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-multiple requests
1827 static noinline_for_stack
1828 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1829 struct ext4_buddy *e4b)
1831 struct super_block *sb = ac->ac_sb;
1832 struct ext4_sb_info *sbi = EXT4_SB(sb);
1833 void *bitmap = EXT4_MB_BITMAP(e4b);
1834 struct ext4_free_extent ex;
1835 ext4_fsblk_t first_group_block;
1840 BUG_ON(sbi->s_stripe == 0);
1842 /* find first stripe-aligned block in group */
1843 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1845 a = first_group_block + sbi->s_stripe - 1;
1846 do_div(a, sbi->s_stripe);
1847 i = (a * sbi->s_stripe) - first_group_block;
1849 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1850 if (!mb_test_bit(i, bitmap)) {
1851 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1852 if (max >= sbi->s_stripe) {
1855 ext4_mb_use_best_found(ac, e4b);
1863 /* This is now called BEFORE we load the buddy bitmap. */
1864 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1865 ext4_group_t group, int cr)
1867 unsigned free, fragments;
1868 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1869 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1871 BUG_ON(cr < 0 || cr >= 4);
1873 /* We only do this if the grp has never been initialized */
1874 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1875 int ret = ext4_mb_init_group(ac->ac_sb, group);
1880 free = grp->bb_free;
1881 fragments = grp->bb_fragments;
1889 BUG_ON(ac->ac_2order == 0);
1891 if (grp->bb_largest_free_order < ac->ac_2order)
1894 /* Avoid using the first bg of a flexgroup for data files */
1895 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1896 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1897 ((group % flex_size) == 0))
1902 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1906 if (free >= ac->ac_g_ex.fe_len)
1919 * lock the group_info alloc_sem of all the groups
1920 * belonging to the same buddy cache page. This
1921 * make sure other parallel operation on the buddy
1922 * cache doesn't happen whild holding the buddy cache
1925 int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1929 int blocks_per_page;
1930 int groups_per_page;
1931 ext4_group_t ngroups = ext4_get_groups_count(sb);
1932 ext4_group_t first_group;
1933 struct ext4_group_info *grp;
1935 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1937 * the buddy cache inode stores the block bitmap
1938 * and buddy information in consecutive blocks.
1939 * So for each group we need two blocks.
1942 pnum = block / blocks_per_page;
1943 first_group = pnum * blocks_per_page / 2;
1945 groups_per_page = blocks_per_page >> 1;
1946 if (groups_per_page == 0)
1947 groups_per_page = 1;
1948 /* read all groups the page covers into the cache */
1949 for (i = 0; i < groups_per_page; i++) {
1951 if ((first_group + i) >= ngroups)
1953 grp = ext4_get_group_info(sb, first_group + i);
1954 /* take all groups write allocation
1955 * semaphore. This make sure there is
1956 * no block allocation going on in any
1959 down_write_nested(&grp->alloc_sem, i);
1964 void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1965 ext4_group_t group, int locked_group)
1969 int blocks_per_page;
1970 ext4_group_t first_group;
1971 struct ext4_group_info *grp;
1973 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1975 * the buddy cache inode stores the block bitmap
1976 * and buddy information in consecutive blocks.
1977 * So for each group we need two blocks.
1980 pnum = block / blocks_per_page;
1981 first_group = pnum * blocks_per_page / 2;
1982 /* release locks on all the groups */
1983 for (i = 0; i < locked_group; i++) {
1985 grp = ext4_get_group_info(sb, first_group + i);
1986 /* take all groups write allocation
1987 * semaphore. This make sure there is
1988 * no block allocation going on in any
1991 up_write(&grp->alloc_sem);
1996 static noinline_for_stack int
1997 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1999 ext4_group_t ngroups, group, i;
2002 struct ext4_sb_info *sbi;
2003 struct super_block *sb;
2004 struct ext4_buddy e4b;
2008 ngroups = ext4_get_groups_count(sb);
2009 /* non-extent files are limited to low blocks/groups */
2010 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2011 ngroups = sbi->s_blockfile_groups;
2013 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2015 /* first, try the goal */
2016 err = ext4_mb_find_by_goal(ac, &e4b);
2017 if (err || ac->ac_status == AC_STATUS_FOUND)
2020 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2024 * ac->ac2_order is set only if the fe_len is a power of 2
2025 * if ac2_order is set we also set criteria to 0 so that we
2026 * try exact allocation using buddy.
2028 i = fls(ac->ac_g_ex.fe_len);
2031 * We search using buddy data only if the order of the request
2032 * is greater than equal to the sbi_s_mb_order2_reqs
2033 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2035 if (i >= sbi->s_mb_order2_reqs) {
2037 * This should tell if fe_len is exactly power of 2
2039 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2040 ac->ac_2order = i - 1;
2043 /* if stream allocation is enabled, use global goal */
2044 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2045 /* TBD: may be hot point */
2046 spin_lock(&sbi->s_md_lock);
2047 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2048 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2049 spin_unlock(&sbi->s_md_lock);
2052 /* Let's just scan groups to find more-less suitable blocks */
2053 cr = ac->ac_2order ? 0 : 1;
2055 * cr == 0 try to get exact allocation,
2056 * cr == 3 try to get anything
2059 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2060 ac->ac_criteria = cr;
2062 * searching for the right group start
2063 * from the goal value specified
2065 group = ac->ac_g_ex.fe_group;
2067 for (i = 0; i < ngroups; group++, i++) {
2068 if (group == ngroups)
2071 /* This now checks without needing the buddy page */
2072 if (!ext4_mb_good_group(ac, group, cr))
2075 err = ext4_mb_load_buddy(sb, group, &e4b);
2079 ext4_lock_group(sb, group);
2082 * We need to check again after locking the
2085 if (!ext4_mb_good_group(ac, group, cr)) {
2086 ext4_unlock_group(sb, group);
2087 ext4_mb_unload_buddy(&e4b);
2091 ac->ac_groups_scanned++;
2093 ext4_mb_simple_scan_group(ac, &e4b);
2094 else if (cr == 1 && sbi->s_stripe &&
2095 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2096 ext4_mb_scan_aligned(ac, &e4b);
2098 ext4_mb_complex_scan_group(ac, &e4b);
2100 ext4_unlock_group(sb, group);
2101 ext4_mb_unload_buddy(&e4b);
2103 if (ac->ac_status != AC_STATUS_CONTINUE)
2108 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2109 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2111 * We've been searching too long. Let's try to allocate
2112 * the best chunk we've found so far
2115 ext4_mb_try_best_found(ac, &e4b);
2116 if (ac->ac_status != AC_STATUS_FOUND) {
2118 * Someone more lucky has already allocated it.
2119 * The only thing we can do is just take first
2121 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2123 ac->ac_b_ex.fe_group = 0;
2124 ac->ac_b_ex.fe_start = 0;
2125 ac->ac_b_ex.fe_len = 0;
2126 ac->ac_status = AC_STATUS_CONTINUE;
2127 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2129 atomic_inc(&sbi->s_mb_lost_chunks);
2137 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2139 struct super_block *sb = seq->private;
2142 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2145 return (void *) ((unsigned long) group);
2148 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2150 struct super_block *sb = seq->private;
2154 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2157 return (void *) ((unsigned long) group);
2160 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2162 struct super_block *sb = seq->private;
2163 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2166 struct ext4_buddy e4b;
2168 struct ext4_group_info info;
2169 ext4_grpblk_t counters[16];
2174 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2175 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2176 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2177 "group", "free", "frags", "first",
2178 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2179 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2181 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2182 sizeof(struct ext4_group_info);
2183 err = ext4_mb_load_buddy(sb, group, &e4b);
2185 seq_printf(seq, "#%-5u: I/O error\n", group);
2188 ext4_lock_group(sb, group);
2189 memcpy(&sg, ext4_get_group_info(sb, group), i);
2190 ext4_unlock_group(sb, group);
2191 ext4_mb_unload_buddy(&e4b);
2193 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2194 sg.info.bb_fragments, sg.info.bb_first_free);
2195 for (i = 0; i <= 13; i++)
2196 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2197 sg.info.bb_counters[i] : 0);
2198 seq_printf(seq, " ]\n");
2203 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2207 static const struct seq_operations ext4_mb_seq_groups_ops = {
2208 .start = ext4_mb_seq_groups_start,
2209 .next = ext4_mb_seq_groups_next,
2210 .stop = ext4_mb_seq_groups_stop,
2211 .show = ext4_mb_seq_groups_show,
2214 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2216 struct super_block *sb = PDE(inode)->data;
2219 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2221 struct seq_file *m = file->private_data;
2228 static const struct file_operations ext4_mb_seq_groups_fops = {
2229 .owner = THIS_MODULE,
2230 .open = ext4_mb_seq_groups_open,
2232 .llseek = seq_lseek,
2233 .release = seq_release,
2237 /* Create and initialize ext4_group_info data for the given group. */
2238 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2239 struct ext4_group_desc *desc)
2243 struct ext4_sb_info *sbi = EXT4_SB(sb);
2244 struct ext4_group_info **meta_group_info;
2247 * First check if this group is the first of a reserved block.
2248 * If it's true, we have to allocate a new table of pointers
2249 * to ext4_group_info structures
2251 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2252 metalen = sizeof(*meta_group_info) <<
2253 EXT4_DESC_PER_BLOCK_BITS(sb);
2254 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2255 if (meta_group_info == NULL) {
2256 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2258 goto exit_meta_group_info;
2260 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2265 * calculate needed size. if change bb_counters size,
2266 * don't forget about ext4_mb_generate_buddy()
2268 len = offsetof(typeof(**meta_group_info),
2269 bb_counters[sb->s_blocksize_bits + 2]);
2272 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2273 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2275 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2276 if (meta_group_info[i] == NULL) {
2277 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2278 goto exit_group_info;
2280 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2281 &(meta_group_info[i]->bb_state));
2284 * initialize bb_free to be able to skip
2285 * empty groups without initialization
2287 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2288 meta_group_info[i]->bb_free =
2289 ext4_free_blocks_after_init(sb, group, desc);
2291 meta_group_info[i]->bb_free =
2292 ext4_free_blks_count(sb, desc);
2295 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2296 init_rwsem(&meta_group_info[i]->alloc_sem);
2297 meta_group_info[i]->bb_free_root = RB_ROOT;
2298 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2302 struct buffer_head *bh;
2303 meta_group_info[i]->bb_bitmap =
2304 kmalloc(sb->s_blocksize, GFP_KERNEL);
2305 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2306 bh = ext4_read_block_bitmap(sb, group);
2308 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2317 /* If a meta_group_info table has been allocated, release it now */
2318 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2319 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2320 exit_meta_group_info:
2322 } /* ext4_mb_add_groupinfo */
2324 static int ext4_mb_init_backend(struct super_block *sb)
2326 ext4_group_t ngroups = ext4_get_groups_count(sb);
2328 struct ext4_sb_info *sbi = EXT4_SB(sb);
2329 struct ext4_super_block *es = sbi->s_es;
2330 int num_meta_group_infos;
2331 int num_meta_group_infos_max;
2333 struct ext4_group_desc *desc;
2335 /* This is the number of blocks used by GDT */
2336 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2337 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2340 * This is the total number of blocks used by GDT including
2341 * the number of reserved blocks for GDT.
2342 * The s_group_info array is allocated with this value
2343 * to allow a clean online resize without a complex
2344 * manipulation of pointer.
2345 * The drawback is the unused memory when no resize
2346 * occurs but it's very low in terms of pages
2347 * (see comments below)
2348 * Need to handle this properly when META_BG resizing is allowed
2350 num_meta_group_infos_max = num_meta_group_infos +
2351 le16_to_cpu(es->s_reserved_gdt_blocks);
2354 * array_size is the size of s_group_info array. We round it
2355 * to the next power of two because this approximation is done
2356 * internally by kmalloc so we can have some more memory
2357 * for free here (e.g. may be used for META_BG resize).
2360 while (array_size < sizeof(*sbi->s_group_info) *
2361 num_meta_group_infos_max)
2362 array_size = array_size << 1;
2363 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2364 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2365 * So a two level scheme suffices for now. */
2366 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2367 if (sbi->s_group_info == NULL) {
2368 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2371 sbi->s_buddy_cache = new_inode(sb);
2372 if (sbi->s_buddy_cache == NULL) {
2373 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2376 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2377 for (i = 0; i < ngroups; i++) {
2378 desc = ext4_get_group_desc(sb, i, NULL);
2381 "EXT4-fs: can't read descriptor %u\n", i);
2384 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2392 kfree(ext4_get_group_info(sb, i));
2393 i = num_meta_group_infos;
2395 kfree(sbi->s_group_info[i]);
2396 iput(sbi->s_buddy_cache);
2398 kfree(sbi->s_group_info);
2402 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2404 struct ext4_sb_info *sbi = EXT4_SB(sb);
2410 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2412 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2413 if (sbi->s_mb_offsets == NULL) {
2417 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2418 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2419 if (sbi->s_mb_maxs == NULL) {
2420 kfree(sbi->s_mb_offsets);
2424 /* order 0 is regular bitmap */
2425 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2426 sbi->s_mb_offsets[0] = 0;
2430 max = sb->s_blocksize << 2;
2432 sbi->s_mb_offsets[i] = offset;
2433 sbi->s_mb_maxs[i] = max;
2434 offset += 1 << (sb->s_blocksize_bits - i);
2437 } while (i <= sb->s_blocksize_bits + 1);
2439 /* init file for buddy data */
2440 ret = ext4_mb_init_backend(sb);
2442 kfree(sbi->s_mb_offsets);
2443 kfree(sbi->s_mb_maxs);
2447 spin_lock_init(&sbi->s_md_lock);
2448 spin_lock_init(&sbi->s_bal_lock);
2450 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2451 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2452 sbi->s_mb_stats = MB_DEFAULT_STATS;
2453 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2454 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2455 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2457 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2458 if (sbi->s_locality_groups == NULL) {
2459 kfree(sbi->s_mb_offsets);
2460 kfree(sbi->s_mb_maxs);
2463 for_each_possible_cpu(i) {
2464 struct ext4_locality_group *lg;
2465 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2466 mutex_init(&lg->lg_mutex);
2467 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2468 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2469 spin_lock_init(&lg->lg_prealloc_lock);
2473 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2474 &ext4_mb_seq_groups_fops, sb);
2477 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2481 /* need to called with the ext4 group lock held */
2482 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2484 struct ext4_prealloc_space *pa;
2485 struct list_head *cur, *tmp;
2488 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2489 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2490 list_del(&pa->pa_group_list);
2492 kmem_cache_free(ext4_pspace_cachep, pa);
2495 mb_debug(1, "mballoc: %u PAs left\n", count);
2499 int ext4_mb_release(struct super_block *sb)
2501 ext4_group_t ngroups = ext4_get_groups_count(sb);
2503 int num_meta_group_infos;
2504 struct ext4_group_info *grinfo;
2505 struct ext4_sb_info *sbi = EXT4_SB(sb);
2507 if (sbi->s_group_info) {
2508 for (i = 0; i < ngroups; i++) {
2509 grinfo = ext4_get_group_info(sb, i);
2511 kfree(grinfo->bb_bitmap);
2513 ext4_lock_group(sb, i);
2514 ext4_mb_cleanup_pa(grinfo);
2515 ext4_unlock_group(sb, i);
2518 num_meta_group_infos = (ngroups +
2519 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2520 EXT4_DESC_PER_BLOCK_BITS(sb);
2521 for (i = 0; i < num_meta_group_infos; i++)
2522 kfree(sbi->s_group_info[i]);
2523 kfree(sbi->s_group_info);
2525 kfree(sbi->s_mb_offsets);
2526 kfree(sbi->s_mb_maxs);
2527 if (sbi->s_buddy_cache)
2528 iput(sbi->s_buddy_cache);
2529 if (sbi->s_mb_stats) {
2531 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2532 atomic_read(&sbi->s_bal_allocated),
2533 atomic_read(&sbi->s_bal_reqs),
2534 atomic_read(&sbi->s_bal_success));
2536 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2537 "%u 2^N hits, %u breaks, %u lost\n",
2538 atomic_read(&sbi->s_bal_ex_scanned),
2539 atomic_read(&sbi->s_bal_goals),
2540 atomic_read(&sbi->s_bal_2orders),
2541 atomic_read(&sbi->s_bal_breaks),
2542 atomic_read(&sbi->s_mb_lost_chunks));
2544 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2545 sbi->s_mb_buddies_generated++,
2546 sbi->s_mb_generation_time);
2548 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2549 atomic_read(&sbi->s_mb_preallocated),
2550 atomic_read(&sbi->s_mb_discarded));
2553 free_percpu(sbi->s_locality_groups);
2555 remove_proc_entry("mb_groups", sbi->s_proc);
2560 static inline void ext4_issue_discard(struct super_block *sb,
2561 ext4_group_t block_group, ext4_grpblk_t block, int count)
2564 ext4_fsblk_t discard_block;
2566 discard_block = block + ext4_group_first_block_no(sb, block_group);
2567 trace_ext4_discard_blocks(sb,
2568 (unsigned long long) discard_block, count);
2569 ret = sb_issue_discard(sb, discard_block, count);
2570 if (ret == EOPNOTSUPP) {
2571 ext4_warning(sb, "discard not supported, disabling");
2572 clear_opt(EXT4_SB(sb)->s_mount_opt, DISCARD);
2577 * This function is called by the jbd2 layer once the commit has finished,
2578 * so we know we can free the blocks that were released with that commit.
2580 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2582 struct super_block *sb = journal->j_private;
2583 struct ext4_buddy e4b;
2584 struct ext4_group_info *db;
2585 int err, count = 0, count2 = 0;
2586 struct ext4_free_data *entry;
2587 struct list_head *l, *ltmp;
2589 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2590 entry = list_entry(l, struct ext4_free_data, list);
2592 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2593 entry->count, entry->group, entry);
2595 if (test_opt(sb, DISCARD))
2596 ext4_issue_discard(sb, entry->group,
2597 entry->start_blk, entry->count);
2599 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2600 /* we expect to find existing buddy because it's pinned */
2604 /* there are blocks to put in buddy to make them really free */
2605 count += entry->count;
2607 ext4_lock_group(sb, entry->group);
2608 /* Take it out of per group rb tree */
2609 rb_erase(&entry->node, &(db->bb_free_root));
2610 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2612 if (!db->bb_free_root.rb_node) {
2613 /* No more items in the per group rb tree
2614 * balance refcounts from ext4_mb_free_metadata()
2616 page_cache_release(e4b.bd_buddy_page);
2617 page_cache_release(e4b.bd_bitmap_page);
2619 ext4_unlock_group(sb, entry->group);
2620 kmem_cache_free(ext4_free_ext_cachep, entry);
2621 ext4_mb_unload_buddy(&e4b);
2624 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2627 #ifdef CONFIG_EXT4_DEBUG
2628 u8 mb_enable_debug __read_mostly;
2630 static struct dentry *debugfs_dir;
2631 static struct dentry *debugfs_debug;
2633 static void __init ext4_create_debugfs_entry(void)
2635 debugfs_dir = debugfs_create_dir("ext4", NULL);
2637 debugfs_debug = debugfs_create_u8("mballoc-debug",
2643 static void ext4_remove_debugfs_entry(void)
2645 debugfs_remove(debugfs_debug);
2646 debugfs_remove(debugfs_dir);
2651 static void __init ext4_create_debugfs_entry(void)
2655 static void ext4_remove_debugfs_entry(void)
2661 int __init init_ext4_mballoc(void)
2663 ext4_pspace_cachep =
2664 kmem_cache_create("ext4_prealloc_space",
2665 sizeof(struct ext4_prealloc_space),
2666 0, SLAB_RECLAIM_ACCOUNT, NULL);
2667 if (ext4_pspace_cachep == NULL)
2671 kmem_cache_create("ext4_alloc_context",
2672 sizeof(struct ext4_allocation_context),
2673 0, SLAB_RECLAIM_ACCOUNT, NULL);
2674 if (ext4_ac_cachep == NULL) {
2675 kmem_cache_destroy(ext4_pspace_cachep);
2679 ext4_free_ext_cachep =
2680 kmem_cache_create("ext4_free_block_extents",
2681 sizeof(struct ext4_free_data),
2682 0, SLAB_RECLAIM_ACCOUNT, NULL);
2683 if (ext4_free_ext_cachep == NULL) {
2684 kmem_cache_destroy(ext4_pspace_cachep);
2685 kmem_cache_destroy(ext4_ac_cachep);
2688 ext4_create_debugfs_entry();
2692 void exit_ext4_mballoc(void)
2695 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2696 * before destroying the slab cache.
2699 kmem_cache_destroy(ext4_pspace_cachep);
2700 kmem_cache_destroy(ext4_ac_cachep);
2701 kmem_cache_destroy(ext4_free_ext_cachep);
2702 ext4_remove_debugfs_entry();
2707 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2708 * Returns 0 if success or error code
2710 static noinline_for_stack int
2711 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2712 handle_t *handle, unsigned int reserv_blks)
2714 struct buffer_head *bitmap_bh = NULL;
2715 struct ext4_group_desc *gdp;
2716 struct buffer_head *gdp_bh;
2717 struct ext4_sb_info *sbi;
2718 struct super_block *sb;
2722 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2723 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2729 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2733 err = ext4_journal_get_write_access(handle, bitmap_bh);
2738 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2742 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2743 ext4_free_blks_count(sb, gdp));
2745 err = ext4_journal_get_write_access(handle, gdp_bh);
2749 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2751 len = ac->ac_b_ex.fe_len;
2752 if (!ext4_data_block_valid(sbi, block, len)) {
2753 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2754 "fs metadata\n", block, block+len);
2755 /* File system mounted not to panic on error
2756 * Fix the bitmap and repeat the block allocation
2757 * We leak some of the blocks here.
2759 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2760 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2761 ac->ac_b_ex.fe_len);
2762 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2763 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2769 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2770 #ifdef AGGRESSIVE_CHECK
2773 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2774 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2775 bitmap_bh->b_data));
2779 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
2780 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2781 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2782 ext4_free_blks_set(sb, gdp,
2783 ext4_free_blocks_after_init(sb,
2784 ac->ac_b_ex.fe_group, gdp));
2786 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
2787 ext4_free_blks_set(sb, gdp, len);
2788 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2790 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2791 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
2793 * Now reduce the dirty block count also. Should not go negative
2795 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2796 /* release all the reserved blocks if non delalloc */
2797 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
2799 if (sbi->s_log_groups_per_flex) {
2800 ext4_group_t flex_group = ext4_flex_group(sbi,
2801 ac->ac_b_ex.fe_group);
2802 atomic_sub(ac->ac_b_ex.fe_len,
2803 &sbi->s_flex_groups[flex_group].free_blocks);
2806 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2809 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2812 ext4_mark_super_dirty(sb);
2818 * here we normalize request for locality group
2819 * Group request are normalized to s_strip size if we set the same via mount
2820 * option. If not we set it to s_mb_group_prealloc which can be configured via
2821 * /sys/fs/ext4/<partition>/mb_group_prealloc
2823 * XXX: should we try to preallocate more than the group has now?
2825 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2827 struct super_block *sb = ac->ac_sb;
2828 struct ext4_locality_group *lg = ac->ac_lg;
2831 if (EXT4_SB(sb)->s_stripe)
2832 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
2834 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2835 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2836 current->pid, ac->ac_g_ex.fe_len);
2840 * Normalization means making request better in terms of
2841 * size and alignment
2843 static noinline_for_stack void
2844 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2845 struct ext4_allocation_request *ar)
2849 loff_t size, orig_size, start_off;
2851 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2852 struct ext4_prealloc_space *pa;
2854 /* do normalize only data requests, metadata requests
2855 do not need preallocation */
2856 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2859 /* sometime caller may want exact blocks */
2860 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2863 /* caller may indicate that preallocation isn't
2864 * required (it's a tail, for example) */
2865 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2868 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2869 ext4_mb_normalize_group_request(ac);
2873 bsbits = ac->ac_sb->s_blocksize_bits;
2875 /* first, let's learn actual file size
2876 * given current request is allocated */
2877 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2878 size = size << bsbits;
2879 if (size < i_size_read(ac->ac_inode))
2880 size = i_size_read(ac->ac_inode);
2883 /* max size of free chunks */
2886 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2887 (req <= (size) || max <= (chunk_size))
2889 /* first, try to predict filesize */
2890 /* XXX: should this table be tunable? */
2892 if (size <= 16 * 1024) {
2894 } else if (size <= 32 * 1024) {
2896 } else if (size <= 64 * 1024) {
2898 } else if (size <= 128 * 1024) {
2900 } else if (size <= 256 * 1024) {
2902 } else if (size <= 512 * 1024) {
2904 } else if (size <= 1024 * 1024) {
2906 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2907 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2908 (21 - bsbits)) << 21;
2909 size = 2 * 1024 * 1024;
2910 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2911 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2912 (22 - bsbits)) << 22;
2913 size = 4 * 1024 * 1024;
2914 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2915 (8<<20)>>bsbits, max, 8 * 1024)) {
2916 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2917 (23 - bsbits)) << 23;
2918 size = 8 * 1024 * 1024;
2920 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2921 size = ac->ac_o_ex.fe_len << bsbits;
2923 size = size >> bsbits;
2924 start = start_off >> bsbits;
2926 /* don't cover already allocated blocks in selected range */
2927 if (ar->pleft && start <= ar->lleft) {
2928 size -= ar->lleft + 1 - start;
2929 start = ar->lleft + 1;
2931 if (ar->pright && start + size - 1 >= ar->lright)
2932 size -= start + size - ar->lright;
2936 /* check we don't cross already preallocated blocks */
2938 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2943 spin_lock(&pa->pa_lock);
2944 if (pa->pa_deleted) {
2945 spin_unlock(&pa->pa_lock);
2949 pa_end = pa->pa_lstart + pa->pa_len;
2951 /* PA must not overlap original request */
2952 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2953 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2955 /* skip PAs this normalized request doesn't overlap with */
2956 if (pa->pa_lstart >= end || pa_end <= start) {
2957 spin_unlock(&pa->pa_lock);
2960 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2962 /* adjust start or end to be adjacent to this pa */
2963 if (pa_end <= ac->ac_o_ex.fe_logical) {
2964 BUG_ON(pa_end < start);
2966 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
2967 BUG_ON(pa->pa_lstart > end);
2968 end = pa->pa_lstart;
2970 spin_unlock(&pa->pa_lock);
2975 /* XXX: extra loop to check we really don't overlap preallocations */
2977 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2979 spin_lock(&pa->pa_lock);
2980 if (pa->pa_deleted == 0) {
2981 pa_end = pa->pa_lstart + pa->pa_len;
2982 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
2984 spin_unlock(&pa->pa_lock);
2988 if (start + size <= ac->ac_o_ex.fe_logical &&
2989 start > ac->ac_o_ex.fe_logical) {
2990 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
2991 (unsigned long) start, (unsigned long) size,
2992 (unsigned long) ac->ac_o_ex.fe_logical);
2994 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
2995 start > ac->ac_o_ex.fe_logical);
2996 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
2998 /* now prepare goal request */
3000 /* XXX: is it better to align blocks WRT to logical
3001 * placement or satisfy big request as is */
3002 ac->ac_g_ex.fe_logical = start;
3003 ac->ac_g_ex.fe_len = size;
3005 /* define goal start in order to merge */
3006 if (ar->pright && (ar->lright == (start + size))) {
3007 /* merge to the right */
3008 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3009 &ac->ac_f_ex.fe_group,
3010 &ac->ac_f_ex.fe_start);
3011 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3013 if (ar->pleft && (ar->lleft + 1 == start)) {
3014 /* merge to the left */
3015 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3016 &ac->ac_f_ex.fe_group,
3017 &ac->ac_f_ex.fe_start);
3018 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3021 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3022 (unsigned) orig_size, (unsigned) start);
3025 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3027 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3029 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3030 atomic_inc(&sbi->s_bal_reqs);
3031 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3032 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3033 atomic_inc(&sbi->s_bal_success);
3034 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3035 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3036 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3037 atomic_inc(&sbi->s_bal_goals);
3038 if (ac->ac_found > sbi->s_mb_max_to_scan)
3039 atomic_inc(&sbi->s_bal_breaks);
3042 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3043 trace_ext4_mballoc_alloc(ac);
3045 trace_ext4_mballoc_prealloc(ac);
3049 * Called on failure; free up any blocks from the inode PA for this
3050 * context. We don't need this for MB_GROUP_PA because we only change
3051 * pa_free in ext4_mb_release_context(), but on failure, we've already
3052 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3054 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3056 struct ext4_prealloc_space *pa = ac->ac_pa;
3059 if (pa && pa->pa_type == MB_INODE_PA) {
3060 len = ac->ac_b_ex.fe_len;
3067 * use blocks preallocated to inode
3069 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3070 struct ext4_prealloc_space *pa)
3076 /* found preallocated blocks, use them */
3077 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3078 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3080 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3081 &ac->ac_b_ex.fe_start);
3082 ac->ac_b_ex.fe_len = len;
3083 ac->ac_status = AC_STATUS_FOUND;
3086 BUG_ON(start < pa->pa_pstart);
3087 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3088 BUG_ON(pa->pa_free < len);
3091 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3095 * use blocks preallocated to locality group
3097 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3098 struct ext4_prealloc_space *pa)
3100 unsigned int len = ac->ac_o_ex.fe_len;
3102 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3103 &ac->ac_b_ex.fe_group,
3104 &ac->ac_b_ex.fe_start);
3105 ac->ac_b_ex.fe_len = len;
3106 ac->ac_status = AC_STATUS_FOUND;
3109 /* we don't correct pa_pstart or pa_plen here to avoid
3110 * possible race when the group is being loaded concurrently
3111 * instead we correct pa later, after blocks are marked
3112 * in on-disk bitmap -- see ext4_mb_release_context()
3113 * Other CPUs are prevented from allocating from this pa by lg_mutex
3115 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3119 * Return the prealloc space that have minimal distance
3120 * from the goal block. @cpa is the prealloc
3121 * space that is having currently known minimal distance
3122 * from the goal block.
3124 static struct ext4_prealloc_space *
3125 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3126 struct ext4_prealloc_space *pa,
3127 struct ext4_prealloc_space *cpa)
3129 ext4_fsblk_t cur_distance, new_distance;
3132 atomic_inc(&pa->pa_count);
3135 cur_distance = abs(goal_block - cpa->pa_pstart);
3136 new_distance = abs(goal_block - pa->pa_pstart);
3138 if (cur_distance < new_distance)
3141 /* drop the previous reference */
3142 atomic_dec(&cpa->pa_count);
3143 atomic_inc(&pa->pa_count);
3148 * search goal blocks in preallocated space
3150 static noinline_for_stack int
3151 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3154 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3155 struct ext4_locality_group *lg;
3156 struct ext4_prealloc_space *pa, *cpa = NULL;
3157 ext4_fsblk_t goal_block;
3159 /* only data can be preallocated */
3160 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3163 /* first, try per-file preallocation */
3165 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3167 /* all fields in this condition don't change,
3168 * so we can skip locking for them */
3169 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3170 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3173 /* non-extent files can't have physical blocks past 2^32 */
3174 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3175 pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3178 /* found preallocated blocks, use them */
3179 spin_lock(&pa->pa_lock);
3180 if (pa->pa_deleted == 0 && pa->pa_free) {
3181 atomic_inc(&pa->pa_count);
3182 ext4_mb_use_inode_pa(ac, pa);
3183 spin_unlock(&pa->pa_lock);
3184 ac->ac_criteria = 10;
3188 spin_unlock(&pa->pa_lock);
3192 /* can we use group allocation? */
3193 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3196 /* inode may have no locality group for some reason */
3200 order = fls(ac->ac_o_ex.fe_len) - 1;
3201 if (order > PREALLOC_TB_SIZE - 1)
3202 /* The max size of hash table is PREALLOC_TB_SIZE */
3203 order = PREALLOC_TB_SIZE - 1;
3205 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3207 * search for the prealloc space that is having
3208 * minimal distance from the goal block.
3210 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3212 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3214 spin_lock(&pa->pa_lock);
3215 if (pa->pa_deleted == 0 &&
3216 pa->pa_free >= ac->ac_o_ex.fe_len) {
3218 cpa = ext4_mb_check_group_pa(goal_block,
3221 spin_unlock(&pa->pa_lock);
3226 ext4_mb_use_group_pa(ac, cpa);
3227 ac->ac_criteria = 20;
3234 * the function goes through all block freed in the group
3235 * but not yet committed and marks them used in in-core bitmap.
3236 * buddy must be generated from this bitmap
3237 * Need to be called with the ext4 group lock held
3239 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3243 struct ext4_group_info *grp;
3244 struct ext4_free_data *entry;
3246 grp = ext4_get_group_info(sb, group);
3247 n = rb_first(&(grp->bb_free_root));
3250 entry = rb_entry(n, struct ext4_free_data, node);
3251 mb_set_bits(bitmap, entry->start_blk, entry->count);
3258 * the function goes through all preallocation in this group and marks them
3259 * used in in-core bitmap. buddy must be generated from this bitmap
3260 * Need to be called with ext4 group lock held
3262 static noinline_for_stack
3263 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3266 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3267 struct ext4_prealloc_space *pa;
3268 struct list_head *cur;
3269 ext4_group_t groupnr;
3270 ext4_grpblk_t start;
3271 int preallocated = 0;
3275 /* all form of preallocation discards first load group,
3276 * so the only competing code is preallocation use.
3277 * we don't need any locking here
3278 * notice we do NOT ignore preallocations with pa_deleted
3279 * otherwise we could leave used blocks available for
3280 * allocation in buddy when concurrent ext4_mb_put_pa()
3281 * is dropping preallocation
3283 list_for_each(cur, &grp->bb_prealloc_list) {
3284 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3285 spin_lock(&pa->pa_lock);
3286 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3289 spin_unlock(&pa->pa_lock);
3290 if (unlikely(len == 0))
3292 BUG_ON(groupnr != group);
3293 mb_set_bits(bitmap, start, len);
3294 preallocated += len;
3297 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3300 static void ext4_mb_pa_callback(struct rcu_head *head)
3302 struct ext4_prealloc_space *pa;
3303 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3304 kmem_cache_free(ext4_pspace_cachep, pa);
3308 * drops a reference to preallocated space descriptor
3309 * if this was the last reference and the space is consumed
3311 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3312 struct super_block *sb, struct ext4_prealloc_space *pa)
3315 ext4_fsblk_t grp_blk;
3317 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3320 /* in this short window concurrent discard can set pa_deleted */
3321 spin_lock(&pa->pa_lock);
3322 if (pa->pa_deleted == 1) {
3323 spin_unlock(&pa->pa_lock);
3328 spin_unlock(&pa->pa_lock);
3330 grp_blk = pa->pa_pstart;
3332 * If doing group-based preallocation, pa_pstart may be in the
3333 * next group when pa is used up
3335 if (pa->pa_type == MB_GROUP_PA)
3338 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3343 * P1 (buddy init) P2 (regular allocation)
3344 * find block B in PA
3345 * copy on-disk bitmap to buddy
3346 * mark B in on-disk bitmap
3347 * drop PA from group
3348 * mark all PAs in buddy
3350 * thus, P1 initializes buddy with B available. to prevent this
3351 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3354 ext4_lock_group(sb, grp);
3355 list_del(&pa->pa_group_list);
3356 ext4_unlock_group(sb, grp);
3358 spin_lock(pa->pa_obj_lock);
3359 list_del_rcu(&pa->pa_inode_list);
3360 spin_unlock(pa->pa_obj_lock);
3362 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3366 * creates new preallocated space for given inode
3368 static noinline_for_stack int
3369 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3371 struct super_block *sb = ac->ac_sb;
3372 struct ext4_prealloc_space *pa;
3373 struct ext4_group_info *grp;
3374 struct ext4_inode_info *ei;
3376 /* preallocate only when found space is larger then requested */
3377 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3378 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3379 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3381 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3385 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3391 /* we can't allocate as much as normalizer wants.
3392 * so, found space must get proper lstart
3393 * to cover original request */
3394 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3395 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3397 /* we're limited by original request in that
3398 * logical block must be covered any way
3399 * winl is window we can move our chunk within */
3400 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3402 /* also, we should cover whole original request */
3403 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3405 /* the smallest one defines real window */
3406 win = min(winl, wins);
3408 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3409 if (offs && offs < win)
3412 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3413 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3414 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3417 /* preallocation can change ac_b_ex, thus we store actually
3418 * allocated blocks for history */
3419 ac->ac_f_ex = ac->ac_b_ex;
3421 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3422 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3423 pa->pa_len = ac->ac_b_ex.fe_len;
3424 pa->pa_free = pa->pa_len;
3425 atomic_set(&pa->pa_count, 1);
3426 spin_lock_init(&pa->pa_lock);
3427 INIT_LIST_HEAD(&pa->pa_inode_list);
3428 INIT_LIST_HEAD(&pa->pa_group_list);
3430 pa->pa_type = MB_INODE_PA;
3432 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3433 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3434 trace_ext4_mb_new_inode_pa(ac, pa);
3436 ext4_mb_use_inode_pa(ac, pa);
3437 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3439 ei = EXT4_I(ac->ac_inode);
3440 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3442 pa->pa_obj_lock = &ei->i_prealloc_lock;
3443 pa->pa_inode = ac->ac_inode;
3445 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3446 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3447 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3449 spin_lock(pa->pa_obj_lock);
3450 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3451 spin_unlock(pa->pa_obj_lock);
3457 * creates new preallocated space for locality group inodes belongs to
3459 static noinline_for_stack int
3460 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3462 struct super_block *sb = ac->ac_sb;
3463 struct ext4_locality_group *lg;
3464 struct ext4_prealloc_space *pa;
3465 struct ext4_group_info *grp;
3467 /* preallocate only when found space is larger then requested */
3468 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3469 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3470 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3472 BUG_ON(ext4_pspace_cachep == NULL);
3473 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3477 /* preallocation can change ac_b_ex, thus we store actually
3478 * allocated blocks for history */
3479 ac->ac_f_ex = ac->ac_b_ex;
3481 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3482 pa->pa_lstart = pa->pa_pstart;
3483 pa->pa_len = ac->ac_b_ex.fe_len;
3484 pa->pa_free = pa->pa_len;
3485 atomic_set(&pa->pa_count, 1);
3486 spin_lock_init(&pa->pa_lock);
3487 INIT_LIST_HEAD(&pa->pa_inode_list);
3488 INIT_LIST_HEAD(&pa->pa_group_list);
3490 pa->pa_type = MB_GROUP_PA;
3492 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3493 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3494 trace_ext4_mb_new_group_pa(ac, pa);
3496 ext4_mb_use_group_pa(ac, pa);
3497 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3499 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3503 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3504 pa->pa_inode = NULL;
3506 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3507 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3508 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3511 * We will later add the new pa to the right bucket
3512 * after updating the pa_free in ext4_mb_release_context
3517 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3521 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3522 err = ext4_mb_new_group_pa(ac);
3524 err = ext4_mb_new_inode_pa(ac);
3529 * finds all unused blocks in on-disk bitmap, frees them in
3530 * in-core bitmap and buddy.
3531 * @pa must be unlinked from inode and group lists, so that
3532 * nobody else can find/use it.
3533 * the caller MUST hold group/inode locks.
3534 * TODO: optimize the case when there are no in-core structures yet
3536 static noinline_for_stack int
3537 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3538 struct ext4_prealloc_space *pa,
3539 struct ext4_allocation_context *ac)
3541 struct super_block *sb = e4b->bd_sb;
3542 struct ext4_sb_info *sbi = EXT4_SB(sb);
3547 unsigned long long grp_blk_start;
3551 BUG_ON(pa->pa_deleted == 0);
3552 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3553 grp_blk_start = pa->pa_pstart - bit;
3554 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3555 end = bit + pa->pa_len;
3559 ac->ac_inode = pa->pa_inode;
3563 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3566 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3567 mb_debug(1, " free preallocated %u/%u in group %u\n",
3568 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3569 (unsigned) next - bit, (unsigned) group);
3573 ac->ac_b_ex.fe_group = group;
3574 ac->ac_b_ex.fe_start = bit;
3575 ac->ac_b_ex.fe_len = next - bit;
3576 ac->ac_b_ex.fe_logical = 0;
3577 trace_ext4_mballoc_discard(ac);
3580 trace_ext4_mb_release_inode_pa(sb, ac, pa, grp_blk_start + bit,
3582 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3585 if (free != pa->pa_free) {
3586 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3587 pa, (unsigned long) pa->pa_lstart,
3588 (unsigned long) pa->pa_pstart,
3589 (unsigned long) pa->pa_len);
3590 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3593 * pa is already deleted so we use the value obtained
3594 * from the bitmap and continue.
3597 atomic_add(free, &sbi->s_mb_discarded);
3602 static noinline_for_stack int
3603 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3604 struct ext4_prealloc_space *pa,
3605 struct ext4_allocation_context *ac)
3607 struct super_block *sb = e4b->bd_sb;
3611 trace_ext4_mb_release_group_pa(sb, ac, pa);
3612 BUG_ON(pa->pa_deleted == 0);
3613 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3614 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3615 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3616 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3620 ac->ac_inode = NULL;
3621 ac->ac_b_ex.fe_group = group;
3622 ac->ac_b_ex.fe_start = bit;
3623 ac->ac_b_ex.fe_len = pa->pa_len;
3624 ac->ac_b_ex.fe_logical = 0;
3625 trace_ext4_mballoc_discard(ac);
3632 * releases all preallocations in given group
3634 * first, we need to decide discard policy:
3635 * - when do we discard
3637 * - how many do we discard
3638 * 1) how many requested
3640 static noinline_for_stack int
3641 ext4_mb_discard_group_preallocations(struct super_block *sb,
3642 ext4_group_t group, int needed)
3644 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3645 struct buffer_head *bitmap_bh = NULL;
3646 struct ext4_prealloc_space *pa, *tmp;
3647 struct ext4_allocation_context *ac;
3648 struct list_head list;
3649 struct ext4_buddy e4b;
3654 mb_debug(1, "discard preallocation for group %u\n", group);
3656 if (list_empty(&grp->bb_prealloc_list))
3659 bitmap_bh = ext4_read_block_bitmap(sb, group);
3660 if (bitmap_bh == NULL) {
3661 ext4_error(sb, "Error reading block bitmap for %u", group);
3665 err = ext4_mb_load_buddy(sb, group, &e4b);
3667 ext4_error(sb, "Error loading buddy information for %u", group);
3673 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3675 INIT_LIST_HEAD(&list);
3676 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3680 ext4_lock_group(sb, group);
3681 list_for_each_entry_safe(pa, tmp,
3682 &grp->bb_prealloc_list, pa_group_list) {
3683 spin_lock(&pa->pa_lock);
3684 if (atomic_read(&pa->pa_count)) {
3685 spin_unlock(&pa->pa_lock);
3689 if (pa->pa_deleted) {
3690 spin_unlock(&pa->pa_lock);
3694 /* seems this one can be freed ... */
3697 /* we can trust pa_free ... */
3698 free += pa->pa_free;
3700 spin_unlock(&pa->pa_lock);
3702 list_del(&pa->pa_group_list);
3703 list_add(&pa->u.pa_tmp_list, &list);
3706 /* if we still need more blocks and some PAs were used, try again */
3707 if (free < needed && busy) {
3709 ext4_unlock_group(sb, group);
3711 * Yield the CPU here so that we don't get soft lockup
3712 * in non preempt case.
3718 /* found anything to free? */
3719 if (list_empty(&list)) {
3724 /* now free all selected PAs */
3725 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3727 /* remove from object (inode or locality group) */
3728 spin_lock(pa->pa_obj_lock);
3729 list_del_rcu(&pa->pa_inode_list);
3730 spin_unlock(pa->pa_obj_lock);
3732 if (pa->pa_type == MB_GROUP_PA)
3733 ext4_mb_release_group_pa(&e4b, pa, ac);
3735 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3737 list_del(&pa->u.pa_tmp_list);
3738 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3742 ext4_unlock_group(sb, group);
3744 kmem_cache_free(ext4_ac_cachep, ac);
3745 ext4_mb_unload_buddy(&e4b);
3751 * releases all non-used preallocated blocks for given inode
3753 * It's important to discard preallocations under i_data_sem
3754 * We don't want another block to be served from the prealloc
3755 * space when we are discarding the inode prealloc space.
3757 * FIXME!! Make sure it is valid at all the call sites
3759 void ext4_discard_preallocations(struct inode *inode)
3761 struct ext4_inode_info *ei = EXT4_I(inode);
3762 struct super_block *sb = inode->i_sb;
3763 struct buffer_head *bitmap_bh = NULL;
3764 struct ext4_prealloc_space *pa, *tmp;
3765 struct ext4_allocation_context *ac;
3766 ext4_group_t group = 0;
3767 struct list_head list;
3768 struct ext4_buddy e4b;
3771 if (!S_ISREG(inode->i_mode)) {
3772 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3776 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3777 trace_ext4_discard_preallocations(inode);
3779 INIT_LIST_HEAD(&list);
3781 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3784 ac->ac_inode = inode;
3787 /* first, collect all pa's in the inode */
3788 spin_lock(&ei->i_prealloc_lock);
3789 while (!list_empty(&ei->i_prealloc_list)) {
3790 pa = list_entry(ei->i_prealloc_list.next,
3791 struct ext4_prealloc_space, pa_inode_list);
3792 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3793 spin_lock(&pa->pa_lock);
3794 if (atomic_read(&pa->pa_count)) {
3795 /* this shouldn't happen often - nobody should
3796 * use preallocation while we're discarding it */
3797 spin_unlock(&pa->pa_lock);
3798 spin_unlock(&ei->i_prealloc_lock);
3799 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3801 schedule_timeout_uninterruptible(HZ);
3805 if (pa->pa_deleted == 0) {
3807 spin_unlock(&pa->pa_lock);
3808 list_del_rcu(&pa->pa_inode_list);
3809 list_add(&pa->u.pa_tmp_list, &list);
3813 /* someone is deleting pa right now */
3814 spin_unlock(&pa->pa_lock);
3815 spin_unlock(&ei->i_prealloc_lock);
3817 /* we have to wait here because pa_deleted
3818 * doesn't mean pa is already unlinked from
3819 * the list. as we might be called from
3820 * ->clear_inode() the inode will get freed
3821 * and concurrent thread which is unlinking
3822 * pa from inode's list may access already
3823 * freed memory, bad-bad-bad */
3825 /* XXX: if this happens too often, we can
3826 * add a flag to force wait only in case
3827 * of ->clear_inode(), but not in case of
3828 * regular truncate */
3829 schedule_timeout_uninterruptible(HZ);
3832 spin_unlock(&ei->i_prealloc_lock);
3834 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3835 BUG_ON(pa->pa_type != MB_INODE_PA);
3836 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3838 err = ext4_mb_load_buddy(sb, group, &e4b);
3840 ext4_error(sb, "Error loading buddy information for %u",
3845 bitmap_bh = ext4_read_block_bitmap(sb, group);
3846 if (bitmap_bh == NULL) {
3847 ext4_error(sb, "Error reading block bitmap for %u",
3849 ext4_mb_unload_buddy(&e4b);
3853 ext4_lock_group(sb, group);
3854 list_del(&pa->pa_group_list);
3855 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3856 ext4_unlock_group(sb, group);
3858 ext4_mb_unload_buddy(&e4b);
3861 list_del(&pa->u.pa_tmp_list);
3862 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3865 kmem_cache_free(ext4_ac_cachep, ac);
3869 * finds all preallocated spaces and return blocks being freed to them
3870 * if preallocated space becomes full (no block is used from the space)
3871 * then the function frees space in buddy
3872 * XXX: at the moment, truncate (which is the only way to free blocks)
3873 * discards all preallocations
3875 static void ext4_mb_return_to_preallocation(struct inode *inode,
3876 struct ext4_buddy *e4b,
3877 sector_t block, int count)
3879 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
3881 #ifdef CONFIG_EXT4_DEBUG
3882 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3884 struct super_block *sb = ac->ac_sb;
3885 ext4_group_t ngroups, i;
3887 if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
3890 printk(KERN_ERR "EXT4-fs: Can't allocate:"
3891 " Allocation context details:\n");
3892 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
3893 ac->ac_status, ac->ac_flags);
3894 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3895 "best %lu/%lu/%lu@%lu cr %d\n",
3896 (unsigned long)ac->ac_o_ex.fe_group,
3897 (unsigned long)ac->ac_o_ex.fe_start,
3898 (unsigned long)ac->ac_o_ex.fe_len,
3899 (unsigned long)ac->ac_o_ex.fe_logical,
3900 (unsigned long)ac->ac_g_ex.fe_group,
3901 (unsigned long)ac->ac_g_ex.fe_start,
3902 (unsigned long)ac->ac_g_ex.fe_len,
3903 (unsigned long)ac->ac_g_ex.fe_logical,
3904 (unsigned long)ac->ac_b_ex.fe_group,
3905 (unsigned long)ac->ac_b_ex.fe_start,
3906 (unsigned long)ac->ac_b_ex.fe_len,
3907 (unsigned long)ac->ac_b_ex.fe_logical,
3908 (int)ac->ac_criteria);
3909 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
3911 printk(KERN_ERR "EXT4-fs: groups: \n");
3912 ngroups = ext4_get_groups_count(sb);
3913 for (i = 0; i < ngroups; i++) {
3914 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3915 struct ext4_prealloc_space *pa;
3916 ext4_grpblk_t start;
3917 struct list_head *cur;
3918 ext4_lock_group(sb, i);
3919 list_for_each(cur, &grp->bb_prealloc_list) {
3920 pa = list_entry(cur, struct ext4_prealloc_space,
3922 spin_lock(&pa->pa_lock);
3923 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3925 spin_unlock(&pa->pa_lock);
3926 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3929 ext4_unlock_group(sb, i);
3931 if (grp->bb_free == 0)
3933 printk(KERN_ERR "%u: %d/%d \n",
3934 i, grp->bb_free, grp->bb_fragments);
3936 printk(KERN_ERR "\n");
3939 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3946 * We use locality group preallocation for small size file. The size of the
3947 * file is determined by the current size or the resulting size after
3948 * allocation which ever is larger
3950 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3952 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3954 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3955 int bsbits = ac->ac_sb->s_blocksize_bits;
3958 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3961 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3964 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3965 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3968 if ((size == isize) &&
3969 !ext4_fs_is_busy(sbi) &&
3970 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3971 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3975 /* don't use group allocation for large files */
3976 size = max(size, isize);
3977 if (size > sbi->s_mb_stream_request) {
3978 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3982 BUG_ON(ac->ac_lg != NULL);
3984 * locality group prealloc space are per cpu. The reason for having
3985 * per cpu locality group is to reduce the contention between block
3986 * request from multiple CPUs.
3988 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
3990 /* we're going to use group allocation */
3991 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3993 /* serialize all allocations in the group */
3994 mutex_lock(&ac->ac_lg->lg_mutex);
3997 static noinline_for_stack int
3998 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
3999 struct ext4_allocation_request *ar)
4001 struct super_block *sb = ar->inode->i_sb;
4002 struct ext4_sb_info *sbi = EXT4_SB(sb);
4003 struct ext4_super_block *es = sbi->s_es;
4007 ext4_grpblk_t block;
4009 /* we can't allocate > group size */
4012 /* just a dirty hack to filter too big requests */
4013 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4014 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4016 /* start searching from the goal */
4018 if (goal < le32_to_cpu(es->s_first_data_block) ||
4019 goal >= ext4_blocks_count(es))
4020 goal = le32_to_cpu(es->s_first_data_block);
4021 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4023 /* set up allocation goals */
4024 memset(ac, 0, sizeof(struct ext4_allocation_context));
4025 ac->ac_b_ex.fe_logical = ar->logical;
4026 ac->ac_status = AC_STATUS_CONTINUE;
4028 ac->ac_inode = ar->inode;
4029 ac->ac_o_ex.fe_logical = ar->logical;
4030 ac->ac_o_ex.fe_group = group;
4031 ac->ac_o_ex.fe_start = block;
4032 ac->ac_o_ex.fe_len = len;
4033 ac->ac_g_ex.fe_logical = ar->logical;
4034 ac->ac_g_ex.fe_group = group;
4035 ac->ac_g_ex.fe_start = block;
4036 ac->ac_g_ex.fe_len = len;
4037 ac->ac_flags = ar->flags;
4039 /* we have to define context: we'll we work with a file or
4040 * locality group. this is a policy, actually */
4041 ext4_mb_group_or_file(ac);
4043 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4044 "left: %u/%u, right %u/%u to %swritable\n",
4045 (unsigned) ar->len, (unsigned) ar->logical,
4046 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4047 (unsigned) ar->lleft, (unsigned) ar->pleft,
4048 (unsigned) ar->lright, (unsigned) ar->pright,
4049 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4054 static noinline_for_stack void
4055 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4056 struct ext4_locality_group *lg,
4057 int order, int total_entries)
4059 ext4_group_t group = 0;
4060 struct ext4_buddy e4b;
4061 struct list_head discard_list;
4062 struct ext4_prealloc_space *pa, *tmp;
4063 struct ext4_allocation_context *ac;
4065 mb_debug(1, "discard locality group preallocation\n");
4067 INIT_LIST_HEAD(&discard_list);
4068 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4072 spin_lock(&lg->lg_prealloc_lock);
4073 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4075 spin_lock(&pa->pa_lock);
4076 if (atomic_read(&pa->pa_count)) {
4078 * This is the pa that we just used
4079 * for block allocation. So don't
4082 spin_unlock(&pa->pa_lock);
4085 if (pa->pa_deleted) {
4086 spin_unlock(&pa->pa_lock);
4089 /* only lg prealloc space */
4090 BUG_ON(pa->pa_type != MB_GROUP_PA);
4092 /* seems this one can be freed ... */
4094 spin_unlock(&pa->pa_lock);
4096 list_del_rcu(&pa->pa_inode_list);
4097 list_add(&pa->u.pa_tmp_list, &discard_list);
4100 if (total_entries <= 5) {
4102 * we want to keep only 5 entries
4103 * allowing it to grow to 8. This
4104 * mak sure we don't call discard
4105 * soon for this list.
4110 spin_unlock(&lg->lg_prealloc_lock);
4112 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4114 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4115 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4116 ext4_error(sb, "Error loading buddy information for %u",
4120 ext4_lock_group(sb, group);
4121 list_del(&pa->pa_group_list);
4122 ext4_mb_release_group_pa(&e4b, pa, ac);
4123 ext4_unlock_group(sb, group);
4125 ext4_mb_unload_buddy(&e4b);
4126 list_del(&pa->u.pa_tmp_list);
4127 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4130 kmem_cache_free(ext4_ac_cachep, ac);
4134 * We have incremented pa_count. So it cannot be freed at this
4135 * point. Also we hold lg_mutex. So no parallel allocation is
4136 * possible from this lg. That means pa_free cannot be updated.
4138 * A parallel ext4_mb_discard_group_preallocations is possible.
4139 * which can cause the lg_prealloc_list to be updated.
4142 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4144 int order, added = 0, lg_prealloc_count = 1;
4145 struct super_block *sb = ac->ac_sb;
4146 struct ext4_locality_group *lg = ac->ac_lg;
4147 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4149 order = fls(pa->pa_free) - 1;
4150 if (order > PREALLOC_TB_SIZE - 1)
4151 /* The max size of hash table is PREALLOC_TB_SIZE */
4152 order = PREALLOC_TB_SIZE - 1;
4153 /* Add the prealloc space to lg */
4155 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4157 spin_lock(&tmp_pa->pa_lock);
4158 if (tmp_pa->pa_deleted) {
4159 spin_unlock(&tmp_pa->pa_lock);
4162 if (!added && pa->pa_free < tmp_pa->pa_free) {
4163 /* Add to the tail of the previous entry */
4164 list_add_tail_rcu(&pa->pa_inode_list,
4165 &tmp_pa->pa_inode_list);
4168 * we want to count the total
4169 * number of entries in the list
4172 spin_unlock(&tmp_pa->pa_lock);
4173 lg_prealloc_count++;
4176 list_add_tail_rcu(&pa->pa_inode_list,
4177 &lg->lg_prealloc_list[order]);
4180 /* Now trim the list to be not more than 8 elements */
4181 if (lg_prealloc_count > 8) {
4182 ext4_mb_discard_lg_preallocations(sb, lg,
4183 order, lg_prealloc_count);
4190 * release all resource we used in allocation
4192 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4194 struct ext4_prealloc_space *pa = ac->ac_pa;
4196 if (pa->pa_type == MB_GROUP_PA) {
4197 /* see comment in ext4_mb_use_group_pa() */
4198 spin_lock(&pa->pa_lock);
4199 pa->pa_pstart += ac->ac_b_ex.fe_len;
4200 pa->pa_lstart += ac->ac_b_ex.fe_len;
4201 pa->pa_free -= ac->ac_b_ex.fe_len;
4202 pa->pa_len -= ac->ac_b_ex.fe_len;
4203 spin_unlock(&pa->pa_lock);
4207 up_read(ac->alloc_semp);
4210 * We want to add the pa to the right bucket.
4211 * Remove it from the list and while adding
4212 * make sure the list to which we are adding
4213 * doesn't grow big. We need to release
4214 * alloc_semp before calling ext4_mb_add_n_trim()
4216 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4217 spin_lock(pa->pa_obj_lock);
4218 list_del_rcu(&pa->pa_inode_list);
4219 spin_unlock(pa->pa_obj_lock);
4220 ext4_mb_add_n_trim(ac);
4222 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4224 if (ac->ac_bitmap_page)
4225 page_cache_release(ac->ac_bitmap_page);
4226 if (ac->ac_buddy_page)
4227 page_cache_release(ac->ac_buddy_page);
4228 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4229 mutex_unlock(&ac->ac_lg->lg_mutex);
4230 ext4_mb_collect_stats(ac);
4234 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4236 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4240 trace_ext4_mb_discard_preallocations(sb, needed);
4241 for (i = 0; i < ngroups && needed > 0; i++) {
4242 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4251 * Main entry point into mballoc to allocate blocks
4252 * it tries to use preallocation first, then falls back
4253 * to usual allocation
4255 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4256 struct ext4_allocation_request *ar, int *errp)
4259 struct ext4_allocation_context *ac = NULL;
4260 struct ext4_sb_info *sbi;
4261 struct super_block *sb;
4262 ext4_fsblk_t block = 0;
4263 unsigned int inquota = 0;
4264 unsigned int reserv_blks = 0;
4266 sb = ar->inode->i_sb;
4269 trace_ext4_request_blocks(ar);
4272 * For delayed allocation, we could skip the ENOSPC and
4273 * EDQUOT check, as blocks and quotas have been already
4274 * reserved when data being copied into pagecache.
4276 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4277 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4279 /* Without delayed allocation we need to verify
4280 * there is enough free blocks to do block allocation
4281 * and verify allocation doesn't exceed the quota limits.
4283 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4284 /* let others to free the space */
4286 ar->len = ar->len >> 1;
4292 reserv_blks = ar->len;
4293 while (ar->len && dquot_alloc_block(ar->inode, ar->len)) {
4294 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4304 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4311 *errp = ext4_mb_initialize_context(ac, ar);
4317 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4318 if (!ext4_mb_use_preallocated(ac)) {
4319 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4320 ext4_mb_normalize_request(ac, ar);
4322 /* allocate space in core */
4323 *errp = ext4_mb_regular_allocator(ac);
4327 /* as we've just preallocated more space than
4328 * user requested orinally, we store allocated
4329 * space in a special descriptor */
4330 if (ac->ac_status == AC_STATUS_FOUND &&
4331 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4332 ext4_mb_new_preallocation(ac);
4334 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4335 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4336 if (*errp == -EAGAIN) {
4338 * drop the reference that we took
4339 * in ext4_mb_use_best_found
4341 ext4_mb_release_context(ac);
4342 ac->ac_b_ex.fe_group = 0;
4343 ac->ac_b_ex.fe_start = 0;
4344 ac->ac_b_ex.fe_len = 0;
4345 ac->ac_status = AC_STATUS_CONTINUE;
4349 ext4_discard_allocated_blocks(ac);
4351 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4352 ar->len = ac->ac_b_ex.fe_len;
4355 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4362 ac->ac_b_ex.fe_len = 0;
4364 ext4_mb_show_ac(ac);
4366 ext4_mb_release_context(ac);
4369 kmem_cache_free(ext4_ac_cachep, ac);
4370 if (inquota && ar->len < inquota)
4371 dquot_free_block(ar->inode, inquota - ar->len);
4373 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4374 /* release all the reserved blocks if non delalloc */
4375 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4379 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4385 * We can merge two free data extents only if the physical blocks
4386 * are contiguous, AND the extents were freed by the same transaction,
4387 * AND the blocks are associated with the same group.
4389 static int can_merge(struct ext4_free_data *entry1,
4390 struct ext4_free_data *entry2)
4392 if ((entry1->t_tid == entry2->t_tid) &&
4393 (entry1->group == entry2->group) &&
4394 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4399 static noinline_for_stack int
4400 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4401 struct ext4_free_data *new_entry)
4403 ext4_group_t group = e4b->bd_group;
4404 ext4_grpblk_t block;
4405 struct ext4_free_data *entry;
4406 struct ext4_group_info *db = e4b->bd_info;
4407 struct super_block *sb = e4b->bd_sb;
4408 struct ext4_sb_info *sbi = EXT4_SB(sb);
4409 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4410 struct rb_node *parent = NULL, *new_node;
4412 BUG_ON(!ext4_handle_valid(handle));
4413 BUG_ON(e4b->bd_bitmap_page == NULL);
4414 BUG_ON(e4b->bd_buddy_page == NULL);
4416 new_node = &new_entry->node;
4417 block = new_entry->start_blk;
4420 /* first free block exent. We need to
4421 protect buddy cache from being freed,
4422 * otherwise we'll refresh it from
4423 * on-disk bitmap and lose not-yet-available
4425 page_cache_get(e4b->bd_buddy_page);
4426 page_cache_get(e4b->bd_bitmap_page);
4430 entry = rb_entry(parent, struct ext4_free_data, node);
4431 if (block < entry->start_blk)
4433 else if (block >= (entry->start_blk + entry->count))
4434 n = &(*n)->rb_right;
4436 ext4_grp_locked_error(sb, group, 0,
4437 ext4_group_first_block_no(sb, group) + block,
4438 "Block already on to-be-freed list");
4443 rb_link_node(new_node, parent, n);
4444 rb_insert_color(new_node, &db->bb_free_root);
4446 /* Now try to see the extent can be merged to left and right */
4447 node = rb_prev(new_node);
4449 entry = rb_entry(node, struct ext4_free_data, node);
4450 if (can_merge(entry, new_entry)) {
4451 new_entry->start_blk = entry->start_blk;
4452 new_entry->count += entry->count;
4453 rb_erase(node, &(db->bb_free_root));
4454 spin_lock(&sbi->s_md_lock);
4455 list_del(&entry->list);
4456 spin_unlock(&sbi->s_md_lock);
4457 kmem_cache_free(ext4_free_ext_cachep, entry);
4461 node = rb_next(new_node);
4463 entry = rb_entry(node, struct ext4_free_data, node);
4464 if (can_merge(new_entry, entry)) {
4465 new_entry->count += entry->count;
4466 rb_erase(node, &(db->bb_free_root));
4467 spin_lock(&sbi->s_md_lock);
4468 list_del(&entry->list);
4469 spin_unlock(&sbi->s_md_lock);
4470 kmem_cache_free(ext4_free_ext_cachep, entry);
4473 /* Add the extent to transaction's private list */
4474 spin_lock(&sbi->s_md_lock);
4475 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4476 spin_unlock(&sbi->s_md_lock);
4481 * ext4_free_blocks() -- Free given blocks and update quota
4482 * @handle: handle for this transaction
4484 * @block: start physical block to free
4485 * @count: number of blocks to count
4486 * @metadata: Are these metadata blocks
4488 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4489 struct buffer_head *bh, ext4_fsblk_t block,
4490 unsigned long count, int flags)
4492 struct buffer_head *bitmap_bh = NULL;
4493 struct super_block *sb = inode->i_sb;
4494 struct ext4_allocation_context *ac = NULL;
4495 struct ext4_group_desc *gdp;
4496 unsigned long freed = 0;
4497 unsigned int overflow;
4499 struct buffer_head *gd_bh;
4500 ext4_group_t block_group;
4501 struct ext4_sb_info *sbi;
4502 struct ext4_buddy e4b;
4508 BUG_ON(block != bh->b_blocknr);
4510 block = bh->b_blocknr;
4514 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4515 !ext4_data_block_valid(sbi, block, count)) {
4516 ext4_error(sb, "Freeing blocks not in datazone - "
4517 "block = %llu, count = %lu", block, count);
4521 ext4_debug("freeing block %llu\n", block);
4522 trace_ext4_free_blocks(inode, block, count, flags);
4524 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4525 struct buffer_head *tbh = bh;
4528 BUG_ON(bh && (count > 1));
4530 for (i = 0; i < count; i++) {
4532 tbh = sb_find_get_block(inode->i_sb,
4534 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4535 inode, tbh, block + i);
4540 * We need to make sure we don't reuse the freed block until
4541 * after the transaction is committed, which we can do by
4542 * treating the block as metadata, below. We make an
4543 * exception if the inode is to be written in writeback mode
4544 * since writeback mode has weak data consistency guarantees.
4546 if (!ext4_should_writeback_data(inode))
4547 flags |= EXT4_FREE_BLOCKS_METADATA;
4549 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4551 ac->ac_inode = inode;
4557 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4560 * Check to see if we are freeing blocks across a group
4563 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4564 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4567 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4572 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4578 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4579 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4580 in_range(block, ext4_inode_table(sb, gdp),
4581 EXT4_SB(sb)->s_itb_per_group) ||
4582 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4583 EXT4_SB(sb)->s_itb_per_group)) {
4585 ext4_error(sb, "Freeing blocks in system zone - "
4586 "Block = %llu, count = %lu", block, count);
4587 /* err = 0. ext4_std_error should be a no op */
4591 BUFFER_TRACE(bitmap_bh, "getting write access");
4592 err = ext4_journal_get_write_access(handle, bitmap_bh);
4597 * We are about to modify some metadata. Call the journal APIs
4598 * to unshare ->b_data if a currently-committing transaction is
4601 BUFFER_TRACE(gd_bh, "get_write_access");
4602 err = ext4_journal_get_write_access(handle, gd_bh);
4605 #ifdef AGGRESSIVE_CHECK
4608 for (i = 0; i < count; i++)
4609 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4613 ac->ac_b_ex.fe_group = block_group;
4614 ac->ac_b_ex.fe_start = bit;
4615 ac->ac_b_ex.fe_len = count;
4616 trace_ext4_mballoc_free(ac);
4619 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4623 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4624 struct ext4_free_data *new_entry;
4626 * blocks being freed are metadata. these blocks shouldn't
4627 * be used until this transaction is committed
4629 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4630 new_entry->start_blk = bit;
4631 new_entry->group = block_group;
4632 new_entry->count = count;
4633 new_entry->t_tid = handle->h_transaction->t_tid;
4635 ext4_lock_group(sb, block_group);
4636 mb_clear_bits(bitmap_bh->b_data, bit, count);
4637 ext4_mb_free_metadata(handle, &e4b, new_entry);
4639 /* need to update group_info->bb_free and bitmap
4640 * with group lock held. generate_buddy look at
4641 * them with group lock_held
4643 ext4_lock_group(sb, block_group);
4644 mb_clear_bits(bitmap_bh->b_data, bit, count);
4645 mb_free_blocks(inode, &e4b, bit, count);
4646 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4647 if (test_opt(sb, DISCARD))
4648 ext4_issue_discard(sb, block_group, bit, count);
4651 ret = ext4_free_blks_count(sb, gdp) + count;
4652 ext4_free_blks_set(sb, gdp, ret);
4653 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4654 ext4_unlock_group(sb, block_group);
4655 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4657 if (sbi->s_log_groups_per_flex) {
4658 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4659 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4662 ext4_mb_unload_buddy(&e4b);
4666 /* We dirtied the bitmap block */
4667 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4668 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4670 /* And the group descriptor block */
4671 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4672 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4676 if (overflow && !err) {
4682 ext4_mark_super_dirty(sb);
4685 dquot_free_block(inode, freed);
4687 ext4_std_error(sb, err);
4689 kmem_cache_free(ext4_ac_cachep, ac);