2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
25 #include <linux/debugfs.h>
26 #include <trace/events/ext4.h>
30 * - test ext4_ext_search_left() and ext4_ext_search_right()
31 * - search for metadata in few groups
34 * - normalization should take into account whether file is still open
35 * - discard preallocations if no free space left (policy?)
36 * - don't normalize tails
38 * - reservation for superuser
41 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
42 * - track min/max extents in each group for better group selection
43 * - mb_mark_used() may allocate chunk right after splitting buddy
44 * - tree of groups sorted by number of free blocks
49 * The allocation request involve request for multiple number of blocks
50 * near to the goal(block) value specified.
52 * During initialization phase of the allocator we decide to use the
53 * group preallocation or inode preallocation depending on the size of
54 * the file. The size of the file could be the resulting file size we
55 * would have after allocation, or the current file size, which ever
56 * is larger. If the size is less than sbi->s_mb_stream_request we
57 * select to use the group preallocation. The default value of
58 * s_mb_stream_request is 16 blocks. This can also be tuned via
59 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
60 * terms of number of blocks.
62 * The main motivation for having small file use group preallocation is to
63 * ensure that we have small files closer together on the disk.
65 * First stage the allocator looks at the inode prealloc list,
66 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
67 * spaces for this particular inode. The inode prealloc space is
70 * pa_lstart -> the logical start block for this prealloc space
71 * pa_pstart -> the physical start block for this prealloc space
72 * pa_len -> lenght for this prealloc space
73 * pa_free -> free space available in this prealloc space
75 * The inode preallocation space is used looking at the _logical_ start
76 * block. If only the logical file block falls within the range of prealloc
77 * space we will consume the particular prealloc space. This make sure that
78 * that the we have contiguous physical blocks representing the file blocks
80 * The important thing to be noted in case of inode prealloc space is that
81 * we don't modify the values associated to inode prealloc space except
84 * If we are not able to find blocks in the inode prealloc space and if we
85 * have the group allocation flag set then we look at the locality group
86 * prealloc space. These are per CPU prealloc list repreasented as
88 * ext4_sb_info.s_locality_groups[smp_processor_id()]
90 * The reason for having a per cpu locality group is to reduce the contention
91 * between CPUs. It is possible to get scheduled at this point.
93 * The locality group prealloc space is used looking at whether we have
94 * enough free space (pa_free) withing the prealloc space.
96 * If we can't allocate blocks via inode prealloc or/and locality group
97 * prealloc then we look at the buddy cache. The buddy cache is represented
98 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
99 * mapped to the buddy and bitmap information regarding different
100 * groups. The buddy information is attached to buddy cache inode so that
101 * we can access them through the page cache. The information regarding
102 * each group is loaded via ext4_mb_load_buddy. The information involve
103 * block bitmap and buddy information. The information are stored in the
107 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
110 * one block each for bitmap and buddy information. So for each group we
111 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
112 * blocksize) blocks. So it can have information regarding groups_per_page
113 * which is blocks_per_page/2
115 * The buddy cache inode is not stored on disk. The inode is thrown
116 * away when the filesystem is unmounted.
118 * We look for count number of blocks in the buddy cache. If we were able
119 * to locate that many free blocks we return with additional information
120 * regarding rest of the contiguous physical block available
122 * Before allocating blocks via buddy cache we normalize the request
123 * blocks. This ensure we ask for more blocks that we needed. The extra
124 * blocks that we get after allocation is added to the respective prealloc
125 * list. In case of inode preallocation we follow a list of heuristics
126 * based on file size. This can be found in ext4_mb_normalize_request. If
127 * we are doing a group prealloc we try to normalize the request to
128 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
129 * 512 blocks. This can be tuned via
130 * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
131 * terms of number of blocks. If we have mounted the file system with -O
132 * stripe=<value> option the group prealloc request is normalized to the
133 * stripe value (sbi->s_stripe)
135 * The regular allocator(using the buddy cache) supports few tunables.
137 * /sys/fs/ext4/<partition>/mb_min_to_scan
138 * /sys/fs/ext4/<partition>/mb_max_to_scan
139 * /sys/fs/ext4/<partition>/mb_order2_req
141 * The regular allocator uses buddy scan only if the request len is power of
142 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
143 * value of s_mb_order2_reqs can be tuned via
144 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
145 * stripe size (sbi->s_stripe), we try to search for contigous block in
146 * stripe size. This should result in better allocation on RAID setups. If
147 * not, we search in the specific group using bitmap for best extents. The
148 * tunable min_to_scan and max_to_scan control the behaviour here.
149 * min_to_scan indicate how long the mballoc __must__ look for a best
150 * extent and max_to_scan indicates how long the mballoc __can__ look for a
151 * best extent in the found extents. Searching for the blocks starts with
152 * the group specified as the goal value in allocation context via
153 * ac_g_ex. Each group is first checked based on the criteria whether it
154 * can used for allocation. ext4_mb_good_group explains how the groups are
157 * Both the prealloc space are getting populated as above. So for the first
158 * request we will hit the buddy cache which will result in this prealloc
159 * space getting filled. The prealloc space is then later used for the
160 * subsequent request.
164 * mballoc operates on the following data:
166 * - in-core buddy (actually includes buddy and bitmap)
167 * - preallocation descriptors (PAs)
169 * there are two types of preallocations:
171 * assiged to specific inode and can be used for this inode only.
172 * it describes part of inode's space preallocated to specific
173 * physical blocks. any block from that preallocated can be used
174 * independent. the descriptor just tracks number of blocks left
175 * unused. so, before taking some block from descriptor, one must
176 * make sure corresponded logical block isn't allocated yet. this
177 * also means that freeing any block within descriptor's range
178 * must discard all preallocated blocks.
180 * assigned to specific locality group which does not translate to
181 * permanent set of inodes: inode can join and leave group. space
182 * from this type of preallocation can be used for any inode. thus
183 * it's consumed from the beginning to the end.
185 * relation between them can be expressed as:
186 * in-core buddy = on-disk bitmap + preallocation descriptors
188 * this mean blocks mballoc considers used are:
189 * - allocated blocks (persistent)
190 * - preallocated blocks (non-persistent)
192 * consistency in mballoc world means that at any time a block is either
193 * free or used in ALL structures. notice: "any time" should not be read
194 * literally -- time is discrete and delimited by locks.
196 * to keep it simple, we don't use block numbers, instead we count number of
197 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
199 * all operations can be expressed as:
200 * - init buddy: buddy = on-disk + PAs
201 * - new PA: buddy += N; PA = N
202 * - use inode PA: on-disk += N; PA -= N
203 * - discard inode PA buddy -= on-disk - PA; PA = 0
204 * - use locality group PA on-disk += N; PA -= N
205 * - discard locality group PA buddy -= PA; PA = 0
206 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
207 * is used in real operation because we can't know actual used
208 * bits from PA, only from on-disk bitmap
210 * if we follow this strict logic, then all operations above should be atomic.
211 * given some of them can block, we'd have to use something like semaphores
212 * killing performance on high-end SMP hardware. let's try to relax it using
213 * the following knowledge:
214 * 1) if buddy is referenced, it's already initialized
215 * 2) while block is used in buddy and the buddy is referenced,
216 * nobody can re-allocate that block
217 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
218 * bit set and PA claims same block, it's OK. IOW, one can set bit in
219 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
222 * so, now we're building a concurrency table:
225 * blocks for PA are allocated in the buddy, buddy must be referenced
226 * until PA is linked to allocation group to avoid concurrent buddy init
228 * we need to make sure that either on-disk bitmap or PA has uptodate data
229 * given (3) we care that PA-=N operation doesn't interfere with init
231 * the simplest way would be to have buddy initialized by the discard
232 * - use locality group PA
233 * again PA-=N must be serialized with init
234 * - discard locality group PA
235 * the simplest way would be to have buddy initialized by the discard
238 * i_data_sem serializes them
240 * discard process must wait until PA isn't used by another process
241 * - use locality group PA
242 * some mutex should serialize them
243 * - discard locality group PA
244 * discard process must wait until PA isn't used by another process
247 * i_data_sem or another mutex should serializes them
249 * discard process must wait until PA isn't used by another process
250 * - use locality group PA
251 * nothing wrong here -- they're different PAs covering different blocks
252 * - discard locality group PA
253 * discard process must wait until PA isn't used by another process
255 * now we're ready to make few consequences:
256 * - PA is referenced and while it is no discard is possible
257 * - PA is referenced until block isn't marked in on-disk bitmap
258 * - PA changes only after on-disk bitmap
259 * - discard must not compete with init. either init is done before
260 * any discard or they're serialized somehow
261 * - buddy init as sum of on-disk bitmap and PAs is done atomically
263 * a special case when we've used PA to emptiness. no need to modify buddy
264 * in this case, but we should care about concurrent init
269 * Logic in few words:
274 * mark bits in on-disk bitmap
277 * - use preallocation:
278 * find proper PA (per-inode or group)
280 * mark bits in on-disk bitmap
286 * mark bits in on-disk bitmap
289 * - discard preallocations in group:
291 * move them onto local list
292 * load on-disk bitmap
294 * remove PA from object (inode or locality group)
295 * mark free blocks in-core
297 * - discard inode's preallocations:
304 * - bitlock on a group (group)
305 * - object (inode/locality) (object)
316 * - release consumed pa:
321 * - generate in-core bitmap:
325 * - discard all for given object (inode, locality group):
330 * - discard all for given group:
337 static struct kmem_cache *ext4_pspace_cachep;
338 static struct kmem_cache *ext4_ac_cachep;
339 static struct kmem_cache *ext4_free_ext_cachep;
340 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
342 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
344 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
346 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
348 #if BITS_PER_LONG == 64
349 *bit += ((unsigned long) addr & 7UL) << 3;
350 addr = (void *) ((unsigned long) addr & ~7UL);
351 #elif BITS_PER_LONG == 32
352 *bit += ((unsigned long) addr & 3UL) << 3;
353 addr = (void *) ((unsigned long) addr & ~3UL);
355 #error "how many bits you are?!"
360 static inline int mb_test_bit(int bit, void *addr)
363 * ext4_test_bit on architecture like powerpc
364 * needs unsigned long aligned address
366 addr = mb_correct_addr_and_bit(&bit, addr);
367 return ext4_test_bit(bit, addr);
370 static inline void mb_set_bit(int bit, void *addr)
372 addr = mb_correct_addr_and_bit(&bit, addr);
373 ext4_set_bit(bit, addr);
376 static inline void mb_clear_bit(int bit, void *addr)
378 addr = mb_correct_addr_and_bit(&bit, addr);
379 ext4_clear_bit(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 assert_spin_locked(ext4_group_lock_ptr(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);
448 ext4_grp_locked_error(sb, e4b->bd_group,
449 __func__, "double-free of inode"
450 " %lu's block %llu(bit %u in group %u)",
451 inode ? inode->i_ino : 0, blocknr,
452 first + i, e4b->bd_group);
454 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
458 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
462 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
464 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
465 for (i = 0; i < count; i++) {
466 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
467 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
471 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
473 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
474 unsigned char *b1, *b2;
476 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
477 b2 = (unsigned char *) bitmap;
478 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
479 if (b1[i] != b2[i]) {
480 printk(KERN_ERR "corruption in group %u "
481 "at byte %u(%u): %x in copy != %x "
482 "on disk/prealloc\n",
483 e4b->bd_group, i, i * 8, b1[i], b2[i]);
491 static inline void mb_free_blocks_double(struct inode *inode,
492 struct ext4_buddy *e4b, int first, int count)
496 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
497 int first, int count)
501 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
507 #ifdef AGGRESSIVE_CHECK
509 #define MB_CHECK_ASSERT(assert) \
513 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
514 function, file, line, # assert); \
519 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
520 const char *function, int line)
522 struct super_block *sb = e4b->bd_sb;
523 int order = e4b->bd_blkbits + 1;
530 struct ext4_group_info *grp;
533 struct list_head *cur;
538 static int mb_check_counter;
539 if (mb_check_counter++ % 100 != 0)
544 buddy = mb_find_buddy(e4b, order, &max);
545 MB_CHECK_ASSERT(buddy);
546 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
547 MB_CHECK_ASSERT(buddy2);
548 MB_CHECK_ASSERT(buddy != buddy2);
549 MB_CHECK_ASSERT(max * 2 == max2);
552 for (i = 0; i < max; i++) {
554 if (mb_test_bit(i, buddy)) {
555 /* only single bit in buddy2 may be 1 */
556 if (!mb_test_bit(i << 1, buddy2)) {
558 mb_test_bit((i<<1)+1, buddy2));
559 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
561 mb_test_bit(i << 1, buddy2));
566 /* both bits in buddy2 must be 0 */
567 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
568 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
570 for (j = 0; j < (1 << order); j++) {
571 k = (i * (1 << order)) + j;
573 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
577 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
582 buddy = mb_find_buddy(e4b, 0, &max);
583 for (i = 0; i < max; i++) {
584 if (!mb_test_bit(i, buddy)) {
585 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
593 /* check used bits only */
594 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
595 buddy2 = mb_find_buddy(e4b, j, &max2);
597 MB_CHECK_ASSERT(k < max2);
598 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
601 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
602 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
604 grp = ext4_get_group_info(sb, e4b->bd_group);
605 buddy = mb_find_buddy(e4b, 0, &max);
606 list_for_each(cur, &grp->bb_prealloc_list) {
607 ext4_group_t groupnr;
608 struct ext4_prealloc_space *pa;
609 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
610 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
611 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
612 for (i = 0; i < pa->pa_len; i++)
613 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
617 #undef MB_CHECK_ASSERT
618 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
619 __FILE__, __func__, __LINE__)
621 #define mb_check_buddy(e4b)
624 /* FIXME!! need more doc */
625 static void ext4_mb_mark_free_simple(struct super_block *sb,
626 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
627 struct ext4_group_info *grp)
629 struct ext4_sb_info *sbi = EXT4_SB(sb);
633 unsigned short border;
635 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
637 border = 2 << sb->s_blocksize_bits;
640 /* find how many blocks can be covered since this position */
641 max = ffs(first | border) - 1;
643 /* find how many blocks of power 2 we need to mark */
650 /* mark multiblock chunks only */
651 grp->bb_counters[min]++;
653 mb_clear_bit(first >> min,
654 buddy + sbi->s_mb_offsets[min]);
661 static noinline_for_stack
662 void ext4_mb_generate_buddy(struct super_block *sb,
663 void *buddy, void *bitmap, ext4_group_t group)
665 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
666 ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb);
671 unsigned fragments = 0;
672 unsigned long long period = get_cycles();
674 /* initialize buddy from bitmap which is aggregation
675 * of on-disk bitmap and preallocations */
676 i = mb_find_next_zero_bit(bitmap, max, 0);
677 grp->bb_first_free = i;
681 i = mb_find_next_bit(bitmap, max, i);
685 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
687 grp->bb_counters[0]++;
689 i = mb_find_next_zero_bit(bitmap, max, i);
691 grp->bb_fragments = fragments;
693 if (free != grp->bb_free) {
694 ext4_grp_locked_error(sb, group, __func__,
695 "EXT4-fs: group %u: %u blocks in bitmap, %u in gd",
696 group, free, grp->bb_free);
698 * If we intent to continue, we consider group descritor
699 * corrupt and update bb_free using bitmap value
704 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
706 period = get_cycles() - period;
707 spin_lock(&EXT4_SB(sb)->s_bal_lock);
708 EXT4_SB(sb)->s_mb_buddies_generated++;
709 EXT4_SB(sb)->s_mb_generation_time += period;
710 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
713 /* The buddy information is attached the buddy cache inode
714 * for convenience. The information regarding each group
715 * is loaded via ext4_mb_load_buddy. The information involve
716 * block bitmap and buddy information. The information are
717 * stored in the inode as
720 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
723 * one block each for bitmap and buddy information.
724 * So for each group we take up 2 blocks. A page can
725 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
726 * So it can have information regarding groups_per_page which
727 * is blocks_per_page/2
730 static int ext4_mb_init_cache(struct page *page, char *incore)
732 ext4_group_t ngroups;
738 ext4_group_t first_group;
740 struct super_block *sb;
741 struct buffer_head *bhs;
742 struct buffer_head **bh;
747 mb_debug(1, "init page %lu\n", page->index);
749 inode = page->mapping->host;
751 ngroups = ext4_get_groups_count(sb);
752 blocksize = 1 << inode->i_blkbits;
753 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
755 groups_per_page = blocks_per_page >> 1;
756 if (groups_per_page == 0)
759 /* allocate buffer_heads to read bitmaps */
760 if (groups_per_page > 1) {
762 i = sizeof(struct buffer_head *) * groups_per_page;
763 bh = kzalloc(i, GFP_NOFS);
769 first_group = page->index * blocks_per_page / 2;
771 /* read all groups the page covers into the cache */
772 for (i = 0; i < groups_per_page; i++) {
773 struct ext4_group_desc *desc;
775 if (first_group + i >= ngroups)
779 desc = ext4_get_group_desc(sb, first_group + i, NULL);
784 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
788 if (bitmap_uptodate(bh[i]))
792 if (bitmap_uptodate(bh[i])) {
793 unlock_buffer(bh[i]);
796 ext4_lock_group(sb, first_group + i);
797 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
798 ext4_init_block_bitmap(sb, bh[i],
799 first_group + i, desc);
800 set_bitmap_uptodate(bh[i]);
801 set_buffer_uptodate(bh[i]);
802 ext4_unlock_group(sb, first_group + i);
803 unlock_buffer(bh[i]);
806 ext4_unlock_group(sb, first_group + i);
807 if (buffer_uptodate(bh[i])) {
809 * if not uninit if bh is uptodate,
810 * bitmap is also uptodate
812 set_bitmap_uptodate(bh[i]);
813 unlock_buffer(bh[i]);
818 * submit the buffer_head for read. We can
819 * safely mark the bitmap as uptodate now.
820 * We do it here so the bitmap uptodate bit
821 * get set with buffer lock held.
823 set_bitmap_uptodate(bh[i]);
824 bh[i]->b_end_io = end_buffer_read_sync;
825 submit_bh(READ, bh[i]);
826 mb_debug(1, "read bitmap for group %u\n", first_group + i);
829 /* wait for I/O completion */
830 for (i = 0; i < groups_per_page && bh[i]; i++)
831 wait_on_buffer(bh[i]);
834 for (i = 0; i < groups_per_page && bh[i]; i++)
835 if (!buffer_uptodate(bh[i]))
839 first_block = page->index * blocks_per_page;
841 memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
842 for (i = 0; i < blocks_per_page; i++) {
844 struct ext4_group_info *grinfo;
846 group = (first_block + i) >> 1;
847 if (group >= ngroups)
851 * data carry information regarding this
852 * particular group in the format specified
856 data = page_address(page) + (i * blocksize);
857 bitmap = bh[group - first_group]->b_data;
860 * We place the buddy block and bitmap block
863 if ((first_block + i) & 1) {
864 /* this is block of buddy */
865 BUG_ON(incore == NULL);
866 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
867 group, page->index, i * blocksize);
868 grinfo = ext4_get_group_info(sb, group);
869 grinfo->bb_fragments = 0;
870 memset(grinfo->bb_counters, 0,
871 sizeof(*grinfo->bb_counters) *
872 (sb->s_blocksize_bits+2));
874 * incore got set to the group block bitmap below
876 ext4_lock_group(sb, group);
877 ext4_mb_generate_buddy(sb, data, incore, group);
878 ext4_unlock_group(sb, group);
881 /* this is block of bitmap */
882 BUG_ON(incore != NULL);
883 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
884 group, page->index, i * blocksize);
886 /* see comments in ext4_mb_put_pa() */
887 ext4_lock_group(sb, group);
888 memcpy(data, bitmap, blocksize);
890 /* mark all preallocated blks used in in-core bitmap */
891 ext4_mb_generate_from_pa(sb, data, group);
892 ext4_mb_generate_from_freelist(sb, data, group);
893 ext4_unlock_group(sb, group);
895 /* set incore so that the buddy information can be
896 * generated using this
901 SetPageUptodate(page);
905 for (i = 0; i < groups_per_page && bh[i]; i++)
913 static noinline_for_stack
914 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
920 int block, pnum, poff;
921 int num_grp_locked = 0;
922 struct ext4_group_info *this_grp;
923 struct ext4_sb_info *sbi = EXT4_SB(sb);
924 struct inode *inode = sbi->s_buddy_cache;
925 struct page *page = NULL, *bitmap_page = NULL;
927 mb_debug(1, "init group %u\n", group);
928 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
929 this_grp = ext4_get_group_info(sb, group);
931 * This ensures that we don't reinit the buddy cache
932 * page which map to the group from which we are already
933 * allocating. If we are looking at the buddy cache we would
934 * have taken a reference using ext4_mb_load_buddy and that
935 * would have taken the alloc_sem lock.
937 num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group);
938 if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
940 * somebody initialized the group
941 * return without doing anything
947 * the buddy cache inode stores the block bitmap
948 * and buddy information in consecutive blocks.
949 * So for each group we need two blocks.
952 pnum = block / blocks_per_page;
953 poff = block % blocks_per_page;
954 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
956 BUG_ON(page->mapping != inode->i_mapping);
957 ret = ext4_mb_init_cache(page, NULL);
964 if (page == NULL || !PageUptodate(page)) {
968 mark_page_accessed(page);
970 bitmap = page_address(page) + (poff * sb->s_blocksize);
972 /* init buddy cache */
974 pnum = block / blocks_per_page;
975 poff = block % blocks_per_page;
976 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
977 if (page == bitmap_page) {
979 * If both the bitmap and buddy are in
980 * the same page we don't need to force
985 BUG_ON(page->mapping != inode->i_mapping);
986 ret = ext4_mb_init_cache(page, bitmap);
993 if (page == NULL || !PageUptodate(page)) {
997 mark_page_accessed(page);
999 ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
1001 page_cache_release(bitmap_page);
1003 page_cache_release(page);
1007 static noinline_for_stack int
1008 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1009 struct ext4_buddy *e4b)
1011 int blocks_per_page;
1017 struct ext4_group_info *grp;
1018 struct ext4_sb_info *sbi = EXT4_SB(sb);
1019 struct inode *inode = sbi->s_buddy_cache;
1021 mb_debug(1, "load group %u\n", group);
1023 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1024 grp = ext4_get_group_info(sb, group);
1026 e4b->bd_blkbits = sb->s_blocksize_bits;
1027 e4b->bd_info = ext4_get_group_info(sb, group);
1029 e4b->bd_group = group;
1030 e4b->bd_buddy_page = NULL;
1031 e4b->bd_bitmap_page = NULL;
1032 e4b->alloc_semp = &grp->alloc_sem;
1034 /* Take the read lock on the group alloc
1035 * sem. This would make sure a parallel
1036 * ext4_mb_init_group happening on other
1037 * groups mapped by the page is blocked
1038 * till we are done with allocation
1041 down_read(e4b->alloc_semp);
1043 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1044 /* we need to check for group need init flag
1045 * with alloc_semp held so that we can be sure
1046 * that new blocks didn't get added to the group
1047 * when we are loading the buddy cache
1049 up_read(e4b->alloc_semp);
1051 * we need full data about the group
1052 * to make a good selection
1054 ret = ext4_mb_init_group(sb, group);
1057 goto repeat_load_buddy;
1061 * the buddy cache inode stores the block bitmap
1062 * and buddy information in consecutive blocks.
1063 * So for each group we need two blocks.
1066 pnum = block / blocks_per_page;
1067 poff = block % blocks_per_page;
1069 /* we could use find_or_create_page(), but it locks page
1070 * what we'd like to avoid in fast path ... */
1071 page = find_get_page(inode->i_mapping, pnum);
1072 if (page == NULL || !PageUptodate(page)) {
1075 * drop the page reference and try
1076 * to get the page with lock. If we
1077 * are not uptodate that implies
1078 * somebody just created the page but
1079 * is yet to initialize the same. So
1080 * wait for it to initialize.
1082 page_cache_release(page);
1083 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1085 BUG_ON(page->mapping != inode->i_mapping);
1086 if (!PageUptodate(page)) {
1087 ret = ext4_mb_init_cache(page, NULL);
1092 mb_cmp_bitmaps(e4b, page_address(page) +
1093 (poff * sb->s_blocksize));
1098 if (page == NULL || !PageUptodate(page)) {
1102 e4b->bd_bitmap_page = page;
1103 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1104 mark_page_accessed(page);
1107 pnum = block / blocks_per_page;
1108 poff = block % blocks_per_page;
1110 page = find_get_page(inode->i_mapping, pnum);
1111 if (page == NULL || !PageUptodate(page)) {
1113 page_cache_release(page);
1114 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1116 BUG_ON(page->mapping != inode->i_mapping);
1117 if (!PageUptodate(page)) {
1118 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1127 if (page == NULL || !PageUptodate(page)) {
1131 e4b->bd_buddy_page = page;
1132 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1133 mark_page_accessed(page);
1135 BUG_ON(e4b->bd_bitmap_page == NULL);
1136 BUG_ON(e4b->bd_buddy_page == NULL);
1141 if (e4b->bd_bitmap_page)
1142 page_cache_release(e4b->bd_bitmap_page);
1143 if (e4b->bd_buddy_page)
1144 page_cache_release(e4b->bd_buddy_page);
1145 e4b->bd_buddy = NULL;
1146 e4b->bd_bitmap = NULL;
1148 /* Done with the buddy cache */
1149 up_read(e4b->alloc_semp);
1153 static void ext4_mb_release_desc(struct ext4_buddy *e4b)
1155 if (e4b->bd_bitmap_page)
1156 page_cache_release(e4b->bd_bitmap_page);
1157 if (e4b->bd_buddy_page)
1158 page_cache_release(e4b->bd_buddy_page);
1159 /* Done with the buddy cache */
1160 if (e4b->alloc_semp)
1161 up_read(e4b->alloc_semp);
1165 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1170 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1171 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1173 bb = EXT4_MB_BUDDY(e4b);
1174 while (order <= e4b->bd_blkbits + 1) {
1176 if (!mb_test_bit(block, bb)) {
1177 /* this block is part of buddy of order 'order' */
1180 bb += 1 << (e4b->bd_blkbits - order);
1186 static void mb_clear_bits(void *bm, int cur, int len)
1192 if ((cur & 31) == 0 && (len - cur) >= 32) {
1193 /* fast path: clear whole word at once */
1194 addr = bm + (cur >> 3);
1199 mb_clear_bit(cur, bm);
1204 static void mb_set_bits(void *bm, int cur, int len)
1210 if ((cur & 31) == 0 && (len - cur) >= 32) {
1211 /* fast path: set whole word at once */
1212 addr = bm + (cur >> 3);
1217 mb_set_bit(cur, bm);
1222 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1223 int first, int count)
1230 struct super_block *sb = e4b->bd_sb;
1232 BUG_ON(first + count > (sb->s_blocksize << 3));
1233 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1234 mb_check_buddy(e4b);
1235 mb_free_blocks_double(inode, e4b, first, count);
1237 e4b->bd_info->bb_free += count;
1238 if (first < e4b->bd_info->bb_first_free)
1239 e4b->bd_info->bb_first_free = first;
1241 /* let's maintain fragments counter */
1243 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1244 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1245 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1247 e4b->bd_info->bb_fragments--;
1248 else if (!block && !max)
1249 e4b->bd_info->bb_fragments++;
1251 /* let's maintain buddy itself */
1252 while (count-- > 0) {
1256 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1257 ext4_fsblk_t blocknr;
1258 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1261 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
1262 ext4_grp_locked_error(sb, e4b->bd_group,
1263 __func__, "double-free of inode"
1264 " %lu's block %llu(bit %u in group %u)",
1265 inode ? inode->i_ino : 0, blocknr, block,
1268 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1269 e4b->bd_info->bb_counters[order]++;
1271 /* start of the buddy */
1272 buddy = mb_find_buddy(e4b, order, &max);
1276 if (mb_test_bit(block, buddy) ||
1277 mb_test_bit(block + 1, buddy))
1280 /* both the buddies are free, try to coalesce them */
1281 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1287 /* for special purposes, we don't set
1288 * free bits in bitmap */
1289 mb_set_bit(block, buddy);
1290 mb_set_bit(block + 1, buddy);
1292 e4b->bd_info->bb_counters[order]--;
1293 e4b->bd_info->bb_counters[order]--;
1297 e4b->bd_info->bb_counters[order]++;
1299 mb_clear_bit(block, buddy2);
1303 mb_check_buddy(e4b);
1306 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1307 int needed, struct ext4_free_extent *ex)
1314 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1317 buddy = mb_find_buddy(e4b, order, &max);
1318 BUG_ON(buddy == NULL);
1319 BUG_ON(block >= max);
1320 if (mb_test_bit(block, buddy)) {
1327 /* FIXME dorp order completely ? */
1328 if (likely(order == 0)) {
1329 /* find actual order */
1330 order = mb_find_order_for_block(e4b, block);
1331 block = block >> order;
1334 ex->fe_len = 1 << order;
1335 ex->fe_start = block << order;
1336 ex->fe_group = e4b->bd_group;
1338 /* calc difference from given start */
1339 next = next - ex->fe_start;
1341 ex->fe_start += next;
1343 while (needed > ex->fe_len &&
1344 (buddy = mb_find_buddy(e4b, order, &max))) {
1346 if (block + 1 >= max)
1349 next = (block + 1) * (1 << order);
1350 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1353 ord = mb_find_order_for_block(e4b, next);
1356 block = next >> order;
1357 ex->fe_len += 1 << order;
1360 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1364 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1370 int start = ex->fe_start;
1371 int len = ex->fe_len;
1376 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1377 BUG_ON(e4b->bd_group != ex->fe_group);
1378 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1379 mb_check_buddy(e4b);
1380 mb_mark_used_double(e4b, start, len);
1382 e4b->bd_info->bb_free -= len;
1383 if (e4b->bd_info->bb_first_free == start)
1384 e4b->bd_info->bb_first_free += len;
1386 /* let's maintain fragments counter */
1388 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1389 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1390 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1392 e4b->bd_info->bb_fragments++;
1393 else if (!mlen && !max)
1394 e4b->bd_info->bb_fragments--;
1396 /* let's maintain buddy itself */
1398 ord = mb_find_order_for_block(e4b, start);
1400 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1401 /* the whole chunk may be allocated at once! */
1403 buddy = mb_find_buddy(e4b, ord, &max);
1404 BUG_ON((start >> ord) >= max);
1405 mb_set_bit(start >> ord, buddy);
1406 e4b->bd_info->bb_counters[ord]--;
1413 /* store for history */
1415 ret = len | (ord << 16);
1417 /* we have to split large buddy */
1419 buddy = mb_find_buddy(e4b, ord, &max);
1420 mb_set_bit(start >> ord, buddy);
1421 e4b->bd_info->bb_counters[ord]--;
1424 cur = (start >> ord) & ~1U;
1425 buddy = mb_find_buddy(e4b, ord, &max);
1426 mb_clear_bit(cur, buddy);
1427 mb_clear_bit(cur + 1, buddy);
1428 e4b->bd_info->bb_counters[ord]++;
1429 e4b->bd_info->bb_counters[ord]++;
1432 mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1433 mb_check_buddy(e4b);
1439 * Must be called under group lock!
1441 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1442 struct ext4_buddy *e4b)
1444 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1447 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1448 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1450 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1451 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1452 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1454 /* preallocation can change ac_b_ex, thus we store actually
1455 * allocated blocks for history */
1456 ac->ac_f_ex = ac->ac_b_ex;
1458 ac->ac_status = AC_STATUS_FOUND;
1459 ac->ac_tail = ret & 0xffff;
1460 ac->ac_buddy = ret >> 16;
1463 * take the page reference. We want the page to be pinned
1464 * so that we don't get a ext4_mb_init_cache_call for this
1465 * group until we update the bitmap. That would mean we
1466 * double allocate blocks. The reference is dropped
1467 * in ext4_mb_release_context
1469 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1470 get_page(ac->ac_bitmap_page);
1471 ac->ac_buddy_page = e4b->bd_buddy_page;
1472 get_page(ac->ac_buddy_page);
1473 /* on allocation we use ac to track the held semaphore */
1474 ac->alloc_semp = e4b->alloc_semp;
1475 e4b->alloc_semp = NULL;
1476 /* store last allocated for subsequent stream allocation */
1477 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1478 spin_lock(&sbi->s_md_lock);
1479 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1480 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1481 spin_unlock(&sbi->s_md_lock);
1486 * regular allocator, for general purposes allocation
1489 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1490 struct ext4_buddy *e4b,
1493 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1494 struct ext4_free_extent *bex = &ac->ac_b_ex;
1495 struct ext4_free_extent *gex = &ac->ac_g_ex;
1496 struct ext4_free_extent ex;
1499 if (ac->ac_status == AC_STATUS_FOUND)
1502 * We don't want to scan for a whole year
1504 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1505 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1506 ac->ac_status = AC_STATUS_BREAK;
1511 * Haven't found good chunk so far, let's continue
1513 if (bex->fe_len < gex->fe_len)
1516 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1517 && bex->fe_group == e4b->bd_group) {
1518 /* recheck chunk's availability - we don't know
1519 * when it was found (within this lock-unlock
1521 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1522 if (max >= gex->fe_len) {
1523 ext4_mb_use_best_found(ac, e4b);
1530 * The routine checks whether found extent is good enough. If it is,
1531 * then the extent gets marked used and flag is set to the context
1532 * to stop scanning. Otherwise, the extent is compared with the
1533 * previous found extent and if new one is better, then it's stored
1534 * in the context. Later, the best found extent will be used, if
1535 * mballoc can't find good enough extent.
1537 * FIXME: real allocation policy is to be designed yet!
1539 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1540 struct ext4_free_extent *ex,
1541 struct ext4_buddy *e4b)
1543 struct ext4_free_extent *bex = &ac->ac_b_ex;
1544 struct ext4_free_extent *gex = &ac->ac_g_ex;
1546 BUG_ON(ex->fe_len <= 0);
1547 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1548 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1549 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1554 * The special case - take what you catch first
1556 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1558 ext4_mb_use_best_found(ac, e4b);
1563 * Let's check whether the chuck is good enough
1565 if (ex->fe_len == gex->fe_len) {
1567 ext4_mb_use_best_found(ac, e4b);
1572 * If this is first found extent, just store it in the context
1574 if (bex->fe_len == 0) {
1580 * If new found extent is better, store it in the context
1582 if (bex->fe_len < gex->fe_len) {
1583 /* if the request isn't satisfied, any found extent
1584 * larger than previous best one is better */
1585 if (ex->fe_len > bex->fe_len)
1587 } else if (ex->fe_len > gex->fe_len) {
1588 /* if the request is satisfied, then we try to find
1589 * an extent that still satisfy the request, but is
1590 * smaller than previous one */
1591 if (ex->fe_len < bex->fe_len)
1595 ext4_mb_check_limits(ac, e4b, 0);
1598 static noinline_for_stack
1599 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1600 struct ext4_buddy *e4b)
1602 struct ext4_free_extent ex = ac->ac_b_ex;
1603 ext4_group_t group = ex.fe_group;
1607 BUG_ON(ex.fe_len <= 0);
1608 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1612 ext4_lock_group(ac->ac_sb, group);
1613 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1617 ext4_mb_use_best_found(ac, e4b);
1620 ext4_unlock_group(ac->ac_sb, group);
1621 ext4_mb_release_desc(e4b);
1626 static noinline_for_stack
1627 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1628 struct ext4_buddy *e4b)
1630 ext4_group_t group = ac->ac_g_ex.fe_group;
1633 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1634 struct ext4_super_block *es = sbi->s_es;
1635 struct ext4_free_extent ex;
1637 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1640 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1644 ext4_lock_group(ac->ac_sb, group);
1645 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1646 ac->ac_g_ex.fe_len, &ex);
1648 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1651 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1652 ex.fe_start + le32_to_cpu(es->s_first_data_block);
1653 /* use do_div to get remainder (would be 64-bit modulo) */
1654 if (do_div(start, sbi->s_stripe) == 0) {
1657 ext4_mb_use_best_found(ac, e4b);
1659 } else if (max >= ac->ac_g_ex.fe_len) {
1660 BUG_ON(ex.fe_len <= 0);
1661 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1662 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1665 ext4_mb_use_best_found(ac, e4b);
1666 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1667 /* Sometimes, caller may want to merge even small
1668 * number of blocks to an existing extent */
1669 BUG_ON(ex.fe_len <= 0);
1670 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1671 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1674 ext4_mb_use_best_found(ac, e4b);
1676 ext4_unlock_group(ac->ac_sb, group);
1677 ext4_mb_release_desc(e4b);
1683 * The routine scans buddy structures (not bitmap!) from given order
1684 * to max order and tries to find big enough chunk to satisfy the req
1686 static noinline_for_stack
1687 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1688 struct ext4_buddy *e4b)
1690 struct super_block *sb = ac->ac_sb;
1691 struct ext4_group_info *grp = e4b->bd_info;
1697 BUG_ON(ac->ac_2order <= 0);
1698 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1699 if (grp->bb_counters[i] == 0)
1702 buddy = mb_find_buddy(e4b, i, &max);
1703 BUG_ON(buddy == NULL);
1705 k = mb_find_next_zero_bit(buddy, max, 0);
1710 ac->ac_b_ex.fe_len = 1 << i;
1711 ac->ac_b_ex.fe_start = k << i;
1712 ac->ac_b_ex.fe_group = e4b->bd_group;
1714 ext4_mb_use_best_found(ac, e4b);
1716 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1718 if (EXT4_SB(sb)->s_mb_stats)
1719 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1726 * The routine scans the group and measures all found extents.
1727 * In order to optimize scanning, caller must pass number of
1728 * free blocks in the group, so the routine can know upper limit.
1730 static noinline_for_stack
1731 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1732 struct ext4_buddy *e4b)
1734 struct super_block *sb = ac->ac_sb;
1735 void *bitmap = EXT4_MB_BITMAP(e4b);
1736 struct ext4_free_extent ex;
1740 free = e4b->bd_info->bb_free;
1743 i = e4b->bd_info->bb_first_free;
1745 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1746 i = mb_find_next_zero_bit(bitmap,
1747 EXT4_BLOCKS_PER_GROUP(sb), i);
1748 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1750 * IF we have corrupt bitmap, we won't find any
1751 * free blocks even though group info says we
1752 * we have free blocks
1754 ext4_grp_locked_error(sb, e4b->bd_group,
1755 __func__, "%d free blocks as per "
1756 "group info. But bitmap says 0",
1761 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1762 BUG_ON(ex.fe_len <= 0);
1763 if (free < ex.fe_len) {
1764 ext4_grp_locked_error(sb, e4b->bd_group,
1765 __func__, "%d free blocks as per "
1766 "group info. But got %d blocks",
1769 * The number of free blocks differs. This mostly
1770 * indicate that the bitmap is corrupt. So exit
1771 * without claiming the space.
1776 ext4_mb_measure_extent(ac, &ex, e4b);
1782 ext4_mb_check_limits(ac, e4b, 1);
1786 * This is a special case for storages like raid5
1787 * we try to find stripe-aligned chunks for stripe-size requests
1788 * XXX should do so at least for multiples of stripe size as well
1790 static noinline_for_stack
1791 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1792 struct ext4_buddy *e4b)
1794 struct super_block *sb = ac->ac_sb;
1795 struct ext4_sb_info *sbi = EXT4_SB(sb);
1796 void *bitmap = EXT4_MB_BITMAP(e4b);
1797 struct ext4_free_extent ex;
1798 ext4_fsblk_t first_group_block;
1803 BUG_ON(sbi->s_stripe == 0);
1805 /* find first stripe-aligned block in group */
1806 first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1807 + le32_to_cpu(sbi->s_es->s_first_data_block);
1808 a = first_group_block + sbi->s_stripe - 1;
1809 do_div(a, sbi->s_stripe);
1810 i = (a * sbi->s_stripe) - first_group_block;
1812 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1813 if (!mb_test_bit(i, bitmap)) {
1814 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1815 if (max >= sbi->s_stripe) {
1818 ext4_mb_use_best_found(ac, e4b);
1826 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1827 ext4_group_t group, int cr)
1829 unsigned free, fragments;
1831 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1832 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1834 BUG_ON(cr < 0 || cr >= 4);
1835 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1837 free = grp->bb_free;
1838 fragments = grp->bb_fragments;
1846 BUG_ON(ac->ac_2order == 0);
1848 /* Avoid using the first bg of a flexgroup for data files */
1849 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1850 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1851 ((group % flex_size) == 0))
1854 bits = ac->ac_sb->s_blocksize_bits + 1;
1855 for (i = ac->ac_2order; i <= bits; i++)
1856 if (grp->bb_counters[i] > 0)
1860 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1864 if (free >= ac->ac_g_ex.fe_len)
1877 * lock the group_info alloc_sem of all the groups
1878 * belonging to the same buddy cache page. This
1879 * make sure other parallel operation on the buddy
1880 * cache doesn't happen whild holding the buddy cache
1883 int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1887 int blocks_per_page;
1888 int groups_per_page;
1889 ext4_group_t ngroups = ext4_get_groups_count(sb);
1890 ext4_group_t first_group;
1891 struct ext4_group_info *grp;
1893 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1895 * the buddy cache inode stores the block bitmap
1896 * and buddy information in consecutive blocks.
1897 * So for each group we need two blocks.
1900 pnum = block / blocks_per_page;
1901 first_group = pnum * blocks_per_page / 2;
1903 groups_per_page = blocks_per_page >> 1;
1904 if (groups_per_page == 0)
1905 groups_per_page = 1;
1906 /* read all groups the page covers into the cache */
1907 for (i = 0; i < groups_per_page; i++) {
1909 if ((first_group + i) >= ngroups)
1911 grp = ext4_get_group_info(sb, first_group + i);
1912 /* take all groups write allocation
1913 * semaphore. This make sure there is
1914 * no block allocation going on in any
1917 down_write_nested(&grp->alloc_sem, i);
1922 void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1923 ext4_group_t group, int locked_group)
1927 int blocks_per_page;
1928 ext4_group_t first_group;
1929 struct ext4_group_info *grp;
1931 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1933 * the buddy cache inode stores the block bitmap
1934 * and buddy information in consecutive blocks.
1935 * So for each group we need two blocks.
1938 pnum = block / blocks_per_page;
1939 first_group = pnum * blocks_per_page / 2;
1940 /* release locks on all the groups */
1941 for (i = 0; i < locked_group; i++) {
1943 grp = ext4_get_group_info(sb, first_group + i);
1944 /* take all groups write allocation
1945 * semaphore. This make sure there is
1946 * no block allocation going on in any
1949 up_write(&grp->alloc_sem);
1954 static noinline_for_stack int
1955 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1957 ext4_group_t ngroups, group, i;
1961 struct ext4_sb_info *sbi;
1962 struct super_block *sb;
1963 struct ext4_buddy e4b;
1967 ngroups = ext4_get_groups_count(sb);
1968 /* non-extent files are limited to low blocks/groups */
1969 if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL))
1970 ngroups = sbi->s_blockfile_groups;
1972 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1974 /* first, try the goal */
1975 err = ext4_mb_find_by_goal(ac, &e4b);
1976 if (err || ac->ac_status == AC_STATUS_FOUND)
1979 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1983 * ac->ac2_order is set only if the fe_len is a power of 2
1984 * if ac2_order is set we also set criteria to 0 so that we
1985 * try exact allocation using buddy.
1987 i = fls(ac->ac_g_ex.fe_len);
1990 * We search using buddy data only if the order of the request
1991 * is greater than equal to the sbi_s_mb_order2_reqs
1992 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1994 if (i >= sbi->s_mb_order2_reqs) {
1996 * This should tell if fe_len is exactly power of 2
1998 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1999 ac->ac_2order = i - 1;
2002 bsbits = ac->ac_sb->s_blocksize_bits;
2004 /* if stream allocation is enabled, use global goal */
2005 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2006 /* TBD: may be hot point */
2007 spin_lock(&sbi->s_md_lock);
2008 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2009 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2010 spin_unlock(&sbi->s_md_lock);
2013 /* Let's just scan groups to find more-less suitable blocks */
2014 cr = ac->ac_2order ? 0 : 1;
2016 * cr == 0 try to get exact allocation,
2017 * cr == 3 try to get anything
2020 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2021 ac->ac_criteria = cr;
2023 * searching for the right group start
2024 * from the goal value specified
2026 group = ac->ac_g_ex.fe_group;
2028 for (i = 0; i < ngroups; group++, i++) {
2029 struct ext4_group_info *grp;
2030 struct ext4_group_desc *desc;
2032 if (group == ngroups)
2035 /* quick check to skip empty groups */
2036 grp = ext4_get_group_info(sb, group);
2037 if (grp->bb_free == 0)
2040 err = ext4_mb_load_buddy(sb, group, &e4b);
2044 ext4_lock_group(sb, group);
2045 if (!ext4_mb_good_group(ac, group, cr)) {
2046 /* someone did allocation from this group */
2047 ext4_unlock_group(sb, group);
2048 ext4_mb_release_desc(&e4b);
2052 ac->ac_groups_scanned++;
2053 desc = ext4_get_group_desc(sb, group, NULL);
2055 ext4_mb_simple_scan_group(ac, &e4b);
2057 ac->ac_g_ex.fe_len == sbi->s_stripe)
2058 ext4_mb_scan_aligned(ac, &e4b);
2060 ext4_mb_complex_scan_group(ac, &e4b);
2062 ext4_unlock_group(sb, group);
2063 ext4_mb_release_desc(&e4b);
2065 if (ac->ac_status != AC_STATUS_CONTINUE)
2070 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2071 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2073 * We've been searching too long. Let's try to allocate
2074 * the best chunk we've found so far
2077 ext4_mb_try_best_found(ac, &e4b);
2078 if (ac->ac_status != AC_STATUS_FOUND) {
2080 * Someone more lucky has already allocated it.
2081 * The only thing we can do is just take first
2083 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2085 ac->ac_b_ex.fe_group = 0;
2086 ac->ac_b_ex.fe_start = 0;
2087 ac->ac_b_ex.fe_len = 0;
2088 ac->ac_status = AC_STATUS_CONTINUE;
2089 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2091 atomic_inc(&sbi->s_mb_lost_chunks);
2099 #ifdef EXT4_MB_HISTORY
2100 struct ext4_mb_proc_session {
2101 struct ext4_mb_history *history;
2102 struct super_block *sb;
2107 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
2108 struct ext4_mb_history *hs,
2111 if (hs == s->history + s->max)
2113 if (!first && hs == s->history + s->start)
2115 while (hs->orig.fe_len == 0) {
2117 if (hs == s->history + s->max)
2119 if (hs == s->history + s->start)
2125 static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
2127 struct ext4_mb_proc_session *s = seq->private;
2128 struct ext4_mb_history *hs;
2132 return SEQ_START_TOKEN;
2133 hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2136 while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
2140 static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
2143 struct ext4_mb_proc_session *s = seq->private;
2144 struct ext4_mb_history *hs = v;
2147 if (v == SEQ_START_TOKEN)
2148 return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2150 return ext4_mb_history_skip_empty(s, ++hs, 0);
2153 static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
2155 char buf[25], buf2[25], buf3[25], *fmt;
2156 struct ext4_mb_history *hs = v;
2158 if (v == SEQ_START_TOKEN) {
2159 seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
2160 "%-5s %-2s %-6s %-5s %-5s %-6s\n",
2161 "pid", "inode", "original", "goal", "result", "found",
2162 "grps", "cr", "flags", "merge", "tail", "broken");
2166 if (hs->op == EXT4_MB_HISTORY_ALLOC) {
2167 fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2168 "0x%04x %-5s %-5u %-6u\n";
2169 sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group,
2170 hs->result.fe_start, hs->result.fe_len,
2171 hs->result.fe_logical);
2172 sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group,
2173 hs->orig.fe_start, hs->orig.fe_len,
2174 hs->orig.fe_logical);
2175 sprintf(buf3, "%u/%d/%u@%u", hs->goal.fe_group,
2176 hs->goal.fe_start, hs->goal.fe_len,
2177 hs->goal.fe_logical);
2178 seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
2179 hs->found, hs->groups, hs->cr, hs->flags,
2180 hs->merged ? "M" : "", hs->tail,
2181 hs->buddy ? 1 << hs->buddy : 0);
2182 } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
2183 fmt = "%-5u %-8u %-23s %-23s %-23s\n";
2184 sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group,
2185 hs->result.fe_start, hs->result.fe_len,
2186 hs->result.fe_logical);
2187 sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group,
2188 hs->orig.fe_start, hs->orig.fe_len,
2189 hs->orig.fe_logical);
2190 seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
2191 } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
2192 sprintf(buf2, "%u/%d/%u", hs->result.fe_group,
2193 hs->result.fe_start, hs->result.fe_len);
2194 seq_printf(seq, "%-5u %-8u %-23s discard\n",
2195 hs->pid, hs->ino, buf2);
2196 } else if (hs->op == EXT4_MB_HISTORY_FREE) {
2197 sprintf(buf2, "%u/%d/%u", hs->result.fe_group,
2198 hs->result.fe_start, hs->result.fe_len);
2199 seq_printf(seq, "%-5u %-8u %-23s free\n",
2200 hs->pid, hs->ino, buf2);
2205 static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
2209 static const struct seq_operations ext4_mb_seq_history_ops = {
2210 .start = ext4_mb_seq_history_start,
2211 .next = ext4_mb_seq_history_next,
2212 .stop = ext4_mb_seq_history_stop,
2213 .show = ext4_mb_seq_history_show,
2216 static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
2218 struct super_block *sb = PDE(inode)->data;
2219 struct ext4_sb_info *sbi = EXT4_SB(sb);
2220 struct ext4_mb_proc_session *s;
2224 if (unlikely(sbi->s_mb_history == NULL))
2226 s = kmalloc(sizeof(*s), GFP_KERNEL);
2230 size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
2231 s->history = kmalloc(size, GFP_KERNEL);
2232 if (s->history == NULL) {
2237 spin_lock(&sbi->s_mb_history_lock);
2238 memcpy(s->history, sbi->s_mb_history, size);
2239 s->max = sbi->s_mb_history_max;
2240 s->start = sbi->s_mb_history_cur % s->max;
2241 spin_unlock(&sbi->s_mb_history_lock);
2243 rc = seq_open(file, &ext4_mb_seq_history_ops);
2245 struct seq_file *m = (struct seq_file *)file->private_data;
2255 static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
2257 struct seq_file *seq = (struct seq_file *)file->private_data;
2258 struct ext4_mb_proc_session *s = seq->private;
2261 return seq_release(inode, file);
2264 static ssize_t ext4_mb_seq_history_write(struct file *file,
2265 const char __user *buffer,
2266 size_t count, loff_t *ppos)
2268 struct seq_file *seq = (struct seq_file *)file->private_data;
2269 struct ext4_mb_proc_session *s = seq->private;
2270 struct super_block *sb = s->sb;
2274 if (count >= sizeof(str)) {
2275 printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
2276 "mb_history", (int)sizeof(str));
2280 if (copy_from_user(str, buffer, count))
2283 value = simple_strtol(str, NULL, 0);
2286 EXT4_SB(sb)->s_mb_history_filter = value;
2291 static const struct file_operations ext4_mb_seq_history_fops = {
2292 .owner = THIS_MODULE,
2293 .open = ext4_mb_seq_history_open,
2295 .write = ext4_mb_seq_history_write,
2296 .llseek = seq_lseek,
2297 .release = ext4_mb_seq_history_release,
2300 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2302 struct super_block *sb = seq->private;
2305 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2308 return (void *) ((unsigned long) group);
2311 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2313 struct super_block *sb = seq->private;
2317 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2320 return (void *) ((unsigned long) group);
2323 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2325 struct super_block *sb = seq->private;
2326 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2329 struct ext4_buddy e4b;
2331 struct ext4_group_info info;
2332 ext4_grpblk_t counters[16];
2337 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2338 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2339 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2340 "group", "free", "frags", "first",
2341 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2342 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2344 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2345 sizeof(struct ext4_group_info);
2346 err = ext4_mb_load_buddy(sb, group, &e4b);
2348 seq_printf(seq, "#%-5u: I/O error\n", group);
2351 ext4_lock_group(sb, group);
2352 memcpy(&sg, ext4_get_group_info(sb, group), i);
2353 ext4_unlock_group(sb, group);
2354 ext4_mb_release_desc(&e4b);
2356 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2357 sg.info.bb_fragments, sg.info.bb_first_free);
2358 for (i = 0; i <= 13; i++)
2359 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2360 sg.info.bb_counters[i] : 0);
2361 seq_printf(seq, " ]\n");
2366 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2370 static const struct seq_operations ext4_mb_seq_groups_ops = {
2371 .start = ext4_mb_seq_groups_start,
2372 .next = ext4_mb_seq_groups_next,
2373 .stop = ext4_mb_seq_groups_stop,
2374 .show = ext4_mb_seq_groups_show,
2377 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2379 struct super_block *sb = PDE(inode)->data;
2382 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2384 struct seq_file *m = (struct seq_file *)file->private_data;
2391 static const struct file_operations ext4_mb_seq_groups_fops = {
2392 .owner = THIS_MODULE,
2393 .open = ext4_mb_seq_groups_open,
2395 .llseek = seq_lseek,
2396 .release = seq_release,
2399 static void ext4_mb_history_release(struct super_block *sb)
2401 struct ext4_sb_info *sbi = EXT4_SB(sb);
2403 if (sbi->s_proc != NULL) {
2404 remove_proc_entry("mb_groups", sbi->s_proc);
2405 if (sbi->s_mb_history_max)
2406 remove_proc_entry("mb_history", sbi->s_proc);
2408 kfree(sbi->s_mb_history);
2411 static void ext4_mb_history_init(struct super_block *sb)
2413 struct ext4_sb_info *sbi = EXT4_SB(sb);
2416 if (sbi->s_proc != NULL) {
2417 if (sbi->s_mb_history_max)
2418 proc_create_data("mb_history", S_IRUGO, sbi->s_proc,
2419 &ext4_mb_seq_history_fops, sb);
2420 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2421 &ext4_mb_seq_groups_fops, sb);
2424 sbi->s_mb_history_cur = 0;
2425 spin_lock_init(&sbi->s_mb_history_lock);
2426 i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
2427 sbi->s_mb_history = i ? kzalloc(i, GFP_KERNEL) : NULL;
2428 /* if we can't allocate history, then we simple won't use it */
2431 static noinline_for_stack void
2432 ext4_mb_store_history(struct ext4_allocation_context *ac)
2434 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2435 struct ext4_mb_history h;
2437 if (sbi->s_mb_history == NULL)
2440 if (!(ac->ac_op & sbi->s_mb_history_filter))
2444 h.pid = current->pid;
2445 h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2446 h.orig = ac->ac_o_ex;
2447 h.result = ac->ac_b_ex;
2448 h.flags = ac->ac_flags;
2449 h.found = ac->ac_found;
2450 h.groups = ac->ac_groups_scanned;
2451 h.cr = ac->ac_criteria;
2452 h.tail = ac->ac_tail;
2453 h.buddy = ac->ac_buddy;
2455 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
2456 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2457 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2459 h.goal = ac->ac_g_ex;
2460 h.result = ac->ac_f_ex;
2463 spin_lock(&sbi->s_mb_history_lock);
2464 memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2465 if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2466 sbi->s_mb_history_cur = 0;
2467 spin_unlock(&sbi->s_mb_history_lock);
2471 #define ext4_mb_history_release(sb)
2472 #define ext4_mb_history_init(sb)
2476 /* Create and initialize ext4_group_info data for the given group. */
2477 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2478 struct ext4_group_desc *desc)
2482 struct ext4_sb_info *sbi = EXT4_SB(sb);
2483 struct ext4_group_info **meta_group_info;
2486 * First check if this group is the first of a reserved block.
2487 * If it's true, we have to allocate a new table of pointers
2488 * to ext4_group_info structures
2490 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2491 metalen = sizeof(*meta_group_info) <<
2492 EXT4_DESC_PER_BLOCK_BITS(sb);
2493 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2494 if (meta_group_info == NULL) {
2495 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2497 goto exit_meta_group_info;
2499 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2504 * calculate needed size. if change bb_counters size,
2505 * don't forget about ext4_mb_generate_buddy()
2507 len = offsetof(typeof(**meta_group_info),
2508 bb_counters[sb->s_blocksize_bits + 2]);
2511 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2512 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2514 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2515 if (meta_group_info[i] == NULL) {
2516 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2517 goto exit_group_info;
2519 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2520 &(meta_group_info[i]->bb_state));
2523 * initialize bb_free to be able to skip
2524 * empty groups without initialization
2526 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2527 meta_group_info[i]->bb_free =
2528 ext4_free_blocks_after_init(sb, group, desc);
2530 meta_group_info[i]->bb_free =
2531 ext4_free_blks_count(sb, desc);
2534 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2535 init_rwsem(&meta_group_info[i]->alloc_sem);
2536 meta_group_info[i]->bb_free_root.rb_node = NULL;
2540 struct buffer_head *bh;
2541 meta_group_info[i]->bb_bitmap =
2542 kmalloc(sb->s_blocksize, GFP_KERNEL);
2543 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2544 bh = ext4_read_block_bitmap(sb, group);
2546 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2555 /* If a meta_group_info table has been allocated, release it now */
2556 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2557 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2558 exit_meta_group_info:
2560 } /* ext4_mb_add_groupinfo */
2562 static int ext4_mb_init_backend(struct super_block *sb)
2564 ext4_group_t ngroups = ext4_get_groups_count(sb);
2566 struct ext4_sb_info *sbi = EXT4_SB(sb);
2567 struct ext4_super_block *es = sbi->s_es;
2568 int num_meta_group_infos;
2569 int num_meta_group_infos_max;
2571 struct ext4_group_desc *desc;
2573 /* This is the number of blocks used by GDT */
2574 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2575 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2578 * This is the total number of blocks used by GDT including
2579 * the number of reserved blocks for GDT.
2580 * The s_group_info array is allocated with this value
2581 * to allow a clean online resize without a complex
2582 * manipulation of pointer.
2583 * The drawback is the unused memory when no resize
2584 * occurs but it's very low in terms of pages
2585 * (see comments below)
2586 * Need to handle this properly when META_BG resizing is allowed
2588 num_meta_group_infos_max = num_meta_group_infos +
2589 le16_to_cpu(es->s_reserved_gdt_blocks);
2592 * array_size is the size of s_group_info array. We round it
2593 * to the next power of two because this approximation is done
2594 * internally by kmalloc so we can have some more memory
2595 * for free here (e.g. may be used for META_BG resize).
2598 while (array_size < sizeof(*sbi->s_group_info) *
2599 num_meta_group_infos_max)
2600 array_size = array_size << 1;
2601 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2602 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2603 * So a two level scheme suffices for now. */
2604 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2605 if (sbi->s_group_info == NULL) {
2606 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2609 sbi->s_buddy_cache = new_inode(sb);
2610 if (sbi->s_buddy_cache == NULL) {
2611 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2614 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2615 for (i = 0; i < ngroups; i++) {
2616 desc = ext4_get_group_desc(sb, i, NULL);
2619 "EXT4-fs: can't read descriptor %u\n", i);
2622 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2630 kfree(ext4_get_group_info(sb, i));
2631 i = num_meta_group_infos;
2633 kfree(sbi->s_group_info[i]);
2634 iput(sbi->s_buddy_cache);
2636 kfree(sbi->s_group_info);
2640 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2642 struct ext4_sb_info *sbi = EXT4_SB(sb);
2648 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2650 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2651 if (sbi->s_mb_offsets == NULL) {
2655 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2656 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2657 if (sbi->s_mb_maxs == NULL) {
2658 kfree(sbi->s_mb_offsets);
2662 /* order 0 is regular bitmap */
2663 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2664 sbi->s_mb_offsets[0] = 0;
2668 max = sb->s_blocksize << 2;
2670 sbi->s_mb_offsets[i] = offset;
2671 sbi->s_mb_maxs[i] = max;
2672 offset += 1 << (sb->s_blocksize_bits - i);
2675 } while (i <= sb->s_blocksize_bits + 1);
2677 /* init file for buddy data */
2678 ret = ext4_mb_init_backend(sb);
2680 kfree(sbi->s_mb_offsets);
2681 kfree(sbi->s_mb_maxs);
2685 spin_lock_init(&sbi->s_md_lock);
2686 spin_lock_init(&sbi->s_bal_lock);
2688 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2689 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2690 sbi->s_mb_stats = MB_DEFAULT_STATS;
2691 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2692 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2693 sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
2694 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2696 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2697 if (sbi->s_locality_groups == NULL) {
2698 kfree(sbi->s_mb_offsets);
2699 kfree(sbi->s_mb_maxs);
2702 for_each_possible_cpu(i) {
2703 struct ext4_locality_group *lg;
2704 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2705 mutex_init(&lg->lg_mutex);
2706 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2707 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2708 spin_lock_init(&lg->lg_prealloc_lock);
2711 ext4_mb_history_init(sb);
2714 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2718 /* need to called with the ext4 group lock held */
2719 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2721 struct ext4_prealloc_space *pa;
2722 struct list_head *cur, *tmp;
2725 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2726 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2727 list_del(&pa->pa_group_list);
2729 kmem_cache_free(ext4_pspace_cachep, pa);
2732 mb_debug(1, "mballoc: %u PAs left\n", count);
2736 int ext4_mb_release(struct super_block *sb)
2738 ext4_group_t ngroups = ext4_get_groups_count(sb);
2740 int num_meta_group_infos;
2741 struct ext4_group_info *grinfo;
2742 struct ext4_sb_info *sbi = EXT4_SB(sb);
2744 if (sbi->s_group_info) {
2745 for (i = 0; i < ngroups; i++) {
2746 grinfo = ext4_get_group_info(sb, i);
2748 kfree(grinfo->bb_bitmap);
2750 ext4_lock_group(sb, i);
2751 ext4_mb_cleanup_pa(grinfo);
2752 ext4_unlock_group(sb, i);
2755 num_meta_group_infos = (ngroups +
2756 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2757 EXT4_DESC_PER_BLOCK_BITS(sb);
2758 for (i = 0; i < num_meta_group_infos; i++)
2759 kfree(sbi->s_group_info[i]);
2760 kfree(sbi->s_group_info);
2762 kfree(sbi->s_mb_offsets);
2763 kfree(sbi->s_mb_maxs);
2764 if (sbi->s_buddy_cache)
2765 iput(sbi->s_buddy_cache);
2766 if (sbi->s_mb_stats) {
2768 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2769 atomic_read(&sbi->s_bal_allocated),
2770 atomic_read(&sbi->s_bal_reqs),
2771 atomic_read(&sbi->s_bal_success));
2773 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2774 "%u 2^N hits, %u breaks, %u lost\n",
2775 atomic_read(&sbi->s_bal_ex_scanned),
2776 atomic_read(&sbi->s_bal_goals),
2777 atomic_read(&sbi->s_bal_2orders),
2778 atomic_read(&sbi->s_bal_breaks),
2779 atomic_read(&sbi->s_mb_lost_chunks));
2781 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2782 sbi->s_mb_buddies_generated++,
2783 sbi->s_mb_generation_time);
2785 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2786 atomic_read(&sbi->s_mb_preallocated),
2787 atomic_read(&sbi->s_mb_discarded));
2790 free_percpu(sbi->s_locality_groups);
2791 ext4_mb_history_release(sb);
2797 * This function is called by the jbd2 layer once the commit has finished,
2798 * so we know we can free the blocks that were released with that commit.
2800 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2802 struct super_block *sb = journal->j_private;
2803 struct ext4_buddy e4b;
2804 struct ext4_group_info *db;
2805 int err, count = 0, count2 = 0;
2806 struct ext4_free_data *entry;
2807 ext4_fsblk_t discard_block;
2808 struct list_head *l, *ltmp;
2810 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2811 entry = list_entry(l, struct ext4_free_data, list);
2813 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2814 entry->count, entry->group, entry);
2816 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2817 /* we expect to find existing buddy because it's pinned */
2821 /* there are blocks to put in buddy to make them really free */
2822 count += entry->count;
2824 ext4_lock_group(sb, entry->group);
2825 /* Take it out of per group rb tree */
2826 rb_erase(&entry->node, &(db->bb_free_root));
2827 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2829 if (!db->bb_free_root.rb_node) {
2830 /* No more items in the per group rb tree
2831 * balance refcounts from ext4_mb_free_metadata()
2833 page_cache_release(e4b.bd_buddy_page);
2834 page_cache_release(e4b.bd_bitmap_page);
2836 ext4_unlock_group(sb, entry->group);
2837 discard_block = (ext4_fsblk_t) entry->group * EXT4_BLOCKS_PER_GROUP(sb)
2839 + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
2840 trace_ext4_discard_blocks(sb, (unsigned long long)discard_block,
2842 sb_issue_discard(sb, discard_block, entry->count);
2844 kmem_cache_free(ext4_free_ext_cachep, entry);
2845 ext4_mb_release_desc(&e4b);
2848 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2851 #ifdef CONFIG_EXT4_DEBUG
2852 u8 mb_enable_debug __read_mostly;
2854 static struct dentry *debugfs_dir;
2855 static struct dentry *debugfs_debug;
2857 static void __init ext4_create_debugfs_entry(void)
2859 debugfs_dir = debugfs_create_dir("ext4", NULL);
2861 debugfs_debug = debugfs_create_u8("mballoc-debug",
2867 static void ext4_remove_debugfs_entry(void)
2869 debugfs_remove(debugfs_debug);
2870 debugfs_remove(debugfs_dir);
2875 static void __init ext4_create_debugfs_entry(void)
2879 static void ext4_remove_debugfs_entry(void)
2885 int __init init_ext4_mballoc(void)
2887 ext4_pspace_cachep =
2888 kmem_cache_create("ext4_prealloc_space",
2889 sizeof(struct ext4_prealloc_space),
2890 0, SLAB_RECLAIM_ACCOUNT, NULL);
2891 if (ext4_pspace_cachep == NULL)
2895 kmem_cache_create("ext4_alloc_context",
2896 sizeof(struct ext4_allocation_context),
2897 0, SLAB_RECLAIM_ACCOUNT, NULL);
2898 if (ext4_ac_cachep == NULL) {
2899 kmem_cache_destroy(ext4_pspace_cachep);
2903 ext4_free_ext_cachep =
2904 kmem_cache_create("ext4_free_block_extents",
2905 sizeof(struct ext4_free_data),
2906 0, SLAB_RECLAIM_ACCOUNT, NULL);
2907 if (ext4_free_ext_cachep == NULL) {
2908 kmem_cache_destroy(ext4_pspace_cachep);
2909 kmem_cache_destroy(ext4_ac_cachep);
2912 ext4_create_debugfs_entry();
2916 void exit_ext4_mballoc(void)
2919 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2920 * before destroying the slab cache.
2923 kmem_cache_destroy(ext4_pspace_cachep);
2924 kmem_cache_destroy(ext4_ac_cachep);
2925 kmem_cache_destroy(ext4_free_ext_cachep);
2926 ext4_remove_debugfs_entry();
2931 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2932 * Returns 0 if success or error code
2934 static noinline_for_stack int
2935 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2936 handle_t *handle, unsigned int reserv_blks)
2938 struct buffer_head *bitmap_bh = NULL;
2939 struct ext4_super_block *es;
2940 struct ext4_group_desc *gdp;
2941 struct buffer_head *gdp_bh;
2942 struct ext4_sb_info *sbi;
2943 struct super_block *sb;
2947 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2948 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2956 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2960 err = ext4_journal_get_write_access(handle, bitmap_bh);
2965 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2969 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2970 ext4_free_blks_count(sb, gdp));
2972 err = ext4_journal_get_write_access(handle, gdp_bh);
2976 block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
2977 + ac->ac_b_ex.fe_start
2978 + le32_to_cpu(es->s_first_data_block);
2980 len = ac->ac_b_ex.fe_len;
2981 if (!ext4_data_block_valid(sbi, block, len)) {
2982 ext4_error(sb, __func__,
2983 "Allocating blocks %llu-%llu which overlap "
2984 "fs metadata\n", block, block+len);
2985 /* File system mounted not to panic on error
2986 * Fix the bitmap and repeat the block allocation
2987 * We leak some of the blocks here.
2989 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2990 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2991 ac->ac_b_ex.fe_len);
2992 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2993 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2999 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3000 #ifdef AGGRESSIVE_CHECK
3003 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3004 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3005 bitmap_bh->b_data));
3009 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
3010 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
3011 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3012 ext4_free_blks_set(sb, gdp,
3013 ext4_free_blocks_after_init(sb,
3014 ac->ac_b_ex.fe_group, gdp));
3016 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
3017 ext4_free_blks_set(sb, gdp, len);
3018 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
3020 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3021 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
3023 * Now reduce the dirty block count also. Should not go negative
3025 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3026 /* release all the reserved blocks if non delalloc */
3027 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
3029 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
3030 ac->ac_b_ex.fe_len);
3031 /* convert reserved quota blocks to real quota blocks */
3032 vfs_dq_claim_block(ac->ac_inode, ac->ac_b_ex.fe_len);
3035 if (sbi->s_log_groups_per_flex) {
3036 ext4_group_t flex_group = ext4_flex_group(sbi,
3037 ac->ac_b_ex.fe_group);
3038 atomic_sub(ac->ac_b_ex.fe_len,
3039 &sbi->s_flex_groups[flex_group].free_blocks);
3042 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3045 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3054 * here we normalize request for locality group
3055 * Group request are normalized to s_strip size if we set the same via mount
3056 * option. If not we set it to s_mb_group_prealloc which can be configured via
3057 * /sys/fs/ext4/<partition>/mb_group_prealloc
3059 * XXX: should we try to preallocate more than the group has now?
3061 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3063 struct super_block *sb = ac->ac_sb;
3064 struct ext4_locality_group *lg = ac->ac_lg;
3067 if (EXT4_SB(sb)->s_stripe)
3068 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
3070 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3071 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3072 current->pid, ac->ac_g_ex.fe_len);
3076 * Normalization means making request better in terms of
3077 * size and alignment
3079 static noinline_for_stack void
3080 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3081 struct ext4_allocation_request *ar)
3085 loff_t size, orig_size, start_off;
3086 ext4_lblk_t start, orig_start;
3087 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3088 struct ext4_prealloc_space *pa;
3090 /* do normalize only data requests, metadata requests
3091 do not need preallocation */
3092 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3095 /* sometime caller may want exact blocks */
3096 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3099 /* caller may indicate that preallocation isn't
3100 * required (it's a tail, for example) */
3101 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3104 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3105 ext4_mb_normalize_group_request(ac);
3109 bsbits = ac->ac_sb->s_blocksize_bits;
3111 /* first, let's learn actual file size
3112 * given current request is allocated */
3113 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3114 size = size << bsbits;
3115 if (size < i_size_read(ac->ac_inode))
3116 size = i_size_read(ac->ac_inode);
3118 /* max size of free chunks */
3121 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3122 (req <= (size) || max <= (chunk_size))
3124 /* first, try to predict filesize */
3125 /* XXX: should this table be tunable? */
3127 if (size <= 16 * 1024) {
3129 } else if (size <= 32 * 1024) {
3131 } else if (size <= 64 * 1024) {
3133 } else if (size <= 128 * 1024) {
3135 } else if (size <= 256 * 1024) {
3137 } else if (size <= 512 * 1024) {
3139 } else if (size <= 1024 * 1024) {
3141 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3142 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3143 (21 - bsbits)) << 21;
3144 size = 2 * 1024 * 1024;
3145 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3146 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3147 (22 - bsbits)) << 22;
3148 size = 4 * 1024 * 1024;
3149 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3150 (8<<20)>>bsbits, max, 8 * 1024)) {
3151 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3152 (23 - bsbits)) << 23;
3153 size = 8 * 1024 * 1024;
3155 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3156 size = ac->ac_o_ex.fe_len << bsbits;
3158 orig_size = size = size >> bsbits;
3159 orig_start = start = start_off >> bsbits;
3161 /* don't cover already allocated blocks in selected range */
3162 if (ar->pleft && start <= ar->lleft) {
3163 size -= ar->lleft + 1 - start;
3164 start = ar->lleft + 1;
3166 if (ar->pright && start + size - 1 >= ar->lright)
3167 size -= start + size - ar->lright;
3171 /* check we don't cross already preallocated blocks */
3173 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3178 spin_lock(&pa->pa_lock);
3179 if (pa->pa_deleted) {
3180 spin_unlock(&pa->pa_lock);
3184 pa_end = pa->pa_lstart + pa->pa_len;
3186 /* PA must not overlap original request */
3187 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3188 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3190 /* skip PAs this normalized request doesn't overlap with */
3191 if (pa->pa_lstart >= end || pa_end <= start) {
3192 spin_unlock(&pa->pa_lock);
3195 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3197 /* adjust start or end to be adjacent to this pa */
3198 if (pa_end <= ac->ac_o_ex.fe_logical) {
3199 BUG_ON(pa_end < start);
3201 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3202 BUG_ON(pa->pa_lstart > end);
3203 end = pa->pa_lstart;
3205 spin_unlock(&pa->pa_lock);
3210 /* XXX: extra loop to check we really don't overlap preallocations */
3212 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3214 spin_lock(&pa->pa_lock);
3215 if (pa->pa_deleted == 0) {
3216 pa_end = pa->pa_lstart + pa->pa_len;
3217 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3219 spin_unlock(&pa->pa_lock);
3223 if (start + size <= ac->ac_o_ex.fe_logical &&
3224 start > ac->ac_o_ex.fe_logical) {
3225 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3226 (unsigned long) start, (unsigned long) size,
3227 (unsigned long) ac->ac_o_ex.fe_logical);
3229 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3230 start > ac->ac_o_ex.fe_logical);
3231 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3233 /* now prepare goal request */
3235 /* XXX: is it better to align blocks WRT to logical
3236 * placement or satisfy big request as is */
3237 ac->ac_g_ex.fe_logical = start;
3238 ac->ac_g_ex.fe_len = size;
3240 /* define goal start in order to merge */
3241 if (ar->pright && (ar->lright == (start + size))) {
3242 /* merge to the right */
3243 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3244 &ac->ac_f_ex.fe_group,
3245 &ac->ac_f_ex.fe_start);
3246 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3248 if (ar->pleft && (ar->lleft + 1 == start)) {
3249 /* merge to the left */
3250 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3251 &ac->ac_f_ex.fe_group,
3252 &ac->ac_f_ex.fe_start);
3253 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3256 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3257 (unsigned) orig_size, (unsigned) start);
3260 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3262 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3264 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3265 atomic_inc(&sbi->s_bal_reqs);
3266 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3267 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3268 atomic_inc(&sbi->s_bal_success);
3269 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3270 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3271 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3272 atomic_inc(&sbi->s_bal_goals);
3273 if (ac->ac_found > sbi->s_mb_max_to_scan)
3274 atomic_inc(&sbi->s_bal_breaks);
3277 ext4_mb_store_history(ac);
3281 * use blocks preallocated to inode
3283 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3284 struct ext4_prealloc_space *pa)
3290 /* found preallocated blocks, use them */
3291 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3292 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3294 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3295 &ac->ac_b_ex.fe_start);
3296 ac->ac_b_ex.fe_len = len;
3297 ac->ac_status = AC_STATUS_FOUND;
3300 BUG_ON(start < pa->pa_pstart);
3301 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3302 BUG_ON(pa->pa_free < len);
3305 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3309 * use blocks preallocated to locality group
3311 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3312 struct ext4_prealloc_space *pa)
3314 unsigned int len = ac->ac_o_ex.fe_len;
3316 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3317 &ac->ac_b_ex.fe_group,
3318 &ac->ac_b_ex.fe_start);
3319 ac->ac_b_ex.fe_len = len;
3320 ac->ac_status = AC_STATUS_FOUND;
3323 /* we don't correct pa_pstart or pa_plen here to avoid
3324 * possible race when the group is being loaded concurrently
3325 * instead we correct pa later, after blocks are marked
3326 * in on-disk bitmap -- see ext4_mb_release_context()
3327 * Other CPUs are prevented from allocating from this pa by lg_mutex
3329 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3333 * Return the prealloc space that have minimal distance
3334 * from the goal block. @cpa is the prealloc
3335 * space that is having currently known minimal distance
3336 * from the goal block.
3338 static struct ext4_prealloc_space *
3339 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3340 struct ext4_prealloc_space *pa,
3341 struct ext4_prealloc_space *cpa)
3343 ext4_fsblk_t cur_distance, new_distance;
3346 atomic_inc(&pa->pa_count);
3349 cur_distance = abs(goal_block - cpa->pa_pstart);
3350 new_distance = abs(goal_block - pa->pa_pstart);
3352 if (cur_distance < new_distance)
3355 /* drop the previous reference */
3356 atomic_dec(&cpa->pa_count);
3357 atomic_inc(&pa->pa_count);
3362 * search goal blocks in preallocated space
3364 static noinline_for_stack int
3365 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3368 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3369 struct ext4_locality_group *lg;
3370 struct ext4_prealloc_space *pa, *cpa = NULL;
3371 ext4_fsblk_t goal_block;
3373 /* only data can be preallocated */
3374 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3377 /* first, try per-file preallocation */
3379 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3381 /* all fields in this condition don't change,
3382 * so we can skip locking for them */
3383 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3384 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3387 /* non-extent files can't have physical blocks past 2^32 */
3388 if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL) &&
3389 pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3392 /* found preallocated blocks, use them */
3393 spin_lock(&pa->pa_lock);
3394 if (pa->pa_deleted == 0 && pa->pa_free) {
3395 atomic_inc(&pa->pa_count);
3396 ext4_mb_use_inode_pa(ac, pa);
3397 spin_unlock(&pa->pa_lock);
3398 ac->ac_criteria = 10;
3402 spin_unlock(&pa->pa_lock);
3406 /* can we use group allocation? */
3407 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3410 /* inode may have no locality group for some reason */
3414 order = fls(ac->ac_o_ex.fe_len) - 1;
3415 if (order > PREALLOC_TB_SIZE - 1)
3416 /* The max size of hash table is PREALLOC_TB_SIZE */
3417 order = PREALLOC_TB_SIZE - 1;
3419 goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) +
3420 ac->ac_g_ex.fe_start +
3421 le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block);
3423 * search for the prealloc space that is having
3424 * minimal distance from the goal block.
3426 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3428 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3430 spin_lock(&pa->pa_lock);
3431 if (pa->pa_deleted == 0 &&
3432 pa->pa_free >= ac->ac_o_ex.fe_len) {
3434 cpa = ext4_mb_check_group_pa(goal_block,
3437 spin_unlock(&pa->pa_lock);
3442 ext4_mb_use_group_pa(ac, cpa);
3443 ac->ac_criteria = 20;
3450 * the function goes through all block freed in the group
3451 * but not yet committed and marks them used in in-core bitmap.
3452 * buddy must be generated from this bitmap
3453 * Need to be called with the ext4 group lock held
3455 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3459 struct ext4_group_info *grp;
3460 struct ext4_free_data *entry;
3462 grp = ext4_get_group_info(sb, group);
3463 n = rb_first(&(grp->bb_free_root));
3466 entry = rb_entry(n, struct ext4_free_data, node);
3467 mb_set_bits(bitmap, entry->start_blk, entry->count);
3474 * the function goes through all preallocation in this group and marks them
3475 * used in in-core bitmap. buddy must be generated from this bitmap
3476 * Need to be called with ext4 group lock held
3478 static noinline_for_stack
3479 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3482 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3483 struct ext4_prealloc_space *pa;
3484 struct list_head *cur;
3485 ext4_group_t groupnr;
3486 ext4_grpblk_t start;
3487 int preallocated = 0;
3491 /* all form of preallocation discards first load group,
3492 * so the only competing code is preallocation use.
3493 * we don't need any locking here
3494 * notice we do NOT ignore preallocations with pa_deleted
3495 * otherwise we could leave used blocks available for
3496 * allocation in buddy when concurrent ext4_mb_put_pa()
3497 * is dropping preallocation
3499 list_for_each(cur, &grp->bb_prealloc_list) {
3500 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3501 spin_lock(&pa->pa_lock);
3502 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3505 spin_unlock(&pa->pa_lock);
3506 if (unlikely(len == 0))
3508 BUG_ON(groupnr != group);
3509 mb_set_bits(bitmap, start, len);
3510 preallocated += len;
3513 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3516 static void ext4_mb_pa_callback(struct rcu_head *head)
3518 struct ext4_prealloc_space *pa;
3519 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3520 kmem_cache_free(ext4_pspace_cachep, pa);
3524 * drops a reference to preallocated space descriptor
3525 * if this was the last reference and the space is consumed
3527 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3528 struct super_block *sb, struct ext4_prealloc_space *pa)
3531 ext4_fsblk_t grp_blk;
3533 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3536 /* in this short window concurrent discard can set pa_deleted */
3537 spin_lock(&pa->pa_lock);
3538 if (pa->pa_deleted == 1) {
3539 spin_unlock(&pa->pa_lock);
3544 spin_unlock(&pa->pa_lock);
3546 grp_blk = pa->pa_pstart;
3548 * If doing group-based preallocation, pa_pstart may be in the
3549 * next group when pa is used up
3551 if (pa->pa_type == MB_GROUP_PA)
3554 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3559 * P1 (buddy init) P2 (regular allocation)
3560 * find block B in PA
3561 * copy on-disk bitmap to buddy
3562 * mark B in on-disk bitmap
3563 * drop PA from group
3564 * mark all PAs in buddy
3566 * thus, P1 initializes buddy with B available. to prevent this
3567 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3570 ext4_lock_group(sb, grp);
3571 list_del(&pa->pa_group_list);
3572 ext4_unlock_group(sb, grp);
3574 spin_lock(pa->pa_obj_lock);
3575 list_del_rcu(&pa->pa_inode_list);
3576 spin_unlock(pa->pa_obj_lock);
3578 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3582 * creates new preallocated space for given inode
3584 static noinline_for_stack int
3585 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3587 struct super_block *sb = ac->ac_sb;
3588 struct ext4_prealloc_space *pa;
3589 struct ext4_group_info *grp;
3590 struct ext4_inode_info *ei;
3592 /* preallocate only when found space is larger then requested */
3593 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3594 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3595 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3597 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3601 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3607 /* we can't allocate as much as normalizer wants.
3608 * so, found space must get proper lstart
3609 * to cover original request */
3610 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3611 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3613 /* we're limited by original request in that
3614 * logical block must be covered any way
3615 * winl is window we can move our chunk within */
3616 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3618 /* also, we should cover whole original request */
3619 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3621 /* the smallest one defines real window */
3622 win = min(winl, wins);
3624 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3625 if (offs && offs < win)
3628 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3629 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3630 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3633 /* preallocation can change ac_b_ex, thus we store actually
3634 * allocated blocks for history */
3635 ac->ac_f_ex = ac->ac_b_ex;
3637 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3638 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3639 pa->pa_len = ac->ac_b_ex.fe_len;
3640 pa->pa_free = pa->pa_len;
3641 atomic_set(&pa->pa_count, 1);
3642 spin_lock_init(&pa->pa_lock);
3643 INIT_LIST_HEAD(&pa->pa_inode_list);
3644 INIT_LIST_HEAD(&pa->pa_group_list);
3646 pa->pa_type = MB_INODE_PA;
3648 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3649 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3650 trace_ext4_mb_new_inode_pa(ac, pa);
3652 ext4_mb_use_inode_pa(ac, pa);
3653 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3655 ei = EXT4_I(ac->ac_inode);
3656 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3658 pa->pa_obj_lock = &ei->i_prealloc_lock;
3659 pa->pa_inode = ac->ac_inode;
3661 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3662 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3663 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3665 spin_lock(pa->pa_obj_lock);
3666 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3667 spin_unlock(pa->pa_obj_lock);
3673 * creates new preallocated space for locality group inodes belongs to
3675 static noinline_for_stack int
3676 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3678 struct super_block *sb = ac->ac_sb;
3679 struct ext4_locality_group *lg;
3680 struct ext4_prealloc_space *pa;
3681 struct ext4_group_info *grp;
3683 /* preallocate only when found space is larger then requested */
3684 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3685 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3686 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3688 BUG_ON(ext4_pspace_cachep == NULL);
3689 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3693 /* preallocation can change ac_b_ex, thus we store actually
3694 * allocated blocks for history */
3695 ac->ac_f_ex = ac->ac_b_ex;
3697 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3698 pa->pa_lstart = pa->pa_pstart;
3699 pa->pa_len = ac->ac_b_ex.fe_len;
3700 pa->pa_free = pa->pa_len;
3701 atomic_set(&pa->pa_count, 1);
3702 spin_lock_init(&pa->pa_lock);
3703 INIT_LIST_HEAD(&pa->pa_inode_list);
3704 INIT_LIST_HEAD(&pa->pa_group_list);
3706 pa->pa_type = MB_GROUP_PA;
3708 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3709 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3710 trace_ext4_mb_new_group_pa(ac, pa);
3712 ext4_mb_use_group_pa(ac, pa);
3713 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3715 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3719 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3720 pa->pa_inode = NULL;
3722 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3723 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3724 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3727 * We will later add the new pa to the right bucket
3728 * after updating the pa_free in ext4_mb_release_context
3733 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3737 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3738 err = ext4_mb_new_group_pa(ac);
3740 err = ext4_mb_new_inode_pa(ac);
3745 * finds all unused blocks in on-disk bitmap, frees them in
3746 * in-core bitmap and buddy.
3747 * @pa must be unlinked from inode and group lists, so that
3748 * nobody else can find/use it.
3749 * the caller MUST hold group/inode locks.
3750 * TODO: optimize the case when there are no in-core structures yet
3752 static noinline_for_stack int
3753 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3754 struct ext4_prealloc_space *pa,
3755 struct ext4_allocation_context *ac)
3757 struct super_block *sb = e4b->bd_sb;
3758 struct ext4_sb_info *sbi = EXT4_SB(sb);
3763 unsigned long long grp_blk_start;
3768 BUG_ON(pa->pa_deleted == 0);
3769 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3770 grp_blk_start = pa->pa_pstart - bit;
3771 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3772 end = bit + pa->pa_len;
3776 ac->ac_inode = pa->pa_inode;
3777 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3781 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3784 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3785 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3786 le32_to_cpu(sbi->s_es->s_first_data_block);
3787 mb_debug(1, " free preallocated %u/%u in group %u\n",
3788 (unsigned) start, (unsigned) next - bit,
3793 ac->ac_b_ex.fe_group = group;
3794 ac->ac_b_ex.fe_start = bit;
3795 ac->ac_b_ex.fe_len = next - bit;
3796 ac->ac_b_ex.fe_logical = 0;
3797 ext4_mb_store_history(ac);
3800 trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit,
3802 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3805 if (free != pa->pa_free) {
3806 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3807 pa, (unsigned long) pa->pa_lstart,
3808 (unsigned long) pa->pa_pstart,
3809 (unsigned long) pa->pa_len);
3810 ext4_grp_locked_error(sb, group,
3811 __func__, "free %u, pa_free %u",
3814 * pa is already deleted so we use the value obtained
3815 * from the bitmap and continue.
3818 atomic_add(free, &sbi->s_mb_discarded);
3823 static noinline_for_stack int
3824 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3825 struct ext4_prealloc_space *pa,
3826 struct ext4_allocation_context *ac)
3828 struct super_block *sb = e4b->bd_sb;
3833 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3835 trace_ext4_mb_release_group_pa(ac, pa);
3836 BUG_ON(pa->pa_deleted == 0);
3837 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3838 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3839 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3840 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3844 ac->ac_inode = NULL;
3845 ac->ac_b_ex.fe_group = group;
3846 ac->ac_b_ex.fe_start = bit;
3847 ac->ac_b_ex.fe_len = pa->pa_len;
3848 ac->ac_b_ex.fe_logical = 0;
3849 ext4_mb_store_history(ac);
3856 * releases all preallocations in given group
3858 * first, we need to decide discard policy:
3859 * - when do we discard
3861 * - how many do we discard
3862 * 1) how many requested
3864 static noinline_for_stack int
3865 ext4_mb_discard_group_preallocations(struct super_block *sb,
3866 ext4_group_t group, int needed)
3868 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3869 struct buffer_head *bitmap_bh = NULL;
3870 struct ext4_prealloc_space *pa, *tmp;
3871 struct ext4_allocation_context *ac;
3872 struct list_head list;
3873 struct ext4_buddy e4b;
3878 mb_debug(1, "discard preallocation for group %u\n", group);
3880 if (list_empty(&grp->bb_prealloc_list))
3883 bitmap_bh = ext4_read_block_bitmap(sb, group);
3884 if (bitmap_bh == NULL) {
3885 ext4_error(sb, __func__, "Error in reading block "
3886 "bitmap for %u", group);
3890 err = ext4_mb_load_buddy(sb, group, &e4b);
3892 ext4_error(sb, __func__, "Error in loading buddy "
3893 "information for %u", group);
3899 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3901 INIT_LIST_HEAD(&list);
3902 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3906 ext4_lock_group(sb, group);
3907 list_for_each_entry_safe(pa, tmp,
3908 &grp->bb_prealloc_list, pa_group_list) {
3909 spin_lock(&pa->pa_lock);
3910 if (atomic_read(&pa->pa_count)) {
3911 spin_unlock(&pa->pa_lock);
3915 if (pa->pa_deleted) {
3916 spin_unlock(&pa->pa_lock);
3920 /* seems this one can be freed ... */
3923 /* we can trust pa_free ... */
3924 free += pa->pa_free;
3926 spin_unlock(&pa->pa_lock);
3928 list_del(&pa->pa_group_list);
3929 list_add(&pa->u.pa_tmp_list, &list);
3932 /* if we still need more blocks and some PAs were used, try again */
3933 if (free < needed && busy) {
3935 ext4_unlock_group(sb, group);
3937 * Yield the CPU here so that we don't get soft lockup
3938 * in non preempt case.
3944 /* found anything to free? */
3945 if (list_empty(&list)) {
3950 /* now free all selected PAs */
3951 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3953 /* remove from object (inode or locality group) */
3954 spin_lock(pa->pa_obj_lock);
3955 list_del_rcu(&pa->pa_inode_list);
3956 spin_unlock(pa->pa_obj_lock);
3958 if (pa->pa_type == MB_GROUP_PA)
3959 ext4_mb_release_group_pa(&e4b, pa, ac);
3961 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3963 list_del(&pa->u.pa_tmp_list);
3964 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3968 ext4_unlock_group(sb, group);
3970 kmem_cache_free(ext4_ac_cachep, ac);
3971 ext4_mb_release_desc(&e4b);
3977 * releases all non-used preallocated blocks for given inode
3979 * It's important to discard preallocations under i_data_sem
3980 * We don't want another block to be served from the prealloc
3981 * space when we are discarding the inode prealloc space.
3983 * FIXME!! Make sure it is valid at all the call sites
3985 void ext4_discard_preallocations(struct inode *inode)
3987 struct ext4_inode_info *ei = EXT4_I(inode);
3988 struct super_block *sb = inode->i_sb;
3989 struct buffer_head *bitmap_bh = NULL;
3990 struct ext4_prealloc_space *pa, *tmp;
3991 struct ext4_allocation_context *ac;
3992 ext4_group_t group = 0;
3993 struct list_head list;
3994 struct ext4_buddy e4b;
3997 if (!S_ISREG(inode->i_mode)) {
3998 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4002 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
4003 trace_ext4_discard_preallocations(inode);
4005 INIT_LIST_HEAD(&list);
4007 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4010 ac->ac_inode = inode;
4013 /* first, collect all pa's in the inode */
4014 spin_lock(&ei->i_prealloc_lock);
4015 while (!list_empty(&ei->i_prealloc_list)) {
4016 pa = list_entry(ei->i_prealloc_list.next,
4017 struct ext4_prealloc_space, pa_inode_list);
4018 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4019 spin_lock(&pa->pa_lock);
4020 if (atomic_read(&pa->pa_count)) {
4021 /* this shouldn't happen often - nobody should
4022 * use preallocation while we're discarding it */
4023 spin_unlock(&pa->pa_lock);
4024 spin_unlock(&ei->i_prealloc_lock);
4025 printk(KERN_ERR "uh-oh! used pa while discarding\n");
4027 schedule_timeout_uninterruptible(HZ);
4031 if (pa->pa_deleted == 0) {
4033 spin_unlock(&pa->pa_lock);
4034 list_del_rcu(&pa->pa_inode_list);
4035 list_add(&pa->u.pa_tmp_list, &list);
4039 /* someone is deleting pa right now */
4040 spin_unlock(&pa->pa_lock);
4041 spin_unlock(&ei->i_prealloc_lock);
4043 /* we have to wait here because pa_deleted
4044 * doesn't mean pa is already unlinked from
4045 * the list. as we might be called from
4046 * ->clear_inode() the inode will get freed
4047 * and concurrent thread which is unlinking
4048 * pa from inode's list may access already
4049 * freed memory, bad-bad-bad */
4051 /* XXX: if this happens too often, we can
4052 * add a flag to force wait only in case
4053 * of ->clear_inode(), but not in case of
4054 * regular truncate */
4055 schedule_timeout_uninterruptible(HZ);
4058 spin_unlock(&ei->i_prealloc_lock);
4060 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4061 BUG_ON(pa->pa_type != MB_INODE_PA);
4062 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4064 err = ext4_mb_load_buddy(sb, group, &e4b);
4066 ext4_error(sb, __func__, "Error in loading buddy "
4067 "information for %u", group);
4071 bitmap_bh = ext4_read_block_bitmap(sb, group);
4072 if (bitmap_bh == NULL) {
4073 ext4_error(sb, __func__, "Error in reading block "
4074 "bitmap for %u", group);
4075 ext4_mb_release_desc(&e4b);
4079 ext4_lock_group(sb, group);
4080 list_del(&pa->pa_group_list);
4081 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
4082 ext4_unlock_group(sb, group);
4084 ext4_mb_release_desc(&e4b);
4087 list_del(&pa->u.pa_tmp_list);
4088 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4091 kmem_cache_free(ext4_ac_cachep, ac);
4095 * finds all preallocated spaces and return blocks being freed to them
4096 * if preallocated space becomes full (no block is used from the space)
4097 * then the function frees space in buddy
4098 * XXX: at the moment, truncate (which is the only way to free blocks)
4099 * discards all preallocations
4101 static void ext4_mb_return_to_preallocation(struct inode *inode,
4102 struct ext4_buddy *e4b,
4103 sector_t block, int count)
4105 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
4107 #ifdef CONFIG_EXT4_DEBUG
4108 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4110 struct super_block *sb = ac->ac_sb;
4111 ext4_group_t ngroups, i;
4113 printk(KERN_ERR "EXT4-fs: Can't allocate:"
4114 " Allocation context details:\n");
4115 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
4116 ac->ac_status, ac->ac_flags);
4117 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4118 "best %lu/%lu/%lu@%lu cr %d\n",
4119 (unsigned long)ac->ac_o_ex.fe_group,
4120 (unsigned long)ac->ac_o_ex.fe_start,
4121 (unsigned long)ac->ac_o_ex.fe_len,
4122 (unsigned long)ac->ac_o_ex.fe_logical,
4123 (unsigned long)ac->ac_g_ex.fe_group,
4124 (unsigned long)ac->ac_g_ex.fe_start,
4125 (unsigned long)ac->ac_g_ex.fe_len,
4126 (unsigned long)ac->ac_g_ex.fe_logical,
4127 (unsigned long)ac->ac_b_ex.fe_group,
4128 (unsigned long)ac->ac_b_ex.fe_start,
4129 (unsigned long)ac->ac_b_ex.fe_len,
4130 (unsigned long)ac->ac_b_ex.fe_logical,
4131 (int)ac->ac_criteria);
4132 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
4134 printk(KERN_ERR "EXT4-fs: groups: \n");
4135 ngroups = ext4_get_groups_count(sb);
4136 for (i = 0; i < ngroups; i++) {
4137 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4138 struct ext4_prealloc_space *pa;
4139 ext4_grpblk_t start;
4140 struct list_head *cur;
4141 ext4_lock_group(sb, i);
4142 list_for_each(cur, &grp->bb_prealloc_list) {
4143 pa = list_entry(cur, struct ext4_prealloc_space,
4145 spin_lock(&pa->pa_lock);
4146 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4148 spin_unlock(&pa->pa_lock);
4149 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4152 ext4_unlock_group(sb, i);
4154 if (grp->bb_free == 0)
4156 printk(KERN_ERR "%u: %d/%d \n",
4157 i, grp->bb_free, grp->bb_fragments);
4159 printk(KERN_ERR "\n");
4162 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4169 * We use locality group preallocation for small size file. The size of the
4170 * file is determined by the current size or the resulting size after
4171 * allocation which ever is larger
4173 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4175 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4177 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4178 int bsbits = ac->ac_sb->s_blocksize_bits;
4181 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4184 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4187 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4188 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4191 if ((size == isize) &&
4192 !ext4_fs_is_busy(sbi) &&
4193 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4194 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4198 /* don't use group allocation for large files */
4199 size = max(size, isize);
4200 if (size >= sbi->s_mb_stream_request) {
4201 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4205 BUG_ON(ac->ac_lg != NULL);
4207 * locality group prealloc space are per cpu. The reason for having
4208 * per cpu locality group is to reduce the contention between block
4209 * request from multiple CPUs.
4211 ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id());
4213 /* we're going to use group allocation */
4214 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4216 /* serialize all allocations in the group */
4217 mutex_lock(&ac->ac_lg->lg_mutex);
4220 static noinline_for_stack int
4221 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4222 struct ext4_allocation_request *ar)
4224 struct super_block *sb = ar->inode->i_sb;
4225 struct ext4_sb_info *sbi = EXT4_SB(sb);
4226 struct ext4_super_block *es = sbi->s_es;
4230 ext4_grpblk_t block;
4232 /* we can't allocate > group size */
4235 /* just a dirty hack to filter too big requests */
4236 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4237 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4239 /* start searching from the goal */
4241 if (goal < le32_to_cpu(es->s_first_data_block) ||
4242 goal >= ext4_blocks_count(es))
4243 goal = le32_to_cpu(es->s_first_data_block);
4244 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4246 /* set up allocation goals */
4247 memset(ac, 0, sizeof(struct ext4_allocation_context));
4248 ac->ac_b_ex.fe_logical = ar->logical;
4249 ac->ac_status = AC_STATUS_CONTINUE;
4251 ac->ac_inode = ar->inode;
4252 ac->ac_o_ex.fe_logical = ar->logical;
4253 ac->ac_o_ex.fe_group = group;
4254 ac->ac_o_ex.fe_start = block;
4255 ac->ac_o_ex.fe_len = len;
4256 ac->ac_g_ex.fe_logical = ar->logical;
4257 ac->ac_g_ex.fe_group = group;
4258 ac->ac_g_ex.fe_start = block;
4259 ac->ac_g_ex.fe_len = len;
4260 ac->ac_flags = ar->flags;
4262 /* we have to define context: we'll we work with a file or
4263 * locality group. this is a policy, actually */
4264 ext4_mb_group_or_file(ac);
4266 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4267 "left: %u/%u, right %u/%u to %swritable\n",
4268 (unsigned) ar->len, (unsigned) ar->logical,
4269 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4270 (unsigned) ar->lleft, (unsigned) ar->pleft,
4271 (unsigned) ar->lright, (unsigned) ar->pright,
4272 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4277 static noinline_for_stack void
4278 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4279 struct ext4_locality_group *lg,
4280 int order, int total_entries)
4282 ext4_group_t group = 0;
4283 struct ext4_buddy e4b;
4284 struct list_head discard_list;
4285 struct ext4_prealloc_space *pa, *tmp;
4286 struct ext4_allocation_context *ac;
4288 mb_debug(1, "discard locality group preallocation\n");
4290 INIT_LIST_HEAD(&discard_list);
4291 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4295 spin_lock(&lg->lg_prealloc_lock);
4296 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4298 spin_lock(&pa->pa_lock);
4299 if (atomic_read(&pa->pa_count)) {
4301 * This is the pa that we just used
4302 * for block allocation. So don't
4305 spin_unlock(&pa->pa_lock);
4308 if (pa->pa_deleted) {
4309 spin_unlock(&pa->pa_lock);
4312 /* only lg prealloc space */
4313 BUG_ON(pa->pa_type != MB_GROUP_PA);
4315 /* seems this one can be freed ... */
4317 spin_unlock(&pa->pa_lock);
4319 list_del_rcu(&pa->pa_inode_list);
4320 list_add(&pa->u.pa_tmp_list, &discard_list);
4323 if (total_entries <= 5) {
4325 * we want to keep only 5 entries
4326 * allowing it to grow to 8. This
4327 * mak sure we don't call discard
4328 * soon for this list.
4333 spin_unlock(&lg->lg_prealloc_lock);
4335 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4337 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4338 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4339 ext4_error(sb, __func__, "Error in loading buddy "
4340 "information for %u", group);
4343 ext4_lock_group(sb, group);
4344 list_del(&pa->pa_group_list);
4345 ext4_mb_release_group_pa(&e4b, pa, ac);
4346 ext4_unlock_group(sb, group);
4348 ext4_mb_release_desc(&e4b);
4349 list_del(&pa->u.pa_tmp_list);
4350 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4353 kmem_cache_free(ext4_ac_cachep, ac);
4357 * We have incremented pa_count. So it cannot be freed at this
4358 * point. Also we hold lg_mutex. So no parallel allocation is
4359 * possible from this lg. That means pa_free cannot be updated.
4361 * A parallel ext4_mb_discard_group_preallocations is possible.
4362 * which can cause the lg_prealloc_list to be updated.
4365 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4367 int order, added = 0, lg_prealloc_count = 1;
4368 struct super_block *sb = ac->ac_sb;
4369 struct ext4_locality_group *lg = ac->ac_lg;
4370 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4372 order = fls(pa->pa_free) - 1;
4373 if (order > PREALLOC_TB_SIZE - 1)
4374 /* The max size of hash table is PREALLOC_TB_SIZE */
4375 order = PREALLOC_TB_SIZE - 1;
4376 /* Add the prealloc space to lg */
4378 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4380 spin_lock(&tmp_pa->pa_lock);
4381 if (tmp_pa->pa_deleted) {
4382 spin_unlock(&tmp_pa->pa_lock);
4385 if (!added && pa->pa_free < tmp_pa->pa_free) {
4386 /* Add to the tail of the previous entry */
4387 list_add_tail_rcu(&pa->pa_inode_list,
4388 &tmp_pa->pa_inode_list);
4391 * we want to count the total
4392 * number of entries in the list
4395 spin_unlock(&tmp_pa->pa_lock);
4396 lg_prealloc_count++;
4399 list_add_tail_rcu(&pa->pa_inode_list,
4400 &lg->lg_prealloc_list[order]);
4403 /* Now trim the list to be not more than 8 elements */
4404 if (lg_prealloc_count > 8) {
4405 ext4_mb_discard_lg_preallocations(sb, lg,
4406 order, lg_prealloc_count);
4413 * release all resource we used in allocation
4415 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4417 struct ext4_prealloc_space *pa = ac->ac_pa;
4419 if (pa->pa_type == MB_GROUP_PA) {
4420 /* see comment in ext4_mb_use_group_pa() */
4421 spin_lock(&pa->pa_lock);
4422 pa->pa_pstart += ac->ac_b_ex.fe_len;
4423 pa->pa_lstart += ac->ac_b_ex.fe_len;
4424 pa->pa_free -= ac->ac_b_ex.fe_len;
4425 pa->pa_len -= ac->ac_b_ex.fe_len;
4426 spin_unlock(&pa->pa_lock);
4430 up_read(ac->alloc_semp);
4433 * We want to add the pa to the right bucket.
4434 * Remove it from the list and while adding
4435 * make sure the list to which we are adding
4436 * doesn't grow big. We need to release
4437 * alloc_semp before calling ext4_mb_add_n_trim()
4439 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4440 spin_lock(pa->pa_obj_lock);
4441 list_del_rcu(&pa->pa_inode_list);
4442 spin_unlock(pa->pa_obj_lock);
4443 ext4_mb_add_n_trim(ac);
4445 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4447 if (ac->ac_bitmap_page)
4448 page_cache_release(ac->ac_bitmap_page);
4449 if (ac->ac_buddy_page)
4450 page_cache_release(ac->ac_buddy_page);
4451 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4452 mutex_unlock(&ac->ac_lg->lg_mutex);
4453 ext4_mb_collect_stats(ac);
4457 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4459 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4463 trace_ext4_mb_discard_preallocations(sb, needed);
4464 for (i = 0; i < ngroups && needed > 0; i++) {
4465 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4474 * Main entry point into mballoc to allocate blocks
4475 * it tries to use preallocation first, then falls back
4476 * to usual allocation
4478 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4479 struct ext4_allocation_request *ar, int *errp)
4482 struct ext4_allocation_context *ac = NULL;
4483 struct ext4_sb_info *sbi;
4484 struct super_block *sb;
4485 ext4_fsblk_t block = 0;
4486 unsigned int inquota = 0;
4487 unsigned int reserv_blks = 0;
4489 sb = ar->inode->i_sb;
4492 trace_ext4_request_blocks(ar);
4495 * For delayed allocation, we could skip the ENOSPC and
4496 * EDQUOT check, as blocks and quotas have been already
4497 * reserved when data being copied into pagecache.
4499 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4500 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4502 /* Without delayed allocation we need to verify
4503 * there is enough free blocks to do block allocation
4504 * and verify allocation doesn't exceed the quota limits.
4506 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4507 /* let others to free the space */
4509 ar->len = ar->len >> 1;
4515 reserv_blks = ar->len;
4516 while (ar->len && vfs_dq_alloc_block(ar->inode, ar->len)) {
4517 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4527 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4534 *errp = ext4_mb_initialize_context(ac, ar);
4540 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4541 if (!ext4_mb_use_preallocated(ac)) {
4542 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4543 ext4_mb_normalize_request(ac, ar);
4545 /* allocate space in core */
4546 ext4_mb_regular_allocator(ac);
4548 /* as we've just preallocated more space than
4549 * user requested orinally, we store allocated
4550 * space in a special descriptor */
4551 if (ac->ac_status == AC_STATUS_FOUND &&
4552 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4553 ext4_mb_new_preallocation(ac);
4555 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4556 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4557 if (*errp == -EAGAIN) {
4559 * drop the reference that we took
4560 * in ext4_mb_use_best_found
4562 ext4_mb_release_context(ac);
4563 ac->ac_b_ex.fe_group = 0;
4564 ac->ac_b_ex.fe_start = 0;
4565 ac->ac_b_ex.fe_len = 0;
4566 ac->ac_status = AC_STATUS_CONTINUE;
4569 ac->ac_b_ex.fe_len = 0;
4571 ext4_mb_show_ac(ac);
4573 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4574 ar->len = ac->ac_b_ex.fe_len;
4577 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4581 ac->ac_b_ex.fe_len = 0;
4583 ext4_mb_show_ac(ac);
4586 ext4_mb_release_context(ac);
4589 kmem_cache_free(ext4_ac_cachep, ac);
4591 if (inquota && ar->len < inquota)
4592 vfs_dq_free_block(ar->inode, inquota - ar->len);
4595 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4596 /* release all the reserved blocks if non delalloc */
4597 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4601 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4607 * We can merge two free data extents only if the physical blocks
4608 * are contiguous, AND the extents were freed by the same transaction,
4609 * AND the blocks are associated with the same group.
4611 static int can_merge(struct ext4_free_data *entry1,
4612 struct ext4_free_data *entry2)
4614 if ((entry1->t_tid == entry2->t_tid) &&
4615 (entry1->group == entry2->group) &&
4616 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4621 static noinline_for_stack int
4622 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4623 struct ext4_free_data *new_entry)
4625 ext4_grpblk_t block;
4626 struct ext4_free_data *entry;
4627 struct ext4_group_info *db = e4b->bd_info;
4628 struct super_block *sb = e4b->bd_sb;
4629 struct ext4_sb_info *sbi = EXT4_SB(sb);
4630 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4631 struct rb_node *parent = NULL, *new_node;
4633 BUG_ON(!ext4_handle_valid(handle));
4634 BUG_ON(e4b->bd_bitmap_page == NULL);
4635 BUG_ON(e4b->bd_buddy_page == NULL);
4637 new_node = &new_entry->node;
4638 block = new_entry->start_blk;
4641 /* first free block exent. We need to
4642 protect buddy cache from being freed,
4643 * otherwise we'll refresh it from
4644 * on-disk bitmap and lose not-yet-available
4646 page_cache_get(e4b->bd_buddy_page);
4647 page_cache_get(e4b->bd_bitmap_page);
4651 entry = rb_entry(parent, struct ext4_free_data, node);
4652 if (block < entry->start_blk)
4654 else if (block >= (entry->start_blk + entry->count))
4655 n = &(*n)->rb_right;
4657 ext4_grp_locked_error(sb, e4b->bd_group, __func__,
4658 "Double free of blocks %d (%d %d)",
4659 block, entry->start_blk, entry->count);
4664 rb_link_node(new_node, parent, n);
4665 rb_insert_color(new_node, &db->bb_free_root);
4667 /* Now try to see the extent can be merged to left and right */
4668 node = rb_prev(new_node);
4670 entry = rb_entry(node, struct ext4_free_data, node);
4671 if (can_merge(entry, new_entry)) {
4672 new_entry->start_blk = entry->start_blk;
4673 new_entry->count += entry->count;
4674 rb_erase(node, &(db->bb_free_root));
4675 spin_lock(&sbi->s_md_lock);
4676 list_del(&entry->list);
4677 spin_unlock(&sbi->s_md_lock);
4678 kmem_cache_free(ext4_free_ext_cachep, entry);
4682 node = rb_next(new_node);
4684 entry = rb_entry(node, struct ext4_free_data, node);
4685 if (can_merge(new_entry, entry)) {
4686 new_entry->count += entry->count;
4687 rb_erase(node, &(db->bb_free_root));
4688 spin_lock(&sbi->s_md_lock);
4689 list_del(&entry->list);
4690 spin_unlock(&sbi->s_md_lock);
4691 kmem_cache_free(ext4_free_ext_cachep, entry);
4694 /* Add the extent to transaction's private list */
4695 spin_lock(&sbi->s_md_lock);
4696 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4697 spin_unlock(&sbi->s_md_lock);
4702 * Main entry point into mballoc to free blocks
4704 void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
4705 ext4_fsblk_t block, unsigned long count,
4706 int metadata, unsigned long *freed)
4708 struct buffer_head *bitmap_bh = NULL;
4709 struct super_block *sb = inode->i_sb;
4710 struct ext4_allocation_context *ac = NULL;
4711 struct ext4_group_desc *gdp;
4712 struct ext4_super_block *es;
4713 unsigned int overflow;
4715 struct buffer_head *gd_bh;
4716 ext4_group_t block_group;
4717 struct ext4_sb_info *sbi;
4718 struct ext4_buddy e4b;
4725 es = EXT4_SB(sb)->s_es;
4726 if (block < le32_to_cpu(es->s_first_data_block) ||
4727 block + count < block ||
4728 block + count > ext4_blocks_count(es)) {
4729 ext4_error(sb, __func__,
4730 "Freeing blocks not in datazone - "
4731 "block = %llu, count = %lu", block, count);
4735 ext4_debug("freeing block %llu\n", block);
4736 trace_ext4_free_blocks(inode, block, count, metadata);
4738 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4740 ac->ac_op = EXT4_MB_HISTORY_FREE;
4741 ac->ac_inode = inode;
4747 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4750 * Check to see if we are freeing blocks across a group
4753 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4754 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4757 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4762 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4768 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4769 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4770 in_range(block, ext4_inode_table(sb, gdp),
4771 EXT4_SB(sb)->s_itb_per_group) ||
4772 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4773 EXT4_SB(sb)->s_itb_per_group)) {
4775 ext4_error(sb, __func__,
4776 "Freeing blocks in system zone - "
4777 "Block = %llu, count = %lu", block, count);
4778 /* err = 0. ext4_std_error should be a no op */
4782 BUFFER_TRACE(bitmap_bh, "getting write access");
4783 err = ext4_journal_get_write_access(handle, bitmap_bh);
4788 * We are about to modify some metadata. Call the journal APIs
4789 * to unshare ->b_data if a currently-committing transaction is
4792 BUFFER_TRACE(gd_bh, "get_write_access");
4793 err = ext4_journal_get_write_access(handle, gd_bh);
4796 #ifdef AGGRESSIVE_CHECK
4799 for (i = 0; i < count; i++)
4800 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4804 ac->ac_b_ex.fe_group = block_group;
4805 ac->ac_b_ex.fe_start = bit;
4806 ac->ac_b_ex.fe_len = count;
4807 ext4_mb_store_history(ac);
4810 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4813 if (metadata && ext4_handle_valid(handle)) {
4814 struct ext4_free_data *new_entry;
4816 * blocks being freed are metadata. these blocks shouldn't
4817 * be used until this transaction is committed
4819 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4820 new_entry->start_blk = bit;
4821 new_entry->group = block_group;
4822 new_entry->count = count;
4823 new_entry->t_tid = handle->h_transaction->t_tid;
4825 ext4_lock_group(sb, block_group);
4826 mb_clear_bits(bitmap_bh->b_data, bit, count);
4827 ext4_mb_free_metadata(handle, &e4b, new_entry);
4829 /* need to update group_info->bb_free and bitmap
4830 * with group lock held. generate_buddy look at
4831 * them with group lock_held
4833 ext4_lock_group(sb, block_group);
4834 mb_clear_bits(bitmap_bh->b_data, bit, count);
4835 mb_free_blocks(inode, &e4b, bit, count);
4836 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4839 ret = ext4_free_blks_count(sb, gdp) + count;
4840 ext4_free_blks_set(sb, gdp, ret);
4841 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4842 ext4_unlock_group(sb, block_group);
4843 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4845 if (sbi->s_log_groups_per_flex) {
4846 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4847 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4850 ext4_mb_release_desc(&e4b);
4854 /* We dirtied the bitmap block */
4855 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4856 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4858 /* And the group descriptor block */
4859 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4860 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4864 if (overflow && !err) {
4873 ext4_std_error(sb, err);
4875 kmem_cache_free(ext4_ac_cachep, ac);