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 we don't add group
932 * to this buddy cache via resize
934 num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group);
935 if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
937 * somebody initialized the group
938 * return without doing anything
944 * the buddy cache inode stores the block bitmap
945 * and buddy information in consecutive blocks.
946 * So for each group we need two blocks.
949 pnum = block / blocks_per_page;
950 poff = block % blocks_per_page;
951 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
953 BUG_ON(page->mapping != inode->i_mapping);
954 ret = ext4_mb_init_cache(page, NULL);
961 if (page == NULL || !PageUptodate(page)) {
965 mark_page_accessed(page);
967 bitmap = page_address(page) + (poff * sb->s_blocksize);
969 /* init buddy cache */
971 pnum = block / blocks_per_page;
972 poff = block % blocks_per_page;
973 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
974 if (page == bitmap_page) {
976 * If both the bitmap and buddy are in
977 * the same page we don't need to force
982 BUG_ON(page->mapping != inode->i_mapping);
983 ret = ext4_mb_init_cache(page, bitmap);
990 if (page == NULL || !PageUptodate(page)) {
994 mark_page_accessed(page);
996 ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
998 page_cache_release(bitmap_page);
1000 page_cache_release(page);
1004 static noinline_for_stack int
1005 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1006 struct ext4_buddy *e4b)
1008 int blocks_per_page;
1014 struct ext4_group_info *grp;
1015 struct ext4_sb_info *sbi = EXT4_SB(sb);
1016 struct inode *inode = sbi->s_buddy_cache;
1018 mb_debug(1, "load group %u\n", group);
1020 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1021 grp = ext4_get_group_info(sb, group);
1023 e4b->bd_blkbits = sb->s_blocksize_bits;
1024 e4b->bd_info = ext4_get_group_info(sb, group);
1026 e4b->bd_group = group;
1027 e4b->bd_buddy_page = NULL;
1028 e4b->bd_bitmap_page = NULL;
1029 e4b->alloc_semp = &grp->alloc_sem;
1031 /* Take the read lock on the group alloc
1032 * sem. This would make sure a parallel
1033 * ext4_mb_init_group happening on other
1034 * groups mapped by the page is blocked
1035 * till we are done with allocation
1037 down_read(e4b->alloc_semp);
1040 * the buddy cache inode stores the block bitmap
1041 * and buddy information in consecutive blocks.
1042 * So for each group we need two blocks.
1045 pnum = block / blocks_per_page;
1046 poff = block % blocks_per_page;
1048 /* we could use find_or_create_page(), but it locks page
1049 * what we'd like to avoid in fast path ... */
1050 page = find_get_page(inode->i_mapping, pnum);
1051 if (page == NULL || !PageUptodate(page)) {
1054 * drop the page reference and try
1055 * to get the page with lock. If we
1056 * are not uptodate that implies
1057 * somebody just created the page but
1058 * is yet to initialize the same. So
1059 * wait for it to initialize.
1061 page_cache_release(page);
1062 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1064 BUG_ON(page->mapping != inode->i_mapping);
1065 if (!PageUptodate(page)) {
1066 ret = ext4_mb_init_cache(page, NULL);
1071 mb_cmp_bitmaps(e4b, page_address(page) +
1072 (poff * sb->s_blocksize));
1077 if (page == NULL || !PageUptodate(page)) {
1081 e4b->bd_bitmap_page = page;
1082 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1083 mark_page_accessed(page);
1086 pnum = block / blocks_per_page;
1087 poff = block % blocks_per_page;
1089 page = find_get_page(inode->i_mapping, pnum);
1090 if (page == NULL || !PageUptodate(page)) {
1092 page_cache_release(page);
1093 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1095 BUG_ON(page->mapping != inode->i_mapping);
1096 if (!PageUptodate(page)) {
1097 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1106 if (page == NULL || !PageUptodate(page)) {
1110 e4b->bd_buddy_page = page;
1111 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1112 mark_page_accessed(page);
1114 BUG_ON(e4b->bd_bitmap_page == NULL);
1115 BUG_ON(e4b->bd_buddy_page == NULL);
1120 if (e4b->bd_bitmap_page)
1121 page_cache_release(e4b->bd_bitmap_page);
1122 if (e4b->bd_buddy_page)
1123 page_cache_release(e4b->bd_buddy_page);
1124 e4b->bd_buddy = NULL;
1125 e4b->bd_bitmap = NULL;
1127 /* Done with the buddy cache */
1128 up_read(e4b->alloc_semp);
1132 static void ext4_mb_release_desc(struct ext4_buddy *e4b)
1134 if (e4b->bd_bitmap_page)
1135 page_cache_release(e4b->bd_bitmap_page);
1136 if (e4b->bd_buddy_page)
1137 page_cache_release(e4b->bd_buddy_page);
1138 /* Done with the buddy cache */
1139 if (e4b->alloc_semp)
1140 up_read(e4b->alloc_semp);
1144 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1149 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1150 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1152 bb = EXT4_MB_BUDDY(e4b);
1153 while (order <= e4b->bd_blkbits + 1) {
1155 if (!mb_test_bit(block, bb)) {
1156 /* this block is part of buddy of order 'order' */
1159 bb += 1 << (e4b->bd_blkbits - order);
1165 static void mb_clear_bits(void *bm, int cur, int len)
1171 if ((cur & 31) == 0 && (len - cur) >= 32) {
1172 /* fast path: clear whole word at once */
1173 addr = bm + (cur >> 3);
1178 mb_clear_bit(cur, bm);
1183 static void mb_set_bits(void *bm, int cur, int len)
1189 if ((cur & 31) == 0 && (len - cur) >= 32) {
1190 /* fast path: set whole word at once */
1191 addr = bm + (cur >> 3);
1196 mb_set_bit(cur, bm);
1201 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1202 int first, int count)
1209 struct super_block *sb = e4b->bd_sb;
1211 BUG_ON(first + count > (sb->s_blocksize << 3));
1212 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1213 mb_check_buddy(e4b);
1214 mb_free_blocks_double(inode, e4b, first, count);
1216 e4b->bd_info->bb_free += count;
1217 if (first < e4b->bd_info->bb_first_free)
1218 e4b->bd_info->bb_first_free = first;
1220 /* let's maintain fragments counter */
1222 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1223 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1224 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1226 e4b->bd_info->bb_fragments--;
1227 else if (!block && !max)
1228 e4b->bd_info->bb_fragments++;
1230 /* let's maintain buddy itself */
1231 while (count-- > 0) {
1235 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1236 ext4_fsblk_t blocknr;
1237 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1240 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
1241 ext4_grp_locked_error(sb, e4b->bd_group,
1242 __func__, "double-free of inode"
1243 " %lu's block %llu(bit %u in group %u)",
1244 inode ? inode->i_ino : 0, blocknr, block,
1247 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1248 e4b->bd_info->bb_counters[order]++;
1250 /* start of the buddy */
1251 buddy = mb_find_buddy(e4b, order, &max);
1255 if (mb_test_bit(block, buddy) ||
1256 mb_test_bit(block + 1, buddy))
1259 /* both the buddies are free, try to coalesce them */
1260 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1266 /* for special purposes, we don't set
1267 * free bits in bitmap */
1268 mb_set_bit(block, buddy);
1269 mb_set_bit(block + 1, buddy);
1271 e4b->bd_info->bb_counters[order]--;
1272 e4b->bd_info->bb_counters[order]--;
1276 e4b->bd_info->bb_counters[order]++;
1278 mb_clear_bit(block, buddy2);
1282 mb_check_buddy(e4b);
1285 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1286 int needed, struct ext4_free_extent *ex)
1293 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1296 buddy = mb_find_buddy(e4b, order, &max);
1297 BUG_ON(buddy == NULL);
1298 BUG_ON(block >= max);
1299 if (mb_test_bit(block, buddy)) {
1306 /* FIXME dorp order completely ? */
1307 if (likely(order == 0)) {
1308 /* find actual order */
1309 order = mb_find_order_for_block(e4b, block);
1310 block = block >> order;
1313 ex->fe_len = 1 << order;
1314 ex->fe_start = block << order;
1315 ex->fe_group = e4b->bd_group;
1317 /* calc difference from given start */
1318 next = next - ex->fe_start;
1320 ex->fe_start += next;
1322 while (needed > ex->fe_len &&
1323 (buddy = mb_find_buddy(e4b, order, &max))) {
1325 if (block + 1 >= max)
1328 next = (block + 1) * (1 << order);
1329 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1332 ord = mb_find_order_for_block(e4b, next);
1335 block = next >> order;
1336 ex->fe_len += 1 << order;
1339 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1343 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1349 int start = ex->fe_start;
1350 int len = ex->fe_len;
1355 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1356 BUG_ON(e4b->bd_group != ex->fe_group);
1357 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1358 mb_check_buddy(e4b);
1359 mb_mark_used_double(e4b, start, len);
1361 e4b->bd_info->bb_free -= len;
1362 if (e4b->bd_info->bb_first_free == start)
1363 e4b->bd_info->bb_first_free += len;
1365 /* let's maintain fragments counter */
1367 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1368 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1369 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1371 e4b->bd_info->bb_fragments++;
1372 else if (!mlen && !max)
1373 e4b->bd_info->bb_fragments--;
1375 /* let's maintain buddy itself */
1377 ord = mb_find_order_for_block(e4b, start);
1379 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1380 /* the whole chunk may be allocated at once! */
1382 buddy = mb_find_buddy(e4b, ord, &max);
1383 BUG_ON((start >> ord) >= max);
1384 mb_set_bit(start >> ord, buddy);
1385 e4b->bd_info->bb_counters[ord]--;
1392 /* store for history */
1394 ret = len | (ord << 16);
1396 /* we have to split large buddy */
1398 buddy = mb_find_buddy(e4b, ord, &max);
1399 mb_set_bit(start >> ord, buddy);
1400 e4b->bd_info->bb_counters[ord]--;
1403 cur = (start >> ord) & ~1U;
1404 buddy = mb_find_buddy(e4b, ord, &max);
1405 mb_clear_bit(cur, buddy);
1406 mb_clear_bit(cur + 1, buddy);
1407 e4b->bd_info->bb_counters[ord]++;
1408 e4b->bd_info->bb_counters[ord]++;
1411 mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1412 mb_check_buddy(e4b);
1418 * Must be called under group lock!
1420 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1421 struct ext4_buddy *e4b)
1423 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1426 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1427 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1429 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1430 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1431 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1433 /* preallocation can change ac_b_ex, thus we store actually
1434 * allocated blocks for history */
1435 ac->ac_f_ex = ac->ac_b_ex;
1437 ac->ac_status = AC_STATUS_FOUND;
1438 ac->ac_tail = ret & 0xffff;
1439 ac->ac_buddy = ret >> 16;
1442 * take the page reference. We want the page to be pinned
1443 * so that we don't get a ext4_mb_init_cache_call for this
1444 * group until we update the bitmap. That would mean we
1445 * double allocate blocks. The reference is dropped
1446 * in ext4_mb_release_context
1448 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1449 get_page(ac->ac_bitmap_page);
1450 ac->ac_buddy_page = e4b->bd_buddy_page;
1451 get_page(ac->ac_buddy_page);
1452 /* on allocation we use ac to track the held semaphore */
1453 ac->alloc_semp = e4b->alloc_semp;
1454 e4b->alloc_semp = NULL;
1455 /* store last allocated for subsequent stream allocation */
1456 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1457 spin_lock(&sbi->s_md_lock);
1458 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1459 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1460 spin_unlock(&sbi->s_md_lock);
1465 * regular allocator, for general purposes allocation
1468 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1469 struct ext4_buddy *e4b,
1472 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1473 struct ext4_free_extent *bex = &ac->ac_b_ex;
1474 struct ext4_free_extent *gex = &ac->ac_g_ex;
1475 struct ext4_free_extent ex;
1478 if (ac->ac_status == AC_STATUS_FOUND)
1481 * We don't want to scan for a whole year
1483 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1484 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1485 ac->ac_status = AC_STATUS_BREAK;
1490 * Haven't found good chunk so far, let's continue
1492 if (bex->fe_len < gex->fe_len)
1495 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1496 && bex->fe_group == e4b->bd_group) {
1497 /* recheck chunk's availability - we don't know
1498 * when it was found (within this lock-unlock
1500 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1501 if (max >= gex->fe_len) {
1502 ext4_mb_use_best_found(ac, e4b);
1509 * The routine checks whether found extent is good enough. If it is,
1510 * then the extent gets marked used and flag is set to the context
1511 * to stop scanning. Otherwise, the extent is compared with the
1512 * previous found extent and if new one is better, then it's stored
1513 * in the context. Later, the best found extent will be used, if
1514 * mballoc can't find good enough extent.
1516 * FIXME: real allocation policy is to be designed yet!
1518 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1519 struct ext4_free_extent *ex,
1520 struct ext4_buddy *e4b)
1522 struct ext4_free_extent *bex = &ac->ac_b_ex;
1523 struct ext4_free_extent *gex = &ac->ac_g_ex;
1525 BUG_ON(ex->fe_len <= 0);
1526 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1527 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1528 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1533 * The special case - take what you catch first
1535 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1537 ext4_mb_use_best_found(ac, e4b);
1542 * Let's check whether the chuck is good enough
1544 if (ex->fe_len == gex->fe_len) {
1546 ext4_mb_use_best_found(ac, e4b);
1551 * If this is first found extent, just store it in the context
1553 if (bex->fe_len == 0) {
1559 * If new found extent is better, store it in the context
1561 if (bex->fe_len < gex->fe_len) {
1562 /* if the request isn't satisfied, any found extent
1563 * larger than previous best one is better */
1564 if (ex->fe_len > bex->fe_len)
1566 } else if (ex->fe_len > gex->fe_len) {
1567 /* if the request is satisfied, then we try to find
1568 * an extent that still satisfy the request, but is
1569 * smaller than previous one */
1570 if (ex->fe_len < bex->fe_len)
1574 ext4_mb_check_limits(ac, e4b, 0);
1577 static noinline_for_stack
1578 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1579 struct ext4_buddy *e4b)
1581 struct ext4_free_extent ex = ac->ac_b_ex;
1582 ext4_group_t group = ex.fe_group;
1586 BUG_ON(ex.fe_len <= 0);
1587 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1591 ext4_lock_group(ac->ac_sb, group);
1592 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1596 ext4_mb_use_best_found(ac, e4b);
1599 ext4_unlock_group(ac->ac_sb, group);
1600 ext4_mb_release_desc(e4b);
1605 static noinline_for_stack
1606 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1607 struct ext4_buddy *e4b)
1609 ext4_group_t group = ac->ac_g_ex.fe_group;
1612 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1613 struct ext4_super_block *es = sbi->s_es;
1614 struct ext4_free_extent ex;
1616 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1619 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1623 ext4_lock_group(ac->ac_sb, group);
1624 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1625 ac->ac_g_ex.fe_len, &ex);
1627 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1630 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1631 ex.fe_start + le32_to_cpu(es->s_first_data_block);
1632 /* use do_div to get remainder (would be 64-bit modulo) */
1633 if (do_div(start, sbi->s_stripe) == 0) {
1636 ext4_mb_use_best_found(ac, e4b);
1638 } else if (max >= ac->ac_g_ex.fe_len) {
1639 BUG_ON(ex.fe_len <= 0);
1640 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1641 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1644 ext4_mb_use_best_found(ac, e4b);
1645 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1646 /* Sometimes, caller may want to merge even small
1647 * number of blocks to an existing extent */
1648 BUG_ON(ex.fe_len <= 0);
1649 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1650 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1653 ext4_mb_use_best_found(ac, e4b);
1655 ext4_unlock_group(ac->ac_sb, group);
1656 ext4_mb_release_desc(e4b);
1662 * The routine scans buddy structures (not bitmap!) from given order
1663 * to max order and tries to find big enough chunk to satisfy the req
1665 static noinline_for_stack
1666 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1667 struct ext4_buddy *e4b)
1669 struct super_block *sb = ac->ac_sb;
1670 struct ext4_group_info *grp = e4b->bd_info;
1676 BUG_ON(ac->ac_2order <= 0);
1677 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1678 if (grp->bb_counters[i] == 0)
1681 buddy = mb_find_buddy(e4b, i, &max);
1682 BUG_ON(buddy == NULL);
1684 k = mb_find_next_zero_bit(buddy, max, 0);
1689 ac->ac_b_ex.fe_len = 1 << i;
1690 ac->ac_b_ex.fe_start = k << i;
1691 ac->ac_b_ex.fe_group = e4b->bd_group;
1693 ext4_mb_use_best_found(ac, e4b);
1695 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1697 if (EXT4_SB(sb)->s_mb_stats)
1698 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1705 * The routine scans the group and measures all found extents.
1706 * In order to optimize scanning, caller must pass number of
1707 * free blocks in the group, so the routine can know upper limit.
1709 static noinline_for_stack
1710 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1711 struct ext4_buddy *e4b)
1713 struct super_block *sb = ac->ac_sb;
1714 void *bitmap = EXT4_MB_BITMAP(e4b);
1715 struct ext4_free_extent ex;
1719 free = e4b->bd_info->bb_free;
1722 i = e4b->bd_info->bb_first_free;
1724 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1725 i = mb_find_next_zero_bit(bitmap,
1726 EXT4_BLOCKS_PER_GROUP(sb), i);
1727 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1729 * IF we have corrupt bitmap, we won't find any
1730 * free blocks even though group info says we
1731 * we have free blocks
1733 ext4_grp_locked_error(sb, e4b->bd_group,
1734 __func__, "%d free blocks as per "
1735 "group info. But bitmap says 0",
1740 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1741 BUG_ON(ex.fe_len <= 0);
1742 if (free < ex.fe_len) {
1743 ext4_grp_locked_error(sb, e4b->bd_group,
1744 __func__, "%d free blocks as per "
1745 "group info. But got %d blocks",
1748 * The number of free blocks differs. This mostly
1749 * indicate that the bitmap is corrupt. So exit
1750 * without claiming the space.
1755 ext4_mb_measure_extent(ac, &ex, e4b);
1761 ext4_mb_check_limits(ac, e4b, 1);
1765 * This is a special case for storages like raid5
1766 * we try to find stripe-aligned chunks for stripe-size requests
1767 * XXX should do so at least for multiples of stripe size as well
1769 static noinline_for_stack
1770 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1771 struct ext4_buddy *e4b)
1773 struct super_block *sb = ac->ac_sb;
1774 struct ext4_sb_info *sbi = EXT4_SB(sb);
1775 void *bitmap = EXT4_MB_BITMAP(e4b);
1776 struct ext4_free_extent ex;
1777 ext4_fsblk_t first_group_block;
1782 BUG_ON(sbi->s_stripe == 0);
1784 /* find first stripe-aligned block in group */
1785 first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1786 + le32_to_cpu(sbi->s_es->s_first_data_block);
1787 a = first_group_block + sbi->s_stripe - 1;
1788 do_div(a, sbi->s_stripe);
1789 i = (a * sbi->s_stripe) - first_group_block;
1791 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1792 if (!mb_test_bit(i, bitmap)) {
1793 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1794 if (max >= sbi->s_stripe) {
1797 ext4_mb_use_best_found(ac, e4b);
1805 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1806 ext4_group_t group, int cr)
1808 unsigned free, fragments;
1810 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1811 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1813 BUG_ON(cr < 0 || cr >= 4);
1814 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1816 free = grp->bb_free;
1817 fragments = grp->bb_fragments;
1825 BUG_ON(ac->ac_2order == 0);
1827 /* Avoid using the first bg of a flexgroup for data files */
1828 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1829 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1830 ((group % flex_size) == 0))
1833 bits = ac->ac_sb->s_blocksize_bits + 1;
1834 for (i = ac->ac_2order; i <= bits; i++)
1835 if (grp->bb_counters[i] > 0)
1839 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1843 if (free >= ac->ac_g_ex.fe_len)
1856 * lock the group_info alloc_sem of all the groups
1857 * belonging to the same buddy cache page. This
1858 * make sure other parallel operation on the buddy
1859 * cache doesn't happen whild holding the buddy cache
1862 int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1866 int blocks_per_page;
1867 int groups_per_page;
1868 ext4_group_t ngroups = ext4_get_groups_count(sb);
1869 ext4_group_t first_group;
1870 struct ext4_group_info *grp;
1872 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1874 * the buddy cache inode stores the block bitmap
1875 * and buddy information in consecutive blocks.
1876 * So for each group we need two blocks.
1879 pnum = block / blocks_per_page;
1880 first_group = pnum * blocks_per_page / 2;
1882 groups_per_page = blocks_per_page >> 1;
1883 if (groups_per_page == 0)
1884 groups_per_page = 1;
1885 /* read all groups the page covers into the cache */
1886 for (i = 0; i < groups_per_page; i++) {
1888 if ((first_group + i) >= ngroups)
1890 grp = ext4_get_group_info(sb, first_group + i);
1891 /* take all groups write allocation
1892 * semaphore. This make sure there is
1893 * no block allocation going on in any
1896 down_write_nested(&grp->alloc_sem, i);
1901 void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1902 ext4_group_t group, int locked_group)
1906 int blocks_per_page;
1907 ext4_group_t first_group;
1908 struct ext4_group_info *grp;
1910 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1912 * the buddy cache inode stores the block bitmap
1913 * and buddy information in consecutive blocks.
1914 * So for each group we need two blocks.
1917 pnum = block / blocks_per_page;
1918 first_group = pnum * blocks_per_page / 2;
1919 /* release locks on all the groups */
1920 for (i = 0; i < locked_group; i++) {
1922 grp = ext4_get_group_info(sb, first_group + i);
1923 /* take all groups write allocation
1924 * semaphore. This make sure there is
1925 * no block allocation going on in any
1928 up_write(&grp->alloc_sem);
1933 static noinline_for_stack int
1934 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1936 ext4_group_t ngroups, group, i;
1940 struct ext4_sb_info *sbi;
1941 struct super_block *sb;
1942 struct ext4_buddy e4b;
1946 ngroups = ext4_get_groups_count(sb);
1947 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1949 /* first, try the goal */
1950 err = ext4_mb_find_by_goal(ac, &e4b);
1951 if (err || ac->ac_status == AC_STATUS_FOUND)
1954 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1958 * ac->ac2_order is set only if the fe_len is a power of 2
1959 * if ac2_order is set we also set criteria to 0 so that we
1960 * try exact allocation using buddy.
1962 i = fls(ac->ac_g_ex.fe_len);
1965 * We search using buddy data only if the order of the request
1966 * is greater than equal to the sbi_s_mb_order2_reqs
1967 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1969 if (i >= sbi->s_mb_order2_reqs) {
1971 * This should tell if fe_len is exactly power of 2
1973 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1974 ac->ac_2order = i - 1;
1977 bsbits = ac->ac_sb->s_blocksize_bits;
1979 /* if stream allocation is enabled, use global goal */
1980 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1981 /* TBD: may be hot point */
1982 spin_lock(&sbi->s_md_lock);
1983 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1984 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1985 spin_unlock(&sbi->s_md_lock);
1988 /* Let's just scan groups to find more-less suitable blocks */
1989 cr = ac->ac_2order ? 0 : 1;
1991 * cr == 0 try to get exact allocation,
1992 * cr == 3 try to get anything
1995 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1996 ac->ac_criteria = cr;
1998 * searching for the right group start
1999 * from the goal value specified
2001 group = ac->ac_g_ex.fe_group;
2003 for (i = 0; i < ngroups; group++, i++) {
2004 struct ext4_group_info *grp;
2005 struct ext4_group_desc *desc;
2007 if (group == ngroups)
2010 /* quick check to skip empty groups */
2011 grp = ext4_get_group_info(sb, group);
2012 if (grp->bb_free == 0)
2016 * if the group is already init we check whether it is
2017 * a good group and if not we don't load the buddy
2019 if (EXT4_MB_GRP_NEED_INIT(grp)) {
2021 * we need full data about the group
2022 * to make a good selection
2024 err = ext4_mb_init_group(sb, group);
2030 * If the particular group doesn't satisfy our
2031 * criteria we continue with the next group
2033 if (!ext4_mb_good_group(ac, group, cr))
2036 err = ext4_mb_load_buddy(sb, group, &e4b);
2040 ext4_lock_group(sb, group);
2041 if (!ext4_mb_good_group(ac, group, cr)) {
2042 /* someone did allocation from this group */
2043 ext4_unlock_group(sb, group);
2044 ext4_mb_release_desc(&e4b);
2048 ac->ac_groups_scanned++;
2049 desc = ext4_get_group_desc(sb, group, NULL);
2051 ext4_mb_simple_scan_group(ac, &e4b);
2053 ac->ac_g_ex.fe_len == sbi->s_stripe)
2054 ext4_mb_scan_aligned(ac, &e4b);
2056 ext4_mb_complex_scan_group(ac, &e4b);
2058 ext4_unlock_group(sb, group);
2059 ext4_mb_release_desc(&e4b);
2061 if (ac->ac_status != AC_STATUS_CONTINUE)
2066 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2067 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2069 * We've been searching too long. Let's try to allocate
2070 * the best chunk we've found so far
2073 ext4_mb_try_best_found(ac, &e4b);
2074 if (ac->ac_status != AC_STATUS_FOUND) {
2076 * Someone more lucky has already allocated it.
2077 * The only thing we can do is just take first
2079 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2081 ac->ac_b_ex.fe_group = 0;
2082 ac->ac_b_ex.fe_start = 0;
2083 ac->ac_b_ex.fe_len = 0;
2084 ac->ac_status = AC_STATUS_CONTINUE;
2085 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2087 atomic_inc(&sbi->s_mb_lost_chunks);
2095 #ifdef EXT4_MB_HISTORY
2096 struct ext4_mb_proc_session {
2097 struct ext4_mb_history *history;
2098 struct super_block *sb;
2103 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
2104 struct ext4_mb_history *hs,
2107 if (hs == s->history + s->max)
2109 if (!first && hs == s->history + s->start)
2111 while (hs->orig.fe_len == 0) {
2113 if (hs == s->history + s->max)
2115 if (hs == s->history + s->start)
2121 static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
2123 struct ext4_mb_proc_session *s = seq->private;
2124 struct ext4_mb_history *hs;
2128 return SEQ_START_TOKEN;
2129 hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2132 while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
2136 static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
2139 struct ext4_mb_proc_session *s = seq->private;
2140 struct ext4_mb_history *hs = v;
2143 if (v == SEQ_START_TOKEN)
2144 return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2146 return ext4_mb_history_skip_empty(s, ++hs, 0);
2149 static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
2151 char buf[25], buf2[25], buf3[25], *fmt;
2152 struct ext4_mb_history *hs = v;
2154 if (v == SEQ_START_TOKEN) {
2155 seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
2156 "%-5s %-2s %-6s %-5s %-5s %-6s\n",
2157 "pid", "inode", "original", "goal", "result", "found",
2158 "grps", "cr", "flags", "merge", "tail", "broken");
2162 if (hs->op == EXT4_MB_HISTORY_ALLOC) {
2163 fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2164 "0x%04x %-5s %-5u %-6u\n";
2165 sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group,
2166 hs->result.fe_start, hs->result.fe_len,
2167 hs->result.fe_logical);
2168 sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group,
2169 hs->orig.fe_start, hs->orig.fe_len,
2170 hs->orig.fe_logical);
2171 sprintf(buf3, "%u/%d/%u@%u", hs->goal.fe_group,
2172 hs->goal.fe_start, hs->goal.fe_len,
2173 hs->goal.fe_logical);
2174 seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
2175 hs->found, hs->groups, hs->cr, hs->flags,
2176 hs->merged ? "M" : "", hs->tail,
2177 hs->buddy ? 1 << hs->buddy : 0);
2178 } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
2179 fmt = "%-5u %-8u %-23s %-23s %-23s\n";
2180 sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group,
2181 hs->result.fe_start, hs->result.fe_len,
2182 hs->result.fe_logical);
2183 sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group,
2184 hs->orig.fe_start, hs->orig.fe_len,
2185 hs->orig.fe_logical);
2186 seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
2187 } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
2188 sprintf(buf2, "%u/%d/%u", hs->result.fe_group,
2189 hs->result.fe_start, hs->result.fe_len);
2190 seq_printf(seq, "%-5u %-8u %-23s discard\n",
2191 hs->pid, hs->ino, buf2);
2192 } else if (hs->op == EXT4_MB_HISTORY_FREE) {
2193 sprintf(buf2, "%u/%d/%u", hs->result.fe_group,
2194 hs->result.fe_start, hs->result.fe_len);
2195 seq_printf(seq, "%-5u %-8u %-23s free\n",
2196 hs->pid, hs->ino, buf2);
2201 static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
2205 static const struct seq_operations ext4_mb_seq_history_ops = {
2206 .start = ext4_mb_seq_history_start,
2207 .next = ext4_mb_seq_history_next,
2208 .stop = ext4_mb_seq_history_stop,
2209 .show = ext4_mb_seq_history_show,
2212 static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
2214 struct super_block *sb = PDE(inode)->data;
2215 struct ext4_sb_info *sbi = EXT4_SB(sb);
2216 struct ext4_mb_proc_session *s;
2220 if (unlikely(sbi->s_mb_history == NULL))
2222 s = kmalloc(sizeof(*s), GFP_KERNEL);
2226 size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
2227 s->history = kmalloc(size, GFP_KERNEL);
2228 if (s->history == NULL) {
2233 spin_lock(&sbi->s_mb_history_lock);
2234 memcpy(s->history, sbi->s_mb_history, size);
2235 s->max = sbi->s_mb_history_max;
2236 s->start = sbi->s_mb_history_cur % s->max;
2237 spin_unlock(&sbi->s_mb_history_lock);
2239 rc = seq_open(file, &ext4_mb_seq_history_ops);
2241 struct seq_file *m = (struct seq_file *)file->private_data;
2251 static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
2253 struct seq_file *seq = (struct seq_file *)file->private_data;
2254 struct ext4_mb_proc_session *s = seq->private;
2257 return seq_release(inode, file);
2260 static ssize_t ext4_mb_seq_history_write(struct file *file,
2261 const char __user *buffer,
2262 size_t count, loff_t *ppos)
2264 struct seq_file *seq = (struct seq_file *)file->private_data;
2265 struct ext4_mb_proc_session *s = seq->private;
2266 struct super_block *sb = s->sb;
2270 if (count >= sizeof(str)) {
2271 printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
2272 "mb_history", (int)sizeof(str));
2276 if (copy_from_user(str, buffer, count))
2279 value = simple_strtol(str, NULL, 0);
2282 EXT4_SB(sb)->s_mb_history_filter = value;
2287 static const struct file_operations ext4_mb_seq_history_fops = {
2288 .owner = THIS_MODULE,
2289 .open = ext4_mb_seq_history_open,
2291 .write = ext4_mb_seq_history_write,
2292 .llseek = seq_lseek,
2293 .release = ext4_mb_seq_history_release,
2296 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2298 struct super_block *sb = seq->private;
2301 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2304 return (void *) ((unsigned long) group);
2307 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2309 struct super_block *sb = seq->private;
2313 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2316 return (void *) ((unsigned long) group);
2319 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2321 struct super_block *sb = seq->private;
2322 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2325 struct ext4_buddy e4b;
2327 struct ext4_group_info info;
2328 ext4_grpblk_t counters[16];
2333 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2334 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2335 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2336 "group", "free", "frags", "first",
2337 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2338 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2340 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2341 sizeof(struct ext4_group_info);
2342 err = ext4_mb_load_buddy(sb, group, &e4b);
2344 seq_printf(seq, "#%-5u: I/O error\n", group);
2347 ext4_lock_group(sb, group);
2348 memcpy(&sg, ext4_get_group_info(sb, group), i);
2349 ext4_unlock_group(sb, group);
2350 ext4_mb_release_desc(&e4b);
2352 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2353 sg.info.bb_fragments, sg.info.bb_first_free);
2354 for (i = 0; i <= 13; i++)
2355 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2356 sg.info.bb_counters[i] : 0);
2357 seq_printf(seq, " ]\n");
2362 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2366 static const struct seq_operations ext4_mb_seq_groups_ops = {
2367 .start = ext4_mb_seq_groups_start,
2368 .next = ext4_mb_seq_groups_next,
2369 .stop = ext4_mb_seq_groups_stop,
2370 .show = ext4_mb_seq_groups_show,
2373 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2375 struct super_block *sb = PDE(inode)->data;
2378 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2380 struct seq_file *m = (struct seq_file *)file->private_data;
2387 static const struct file_operations ext4_mb_seq_groups_fops = {
2388 .owner = THIS_MODULE,
2389 .open = ext4_mb_seq_groups_open,
2391 .llseek = seq_lseek,
2392 .release = seq_release,
2395 static void ext4_mb_history_release(struct super_block *sb)
2397 struct ext4_sb_info *sbi = EXT4_SB(sb);
2399 if (sbi->s_proc != NULL) {
2400 remove_proc_entry("mb_groups", sbi->s_proc);
2401 if (sbi->s_mb_history_max)
2402 remove_proc_entry("mb_history", sbi->s_proc);
2404 kfree(sbi->s_mb_history);
2407 static void ext4_mb_history_init(struct super_block *sb)
2409 struct ext4_sb_info *sbi = EXT4_SB(sb);
2412 if (sbi->s_proc != NULL) {
2413 if (sbi->s_mb_history_max)
2414 proc_create_data("mb_history", S_IRUGO, sbi->s_proc,
2415 &ext4_mb_seq_history_fops, sb);
2416 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2417 &ext4_mb_seq_groups_fops, sb);
2420 sbi->s_mb_history_cur = 0;
2421 spin_lock_init(&sbi->s_mb_history_lock);
2422 i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
2423 sbi->s_mb_history = i ? kzalloc(i, GFP_KERNEL) : NULL;
2424 /* if we can't allocate history, then we simple won't use it */
2427 static noinline_for_stack void
2428 ext4_mb_store_history(struct ext4_allocation_context *ac)
2430 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2431 struct ext4_mb_history h;
2433 if (sbi->s_mb_history == NULL)
2436 if (!(ac->ac_op & sbi->s_mb_history_filter))
2440 h.pid = current->pid;
2441 h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2442 h.orig = ac->ac_o_ex;
2443 h.result = ac->ac_b_ex;
2444 h.flags = ac->ac_flags;
2445 h.found = ac->ac_found;
2446 h.groups = ac->ac_groups_scanned;
2447 h.cr = ac->ac_criteria;
2448 h.tail = ac->ac_tail;
2449 h.buddy = ac->ac_buddy;
2451 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
2452 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2453 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2455 h.goal = ac->ac_g_ex;
2456 h.result = ac->ac_f_ex;
2459 spin_lock(&sbi->s_mb_history_lock);
2460 memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2461 if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2462 sbi->s_mb_history_cur = 0;
2463 spin_unlock(&sbi->s_mb_history_lock);
2467 #define ext4_mb_history_release(sb)
2468 #define ext4_mb_history_init(sb)
2472 /* Create and initialize ext4_group_info data for the given group. */
2473 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2474 struct ext4_group_desc *desc)
2478 struct ext4_sb_info *sbi = EXT4_SB(sb);
2479 struct ext4_group_info **meta_group_info;
2482 * First check if this group is the first of a reserved block.
2483 * If it's true, we have to allocate a new table of pointers
2484 * to ext4_group_info structures
2486 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2487 metalen = sizeof(*meta_group_info) <<
2488 EXT4_DESC_PER_BLOCK_BITS(sb);
2489 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2490 if (meta_group_info == NULL) {
2491 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2493 goto exit_meta_group_info;
2495 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2500 * calculate needed size. if change bb_counters size,
2501 * don't forget about ext4_mb_generate_buddy()
2503 len = offsetof(typeof(**meta_group_info),
2504 bb_counters[sb->s_blocksize_bits + 2]);
2507 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2508 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2510 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2511 if (meta_group_info[i] == NULL) {
2512 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2513 goto exit_group_info;
2515 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2516 &(meta_group_info[i]->bb_state));
2519 * initialize bb_free to be able to skip
2520 * empty groups without initialization
2522 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2523 meta_group_info[i]->bb_free =
2524 ext4_free_blocks_after_init(sb, group, desc);
2526 meta_group_info[i]->bb_free =
2527 ext4_free_blks_count(sb, desc);
2530 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2531 init_rwsem(&meta_group_info[i]->alloc_sem);
2532 meta_group_info[i]->bb_free_root.rb_node = NULL;
2536 struct buffer_head *bh;
2537 meta_group_info[i]->bb_bitmap =
2538 kmalloc(sb->s_blocksize, GFP_KERNEL);
2539 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2540 bh = ext4_read_block_bitmap(sb, group);
2542 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2551 /* If a meta_group_info table has been allocated, release it now */
2552 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2553 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2554 exit_meta_group_info:
2556 } /* ext4_mb_add_groupinfo */
2558 static int ext4_mb_init_backend(struct super_block *sb)
2560 ext4_group_t ngroups = ext4_get_groups_count(sb);
2562 struct ext4_sb_info *sbi = EXT4_SB(sb);
2563 struct ext4_super_block *es = sbi->s_es;
2564 int num_meta_group_infos;
2565 int num_meta_group_infos_max;
2567 struct ext4_group_desc *desc;
2569 /* This is the number of blocks used by GDT */
2570 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2571 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2574 * This is the total number of blocks used by GDT including
2575 * the number of reserved blocks for GDT.
2576 * The s_group_info array is allocated with this value
2577 * to allow a clean online resize without a complex
2578 * manipulation of pointer.
2579 * The drawback is the unused memory when no resize
2580 * occurs but it's very low in terms of pages
2581 * (see comments below)
2582 * Need to handle this properly when META_BG resizing is allowed
2584 num_meta_group_infos_max = num_meta_group_infos +
2585 le16_to_cpu(es->s_reserved_gdt_blocks);
2588 * array_size is the size of s_group_info array. We round it
2589 * to the next power of two because this approximation is done
2590 * internally by kmalloc so we can have some more memory
2591 * for free here (e.g. may be used for META_BG resize).
2594 while (array_size < sizeof(*sbi->s_group_info) *
2595 num_meta_group_infos_max)
2596 array_size = array_size << 1;
2597 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2598 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2599 * So a two level scheme suffices for now. */
2600 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2601 if (sbi->s_group_info == NULL) {
2602 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2605 sbi->s_buddy_cache = new_inode(sb);
2606 if (sbi->s_buddy_cache == NULL) {
2607 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2610 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2611 for (i = 0; i < ngroups; i++) {
2612 desc = ext4_get_group_desc(sb, i, NULL);
2615 "EXT4-fs: can't read descriptor %u\n", i);
2618 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2626 kfree(ext4_get_group_info(sb, i));
2627 i = num_meta_group_infos;
2629 kfree(sbi->s_group_info[i]);
2630 iput(sbi->s_buddy_cache);
2632 kfree(sbi->s_group_info);
2636 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2638 struct ext4_sb_info *sbi = EXT4_SB(sb);
2644 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2646 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2647 if (sbi->s_mb_offsets == NULL) {
2651 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2652 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2653 if (sbi->s_mb_maxs == NULL) {
2654 kfree(sbi->s_mb_offsets);
2658 /* order 0 is regular bitmap */
2659 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2660 sbi->s_mb_offsets[0] = 0;
2664 max = sb->s_blocksize << 2;
2666 sbi->s_mb_offsets[i] = offset;
2667 sbi->s_mb_maxs[i] = max;
2668 offset += 1 << (sb->s_blocksize_bits - i);
2671 } while (i <= sb->s_blocksize_bits + 1);
2673 /* init file for buddy data */
2674 ret = ext4_mb_init_backend(sb);
2676 kfree(sbi->s_mb_offsets);
2677 kfree(sbi->s_mb_maxs);
2681 spin_lock_init(&sbi->s_md_lock);
2682 spin_lock_init(&sbi->s_bal_lock);
2684 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2685 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2686 sbi->s_mb_stats = MB_DEFAULT_STATS;
2687 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2688 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2689 sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
2690 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2692 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2693 if (sbi->s_locality_groups == NULL) {
2694 kfree(sbi->s_mb_offsets);
2695 kfree(sbi->s_mb_maxs);
2698 for_each_possible_cpu(i) {
2699 struct ext4_locality_group *lg;
2700 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2701 mutex_init(&lg->lg_mutex);
2702 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2703 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2704 spin_lock_init(&lg->lg_prealloc_lock);
2707 ext4_mb_history_init(sb);
2710 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2712 printk(KERN_INFO "EXT4-fs: mballoc enabled\n");
2716 /* need to called with the ext4 group lock held */
2717 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2719 struct ext4_prealloc_space *pa;
2720 struct list_head *cur, *tmp;
2723 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2724 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2725 list_del(&pa->pa_group_list);
2727 kmem_cache_free(ext4_pspace_cachep, pa);
2730 mb_debug(1, "mballoc: %u PAs left\n", count);
2734 int ext4_mb_release(struct super_block *sb)
2736 ext4_group_t ngroups = ext4_get_groups_count(sb);
2738 int num_meta_group_infos;
2739 struct ext4_group_info *grinfo;
2740 struct ext4_sb_info *sbi = EXT4_SB(sb);
2742 if (sbi->s_group_info) {
2743 for (i = 0; i < ngroups; i++) {
2744 grinfo = ext4_get_group_info(sb, i);
2746 kfree(grinfo->bb_bitmap);
2748 ext4_lock_group(sb, i);
2749 ext4_mb_cleanup_pa(grinfo);
2750 ext4_unlock_group(sb, i);
2753 num_meta_group_infos = (ngroups +
2754 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2755 EXT4_DESC_PER_BLOCK_BITS(sb);
2756 for (i = 0; i < num_meta_group_infos; i++)
2757 kfree(sbi->s_group_info[i]);
2758 kfree(sbi->s_group_info);
2760 kfree(sbi->s_mb_offsets);
2761 kfree(sbi->s_mb_maxs);
2762 if (sbi->s_buddy_cache)
2763 iput(sbi->s_buddy_cache);
2764 if (sbi->s_mb_stats) {
2766 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2767 atomic_read(&sbi->s_bal_allocated),
2768 atomic_read(&sbi->s_bal_reqs),
2769 atomic_read(&sbi->s_bal_success));
2771 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2772 "%u 2^N hits, %u breaks, %u lost\n",
2773 atomic_read(&sbi->s_bal_ex_scanned),
2774 atomic_read(&sbi->s_bal_goals),
2775 atomic_read(&sbi->s_bal_2orders),
2776 atomic_read(&sbi->s_bal_breaks),
2777 atomic_read(&sbi->s_mb_lost_chunks));
2779 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2780 sbi->s_mb_buddies_generated++,
2781 sbi->s_mb_generation_time);
2783 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2784 atomic_read(&sbi->s_mb_preallocated),
2785 atomic_read(&sbi->s_mb_discarded));
2788 free_percpu(sbi->s_locality_groups);
2789 ext4_mb_history_release(sb);
2795 * This function is called by the jbd2 layer once the commit has finished,
2796 * so we know we can free the blocks that were released with that commit.
2798 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2800 struct super_block *sb = journal->j_private;
2801 struct ext4_buddy e4b;
2802 struct ext4_group_info *db;
2803 int err, count = 0, count2 = 0;
2804 struct ext4_free_data *entry;
2805 ext4_fsblk_t discard_block;
2806 struct list_head *l, *ltmp;
2808 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2809 entry = list_entry(l, struct ext4_free_data, list);
2811 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2812 entry->count, entry->group, entry);
2814 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2815 /* we expect to find existing buddy because it's pinned */
2819 /* there are blocks to put in buddy to make them really free */
2820 count += entry->count;
2822 ext4_lock_group(sb, entry->group);
2823 /* Take it out of per group rb tree */
2824 rb_erase(&entry->node, &(db->bb_free_root));
2825 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2827 if (!db->bb_free_root.rb_node) {
2828 /* No more items in the per group rb tree
2829 * balance refcounts from ext4_mb_free_metadata()
2831 page_cache_release(e4b.bd_buddy_page);
2832 page_cache_release(e4b.bd_bitmap_page);
2834 ext4_unlock_group(sb, entry->group);
2835 discard_block = (ext4_fsblk_t) entry->group * EXT4_BLOCKS_PER_GROUP(sb)
2837 + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
2838 trace_ext4_discard_blocks(sb, (unsigned long long)discard_block,
2840 sb_issue_discard(sb, discard_block, entry->count);
2842 kmem_cache_free(ext4_free_ext_cachep, entry);
2843 ext4_mb_release_desc(&e4b);
2846 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2849 #ifdef CONFIG_EXT4_DEBUG
2850 u8 mb_enable_debug __read_mostly;
2852 static struct dentry *debugfs_dir;
2853 static struct dentry *debugfs_debug;
2855 static void __init ext4_create_debugfs_entry(void)
2857 debugfs_dir = debugfs_create_dir("ext4", NULL);
2859 debugfs_debug = debugfs_create_u8("mballoc-debug",
2865 static void ext4_remove_debugfs_entry(void)
2867 debugfs_remove(debugfs_debug);
2868 debugfs_remove(debugfs_dir);
2873 static void __init ext4_create_debugfs_entry(void)
2877 static void ext4_remove_debugfs_entry(void)
2883 int __init init_ext4_mballoc(void)
2885 ext4_pspace_cachep =
2886 kmem_cache_create("ext4_prealloc_space",
2887 sizeof(struct ext4_prealloc_space),
2888 0, SLAB_RECLAIM_ACCOUNT, NULL);
2889 if (ext4_pspace_cachep == NULL)
2893 kmem_cache_create("ext4_alloc_context",
2894 sizeof(struct ext4_allocation_context),
2895 0, SLAB_RECLAIM_ACCOUNT, NULL);
2896 if (ext4_ac_cachep == NULL) {
2897 kmem_cache_destroy(ext4_pspace_cachep);
2901 ext4_free_ext_cachep =
2902 kmem_cache_create("ext4_free_block_extents",
2903 sizeof(struct ext4_free_data),
2904 0, SLAB_RECLAIM_ACCOUNT, NULL);
2905 if (ext4_free_ext_cachep == NULL) {
2906 kmem_cache_destroy(ext4_pspace_cachep);
2907 kmem_cache_destroy(ext4_ac_cachep);
2910 ext4_create_debugfs_entry();
2914 void exit_ext4_mballoc(void)
2917 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2918 * before destroying the slab cache.
2921 kmem_cache_destroy(ext4_pspace_cachep);
2922 kmem_cache_destroy(ext4_ac_cachep);
2923 kmem_cache_destroy(ext4_free_ext_cachep);
2924 ext4_remove_debugfs_entry();
2929 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2930 * Returns 0 if success or error code
2932 static noinline_for_stack int
2933 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2934 handle_t *handle, unsigned int reserv_blks)
2936 struct buffer_head *bitmap_bh = NULL;
2937 struct ext4_super_block *es;
2938 struct ext4_group_desc *gdp;
2939 struct buffer_head *gdp_bh;
2940 struct ext4_sb_info *sbi;
2941 struct super_block *sb;
2945 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2946 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2954 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2958 err = ext4_journal_get_write_access(handle, bitmap_bh);
2963 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2967 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2968 ext4_free_blks_count(sb, gdp));
2970 err = ext4_journal_get_write_access(handle, gdp_bh);
2974 block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
2975 + ac->ac_b_ex.fe_start
2976 + le32_to_cpu(es->s_first_data_block);
2978 len = ac->ac_b_ex.fe_len;
2979 if (!ext4_data_block_valid(sbi, block, len)) {
2980 ext4_error(sb, __func__,
2981 "Allocating blocks %llu-%llu which overlap "
2982 "fs metadata\n", block, block+len);
2983 /* File system mounted not to panic on error
2984 * Fix the bitmap and repeat the block allocation
2985 * We leak some of the blocks here.
2987 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2988 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2989 ac->ac_b_ex.fe_len);
2990 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2991 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2997 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2998 #ifdef AGGRESSIVE_CHECK
3001 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3002 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3003 bitmap_bh->b_data));
3007 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
3008 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
3009 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3010 ext4_free_blks_set(sb, gdp,
3011 ext4_free_blocks_after_init(sb,
3012 ac->ac_b_ex.fe_group, gdp));
3014 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
3015 ext4_free_blks_set(sb, gdp, len);
3016 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
3018 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3019 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
3021 * Now reduce the dirty block count also. Should not go negative
3023 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3024 /* release all the reserved blocks if non delalloc */
3025 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
3027 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
3028 ac->ac_b_ex.fe_len);
3029 /* convert reserved quota blocks to real quota blocks */
3030 vfs_dq_claim_block(ac->ac_inode, ac->ac_b_ex.fe_len);
3033 if (sbi->s_log_groups_per_flex) {
3034 ext4_group_t flex_group = ext4_flex_group(sbi,
3035 ac->ac_b_ex.fe_group);
3036 atomic_sub(ac->ac_b_ex.fe_len,
3037 &sbi->s_flex_groups[flex_group].free_blocks);
3040 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3043 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3052 * here we normalize request for locality group
3053 * Group request are normalized to s_strip size if we set the same via mount
3054 * option. If not we set it to s_mb_group_prealloc which can be configured via
3055 * /sys/fs/ext4/<partition>/mb_group_prealloc
3057 * XXX: should we try to preallocate more than the group has now?
3059 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3061 struct super_block *sb = ac->ac_sb;
3062 struct ext4_locality_group *lg = ac->ac_lg;
3065 if (EXT4_SB(sb)->s_stripe)
3066 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
3068 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3069 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3070 current->pid, ac->ac_g_ex.fe_len);
3074 * Normalization means making request better in terms of
3075 * size and alignment
3077 static noinline_for_stack void
3078 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3079 struct ext4_allocation_request *ar)
3083 loff_t size, orig_size, start_off;
3084 ext4_lblk_t start, orig_start;
3085 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3086 struct ext4_prealloc_space *pa;
3088 /* do normalize only data requests, metadata requests
3089 do not need preallocation */
3090 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3093 /* sometime caller may want exact blocks */
3094 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3097 /* caller may indicate that preallocation isn't
3098 * required (it's a tail, for example) */
3099 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3102 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3103 ext4_mb_normalize_group_request(ac);
3107 bsbits = ac->ac_sb->s_blocksize_bits;
3109 /* first, let's learn actual file size
3110 * given current request is allocated */
3111 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3112 size = size << bsbits;
3113 if (size < i_size_read(ac->ac_inode))
3114 size = i_size_read(ac->ac_inode);
3116 /* max size of free chunks */
3119 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3120 (req <= (size) || max <= (chunk_size))
3122 /* first, try to predict filesize */
3123 /* XXX: should this table be tunable? */
3125 if (size <= 16 * 1024) {
3127 } else if (size <= 32 * 1024) {
3129 } else if (size <= 64 * 1024) {
3131 } else if (size <= 128 * 1024) {
3133 } else if (size <= 256 * 1024) {
3135 } else if (size <= 512 * 1024) {
3137 } else if (size <= 1024 * 1024) {
3139 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3140 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3141 (21 - bsbits)) << 21;
3142 size = 2 * 1024 * 1024;
3143 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3144 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3145 (22 - bsbits)) << 22;
3146 size = 4 * 1024 * 1024;
3147 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3148 (8<<20)>>bsbits, max, 8 * 1024)) {
3149 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3150 (23 - bsbits)) << 23;
3151 size = 8 * 1024 * 1024;
3153 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3154 size = ac->ac_o_ex.fe_len << bsbits;
3156 orig_size = size = size >> bsbits;
3157 orig_start = start = start_off >> bsbits;
3159 /* don't cover already allocated blocks in selected range */
3160 if (ar->pleft && start <= ar->lleft) {
3161 size -= ar->lleft + 1 - start;
3162 start = ar->lleft + 1;
3164 if (ar->pright && start + size - 1 >= ar->lright)
3165 size -= start + size - ar->lright;
3169 /* check we don't cross already preallocated blocks */
3171 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3176 spin_lock(&pa->pa_lock);
3177 if (pa->pa_deleted) {
3178 spin_unlock(&pa->pa_lock);
3182 pa_end = pa->pa_lstart + pa->pa_len;
3184 /* PA must not overlap original request */
3185 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3186 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3188 /* skip PAs this normalized request doesn't overlap with */
3189 if (pa->pa_lstart >= end || pa_end <= start) {
3190 spin_unlock(&pa->pa_lock);
3193 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3195 /* adjust start or end to be adjacent to this pa */
3196 if (pa_end <= ac->ac_o_ex.fe_logical) {
3197 BUG_ON(pa_end < start);
3199 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3200 BUG_ON(pa->pa_lstart > end);
3201 end = pa->pa_lstart;
3203 spin_unlock(&pa->pa_lock);
3208 /* XXX: extra loop to check we really don't overlap preallocations */
3210 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3212 spin_lock(&pa->pa_lock);
3213 if (pa->pa_deleted == 0) {
3214 pa_end = pa->pa_lstart + pa->pa_len;
3215 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3217 spin_unlock(&pa->pa_lock);
3221 if (start + size <= ac->ac_o_ex.fe_logical &&
3222 start > ac->ac_o_ex.fe_logical) {
3223 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3224 (unsigned long) start, (unsigned long) size,
3225 (unsigned long) ac->ac_o_ex.fe_logical);
3227 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3228 start > ac->ac_o_ex.fe_logical);
3229 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3231 /* now prepare goal request */
3233 /* XXX: is it better to align blocks WRT to logical
3234 * placement or satisfy big request as is */
3235 ac->ac_g_ex.fe_logical = start;
3236 ac->ac_g_ex.fe_len = size;
3238 /* define goal start in order to merge */
3239 if (ar->pright && (ar->lright == (start + size))) {
3240 /* merge to the right */
3241 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3242 &ac->ac_f_ex.fe_group,
3243 &ac->ac_f_ex.fe_start);
3244 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3246 if (ar->pleft && (ar->lleft + 1 == start)) {
3247 /* merge to the left */
3248 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3249 &ac->ac_f_ex.fe_group,
3250 &ac->ac_f_ex.fe_start);
3251 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3254 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3255 (unsigned) orig_size, (unsigned) start);
3258 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3260 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3262 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3263 atomic_inc(&sbi->s_bal_reqs);
3264 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3265 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3266 atomic_inc(&sbi->s_bal_success);
3267 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3268 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3269 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3270 atomic_inc(&sbi->s_bal_goals);
3271 if (ac->ac_found > sbi->s_mb_max_to_scan)
3272 atomic_inc(&sbi->s_bal_breaks);
3275 ext4_mb_store_history(ac);
3279 * use blocks preallocated to inode
3281 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3282 struct ext4_prealloc_space *pa)
3288 /* found preallocated blocks, use them */
3289 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3290 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3292 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3293 &ac->ac_b_ex.fe_start);
3294 ac->ac_b_ex.fe_len = len;
3295 ac->ac_status = AC_STATUS_FOUND;
3298 BUG_ON(start < pa->pa_pstart);
3299 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3300 BUG_ON(pa->pa_free < len);
3303 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3307 * use blocks preallocated to locality group
3309 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3310 struct ext4_prealloc_space *pa)
3312 unsigned int len = ac->ac_o_ex.fe_len;
3314 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3315 &ac->ac_b_ex.fe_group,
3316 &ac->ac_b_ex.fe_start);
3317 ac->ac_b_ex.fe_len = len;
3318 ac->ac_status = AC_STATUS_FOUND;
3321 /* we don't correct pa_pstart or pa_plen here to avoid
3322 * possible race when the group is being loaded concurrently
3323 * instead we correct pa later, after blocks are marked
3324 * in on-disk bitmap -- see ext4_mb_release_context()
3325 * Other CPUs are prevented from allocating from this pa by lg_mutex
3327 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3331 * Return the prealloc space that have minimal distance
3332 * from the goal block. @cpa is the prealloc
3333 * space that is having currently known minimal distance
3334 * from the goal block.
3336 static struct ext4_prealloc_space *
3337 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3338 struct ext4_prealloc_space *pa,
3339 struct ext4_prealloc_space *cpa)
3341 ext4_fsblk_t cur_distance, new_distance;
3344 atomic_inc(&pa->pa_count);
3347 cur_distance = abs(goal_block - cpa->pa_pstart);
3348 new_distance = abs(goal_block - pa->pa_pstart);
3350 if (cur_distance < new_distance)
3353 /* drop the previous reference */
3354 atomic_dec(&cpa->pa_count);
3355 atomic_inc(&pa->pa_count);
3360 * search goal blocks in preallocated space
3362 static noinline_for_stack int
3363 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3366 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3367 struct ext4_locality_group *lg;
3368 struct ext4_prealloc_space *pa, *cpa = NULL;
3369 ext4_fsblk_t goal_block;
3371 /* only data can be preallocated */
3372 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3375 /* first, try per-file preallocation */
3377 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3379 /* all fields in this condition don't change,
3380 * so we can skip locking for them */
3381 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3382 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3385 /* found preallocated blocks, use them */
3386 spin_lock(&pa->pa_lock);
3387 if (pa->pa_deleted == 0 && pa->pa_free) {
3388 atomic_inc(&pa->pa_count);
3389 ext4_mb_use_inode_pa(ac, pa);
3390 spin_unlock(&pa->pa_lock);
3391 ac->ac_criteria = 10;
3395 spin_unlock(&pa->pa_lock);
3399 /* can we use group allocation? */
3400 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3403 /* inode may have no locality group for some reason */
3407 order = fls(ac->ac_o_ex.fe_len) - 1;
3408 if (order > PREALLOC_TB_SIZE - 1)
3409 /* The max size of hash table is PREALLOC_TB_SIZE */
3410 order = PREALLOC_TB_SIZE - 1;
3412 goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) +
3413 ac->ac_g_ex.fe_start +
3414 le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block);
3416 * search for the prealloc space that is having
3417 * minimal distance from the goal block.
3419 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3421 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3423 spin_lock(&pa->pa_lock);
3424 if (pa->pa_deleted == 0 &&
3425 pa->pa_free >= ac->ac_o_ex.fe_len) {
3427 cpa = ext4_mb_check_group_pa(goal_block,
3430 spin_unlock(&pa->pa_lock);
3435 ext4_mb_use_group_pa(ac, cpa);
3436 ac->ac_criteria = 20;
3443 * the function goes through all block freed in the group
3444 * but not yet committed and marks them used in in-core bitmap.
3445 * buddy must be generated from this bitmap
3446 * Need to be called with the ext4 group lock held
3448 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3452 struct ext4_group_info *grp;
3453 struct ext4_free_data *entry;
3455 grp = ext4_get_group_info(sb, group);
3456 n = rb_first(&(grp->bb_free_root));
3459 entry = rb_entry(n, struct ext4_free_data, node);
3460 mb_set_bits(bitmap, entry->start_blk, entry->count);
3467 * the function goes through all preallocation in this group and marks them
3468 * used in in-core bitmap. buddy must be generated from this bitmap
3469 * Need to be called with ext4 group lock held
3471 static noinline_for_stack
3472 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3475 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3476 struct ext4_prealloc_space *pa;
3477 struct list_head *cur;
3478 ext4_group_t groupnr;
3479 ext4_grpblk_t start;
3480 int preallocated = 0;
3484 /* all form of preallocation discards first load group,
3485 * so the only competing code is preallocation use.
3486 * we don't need any locking here
3487 * notice we do NOT ignore preallocations with pa_deleted
3488 * otherwise we could leave used blocks available for
3489 * allocation in buddy when concurrent ext4_mb_put_pa()
3490 * is dropping preallocation
3492 list_for_each(cur, &grp->bb_prealloc_list) {
3493 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3494 spin_lock(&pa->pa_lock);
3495 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3498 spin_unlock(&pa->pa_lock);
3499 if (unlikely(len == 0))
3501 BUG_ON(groupnr != group);
3502 mb_set_bits(bitmap, start, len);
3503 preallocated += len;
3506 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3509 static void ext4_mb_pa_callback(struct rcu_head *head)
3511 struct ext4_prealloc_space *pa;
3512 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3513 kmem_cache_free(ext4_pspace_cachep, pa);
3517 * drops a reference to preallocated space descriptor
3518 * if this was the last reference and the space is consumed
3520 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3521 struct super_block *sb, struct ext4_prealloc_space *pa)
3524 ext4_fsblk_t grp_blk;
3526 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3529 /* in this short window concurrent discard can set pa_deleted */
3530 spin_lock(&pa->pa_lock);
3531 if (pa->pa_deleted == 1) {
3532 spin_unlock(&pa->pa_lock);
3537 spin_unlock(&pa->pa_lock);
3539 grp_blk = pa->pa_pstart;
3541 * If doing group-based preallocation, pa_pstart may be in the
3542 * next group when pa is used up
3544 if (pa->pa_type == MB_GROUP_PA)
3547 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3552 * P1 (buddy init) P2 (regular allocation)
3553 * find block B in PA
3554 * copy on-disk bitmap to buddy
3555 * mark B in on-disk bitmap
3556 * drop PA from group
3557 * mark all PAs in buddy
3559 * thus, P1 initializes buddy with B available. to prevent this
3560 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3563 ext4_lock_group(sb, grp);
3564 list_del(&pa->pa_group_list);
3565 ext4_unlock_group(sb, grp);
3567 spin_lock(pa->pa_obj_lock);
3568 list_del_rcu(&pa->pa_inode_list);
3569 spin_unlock(pa->pa_obj_lock);
3571 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3575 * creates new preallocated space for given inode
3577 static noinline_for_stack int
3578 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3580 struct super_block *sb = ac->ac_sb;
3581 struct ext4_prealloc_space *pa;
3582 struct ext4_group_info *grp;
3583 struct ext4_inode_info *ei;
3585 /* preallocate only when found space is larger then requested */
3586 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3587 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3588 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3590 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3594 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3600 /* we can't allocate as much as normalizer wants.
3601 * so, found space must get proper lstart
3602 * to cover original request */
3603 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3604 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3606 /* we're limited by original request in that
3607 * logical block must be covered any way
3608 * winl is window we can move our chunk within */
3609 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3611 /* also, we should cover whole original request */
3612 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3614 /* the smallest one defines real window */
3615 win = min(winl, wins);
3617 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3618 if (offs && offs < win)
3621 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3622 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3623 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3626 /* preallocation can change ac_b_ex, thus we store actually
3627 * allocated blocks for history */
3628 ac->ac_f_ex = ac->ac_b_ex;
3630 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3631 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3632 pa->pa_len = ac->ac_b_ex.fe_len;
3633 pa->pa_free = pa->pa_len;
3634 atomic_set(&pa->pa_count, 1);
3635 spin_lock_init(&pa->pa_lock);
3636 INIT_LIST_HEAD(&pa->pa_inode_list);
3637 INIT_LIST_HEAD(&pa->pa_group_list);
3639 pa->pa_type = MB_INODE_PA;
3641 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3642 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3643 trace_ext4_mb_new_inode_pa(ac, pa);
3645 ext4_mb_use_inode_pa(ac, pa);
3646 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3648 ei = EXT4_I(ac->ac_inode);
3649 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3651 pa->pa_obj_lock = &ei->i_prealloc_lock;
3652 pa->pa_inode = ac->ac_inode;
3654 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3655 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3656 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3658 spin_lock(pa->pa_obj_lock);
3659 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3660 spin_unlock(pa->pa_obj_lock);
3666 * creates new preallocated space for locality group inodes belongs to
3668 static noinline_for_stack int
3669 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3671 struct super_block *sb = ac->ac_sb;
3672 struct ext4_locality_group *lg;
3673 struct ext4_prealloc_space *pa;
3674 struct ext4_group_info *grp;
3676 /* preallocate only when found space is larger then requested */
3677 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3678 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3679 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3681 BUG_ON(ext4_pspace_cachep == NULL);
3682 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3686 /* preallocation can change ac_b_ex, thus we store actually
3687 * allocated blocks for history */
3688 ac->ac_f_ex = ac->ac_b_ex;
3690 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3691 pa->pa_lstart = pa->pa_pstart;
3692 pa->pa_len = ac->ac_b_ex.fe_len;
3693 pa->pa_free = pa->pa_len;
3694 atomic_set(&pa->pa_count, 1);
3695 spin_lock_init(&pa->pa_lock);
3696 INIT_LIST_HEAD(&pa->pa_inode_list);
3697 INIT_LIST_HEAD(&pa->pa_group_list);
3699 pa->pa_type = MB_GROUP_PA;
3701 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3702 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3703 trace_ext4_mb_new_group_pa(ac, pa);
3705 ext4_mb_use_group_pa(ac, pa);
3706 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3708 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3712 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3713 pa->pa_inode = NULL;
3715 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3716 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3717 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3720 * We will later add the new pa to the right bucket
3721 * after updating the pa_free in ext4_mb_release_context
3726 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3730 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3731 err = ext4_mb_new_group_pa(ac);
3733 err = ext4_mb_new_inode_pa(ac);
3738 * finds all unused blocks in on-disk bitmap, frees them in
3739 * in-core bitmap and buddy.
3740 * @pa must be unlinked from inode and group lists, so that
3741 * nobody else can find/use it.
3742 * the caller MUST hold group/inode locks.
3743 * TODO: optimize the case when there are no in-core structures yet
3745 static noinline_for_stack int
3746 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3747 struct ext4_prealloc_space *pa,
3748 struct ext4_allocation_context *ac)
3750 struct super_block *sb = e4b->bd_sb;
3751 struct ext4_sb_info *sbi = EXT4_SB(sb);
3756 unsigned long long grp_blk_start;
3761 BUG_ON(pa->pa_deleted == 0);
3762 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3763 grp_blk_start = pa->pa_pstart - bit;
3764 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3765 end = bit + pa->pa_len;
3769 ac->ac_inode = pa->pa_inode;
3770 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3774 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3777 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3778 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3779 le32_to_cpu(sbi->s_es->s_first_data_block);
3780 mb_debug(1, " free preallocated %u/%u in group %u\n",
3781 (unsigned) start, (unsigned) next - bit,
3786 ac->ac_b_ex.fe_group = group;
3787 ac->ac_b_ex.fe_start = bit;
3788 ac->ac_b_ex.fe_len = next - bit;
3789 ac->ac_b_ex.fe_logical = 0;
3790 ext4_mb_store_history(ac);
3793 trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit,
3795 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3798 if (free != pa->pa_free) {
3799 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3800 pa, (unsigned long) pa->pa_lstart,
3801 (unsigned long) pa->pa_pstart,
3802 (unsigned long) pa->pa_len);
3803 ext4_grp_locked_error(sb, group,
3804 __func__, "free %u, pa_free %u",
3807 * pa is already deleted so we use the value obtained
3808 * from the bitmap and continue.
3811 atomic_add(free, &sbi->s_mb_discarded);
3816 static noinline_for_stack int
3817 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3818 struct ext4_prealloc_space *pa,
3819 struct ext4_allocation_context *ac)
3821 struct super_block *sb = e4b->bd_sb;
3826 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3828 trace_ext4_mb_release_group_pa(ac, pa);
3829 BUG_ON(pa->pa_deleted == 0);
3830 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3831 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3832 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3833 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3837 ac->ac_inode = NULL;
3838 ac->ac_b_ex.fe_group = group;
3839 ac->ac_b_ex.fe_start = bit;
3840 ac->ac_b_ex.fe_len = pa->pa_len;
3841 ac->ac_b_ex.fe_logical = 0;
3842 ext4_mb_store_history(ac);
3849 * releases all preallocations in given group
3851 * first, we need to decide discard policy:
3852 * - when do we discard
3854 * - how many do we discard
3855 * 1) how many requested
3857 static noinline_for_stack int
3858 ext4_mb_discard_group_preallocations(struct super_block *sb,
3859 ext4_group_t group, int needed)
3861 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3862 struct buffer_head *bitmap_bh = NULL;
3863 struct ext4_prealloc_space *pa, *tmp;
3864 struct ext4_allocation_context *ac;
3865 struct list_head list;
3866 struct ext4_buddy e4b;
3871 mb_debug(1, "discard preallocation for group %u\n", group);
3873 if (list_empty(&grp->bb_prealloc_list))
3876 bitmap_bh = ext4_read_block_bitmap(sb, group);
3877 if (bitmap_bh == NULL) {
3878 ext4_error(sb, __func__, "Error in reading block "
3879 "bitmap for %u", group);
3883 err = ext4_mb_load_buddy(sb, group, &e4b);
3885 ext4_error(sb, __func__, "Error in loading buddy "
3886 "information for %u", group);
3892 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3894 INIT_LIST_HEAD(&list);
3895 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3899 ext4_lock_group(sb, group);
3900 list_for_each_entry_safe(pa, tmp,
3901 &grp->bb_prealloc_list, pa_group_list) {
3902 spin_lock(&pa->pa_lock);
3903 if (atomic_read(&pa->pa_count)) {
3904 spin_unlock(&pa->pa_lock);
3908 if (pa->pa_deleted) {
3909 spin_unlock(&pa->pa_lock);
3913 /* seems this one can be freed ... */
3916 /* we can trust pa_free ... */
3917 free += pa->pa_free;
3919 spin_unlock(&pa->pa_lock);
3921 list_del(&pa->pa_group_list);
3922 list_add(&pa->u.pa_tmp_list, &list);
3925 /* if we still need more blocks and some PAs were used, try again */
3926 if (free < needed && busy) {
3928 ext4_unlock_group(sb, group);
3930 * Yield the CPU here so that we don't get soft lockup
3931 * in non preempt case.
3937 /* found anything to free? */
3938 if (list_empty(&list)) {
3943 /* now free all selected PAs */
3944 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3946 /* remove from object (inode or locality group) */
3947 spin_lock(pa->pa_obj_lock);
3948 list_del_rcu(&pa->pa_inode_list);
3949 spin_unlock(pa->pa_obj_lock);
3951 if (pa->pa_type == MB_GROUP_PA)
3952 ext4_mb_release_group_pa(&e4b, pa, ac);
3954 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3956 list_del(&pa->u.pa_tmp_list);
3957 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3961 ext4_unlock_group(sb, group);
3963 kmem_cache_free(ext4_ac_cachep, ac);
3964 ext4_mb_release_desc(&e4b);
3970 * releases all non-used preallocated blocks for given inode
3972 * It's important to discard preallocations under i_data_sem
3973 * We don't want another block to be served from the prealloc
3974 * space when we are discarding the inode prealloc space.
3976 * FIXME!! Make sure it is valid at all the call sites
3978 void ext4_discard_preallocations(struct inode *inode)
3980 struct ext4_inode_info *ei = EXT4_I(inode);
3981 struct super_block *sb = inode->i_sb;
3982 struct buffer_head *bitmap_bh = NULL;
3983 struct ext4_prealloc_space *pa, *tmp;
3984 struct ext4_allocation_context *ac;
3985 ext4_group_t group = 0;
3986 struct list_head list;
3987 struct ext4_buddy e4b;
3990 if (!S_ISREG(inode->i_mode)) {
3991 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3995 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3996 trace_ext4_discard_preallocations(inode);
3998 INIT_LIST_HEAD(&list);
4000 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4003 ac->ac_inode = inode;
4006 /* first, collect all pa's in the inode */
4007 spin_lock(&ei->i_prealloc_lock);
4008 while (!list_empty(&ei->i_prealloc_list)) {
4009 pa = list_entry(ei->i_prealloc_list.next,
4010 struct ext4_prealloc_space, pa_inode_list);
4011 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4012 spin_lock(&pa->pa_lock);
4013 if (atomic_read(&pa->pa_count)) {
4014 /* this shouldn't happen often - nobody should
4015 * use preallocation while we're discarding it */
4016 spin_unlock(&pa->pa_lock);
4017 spin_unlock(&ei->i_prealloc_lock);
4018 printk(KERN_ERR "uh-oh! used pa while discarding\n");
4020 schedule_timeout_uninterruptible(HZ);
4024 if (pa->pa_deleted == 0) {
4026 spin_unlock(&pa->pa_lock);
4027 list_del_rcu(&pa->pa_inode_list);
4028 list_add(&pa->u.pa_tmp_list, &list);
4032 /* someone is deleting pa right now */
4033 spin_unlock(&pa->pa_lock);
4034 spin_unlock(&ei->i_prealloc_lock);
4036 /* we have to wait here because pa_deleted
4037 * doesn't mean pa is already unlinked from
4038 * the list. as we might be called from
4039 * ->clear_inode() the inode will get freed
4040 * and concurrent thread which is unlinking
4041 * pa from inode's list may access already
4042 * freed memory, bad-bad-bad */
4044 /* XXX: if this happens too often, we can
4045 * add a flag to force wait only in case
4046 * of ->clear_inode(), but not in case of
4047 * regular truncate */
4048 schedule_timeout_uninterruptible(HZ);
4051 spin_unlock(&ei->i_prealloc_lock);
4053 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4054 BUG_ON(pa->pa_type != MB_INODE_PA);
4055 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4057 err = ext4_mb_load_buddy(sb, group, &e4b);
4059 ext4_error(sb, __func__, "Error in loading buddy "
4060 "information for %u", group);
4064 bitmap_bh = ext4_read_block_bitmap(sb, group);
4065 if (bitmap_bh == NULL) {
4066 ext4_error(sb, __func__, "Error in reading block "
4067 "bitmap for %u", group);
4068 ext4_mb_release_desc(&e4b);
4072 ext4_lock_group(sb, group);
4073 list_del(&pa->pa_group_list);
4074 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
4075 ext4_unlock_group(sb, group);
4077 ext4_mb_release_desc(&e4b);
4080 list_del(&pa->u.pa_tmp_list);
4081 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4084 kmem_cache_free(ext4_ac_cachep, ac);
4088 * finds all preallocated spaces and return blocks being freed to them
4089 * if preallocated space becomes full (no block is used from the space)
4090 * then the function frees space in buddy
4091 * XXX: at the moment, truncate (which is the only way to free blocks)
4092 * discards all preallocations
4094 static void ext4_mb_return_to_preallocation(struct inode *inode,
4095 struct ext4_buddy *e4b,
4096 sector_t block, int count)
4098 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
4100 #ifdef CONFIG_EXT4_DEBUG
4101 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4103 struct super_block *sb = ac->ac_sb;
4104 ext4_group_t ngroups, i;
4106 printk(KERN_ERR "EXT4-fs: Can't allocate:"
4107 " Allocation context details:\n");
4108 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
4109 ac->ac_status, ac->ac_flags);
4110 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4111 "best %lu/%lu/%lu@%lu cr %d\n",
4112 (unsigned long)ac->ac_o_ex.fe_group,
4113 (unsigned long)ac->ac_o_ex.fe_start,
4114 (unsigned long)ac->ac_o_ex.fe_len,
4115 (unsigned long)ac->ac_o_ex.fe_logical,
4116 (unsigned long)ac->ac_g_ex.fe_group,
4117 (unsigned long)ac->ac_g_ex.fe_start,
4118 (unsigned long)ac->ac_g_ex.fe_len,
4119 (unsigned long)ac->ac_g_ex.fe_logical,
4120 (unsigned long)ac->ac_b_ex.fe_group,
4121 (unsigned long)ac->ac_b_ex.fe_start,
4122 (unsigned long)ac->ac_b_ex.fe_len,
4123 (unsigned long)ac->ac_b_ex.fe_logical,
4124 (int)ac->ac_criteria);
4125 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
4127 printk(KERN_ERR "EXT4-fs: groups: \n");
4128 ngroups = ext4_get_groups_count(sb);
4129 for (i = 0; i < ngroups; i++) {
4130 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4131 struct ext4_prealloc_space *pa;
4132 ext4_grpblk_t start;
4133 struct list_head *cur;
4134 ext4_lock_group(sb, i);
4135 list_for_each(cur, &grp->bb_prealloc_list) {
4136 pa = list_entry(cur, struct ext4_prealloc_space,
4138 spin_lock(&pa->pa_lock);
4139 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4141 spin_unlock(&pa->pa_lock);
4142 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4145 ext4_unlock_group(sb, i);
4147 if (grp->bb_free == 0)
4149 printk(KERN_ERR "%u: %d/%d \n",
4150 i, grp->bb_free, grp->bb_fragments);
4152 printk(KERN_ERR "\n");
4155 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4162 * We use locality group preallocation for small size file. The size of the
4163 * file is determined by the current size or the resulting size after
4164 * allocation which ever is larger
4166 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4168 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4170 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4171 int bsbits = ac->ac_sb->s_blocksize_bits;
4174 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4177 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4180 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4181 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4183 size = max(size, isize);
4185 if ((size == isize) &&
4186 !ext4_fs_is_busy(sbi) &&
4187 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4188 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4192 /* don't use group allocation for large files */
4193 if (size >= sbi->s_mb_stream_request) {
4194 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4198 BUG_ON(ac->ac_lg != NULL);
4200 * locality group prealloc space are per cpu. The reason for having
4201 * per cpu locality group is to reduce the contention between block
4202 * request from multiple CPUs.
4204 ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id());
4206 /* we're going to use group allocation */
4207 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4209 /* serialize all allocations in the group */
4210 mutex_lock(&ac->ac_lg->lg_mutex);
4213 static noinline_for_stack int
4214 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4215 struct ext4_allocation_request *ar)
4217 struct super_block *sb = ar->inode->i_sb;
4218 struct ext4_sb_info *sbi = EXT4_SB(sb);
4219 struct ext4_super_block *es = sbi->s_es;
4223 ext4_grpblk_t block;
4225 /* we can't allocate > group size */
4228 /* just a dirty hack to filter too big requests */
4229 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4230 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4232 /* start searching from the goal */
4234 if (goal < le32_to_cpu(es->s_first_data_block) ||
4235 goal >= ext4_blocks_count(es))
4236 goal = le32_to_cpu(es->s_first_data_block);
4237 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4239 /* set up allocation goals */
4240 memset(ac, 0, sizeof(struct ext4_allocation_context));
4241 ac->ac_b_ex.fe_logical = ar->logical;
4242 ac->ac_status = AC_STATUS_CONTINUE;
4244 ac->ac_inode = ar->inode;
4245 ac->ac_o_ex.fe_logical = ar->logical;
4246 ac->ac_o_ex.fe_group = group;
4247 ac->ac_o_ex.fe_start = block;
4248 ac->ac_o_ex.fe_len = len;
4249 ac->ac_g_ex.fe_logical = ar->logical;
4250 ac->ac_g_ex.fe_group = group;
4251 ac->ac_g_ex.fe_start = block;
4252 ac->ac_g_ex.fe_len = len;
4253 ac->ac_flags = ar->flags;
4255 /* we have to define context: we'll we work with a file or
4256 * locality group. this is a policy, actually */
4257 ext4_mb_group_or_file(ac);
4259 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4260 "left: %u/%u, right %u/%u to %swritable\n",
4261 (unsigned) ar->len, (unsigned) ar->logical,
4262 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4263 (unsigned) ar->lleft, (unsigned) ar->pleft,
4264 (unsigned) ar->lright, (unsigned) ar->pright,
4265 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4270 static noinline_for_stack void
4271 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4272 struct ext4_locality_group *lg,
4273 int order, int total_entries)
4275 ext4_group_t group = 0;
4276 struct ext4_buddy e4b;
4277 struct list_head discard_list;
4278 struct ext4_prealloc_space *pa, *tmp;
4279 struct ext4_allocation_context *ac;
4281 mb_debug(1, "discard locality group preallocation\n");
4283 INIT_LIST_HEAD(&discard_list);
4284 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4288 spin_lock(&lg->lg_prealloc_lock);
4289 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4291 spin_lock(&pa->pa_lock);
4292 if (atomic_read(&pa->pa_count)) {
4294 * This is the pa that we just used
4295 * for block allocation. So don't
4298 spin_unlock(&pa->pa_lock);
4301 if (pa->pa_deleted) {
4302 spin_unlock(&pa->pa_lock);
4305 /* only lg prealloc space */
4306 BUG_ON(pa->pa_type != MB_GROUP_PA);
4308 /* seems this one can be freed ... */
4310 spin_unlock(&pa->pa_lock);
4312 list_del_rcu(&pa->pa_inode_list);
4313 list_add(&pa->u.pa_tmp_list, &discard_list);
4316 if (total_entries <= 5) {
4318 * we want to keep only 5 entries
4319 * allowing it to grow to 8. This
4320 * mak sure we don't call discard
4321 * soon for this list.
4326 spin_unlock(&lg->lg_prealloc_lock);
4328 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4330 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4331 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4332 ext4_error(sb, __func__, "Error in loading buddy "
4333 "information for %u", group);
4336 ext4_lock_group(sb, group);
4337 list_del(&pa->pa_group_list);
4338 ext4_mb_release_group_pa(&e4b, pa, ac);
4339 ext4_unlock_group(sb, group);
4341 ext4_mb_release_desc(&e4b);
4342 list_del(&pa->u.pa_tmp_list);
4343 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4346 kmem_cache_free(ext4_ac_cachep, ac);
4350 * We have incremented pa_count. So it cannot be freed at this
4351 * point. Also we hold lg_mutex. So no parallel allocation is
4352 * possible from this lg. That means pa_free cannot be updated.
4354 * A parallel ext4_mb_discard_group_preallocations is possible.
4355 * which can cause the lg_prealloc_list to be updated.
4358 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4360 int order, added = 0, lg_prealloc_count = 1;
4361 struct super_block *sb = ac->ac_sb;
4362 struct ext4_locality_group *lg = ac->ac_lg;
4363 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4365 order = fls(pa->pa_free) - 1;
4366 if (order > PREALLOC_TB_SIZE - 1)
4367 /* The max size of hash table is PREALLOC_TB_SIZE */
4368 order = PREALLOC_TB_SIZE - 1;
4369 /* Add the prealloc space to lg */
4371 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4373 spin_lock(&tmp_pa->pa_lock);
4374 if (tmp_pa->pa_deleted) {
4375 spin_unlock(&tmp_pa->pa_lock);
4378 if (!added && pa->pa_free < tmp_pa->pa_free) {
4379 /* Add to the tail of the previous entry */
4380 list_add_tail_rcu(&pa->pa_inode_list,
4381 &tmp_pa->pa_inode_list);
4384 * we want to count the total
4385 * number of entries in the list
4388 spin_unlock(&tmp_pa->pa_lock);
4389 lg_prealloc_count++;
4392 list_add_tail_rcu(&pa->pa_inode_list,
4393 &lg->lg_prealloc_list[order]);
4396 /* Now trim the list to be not more than 8 elements */
4397 if (lg_prealloc_count > 8) {
4398 ext4_mb_discard_lg_preallocations(sb, lg,
4399 order, lg_prealloc_count);
4406 * release all resource we used in allocation
4408 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4410 struct ext4_prealloc_space *pa = ac->ac_pa;
4412 if (pa->pa_type == MB_GROUP_PA) {
4413 /* see comment in ext4_mb_use_group_pa() */
4414 spin_lock(&pa->pa_lock);
4415 pa->pa_pstart += ac->ac_b_ex.fe_len;
4416 pa->pa_lstart += ac->ac_b_ex.fe_len;
4417 pa->pa_free -= ac->ac_b_ex.fe_len;
4418 pa->pa_len -= ac->ac_b_ex.fe_len;
4419 spin_unlock(&pa->pa_lock);
4423 up_read(ac->alloc_semp);
4426 * We want to add the pa to the right bucket.
4427 * Remove it from the list and while adding
4428 * make sure the list to which we are adding
4429 * doesn't grow big. We need to release
4430 * alloc_semp before calling ext4_mb_add_n_trim()
4432 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4433 spin_lock(pa->pa_obj_lock);
4434 list_del_rcu(&pa->pa_inode_list);
4435 spin_unlock(pa->pa_obj_lock);
4436 ext4_mb_add_n_trim(ac);
4438 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4440 if (ac->ac_bitmap_page)
4441 page_cache_release(ac->ac_bitmap_page);
4442 if (ac->ac_buddy_page)
4443 page_cache_release(ac->ac_buddy_page);
4444 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4445 mutex_unlock(&ac->ac_lg->lg_mutex);
4446 ext4_mb_collect_stats(ac);
4450 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4452 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4456 trace_ext4_mb_discard_preallocations(sb, needed);
4457 for (i = 0; i < ngroups && needed > 0; i++) {
4458 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4467 * Main entry point into mballoc to allocate blocks
4468 * it tries to use preallocation first, then falls back
4469 * to usual allocation
4471 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4472 struct ext4_allocation_request *ar, int *errp)
4475 struct ext4_allocation_context *ac = NULL;
4476 struct ext4_sb_info *sbi;
4477 struct super_block *sb;
4478 ext4_fsblk_t block = 0;
4479 unsigned int inquota = 0;
4480 unsigned int reserv_blks = 0;
4482 sb = ar->inode->i_sb;
4485 trace_ext4_request_blocks(ar);
4488 * For delayed allocation, we could skip the ENOSPC and
4489 * EDQUOT check, as blocks and quotas have been already
4490 * reserved when data being copied into pagecache.
4492 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4493 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4495 /* Without delayed allocation we need to verify
4496 * there is enough free blocks to do block allocation
4497 * and verify allocation doesn't exceed the quota limits.
4499 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4500 /* let others to free the space */
4502 ar->len = ar->len >> 1;
4508 reserv_blks = ar->len;
4509 while (ar->len && vfs_dq_alloc_block(ar->inode, ar->len)) {
4510 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4520 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4527 *errp = ext4_mb_initialize_context(ac, ar);
4533 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4534 if (!ext4_mb_use_preallocated(ac)) {
4535 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4536 ext4_mb_normalize_request(ac, ar);
4538 /* allocate space in core */
4539 ext4_mb_regular_allocator(ac);
4541 /* as we've just preallocated more space than
4542 * user requested orinally, we store allocated
4543 * space in a special descriptor */
4544 if (ac->ac_status == AC_STATUS_FOUND &&
4545 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4546 ext4_mb_new_preallocation(ac);
4548 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4549 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4550 if (*errp == -EAGAIN) {
4552 * drop the reference that we took
4553 * in ext4_mb_use_best_found
4555 ext4_mb_release_context(ac);
4556 ac->ac_b_ex.fe_group = 0;
4557 ac->ac_b_ex.fe_start = 0;
4558 ac->ac_b_ex.fe_len = 0;
4559 ac->ac_status = AC_STATUS_CONTINUE;
4562 ac->ac_b_ex.fe_len = 0;
4564 ext4_mb_show_ac(ac);
4566 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4567 ar->len = ac->ac_b_ex.fe_len;
4570 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4574 ac->ac_b_ex.fe_len = 0;
4576 ext4_mb_show_ac(ac);
4579 ext4_mb_release_context(ac);
4582 kmem_cache_free(ext4_ac_cachep, ac);
4584 if (inquota && ar->len < inquota)
4585 vfs_dq_free_block(ar->inode, inquota - ar->len);
4588 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4589 /* release all the reserved blocks if non delalloc */
4590 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4594 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4600 * We can merge two free data extents only if the physical blocks
4601 * are contiguous, AND the extents were freed by the same transaction,
4602 * AND the blocks are associated with the same group.
4604 static int can_merge(struct ext4_free_data *entry1,
4605 struct ext4_free_data *entry2)
4607 if ((entry1->t_tid == entry2->t_tid) &&
4608 (entry1->group == entry2->group) &&
4609 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4614 static noinline_for_stack int
4615 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4616 struct ext4_free_data *new_entry)
4618 ext4_grpblk_t block;
4619 struct ext4_free_data *entry;
4620 struct ext4_group_info *db = e4b->bd_info;
4621 struct super_block *sb = e4b->bd_sb;
4622 struct ext4_sb_info *sbi = EXT4_SB(sb);
4623 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4624 struct rb_node *parent = NULL, *new_node;
4626 BUG_ON(!ext4_handle_valid(handle));
4627 BUG_ON(e4b->bd_bitmap_page == NULL);
4628 BUG_ON(e4b->bd_buddy_page == NULL);
4630 new_node = &new_entry->node;
4631 block = new_entry->start_blk;
4634 /* first free block exent. We need to
4635 protect buddy cache from being freed,
4636 * otherwise we'll refresh it from
4637 * on-disk bitmap and lose not-yet-available
4639 page_cache_get(e4b->bd_buddy_page);
4640 page_cache_get(e4b->bd_bitmap_page);
4644 entry = rb_entry(parent, struct ext4_free_data, node);
4645 if (block < entry->start_blk)
4647 else if (block >= (entry->start_blk + entry->count))
4648 n = &(*n)->rb_right;
4650 ext4_grp_locked_error(sb, e4b->bd_group, __func__,
4651 "Double free of blocks %d (%d %d)",
4652 block, entry->start_blk, entry->count);
4657 rb_link_node(new_node, parent, n);
4658 rb_insert_color(new_node, &db->bb_free_root);
4660 /* Now try to see the extent can be merged to left and right */
4661 node = rb_prev(new_node);
4663 entry = rb_entry(node, struct ext4_free_data, node);
4664 if (can_merge(entry, new_entry)) {
4665 new_entry->start_blk = entry->start_blk;
4666 new_entry->count += entry->count;
4667 rb_erase(node, &(db->bb_free_root));
4668 spin_lock(&sbi->s_md_lock);
4669 list_del(&entry->list);
4670 spin_unlock(&sbi->s_md_lock);
4671 kmem_cache_free(ext4_free_ext_cachep, entry);
4675 node = rb_next(new_node);
4677 entry = rb_entry(node, struct ext4_free_data, node);
4678 if (can_merge(new_entry, entry)) {
4679 new_entry->count += entry->count;
4680 rb_erase(node, &(db->bb_free_root));
4681 spin_lock(&sbi->s_md_lock);
4682 list_del(&entry->list);
4683 spin_unlock(&sbi->s_md_lock);
4684 kmem_cache_free(ext4_free_ext_cachep, entry);
4687 /* Add the extent to transaction's private list */
4688 spin_lock(&sbi->s_md_lock);
4689 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4690 spin_unlock(&sbi->s_md_lock);
4695 * Main entry point into mballoc to free blocks
4697 void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
4698 ext4_fsblk_t block, unsigned long count,
4699 int metadata, unsigned long *freed)
4701 struct buffer_head *bitmap_bh = NULL;
4702 struct super_block *sb = inode->i_sb;
4703 struct ext4_allocation_context *ac = NULL;
4704 struct ext4_group_desc *gdp;
4705 struct ext4_super_block *es;
4706 unsigned int overflow;
4708 struct buffer_head *gd_bh;
4709 ext4_group_t block_group;
4710 struct ext4_sb_info *sbi;
4711 struct ext4_buddy e4b;
4718 es = EXT4_SB(sb)->s_es;
4719 if (block < le32_to_cpu(es->s_first_data_block) ||
4720 block + count < block ||
4721 block + count > ext4_blocks_count(es)) {
4722 ext4_error(sb, __func__,
4723 "Freeing blocks not in datazone - "
4724 "block = %llu, count = %lu", block, count);
4728 ext4_debug("freeing block %llu\n", block);
4729 trace_ext4_free_blocks(inode, block, count, metadata);
4731 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4733 ac->ac_op = EXT4_MB_HISTORY_FREE;
4734 ac->ac_inode = inode;
4740 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4743 * Check to see if we are freeing blocks across a group
4746 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4747 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4750 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4755 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4761 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4762 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4763 in_range(block, ext4_inode_table(sb, gdp),
4764 EXT4_SB(sb)->s_itb_per_group) ||
4765 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4766 EXT4_SB(sb)->s_itb_per_group)) {
4768 ext4_error(sb, __func__,
4769 "Freeing blocks in system zone - "
4770 "Block = %llu, count = %lu", block, count);
4771 /* err = 0. ext4_std_error should be a no op */
4775 BUFFER_TRACE(bitmap_bh, "getting write access");
4776 err = ext4_journal_get_write_access(handle, bitmap_bh);
4781 * We are about to modify some metadata. Call the journal APIs
4782 * to unshare ->b_data if a currently-committing transaction is
4785 BUFFER_TRACE(gd_bh, "get_write_access");
4786 err = ext4_journal_get_write_access(handle, gd_bh);
4789 #ifdef AGGRESSIVE_CHECK
4792 for (i = 0; i < count; i++)
4793 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4797 ac->ac_b_ex.fe_group = block_group;
4798 ac->ac_b_ex.fe_start = bit;
4799 ac->ac_b_ex.fe_len = count;
4800 ext4_mb_store_history(ac);
4803 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4806 if (metadata && ext4_handle_valid(handle)) {
4807 struct ext4_free_data *new_entry;
4809 * blocks being freed are metadata. these blocks shouldn't
4810 * be used until this transaction is committed
4812 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4813 new_entry->start_blk = bit;
4814 new_entry->group = block_group;
4815 new_entry->count = count;
4816 new_entry->t_tid = handle->h_transaction->t_tid;
4818 ext4_lock_group(sb, block_group);
4819 mb_clear_bits(bitmap_bh->b_data, bit, count);
4820 ext4_mb_free_metadata(handle, &e4b, new_entry);
4822 /* need to update group_info->bb_free and bitmap
4823 * with group lock held. generate_buddy look at
4824 * them with group lock_held
4826 ext4_lock_group(sb, block_group);
4827 mb_clear_bits(bitmap_bh->b_data, bit, count);
4828 mb_free_blocks(inode, &e4b, bit, count);
4829 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4832 ret = ext4_free_blks_count(sb, gdp) + count;
4833 ext4_free_blks_set(sb, gdp, ret);
4834 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4835 ext4_unlock_group(sb, block_group);
4836 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4838 if (sbi->s_log_groups_per_flex) {
4839 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4840 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4843 ext4_mb_release_desc(&e4b);
4847 /* We dirtied the bitmap block */
4848 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4849 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4851 /* And the group descriptor block */
4852 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4853 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4857 if (overflow && !err) {
4866 ext4_std_error(sb, err);
4868 kmem_cache_free(ext4_ac_cachep, ac);