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
27 * - test ext4_ext_search_left() and ext4_ext_search_right()
28 * - search for metadata in few groups
31 * - normalization should take into account whether file is still open
32 * - discard preallocations if no free space left (policy?)
33 * - don't normalize tails
35 * - reservation for superuser
38 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
39 * - track min/max extents in each group for better group selection
40 * - mb_mark_used() may allocate chunk right after splitting buddy
41 * - tree of groups sorted by number of free blocks
46 * The allocation request involve request for multiple number of blocks
47 * near to the goal(block) value specified.
49 * During initialization phase of the allocator we decide to use the group
50 * preallocation or inode preallocation depending on the size file. The
51 * size of the file could be the resulting file size we would have after
52 * allocation or the current file size which ever is larger. If the size is
53 * less that sbi->s_mb_stream_request we select the group
54 * preallocation. The default value of s_mb_stream_request is 16
55 * blocks. This can also be tuned via
56 * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms
57 * of number of blocks.
59 * The main motivation for having small file use group preallocation is to
60 * ensure that we have small file closer in the disk.
62 * First stage the allocator looks at the inode prealloc list
63 * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for
64 * this particular inode. The inode prealloc space is represented as:
66 * pa_lstart -> the logical start block for this prealloc space
67 * pa_pstart -> the physical start block for this prealloc space
68 * pa_len -> lenght for this prealloc space
69 * pa_free -> free space available in this prealloc space
71 * The inode preallocation space is used looking at the _logical_ start
72 * block. If only the logical file block falls within the range of prealloc
73 * space we will consume the particular prealloc space. This make sure that
74 * that the we have contiguous physical blocks representing the file blocks
76 * The important thing to be noted in case of inode prealloc space is that
77 * we don't modify the values associated to inode prealloc space except
80 * If we are not able to find blocks in the inode prealloc space and if we
81 * have the group allocation flag set then we look at the locality group
82 * prealloc space. These are per CPU prealloc list repreasented as
84 * ext4_sb_info.s_locality_groups[smp_processor_id()]
86 * The reason for having a per cpu locality group is to reduce the contention
87 * between CPUs. It is possible to get scheduled at this point.
89 * The locality group prealloc space is used looking at whether we have
90 * enough free space (pa_free) withing the prealloc space.
92 * If we can't allocate blocks via inode prealloc or/and locality group
93 * prealloc then we look at the buddy cache. The buddy cache is represented
94 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
95 * mapped to the buddy and bitmap information regarding different
96 * groups. The buddy information is attached to buddy cache inode so that
97 * we can access them through the page cache. The information regarding
98 * each group is loaded via ext4_mb_load_buddy. The information involve
99 * block bitmap and buddy information. The information are stored in the
103 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
106 * one block each for bitmap and buddy information. So for each group we
107 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
108 * blocksize) blocks. So it can have information regarding groups_per_page
109 * which is blocks_per_page/2
111 * The buddy cache inode is not stored on disk. The inode is thrown
112 * away when the filesystem is unmounted.
114 * We look for count number of blocks in the buddy cache. If we were able
115 * to locate that many free blocks we return with additional information
116 * regarding rest of the contiguous physical block available
118 * Before allocating blocks via buddy cache we normalize the request
119 * blocks. This ensure we ask for more blocks that we needed. The extra
120 * blocks that we get after allocation is added to the respective prealloc
121 * list. In case of inode preallocation we follow a list of heuristics
122 * based on file size. This can be found in ext4_mb_normalize_request. If
123 * we are doing a group prealloc we try to normalize the request to
124 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is set to
125 * 512 blocks. This can be tuned via
126 * /proc/fs/ext4/<partition/group_prealloc. The value is represented in
127 * terms of number of blocks. If we have mounted the file system with -O
128 * stripe=<value> option the group prealloc request is normalized to the
129 * stripe value (sbi->s_stripe)
131 * The regular allocator(using the buddy cache) support few tunables.
133 * /proc/fs/ext4/<partition>/min_to_scan
134 * /proc/fs/ext4/<partition>/max_to_scan
135 * /proc/fs/ext4/<partition>/order2_req
137 * The regular allocator use buddy scan only if the request len is power of
138 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
139 * value of s_mb_order2_reqs can be tuned via
140 * /proc/fs/ext4/<partition>/order2_req. If the request len is equal to
141 * stripe size (sbi->s_stripe), we try to search for contigous block in
142 * stripe size. This should result in better allocation on RAID setup. If
143 * not we search in the specific group using bitmap for best extents. The
144 * tunable min_to_scan and max_to_scan controll the behaviour here.
145 * min_to_scan indicate how long the mballoc __must__ look for a best
146 * extent and max_to_scanindicate how long the mballoc __can__ look for a
147 * best extent in the found extents. Searching for the blocks starts with
148 * the group specified as the goal value in allocation context via
149 * ac_g_ex. Each group is first checked based on the criteria whether it
150 * can used for allocation. ext4_mb_good_group explains how the groups are
153 * Both the prealloc space are getting populated as above. So for the first
154 * request we will hit the buddy cache which will result in this prealloc
155 * space getting filled. The prealloc space is then later used for the
156 * subsequent request.
160 * mballoc operates on the following data:
162 * - in-core buddy (actually includes buddy and bitmap)
163 * - preallocation descriptors (PAs)
165 * there are two types of preallocations:
167 * assiged to specific inode and can be used for this inode only.
168 * it describes part of inode's space preallocated to specific
169 * physical blocks. any block from that preallocated can be used
170 * independent. the descriptor just tracks number of blocks left
171 * unused. so, before taking some block from descriptor, one must
172 * make sure corresponded logical block isn't allocated yet. this
173 * also means that freeing any block within descriptor's range
174 * must discard all preallocated blocks.
176 * assigned to specific locality group which does not translate to
177 * permanent set of inodes: inode can join and leave group. space
178 * from this type of preallocation can be used for any inode. thus
179 * it's consumed from the beginning to the end.
181 * relation between them can be expressed as:
182 * in-core buddy = on-disk bitmap + preallocation descriptors
184 * this mean blocks mballoc considers used are:
185 * - allocated blocks (persistent)
186 * - preallocated blocks (non-persistent)
188 * consistency in mballoc world means that at any time a block is either
189 * free or used in ALL structures. notice: "any time" should not be read
190 * literally -- time is discrete and delimited by locks.
192 * to keep it simple, we don't use block numbers, instead we count number of
193 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
195 * all operations can be expressed as:
196 * - init buddy: buddy = on-disk + PAs
197 * - new PA: buddy += N; PA = N
198 * - use inode PA: on-disk += N; PA -= N
199 * - discard inode PA buddy -= on-disk - PA; PA = 0
200 * - use locality group PA on-disk += N; PA -= N
201 * - discard locality group PA buddy -= PA; PA = 0
202 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
203 * is used in real operation because we can't know actual used
204 * bits from PA, only from on-disk bitmap
206 * if we follow this strict logic, then all operations above should be atomic.
207 * given some of them can block, we'd have to use something like semaphores
208 * killing performance on high-end SMP hardware. let's try to relax it using
209 * the following knowledge:
210 * 1) if buddy is referenced, it's already initialized
211 * 2) while block is used in buddy and the buddy is referenced,
212 * nobody can re-allocate that block
213 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
214 * bit set and PA claims same block, it's OK. IOW, one can set bit in
215 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
218 * so, now we're building a concurrency table:
221 * blocks for PA are allocated in the buddy, buddy must be referenced
222 * until PA is linked to allocation group to avoid concurrent buddy init
224 * we need to make sure that either on-disk bitmap or PA has uptodate data
225 * given (3) we care that PA-=N operation doesn't interfere with init
227 * the simplest way would be to have buddy initialized by the discard
228 * - use locality group PA
229 * again PA-=N must be serialized with init
230 * - discard locality group PA
231 * the simplest way would be to have buddy initialized by the discard
234 * i_data_sem serializes them
236 * discard process must wait until PA isn't used by another process
237 * - use locality group PA
238 * some mutex should serialize them
239 * - discard locality group PA
240 * discard process must wait until PA isn't used by another process
243 * i_data_sem or another mutex should serializes them
245 * discard process must wait until PA isn't used by another process
246 * - use locality group PA
247 * nothing wrong here -- they're different PAs covering different blocks
248 * - discard locality group PA
249 * discard process must wait until PA isn't used by another process
251 * now we're ready to make few consequences:
252 * - PA is referenced and while it is no discard is possible
253 * - PA is referenced until block isn't marked in on-disk bitmap
254 * - PA changes only after on-disk bitmap
255 * - discard must not compete with init. either init is done before
256 * any discard or they're serialized somehow
257 * - buddy init as sum of on-disk bitmap and PAs is done atomically
259 * a special case when we've used PA to emptiness. no need to modify buddy
260 * in this case, but we should care about concurrent init
265 * Logic in few words:
270 * mark bits in on-disk bitmap
273 * - use preallocation:
274 * find proper PA (per-inode or group)
276 * mark bits in on-disk bitmap
282 * mark bits in on-disk bitmap
285 * - discard preallocations in group:
287 * move them onto local list
288 * load on-disk bitmap
290 * remove PA from object (inode or locality group)
291 * mark free blocks in-core
293 * - discard inode's preallocations:
300 * - bitlock on a group (group)
301 * - object (inode/locality) (object)
312 * - release consumed pa:
317 * - generate in-core bitmap:
321 * - discard all for given object (inode, locality group):
326 * - discard all for given group:
334 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
336 #if BITS_PER_LONG == 64
337 *bit += ((unsigned long) addr & 7UL) << 3;
338 addr = (void *) ((unsigned long) addr & ~7UL);
339 #elif BITS_PER_LONG == 32
340 *bit += ((unsigned long) addr & 3UL) << 3;
341 addr = (void *) ((unsigned long) addr & ~3UL);
343 #error "how many bits you are?!"
348 static inline int mb_test_bit(int bit, void *addr)
351 * ext4_test_bit on architecture like powerpc
352 * needs unsigned long aligned address
354 addr = mb_correct_addr_and_bit(&bit, addr);
355 return ext4_test_bit(bit, addr);
358 static inline void mb_set_bit(int bit, void *addr)
360 addr = mb_correct_addr_and_bit(&bit, addr);
361 ext4_set_bit(bit, addr);
364 static inline void mb_set_bit_atomic(spinlock_t *lock, int bit, void *addr)
366 addr = mb_correct_addr_and_bit(&bit, addr);
367 ext4_set_bit_atomic(lock, bit, addr);
370 static inline void mb_clear_bit(int bit, void *addr)
372 addr = mb_correct_addr_and_bit(&bit, addr);
373 ext4_clear_bit(bit, addr);
376 static inline void mb_clear_bit_atomic(spinlock_t *lock, int bit, void *addr)
378 addr = mb_correct_addr_and_bit(&bit, addr);
379 ext4_clear_bit_atomic(lock, bit, addr);
382 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
384 int fix = 0, ret, tmpmax;
385 addr = mb_correct_addr_and_bit(&fix, addr);
389 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
395 static inline int mb_find_next_bit(void *addr, int max, int start)
397 int fix = 0, ret, tmpmax;
398 addr = mb_correct_addr_and_bit(&fix, addr);
402 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
408 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
412 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
415 if (order > e4b->bd_blkbits + 1) {
420 /* at order 0 we see each particular block */
421 *max = 1 << (e4b->bd_blkbits + 3);
423 return EXT4_MB_BITMAP(e4b);
425 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
426 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
432 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
433 int first, int count)
436 struct super_block *sb = e4b->bd_sb;
438 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
440 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
441 for (i = 0; i < count; i++) {
442 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
443 ext4_fsblk_t blocknr;
444 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
445 blocknr += first + i;
447 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
449 ext4_error(sb, __func__, "double-free of inode"
450 " %lu's block %llu(bit %u in group %lu)\n",
451 inode ? inode->i_ino : 0, blocknr,
452 first + i, e4b->bd_group);
454 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
458 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
462 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
464 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
465 for (i = 0; i < count; i++) {
466 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
467 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
471 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
473 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
474 unsigned char *b1, *b2;
476 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
477 b2 = (unsigned char *) bitmap;
478 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
479 if (b1[i] != b2[i]) {
480 printk(KERN_ERR "corruption in group %lu "
481 "at byte %u(%u): %x in copy != %x "
482 "on disk/prealloc\n",
483 e4b->bd_group, i, i * 8, b1[i], b2[i]);
491 static inline void mb_free_blocks_double(struct inode *inode,
492 struct ext4_buddy *e4b, int first, int count)
496 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
497 int first, int count)
501 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
507 #ifdef AGGRESSIVE_CHECK
509 #define MB_CHECK_ASSERT(assert) \
513 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
514 function, file, line, # assert); \
519 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
520 const char *function, int line)
522 struct super_block *sb = e4b->bd_sb;
523 int order = e4b->bd_blkbits + 1;
530 struct ext4_group_info *grp;
533 struct list_head *cur;
538 static int mb_check_counter;
539 if (mb_check_counter++ % 100 != 0)
544 buddy = mb_find_buddy(e4b, order, &max);
545 MB_CHECK_ASSERT(buddy);
546 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
547 MB_CHECK_ASSERT(buddy2);
548 MB_CHECK_ASSERT(buddy != buddy2);
549 MB_CHECK_ASSERT(max * 2 == max2);
552 for (i = 0; i < max; i++) {
554 if (mb_test_bit(i, buddy)) {
555 /* only single bit in buddy2 may be 1 */
556 if (!mb_test_bit(i << 1, buddy2)) {
558 mb_test_bit((i<<1)+1, buddy2));
559 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
561 mb_test_bit(i << 1, buddy2));
566 /* both bits in buddy2 must be 0 */
567 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
568 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
570 for (j = 0; j < (1 << order); j++) {
571 k = (i * (1 << order)) + j;
573 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
577 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
582 buddy = mb_find_buddy(e4b, 0, &max);
583 for (i = 0; i < max; i++) {
584 if (!mb_test_bit(i, buddy)) {
585 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
593 /* check used bits only */
594 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
595 buddy2 = mb_find_buddy(e4b, j, &max2);
597 MB_CHECK_ASSERT(k < max2);
598 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
601 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
602 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
604 grp = ext4_get_group_info(sb, e4b->bd_group);
605 buddy = mb_find_buddy(e4b, 0, &max);
606 list_for_each(cur, &grp->bb_prealloc_list) {
607 ext4_group_t groupnr;
608 struct ext4_prealloc_space *pa;
609 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
610 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
611 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
612 for (i = 0; i < pa->pa_len; i++)
613 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
617 #undef MB_CHECK_ASSERT
618 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
619 __FILE__, __func__, __LINE__)
621 #define mb_check_buddy(e4b)
624 /* FIXME!! need more doc */
625 static void ext4_mb_mark_free_simple(struct super_block *sb,
626 void *buddy, unsigned first, int len,
627 struct ext4_group_info *grp)
629 struct ext4_sb_info *sbi = EXT4_SB(sb);
632 unsigned short chunk;
633 unsigned short border;
635 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
637 border = 2 << sb->s_blocksize_bits;
640 /* find how many blocks can be covered since this position */
641 max = ffs(first | border) - 1;
643 /* find how many blocks of power 2 we need to mark */
650 /* mark multiblock chunks only */
651 grp->bb_counters[min]++;
653 mb_clear_bit(first >> min,
654 buddy + sbi->s_mb_offsets[min]);
661 static void ext4_mb_generate_buddy(struct super_block *sb,
662 void *buddy, void *bitmap, ext4_group_t group)
664 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
665 unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
666 unsigned short i = 0;
667 unsigned short first;
670 unsigned fragments = 0;
671 unsigned long long period = get_cycles();
673 /* initialize buddy from bitmap which is aggregation
674 * of on-disk bitmap and preallocations */
675 i = mb_find_next_zero_bit(bitmap, max, 0);
676 grp->bb_first_free = i;
680 i = mb_find_next_bit(bitmap, max, i);
684 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
686 grp->bb_counters[0]++;
688 i = mb_find_next_zero_bit(bitmap, max, i);
690 grp->bb_fragments = fragments;
692 if (free != grp->bb_free) {
693 ext4_error(sb, __func__,
694 "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
695 group, free, grp->bb_free);
697 * If we intent to continue, we consider group descritor
698 * corrupt and update bb_free using bitmap value
703 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
705 period = get_cycles() - period;
706 spin_lock(&EXT4_SB(sb)->s_bal_lock);
707 EXT4_SB(sb)->s_mb_buddies_generated++;
708 EXT4_SB(sb)->s_mb_generation_time += period;
709 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
712 /* The buddy information is attached the buddy cache inode
713 * for convenience. The information regarding each group
714 * is loaded via ext4_mb_load_buddy. The information involve
715 * block bitmap and buddy information. The information are
716 * stored in the inode as
719 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
722 * one block each for bitmap and buddy information.
723 * So for each group we take up 2 blocks. A page can
724 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
725 * So it can have information regarding groups_per_page which
726 * is blocks_per_page/2
729 static int ext4_mb_init_cache(struct page *page, char *incore)
736 ext4_group_t first_group;
738 struct super_block *sb;
739 struct buffer_head *bhs;
740 struct buffer_head **bh;
745 mb_debug("init page %lu\n", page->index);
747 inode = page->mapping->host;
749 blocksize = 1 << inode->i_blkbits;
750 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
752 groups_per_page = blocks_per_page >> 1;
753 if (groups_per_page == 0)
756 /* allocate buffer_heads to read bitmaps */
757 if (groups_per_page > 1) {
759 i = sizeof(struct buffer_head *) * groups_per_page;
760 bh = kzalloc(i, GFP_NOFS);
766 first_group = page->index * blocks_per_page / 2;
768 /* read all groups the page covers into the cache */
769 for (i = 0; i < groups_per_page; i++) {
770 struct ext4_group_desc *desc;
772 if (first_group + i >= EXT4_SB(sb)->s_groups_count)
776 desc = ext4_get_group_desc(sb, first_group + i, NULL);
781 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
785 if (buffer_uptodate(bh[i]) &&
786 !(desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))
790 spin_lock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
791 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
792 ext4_init_block_bitmap(sb, bh[i],
793 first_group + i, desc);
794 set_buffer_uptodate(bh[i]);
795 unlock_buffer(bh[i]);
796 spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
799 spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
801 bh[i]->b_end_io = end_buffer_read_sync;
802 submit_bh(READ, bh[i]);
803 mb_debug("read bitmap for group %lu\n", first_group + i);
806 /* wait for I/O completion */
807 for (i = 0; i < groups_per_page && bh[i]; i++)
808 wait_on_buffer(bh[i]);
811 for (i = 0; i < groups_per_page && bh[i]; i++)
812 if (!buffer_uptodate(bh[i]))
816 first_block = page->index * blocks_per_page;
817 for (i = 0; i < blocks_per_page; i++) {
819 struct ext4_group_info *grinfo;
821 group = (first_block + i) >> 1;
822 if (group >= EXT4_SB(sb)->s_groups_count)
826 * data carry information regarding this
827 * particular group in the format specified
831 data = page_address(page) + (i * blocksize);
832 bitmap = bh[group - first_group]->b_data;
835 * We place the buddy block and bitmap block
838 if ((first_block + i) & 1) {
839 /* this is block of buddy */
840 BUG_ON(incore == NULL);
841 mb_debug("put buddy for group %u in page %lu/%x\n",
842 group, page->index, i * blocksize);
843 memset(data, 0xff, blocksize);
844 grinfo = ext4_get_group_info(sb, group);
845 grinfo->bb_fragments = 0;
846 memset(grinfo->bb_counters, 0,
847 sizeof(unsigned short)*(sb->s_blocksize_bits+2));
849 * incore got set to the group block bitmap below
851 ext4_mb_generate_buddy(sb, data, incore, group);
854 /* this is block of bitmap */
855 BUG_ON(incore != NULL);
856 mb_debug("put bitmap for group %u in page %lu/%x\n",
857 group, page->index, i * blocksize);
859 /* see comments in ext4_mb_put_pa() */
860 ext4_lock_group(sb, group);
861 memcpy(data, bitmap, blocksize);
863 /* mark all preallocated blks used in in-core bitmap */
864 ext4_mb_generate_from_pa(sb, data, group);
865 ext4_unlock_group(sb, group);
867 /* set incore so that the buddy information can be
868 * generated using this
873 SetPageUptodate(page);
877 for (i = 0; i < groups_per_page && bh[i]; i++)
885 static noinline_for_stack int
886 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
887 struct ext4_buddy *e4b)
889 struct ext4_sb_info *sbi = EXT4_SB(sb);
890 struct inode *inode = sbi->s_buddy_cache;
898 mb_debug("load group %lu\n", group);
900 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
902 e4b->bd_blkbits = sb->s_blocksize_bits;
903 e4b->bd_info = ext4_get_group_info(sb, group);
905 e4b->bd_group = group;
906 e4b->bd_buddy_page = NULL;
907 e4b->bd_bitmap_page = NULL;
910 * the buddy cache inode stores the block bitmap
911 * and buddy information in consecutive blocks.
912 * So for each group we need two blocks.
915 pnum = block / blocks_per_page;
916 poff = block % blocks_per_page;
918 /* we could use find_or_create_page(), but it locks page
919 * what we'd like to avoid in fast path ... */
920 page = find_get_page(inode->i_mapping, pnum);
921 if (page == NULL || !PageUptodate(page)) {
923 page_cache_release(page);
924 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
926 BUG_ON(page->mapping != inode->i_mapping);
927 if (!PageUptodate(page)) {
928 ret = ext4_mb_init_cache(page, NULL);
933 mb_cmp_bitmaps(e4b, page_address(page) +
934 (poff * sb->s_blocksize));
939 if (page == NULL || !PageUptodate(page)) {
943 e4b->bd_bitmap_page = page;
944 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
945 mark_page_accessed(page);
948 pnum = block / blocks_per_page;
949 poff = block % blocks_per_page;
951 page = find_get_page(inode->i_mapping, pnum);
952 if (page == NULL || !PageUptodate(page)) {
954 page_cache_release(page);
955 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
957 BUG_ON(page->mapping != inode->i_mapping);
958 if (!PageUptodate(page)) {
959 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
968 if (page == NULL || !PageUptodate(page)) {
972 e4b->bd_buddy_page = page;
973 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
974 mark_page_accessed(page);
976 BUG_ON(e4b->bd_bitmap_page == NULL);
977 BUG_ON(e4b->bd_buddy_page == NULL);
982 if (e4b->bd_bitmap_page)
983 page_cache_release(e4b->bd_bitmap_page);
984 if (e4b->bd_buddy_page)
985 page_cache_release(e4b->bd_buddy_page);
986 e4b->bd_buddy = NULL;
987 e4b->bd_bitmap = NULL;
991 static void ext4_mb_release_desc(struct ext4_buddy *e4b)
993 if (e4b->bd_bitmap_page)
994 page_cache_release(e4b->bd_bitmap_page);
995 if (e4b->bd_buddy_page)
996 page_cache_release(e4b->bd_buddy_page);
1000 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1005 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1006 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1008 bb = EXT4_MB_BUDDY(e4b);
1009 while (order <= e4b->bd_blkbits + 1) {
1011 if (!mb_test_bit(block, bb)) {
1012 /* this block is part of buddy of order 'order' */
1015 bb += 1 << (e4b->bd_blkbits - order);
1021 static void mb_clear_bits(spinlock_t *lock, void *bm, int cur, int len)
1027 if ((cur & 31) == 0 && (len - cur) >= 32) {
1028 /* fast path: clear whole word at once */
1029 addr = bm + (cur >> 3);
1034 mb_clear_bit_atomic(lock, cur, bm);
1039 static void mb_set_bits(spinlock_t *lock, void *bm, int cur, int len)
1045 if ((cur & 31) == 0 && (len - cur) >= 32) {
1046 /* fast path: set whole word at once */
1047 addr = bm + (cur >> 3);
1052 mb_set_bit_atomic(lock, cur, bm);
1057 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1058 int first, int count)
1065 struct super_block *sb = e4b->bd_sb;
1067 BUG_ON(first + count > (sb->s_blocksize << 3));
1068 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
1069 mb_check_buddy(e4b);
1070 mb_free_blocks_double(inode, e4b, first, count);
1072 e4b->bd_info->bb_free += count;
1073 if (first < e4b->bd_info->bb_first_free)
1074 e4b->bd_info->bb_first_free = first;
1076 /* let's maintain fragments counter */
1078 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1079 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1080 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1082 e4b->bd_info->bb_fragments--;
1083 else if (!block && !max)
1084 e4b->bd_info->bb_fragments++;
1086 /* let's maintain buddy itself */
1087 while (count-- > 0) {
1091 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1092 ext4_fsblk_t blocknr;
1093 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1096 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
1097 ext4_unlock_group(sb, e4b->bd_group);
1098 ext4_error(sb, __func__, "double-free of inode"
1099 " %lu's block %llu(bit %u in group %lu)\n",
1100 inode ? inode->i_ino : 0, blocknr, block,
1102 ext4_lock_group(sb, e4b->bd_group);
1104 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1105 e4b->bd_info->bb_counters[order]++;
1107 /* start of the buddy */
1108 buddy = mb_find_buddy(e4b, order, &max);
1112 if (mb_test_bit(block, buddy) ||
1113 mb_test_bit(block + 1, buddy))
1116 /* both the buddies are free, try to coalesce them */
1117 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1123 /* for special purposes, we don't set
1124 * free bits in bitmap */
1125 mb_set_bit(block, buddy);
1126 mb_set_bit(block + 1, buddy);
1128 e4b->bd_info->bb_counters[order]--;
1129 e4b->bd_info->bb_counters[order]--;
1133 e4b->bd_info->bb_counters[order]++;
1135 mb_clear_bit(block, buddy2);
1139 mb_check_buddy(e4b);
1142 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1143 int needed, struct ext4_free_extent *ex)
1150 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1153 buddy = mb_find_buddy(e4b, order, &max);
1154 BUG_ON(buddy == NULL);
1155 BUG_ON(block >= max);
1156 if (mb_test_bit(block, buddy)) {
1163 /* FIXME dorp order completely ? */
1164 if (likely(order == 0)) {
1165 /* find actual order */
1166 order = mb_find_order_for_block(e4b, block);
1167 block = block >> order;
1170 ex->fe_len = 1 << order;
1171 ex->fe_start = block << order;
1172 ex->fe_group = e4b->bd_group;
1174 /* calc difference from given start */
1175 next = next - ex->fe_start;
1177 ex->fe_start += next;
1179 while (needed > ex->fe_len &&
1180 (buddy = mb_find_buddy(e4b, order, &max))) {
1182 if (block + 1 >= max)
1185 next = (block + 1) * (1 << order);
1186 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1189 ord = mb_find_order_for_block(e4b, next);
1192 block = next >> order;
1193 ex->fe_len += 1 << order;
1196 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1200 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1206 int start = ex->fe_start;
1207 int len = ex->fe_len;
1212 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1213 BUG_ON(e4b->bd_group != ex->fe_group);
1214 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1215 mb_check_buddy(e4b);
1216 mb_mark_used_double(e4b, start, len);
1218 e4b->bd_info->bb_free -= len;
1219 if (e4b->bd_info->bb_first_free == start)
1220 e4b->bd_info->bb_first_free += len;
1222 /* let's maintain fragments counter */
1224 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1225 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1226 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1228 e4b->bd_info->bb_fragments++;
1229 else if (!mlen && !max)
1230 e4b->bd_info->bb_fragments--;
1232 /* let's maintain buddy itself */
1234 ord = mb_find_order_for_block(e4b, start);
1236 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1237 /* the whole chunk may be allocated at once! */
1239 buddy = mb_find_buddy(e4b, ord, &max);
1240 BUG_ON((start >> ord) >= max);
1241 mb_set_bit(start >> ord, buddy);
1242 e4b->bd_info->bb_counters[ord]--;
1249 /* store for history */
1251 ret = len | (ord << 16);
1253 /* we have to split large buddy */
1255 buddy = mb_find_buddy(e4b, ord, &max);
1256 mb_set_bit(start >> ord, buddy);
1257 e4b->bd_info->bb_counters[ord]--;
1260 cur = (start >> ord) & ~1U;
1261 buddy = mb_find_buddy(e4b, ord, &max);
1262 mb_clear_bit(cur, buddy);
1263 mb_clear_bit(cur + 1, buddy);
1264 e4b->bd_info->bb_counters[ord]++;
1265 e4b->bd_info->bb_counters[ord]++;
1268 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b->bd_sb), ex->fe_group),
1269 EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1270 mb_check_buddy(e4b);
1276 * Must be called under group lock!
1278 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1279 struct ext4_buddy *e4b)
1281 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1284 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1285 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1287 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1288 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1289 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1291 /* preallocation can change ac_b_ex, thus we store actually
1292 * allocated blocks for history */
1293 ac->ac_f_ex = ac->ac_b_ex;
1295 ac->ac_status = AC_STATUS_FOUND;
1296 ac->ac_tail = ret & 0xffff;
1297 ac->ac_buddy = ret >> 16;
1299 /* XXXXXXX: SUCH A HORRIBLE **CK */
1301 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1302 get_page(ac->ac_bitmap_page);
1303 ac->ac_buddy_page = e4b->bd_buddy_page;
1304 get_page(ac->ac_buddy_page);
1306 /* store last allocated for subsequent stream allocation */
1307 if ((ac->ac_flags & EXT4_MB_HINT_DATA)) {
1308 spin_lock(&sbi->s_md_lock);
1309 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1310 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1311 spin_unlock(&sbi->s_md_lock);
1316 * regular allocator, for general purposes allocation
1319 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1320 struct ext4_buddy *e4b,
1323 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1324 struct ext4_free_extent *bex = &ac->ac_b_ex;
1325 struct ext4_free_extent *gex = &ac->ac_g_ex;
1326 struct ext4_free_extent ex;
1330 * We don't want to scan for a whole year
1332 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1333 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1334 ac->ac_status = AC_STATUS_BREAK;
1339 * Haven't found good chunk so far, let's continue
1341 if (bex->fe_len < gex->fe_len)
1344 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1345 && bex->fe_group == e4b->bd_group) {
1346 /* recheck chunk's availability - we don't know
1347 * when it was found (within this lock-unlock
1349 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1350 if (max >= gex->fe_len) {
1351 ext4_mb_use_best_found(ac, e4b);
1358 * The routine checks whether found extent is good enough. If it is,
1359 * then the extent gets marked used and flag is set to the context
1360 * to stop scanning. Otherwise, the extent is compared with the
1361 * previous found extent and if new one is better, then it's stored
1362 * in the context. Later, the best found extent will be used, if
1363 * mballoc can't find good enough extent.
1365 * FIXME: real allocation policy is to be designed yet!
1367 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1368 struct ext4_free_extent *ex,
1369 struct ext4_buddy *e4b)
1371 struct ext4_free_extent *bex = &ac->ac_b_ex;
1372 struct ext4_free_extent *gex = &ac->ac_g_ex;
1374 BUG_ON(ex->fe_len <= 0);
1375 BUG_ON(ex->fe_len >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1376 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1377 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1382 * The special case - take what you catch first
1384 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1386 ext4_mb_use_best_found(ac, e4b);
1391 * Let's check whether the chuck is good enough
1393 if (ex->fe_len == gex->fe_len) {
1395 ext4_mb_use_best_found(ac, e4b);
1400 * If this is first found extent, just store it in the context
1402 if (bex->fe_len == 0) {
1408 * If new found extent is better, store it in the context
1410 if (bex->fe_len < gex->fe_len) {
1411 /* if the request isn't satisfied, any found extent
1412 * larger than previous best one is better */
1413 if (ex->fe_len > bex->fe_len)
1415 } else if (ex->fe_len > gex->fe_len) {
1416 /* if the request is satisfied, then we try to find
1417 * an extent that still satisfy the request, but is
1418 * smaller than previous one */
1419 if (ex->fe_len < bex->fe_len)
1423 ext4_mb_check_limits(ac, e4b, 0);
1426 static int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1427 struct ext4_buddy *e4b)
1429 struct ext4_free_extent ex = ac->ac_b_ex;
1430 ext4_group_t group = ex.fe_group;
1434 BUG_ON(ex.fe_len <= 0);
1435 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1439 ext4_lock_group(ac->ac_sb, group);
1440 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1444 ext4_mb_use_best_found(ac, e4b);
1447 ext4_unlock_group(ac->ac_sb, group);
1448 ext4_mb_release_desc(e4b);
1453 static int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1454 struct ext4_buddy *e4b)
1456 ext4_group_t group = ac->ac_g_ex.fe_group;
1459 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1460 struct ext4_super_block *es = sbi->s_es;
1461 struct ext4_free_extent ex;
1463 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1466 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1470 ext4_lock_group(ac->ac_sb, group);
1471 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1472 ac->ac_g_ex.fe_len, &ex);
1474 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1477 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1478 ex.fe_start + le32_to_cpu(es->s_first_data_block);
1479 /* use do_div to get remainder (would be 64-bit modulo) */
1480 if (do_div(start, sbi->s_stripe) == 0) {
1483 ext4_mb_use_best_found(ac, e4b);
1485 } else if (max >= ac->ac_g_ex.fe_len) {
1486 BUG_ON(ex.fe_len <= 0);
1487 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1488 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1491 ext4_mb_use_best_found(ac, e4b);
1492 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1493 /* Sometimes, caller may want to merge even small
1494 * number of blocks to an existing extent */
1495 BUG_ON(ex.fe_len <= 0);
1496 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1497 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1500 ext4_mb_use_best_found(ac, e4b);
1502 ext4_unlock_group(ac->ac_sb, group);
1503 ext4_mb_release_desc(e4b);
1509 * The routine scans buddy structures (not bitmap!) from given order
1510 * to max order and tries to find big enough chunk to satisfy the req
1512 static void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1513 struct ext4_buddy *e4b)
1515 struct super_block *sb = ac->ac_sb;
1516 struct ext4_group_info *grp = e4b->bd_info;
1522 BUG_ON(ac->ac_2order <= 0);
1523 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1524 if (grp->bb_counters[i] == 0)
1527 buddy = mb_find_buddy(e4b, i, &max);
1528 BUG_ON(buddy == NULL);
1530 k = mb_find_next_zero_bit(buddy, max, 0);
1535 ac->ac_b_ex.fe_len = 1 << i;
1536 ac->ac_b_ex.fe_start = k << i;
1537 ac->ac_b_ex.fe_group = e4b->bd_group;
1539 ext4_mb_use_best_found(ac, e4b);
1541 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1543 if (EXT4_SB(sb)->s_mb_stats)
1544 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1551 * The routine scans the group and measures all found extents.
1552 * In order to optimize scanning, caller must pass number of
1553 * free blocks in the group, so the routine can know upper limit.
1555 static void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1556 struct ext4_buddy *e4b)
1558 struct super_block *sb = ac->ac_sb;
1559 void *bitmap = EXT4_MB_BITMAP(e4b);
1560 struct ext4_free_extent ex;
1564 free = e4b->bd_info->bb_free;
1567 i = e4b->bd_info->bb_first_free;
1569 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1570 i = mb_find_next_zero_bit(bitmap,
1571 EXT4_BLOCKS_PER_GROUP(sb), i);
1572 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1574 * IF we have corrupt bitmap, we won't find any
1575 * free blocks even though group info says we
1576 * we have free blocks
1578 ext4_error(sb, __func__, "%d free blocks as per "
1579 "group info. But bitmap says 0\n",
1584 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1585 BUG_ON(ex.fe_len <= 0);
1586 if (free < ex.fe_len) {
1587 ext4_error(sb, __func__, "%d free blocks as per "
1588 "group info. But got %d blocks\n",
1591 * The number of free blocks differs. This mostly
1592 * indicate that the bitmap is corrupt. So exit
1593 * without claiming the space.
1598 ext4_mb_measure_extent(ac, &ex, e4b);
1604 ext4_mb_check_limits(ac, e4b, 1);
1608 * This is a special case for storages like raid5
1609 * we try to find stripe-aligned chunks for stripe-size requests
1610 * XXX should do so at least for multiples of stripe size as well
1612 static void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1613 struct ext4_buddy *e4b)
1615 struct super_block *sb = ac->ac_sb;
1616 struct ext4_sb_info *sbi = EXT4_SB(sb);
1617 void *bitmap = EXT4_MB_BITMAP(e4b);
1618 struct ext4_free_extent ex;
1619 ext4_fsblk_t first_group_block;
1624 BUG_ON(sbi->s_stripe == 0);
1626 /* find first stripe-aligned block in group */
1627 first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1628 + le32_to_cpu(sbi->s_es->s_first_data_block);
1629 a = first_group_block + sbi->s_stripe - 1;
1630 do_div(a, sbi->s_stripe);
1631 i = (a * sbi->s_stripe) - first_group_block;
1633 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1634 if (!mb_test_bit(i, bitmap)) {
1635 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1636 if (max >= sbi->s_stripe) {
1639 ext4_mb_use_best_found(ac, e4b);
1647 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1648 ext4_group_t group, int cr)
1650 unsigned free, fragments;
1652 struct ext4_group_desc *desc;
1653 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1655 BUG_ON(cr < 0 || cr >= 4);
1656 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1658 free = grp->bb_free;
1659 fragments = grp->bb_fragments;
1667 BUG_ON(ac->ac_2order == 0);
1668 /* If this group is uninitialized, skip it initially */
1669 desc = ext4_get_group_desc(ac->ac_sb, group, NULL);
1670 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
1673 bits = ac->ac_sb->s_blocksize_bits + 1;
1674 for (i = ac->ac_2order; i <= bits; i++)
1675 if (grp->bb_counters[i] > 0)
1679 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1683 if (free >= ac->ac_g_ex.fe_len)
1695 static noinline_for_stack int
1696 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1703 struct ext4_sb_info *sbi;
1704 struct super_block *sb;
1705 struct ext4_buddy e4b;
1710 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1712 /* first, try the goal */
1713 err = ext4_mb_find_by_goal(ac, &e4b);
1714 if (err || ac->ac_status == AC_STATUS_FOUND)
1717 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1721 * ac->ac2_order is set only if the fe_len is a power of 2
1722 * if ac2_order is set we also set criteria to 0 so that we
1723 * try exact allocation using buddy.
1725 i = fls(ac->ac_g_ex.fe_len);
1728 * We search using buddy data only if the order of the request
1729 * is greater than equal to the sbi_s_mb_order2_reqs
1730 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1732 if (i >= sbi->s_mb_order2_reqs) {
1734 * This should tell if fe_len is exactly power of 2
1736 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1737 ac->ac_2order = i - 1;
1740 bsbits = ac->ac_sb->s_blocksize_bits;
1741 /* if stream allocation is enabled, use global goal */
1742 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
1743 isize = i_size_read(ac->ac_inode) >> bsbits;
1747 if (size < sbi->s_mb_stream_request &&
1748 (ac->ac_flags & EXT4_MB_HINT_DATA)) {
1749 /* TBD: may be hot point */
1750 spin_lock(&sbi->s_md_lock);
1751 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1752 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1753 spin_unlock(&sbi->s_md_lock);
1755 /* Let's just scan groups to find more-less suitable blocks */
1756 cr = ac->ac_2order ? 0 : 1;
1758 * cr == 0 try to get exact allocation,
1759 * cr == 3 try to get anything
1762 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1763 ac->ac_criteria = cr;
1765 * searching for the right group start
1766 * from the goal value specified
1768 group = ac->ac_g_ex.fe_group;
1770 for (i = 0; i < EXT4_SB(sb)->s_groups_count; group++, i++) {
1771 struct ext4_group_info *grp;
1772 struct ext4_group_desc *desc;
1774 if (group == EXT4_SB(sb)->s_groups_count)
1777 /* quick check to skip empty groups */
1778 grp = ext4_get_group_info(ac->ac_sb, group);
1779 if (grp->bb_free == 0)
1783 * if the group is already init we check whether it is
1784 * a good group and if not we don't load the buddy
1786 if (EXT4_MB_GRP_NEED_INIT(grp)) {
1788 * we need full data about the group
1789 * to make a good selection
1791 err = ext4_mb_load_buddy(sb, group, &e4b);
1794 ext4_mb_release_desc(&e4b);
1798 * If the particular group doesn't satisfy our
1799 * criteria we continue with the next group
1801 if (!ext4_mb_good_group(ac, group, cr))
1804 err = ext4_mb_load_buddy(sb, group, &e4b);
1808 ext4_lock_group(sb, group);
1809 if (!ext4_mb_good_group(ac, group, cr)) {
1810 /* someone did allocation from this group */
1811 ext4_unlock_group(sb, group);
1812 ext4_mb_release_desc(&e4b);
1816 ac->ac_groups_scanned++;
1817 desc = ext4_get_group_desc(sb, group, NULL);
1818 if (cr == 0 || (desc->bg_flags &
1819 cpu_to_le16(EXT4_BG_BLOCK_UNINIT) &&
1820 ac->ac_2order != 0))
1821 ext4_mb_simple_scan_group(ac, &e4b);
1823 ac->ac_g_ex.fe_len == sbi->s_stripe)
1824 ext4_mb_scan_aligned(ac, &e4b);
1826 ext4_mb_complex_scan_group(ac, &e4b);
1828 ext4_unlock_group(sb, group);
1829 ext4_mb_release_desc(&e4b);
1831 if (ac->ac_status != AC_STATUS_CONTINUE)
1836 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
1837 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1839 * We've been searching too long. Let's try to allocate
1840 * the best chunk we've found so far
1843 ext4_mb_try_best_found(ac, &e4b);
1844 if (ac->ac_status != AC_STATUS_FOUND) {
1846 * Someone more lucky has already allocated it.
1847 * The only thing we can do is just take first
1849 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
1851 ac->ac_b_ex.fe_group = 0;
1852 ac->ac_b_ex.fe_start = 0;
1853 ac->ac_b_ex.fe_len = 0;
1854 ac->ac_status = AC_STATUS_CONTINUE;
1855 ac->ac_flags |= EXT4_MB_HINT_FIRST;
1857 atomic_inc(&sbi->s_mb_lost_chunks);
1865 #ifdef EXT4_MB_HISTORY
1866 struct ext4_mb_proc_session {
1867 struct ext4_mb_history *history;
1868 struct super_block *sb;
1873 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
1874 struct ext4_mb_history *hs,
1877 if (hs == s->history + s->max)
1879 if (!first && hs == s->history + s->start)
1881 while (hs->orig.fe_len == 0) {
1883 if (hs == s->history + s->max)
1885 if (hs == s->history + s->start)
1891 static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
1893 struct ext4_mb_proc_session *s = seq->private;
1894 struct ext4_mb_history *hs;
1898 return SEQ_START_TOKEN;
1899 hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
1902 while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
1906 static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
1909 struct ext4_mb_proc_session *s = seq->private;
1910 struct ext4_mb_history *hs = v;
1913 if (v == SEQ_START_TOKEN)
1914 return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
1916 return ext4_mb_history_skip_empty(s, ++hs, 0);
1919 static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
1921 char buf[25], buf2[25], buf3[25], *fmt;
1922 struct ext4_mb_history *hs = v;
1924 if (v == SEQ_START_TOKEN) {
1925 seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
1926 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
1927 "pid", "inode", "original", "goal", "result", "found",
1928 "grps", "cr", "flags", "merge", "tail", "broken");
1932 if (hs->op == EXT4_MB_HISTORY_ALLOC) {
1933 fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
1934 "%-5u %-5s %-5u %-6u\n";
1935 sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
1936 hs->result.fe_start, hs->result.fe_len,
1937 hs->result.fe_logical);
1938 sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
1939 hs->orig.fe_start, hs->orig.fe_len,
1940 hs->orig.fe_logical);
1941 sprintf(buf3, "%lu/%d/%u@%u", hs->goal.fe_group,
1942 hs->goal.fe_start, hs->goal.fe_len,
1943 hs->goal.fe_logical);
1944 seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
1945 hs->found, hs->groups, hs->cr, hs->flags,
1946 hs->merged ? "M" : "", hs->tail,
1947 hs->buddy ? 1 << hs->buddy : 0);
1948 } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
1949 fmt = "%-5u %-8u %-23s %-23s %-23s\n";
1950 sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
1951 hs->result.fe_start, hs->result.fe_len,
1952 hs->result.fe_logical);
1953 sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
1954 hs->orig.fe_start, hs->orig.fe_len,
1955 hs->orig.fe_logical);
1956 seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
1957 } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
1958 sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
1959 hs->result.fe_start, hs->result.fe_len);
1960 seq_printf(seq, "%-5u %-8u %-23s discard\n",
1961 hs->pid, hs->ino, buf2);
1962 } else if (hs->op == EXT4_MB_HISTORY_FREE) {
1963 sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
1964 hs->result.fe_start, hs->result.fe_len);
1965 seq_printf(seq, "%-5u %-8u %-23s free\n",
1966 hs->pid, hs->ino, buf2);
1971 static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
1975 static struct seq_operations ext4_mb_seq_history_ops = {
1976 .start = ext4_mb_seq_history_start,
1977 .next = ext4_mb_seq_history_next,
1978 .stop = ext4_mb_seq_history_stop,
1979 .show = ext4_mb_seq_history_show,
1982 static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
1984 struct super_block *sb = PDE(inode)->data;
1985 struct ext4_sb_info *sbi = EXT4_SB(sb);
1986 struct ext4_mb_proc_session *s;
1990 if (unlikely(sbi->s_mb_history == NULL))
1992 s = kmalloc(sizeof(*s), GFP_KERNEL);
1996 size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
1997 s->history = kmalloc(size, GFP_KERNEL);
1998 if (s->history == NULL) {
2003 spin_lock(&sbi->s_mb_history_lock);
2004 memcpy(s->history, sbi->s_mb_history, size);
2005 s->max = sbi->s_mb_history_max;
2006 s->start = sbi->s_mb_history_cur % s->max;
2007 spin_unlock(&sbi->s_mb_history_lock);
2009 rc = seq_open(file, &ext4_mb_seq_history_ops);
2011 struct seq_file *m = (struct seq_file *)file->private_data;
2021 static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
2023 struct seq_file *seq = (struct seq_file *)file->private_data;
2024 struct ext4_mb_proc_session *s = seq->private;
2027 return seq_release(inode, file);
2030 static ssize_t ext4_mb_seq_history_write(struct file *file,
2031 const char __user *buffer,
2032 size_t count, loff_t *ppos)
2034 struct seq_file *seq = (struct seq_file *)file->private_data;
2035 struct ext4_mb_proc_session *s = seq->private;
2036 struct super_block *sb = s->sb;
2040 if (count >= sizeof(str)) {
2041 printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
2042 "mb_history", (int)sizeof(str));
2046 if (copy_from_user(str, buffer, count))
2049 value = simple_strtol(str, NULL, 0);
2052 EXT4_SB(sb)->s_mb_history_filter = value;
2057 static struct file_operations ext4_mb_seq_history_fops = {
2058 .owner = THIS_MODULE,
2059 .open = ext4_mb_seq_history_open,
2061 .write = ext4_mb_seq_history_write,
2062 .llseek = seq_lseek,
2063 .release = ext4_mb_seq_history_release,
2066 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2068 struct super_block *sb = seq->private;
2069 struct ext4_sb_info *sbi = EXT4_SB(sb);
2072 if (*pos < 0 || *pos >= sbi->s_groups_count)
2076 return (void *) group;
2079 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2081 struct super_block *sb = seq->private;
2082 struct ext4_sb_info *sbi = EXT4_SB(sb);
2086 if (*pos < 0 || *pos >= sbi->s_groups_count)
2089 return (void *) group;;
2092 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2094 struct super_block *sb = seq->private;
2095 long group = (long) v;
2098 struct ext4_buddy e4b;
2100 struct ext4_group_info info;
2101 unsigned short counters[16];
2106 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2107 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2108 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2109 "group", "free", "frags", "first",
2110 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2111 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2113 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2114 sizeof(struct ext4_group_info);
2115 err = ext4_mb_load_buddy(sb, group, &e4b);
2117 seq_printf(seq, "#%-5lu: I/O error\n", group);
2120 ext4_lock_group(sb, group);
2121 memcpy(&sg, ext4_get_group_info(sb, group), i);
2122 ext4_unlock_group(sb, group);
2123 ext4_mb_release_desc(&e4b);
2125 seq_printf(seq, "#%-5lu: %-5u %-5u %-5u [", group, sg.info.bb_free,
2126 sg.info.bb_fragments, sg.info.bb_first_free);
2127 for (i = 0; i <= 13; i++)
2128 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2129 sg.info.bb_counters[i] : 0);
2130 seq_printf(seq, " ]\n");
2135 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2139 static struct seq_operations ext4_mb_seq_groups_ops = {
2140 .start = ext4_mb_seq_groups_start,
2141 .next = ext4_mb_seq_groups_next,
2142 .stop = ext4_mb_seq_groups_stop,
2143 .show = ext4_mb_seq_groups_show,
2146 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2148 struct super_block *sb = PDE(inode)->data;
2151 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2153 struct seq_file *m = (struct seq_file *)file->private_data;
2160 static struct file_operations ext4_mb_seq_groups_fops = {
2161 .owner = THIS_MODULE,
2162 .open = ext4_mb_seq_groups_open,
2164 .llseek = seq_lseek,
2165 .release = seq_release,
2168 static void ext4_mb_history_release(struct super_block *sb)
2170 struct ext4_sb_info *sbi = EXT4_SB(sb);
2172 if (sbi->s_proc != NULL) {
2173 remove_proc_entry("mb_groups", sbi->s_proc);
2174 remove_proc_entry("mb_history", sbi->s_proc);
2176 kfree(sbi->s_mb_history);
2179 static void ext4_mb_history_init(struct super_block *sb)
2181 struct ext4_sb_info *sbi = EXT4_SB(sb);
2184 if (sbi->s_proc != NULL) {
2185 proc_create_data("mb_history", S_IRUGO, sbi->s_proc,
2186 &ext4_mb_seq_history_fops, sb);
2187 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2188 &ext4_mb_seq_groups_fops, sb);
2191 sbi->s_mb_history_max = 1000;
2192 sbi->s_mb_history_cur = 0;
2193 spin_lock_init(&sbi->s_mb_history_lock);
2194 i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
2195 sbi->s_mb_history = kzalloc(i, GFP_KERNEL);
2196 /* if we can't allocate history, then we simple won't use it */
2199 static noinline_for_stack void
2200 ext4_mb_store_history(struct ext4_allocation_context *ac)
2202 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2203 struct ext4_mb_history h;
2205 if (unlikely(sbi->s_mb_history == NULL))
2208 if (!(ac->ac_op & sbi->s_mb_history_filter))
2212 h.pid = current->pid;
2213 h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2214 h.orig = ac->ac_o_ex;
2215 h.result = ac->ac_b_ex;
2216 h.flags = ac->ac_flags;
2217 h.found = ac->ac_found;
2218 h.groups = ac->ac_groups_scanned;
2219 h.cr = ac->ac_criteria;
2220 h.tail = ac->ac_tail;
2221 h.buddy = ac->ac_buddy;
2223 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
2224 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2225 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2227 h.goal = ac->ac_g_ex;
2228 h.result = ac->ac_f_ex;
2231 spin_lock(&sbi->s_mb_history_lock);
2232 memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2233 if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2234 sbi->s_mb_history_cur = 0;
2235 spin_unlock(&sbi->s_mb_history_lock);
2239 #define ext4_mb_history_release(sb)
2240 #define ext4_mb_history_init(sb)
2244 /* Create and initialize ext4_group_info data for the given group. */
2245 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2246 struct ext4_group_desc *desc)
2250 struct ext4_sb_info *sbi = EXT4_SB(sb);
2251 struct ext4_group_info **meta_group_info;
2254 * First check if this group is the first of a reserved block.
2255 * If it's true, we have to allocate a new table of pointers
2256 * to ext4_group_info structures
2258 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2259 metalen = sizeof(*meta_group_info) <<
2260 EXT4_DESC_PER_BLOCK_BITS(sb);
2261 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2262 if (meta_group_info == NULL) {
2263 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2265 goto exit_meta_group_info;
2267 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2272 * calculate needed size. if change bb_counters size,
2273 * don't forget about ext4_mb_generate_buddy()
2275 len = offsetof(typeof(**meta_group_info),
2276 bb_counters[sb->s_blocksize_bits + 2]);
2279 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2280 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2282 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2283 if (meta_group_info[i] == NULL) {
2284 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2285 goto exit_group_info;
2287 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2288 &(meta_group_info[i]->bb_state));
2291 * initialize bb_free to be able to skip
2292 * empty groups without initialization
2294 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2295 meta_group_info[i]->bb_free =
2296 ext4_free_blocks_after_init(sb, group, desc);
2298 meta_group_info[i]->bb_free =
2299 le16_to_cpu(desc->bg_free_blocks_count);
2302 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2303 meta_group_info[i]->bb_free_root.rb_node = NULL;;
2307 struct buffer_head *bh;
2308 meta_group_info[i]->bb_bitmap =
2309 kmalloc(sb->s_blocksize, GFP_KERNEL);
2310 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2311 bh = ext4_read_block_bitmap(sb, group);
2313 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2322 /* If a meta_group_info table has been allocated, release it now */
2323 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2324 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2325 exit_meta_group_info:
2327 } /* ext4_mb_add_groupinfo */
2330 * Add a group to the existing groups.
2331 * This function is used for online resize
2333 int ext4_mb_add_more_groupinfo(struct super_block *sb, ext4_group_t group,
2334 struct ext4_group_desc *desc)
2336 struct ext4_sb_info *sbi = EXT4_SB(sb);
2337 struct inode *inode = sbi->s_buddy_cache;
2338 int blocks_per_page;
2344 /* Add group based on group descriptor*/
2345 err = ext4_mb_add_groupinfo(sb, group, desc);
2350 * Cache pages containing dynamic mb_alloc datas (buddy and bitmap
2351 * datas) are set not up to date so that they will be re-initilaized
2352 * during the next call to ext4_mb_load_buddy
2355 /* Set buddy page as not up to date */
2356 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
2358 pnum = block / blocks_per_page;
2359 page = find_get_page(inode->i_mapping, pnum);
2361 ClearPageUptodate(page);
2362 page_cache_release(page);
2365 /* Set bitmap page as not up to date */
2367 pnum = block / blocks_per_page;
2368 page = find_get_page(inode->i_mapping, pnum);
2370 ClearPageUptodate(page);
2371 page_cache_release(page);
2378 * Update an existing group.
2379 * This function is used for online resize
2381 void ext4_mb_update_group_info(struct ext4_group_info *grp, ext4_grpblk_t add)
2383 grp->bb_free += add;
2386 static int ext4_mb_init_backend(struct super_block *sb)
2390 struct ext4_sb_info *sbi = EXT4_SB(sb);
2391 struct ext4_super_block *es = sbi->s_es;
2392 int num_meta_group_infos;
2393 int num_meta_group_infos_max;
2395 struct ext4_group_info **meta_group_info;
2396 struct ext4_group_desc *desc;
2398 /* This is the number of blocks used by GDT */
2399 num_meta_group_infos = (sbi->s_groups_count + EXT4_DESC_PER_BLOCK(sb) -
2400 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2403 * This is the total number of blocks used by GDT including
2404 * the number of reserved blocks for GDT.
2405 * The s_group_info array is allocated with this value
2406 * to allow a clean online resize without a complex
2407 * manipulation of pointer.
2408 * The drawback is the unused memory when no resize
2409 * occurs but it's very low in terms of pages
2410 * (see comments below)
2411 * Need to handle this properly when META_BG resizing is allowed
2413 num_meta_group_infos_max = num_meta_group_infos +
2414 le16_to_cpu(es->s_reserved_gdt_blocks);
2417 * array_size is the size of s_group_info array. We round it
2418 * to the next power of two because this approximation is done
2419 * internally by kmalloc so we can have some more memory
2420 * for free here (e.g. may be used for META_BG resize).
2423 while (array_size < sizeof(*sbi->s_group_info) *
2424 num_meta_group_infos_max)
2425 array_size = array_size << 1;
2426 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2427 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2428 * So a two level scheme suffices for now. */
2429 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2430 if (sbi->s_group_info == NULL) {
2431 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2434 sbi->s_buddy_cache = new_inode(sb);
2435 if (sbi->s_buddy_cache == NULL) {
2436 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2439 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2441 metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb);
2442 for (i = 0; i < num_meta_group_infos; i++) {
2443 if ((i + 1) == num_meta_group_infos)
2444 metalen = sizeof(*meta_group_info) *
2445 (sbi->s_groups_count -
2446 (i << EXT4_DESC_PER_BLOCK_BITS(sb)));
2447 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2448 if (meta_group_info == NULL) {
2449 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2453 sbi->s_group_info[i] = meta_group_info;
2456 for (i = 0; i < sbi->s_groups_count; i++) {
2457 desc = ext4_get_group_desc(sb, i, NULL);
2460 "EXT4-fs: can't read descriptor %lu\n", i);
2463 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2471 kfree(ext4_get_group_info(sb, i));
2472 i = num_meta_group_infos;
2475 kfree(sbi->s_group_info[i]);
2476 iput(sbi->s_buddy_cache);
2478 kfree(sbi->s_group_info);
2482 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2484 struct ext4_sb_info *sbi = EXT4_SB(sb);
2490 i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
2492 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2493 if (sbi->s_mb_offsets == NULL) {
2496 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2497 if (sbi->s_mb_maxs == NULL) {
2498 kfree(sbi->s_mb_maxs);
2502 /* order 0 is regular bitmap */
2503 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2504 sbi->s_mb_offsets[0] = 0;
2508 max = sb->s_blocksize << 2;
2510 sbi->s_mb_offsets[i] = offset;
2511 sbi->s_mb_maxs[i] = max;
2512 offset += 1 << (sb->s_blocksize_bits - i);
2515 } while (i <= sb->s_blocksize_bits + 1);
2517 /* init file for buddy data */
2518 ret = ext4_mb_init_backend(sb);
2520 kfree(sbi->s_mb_offsets);
2521 kfree(sbi->s_mb_maxs);
2525 spin_lock_init(&sbi->s_md_lock);
2526 INIT_LIST_HEAD(&sbi->s_active_transaction);
2527 INIT_LIST_HEAD(&sbi->s_closed_transaction);
2528 INIT_LIST_HEAD(&sbi->s_committed_transaction);
2529 spin_lock_init(&sbi->s_bal_lock);
2531 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2532 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2533 sbi->s_mb_stats = MB_DEFAULT_STATS;
2534 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2535 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2536 sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
2537 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2539 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2540 if (sbi->s_locality_groups == NULL) {
2541 kfree(sbi->s_mb_offsets);
2542 kfree(sbi->s_mb_maxs);
2545 for_each_possible_cpu(i) {
2546 struct ext4_locality_group *lg;
2547 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2548 mutex_init(&lg->lg_mutex);
2549 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2550 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2551 spin_lock_init(&lg->lg_prealloc_lock);
2554 ext4_mb_init_per_dev_proc(sb);
2555 ext4_mb_history_init(sb);
2557 printk(KERN_INFO "EXT4-fs: mballoc enabled\n");
2561 /* need to called with ext4 group lock (ext4_lock_group) */
2562 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2564 struct ext4_prealloc_space *pa;
2565 struct list_head *cur, *tmp;
2568 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2569 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2570 list_del(&pa->pa_group_list);
2572 kmem_cache_free(ext4_pspace_cachep, pa);
2575 mb_debug("mballoc: %u PAs left\n", count);
2579 int ext4_mb_release(struct super_block *sb)
2582 int num_meta_group_infos;
2583 struct ext4_group_info *grinfo;
2584 struct ext4_sb_info *sbi = EXT4_SB(sb);
2586 /* release freed, non-committed blocks */
2587 spin_lock(&sbi->s_md_lock);
2588 list_splice_init(&sbi->s_closed_transaction,
2589 &sbi->s_committed_transaction);
2590 list_splice_init(&sbi->s_active_transaction,
2591 &sbi->s_committed_transaction);
2592 spin_unlock(&sbi->s_md_lock);
2593 ext4_mb_free_committed_blocks(sb);
2595 if (sbi->s_group_info) {
2596 for (i = 0; i < sbi->s_groups_count; i++) {
2597 grinfo = ext4_get_group_info(sb, i);
2599 kfree(grinfo->bb_bitmap);
2601 ext4_lock_group(sb, i);
2602 ext4_mb_cleanup_pa(grinfo);
2603 ext4_unlock_group(sb, i);
2606 num_meta_group_infos = (sbi->s_groups_count +
2607 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2608 EXT4_DESC_PER_BLOCK_BITS(sb);
2609 for (i = 0; i < num_meta_group_infos; i++)
2610 kfree(sbi->s_group_info[i]);
2611 kfree(sbi->s_group_info);
2613 kfree(sbi->s_mb_offsets);
2614 kfree(sbi->s_mb_maxs);
2615 if (sbi->s_buddy_cache)
2616 iput(sbi->s_buddy_cache);
2617 if (sbi->s_mb_stats) {
2619 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2620 atomic_read(&sbi->s_bal_allocated),
2621 atomic_read(&sbi->s_bal_reqs),
2622 atomic_read(&sbi->s_bal_success));
2624 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2625 "%u 2^N hits, %u breaks, %u lost\n",
2626 atomic_read(&sbi->s_bal_ex_scanned),
2627 atomic_read(&sbi->s_bal_goals),
2628 atomic_read(&sbi->s_bal_2orders),
2629 atomic_read(&sbi->s_bal_breaks),
2630 atomic_read(&sbi->s_mb_lost_chunks));
2632 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2633 sbi->s_mb_buddies_generated++,
2634 sbi->s_mb_generation_time);
2636 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2637 atomic_read(&sbi->s_mb_preallocated),
2638 atomic_read(&sbi->s_mb_discarded));
2641 free_percpu(sbi->s_locality_groups);
2642 ext4_mb_history_release(sb);
2643 ext4_mb_destroy_per_dev_proc(sb);
2648 static noinline_for_stack void
2649 ext4_mb_free_committed_blocks(struct super_block *sb)
2651 struct ext4_buddy e4b;
2652 struct ext4_group_info *db;
2653 struct ext4_sb_info *sbi = EXT4_SB(sb);
2654 int err, count = 0, count2 = 0;
2655 struct ext4_free_data *entry;
2657 if (list_empty(&sbi->s_committed_transaction))
2660 /* there is committed blocks to be freed yet */
2662 /* get next array of blocks */
2664 spin_lock(&sbi->s_md_lock);
2665 if (!list_empty(&sbi->s_committed_transaction)) {
2666 entry = list_entry(sbi->s_committed_transaction.next,
2667 struct ext4_free_data, list);
2668 list_del(&entry->list);
2670 spin_unlock(&sbi->s_md_lock);
2675 mb_debug("gonna free %u blocks in group %lu (0x%p):",
2676 entry->count, entry->group, entry);
2678 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2679 /* we expect to find existing buddy because it's pinned */
2683 /* there are blocks to put in buddy to make them really free */
2684 count += entry->count;
2686 ext4_lock_group(sb, entry->group);
2687 /* Take it out of per group rb tree */
2688 rb_erase(&entry->node, &(db->bb_free_root));
2689 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2691 if (!db->bb_free_root.rb_node) {
2692 /* No more items in the per group rb tree
2693 * balance refcounts from ext4_mb_free_metadata()
2695 page_cache_release(e4b.bd_buddy_page);
2696 page_cache_release(e4b.bd_bitmap_page);
2698 ext4_unlock_group(sb, entry->group);
2700 kmem_cache_free(ext4_free_ext_cachep, entry);
2701 ext4_mb_release_desc(&e4b);
2704 mb_debug("freed %u blocks in %u structures\n", count, count2);
2707 #define EXT4_MB_STATS_NAME "stats"
2708 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2709 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2710 #define EXT4_MB_ORDER2_REQ "order2_req"
2711 #define EXT4_MB_STREAM_REQ "stream_req"
2712 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2714 static int ext4_mb_init_per_dev_proc(struct super_block *sb)
2716 mode_t mode = S_IFREG | S_IRUGO | S_IWUSR;
2717 struct ext4_sb_info *sbi = EXT4_SB(sb);
2718 struct proc_dir_entry *proc;
2720 if (sbi->s_proc == NULL)
2723 EXT4_PROC_HANDLER(EXT4_MB_STATS_NAME, mb_stats);
2724 EXT4_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME, mb_max_to_scan);
2725 EXT4_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME, mb_min_to_scan);
2726 EXT4_PROC_HANDLER(EXT4_MB_ORDER2_REQ, mb_order2_reqs);
2727 EXT4_PROC_HANDLER(EXT4_MB_STREAM_REQ, mb_stream_request);
2728 EXT4_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC, mb_group_prealloc);
2732 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_proc);
2733 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_proc);
2734 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_proc);
2735 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_proc);
2736 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_proc);
2737 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_proc);
2741 static int ext4_mb_destroy_per_dev_proc(struct super_block *sb)
2743 struct ext4_sb_info *sbi = EXT4_SB(sb);
2745 if (sbi->s_proc == NULL)
2748 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_proc);
2749 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_proc);
2750 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_proc);
2751 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_proc);
2752 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_proc);
2753 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_proc);
2758 int __init init_ext4_mballoc(void)
2760 ext4_pspace_cachep =
2761 kmem_cache_create("ext4_prealloc_space",
2762 sizeof(struct ext4_prealloc_space),
2763 0, SLAB_RECLAIM_ACCOUNT, NULL);
2764 if (ext4_pspace_cachep == NULL)
2768 kmem_cache_create("ext4_alloc_context",
2769 sizeof(struct ext4_allocation_context),
2770 0, SLAB_RECLAIM_ACCOUNT, NULL);
2771 if (ext4_ac_cachep == NULL) {
2772 kmem_cache_destroy(ext4_pspace_cachep);
2776 ext4_free_ext_cachep =
2777 kmem_cache_create("ext4_free_block_extents",
2778 sizeof(struct ext4_free_data),
2779 0, SLAB_RECLAIM_ACCOUNT, NULL);
2780 if (ext4_free_ext_cachep == NULL) {
2781 kmem_cache_destroy(ext4_pspace_cachep);
2782 kmem_cache_destroy(ext4_ac_cachep);
2788 void exit_ext4_mballoc(void)
2790 /* XXX: synchronize_rcu(); */
2791 kmem_cache_destroy(ext4_pspace_cachep);
2792 kmem_cache_destroy(ext4_ac_cachep);
2793 kmem_cache_destroy(ext4_free_ext_cachep);
2798 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2799 * Returns 0 if success or error code
2801 static noinline_for_stack int
2802 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2803 handle_t *handle, unsigned long reserv_blks)
2805 struct buffer_head *bitmap_bh = NULL;
2806 struct ext4_super_block *es;
2807 struct ext4_group_desc *gdp;
2808 struct buffer_head *gdp_bh;
2809 struct ext4_sb_info *sbi;
2810 struct super_block *sb;
2814 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2815 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2823 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2827 err = ext4_journal_get_write_access(handle, bitmap_bh);
2832 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2836 ext4_debug("using block group %lu(%d)\n", ac->ac_b_ex.fe_group,
2837 gdp->bg_free_blocks_count);
2839 err = ext4_journal_get_write_access(handle, gdp_bh);
2843 block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
2844 + ac->ac_b_ex.fe_start
2845 + le32_to_cpu(es->s_first_data_block);
2847 len = ac->ac_b_ex.fe_len;
2848 if (in_range(ext4_block_bitmap(sb, gdp), block, len) ||
2849 in_range(ext4_inode_bitmap(sb, gdp), block, len) ||
2850 in_range(block, ext4_inode_table(sb, gdp),
2851 EXT4_SB(sb)->s_itb_per_group) ||
2852 in_range(block + len - 1, ext4_inode_table(sb, gdp),
2853 EXT4_SB(sb)->s_itb_per_group)) {
2854 ext4_error(sb, __func__,
2855 "Allocating block in system zone - block = %llu",
2857 /* File system mounted not to panic on error
2858 * Fix the bitmap and repeat the block allocation
2859 * We leak some of the blocks here.
2861 mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group),
2862 bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2863 ac->ac_b_ex.fe_len);
2864 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
2869 #ifdef AGGRESSIVE_CHECK
2872 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2873 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2874 bitmap_bh->b_data));
2878 mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), bitmap_bh->b_data,
2879 ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
2881 spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
2882 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2883 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2884 gdp->bg_free_blocks_count =
2885 cpu_to_le16(ext4_free_blocks_after_init(sb,
2886 ac->ac_b_ex.fe_group,
2889 le16_add_cpu(&gdp->bg_free_blocks_count, -ac->ac_b_ex.fe_len);
2890 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2891 spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
2892 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
2894 * Now reduce the dirty block count also. Should not go negative
2896 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2897 /* release all the reserved blocks if non delalloc */
2898 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
2900 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
2901 ac->ac_b_ex.fe_len);
2903 if (sbi->s_log_groups_per_flex) {
2904 ext4_group_t flex_group = ext4_flex_group(sbi,
2905 ac->ac_b_ex.fe_group);
2906 spin_lock(sb_bgl_lock(sbi, flex_group));
2907 sbi->s_flex_groups[flex_group].free_blocks -= ac->ac_b_ex.fe_len;
2908 spin_unlock(sb_bgl_lock(sbi, flex_group));
2911 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
2914 err = ext4_journal_dirty_metadata(handle, gdp_bh);
2923 * here we normalize request for locality group
2924 * Group request are normalized to s_strip size if we set the same via mount
2925 * option. If not we set it to s_mb_group_prealloc which can be configured via
2926 * /proc/fs/ext4/<partition>/group_prealloc
2928 * XXX: should we try to preallocate more than the group has now?
2930 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2932 struct super_block *sb = ac->ac_sb;
2933 struct ext4_locality_group *lg = ac->ac_lg;
2936 if (EXT4_SB(sb)->s_stripe)
2937 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
2939 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2940 mb_debug("#%u: goal %u blocks for locality group\n",
2941 current->pid, ac->ac_g_ex.fe_len);
2945 * Normalization means making request better in terms of
2946 * size and alignment
2948 static noinline_for_stack void
2949 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2950 struct ext4_allocation_request *ar)
2954 loff_t size, orig_size, start_off;
2955 ext4_lblk_t start, orig_start;
2956 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2957 struct ext4_prealloc_space *pa;
2959 /* do normalize only data requests, metadata requests
2960 do not need preallocation */
2961 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2964 /* sometime caller may want exact blocks */
2965 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2968 /* caller may indicate that preallocation isn't
2969 * required (it's a tail, for example) */
2970 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2973 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2974 ext4_mb_normalize_group_request(ac);
2978 bsbits = ac->ac_sb->s_blocksize_bits;
2980 /* first, let's learn actual file size
2981 * given current request is allocated */
2982 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2983 size = size << bsbits;
2984 if (size < i_size_read(ac->ac_inode))
2985 size = i_size_read(ac->ac_inode);
2987 /* max size of free chunks */
2990 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2991 (req <= (size) || max <= (chunk_size))
2993 /* first, try to predict filesize */
2994 /* XXX: should this table be tunable? */
2996 if (size <= 16 * 1024) {
2998 } else if (size <= 32 * 1024) {
3000 } else if (size <= 64 * 1024) {
3002 } else if (size <= 128 * 1024) {
3004 } else if (size <= 256 * 1024) {
3006 } else if (size <= 512 * 1024) {
3008 } else if (size <= 1024 * 1024) {
3010 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3011 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3012 (21 - bsbits)) << 21;
3013 size = 2 * 1024 * 1024;
3014 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3015 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3016 (22 - bsbits)) << 22;
3017 size = 4 * 1024 * 1024;
3018 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3019 (8<<20)>>bsbits, max, 8 * 1024)) {
3020 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3021 (23 - bsbits)) << 23;
3022 size = 8 * 1024 * 1024;
3024 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3025 size = ac->ac_o_ex.fe_len << bsbits;
3027 orig_size = size = size >> bsbits;
3028 orig_start = start = start_off >> bsbits;
3030 /* don't cover already allocated blocks in selected range */
3031 if (ar->pleft && start <= ar->lleft) {
3032 size -= ar->lleft + 1 - start;
3033 start = ar->lleft + 1;
3035 if (ar->pright && start + size - 1 >= ar->lright)
3036 size -= start + size - ar->lright;
3040 /* check we don't cross already preallocated blocks */
3042 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3043 unsigned long pa_end;
3047 spin_lock(&pa->pa_lock);
3048 if (pa->pa_deleted) {
3049 spin_unlock(&pa->pa_lock);
3053 pa_end = pa->pa_lstart + pa->pa_len;
3055 /* PA must not overlap original request */
3056 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3057 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3059 /* skip PA normalized request doesn't overlap with */
3060 if (pa->pa_lstart >= end) {
3061 spin_unlock(&pa->pa_lock);
3064 if (pa_end <= start) {
3065 spin_unlock(&pa->pa_lock);
3068 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3070 if (pa_end <= ac->ac_o_ex.fe_logical) {
3071 BUG_ON(pa_end < start);
3075 if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3076 BUG_ON(pa->pa_lstart > end);
3077 end = pa->pa_lstart;
3079 spin_unlock(&pa->pa_lock);
3084 /* XXX: extra loop to check we really don't overlap preallocations */
3086 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3087 unsigned long pa_end;
3088 spin_lock(&pa->pa_lock);
3089 if (pa->pa_deleted == 0) {
3090 pa_end = pa->pa_lstart + pa->pa_len;
3091 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3093 spin_unlock(&pa->pa_lock);
3097 if (start + size <= ac->ac_o_ex.fe_logical &&
3098 start > ac->ac_o_ex.fe_logical) {
3099 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3100 (unsigned long) start, (unsigned long) size,
3101 (unsigned long) ac->ac_o_ex.fe_logical);
3103 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3104 start > ac->ac_o_ex.fe_logical);
3105 BUG_ON(size <= 0 || size >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3107 /* now prepare goal request */
3109 /* XXX: is it better to align blocks WRT to logical
3110 * placement or satisfy big request as is */
3111 ac->ac_g_ex.fe_logical = start;
3112 ac->ac_g_ex.fe_len = size;
3114 /* define goal start in order to merge */
3115 if (ar->pright && (ar->lright == (start + size))) {
3116 /* merge to the right */
3117 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3118 &ac->ac_f_ex.fe_group,
3119 &ac->ac_f_ex.fe_start);
3120 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3122 if (ar->pleft && (ar->lleft + 1 == start)) {
3123 /* merge to the left */
3124 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3125 &ac->ac_f_ex.fe_group,
3126 &ac->ac_f_ex.fe_start);
3127 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3130 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
3131 (unsigned) orig_size, (unsigned) start);
3134 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3136 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3138 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3139 atomic_inc(&sbi->s_bal_reqs);
3140 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3141 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3142 atomic_inc(&sbi->s_bal_success);
3143 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3144 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3145 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3146 atomic_inc(&sbi->s_bal_goals);
3147 if (ac->ac_found > sbi->s_mb_max_to_scan)
3148 atomic_inc(&sbi->s_bal_breaks);
3151 ext4_mb_store_history(ac);
3155 * use blocks preallocated to inode
3157 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3158 struct ext4_prealloc_space *pa)
3164 /* found preallocated blocks, use them */
3165 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3166 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3168 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3169 &ac->ac_b_ex.fe_start);
3170 ac->ac_b_ex.fe_len = len;
3171 ac->ac_status = AC_STATUS_FOUND;
3174 BUG_ON(start < pa->pa_pstart);
3175 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3176 BUG_ON(pa->pa_free < len);
3179 mb_debug("use %llu/%u from inode pa %p\n", start, len, pa);
3183 * use blocks preallocated to locality group
3185 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3186 struct ext4_prealloc_space *pa)
3188 unsigned int len = ac->ac_o_ex.fe_len;
3190 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3191 &ac->ac_b_ex.fe_group,
3192 &ac->ac_b_ex.fe_start);
3193 ac->ac_b_ex.fe_len = len;
3194 ac->ac_status = AC_STATUS_FOUND;
3197 /* we don't correct pa_pstart or pa_plen here to avoid
3198 * possible race when the group is being loaded concurrently
3199 * instead we correct pa later, after blocks are marked
3200 * in on-disk bitmap -- see ext4_mb_release_context()
3201 * Other CPUs are prevented from allocating from this pa by lg_mutex
3203 mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3207 * Return the prealloc space that have minimal distance
3208 * from the goal block. @cpa is the prealloc
3209 * space that is having currently known minimal distance
3210 * from the goal block.
3212 static struct ext4_prealloc_space *
3213 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3214 struct ext4_prealloc_space *pa,
3215 struct ext4_prealloc_space *cpa)
3217 ext4_fsblk_t cur_distance, new_distance;
3220 atomic_inc(&pa->pa_count);
3223 cur_distance = abs(goal_block - cpa->pa_pstart);
3224 new_distance = abs(goal_block - pa->pa_pstart);
3226 if (cur_distance < new_distance)
3229 /* drop the previous reference */
3230 atomic_dec(&cpa->pa_count);
3231 atomic_inc(&pa->pa_count);
3236 * search goal blocks in preallocated space
3238 static noinline_for_stack int
3239 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3242 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3243 struct ext4_locality_group *lg;
3244 struct ext4_prealloc_space *pa, *cpa = NULL;
3245 ext4_fsblk_t goal_block;
3247 /* only data can be preallocated */
3248 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3251 /* first, try per-file preallocation */
3253 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3255 /* all fields in this condition don't change,
3256 * so we can skip locking for them */
3257 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3258 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3261 /* found preallocated blocks, use them */
3262 spin_lock(&pa->pa_lock);
3263 if (pa->pa_deleted == 0 && pa->pa_free) {
3264 atomic_inc(&pa->pa_count);
3265 ext4_mb_use_inode_pa(ac, pa);
3266 spin_unlock(&pa->pa_lock);
3267 ac->ac_criteria = 10;
3271 spin_unlock(&pa->pa_lock);
3275 /* can we use group allocation? */
3276 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3279 /* inode may have no locality group for some reason */
3283 order = fls(ac->ac_o_ex.fe_len) - 1;
3284 if (order > PREALLOC_TB_SIZE - 1)
3285 /* The max size of hash table is PREALLOC_TB_SIZE */
3286 order = PREALLOC_TB_SIZE - 1;
3288 goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) +
3289 ac->ac_g_ex.fe_start +
3290 le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block);
3292 * search for the prealloc space that is having
3293 * minimal distance from the goal block.
3295 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3297 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3299 spin_lock(&pa->pa_lock);
3300 if (pa->pa_deleted == 0 &&
3301 pa->pa_free >= ac->ac_o_ex.fe_len) {
3303 cpa = ext4_mb_check_group_pa(goal_block,
3306 spin_unlock(&pa->pa_lock);
3311 ext4_mb_use_group_pa(ac, cpa);
3312 ac->ac_criteria = 20;
3319 * the function goes through all preallocation in this group and marks them
3320 * used in in-core bitmap. buddy must be generated from this bitmap
3321 * Need to be called with ext4 group lock (ext4_lock_group)
3323 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3326 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3327 struct ext4_prealloc_space *pa;
3328 struct list_head *cur;
3329 ext4_group_t groupnr;
3330 ext4_grpblk_t start;
3331 int preallocated = 0;
3335 /* all form of preallocation discards first load group,
3336 * so the only competing code is preallocation use.
3337 * we don't need any locking here
3338 * notice we do NOT ignore preallocations with pa_deleted
3339 * otherwise we could leave used blocks available for
3340 * allocation in buddy when concurrent ext4_mb_put_pa()
3341 * is dropping preallocation
3343 list_for_each(cur, &grp->bb_prealloc_list) {
3344 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3345 spin_lock(&pa->pa_lock);
3346 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3349 spin_unlock(&pa->pa_lock);
3350 if (unlikely(len == 0))
3352 BUG_ON(groupnr != group);
3353 mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
3354 bitmap, start, len);
3355 preallocated += len;
3358 mb_debug("prellocated %u for group %lu\n", preallocated, group);
3361 static void ext4_mb_pa_callback(struct rcu_head *head)
3363 struct ext4_prealloc_space *pa;
3364 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3365 kmem_cache_free(ext4_pspace_cachep, pa);
3369 * drops a reference to preallocated space descriptor
3370 * if this was the last reference and the space is consumed
3372 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3373 struct super_block *sb, struct ext4_prealloc_space *pa)
3377 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3380 /* in this short window concurrent discard can set pa_deleted */
3381 spin_lock(&pa->pa_lock);
3382 if (pa->pa_deleted == 1) {
3383 spin_unlock(&pa->pa_lock);
3388 spin_unlock(&pa->pa_lock);
3390 /* -1 is to protect from crossing allocation group */
3391 ext4_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL);
3396 * P1 (buddy init) P2 (regular allocation)
3397 * find block B in PA
3398 * copy on-disk bitmap to buddy
3399 * mark B in on-disk bitmap
3400 * drop PA from group
3401 * mark all PAs in buddy
3403 * thus, P1 initializes buddy with B available. to prevent this
3404 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3407 ext4_lock_group(sb, grp);
3408 list_del(&pa->pa_group_list);
3409 ext4_unlock_group(sb, grp);
3411 spin_lock(pa->pa_obj_lock);
3412 list_del_rcu(&pa->pa_inode_list);
3413 spin_unlock(pa->pa_obj_lock);
3415 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3419 * creates new preallocated space for given inode
3421 static noinline_for_stack int
3422 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3424 struct super_block *sb = ac->ac_sb;
3425 struct ext4_prealloc_space *pa;
3426 struct ext4_group_info *grp;
3427 struct ext4_inode_info *ei;
3429 /* preallocate only when found space is larger then requested */
3430 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3431 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3432 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3434 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3438 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3444 /* we can't allocate as much as normalizer wants.
3445 * so, found space must get proper lstart
3446 * to cover original request */
3447 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3448 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3450 /* we're limited by original request in that
3451 * logical block must be covered any way
3452 * winl is window we can move our chunk within */
3453 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3455 /* also, we should cover whole original request */
3456 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3458 /* the smallest one defines real window */
3459 win = min(winl, wins);
3461 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3462 if (offs && offs < win)
3465 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3466 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3467 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3470 /* preallocation can change ac_b_ex, thus we store actually
3471 * allocated blocks for history */
3472 ac->ac_f_ex = ac->ac_b_ex;
3474 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3475 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3476 pa->pa_len = ac->ac_b_ex.fe_len;
3477 pa->pa_free = pa->pa_len;
3478 atomic_set(&pa->pa_count, 1);
3479 spin_lock_init(&pa->pa_lock);
3483 mb_debug("new inode pa %p: %llu/%u for %u\n", pa,
3484 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3486 ext4_mb_use_inode_pa(ac, pa);
3487 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3489 ei = EXT4_I(ac->ac_inode);
3490 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3492 pa->pa_obj_lock = &ei->i_prealloc_lock;
3493 pa->pa_inode = ac->ac_inode;
3495 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3496 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3497 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3499 spin_lock(pa->pa_obj_lock);
3500 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3501 spin_unlock(pa->pa_obj_lock);
3507 * creates new preallocated space for locality group inodes belongs to
3509 static noinline_for_stack int
3510 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3512 struct super_block *sb = ac->ac_sb;
3513 struct ext4_locality_group *lg;
3514 struct ext4_prealloc_space *pa;
3515 struct ext4_group_info *grp;
3517 /* preallocate only when found space is larger then requested */
3518 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3519 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3520 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3522 BUG_ON(ext4_pspace_cachep == NULL);
3523 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3527 /* preallocation can change ac_b_ex, thus we store actually
3528 * allocated blocks for history */
3529 ac->ac_f_ex = ac->ac_b_ex;
3531 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3532 pa->pa_lstart = pa->pa_pstart;
3533 pa->pa_len = ac->ac_b_ex.fe_len;
3534 pa->pa_free = pa->pa_len;
3535 atomic_set(&pa->pa_count, 1);
3536 spin_lock_init(&pa->pa_lock);
3537 INIT_LIST_HEAD(&pa->pa_inode_list);
3541 mb_debug("new group pa %p: %llu/%u for %u\n", pa,
3542 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3544 ext4_mb_use_group_pa(ac, pa);
3545 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3547 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3551 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3552 pa->pa_inode = NULL;
3554 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3555 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3556 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3559 * We will later add the new pa to the right bucket
3560 * after updating the pa_free in ext4_mb_release_context
3565 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3569 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3570 err = ext4_mb_new_group_pa(ac);
3572 err = ext4_mb_new_inode_pa(ac);
3577 * finds all unused blocks in on-disk bitmap, frees them in
3578 * in-core bitmap and buddy.
3579 * @pa must be unlinked from inode and group lists, so that
3580 * nobody else can find/use it.
3581 * the caller MUST hold group/inode locks.
3582 * TODO: optimize the case when there are no in-core structures yet
3584 static noinline_for_stack int
3585 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3586 struct ext4_prealloc_space *pa,
3587 struct ext4_allocation_context *ac)
3589 struct super_block *sb = e4b->bd_sb;
3590 struct ext4_sb_info *sbi = EXT4_SB(sb);
3599 BUG_ON(pa->pa_deleted == 0);
3600 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3601 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3602 end = bit + pa->pa_len;
3606 ac->ac_inode = pa->pa_inode;
3607 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3611 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3614 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3615 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3616 le32_to_cpu(sbi->s_es->s_first_data_block);
3617 mb_debug(" free preallocated %u/%u in group %u\n",
3618 (unsigned) start, (unsigned) next - bit,
3623 ac->ac_b_ex.fe_group = group;
3624 ac->ac_b_ex.fe_start = bit;
3625 ac->ac_b_ex.fe_len = next - bit;
3626 ac->ac_b_ex.fe_logical = 0;
3627 ext4_mb_store_history(ac);
3630 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3633 if (free != pa->pa_free) {
3634 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3635 pa, (unsigned long) pa->pa_lstart,
3636 (unsigned long) pa->pa_pstart,
3637 (unsigned long) pa->pa_len);
3638 ext4_error(sb, __func__, "free %u, pa_free %u\n",
3641 * pa is already deleted so we use the value obtained
3642 * from the bitmap and continue.
3645 atomic_add(free, &sbi->s_mb_discarded);
3650 static noinline_for_stack int
3651 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3652 struct ext4_prealloc_space *pa,
3653 struct ext4_allocation_context *ac)
3655 struct super_block *sb = e4b->bd_sb;
3660 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3662 BUG_ON(pa->pa_deleted == 0);
3663 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3664 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3665 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3666 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3670 ac->ac_inode = NULL;
3671 ac->ac_b_ex.fe_group = group;
3672 ac->ac_b_ex.fe_start = bit;
3673 ac->ac_b_ex.fe_len = pa->pa_len;
3674 ac->ac_b_ex.fe_logical = 0;
3675 ext4_mb_store_history(ac);
3682 * releases all preallocations in given group
3684 * first, we need to decide discard policy:
3685 * - when do we discard
3687 * - how many do we discard
3688 * 1) how many requested
3690 static noinline_for_stack int
3691 ext4_mb_discard_group_preallocations(struct super_block *sb,
3692 ext4_group_t group, int needed)
3694 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3695 struct buffer_head *bitmap_bh = NULL;
3696 struct ext4_prealloc_space *pa, *tmp;
3697 struct ext4_allocation_context *ac;
3698 struct list_head list;
3699 struct ext4_buddy e4b;
3704 mb_debug("discard preallocation for group %lu\n", group);
3706 if (list_empty(&grp->bb_prealloc_list))
3709 bitmap_bh = ext4_read_block_bitmap(sb, group);
3710 if (bitmap_bh == NULL) {
3711 ext4_error(sb, __func__, "Error in reading block "
3712 "bitmap for %lu\n", group);
3716 err = ext4_mb_load_buddy(sb, group, &e4b);
3718 ext4_error(sb, __func__, "Error in loading buddy "
3719 "information for %lu\n", group);
3725 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3727 INIT_LIST_HEAD(&list);
3728 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3730 ext4_lock_group(sb, group);
3731 list_for_each_entry_safe(pa, tmp,
3732 &grp->bb_prealloc_list, pa_group_list) {
3733 spin_lock(&pa->pa_lock);
3734 if (atomic_read(&pa->pa_count)) {
3735 spin_unlock(&pa->pa_lock);
3739 if (pa->pa_deleted) {
3740 spin_unlock(&pa->pa_lock);
3744 /* seems this one can be freed ... */
3747 /* we can trust pa_free ... */
3748 free += pa->pa_free;
3750 spin_unlock(&pa->pa_lock);
3752 list_del(&pa->pa_group_list);
3753 list_add(&pa->u.pa_tmp_list, &list);
3756 /* if we still need more blocks and some PAs were used, try again */
3757 if (free < needed && busy) {
3759 ext4_unlock_group(sb, group);
3761 * Yield the CPU here so that we don't get soft lockup
3762 * in non preempt case.
3768 /* found anything to free? */
3769 if (list_empty(&list)) {
3774 /* now free all selected PAs */
3775 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3777 /* remove from object (inode or locality group) */
3778 spin_lock(pa->pa_obj_lock);
3779 list_del_rcu(&pa->pa_inode_list);
3780 spin_unlock(pa->pa_obj_lock);
3783 ext4_mb_release_group_pa(&e4b, pa, ac);
3785 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3787 list_del(&pa->u.pa_tmp_list);
3788 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3792 ext4_unlock_group(sb, group);
3794 kmem_cache_free(ext4_ac_cachep, ac);
3795 ext4_mb_release_desc(&e4b);
3801 * releases all non-used preallocated blocks for given inode
3803 * It's important to discard preallocations under i_data_sem
3804 * We don't want another block to be served from the prealloc
3805 * space when we are discarding the inode prealloc space.
3807 * FIXME!! Make sure it is valid at all the call sites
3809 void ext4_discard_preallocations(struct inode *inode)
3811 struct ext4_inode_info *ei = EXT4_I(inode);
3812 struct super_block *sb = inode->i_sb;
3813 struct buffer_head *bitmap_bh = NULL;
3814 struct ext4_prealloc_space *pa, *tmp;
3815 struct ext4_allocation_context *ac;
3816 ext4_group_t group = 0;
3817 struct list_head list;
3818 struct ext4_buddy e4b;
3821 if (!S_ISREG(inode->i_mode)) {
3822 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3826 mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
3828 INIT_LIST_HEAD(&list);
3830 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3832 /* first, collect all pa's in the inode */
3833 spin_lock(&ei->i_prealloc_lock);
3834 while (!list_empty(&ei->i_prealloc_list)) {
3835 pa = list_entry(ei->i_prealloc_list.next,
3836 struct ext4_prealloc_space, pa_inode_list);
3837 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3838 spin_lock(&pa->pa_lock);
3839 if (atomic_read(&pa->pa_count)) {
3840 /* this shouldn't happen often - nobody should
3841 * use preallocation while we're discarding it */
3842 spin_unlock(&pa->pa_lock);
3843 spin_unlock(&ei->i_prealloc_lock);
3844 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3846 schedule_timeout_uninterruptible(HZ);
3850 if (pa->pa_deleted == 0) {
3852 spin_unlock(&pa->pa_lock);
3853 list_del_rcu(&pa->pa_inode_list);
3854 list_add(&pa->u.pa_tmp_list, &list);
3858 /* someone is deleting pa right now */
3859 spin_unlock(&pa->pa_lock);
3860 spin_unlock(&ei->i_prealloc_lock);
3862 /* we have to wait here because pa_deleted
3863 * doesn't mean pa is already unlinked from
3864 * the list. as we might be called from
3865 * ->clear_inode() the inode will get freed
3866 * and concurrent thread which is unlinking
3867 * pa from inode's list may access already
3868 * freed memory, bad-bad-bad */
3870 /* XXX: if this happens too often, we can
3871 * add a flag to force wait only in case
3872 * of ->clear_inode(), but not in case of
3873 * regular truncate */
3874 schedule_timeout_uninterruptible(HZ);
3877 spin_unlock(&ei->i_prealloc_lock);
3879 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3880 BUG_ON(pa->pa_linear != 0);
3881 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3883 err = ext4_mb_load_buddy(sb, group, &e4b);
3885 ext4_error(sb, __func__, "Error in loading buddy "
3886 "information for %lu\n", group);
3890 bitmap_bh = ext4_read_block_bitmap(sb, group);
3891 if (bitmap_bh == NULL) {
3892 ext4_error(sb, __func__, "Error in reading block "
3893 "bitmap for %lu\n", group);
3894 ext4_mb_release_desc(&e4b);
3898 ext4_lock_group(sb, group);
3899 list_del(&pa->pa_group_list);
3900 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3901 ext4_unlock_group(sb, group);
3903 ext4_mb_release_desc(&e4b);
3906 list_del(&pa->u.pa_tmp_list);
3907 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3910 kmem_cache_free(ext4_ac_cachep, ac);
3914 * finds all preallocated spaces and return blocks being freed to them
3915 * if preallocated space becomes full (no block is used from the space)
3916 * then the function frees space in buddy
3917 * XXX: at the moment, truncate (which is the only way to free blocks)
3918 * discards all preallocations
3920 static void ext4_mb_return_to_preallocation(struct inode *inode,
3921 struct ext4_buddy *e4b,
3922 sector_t block, int count)
3924 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
3927 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3929 struct super_block *sb = ac->ac_sb;
3932 printk(KERN_ERR "EXT4-fs: Can't allocate:"
3933 " Allocation context details:\n");
3934 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
3935 ac->ac_status, ac->ac_flags);
3936 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3937 "best %lu/%lu/%lu@%lu cr %d\n",
3938 (unsigned long)ac->ac_o_ex.fe_group,
3939 (unsigned long)ac->ac_o_ex.fe_start,
3940 (unsigned long)ac->ac_o_ex.fe_len,
3941 (unsigned long)ac->ac_o_ex.fe_logical,
3942 (unsigned long)ac->ac_g_ex.fe_group,
3943 (unsigned long)ac->ac_g_ex.fe_start,
3944 (unsigned long)ac->ac_g_ex.fe_len,
3945 (unsigned long)ac->ac_g_ex.fe_logical,
3946 (unsigned long)ac->ac_b_ex.fe_group,
3947 (unsigned long)ac->ac_b_ex.fe_start,
3948 (unsigned long)ac->ac_b_ex.fe_len,
3949 (unsigned long)ac->ac_b_ex.fe_logical,
3950 (int)ac->ac_criteria);
3951 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
3953 printk(KERN_ERR "EXT4-fs: groups: \n");
3954 for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
3955 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3956 struct ext4_prealloc_space *pa;
3957 ext4_grpblk_t start;
3958 struct list_head *cur;
3959 ext4_lock_group(sb, i);
3960 list_for_each(cur, &grp->bb_prealloc_list) {
3961 pa = list_entry(cur, struct ext4_prealloc_space,
3963 spin_lock(&pa->pa_lock);
3964 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3966 spin_unlock(&pa->pa_lock);
3967 printk(KERN_ERR "PA:%lu:%d:%u \n", i,
3970 ext4_unlock_group(sb, i);
3972 if (grp->bb_free == 0)
3974 printk(KERN_ERR "%lu: %d/%d \n",
3975 i, grp->bb_free, grp->bb_fragments);
3977 printk(KERN_ERR "\n");
3980 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3987 * We use locality group preallocation for small size file. The size of the
3988 * file is determined by the current size or the resulting size after
3989 * allocation which ever is larger
3991 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
3993 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3995 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3996 int bsbits = ac->ac_sb->s_blocksize_bits;
3999 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4002 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4003 isize = i_size_read(ac->ac_inode) >> bsbits;
4004 size = max(size, isize);
4006 /* don't use group allocation for large files */
4007 if (size >= sbi->s_mb_stream_request)
4010 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4013 BUG_ON(ac->ac_lg != NULL);
4015 * locality group prealloc space are per cpu. The reason for having
4016 * per cpu locality group is to reduce the contention between block
4017 * request from multiple CPUs.
4019 ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id());
4021 /* we're going to use group allocation */
4022 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4024 /* serialize all allocations in the group */
4025 mutex_lock(&ac->ac_lg->lg_mutex);
4028 static noinline_for_stack int
4029 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4030 struct ext4_allocation_request *ar)
4032 struct super_block *sb = ar->inode->i_sb;
4033 struct ext4_sb_info *sbi = EXT4_SB(sb);
4034 struct ext4_super_block *es = sbi->s_es;
4038 ext4_grpblk_t block;
4040 /* we can't allocate > group size */
4043 /* just a dirty hack to filter too big requests */
4044 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4045 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4047 /* start searching from the goal */
4049 if (goal < le32_to_cpu(es->s_first_data_block) ||
4050 goal >= ext4_blocks_count(es))
4051 goal = le32_to_cpu(es->s_first_data_block);
4052 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4054 /* set up allocation goals */
4055 ac->ac_b_ex.fe_logical = ar->logical;
4056 ac->ac_b_ex.fe_group = 0;
4057 ac->ac_b_ex.fe_start = 0;
4058 ac->ac_b_ex.fe_len = 0;
4059 ac->ac_status = AC_STATUS_CONTINUE;
4060 ac->ac_groups_scanned = 0;
4061 ac->ac_ex_scanned = 0;
4064 ac->ac_inode = ar->inode;
4065 ac->ac_o_ex.fe_logical = ar->logical;
4066 ac->ac_o_ex.fe_group = group;
4067 ac->ac_o_ex.fe_start = block;
4068 ac->ac_o_ex.fe_len = len;
4069 ac->ac_g_ex.fe_logical = ar->logical;
4070 ac->ac_g_ex.fe_group = group;
4071 ac->ac_g_ex.fe_start = block;
4072 ac->ac_g_ex.fe_len = len;
4073 ac->ac_f_ex.fe_len = 0;
4074 ac->ac_flags = ar->flags;
4076 ac->ac_criteria = 0;
4078 ac->ac_bitmap_page = NULL;
4079 ac->ac_buddy_page = NULL;
4082 /* we have to define context: we'll we work with a file or
4083 * locality group. this is a policy, actually */
4084 ext4_mb_group_or_file(ac);
4086 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4087 "left: %u/%u, right %u/%u to %swritable\n",
4088 (unsigned) ar->len, (unsigned) ar->logical,
4089 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4090 (unsigned) ar->lleft, (unsigned) ar->pleft,
4091 (unsigned) ar->lright, (unsigned) ar->pright,
4092 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4097 static noinline_for_stack void
4098 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4099 struct ext4_locality_group *lg,
4100 int order, int total_entries)
4102 ext4_group_t group = 0;
4103 struct ext4_buddy e4b;
4104 struct list_head discard_list;
4105 struct ext4_prealloc_space *pa, *tmp;
4106 struct ext4_allocation_context *ac;
4108 mb_debug("discard locality group preallocation\n");
4110 INIT_LIST_HEAD(&discard_list);
4111 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4113 spin_lock(&lg->lg_prealloc_lock);
4114 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4116 spin_lock(&pa->pa_lock);
4117 if (atomic_read(&pa->pa_count)) {
4119 * This is the pa that we just used
4120 * for block allocation. So don't
4123 spin_unlock(&pa->pa_lock);
4126 if (pa->pa_deleted) {
4127 spin_unlock(&pa->pa_lock);
4130 /* only lg prealloc space */
4131 BUG_ON(!pa->pa_linear);
4133 /* seems this one can be freed ... */
4135 spin_unlock(&pa->pa_lock);
4137 list_del_rcu(&pa->pa_inode_list);
4138 list_add(&pa->u.pa_tmp_list, &discard_list);
4141 if (total_entries <= 5) {
4143 * we want to keep only 5 entries
4144 * allowing it to grow to 8. This
4145 * mak sure we don't call discard
4146 * soon for this list.
4151 spin_unlock(&lg->lg_prealloc_lock);
4153 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4155 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4156 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4157 ext4_error(sb, __func__, "Error in loading buddy "
4158 "information for %lu\n", group);
4161 ext4_lock_group(sb, group);
4162 list_del(&pa->pa_group_list);
4163 ext4_mb_release_group_pa(&e4b, pa, ac);
4164 ext4_unlock_group(sb, group);
4166 ext4_mb_release_desc(&e4b);
4167 list_del(&pa->u.pa_tmp_list);
4168 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4171 kmem_cache_free(ext4_ac_cachep, ac);
4175 * We have incremented pa_count. So it cannot be freed at this
4176 * point. Also we hold lg_mutex. So no parallel allocation is
4177 * possible from this lg. That means pa_free cannot be updated.
4179 * A parallel ext4_mb_discard_group_preallocations is possible.
4180 * which can cause the lg_prealloc_list to be updated.
4183 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4185 int order, added = 0, lg_prealloc_count = 1;
4186 struct super_block *sb = ac->ac_sb;
4187 struct ext4_locality_group *lg = ac->ac_lg;
4188 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4190 order = fls(pa->pa_free) - 1;
4191 if (order > PREALLOC_TB_SIZE - 1)
4192 /* The max size of hash table is PREALLOC_TB_SIZE */
4193 order = PREALLOC_TB_SIZE - 1;
4194 /* Add the prealloc space to lg */
4196 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4198 spin_lock(&tmp_pa->pa_lock);
4199 if (tmp_pa->pa_deleted) {
4200 spin_unlock(&pa->pa_lock);
4203 if (!added && pa->pa_free < tmp_pa->pa_free) {
4204 /* Add to the tail of the previous entry */
4205 list_add_tail_rcu(&pa->pa_inode_list,
4206 &tmp_pa->pa_inode_list);
4209 * we want to count the total
4210 * number of entries in the list
4213 spin_unlock(&tmp_pa->pa_lock);
4214 lg_prealloc_count++;
4217 list_add_tail_rcu(&pa->pa_inode_list,
4218 &lg->lg_prealloc_list[order]);
4221 /* Now trim the list to be not more than 8 elements */
4222 if (lg_prealloc_count > 8) {
4223 ext4_mb_discard_lg_preallocations(sb, lg,
4224 order, lg_prealloc_count);
4231 * release all resource we used in allocation
4233 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4235 struct ext4_prealloc_space *pa = ac->ac_pa;
4237 if (pa->pa_linear) {
4238 /* see comment in ext4_mb_use_group_pa() */
4239 spin_lock(&pa->pa_lock);
4240 pa->pa_pstart += ac->ac_b_ex.fe_len;
4241 pa->pa_lstart += ac->ac_b_ex.fe_len;
4242 pa->pa_free -= ac->ac_b_ex.fe_len;
4243 pa->pa_len -= ac->ac_b_ex.fe_len;
4244 spin_unlock(&pa->pa_lock);
4246 * We want to add the pa to the right bucket.
4247 * Remove it from the list and while adding
4248 * make sure the list to which we are adding
4251 if (likely(pa->pa_free)) {
4252 spin_lock(pa->pa_obj_lock);
4253 list_del_rcu(&pa->pa_inode_list);
4254 spin_unlock(pa->pa_obj_lock);
4255 ext4_mb_add_n_trim(ac);
4258 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4260 if (ac->ac_bitmap_page)
4261 page_cache_release(ac->ac_bitmap_page);
4262 if (ac->ac_buddy_page)
4263 page_cache_release(ac->ac_buddy_page);
4264 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4265 mutex_unlock(&ac->ac_lg->lg_mutex);
4266 ext4_mb_collect_stats(ac);
4270 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4276 for (i = 0; i < EXT4_SB(sb)->s_groups_count && needed > 0; i++) {
4277 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4286 * Main entry point into mballoc to allocate blocks
4287 * it tries to use preallocation first, then falls back
4288 * to usual allocation
4290 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4291 struct ext4_allocation_request *ar, int *errp)
4294 struct ext4_allocation_context *ac = NULL;
4295 struct ext4_sb_info *sbi;
4296 struct super_block *sb;
4297 ext4_fsblk_t block = 0;
4298 unsigned long inquota;
4299 unsigned long reserv_blks = 0;
4301 sb = ar->inode->i_sb;
4304 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) {
4306 * With delalloc we already reserved the blocks
4308 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4309 /* let others to free the space */
4311 ar->len = ar->len >> 1;
4317 reserv_blks = ar->len;
4319 while (ar->len && DQUOT_ALLOC_BLOCK(ar->inode, ar->len)) {
4320 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4329 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4330 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4332 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4339 ext4_mb_poll_new_transaction(sb, handle);
4341 *errp = ext4_mb_initialize_context(ac, ar);
4347 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4348 if (!ext4_mb_use_preallocated(ac)) {
4349 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4350 ext4_mb_normalize_request(ac, ar);
4352 /* allocate space in core */
4353 ext4_mb_regular_allocator(ac);
4355 /* as we've just preallocated more space than
4356 * user requested orinally, we store allocated
4357 * space in a special descriptor */
4358 if (ac->ac_status == AC_STATUS_FOUND &&
4359 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4360 ext4_mb_new_preallocation(ac);
4363 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4364 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4365 if (*errp == -EAGAIN) {
4366 ac->ac_b_ex.fe_group = 0;
4367 ac->ac_b_ex.fe_start = 0;
4368 ac->ac_b_ex.fe_len = 0;
4369 ac->ac_status = AC_STATUS_CONTINUE;
4372 ac->ac_b_ex.fe_len = 0;
4374 ext4_mb_show_ac(ac);
4376 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4377 ar->len = ac->ac_b_ex.fe_len;
4380 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4384 ac->ac_b_ex.fe_len = 0;
4386 ext4_mb_show_ac(ac);
4389 ext4_mb_release_context(ac);
4392 kmem_cache_free(ext4_ac_cachep, ac);
4394 if (ar->len < inquota)
4395 DQUOT_FREE_BLOCK(ar->inode, inquota - ar->len);
4399 static void ext4_mb_poll_new_transaction(struct super_block *sb,
4402 struct ext4_sb_info *sbi = EXT4_SB(sb);
4404 if (sbi->s_last_transaction == handle->h_transaction->t_tid)
4407 /* new transaction! time to close last one and free blocks for
4408 * committed transaction. we know that only transaction can be
4409 * active, so previos transaction can be being logged and we
4410 * know that transaction before previous is known to be already
4411 * logged. this means that now we may free blocks freed in all
4412 * transactions before previous one. hope I'm clear enough ... */
4414 spin_lock(&sbi->s_md_lock);
4415 if (sbi->s_last_transaction != handle->h_transaction->t_tid) {
4416 mb_debug("new transaction %lu, old %lu\n",
4417 (unsigned long) handle->h_transaction->t_tid,
4418 (unsigned long) sbi->s_last_transaction);
4419 list_splice_init(&sbi->s_closed_transaction,
4420 &sbi->s_committed_transaction);
4421 list_splice_init(&sbi->s_active_transaction,
4422 &sbi->s_closed_transaction);
4423 sbi->s_last_transaction = handle->h_transaction->t_tid;
4425 spin_unlock(&sbi->s_md_lock);
4427 ext4_mb_free_committed_blocks(sb);
4431 * We can merge two free data extents only if the physical blocks
4432 * are contiguous, AND the extents were freed by the same transaction,
4433 * AND the blocks are associated with the same group.
4435 static int can_merge(struct ext4_free_data *entry1,
4436 struct ext4_free_data *entry2)
4438 if ((entry1->t_tid == entry2->t_tid) &&
4439 (entry1->group == entry2->group) &&
4440 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4445 static noinline_for_stack int
4446 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4447 ext4_group_t group, ext4_grpblk_t block, int count)
4449 struct ext4_group_info *db = e4b->bd_info;
4450 struct super_block *sb = e4b->bd_sb;
4451 struct ext4_sb_info *sbi = EXT4_SB(sb);
4452 struct ext4_free_data *entry, *new_entry;
4453 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4454 struct rb_node *parent = NULL, *new_node;
4457 BUG_ON(e4b->bd_bitmap_page == NULL);
4458 BUG_ON(e4b->bd_buddy_page == NULL);
4460 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4461 new_entry->start_blk = block;
4462 new_entry->group = group;
4463 new_entry->count = count;
4464 new_entry->t_tid = handle->h_transaction->t_tid;
4465 new_node = &new_entry->node;
4467 ext4_lock_group(sb, group);
4469 /* first free block exent. We need to
4470 protect buddy cache from being freed,
4471 * otherwise we'll refresh it from
4472 * on-disk bitmap and lose not-yet-available
4474 page_cache_get(e4b->bd_buddy_page);
4475 page_cache_get(e4b->bd_bitmap_page);
4479 entry = rb_entry(parent, struct ext4_free_data, node);
4480 if (block < entry->start_blk)
4482 else if (block >= (entry->start_blk + entry->count))
4483 n = &(*n)->rb_right;
4485 ext4_error(sb, __func__,
4486 "Double free of blocks %d (%d %d)\n",
4487 block, entry->start_blk, entry->count);
4492 rb_link_node(new_node, parent, n);
4493 rb_insert_color(new_node, &db->bb_free_root);
4495 /* Now try to see the extent can be merged to left and right */
4496 node = rb_prev(new_node);
4498 entry = rb_entry(node, struct ext4_free_data, node);
4499 if (can_merge(entry, new_entry)) {
4500 new_entry->start_blk = entry->start_blk;
4501 new_entry->count += entry->count;
4502 rb_erase(node, &(db->bb_free_root));
4503 spin_lock(&sbi->s_md_lock);
4504 list_del(&entry->list);
4505 spin_unlock(&sbi->s_md_lock);
4506 kmem_cache_free(ext4_free_ext_cachep, entry);
4510 node = rb_next(new_node);
4512 entry = rb_entry(node, struct ext4_free_data, node);
4513 if (can_merge(new_entry, entry)) {
4514 new_entry->count += entry->count;
4515 rb_erase(node, &(db->bb_free_root));
4516 spin_lock(&sbi->s_md_lock);
4517 list_del(&entry->list);
4518 spin_unlock(&sbi->s_md_lock);
4519 kmem_cache_free(ext4_free_ext_cachep, entry);
4522 /* Add the extent to active_transaction list */
4523 spin_lock(&sbi->s_md_lock);
4524 list_add(&new_entry->list, &sbi->s_active_transaction);
4525 spin_unlock(&sbi->s_md_lock);
4526 ext4_unlock_group(sb, group);
4531 * Main entry point into mballoc to free blocks
4533 void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
4534 unsigned long block, unsigned long count,
4535 int metadata, unsigned long *freed)
4537 struct buffer_head *bitmap_bh = NULL;
4538 struct super_block *sb = inode->i_sb;
4539 struct ext4_allocation_context *ac = NULL;
4540 struct ext4_group_desc *gdp;
4541 struct ext4_super_block *es;
4542 unsigned long overflow;
4544 struct buffer_head *gd_bh;
4545 ext4_group_t block_group;
4546 struct ext4_sb_info *sbi;
4547 struct ext4_buddy e4b;
4553 ext4_mb_poll_new_transaction(sb, handle);
4556 es = EXT4_SB(sb)->s_es;
4557 if (block < le32_to_cpu(es->s_first_data_block) ||
4558 block + count < block ||
4559 block + count > ext4_blocks_count(es)) {
4560 ext4_error(sb, __func__,
4561 "Freeing blocks not in datazone - "
4562 "block = %lu, count = %lu", block, count);
4566 ext4_debug("freeing block %lu\n", block);
4568 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4570 ac->ac_op = EXT4_MB_HISTORY_FREE;
4571 ac->ac_inode = inode;
4577 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4580 * Check to see if we are freeing blocks across a group
4583 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4584 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4587 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4592 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4598 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4599 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4600 in_range(block, ext4_inode_table(sb, gdp),
4601 EXT4_SB(sb)->s_itb_per_group) ||
4602 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4603 EXT4_SB(sb)->s_itb_per_group)) {
4605 ext4_error(sb, __func__,
4606 "Freeing blocks in system zone - "
4607 "Block = %lu, count = %lu", block, count);
4608 /* err = 0. ext4_std_error should be a no op */
4612 BUFFER_TRACE(bitmap_bh, "getting write access");
4613 err = ext4_journal_get_write_access(handle, bitmap_bh);
4618 * We are about to modify some metadata. Call the journal APIs
4619 * to unshare ->b_data if a currently-committing transaction is
4622 BUFFER_TRACE(gd_bh, "get_write_access");
4623 err = ext4_journal_get_write_access(handle, gd_bh);
4627 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4631 #ifdef AGGRESSIVE_CHECK
4634 for (i = 0; i < count; i++)
4635 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4638 mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data,
4641 /* We dirtied the bitmap block */
4642 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4643 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
4646 ac->ac_b_ex.fe_group = block_group;
4647 ac->ac_b_ex.fe_start = bit;
4648 ac->ac_b_ex.fe_len = count;
4649 ext4_mb_store_history(ac);
4653 /* blocks being freed are metadata. these blocks shouldn't
4654 * be used until this transaction is committed */
4655 ext4_mb_free_metadata(handle, &e4b, block_group, bit, count);
4657 ext4_lock_group(sb, block_group);
4658 mb_free_blocks(inode, &e4b, bit, count);
4659 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4660 ext4_unlock_group(sb, block_group);
4663 spin_lock(sb_bgl_lock(sbi, block_group));
4664 le16_add_cpu(&gdp->bg_free_blocks_count, count);
4665 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4666 spin_unlock(sb_bgl_lock(sbi, block_group));
4667 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4669 if (sbi->s_log_groups_per_flex) {
4670 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4671 spin_lock(sb_bgl_lock(sbi, flex_group));
4672 sbi->s_flex_groups[flex_group].free_blocks += count;
4673 spin_unlock(sb_bgl_lock(sbi, flex_group));
4676 ext4_mb_release_desc(&e4b);
4680 /* And the group descriptor block */
4681 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4682 ret = ext4_journal_dirty_metadata(handle, gd_bh);
4686 if (overflow && !err) {
4695 ext4_std_error(sb, err);
4697 kmem_cache_free(ext4_ac_cachep, ac);