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 <trace/events/ext4.h>
29 * - test ext4_ext_search_left() and ext4_ext_search_right()
30 * - search for metadata in few groups
33 * - normalization should take into account whether file is still open
34 * - discard preallocations if no free space left (policy?)
35 * - don't normalize tails
37 * - reservation for superuser
40 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
41 * - track min/max extents in each group for better group selection
42 * - mb_mark_used() may allocate chunk right after splitting buddy
43 * - tree of groups sorted by number of free blocks
48 * The allocation request involve request for multiple number of blocks
49 * near to the goal(block) value specified.
51 * During initialization phase of the allocator we decide to use the
52 * group preallocation or inode preallocation depending on the size of
53 * the file. The size of the file could be the resulting file size we
54 * would have after allocation, or the current file size, which ever
55 * is larger. If the size is less than sbi->s_mb_stream_request we
56 * select to use the group preallocation. The default value of
57 * s_mb_stream_request is 16 blocks. This can also be tuned via
58 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
59 * terms of number of blocks.
61 * The main motivation for having small file use group preallocation is to
62 * ensure that we have small files closer together on the disk.
64 * First stage the allocator looks at the inode prealloc list,
65 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
66 * spaces for this particular inode. The inode prealloc space is
69 * pa_lstart -> the logical start block for this prealloc space
70 * pa_pstart -> the physical start block for this prealloc space
71 * pa_len -> lenght for this prealloc space
72 * pa_free -> free space available in this prealloc space
74 * The inode preallocation space is used looking at the _logical_ start
75 * block. If only the logical file block falls within the range of prealloc
76 * space we will consume the particular prealloc space. This make sure that
77 * that the we have contiguous physical blocks representing the file blocks
79 * The important thing to be noted in case of inode prealloc space is that
80 * we don't modify the values associated to inode prealloc space except
83 * If we are not able to find blocks in the inode prealloc space and if we
84 * have the group allocation flag set then we look at the locality group
85 * prealloc space. These are per CPU prealloc list repreasented as
87 * ext4_sb_info.s_locality_groups[smp_processor_id()]
89 * The reason for having a per cpu locality group is to reduce the contention
90 * between CPUs. It is possible to get scheduled at this point.
92 * The locality group prealloc space is used looking at whether we have
93 * enough free space (pa_free) withing the prealloc space.
95 * If we can't allocate blocks via inode prealloc or/and locality group
96 * prealloc then we look at the buddy cache. The buddy cache is represented
97 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
98 * mapped to the buddy and bitmap information regarding different
99 * groups. The buddy information is attached to buddy cache inode so that
100 * we can access them through the page cache. The information regarding
101 * each group is loaded via ext4_mb_load_buddy. The information involve
102 * block bitmap and buddy information. The information are stored in the
106 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
109 * one block each for bitmap and buddy information. So for each group we
110 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
111 * blocksize) blocks. So it can have information regarding groups_per_page
112 * which is blocks_per_page/2
114 * The buddy cache inode is not stored on disk. The inode is thrown
115 * away when the filesystem is unmounted.
117 * We look for count number of blocks in the buddy cache. If we were able
118 * to locate that many free blocks we return with additional information
119 * regarding rest of the contiguous physical block available
121 * Before allocating blocks via buddy cache we normalize the request
122 * blocks. This ensure we ask for more blocks that we needed. The extra
123 * blocks that we get after allocation is added to the respective prealloc
124 * list. In case of inode preallocation we follow a list of heuristics
125 * based on file size. This can be found in ext4_mb_normalize_request. If
126 * we are doing a group prealloc we try to normalize the request to
127 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
128 * 512 blocks. This can be tuned via
129 * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
130 * terms of number of blocks. If we have mounted the file system with -O
131 * stripe=<value> option the group prealloc request is normalized to the
132 * stripe value (sbi->s_stripe)
134 * The regular allocator(using the buddy cache) supports few tunables.
136 * /sys/fs/ext4/<partition>/mb_min_to_scan
137 * /sys/fs/ext4/<partition>/mb_max_to_scan
138 * /sys/fs/ext4/<partition>/mb_order2_req
140 * The regular allocator uses buddy scan only if the request len is power of
141 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
142 * value of s_mb_order2_reqs can be tuned via
143 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
144 * stripe size (sbi->s_stripe), we try to search for contigous block in
145 * stripe size. This should result in better allocation on RAID setups. If
146 * not, we search in the specific group using bitmap for best extents. The
147 * tunable min_to_scan and max_to_scan control the behaviour here.
148 * min_to_scan indicate how long the mballoc __must__ look for a best
149 * extent and max_to_scan indicates how long the mballoc __can__ look for a
150 * best extent in the found extents. Searching for the blocks starts with
151 * the group specified as the goal value in allocation context via
152 * ac_g_ex. Each group is first checked based on the criteria whether it
153 * can used for allocation. ext4_mb_good_group explains how the groups are
156 * Both the prealloc space are getting populated as above. So for the first
157 * request we will hit the buddy cache which will result in this prealloc
158 * space getting filled. The prealloc space is then later used for the
159 * subsequent request.
163 * mballoc operates on the following data:
165 * - in-core buddy (actually includes buddy and bitmap)
166 * - preallocation descriptors (PAs)
168 * there are two types of preallocations:
170 * assiged to specific inode and can be used for this inode only.
171 * it describes part of inode's space preallocated to specific
172 * physical blocks. any block from that preallocated can be used
173 * independent. the descriptor just tracks number of blocks left
174 * unused. so, before taking some block from descriptor, one must
175 * make sure corresponded logical block isn't allocated yet. this
176 * also means that freeing any block within descriptor's range
177 * must discard all preallocated blocks.
179 * assigned to specific locality group which does not translate to
180 * permanent set of inodes: inode can join and leave group. space
181 * from this type of preallocation can be used for any inode. thus
182 * it's consumed from the beginning to the end.
184 * relation between them can be expressed as:
185 * in-core buddy = on-disk bitmap + preallocation descriptors
187 * this mean blocks mballoc considers used are:
188 * - allocated blocks (persistent)
189 * - preallocated blocks (non-persistent)
191 * consistency in mballoc world means that at any time a block is either
192 * free or used in ALL structures. notice: "any time" should not be read
193 * literally -- time is discrete and delimited by locks.
195 * to keep it simple, we don't use block numbers, instead we count number of
196 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
198 * all operations can be expressed as:
199 * - init buddy: buddy = on-disk + PAs
200 * - new PA: buddy += N; PA = N
201 * - use inode PA: on-disk += N; PA -= N
202 * - discard inode PA buddy -= on-disk - PA; PA = 0
203 * - use locality group PA on-disk += N; PA -= N
204 * - discard locality group PA buddy -= PA; PA = 0
205 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
206 * is used in real operation because we can't know actual used
207 * bits from PA, only from on-disk bitmap
209 * if we follow this strict logic, then all operations above should be atomic.
210 * given some of them can block, we'd have to use something like semaphores
211 * killing performance on high-end SMP hardware. let's try to relax it using
212 * the following knowledge:
213 * 1) if buddy is referenced, it's already initialized
214 * 2) while block is used in buddy and the buddy is referenced,
215 * nobody can re-allocate that block
216 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
217 * bit set and PA claims same block, it's OK. IOW, one can set bit in
218 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
221 * so, now we're building a concurrency table:
224 * blocks for PA are allocated in the buddy, buddy must be referenced
225 * until PA is linked to allocation group to avoid concurrent buddy init
227 * we need to make sure that either on-disk bitmap or PA has uptodate data
228 * given (3) we care that PA-=N operation doesn't interfere with init
230 * the simplest way would be to have buddy initialized by the discard
231 * - use locality group PA
232 * again PA-=N must be serialized with init
233 * - discard locality group PA
234 * the simplest way would be to have buddy initialized by the discard
237 * i_data_sem serializes them
239 * discard process must wait until PA isn't used by another process
240 * - use locality group PA
241 * some mutex should serialize them
242 * - discard locality group PA
243 * discard process must wait until PA isn't used by another process
246 * i_data_sem or another mutex should serializes them
248 * discard process must wait until PA isn't used by another process
249 * - use locality group PA
250 * nothing wrong here -- they're different PAs covering different blocks
251 * - discard locality group PA
252 * discard process must wait until PA isn't used by another process
254 * now we're ready to make few consequences:
255 * - PA is referenced and while it is no discard is possible
256 * - PA is referenced until block isn't marked in on-disk bitmap
257 * - PA changes only after on-disk bitmap
258 * - discard must not compete with init. either init is done before
259 * any discard or they're serialized somehow
260 * - buddy init as sum of on-disk bitmap and PAs is done atomically
262 * a special case when we've used PA to emptiness. no need to modify buddy
263 * in this case, but we should care about concurrent init
268 * Logic in few words:
273 * mark bits in on-disk bitmap
276 * - use preallocation:
277 * find proper PA (per-inode or group)
279 * mark bits in on-disk bitmap
285 * mark bits in on-disk bitmap
288 * - discard preallocations in group:
290 * move them onto local list
291 * load on-disk bitmap
293 * remove PA from object (inode or locality group)
294 * mark free blocks in-core
296 * - discard inode's preallocations:
303 * - bitlock on a group (group)
304 * - object (inode/locality) (object)
315 * - release consumed pa:
320 * - generate in-core bitmap:
324 * - discard all for given object (inode, locality group):
329 * - discard all for given group:
336 static struct kmem_cache *ext4_pspace_cachep;
337 static struct kmem_cache *ext4_ac_cachep;
338 static struct kmem_cache *ext4_free_ext_cachep;
339 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
341 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
343 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
345 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
347 #if BITS_PER_LONG == 64
348 *bit += ((unsigned long) addr & 7UL) << 3;
349 addr = (void *) ((unsigned long) addr & ~7UL);
350 #elif BITS_PER_LONG == 32
351 *bit += ((unsigned long) addr & 3UL) << 3;
352 addr = (void *) ((unsigned long) addr & ~3UL);
354 #error "how many bits you are?!"
359 static inline int mb_test_bit(int bit, void *addr)
362 * ext4_test_bit on architecture like powerpc
363 * needs unsigned long aligned address
365 addr = mb_correct_addr_and_bit(&bit, addr);
366 return ext4_test_bit(bit, addr);
369 static inline void mb_set_bit(int bit, void *addr)
371 addr = mb_correct_addr_and_bit(&bit, addr);
372 ext4_set_bit(bit, addr);
375 static inline void mb_clear_bit(int bit, void *addr)
377 addr = mb_correct_addr_and_bit(&bit, addr);
378 ext4_clear_bit(bit, addr);
381 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
383 int fix = 0, ret, tmpmax;
384 addr = mb_correct_addr_and_bit(&fix, addr);
388 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
394 static inline int mb_find_next_bit(void *addr, int max, int start)
396 int fix = 0, ret, tmpmax;
397 addr = mb_correct_addr_and_bit(&fix, addr);
401 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
407 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
411 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
414 if (order > e4b->bd_blkbits + 1) {
419 /* at order 0 we see each particular block */
420 *max = 1 << (e4b->bd_blkbits + 3);
422 return EXT4_MB_BITMAP(e4b);
424 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
425 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
431 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
432 int first, int count)
435 struct super_block *sb = e4b->bd_sb;
437 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
439 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
440 for (i = 0; i < count; i++) {
441 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
442 ext4_fsblk_t blocknr;
443 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
444 blocknr += first + i;
446 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
447 ext4_grp_locked_error(sb, e4b->bd_group,
448 __func__, "double-free of inode"
449 " %lu's block %llu(bit %u in group %u)",
450 inode ? inode->i_ino : 0, blocknr,
451 first + i, e4b->bd_group);
453 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
457 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
461 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
463 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
464 for (i = 0; i < count; i++) {
465 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
466 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
470 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
472 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
473 unsigned char *b1, *b2;
475 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
476 b2 = (unsigned char *) bitmap;
477 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
478 if (b1[i] != b2[i]) {
479 printk(KERN_ERR "corruption in group %u "
480 "at byte %u(%u): %x in copy != %x "
481 "on disk/prealloc\n",
482 e4b->bd_group, i, i * 8, b1[i], b2[i]);
490 static inline void mb_free_blocks_double(struct inode *inode,
491 struct ext4_buddy *e4b, int first, int count)
495 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
496 int first, int count)
500 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
506 #ifdef AGGRESSIVE_CHECK
508 #define MB_CHECK_ASSERT(assert) \
512 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
513 function, file, line, # assert); \
518 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
519 const char *function, int line)
521 struct super_block *sb = e4b->bd_sb;
522 int order = e4b->bd_blkbits + 1;
529 struct ext4_group_info *grp;
532 struct list_head *cur;
537 static int mb_check_counter;
538 if (mb_check_counter++ % 100 != 0)
543 buddy = mb_find_buddy(e4b, order, &max);
544 MB_CHECK_ASSERT(buddy);
545 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
546 MB_CHECK_ASSERT(buddy2);
547 MB_CHECK_ASSERT(buddy != buddy2);
548 MB_CHECK_ASSERT(max * 2 == max2);
551 for (i = 0; i < max; i++) {
553 if (mb_test_bit(i, buddy)) {
554 /* only single bit in buddy2 may be 1 */
555 if (!mb_test_bit(i << 1, buddy2)) {
557 mb_test_bit((i<<1)+1, buddy2));
558 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
560 mb_test_bit(i << 1, buddy2));
565 /* both bits in buddy2 must be 0 */
566 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
567 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
569 for (j = 0; j < (1 << order); j++) {
570 k = (i * (1 << order)) + j;
572 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
576 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
581 buddy = mb_find_buddy(e4b, 0, &max);
582 for (i = 0; i < max; i++) {
583 if (!mb_test_bit(i, buddy)) {
584 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
592 /* check used bits only */
593 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
594 buddy2 = mb_find_buddy(e4b, j, &max2);
596 MB_CHECK_ASSERT(k < max2);
597 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
600 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
601 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
603 grp = ext4_get_group_info(sb, e4b->bd_group);
604 buddy = mb_find_buddy(e4b, 0, &max);
605 list_for_each(cur, &grp->bb_prealloc_list) {
606 ext4_group_t groupnr;
607 struct ext4_prealloc_space *pa;
608 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
609 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
610 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
611 for (i = 0; i < pa->pa_len; i++)
612 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
616 #undef MB_CHECK_ASSERT
617 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
618 __FILE__, __func__, __LINE__)
620 #define mb_check_buddy(e4b)
623 /* FIXME!! need more doc */
624 static void ext4_mb_mark_free_simple(struct super_block *sb,
625 void *buddy, unsigned first, int len,
626 struct ext4_group_info *grp)
628 struct ext4_sb_info *sbi = EXT4_SB(sb);
631 unsigned short chunk;
632 unsigned short border;
634 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
636 border = 2 << sb->s_blocksize_bits;
639 /* find how many blocks can be covered since this position */
640 max = ffs(first | border) - 1;
642 /* find how many blocks of power 2 we need to mark */
649 /* mark multiblock chunks only */
650 grp->bb_counters[min]++;
652 mb_clear_bit(first >> min,
653 buddy + sbi->s_mb_offsets[min]);
660 static noinline_for_stack
661 void ext4_mb_generate_buddy(struct super_block *sb,
662 void *buddy, void *bitmap, ext4_group_t group)
664 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
665 unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
666 unsigned short i = 0;
667 unsigned short first;
670 unsigned fragments = 0;
671 unsigned long long period = get_cycles();
673 /* initialize buddy from bitmap which is aggregation
674 * of on-disk bitmap and preallocations */
675 i = mb_find_next_zero_bit(bitmap, max, 0);
676 grp->bb_first_free = i;
680 i = mb_find_next_bit(bitmap, max, i);
684 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
686 grp->bb_counters[0]++;
688 i = mb_find_next_zero_bit(bitmap, max, i);
690 grp->bb_fragments = fragments;
692 if (free != grp->bb_free) {
693 ext4_grp_locked_error(sb, group, __func__,
694 "EXT4-fs: group %u: %u blocks in bitmap, %u in gd",
695 group, free, grp->bb_free);
697 * If we intent to continue, we consider group descritor
698 * corrupt and update bb_free using bitmap value
703 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
705 period = get_cycles() - period;
706 spin_lock(&EXT4_SB(sb)->s_bal_lock);
707 EXT4_SB(sb)->s_mb_buddies_generated++;
708 EXT4_SB(sb)->s_mb_generation_time += period;
709 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
712 /* The buddy information is attached the buddy cache inode
713 * for convenience. The information regarding each group
714 * is loaded via ext4_mb_load_buddy. The information involve
715 * block bitmap and buddy information. The information are
716 * stored in the inode as
719 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
722 * one block each for bitmap and buddy information.
723 * So for each group we take up 2 blocks. A page can
724 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
725 * So it can have information regarding groups_per_page which
726 * is blocks_per_page/2
729 static int ext4_mb_init_cache(struct page *page, char *incore)
731 ext4_group_t ngroups;
737 ext4_group_t first_group;
739 struct super_block *sb;
740 struct buffer_head *bhs;
741 struct buffer_head **bh;
746 mb_debug("init page %lu\n", page->index);
748 inode = page->mapping->host;
750 ngroups = ext4_get_groups_count(sb);
751 blocksize = 1 << inode->i_blkbits;
752 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
754 groups_per_page = blocks_per_page >> 1;
755 if (groups_per_page == 0)
758 /* allocate buffer_heads to read bitmaps */
759 if (groups_per_page > 1) {
761 i = sizeof(struct buffer_head *) * groups_per_page;
762 bh = kzalloc(i, GFP_NOFS);
768 first_group = page->index * blocks_per_page / 2;
770 /* read all groups the page covers into the cache */
771 for (i = 0; i < groups_per_page; i++) {
772 struct ext4_group_desc *desc;
774 if (first_group + i >= ngroups)
778 desc = ext4_get_group_desc(sb, first_group + i, NULL);
783 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
787 if (bitmap_uptodate(bh[i]))
791 if (bitmap_uptodate(bh[i])) {
792 unlock_buffer(bh[i]);
795 ext4_lock_group(sb, first_group + i);
796 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
797 ext4_init_block_bitmap(sb, bh[i],
798 first_group + i, desc);
799 set_bitmap_uptodate(bh[i]);
800 set_buffer_uptodate(bh[i]);
801 ext4_unlock_group(sb, first_group + i);
802 unlock_buffer(bh[i]);
805 ext4_unlock_group(sb, first_group + i);
806 if (buffer_uptodate(bh[i])) {
808 * if not uninit if bh is uptodate,
809 * bitmap is also uptodate
811 set_bitmap_uptodate(bh[i]);
812 unlock_buffer(bh[i]);
817 * submit the buffer_head for read. We can
818 * safely mark the bitmap as uptodate now.
819 * We do it here so the bitmap uptodate bit
820 * get set with buffer lock held.
822 set_bitmap_uptodate(bh[i]);
823 bh[i]->b_end_io = end_buffer_read_sync;
824 submit_bh(READ, bh[i]);
825 mb_debug("read bitmap for group %u\n", first_group + i);
828 /* wait for I/O completion */
829 for (i = 0; i < groups_per_page && bh[i]; i++)
830 wait_on_buffer(bh[i]);
833 for (i = 0; i < groups_per_page && bh[i]; i++)
834 if (!buffer_uptodate(bh[i]))
838 first_block = page->index * blocks_per_page;
840 memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
841 for (i = 0; i < blocks_per_page; i++) {
843 struct ext4_group_info *grinfo;
845 group = (first_block + i) >> 1;
846 if (group >= ngroups)
850 * data carry information regarding this
851 * particular group in the format specified
855 data = page_address(page) + (i * blocksize);
856 bitmap = bh[group - first_group]->b_data;
859 * We place the buddy block and bitmap block
862 if ((first_block + i) & 1) {
863 /* this is block of buddy */
864 BUG_ON(incore == NULL);
865 mb_debug("put buddy for group %u in page %lu/%x\n",
866 group, page->index, i * blocksize);
867 grinfo = ext4_get_group_info(sb, group);
868 grinfo->bb_fragments = 0;
869 memset(grinfo->bb_counters, 0,
870 sizeof(unsigned short)*(sb->s_blocksize_bits+2));
872 * incore got set to the group block bitmap below
874 ext4_lock_group(sb, group);
875 ext4_mb_generate_buddy(sb, data, incore, group);
876 ext4_unlock_group(sb, group);
879 /* this is block of bitmap */
880 BUG_ON(incore != NULL);
881 mb_debug("put bitmap for group %u in page %lu/%x\n",
882 group, page->index, i * blocksize);
884 /* see comments in ext4_mb_put_pa() */
885 ext4_lock_group(sb, group);
886 memcpy(data, bitmap, blocksize);
888 /* mark all preallocated blks used in in-core bitmap */
889 ext4_mb_generate_from_pa(sb, data, group);
890 ext4_mb_generate_from_freelist(sb, data, group);
891 ext4_unlock_group(sb, group);
893 /* set incore so that the buddy information can be
894 * generated using this
899 SetPageUptodate(page);
903 for (i = 0; i < groups_per_page && bh[i]; i++)
911 static noinline_for_stack int
912 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
913 struct ext4_buddy *e4b)
921 struct ext4_group_info *grp;
922 struct ext4_sb_info *sbi = EXT4_SB(sb);
923 struct inode *inode = sbi->s_buddy_cache;
925 mb_debug("load group %u\n", group);
927 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
928 grp = ext4_get_group_info(sb, group);
930 e4b->bd_blkbits = sb->s_blocksize_bits;
931 e4b->bd_info = ext4_get_group_info(sb, group);
933 e4b->bd_group = group;
934 e4b->bd_buddy_page = NULL;
935 e4b->bd_bitmap_page = NULL;
936 e4b->alloc_semp = &grp->alloc_sem;
938 /* Take the read lock on the group alloc
939 * sem. This would make sure a parallel
940 * ext4_mb_init_group happening on other
941 * groups mapped by the page is blocked
942 * till we are done with allocation
944 down_read(e4b->alloc_semp);
947 * the buddy cache inode stores the block bitmap
948 * and buddy information in consecutive blocks.
949 * So for each group we need two blocks.
952 pnum = block / blocks_per_page;
953 poff = block % blocks_per_page;
955 /* we could use find_or_create_page(), but it locks page
956 * what we'd like to avoid in fast path ... */
957 page = find_get_page(inode->i_mapping, pnum);
958 if (page == NULL || !PageUptodate(page)) {
961 * drop the page reference and try
962 * to get the page with lock. If we
963 * are not uptodate that implies
964 * somebody just created the page but
965 * is yet to initialize the same. So
966 * wait for it to initialize.
968 page_cache_release(page);
969 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
971 BUG_ON(page->mapping != inode->i_mapping);
972 if (!PageUptodate(page)) {
973 ret = ext4_mb_init_cache(page, NULL);
978 mb_cmp_bitmaps(e4b, page_address(page) +
979 (poff * sb->s_blocksize));
984 if (page == NULL || !PageUptodate(page)) {
988 e4b->bd_bitmap_page = page;
989 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
990 mark_page_accessed(page);
993 pnum = block / blocks_per_page;
994 poff = block % blocks_per_page;
996 page = find_get_page(inode->i_mapping, pnum);
997 if (page == NULL || !PageUptodate(page)) {
999 page_cache_release(page);
1000 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1002 BUG_ON(page->mapping != inode->i_mapping);
1003 if (!PageUptodate(page)) {
1004 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1013 if (page == NULL || !PageUptodate(page)) {
1017 e4b->bd_buddy_page = page;
1018 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1019 mark_page_accessed(page);
1021 BUG_ON(e4b->bd_bitmap_page == NULL);
1022 BUG_ON(e4b->bd_buddy_page == NULL);
1027 if (e4b->bd_bitmap_page)
1028 page_cache_release(e4b->bd_bitmap_page);
1029 if (e4b->bd_buddy_page)
1030 page_cache_release(e4b->bd_buddy_page);
1031 e4b->bd_buddy = NULL;
1032 e4b->bd_bitmap = NULL;
1034 /* Done with the buddy cache */
1035 up_read(e4b->alloc_semp);
1039 static void ext4_mb_release_desc(struct ext4_buddy *e4b)
1041 if (e4b->bd_bitmap_page)
1042 page_cache_release(e4b->bd_bitmap_page);
1043 if (e4b->bd_buddy_page)
1044 page_cache_release(e4b->bd_buddy_page);
1045 /* Done with the buddy cache */
1046 if (e4b->alloc_semp)
1047 up_read(e4b->alloc_semp);
1051 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1056 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1057 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1059 bb = EXT4_MB_BUDDY(e4b);
1060 while (order <= e4b->bd_blkbits + 1) {
1062 if (!mb_test_bit(block, bb)) {
1063 /* this block is part of buddy of order 'order' */
1066 bb += 1 << (e4b->bd_blkbits - order);
1072 static void mb_clear_bits(void *bm, int cur, int len)
1078 if ((cur & 31) == 0 && (len - cur) >= 32) {
1079 /* fast path: clear whole word at once */
1080 addr = bm + (cur >> 3);
1085 mb_clear_bit(cur, bm);
1090 static void mb_set_bits(void *bm, int cur, int len)
1096 if ((cur & 31) == 0 && (len - cur) >= 32) {
1097 /* fast path: set whole word at once */
1098 addr = bm + (cur >> 3);
1103 mb_set_bit(cur, bm);
1108 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1109 int first, int count)
1116 struct super_block *sb = e4b->bd_sb;
1118 BUG_ON(first + count > (sb->s_blocksize << 3));
1119 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1120 mb_check_buddy(e4b);
1121 mb_free_blocks_double(inode, e4b, first, count);
1123 e4b->bd_info->bb_free += count;
1124 if (first < e4b->bd_info->bb_first_free)
1125 e4b->bd_info->bb_first_free = first;
1127 /* let's maintain fragments counter */
1129 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1130 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1131 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1133 e4b->bd_info->bb_fragments--;
1134 else if (!block && !max)
1135 e4b->bd_info->bb_fragments++;
1137 /* let's maintain buddy itself */
1138 while (count-- > 0) {
1142 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1143 ext4_fsblk_t blocknr;
1144 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1147 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
1148 ext4_grp_locked_error(sb, e4b->bd_group,
1149 __func__, "double-free of inode"
1150 " %lu's block %llu(bit %u in group %u)",
1151 inode ? inode->i_ino : 0, blocknr, block,
1154 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1155 e4b->bd_info->bb_counters[order]++;
1157 /* start of the buddy */
1158 buddy = mb_find_buddy(e4b, order, &max);
1162 if (mb_test_bit(block, buddy) ||
1163 mb_test_bit(block + 1, buddy))
1166 /* both the buddies are free, try to coalesce them */
1167 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1173 /* for special purposes, we don't set
1174 * free bits in bitmap */
1175 mb_set_bit(block, buddy);
1176 mb_set_bit(block + 1, buddy);
1178 e4b->bd_info->bb_counters[order]--;
1179 e4b->bd_info->bb_counters[order]--;
1183 e4b->bd_info->bb_counters[order]++;
1185 mb_clear_bit(block, buddy2);
1189 mb_check_buddy(e4b);
1192 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1193 int needed, struct ext4_free_extent *ex)
1200 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1203 buddy = mb_find_buddy(e4b, order, &max);
1204 BUG_ON(buddy == NULL);
1205 BUG_ON(block >= max);
1206 if (mb_test_bit(block, buddy)) {
1213 /* FIXME dorp order completely ? */
1214 if (likely(order == 0)) {
1215 /* find actual order */
1216 order = mb_find_order_for_block(e4b, block);
1217 block = block >> order;
1220 ex->fe_len = 1 << order;
1221 ex->fe_start = block << order;
1222 ex->fe_group = e4b->bd_group;
1224 /* calc difference from given start */
1225 next = next - ex->fe_start;
1227 ex->fe_start += next;
1229 while (needed > ex->fe_len &&
1230 (buddy = mb_find_buddy(e4b, order, &max))) {
1232 if (block + 1 >= max)
1235 next = (block + 1) * (1 << order);
1236 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1239 ord = mb_find_order_for_block(e4b, next);
1242 block = next >> order;
1243 ex->fe_len += 1 << order;
1246 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1250 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1256 int start = ex->fe_start;
1257 int len = ex->fe_len;
1262 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1263 BUG_ON(e4b->bd_group != ex->fe_group);
1264 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1265 mb_check_buddy(e4b);
1266 mb_mark_used_double(e4b, start, len);
1268 e4b->bd_info->bb_free -= len;
1269 if (e4b->bd_info->bb_first_free == start)
1270 e4b->bd_info->bb_first_free += len;
1272 /* let's maintain fragments counter */
1274 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1275 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1276 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1278 e4b->bd_info->bb_fragments++;
1279 else if (!mlen && !max)
1280 e4b->bd_info->bb_fragments--;
1282 /* let's maintain buddy itself */
1284 ord = mb_find_order_for_block(e4b, start);
1286 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1287 /* the whole chunk may be allocated at once! */
1289 buddy = mb_find_buddy(e4b, ord, &max);
1290 BUG_ON((start >> ord) >= max);
1291 mb_set_bit(start >> ord, buddy);
1292 e4b->bd_info->bb_counters[ord]--;
1299 /* store for history */
1301 ret = len | (ord << 16);
1303 /* we have to split large buddy */
1305 buddy = mb_find_buddy(e4b, ord, &max);
1306 mb_set_bit(start >> ord, buddy);
1307 e4b->bd_info->bb_counters[ord]--;
1310 cur = (start >> ord) & ~1U;
1311 buddy = mb_find_buddy(e4b, ord, &max);
1312 mb_clear_bit(cur, buddy);
1313 mb_clear_bit(cur + 1, buddy);
1314 e4b->bd_info->bb_counters[ord]++;
1315 e4b->bd_info->bb_counters[ord]++;
1318 mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1319 mb_check_buddy(e4b);
1325 * Must be called under group lock!
1327 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1328 struct ext4_buddy *e4b)
1330 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1333 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1334 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1336 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1337 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1338 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1340 /* preallocation can change ac_b_ex, thus we store actually
1341 * allocated blocks for history */
1342 ac->ac_f_ex = ac->ac_b_ex;
1344 ac->ac_status = AC_STATUS_FOUND;
1345 ac->ac_tail = ret & 0xffff;
1346 ac->ac_buddy = ret >> 16;
1349 * take the page reference. We want the page to be pinned
1350 * so that we don't get a ext4_mb_init_cache_call for this
1351 * group until we update the bitmap. That would mean we
1352 * double allocate blocks. The reference is dropped
1353 * in ext4_mb_release_context
1355 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1356 get_page(ac->ac_bitmap_page);
1357 ac->ac_buddy_page = e4b->bd_buddy_page;
1358 get_page(ac->ac_buddy_page);
1359 /* on allocation we use ac to track the held semaphore */
1360 ac->alloc_semp = e4b->alloc_semp;
1361 e4b->alloc_semp = NULL;
1362 /* store last allocated for subsequent stream allocation */
1363 if ((ac->ac_flags & EXT4_MB_HINT_DATA)) {
1364 spin_lock(&sbi->s_md_lock);
1365 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1366 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1367 spin_unlock(&sbi->s_md_lock);
1372 * regular allocator, for general purposes allocation
1375 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1376 struct ext4_buddy *e4b,
1379 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1380 struct ext4_free_extent *bex = &ac->ac_b_ex;
1381 struct ext4_free_extent *gex = &ac->ac_g_ex;
1382 struct ext4_free_extent ex;
1385 if (ac->ac_status == AC_STATUS_FOUND)
1388 * We don't want to scan for a whole year
1390 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1391 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1392 ac->ac_status = AC_STATUS_BREAK;
1397 * Haven't found good chunk so far, let's continue
1399 if (bex->fe_len < gex->fe_len)
1402 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1403 && bex->fe_group == e4b->bd_group) {
1404 /* recheck chunk's availability - we don't know
1405 * when it was found (within this lock-unlock
1407 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1408 if (max >= gex->fe_len) {
1409 ext4_mb_use_best_found(ac, e4b);
1416 * The routine checks whether found extent is good enough. If it is,
1417 * then the extent gets marked used and flag is set to the context
1418 * to stop scanning. Otherwise, the extent is compared with the
1419 * previous found extent and if new one is better, then it's stored
1420 * in the context. Later, the best found extent will be used, if
1421 * mballoc can't find good enough extent.
1423 * FIXME: real allocation policy is to be designed yet!
1425 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1426 struct ext4_free_extent *ex,
1427 struct ext4_buddy *e4b)
1429 struct ext4_free_extent *bex = &ac->ac_b_ex;
1430 struct ext4_free_extent *gex = &ac->ac_g_ex;
1432 BUG_ON(ex->fe_len <= 0);
1433 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1434 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1435 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1440 * The special case - take what you catch first
1442 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1444 ext4_mb_use_best_found(ac, e4b);
1449 * Let's check whether the chuck is good enough
1451 if (ex->fe_len == gex->fe_len) {
1453 ext4_mb_use_best_found(ac, e4b);
1458 * If this is first found extent, just store it in the context
1460 if (bex->fe_len == 0) {
1466 * If new found extent is better, store it in the context
1468 if (bex->fe_len < gex->fe_len) {
1469 /* if the request isn't satisfied, any found extent
1470 * larger than previous best one is better */
1471 if (ex->fe_len > bex->fe_len)
1473 } else if (ex->fe_len > gex->fe_len) {
1474 /* if the request is satisfied, then we try to find
1475 * an extent that still satisfy the request, but is
1476 * smaller than previous one */
1477 if (ex->fe_len < bex->fe_len)
1481 ext4_mb_check_limits(ac, e4b, 0);
1484 static noinline_for_stack
1485 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1486 struct ext4_buddy *e4b)
1488 struct ext4_free_extent ex = ac->ac_b_ex;
1489 ext4_group_t group = ex.fe_group;
1493 BUG_ON(ex.fe_len <= 0);
1494 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1498 ext4_lock_group(ac->ac_sb, group);
1499 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1503 ext4_mb_use_best_found(ac, e4b);
1506 ext4_unlock_group(ac->ac_sb, group);
1507 ext4_mb_release_desc(e4b);
1512 static noinline_for_stack
1513 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1514 struct ext4_buddy *e4b)
1516 ext4_group_t group = ac->ac_g_ex.fe_group;
1519 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1520 struct ext4_super_block *es = sbi->s_es;
1521 struct ext4_free_extent ex;
1523 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1526 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1530 ext4_lock_group(ac->ac_sb, group);
1531 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1532 ac->ac_g_ex.fe_len, &ex);
1534 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1537 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1538 ex.fe_start + le32_to_cpu(es->s_first_data_block);
1539 /* use do_div to get remainder (would be 64-bit modulo) */
1540 if (do_div(start, sbi->s_stripe) == 0) {
1543 ext4_mb_use_best_found(ac, e4b);
1545 } else if (max >= ac->ac_g_ex.fe_len) {
1546 BUG_ON(ex.fe_len <= 0);
1547 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1548 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1551 ext4_mb_use_best_found(ac, e4b);
1552 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1553 /* Sometimes, caller may want to merge even small
1554 * number of blocks to an existing extent */
1555 BUG_ON(ex.fe_len <= 0);
1556 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1557 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1560 ext4_mb_use_best_found(ac, e4b);
1562 ext4_unlock_group(ac->ac_sb, group);
1563 ext4_mb_release_desc(e4b);
1569 * The routine scans buddy structures (not bitmap!) from given order
1570 * to max order and tries to find big enough chunk to satisfy the req
1572 static noinline_for_stack
1573 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1574 struct ext4_buddy *e4b)
1576 struct super_block *sb = ac->ac_sb;
1577 struct ext4_group_info *grp = e4b->bd_info;
1583 BUG_ON(ac->ac_2order <= 0);
1584 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1585 if (grp->bb_counters[i] == 0)
1588 buddy = mb_find_buddy(e4b, i, &max);
1589 BUG_ON(buddy == NULL);
1591 k = mb_find_next_zero_bit(buddy, max, 0);
1596 ac->ac_b_ex.fe_len = 1 << i;
1597 ac->ac_b_ex.fe_start = k << i;
1598 ac->ac_b_ex.fe_group = e4b->bd_group;
1600 ext4_mb_use_best_found(ac, e4b);
1602 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1604 if (EXT4_SB(sb)->s_mb_stats)
1605 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1612 * The routine scans the group and measures all found extents.
1613 * In order to optimize scanning, caller must pass number of
1614 * free blocks in the group, so the routine can know upper limit.
1616 static noinline_for_stack
1617 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1618 struct ext4_buddy *e4b)
1620 struct super_block *sb = ac->ac_sb;
1621 void *bitmap = EXT4_MB_BITMAP(e4b);
1622 struct ext4_free_extent ex;
1626 free = e4b->bd_info->bb_free;
1629 i = e4b->bd_info->bb_first_free;
1631 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1632 i = mb_find_next_zero_bit(bitmap,
1633 EXT4_BLOCKS_PER_GROUP(sb), i);
1634 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1636 * IF we have corrupt bitmap, we won't find any
1637 * free blocks even though group info says we
1638 * we have free blocks
1640 ext4_grp_locked_error(sb, e4b->bd_group,
1641 __func__, "%d free blocks as per "
1642 "group info. But bitmap says 0",
1647 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1648 BUG_ON(ex.fe_len <= 0);
1649 if (free < ex.fe_len) {
1650 ext4_grp_locked_error(sb, e4b->bd_group,
1651 __func__, "%d free blocks as per "
1652 "group info. But got %d blocks",
1655 * The number of free blocks differs. This mostly
1656 * indicate that the bitmap is corrupt. So exit
1657 * without claiming the space.
1662 ext4_mb_measure_extent(ac, &ex, e4b);
1668 ext4_mb_check_limits(ac, e4b, 1);
1672 * This is a special case for storages like raid5
1673 * we try to find stripe-aligned chunks for stripe-size requests
1674 * XXX should do so at least for multiples of stripe size as well
1676 static noinline_for_stack
1677 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1678 struct ext4_buddy *e4b)
1680 struct super_block *sb = ac->ac_sb;
1681 struct ext4_sb_info *sbi = EXT4_SB(sb);
1682 void *bitmap = EXT4_MB_BITMAP(e4b);
1683 struct ext4_free_extent ex;
1684 ext4_fsblk_t first_group_block;
1689 BUG_ON(sbi->s_stripe == 0);
1691 /* find first stripe-aligned block in group */
1692 first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1693 + le32_to_cpu(sbi->s_es->s_first_data_block);
1694 a = first_group_block + sbi->s_stripe - 1;
1695 do_div(a, sbi->s_stripe);
1696 i = (a * sbi->s_stripe) - first_group_block;
1698 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1699 if (!mb_test_bit(i, bitmap)) {
1700 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1701 if (max >= sbi->s_stripe) {
1704 ext4_mb_use_best_found(ac, e4b);
1712 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1713 ext4_group_t group, int cr)
1715 unsigned free, fragments;
1717 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1718 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1720 BUG_ON(cr < 0 || cr >= 4);
1721 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1723 free = grp->bb_free;
1724 fragments = grp->bb_fragments;
1732 BUG_ON(ac->ac_2order == 0);
1734 /* Avoid using the first bg of a flexgroup for data files */
1735 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1736 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1737 ((group % flex_size) == 0))
1740 bits = ac->ac_sb->s_blocksize_bits + 1;
1741 for (i = ac->ac_2order; i <= bits; i++)
1742 if (grp->bb_counters[i] > 0)
1746 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1750 if (free >= ac->ac_g_ex.fe_len)
1763 * lock the group_info alloc_sem of all the groups
1764 * belonging to the same buddy cache page. This
1765 * make sure other parallel operation on the buddy
1766 * cache doesn't happen whild holding the buddy cache
1769 int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1773 int blocks_per_page;
1774 int groups_per_page;
1775 ext4_group_t ngroups = ext4_get_groups_count(sb);
1776 ext4_group_t first_group;
1777 struct ext4_group_info *grp;
1779 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1781 * the buddy cache inode stores the block bitmap
1782 * and buddy information in consecutive blocks.
1783 * So for each group we need two blocks.
1786 pnum = block / blocks_per_page;
1787 first_group = pnum * blocks_per_page / 2;
1789 groups_per_page = blocks_per_page >> 1;
1790 if (groups_per_page == 0)
1791 groups_per_page = 1;
1792 /* read all groups the page covers into the cache */
1793 for (i = 0; i < groups_per_page; i++) {
1795 if ((first_group + i) >= ngroups)
1797 grp = ext4_get_group_info(sb, first_group + i);
1798 /* take all groups write allocation
1799 * semaphore. This make sure there is
1800 * no block allocation going on in any
1803 down_write_nested(&grp->alloc_sem, i);
1808 void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1809 ext4_group_t group, int locked_group)
1813 int blocks_per_page;
1814 ext4_group_t first_group;
1815 struct ext4_group_info *grp;
1817 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1819 * the buddy cache inode stores the block bitmap
1820 * and buddy information in consecutive blocks.
1821 * So for each group we need two blocks.
1824 pnum = block / blocks_per_page;
1825 first_group = pnum * blocks_per_page / 2;
1826 /* release locks on all the groups */
1827 for (i = 0; i < locked_group; i++) {
1829 grp = ext4_get_group_info(sb, first_group + i);
1830 /* take all groups write allocation
1831 * semaphore. This make sure there is
1832 * no block allocation going on in any
1835 up_write(&grp->alloc_sem);
1840 static noinline_for_stack
1841 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1846 int blocks_per_page;
1847 int block, pnum, poff;
1848 int num_grp_locked = 0;
1849 struct ext4_group_info *this_grp;
1850 struct ext4_sb_info *sbi = EXT4_SB(sb);
1851 struct inode *inode = sbi->s_buddy_cache;
1852 struct page *page = NULL, *bitmap_page = NULL;
1854 mb_debug("init group %lu\n", group);
1855 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1856 this_grp = ext4_get_group_info(sb, group);
1858 * This ensures we don't add group
1859 * to this buddy cache via resize
1861 num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group);
1862 if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
1864 * somebody initialized the group
1865 * return without doing anything
1871 * the buddy cache inode stores the block bitmap
1872 * and buddy information in consecutive blocks.
1873 * So for each group we need two blocks.
1876 pnum = block / blocks_per_page;
1877 poff = block % blocks_per_page;
1878 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1880 BUG_ON(page->mapping != inode->i_mapping);
1881 ret = ext4_mb_init_cache(page, NULL);
1888 if (page == NULL || !PageUptodate(page)) {
1892 mark_page_accessed(page);
1894 bitmap = page_address(page) + (poff * sb->s_blocksize);
1896 /* init buddy cache */
1898 pnum = block / blocks_per_page;
1899 poff = block % blocks_per_page;
1900 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1901 if (page == bitmap_page) {
1903 * If both the bitmap and buddy are in
1904 * the same page we don't need to force
1909 BUG_ON(page->mapping != inode->i_mapping);
1910 ret = ext4_mb_init_cache(page, bitmap);
1917 if (page == NULL || !PageUptodate(page)) {
1921 mark_page_accessed(page);
1923 ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
1925 page_cache_release(bitmap_page);
1927 page_cache_release(page);
1931 static noinline_for_stack int
1932 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1934 ext4_group_t ngroups, group, i;
1938 struct ext4_sb_info *sbi;
1939 struct super_block *sb;
1940 struct ext4_buddy e4b;
1945 ngroups = ext4_get_groups_count(sb);
1946 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1948 /* first, try the goal */
1949 err = ext4_mb_find_by_goal(ac, &e4b);
1950 if (err || ac->ac_status == AC_STATUS_FOUND)
1953 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1957 * ac->ac2_order is set only if the fe_len is a power of 2
1958 * if ac2_order is set we also set criteria to 0 so that we
1959 * try exact allocation using buddy.
1961 i = fls(ac->ac_g_ex.fe_len);
1964 * We search using buddy data only if the order of the request
1965 * is greater than equal to the sbi_s_mb_order2_reqs
1966 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1968 if (i >= sbi->s_mb_order2_reqs) {
1970 * This should tell if fe_len is exactly power of 2
1972 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1973 ac->ac_2order = i - 1;
1976 bsbits = ac->ac_sb->s_blocksize_bits;
1977 /* if stream allocation is enabled, use global goal */
1978 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
1979 isize = i_size_read(ac->ac_inode) >> bsbits;
1983 if (size < sbi->s_mb_stream_request &&
1984 (ac->ac_flags & EXT4_MB_HINT_DATA)) {
1985 /* TBD: may be hot point */
1986 spin_lock(&sbi->s_md_lock);
1987 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1988 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1989 spin_unlock(&sbi->s_md_lock);
1991 /* Let's just scan groups to find more-less suitable blocks */
1992 cr = ac->ac_2order ? 0 : 1;
1994 * cr == 0 try to get exact allocation,
1995 * cr == 3 try to get anything
1998 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1999 ac->ac_criteria = cr;
2001 * searching for the right group start
2002 * from the goal value specified
2004 group = ac->ac_g_ex.fe_group;
2006 for (i = 0; i < ngroups; group++, i++) {
2007 struct ext4_group_info *grp;
2008 struct ext4_group_desc *desc;
2010 if (group == ngroups)
2013 /* quick check to skip empty groups */
2014 grp = ext4_get_group_info(sb, group);
2015 if (grp->bb_free == 0)
2019 * if the group is already init we check whether it is
2020 * a good group and if not we don't load the buddy
2022 if (EXT4_MB_GRP_NEED_INIT(grp)) {
2024 * we need full data about the group
2025 * to make a good selection
2027 err = ext4_mb_init_group(sb, group);
2033 * If the particular group doesn't satisfy our
2034 * criteria we continue with the next group
2036 if (!ext4_mb_good_group(ac, group, cr))
2039 err = ext4_mb_load_buddy(sb, group, &e4b);
2043 ext4_lock_group(sb, group);
2044 if (!ext4_mb_good_group(ac, group, cr)) {
2045 /* someone did allocation from this group */
2046 ext4_unlock_group(sb, group);
2047 ext4_mb_release_desc(&e4b);
2051 ac->ac_groups_scanned++;
2052 desc = ext4_get_group_desc(sb, group, NULL);
2054 ext4_mb_simple_scan_group(ac, &e4b);
2056 ac->ac_g_ex.fe_len == sbi->s_stripe)
2057 ext4_mb_scan_aligned(ac, &e4b);
2059 ext4_mb_complex_scan_group(ac, &e4b);
2061 ext4_unlock_group(sb, group);
2062 ext4_mb_release_desc(&e4b);
2064 if (ac->ac_status != AC_STATUS_CONTINUE)
2069 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2070 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2072 * We've been searching too long. Let's try to allocate
2073 * the best chunk we've found so far
2076 ext4_mb_try_best_found(ac, &e4b);
2077 if (ac->ac_status != AC_STATUS_FOUND) {
2079 * Someone more lucky has already allocated it.
2080 * The only thing we can do is just take first
2082 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2084 ac->ac_b_ex.fe_group = 0;
2085 ac->ac_b_ex.fe_start = 0;
2086 ac->ac_b_ex.fe_len = 0;
2087 ac->ac_status = AC_STATUS_CONTINUE;
2088 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2090 atomic_inc(&sbi->s_mb_lost_chunks);
2098 #ifdef EXT4_MB_HISTORY
2099 struct ext4_mb_proc_session {
2100 struct ext4_mb_history *history;
2101 struct super_block *sb;
2106 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
2107 struct ext4_mb_history *hs,
2110 if (hs == s->history + s->max)
2112 if (!first && hs == s->history + s->start)
2114 while (hs->orig.fe_len == 0) {
2116 if (hs == s->history + s->max)
2118 if (hs == s->history + s->start)
2124 static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
2126 struct ext4_mb_proc_session *s = seq->private;
2127 struct ext4_mb_history *hs;
2131 return SEQ_START_TOKEN;
2132 hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2135 while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
2139 static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
2142 struct ext4_mb_proc_session *s = seq->private;
2143 struct ext4_mb_history *hs = v;
2146 if (v == SEQ_START_TOKEN)
2147 return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2149 return ext4_mb_history_skip_empty(s, ++hs, 0);
2152 static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
2154 char buf[25], buf2[25], buf3[25], *fmt;
2155 struct ext4_mb_history *hs = v;
2157 if (v == SEQ_START_TOKEN) {
2158 seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
2159 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
2160 "pid", "inode", "original", "goal", "result", "found",
2161 "grps", "cr", "flags", "merge", "tail", "broken");
2165 if (hs->op == EXT4_MB_HISTORY_ALLOC) {
2166 fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2167 "%-5u %-5s %-5u %-6u\n";
2168 sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group,
2169 hs->result.fe_start, hs->result.fe_len,
2170 hs->result.fe_logical);
2171 sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group,
2172 hs->orig.fe_start, hs->orig.fe_len,
2173 hs->orig.fe_logical);
2174 sprintf(buf3, "%u/%d/%u@%u", hs->goal.fe_group,
2175 hs->goal.fe_start, hs->goal.fe_len,
2176 hs->goal.fe_logical);
2177 seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
2178 hs->found, hs->groups, hs->cr, hs->flags,
2179 hs->merged ? "M" : "", hs->tail,
2180 hs->buddy ? 1 << hs->buddy : 0);
2181 } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
2182 fmt = "%-5u %-8u %-23s %-23s %-23s\n";
2183 sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group,
2184 hs->result.fe_start, hs->result.fe_len,
2185 hs->result.fe_logical);
2186 sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group,
2187 hs->orig.fe_start, hs->orig.fe_len,
2188 hs->orig.fe_logical);
2189 seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
2190 } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
2191 sprintf(buf2, "%u/%d/%u", hs->result.fe_group,
2192 hs->result.fe_start, hs->result.fe_len);
2193 seq_printf(seq, "%-5u %-8u %-23s discard\n",
2194 hs->pid, hs->ino, buf2);
2195 } else if (hs->op == EXT4_MB_HISTORY_FREE) {
2196 sprintf(buf2, "%u/%d/%u", hs->result.fe_group,
2197 hs->result.fe_start, hs->result.fe_len);
2198 seq_printf(seq, "%-5u %-8u %-23s free\n",
2199 hs->pid, hs->ino, buf2);
2204 static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
2208 static struct seq_operations ext4_mb_seq_history_ops = {
2209 .start = ext4_mb_seq_history_start,
2210 .next = ext4_mb_seq_history_next,
2211 .stop = ext4_mb_seq_history_stop,
2212 .show = ext4_mb_seq_history_show,
2215 static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
2217 struct super_block *sb = PDE(inode)->data;
2218 struct ext4_sb_info *sbi = EXT4_SB(sb);
2219 struct ext4_mb_proc_session *s;
2223 if (unlikely(sbi->s_mb_history == NULL))
2225 s = kmalloc(sizeof(*s), GFP_KERNEL);
2229 size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
2230 s->history = kmalloc(size, GFP_KERNEL);
2231 if (s->history == NULL) {
2236 spin_lock(&sbi->s_mb_history_lock);
2237 memcpy(s->history, sbi->s_mb_history, size);
2238 s->max = sbi->s_mb_history_max;
2239 s->start = sbi->s_mb_history_cur % s->max;
2240 spin_unlock(&sbi->s_mb_history_lock);
2242 rc = seq_open(file, &ext4_mb_seq_history_ops);
2244 struct seq_file *m = (struct seq_file *)file->private_data;
2254 static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
2256 struct seq_file *seq = (struct seq_file *)file->private_data;
2257 struct ext4_mb_proc_session *s = seq->private;
2260 return seq_release(inode, file);
2263 static ssize_t ext4_mb_seq_history_write(struct file *file,
2264 const char __user *buffer,
2265 size_t count, loff_t *ppos)
2267 struct seq_file *seq = (struct seq_file *)file->private_data;
2268 struct ext4_mb_proc_session *s = seq->private;
2269 struct super_block *sb = s->sb;
2273 if (count >= sizeof(str)) {
2274 printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
2275 "mb_history", (int)sizeof(str));
2279 if (copy_from_user(str, buffer, count))
2282 value = simple_strtol(str, NULL, 0);
2285 EXT4_SB(sb)->s_mb_history_filter = value;
2290 static struct file_operations ext4_mb_seq_history_fops = {
2291 .owner = THIS_MODULE,
2292 .open = ext4_mb_seq_history_open,
2294 .write = ext4_mb_seq_history_write,
2295 .llseek = seq_lseek,
2296 .release = ext4_mb_seq_history_release,
2299 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2301 struct super_block *sb = seq->private;
2304 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2307 return (void *) ((unsigned long) group);
2310 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2312 struct super_block *sb = seq->private;
2316 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2319 return (void *) ((unsigned long) group);
2322 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2324 struct super_block *sb = seq->private;
2325 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2328 struct ext4_buddy e4b;
2330 struct ext4_group_info info;
2331 unsigned short counters[16];
2336 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2337 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2338 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2339 "group", "free", "frags", "first",
2340 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2341 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2343 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2344 sizeof(struct ext4_group_info);
2345 err = ext4_mb_load_buddy(sb, group, &e4b);
2347 seq_printf(seq, "#%-5u: I/O error\n", group);
2350 ext4_lock_group(sb, group);
2351 memcpy(&sg, ext4_get_group_info(sb, group), i);
2352 ext4_unlock_group(sb, group);
2353 ext4_mb_release_desc(&e4b);
2355 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2356 sg.info.bb_fragments, sg.info.bb_first_free);
2357 for (i = 0; i <= 13; i++)
2358 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2359 sg.info.bb_counters[i] : 0);
2360 seq_printf(seq, " ]\n");
2365 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2369 static struct seq_operations ext4_mb_seq_groups_ops = {
2370 .start = ext4_mb_seq_groups_start,
2371 .next = ext4_mb_seq_groups_next,
2372 .stop = ext4_mb_seq_groups_stop,
2373 .show = ext4_mb_seq_groups_show,
2376 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2378 struct super_block *sb = PDE(inode)->data;
2381 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2383 struct seq_file *m = (struct seq_file *)file->private_data;
2390 static struct file_operations ext4_mb_seq_groups_fops = {
2391 .owner = THIS_MODULE,
2392 .open = ext4_mb_seq_groups_open,
2394 .llseek = seq_lseek,
2395 .release = seq_release,
2398 static void ext4_mb_history_release(struct super_block *sb)
2400 struct ext4_sb_info *sbi = EXT4_SB(sb);
2402 if (sbi->s_proc != NULL) {
2403 remove_proc_entry("mb_groups", sbi->s_proc);
2404 if (sbi->s_mb_history_max)
2405 remove_proc_entry("mb_history", sbi->s_proc);
2407 kfree(sbi->s_mb_history);
2410 static void ext4_mb_history_init(struct super_block *sb)
2412 struct ext4_sb_info *sbi = EXT4_SB(sb);
2415 if (sbi->s_proc != NULL) {
2416 if (sbi->s_mb_history_max)
2417 proc_create_data("mb_history", S_IRUGO, sbi->s_proc,
2418 &ext4_mb_seq_history_fops, sb);
2419 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2420 &ext4_mb_seq_groups_fops, sb);
2423 sbi->s_mb_history_cur = 0;
2424 spin_lock_init(&sbi->s_mb_history_lock);
2425 i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
2426 sbi->s_mb_history = i ? kzalloc(i, GFP_KERNEL) : NULL;
2427 /* if we can't allocate history, then we simple won't use it */
2430 static noinline_for_stack void
2431 ext4_mb_store_history(struct ext4_allocation_context *ac)
2433 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2434 struct ext4_mb_history h;
2436 if (sbi->s_mb_history == NULL)
2439 if (!(ac->ac_op & sbi->s_mb_history_filter))
2443 h.pid = current->pid;
2444 h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2445 h.orig = ac->ac_o_ex;
2446 h.result = ac->ac_b_ex;
2447 h.flags = ac->ac_flags;
2448 h.found = ac->ac_found;
2449 h.groups = ac->ac_groups_scanned;
2450 h.cr = ac->ac_criteria;
2451 h.tail = ac->ac_tail;
2452 h.buddy = ac->ac_buddy;
2454 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
2455 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2456 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2458 h.goal = ac->ac_g_ex;
2459 h.result = ac->ac_f_ex;
2462 spin_lock(&sbi->s_mb_history_lock);
2463 memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2464 if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2465 sbi->s_mb_history_cur = 0;
2466 spin_unlock(&sbi->s_mb_history_lock);
2470 #define ext4_mb_history_release(sb)
2471 #define ext4_mb_history_init(sb)
2475 /* Create and initialize ext4_group_info data for the given group. */
2476 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2477 struct ext4_group_desc *desc)
2481 struct ext4_sb_info *sbi = EXT4_SB(sb);
2482 struct ext4_group_info **meta_group_info;
2485 * First check if this group is the first of a reserved block.
2486 * If it's true, we have to allocate a new table of pointers
2487 * to ext4_group_info structures
2489 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2490 metalen = sizeof(*meta_group_info) <<
2491 EXT4_DESC_PER_BLOCK_BITS(sb);
2492 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2493 if (meta_group_info == NULL) {
2494 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2496 goto exit_meta_group_info;
2498 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2503 * calculate needed size. if change bb_counters size,
2504 * don't forget about ext4_mb_generate_buddy()
2506 len = offsetof(typeof(**meta_group_info),
2507 bb_counters[sb->s_blocksize_bits + 2]);
2510 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2511 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2513 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2514 if (meta_group_info[i] == NULL) {
2515 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2516 goto exit_group_info;
2518 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2519 &(meta_group_info[i]->bb_state));
2522 * initialize bb_free to be able to skip
2523 * empty groups without initialization
2525 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2526 meta_group_info[i]->bb_free =
2527 ext4_free_blocks_after_init(sb, group, desc);
2529 meta_group_info[i]->bb_free =
2530 ext4_free_blks_count(sb, desc);
2533 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2534 init_rwsem(&meta_group_info[i]->alloc_sem);
2535 meta_group_info[i]->bb_free_root.rb_node = NULL;;
2539 struct buffer_head *bh;
2540 meta_group_info[i]->bb_bitmap =
2541 kmalloc(sb->s_blocksize, GFP_KERNEL);
2542 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2543 bh = ext4_read_block_bitmap(sb, group);
2545 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2554 /* If a meta_group_info table has been allocated, release it now */
2555 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2556 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2557 exit_meta_group_info:
2559 } /* ext4_mb_add_groupinfo */
2562 * Update an existing group.
2563 * This function is used for online resize
2565 void ext4_mb_update_group_info(struct ext4_group_info *grp, ext4_grpblk_t add)
2567 grp->bb_free += add;
2570 static int ext4_mb_init_backend(struct super_block *sb)
2572 ext4_group_t ngroups = ext4_get_groups_count(sb);
2575 struct ext4_sb_info *sbi = EXT4_SB(sb);
2576 struct ext4_super_block *es = sbi->s_es;
2577 int num_meta_group_infos;
2578 int num_meta_group_infos_max;
2580 struct ext4_group_info **meta_group_info;
2581 struct ext4_group_desc *desc;
2583 /* This is the number of blocks used by GDT */
2584 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2585 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2588 * This is the total number of blocks used by GDT including
2589 * the number of reserved blocks for GDT.
2590 * The s_group_info array is allocated with this value
2591 * to allow a clean online resize without a complex
2592 * manipulation of pointer.
2593 * The drawback is the unused memory when no resize
2594 * occurs but it's very low in terms of pages
2595 * (see comments below)
2596 * Need to handle this properly when META_BG resizing is allowed
2598 num_meta_group_infos_max = num_meta_group_infos +
2599 le16_to_cpu(es->s_reserved_gdt_blocks);
2602 * array_size is the size of s_group_info array. We round it
2603 * to the next power of two because this approximation is done
2604 * internally by kmalloc so we can have some more memory
2605 * for free here (e.g. may be used for META_BG resize).
2608 while (array_size < sizeof(*sbi->s_group_info) *
2609 num_meta_group_infos_max)
2610 array_size = array_size << 1;
2611 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2612 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2613 * So a two level scheme suffices for now. */
2614 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2615 if (sbi->s_group_info == NULL) {
2616 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2619 sbi->s_buddy_cache = new_inode(sb);
2620 if (sbi->s_buddy_cache == NULL) {
2621 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2624 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2626 metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb);
2627 for (i = 0; i < num_meta_group_infos; i++) {
2628 if ((i + 1) == num_meta_group_infos)
2629 metalen = sizeof(*meta_group_info) *
2631 (i << EXT4_DESC_PER_BLOCK_BITS(sb)));
2632 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2633 if (meta_group_info == NULL) {
2634 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2638 sbi->s_group_info[i] = meta_group_info;
2641 for (i = 0; i < ngroups; i++) {
2642 desc = ext4_get_group_desc(sb, i, NULL);
2645 "EXT4-fs: can't read descriptor %u\n", i);
2648 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2656 kfree(ext4_get_group_info(sb, i));
2657 i = num_meta_group_infos;
2660 kfree(sbi->s_group_info[i]);
2661 iput(sbi->s_buddy_cache);
2663 kfree(sbi->s_group_info);
2667 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2669 struct ext4_sb_info *sbi = EXT4_SB(sb);
2675 i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
2677 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2678 if (sbi->s_mb_offsets == NULL) {
2682 i = (sb->s_blocksize_bits + 2) * sizeof(unsigned int);
2683 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2684 if (sbi->s_mb_maxs == NULL) {
2685 kfree(sbi->s_mb_offsets);
2689 /* order 0 is regular bitmap */
2690 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2691 sbi->s_mb_offsets[0] = 0;
2695 max = sb->s_blocksize << 2;
2697 sbi->s_mb_offsets[i] = offset;
2698 sbi->s_mb_maxs[i] = max;
2699 offset += 1 << (sb->s_blocksize_bits - i);
2702 } while (i <= sb->s_blocksize_bits + 1);
2704 /* init file for buddy data */
2705 ret = ext4_mb_init_backend(sb);
2707 kfree(sbi->s_mb_offsets);
2708 kfree(sbi->s_mb_maxs);
2712 spin_lock_init(&sbi->s_md_lock);
2713 spin_lock_init(&sbi->s_bal_lock);
2715 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2716 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2717 sbi->s_mb_stats = MB_DEFAULT_STATS;
2718 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2719 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2720 sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
2721 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2723 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2724 if (sbi->s_locality_groups == NULL) {
2725 kfree(sbi->s_mb_offsets);
2726 kfree(sbi->s_mb_maxs);
2729 for_each_possible_cpu(i) {
2730 struct ext4_locality_group *lg;
2731 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2732 mutex_init(&lg->lg_mutex);
2733 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2734 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2735 spin_lock_init(&lg->lg_prealloc_lock);
2738 ext4_mb_history_init(sb);
2741 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2743 printk(KERN_INFO "EXT4-fs: mballoc enabled\n");
2747 /* need to called with the ext4 group lock held */
2748 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2750 struct ext4_prealloc_space *pa;
2751 struct list_head *cur, *tmp;
2754 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2755 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2756 list_del(&pa->pa_group_list);
2758 kmem_cache_free(ext4_pspace_cachep, pa);
2761 mb_debug("mballoc: %u PAs left\n", count);
2765 int ext4_mb_release(struct super_block *sb)
2767 ext4_group_t ngroups = ext4_get_groups_count(sb);
2769 int num_meta_group_infos;
2770 struct ext4_group_info *grinfo;
2771 struct ext4_sb_info *sbi = EXT4_SB(sb);
2773 if (sbi->s_group_info) {
2774 for (i = 0; i < ngroups; i++) {
2775 grinfo = ext4_get_group_info(sb, i);
2777 kfree(grinfo->bb_bitmap);
2779 ext4_lock_group(sb, i);
2780 ext4_mb_cleanup_pa(grinfo);
2781 ext4_unlock_group(sb, i);
2784 num_meta_group_infos = (ngroups +
2785 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2786 EXT4_DESC_PER_BLOCK_BITS(sb);
2787 for (i = 0; i < num_meta_group_infos; i++)
2788 kfree(sbi->s_group_info[i]);
2789 kfree(sbi->s_group_info);
2791 kfree(sbi->s_mb_offsets);
2792 kfree(sbi->s_mb_maxs);
2793 if (sbi->s_buddy_cache)
2794 iput(sbi->s_buddy_cache);
2795 if (sbi->s_mb_stats) {
2797 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2798 atomic_read(&sbi->s_bal_allocated),
2799 atomic_read(&sbi->s_bal_reqs),
2800 atomic_read(&sbi->s_bal_success));
2802 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2803 "%u 2^N hits, %u breaks, %u lost\n",
2804 atomic_read(&sbi->s_bal_ex_scanned),
2805 atomic_read(&sbi->s_bal_goals),
2806 atomic_read(&sbi->s_bal_2orders),
2807 atomic_read(&sbi->s_bal_breaks),
2808 atomic_read(&sbi->s_mb_lost_chunks));
2810 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2811 sbi->s_mb_buddies_generated++,
2812 sbi->s_mb_generation_time);
2814 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2815 atomic_read(&sbi->s_mb_preallocated),
2816 atomic_read(&sbi->s_mb_discarded));
2819 free_percpu(sbi->s_locality_groups);
2820 ext4_mb_history_release(sb);
2826 * This function is called by the jbd2 layer once the commit has finished,
2827 * so we know we can free the blocks that were released with that commit.
2829 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2831 struct super_block *sb = journal->j_private;
2832 struct ext4_buddy e4b;
2833 struct ext4_group_info *db;
2834 int err, count = 0, count2 = 0;
2835 struct ext4_free_data *entry;
2836 ext4_fsblk_t discard_block;
2837 struct list_head *l, *ltmp;
2839 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2840 entry = list_entry(l, struct ext4_free_data, list);
2842 mb_debug("gonna free %u blocks in group %u (0x%p):",
2843 entry->count, entry->group, entry);
2845 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2846 /* we expect to find existing buddy because it's pinned */
2850 /* there are blocks to put in buddy to make them really free */
2851 count += entry->count;
2853 ext4_lock_group(sb, entry->group);
2854 /* Take it out of per group rb tree */
2855 rb_erase(&entry->node, &(db->bb_free_root));
2856 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2858 if (!db->bb_free_root.rb_node) {
2859 /* No more items in the per group rb tree
2860 * balance refcounts from ext4_mb_free_metadata()
2862 page_cache_release(e4b.bd_buddy_page);
2863 page_cache_release(e4b.bd_bitmap_page);
2865 ext4_unlock_group(sb, entry->group);
2866 discard_block = (ext4_fsblk_t) entry->group * EXT4_BLOCKS_PER_GROUP(sb)
2868 + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
2869 trace_ext4_discard_blocks(sb, (unsigned long long)discard_block,
2871 sb_issue_discard(sb, discard_block, entry->count);
2873 kmem_cache_free(ext4_free_ext_cachep, entry);
2874 ext4_mb_release_desc(&e4b);
2877 mb_debug("freed %u blocks in %u structures\n", count, count2);
2880 int __init init_ext4_mballoc(void)
2882 ext4_pspace_cachep =
2883 kmem_cache_create("ext4_prealloc_space",
2884 sizeof(struct ext4_prealloc_space),
2885 0, SLAB_RECLAIM_ACCOUNT, NULL);
2886 if (ext4_pspace_cachep == NULL)
2890 kmem_cache_create("ext4_alloc_context",
2891 sizeof(struct ext4_allocation_context),
2892 0, SLAB_RECLAIM_ACCOUNT, NULL);
2893 if (ext4_ac_cachep == NULL) {
2894 kmem_cache_destroy(ext4_pspace_cachep);
2898 ext4_free_ext_cachep =
2899 kmem_cache_create("ext4_free_block_extents",
2900 sizeof(struct ext4_free_data),
2901 0, SLAB_RECLAIM_ACCOUNT, NULL);
2902 if (ext4_free_ext_cachep == NULL) {
2903 kmem_cache_destroy(ext4_pspace_cachep);
2904 kmem_cache_destroy(ext4_ac_cachep);
2910 void exit_ext4_mballoc(void)
2912 /* XXX: synchronize_rcu(); */
2913 kmem_cache_destroy(ext4_pspace_cachep);
2914 kmem_cache_destroy(ext4_ac_cachep);
2915 kmem_cache_destroy(ext4_free_ext_cachep);
2920 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2921 * Returns 0 if success or error code
2923 static noinline_for_stack int
2924 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2925 handle_t *handle, unsigned int reserv_blks)
2927 struct buffer_head *bitmap_bh = NULL;
2928 struct ext4_super_block *es;
2929 struct ext4_group_desc *gdp;
2930 struct buffer_head *gdp_bh;
2931 struct ext4_sb_info *sbi;
2932 struct super_block *sb;
2936 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2937 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2945 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2949 err = ext4_journal_get_write_access(handle, bitmap_bh);
2954 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2958 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2959 ext4_free_blks_count(sb, gdp));
2961 err = ext4_journal_get_write_access(handle, gdp_bh);
2965 block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
2966 + ac->ac_b_ex.fe_start
2967 + le32_to_cpu(es->s_first_data_block);
2969 len = ac->ac_b_ex.fe_len;
2970 if (!ext4_data_block_valid(sbi, block, len)) {
2971 ext4_error(sb, __func__,
2972 "Allocating blocks %llu-%llu which overlap "
2973 "fs metadata\n", block, block+len);
2974 /* File system mounted not to panic on error
2975 * Fix the bitmap and repeat the block allocation
2976 * We leak some of the blocks here.
2978 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2979 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2980 ac->ac_b_ex.fe_len);
2981 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2982 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2988 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2989 #ifdef AGGRESSIVE_CHECK
2992 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2993 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2994 bitmap_bh->b_data));
2998 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
2999 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
3000 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3001 ext4_free_blks_set(sb, gdp,
3002 ext4_free_blocks_after_init(sb,
3003 ac->ac_b_ex.fe_group, gdp));
3005 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
3006 ext4_free_blks_set(sb, gdp, len);
3007 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
3009 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3010 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
3012 * Now reduce the dirty block count also. Should not go negative
3014 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3015 /* release all the reserved blocks if non delalloc */
3016 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
3018 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
3019 ac->ac_b_ex.fe_len);
3020 /* convert reserved quota blocks to real quota blocks */
3021 vfs_dq_claim_block(ac->ac_inode, ac->ac_b_ex.fe_len);
3024 if (sbi->s_log_groups_per_flex) {
3025 ext4_group_t flex_group = ext4_flex_group(sbi,
3026 ac->ac_b_ex.fe_group);
3027 atomic_sub(ac->ac_b_ex.fe_len,
3028 &sbi->s_flex_groups[flex_group].free_blocks);
3031 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3034 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3043 * here we normalize request for locality group
3044 * Group request are normalized to s_strip size if we set the same via mount
3045 * option. If not we set it to s_mb_group_prealloc which can be configured via
3046 * /sys/fs/ext4/<partition>/mb_group_prealloc
3048 * XXX: should we try to preallocate more than the group has now?
3050 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3052 struct super_block *sb = ac->ac_sb;
3053 struct ext4_locality_group *lg = ac->ac_lg;
3056 if (EXT4_SB(sb)->s_stripe)
3057 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
3059 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3060 mb_debug("#%u: goal %u blocks for locality group\n",
3061 current->pid, ac->ac_g_ex.fe_len);
3065 * Normalization means making request better in terms of
3066 * size and alignment
3068 static noinline_for_stack void
3069 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3070 struct ext4_allocation_request *ar)
3074 loff_t size, orig_size, start_off;
3075 ext4_lblk_t start, orig_start;
3076 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3077 struct ext4_prealloc_space *pa;
3079 /* do normalize only data requests, metadata requests
3080 do not need preallocation */
3081 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3084 /* sometime caller may want exact blocks */
3085 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3088 /* caller may indicate that preallocation isn't
3089 * required (it's a tail, for example) */
3090 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3093 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3094 ext4_mb_normalize_group_request(ac);
3098 bsbits = ac->ac_sb->s_blocksize_bits;
3100 /* first, let's learn actual file size
3101 * given current request is allocated */
3102 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3103 size = size << bsbits;
3104 if (size < i_size_read(ac->ac_inode))
3105 size = i_size_read(ac->ac_inode);
3107 /* max size of free chunks */
3110 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3111 (req <= (size) || max <= (chunk_size))
3113 /* first, try to predict filesize */
3114 /* XXX: should this table be tunable? */
3116 if (size <= 16 * 1024) {
3118 } else if (size <= 32 * 1024) {
3120 } else if (size <= 64 * 1024) {
3122 } else if (size <= 128 * 1024) {
3124 } else if (size <= 256 * 1024) {
3126 } else if (size <= 512 * 1024) {
3128 } else if (size <= 1024 * 1024) {
3130 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3131 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3132 (21 - bsbits)) << 21;
3133 size = 2 * 1024 * 1024;
3134 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3135 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3136 (22 - bsbits)) << 22;
3137 size = 4 * 1024 * 1024;
3138 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3139 (8<<20)>>bsbits, max, 8 * 1024)) {
3140 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3141 (23 - bsbits)) << 23;
3142 size = 8 * 1024 * 1024;
3144 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3145 size = ac->ac_o_ex.fe_len << bsbits;
3147 orig_size = size = size >> bsbits;
3148 orig_start = start = start_off >> bsbits;
3150 /* don't cover already allocated blocks in selected range */
3151 if (ar->pleft && start <= ar->lleft) {
3152 size -= ar->lleft + 1 - start;
3153 start = ar->lleft + 1;
3155 if (ar->pright && start + size - 1 >= ar->lright)
3156 size -= start + size - ar->lright;
3160 /* check we don't cross already preallocated blocks */
3162 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3167 spin_lock(&pa->pa_lock);
3168 if (pa->pa_deleted) {
3169 spin_unlock(&pa->pa_lock);
3173 pa_end = pa->pa_lstart + pa->pa_len;
3175 /* PA must not overlap original request */
3176 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3177 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3179 /* skip PA normalized request doesn't overlap with */
3180 if (pa->pa_lstart >= end) {
3181 spin_unlock(&pa->pa_lock);
3184 if (pa_end <= start) {
3185 spin_unlock(&pa->pa_lock);
3188 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3190 if (pa_end <= ac->ac_o_ex.fe_logical) {
3191 BUG_ON(pa_end < start);
3195 if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3196 BUG_ON(pa->pa_lstart > end);
3197 end = pa->pa_lstart;
3199 spin_unlock(&pa->pa_lock);
3204 /* XXX: extra loop to check we really don't overlap preallocations */
3206 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3208 spin_lock(&pa->pa_lock);
3209 if (pa->pa_deleted == 0) {
3210 pa_end = pa->pa_lstart + pa->pa_len;
3211 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3213 spin_unlock(&pa->pa_lock);
3217 if (start + size <= ac->ac_o_ex.fe_logical &&
3218 start > ac->ac_o_ex.fe_logical) {
3219 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3220 (unsigned long) start, (unsigned long) size,
3221 (unsigned long) ac->ac_o_ex.fe_logical);
3223 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3224 start > ac->ac_o_ex.fe_logical);
3225 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3227 /* now prepare goal request */
3229 /* XXX: is it better to align blocks WRT to logical
3230 * placement or satisfy big request as is */
3231 ac->ac_g_ex.fe_logical = start;
3232 ac->ac_g_ex.fe_len = size;
3234 /* define goal start in order to merge */
3235 if (ar->pright && (ar->lright == (start + size))) {
3236 /* merge to the right */
3237 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3238 &ac->ac_f_ex.fe_group,
3239 &ac->ac_f_ex.fe_start);
3240 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3242 if (ar->pleft && (ar->lleft + 1 == start)) {
3243 /* merge to the left */
3244 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3245 &ac->ac_f_ex.fe_group,
3246 &ac->ac_f_ex.fe_start);
3247 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3250 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
3251 (unsigned) orig_size, (unsigned) start);
3254 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3256 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3258 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3259 atomic_inc(&sbi->s_bal_reqs);
3260 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3261 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3262 atomic_inc(&sbi->s_bal_success);
3263 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3264 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3265 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3266 atomic_inc(&sbi->s_bal_goals);
3267 if (ac->ac_found > sbi->s_mb_max_to_scan)
3268 atomic_inc(&sbi->s_bal_breaks);
3271 ext4_mb_store_history(ac);
3275 * use blocks preallocated to inode
3277 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3278 struct ext4_prealloc_space *pa)
3284 /* found preallocated blocks, use them */
3285 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3286 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3288 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3289 &ac->ac_b_ex.fe_start);
3290 ac->ac_b_ex.fe_len = len;
3291 ac->ac_status = AC_STATUS_FOUND;
3294 BUG_ON(start < pa->pa_pstart);
3295 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3296 BUG_ON(pa->pa_free < len);
3299 mb_debug("use %llu/%u from inode pa %p\n", start, len, pa);
3303 * use blocks preallocated to locality group
3305 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3306 struct ext4_prealloc_space *pa)
3308 unsigned int len = ac->ac_o_ex.fe_len;
3310 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3311 &ac->ac_b_ex.fe_group,
3312 &ac->ac_b_ex.fe_start);
3313 ac->ac_b_ex.fe_len = len;
3314 ac->ac_status = AC_STATUS_FOUND;
3317 /* we don't correct pa_pstart or pa_plen here to avoid
3318 * possible race when the group is being loaded concurrently
3319 * instead we correct pa later, after blocks are marked
3320 * in on-disk bitmap -- see ext4_mb_release_context()
3321 * Other CPUs are prevented from allocating from this pa by lg_mutex
3323 mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3327 * Return the prealloc space that have minimal distance
3328 * from the goal block. @cpa is the prealloc
3329 * space that is having currently known minimal distance
3330 * from the goal block.
3332 static struct ext4_prealloc_space *
3333 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3334 struct ext4_prealloc_space *pa,
3335 struct ext4_prealloc_space *cpa)
3337 ext4_fsblk_t cur_distance, new_distance;
3340 atomic_inc(&pa->pa_count);
3343 cur_distance = abs(goal_block - cpa->pa_pstart);
3344 new_distance = abs(goal_block - pa->pa_pstart);
3346 if (cur_distance < new_distance)
3349 /* drop the previous reference */
3350 atomic_dec(&cpa->pa_count);
3351 atomic_inc(&pa->pa_count);
3356 * search goal blocks in preallocated space
3358 static noinline_for_stack int
3359 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3362 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3363 struct ext4_locality_group *lg;
3364 struct ext4_prealloc_space *pa, *cpa = NULL;
3365 ext4_fsblk_t goal_block;
3367 /* only data can be preallocated */
3368 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3371 /* first, try per-file preallocation */
3373 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3375 /* all fields in this condition don't change,
3376 * so we can skip locking for them */
3377 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3378 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3381 /* found preallocated blocks, use them */
3382 spin_lock(&pa->pa_lock);
3383 if (pa->pa_deleted == 0 && pa->pa_free) {
3384 atomic_inc(&pa->pa_count);
3385 ext4_mb_use_inode_pa(ac, pa);
3386 spin_unlock(&pa->pa_lock);
3387 ac->ac_criteria = 10;
3391 spin_unlock(&pa->pa_lock);
3395 /* can we use group allocation? */
3396 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3399 /* inode may have no locality group for some reason */
3403 order = fls(ac->ac_o_ex.fe_len) - 1;
3404 if (order > PREALLOC_TB_SIZE - 1)
3405 /* The max size of hash table is PREALLOC_TB_SIZE */
3406 order = PREALLOC_TB_SIZE - 1;
3408 goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) +
3409 ac->ac_g_ex.fe_start +
3410 le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block);
3412 * search for the prealloc space that is having
3413 * minimal distance from the goal block.
3415 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3417 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3419 spin_lock(&pa->pa_lock);
3420 if (pa->pa_deleted == 0 &&
3421 pa->pa_free >= ac->ac_o_ex.fe_len) {
3423 cpa = ext4_mb_check_group_pa(goal_block,
3426 spin_unlock(&pa->pa_lock);
3431 ext4_mb_use_group_pa(ac, cpa);
3432 ac->ac_criteria = 20;
3439 * the function goes through all block freed in the group
3440 * but not yet committed and marks them used in in-core bitmap.
3441 * buddy must be generated from this bitmap
3442 * Need to be called with the ext4 group lock held
3444 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3448 struct ext4_group_info *grp;
3449 struct ext4_free_data *entry;
3451 grp = ext4_get_group_info(sb, group);
3452 n = rb_first(&(grp->bb_free_root));
3455 entry = rb_entry(n, struct ext4_free_data, node);
3456 mb_set_bits(bitmap, entry->start_blk, entry->count);
3463 * the function goes through all preallocation in this group and marks them
3464 * used in in-core bitmap. buddy must be generated from this bitmap
3465 * Need to be called with ext4 group lock held
3467 static noinline_for_stack
3468 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3471 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3472 struct ext4_prealloc_space *pa;
3473 struct list_head *cur;
3474 ext4_group_t groupnr;
3475 ext4_grpblk_t start;
3476 int preallocated = 0;
3480 /* all form of preallocation discards first load group,
3481 * so the only competing code is preallocation use.
3482 * we don't need any locking here
3483 * notice we do NOT ignore preallocations with pa_deleted
3484 * otherwise we could leave used blocks available for
3485 * allocation in buddy when concurrent ext4_mb_put_pa()
3486 * is dropping preallocation
3488 list_for_each(cur, &grp->bb_prealloc_list) {
3489 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3490 spin_lock(&pa->pa_lock);
3491 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3494 spin_unlock(&pa->pa_lock);
3495 if (unlikely(len == 0))
3497 BUG_ON(groupnr != group);
3498 mb_set_bits(bitmap, start, len);
3499 preallocated += len;
3502 mb_debug("prellocated %u for group %u\n", preallocated, group);
3505 static void ext4_mb_pa_callback(struct rcu_head *head)
3507 struct ext4_prealloc_space *pa;
3508 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3509 kmem_cache_free(ext4_pspace_cachep, pa);
3513 * drops a reference to preallocated space descriptor
3514 * if this was the last reference and the space is consumed
3516 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3517 struct super_block *sb, struct ext4_prealloc_space *pa)
3520 ext4_fsblk_t grp_blk;
3522 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3525 /* in this short window concurrent discard can set pa_deleted */
3526 spin_lock(&pa->pa_lock);
3527 if (pa->pa_deleted == 1) {
3528 spin_unlock(&pa->pa_lock);
3533 spin_unlock(&pa->pa_lock);
3535 grp_blk = pa->pa_pstart;
3537 * If doing group-based preallocation, pa_pstart may be in the
3538 * next group when pa is used up
3540 if (pa->pa_type == MB_GROUP_PA)
3543 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3548 * P1 (buddy init) P2 (regular allocation)
3549 * find block B in PA
3550 * copy on-disk bitmap to buddy
3551 * mark B in on-disk bitmap
3552 * drop PA from group
3553 * mark all PAs in buddy
3555 * thus, P1 initializes buddy with B available. to prevent this
3556 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3559 ext4_lock_group(sb, grp);
3560 list_del(&pa->pa_group_list);
3561 ext4_unlock_group(sb, grp);
3563 spin_lock(pa->pa_obj_lock);
3564 list_del_rcu(&pa->pa_inode_list);
3565 spin_unlock(pa->pa_obj_lock);
3567 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3571 * creates new preallocated space for given inode
3573 static noinline_for_stack int
3574 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3576 struct super_block *sb = ac->ac_sb;
3577 struct ext4_prealloc_space *pa;
3578 struct ext4_group_info *grp;
3579 struct ext4_inode_info *ei;
3581 /* preallocate only when found space is larger then requested */
3582 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3583 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3584 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3586 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3590 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3596 /* we can't allocate as much as normalizer wants.
3597 * so, found space must get proper lstart
3598 * to cover original request */
3599 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3600 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3602 /* we're limited by original request in that
3603 * logical block must be covered any way
3604 * winl is window we can move our chunk within */
3605 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3607 /* also, we should cover whole original request */
3608 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3610 /* the smallest one defines real window */
3611 win = min(winl, wins);
3613 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3614 if (offs && offs < win)
3617 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3618 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3619 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3622 /* preallocation can change ac_b_ex, thus we store actually
3623 * allocated blocks for history */
3624 ac->ac_f_ex = ac->ac_b_ex;
3626 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3627 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3628 pa->pa_len = ac->ac_b_ex.fe_len;
3629 pa->pa_free = pa->pa_len;
3630 atomic_set(&pa->pa_count, 1);
3631 spin_lock_init(&pa->pa_lock);
3632 INIT_LIST_HEAD(&pa->pa_inode_list);
3633 INIT_LIST_HEAD(&pa->pa_group_list);
3635 pa->pa_type = MB_INODE_PA;
3637 mb_debug("new inode pa %p: %llu/%u for %u\n", pa,
3638 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3639 trace_ext4_mb_new_inode_pa(ac, pa);
3641 ext4_mb_use_inode_pa(ac, pa);
3642 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3644 ei = EXT4_I(ac->ac_inode);
3645 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3647 pa->pa_obj_lock = &ei->i_prealloc_lock;
3648 pa->pa_inode = ac->ac_inode;
3650 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3651 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3652 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3654 spin_lock(pa->pa_obj_lock);
3655 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3656 spin_unlock(pa->pa_obj_lock);
3662 * creates new preallocated space for locality group inodes belongs to
3664 static noinline_for_stack int
3665 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3667 struct super_block *sb = ac->ac_sb;
3668 struct ext4_locality_group *lg;
3669 struct ext4_prealloc_space *pa;
3670 struct ext4_group_info *grp;
3672 /* preallocate only when found space is larger then requested */
3673 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3674 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3675 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3677 BUG_ON(ext4_pspace_cachep == NULL);
3678 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3682 /* preallocation can change ac_b_ex, thus we store actually
3683 * allocated blocks for history */
3684 ac->ac_f_ex = ac->ac_b_ex;
3686 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3687 pa->pa_lstart = pa->pa_pstart;
3688 pa->pa_len = ac->ac_b_ex.fe_len;
3689 pa->pa_free = pa->pa_len;
3690 atomic_set(&pa->pa_count, 1);
3691 spin_lock_init(&pa->pa_lock);
3692 INIT_LIST_HEAD(&pa->pa_inode_list);
3693 INIT_LIST_HEAD(&pa->pa_group_list);
3695 pa->pa_type = MB_GROUP_PA;
3697 mb_debug("new group pa %p: %llu/%u for %u\n", pa,
3698 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3699 trace_ext4_mb_new_group_pa(ac, pa);
3701 ext4_mb_use_group_pa(ac, pa);
3702 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3704 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3708 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3709 pa->pa_inode = NULL;
3711 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3712 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3713 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3716 * We will later add the new pa to the right bucket
3717 * after updating the pa_free in ext4_mb_release_context
3722 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3726 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3727 err = ext4_mb_new_group_pa(ac);
3729 err = ext4_mb_new_inode_pa(ac);
3734 * finds all unused blocks in on-disk bitmap, frees them in
3735 * in-core bitmap and buddy.
3736 * @pa must be unlinked from inode and group lists, so that
3737 * nobody else can find/use it.
3738 * the caller MUST hold group/inode locks.
3739 * TODO: optimize the case when there are no in-core structures yet
3741 static noinline_for_stack int
3742 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3743 struct ext4_prealloc_space *pa,
3744 struct ext4_allocation_context *ac)
3746 struct super_block *sb = e4b->bd_sb;
3747 struct ext4_sb_info *sbi = EXT4_SB(sb);
3752 unsigned long long grp_blk_start;
3757 BUG_ON(pa->pa_deleted == 0);
3758 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3759 grp_blk_start = pa->pa_pstart - bit;
3760 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3761 end = bit + pa->pa_len;
3765 ac->ac_inode = pa->pa_inode;
3766 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3770 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3773 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3774 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3775 le32_to_cpu(sbi->s_es->s_first_data_block);
3776 mb_debug(" free preallocated %u/%u in group %u\n",
3777 (unsigned) start, (unsigned) next - bit,
3782 ac->ac_b_ex.fe_group = group;
3783 ac->ac_b_ex.fe_start = bit;
3784 ac->ac_b_ex.fe_len = next - bit;
3785 ac->ac_b_ex.fe_logical = 0;
3786 ext4_mb_store_history(ac);
3789 trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit,
3791 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3794 if (free != pa->pa_free) {
3795 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3796 pa, (unsigned long) pa->pa_lstart,
3797 (unsigned long) pa->pa_pstart,
3798 (unsigned long) pa->pa_len);
3799 ext4_grp_locked_error(sb, group,
3800 __func__, "free %u, pa_free %u",
3803 * pa is already deleted so we use the value obtained
3804 * from the bitmap and continue.
3807 atomic_add(free, &sbi->s_mb_discarded);
3812 static noinline_for_stack int
3813 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3814 struct ext4_prealloc_space *pa,
3815 struct ext4_allocation_context *ac)
3817 struct super_block *sb = e4b->bd_sb;
3822 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3824 trace_ext4_mb_release_group_pa(ac, pa);
3825 BUG_ON(pa->pa_deleted == 0);
3826 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3827 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3828 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3829 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3833 ac->ac_inode = NULL;
3834 ac->ac_b_ex.fe_group = group;
3835 ac->ac_b_ex.fe_start = bit;
3836 ac->ac_b_ex.fe_len = pa->pa_len;
3837 ac->ac_b_ex.fe_logical = 0;
3838 ext4_mb_store_history(ac);
3845 * releases all preallocations in given group
3847 * first, we need to decide discard policy:
3848 * - when do we discard
3850 * - how many do we discard
3851 * 1) how many requested
3853 static noinline_for_stack int
3854 ext4_mb_discard_group_preallocations(struct super_block *sb,
3855 ext4_group_t group, int needed)
3857 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3858 struct buffer_head *bitmap_bh = NULL;
3859 struct ext4_prealloc_space *pa, *tmp;
3860 struct ext4_allocation_context *ac;
3861 struct list_head list;
3862 struct ext4_buddy e4b;
3867 mb_debug("discard preallocation for group %u\n", group);
3869 if (list_empty(&grp->bb_prealloc_list))
3872 bitmap_bh = ext4_read_block_bitmap(sb, group);
3873 if (bitmap_bh == NULL) {
3874 ext4_error(sb, __func__, "Error in reading block "
3875 "bitmap for %u", group);
3879 err = ext4_mb_load_buddy(sb, group, &e4b);
3881 ext4_error(sb, __func__, "Error in loading buddy "
3882 "information for %u", group);
3888 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3890 INIT_LIST_HEAD(&list);
3891 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3895 ext4_lock_group(sb, group);
3896 list_for_each_entry_safe(pa, tmp,
3897 &grp->bb_prealloc_list, pa_group_list) {
3898 spin_lock(&pa->pa_lock);
3899 if (atomic_read(&pa->pa_count)) {
3900 spin_unlock(&pa->pa_lock);
3904 if (pa->pa_deleted) {
3905 spin_unlock(&pa->pa_lock);
3909 /* seems this one can be freed ... */
3912 /* we can trust pa_free ... */
3913 free += pa->pa_free;
3915 spin_unlock(&pa->pa_lock);
3917 list_del(&pa->pa_group_list);
3918 list_add(&pa->u.pa_tmp_list, &list);
3921 /* if we still need more blocks and some PAs were used, try again */
3922 if (free < needed && busy) {
3924 ext4_unlock_group(sb, group);
3926 * Yield the CPU here so that we don't get soft lockup
3927 * in non preempt case.
3933 /* found anything to free? */
3934 if (list_empty(&list)) {
3939 /* now free all selected PAs */
3940 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3942 /* remove from object (inode or locality group) */
3943 spin_lock(pa->pa_obj_lock);
3944 list_del_rcu(&pa->pa_inode_list);
3945 spin_unlock(pa->pa_obj_lock);
3947 if (pa->pa_type == MB_GROUP_PA)
3948 ext4_mb_release_group_pa(&e4b, pa, ac);
3950 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3952 list_del(&pa->u.pa_tmp_list);
3953 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3957 ext4_unlock_group(sb, group);
3959 kmem_cache_free(ext4_ac_cachep, ac);
3960 ext4_mb_release_desc(&e4b);
3966 * releases all non-used preallocated blocks for given inode
3968 * It's important to discard preallocations under i_data_sem
3969 * We don't want another block to be served from the prealloc
3970 * space when we are discarding the inode prealloc space.
3972 * FIXME!! Make sure it is valid at all the call sites
3974 void ext4_discard_preallocations(struct inode *inode)
3976 struct ext4_inode_info *ei = EXT4_I(inode);
3977 struct super_block *sb = inode->i_sb;
3978 struct buffer_head *bitmap_bh = NULL;
3979 struct ext4_prealloc_space *pa, *tmp;
3980 struct ext4_allocation_context *ac;
3981 ext4_group_t group = 0;
3982 struct list_head list;
3983 struct ext4_buddy e4b;
3986 if (!S_ISREG(inode->i_mode)) {
3987 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3991 mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
3992 trace_ext4_discard_preallocations(inode);
3994 INIT_LIST_HEAD(&list);
3996 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3999 ac->ac_inode = inode;
4002 /* first, collect all pa's in the inode */
4003 spin_lock(&ei->i_prealloc_lock);
4004 while (!list_empty(&ei->i_prealloc_list)) {
4005 pa = list_entry(ei->i_prealloc_list.next,
4006 struct ext4_prealloc_space, pa_inode_list);
4007 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4008 spin_lock(&pa->pa_lock);
4009 if (atomic_read(&pa->pa_count)) {
4010 /* this shouldn't happen often - nobody should
4011 * use preallocation while we're discarding it */
4012 spin_unlock(&pa->pa_lock);
4013 spin_unlock(&ei->i_prealloc_lock);
4014 printk(KERN_ERR "uh-oh! used pa while discarding\n");
4016 schedule_timeout_uninterruptible(HZ);
4020 if (pa->pa_deleted == 0) {
4022 spin_unlock(&pa->pa_lock);
4023 list_del_rcu(&pa->pa_inode_list);
4024 list_add(&pa->u.pa_tmp_list, &list);
4028 /* someone is deleting pa right now */
4029 spin_unlock(&pa->pa_lock);
4030 spin_unlock(&ei->i_prealloc_lock);
4032 /* we have to wait here because pa_deleted
4033 * doesn't mean pa is already unlinked from
4034 * the list. as we might be called from
4035 * ->clear_inode() the inode will get freed
4036 * and concurrent thread which is unlinking
4037 * pa from inode's list may access already
4038 * freed memory, bad-bad-bad */
4040 /* XXX: if this happens too often, we can
4041 * add a flag to force wait only in case
4042 * of ->clear_inode(), but not in case of
4043 * regular truncate */
4044 schedule_timeout_uninterruptible(HZ);
4047 spin_unlock(&ei->i_prealloc_lock);
4049 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4050 BUG_ON(pa->pa_type != MB_INODE_PA);
4051 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4053 err = ext4_mb_load_buddy(sb, group, &e4b);
4055 ext4_error(sb, __func__, "Error in loading buddy "
4056 "information for %u", group);
4060 bitmap_bh = ext4_read_block_bitmap(sb, group);
4061 if (bitmap_bh == NULL) {
4062 ext4_error(sb, __func__, "Error in reading block "
4063 "bitmap for %u", group);
4064 ext4_mb_release_desc(&e4b);
4068 ext4_lock_group(sb, group);
4069 list_del(&pa->pa_group_list);
4070 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
4071 ext4_unlock_group(sb, group);
4073 ext4_mb_release_desc(&e4b);
4076 list_del(&pa->u.pa_tmp_list);
4077 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4080 kmem_cache_free(ext4_ac_cachep, ac);
4084 * finds all preallocated spaces and return blocks being freed to them
4085 * if preallocated space becomes full (no block is used from the space)
4086 * then the function frees space in buddy
4087 * XXX: at the moment, truncate (which is the only way to free blocks)
4088 * discards all preallocations
4090 static void ext4_mb_return_to_preallocation(struct inode *inode,
4091 struct ext4_buddy *e4b,
4092 sector_t block, int count)
4094 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
4097 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4099 struct super_block *sb = ac->ac_sb;
4100 ext4_group_t ngroups, i;
4102 printk(KERN_ERR "EXT4-fs: Can't allocate:"
4103 " Allocation context details:\n");
4104 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
4105 ac->ac_status, ac->ac_flags);
4106 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4107 "best %lu/%lu/%lu@%lu cr %d\n",
4108 (unsigned long)ac->ac_o_ex.fe_group,
4109 (unsigned long)ac->ac_o_ex.fe_start,
4110 (unsigned long)ac->ac_o_ex.fe_len,
4111 (unsigned long)ac->ac_o_ex.fe_logical,
4112 (unsigned long)ac->ac_g_ex.fe_group,
4113 (unsigned long)ac->ac_g_ex.fe_start,
4114 (unsigned long)ac->ac_g_ex.fe_len,
4115 (unsigned long)ac->ac_g_ex.fe_logical,
4116 (unsigned long)ac->ac_b_ex.fe_group,
4117 (unsigned long)ac->ac_b_ex.fe_start,
4118 (unsigned long)ac->ac_b_ex.fe_len,
4119 (unsigned long)ac->ac_b_ex.fe_logical,
4120 (int)ac->ac_criteria);
4121 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
4123 printk(KERN_ERR "EXT4-fs: groups: \n");
4124 ngroups = ext4_get_groups_count(sb);
4125 for (i = 0; i < ngroups; i++) {
4126 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4127 struct ext4_prealloc_space *pa;
4128 ext4_grpblk_t start;
4129 struct list_head *cur;
4130 ext4_lock_group(sb, i);
4131 list_for_each(cur, &grp->bb_prealloc_list) {
4132 pa = list_entry(cur, struct ext4_prealloc_space,
4134 spin_lock(&pa->pa_lock);
4135 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4137 spin_unlock(&pa->pa_lock);
4138 printk(KERN_ERR "PA:%lu:%d:%u \n", i,
4141 ext4_unlock_group(sb, i);
4143 if (grp->bb_free == 0)
4145 printk(KERN_ERR "%lu: %d/%d \n",
4146 i, grp->bb_free, grp->bb_fragments);
4148 printk(KERN_ERR "\n");
4151 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4158 * We use locality group preallocation for small size file. The size of the
4159 * file is determined by the current size or the resulting size after
4160 * allocation which ever is larger
4162 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4164 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4166 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4167 int bsbits = ac->ac_sb->s_blocksize_bits;
4170 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4173 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4174 isize = i_size_read(ac->ac_inode) >> bsbits;
4175 size = max(size, isize);
4177 /* don't use group allocation for large files */
4178 if (size >= sbi->s_mb_stream_request)
4181 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4184 BUG_ON(ac->ac_lg != NULL);
4186 * locality group prealloc space are per cpu. The reason for having
4187 * per cpu locality group is to reduce the contention between block
4188 * request from multiple CPUs.
4190 ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id());
4192 /* we're going to use group allocation */
4193 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4195 /* serialize all allocations in the group */
4196 mutex_lock(&ac->ac_lg->lg_mutex);
4199 static noinline_for_stack int
4200 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4201 struct ext4_allocation_request *ar)
4203 struct super_block *sb = ar->inode->i_sb;
4204 struct ext4_sb_info *sbi = EXT4_SB(sb);
4205 struct ext4_super_block *es = sbi->s_es;
4209 ext4_grpblk_t block;
4211 /* we can't allocate > group size */
4214 /* just a dirty hack to filter too big requests */
4215 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4216 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4218 /* start searching from the goal */
4220 if (goal < le32_to_cpu(es->s_first_data_block) ||
4221 goal >= ext4_blocks_count(es))
4222 goal = le32_to_cpu(es->s_first_data_block);
4223 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4225 /* set up allocation goals */
4226 ac->ac_b_ex.fe_logical = ar->logical;
4227 ac->ac_b_ex.fe_group = 0;
4228 ac->ac_b_ex.fe_start = 0;
4229 ac->ac_b_ex.fe_len = 0;
4230 ac->ac_status = AC_STATUS_CONTINUE;
4231 ac->ac_groups_scanned = 0;
4232 ac->ac_ex_scanned = 0;
4235 ac->ac_inode = ar->inode;
4236 ac->ac_o_ex.fe_logical = ar->logical;
4237 ac->ac_o_ex.fe_group = group;
4238 ac->ac_o_ex.fe_start = block;
4239 ac->ac_o_ex.fe_len = len;
4240 ac->ac_g_ex.fe_logical = ar->logical;
4241 ac->ac_g_ex.fe_group = group;
4242 ac->ac_g_ex.fe_start = block;
4243 ac->ac_g_ex.fe_len = len;
4244 ac->ac_f_ex.fe_len = 0;
4245 ac->ac_flags = ar->flags;
4247 ac->ac_criteria = 0;
4249 ac->ac_bitmap_page = NULL;
4250 ac->ac_buddy_page = NULL;
4251 ac->alloc_semp = NULL;
4254 /* we have to define context: we'll we work with a file or
4255 * locality group. this is a policy, actually */
4256 ext4_mb_group_or_file(ac);
4258 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4259 "left: %u/%u, right %u/%u to %swritable\n",
4260 (unsigned) ar->len, (unsigned) ar->logical,
4261 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4262 (unsigned) ar->lleft, (unsigned) ar->pleft,
4263 (unsigned) ar->lright, (unsigned) ar->pright,
4264 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4269 static noinline_for_stack void
4270 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4271 struct ext4_locality_group *lg,
4272 int order, int total_entries)
4274 ext4_group_t group = 0;
4275 struct ext4_buddy e4b;
4276 struct list_head discard_list;
4277 struct ext4_prealloc_space *pa, *tmp;
4278 struct ext4_allocation_context *ac;
4280 mb_debug("discard locality group preallocation\n");
4282 INIT_LIST_HEAD(&discard_list);
4283 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4287 spin_lock(&lg->lg_prealloc_lock);
4288 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4290 spin_lock(&pa->pa_lock);
4291 if (atomic_read(&pa->pa_count)) {
4293 * This is the pa that we just used
4294 * for block allocation. So don't
4297 spin_unlock(&pa->pa_lock);
4300 if (pa->pa_deleted) {
4301 spin_unlock(&pa->pa_lock);
4304 /* only lg prealloc space */
4305 BUG_ON(pa->pa_type != MB_GROUP_PA);
4307 /* seems this one can be freed ... */
4309 spin_unlock(&pa->pa_lock);
4311 list_del_rcu(&pa->pa_inode_list);
4312 list_add(&pa->u.pa_tmp_list, &discard_list);
4315 if (total_entries <= 5) {
4317 * we want to keep only 5 entries
4318 * allowing it to grow to 8. This
4319 * mak sure we don't call discard
4320 * soon for this list.
4325 spin_unlock(&lg->lg_prealloc_lock);
4327 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4329 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4330 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4331 ext4_error(sb, __func__, "Error in loading buddy "
4332 "information for %u", group);
4335 ext4_lock_group(sb, group);
4336 list_del(&pa->pa_group_list);
4337 ext4_mb_release_group_pa(&e4b, pa, ac);
4338 ext4_unlock_group(sb, group);
4340 ext4_mb_release_desc(&e4b);
4341 list_del(&pa->u.pa_tmp_list);
4342 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4345 kmem_cache_free(ext4_ac_cachep, ac);
4349 * We have incremented pa_count. So it cannot be freed at this
4350 * point. Also we hold lg_mutex. So no parallel allocation is
4351 * possible from this lg. That means pa_free cannot be updated.
4353 * A parallel ext4_mb_discard_group_preallocations is possible.
4354 * which can cause the lg_prealloc_list to be updated.
4357 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4359 int order, added = 0, lg_prealloc_count = 1;
4360 struct super_block *sb = ac->ac_sb;
4361 struct ext4_locality_group *lg = ac->ac_lg;
4362 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4364 order = fls(pa->pa_free) - 1;
4365 if (order > PREALLOC_TB_SIZE - 1)
4366 /* The max size of hash table is PREALLOC_TB_SIZE */
4367 order = PREALLOC_TB_SIZE - 1;
4368 /* Add the prealloc space to lg */
4370 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4372 spin_lock(&tmp_pa->pa_lock);
4373 if (tmp_pa->pa_deleted) {
4374 spin_unlock(&tmp_pa->pa_lock);
4377 if (!added && pa->pa_free < tmp_pa->pa_free) {
4378 /* Add to the tail of the previous entry */
4379 list_add_tail_rcu(&pa->pa_inode_list,
4380 &tmp_pa->pa_inode_list);
4383 * we want to count the total
4384 * number of entries in the list
4387 spin_unlock(&tmp_pa->pa_lock);
4388 lg_prealloc_count++;
4391 list_add_tail_rcu(&pa->pa_inode_list,
4392 &lg->lg_prealloc_list[order]);
4395 /* Now trim the list to be not more than 8 elements */
4396 if (lg_prealloc_count > 8) {
4397 ext4_mb_discard_lg_preallocations(sb, lg,
4398 order, lg_prealloc_count);
4405 * release all resource we used in allocation
4407 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4409 struct ext4_prealloc_space *pa = ac->ac_pa;
4411 if (pa->pa_type == MB_GROUP_PA) {
4412 /* see comment in ext4_mb_use_group_pa() */
4413 spin_lock(&pa->pa_lock);
4414 pa->pa_pstart += ac->ac_b_ex.fe_len;
4415 pa->pa_lstart += ac->ac_b_ex.fe_len;
4416 pa->pa_free -= ac->ac_b_ex.fe_len;
4417 pa->pa_len -= ac->ac_b_ex.fe_len;
4418 spin_unlock(&pa->pa_lock);
4422 up_read(ac->alloc_semp);
4425 * We want to add the pa to the right bucket.
4426 * Remove it from the list and while adding
4427 * make sure the list to which we are adding
4428 * doesn't grow big. We need to release
4429 * alloc_semp before calling ext4_mb_add_n_trim()
4431 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4432 spin_lock(pa->pa_obj_lock);
4433 list_del_rcu(&pa->pa_inode_list);
4434 spin_unlock(pa->pa_obj_lock);
4435 ext4_mb_add_n_trim(ac);
4437 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4439 if (ac->ac_bitmap_page)
4440 page_cache_release(ac->ac_bitmap_page);
4441 if (ac->ac_buddy_page)
4442 page_cache_release(ac->ac_buddy_page);
4443 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4444 mutex_unlock(&ac->ac_lg->lg_mutex);
4445 ext4_mb_collect_stats(ac);
4449 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4451 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4455 trace_ext4_mb_discard_preallocations(sb, needed);
4456 for (i = 0; i < ngroups && needed > 0; i++) {
4457 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4466 * Main entry point into mballoc to allocate blocks
4467 * it tries to use preallocation first, then falls back
4468 * to usual allocation
4470 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4471 struct ext4_allocation_request *ar, int *errp)
4474 struct ext4_allocation_context *ac = NULL;
4475 struct ext4_sb_info *sbi;
4476 struct super_block *sb;
4477 ext4_fsblk_t block = 0;
4478 unsigned int inquota = 0;
4479 unsigned int reserv_blks = 0;
4481 sb = ar->inode->i_sb;
4484 trace_ext4_request_blocks(ar);
4487 * For delayed allocation, we could skip the ENOSPC and
4488 * EDQUOT check, as blocks and quotas have been already
4489 * reserved when data being copied into pagecache.
4491 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4492 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4494 /* Without delayed allocation we need to verify
4495 * there is enough free blocks to do block allocation
4496 * and verify allocation doesn't exceed the quota limits.
4498 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4499 /* let others to free the space */
4501 ar->len = ar->len >> 1;
4507 reserv_blks = ar->len;
4508 while (ar->len && vfs_dq_alloc_block(ar->inode, ar->len)) {
4509 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4519 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4522 ac->ac_inode = ar->inode;
4529 *errp = ext4_mb_initialize_context(ac, ar);
4535 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4536 if (!ext4_mb_use_preallocated(ac)) {
4537 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4538 ext4_mb_normalize_request(ac, ar);
4540 /* allocate space in core */
4541 ext4_mb_regular_allocator(ac);
4543 /* as we've just preallocated more space than
4544 * user requested orinally, we store allocated
4545 * space in a special descriptor */
4546 if (ac->ac_status == AC_STATUS_FOUND &&
4547 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4548 ext4_mb_new_preallocation(ac);
4550 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4551 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4552 if (*errp == -EAGAIN) {
4554 * drop the reference that we took
4555 * in ext4_mb_use_best_found
4557 ext4_mb_release_context(ac);
4558 ac->ac_b_ex.fe_group = 0;
4559 ac->ac_b_ex.fe_start = 0;
4560 ac->ac_b_ex.fe_len = 0;
4561 ac->ac_status = AC_STATUS_CONTINUE;
4564 ac->ac_b_ex.fe_len = 0;
4566 ext4_mb_show_ac(ac);
4568 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4569 ar->len = ac->ac_b_ex.fe_len;
4572 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4576 ac->ac_b_ex.fe_len = 0;
4578 ext4_mb_show_ac(ac);
4581 ext4_mb_release_context(ac);
4584 kmem_cache_free(ext4_ac_cachep, ac);
4586 if (inquota && ar->len < inquota)
4587 vfs_dq_free_block(ar->inode, inquota - ar->len);
4590 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4591 /* release all the reserved blocks if non delalloc */
4592 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4596 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4602 * We can merge two free data extents only if the physical blocks
4603 * are contiguous, AND the extents were freed by the same transaction,
4604 * AND the blocks are associated with the same group.
4606 static int can_merge(struct ext4_free_data *entry1,
4607 struct ext4_free_data *entry2)
4609 if ((entry1->t_tid == entry2->t_tid) &&
4610 (entry1->group == entry2->group) &&
4611 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4616 static noinline_for_stack int
4617 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4618 struct ext4_free_data *new_entry)
4620 ext4_grpblk_t block;
4621 struct ext4_free_data *entry;
4622 struct ext4_group_info *db = e4b->bd_info;
4623 struct super_block *sb = e4b->bd_sb;
4624 struct ext4_sb_info *sbi = EXT4_SB(sb);
4625 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4626 struct rb_node *parent = NULL, *new_node;
4628 BUG_ON(!ext4_handle_valid(handle));
4629 BUG_ON(e4b->bd_bitmap_page == NULL);
4630 BUG_ON(e4b->bd_buddy_page == NULL);
4632 new_node = &new_entry->node;
4633 block = new_entry->start_blk;
4636 /* first free block exent. We need to
4637 protect buddy cache from being freed,
4638 * otherwise we'll refresh it from
4639 * on-disk bitmap and lose not-yet-available
4641 page_cache_get(e4b->bd_buddy_page);
4642 page_cache_get(e4b->bd_bitmap_page);
4646 entry = rb_entry(parent, struct ext4_free_data, node);
4647 if (block < entry->start_blk)
4649 else if (block >= (entry->start_blk + entry->count))
4650 n = &(*n)->rb_right;
4652 ext4_grp_locked_error(sb, e4b->bd_group, __func__,
4653 "Double free of blocks %d (%d %d)",
4654 block, entry->start_blk, entry->count);
4659 rb_link_node(new_node, parent, n);
4660 rb_insert_color(new_node, &db->bb_free_root);
4662 /* Now try to see the extent can be merged to left and right */
4663 node = rb_prev(new_node);
4665 entry = rb_entry(node, struct ext4_free_data, node);
4666 if (can_merge(entry, new_entry)) {
4667 new_entry->start_blk = entry->start_blk;
4668 new_entry->count += entry->count;
4669 rb_erase(node, &(db->bb_free_root));
4670 spin_lock(&sbi->s_md_lock);
4671 list_del(&entry->list);
4672 spin_unlock(&sbi->s_md_lock);
4673 kmem_cache_free(ext4_free_ext_cachep, entry);
4677 node = rb_next(new_node);
4679 entry = rb_entry(node, struct ext4_free_data, node);
4680 if (can_merge(new_entry, entry)) {
4681 new_entry->count += entry->count;
4682 rb_erase(node, &(db->bb_free_root));
4683 spin_lock(&sbi->s_md_lock);
4684 list_del(&entry->list);
4685 spin_unlock(&sbi->s_md_lock);
4686 kmem_cache_free(ext4_free_ext_cachep, entry);
4689 /* Add the extent to transaction's private list */
4690 spin_lock(&sbi->s_md_lock);
4691 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4692 spin_unlock(&sbi->s_md_lock);
4697 * Main entry point into mballoc to free blocks
4699 void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
4700 ext4_fsblk_t block, unsigned long count,
4701 int metadata, unsigned long *freed)
4703 struct buffer_head *bitmap_bh = NULL;
4704 struct super_block *sb = inode->i_sb;
4705 struct ext4_allocation_context *ac = NULL;
4706 struct ext4_group_desc *gdp;
4707 struct ext4_super_block *es;
4708 unsigned int overflow;
4710 struct buffer_head *gd_bh;
4711 ext4_group_t block_group;
4712 struct ext4_sb_info *sbi;
4713 struct ext4_buddy e4b;
4720 es = EXT4_SB(sb)->s_es;
4721 if (block < le32_to_cpu(es->s_first_data_block) ||
4722 block + count < block ||
4723 block + count > ext4_blocks_count(es)) {
4724 ext4_error(sb, __func__,
4725 "Freeing blocks not in datazone - "
4726 "block = %llu, count = %lu", block, count);
4730 ext4_debug("freeing block %llu\n", block);
4731 trace_ext4_free_blocks(inode, block, count, metadata);
4733 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4735 ac->ac_op = EXT4_MB_HISTORY_FREE;
4736 ac->ac_inode = inode;
4742 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4745 * Check to see if we are freeing blocks across a group
4748 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4749 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4752 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4757 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4763 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4764 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4765 in_range(block, ext4_inode_table(sb, gdp),
4766 EXT4_SB(sb)->s_itb_per_group) ||
4767 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4768 EXT4_SB(sb)->s_itb_per_group)) {
4770 ext4_error(sb, __func__,
4771 "Freeing blocks in system zone - "
4772 "Block = %llu, count = %lu", block, count);
4773 /* err = 0. ext4_std_error should be a no op */
4777 BUFFER_TRACE(bitmap_bh, "getting write access");
4778 err = ext4_journal_get_write_access(handle, bitmap_bh);
4783 * We are about to modify some metadata. Call the journal APIs
4784 * to unshare ->b_data if a currently-committing transaction is
4787 BUFFER_TRACE(gd_bh, "get_write_access");
4788 err = ext4_journal_get_write_access(handle, gd_bh);
4791 #ifdef AGGRESSIVE_CHECK
4794 for (i = 0; i < count; i++)
4795 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4799 ac->ac_b_ex.fe_group = block_group;
4800 ac->ac_b_ex.fe_start = bit;
4801 ac->ac_b_ex.fe_len = count;
4802 ext4_mb_store_history(ac);
4805 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4808 if (metadata && ext4_handle_valid(handle)) {
4809 struct ext4_free_data *new_entry;
4811 * blocks being freed are metadata. these blocks shouldn't
4812 * be used until this transaction is committed
4814 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4815 new_entry->start_blk = bit;
4816 new_entry->group = block_group;
4817 new_entry->count = count;
4818 new_entry->t_tid = handle->h_transaction->t_tid;
4820 ext4_lock_group(sb, block_group);
4821 mb_clear_bits(bitmap_bh->b_data, bit, count);
4822 ext4_mb_free_metadata(handle, &e4b, new_entry);
4824 /* need to update group_info->bb_free and bitmap
4825 * with group lock held. generate_buddy look at
4826 * them with group lock_held
4828 ext4_lock_group(sb, block_group);
4829 mb_clear_bits(bitmap_bh->b_data, bit, count);
4830 mb_free_blocks(inode, &e4b, bit, count);
4831 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4834 ret = ext4_free_blks_count(sb, gdp) + count;
4835 ext4_free_blks_set(sb, gdp, ret);
4836 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4837 ext4_unlock_group(sb, block_group);
4838 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4840 if (sbi->s_log_groups_per_flex) {
4841 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4842 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4845 ext4_mb_release_desc(&e4b);
4849 /* We dirtied the bitmap block */
4850 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4851 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4853 /* And the group descriptor block */
4854 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4855 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4859 if (overflow && !err) {
4868 ext4_std_error(sb, err);
4870 kmem_cache_free(ext4_ac_cachep, ac);