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
25 #include <linux/debugfs.h>
26 #include <linux/slab.h>
27 #include <trace/events/ext4.h>
31 * - test ext4_ext_search_left() and ext4_ext_search_right()
32 * - search for metadata in few groups
35 * - normalization should take into account whether file is still open
36 * - discard preallocations if no free space left (policy?)
37 * - don't normalize tails
39 * - reservation for superuser
42 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
43 * - track min/max extents in each group for better group selection
44 * - mb_mark_used() may allocate chunk right after splitting buddy
45 * - tree of groups sorted by number of free blocks
50 * The allocation request involve request for multiple number of blocks
51 * near to the goal(block) value specified.
53 * During initialization phase of the allocator we decide to use the
54 * group preallocation or inode preallocation depending on the size of
55 * the file. The size of the file could be the resulting file size we
56 * would have after allocation, or the current file size, which ever
57 * is larger. If the size is less than sbi->s_mb_stream_request we
58 * select to use the group preallocation. The default value of
59 * s_mb_stream_request is 16 blocks. This can also be tuned via
60 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
61 * terms of number of blocks.
63 * The main motivation for having small file use group preallocation is to
64 * ensure that we have small files closer together on the disk.
66 * First stage the allocator looks at the inode prealloc list,
67 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
68 * spaces for this particular inode. The inode prealloc space is
71 * pa_lstart -> the logical start block for this prealloc space
72 * pa_pstart -> the physical start block for this prealloc space
73 * pa_len -> length for this prealloc space
74 * pa_free -> free space available in this prealloc space
76 * The inode preallocation space is used looking at the _logical_ start
77 * block. If only the logical file block falls within the range of prealloc
78 * space we will consume the particular prealloc space. This make sure that
79 * that the we have contiguous physical blocks representing the file blocks
81 * The important thing to be noted in case of inode prealloc space is that
82 * we don't modify the values associated to inode prealloc space except
85 * If we are not able to find blocks in the inode prealloc space and if we
86 * have the group allocation flag set then we look at the locality group
87 * prealloc space. These are per CPU prealloc list repreasented as
89 * ext4_sb_info.s_locality_groups[smp_processor_id()]
91 * The reason for having a per cpu locality group is to reduce the contention
92 * between CPUs. It is possible to get scheduled at this point.
94 * The locality group prealloc space is used looking at whether we have
95 * enough free space (pa_free) withing the prealloc space.
97 * If we can't allocate blocks via inode prealloc or/and locality group
98 * prealloc then we look at the buddy cache. The buddy cache is represented
99 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
100 * mapped to the buddy and bitmap information regarding different
101 * groups. The buddy information is attached to buddy cache inode so that
102 * we can access them through the page cache. The information regarding
103 * each group is loaded via ext4_mb_load_buddy. The information involve
104 * block bitmap and buddy information. The information are stored in the
108 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
111 * one block each for bitmap and buddy information. So for each group we
112 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
113 * blocksize) blocks. So it can have information regarding groups_per_page
114 * which is blocks_per_page/2
116 * The buddy cache inode is not stored on disk. The inode is thrown
117 * away when the filesystem is unmounted.
119 * We look for count number of blocks in the buddy cache. If we were able
120 * to locate that many free blocks we return with additional information
121 * regarding rest of the contiguous physical block available
123 * Before allocating blocks via buddy cache we normalize the request
124 * blocks. This ensure we ask for more blocks that we needed. The extra
125 * blocks that we get after allocation is added to the respective prealloc
126 * list. In case of inode preallocation we follow a list of heuristics
127 * based on file size. This can be found in ext4_mb_normalize_request. If
128 * we are doing a group prealloc we try to normalize the request to
129 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
130 * 512 blocks. This can be tuned via
131 * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
132 * terms of number of blocks. If we have mounted the file system with -O
133 * stripe=<value> option the group prealloc request is normalized to the
134 * stripe value (sbi->s_stripe)
136 * The regular allocator(using the buddy cache) supports few tunables.
138 * /sys/fs/ext4/<partition>/mb_min_to_scan
139 * /sys/fs/ext4/<partition>/mb_max_to_scan
140 * /sys/fs/ext4/<partition>/mb_order2_req
142 * The regular allocator uses buddy scan only if the request len is power of
143 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
144 * value of s_mb_order2_reqs can be tuned via
145 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
146 * stripe size (sbi->s_stripe), we try to search for contiguous block in
147 * stripe size. This should result in better allocation on RAID setups. If
148 * not, we search in the specific group using bitmap for best extents. The
149 * tunable min_to_scan and max_to_scan control the behaviour here.
150 * min_to_scan indicate how long the mballoc __must__ look for a best
151 * extent and max_to_scan indicates how long the mballoc __can__ look for a
152 * best extent in the found extents. Searching for the blocks starts with
153 * the group specified as the goal value in allocation context via
154 * ac_g_ex. Each group is first checked based on the criteria whether it
155 * can used for allocation. ext4_mb_good_group explains how the groups are
158 * Both the prealloc space are getting populated as above. So for the first
159 * request we will hit the buddy cache which will result in this prealloc
160 * space getting filled. The prealloc space is then later used for the
161 * subsequent request.
165 * mballoc operates on the following data:
167 * - in-core buddy (actually includes buddy and bitmap)
168 * - preallocation descriptors (PAs)
170 * there are two types of preallocations:
172 * assiged to specific inode and can be used for this inode only.
173 * it describes part of inode's space preallocated to specific
174 * physical blocks. any block from that preallocated can be used
175 * independent. the descriptor just tracks number of blocks left
176 * unused. so, before taking some block from descriptor, one must
177 * make sure corresponded logical block isn't allocated yet. this
178 * also means that freeing any block within descriptor's range
179 * must discard all preallocated blocks.
181 * assigned to specific locality group which does not translate to
182 * permanent set of inodes: inode can join and leave group. space
183 * from this type of preallocation can be used for any inode. thus
184 * it's consumed from the beginning to the end.
186 * relation between them can be expressed as:
187 * in-core buddy = on-disk bitmap + preallocation descriptors
189 * this mean blocks mballoc considers used are:
190 * - allocated blocks (persistent)
191 * - preallocated blocks (non-persistent)
193 * consistency in mballoc world means that at any time a block is either
194 * free or used in ALL structures. notice: "any time" should not be read
195 * literally -- time is discrete and delimited by locks.
197 * to keep it simple, we don't use block numbers, instead we count number of
198 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
200 * all operations can be expressed as:
201 * - init buddy: buddy = on-disk + PAs
202 * - new PA: buddy += N; PA = N
203 * - use inode PA: on-disk += N; PA -= N
204 * - discard inode PA buddy -= on-disk - PA; PA = 0
205 * - use locality group PA on-disk += N; PA -= N
206 * - discard locality group PA buddy -= PA; PA = 0
207 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
208 * is used in real operation because we can't know actual used
209 * bits from PA, only from on-disk bitmap
211 * if we follow this strict logic, then all operations above should be atomic.
212 * given some of them can block, we'd have to use something like semaphores
213 * killing performance on high-end SMP hardware. let's try to relax it using
214 * the following knowledge:
215 * 1) if buddy is referenced, it's already initialized
216 * 2) while block is used in buddy and the buddy is referenced,
217 * nobody can re-allocate that block
218 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
219 * bit set and PA claims same block, it's OK. IOW, one can set bit in
220 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
223 * so, now we're building a concurrency table:
226 * blocks for PA are allocated in the buddy, buddy must be referenced
227 * until PA is linked to allocation group to avoid concurrent buddy init
229 * we need to make sure that either on-disk bitmap or PA has uptodate data
230 * given (3) we care that PA-=N operation doesn't interfere with init
232 * the simplest way would be to have buddy initialized by the discard
233 * - use locality group PA
234 * again PA-=N must be serialized with init
235 * - discard locality group PA
236 * the simplest way would be to have buddy initialized by the discard
239 * i_data_sem serializes them
241 * discard process must wait until PA isn't used by another process
242 * - use locality group PA
243 * some mutex should serialize them
244 * - discard locality group PA
245 * discard process must wait until PA isn't used by another process
248 * i_data_sem or another mutex should serializes them
250 * discard process must wait until PA isn't used by another process
251 * - use locality group PA
252 * nothing wrong here -- they're different PAs covering different blocks
253 * - discard locality group PA
254 * discard process must wait until PA isn't used by another process
256 * now we're ready to make few consequences:
257 * - PA is referenced and while it is no discard is possible
258 * - PA is referenced until block isn't marked in on-disk bitmap
259 * - PA changes only after on-disk bitmap
260 * - discard must not compete with init. either init is done before
261 * any discard or they're serialized somehow
262 * - buddy init as sum of on-disk bitmap and PAs is done atomically
264 * a special case when we've used PA to emptiness. no need to modify buddy
265 * in this case, but we should care about concurrent init
270 * Logic in few words:
275 * mark bits in on-disk bitmap
278 * - use preallocation:
279 * find proper PA (per-inode or group)
281 * mark bits in on-disk bitmap
287 * mark bits in on-disk bitmap
290 * - discard preallocations in group:
292 * move them onto local list
293 * load on-disk bitmap
295 * remove PA from object (inode or locality group)
296 * mark free blocks in-core
298 * - discard inode's preallocations:
305 * - bitlock on a group (group)
306 * - object (inode/locality) (object)
317 * - release consumed pa:
322 * - generate in-core bitmap:
326 * - discard all for given object (inode, locality group):
331 * - discard all for given group:
338 static struct kmem_cache *ext4_pspace_cachep;
339 static struct kmem_cache *ext4_ac_cachep;
340 static struct kmem_cache *ext4_free_ext_cachep;
341 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
343 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
345 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
347 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
349 #if BITS_PER_LONG == 64
350 *bit += ((unsigned long) addr & 7UL) << 3;
351 addr = (void *) ((unsigned long) addr & ~7UL);
352 #elif BITS_PER_LONG == 32
353 *bit += ((unsigned long) addr & 3UL) << 3;
354 addr = (void *) ((unsigned long) addr & ~3UL);
356 #error "how many bits you are?!"
361 static inline int mb_test_bit(int bit, void *addr)
364 * ext4_test_bit on architecture like powerpc
365 * needs unsigned long aligned address
367 addr = mb_correct_addr_and_bit(&bit, addr);
368 return ext4_test_bit(bit, addr);
371 static inline void mb_set_bit(int bit, void *addr)
373 addr = mb_correct_addr_and_bit(&bit, addr);
374 ext4_set_bit(bit, addr);
377 static inline void mb_clear_bit(int bit, void *addr)
379 addr = mb_correct_addr_and_bit(&bit, addr);
380 ext4_clear_bit(bit, addr);
383 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
385 int fix = 0, ret, tmpmax;
386 addr = mb_correct_addr_and_bit(&fix, addr);
390 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
396 static inline int mb_find_next_bit(void *addr, int max, int start)
398 int fix = 0, ret, tmpmax;
399 addr = mb_correct_addr_and_bit(&fix, addr);
403 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
409 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
413 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
416 if (order > e4b->bd_blkbits + 1) {
421 /* at order 0 we see each particular block */
422 *max = 1 << (e4b->bd_blkbits + 3);
424 return EXT4_MB_BITMAP(e4b);
426 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
427 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
433 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
434 int first, int count)
437 struct super_block *sb = e4b->bd_sb;
439 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
441 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
442 for (i = 0; i < count; i++) {
443 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
444 ext4_fsblk_t blocknr;
446 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
447 blocknr += first + i;
448 ext4_grp_locked_error(sb, e4b->bd_group,
449 __func__, "double-free of inode"
450 " %lu's block %llu(bit %u in group %u)",
451 inode ? inode->i_ino : 0, blocknr,
452 first + i, e4b->bd_group);
454 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
458 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
462 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
464 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
465 for (i = 0; i < count; i++) {
466 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
467 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
471 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
473 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
474 unsigned char *b1, *b2;
476 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
477 b2 = (unsigned char *) bitmap;
478 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
479 if (b1[i] != b2[i]) {
480 printk(KERN_ERR "corruption in group %u "
481 "at byte %u(%u): %x in copy != %x "
482 "on disk/prealloc\n",
483 e4b->bd_group, i, i * 8, b1[i], b2[i]);
491 static inline void mb_free_blocks_double(struct inode *inode,
492 struct ext4_buddy *e4b, int first, int count)
496 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
497 int first, int count)
501 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
507 #ifdef AGGRESSIVE_CHECK
509 #define MB_CHECK_ASSERT(assert) \
513 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
514 function, file, line, # assert); \
519 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
520 const char *function, int line)
522 struct super_block *sb = e4b->bd_sb;
523 int order = e4b->bd_blkbits + 1;
530 struct ext4_group_info *grp;
533 struct list_head *cur;
538 static int mb_check_counter;
539 if (mb_check_counter++ % 100 != 0)
544 buddy = mb_find_buddy(e4b, order, &max);
545 MB_CHECK_ASSERT(buddy);
546 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
547 MB_CHECK_ASSERT(buddy2);
548 MB_CHECK_ASSERT(buddy != buddy2);
549 MB_CHECK_ASSERT(max * 2 == max2);
552 for (i = 0; i < max; i++) {
554 if (mb_test_bit(i, buddy)) {
555 /* only single bit in buddy2 may be 1 */
556 if (!mb_test_bit(i << 1, buddy2)) {
558 mb_test_bit((i<<1)+1, buddy2));
559 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
561 mb_test_bit(i << 1, buddy2));
566 /* both bits in buddy2 must be 0 */
567 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
568 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
570 for (j = 0; j < (1 << order); j++) {
571 k = (i * (1 << order)) + j;
573 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
577 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
582 buddy = mb_find_buddy(e4b, 0, &max);
583 for (i = 0; i < max; i++) {
584 if (!mb_test_bit(i, buddy)) {
585 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
593 /* check used bits only */
594 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
595 buddy2 = mb_find_buddy(e4b, j, &max2);
597 MB_CHECK_ASSERT(k < max2);
598 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
601 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
602 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
604 grp = ext4_get_group_info(sb, e4b->bd_group);
605 buddy = mb_find_buddy(e4b, 0, &max);
606 list_for_each(cur, &grp->bb_prealloc_list) {
607 ext4_group_t groupnr;
608 struct ext4_prealloc_space *pa;
609 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
610 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
611 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
612 for (i = 0; i < pa->pa_len; i++)
613 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
617 #undef MB_CHECK_ASSERT
618 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
619 __FILE__, __func__, __LINE__)
621 #define mb_check_buddy(e4b)
624 /* FIXME!! need more doc */
625 static void ext4_mb_mark_free_simple(struct super_block *sb,
626 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
627 struct ext4_group_info *grp)
629 struct ext4_sb_info *sbi = EXT4_SB(sb);
633 unsigned short border;
635 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
637 border = 2 << sb->s_blocksize_bits;
640 /* find how many blocks can be covered since this position */
641 max = ffs(first | border) - 1;
643 /* find how many blocks of power 2 we need to mark */
650 /* mark multiblock chunks only */
651 grp->bb_counters[min]++;
653 mb_clear_bit(first >> min,
654 buddy + sbi->s_mb_offsets[min]);
662 * Cache the order of the largest free extent we have available in this block
666 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
671 grp->bb_largest_free_order = -1; /* uninit */
673 bits = sb->s_blocksize_bits + 1;
674 for (i = bits; i >= 0; i--) {
675 if (grp->bb_counters[i] > 0) {
676 grp->bb_largest_free_order = i;
682 static noinline_for_stack
683 void ext4_mb_generate_buddy(struct super_block *sb,
684 void *buddy, void *bitmap, ext4_group_t group)
686 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
687 ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb);
692 unsigned fragments = 0;
693 unsigned long long period = get_cycles();
695 /* initialize buddy from bitmap which is aggregation
696 * of on-disk bitmap and preallocations */
697 i = mb_find_next_zero_bit(bitmap, max, 0);
698 grp->bb_first_free = i;
702 i = mb_find_next_bit(bitmap, max, i);
706 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
708 grp->bb_counters[0]++;
710 i = mb_find_next_zero_bit(bitmap, max, i);
712 grp->bb_fragments = fragments;
714 if (free != grp->bb_free) {
715 ext4_grp_locked_error(sb, group, __func__,
716 "EXT4-fs: group %u: %u blocks in bitmap, %u in gd",
717 group, free, grp->bb_free);
719 * If we intent to continue, we consider group descritor
720 * corrupt and update bb_free using bitmap value
724 mb_set_largest_free_order(sb, grp);
726 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
728 period = get_cycles() - period;
729 spin_lock(&EXT4_SB(sb)->s_bal_lock);
730 EXT4_SB(sb)->s_mb_buddies_generated++;
731 EXT4_SB(sb)->s_mb_generation_time += period;
732 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
735 /* The buddy information is attached the buddy cache inode
736 * for convenience. The information regarding each group
737 * is loaded via ext4_mb_load_buddy. The information involve
738 * block bitmap and buddy information. The information are
739 * stored in the inode as
742 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
745 * one block each for bitmap and buddy information.
746 * So for each group we take up 2 blocks. A page can
747 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
748 * So it can have information regarding groups_per_page which
749 * is blocks_per_page/2
751 * Locking note: This routine takes the block group lock of all groups
752 * for this page; do not hold this lock when calling this routine!
755 static int ext4_mb_init_cache(struct page *page, char *incore)
757 ext4_group_t ngroups;
763 ext4_group_t first_group;
765 struct super_block *sb;
766 struct buffer_head *bhs;
767 struct buffer_head **bh;
772 mb_debug(1, "init page %lu\n", page->index);
774 inode = page->mapping->host;
776 ngroups = ext4_get_groups_count(sb);
777 blocksize = 1 << inode->i_blkbits;
778 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
780 groups_per_page = blocks_per_page >> 1;
781 if (groups_per_page == 0)
784 /* allocate buffer_heads to read bitmaps */
785 if (groups_per_page > 1) {
787 i = sizeof(struct buffer_head *) * groups_per_page;
788 bh = kzalloc(i, GFP_NOFS);
794 first_group = page->index * blocks_per_page / 2;
796 /* read all groups the page covers into the cache */
797 for (i = 0; i < groups_per_page; i++) {
798 struct ext4_group_desc *desc;
800 if (first_group + i >= ngroups)
804 desc = ext4_get_group_desc(sb, first_group + i, NULL);
809 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
813 if (bitmap_uptodate(bh[i]))
817 if (bitmap_uptodate(bh[i])) {
818 unlock_buffer(bh[i]);
821 ext4_lock_group(sb, first_group + i);
822 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
823 ext4_init_block_bitmap(sb, bh[i],
824 first_group + i, desc);
825 set_bitmap_uptodate(bh[i]);
826 set_buffer_uptodate(bh[i]);
827 ext4_unlock_group(sb, first_group + i);
828 unlock_buffer(bh[i]);
831 ext4_unlock_group(sb, first_group + i);
832 if (buffer_uptodate(bh[i])) {
834 * if not uninit if bh is uptodate,
835 * bitmap is also uptodate
837 set_bitmap_uptodate(bh[i]);
838 unlock_buffer(bh[i]);
843 * submit the buffer_head for read. We can
844 * safely mark the bitmap as uptodate now.
845 * We do it here so the bitmap uptodate bit
846 * get set with buffer lock held.
848 set_bitmap_uptodate(bh[i]);
849 bh[i]->b_end_io = end_buffer_read_sync;
850 submit_bh(READ, bh[i]);
851 mb_debug(1, "read bitmap for group %u\n", first_group + i);
854 /* wait for I/O completion */
855 for (i = 0; i < groups_per_page && bh[i]; i++)
856 wait_on_buffer(bh[i]);
859 for (i = 0; i < groups_per_page && bh[i]; i++)
860 if (!buffer_uptodate(bh[i]))
864 first_block = page->index * blocks_per_page;
866 memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
867 for (i = 0; i < blocks_per_page; i++) {
869 struct ext4_group_info *grinfo;
871 group = (first_block + i) >> 1;
872 if (group >= ngroups)
876 * data carry information regarding this
877 * particular group in the format specified
881 data = page_address(page) + (i * blocksize);
882 bitmap = bh[group - first_group]->b_data;
885 * We place the buddy block and bitmap block
888 if ((first_block + i) & 1) {
889 /* this is block of buddy */
890 BUG_ON(incore == NULL);
891 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
892 group, page->index, i * blocksize);
893 trace_ext4_mb_buddy_bitmap_load(sb, group);
894 grinfo = ext4_get_group_info(sb, group);
895 grinfo->bb_fragments = 0;
896 memset(grinfo->bb_counters, 0,
897 sizeof(*grinfo->bb_counters) *
898 (sb->s_blocksize_bits+2));
900 * incore got set to the group block bitmap below
902 ext4_lock_group(sb, group);
903 ext4_mb_generate_buddy(sb, data, incore, group);
904 ext4_unlock_group(sb, group);
907 /* this is block of bitmap */
908 BUG_ON(incore != NULL);
909 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
910 group, page->index, i * blocksize);
911 trace_ext4_mb_bitmap_load(sb, group);
913 /* see comments in ext4_mb_put_pa() */
914 ext4_lock_group(sb, group);
915 memcpy(data, bitmap, blocksize);
917 /* mark all preallocated blks used in in-core bitmap */
918 ext4_mb_generate_from_pa(sb, data, group);
919 ext4_mb_generate_from_freelist(sb, data, group);
920 ext4_unlock_group(sb, group);
922 /* set incore so that the buddy information can be
923 * generated using this
928 SetPageUptodate(page);
932 for (i = 0; i < groups_per_page && bh[i]; i++)
941 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
942 * block group lock of all groups for this page; do not hold the BG lock when
943 * calling this routine!
945 static noinline_for_stack
946 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
952 int block, pnum, poff;
953 int num_grp_locked = 0;
954 struct ext4_group_info *this_grp;
955 struct ext4_sb_info *sbi = EXT4_SB(sb);
956 struct inode *inode = sbi->s_buddy_cache;
957 struct page *page = NULL, *bitmap_page = NULL;
959 mb_debug(1, "init group %u\n", group);
960 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
961 this_grp = ext4_get_group_info(sb, group);
963 * This ensures that we don't reinit the buddy cache
964 * page which map to the group from which we are already
965 * allocating. If we are looking at the buddy cache we would
966 * have taken a reference using ext4_mb_load_buddy and that
967 * would have taken the alloc_sem lock.
969 num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group);
970 if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
972 * somebody initialized the group
973 * return without doing anything
979 * the buddy cache inode stores the block bitmap
980 * and buddy information in consecutive blocks.
981 * So for each group we need two blocks.
984 pnum = block / blocks_per_page;
985 poff = block % blocks_per_page;
986 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
988 BUG_ON(page->mapping != inode->i_mapping);
989 ret = ext4_mb_init_cache(page, NULL);
996 if (page == NULL || !PageUptodate(page)) {
1000 mark_page_accessed(page);
1002 bitmap = page_address(page) + (poff * sb->s_blocksize);
1004 /* init buddy cache */
1006 pnum = block / blocks_per_page;
1007 poff = block % blocks_per_page;
1008 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1009 if (page == bitmap_page) {
1011 * If both the bitmap and buddy are in
1012 * the same page we don't need to force
1017 BUG_ON(page->mapping != inode->i_mapping);
1018 ret = ext4_mb_init_cache(page, bitmap);
1025 if (page == NULL || !PageUptodate(page)) {
1029 mark_page_accessed(page);
1031 ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
1033 page_cache_release(bitmap_page);
1035 page_cache_release(page);
1040 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1041 * block group lock of all groups for this page; do not hold the BG lock when
1042 * calling this routine!
1044 static noinline_for_stack int
1045 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1046 struct ext4_buddy *e4b)
1048 int blocks_per_page;
1054 struct ext4_group_info *grp;
1055 struct ext4_sb_info *sbi = EXT4_SB(sb);
1056 struct inode *inode = sbi->s_buddy_cache;
1058 mb_debug(1, "load group %u\n", group);
1060 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1061 grp = ext4_get_group_info(sb, group);
1063 e4b->bd_blkbits = sb->s_blocksize_bits;
1064 e4b->bd_info = ext4_get_group_info(sb, group);
1066 e4b->bd_group = group;
1067 e4b->bd_buddy_page = NULL;
1068 e4b->bd_bitmap_page = NULL;
1069 e4b->alloc_semp = &grp->alloc_sem;
1071 /* Take the read lock on the group alloc
1072 * sem. This would make sure a parallel
1073 * ext4_mb_init_group happening on other
1074 * groups mapped by the page is blocked
1075 * till we are done with allocation
1078 down_read(e4b->alloc_semp);
1080 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1081 /* we need to check for group need init flag
1082 * with alloc_semp held so that we can be sure
1083 * that new blocks didn't get added to the group
1084 * when we are loading the buddy cache
1086 up_read(e4b->alloc_semp);
1088 * we need full data about the group
1089 * to make a good selection
1091 ret = ext4_mb_init_group(sb, group);
1094 goto repeat_load_buddy;
1098 * the buddy cache inode stores the block bitmap
1099 * and buddy information in consecutive blocks.
1100 * So for each group we need two blocks.
1103 pnum = block / blocks_per_page;
1104 poff = block % blocks_per_page;
1106 /* we could use find_or_create_page(), but it locks page
1107 * what we'd like to avoid in fast path ... */
1108 page = find_get_page(inode->i_mapping, pnum);
1109 if (page == NULL || !PageUptodate(page)) {
1112 * drop the page reference and try
1113 * to get the page with lock. If we
1114 * are not uptodate that implies
1115 * somebody just created the page but
1116 * is yet to initialize the same. So
1117 * wait for it to initialize.
1119 page_cache_release(page);
1120 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1122 BUG_ON(page->mapping != inode->i_mapping);
1123 if (!PageUptodate(page)) {
1124 ret = ext4_mb_init_cache(page, NULL);
1129 mb_cmp_bitmaps(e4b, page_address(page) +
1130 (poff * sb->s_blocksize));
1135 if (page == NULL || !PageUptodate(page)) {
1139 e4b->bd_bitmap_page = page;
1140 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1141 mark_page_accessed(page);
1144 pnum = block / blocks_per_page;
1145 poff = block % blocks_per_page;
1147 page = find_get_page(inode->i_mapping, pnum);
1148 if (page == NULL || !PageUptodate(page)) {
1150 page_cache_release(page);
1151 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1153 BUG_ON(page->mapping != inode->i_mapping);
1154 if (!PageUptodate(page)) {
1155 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1164 if (page == NULL || !PageUptodate(page)) {
1168 e4b->bd_buddy_page = page;
1169 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1170 mark_page_accessed(page);
1172 BUG_ON(e4b->bd_bitmap_page == NULL);
1173 BUG_ON(e4b->bd_buddy_page == NULL);
1178 if (e4b->bd_bitmap_page)
1179 page_cache_release(e4b->bd_bitmap_page);
1180 if (e4b->bd_buddy_page)
1181 page_cache_release(e4b->bd_buddy_page);
1182 e4b->bd_buddy = NULL;
1183 e4b->bd_bitmap = NULL;
1185 /* Done with the buddy cache */
1186 up_read(e4b->alloc_semp);
1190 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1192 if (e4b->bd_bitmap_page)
1193 page_cache_release(e4b->bd_bitmap_page);
1194 if (e4b->bd_buddy_page)
1195 page_cache_release(e4b->bd_buddy_page);
1196 /* Done with the buddy cache */
1197 if (e4b->alloc_semp)
1198 up_read(e4b->alloc_semp);
1202 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1207 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1208 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1210 bb = EXT4_MB_BUDDY(e4b);
1211 while (order <= e4b->bd_blkbits + 1) {
1213 if (!mb_test_bit(block, bb)) {
1214 /* this block is part of buddy of order 'order' */
1217 bb += 1 << (e4b->bd_blkbits - order);
1223 static void mb_clear_bits(void *bm, int cur, int len)
1229 if ((cur & 31) == 0 && (len - cur) >= 32) {
1230 /* fast path: clear whole word at once */
1231 addr = bm + (cur >> 3);
1236 mb_clear_bit(cur, bm);
1241 static void mb_set_bits(void *bm, int cur, int len)
1247 if ((cur & 31) == 0 && (len - cur) >= 32) {
1248 /* fast path: set whole word at once */
1249 addr = bm + (cur >> 3);
1254 mb_set_bit(cur, bm);
1259 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1260 int first, int count)
1267 struct super_block *sb = e4b->bd_sb;
1269 BUG_ON(first + count > (sb->s_blocksize << 3));
1270 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1271 mb_check_buddy(e4b);
1272 mb_free_blocks_double(inode, e4b, first, count);
1274 e4b->bd_info->bb_free += count;
1275 if (first < e4b->bd_info->bb_first_free)
1276 e4b->bd_info->bb_first_free = first;
1278 /* let's maintain fragments counter */
1280 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1281 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1282 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1284 e4b->bd_info->bb_fragments--;
1285 else if (!block && !max)
1286 e4b->bd_info->bb_fragments++;
1288 /* let's maintain buddy itself */
1289 while (count-- > 0) {
1293 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1294 ext4_fsblk_t blocknr;
1296 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1298 ext4_grp_locked_error(sb, e4b->bd_group,
1299 __func__, "double-free of inode"
1300 " %lu's block %llu(bit %u in group %u)",
1301 inode ? inode->i_ino : 0, blocknr, block,
1304 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1305 e4b->bd_info->bb_counters[order]++;
1307 /* start of the buddy */
1308 buddy = mb_find_buddy(e4b, order, &max);
1312 if (mb_test_bit(block, buddy) ||
1313 mb_test_bit(block + 1, buddy))
1316 /* both the buddies are free, try to coalesce them */
1317 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1323 /* for special purposes, we don't set
1324 * free bits in bitmap */
1325 mb_set_bit(block, buddy);
1326 mb_set_bit(block + 1, buddy);
1328 e4b->bd_info->bb_counters[order]--;
1329 e4b->bd_info->bb_counters[order]--;
1333 e4b->bd_info->bb_counters[order]++;
1335 mb_clear_bit(block, buddy2);
1339 mb_set_largest_free_order(sb, e4b->bd_info);
1340 mb_check_buddy(e4b);
1343 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1344 int needed, struct ext4_free_extent *ex)
1351 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1354 buddy = mb_find_buddy(e4b, order, &max);
1355 BUG_ON(buddy == NULL);
1356 BUG_ON(block >= max);
1357 if (mb_test_bit(block, buddy)) {
1364 /* FIXME dorp order completely ? */
1365 if (likely(order == 0)) {
1366 /* find actual order */
1367 order = mb_find_order_for_block(e4b, block);
1368 block = block >> order;
1371 ex->fe_len = 1 << order;
1372 ex->fe_start = block << order;
1373 ex->fe_group = e4b->bd_group;
1375 /* calc difference from given start */
1376 next = next - ex->fe_start;
1378 ex->fe_start += next;
1380 while (needed > ex->fe_len &&
1381 (buddy = mb_find_buddy(e4b, order, &max))) {
1383 if (block + 1 >= max)
1386 next = (block + 1) * (1 << order);
1387 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1390 ord = mb_find_order_for_block(e4b, next);
1393 block = next >> order;
1394 ex->fe_len += 1 << order;
1397 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1401 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1407 int start = ex->fe_start;
1408 int len = ex->fe_len;
1413 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1414 BUG_ON(e4b->bd_group != ex->fe_group);
1415 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1416 mb_check_buddy(e4b);
1417 mb_mark_used_double(e4b, start, len);
1419 e4b->bd_info->bb_free -= len;
1420 if (e4b->bd_info->bb_first_free == start)
1421 e4b->bd_info->bb_first_free += len;
1423 /* let's maintain fragments counter */
1425 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1426 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1427 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1429 e4b->bd_info->bb_fragments++;
1430 else if (!mlen && !max)
1431 e4b->bd_info->bb_fragments--;
1433 /* let's maintain buddy itself */
1435 ord = mb_find_order_for_block(e4b, start);
1437 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1438 /* the whole chunk may be allocated at once! */
1440 buddy = mb_find_buddy(e4b, ord, &max);
1441 BUG_ON((start >> ord) >= max);
1442 mb_set_bit(start >> ord, buddy);
1443 e4b->bd_info->bb_counters[ord]--;
1450 /* store for history */
1452 ret = len | (ord << 16);
1454 /* we have to split large buddy */
1456 buddy = mb_find_buddy(e4b, ord, &max);
1457 mb_set_bit(start >> ord, buddy);
1458 e4b->bd_info->bb_counters[ord]--;
1461 cur = (start >> ord) & ~1U;
1462 buddy = mb_find_buddy(e4b, ord, &max);
1463 mb_clear_bit(cur, buddy);
1464 mb_clear_bit(cur + 1, buddy);
1465 e4b->bd_info->bb_counters[ord]++;
1466 e4b->bd_info->bb_counters[ord]++;
1468 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1470 mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1471 mb_check_buddy(e4b);
1477 * Must be called under group lock!
1479 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1480 struct ext4_buddy *e4b)
1482 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1485 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1486 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1488 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1489 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1490 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1492 /* preallocation can change ac_b_ex, thus we store actually
1493 * allocated blocks for history */
1494 ac->ac_f_ex = ac->ac_b_ex;
1496 ac->ac_status = AC_STATUS_FOUND;
1497 ac->ac_tail = ret & 0xffff;
1498 ac->ac_buddy = ret >> 16;
1501 * take the page reference. We want the page to be pinned
1502 * so that we don't get a ext4_mb_init_cache_call for this
1503 * group until we update the bitmap. That would mean we
1504 * double allocate blocks. The reference is dropped
1505 * in ext4_mb_release_context
1507 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1508 get_page(ac->ac_bitmap_page);
1509 ac->ac_buddy_page = e4b->bd_buddy_page;
1510 get_page(ac->ac_buddy_page);
1511 /* on allocation we use ac to track the held semaphore */
1512 ac->alloc_semp = e4b->alloc_semp;
1513 e4b->alloc_semp = NULL;
1514 /* store last allocated for subsequent stream allocation */
1515 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1516 spin_lock(&sbi->s_md_lock);
1517 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1518 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1519 spin_unlock(&sbi->s_md_lock);
1524 * regular allocator, for general purposes allocation
1527 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1528 struct ext4_buddy *e4b,
1531 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1532 struct ext4_free_extent *bex = &ac->ac_b_ex;
1533 struct ext4_free_extent *gex = &ac->ac_g_ex;
1534 struct ext4_free_extent ex;
1537 if (ac->ac_status == AC_STATUS_FOUND)
1540 * We don't want to scan for a whole year
1542 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1543 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1544 ac->ac_status = AC_STATUS_BREAK;
1549 * Haven't found good chunk so far, let's continue
1551 if (bex->fe_len < gex->fe_len)
1554 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1555 && bex->fe_group == e4b->bd_group) {
1556 /* recheck chunk's availability - we don't know
1557 * when it was found (within this lock-unlock
1559 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1560 if (max >= gex->fe_len) {
1561 ext4_mb_use_best_found(ac, e4b);
1568 * The routine checks whether found extent is good enough. If it is,
1569 * then the extent gets marked used and flag is set to the context
1570 * to stop scanning. Otherwise, the extent is compared with the
1571 * previous found extent and if new one is better, then it's stored
1572 * in the context. Later, the best found extent will be used, if
1573 * mballoc can't find good enough extent.
1575 * FIXME: real allocation policy is to be designed yet!
1577 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1578 struct ext4_free_extent *ex,
1579 struct ext4_buddy *e4b)
1581 struct ext4_free_extent *bex = &ac->ac_b_ex;
1582 struct ext4_free_extent *gex = &ac->ac_g_ex;
1584 BUG_ON(ex->fe_len <= 0);
1585 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1586 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1587 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1592 * The special case - take what you catch first
1594 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1596 ext4_mb_use_best_found(ac, e4b);
1601 * Let's check whether the chuck is good enough
1603 if (ex->fe_len == gex->fe_len) {
1605 ext4_mb_use_best_found(ac, e4b);
1610 * If this is first found extent, just store it in the context
1612 if (bex->fe_len == 0) {
1618 * If new found extent is better, store it in the context
1620 if (bex->fe_len < gex->fe_len) {
1621 /* if the request isn't satisfied, any found extent
1622 * larger than previous best one is better */
1623 if (ex->fe_len > bex->fe_len)
1625 } else if (ex->fe_len > gex->fe_len) {
1626 /* if the request is satisfied, then we try to find
1627 * an extent that still satisfy the request, but is
1628 * smaller than previous one */
1629 if (ex->fe_len < bex->fe_len)
1633 ext4_mb_check_limits(ac, e4b, 0);
1636 static noinline_for_stack
1637 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1638 struct ext4_buddy *e4b)
1640 struct ext4_free_extent ex = ac->ac_b_ex;
1641 ext4_group_t group = ex.fe_group;
1645 BUG_ON(ex.fe_len <= 0);
1646 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1650 ext4_lock_group(ac->ac_sb, group);
1651 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1655 ext4_mb_use_best_found(ac, e4b);
1658 ext4_unlock_group(ac->ac_sb, group);
1659 ext4_mb_unload_buddy(e4b);
1664 static noinline_for_stack
1665 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1666 struct ext4_buddy *e4b)
1668 ext4_group_t group = ac->ac_g_ex.fe_group;
1671 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1672 struct ext4_free_extent ex;
1674 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1677 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1681 ext4_lock_group(ac->ac_sb, group);
1682 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1683 ac->ac_g_ex.fe_len, &ex);
1685 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1688 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1690 /* use do_div to get remainder (would be 64-bit modulo) */
1691 if (do_div(start, sbi->s_stripe) == 0) {
1694 ext4_mb_use_best_found(ac, e4b);
1696 } else if (max >= ac->ac_g_ex.fe_len) {
1697 BUG_ON(ex.fe_len <= 0);
1698 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1699 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1702 ext4_mb_use_best_found(ac, e4b);
1703 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1704 /* Sometimes, caller may want to merge even small
1705 * number of blocks to an existing extent */
1706 BUG_ON(ex.fe_len <= 0);
1707 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1708 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1711 ext4_mb_use_best_found(ac, e4b);
1713 ext4_unlock_group(ac->ac_sb, group);
1714 ext4_mb_unload_buddy(e4b);
1720 * The routine scans buddy structures (not bitmap!) from given order
1721 * to max order and tries to find big enough chunk to satisfy the req
1723 static noinline_for_stack
1724 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1725 struct ext4_buddy *e4b)
1727 struct super_block *sb = ac->ac_sb;
1728 struct ext4_group_info *grp = e4b->bd_info;
1734 BUG_ON(ac->ac_2order <= 0);
1735 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1736 if (grp->bb_counters[i] == 0)
1739 buddy = mb_find_buddy(e4b, i, &max);
1740 BUG_ON(buddy == NULL);
1742 k = mb_find_next_zero_bit(buddy, max, 0);
1747 ac->ac_b_ex.fe_len = 1 << i;
1748 ac->ac_b_ex.fe_start = k << i;
1749 ac->ac_b_ex.fe_group = e4b->bd_group;
1751 ext4_mb_use_best_found(ac, e4b);
1753 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1755 if (EXT4_SB(sb)->s_mb_stats)
1756 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1763 * The routine scans the group and measures all found extents.
1764 * In order to optimize scanning, caller must pass number of
1765 * free blocks in the group, so the routine can know upper limit.
1767 static noinline_for_stack
1768 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1769 struct ext4_buddy *e4b)
1771 struct super_block *sb = ac->ac_sb;
1772 void *bitmap = EXT4_MB_BITMAP(e4b);
1773 struct ext4_free_extent ex;
1777 free = e4b->bd_info->bb_free;
1780 i = e4b->bd_info->bb_first_free;
1782 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1783 i = mb_find_next_zero_bit(bitmap,
1784 EXT4_BLOCKS_PER_GROUP(sb), i);
1785 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1787 * IF we have corrupt bitmap, we won't find any
1788 * free blocks even though group info says we
1789 * we have free blocks
1791 ext4_grp_locked_error(sb, e4b->bd_group,
1792 __func__, "%d free blocks as per "
1793 "group info. But bitmap says 0",
1798 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1799 BUG_ON(ex.fe_len <= 0);
1800 if (free < ex.fe_len) {
1801 ext4_grp_locked_error(sb, e4b->bd_group,
1802 __func__, "%d free blocks as per "
1803 "group info. But got %d blocks",
1806 * The number of free blocks differs. This mostly
1807 * indicate that the bitmap is corrupt. So exit
1808 * without claiming the space.
1813 ext4_mb_measure_extent(ac, &ex, e4b);
1819 ext4_mb_check_limits(ac, e4b, 1);
1823 * This is a special case for storages like raid5
1824 * we try to find stripe-aligned chunks for stripe-size requests
1825 * XXX should do so at least for multiples of stripe size as well
1827 static noinline_for_stack
1828 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1829 struct ext4_buddy *e4b)
1831 struct super_block *sb = ac->ac_sb;
1832 struct ext4_sb_info *sbi = EXT4_SB(sb);
1833 void *bitmap = EXT4_MB_BITMAP(e4b);
1834 struct ext4_free_extent ex;
1835 ext4_fsblk_t first_group_block;
1840 BUG_ON(sbi->s_stripe == 0);
1842 /* find first stripe-aligned block in group */
1843 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1845 a = first_group_block + sbi->s_stripe - 1;
1846 do_div(a, sbi->s_stripe);
1847 i = (a * sbi->s_stripe) - first_group_block;
1849 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1850 if (!mb_test_bit(i, bitmap)) {
1851 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1852 if (max >= sbi->s_stripe) {
1855 ext4_mb_use_best_found(ac, e4b);
1863 /* This is now called BEFORE we load the buddy bitmap. */
1864 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1865 ext4_group_t group, int cr)
1867 unsigned free, fragments;
1868 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1869 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1871 BUG_ON(cr < 0 || cr >= 4);
1873 /* We only do this if the grp has never been initialized */
1874 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1875 int ret = ext4_mb_init_group(ac->ac_sb, group);
1880 free = grp->bb_free;
1881 fragments = grp->bb_fragments;
1889 BUG_ON(ac->ac_2order == 0);
1891 if (grp->bb_largest_free_order < ac->ac_2order)
1894 /* Avoid using the first bg of a flexgroup for data files */
1895 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1896 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1897 ((group % flex_size) == 0))
1902 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1906 if (free >= ac->ac_g_ex.fe_len)
1919 * lock the group_info alloc_sem of all the groups
1920 * belonging to the same buddy cache page. This
1921 * make sure other parallel operation on the buddy
1922 * cache doesn't happen whild holding the buddy cache
1925 int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1929 int blocks_per_page;
1930 int groups_per_page;
1931 ext4_group_t ngroups = ext4_get_groups_count(sb);
1932 ext4_group_t first_group;
1933 struct ext4_group_info *grp;
1935 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1937 * the buddy cache inode stores the block bitmap
1938 * and buddy information in consecutive blocks.
1939 * So for each group we need two blocks.
1942 pnum = block / blocks_per_page;
1943 first_group = pnum * blocks_per_page / 2;
1945 groups_per_page = blocks_per_page >> 1;
1946 if (groups_per_page == 0)
1947 groups_per_page = 1;
1948 /* read all groups the page covers into the cache */
1949 for (i = 0; i < groups_per_page; i++) {
1951 if ((first_group + i) >= ngroups)
1953 grp = ext4_get_group_info(sb, first_group + i);
1954 /* take all groups write allocation
1955 * semaphore. This make sure there is
1956 * no block allocation going on in any
1959 down_write_nested(&grp->alloc_sem, i);
1964 void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1965 ext4_group_t group, int locked_group)
1969 int blocks_per_page;
1970 ext4_group_t first_group;
1971 struct ext4_group_info *grp;
1973 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1975 * the buddy cache inode stores the block bitmap
1976 * and buddy information in consecutive blocks.
1977 * So for each group we need two blocks.
1980 pnum = block / blocks_per_page;
1981 first_group = pnum * blocks_per_page / 2;
1982 /* release locks on all the groups */
1983 for (i = 0; i < locked_group; i++) {
1985 grp = ext4_get_group_info(sb, first_group + i);
1986 /* take all groups write allocation
1987 * semaphore. This make sure there is
1988 * no block allocation going on in any
1991 up_write(&grp->alloc_sem);
1996 static noinline_for_stack int
1997 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1999 ext4_group_t ngroups, group, i;
2003 struct ext4_sb_info *sbi;
2004 struct super_block *sb;
2005 struct ext4_buddy e4b;
2009 ngroups = ext4_get_groups_count(sb);
2010 /* non-extent files are limited to low blocks/groups */
2011 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2012 ngroups = sbi->s_blockfile_groups;
2014 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2016 /* first, try the goal */
2017 err = ext4_mb_find_by_goal(ac, &e4b);
2018 if (err || ac->ac_status == AC_STATUS_FOUND)
2021 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2025 * ac->ac2_order is set only if the fe_len is a power of 2
2026 * if ac2_order is set we also set criteria to 0 so that we
2027 * try exact allocation using buddy.
2029 i = fls(ac->ac_g_ex.fe_len);
2032 * We search using buddy data only if the order of the request
2033 * is greater than equal to the sbi_s_mb_order2_reqs
2034 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2036 if (i >= sbi->s_mb_order2_reqs) {
2038 * This should tell if fe_len is exactly power of 2
2040 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2041 ac->ac_2order = i - 1;
2044 bsbits = ac->ac_sb->s_blocksize_bits;
2046 /* if stream allocation is enabled, use global goal */
2047 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2048 /* TBD: may be hot point */
2049 spin_lock(&sbi->s_md_lock);
2050 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2051 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2052 spin_unlock(&sbi->s_md_lock);
2055 /* Let's just scan groups to find more-less suitable blocks */
2056 cr = ac->ac_2order ? 0 : 1;
2058 * cr == 0 try to get exact allocation,
2059 * cr == 3 try to get anything
2062 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2063 ac->ac_criteria = cr;
2065 * searching for the right group start
2066 * from the goal value specified
2068 group = ac->ac_g_ex.fe_group;
2070 for (i = 0; i < ngroups; group++, i++) {
2071 if (group == ngroups)
2074 /* This now checks without needing the buddy page */
2075 if (!ext4_mb_good_group(ac, group, cr))
2078 err = ext4_mb_load_buddy(sb, group, &e4b);
2082 ext4_lock_group(sb, group);
2085 * We need to check again after locking the
2088 if (!ext4_mb_good_group(ac, group, cr)) {
2089 ext4_unlock_group(sb, group);
2090 ext4_mb_unload_buddy(&e4b);
2094 ac->ac_groups_scanned++;
2096 ext4_mb_simple_scan_group(ac, &e4b);
2098 ac->ac_g_ex.fe_len == sbi->s_stripe)
2099 ext4_mb_scan_aligned(ac, &e4b);
2101 ext4_mb_complex_scan_group(ac, &e4b);
2103 ext4_unlock_group(sb, group);
2104 ext4_mb_unload_buddy(&e4b);
2106 if (ac->ac_status != AC_STATUS_CONTINUE)
2111 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2112 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2114 * We've been searching too long. Let's try to allocate
2115 * the best chunk we've found so far
2118 ext4_mb_try_best_found(ac, &e4b);
2119 if (ac->ac_status != AC_STATUS_FOUND) {
2121 * Someone more lucky has already allocated it.
2122 * The only thing we can do is just take first
2124 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2126 ac->ac_b_ex.fe_group = 0;
2127 ac->ac_b_ex.fe_start = 0;
2128 ac->ac_b_ex.fe_len = 0;
2129 ac->ac_status = AC_STATUS_CONTINUE;
2130 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2132 atomic_inc(&sbi->s_mb_lost_chunks);
2140 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2142 struct super_block *sb = seq->private;
2145 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2148 return (void *) ((unsigned long) group);
2151 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2153 struct super_block *sb = seq->private;
2157 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2160 return (void *) ((unsigned long) group);
2163 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2165 struct super_block *sb = seq->private;
2166 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2169 struct ext4_buddy e4b;
2171 struct ext4_group_info info;
2172 ext4_grpblk_t counters[16];
2177 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2178 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2179 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2180 "group", "free", "frags", "first",
2181 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2182 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2184 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2185 sizeof(struct ext4_group_info);
2186 err = ext4_mb_load_buddy(sb, group, &e4b);
2188 seq_printf(seq, "#%-5u: I/O error\n", group);
2191 ext4_lock_group(sb, group);
2192 memcpy(&sg, ext4_get_group_info(sb, group), i);
2193 ext4_unlock_group(sb, group);
2194 ext4_mb_unload_buddy(&e4b);
2196 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2197 sg.info.bb_fragments, sg.info.bb_first_free);
2198 for (i = 0; i <= 13; i++)
2199 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2200 sg.info.bb_counters[i] : 0);
2201 seq_printf(seq, " ]\n");
2206 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2210 static const struct seq_operations ext4_mb_seq_groups_ops = {
2211 .start = ext4_mb_seq_groups_start,
2212 .next = ext4_mb_seq_groups_next,
2213 .stop = ext4_mb_seq_groups_stop,
2214 .show = ext4_mb_seq_groups_show,
2217 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2219 struct super_block *sb = PDE(inode)->data;
2222 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2224 struct seq_file *m = (struct seq_file *)file->private_data;
2231 static const struct file_operations ext4_mb_seq_groups_fops = {
2232 .owner = THIS_MODULE,
2233 .open = ext4_mb_seq_groups_open,
2235 .llseek = seq_lseek,
2236 .release = seq_release,
2240 /* Create and initialize ext4_group_info data for the given group. */
2241 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2242 struct ext4_group_desc *desc)
2246 struct ext4_sb_info *sbi = EXT4_SB(sb);
2247 struct ext4_group_info **meta_group_info;
2250 * First check if this group is the first of a reserved block.
2251 * If it's true, we have to allocate a new table of pointers
2252 * to ext4_group_info structures
2254 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2255 metalen = sizeof(*meta_group_info) <<
2256 EXT4_DESC_PER_BLOCK_BITS(sb);
2257 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2258 if (meta_group_info == NULL) {
2259 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2261 goto exit_meta_group_info;
2263 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2268 * calculate needed size. if change bb_counters size,
2269 * don't forget about ext4_mb_generate_buddy()
2271 len = offsetof(typeof(**meta_group_info),
2272 bb_counters[sb->s_blocksize_bits + 2]);
2275 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2276 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2278 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2279 if (meta_group_info[i] == NULL) {
2280 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2281 goto exit_group_info;
2283 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2284 &(meta_group_info[i]->bb_state));
2287 * initialize bb_free to be able to skip
2288 * empty groups without initialization
2290 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2291 meta_group_info[i]->bb_free =
2292 ext4_free_blocks_after_init(sb, group, desc);
2294 meta_group_info[i]->bb_free =
2295 ext4_free_blks_count(sb, desc);
2298 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2299 init_rwsem(&meta_group_info[i]->alloc_sem);
2300 meta_group_info[i]->bb_free_root = RB_ROOT;
2301 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2305 struct buffer_head *bh;
2306 meta_group_info[i]->bb_bitmap =
2307 kmalloc(sb->s_blocksize, GFP_KERNEL);
2308 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2309 bh = ext4_read_block_bitmap(sb, group);
2311 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2320 /* If a meta_group_info table has been allocated, release it now */
2321 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2322 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2323 exit_meta_group_info:
2325 } /* ext4_mb_add_groupinfo */
2327 static int ext4_mb_init_backend(struct super_block *sb)
2329 ext4_group_t ngroups = ext4_get_groups_count(sb);
2331 struct ext4_sb_info *sbi = EXT4_SB(sb);
2332 struct ext4_super_block *es = sbi->s_es;
2333 int num_meta_group_infos;
2334 int num_meta_group_infos_max;
2336 struct ext4_group_desc *desc;
2338 /* This is the number of blocks used by GDT */
2339 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2340 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2343 * This is the total number of blocks used by GDT including
2344 * the number of reserved blocks for GDT.
2345 * The s_group_info array is allocated with this value
2346 * to allow a clean online resize without a complex
2347 * manipulation of pointer.
2348 * The drawback is the unused memory when no resize
2349 * occurs but it's very low in terms of pages
2350 * (see comments below)
2351 * Need to handle this properly when META_BG resizing is allowed
2353 num_meta_group_infos_max = num_meta_group_infos +
2354 le16_to_cpu(es->s_reserved_gdt_blocks);
2357 * array_size is the size of s_group_info array. We round it
2358 * to the next power of two because this approximation is done
2359 * internally by kmalloc so we can have some more memory
2360 * for free here (e.g. may be used for META_BG resize).
2363 while (array_size < sizeof(*sbi->s_group_info) *
2364 num_meta_group_infos_max)
2365 array_size = array_size << 1;
2366 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2367 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2368 * So a two level scheme suffices for now. */
2369 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2370 if (sbi->s_group_info == NULL) {
2371 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2374 sbi->s_buddy_cache = new_inode(sb);
2375 if (sbi->s_buddy_cache == NULL) {
2376 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2379 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2380 for (i = 0; i < ngroups; i++) {
2381 desc = ext4_get_group_desc(sb, i, NULL);
2384 "EXT4-fs: can't read descriptor %u\n", i);
2387 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2395 kfree(ext4_get_group_info(sb, i));
2396 i = num_meta_group_infos;
2398 kfree(sbi->s_group_info[i]);
2399 iput(sbi->s_buddy_cache);
2401 kfree(sbi->s_group_info);
2405 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2407 struct ext4_sb_info *sbi = EXT4_SB(sb);
2413 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2415 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2416 if (sbi->s_mb_offsets == NULL) {
2420 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2421 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2422 if (sbi->s_mb_maxs == NULL) {
2423 kfree(sbi->s_mb_offsets);
2427 /* order 0 is regular bitmap */
2428 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2429 sbi->s_mb_offsets[0] = 0;
2433 max = sb->s_blocksize << 2;
2435 sbi->s_mb_offsets[i] = offset;
2436 sbi->s_mb_maxs[i] = max;
2437 offset += 1 << (sb->s_blocksize_bits - i);
2440 } while (i <= sb->s_blocksize_bits + 1);
2442 /* init file for buddy data */
2443 ret = ext4_mb_init_backend(sb);
2445 kfree(sbi->s_mb_offsets);
2446 kfree(sbi->s_mb_maxs);
2450 spin_lock_init(&sbi->s_md_lock);
2451 spin_lock_init(&sbi->s_bal_lock);
2453 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2454 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2455 sbi->s_mb_stats = MB_DEFAULT_STATS;
2456 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2457 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2458 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2460 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2461 if (sbi->s_locality_groups == NULL) {
2462 kfree(sbi->s_mb_offsets);
2463 kfree(sbi->s_mb_maxs);
2466 for_each_possible_cpu(i) {
2467 struct ext4_locality_group *lg;
2468 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2469 mutex_init(&lg->lg_mutex);
2470 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2471 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2472 spin_lock_init(&lg->lg_prealloc_lock);
2476 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2477 &ext4_mb_seq_groups_fops, sb);
2480 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2484 /* need to called with the ext4 group lock held */
2485 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2487 struct ext4_prealloc_space *pa;
2488 struct list_head *cur, *tmp;
2491 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2492 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2493 list_del(&pa->pa_group_list);
2495 kmem_cache_free(ext4_pspace_cachep, pa);
2498 mb_debug(1, "mballoc: %u PAs left\n", count);
2502 int ext4_mb_release(struct super_block *sb)
2504 ext4_group_t ngroups = ext4_get_groups_count(sb);
2506 int num_meta_group_infos;
2507 struct ext4_group_info *grinfo;
2508 struct ext4_sb_info *sbi = EXT4_SB(sb);
2510 if (sbi->s_group_info) {
2511 for (i = 0; i < ngroups; i++) {
2512 grinfo = ext4_get_group_info(sb, i);
2514 kfree(grinfo->bb_bitmap);
2516 ext4_lock_group(sb, i);
2517 ext4_mb_cleanup_pa(grinfo);
2518 ext4_unlock_group(sb, i);
2521 num_meta_group_infos = (ngroups +
2522 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2523 EXT4_DESC_PER_BLOCK_BITS(sb);
2524 for (i = 0; i < num_meta_group_infos; i++)
2525 kfree(sbi->s_group_info[i]);
2526 kfree(sbi->s_group_info);
2528 kfree(sbi->s_mb_offsets);
2529 kfree(sbi->s_mb_maxs);
2530 if (sbi->s_buddy_cache)
2531 iput(sbi->s_buddy_cache);
2532 if (sbi->s_mb_stats) {
2534 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2535 atomic_read(&sbi->s_bal_allocated),
2536 atomic_read(&sbi->s_bal_reqs),
2537 atomic_read(&sbi->s_bal_success));
2539 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2540 "%u 2^N hits, %u breaks, %u lost\n",
2541 atomic_read(&sbi->s_bal_ex_scanned),
2542 atomic_read(&sbi->s_bal_goals),
2543 atomic_read(&sbi->s_bal_2orders),
2544 atomic_read(&sbi->s_bal_breaks),
2545 atomic_read(&sbi->s_mb_lost_chunks));
2547 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2548 sbi->s_mb_buddies_generated++,
2549 sbi->s_mb_generation_time);
2551 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2552 atomic_read(&sbi->s_mb_preallocated),
2553 atomic_read(&sbi->s_mb_discarded));
2556 free_percpu(sbi->s_locality_groups);
2558 remove_proc_entry("mb_groups", sbi->s_proc);
2564 * This function is called by the jbd2 layer once the commit has finished,
2565 * so we know we can free the blocks that were released with that commit.
2567 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2569 struct super_block *sb = journal->j_private;
2570 struct ext4_buddy e4b;
2571 struct ext4_group_info *db;
2572 int err, count = 0, count2 = 0;
2573 struct ext4_free_data *entry;
2574 struct list_head *l, *ltmp;
2576 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2577 entry = list_entry(l, struct ext4_free_data, list);
2579 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2580 entry->count, entry->group, entry);
2582 if (test_opt(sb, DISCARD)) {
2584 ext4_fsblk_t discard_block;
2586 discard_block = entry->start_blk +
2587 ext4_group_first_block_no(sb, entry->group);
2588 trace_ext4_discard_blocks(sb,
2589 (unsigned long long)discard_block,
2591 ret = sb_issue_discard(sb, discard_block, entry->count);
2592 if (ret == EOPNOTSUPP) {
2594 "discard not supported, disabling");
2595 clear_opt(EXT4_SB(sb)->s_mount_opt, DISCARD);
2599 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2600 /* we expect to find existing buddy because it's pinned */
2604 /* there are blocks to put in buddy to make them really free */
2605 count += entry->count;
2607 ext4_lock_group(sb, entry->group);
2608 /* Take it out of per group rb tree */
2609 rb_erase(&entry->node, &(db->bb_free_root));
2610 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2612 if (!db->bb_free_root.rb_node) {
2613 /* No more items in the per group rb tree
2614 * balance refcounts from ext4_mb_free_metadata()
2616 page_cache_release(e4b.bd_buddy_page);
2617 page_cache_release(e4b.bd_bitmap_page);
2619 ext4_unlock_group(sb, entry->group);
2620 kmem_cache_free(ext4_free_ext_cachep, entry);
2621 ext4_mb_unload_buddy(&e4b);
2624 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2627 #ifdef CONFIG_EXT4_DEBUG
2628 u8 mb_enable_debug __read_mostly;
2630 static struct dentry *debugfs_dir;
2631 static struct dentry *debugfs_debug;
2633 static void __init ext4_create_debugfs_entry(void)
2635 debugfs_dir = debugfs_create_dir("ext4", NULL);
2637 debugfs_debug = debugfs_create_u8("mballoc-debug",
2643 static void ext4_remove_debugfs_entry(void)
2645 debugfs_remove(debugfs_debug);
2646 debugfs_remove(debugfs_dir);
2651 static void __init ext4_create_debugfs_entry(void)
2655 static void ext4_remove_debugfs_entry(void)
2661 int __init init_ext4_mballoc(void)
2663 ext4_pspace_cachep =
2664 kmem_cache_create("ext4_prealloc_space",
2665 sizeof(struct ext4_prealloc_space),
2666 0, SLAB_RECLAIM_ACCOUNT, NULL);
2667 if (ext4_pspace_cachep == NULL)
2671 kmem_cache_create("ext4_alloc_context",
2672 sizeof(struct ext4_allocation_context),
2673 0, SLAB_RECLAIM_ACCOUNT, NULL);
2674 if (ext4_ac_cachep == NULL) {
2675 kmem_cache_destroy(ext4_pspace_cachep);
2679 ext4_free_ext_cachep =
2680 kmem_cache_create("ext4_free_block_extents",
2681 sizeof(struct ext4_free_data),
2682 0, SLAB_RECLAIM_ACCOUNT, NULL);
2683 if (ext4_free_ext_cachep == NULL) {
2684 kmem_cache_destroy(ext4_pspace_cachep);
2685 kmem_cache_destroy(ext4_ac_cachep);
2688 ext4_create_debugfs_entry();
2692 void exit_ext4_mballoc(void)
2695 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2696 * before destroying the slab cache.
2699 kmem_cache_destroy(ext4_pspace_cachep);
2700 kmem_cache_destroy(ext4_ac_cachep);
2701 kmem_cache_destroy(ext4_free_ext_cachep);
2702 ext4_remove_debugfs_entry();
2707 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2708 * Returns 0 if success or error code
2710 static noinline_for_stack int
2711 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2712 handle_t *handle, unsigned int reserv_blks)
2714 struct buffer_head *bitmap_bh = NULL;
2715 struct ext4_super_block *es;
2716 struct ext4_group_desc *gdp;
2717 struct buffer_head *gdp_bh;
2718 struct ext4_sb_info *sbi;
2719 struct super_block *sb;
2723 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2724 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2732 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2736 err = ext4_journal_get_write_access(handle, bitmap_bh);
2741 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2745 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2746 ext4_free_blks_count(sb, gdp));
2748 err = ext4_journal_get_write_access(handle, gdp_bh);
2752 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2754 len = ac->ac_b_ex.fe_len;
2755 if (!ext4_data_block_valid(sbi, block, len)) {
2756 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2757 "fs metadata\n", block, block+len);
2758 /* File system mounted not to panic on error
2759 * Fix the bitmap and repeat the block allocation
2760 * We leak some of the blocks here.
2762 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2763 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2764 ac->ac_b_ex.fe_len);
2765 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2766 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2772 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2773 #ifdef AGGRESSIVE_CHECK
2776 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2777 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2778 bitmap_bh->b_data));
2782 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
2783 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2784 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2785 ext4_free_blks_set(sb, gdp,
2786 ext4_free_blocks_after_init(sb,
2787 ac->ac_b_ex.fe_group, gdp));
2789 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
2790 ext4_free_blks_set(sb, gdp, len);
2791 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2793 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2794 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
2796 * Now reduce the dirty block count also. Should not go negative
2798 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2799 /* release all the reserved blocks if non delalloc */
2800 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
2802 if (sbi->s_log_groups_per_flex) {
2803 ext4_group_t flex_group = ext4_flex_group(sbi,
2804 ac->ac_b_ex.fe_group);
2805 atomic_sub(ac->ac_b_ex.fe_len,
2806 &sbi->s_flex_groups[flex_group].free_blocks);
2809 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2812 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2821 * here we normalize request for locality group
2822 * Group request are normalized to s_strip size if we set the same via mount
2823 * option. If not we set it to s_mb_group_prealloc which can be configured via
2824 * /sys/fs/ext4/<partition>/mb_group_prealloc
2826 * XXX: should we try to preallocate more than the group has now?
2828 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2830 struct super_block *sb = ac->ac_sb;
2831 struct ext4_locality_group *lg = ac->ac_lg;
2834 if (EXT4_SB(sb)->s_stripe)
2835 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
2837 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2838 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2839 current->pid, ac->ac_g_ex.fe_len);
2843 * Normalization means making request better in terms of
2844 * size and alignment
2846 static noinline_for_stack void
2847 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2848 struct ext4_allocation_request *ar)
2852 loff_t size, orig_size, start_off;
2853 ext4_lblk_t start, orig_start;
2854 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2855 struct ext4_prealloc_space *pa;
2857 /* do normalize only data requests, metadata requests
2858 do not need preallocation */
2859 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2862 /* sometime caller may want exact blocks */
2863 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2866 /* caller may indicate that preallocation isn't
2867 * required (it's a tail, for example) */
2868 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2871 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2872 ext4_mb_normalize_group_request(ac);
2876 bsbits = ac->ac_sb->s_blocksize_bits;
2878 /* first, let's learn actual file size
2879 * given current request is allocated */
2880 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2881 size = size << bsbits;
2882 if (size < i_size_read(ac->ac_inode))
2883 size = i_size_read(ac->ac_inode);
2885 /* max size of free chunks */
2888 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2889 (req <= (size) || max <= (chunk_size))
2891 /* first, try to predict filesize */
2892 /* XXX: should this table be tunable? */
2894 if (size <= 16 * 1024) {
2896 } else if (size <= 32 * 1024) {
2898 } else if (size <= 64 * 1024) {
2900 } else if (size <= 128 * 1024) {
2902 } else if (size <= 256 * 1024) {
2904 } else if (size <= 512 * 1024) {
2906 } else if (size <= 1024 * 1024) {
2908 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2909 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2910 (21 - bsbits)) << 21;
2911 size = 2 * 1024 * 1024;
2912 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2913 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2914 (22 - bsbits)) << 22;
2915 size = 4 * 1024 * 1024;
2916 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2917 (8<<20)>>bsbits, max, 8 * 1024)) {
2918 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2919 (23 - bsbits)) << 23;
2920 size = 8 * 1024 * 1024;
2922 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2923 size = ac->ac_o_ex.fe_len << bsbits;
2925 orig_size = size = size >> bsbits;
2926 orig_start = start = start_off >> bsbits;
2928 /* don't cover already allocated blocks in selected range */
2929 if (ar->pleft && start <= ar->lleft) {
2930 size -= ar->lleft + 1 - start;
2931 start = ar->lleft + 1;
2933 if (ar->pright && start + size - 1 >= ar->lright)
2934 size -= start + size - ar->lright;
2938 /* check we don't cross already preallocated blocks */
2940 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2945 spin_lock(&pa->pa_lock);
2946 if (pa->pa_deleted) {
2947 spin_unlock(&pa->pa_lock);
2951 pa_end = pa->pa_lstart + pa->pa_len;
2953 /* PA must not overlap original request */
2954 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2955 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2957 /* skip PAs this normalized request doesn't overlap with */
2958 if (pa->pa_lstart >= end || pa_end <= start) {
2959 spin_unlock(&pa->pa_lock);
2962 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2964 /* adjust start or end to be adjacent to this pa */
2965 if (pa_end <= ac->ac_o_ex.fe_logical) {
2966 BUG_ON(pa_end < start);
2968 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
2969 BUG_ON(pa->pa_lstart > end);
2970 end = pa->pa_lstart;
2972 spin_unlock(&pa->pa_lock);
2977 /* XXX: extra loop to check we really don't overlap preallocations */
2979 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2981 spin_lock(&pa->pa_lock);
2982 if (pa->pa_deleted == 0) {
2983 pa_end = pa->pa_lstart + pa->pa_len;
2984 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
2986 spin_unlock(&pa->pa_lock);
2990 if (start + size <= ac->ac_o_ex.fe_logical &&
2991 start > ac->ac_o_ex.fe_logical) {
2992 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
2993 (unsigned long) start, (unsigned long) size,
2994 (unsigned long) ac->ac_o_ex.fe_logical);
2996 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
2997 start > ac->ac_o_ex.fe_logical);
2998 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3000 /* now prepare goal request */
3002 /* XXX: is it better to align blocks WRT to logical
3003 * placement or satisfy big request as is */
3004 ac->ac_g_ex.fe_logical = start;
3005 ac->ac_g_ex.fe_len = size;
3007 /* define goal start in order to merge */
3008 if (ar->pright && (ar->lright == (start + size))) {
3009 /* merge to the right */
3010 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3011 &ac->ac_f_ex.fe_group,
3012 &ac->ac_f_ex.fe_start);
3013 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3015 if (ar->pleft && (ar->lleft + 1 == start)) {
3016 /* merge to the left */
3017 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3018 &ac->ac_f_ex.fe_group,
3019 &ac->ac_f_ex.fe_start);
3020 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3023 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3024 (unsigned) orig_size, (unsigned) start);
3027 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3029 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3031 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3032 atomic_inc(&sbi->s_bal_reqs);
3033 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3034 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3035 atomic_inc(&sbi->s_bal_success);
3036 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3037 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3038 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3039 atomic_inc(&sbi->s_bal_goals);
3040 if (ac->ac_found > sbi->s_mb_max_to_scan)
3041 atomic_inc(&sbi->s_bal_breaks);
3044 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3045 trace_ext4_mballoc_alloc(ac);
3047 trace_ext4_mballoc_prealloc(ac);
3051 * Called on failure; free up any blocks from the inode PA for this
3052 * context. We don't need this for MB_GROUP_PA because we only change
3053 * pa_free in ext4_mb_release_context(), but on failure, we've already
3054 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3056 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3058 struct ext4_prealloc_space *pa = ac->ac_pa;
3061 if (pa && pa->pa_type == MB_INODE_PA) {
3062 len = ac->ac_b_ex.fe_len;
3069 * use blocks preallocated to inode
3071 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3072 struct ext4_prealloc_space *pa)
3078 /* found preallocated blocks, use them */
3079 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3080 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3082 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3083 &ac->ac_b_ex.fe_start);
3084 ac->ac_b_ex.fe_len = len;
3085 ac->ac_status = AC_STATUS_FOUND;
3088 BUG_ON(start < pa->pa_pstart);
3089 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3090 BUG_ON(pa->pa_free < len);
3093 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3097 * use blocks preallocated to locality group
3099 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3100 struct ext4_prealloc_space *pa)
3102 unsigned int len = ac->ac_o_ex.fe_len;
3104 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3105 &ac->ac_b_ex.fe_group,
3106 &ac->ac_b_ex.fe_start);
3107 ac->ac_b_ex.fe_len = len;
3108 ac->ac_status = AC_STATUS_FOUND;
3111 /* we don't correct pa_pstart or pa_plen here to avoid
3112 * possible race when the group is being loaded concurrently
3113 * instead we correct pa later, after blocks are marked
3114 * in on-disk bitmap -- see ext4_mb_release_context()
3115 * Other CPUs are prevented from allocating from this pa by lg_mutex
3117 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3121 * Return the prealloc space that have minimal distance
3122 * from the goal block. @cpa is the prealloc
3123 * space that is having currently known minimal distance
3124 * from the goal block.
3126 static struct ext4_prealloc_space *
3127 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3128 struct ext4_prealloc_space *pa,
3129 struct ext4_prealloc_space *cpa)
3131 ext4_fsblk_t cur_distance, new_distance;
3134 atomic_inc(&pa->pa_count);
3137 cur_distance = abs(goal_block - cpa->pa_pstart);
3138 new_distance = abs(goal_block - pa->pa_pstart);
3140 if (cur_distance < new_distance)
3143 /* drop the previous reference */
3144 atomic_dec(&cpa->pa_count);
3145 atomic_inc(&pa->pa_count);
3150 * search goal blocks in preallocated space
3152 static noinline_for_stack int
3153 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3156 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3157 struct ext4_locality_group *lg;
3158 struct ext4_prealloc_space *pa, *cpa = NULL;
3159 ext4_fsblk_t goal_block;
3161 /* only data can be preallocated */
3162 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3165 /* first, try per-file preallocation */
3167 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3169 /* all fields in this condition don't change,
3170 * so we can skip locking for them */
3171 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3172 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3175 /* non-extent files can't have physical blocks past 2^32 */
3176 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3177 pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3180 /* found preallocated blocks, use them */
3181 spin_lock(&pa->pa_lock);
3182 if (pa->pa_deleted == 0 && pa->pa_free) {
3183 atomic_inc(&pa->pa_count);
3184 ext4_mb_use_inode_pa(ac, pa);
3185 spin_unlock(&pa->pa_lock);
3186 ac->ac_criteria = 10;
3190 spin_unlock(&pa->pa_lock);
3194 /* can we use group allocation? */
3195 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3198 /* inode may have no locality group for some reason */
3202 order = fls(ac->ac_o_ex.fe_len) - 1;
3203 if (order > PREALLOC_TB_SIZE - 1)
3204 /* The max size of hash table is PREALLOC_TB_SIZE */
3205 order = PREALLOC_TB_SIZE - 1;
3207 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3209 * search for the prealloc space that is having
3210 * minimal distance from the goal block.
3212 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3214 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3216 spin_lock(&pa->pa_lock);
3217 if (pa->pa_deleted == 0 &&
3218 pa->pa_free >= ac->ac_o_ex.fe_len) {
3220 cpa = ext4_mb_check_group_pa(goal_block,
3223 spin_unlock(&pa->pa_lock);
3228 ext4_mb_use_group_pa(ac, cpa);
3229 ac->ac_criteria = 20;
3236 * the function goes through all block freed in the group
3237 * but not yet committed and marks them used in in-core bitmap.
3238 * buddy must be generated from this bitmap
3239 * Need to be called with the ext4 group lock held
3241 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3245 struct ext4_group_info *grp;
3246 struct ext4_free_data *entry;
3248 grp = ext4_get_group_info(sb, group);
3249 n = rb_first(&(grp->bb_free_root));
3252 entry = rb_entry(n, struct ext4_free_data, node);
3253 mb_set_bits(bitmap, entry->start_blk, entry->count);
3260 * the function goes through all preallocation in this group and marks them
3261 * used in in-core bitmap. buddy must be generated from this bitmap
3262 * Need to be called with ext4 group lock held
3264 static noinline_for_stack
3265 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3268 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3269 struct ext4_prealloc_space *pa;
3270 struct list_head *cur;
3271 ext4_group_t groupnr;
3272 ext4_grpblk_t start;
3273 int preallocated = 0;
3277 /* all form of preallocation discards first load group,
3278 * so the only competing code is preallocation use.
3279 * we don't need any locking here
3280 * notice we do NOT ignore preallocations with pa_deleted
3281 * otherwise we could leave used blocks available for
3282 * allocation in buddy when concurrent ext4_mb_put_pa()
3283 * is dropping preallocation
3285 list_for_each(cur, &grp->bb_prealloc_list) {
3286 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3287 spin_lock(&pa->pa_lock);
3288 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3291 spin_unlock(&pa->pa_lock);
3292 if (unlikely(len == 0))
3294 BUG_ON(groupnr != group);
3295 mb_set_bits(bitmap, start, len);
3296 preallocated += len;
3299 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3302 static void ext4_mb_pa_callback(struct rcu_head *head)
3304 struct ext4_prealloc_space *pa;
3305 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3306 kmem_cache_free(ext4_pspace_cachep, pa);
3310 * drops a reference to preallocated space descriptor
3311 * if this was the last reference and the space is consumed
3313 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3314 struct super_block *sb, struct ext4_prealloc_space *pa)
3317 ext4_fsblk_t grp_blk;
3319 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3322 /* in this short window concurrent discard can set pa_deleted */
3323 spin_lock(&pa->pa_lock);
3324 if (pa->pa_deleted == 1) {
3325 spin_unlock(&pa->pa_lock);
3330 spin_unlock(&pa->pa_lock);
3332 grp_blk = pa->pa_pstart;
3334 * If doing group-based preallocation, pa_pstart may be in the
3335 * next group when pa is used up
3337 if (pa->pa_type == MB_GROUP_PA)
3340 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3345 * P1 (buddy init) P2 (regular allocation)
3346 * find block B in PA
3347 * copy on-disk bitmap to buddy
3348 * mark B in on-disk bitmap
3349 * drop PA from group
3350 * mark all PAs in buddy
3352 * thus, P1 initializes buddy with B available. to prevent this
3353 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3356 ext4_lock_group(sb, grp);
3357 list_del(&pa->pa_group_list);
3358 ext4_unlock_group(sb, grp);
3360 spin_lock(pa->pa_obj_lock);
3361 list_del_rcu(&pa->pa_inode_list);
3362 spin_unlock(pa->pa_obj_lock);
3364 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3368 * creates new preallocated space for given inode
3370 static noinline_for_stack int
3371 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3373 struct super_block *sb = ac->ac_sb;
3374 struct ext4_prealloc_space *pa;
3375 struct ext4_group_info *grp;
3376 struct ext4_inode_info *ei;
3378 /* preallocate only when found space is larger then requested */
3379 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3380 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3381 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3383 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3387 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3393 /* we can't allocate as much as normalizer wants.
3394 * so, found space must get proper lstart
3395 * to cover original request */
3396 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3397 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3399 /* we're limited by original request in that
3400 * logical block must be covered any way
3401 * winl is window we can move our chunk within */
3402 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3404 /* also, we should cover whole original request */
3405 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3407 /* the smallest one defines real window */
3408 win = min(winl, wins);
3410 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3411 if (offs && offs < win)
3414 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3415 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3416 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3419 /* preallocation can change ac_b_ex, thus we store actually
3420 * allocated blocks for history */
3421 ac->ac_f_ex = ac->ac_b_ex;
3423 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3424 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3425 pa->pa_len = ac->ac_b_ex.fe_len;
3426 pa->pa_free = pa->pa_len;
3427 atomic_set(&pa->pa_count, 1);
3428 spin_lock_init(&pa->pa_lock);
3429 INIT_LIST_HEAD(&pa->pa_inode_list);
3430 INIT_LIST_HEAD(&pa->pa_group_list);
3432 pa->pa_type = MB_INODE_PA;
3434 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3435 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3436 trace_ext4_mb_new_inode_pa(ac, pa);
3438 ext4_mb_use_inode_pa(ac, pa);
3439 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3441 ei = EXT4_I(ac->ac_inode);
3442 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3444 pa->pa_obj_lock = &ei->i_prealloc_lock;
3445 pa->pa_inode = ac->ac_inode;
3447 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3448 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3449 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3451 spin_lock(pa->pa_obj_lock);
3452 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3453 spin_unlock(pa->pa_obj_lock);
3459 * creates new preallocated space for locality group inodes belongs to
3461 static noinline_for_stack int
3462 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3464 struct super_block *sb = ac->ac_sb;
3465 struct ext4_locality_group *lg;
3466 struct ext4_prealloc_space *pa;
3467 struct ext4_group_info *grp;
3469 /* preallocate only when found space is larger then requested */
3470 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3471 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3472 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3474 BUG_ON(ext4_pspace_cachep == NULL);
3475 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3479 /* preallocation can change ac_b_ex, thus we store actually
3480 * allocated blocks for history */
3481 ac->ac_f_ex = ac->ac_b_ex;
3483 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3484 pa->pa_lstart = pa->pa_pstart;
3485 pa->pa_len = ac->ac_b_ex.fe_len;
3486 pa->pa_free = pa->pa_len;
3487 atomic_set(&pa->pa_count, 1);
3488 spin_lock_init(&pa->pa_lock);
3489 INIT_LIST_HEAD(&pa->pa_inode_list);
3490 INIT_LIST_HEAD(&pa->pa_group_list);
3492 pa->pa_type = MB_GROUP_PA;
3494 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3495 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3496 trace_ext4_mb_new_group_pa(ac, pa);
3498 ext4_mb_use_group_pa(ac, pa);
3499 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3501 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3505 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3506 pa->pa_inode = NULL;
3508 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3509 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3510 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3513 * We will later add the new pa to the right bucket
3514 * after updating the pa_free in ext4_mb_release_context
3519 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3523 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3524 err = ext4_mb_new_group_pa(ac);
3526 err = ext4_mb_new_inode_pa(ac);
3531 * finds all unused blocks in on-disk bitmap, frees them in
3532 * in-core bitmap and buddy.
3533 * @pa must be unlinked from inode and group lists, so that
3534 * nobody else can find/use it.
3535 * the caller MUST hold group/inode locks.
3536 * TODO: optimize the case when there are no in-core structures yet
3538 static noinline_for_stack int
3539 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3540 struct ext4_prealloc_space *pa,
3541 struct ext4_allocation_context *ac)
3543 struct super_block *sb = e4b->bd_sb;
3544 struct ext4_sb_info *sbi = EXT4_SB(sb);
3549 unsigned long long grp_blk_start;
3554 BUG_ON(pa->pa_deleted == 0);
3555 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3556 grp_blk_start = pa->pa_pstart - bit;
3557 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3558 end = bit + pa->pa_len;
3562 ac->ac_inode = pa->pa_inode;
3566 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3569 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3570 start = ext4_group_first_block_no(sb, group) + bit;
3571 mb_debug(1, " free preallocated %u/%u in group %u\n",
3572 (unsigned) start, (unsigned) next - bit,
3577 ac->ac_b_ex.fe_group = group;
3578 ac->ac_b_ex.fe_start = bit;
3579 ac->ac_b_ex.fe_len = next - bit;
3580 ac->ac_b_ex.fe_logical = 0;
3581 trace_ext4_mballoc_discard(ac);
3584 trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit,
3586 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3589 if (free != pa->pa_free) {
3590 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3591 pa, (unsigned long) pa->pa_lstart,
3592 (unsigned long) pa->pa_pstart,
3593 (unsigned long) pa->pa_len);
3594 ext4_grp_locked_error(sb, group,
3595 __func__, "free %u, pa_free %u",
3598 * pa is already deleted so we use the value obtained
3599 * from the bitmap and continue.
3602 atomic_add(free, &sbi->s_mb_discarded);
3607 static noinline_for_stack int
3608 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3609 struct ext4_prealloc_space *pa,
3610 struct ext4_allocation_context *ac)
3612 struct super_block *sb = e4b->bd_sb;
3616 trace_ext4_mb_release_group_pa(ac, pa);
3617 BUG_ON(pa->pa_deleted == 0);
3618 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3619 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3620 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3621 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3625 ac->ac_inode = NULL;
3626 ac->ac_b_ex.fe_group = group;
3627 ac->ac_b_ex.fe_start = bit;
3628 ac->ac_b_ex.fe_len = pa->pa_len;
3629 ac->ac_b_ex.fe_logical = 0;
3630 trace_ext4_mballoc_discard(ac);
3637 * releases all preallocations in given group
3639 * first, we need to decide discard policy:
3640 * - when do we discard
3642 * - how many do we discard
3643 * 1) how many requested
3645 static noinline_for_stack int
3646 ext4_mb_discard_group_preallocations(struct super_block *sb,
3647 ext4_group_t group, int needed)
3649 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3650 struct buffer_head *bitmap_bh = NULL;
3651 struct ext4_prealloc_space *pa, *tmp;
3652 struct ext4_allocation_context *ac;
3653 struct list_head list;
3654 struct ext4_buddy e4b;
3659 mb_debug(1, "discard preallocation for group %u\n", group);
3661 if (list_empty(&grp->bb_prealloc_list))
3664 bitmap_bh = ext4_read_block_bitmap(sb, group);
3665 if (bitmap_bh == NULL) {
3666 ext4_error(sb, "Error reading block bitmap for %u", group);
3670 err = ext4_mb_load_buddy(sb, group, &e4b);
3672 ext4_error(sb, "Error loading buddy information for %u", group);
3678 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3680 INIT_LIST_HEAD(&list);
3681 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3685 ext4_lock_group(sb, group);
3686 list_for_each_entry_safe(pa, tmp,
3687 &grp->bb_prealloc_list, pa_group_list) {
3688 spin_lock(&pa->pa_lock);
3689 if (atomic_read(&pa->pa_count)) {
3690 spin_unlock(&pa->pa_lock);
3694 if (pa->pa_deleted) {
3695 spin_unlock(&pa->pa_lock);
3699 /* seems this one can be freed ... */
3702 /* we can trust pa_free ... */
3703 free += pa->pa_free;
3705 spin_unlock(&pa->pa_lock);
3707 list_del(&pa->pa_group_list);
3708 list_add(&pa->u.pa_tmp_list, &list);
3711 /* if we still need more blocks and some PAs were used, try again */
3712 if (free < needed && busy) {
3714 ext4_unlock_group(sb, group);
3716 * Yield the CPU here so that we don't get soft lockup
3717 * in non preempt case.
3723 /* found anything to free? */
3724 if (list_empty(&list)) {
3729 /* now free all selected PAs */
3730 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3732 /* remove from object (inode or locality group) */
3733 spin_lock(pa->pa_obj_lock);
3734 list_del_rcu(&pa->pa_inode_list);
3735 spin_unlock(pa->pa_obj_lock);
3737 if (pa->pa_type == MB_GROUP_PA)
3738 ext4_mb_release_group_pa(&e4b, pa, ac);
3740 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3742 list_del(&pa->u.pa_tmp_list);
3743 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3747 ext4_unlock_group(sb, group);
3749 kmem_cache_free(ext4_ac_cachep, ac);
3750 ext4_mb_unload_buddy(&e4b);
3756 * releases all non-used preallocated blocks for given inode
3758 * It's important to discard preallocations under i_data_sem
3759 * We don't want another block to be served from the prealloc
3760 * space when we are discarding the inode prealloc space.
3762 * FIXME!! Make sure it is valid at all the call sites
3764 void ext4_discard_preallocations(struct inode *inode)
3766 struct ext4_inode_info *ei = EXT4_I(inode);
3767 struct super_block *sb = inode->i_sb;
3768 struct buffer_head *bitmap_bh = NULL;
3769 struct ext4_prealloc_space *pa, *tmp;
3770 struct ext4_allocation_context *ac;
3771 ext4_group_t group = 0;
3772 struct list_head list;
3773 struct ext4_buddy e4b;
3776 if (!S_ISREG(inode->i_mode)) {
3777 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3781 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3782 trace_ext4_discard_preallocations(inode);
3784 INIT_LIST_HEAD(&list);
3786 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3789 ac->ac_inode = inode;
3792 /* first, collect all pa's in the inode */
3793 spin_lock(&ei->i_prealloc_lock);
3794 while (!list_empty(&ei->i_prealloc_list)) {
3795 pa = list_entry(ei->i_prealloc_list.next,
3796 struct ext4_prealloc_space, pa_inode_list);
3797 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3798 spin_lock(&pa->pa_lock);
3799 if (atomic_read(&pa->pa_count)) {
3800 /* this shouldn't happen often - nobody should
3801 * use preallocation while we're discarding it */
3802 spin_unlock(&pa->pa_lock);
3803 spin_unlock(&ei->i_prealloc_lock);
3804 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3806 schedule_timeout_uninterruptible(HZ);
3810 if (pa->pa_deleted == 0) {
3812 spin_unlock(&pa->pa_lock);
3813 list_del_rcu(&pa->pa_inode_list);
3814 list_add(&pa->u.pa_tmp_list, &list);
3818 /* someone is deleting pa right now */
3819 spin_unlock(&pa->pa_lock);
3820 spin_unlock(&ei->i_prealloc_lock);
3822 /* we have to wait here because pa_deleted
3823 * doesn't mean pa is already unlinked from
3824 * the list. as we might be called from
3825 * ->clear_inode() the inode will get freed
3826 * and concurrent thread which is unlinking
3827 * pa from inode's list may access already
3828 * freed memory, bad-bad-bad */
3830 /* XXX: if this happens too often, we can
3831 * add a flag to force wait only in case
3832 * of ->clear_inode(), but not in case of
3833 * regular truncate */
3834 schedule_timeout_uninterruptible(HZ);
3837 spin_unlock(&ei->i_prealloc_lock);
3839 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3840 BUG_ON(pa->pa_type != MB_INODE_PA);
3841 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3843 err = ext4_mb_load_buddy(sb, group, &e4b);
3845 ext4_error(sb, "Error loading buddy information for %u",
3850 bitmap_bh = ext4_read_block_bitmap(sb, group);
3851 if (bitmap_bh == NULL) {
3852 ext4_error(sb, "Error reading block bitmap for %u",
3854 ext4_mb_unload_buddy(&e4b);
3858 ext4_lock_group(sb, group);
3859 list_del(&pa->pa_group_list);
3860 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3861 ext4_unlock_group(sb, group);
3863 ext4_mb_unload_buddy(&e4b);
3866 list_del(&pa->u.pa_tmp_list);
3867 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3870 kmem_cache_free(ext4_ac_cachep, ac);
3874 * finds all preallocated spaces and return blocks being freed to them
3875 * if preallocated space becomes full (no block is used from the space)
3876 * then the function frees space in buddy
3877 * XXX: at the moment, truncate (which is the only way to free blocks)
3878 * discards all preallocations
3880 static void ext4_mb_return_to_preallocation(struct inode *inode,
3881 struct ext4_buddy *e4b,
3882 sector_t block, int count)
3884 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
3886 #ifdef CONFIG_EXT4_DEBUG
3887 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3889 struct super_block *sb = ac->ac_sb;
3890 ext4_group_t ngroups, i;
3892 printk(KERN_ERR "EXT4-fs: Can't allocate:"
3893 " Allocation context details:\n");
3894 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
3895 ac->ac_status, ac->ac_flags);
3896 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3897 "best %lu/%lu/%lu@%lu cr %d\n",
3898 (unsigned long)ac->ac_o_ex.fe_group,
3899 (unsigned long)ac->ac_o_ex.fe_start,
3900 (unsigned long)ac->ac_o_ex.fe_len,
3901 (unsigned long)ac->ac_o_ex.fe_logical,
3902 (unsigned long)ac->ac_g_ex.fe_group,
3903 (unsigned long)ac->ac_g_ex.fe_start,
3904 (unsigned long)ac->ac_g_ex.fe_len,
3905 (unsigned long)ac->ac_g_ex.fe_logical,
3906 (unsigned long)ac->ac_b_ex.fe_group,
3907 (unsigned long)ac->ac_b_ex.fe_start,
3908 (unsigned long)ac->ac_b_ex.fe_len,
3909 (unsigned long)ac->ac_b_ex.fe_logical,
3910 (int)ac->ac_criteria);
3911 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
3913 printk(KERN_ERR "EXT4-fs: groups: \n");
3914 ngroups = ext4_get_groups_count(sb);
3915 for (i = 0; i < ngroups; i++) {
3916 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3917 struct ext4_prealloc_space *pa;
3918 ext4_grpblk_t start;
3919 struct list_head *cur;
3920 ext4_lock_group(sb, i);
3921 list_for_each(cur, &grp->bb_prealloc_list) {
3922 pa = list_entry(cur, struct ext4_prealloc_space,
3924 spin_lock(&pa->pa_lock);
3925 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3927 spin_unlock(&pa->pa_lock);
3928 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3931 ext4_unlock_group(sb, i);
3933 if (grp->bb_free == 0)
3935 printk(KERN_ERR "%u: %d/%d \n",
3936 i, grp->bb_free, grp->bb_fragments);
3938 printk(KERN_ERR "\n");
3941 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3948 * We use locality group preallocation for small size file. The size of the
3949 * file is determined by the current size or the resulting size after
3950 * allocation which ever is larger
3952 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3954 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3956 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3957 int bsbits = ac->ac_sb->s_blocksize_bits;
3960 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3963 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3966 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3967 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3970 if ((size == isize) &&
3971 !ext4_fs_is_busy(sbi) &&
3972 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3973 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3977 /* don't use group allocation for large files */
3978 size = max(size, isize);
3979 if (size > sbi->s_mb_stream_request) {
3980 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3984 BUG_ON(ac->ac_lg != NULL);
3986 * locality group prealloc space are per cpu. The reason for having
3987 * per cpu locality group is to reduce the contention between block
3988 * request from multiple CPUs.
3990 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
3992 /* we're going to use group allocation */
3993 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3995 /* serialize all allocations in the group */
3996 mutex_lock(&ac->ac_lg->lg_mutex);
3999 static noinline_for_stack int
4000 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4001 struct ext4_allocation_request *ar)
4003 struct super_block *sb = ar->inode->i_sb;
4004 struct ext4_sb_info *sbi = EXT4_SB(sb);
4005 struct ext4_super_block *es = sbi->s_es;
4009 ext4_grpblk_t block;
4011 /* we can't allocate > group size */
4014 /* just a dirty hack to filter too big requests */
4015 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4016 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4018 /* start searching from the goal */
4020 if (goal < le32_to_cpu(es->s_first_data_block) ||
4021 goal >= ext4_blocks_count(es))
4022 goal = le32_to_cpu(es->s_first_data_block);
4023 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4025 /* set up allocation goals */
4026 memset(ac, 0, sizeof(struct ext4_allocation_context));
4027 ac->ac_b_ex.fe_logical = ar->logical;
4028 ac->ac_status = AC_STATUS_CONTINUE;
4030 ac->ac_inode = ar->inode;
4031 ac->ac_o_ex.fe_logical = ar->logical;
4032 ac->ac_o_ex.fe_group = group;
4033 ac->ac_o_ex.fe_start = block;
4034 ac->ac_o_ex.fe_len = len;
4035 ac->ac_g_ex.fe_logical = ar->logical;
4036 ac->ac_g_ex.fe_group = group;
4037 ac->ac_g_ex.fe_start = block;
4038 ac->ac_g_ex.fe_len = len;
4039 ac->ac_flags = ar->flags;
4041 /* we have to define context: we'll we work with a file or
4042 * locality group. this is a policy, actually */
4043 ext4_mb_group_or_file(ac);
4045 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4046 "left: %u/%u, right %u/%u to %swritable\n",
4047 (unsigned) ar->len, (unsigned) ar->logical,
4048 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4049 (unsigned) ar->lleft, (unsigned) ar->pleft,
4050 (unsigned) ar->lright, (unsigned) ar->pright,
4051 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4056 static noinline_for_stack void
4057 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4058 struct ext4_locality_group *lg,
4059 int order, int total_entries)
4061 ext4_group_t group = 0;
4062 struct ext4_buddy e4b;
4063 struct list_head discard_list;
4064 struct ext4_prealloc_space *pa, *tmp;
4065 struct ext4_allocation_context *ac;
4067 mb_debug(1, "discard locality group preallocation\n");
4069 INIT_LIST_HEAD(&discard_list);
4070 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4074 spin_lock(&lg->lg_prealloc_lock);
4075 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4077 spin_lock(&pa->pa_lock);
4078 if (atomic_read(&pa->pa_count)) {
4080 * This is the pa that we just used
4081 * for block allocation. So don't
4084 spin_unlock(&pa->pa_lock);
4087 if (pa->pa_deleted) {
4088 spin_unlock(&pa->pa_lock);
4091 /* only lg prealloc space */
4092 BUG_ON(pa->pa_type != MB_GROUP_PA);
4094 /* seems this one can be freed ... */
4096 spin_unlock(&pa->pa_lock);
4098 list_del_rcu(&pa->pa_inode_list);
4099 list_add(&pa->u.pa_tmp_list, &discard_list);
4102 if (total_entries <= 5) {
4104 * we want to keep only 5 entries
4105 * allowing it to grow to 8. This
4106 * mak sure we don't call discard
4107 * soon for this list.
4112 spin_unlock(&lg->lg_prealloc_lock);
4114 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4116 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4117 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4118 ext4_error(sb, "Error loading buddy information for %u",
4122 ext4_lock_group(sb, group);
4123 list_del(&pa->pa_group_list);
4124 ext4_mb_release_group_pa(&e4b, pa, ac);
4125 ext4_unlock_group(sb, group);
4127 ext4_mb_unload_buddy(&e4b);
4128 list_del(&pa->u.pa_tmp_list);
4129 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4132 kmem_cache_free(ext4_ac_cachep, ac);
4136 * We have incremented pa_count. So it cannot be freed at this
4137 * point. Also we hold lg_mutex. So no parallel allocation is
4138 * possible from this lg. That means pa_free cannot be updated.
4140 * A parallel ext4_mb_discard_group_preallocations is possible.
4141 * which can cause the lg_prealloc_list to be updated.
4144 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4146 int order, added = 0, lg_prealloc_count = 1;
4147 struct super_block *sb = ac->ac_sb;
4148 struct ext4_locality_group *lg = ac->ac_lg;
4149 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4151 order = fls(pa->pa_free) - 1;
4152 if (order > PREALLOC_TB_SIZE - 1)
4153 /* The max size of hash table is PREALLOC_TB_SIZE */
4154 order = PREALLOC_TB_SIZE - 1;
4155 /* Add the prealloc space to lg */
4157 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4159 spin_lock(&tmp_pa->pa_lock);
4160 if (tmp_pa->pa_deleted) {
4161 spin_unlock(&tmp_pa->pa_lock);
4164 if (!added && pa->pa_free < tmp_pa->pa_free) {
4165 /* Add to the tail of the previous entry */
4166 list_add_tail_rcu(&pa->pa_inode_list,
4167 &tmp_pa->pa_inode_list);
4170 * we want to count the total
4171 * number of entries in the list
4174 spin_unlock(&tmp_pa->pa_lock);
4175 lg_prealloc_count++;
4178 list_add_tail_rcu(&pa->pa_inode_list,
4179 &lg->lg_prealloc_list[order]);
4182 /* Now trim the list to be not more than 8 elements */
4183 if (lg_prealloc_count > 8) {
4184 ext4_mb_discard_lg_preallocations(sb, lg,
4185 order, lg_prealloc_count);
4192 * release all resource we used in allocation
4194 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4196 struct ext4_prealloc_space *pa = ac->ac_pa;
4198 if (pa->pa_type == MB_GROUP_PA) {
4199 /* see comment in ext4_mb_use_group_pa() */
4200 spin_lock(&pa->pa_lock);
4201 pa->pa_pstart += ac->ac_b_ex.fe_len;
4202 pa->pa_lstart += ac->ac_b_ex.fe_len;
4203 pa->pa_free -= ac->ac_b_ex.fe_len;
4204 pa->pa_len -= ac->ac_b_ex.fe_len;
4205 spin_unlock(&pa->pa_lock);
4209 up_read(ac->alloc_semp);
4212 * We want to add the pa to the right bucket.
4213 * Remove it from the list and while adding
4214 * make sure the list to which we are adding
4215 * doesn't grow big. We need to release
4216 * alloc_semp before calling ext4_mb_add_n_trim()
4218 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4219 spin_lock(pa->pa_obj_lock);
4220 list_del_rcu(&pa->pa_inode_list);
4221 spin_unlock(pa->pa_obj_lock);
4222 ext4_mb_add_n_trim(ac);
4224 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4226 if (ac->ac_bitmap_page)
4227 page_cache_release(ac->ac_bitmap_page);
4228 if (ac->ac_buddy_page)
4229 page_cache_release(ac->ac_buddy_page);
4230 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4231 mutex_unlock(&ac->ac_lg->lg_mutex);
4232 ext4_mb_collect_stats(ac);
4236 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4238 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4242 trace_ext4_mb_discard_preallocations(sb, needed);
4243 for (i = 0; i < ngroups && needed > 0; i++) {
4244 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4253 * Main entry point into mballoc to allocate blocks
4254 * it tries to use preallocation first, then falls back
4255 * to usual allocation
4257 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4258 struct ext4_allocation_request *ar, int *errp)
4261 struct ext4_allocation_context *ac = NULL;
4262 struct ext4_sb_info *sbi;
4263 struct super_block *sb;
4264 ext4_fsblk_t block = 0;
4265 unsigned int inquota = 0;
4266 unsigned int reserv_blks = 0;
4268 sb = ar->inode->i_sb;
4271 trace_ext4_request_blocks(ar);
4274 * For delayed allocation, we could skip the ENOSPC and
4275 * EDQUOT check, as blocks and quotas have been already
4276 * reserved when data being copied into pagecache.
4278 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4279 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4281 /* Without delayed allocation we need to verify
4282 * there is enough free blocks to do block allocation
4283 * and verify allocation doesn't exceed the quota limits.
4285 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4286 /* let others to free the space */
4288 ar->len = ar->len >> 1;
4294 reserv_blks = ar->len;
4295 while (ar->len && dquot_alloc_block(ar->inode, ar->len)) {
4296 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4306 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4313 *errp = ext4_mb_initialize_context(ac, ar);
4319 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4320 if (!ext4_mb_use_preallocated(ac)) {
4321 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4322 ext4_mb_normalize_request(ac, ar);
4324 /* allocate space in core */
4325 ext4_mb_regular_allocator(ac);
4327 /* as we've just preallocated more space than
4328 * user requested orinally, we store allocated
4329 * space in a special descriptor */
4330 if (ac->ac_status == AC_STATUS_FOUND &&
4331 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4332 ext4_mb_new_preallocation(ac);
4334 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4335 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4336 if (*errp == -EAGAIN) {
4338 * drop the reference that we took
4339 * in ext4_mb_use_best_found
4341 ext4_mb_release_context(ac);
4342 ac->ac_b_ex.fe_group = 0;
4343 ac->ac_b_ex.fe_start = 0;
4344 ac->ac_b_ex.fe_len = 0;
4345 ac->ac_status = AC_STATUS_CONTINUE;
4348 ext4_discard_allocated_blocks(ac);
4349 ac->ac_b_ex.fe_len = 0;
4351 ext4_mb_show_ac(ac);
4353 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4354 ar->len = ac->ac_b_ex.fe_len;
4357 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4361 ac->ac_b_ex.fe_len = 0;
4363 ext4_mb_show_ac(ac);
4366 ext4_mb_release_context(ac);
4369 kmem_cache_free(ext4_ac_cachep, ac);
4371 if (inquota && ar->len < inquota)
4372 dquot_free_block(ar->inode, inquota - ar->len);
4375 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4376 /* release all the reserved blocks if non delalloc */
4377 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4381 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4387 * We can merge two free data extents only if the physical blocks
4388 * are contiguous, AND the extents were freed by the same transaction,
4389 * AND the blocks are associated with the same group.
4391 static int can_merge(struct ext4_free_data *entry1,
4392 struct ext4_free_data *entry2)
4394 if ((entry1->t_tid == entry2->t_tid) &&
4395 (entry1->group == entry2->group) &&
4396 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4401 static noinline_for_stack int
4402 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4403 struct ext4_free_data *new_entry)
4405 ext4_grpblk_t block;
4406 struct ext4_free_data *entry;
4407 struct ext4_group_info *db = e4b->bd_info;
4408 struct super_block *sb = e4b->bd_sb;
4409 struct ext4_sb_info *sbi = EXT4_SB(sb);
4410 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4411 struct rb_node *parent = NULL, *new_node;
4413 BUG_ON(!ext4_handle_valid(handle));
4414 BUG_ON(e4b->bd_bitmap_page == NULL);
4415 BUG_ON(e4b->bd_buddy_page == NULL);
4417 new_node = &new_entry->node;
4418 block = new_entry->start_blk;
4421 /* first free block exent. We need to
4422 protect buddy cache from being freed,
4423 * otherwise we'll refresh it from
4424 * on-disk bitmap and lose not-yet-available
4426 page_cache_get(e4b->bd_buddy_page);
4427 page_cache_get(e4b->bd_bitmap_page);
4431 entry = rb_entry(parent, struct ext4_free_data, node);
4432 if (block < entry->start_blk)
4434 else if (block >= (entry->start_blk + entry->count))
4435 n = &(*n)->rb_right;
4437 ext4_grp_locked_error(sb, e4b->bd_group, __func__,
4438 "Double free of blocks %d (%d %d)",
4439 block, entry->start_blk, entry->count);
4444 rb_link_node(new_node, parent, n);
4445 rb_insert_color(new_node, &db->bb_free_root);
4447 /* Now try to see the extent can be merged to left and right */
4448 node = rb_prev(new_node);
4450 entry = rb_entry(node, struct ext4_free_data, node);
4451 if (can_merge(entry, new_entry)) {
4452 new_entry->start_blk = entry->start_blk;
4453 new_entry->count += entry->count;
4454 rb_erase(node, &(db->bb_free_root));
4455 spin_lock(&sbi->s_md_lock);
4456 list_del(&entry->list);
4457 spin_unlock(&sbi->s_md_lock);
4458 kmem_cache_free(ext4_free_ext_cachep, entry);
4462 node = rb_next(new_node);
4464 entry = rb_entry(node, struct ext4_free_data, node);
4465 if (can_merge(new_entry, entry)) {
4466 new_entry->count += entry->count;
4467 rb_erase(node, &(db->bb_free_root));
4468 spin_lock(&sbi->s_md_lock);
4469 list_del(&entry->list);
4470 spin_unlock(&sbi->s_md_lock);
4471 kmem_cache_free(ext4_free_ext_cachep, entry);
4474 /* Add the extent to transaction's private list */
4475 spin_lock(&sbi->s_md_lock);
4476 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4477 spin_unlock(&sbi->s_md_lock);
4482 * ext4_free_blocks() -- Free given blocks and update quota
4483 * @handle: handle for this transaction
4485 * @block: start physical block to free
4486 * @count: number of blocks to count
4487 * @metadata: Are these metadata blocks
4489 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4490 struct buffer_head *bh, ext4_fsblk_t block,
4491 unsigned long count, int flags)
4493 struct buffer_head *bitmap_bh = NULL;
4494 struct super_block *sb = inode->i_sb;
4495 struct ext4_allocation_context *ac = NULL;
4496 struct ext4_group_desc *gdp;
4497 struct ext4_super_block *es;
4498 unsigned long freed = 0;
4499 unsigned int overflow;
4501 struct buffer_head *gd_bh;
4502 ext4_group_t block_group;
4503 struct ext4_sb_info *sbi;
4504 struct ext4_buddy e4b;
4510 BUG_ON(block != bh->b_blocknr);
4512 block = bh->b_blocknr;
4516 es = EXT4_SB(sb)->s_es;
4517 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4518 !ext4_data_block_valid(sbi, block, count)) {
4519 ext4_error(sb, "Freeing blocks not in datazone - "
4520 "block = %llu, count = %lu", block, count);
4524 ext4_debug("freeing block %llu\n", block);
4525 trace_ext4_free_blocks(inode, block, count, flags);
4527 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4528 struct buffer_head *tbh = bh;
4531 BUG_ON(bh && (count > 1));
4533 for (i = 0; i < count; i++) {
4535 tbh = sb_find_get_block(inode->i_sb,
4537 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4538 inode, tbh, block + i);
4543 * We need to make sure we don't reuse the freed block until
4544 * after the transaction is committed, which we can do by
4545 * treating the block as metadata, below. We make an
4546 * exception if the inode is to be written in writeback mode
4547 * since writeback mode has weak data consistency guarantees.
4549 if (!ext4_should_writeback_data(inode))
4550 flags |= EXT4_FREE_BLOCKS_METADATA;
4552 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4554 ac->ac_inode = inode;
4560 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4563 * Check to see if we are freeing blocks across a group
4566 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4567 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4570 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4575 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4581 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4582 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4583 in_range(block, ext4_inode_table(sb, gdp),
4584 EXT4_SB(sb)->s_itb_per_group) ||
4585 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4586 EXT4_SB(sb)->s_itb_per_group)) {
4588 ext4_error(sb, "Freeing blocks in system zone - "
4589 "Block = %llu, count = %lu", block, count);
4590 /* err = 0. ext4_std_error should be a no op */
4594 BUFFER_TRACE(bitmap_bh, "getting write access");
4595 err = ext4_journal_get_write_access(handle, bitmap_bh);
4600 * We are about to modify some metadata. Call the journal APIs
4601 * to unshare ->b_data if a currently-committing transaction is
4604 BUFFER_TRACE(gd_bh, "get_write_access");
4605 err = ext4_journal_get_write_access(handle, gd_bh);
4608 #ifdef AGGRESSIVE_CHECK
4611 for (i = 0; i < count; i++)
4612 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4616 ac->ac_b_ex.fe_group = block_group;
4617 ac->ac_b_ex.fe_start = bit;
4618 ac->ac_b_ex.fe_len = count;
4619 trace_ext4_mballoc_free(ac);
4622 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4626 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4627 struct ext4_free_data *new_entry;
4629 * blocks being freed are metadata. these blocks shouldn't
4630 * be used until this transaction is committed
4632 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4633 new_entry->start_blk = bit;
4634 new_entry->group = block_group;
4635 new_entry->count = count;
4636 new_entry->t_tid = handle->h_transaction->t_tid;
4638 ext4_lock_group(sb, block_group);
4639 mb_clear_bits(bitmap_bh->b_data, bit, count);
4640 ext4_mb_free_metadata(handle, &e4b, new_entry);
4642 /* need to update group_info->bb_free and bitmap
4643 * with group lock held. generate_buddy look at
4644 * them with group lock_held
4646 ext4_lock_group(sb, block_group);
4647 mb_clear_bits(bitmap_bh->b_data, bit, count);
4648 mb_free_blocks(inode, &e4b, bit, count);
4649 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4652 ret = ext4_free_blks_count(sb, gdp) + count;
4653 ext4_free_blks_set(sb, gdp, ret);
4654 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4655 ext4_unlock_group(sb, block_group);
4656 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4658 if (sbi->s_log_groups_per_flex) {
4659 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4660 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4663 ext4_mb_unload_buddy(&e4b);
4667 /* We dirtied the bitmap block */
4668 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4669 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4671 /* And the group descriptor block */
4672 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4673 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4677 if (overflow && !err) {
4686 dquot_free_block(inode, freed);
4688 ext4_std_error(sb, err);
4690 kmem_cache_free(ext4_ac_cachep, ac);