated %u/%u in group %u\n",
3767 (unsigned) start
, (unsigned) next
- bit
,
3772 ac
->ac_b_ex
.fe_group
= group
;
3773 ac
->ac_b_ex
.fe_start
= bit
;
3774 ac
->ac_b_ex
.fe_len
= next
- bit
;
3775 ac
->ac_b_ex
.fe_logical
= 0;
3776 ext4_mb_store_history(ac
);
3779 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3782 if (free
!= pa
->pa_free
) {
3783 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3784 pa
, (unsigned long) pa
->pa_lstart
,
3785 (unsigned long) pa
->pa_pstart
,
3786 (unsigned long) pa
->pa_len
);
3787 ext4_error(sb
, __func__
, "free %u, pa_free %u\n",
3790 * pa is already deleted so we use the value obtained
3791 * from the bitmap and continue.
3794 atomic_add(free
, &sbi
->s_mb_discarded
);
3799 static noinline_for_stack
int
3800 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3801 struct ext4_prealloc_space
*pa
,
3802 struct ext4_allocation_context
*ac
)
3804 struct super_block
*sb
= e4b
->bd_sb
;
3809 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3811 BUG_ON(pa
->pa_deleted
== 0);
3812 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3813 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3814 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3815 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3819 ac
->ac_inode
= NULL
;
3820 ac
->ac_b_ex
.fe_group
= group
;
3821 ac
->ac_b_ex
.fe_start
= bit
;
3822 ac
->ac_b_ex
.fe_len
= pa
->pa_len
;
3823 ac
->ac_b_ex
.fe_logical
= 0;
3824 ext4_mb_store_history(ac
);
3831 * releases all preallocations in given group
3833 * first, we need to decide discard policy:
3834 * - when do we discard
3836 * - how many do we discard
3837 * 1) how many requested
3839 static noinline_for_stack
int
3840 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3841 ext4_group_t group
, int needed
)
3843 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3844 struct buffer_head
*bitmap_bh
= NULL
;
3845 struct ext4_prealloc_space
*pa
, *tmp
;
3846 struct ext4_allocation_context
*ac
;
3847 struct list_head list
;
3848 struct ext4_buddy e4b
;
3853 mb_debug("discard preallocation for group %lu\n", group
);
3855 if (list_empty(&grp
->bb_prealloc_list
))
3858 bitmap_bh
= read_block_bitmap(sb
, group
);
3859 if (bitmap_bh
== NULL
) {
3860 /* error handling here */
3861 ext4_mb_release_desc(&e4b
);
3862 BUG_ON(bitmap_bh
== NULL
);
3865 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3866 BUG_ON(err
!= 0); /* error handling here */
3869 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3871 grp
= ext4_get_group_info(sb
, group
);
3872 INIT_LIST_HEAD(&list
);
3874 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3876 ext4_lock_group(sb
, group
);
3877 list_for_each_entry_safe(pa
, tmp
,
3878 &grp
->bb_prealloc_list
, pa_group_list
) {
3879 spin_lock(&pa
->pa_lock
);
3880 if (atomic_read(&pa
->pa_count
)) {
3881 spin_unlock(&pa
->pa_lock
);
3885 if (pa
->pa_deleted
) {
3886 spin_unlock(&pa
->pa_lock
);
3890 /* seems this one can be freed ... */
3893 /* we can trust pa_free ... */
3894 free
+= pa
->pa_free
;
3896 spin_unlock(&pa
->pa_lock
);
3898 list_del(&pa
->pa_group_list
);
3899 list_add(&pa
->u
.pa_tmp_list
, &list
);
3902 /* if we still need more blocks and some PAs were used, try again */
3903 if (free
< needed
&& busy
) {
3905 ext4_unlock_group(sb
, group
);
3907 * Yield the CPU here so that we don't get soft lockup
3908 * in non preempt case.
3914 /* found anything to free? */
3915 if (list_empty(&list
)) {
3920 /* now free all selected PAs */
3921 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3923 /* remove from object (inode or locality group) */
3924 spin_lock(pa
->pa_obj_lock
);
3925 list_del_rcu(&pa
->pa_inode_list
);
3926 spin_unlock(pa
->pa_obj_lock
);
3929 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
3931 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3933 list_del(&pa
->u
.pa_tmp_list
);
3934 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3938 ext4_unlock_group(sb
, group
);
3940 kmem_cache_free(ext4_ac_cachep
, ac
);
3941 ext4_mb_release_desc(&e4b
);
3947 * releases all non-used preallocated blocks for given inode
3949 * It's important to discard preallocations under i_data_sem
3950 * We don't want another block to be served from the prealloc
3951 * space when we are discarding the inode prealloc space.
3953 * FIXME!! Make sure it is valid at all the call sites
3955 void ext4_mb_discard_inode_preallocations(struct inode
*inode
)
3957 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3958 struct super_block
*sb
= inode
->i_sb
;
3959 struct buffer_head
*bitmap_bh
= NULL
;
3960 struct ext4_prealloc_space
*pa
, *tmp
;
3961 struct ext4_allocation_context
*ac
;
3962 ext4_group_t group
= 0;
3963 struct list_head list
;
3964 struct ext4_buddy e4b
;
3967 if (!test_opt(sb
, MBALLOC
) || !S_ISREG(inode
->i_mode
)) {
3968 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3972 mb_debug("discard preallocation for inode %lu\n", inode
->i_ino
);
3974 INIT_LIST_HEAD(&list
);
3976 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3978 /* first, collect all pa's in the inode */
3979 spin_lock(&ei
->i_prealloc_lock
);
3980 while (!list_empty(&ei
->i_prealloc_list
)) {
3981 pa
= list_entry(ei
->i_prealloc_list
.next
,
3982 struct ext4_prealloc_space
, pa_inode_list
);
3983 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
3984 spin_lock(&pa
->pa_lock
);
3985 if (atomic_read(&pa
->pa_count
)) {
3986 /* this shouldn't happen often - nobody should
3987 * use preallocation while we're discarding it */
3988 spin_unlock(&pa
->pa_lock
);
3989 spin_unlock(&ei
->i_prealloc_lock
);
3990 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
3992 schedule_timeout_uninterruptible(HZ
);
3996 if (pa
->pa_deleted
== 0) {
3998 spin_unlock(&pa
->pa_lock
);
3999 list_del_rcu(&pa
->pa_inode_list
);
4000 list_add(&pa
->u
.pa_tmp_list
, &list
);
4004 /* someone is deleting pa right now */
4005 spin_unlock(&pa
->pa_lock
);
4006 spin_unlock(&ei
->i_prealloc_lock
);
4008 /* we have to wait here because pa_deleted
4009 * doesn't mean pa is already unlinked from
4010 * the list. as we might be called from
4011 * ->clear_inode() the inode will get freed
4012 * and concurrent thread which is unlinking
4013 * pa from inode's list may access already
4014 * freed memory, bad-bad-bad */
4016 /* XXX: if this happens too often, we can
4017 * add a flag to force wait only in case
4018 * of ->clear_inode(), but not in case of
4019 * regular truncate */
4020 schedule_timeout_uninterruptible(HZ
);
4023 spin_unlock(&ei
->i_prealloc_lock
);
4025 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
4026 BUG_ON(pa
->pa_linear
!= 0);
4027 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4029 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
4030 BUG_ON(err
!= 0); /* error handling here */
4032 bitmap_bh
= read_block_bitmap(sb
, group
);
4033 if (bitmap_bh
== NULL
) {
4034 /* error handling here */
4035 ext4_mb_release_desc(&e4b
);
4036 BUG_ON(bitmap_bh
== NULL
);
4039 ext4_lock_group(sb
, group
);
4040 list_del(&pa
->pa_group_list
);
4041 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
4042 ext4_unlock_group(sb
, group
);
4044 ext4_mb_release_desc(&e4b
);
4047 list_del(&pa
->u
.pa_tmp_list
);
4048 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4051 kmem_cache_free(ext4_ac_cachep
, ac
);
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 noinline_for_stack
int
4171 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4172 struct ext4_allocation_request
*ar
)
4174 struct super_block
*sb
= ar
->inode
->i_sb
;
4175 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4176 struct ext4_super_block
*es
= sbi
->s_es
;
4180 ext4_grpblk_t block
;
4182 /* we can't allocate > group size */
4185 /* just a dirty hack to filter too big requests */
4186 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4187 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4189 /* start searching from the goal */
4191 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4192 goal
>= ext4_blocks_count(es
))
4193 goal
= le32_to_cpu(es
->s_first_data_block
);
4194 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4196 /* set up allocation goals */
4197 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4198 ac
->ac_b_ex
.fe_group
= 0;
4199 ac
->ac_b_ex
.fe_start
= 0;
4200 ac
->ac_b_ex
.fe_len
= 0;
4201 ac
->ac_status
= AC_STATUS_CONTINUE
;
4202 ac
->ac_groups_scanned
= 0;
4203 ac
->ac_ex_scanned
= 0;
4206 ac
->ac_inode
= ar
->inode
;
4207 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4208 ac
->ac_o_ex
.fe_group
= group
;
4209 ac
->ac_o_ex
.fe_start
= block
;
4210 ac
->ac_o_ex
.fe_len
= len
;
4211 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4212 ac
->ac_g_ex
.fe_group
= group
;
4213 ac
->ac_g_ex
.fe_start
= block
;
4214 ac
->ac_g_ex
.fe_len
= len
;
4215 ac
->ac_f_ex
.fe_len
= 0;
4216 ac
->ac_flags
= ar
->flags
;
4218 ac
->ac_criteria
= 0;
4220 ac
->ac_bitmap_page
= NULL
;
4221 ac
->ac_buddy_page
= NULL
;
4224 /* we have to define context: we'll we work with a file or
4225 * locality group. this is a policy, actually */
4226 ext4_mb_group_or_file(ac
);
4228 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4229 "left: %u/%u, right %u/%u to %swritable\n",
4230 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4231 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4232 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4233 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4234 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4240 * release all resource we used in allocation
4242 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4245 if (ac
->ac_pa
->pa_linear
) {
4246 /* see comment in ext4_mb_use_group_pa() */
4247 spin_lock(&ac
->ac_pa
->pa_lock
);
4248 ac
->ac_pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4249 ac
->ac_pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4250 ac
->ac_pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4251 ac
->ac_pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4252 spin_unlock(&ac
->ac_pa
->pa_lock
);
4254 ext4_mb_put_pa(ac
, ac
->ac_sb
, ac
->ac_pa
);
4256 if (ac
->ac_bitmap_page
)
4257 page_cache_release(ac
->ac_bitmap_page
);
4258 if (ac
->ac_buddy_page
)
4259 page_cache_release(ac
->ac_buddy_page
);
4260 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4261 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4262 ext4_mb_collect_stats(ac
);
4266 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4272 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
&& needed
> 0; i
++) {
4273 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4282 * Main entry point into mballoc to allocate blocks
4283 * it tries to use preallocation first, then falls back
4284 * to usual allocation
4286 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4287 struct ext4_allocation_request
*ar
, int *errp
)
4289 struct ext4_allocation_context
*ac
= NULL
;
4290 struct ext4_sb_info
*sbi
;
4291 struct super_block
*sb
;
4292 ext4_fsblk_t block
= 0;
4296 sb
= ar
->inode
->i_sb
;
4299 if (!test_opt(sb
, MBALLOC
)) {
4300 block
= ext4_new_blocks_old(handle
, ar
->inode
, ar
->goal
,
4305 while (ar
->len
&& DQUOT_ALLOC_BLOCK(ar
->inode
, ar
->len
)) {
4306 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4315 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4321 ext4_mb_poll_new_transaction(sb
, handle
);
4323 *errp
= ext4_mb_initialize_context(ac
, ar
);
4329 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4330 if (!ext4_mb_use_preallocated(ac
)) {
4332 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4333 ext4_mb_normalize_request(ac
, ar
);
4336 /* allocate space in core */
4337 ext4_mb_regular_allocator(ac
);
4339 /* as we've just preallocated more space than
4340 * user requested orinally, we store allocated
4341 * space in a special descriptor */
4342 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4343 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4344 ext4_mb_new_preallocation(ac
);
4347 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4348 ext4_mb_mark_diskspace_used(ac
, handle
);
4350 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4351 ar
->len
= ac
->ac_b_ex
.fe_len
;
4353 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4357 ac
->ac_b_ex
.fe_len
= 0;
4359 ext4_mb_show_ac(ac
);
4362 ext4_mb_release_context(ac
);
4365 if (ar
->len
< inquota
)
4366 DQUOT_FREE_BLOCK(ar
->inode
, inquota
- ar
->len
);
4368 kmem_cache_free(ext4_ac_cachep
, ac
);
4371 static void ext4_mb_poll_new_transaction(struct super_block
*sb
,
4374 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4376 if (sbi
->s_last_transaction
== handle
->h_transaction
->t_tid
)
4379 /* new transaction! time to close last one and free blocks for
4380 * committed transaction. we know that only transaction can be
4381 * active, so previos transaction can be being logged and we
4382 * know that transaction before previous is known to be already
4383 * logged. this means that now we may free blocks freed in all
4384 * transactions before previous one. hope I'm clear enough ... */
4386 spin_lock(&sbi
->s_md_lock
);
4387 if (sbi
->s_last_transaction
!= handle
->h_transaction
->t_tid
) {
4388 mb_debug("new transaction %lu, old %lu\n",
4389 (unsigned long) handle
->h_transaction
->t_tid
,
4390 (unsigned long) sbi
->s_last_transaction
);
4391 list_splice_init(&sbi
->s_closed_transaction
,
4392 &sbi
->s_committed_transaction
);
4393 list_splice_init(&sbi
->s_active_transaction
,
4394 &sbi
->s_closed_transaction
);
4395 sbi
->s_last_transaction
= handle
->h_transaction
->t_tid
;
4397 spin_unlock(&sbi
->s_md_lock
);
4399 ext4_mb_free_committed_blocks(sb
);
4402 static noinline_for_stack
int
4403 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4404 ext4_group_t group
, ext4_grpblk_t block
, int count
)
4406 struct ext4_group_info
*db
= e4b
->bd_info
;
4407 struct super_block
*sb
= e4b
->bd_sb
;
4408 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4409 struct ext4_free_metadata
*md
;
4412 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4413 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4415 ext4_lock_group(sb
, group
);
4416 for (i
= 0; i
< count
; i
++) {
4418 if (md
&& db
->bb_tid
!= handle
->h_transaction
->t_tid
) {
4419 db
->bb_md_cur
= NULL
;
4424 ext4_unlock_group(sb
, group
);
4425 md
= kmalloc(sizeof(*md
), GFP_NOFS
);
4431 ext4_lock_group(sb
, group
);
4432 if (db
->bb_md_cur
== NULL
) {
4433 spin_lock(&sbi
->s_md_lock
);
4434 list_add(&md
->list
, &sbi
->s_active_transaction
);
4435 spin_unlock(&sbi
->s_md_lock
);
4436 /* protect buddy cache from being freed,
4437 * otherwise we'll refresh it from
4438 * on-disk bitmap and lose not-yet-available
4440 page_cache_get(e4b
->bd_buddy_page
);
4441 page_cache_get(e4b
->bd_bitmap_page
);
4443 db
->bb_tid
= handle
->h_transaction
->t_tid
;
4444 mb_debug("new md 0x%p for group %lu\n",
4452 BUG_ON(md
->num
>= EXT4_BB_MAX_BLOCKS
);
4453 md
->blocks
[md
->num
] = block
+ i
;
4455 if (md
->num
== EXT4_BB_MAX_BLOCKS
) {
4456 /* no more space, put full container on a sb's list */
4457 db
->bb_md_cur
= NULL
;
4460 ext4_unlock_group(sb
, group
);
4465 * Main entry point into mballoc to free blocks
4467 void ext4_mb_free_blocks(handle_t
*handle
, struct inode
*inode
,
4468 unsigned long block
, unsigned long count
,
4469 int metadata
, unsigned long *freed
)
4471 struct buffer_head
*bitmap_bh
= NULL
;
4472 struct super_block
*sb
= inode
->i_sb
;
4473 struct ext4_allocation_context
*ac
= NULL
;
4474 struct ext4_group_desc
*gdp
;
4475 struct ext4_super_block
*es
;
4476 unsigned long overflow
;
4478 struct buffer_head
*gd_bh
;
4479 ext4_group_t block_group
;
4480 struct ext4_sb_info
*sbi
;
4481 struct ext4_buddy e4b
;
4487 ext4_mb_poll_new_transaction(sb
, handle
);
4490 es
= EXT4_SB(sb
)->s_es
;
4491 if (block
< le32_to_cpu(es
->s_first_data_block
) ||
4492 block
+ count
< block
||
4493 block
+ count
> ext4_blocks_count(es
)) {
4494 ext4_error(sb
, __func__
,
4495 "Freeing blocks not in datazone - "
4496 "block = %lu, count = %lu", block
, count
);
4500 ext4_debug("freeing block %lu\n", block
);
4502 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4504 ac
->ac_op
= EXT4_MB_HISTORY_FREE
;
4505 ac
->ac_inode
= inode
;
4511 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4514 * Check to see if we are freeing blocks across a group
4517 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4518 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4521 bitmap_bh
= read_block_bitmap(sb
, block_group
);
4524 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4528 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4529 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4530 in_range(block
, ext4_inode_table(sb
, gdp
),
4531 EXT4_SB(sb
)->s_itb_per_group
) ||
4532 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4533 EXT4_SB(sb
)->s_itb_per_group
)) {
4535 ext4_error(sb
, __func__
,
4536 "Freeing blocks in system zone - "
4537 "Block = %lu, count = %lu", block
, count
);
4540 BUFFER_TRACE(bitmap_bh
, "getting write access");
4541 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4546 * We are about to modify some metadata. Call the journal APIs
4547 * to unshare ->b_data if a currently-committing transaction is
4550 BUFFER_TRACE(gd_bh
, "get_write_access");
4551 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4555 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4559 #ifdef AGGRESSIVE_CHECK
4562 for (i
= 0; i
< count
; i
++)
4563 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4566 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4569 /* We dirtied the bitmap block */
4570 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4571 err
= ext4_journal_dirty_metadata(handle
, bitmap_bh
);
4574 ac
->ac_b_ex
.fe_group
= block_group
;
4575 ac
->ac_b_ex
.fe_start
= bit
;
4576 ac
->ac_b_ex
.fe_len
= count
;
4577 ext4_mb_store_history(ac
);
4581 /* blocks being freed are metadata. these blocks shouldn't
4582 * be used until this transaction is committed */
4583 ext4_mb_free_metadata(handle
, &e4b
, block_group
, bit
, count
);
4585 ext4_lock_group(sb
, block_group
);
4586 err
= mb_free_blocks(inode
, &e4b
, bit
, count
);
4587 ext4_mb_return_to_preallocation(inode
, &e4b
, block
, count
);
4588 ext4_unlock_group(sb
, block_group
);
4592 spin_lock(sb_bgl_lock(sbi
, block_group
));
4593 le16_add_cpu(&gdp
->bg_free_blocks_count
, count
);
4594 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4595 spin_unlock(sb_bgl_lock(sbi
, block_group
));
4596 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4598 ext4_mb_release_desc(&e4b
);
4602 /* And the group descriptor block */
4603 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4604 ret
= ext4_journal_dirty_metadata(handle
, gd_bh
);
4608 if (overflow
&& !err
) {
4617 ext4_std_error(sb
, err
);
4619 kmem_cache_free(ext4_ac_cachep
, ac
);