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 inode ? inode->i_ino : 0,
451 "freeing block already freed "
455 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
459 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
463 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
465 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
466 for (i = 0; i < count; i++) {
467 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
468 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
472 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
474 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
475 unsigned char *b1, *b2;
477 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
478 b2 = (unsigned char *) bitmap;
479 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
480 if (b1[i] != b2[i]) {
481 printk(KERN_ERR "corruption in group %u "
482 "at byte %u(%u): %x in copy != %x "
483 "on disk/prealloc\n",
484 e4b->bd_group, i, i * 8, b1[i], b2[i]);
492 static inline void mb_free_blocks_double(struct inode *inode,
493 struct ext4_buddy *e4b, int first, int count)
497 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
498 int first, int count)
502 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
508 #ifdef AGGRESSIVE_CHECK
510 #define MB_CHECK_ASSERT(assert) \
514 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
515 function, file, line, # assert); \
520 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
521 const char *function, int line)
523 struct super_block *sb = e4b->bd_sb;
524 int order = e4b->bd_blkbits + 1;
531 struct ext4_group_info *grp;
534 struct list_head *cur;
539 static int mb_check_counter;
540 if (mb_check_counter++ % 100 != 0)
545 buddy = mb_find_buddy(e4b, order, &max);
546 MB_CHECK_ASSERT(buddy);
547 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
548 MB_CHECK_ASSERT(buddy2);
549 MB_CHECK_ASSERT(buddy != buddy2);
550 MB_CHECK_ASSERT(max * 2 == max2);
553 for (i = 0; i < max; i++) {
555 if (mb_test_bit(i, buddy)) {
556 /* only single bit in buddy2 may be 1 */
557 if (!mb_test_bit(i << 1, buddy2)) {
559 mb_test_bit((i<<1)+1, buddy2));
560 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
562 mb_test_bit(i << 1, buddy2));
567 /* both bits in buddy2 must be 0 */
568 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
569 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
571 for (j = 0; j < (1 << order); j++) {
572 k = (i * (1 << order)) + j;
574 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
578 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
583 buddy = mb_find_buddy(e4b, 0, &max);
584 for (i = 0; i < max; i++) {
585 if (!mb_test_bit(i, buddy)) {
586 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
594 /* check used bits only */
595 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
596 buddy2 = mb_find_buddy(e4b, j, &max2);
598 MB_CHECK_ASSERT(k < max2);
599 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
602 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
603 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
605 grp = ext4_get_group_info(sb, e4b->bd_group);
606 buddy = mb_find_buddy(e4b, 0, &max);
607 list_for_each(cur, &grp->bb_prealloc_list) {
608 ext4_group_t groupnr;
609 struct ext4_prealloc_space *pa;
610 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
611 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
612 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
613 for (i = 0; i < pa->pa_len; i++)
614 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
618 #undef MB_CHECK_ASSERT
619 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
620 __FILE__, __func__, __LINE__)
622 #define mb_check_buddy(e4b)
625 /* FIXME!! need more doc */
626 static void ext4_mb_mark_free_simple(struct super_block *sb,
627 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
628 struct ext4_group_info *grp)
630 struct ext4_sb_info *sbi = EXT4_SB(sb);
634 unsigned short border;
636 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
638 border = 2 << sb->s_blocksize_bits;
641 /* find how many blocks can be covered since this position */
642 max = ffs(first | border) - 1;
644 /* find how many blocks of power 2 we need to mark */
651 /* mark multiblock chunks only */
652 grp->bb_counters[min]++;
654 mb_clear_bit(first >> min,
655 buddy + sbi->s_mb_offsets[min]);
663 * Cache the order of the largest free extent we have available in this block
667 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
672 grp->bb_largest_free_order = -1; /* uninit */
674 bits = sb->s_blocksize_bits + 1;
675 for (i = bits; i >= 0; i--) {
676 if (grp->bb_counters[i] > 0) {
677 grp->bb_largest_free_order = i;
683 static noinline_for_stack
684 void ext4_mb_generate_buddy(struct super_block *sb,
685 void *buddy, void *bitmap, ext4_group_t group)
687 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
688 ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb);
693 unsigned fragments = 0;
694 unsigned long long period = get_cycles();
696 /* initialize buddy from bitmap which is aggregation
697 * of on-disk bitmap and preallocations */
698 i = mb_find_next_zero_bit(bitmap, max, 0);
699 grp->bb_first_free = i;
703 i = mb_find_next_bit(bitmap, max, i);
707 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
709 grp->bb_counters[0]++;
711 i = mb_find_next_zero_bit(bitmap, max, i);
713 grp->bb_fragments = fragments;
715 if (free != grp->bb_free) {
716 ext4_grp_locked_error(sb, group, 0, 0,
717 "%u blocks in bitmap, %u in gd",
720 * If we intent to continue, we consider group descritor
721 * corrupt and update bb_free using bitmap value
725 mb_set_largest_free_order(sb, grp);
727 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
729 period = get_cycles() - period;
730 spin_lock(&EXT4_SB(sb)->s_bal_lock);
731 EXT4_SB(sb)->s_mb_buddies_generated++;
732 EXT4_SB(sb)->s_mb_generation_time += period;
733 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
736 /* The buddy information is attached the buddy cache inode
737 * for convenience. The information regarding each group
738 * is loaded via ext4_mb_load_buddy. The information involve
739 * block bitmap and buddy information. The information are
740 * stored in the inode as
743 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
746 * one block each for bitmap and buddy information.
747 * So for each group we take up 2 blocks. A page can
748 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
749 * So it can have information regarding groups_per_page which
750 * is blocks_per_page/2
752 * Locking note: This routine takes the block group lock of all groups
753 * for this page; do not hold this lock when calling this routine!
756 static int ext4_mb_init_cache(struct page *page, char *incore)
758 ext4_group_t ngroups;
764 ext4_group_t first_group;
766 struct super_block *sb;
767 struct buffer_head *bhs;
768 struct buffer_head **bh;
773 mb_debug(1, "init page %lu\n", page->index);
775 inode = page->mapping->host;
777 ngroups = ext4_get_groups_count(sb);
778 blocksize = 1 << inode->i_blkbits;
779 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
781 groups_per_page = blocks_per_page >> 1;
782 if (groups_per_page == 0)
785 /* allocate buffer_heads to read bitmaps */
786 if (groups_per_page > 1) {
788 i = sizeof(struct buffer_head *) * groups_per_page;
789 bh = kzalloc(i, GFP_NOFS);
795 first_group = page->index * blocks_per_page / 2;
797 /* read all groups the page covers into the cache */
798 for (i = 0; i < groups_per_page; i++) {
799 struct ext4_group_desc *desc;
801 if (first_group + i >= ngroups)
805 desc = ext4_get_group_desc(sb, first_group + i, NULL);
810 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
814 if (bitmap_uptodate(bh[i]))
818 if (bitmap_uptodate(bh[i])) {
819 unlock_buffer(bh[i]);
822 ext4_lock_group(sb, first_group + i);
823 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
824 ext4_init_block_bitmap(sb, bh[i],
825 first_group + i, desc);
826 set_bitmap_uptodate(bh[i]);
827 set_buffer_uptodate(bh[i]);
828 ext4_unlock_group(sb, first_group + i);
829 unlock_buffer(bh[i]);
832 ext4_unlock_group(sb, first_group + i);
833 if (buffer_uptodate(bh[i])) {
835 * if not uninit if bh is uptodate,
836 * bitmap is also uptodate
838 set_bitmap_uptodate(bh[i]);
839 unlock_buffer(bh[i]);
844 * submit the buffer_head for read. We can
845 * safely mark the bitmap as uptodate now.
846 * We do it here so the bitmap uptodate bit
847 * get set with buffer lock held.
849 set_bitmap_uptodate(bh[i]);
850 bh[i]->b_end_io = end_buffer_read_sync;
851 submit_bh(READ, bh[i]);
852 mb_debug(1, "read bitmap for group %u\n", first_group + i);
855 /* wait for I/O completion */
856 for (i = 0; i < groups_per_page && bh[i]; i++)
857 wait_on_buffer(bh[i]);
860 for (i = 0; i < groups_per_page && bh[i]; i++)
861 if (!buffer_uptodate(bh[i]))
865 first_block = page->index * blocks_per_page;
867 memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
868 for (i = 0; i < blocks_per_page; i++) {
870 struct ext4_group_info *grinfo;
872 group = (first_block + i) >> 1;
873 if (group >= ngroups)
877 * data carry information regarding this
878 * particular group in the format specified
882 data = page_address(page) + (i * blocksize);
883 bitmap = bh[group - first_group]->b_data;
886 * We place the buddy block and bitmap block
889 if ((first_block + i) & 1) {
890 /* this is block of buddy */
891 BUG_ON(incore == NULL);
892 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
893 group, page->index, i * blocksize);
894 trace_ext4_mb_buddy_bitmap_load(sb, group);
895 grinfo = ext4_get_group_info(sb, group);
896 grinfo->bb_fragments = 0;
897 memset(grinfo->bb_counters, 0,
898 sizeof(*grinfo->bb_counters) *
899 (sb->s_blocksize_bits+2));
901 * incore got set to the group block bitmap below
903 ext4_lock_group(sb, group);
904 ext4_mb_generate_buddy(sb, data, incore, group);
905 ext4_unlock_group(sb, group);
908 /* this is block of bitmap */
909 BUG_ON(incore != NULL);
910 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
911 group, page->index, i * blocksize);
912 trace_ext4_mb_bitmap_load(sb, group);
914 /* see comments in ext4_mb_put_pa() */
915 ext4_lock_group(sb, group);
916 memcpy(data, bitmap, blocksize);
918 /* mark all preallocated blks used in in-core bitmap */
919 ext4_mb_generate_from_pa(sb, data, group);
920 ext4_mb_generate_from_freelist(sb, data, group);
921 ext4_unlock_group(sb, group);
923 /* set incore so that the buddy information can be
924 * generated using this
929 SetPageUptodate(page);
933 for (i = 0; i < groups_per_page && bh[i]; i++)
942 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
943 * block group lock of all groups for this page; do not hold the BG lock when
944 * calling this routine!
946 static noinline_for_stack
947 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
953 int block, pnum, poff;
954 int num_grp_locked = 0;
955 struct ext4_group_info *this_grp;
956 struct ext4_sb_info *sbi = EXT4_SB(sb);
957 struct inode *inode = sbi->s_buddy_cache;
958 struct page *page = NULL, *bitmap_page = NULL;
960 mb_debug(1, "init group %u\n", group);
961 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
962 this_grp = ext4_get_group_info(sb, group);
964 * This ensures that we don't reinit the buddy cache
965 * page which map to the group from which we are already
966 * allocating. If we are looking at the buddy cache we would
967 * have taken a reference using ext4_mb_load_buddy and that
968 * would have taken the alloc_sem lock.
970 num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group);
971 if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
973 * somebody initialized the group
974 * return without doing anything
980 * the buddy cache inode stores the block bitmap
981 * and buddy information in consecutive blocks.
982 * So for each group we need two blocks.
985 pnum = block / blocks_per_page;
986 poff = block % blocks_per_page;
987 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
989 BUG_ON(page->mapping != inode->i_mapping);
990 ret = ext4_mb_init_cache(page, NULL);
997 if (page == NULL || !PageUptodate(page)) {
1001 mark_page_accessed(page);
1003 bitmap = page_address(page) + (poff * sb->s_blocksize);
1005 /* init buddy cache */
1007 pnum = block / blocks_per_page;
1008 poff = block % blocks_per_page;
1009 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1010 if (page == bitmap_page) {
1012 * If both the bitmap and buddy are in
1013 * the same page we don't need to force
1018 BUG_ON(page->mapping != inode->i_mapping);
1019 ret = ext4_mb_init_cache(page, bitmap);
1026 if (page == NULL || !PageUptodate(page)) {
1030 mark_page_accessed(page);
1032 ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
1034 page_cache_release(bitmap_page);
1036 page_cache_release(page);
1041 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1042 * block group lock of all groups for this page; do not hold the BG lock when
1043 * calling this routine!
1045 static noinline_for_stack int
1046 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1047 struct ext4_buddy *e4b)
1049 int blocks_per_page;
1055 struct ext4_group_info *grp;
1056 struct ext4_sb_info *sbi = EXT4_SB(sb);
1057 struct inode *inode = sbi->s_buddy_cache;
1059 mb_debug(1, "load group %u\n", group);
1061 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1062 grp = ext4_get_group_info(sb, group);
1064 e4b->bd_blkbits = sb->s_blocksize_bits;
1065 e4b->bd_info = ext4_get_group_info(sb, group);
1067 e4b->bd_group = group;
1068 e4b->bd_buddy_page = NULL;
1069 e4b->bd_bitmap_page = NULL;
1070 e4b->alloc_semp = &grp->alloc_sem;
1072 /* Take the read lock on the group alloc
1073 * sem. This would make sure a parallel
1074 * ext4_mb_init_group happening on other
1075 * groups mapped by the page is blocked
1076 * till we are done with allocation
1079 down_read(e4b->alloc_semp);
1081 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1082 /* we need to check for group need init flag
1083 * with alloc_semp held so that we can be sure
1084 * that new blocks didn't get added to the group
1085 * when we are loading the buddy cache
1087 up_read(e4b->alloc_semp);
1089 * we need full data about the group
1090 * to make a good selection
1092 ret = ext4_mb_init_group(sb, group);
1095 goto repeat_load_buddy;
1099 * the buddy cache inode stores the block bitmap
1100 * and buddy information in consecutive blocks.
1101 * So for each group we need two blocks.
1104 pnum = block / blocks_per_page;
1105 poff = block % blocks_per_page;
1107 /* we could use find_or_create_page(), but it locks page
1108 * what we'd like to avoid in fast path ... */
1109 page = find_get_page(inode->i_mapping, pnum);
1110 if (page == NULL || !PageUptodate(page)) {
1113 * drop the page reference and try
1114 * to get the page with lock. If we
1115 * are not uptodate that implies
1116 * somebody just created the page but
1117 * is yet to initialize the same. So
1118 * wait for it to initialize.
1120 page_cache_release(page);
1121 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1123 BUG_ON(page->mapping != inode->i_mapping);
1124 if (!PageUptodate(page)) {
1125 ret = ext4_mb_init_cache(page, NULL);
1130 mb_cmp_bitmaps(e4b, page_address(page) +
1131 (poff * sb->s_blocksize));
1136 if (page == NULL || !PageUptodate(page)) {
1140 e4b->bd_bitmap_page = page;
1141 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1142 mark_page_accessed(page);
1145 pnum = block / blocks_per_page;
1146 poff = block % blocks_per_page;
1148 page = find_get_page(inode->i_mapping, pnum);
1149 if (page == NULL || !PageUptodate(page)) {
1151 page_cache_release(page);
1152 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1154 BUG_ON(page->mapping != inode->i_mapping);
1155 if (!PageUptodate(page)) {
1156 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1165 if (page == NULL || !PageUptodate(page)) {
1169 e4b->bd_buddy_page = page;
1170 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1171 mark_page_accessed(page);
1173 BUG_ON(e4b->bd_bitmap_page == NULL);
1174 BUG_ON(e4b->bd_buddy_page == NULL);
1179 if (e4b->bd_bitmap_page)
1180 page_cache_release(e4b->bd_bitmap_page);
1181 if (e4b->bd_buddy_page)
1182 page_cache_release(e4b->bd_buddy_page);
1183 e4b->bd_buddy = NULL;
1184 e4b->bd_bitmap = NULL;
1186 /* Done with the buddy cache */
1187 up_read(e4b->alloc_semp);
1191 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1193 if (e4b->bd_bitmap_page)
1194 page_cache_release(e4b->bd_bitmap_page);
1195 if (e4b->bd_buddy_page)
1196 page_cache_release(e4b->bd_buddy_page);
1197 /* Done with the buddy cache */
1198 if (e4b->alloc_semp)
1199 up_read(e4b->alloc_semp);
1203 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1208 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1209 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1211 bb = EXT4_MB_BUDDY(e4b);
1212 while (order <= e4b->bd_blkbits + 1) {
1214 if (!mb_test_bit(block, bb)) {
1215 /* this block is part of buddy of order 'order' */
1218 bb += 1 << (e4b->bd_blkbits - order);
1224 static void mb_clear_bits(void *bm, int cur, int len)
1230 if ((cur & 31) == 0 && (len - cur) >= 32) {
1231 /* fast path: clear whole word at once */
1232 addr = bm + (cur >> 3);
1237 mb_clear_bit(cur, bm);
1242 static void mb_set_bits(void *bm, int cur, int len)
1248 if ((cur & 31) == 0 && (len - cur) >= 32) {
1249 /* fast path: set whole word at once */
1250 addr = bm + (cur >> 3);
1255 mb_set_bit(cur, bm);
1260 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1261 int first, int count)
1268 struct super_block *sb = e4b->bd_sb;
1270 BUG_ON(first + count > (sb->s_blocksize << 3));
1271 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1272 mb_check_buddy(e4b);
1273 mb_free_blocks_double(inode, e4b, first, count);
1275 e4b->bd_info->bb_free += count;
1276 if (first < e4b->bd_info->bb_first_free)
1277 e4b->bd_info->bb_first_free = first;
1279 /* let's maintain fragments counter */
1281 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1282 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1283 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1285 e4b->bd_info->bb_fragments--;
1286 else if (!block && !max)
1287 e4b->bd_info->bb_fragments++;
1289 /* let's maintain buddy itself */
1290 while (count-- > 0) {
1294 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1295 ext4_fsblk_t blocknr;
1297 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1299 ext4_grp_locked_error(sb, e4b->bd_group,
1300 inode ? inode->i_ino : 0,
1302 "freeing already freed block "
1305 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1306 e4b->bd_info->bb_counters[order]++;
1308 /* start of the buddy */
1309 buddy = mb_find_buddy(e4b, order, &max);
1313 if (mb_test_bit(block, buddy) ||
1314 mb_test_bit(block + 1, buddy))
1317 /* both the buddies are free, try to coalesce them */
1318 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1324 /* for special purposes, we don't set
1325 * free bits in bitmap */
1326 mb_set_bit(block, buddy);
1327 mb_set_bit(block + 1, buddy);
1329 e4b->bd_info->bb_counters[order]--;
1330 e4b->bd_info->bb_counters[order]--;
1334 e4b->bd_info->bb_counters[order]++;
1336 mb_clear_bit(block, buddy2);
1340 mb_set_largest_free_order(sb, e4b->bd_info);
1341 mb_check_buddy(e4b);
1344 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1345 int needed, struct ext4_free_extent *ex)
1352 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1355 buddy = mb_find_buddy(e4b, order, &max);
1356 BUG_ON(buddy == NULL);
1357 BUG_ON(block >= max);
1358 if (mb_test_bit(block, buddy)) {
1365 /* FIXME dorp order completely ? */
1366 if (likely(order == 0)) {
1367 /* find actual order */
1368 order = mb_find_order_for_block(e4b, block);
1369 block = block >> order;
1372 ex->fe_len = 1 << order;
1373 ex->fe_start = block << order;
1374 ex->fe_group = e4b->bd_group;
1376 /* calc difference from given start */
1377 next = next - ex->fe_start;
1379 ex->fe_start += next;
1381 while (needed > ex->fe_len &&
1382 (buddy = mb_find_buddy(e4b, order, &max))) {
1384 if (block + 1 >= max)
1387 next = (block + 1) * (1 << order);
1388 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1391 ord = mb_find_order_for_block(e4b, next);
1394 block = next >> order;
1395 ex->fe_len += 1 << order;
1398 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1402 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1408 int start = ex->fe_start;
1409 int len = ex->fe_len;
1414 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1415 BUG_ON(e4b->bd_group != ex->fe_group);
1416 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1417 mb_check_buddy(e4b);
1418 mb_mark_used_double(e4b, start, len);
1420 e4b->bd_info->bb_free -= len;
1421 if (e4b->bd_info->bb_first_free == start)
1422 e4b->bd_info->bb_first_free += len;
1424 /* let's maintain fragments counter */
1426 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1427 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1428 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1430 e4b->bd_info->bb_fragments++;
1431 else if (!mlen && !max)
1432 e4b->bd_info->bb_fragments--;
1434 /* let's maintain buddy itself */
1436 ord = mb_find_order_for_block(e4b, start);
1438 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1439 /* the whole chunk may be allocated at once! */
1441 buddy = mb_find_buddy(e4b, ord, &max);
1442 BUG_ON((start >> ord) >= max);
1443 mb_set_bit(start >> ord, buddy);
1444 e4b->bd_info->bb_counters[ord]--;
1451 /* store for history */
1453 ret = len | (ord << 16);
1455 /* we have to split large buddy */
1457 buddy = mb_find_buddy(e4b, ord, &max);
1458 mb_set_bit(start >> ord, buddy);
1459 e4b->bd_info->bb_counters[ord]--;
1462 cur = (start >> ord) & ~1U;
1463 buddy = mb_find_buddy(e4b, ord, &max);
1464 mb_clear_bit(cur, buddy);
1465 mb_clear_bit(cur + 1, buddy);
1466 e4b->bd_info->bb_counters[ord]++;
1467 e4b->bd_info->bb_counters[ord]++;
1469 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1471 mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1472 mb_check_buddy(e4b);
1478 * Must be called under group lock!
1480 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1481 struct ext4_buddy *e4b)
1483 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1486 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1487 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1489 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1490 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1491 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1493 /* preallocation can change ac_b_ex, thus we store actually
1494 * allocated blocks for history */
1495 ac->ac_f_ex = ac->ac_b_ex;
1497 ac->ac_status = AC_STATUS_FOUND;
1498 ac->ac_tail = ret & 0xffff;
1499 ac->ac_buddy = ret >> 16;
1502 * take the page reference. We want the page to be pinned
1503 * so that we don't get a ext4_mb_init_cache_call for this
1504 * group until we update the bitmap. That would mean we
1505 * double allocate blocks. The reference is dropped
1506 * in ext4_mb_release_context
1508 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1509 get_page(ac->ac_bitmap_page);
1510 ac->ac_buddy_page = e4b->bd_buddy_page;
1511 get_page(ac->ac_buddy_page);
1512 /* on allocation we use ac to track the held semaphore */
1513 ac->alloc_semp = e4b->alloc_semp;
1514 e4b->alloc_semp = NULL;
1515 /* store last allocated for subsequent stream allocation */
1516 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1517 spin_lock(&sbi->s_md_lock);
1518 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1519 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1520 spin_unlock(&sbi->s_md_lock);
1525 * regular allocator, for general purposes allocation
1528 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1529 struct ext4_buddy *e4b,
1532 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1533 struct ext4_free_extent *bex = &ac->ac_b_ex;
1534 struct ext4_free_extent *gex = &ac->ac_g_ex;
1535 struct ext4_free_extent ex;
1538 if (ac->ac_status == AC_STATUS_FOUND)
1541 * We don't want to scan for a whole year
1543 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1544 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1545 ac->ac_status = AC_STATUS_BREAK;
1550 * Haven't found good chunk so far, let's continue
1552 if (bex->fe_len < gex->fe_len)
1555 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1556 && bex->fe_group == e4b->bd_group) {
1557 /* recheck chunk's availability - we don't know
1558 * when it was found (within this lock-unlock
1560 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1561 if (max >= gex->fe_len) {
1562 ext4_mb_use_best_found(ac, e4b);
1569 * The routine checks whether found extent is good enough. If it is,
1570 * then the extent gets marked used and flag is set to the context
1571 * to stop scanning. Otherwise, the extent is compared with the
1572 * previous found extent and if new one is better, then it's stored
1573 * in the context. Later, the best found extent will be used, if
1574 * mballoc can't find good enough extent.
1576 * FIXME: real allocation policy is to be designed yet!
1578 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1579 struct ext4_free_extent *ex,
1580 struct ext4_buddy *e4b)
1582 struct ext4_free_extent *bex = &ac->ac_b_ex;
1583 struct ext4_free_extent *gex = &ac->ac_g_ex;
1585 BUG_ON(ex->fe_len <= 0);
1586 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1587 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1588 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1593 * The special case - take what you catch first
1595 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1597 ext4_mb_use_best_found(ac, e4b);
1602 * Let's check whether the chuck is good enough
1604 if (ex->fe_len == gex->fe_len) {
1606 ext4_mb_use_best_found(ac, e4b);
1611 * If this is first found extent, just store it in the context
1613 if (bex->fe_len == 0) {
1619 * If new found extent is better, store it in the context
1621 if (bex->fe_len < gex->fe_len) {
1622 /* if the request isn't satisfied, any found extent
1623 * larger than previous best one is better */
1624 if (ex->fe_len > bex->fe_len)
1626 } else if (ex->fe_len > gex->fe_len) {
1627 /* if the request is satisfied, then we try to find
1628 * an extent that still satisfy the request, but is
1629 * smaller than previous one */
1630 if (ex->fe_len < bex->fe_len)
1634 ext4_mb_check_limits(ac, e4b, 0);
1637 static noinline_for_stack
1638 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1639 struct ext4_buddy *e4b)
1641 struct ext4_free_extent ex = ac->ac_b_ex;
1642 ext4_group_t group = ex.fe_group;
1646 BUG_ON(ex.fe_len <= 0);
1647 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1651 ext4_lock_group(ac->ac_sb, group);
1652 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1656 ext4_mb_use_best_found(ac, e4b);
1659 ext4_unlock_group(ac->ac_sb, group);
1660 ext4_mb_unload_buddy(e4b);
1665 static noinline_for_stack
1666 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1667 struct ext4_buddy *e4b)
1669 ext4_group_t group = ac->ac_g_ex.fe_group;
1672 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1673 struct ext4_free_extent ex;
1675 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1678 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1682 ext4_lock_group(ac->ac_sb, group);
1683 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1684 ac->ac_g_ex.fe_len, &ex);
1686 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1689 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1691 /* use do_div to get remainder (would be 64-bit modulo) */
1692 if (do_div(start, sbi->s_stripe) == 0) {
1695 ext4_mb_use_best_found(ac, e4b);
1697 } else if (max >= ac->ac_g_ex.fe_len) {
1698 BUG_ON(ex.fe_len <= 0);
1699 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1700 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1703 ext4_mb_use_best_found(ac, e4b);
1704 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1705 /* Sometimes, caller may want to merge even small
1706 * number of blocks to an existing extent */
1707 BUG_ON(ex.fe_len <= 0);
1708 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1709 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1712 ext4_mb_use_best_found(ac, e4b);
1714 ext4_unlock_group(ac->ac_sb, group);
1715 ext4_mb_unload_buddy(e4b);
1721 * The routine scans buddy structures (not bitmap!) from given order
1722 * to max order and tries to find big enough chunk to satisfy the req
1724 static noinline_for_stack
1725 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1726 struct ext4_buddy *e4b)
1728 struct super_block *sb = ac->ac_sb;
1729 struct ext4_group_info *grp = e4b->bd_info;
1735 BUG_ON(ac->ac_2order <= 0);
1736 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1737 if (grp->bb_counters[i] == 0)
1740 buddy = mb_find_buddy(e4b, i, &max);
1741 BUG_ON(buddy == NULL);
1743 k = mb_find_next_zero_bit(buddy, max, 0);
1748 ac->ac_b_ex.fe_len = 1 << i;
1749 ac->ac_b_ex.fe_start = k << i;
1750 ac->ac_b_ex.fe_group = e4b->bd_group;
1752 ext4_mb_use_best_found(ac, e4b);
1754 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1756 if (EXT4_SB(sb)->s_mb_stats)
1757 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1764 * The routine scans the group and measures all found extents.
1765 * In order to optimize scanning, caller must pass number of
1766 * free blocks in the group, so the routine can know upper limit.
1768 static noinline_for_stack
1769 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1770 struct ext4_buddy *e4b)
1772 struct super_block *sb = ac->ac_sb;
1773 void *bitmap = EXT4_MB_BITMAP(e4b);
1774 struct ext4_free_extent ex;
1778 free = e4b->bd_info->bb_free;
1781 i = e4b->bd_info->bb_first_free;
1783 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1784 i = mb_find_next_zero_bit(bitmap,
1785 EXT4_BLOCKS_PER_GROUP(sb), i);
1786 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1788 * IF we have corrupt bitmap, we won't find any
1789 * free blocks even though group info says we
1790 * we have free blocks
1792 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1793 "%d free blocks as per "
1794 "group info. But bitmap says 0",
1799 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1800 BUG_ON(ex.fe_len <= 0);
1801 if (free < ex.fe_len) {
1802 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1803 "%d free blocks as per "
1804 "group info. But got %d blocks",
1807 * The number of free blocks differs. This mostly
1808 * indicate that the bitmap is corrupt. So exit
1809 * without claiming the space.
1814 ext4_mb_measure_extent(ac, &ex, e4b);
1820 ext4_mb_check_limits(ac, e4b, 1);
1824 * This is a special case for storages like raid5
1825 * we try to find stripe-aligned chunks for stripe-size requests
1826 * XXX should do so at least for multiples of stripe size as well
1828 static noinline_for_stack
1829 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1830 struct ext4_buddy *e4b)
1832 struct super_block *sb = ac->ac_sb;
1833 struct ext4_sb_info *sbi = EXT4_SB(sb);
1834 void *bitmap = EXT4_MB_BITMAP(e4b);
1835 struct ext4_free_extent ex;
1836 ext4_fsblk_t first_group_block;
1841 BUG_ON(sbi->s_stripe == 0);
1843 /* find first stripe-aligned block in group */
1844 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1846 a = first_group_block + sbi->s_stripe - 1;
1847 do_div(a, sbi->s_stripe);
1848 i = (a * sbi->s_stripe) - first_group_block;
1850 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1851 if (!mb_test_bit(i, bitmap)) {
1852 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1853 if (max >= sbi->s_stripe) {
1856 ext4_mb_use_best_found(ac, e4b);
1864 /* This is now called BEFORE we load the buddy bitmap. */
1865 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1866 ext4_group_t group, int cr)
1868 unsigned free, fragments;
1869 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1870 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1872 BUG_ON(cr < 0 || cr >= 4);
1874 /* We only do this if the grp has never been initialized */
1875 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1876 int ret = ext4_mb_init_group(ac->ac_sb, group);
1881 free = grp->bb_free;
1882 fragments = grp->bb_fragments;
1890 BUG_ON(ac->ac_2order == 0);
1892 if (grp->bb_largest_free_order < ac->ac_2order)
1895 /* Avoid using the first bg of a flexgroup for data files */
1896 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1897 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1898 ((group % flex_size) == 0))
1903 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1907 if (free >= ac->ac_g_ex.fe_len)
1920 * lock the group_info alloc_sem of all the groups
1921 * belonging to the same buddy cache page. This
1922 * make sure other parallel operation on the buddy
1923 * cache doesn't happen whild holding the buddy cache
1926 int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1930 int blocks_per_page;
1931 int groups_per_page;
1932 ext4_group_t ngroups = ext4_get_groups_count(sb);
1933 ext4_group_t first_group;
1934 struct ext4_group_info *grp;
1936 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1938 * the buddy cache inode stores the block bitmap
1939 * and buddy information in consecutive blocks.
1940 * So for each group we need two blocks.
1943 pnum = block / blocks_per_page;
1944 first_group = pnum * blocks_per_page / 2;
1946 groups_per_page = blocks_per_page >> 1;
1947 if (groups_per_page == 0)
1948 groups_per_page = 1;
1949 /* read all groups the page covers into the cache */
1950 for (i = 0; i < groups_per_page; i++) {
1952 if ((first_group + i) >= ngroups)
1954 grp = ext4_get_group_info(sb, first_group + i);
1955 /* take all groups write allocation
1956 * semaphore. This make sure there is
1957 * no block allocation going on in any
1960 down_write_nested(&grp->alloc_sem, i);
1965 void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1966 ext4_group_t group, int locked_group)
1970 int blocks_per_page;
1971 ext4_group_t first_group;
1972 struct ext4_group_info *grp;
1974 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1976 * the buddy cache inode stores the block bitmap
1977 * and buddy information in consecutive blocks.
1978 * So for each group we need two blocks.
1981 pnum = block / blocks_per_page;
1982 first_group = pnum * blocks_per_page / 2;
1983 /* release locks on all the groups */
1984 for (i = 0; i < locked_group; i++) {
1986 grp = ext4_get_group_info(sb, first_group + i);
1987 /* take all groups write allocation
1988 * semaphore. This make sure there is
1989 * no block allocation going on in any
1992 up_write(&grp->alloc_sem);
1997 static noinline_for_stack int
1998 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2000 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 /* if stream allocation is enabled, use global goal */
2045 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2046 /* TBD: may be hot point */
2047 spin_lock(&sbi->s_md_lock);
2048 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2049 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2050 spin_unlock(&sbi->s_md_lock);
2053 /* Let's just scan groups to find more-less suitable blocks */
2054 cr = ac->ac_2order ? 0 : 1;
2056 * cr == 0 try to get exact allocation,
2057 * cr == 3 try to get anything
2060 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2061 ac->ac_criteria = cr;
2063 * searching for the right group start
2064 * from the goal value specified
2066 group = ac->ac_g_ex.fe_group;
2068 for (i = 0; i < ngroups; group++, i++) {
2069 if (group == ngroups)
2072 /* This now checks without needing the buddy page */
2073 if (!ext4_mb_good_group(ac, group, cr))
2076 err = ext4_mb_load_buddy(sb, group, &e4b);
2080 ext4_lock_group(sb, group);
2083 * We need to check again after locking the
2086 if (!ext4_mb_good_group(ac, group, cr)) {
2087 ext4_unlock_group(sb, group);
2088 ext4_mb_unload_buddy(&e4b);
2092 ac->ac_groups_scanned++;
2094 ext4_mb_simple_scan_group(ac, &e4b);
2096 ac->ac_g_ex.fe_len == sbi->s_stripe)
2097 ext4_mb_scan_aligned(ac, &e4b);
2099 ext4_mb_complex_scan_group(ac, &e4b);
2101 ext4_unlock_group(sb, group);
2102 ext4_mb_unload_buddy(&e4b);
2104 if (ac->ac_status != AC_STATUS_CONTINUE)
2109 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2110 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2112 * We've been searching too long. Let's try to allocate
2113 * the best chunk we've found so far
2116 ext4_mb_try_best_found(ac, &e4b);
2117 if (ac->ac_status != AC_STATUS_FOUND) {
2119 * Someone more lucky has already allocated it.
2120 * The only thing we can do is just take first
2122 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2124 ac->ac_b_ex.fe_group = 0;
2125 ac->ac_b_ex.fe_start = 0;
2126 ac->ac_b_ex.fe_len = 0;
2127 ac->ac_status = AC_STATUS_CONTINUE;
2128 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2130 atomic_inc(&sbi->s_mb_lost_chunks);
2138 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2140 struct super_block *sb = seq->private;
2143 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2146 return (void *) ((unsigned long) group);
2149 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2151 struct super_block *sb = seq->private;
2155 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2158 return (void *) ((unsigned long) group);
2161 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2163 struct super_block *sb = seq->private;
2164 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2167 struct ext4_buddy e4b;
2169 struct ext4_group_info info;
2170 ext4_grpblk_t counters[16];
2175 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2176 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2177 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2178 "group", "free", "frags", "first",
2179 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2180 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2182 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2183 sizeof(struct ext4_group_info);
2184 err = ext4_mb_load_buddy(sb, group, &e4b);
2186 seq_printf(seq, "#%-5u: I/O error\n", group);
2189 ext4_lock_group(sb, group);
2190 memcpy(&sg, ext4_get_group_info(sb, group), i);
2191 ext4_unlock_group(sb, group);
2192 ext4_mb_unload_buddy(&e4b);
2194 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2195 sg.info.bb_fragments, sg.info.bb_first_free);
2196 for (i = 0; i <= 13; i++)
2197 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2198 sg.info.bb_counters[i] : 0);
2199 seq_printf(seq, " ]\n");
2204 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2208 static const struct seq_operations ext4_mb_seq_groups_ops = {
2209 .start = ext4_mb_seq_groups_start,
2210 .next = ext4_mb_seq_groups_next,
2211 .stop = ext4_mb_seq_groups_stop,
2212 .show = ext4_mb_seq_groups_show,
2215 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2217 struct super_block *sb = PDE(inode)->data;
2220 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2222 struct seq_file *m = (struct seq_file *)file->private_data;
2229 static const struct file_operations ext4_mb_seq_groups_fops = {
2230 .owner = THIS_MODULE,
2231 .open = ext4_mb_seq_groups_open,
2233 .llseek = seq_lseek,
2234 .release = seq_release,
2238 /* Create and initialize ext4_group_info data for the given group. */
2239 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2240 struct ext4_group_desc *desc)
2244 struct ext4_sb_info *sbi = EXT4_SB(sb);
2245 struct ext4_group_info **meta_group_info;
2248 * First check if this group is the first of a reserved block.
2249 * If it's true, we have to allocate a new table of pointers
2250 * to ext4_group_info structures
2252 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2253 metalen = sizeof(*meta_group_info) <<
2254 EXT4_DESC_PER_BLOCK_BITS(sb);
2255 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2256 if (meta_group_info == NULL) {
2257 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2259 goto exit_meta_group_info;
2261 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2266 * calculate needed size. if change bb_counters size,
2267 * don't forget about ext4_mb_generate_buddy()
2269 len = offsetof(typeof(**meta_group_info),
2270 bb_counters[sb->s_blocksize_bits + 2]);
2273 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2274 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2276 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2277 if (meta_group_info[i] == NULL) {
2278 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2279 goto exit_group_info;
2281 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2282 &(meta_group_info[i]->bb_state));
2285 * initialize bb_free to be able to skip
2286 * empty groups without initialization
2288 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2289 meta_group_info[i]->bb_free =
2290 ext4_free_blocks_after_init(sb, group, desc);
2292 meta_group_info[i]->bb_free =
2293 ext4_free_blks_count(sb, desc);
2296 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2297 init_rwsem(&meta_group_info[i]->alloc_sem);
2298 meta_group_info[i]->bb_free_root = RB_ROOT;
2299 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2303 struct buffer_head *bh;
2304 meta_group_info[i]->bb_bitmap =
2305 kmalloc(sb->s_blocksize, GFP_KERNEL);
2306 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2307 bh = ext4_read_block_bitmap(sb, group);
2309 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2318 /* If a meta_group_info table has been allocated, release it now */
2319 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2320 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2321 exit_meta_group_info:
2323 } /* ext4_mb_add_groupinfo */
2325 static int ext4_mb_init_backend(struct super_block *sb)
2327 ext4_group_t ngroups = ext4_get_groups_count(sb);
2329 struct ext4_sb_info *sbi = EXT4_SB(sb);
2330 struct ext4_super_block *es = sbi->s_es;
2331 int num_meta_group_infos;
2332 int num_meta_group_infos_max;
2334 struct ext4_group_desc *desc;
2336 /* This is the number of blocks used by GDT */
2337 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2338 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2341 * This is the total number of blocks used by GDT including
2342 * the number of reserved blocks for GDT.
2343 * The s_group_info array is allocated with this value
2344 * to allow a clean online resize without a complex
2345 * manipulation of pointer.
2346 * The drawback is the unused memory when no resize
2347 * occurs but it's very low in terms of pages
2348 * (see comments below)
2349 * Need to handle this properly when META_BG resizing is allowed
2351 num_meta_group_infos_max = num_meta_group_infos +
2352 le16_to_cpu(es->s_reserved_gdt_blocks);
2355 * array_size is the size of s_group_info array. We round it
2356 * to the next power of two because this approximation is done
2357 * internally by kmalloc so we can have some more memory
2358 * for free here (e.g. may be used for META_BG resize).
2361 while (array_size < sizeof(*sbi->s_group_info) *
2362 num_meta_group_infos_max)
2363 array_size = array_size << 1;
2364 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2365 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2366 * So a two level scheme suffices for now. */
2367 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2368 if (sbi->s_group_info == NULL) {
2369 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2372 sbi->s_buddy_cache = new_inode(sb);
2373 if (sbi->s_buddy_cache == NULL) {
2374 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2377 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2378 for (i = 0; i < ngroups; i++) {
2379 desc = ext4_get_group_desc(sb, i, NULL);
2382 "EXT4-fs: can't read descriptor %u\n", i);
2385 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2393 kfree(ext4_get_group_info(sb, i));
2394 i = num_meta_group_infos;
2396 kfree(sbi->s_group_info[i]);
2397 iput(sbi->s_buddy_cache);
2399 kfree(sbi->s_group_info);
2403 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2405 struct ext4_sb_info *sbi = EXT4_SB(sb);
2411 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2413 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2414 if (sbi->s_mb_offsets == NULL) {
2418 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2419 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2420 if (sbi->s_mb_maxs == NULL) {
2421 kfree(sbi->s_mb_offsets);
2425 /* order 0 is regular bitmap */
2426 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2427 sbi->s_mb_offsets[0] = 0;
2431 max = sb->s_blocksize << 2;
2433 sbi->s_mb_offsets[i] = offset;
2434 sbi->s_mb_maxs[i] = max;
2435 offset += 1 << (sb->s_blocksize_bits - i);
2438 } while (i <= sb->s_blocksize_bits + 1);
2440 /* init file for buddy data */
2441 ret = ext4_mb_init_backend(sb);
2443 kfree(sbi->s_mb_offsets);
2444 kfree(sbi->s_mb_maxs);
2448 spin_lock_init(&sbi->s_md_lock);
2449 spin_lock_init(&sbi->s_bal_lock);
2451 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2452 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2453 sbi->s_mb_stats = MB_DEFAULT_STATS;
2454 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2455 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2456 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2458 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2459 if (sbi->s_locality_groups == NULL) {
2460 kfree(sbi->s_mb_offsets);
2461 kfree(sbi->s_mb_maxs);
2464 for_each_possible_cpu(i) {
2465 struct ext4_locality_group *lg;
2466 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2467 mutex_init(&lg->lg_mutex);
2468 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2469 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2470 spin_lock_init(&lg->lg_prealloc_lock);
2474 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2475 &ext4_mb_seq_groups_fops, sb);
2478 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2482 /* need to called with the ext4 group lock held */
2483 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2485 struct ext4_prealloc_space *pa;
2486 struct list_head *cur, *tmp;
2489 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2490 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2491 list_del(&pa->pa_group_list);
2493 kmem_cache_free(ext4_pspace_cachep, pa);
2496 mb_debug(1, "mballoc: %u PAs left\n", count);
2500 int ext4_mb_release(struct super_block *sb)
2502 ext4_group_t ngroups = ext4_get_groups_count(sb);
2504 int num_meta_group_infos;
2505 struct ext4_group_info *grinfo;
2506 struct ext4_sb_info *sbi = EXT4_SB(sb);
2508 if (sbi->s_group_info) {
2509 for (i = 0; i < ngroups; i++) {
2510 grinfo = ext4_get_group_info(sb, i);
2512 kfree(grinfo->bb_bitmap);
2514 ext4_lock_group(sb, i);
2515 ext4_mb_cleanup_pa(grinfo);
2516 ext4_unlock_group(sb, i);
2519 num_meta_group_infos = (ngroups +
2520 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2521 EXT4_DESC_PER_BLOCK_BITS(sb);
2522 for (i = 0; i < num_meta_group_infos; i++)
2523 kfree(sbi->s_group_info[i]);
2524 kfree(sbi->s_group_info);
2526 kfree(sbi->s_mb_offsets);
2527 kfree(sbi->s_mb_maxs);
2528 if (sbi->s_buddy_cache)
2529 iput(sbi->s_buddy_cache);
2530 if (sbi->s_mb_stats) {
2532 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2533 atomic_read(&sbi->s_bal_allocated),
2534 atomic_read(&sbi->s_bal_reqs),
2535 atomic_read(&sbi->s_bal_success));
2537 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2538 "%u 2^N hits, %u breaks, %u lost\n",
2539 atomic_read(&sbi->s_bal_ex_scanned),
2540 atomic_read(&sbi->s_bal_goals),
2541 atomic_read(&sbi->s_bal_2orders),
2542 atomic_read(&sbi->s_bal_breaks),
2543 atomic_read(&sbi->s_mb_lost_chunks));
2545 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2546 sbi->s_mb_buddies_generated++,
2547 sbi->s_mb_generation_time);
2549 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2550 atomic_read(&sbi->s_mb_preallocated),
2551 atomic_read(&sbi->s_mb_discarded));
2554 free_percpu(sbi->s_locality_groups);
2556 remove_proc_entry("mb_groups", sbi->s_proc);
2562 * This function is called by the jbd2 layer once the commit has finished,
2563 * so we know we can free the blocks that were released with that commit.
2565 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2567 struct super_block *sb = journal->j_private;
2568 struct ext4_buddy e4b;
2569 struct ext4_group_info *db;
2570 int err, count = 0, count2 = 0;
2571 struct ext4_free_data *entry;
2572 struct list_head *l, *ltmp;
2574 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2575 entry = list_entry(l, struct ext4_free_data, list);
2577 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2578 entry->count, entry->group, entry);
2580 if (test_opt(sb, DISCARD)) {
2582 ext4_fsblk_t discard_block;
2584 discard_block = entry->start_blk +
2585 ext4_group_first_block_no(sb, entry->group);
2586 trace_ext4_discard_blocks(sb,
2587 (unsigned long long)discard_block,
2589 ret = sb_issue_discard(sb, discard_block, entry->count);
2590 if (ret == EOPNOTSUPP) {
2592 "discard not supported, disabling");
2593 clear_opt(EXT4_SB(sb)->s_mount_opt, DISCARD);
2597 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2598 /* we expect to find existing buddy because it's pinned */
2602 /* there are blocks to put in buddy to make them really free */
2603 count += entry->count;
2605 ext4_lock_group(sb, entry->group);
2606 /* Take it out of per group rb tree */
2607 rb_erase(&entry->node, &(db->bb_free_root));
2608 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2610 if (!db->bb_free_root.rb_node) {
2611 /* No more items in the per group rb tree
2612 * balance refcounts from ext4_mb_free_metadata()
2614 page_cache_release(e4b.bd_buddy_page);
2615 page_cache_release(e4b.bd_bitmap_page);
2617 ext4_unlock_group(sb, entry->group);
2618 kmem_cache_free(ext4_free_ext_cachep, entry);
2619 ext4_mb_unload_buddy(&e4b);
2622 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2625 #ifdef CONFIG_EXT4_DEBUG
2626 u8 mb_enable_debug __read_mostly;
2628 static struct dentry *debugfs_dir;
2629 static struct dentry *debugfs_debug;
2631 static void __init ext4_create_debugfs_entry(void)
2633 debugfs_dir = debugfs_create_dir("ext4", NULL);
2635 debugfs_debug = debugfs_create_u8("mballoc-debug",
2641 static void ext4_remove_debugfs_entry(void)
2643 debugfs_remove(debugfs_debug);
2644 debugfs_remove(debugfs_dir);
2649 static void __init ext4_create_debugfs_entry(void)
2653 static void ext4_remove_debugfs_entry(void)
2659 int __init init_ext4_mballoc(void)
2661 ext4_pspace_cachep =
2662 kmem_cache_create("ext4_prealloc_space",
2663 sizeof(struct ext4_prealloc_space),
2664 0, SLAB_RECLAIM_ACCOUNT, NULL);
2665 if (ext4_pspace_cachep == NULL)
2669 kmem_cache_create("ext4_alloc_context",
2670 sizeof(struct ext4_allocation_context),
2671 0, SLAB_RECLAIM_ACCOUNT, NULL);
2672 if (ext4_ac_cachep == NULL) {
2673 kmem_cache_destroy(ext4_pspace_cachep);
2677 ext4_free_ext_cachep =
2678 kmem_cache_create("ext4_free_block_extents",
2679 sizeof(struct ext4_free_data),
2680 0, SLAB_RECLAIM_ACCOUNT, NULL);
2681 if (ext4_free_ext_cachep == NULL) {
2682 kmem_cache_destroy(ext4_pspace_cachep);
2683 kmem_cache_destroy(ext4_ac_cachep);
2686 ext4_create_debugfs_entry();
2690 void exit_ext4_mballoc(void)
2693 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2694 * before destroying the slab cache.
2697 kmem_cache_destroy(ext4_pspace_cachep);
2698 kmem_cache_destroy(ext4_ac_cachep);
2699 kmem_cache_destroy(ext4_free_ext_cachep);
2700 ext4_remove_debugfs_entry();
2705 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2706 * Returns 0 if success or error code
2708 static noinline_for_stack int
2709 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2710 handle_t *handle, unsigned int reserv_blks)
2712 struct buffer_head *bitmap_bh = NULL;
2713 struct ext4_group_desc *gdp;
2714 struct buffer_head *gdp_bh;
2715 struct ext4_sb_info *sbi;
2716 struct super_block *sb;
2720 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2721 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2727 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2731 err = ext4_journal_get_write_access(handle, bitmap_bh);
2736 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2740 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2741 ext4_free_blks_count(sb, gdp));
2743 err = ext4_journal_get_write_access(handle, gdp_bh);
2747 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2749 len = ac->ac_b_ex.fe_len;
2750 if (!ext4_data_block_valid(sbi, block, len)) {
2751 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2752 "fs metadata\n", block, block+len);
2753 /* File system mounted not to panic on error
2754 * Fix the bitmap and repeat the block allocation
2755 * We leak some of the blocks here.
2757 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2758 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2759 ac->ac_b_ex.fe_len);
2760 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2761 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2767 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2768 #ifdef AGGRESSIVE_CHECK
2771 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2772 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2773 bitmap_bh->b_data));
2777 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
2778 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2779 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2780 ext4_free_blks_set(sb, gdp,
2781 ext4_free_blocks_after_init(sb,
2782 ac->ac_b_ex.fe_group, gdp));
2784 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
2785 ext4_free_blks_set(sb, gdp, len);
2786 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2788 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2789 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
2791 * Now reduce the dirty block count also. Should not go negative
2793 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2794 /* release all the reserved blocks if non delalloc */
2795 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
2797 if (sbi->s_log_groups_per_flex) {
2798 ext4_group_t flex_group = ext4_flex_group(sbi,
2799 ac->ac_b_ex.fe_group);
2800 atomic_sub(ac->ac_b_ex.fe_len,
2801 &sbi->s_flex_groups[flex_group].free_blocks);
2804 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2807 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2810 ext4_mark_super_dirty(sb);
2816 * here we normalize request for locality group
2817 * Group request are normalized to s_strip size if we set the same via mount
2818 * option. If not we set it to s_mb_group_prealloc which can be configured via
2819 * /sys/fs/ext4/<partition>/mb_group_prealloc
2821 * XXX: should we try to preallocate more than the group has now?
2823 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2825 struct super_block *sb = ac->ac_sb;
2826 struct ext4_locality_group *lg = ac->ac_lg;
2829 if (EXT4_SB(sb)->s_stripe)
2830 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
2832 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2833 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2834 current->pid, ac->ac_g_ex.fe_len);
2838 * Normalization means making request better in terms of
2839 * size and alignment
2841 static noinline_for_stack void
2842 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2843 struct ext4_allocation_request *ar)
2847 loff_t size, orig_size, start_off;
2849 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2850 struct ext4_prealloc_space *pa;
2852 /* do normalize only data requests, metadata requests
2853 do not need preallocation */
2854 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2857 /* sometime caller may want exact blocks */
2858 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2861 /* caller may indicate that preallocation isn't
2862 * required (it's a tail, for example) */
2863 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2866 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2867 ext4_mb_normalize_group_request(ac);
2871 bsbits = ac->ac_sb->s_blocksize_bits;
2873 /* first, let's learn actual file size
2874 * given current request is allocated */
2875 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2876 size = size << bsbits;
2877 if (size < i_size_read(ac->ac_inode))
2878 size = i_size_read(ac->ac_inode);
2881 /* max size of free chunks */
2884 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2885 (req <= (size) || max <= (chunk_size))
2887 /* first, try to predict filesize */
2888 /* XXX: should this table be tunable? */
2890 if (size <= 16 * 1024) {
2892 } else if (size <= 32 * 1024) {
2894 } else if (size <= 64 * 1024) {
2896 } else if (size <= 128 * 1024) {
2898 } else if (size <= 256 * 1024) {
2900 } else if (size <= 512 * 1024) {
2902 } else if (size <= 1024 * 1024) {
2904 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2905 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2906 (21 - bsbits)) << 21;
2907 size = 2 * 1024 * 1024;
2908 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2909 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2910 (22 - bsbits)) << 22;
2911 size = 4 * 1024 * 1024;
2912 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2913 (8<<20)>>bsbits, max, 8 * 1024)) {
2914 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2915 (23 - bsbits)) << 23;
2916 size = 8 * 1024 * 1024;
2918 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2919 size = ac->ac_o_ex.fe_len << bsbits;
2921 size = size >> bsbits;
2922 start = start_off >> bsbits;
2924 /* don't cover already allocated blocks in selected range */
2925 if (ar->pleft && start <= ar->lleft) {
2926 size -= ar->lleft + 1 - start;
2927 start = ar->lleft + 1;
2929 if (ar->pright && start + size - 1 >= ar->lright)
2930 size -= start + size - ar->lright;
2934 /* check we don't cross already preallocated blocks */
2936 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2941 spin_lock(&pa->pa_lock);
2942 if (pa->pa_deleted) {
2943 spin_unlock(&pa->pa_lock);
2947 pa_end = pa->pa_lstart + pa->pa_len;
2949 /* PA must not overlap original request */
2950 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2951 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2953 /* skip PAs this normalized request doesn't overlap with */
2954 if (pa->pa_lstart >= end || pa_end <= start) {
2955 spin_unlock(&pa->pa_lock);
2958 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2960 /* adjust start or end to be adjacent to this pa */
2961 if (pa_end <= ac->ac_o_ex.fe_logical) {
2962 BUG_ON(pa_end < start);
2964 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
2965 BUG_ON(pa->pa_lstart > end);
2966 end = pa->pa_lstart;
2968 spin_unlock(&pa->pa_lock);
2973 /* XXX: extra loop to check we really don't overlap preallocations */
2975 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2977 spin_lock(&pa->pa_lock);
2978 if (pa->pa_deleted == 0) {
2979 pa_end = pa->pa_lstart + pa->pa_len;
2980 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
2982 spin_unlock(&pa->pa_lock);
2986 if (start + size <= ac->ac_o_ex.fe_logical &&
2987 start > ac->ac_o_ex.fe_logical) {
2988 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
2989 (unsigned long) start, (unsigned long) size,
2990 (unsigned long) ac->ac_o_ex.fe_logical);
2992 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
2993 start > ac->ac_o_ex.fe_logical);
2994 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
2996 /* now prepare goal request */
2998 /* XXX: is it better to align blocks WRT to logical
2999 * placement or satisfy big request as is */
3000 ac->ac_g_ex.fe_logical = start;
3001 ac->ac_g_ex.fe_len = size;
3003 /* define goal start in order to merge */
3004 if (ar->pright && (ar->lright == (start + size))) {
3005 /* merge to the right */
3006 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3007 &ac->ac_f_ex.fe_group,
3008 &ac->ac_f_ex.fe_start);
3009 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3011 if (ar->pleft && (ar->lleft + 1 == start)) {
3012 /* merge to the left */
3013 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3014 &ac->ac_f_ex.fe_group,
3015 &ac->ac_f_ex.fe_start);
3016 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3019 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3020 (unsigned) orig_size, (unsigned) start);
3023 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3025 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3027 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3028 atomic_inc(&sbi->s_bal_reqs);
3029 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3030 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3031 atomic_inc(&sbi->s_bal_success);
3032 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3033 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3034 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3035 atomic_inc(&sbi->s_bal_goals);
3036 if (ac->ac_found > sbi->s_mb_max_to_scan)
3037 atomic_inc(&sbi->s_bal_breaks);
3040 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3041 trace_ext4_mballoc_alloc(ac);
3043 trace_ext4_mballoc_prealloc(ac);
3047 * Called on failure; free up any blocks from the inode PA for this
3048 * context. We don't need this for MB_GROUP_PA because we only change
3049 * pa_free in ext4_mb_release_context(), but on failure, we've already
3050 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3052 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3054 struct ext4_prealloc_space *pa = ac->ac_pa;
3057 if (pa && pa->pa_type == MB_INODE_PA) {
3058 len = ac->ac_b_ex.fe_len;
3065 * use blocks preallocated to inode
3067 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3068 struct ext4_prealloc_space *pa)
3074 /* found preallocated blocks, use them */
3075 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3076 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3078 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3079 &ac->ac_b_ex.fe_start);
3080 ac->ac_b_ex.fe_len = len;
3081 ac->ac_status = AC_STATUS_FOUND;
3084 BUG_ON(start < pa->pa_pstart);
3085 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3086 BUG_ON(pa->pa_free < len);
3089 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3093 * use blocks preallocated to locality group
3095 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3096 struct ext4_prealloc_space *pa)
3098 unsigned int len = ac->ac_o_ex.fe_len;
3100 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3101 &ac->ac_b_ex.fe_group,
3102 &ac->ac_b_ex.fe_start);
3103 ac->ac_b_ex.fe_len = len;
3104 ac->ac_status = AC_STATUS_FOUND;
3107 /* we don't correct pa_pstart or pa_plen here to avoid
3108 * possible race when the group is being loaded concurrently
3109 * instead we correct pa later, after blocks are marked
3110 * in on-disk bitmap -- see ext4_mb_release_context()
3111 * Other CPUs are prevented from allocating from this pa by lg_mutex
3113 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3117 * Return the prealloc space that have minimal distance
3118 * from the goal block. @cpa is the prealloc
3119 * space that is having currently known minimal distance
3120 * from the goal block.
3122 static struct ext4_prealloc_space *
3123 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3124 struct ext4_prealloc_space *pa,
3125 struct ext4_prealloc_space *cpa)
3127 ext4_fsblk_t cur_distance, new_distance;
3130 atomic_inc(&pa->pa_count);
3133 cur_distance = abs(goal_block - cpa->pa_pstart);
3134 new_distance = abs(goal_block - pa->pa_pstart);
3136 if (cur_distance < new_distance)
3139 /* drop the previous reference */
3140 atomic_dec(&cpa->pa_count);
3141 atomic_inc(&pa->pa_count);
3146 * search goal blocks in preallocated space
3148 static noinline_for_stack int
3149 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3152 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3153 struct ext4_locality_group *lg;
3154 struct ext4_prealloc_space *pa, *cpa = NULL;
3155 ext4_fsblk_t goal_block;
3157 /* only data can be preallocated */
3158 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3161 /* first, try per-file preallocation */
3163 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3165 /* all fields in this condition don't change,
3166 * so we can skip locking for them */
3167 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3168 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3171 /* non-extent files can't have physical blocks past 2^32 */
3172 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3173 pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3176 /* found preallocated blocks, use them */
3177 spin_lock(&pa->pa_lock);
3178 if (pa->pa_deleted == 0 && pa->pa_free) {
3179 atomic_inc(&pa->pa_count);
3180 ext4_mb_use_inode_pa(ac, pa);
3181 spin_unlock(&pa->pa_lock);
3182 ac->ac_criteria = 10;
3186 spin_unlock(&pa->pa_lock);
3190 /* can we use group allocation? */
3191 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3194 /* inode may have no locality group for some reason */
3198 order = fls(ac->ac_o_ex.fe_len) - 1;
3199 if (order > PREALLOC_TB_SIZE - 1)
3200 /* The max size of hash table is PREALLOC_TB_SIZE */
3201 order = PREALLOC_TB_SIZE - 1;
3203 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3205 * search for the prealloc space that is having
3206 * minimal distance from the goal block.
3208 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3210 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3212 spin_lock(&pa->pa_lock);
3213 if (pa->pa_deleted == 0 &&
3214 pa->pa_free >= ac->ac_o_ex.fe_len) {
3216 cpa = ext4_mb_check_group_pa(goal_block,
3219 spin_unlock(&pa->pa_lock);
3224 ext4_mb_use_group_pa(ac, cpa);
3225 ac->ac_criteria = 20;
3232 * the function goes through all block freed in the group
3233 * but not yet committed and marks them used in in-core bitmap.
3234 * buddy must be generated from this bitmap
3235 * Need to be called with the ext4 group lock held
3237 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3241 struct ext4_group_info *grp;
3242 struct ext4_free_data *entry;
3244 grp = ext4_get_group_info(sb, group);
3245 n = rb_first(&(grp->bb_free_root));
3248 entry = rb_entry(n, struct ext4_free_data, node);
3249 mb_set_bits(bitmap, entry->start_blk, entry->count);
3256 * the function goes through all preallocation in this group and marks them
3257 * used in in-core bitmap. buddy must be generated from this bitmap
3258 * Need to be called with ext4 group lock held
3260 static noinline_for_stack
3261 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3264 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3265 struct ext4_prealloc_space *pa;
3266 struct list_head *cur;
3267 ext4_group_t groupnr;
3268 ext4_grpblk_t start;
3269 int preallocated = 0;
3273 /* all form of preallocation discards first load group,
3274 * so the only competing code is preallocation use.
3275 * we don't need any locking here
3276 * notice we do NOT ignore preallocations with pa_deleted
3277 * otherwise we could leave used blocks available for
3278 * allocation in buddy when concurrent ext4_mb_put_pa()
3279 * is dropping preallocation
3281 list_for_each(cur, &grp->bb_prealloc_list) {
3282 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3283 spin_lock(&pa->pa_lock);
3284 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3287 spin_unlock(&pa->pa_lock);
3288 if (unlikely(len == 0))
3290 BUG_ON(groupnr != group);
3291 mb_set_bits(bitmap, start, len);
3292 preallocated += len;
3295 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3298 static void ext4_mb_pa_callback(struct rcu_head *head)
3300 struct ext4_prealloc_space *pa;
3301 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3302 kmem_cache_free(ext4_pspace_cachep, pa);
3306 * drops a reference to preallocated space descriptor
3307 * if this was the last reference and the space is consumed
3309 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3310 struct super_block *sb, struct ext4_prealloc_space *pa)
3313 ext4_fsblk_t grp_blk;
3315 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3318 /* in this short window concurrent discard can set pa_deleted */
3319 spin_lock(&pa->pa_lock);
3320 if (pa->pa_deleted == 1) {
3321 spin_unlock(&pa->pa_lock);
3326 spin_unlock(&pa->pa_lock);
3328 grp_blk = pa->pa_pstart;
3330 * If doing group-based preallocation, pa_pstart may be in the
3331 * next group when pa is used up
3333 if (pa->pa_type == MB_GROUP_PA)
3336 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3341 * P1 (buddy init) P2 (regular allocation)
3342 * find block B in PA
3343 * copy on-disk bitmap to buddy
3344 * mark B in on-disk bitmap
3345 * drop PA from group
3346 * mark all PAs in buddy
3348 * thus, P1 initializes buddy with B available. to prevent this
3349 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3352 ext4_lock_group(sb, grp);
3353 list_del(&pa->pa_group_list);
3354 ext4_unlock_group(sb, grp);
3356 spin_lock(pa->pa_obj_lock);
3357 list_del_rcu(&pa->pa_inode_list);
3358 spin_unlock(pa->pa_obj_lock);
3360 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3364 * creates new preallocated space for given inode
3366 static noinline_for_stack int
3367 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3369 struct super_block *sb = ac->ac_sb;
3370 struct ext4_prealloc_space *pa;
3371 struct ext4_group_info *grp;
3372 struct ext4_inode_info *ei;
3374 /* preallocate only when found space is larger then requested */
3375 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3376 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3377 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3379 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3383 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3389 /* we can't allocate as much as normalizer wants.
3390 * so, found space must get proper lstart
3391 * to cover original request */
3392 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3393 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3395 /* we're limited by original request in that
3396 * logical block must be covered any way
3397 * winl is window we can move our chunk within */
3398 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3400 /* also, we should cover whole original request */
3401 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3403 /* the smallest one defines real window */
3404 win = min(winl, wins);
3406 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3407 if (offs && offs < win)
3410 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3411 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3412 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3415 /* preallocation can change ac_b_ex, thus we store actually
3416 * allocated blocks for history */
3417 ac->ac_f_ex = ac->ac_b_ex;
3419 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3420 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3421 pa->pa_len = ac->ac_b_ex.fe_len;
3422 pa->pa_free = pa->pa_len;
3423 atomic_set(&pa->pa_count, 1);
3424 spin_lock_init(&pa->pa_lock);
3425 INIT_LIST_HEAD(&pa->pa_inode_list);
3426 INIT_LIST_HEAD(&pa->pa_group_list);
3428 pa->pa_type = MB_INODE_PA;
3430 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3431 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3432 trace_ext4_mb_new_inode_pa(ac, pa);
3434 ext4_mb_use_inode_pa(ac, pa);
3435 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3437 ei = EXT4_I(ac->ac_inode);
3438 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3440 pa->pa_obj_lock = &ei->i_prealloc_lock;
3441 pa->pa_inode = ac->ac_inode;
3443 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3444 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3445 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3447 spin_lock(pa->pa_obj_lock);
3448 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3449 spin_unlock(pa->pa_obj_lock);
3455 * creates new preallocated space for locality group inodes belongs to
3457 static noinline_for_stack int
3458 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3460 struct super_block *sb = ac->ac_sb;
3461 struct ext4_locality_group *lg;
3462 struct ext4_prealloc_space *pa;
3463 struct ext4_group_info *grp;
3465 /* preallocate only when found space is larger then requested */
3466 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3467 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3468 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3470 BUG_ON(ext4_pspace_cachep == NULL);
3471 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3475 /* preallocation can change ac_b_ex, thus we store actually
3476 * allocated blocks for history */
3477 ac->ac_f_ex = ac->ac_b_ex;
3479 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3480 pa->pa_lstart = pa->pa_pstart;
3481 pa->pa_len = ac->ac_b_ex.fe_len;
3482 pa->pa_free = pa->pa_len;
3483 atomic_set(&pa->pa_count, 1);
3484 spin_lock_init(&pa->pa_lock);
3485 INIT_LIST_HEAD(&pa->pa_inode_list);
3486 INIT_LIST_HEAD(&pa->pa_group_list);
3488 pa->pa_type = MB_GROUP_PA;
3490 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3491 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3492 trace_ext4_mb_new_group_pa(ac, pa);
3494 ext4_mb_use_group_pa(ac, pa);
3495 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3497 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3501 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3502 pa->pa_inode = NULL;
3504 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3505 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3506 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3509 * We will later add the new pa to the right bucket
3510 * after updating the pa_free in ext4_mb_release_context
3515 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3519 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3520 err = ext4_mb_new_group_pa(ac);
3522 err = ext4_mb_new_inode_pa(ac);
3527 * finds all unused blocks in on-disk bitmap, frees them in
3528 * in-core bitmap and buddy.
3529 * @pa must be unlinked from inode and group lists, so that
3530 * nobody else can find/use it.
3531 * the caller MUST hold group/inode locks.
3532 * TODO: optimize the case when there are no in-core structures yet
3534 static noinline_for_stack int
3535 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3536 struct ext4_prealloc_space *pa,
3537 struct ext4_allocation_context *ac)
3539 struct super_block *sb = e4b->bd_sb;
3540 struct ext4_sb_info *sbi = EXT4_SB(sb);
3545 unsigned long long grp_blk_start;
3549 BUG_ON(pa->pa_deleted == 0);
3550 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3551 grp_blk_start = pa->pa_pstart - bit;
3552 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3553 end = bit + pa->pa_len;
3557 ac->ac_inode = pa->pa_inode;
3561 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3564 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3565 mb_debug(1, " free preallocated %u/%u in group %u\n",
3566 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3567 (unsigned) next - bit, (unsigned) group);
3571 ac->ac_b_ex.fe_group = group;
3572 ac->ac_b_ex.fe_start = bit;
3573 ac->ac_b_ex.fe_len = next - bit;
3574 ac->ac_b_ex.fe_logical = 0;
3575 trace_ext4_mballoc_discard(ac);
3578 trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit,
3580 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3583 if (free != pa->pa_free) {
3584 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3585 pa, (unsigned long) pa->pa_lstart,
3586 (unsigned long) pa->pa_pstart,
3587 (unsigned long) pa->pa_len);
3588 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3591 * pa is already deleted so we use the value obtained
3592 * from the bitmap and continue.
3595 atomic_add(free, &sbi->s_mb_discarded);
3600 static noinline_for_stack int
3601 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3602 struct ext4_prealloc_space *pa,
3603 struct ext4_allocation_context *ac)
3605 struct super_block *sb = e4b->bd_sb;
3609 trace_ext4_mb_release_group_pa(ac, pa);
3610 BUG_ON(pa->pa_deleted == 0);
3611 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3612 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3613 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3614 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3618 ac->ac_inode = NULL;
3619 ac->ac_b_ex.fe_group = group;
3620 ac->ac_b_ex.fe_start = bit;
3621 ac->ac_b_ex.fe_len = pa->pa_len;
3622 ac->ac_b_ex.fe_logical = 0;
3623 trace_ext4_mballoc_discard(ac);
3630 * releases all preallocations in given group
3632 * first, we need to decide discard policy:
3633 * - when do we discard
3635 * - how many do we discard
3636 * 1) how many requested
3638 static noinline_for_stack int
3639 ext4_mb_discard_group_preallocations(struct super_block *sb,
3640 ext4_group_t group, int needed)
3642 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3643 struct buffer_head *bitmap_bh = NULL;
3644 struct ext4_prealloc_space *pa, *tmp;
3645 struct ext4_allocation_context *ac;
3646 struct list_head list;
3647 struct ext4_buddy e4b;
3652 mb_debug(1, "discard preallocation for group %u\n", group);
3654 if (list_empty(&grp->bb_prealloc_list))
3657 bitmap_bh = ext4_read_block_bitmap(sb, group);
3658 if (bitmap_bh == NULL) {
3659 ext4_error(sb, "Error reading block bitmap for %u", group);
3663 err = ext4_mb_load_buddy(sb, group, &e4b);
3665 ext4_error(sb, "Error loading buddy information for %u", group);
3671 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3673 INIT_LIST_HEAD(&list);
3674 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3678 ext4_lock_group(sb, group);
3679 list_for_each_entry_safe(pa, tmp,
3680 &grp->bb_prealloc_list, pa_group_list) {
3681 spin_lock(&pa->pa_lock);
3682 if (atomic_read(&pa->pa_count)) {
3683 spin_unlock(&pa->pa_lock);
3687 if (pa->pa_deleted) {
3688 spin_unlock(&pa->pa_lock);
3692 /* seems this one can be freed ... */
3695 /* we can trust pa_free ... */
3696 free += pa->pa_free;
3698 spin_unlock(&pa->pa_lock);
3700 list_del(&pa->pa_group_list);
3701 list_add(&pa->u.pa_tmp_list, &list);
3704 /* if we still need more blocks and some PAs were used, try again */
3705 if (free < needed && busy) {
3707 ext4_unlock_group(sb, group);
3709 * Yield the CPU here so that we don't get soft lockup
3710 * in non preempt case.
3716 /* found anything to free? */
3717 if (list_empty(&list)) {
3722 /* now free all selected PAs */
3723 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3725 /* remove from object (inode or locality group) */
3726 spin_lock(pa->pa_obj_lock);
3727 list_del_rcu(&pa->pa_inode_list);
3728 spin_unlock(pa->pa_obj_lock);
3730 if (pa->pa_type == MB_GROUP_PA)
3731 ext4_mb_release_group_pa(&e4b, pa, ac);
3733 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3735 list_del(&pa->u.pa_tmp_list);
3736 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3740 ext4_unlock_group(sb, group);
3742 kmem_cache_free(ext4_ac_cachep, ac);
3743 ext4_mb_unload_buddy(&e4b);
3749 * releases all non-used preallocated blocks for given inode
3751 * It's important to discard preallocations under i_data_sem
3752 * We don't want another block to be served from the prealloc
3753 * space when we are discarding the inode prealloc space.
3755 * FIXME!! Make sure it is valid at all the call sites
3757 void ext4_discard_preallocations(struct inode *inode)
3759 struct ext4_inode_info *ei = EXT4_I(inode);
3760 struct super_block *sb = inode->i_sb;
3761 struct buffer_head *bitmap_bh = NULL;
3762 struct ext4_prealloc_space *pa, *tmp;
3763 struct ext4_allocation_context *ac;
3764 ext4_group_t group = 0;
3765 struct list_head list;
3766 struct ext4_buddy e4b;
3769 if (!S_ISREG(inode->i_mode)) {
3770 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3774 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3775 trace_ext4_discard_preallocations(inode);
3777 INIT_LIST_HEAD(&list);
3779 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3782 ac->ac_inode = inode;
3785 /* first, collect all pa's in the inode */
3786 spin_lock(&ei->i_prealloc_lock);
3787 while (!list_empty(&ei->i_prealloc_list)) {
3788 pa = list_entry(ei->i_prealloc_list.next,
3789 struct ext4_prealloc_space, pa_inode_list);
3790 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3791 spin_lock(&pa->pa_lock);
3792 if (atomic_read(&pa->pa_count)) {
3793 /* this shouldn't happen often - nobody should
3794 * use preallocation while we're discarding it */
3795 spin_unlock(&pa->pa_lock);
3796 spin_unlock(&ei->i_prealloc_lock);
3797 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3799 schedule_timeout_uninterruptible(HZ);
3803 if (pa->pa_deleted == 0) {
3805 spin_unlock(&pa->pa_lock);
3806 list_del_rcu(&pa->pa_inode_list);
3807 list_add(&pa->u.pa_tmp_list, &list);
3811 /* someone is deleting pa right now */
3812 spin_unlock(&pa->pa_lock);
3813 spin_unlock(&ei->i_prealloc_lock);
3815 /* we have to wait here because pa_deleted
3816 * doesn't mean pa is already unlinked from
3817 * the list. as we might be called from
3818 * ->clear_inode() the inode will get freed
3819 * and concurrent thread which is unlinking
3820 * pa from inode's list may access already
3821 * freed memory, bad-bad-bad */
3823 /* XXX: if this happens too often, we can
3824 * add a flag to force wait only in case
3825 * of ->clear_inode(), but not in case of
3826 * regular truncate */
3827 schedule_timeout_uninterruptible(HZ);
3830 spin_unlock(&ei->i_prealloc_lock);
3832 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3833 BUG_ON(pa->pa_type != MB_INODE_PA);
3834 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3836 err = ext4_mb_load_buddy(sb, group, &e4b);
3838 ext4_error(sb, "Error loading buddy information for %u",
3843 bitmap_bh = ext4_read_block_bitmap(sb, group);
3844 if (bitmap_bh == NULL) {
3845 ext4_error(sb, "Error reading block bitmap for %u",
3847 ext4_mb_unload_buddy(&e4b);
3851 ext4_lock_group(sb, group);
3852 list_del(&pa->pa_group_list);
3853 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3854 ext4_unlock_group(sb, group);
3856 ext4_mb_unload_buddy(&e4b);
3859 list_del(&pa->u.pa_tmp_list);
3860 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3863 kmem_cache_free(ext4_ac_cachep, ac);
3867 * finds all preallocated spaces and return blocks being freed to them
3868 * if preallocated space becomes full (no block is used from the space)
3869 * then the function frees space in buddy
3870 * XXX: at the moment, truncate (which is the only way to free blocks)
3871 * discards all preallocations
3873 static void ext4_mb_return_to_preallocation(struct inode *inode,
3874 struct ext4_buddy *e4b,
3875 sector_t block, int count)
3877 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
3879 #ifdef CONFIG_EXT4_DEBUG
3880 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3882 struct super_block *sb = ac->ac_sb;
3883 ext4_group_t ngroups, i;
3885 printk(KERN_ERR "EXT4-fs: Can't allocate:"
3886 " Allocation context details:\n");
3887 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
3888 ac->ac_status, ac->ac_flags);
3889 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3890 "best %lu/%lu/%lu@%lu cr %d\n",
3891 (unsigned long)ac->ac_o_ex.fe_group,
3892 (unsigned long)ac->ac_o_ex.fe_start,
3893 (unsigned long)ac->ac_o_ex.fe_len,
3894 (unsigned long)ac->ac_o_ex.fe_logical,
3895 (unsigned long)ac->ac_g_ex.fe_group,
3896 (unsigned long)ac->ac_g_ex.fe_start,
3897 (unsigned long)ac->ac_g_ex.fe_len,
3898 (unsigned long)ac->ac_g_ex.fe_logical,
3899 (unsigned long)ac->ac_b_ex.fe_group,
3900 (unsigned long)ac->ac_b_ex.fe_start,
3901 (unsigned long)ac->ac_b_ex.fe_len,
3902 (unsigned long)ac->ac_b_ex.fe_logical,
3903 (int)ac->ac_criteria);
3904 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
3906 printk(KERN_ERR "EXT4-fs: groups: \n");
3907 ngroups = ext4_get_groups_count(sb);
3908 for (i = 0; i < ngroups; i++) {
3909 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3910 struct ext4_prealloc_space *pa;
3911 ext4_grpblk_t start;
3912 struct list_head *cur;
3913 ext4_lock_group(sb, i);
3914 list_for_each(cur, &grp->bb_prealloc_list) {
3915 pa = list_entry(cur, struct ext4_prealloc_space,
3917 spin_lock(&pa->pa_lock);
3918 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3920 spin_unlock(&pa->pa_lock);
3921 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3924 ext4_unlock_group(sb, i);
3926 if (grp->bb_free == 0)
3928 printk(KERN_ERR "%u: %d/%d \n",
3929 i, grp->bb_free, grp->bb_fragments);
3931 printk(KERN_ERR "\n");
3934 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3941 * We use locality group preallocation for small size file. The size of the
3942 * file is determined by the current size or the resulting size after
3943 * allocation which ever is larger
3945 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3947 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3949 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3950 int bsbits = ac->ac_sb->s_blocksize_bits;
3953 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3956 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3959 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3960 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3963 if ((size == isize) &&
3964 !ext4_fs_is_busy(sbi) &&
3965 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3966 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3970 /* don't use group allocation for large files */
3971 size = max(size, isize);
3972 if (size > sbi->s_mb_stream_request) {
3973 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3977 BUG_ON(ac->ac_lg != NULL);
3979 * locality group prealloc space are per cpu. The reason for having
3980 * per cpu locality group is to reduce the contention between block
3981 * request from multiple CPUs.
3983 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
3985 /* we're going to use group allocation */
3986 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3988 /* serialize all allocations in the group */
3989 mutex_lock(&ac->ac_lg->lg_mutex);
3992 static noinline_for_stack int
3993 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
3994 struct ext4_allocation_request *ar)
3996 struct super_block *sb = ar->inode->i_sb;
3997 struct ext4_sb_info *sbi = EXT4_SB(sb);
3998 struct ext4_super_block *es = sbi->s_es;
4002 ext4_grpblk_t block;
4004 /* we can't allocate > group size */
4007 /* just a dirty hack to filter too big requests */
4008 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4009 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4011 /* start searching from the goal */
4013 if (goal < le32_to_cpu(es->s_first_data_block) ||
4014 goal >= ext4_blocks_count(es))
4015 goal = le32_to_cpu(es->s_first_data_block);
4016 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4018 /* set up allocation goals */
4019 memset(ac, 0, sizeof(struct ext4_allocation_context));
4020 ac->ac_b_ex.fe_logical = ar->logical;
4021 ac->ac_status = AC_STATUS_CONTINUE;
4023 ac->ac_inode = ar->inode;
4024 ac->ac_o_ex.fe_logical = ar->logical;
4025 ac->ac_o_ex.fe_group = group;
4026 ac->ac_o_ex.fe_start = block;
4027 ac->ac_o_ex.fe_len = len;
4028 ac->ac_g_ex.fe_logical = ar->logical;
4029 ac->ac_g_ex.fe_group = group;
4030 ac->ac_g_ex.fe_start = block;
4031 ac->ac_g_ex.fe_len = len;
4032 ac->ac_flags = ar->flags;
4034 /* we have to define context: we'll we work with a file or
4035 * locality group. this is a policy, actually */
4036 ext4_mb_group_or_file(ac);
4038 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4039 "left: %u/%u, right %u/%u to %swritable\n",
4040 (unsigned) ar->len, (unsigned) ar->logical,
4041 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4042 (unsigned) ar->lleft, (unsigned) ar->pleft,
4043 (unsigned) ar->lright, (unsigned) ar->pright,
4044 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4049 static noinline_for_stack void
4050 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4051 struct ext4_locality_group *lg,
4052 int order, int total_entries)
4054 ext4_group_t group = 0;
4055 struct ext4_buddy e4b;
4056 struct list_head discard_list;
4057 struct ext4_prealloc_space *pa, *tmp;
4058 struct ext4_allocation_context *ac;
4060 mb_debug(1, "discard locality group preallocation\n");
4062 INIT_LIST_HEAD(&discard_list);
4063 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4067 spin_lock(&lg->lg_prealloc_lock);
4068 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4070 spin_lock(&pa->pa_lock);
4071 if (atomic_read(&pa->pa_count)) {
4073 * This is the pa that we just used
4074 * for block allocation. So don't
4077 spin_unlock(&pa->pa_lock);
4080 if (pa->pa_deleted) {
4081 spin_unlock(&pa->pa_lock);
4084 /* only lg prealloc space */
4085 BUG_ON(pa->pa_type != MB_GROUP_PA);
4087 /* seems this one can be freed ... */
4089 spin_unlock(&pa->pa_lock);
4091 list_del_rcu(&pa->pa_inode_list);
4092 list_add(&pa->u.pa_tmp_list, &discard_list);
4095 if (total_entries <= 5) {
4097 * we want to keep only 5 entries
4098 * allowing it to grow to 8. This
4099 * mak sure we don't call discard
4100 * soon for this list.
4105 spin_unlock(&lg->lg_prealloc_lock);
4107 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4109 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4110 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4111 ext4_error(sb, "Error loading buddy information for %u",
4115 ext4_lock_group(sb, group);
4116 list_del(&pa->pa_group_list);
4117 ext4_mb_release_group_pa(&e4b, pa, ac);
4118 ext4_unlock_group(sb, group);
4120 ext4_mb_unload_buddy(&e4b);
4121 list_del(&pa->u.pa_tmp_list);
4122 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4125 kmem_cache_free(ext4_ac_cachep, ac);
4129 * We have incremented pa_count. So it cannot be freed at this
4130 * point. Also we hold lg_mutex. So no parallel allocation is
4131 * possible from this lg. That means pa_free cannot be updated.
4133 * A parallel ext4_mb_discard_group_preallocations is possible.
4134 * which can cause the lg_prealloc_list to be updated.
4137 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4139 int order, added = 0, lg_prealloc_count = 1;
4140 struct super_block *sb = ac->ac_sb;
4141 struct ext4_locality_group *lg = ac->ac_lg;
4142 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4144 order = fls(pa->pa_free) - 1;
4145 if (order > PREALLOC_TB_SIZE - 1)
4146 /* The max size of hash table is PREALLOC_TB_SIZE */
4147 order = PREALLOC_TB_SIZE - 1;
4148 /* Add the prealloc space to lg */
4150 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4152 spin_lock(&tmp_pa->pa_lock);
4153 if (tmp_pa->pa_deleted) {
4154 spin_unlock(&tmp_pa->pa_lock);
4157 if (!added && pa->pa_free < tmp_pa->pa_free) {
4158 /* Add to the tail of the previous entry */
4159 list_add_tail_rcu(&pa->pa_inode_list,
4160 &tmp_pa->pa_inode_list);
4163 * we want to count the total
4164 * number of entries in the list
4167 spin_unlock(&tmp_pa->pa_lock);
4168 lg_prealloc_count++;
4171 list_add_tail_rcu(&pa->pa_inode_list,
4172 &lg->lg_prealloc_list[order]);
4175 /* Now trim the list to be not more than 8 elements */
4176 if (lg_prealloc_count > 8) {
4177 ext4_mb_discard_lg_preallocations(sb, lg,
4178 order, lg_prealloc_count);
4185 * release all resource we used in allocation
4187 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4189 struct ext4_prealloc_space *pa = ac->ac_pa;
4191 if (pa->pa_type == MB_GROUP_PA) {
4192 /* see comment in ext4_mb_use_group_pa() */
4193 spin_lock(&pa->pa_lock);
4194 pa->pa_pstart += ac->ac_b_ex.fe_len;
4195 pa->pa_lstart += ac->ac_b_ex.fe_len;
4196 pa->pa_free -= ac->ac_b_ex.fe_len;
4197 pa->pa_len -= ac->ac_b_ex.fe_len;
4198 spin_unlock(&pa->pa_lock);
4202 up_read(ac->alloc_semp);
4205 * We want to add the pa to the right bucket.
4206 * Remove it from the list and while adding
4207 * make sure the list to which we are adding
4208 * doesn't grow big. We need to release
4209 * alloc_semp before calling ext4_mb_add_n_trim()
4211 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4212 spin_lock(pa->pa_obj_lock);
4213 list_del_rcu(&pa->pa_inode_list);
4214 spin_unlock(pa->pa_obj_lock);
4215 ext4_mb_add_n_trim(ac);
4217 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4219 if (ac->ac_bitmap_page)
4220 page_cache_release(ac->ac_bitmap_page);
4221 if (ac->ac_buddy_page)
4222 page_cache_release(ac->ac_buddy_page);
4223 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4224 mutex_unlock(&ac->ac_lg->lg_mutex);
4225 ext4_mb_collect_stats(ac);
4229 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4231 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4235 trace_ext4_mb_discard_preallocations(sb, needed);
4236 for (i = 0; i < ngroups && needed > 0; i++) {
4237 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4246 * Main entry point into mballoc to allocate blocks
4247 * it tries to use preallocation first, then falls back
4248 * to usual allocation
4250 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4251 struct ext4_allocation_request *ar, int *errp)
4254 struct ext4_allocation_context *ac = NULL;
4255 struct ext4_sb_info *sbi;
4256 struct super_block *sb;
4257 ext4_fsblk_t block = 0;
4258 unsigned int inquota = 0;
4259 unsigned int reserv_blks = 0;
4261 sb = ar->inode->i_sb;
4264 trace_ext4_request_blocks(ar);
4267 * For delayed allocation, we could skip the ENOSPC and
4268 * EDQUOT check, as blocks and quotas have been already
4269 * reserved when data being copied into pagecache.
4271 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4272 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4274 /* Without delayed allocation we need to verify
4275 * there is enough free blocks to do block allocation
4276 * and verify allocation doesn't exceed the quota limits.
4278 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4279 /* let others to free the space */
4281 ar->len = ar->len >> 1;
4287 reserv_blks = ar->len;
4288 while (ar->len && dquot_alloc_block(ar->inode, ar->len)) {
4289 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4299 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4306 *errp = ext4_mb_initialize_context(ac, ar);
4312 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4313 if (!ext4_mb_use_preallocated(ac)) {
4314 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4315 ext4_mb_normalize_request(ac, ar);
4317 /* allocate space in core */
4318 ext4_mb_regular_allocator(ac);
4320 /* as we've just preallocated more space than
4321 * user requested orinally, we store allocated
4322 * space in a special descriptor */
4323 if (ac->ac_status == AC_STATUS_FOUND &&
4324 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4325 ext4_mb_new_preallocation(ac);
4327 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4328 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4329 if (*errp == -EAGAIN) {
4331 * drop the reference that we took
4332 * in ext4_mb_use_best_found
4334 ext4_mb_release_context(ac);
4335 ac->ac_b_ex.fe_group = 0;
4336 ac->ac_b_ex.fe_start = 0;
4337 ac->ac_b_ex.fe_len = 0;
4338 ac->ac_status = AC_STATUS_CONTINUE;
4341 ext4_discard_allocated_blocks(ac);
4342 ac->ac_b_ex.fe_len = 0;
4344 ext4_mb_show_ac(ac);
4346 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4347 ar->len = ac->ac_b_ex.fe_len;
4350 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4354 ac->ac_b_ex.fe_len = 0;
4356 ext4_mb_show_ac(ac);
4359 ext4_mb_release_context(ac);
4362 kmem_cache_free(ext4_ac_cachep, ac);
4364 if (inquota && ar->len < inquota)
4365 dquot_free_block(ar->inode, inquota - ar->len);
4368 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4369 /* release all the reserved blocks if non delalloc */
4370 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4374 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4380 * We can merge two free data extents only if the physical blocks
4381 * are contiguous, AND the extents were freed by the same transaction,
4382 * AND the blocks are associated with the same group.
4384 static int can_merge(struct ext4_free_data *entry1,
4385 struct ext4_free_data *entry2)
4387 if ((entry1->t_tid == entry2->t_tid) &&
4388 (entry1->group == entry2->group) &&
4389 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4394 static noinline_for_stack int
4395 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4396 struct ext4_free_data *new_entry)
4398 ext4_group_t group = e4b->bd_group;
4399 ext4_grpblk_t block;
4400 struct ext4_free_data *entry;
4401 struct ext4_group_info *db = e4b->bd_info;
4402 struct super_block *sb = e4b->bd_sb;
4403 struct ext4_sb_info *sbi = EXT4_SB(sb);
4404 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4405 struct rb_node *parent = NULL, *new_node;
4407 BUG_ON(!ext4_handle_valid(handle));
4408 BUG_ON(e4b->bd_bitmap_page == NULL);
4409 BUG_ON(e4b->bd_buddy_page == NULL);
4411 new_node = &new_entry->node;
4412 block = new_entry->start_blk;
4415 /* first free block exent. We need to
4416 protect buddy cache from being freed,
4417 * otherwise we'll refresh it from
4418 * on-disk bitmap and lose not-yet-available
4420 page_cache_get(e4b->bd_buddy_page);
4421 page_cache_get(e4b->bd_bitmap_page);
4425 entry = rb_entry(parent, struct ext4_free_data, node);
4426 if (block < entry->start_blk)
4428 else if (block >= (entry->start_blk + entry->count))
4429 n = &(*n)->rb_right;
4431 ext4_grp_locked_error(sb, group, 0,
4432 ext4_group_first_block_no(sb, group) + block,
4433 "Block already on to-be-freed list");
4438 rb_link_node(new_node, parent, n);
4439 rb_insert_color(new_node, &db->bb_free_root);
4441 /* Now try to see the extent can be merged to left and right */
4442 node = rb_prev(new_node);
4444 entry = rb_entry(node, struct ext4_free_data, node);
4445 if (can_merge(entry, new_entry)) {
4446 new_entry->start_blk = entry->start_blk;
4447 new_entry->count += entry->count;
4448 rb_erase(node, &(db->bb_free_root));
4449 spin_lock(&sbi->s_md_lock);
4450 list_del(&entry->list);
4451 spin_unlock(&sbi->s_md_lock);
4452 kmem_cache_free(ext4_free_ext_cachep, entry);
4456 node = rb_next(new_node);
4458 entry = rb_entry(node, struct ext4_free_data, node);
4459 if (can_merge(new_entry, entry)) {
4460 new_entry->count += entry->count;
4461 rb_erase(node, &(db->bb_free_root));
4462 spin_lock(&sbi->s_md_lock);
4463 list_del(&entry->list);
4464 spin_unlock(&sbi->s_md_lock);
4465 kmem_cache_free(ext4_free_ext_cachep, entry);
4468 /* Add the extent to transaction's private list */
4469 spin_lock(&sbi->s_md_lock);
4470 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4471 spin_unlock(&sbi->s_md_lock);
4476 * ext4_free_blocks() -- Free given blocks and update quota
4477 * @handle: handle for this transaction
4479 * @block: start physical block to free
4480 * @count: number of blocks to count
4481 * @metadata: Are these metadata blocks
4483 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4484 struct buffer_head *bh, ext4_fsblk_t block,
4485 unsigned long count, int flags)
4487 struct buffer_head *bitmap_bh = NULL;
4488 struct super_block *sb = inode->i_sb;
4489 struct ext4_allocation_context *ac = NULL;
4490 struct ext4_group_desc *gdp;
4491 unsigned long freed = 0;
4492 unsigned int overflow;
4494 struct buffer_head *gd_bh;
4495 ext4_group_t block_group;
4496 struct ext4_sb_info *sbi;
4497 struct ext4_buddy e4b;
4503 BUG_ON(block != bh->b_blocknr);
4505 block = bh->b_blocknr;
4509 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4510 !ext4_data_block_valid(sbi, block, count)) {
4511 ext4_error(sb, "Freeing blocks not in datazone - "
4512 "block = %llu, count = %lu", block, count);
4516 ext4_debug("freeing block %llu\n", block);
4517 trace_ext4_free_blocks(inode, block, count, flags);
4519 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4520 struct buffer_head *tbh = bh;
4523 BUG_ON(bh && (count > 1));
4525 for (i = 0; i < count; i++) {
4527 tbh = sb_find_get_block(inode->i_sb,
4529 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4530 inode, tbh, block + i);
4535 * We need to make sure we don't reuse the freed block until
4536 * after the transaction is committed, which we can do by
4537 * treating the block as metadata, below. We make an
4538 * exception if the inode is to be written in writeback mode
4539 * since writeback mode has weak data consistency guarantees.
4541 if (!ext4_should_writeback_data(inode))
4542 flags |= EXT4_FREE_BLOCKS_METADATA;
4544 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4546 ac->ac_inode = inode;
4552 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4555 * Check to see if we are freeing blocks across a group
4558 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4559 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4562 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4567 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4573 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4574 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4575 in_range(block, ext4_inode_table(sb, gdp),
4576 EXT4_SB(sb)->s_itb_per_group) ||
4577 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4578 EXT4_SB(sb)->s_itb_per_group)) {
4580 ext4_error(sb, "Freeing blocks in system zone - "
4581 "Block = %llu, count = %lu", block, count);
4582 /* err = 0. ext4_std_error should be a no op */
4586 BUFFER_TRACE(bitmap_bh, "getting write access");
4587 err = ext4_journal_get_write_access(handle, bitmap_bh);
4592 * We are about to modify some metadata. Call the journal APIs
4593 * to unshare ->b_data if a currently-committing transaction is
4596 BUFFER_TRACE(gd_bh, "get_write_access");
4597 err = ext4_journal_get_write_access(handle, gd_bh);
4600 #ifdef AGGRESSIVE_CHECK
4603 for (i = 0; i < count; i++)
4604 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4608 ac->ac_b_ex.fe_group = block_group;
4609 ac->ac_b_ex.fe_start = bit;
4610 ac->ac_b_ex.fe_len = count;
4611 trace_ext4_mballoc_free(ac);
4614 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4618 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4619 struct ext4_free_data *new_entry;
4621 * blocks being freed are metadata. these blocks shouldn't
4622 * be used until this transaction is committed
4624 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4625 new_entry->start_blk = bit;
4626 new_entry->group = block_group;
4627 new_entry->count = count;
4628 new_entry->t_tid = handle->h_transaction->t_tid;
4630 ext4_lock_group(sb, block_group);
4631 mb_clear_bits(bitmap_bh->b_data, bit, count);
4632 ext4_mb_free_metadata(handle, &e4b, new_entry);
4634 /* need to update group_info->bb_free and bitmap
4635 * with group lock held. generate_buddy look at
4636 * them with group lock_held
4638 ext4_lock_group(sb, block_group);
4639 mb_clear_bits(bitmap_bh->b_data, bit, count);
4640 mb_free_blocks(inode, &e4b, bit, count);
4641 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4644 ret = ext4_free_blks_count(sb, gdp) + count;
4645 ext4_free_blks_set(sb, gdp, ret);
4646 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4647 ext4_unlock_group(sb, block_group);
4648 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4650 if (sbi->s_log_groups_per_flex) {
4651 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4652 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4655 ext4_mb_unload_buddy(&e4b);
4659 /* We dirtied the bitmap block */
4660 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4661 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4663 /* And the group descriptor block */
4664 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4665 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4669 if (overflow && !err) {
4675 ext4_mark_super_dirty(sb);
4678 dquot_free_block(inode, freed);
4680 ext4_std_error(sb, err);
4682 kmem_cache_free(ext4_ac_cachep, ac);