2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
27 * - test ext4_ext_search_left() and ext4_ext_search_right()
28 * - search for metadata in few groups
31 * - normalization should take into account whether file is still open
32 * - discard preallocations if no free space left (policy?)
33 * - don't normalize tails
35 * - reservation for superuser
38 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
39 * - track min/max extents in each group for better group selection
40 * - mb_mark_used() may allocate chunk right after splitting buddy
41 * - tree of groups sorted by number of free blocks
46 * The allocation request involve request for multiple number of blocks
47 * near to the goal(block) value specified.
49 * During initialization phase of the allocator we decide to use the group
50 * preallocation or inode preallocation depending on the size file. The
51 * size of the file could be the resulting file size we would have after
52 * allocation or the current file size which ever is larger. If the size is
53 * less that sbi->s_mb_stream_request we select the group
54 * preallocation. The default value of s_mb_stream_request is 16
55 * blocks. This can also be tuned via
56 * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms
57 * of number of blocks.
59 * The main motivation for having small file use group preallocation is to
60 * ensure that we have small file closer in the disk.
62 * First stage the allocator looks at the inode prealloc list
63 * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for
64 * this particular inode. The inode prealloc space is represented as:
66 * pa_lstart -> the logical start block for this prealloc space
67 * pa_pstart -> the physical start block for this prealloc space
68 * pa_len -> lenght for this prealloc space
69 * pa_free -> free space available in this prealloc space
71 * The inode preallocation space is used looking at the _logical_ start
72 * block. If only the logical file block falls within the range of prealloc
73 * space we will consume the particular prealloc space. This make sure that
74 * that the we have contiguous physical blocks representing the file blocks
76 * The important thing to be noted in case of inode prealloc space is that
77 * we don't modify the values associated to inode prealloc space except
80 * If we are not able to find blocks in the inode prealloc space and if we
81 * have the group allocation flag set then we look at the locality group
82 * prealloc space. These are per CPU prealloc list repreasented as
84 * ext4_sb_info.s_locality_groups[smp_processor_id()]
86 * The reason for having a per cpu locality group is to reduce the contention
87 * between CPUs. It is possible to get scheduled at this point.
89 * The locality group prealloc space is used looking at whether we have
90 * enough free space (pa_free) withing the prealloc space.
92 * If we can't allocate blocks via inode prealloc or/and locality group
93 * prealloc then we look at the buddy cache. The buddy cache is represented
94 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
95 * mapped to the buddy and bitmap information regarding different
96 * groups. The buddy information is attached to buddy cache inode so that
97 * we can access them through the page cache. The information regarding
98 * each group is loaded via ext4_mb_load_buddy. The information involve
99 * block bitmap and buddy information. The information are stored in the
103 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
106 * one block each for bitmap and buddy information. So for each group we
107 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
108 * blocksize) blocks. So it can have information regarding groups_per_page
109 * which is blocks_per_page/2
111 * The buddy cache inode is not stored on disk. The inode is thrown
112 * away when the filesystem is unmounted.
114 * We look for count number of blocks in the buddy cache. If we were able
115 * to locate that many free blocks we return with additional information
116 * regarding rest of the contiguous physical block available
118 * Before allocating blocks via buddy cache we normalize the request
119 * blocks. This ensure we ask for more blocks that we needed. The extra
120 * blocks that we get after allocation is added to the respective prealloc
121 * list. In case of inode preallocation we follow a list of heuristics
122 * based on file size. This can be found in ext4_mb_normalize_request. If
123 * we are doing a group prealloc we try to normalize the request to
124 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is set to
125 * 512 blocks. This can be tuned via
126 * /proc/fs/ext4/<partition/group_prealloc. The value is represented in
127 * terms of number of blocks. If we have mounted the file system with -O
128 * stripe=<value> option the group prealloc request is normalized to the
129 * stripe value (sbi->s_stripe)
131 * The regular allocator(using the buddy cache) support few tunables.
133 * /proc/fs/ext4/<partition>/min_to_scan
134 * /proc/fs/ext4/<partition>/max_to_scan
135 * /proc/fs/ext4/<partition>/order2_req
137 * The regular allocator use buddy scan only if the request len is power of
138 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
139 * value of s_mb_order2_reqs can be tuned via
140 * /proc/fs/ext4/<partition>/order2_req. If the request len is equal to
141 * stripe size (sbi->s_stripe), we try to search for contigous block in
142 * stripe size. This should result in better allocation on RAID setup. If
143 * not we search in the specific group using bitmap for best extents. The
144 * tunable min_to_scan and max_to_scan controll the behaviour here.
145 * min_to_scan indicate how long the mballoc __must__ look for a best
146 * extent and max_to_scanindicate how long the mballoc __can__ look for a
147 * best extent in the found extents. Searching for the blocks starts with
148 * the group specified as the goal value in allocation context via
149 * ac_g_ex. Each group is first checked based on the criteria whether it
150 * can used for allocation. ext4_mb_good_group explains how the groups are
153 * Both the prealloc space are getting populated as above. So for the first
154 * request we will hit the buddy cache which will result in this prealloc
155 * space getting filled. The prealloc space is then later used for the
156 * subsequent request.
160 * mballoc operates on the following data:
162 * - in-core buddy (actually includes buddy and bitmap)
163 * - preallocation descriptors (PAs)
165 * there are two types of preallocations:
167 * assiged to specific inode and can be used for this inode only.
168 * it describes part of inode's space preallocated to specific
169 * physical blocks. any block from that preallocated can be used
170 * independent. the descriptor just tracks number of blocks left
171 * unused. so, before taking some block from descriptor, one must
172 * make sure corresponded logical block isn't allocated yet. this
173 * also means that freeing any block within descriptor's range
174 * must discard all preallocated blocks.
176 * assigned to specific locality group which does not translate to
177 * permanent set of inodes: inode can join and leave group. space
178 * from this type of preallocation can be used for any inode. thus
179 * it's consumed from the beginning to the end.
181 * relation between them can be expressed as:
182 * in-core buddy = on-disk bitmap + preallocation descriptors
184 * this mean blocks mballoc considers used are:
185 * - allocated blocks (persistent)
186 * - preallocated blocks (non-persistent)
188 * consistency in mballoc world means that at any time a block is either
189 * free or used in ALL structures. notice: "any time" should not be read
190 * literally -- time is discrete and delimited by locks.
192 * to keep it simple, we don't use block numbers, instead we count number of
193 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
195 * all operations can be expressed as:
196 * - init buddy: buddy = on-disk + PAs
197 * - new PA: buddy += N; PA = N
198 * - use inode PA: on-disk += N; PA -= N
199 * - discard inode PA buddy -= on-disk - PA; PA = 0
200 * - use locality group PA on-disk += N; PA -= N
201 * - discard locality group PA buddy -= PA; PA = 0
202 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
203 * is used in real operation because we can't know actual used
204 * bits from PA, only from on-disk bitmap
206 * if we follow this strict logic, then all operations above should be atomic.
207 * given some of them can block, we'd have to use something like semaphores
208 * killing performance on high-end SMP hardware. let's try to relax it using
209 * the following knowledge:
210 * 1) if buddy is referenced, it's already initialized
211 * 2) while block is used in buddy and the buddy is referenced,
212 * nobody can re-allocate that block
213 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
214 * bit set and PA claims same block, it's OK. IOW, one can set bit in
215 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
218 * so, now we're building a concurrency table:
221 * blocks for PA are allocated in the buddy, buddy must be referenced
222 * until PA is linked to allocation group to avoid concurrent buddy init
224 * we need to make sure that either on-disk bitmap or PA has uptodate data
225 * given (3) we care that PA-=N operation doesn't interfere with init
227 * the simplest way would be to have buddy initialized by the discard
228 * - use locality group PA
229 * again PA-=N must be serialized with init
230 * - discard locality group PA
231 * the simplest way would be to have buddy initialized by the discard
234 * i_data_sem serializes them
236 * discard process must wait until PA isn't used by another process
237 * - use locality group PA
238 * some mutex should serialize them
239 * - discard locality group PA
240 * discard process must wait until PA isn't used by another process
243 * i_data_sem or another mutex should serializes them
245 * discard process must wait until PA isn't used by another process
246 * - use locality group PA
247 * nothing wrong here -- they're different PAs covering different blocks
248 * - discard locality group PA
249 * discard process must wait until PA isn't used by another process
251 * now we're ready to make few consequences:
252 * - PA is referenced and while it is no discard is possible
253 * - PA is referenced until block isn't marked in on-disk bitmap
254 * - PA changes only after on-disk bitmap
255 * - discard must not compete with init. either init is done before
256 * any discard or they're serialized somehow
257 * - buddy init as sum of on-disk bitmap and PAs is done atomically
259 * a special case when we've used PA to emptiness. no need to modify buddy
260 * in this case, but we should care about concurrent init
265 * Logic in few words:
270 * mark bits in on-disk bitmap
273 * - use preallocation:
274 * find proper PA (per-inode or group)
276 * mark bits in on-disk bitmap
282 * mark bits in on-disk bitmap
285 * - discard preallocations in group:
287 * move them onto local list
288 * load on-disk bitmap
290 * remove PA from object (inode or locality group)
291 * mark free blocks in-core
293 * - discard inode's preallocations:
300 * - bitlock on a group (group)
301 * - object (inode/locality) (object)
312 * - release consumed pa:
317 * - generate in-core bitmap:
321 * - discard all for given object (inode, locality group):
326 * - discard all for given group:
333 static struct kmem_cache *ext4_pspace_cachep;
334 static struct kmem_cache *ext4_ac_cachep;
335 static struct kmem_cache *ext4_free_ext_cachep;
336 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
338 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
340 static int ext4_mb_init_per_dev_proc(struct super_block *sb);
341 static int ext4_mb_destroy_per_dev_proc(struct super_block *sb);
342 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
346 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
348 #if BITS_PER_LONG == 64
349 *bit += ((unsigned long) addr & 7UL) << 3;
350 addr = (void *) ((unsigned long) addr & ~7UL);
351 #elif BITS_PER_LONG == 32
352 *bit += ((unsigned long) addr & 3UL) << 3;
353 addr = (void *) ((unsigned long) addr & ~3UL);
355 #error "how many bits you are?!"
360 static inline int mb_test_bit(int bit, void *addr)
363 * ext4_test_bit on architecture like powerpc
364 * needs unsigned long aligned address
366 addr = mb_correct_addr_and_bit(&bit, addr);
367 return ext4_test_bit(bit, addr);
370 static inline void mb_set_bit(int bit, void *addr)
372 addr = mb_correct_addr_and_bit(&bit, addr);
373 ext4_set_bit(bit, addr);
376 static inline void mb_set_bit_atomic(spinlock_t *lock, int bit, void *addr)
378 addr = mb_correct_addr_and_bit(&bit, addr);
379 ext4_set_bit_atomic(lock, bit, addr);
382 static inline void mb_clear_bit(int bit, void *addr)
384 addr = mb_correct_addr_and_bit(&bit, addr);
385 ext4_clear_bit(bit, addr);
388 static inline void mb_clear_bit_atomic(spinlock_t *lock, int bit, void *addr)
390 addr = mb_correct_addr_and_bit(&bit, addr);
391 ext4_clear_bit_atomic(lock, bit, addr);
394 static inline int mb_find_next_zero_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_zero_bit(addr, tmpmax, start) - fix;
407 static inline int mb_find_next_bit(void *addr, int max, int start)
409 int fix = 0, ret, tmpmax;
410 addr = mb_correct_addr_and_bit(&fix, addr);
414 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
420 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
424 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
427 if (order > e4b->bd_blkbits + 1) {
432 /* at order 0 we see each particular block */
433 *max = 1 << (e4b->bd_blkbits + 3);
435 return EXT4_MB_BITMAP(e4b);
437 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
438 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
444 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
445 int first, int count)
448 struct super_block *sb = e4b->bd_sb;
450 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
452 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
453 for (i = 0; i < count; i++) {
454 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
455 ext4_fsblk_t blocknr;
456 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
457 blocknr += first + i;
459 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
460 ext4_grp_locked_error(sb, e4b->bd_group,
461 __func__, "double-free of inode"
462 " %lu's block %llu(bit %u in group %u)",
463 inode ? inode->i_ino : 0, blocknr,
464 first + i, e4b->bd_group);
466 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
470 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
474 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
476 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
477 for (i = 0; i < count; i++) {
478 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
479 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
483 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
485 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
486 unsigned char *b1, *b2;
488 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
489 b2 = (unsigned char *) bitmap;
490 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
491 if (b1[i] != b2[i]) {
492 printk(KERN_ERR "corruption in group %u "
493 "at byte %u(%u): %x in copy != %x "
494 "on disk/prealloc\n",
495 e4b->bd_group, i, i * 8, b1[i], b2[i]);
503 static inline void mb_free_blocks_double(struct inode *inode,
504 struct ext4_buddy *e4b, int first, int count)
508 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
509 int first, int count)
513 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
519 #ifdef AGGRESSIVE_CHECK
521 #define MB_CHECK_ASSERT(assert) \
525 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
526 function, file, line, # assert); \
531 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
532 const char *function, int line)
534 struct super_block *sb = e4b->bd_sb;
535 int order = e4b->bd_blkbits + 1;
542 struct ext4_group_info *grp;
545 struct list_head *cur;
550 static int mb_check_counter;
551 if (mb_check_counter++ % 100 != 0)
556 buddy = mb_find_buddy(e4b, order, &max);
557 MB_CHECK_ASSERT(buddy);
558 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
559 MB_CHECK_ASSERT(buddy2);
560 MB_CHECK_ASSERT(buddy != buddy2);
561 MB_CHECK_ASSERT(max * 2 == max2);
564 for (i = 0; i < max; i++) {
566 if (mb_test_bit(i, buddy)) {
567 /* only single bit in buddy2 may be 1 */
568 if (!mb_test_bit(i << 1, buddy2)) {
570 mb_test_bit((i<<1)+1, buddy2));
571 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
573 mb_test_bit(i << 1, buddy2));
578 /* both bits in buddy2 must be 0 */
579 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
580 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
582 for (j = 0; j < (1 << order); j++) {
583 k = (i * (1 << order)) + j;
585 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
589 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
594 buddy = mb_find_buddy(e4b, 0, &max);
595 for (i = 0; i < max; i++) {
596 if (!mb_test_bit(i, buddy)) {
597 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
605 /* check used bits only */
606 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
607 buddy2 = mb_find_buddy(e4b, j, &max2);
609 MB_CHECK_ASSERT(k < max2);
610 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
613 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
614 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
616 grp = ext4_get_group_info(sb, e4b->bd_group);
617 buddy = mb_find_buddy(e4b, 0, &max);
618 list_for_each(cur, &grp->bb_prealloc_list) {
619 ext4_group_t groupnr;
620 struct ext4_prealloc_space *pa;
621 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
622 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
623 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
624 for (i = 0; i < pa->pa_len; i++)
625 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
629 #undef MB_CHECK_ASSERT
630 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
631 __FILE__, __func__, __LINE__)
633 #define mb_check_buddy(e4b)
636 /* FIXME!! need more doc */
637 static void ext4_mb_mark_free_simple(struct super_block *sb,
638 void *buddy, unsigned first, int len,
639 struct ext4_group_info *grp)
641 struct ext4_sb_info *sbi = EXT4_SB(sb);
644 unsigned short chunk;
645 unsigned short border;
647 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
649 border = 2 << sb->s_blocksize_bits;
652 /* find how many blocks can be covered since this position */
653 max = ffs(first | border) - 1;
655 /* find how many blocks of power 2 we need to mark */
662 /* mark multiblock chunks only */
663 grp->bb_counters[min]++;
665 mb_clear_bit(first >> min,
666 buddy + sbi->s_mb_offsets[min]);
673 static void ext4_mb_generate_buddy(struct super_block *sb,
674 void *buddy, void *bitmap, ext4_group_t group)
676 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
677 unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
678 unsigned short i = 0;
679 unsigned short first;
682 unsigned fragments = 0;
683 unsigned long long period = get_cycles();
685 /* initialize buddy from bitmap which is aggregation
686 * of on-disk bitmap and preallocations */
687 i = mb_find_next_zero_bit(bitmap, max, 0);
688 grp->bb_first_free = i;
692 i = mb_find_next_bit(bitmap, max, i);
696 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
698 grp->bb_counters[0]++;
700 i = mb_find_next_zero_bit(bitmap, max, i);
702 grp->bb_fragments = fragments;
704 if (free != grp->bb_free) {
705 ext4_grp_locked_error(sb, group, __func__,
706 "EXT4-fs: group %u: %u blocks in bitmap, %u in gd",
707 group, free, grp->bb_free);
709 * If we intent to continue, we consider group descritor
710 * corrupt and update bb_free using bitmap value
715 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
717 period = get_cycles() - period;
718 spin_lock(&EXT4_SB(sb)->s_bal_lock);
719 EXT4_SB(sb)->s_mb_buddies_generated++;
720 EXT4_SB(sb)->s_mb_generation_time += period;
721 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
724 /* The buddy information is attached the buddy cache inode
725 * for convenience. The information regarding each group
726 * is loaded via ext4_mb_load_buddy. The information involve
727 * block bitmap and buddy information. The information are
728 * stored in the inode as
731 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
734 * one block each for bitmap and buddy information.
735 * So for each group we take up 2 blocks. A page can
736 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
737 * So it can have information regarding groups_per_page which
738 * is blocks_per_page/2
741 static int ext4_mb_init_cache(struct page *page, char *incore)
748 ext4_group_t first_group;
750 struct super_block *sb;
751 struct buffer_head *bhs;
752 struct buffer_head **bh;
757 mb_debug("init page %lu\n", page->index);
759 inode = page->mapping->host;
761 blocksize = 1 << inode->i_blkbits;
762 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
764 groups_per_page = blocks_per_page >> 1;
765 if (groups_per_page == 0)
768 /* allocate buffer_heads to read bitmaps */
769 if (groups_per_page > 1) {
771 i = sizeof(struct buffer_head *) * groups_per_page;
772 bh = kzalloc(i, GFP_NOFS);
778 first_group = page->index * blocks_per_page / 2;
780 /* read all groups the page covers into the cache */
781 for (i = 0; i < groups_per_page; i++) {
782 struct ext4_group_desc *desc;
784 if (first_group + i >= EXT4_SB(sb)->s_groups_count)
788 desc = ext4_get_group_desc(sb, first_group + i, NULL);
793 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
797 if (bitmap_uptodate(bh[i]))
801 if (bitmap_uptodate(bh[i])) {
802 unlock_buffer(bh[i]);
805 spin_lock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
806 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
807 ext4_init_block_bitmap(sb, bh[i],
808 first_group + i, desc);
809 set_bitmap_uptodate(bh[i]);
810 set_buffer_uptodate(bh[i]);
811 spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
812 unlock_buffer(bh[i]);
815 spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
816 if (buffer_uptodate(bh[i])) {
818 * if not uninit if bh is uptodate,
819 * bitmap is also uptodate
821 set_bitmap_uptodate(bh[i]);
822 unlock_buffer(bh[i]);
827 * submit the buffer_head for read. We can
828 * safely mark the bitmap as uptodate now.
829 * We do it here so the bitmap uptodate bit
830 * get set with buffer lock held.
832 set_bitmap_uptodate(bh[i]);
833 bh[i]->b_end_io = end_buffer_read_sync;
834 submit_bh(READ, bh[i]);
835 mb_debug("read bitmap for group %u\n", first_group + i);
838 /* wait for I/O completion */
839 for (i = 0; i < groups_per_page && bh[i]; i++)
840 wait_on_buffer(bh[i]);
843 for (i = 0; i < groups_per_page && bh[i]; i++)
844 if (!buffer_uptodate(bh[i]))
848 first_block = page->index * blocks_per_page;
850 memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
851 for (i = 0; i < blocks_per_page; i++) {
853 struct ext4_group_info *grinfo;
855 group = (first_block + i) >> 1;
856 if (group >= EXT4_SB(sb)->s_groups_count)
860 * data carry information regarding this
861 * particular group in the format specified
865 data = page_address(page) + (i * blocksize);
866 bitmap = bh[group - first_group]->b_data;
869 * We place the buddy block and bitmap block
872 if ((first_block + i) & 1) {
873 /* this is block of buddy */
874 BUG_ON(incore == NULL);
875 mb_debug("put buddy for group %u in page %lu/%x\n",
876 group, page->index, i * blocksize);
877 grinfo = ext4_get_group_info(sb, group);
878 grinfo->bb_fragments = 0;
879 memset(grinfo->bb_counters, 0,
880 sizeof(unsigned short)*(sb->s_blocksize_bits+2));
882 * incore got set to the group block bitmap below
884 ext4_lock_group(sb, group);
885 ext4_mb_generate_buddy(sb, data, incore, group);
886 ext4_unlock_group(sb, group);
889 /* this is block of bitmap */
890 BUG_ON(incore != NULL);
891 mb_debug("put bitmap for group %u in page %lu/%x\n",
892 group, page->index, i * blocksize);
894 /* see comments in ext4_mb_put_pa() */
895 ext4_lock_group(sb, group);
896 memcpy(data, bitmap, blocksize);
898 /* mark all preallocated blks used in in-core bitmap */
899 ext4_mb_generate_from_pa(sb, data, group);
900 ext4_mb_generate_from_freelist(sb, data, group);
901 ext4_unlock_group(sb, group);
903 /* set incore so that the buddy information can be
904 * generated using this
909 SetPageUptodate(page);
913 for (i = 0; i < groups_per_page && bh[i]; i++)
921 static noinline_for_stack int
922 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
923 struct ext4_buddy *e4b)
931 struct ext4_group_info *grp;
932 struct ext4_sb_info *sbi = EXT4_SB(sb);
933 struct inode *inode = sbi->s_buddy_cache;
935 mb_debug("load group %u\n", group);
937 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
938 grp = ext4_get_group_info(sb, group);
940 e4b->bd_blkbits = sb->s_blocksize_bits;
941 e4b->bd_info = ext4_get_group_info(sb, group);
943 e4b->bd_group = group;
944 e4b->bd_buddy_page = NULL;
945 e4b->bd_bitmap_page = NULL;
946 e4b->alloc_semp = &grp->alloc_sem;
948 /* Take the read lock on the group alloc
949 * sem. This would make sure a parallel
950 * ext4_mb_init_group happening on other
951 * groups mapped by the page is blocked
952 * till we are done with allocation
954 down_read(e4b->alloc_semp);
957 * the buddy cache inode stores the block bitmap
958 * and buddy information in consecutive blocks.
959 * So for each group we need two blocks.
962 pnum = block / blocks_per_page;
963 poff = block % blocks_per_page;
965 /* we could use find_or_create_page(), but it locks page
966 * what we'd like to avoid in fast path ... */
967 page = find_get_page(inode->i_mapping, pnum);
968 if (page == NULL || !PageUptodate(page)) {
971 * drop the page reference and try
972 * to get the page with lock. If we
973 * are not uptodate that implies
974 * somebody just created the page but
975 * is yet to initialize the same. So
976 * wait for it to initialize.
978 page_cache_release(page);
979 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
981 BUG_ON(page->mapping != inode->i_mapping);
982 if (!PageUptodate(page)) {
983 ret = ext4_mb_init_cache(page, NULL);
988 mb_cmp_bitmaps(e4b, page_address(page) +
989 (poff * sb->s_blocksize));
994 if (page == NULL || !PageUptodate(page)) {
998 e4b->bd_bitmap_page = page;
999 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1000 mark_page_accessed(page);
1003 pnum = block / blocks_per_page;
1004 poff = block % blocks_per_page;
1006 page = find_get_page(inode->i_mapping, pnum);
1007 if (page == NULL || !PageUptodate(page)) {
1009 page_cache_release(page);
1010 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1012 BUG_ON(page->mapping != inode->i_mapping);
1013 if (!PageUptodate(page)) {
1014 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1023 if (page == NULL || !PageUptodate(page)) {
1027 e4b->bd_buddy_page = page;
1028 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1029 mark_page_accessed(page);
1031 BUG_ON(e4b->bd_bitmap_page == NULL);
1032 BUG_ON(e4b->bd_buddy_page == NULL);
1037 if (e4b->bd_bitmap_page)
1038 page_cache_release(e4b->bd_bitmap_page);
1039 if (e4b->bd_buddy_page)
1040 page_cache_release(e4b->bd_buddy_page);
1041 e4b->bd_buddy = NULL;
1042 e4b->bd_bitmap = NULL;
1044 /* Done with the buddy cache */
1045 up_read(e4b->alloc_semp);
1049 static void ext4_mb_release_desc(struct ext4_buddy *e4b)
1051 if (e4b->bd_bitmap_page)
1052 page_cache_release(e4b->bd_bitmap_page);
1053 if (e4b->bd_buddy_page)
1054 page_cache_release(e4b->bd_buddy_page);
1055 /* Done with the buddy cache */
1056 if (e4b->alloc_semp)
1057 up_read(e4b->alloc_semp);
1061 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1066 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1067 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1069 bb = EXT4_MB_BUDDY(e4b);
1070 while (order <= e4b->bd_blkbits + 1) {
1072 if (!mb_test_bit(block, bb)) {
1073 /* this block is part of buddy of order 'order' */
1076 bb += 1 << (e4b->bd_blkbits - order);
1082 static void mb_clear_bits(spinlock_t *lock, void *bm, int cur, int len)
1088 if ((cur & 31) == 0 && (len - cur) >= 32) {
1089 /* fast path: clear whole word at once */
1090 addr = bm + (cur >> 3);
1096 mb_clear_bit_atomic(lock, cur, bm);
1098 mb_clear_bit(cur, bm);
1103 static void mb_set_bits(spinlock_t *lock, void *bm, int cur, int len)
1109 if ((cur & 31) == 0 && (len - cur) >= 32) {
1110 /* fast path: set whole word at once */
1111 addr = bm + (cur >> 3);
1117 mb_set_bit_atomic(lock, cur, bm);
1119 mb_set_bit(cur, bm);
1124 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1125 int first, int count)
1132 struct super_block *sb = e4b->bd_sb;
1134 BUG_ON(first + count > (sb->s_blocksize << 3));
1135 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
1136 mb_check_buddy(e4b);
1137 mb_free_blocks_double(inode, e4b, first, count);
1139 e4b->bd_info->bb_free += count;
1140 if (first < e4b->bd_info->bb_first_free)
1141 e4b->bd_info->bb_first_free = first;
1143 /* let's maintain fragments counter */
1145 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1146 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1147 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1149 e4b->bd_info->bb_fragments--;
1150 else if (!block && !max)
1151 e4b->bd_info->bb_fragments++;
1153 /* let's maintain buddy itself */
1154 while (count-- > 0) {
1158 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1159 ext4_fsblk_t blocknr;
1160 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1163 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
1164 ext4_grp_locked_error(sb, e4b->bd_group,
1165 __func__, "double-free of inode"
1166 " %lu's block %llu(bit %u in group %u)",
1167 inode ? inode->i_ino : 0, blocknr, block,
1170 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1171 e4b->bd_info->bb_counters[order]++;
1173 /* start of the buddy */
1174 buddy = mb_find_buddy(e4b, order, &max);
1178 if (mb_test_bit(block, buddy) ||
1179 mb_test_bit(block + 1, buddy))
1182 /* both the buddies are free, try to coalesce them */
1183 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1189 /* for special purposes, we don't set
1190 * free bits in bitmap */
1191 mb_set_bit(block, buddy);
1192 mb_set_bit(block + 1, buddy);
1194 e4b->bd_info->bb_counters[order]--;
1195 e4b->bd_info->bb_counters[order]--;
1199 e4b->bd_info->bb_counters[order]++;
1201 mb_clear_bit(block, buddy2);
1205 mb_check_buddy(e4b);
1208 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1209 int needed, struct ext4_free_extent *ex)
1216 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1219 buddy = mb_find_buddy(e4b, order, &max);
1220 BUG_ON(buddy == NULL);
1221 BUG_ON(block >= max);
1222 if (mb_test_bit(block, buddy)) {
1229 /* FIXME dorp order completely ? */
1230 if (likely(order == 0)) {
1231 /* find actual order */
1232 order = mb_find_order_for_block(e4b, block);
1233 block = block >> order;
1236 ex->fe_len = 1 << order;
1237 ex->fe_start = block << order;
1238 ex->fe_group = e4b->bd_group;
1240 /* calc difference from given start */
1241 next = next - ex->fe_start;
1243 ex->fe_start += next;
1245 while (needed > ex->fe_len &&
1246 (buddy = mb_find_buddy(e4b, order, &max))) {
1248 if (block + 1 >= max)
1251 next = (block + 1) * (1 << order);
1252 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1255 ord = mb_find_order_for_block(e4b, next);
1258 block = next >> order;
1259 ex->fe_len += 1 << order;
1262 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1266 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1272 int start = ex->fe_start;
1273 int len = ex->fe_len;
1278 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1279 BUG_ON(e4b->bd_group != ex->fe_group);
1280 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1281 mb_check_buddy(e4b);
1282 mb_mark_used_double(e4b, start, len);
1284 e4b->bd_info->bb_free -= len;
1285 if (e4b->bd_info->bb_first_free == start)
1286 e4b->bd_info->bb_first_free += len;
1288 /* let's maintain fragments counter */
1290 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1291 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1292 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1294 e4b->bd_info->bb_fragments++;
1295 else if (!mlen && !max)
1296 e4b->bd_info->bb_fragments--;
1298 /* let's maintain buddy itself */
1300 ord = mb_find_order_for_block(e4b, start);
1302 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1303 /* the whole chunk may be allocated at once! */
1305 buddy = mb_find_buddy(e4b, ord, &max);
1306 BUG_ON((start >> ord) >= max);
1307 mb_set_bit(start >> ord, buddy);
1308 e4b->bd_info->bb_counters[ord]--;
1315 /* store for history */
1317 ret = len | (ord << 16);
1319 /* we have to split large buddy */
1321 buddy = mb_find_buddy(e4b, ord, &max);
1322 mb_set_bit(start >> ord, buddy);
1323 e4b->bd_info->bb_counters[ord]--;
1326 cur = (start >> ord) & ~1U;
1327 buddy = mb_find_buddy(e4b, ord, &max);
1328 mb_clear_bit(cur, buddy);
1329 mb_clear_bit(cur + 1, buddy);
1330 e4b->bd_info->bb_counters[ord]++;
1331 e4b->bd_info->bb_counters[ord]++;
1334 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b->bd_sb), ex->fe_group),
1335 EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1336 mb_check_buddy(e4b);
1342 * Must be called under group lock!
1344 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1345 struct ext4_buddy *e4b)
1347 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1350 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1351 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1353 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1354 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1355 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1357 /* preallocation can change ac_b_ex, thus we store actually
1358 * allocated blocks for history */
1359 ac->ac_f_ex = ac->ac_b_ex;
1361 ac->ac_status = AC_STATUS_FOUND;
1362 ac->ac_tail = ret & 0xffff;
1363 ac->ac_buddy = ret >> 16;
1366 * take the page reference. We want the page to be pinned
1367 * so that we don't get a ext4_mb_init_cache_call for this
1368 * group until we update the bitmap. That would mean we
1369 * double allocate blocks. The reference is dropped
1370 * in ext4_mb_release_context
1372 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1373 get_page(ac->ac_bitmap_page);
1374 ac->ac_buddy_page = e4b->bd_buddy_page;
1375 get_page(ac->ac_buddy_page);
1376 /* on allocation we use ac to track the held semaphore */
1377 ac->alloc_semp = e4b->alloc_semp;
1378 e4b->alloc_semp = NULL;
1379 /* store last allocated for subsequent stream allocation */
1380 if ((ac->ac_flags & EXT4_MB_HINT_DATA)) {
1381 spin_lock(&sbi->s_md_lock);
1382 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1383 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1384 spin_unlock(&sbi->s_md_lock);
1389 * regular allocator, for general purposes allocation
1392 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1393 struct ext4_buddy *e4b,
1396 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1397 struct ext4_free_extent *bex = &ac->ac_b_ex;
1398 struct ext4_free_extent *gex = &ac->ac_g_ex;
1399 struct ext4_free_extent ex;
1402 if (ac->ac_status == AC_STATUS_FOUND)
1405 * We don't want to scan for a whole year
1407 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1408 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1409 ac->ac_status = AC_STATUS_BREAK;
1414 * Haven't found good chunk so far, let's continue
1416 if (bex->fe_len < gex->fe_len)
1419 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1420 && bex->fe_group == e4b->bd_group) {
1421 /* recheck chunk's availability - we don't know
1422 * when it was found (within this lock-unlock
1424 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1425 if (max >= gex->fe_len) {
1426 ext4_mb_use_best_found(ac, e4b);
1433 * The routine checks whether found extent is good enough. If it is,
1434 * then the extent gets marked used and flag is set to the context
1435 * to stop scanning. Otherwise, the extent is compared with the
1436 * previous found extent and if new one is better, then it's stored
1437 * in the context. Later, the best found extent will be used, if
1438 * mballoc can't find good enough extent.
1440 * FIXME: real allocation policy is to be designed yet!
1442 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1443 struct ext4_free_extent *ex,
1444 struct ext4_buddy *e4b)
1446 struct ext4_free_extent *bex = &ac->ac_b_ex;
1447 struct ext4_free_extent *gex = &ac->ac_g_ex;
1449 BUG_ON(ex->fe_len <= 0);
1450 BUG_ON(ex->fe_len >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1451 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1452 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1457 * The special case - take what you catch first
1459 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1461 ext4_mb_use_best_found(ac, e4b);
1466 * Let's check whether the chuck is good enough
1468 if (ex->fe_len == gex->fe_len) {
1470 ext4_mb_use_best_found(ac, e4b);
1475 * If this is first found extent, just store it in the context
1477 if (bex->fe_len == 0) {
1483 * If new found extent is better, store it in the context
1485 if (bex->fe_len < gex->fe_len) {
1486 /* if the request isn't satisfied, any found extent
1487 * larger than previous best one is better */
1488 if (ex->fe_len > bex->fe_len)
1490 } else if (ex->fe_len > gex->fe_len) {
1491 /* if the request is satisfied, then we try to find
1492 * an extent that still satisfy the request, but is
1493 * smaller than previous one */
1494 if (ex->fe_len < bex->fe_len)
1498 ext4_mb_check_limits(ac, e4b, 0);
1501 static int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1502 struct ext4_buddy *e4b)
1504 struct ext4_free_extent ex = ac->ac_b_ex;
1505 ext4_group_t group = ex.fe_group;
1509 BUG_ON(ex.fe_len <= 0);
1510 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1514 ext4_lock_group(ac->ac_sb, group);
1515 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1519 ext4_mb_use_best_found(ac, e4b);
1522 ext4_unlock_group(ac->ac_sb, group);
1523 ext4_mb_release_desc(e4b);
1528 static int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1529 struct ext4_buddy *e4b)
1531 ext4_group_t group = ac->ac_g_ex.fe_group;
1534 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1535 struct ext4_super_block *es = sbi->s_es;
1536 struct ext4_free_extent ex;
1538 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1541 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1545 ext4_lock_group(ac->ac_sb, group);
1546 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1547 ac->ac_g_ex.fe_len, &ex);
1549 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1552 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1553 ex.fe_start + le32_to_cpu(es->s_first_data_block);
1554 /* use do_div to get remainder (would be 64-bit modulo) */
1555 if (do_div(start, sbi->s_stripe) == 0) {
1558 ext4_mb_use_best_found(ac, e4b);
1560 } else if (max >= ac->ac_g_ex.fe_len) {
1561 BUG_ON(ex.fe_len <= 0);
1562 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1563 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1566 ext4_mb_use_best_found(ac, e4b);
1567 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1568 /* Sometimes, caller may want to merge even small
1569 * number of blocks to an existing extent */
1570 BUG_ON(ex.fe_len <= 0);
1571 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1572 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1575 ext4_mb_use_best_found(ac, e4b);
1577 ext4_unlock_group(ac->ac_sb, group);
1578 ext4_mb_release_desc(e4b);
1584 * The routine scans buddy structures (not bitmap!) from given order
1585 * to max order and tries to find big enough chunk to satisfy the req
1587 static void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1588 struct ext4_buddy *e4b)
1590 struct super_block *sb = ac->ac_sb;
1591 struct ext4_group_info *grp = e4b->bd_info;
1597 BUG_ON(ac->ac_2order <= 0);
1598 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1599 if (grp->bb_counters[i] == 0)
1602 buddy = mb_find_buddy(e4b, i, &max);
1603 BUG_ON(buddy == NULL);
1605 k = mb_find_next_zero_bit(buddy, max, 0);
1610 ac->ac_b_ex.fe_len = 1 << i;
1611 ac->ac_b_ex.fe_start = k << i;
1612 ac->ac_b_ex.fe_group = e4b->bd_group;
1614 ext4_mb_use_best_found(ac, e4b);
1616 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1618 if (EXT4_SB(sb)->s_mb_stats)
1619 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1626 * The routine scans the group and measures all found extents.
1627 * In order to optimize scanning, caller must pass number of
1628 * free blocks in the group, so the routine can know upper limit.
1630 static void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1631 struct ext4_buddy *e4b)
1633 struct super_block *sb = ac->ac_sb;
1634 void *bitmap = EXT4_MB_BITMAP(e4b);
1635 struct ext4_free_extent ex;
1639 free = e4b->bd_info->bb_free;
1642 i = e4b->bd_info->bb_first_free;
1644 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1645 i = mb_find_next_zero_bit(bitmap,
1646 EXT4_BLOCKS_PER_GROUP(sb), i);
1647 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1649 * IF we have corrupt bitmap, we won't find any
1650 * free blocks even though group info says we
1651 * we have free blocks
1653 ext4_grp_locked_error(sb, e4b->bd_group,
1654 __func__, "%d free blocks as per "
1655 "group info. But bitmap says 0",
1660 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1661 BUG_ON(ex.fe_len <= 0);
1662 if (free < ex.fe_len) {
1663 ext4_grp_locked_error(sb, e4b->bd_group,
1664 __func__, "%d free blocks as per "
1665 "group info. But got %d blocks",
1668 * The number of free blocks differs. This mostly
1669 * indicate that the bitmap is corrupt. So exit
1670 * without claiming the space.
1675 ext4_mb_measure_extent(ac, &ex, e4b);
1681 ext4_mb_check_limits(ac, e4b, 1);
1685 * This is a special case for storages like raid5
1686 * we try to find stripe-aligned chunks for stripe-size requests
1687 * XXX should do so at least for multiples of stripe size as well
1689 static void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1690 struct ext4_buddy *e4b)
1692 struct super_block *sb = ac->ac_sb;
1693 struct ext4_sb_info *sbi = EXT4_SB(sb);
1694 void *bitmap = EXT4_MB_BITMAP(e4b);
1695 struct ext4_free_extent ex;
1696 ext4_fsblk_t first_group_block;
1701 BUG_ON(sbi->s_stripe == 0);
1703 /* find first stripe-aligned block in group */
1704 first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1705 + le32_to_cpu(sbi->s_es->s_first_data_block);
1706 a = first_group_block + sbi->s_stripe - 1;
1707 do_div(a, sbi->s_stripe);
1708 i = (a * sbi->s_stripe) - first_group_block;
1710 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1711 if (!mb_test_bit(i, bitmap)) {
1712 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1713 if (max >= sbi->s_stripe) {
1716 ext4_mb_use_best_found(ac, e4b);
1724 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1725 ext4_group_t group, int cr)
1727 unsigned free, fragments;
1729 struct ext4_group_desc *desc;
1730 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1732 BUG_ON(cr < 0 || cr >= 4);
1733 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1735 free = grp->bb_free;
1736 fragments = grp->bb_fragments;
1744 BUG_ON(ac->ac_2order == 0);
1745 /* If this group is uninitialized, skip it initially */
1746 desc = ext4_get_group_desc(ac->ac_sb, group, NULL);
1747 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
1750 bits = ac->ac_sb->s_blocksize_bits + 1;
1751 for (i = ac->ac_2order; i <= bits; i++)
1752 if (grp->bb_counters[i] > 0)
1756 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1760 if (free >= ac->ac_g_ex.fe_len)
1773 * lock the group_info alloc_sem of all the groups
1774 * belonging to the same buddy cache page. This
1775 * make sure other parallel operation on the buddy
1776 * cache doesn't happen whild holding the buddy cache
1779 int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1783 int blocks_per_page;
1784 int groups_per_page;
1785 ext4_group_t first_group;
1786 struct ext4_group_info *grp;
1788 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1790 * the buddy cache inode stores the block bitmap
1791 * and buddy information in consecutive blocks.
1792 * So for each group we need two blocks.
1795 pnum = block / blocks_per_page;
1796 first_group = pnum * blocks_per_page / 2;
1798 groups_per_page = blocks_per_page >> 1;
1799 if (groups_per_page == 0)
1800 groups_per_page = 1;
1801 /* read all groups the page covers into the cache */
1802 for (i = 0; i < groups_per_page; i++) {
1804 if ((first_group + i) >= EXT4_SB(sb)->s_groups_count)
1806 grp = ext4_get_group_info(sb, first_group + i);
1807 /* take all groups write allocation
1808 * semaphore. This make sure there is
1809 * no block allocation going on in any
1812 down_write_nested(&grp->alloc_sem, i);
1817 void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1818 ext4_group_t group, int locked_group)
1822 int blocks_per_page;
1823 ext4_group_t first_group;
1824 struct ext4_group_info *grp;
1826 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1828 * the buddy cache inode stores the block bitmap
1829 * and buddy information in consecutive blocks.
1830 * So for each group we need two blocks.
1833 pnum = block / blocks_per_page;
1834 first_group = pnum * blocks_per_page / 2;
1835 /* release locks on all the groups */
1836 for (i = 0; i < locked_group; i++) {
1838 grp = ext4_get_group_info(sb, first_group + i);
1839 /* take all groups write allocation
1840 * semaphore. This make sure there is
1841 * no block allocation going on in any
1844 up_write(&grp->alloc_sem);
1849 static int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1854 int blocks_per_page;
1855 int block, pnum, poff;
1856 int num_grp_locked = 0;
1857 struct ext4_group_info *this_grp;
1858 struct ext4_sb_info *sbi = EXT4_SB(sb);
1859 struct inode *inode = sbi->s_buddy_cache;
1860 struct page *page = NULL, *bitmap_page = NULL;
1862 mb_debug("init group %lu\n", group);
1863 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1864 this_grp = ext4_get_group_info(sb, group);
1866 * This ensures we don't add group
1867 * to this buddy cache via resize
1869 num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group);
1870 if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
1872 * somebody initialized the group
1873 * return without doing anything
1879 * the buddy cache inode stores the block bitmap
1880 * and buddy information in consecutive blocks.
1881 * So for each group we need two blocks.
1884 pnum = block / blocks_per_page;
1885 poff = block % blocks_per_page;
1886 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1888 BUG_ON(page->mapping != inode->i_mapping);
1889 ret = ext4_mb_init_cache(page, NULL);
1896 if (page == NULL || !PageUptodate(page)) {
1900 mark_page_accessed(page);
1902 bitmap = page_address(page) + (poff * sb->s_blocksize);
1904 /* init buddy cache */
1906 pnum = block / blocks_per_page;
1907 poff = block % blocks_per_page;
1908 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1909 if (page == bitmap_page) {
1911 * If both the bitmap and buddy are in
1912 * the same page we don't need to force
1917 BUG_ON(page->mapping != inode->i_mapping);
1918 ret = ext4_mb_init_cache(page, bitmap);
1925 if (page == NULL || !PageUptodate(page)) {
1929 mark_page_accessed(page);
1931 ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
1933 page_cache_release(bitmap_page);
1935 page_cache_release(page);
1939 static noinline_for_stack int
1940 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1947 struct ext4_sb_info *sbi;
1948 struct super_block *sb;
1949 struct ext4_buddy e4b;
1954 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1956 /* first, try the goal */
1957 err = ext4_mb_find_by_goal(ac, &e4b);
1958 if (err || ac->ac_status == AC_STATUS_FOUND)
1961 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1965 * ac->ac2_order is set only if the fe_len is a power of 2
1966 * if ac2_order is set we also set criteria to 0 so that we
1967 * try exact allocation using buddy.
1969 i = fls(ac->ac_g_ex.fe_len);
1972 * We search using buddy data only if the order of the request
1973 * is greater than equal to the sbi_s_mb_order2_reqs
1974 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1976 if (i >= sbi->s_mb_order2_reqs) {
1978 * This should tell if fe_len is exactly power of 2
1980 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1981 ac->ac_2order = i - 1;
1984 bsbits = ac->ac_sb->s_blocksize_bits;
1985 /* if stream allocation is enabled, use global goal */
1986 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
1987 isize = i_size_read(ac->ac_inode) >> bsbits;
1991 if (size < sbi->s_mb_stream_request &&
1992 (ac->ac_flags & EXT4_MB_HINT_DATA)) {
1993 /* TBD: may be hot point */
1994 spin_lock(&sbi->s_md_lock);
1995 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1996 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1997 spin_unlock(&sbi->s_md_lock);
1999 /* Let's just scan groups to find more-less suitable blocks */
2000 cr = ac->ac_2order ? 0 : 1;
2002 * cr == 0 try to get exact allocation,
2003 * cr == 3 try to get anything
2006 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2007 ac->ac_criteria = cr;
2009 * searching for the right group start
2010 * from the goal value specified
2012 group = ac->ac_g_ex.fe_group;
2014 for (i = 0; i < EXT4_SB(sb)->s_groups_count; group++, i++) {
2015 struct ext4_group_info *grp;
2016 struct ext4_group_desc *desc;
2018 if (group == EXT4_SB(sb)->s_groups_count)
2021 /* quick check to skip empty groups */
2022 grp = ext4_get_group_info(sb, group);
2023 if (grp->bb_free == 0)
2027 * if the group is already init we check whether it is
2028 * a good group and if not we don't load the buddy
2030 if (EXT4_MB_GRP_NEED_INIT(grp)) {
2032 * we need full data about the group
2033 * to make a good selection
2035 err = ext4_mb_init_group(sb, group);
2041 * If the particular group doesn't satisfy our
2042 * criteria we continue with the next group
2044 if (!ext4_mb_good_group(ac, group, cr))
2047 err = ext4_mb_load_buddy(sb, group, &e4b);
2051 ext4_lock_group(sb, group);
2052 if (!ext4_mb_good_group(ac, group, cr)) {
2053 /* someone did allocation from this group */
2054 ext4_unlock_group(sb, group);
2055 ext4_mb_release_desc(&e4b);
2059 ac->ac_groups_scanned++;
2060 desc = ext4_get_group_desc(sb, group, NULL);
2061 if (cr == 0 || (desc->bg_flags &
2062 cpu_to_le16(EXT4_BG_BLOCK_UNINIT) &&
2063 ac->ac_2order != 0))
2064 ext4_mb_simple_scan_group(ac, &e4b);
2066 ac->ac_g_ex.fe_len == sbi->s_stripe)
2067 ext4_mb_scan_aligned(ac, &e4b);
2069 ext4_mb_complex_scan_group(ac, &e4b);
2071 ext4_unlock_group(sb, group);
2072 ext4_mb_release_desc(&e4b);
2074 if (ac->ac_status != AC_STATUS_CONTINUE)
2079 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2080 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2082 * We've been searching too long. Let's try to allocate
2083 * the best chunk we've found so far
2086 ext4_mb_try_best_found(ac, &e4b);
2087 if (ac->ac_status != AC_STATUS_FOUND) {
2089 * Someone more lucky has already allocated it.
2090 * The only thing we can do is just take first
2092 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2094 ac->ac_b_ex.fe_group = 0;
2095 ac->ac_b_ex.fe_start = 0;
2096 ac->ac_b_ex.fe_len = 0;
2097 ac->ac_status = AC_STATUS_CONTINUE;
2098 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2100 atomic_inc(&sbi->s_mb_lost_chunks);
2108 #ifdef EXT4_MB_HISTORY
2109 struct ext4_mb_proc_session {
2110 struct ext4_mb_history *history;
2111 struct super_block *sb;
2116 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
2117 struct ext4_mb_history *hs,
2120 if (hs == s->history + s->max)
2122 if (!first && hs == s->history + s->start)
2124 while (hs->orig.fe_len == 0) {
2126 if (hs == s->history + s->max)
2128 if (hs == s->history + s->start)
2134 static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
2136 struct ext4_mb_proc_session *s = seq->private;
2137 struct ext4_mb_history *hs;
2141 return SEQ_START_TOKEN;
2142 hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2145 while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
2149 static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
2152 struct ext4_mb_proc_session *s = seq->private;
2153 struct ext4_mb_history *hs = v;
2156 if (v == SEQ_START_TOKEN)
2157 return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2159 return ext4_mb_history_skip_empty(s, ++hs, 0);
2162 static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
2164 char buf[25], buf2[25], buf3[25], *fmt;
2165 struct ext4_mb_history *hs = v;
2167 if (v == SEQ_START_TOKEN) {
2168 seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
2169 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
2170 "pid", "inode", "original", "goal", "result", "found",
2171 "grps", "cr", "flags", "merge", "tail", "broken");
2175 if (hs->op == EXT4_MB_HISTORY_ALLOC) {
2176 fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2177 "%-5u %-5s %-5u %-6u\n";
2178 sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group,
2179 hs->result.fe_start, hs->result.fe_len,
2180 hs->result.fe_logical);
2181 sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group,
2182 hs->orig.fe_start, hs->orig.fe_len,
2183 hs->orig.fe_logical);
2184 sprintf(buf3, "%u/%d/%u@%u", hs->goal.fe_group,
2185 hs->goal.fe_start, hs->goal.fe_len,
2186 hs->goal.fe_logical);
2187 seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
2188 hs->found, hs->groups, hs->cr, hs->flags,
2189 hs->merged ? "M" : "", hs->tail,
2190 hs->buddy ? 1 << hs->buddy : 0);
2191 } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
2192 fmt = "%-5u %-8u %-23s %-23s %-23s\n";
2193 sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group,
2194 hs->result.fe_start, hs->result.fe_len,
2195 hs->result.fe_logical);
2196 sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group,
2197 hs->orig.fe_start, hs->orig.fe_len,
2198 hs->orig.fe_logical);
2199 seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
2200 } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
2201 sprintf(buf2, "%u/%d/%u", hs->result.fe_group,
2202 hs->result.fe_start, hs->result.fe_len);
2203 seq_printf(seq, "%-5u %-8u %-23s discard\n",
2204 hs->pid, hs->ino, buf2);
2205 } else if (hs->op == EXT4_MB_HISTORY_FREE) {
2206 sprintf(buf2, "%u/%d/%u", hs->result.fe_group,
2207 hs->result.fe_start, hs->result.fe_len);
2208 seq_printf(seq, "%-5u %-8u %-23s free\n",
2209 hs->pid, hs->ino, buf2);
2214 static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
2218 static struct seq_operations ext4_mb_seq_history_ops = {
2219 .start = ext4_mb_seq_history_start,
2220 .next = ext4_mb_seq_history_next,
2221 .stop = ext4_mb_seq_history_stop,
2222 .show = ext4_mb_seq_history_show,
2225 static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
2227 struct super_block *sb = PDE(inode)->data;
2228 struct ext4_sb_info *sbi = EXT4_SB(sb);
2229 struct ext4_mb_proc_session *s;
2233 if (unlikely(sbi->s_mb_history == NULL))
2235 s = kmalloc(sizeof(*s), GFP_KERNEL);
2239 size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
2240 s->history = kmalloc(size, GFP_KERNEL);
2241 if (s->history == NULL) {
2246 spin_lock(&sbi->s_mb_history_lock);
2247 memcpy(s->history, sbi->s_mb_history, size);
2248 s->max = sbi->s_mb_history_max;
2249 s->start = sbi->s_mb_history_cur % s->max;
2250 spin_unlock(&sbi->s_mb_history_lock);
2252 rc = seq_open(file, &ext4_mb_seq_history_ops);
2254 struct seq_file *m = (struct seq_file *)file->private_data;
2264 static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
2266 struct seq_file *seq = (struct seq_file *)file->private_data;
2267 struct ext4_mb_proc_session *s = seq->private;
2270 return seq_release(inode, file);
2273 static ssize_t ext4_mb_seq_history_write(struct file *file,
2274 const char __user *buffer,
2275 size_t count, loff_t *ppos)
2277 struct seq_file *seq = (struct seq_file *)file->private_data;
2278 struct ext4_mb_proc_session *s = seq->private;
2279 struct super_block *sb = s->sb;
2283 if (count >= sizeof(str)) {
2284 printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
2285 "mb_history", (int)sizeof(str));
2289 if (copy_from_user(str, buffer, count))
2292 value = simple_strtol(str, NULL, 0);
2295 EXT4_SB(sb)->s_mb_history_filter = value;
2300 static struct file_operations ext4_mb_seq_history_fops = {
2301 .owner = THIS_MODULE,
2302 .open = ext4_mb_seq_history_open,
2304 .write = ext4_mb_seq_history_write,
2305 .llseek = seq_lseek,
2306 .release = ext4_mb_seq_history_release,
2309 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2311 struct super_block *sb = seq->private;
2312 struct ext4_sb_info *sbi = EXT4_SB(sb);
2315 if (*pos < 0 || *pos >= sbi->s_groups_count)
2319 return (void *) ((unsigned long) group);
2322 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2324 struct super_block *sb = seq->private;
2325 struct ext4_sb_info *sbi = EXT4_SB(sb);
2329 if (*pos < 0 || *pos >= sbi->s_groups_count)
2332 return (void *) ((unsigned long) group);
2335 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2337 struct super_block *sb = seq->private;
2338 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2341 struct ext4_buddy e4b;
2343 struct ext4_group_info info;
2344 unsigned short counters[16];
2349 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2350 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2351 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2352 "group", "free", "frags", "first",
2353 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2354 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2356 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2357 sizeof(struct ext4_group_info);
2358 err = ext4_mb_load_buddy(sb, group, &e4b);
2360 seq_printf(seq, "#%-5u: I/O error\n", group);
2363 ext4_lock_group(sb, group);
2364 memcpy(&sg, ext4_get_group_info(sb, group), i);
2365 ext4_unlock_group(sb, group);
2366 ext4_mb_release_desc(&e4b);
2368 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2369 sg.info.bb_fragments, sg.info.bb_first_free);
2370 for (i = 0; i <= 13; i++)
2371 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2372 sg.info.bb_counters[i] : 0);
2373 seq_printf(seq, " ]\n");
2378 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2382 static struct seq_operations ext4_mb_seq_groups_ops = {
2383 .start = ext4_mb_seq_groups_start,
2384 .next = ext4_mb_seq_groups_next,
2385 .stop = ext4_mb_seq_groups_stop,
2386 .show = ext4_mb_seq_groups_show,
2389 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2391 struct super_block *sb = PDE(inode)->data;
2394 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2396 struct seq_file *m = (struct seq_file *)file->private_data;
2403 static struct file_operations ext4_mb_seq_groups_fops = {
2404 .owner = THIS_MODULE,
2405 .open = ext4_mb_seq_groups_open,
2407 .llseek = seq_lseek,
2408 .release = seq_release,
2411 static void ext4_mb_history_release(struct super_block *sb)
2413 struct ext4_sb_info *sbi = EXT4_SB(sb);
2415 if (sbi->s_proc != NULL) {
2416 remove_proc_entry("mb_groups", sbi->s_proc);
2417 remove_proc_entry("mb_history", sbi->s_proc);
2419 kfree(sbi->s_mb_history);
2422 static void ext4_mb_history_init(struct super_block *sb)
2424 struct ext4_sb_info *sbi = EXT4_SB(sb);
2427 if (sbi->s_proc != NULL) {
2428 proc_create_data("mb_history", S_IRUGO, sbi->s_proc,
2429 &ext4_mb_seq_history_fops, sb);
2430 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2431 &ext4_mb_seq_groups_fops, sb);
2434 sbi->s_mb_history_max = 1000;
2435 sbi->s_mb_history_cur = 0;
2436 spin_lock_init(&sbi->s_mb_history_lock);
2437 i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
2438 sbi->s_mb_history = kzalloc(i, GFP_KERNEL);
2439 /* if we can't allocate history, then we simple won't use it */
2442 static noinline_for_stack void
2443 ext4_mb_store_history(struct ext4_allocation_context *ac)
2445 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2446 struct ext4_mb_history h;
2448 if (unlikely(sbi->s_mb_history == NULL))
2451 if (!(ac->ac_op & sbi->s_mb_history_filter))
2455 h.pid = current->pid;
2456 h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2457 h.orig = ac->ac_o_ex;
2458 h.result = ac->ac_b_ex;
2459 h.flags = ac->ac_flags;
2460 h.found = ac->ac_found;
2461 h.groups = ac->ac_groups_scanned;
2462 h.cr = ac->ac_criteria;
2463 h.tail = ac->ac_tail;
2464 h.buddy = ac->ac_buddy;
2466 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
2467 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2468 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2470 h.goal = ac->ac_g_ex;
2471 h.result = ac->ac_f_ex;
2474 spin_lock(&sbi->s_mb_history_lock);
2475 memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2476 if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2477 sbi->s_mb_history_cur = 0;
2478 spin_unlock(&sbi->s_mb_history_lock);
2482 #define ext4_mb_history_release(sb)
2483 #define ext4_mb_history_init(sb)
2487 /* Create and initialize ext4_group_info data for the given group. */
2488 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2489 struct ext4_group_desc *desc)
2493 struct ext4_sb_info *sbi = EXT4_SB(sb);
2494 struct ext4_group_info **meta_group_info;
2497 * First check if this group is the first of a reserved block.
2498 * If it's true, we have to allocate a new table of pointers
2499 * to ext4_group_info structures
2501 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2502 metalen = sizeof(*meta_group_info) <<
2503 EXT4_DESC_PER_BLOCK_BITS(sb);
2504 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2505 if (meta_group_info == NULL) {
2506 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2508 goto exit_meta_group_info;
2510 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2515 * calculate needed size. if change bb_counters size,
2516 * don't forget about ext4_mb_generate_buddy()
2518 len = offsetof(typeof(**meta_group_info),
2519 bb_counters[sb->s_blocksize_bits + 2]);
2522 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2523 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2525 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2526 if (meta_group_info[i] == NULL) {
2527 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2528 goto exit_group_info;
2530 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2531 &(meta_group_info[i]->bb_state));
2534 * initialize bb_free to be able to skip
2535 * empty groups without initialization
2537 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2538 meta_group_info[i]->bb_free =
2539 ext4_free_blocks_after_init(sb, group, desc);
2541 meta_group_info[i]->bb_free =
2542 ext4_free_blks_count(sb, desc);
2545 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2546 init_rwsem(&meta_group_info[i]->alloc_sem);
2547 meta_group_info[i]->bb_free_root.rb_node = NULL;;
2551 struct buffer_head *bh;
2552 meta_group_info[i]->bb_bitmap =
2553 kmalloc(sb->s_blocksize, GFP_KERNEL);
2554 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2555 bh = ext4_read_block_bitmap(sb, group);
2557 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2566 /* If a meta_group_info table has been allocated, release it now */
2567 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2568 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2569 exit_meta_group_info:
2571 } /* ext4_mb_add_groupinfo */
2574 * Update an existing group.
2575 * This function is used for online resize
2577 void ext4_mb_update_group_info(struct ext4_group_info *grp, ext4_grpblk_t add)
2579 grp->bb_free += add;
2582 static int ext4_mb_init_backend(struct super_block *sb)
2586 struct ext4_sb_info *sbi = EXT4_SB(sb);
2587 struct ext4_super_block *es = sbi->s_es;
2588 int num_meta_group_infos;
2589 int num_meta_group_infos_max;
2591 struct ext4_group_info **meta_group_info;
2592 struct ext4_group_desc *desc;
2594 /* This is the number of blocks used by GDT */
2595 num_meta_group_infos = (sbi->s_groups_count + EXT4_DESC_PER_BLOCK(sb) -
2596 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2599 * This is the total number of blocks used by GDT including
2600 * the number of reserved blocks for GDT.
2601 * The s_group_info array is allocated with this value
2602 * to allow a clean online resize without a complex
2603 * manipulation of pointer.
2604 * The drawback is the unused memory when no resize
2605 * occurs but it's very low in terms of pages
2606 * (see comments below)
2607 * Need to handle this properly when META_BG resizing is allowed
2609 num_meta_group_infos_max = num_meta_group_infos +
2610 le16_to_cpu(es->s_reserved_gdt_blocks);
2613 * array_size is the size of s_group_info array. We round it
2614 * to the next power of two because this approximation is done
2615 * internally by kmalloc so we can have some more memory
2616 * for free here (e.g. may be used for META_BG resize).
2619 while (array_size < sizeof(*sbi->s_group_info) *
2620 num_meta_group_infos_max)
2621 array_size = array_size << 1;
2622 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2623 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2624 * So a two level scheme suffices for now. */
2625 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2626 if (sbi->s_group_info == NULL) {
2627 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2630 sbi->s_buddy_cache = new_inode(sb);
2631 if (sbi->s_buddy_cache == NULL) {
2632 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2635 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2637 metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb);
2638 for (i = 0; i < num_meta_group_infos; i++) {
2639 if ((i + 1) == num_meta_group_infos)
2640 metalen = sizeof(*meta_group_info) *
2641 (sbi->s_groups_count -
2642 (i << EXT4_DESC_PER_BLOCK_BITS(sb)));
2643 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2644 if (meta_group_info == NULL) {
2645 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2649 sbi->s_group_info[i] = meta_group_info;
2652 for (i = 0; i < sbi->s_groups_count; i++) {
2653 desc = ext4_get_group_desc(sb, i, NULL);
2656 "EXT4-fs: can't read descriptor %u\n", i);
2659 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2667 kfree(ext4_get_group_info(sb, i));
2668 i = num_meta_group_infos;
2671 kfree(sbi->s_group_info[i]);
2672 iput(sbi->s_buddy_cache);
2674 kfree(sbi->s_group_info);
2678 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2680 struct ext4_sb_info *sbi = EXT4_SB(sb);
2686 i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
2688 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2689 if (sbi->s_mb_offsets == NULL) {
2693 i = (sb->s_blocksize_bits + 2) * sizeof(unsigned int);
2694 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2695 if (sbi->s_mb_maxs == NULL) {
2696 kfree(sbi->s_mb_maxs);
2700 /* order 0 is regular bitmap */
2701 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2702 sbi->s_mb_offsets[0] = 0;
2706 max = sb->s_blocksize << 2;
2708 sbi->s_mb_offsets[i] = offset;
2709 sbi->s_mb_maxs[i] = max;
2710 offset += 1 << (sb->s_blocksize_bits - i);
2713 } while (i <= sb->s_blocksize_bits + 1);
2715 /* init file for buddy data */
2716 ret = ext4_mb_init_backend(sb);
2718 kfree(sbi->s_mb_offsets);
2719 kfree(sbi->s_mb_maxs);
2723 spin_lock_init(&sbi->s_md_lock);
2724 spin_lock_init(&sbi->s_bal_lock);
2726 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2727 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2728 sbi->s_mb_stats = MB_DEFAULT_STATS;
2729 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2730 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2731 sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
2732 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2734 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2735 if (sbi->s_locality_groups == NULL) {
2736 kfree(sbi->s_mb_offsets);
2737 kfree(sbi->s_mb_maxs);
2740 for_each_possible_cpu(i) {
2741 struct ext4_locality_group *lg;
2742 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2743 mutex_init(&lg->lg_mutex);
2744 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2745 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2746 spin_lock_init(&lg->lg_prealloc_lock);
2749 ext4_mb_init_per_dev_proc(sb);
2750 ext4_mb_history_init(sb);
2753 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2755 printk(KERN_INFO "EXT4-fs: mballoc enabled\n");
2759 /* need to called with ext4 group lock (ext4_lock_group) */
2760 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2762 struct ext4_prealloc_space *pa;
2763 struct list_head *cur, *tmp;
2766 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2767 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2768 list_del(&pa->pa_group_list);
2770 kmem_cache_free(ext4_pspace_cachep, pa);
2773 mb_debug("mballoc: %u PAs left\n", count);
2777 int ext4_mb_release(struct super_block *sb)
2780 int num_meta_group_infos;
2781 struct ext4_group_info *grinfo;
2782 struct ext4_sb_info *sbi = EXT4_SB(sb);
2784 if (sbi->s_group_info) {
2785 for (i = 0; i < sbi->s_groups_count; i++) {
2786 grinfo = ext4_get_group_info(sb, i);
2788 kfree(grinfo->bb_bitmap);
2790 ext4_lock_group(sb, i);
2791 ext4_mb_cleanup_pa(grinfo);
2792 ext4_unlock_group(sb, i);
2795 num_meta_group_infos = (sbi->s_groups_count +
2796 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2797 EXT4_DESC_PER_BLOCK_BITS(sb);
2798 for (i = 0; i < num_meta_group_infos; i++)
2799 kfree(sbi->s_group_info[i]);
2800 kfree(sbi->s_group_info);
2802 kfree(sbi->s_mb_offsets);
2803 kfree(sbi->s_mb_maxs);
2804 if (sbi->s_buddy_cache)
2805 iput(sbi->s_buddy_cache);
2806 if (sbi->s_mb_stats) {
2808 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2809 atomic_read(&sbi->s_bal_allocated),
2810 atomic_read(&sbi->s_bal_reqs),
2811 atomic_read(&sbi->s_bal_success));
2813 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2814 "%u 2^N hits, %u breaks, %u lost\n",
2815 atomic_read(&sbi->s_bal_ex_scanned),
2816 atomic_read(&sbi->s_bal_goals),
2817 atomic_read(&sbi->s_bal_2orders),
2818 atomic_read(&sbi->s_bal_breaks),
2819 atomic_read(&sbi->s_mb_lost_chunks));
2821 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2822 sbi->s_mb_buddies_generated++,
2823 sbi->s_mb_generation_time);
2825 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2826 atomic_read(&sbi->s_mb_preallocated),
2827 atomic_read(&sbi->s_mb_discarded));
2830 free_percpu(sbi->s_locality_groups);
2831 ext4_mb_history_release(sb);
2832 ext4_mb_destroy_per_dev_proc(sb);
2838 * This function is called by the jbd2 layer once the commit has finished,
2839 * so we know we can free the blocks that were released with that commit.
2841 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2843 struct super_block *sb = journal->j_private;
2844 struct ext4_buddy e4b;
2845 struct ext4_group_info *db;
2846 int err, count = 0, count2 = 0;
2847 struct ext4_free_data *entry;
2848 ext4_fsblk_t discard_block;
2849 struct list_head *l, *ltmp;
2851 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2852 entry = list_entry(l, struct ext4_free_data, list);
2854 mb_debug("gonna free %u blocks in group %u (0x%p):",
2855 entry->count, entry->group, entry);
2857 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2858 /* we expect to find existing buddy because it's pinned */
2862 /* there are blocks to put in buddy to make them really free */
2863 count += entry->count;
2865 ext4_lock_group(sb, entry->group);
2866 /* Take it out of per group rb tree */
2867 rb_erase(&entry->node, &(db->bb_free_root));
2868 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2870 if (!db->bb_free_root.rb_node) {
2871 /* No more items in the per group rb tree
2872 * balance refcounts from ext4_mb_free_metadata()
2874 page_cache_release(e4b.bd_buddy_page);
2875 page_cache_release(e4b.bd_bitmap_page);
2877 ext4_unlock_group(sb, entry->group);
2878 discard_block = (ext4_fsblk_t) entry->group * EXT4_BLOCKS_PER_GROUP(sb)
2880 + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
2881 trace_mark(ext4_discard_blocks, "dev %s blk %llu count %u", sb->s_id,
2882 (unsigned long long) discard_block, entry->count);
2883 sb_issue_discard(sb, discard_block, entry->count);
2885 kmem_cache_free(ext4_free_ext_cachep, entry);
2886 ext4_mb_release_desc(&e4b);
2889 mb_debug("freed %u blocks in %u structures\n", count, count2);
2892 #define EXT4_MB_STATS_NAME "stats"
2893 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2894 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2895 #define EXT4_MB_ORDER2_REQ "order2_req"
2896 #define EXT4_MB_STREAM_REQ "stream_req"
2897 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2899 static int ext4_mb_init_per_dev_proc(struct super_block *sb)
2901 #ifdef CONFIG_PROC_FS
2902 mode_t mode = S_IFREG | S_IRUGO | S_IWUSR;
2903 struct ext4_sb_info *sbi = EXT4_SB(sb);
2904 struct proc_dir_entry *proc;
2906 if (sbi->s_proc == NULL)
2909 EXT4_PROC_HANDLER(EXT4_MB_STATS_NAME, mb_stats);
2910 EXT4_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME, mb_max_to_scan);
2911 EXT4_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME, mb_min_to_scan);
2912 EXT4_PROC_HANDLER(EXT4_MB_ORDER2_REQ, mb_order2_reqs);
2913 EXT4_PROC_HANDLER(EXT4_MB_STREAM_REQ, mb_stream_request);
2914 EXT4_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC, mb_group_prealloc);
2918 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_proc);
2919 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_proc);
2920 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_proc);
2921 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_proc);
2922 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_proc);
2923 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_proc);
2930 static int ext4_mb_destroy_per_dev_proc(struct super_block *sb)
2932 #ifdef CONFIG_PROC_FS
2933 struct ext4_sb_info *sbi = EXT4_SB(sb);
2935 if (sbi->s_proc == NULL)
2938 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_proc);
2939 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_proc);
2940 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_proc);
2941 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_proc);
2942 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_proc);
2943 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_proc);
2948 int __init init_ext4_mballoc(void)
2950 ext4_pspace_cachep =
2951 kmem_cache_create("ext4_prealloc_space",
2952 sizeof(struct ext4_prealloc_space),
2953 0, SLAB_RECLAIM_ACCOUNT, NULL);
2954 if (ext4_pspace_cachep == NULL)
2958 kmem_cache_create("ext4_alloc_context",
2959 sizeof(struct ext4_allocation_context),
2960 0, SLAB_RECLAIM_ACCOUNT, NULL);
2961 if (ext4_ac_cachep == NULL) {
2962 kmem_cache_destroy(ext4_pspace_cachep);
2966 ext4_free_ext_cachep =
2967 kmem_cache_create("ext4_free_block_extents",
2968 sizeof(struct ext4_free_data),
2969 0, SLAB_RECLAIM_ACCOUNT, NULL);
2970 if (ext4_free_ext_cachep == NULL) {
2971 kmem_cache_destroy(ext4_pspace_cachep);
2972 kmem_cache_destroy(ext4_ac_cachep);
2978 void exit_ext4_mballoc(void)
2980 /* XXX: synchronize_rcu(); */
2981 kmem_cache_destroy(ext4_pspace_cachep);
2982 kmem_cache_destroy(ext4_ac_cachep);
2983 kmem_cache_destroy(ext4_free_ext_cachep);
2988 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2989 * Returns 0 if success or error code
2991 static noinline_for_stack int
2992 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2993 handle_t *handle, unsigned int reserv_blks)
2995 struct buffer_head *bitmap_bh = NULL;
2996 struct ext4_super_block *es;
2997 struct ext4_group_desc *gdp;
2998 struct buffer_head *gdp_bh;
2999 struct ext4_sb_info *sbi;
3000 struct super_block *sb;
3004 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3005 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3013 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3017 err = ext4_journal_get_write_access(handle, bitmap_bh);
3022 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3026 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3027 gdp->bg_free_blocks_count);
3029 err = ext4_journal_get_write_access(handle, gdp_bh);
3033 block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
3034 + ac->ac_b_ex.fe_start
3035 + le32_to_cpu(es->s_first_data_block);
3037 len = ac->ac_b_ex.fe_len;
3038 if (in_range(ext4_block_bitmap(sb, gdp), block, len) ||
3039 in_range(ext4_inode_bitmap(sb, gdp), block, len) ||
3040 in_range(block, ext4_inode_table(sb, gdp),
3041 EXT4_SB(sb)->s_itb_per_group) ||
3042 in_range(block + len - 1, ext4_inode_table(sb, gdp),
3043 EXT4_SB(sb)->s_itb_per_group)) {
3044 ext4_error(sb, __func__,
3045 "Allocating block %llu in system zone of %d group\n",
3046 block, ac->ac_b_ex.fe_group);
3047 /* File system mounted not to panic on error
3048 * Fix the bitmap and repeat the block allocation
3049 * We leak some of the blocks here.
3051 mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group),
3052 bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3053 ac->ac_b_ex.fe_len);
3054 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3059 #ifdef AGGRESSIVE_CHECK
3062 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3063 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3064 bitmap_bh->b_data));
3068 spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
3069 mb_set_bits(NULL, bitmap_bh->b_data,
3070 ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
3071 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
3072 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3073 ext4_free_blks_set(sb, gdp,
3074 ext4_free_blocks_after_init(sb,
3075 ac->ac_b_ex.fe_group, gdp));
3077 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
3078 ext4_free_blks_set(sb, gdp, len);
3079 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
3080 spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
3081 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
3083 * Now reduce the dirty block count also. Should not go negative
3085 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3086 /* release all the reserved blocks if non delalloc */
3087 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
3089 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
3090 ac->ac_b_ex.fe_len);
3092 if (sbi->s_log_groups_per_flex) {
3093 ext4_group_t flex_group = ext4_flex_group(sbi,
3094 ac->ac_b_ex.fe_group);
3095 spin_lock(sb_bgl_lock(sbi, flex_group));
3096 sbi->s_flex_groups[flex_group].free_blocks -= ac->ac_b_ex.fe_len;
3097 spin_unlock(sb_bgl_lock(sbi, flex_group));
3100 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3103 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3112 * here we normalize request for locality group
3113 * Group request are normalized to s_strip size if we set the same via mount
3114 * option. If not we set it to s_mb_group_prealloc which can be configured via
3115 * /proc/fs/ext4/<partition>/group_prealloc
3117 * XXX: should we try to preallocate more than the group has now?
3119 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3121 struct super_block *sb = ac->ac_sb;
3122 struct ext4_locality_group *lg = ac->ac_lg;
3125 if (EXT4_SB(sb)->s_stripe)
3126 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
3128 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3129 mb_debug("#%u: goal %u blocks for locality group\n",
3130 current->pid, ac->ac_g_ex.fe_len);
3134 * Normalization means making request better in terms of
3135 * size and alignment
3137 static noinline_for_stack void
3138 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3139 struct ext4_allocation_request *ar)
3143 loff_t size, orig_size, start_off;
3144 ext4_lblk_t start, orig_start;
3145 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3146 struct ext4_prealloc_space *pa;
3148 /* do normalize only data requests, metadata requests
3149 do not need preallocation */
3150 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3153 /* sometime caller may want exact blocks */
3154 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3157 /* caller may indicate that preallocation isn't
3158 * required (it's a tail, for example) */
3159 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3162 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3163 ext4_mb_normalize_group_request(ac);
3167 bsbits = ac->ac_sb->s_blocksize_bits;
3169 /* first, let's learn actual file size
3170 * given current request is allocated */
3171 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3172 size = size << bsbits;
3173 if (size < i_size_read(ac->ac_inode))
3174 size = i_size_read(ac->ac_inode);
3176 /* max size of free chunks */
3179 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3180 (req <= (size) || max <= (chunk_size))
3182 /* first, try to predict filesize */
3183 /* XXX: should this table be tunable? */
3185 if (size <= 16 * 1024) {
3187 } else if (size <= 32 * 1024) {
3189 } else if (size <= 64 * 1024) {
3191 } else if (size <= 128 * 1024) {
3193 } else if (size <= 256 * 1024) {
3195 } else if (size <= 512 * 1024) {
3197 } else if (size <= 1024 * 1024) {
3199 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3200 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3201 (21 - bsbits)) << 21;
3202 size = 2 * 1024 * 1024;
3203 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3204 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3205 (22 - bsbits)) << 22;
3206 size = 4 * 1024 * 1024;
3207 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3208 (8<<20)>>bsbits, max, 8 * 1024)) {
3209 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3210 (23 - bsbits)) << 23;
3211 size = 8 * 1024 * 1024;
3213 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3214 size = ac->ac_o_ex.fe_len << bsbits;
3216 orig_size = size = size >> bsbits;
3217 orig_start = start = start_off >> bsbits;
3219 /* don't cover already allocated blocks in selected range */
3220 if (ar->pleft && start <= ar->lleft) {
3221 size -= ar->lleft + 1 - start;
3222 start = ar->lleft + 1;
3224 if (ar->pright && start + size - 1 >= ar->lright)
3225 size -= start + size - ar->lright;
3229 /* check we don't cross already preallocated blocks */
3231 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3236 spin_lock(&pa->pa_lock);
3237 if (pa->pa_deleted) {
3238 spin_unlock(&pa->pa_lock);
3242 pa_end = pa->pa_lstart + pa->pa_len;
3244 /* PA must not overlap original request */
3245 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3246 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3248 /* skip PA normalized request doesn't overlap with */
3249 if (pa->pa_lstart >= end) {
3250 spin_unlock(&pa->pa_lock);
3253 if (pa_end <= start) {
3254 spin_unlock(&pa->pa_lock);
3257 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3259 if (pa_end <= ac->ac_o_ex.fe_logical) {
3260 BUG_ON(pa_end < start);
3264 if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3265 BUG_ON(pa->pa_lstart > end);
3266 end = pa->pa_lstart;
3268 spin_unlock(&pa->pa_lock);
3273 /* XXX: extra loop to check we really don't overlap preallocations */
3275 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3277 spin_lock(&pa->pa_lock);
3278 if (pa->pa_deleted == 0) {
3279 pa_end = pa->pa_lstart + pa->pa_len;
3280 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3282 spin_unlock(&pa->pa_lock);
3286 if (start + size <= ac->ac_o_ex.fe_logical &&
3287 start > ac->ac_o_ex.fe_logical) {
3288 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3289 (unsigned long) start, (unsigned long) size,
3290 (unsigned long) ac->ac_o_ex.fe_logical);
3292 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3293 start > ac->ac_o_ex.fe_logical);
3294 BUG_ON(size <= 0 || size >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3296 /* now prepare goal request */
3298 /* XXX: is it better to align blocks WRT to logical
3299 * placement or satisfy big request as is */
3300 ac->ac_g_ex.fe_logical = start;
3301 ac->ac_g_ex.fe_len = size;
3303 /* define goal start in order to merge */
3304 if (ar->pright && (ar->lright == (start + size))) {
3305 /* merge to the right */
3306 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3307 &ac->ac_f_ex.fe_group,
3308 &ac->ac_f_ex.fe_start);
3309 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3311 if (ar->pleft && (ar->lleft + 1 == start)) {
3312 /* merge to the left */
3313 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3314 &ac->ac_f_ex.fe_group,
3315 &ac->ac_f_ex.fe_start);
3316 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3319 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
3320 (unsigned) orig_size, (unsigned) start);
3323 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3325 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3327 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3328 atomic_inc(&sbi->s_bal_reqs);
3329 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3330 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3331 atomic_inc(&sbi->s_bal_success);
3332 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3333 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3334 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3335 atomic_inc(&sbi->s_bal_goals);
3336 if (ac->ac_found > sbi->s_mb_max_to_scan)
3337 atomic_inc(&sbi->s_bal_breaks);
3340 ext4_mb_store_history(ac);
3344 * use blocks preallocated to inode
3346 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3347 struct ext4_prealloc_space *pa)
3353 /* found preallocated blocks, use them */
3354 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3355 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3357 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3358 &ac->ac_b_ex.fe_start);
3359 ac->ac_b_ex.fe_len = len;
3360 ac->ac_status = AC_STATUS_FOUND;
3363 BUG_ON(start < pa->pa_pstart);
3364 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3365 BUG_ON(pa->pa_free < len);
3368 mb_debug("use %llu/%u from inode pa %p\n", start, len, pa);
3372 * use blocks preallocated to locality group
3374 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3375 struct ext4_prealloc_space *pa)
3377 unsigned int len = ac->ac_o_ex.fe_len;
3379 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3380 &ac->ac_b_ex.fe_group,
3381 &ac->ac_b_ex.fe_start);
3382 ac->ac_b_ex.fe_len = len;
3383 ac->ac_status = AC_STATUS_FOUND;
3386 /* we don't correct pa_pstart or pa_plen here to avoid
3387 * possible race when the group is being loaded concurrently
3388 * instead we correct pa later, after blocks are marked
3389 * in on-disk bitmap -- see ext4_mb_release_context()
3390 * Other CPUs are prevented from allocating from this pa by lg_mutex
3392 mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3396 * Return the prealloc space that have minimal distance
3397 * from the goal block. @cpa is the prealloc
3398 * space that is having currently known minimal distance
3399 * from the goal block.
3401 static struct ext4_prealloc_space *
3402 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3403 struct ext4_prealloc_space *pa,
3404 struct ext4_prealloc_space *cpa)
3406 ext4_fsblk_t cur_distance, new_distance;
3409 atomic_inc(&pa->pa_count);
3412 cur_distance = abs(goal_block - cpa->pa_pstart);
3413 new_distance = abs(goal_block - pa->pa_pstart);
3415 if (cur_distance < new_distance)
3418 /* drop the previous reference */
3419 atomic_dec(&cpa->pa_count);
3420 atomic_inc(&pa->pa_count);
3425 * search goal blocks in preallocated space
3427 static noinline_for_stack int
3428 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3431 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3432 struct ext4_locality_group *lg;
3433 struct ext4_prealloc_space *pa, *cpa = NULL;
3434 ext4_fsblk_t goal_block;
3436 /* only data can be preallocated */
3437 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3440 /* first, try per-file preallocation */
3442 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3444 /* all fields in this condition don't change,
3445 * so we can skip locking for them */
3446 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3447 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3450 /* found preallocated blocks, use them */
3451 spin_lock(&pa->pa_lock);
3452 if (pa->pa_deleted == 0 && pa->pa_free) {
3453 atomic_inc(&pa->pa_count);
3454 ext4_mb_use_inode_pa(ac, pa);
3455 spin_unlock(&pa->pa_lock);
3456 ac->ac_criteria = 10;
3460 spin_unlock(&pa->pa_lock);
3464 /* can we use group allocation? */
3465 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3468 /* inode may have no locality group for some reason */
3472 order = fls(ac->ac_o_ex.fe_len) - 1;
3473 if (order > PREALLOC_TB_SIZE - 1)
3474 /* The max size of hash table is PREALLOC_TB_SIZE */
3475 order = PREALLOC_TB_SIZE - 1;
3477 goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) +
3478 ac->ac_g_ex.fe_start +
3479 le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block);
3481 * search for the prealloc space that is having
3482 * minimal distance from the goal block.
3484 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3486 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3488 spin_lock(&pa->pa_lock);
3489 if (pa->pa_deleted == 0 &&
3490 pa->pa_free >= ac->ac_o_ex.fe_len) {
3492 cpa = ext4_mb_check_group_pa(goal_block,
3495 spin_unlock(&pa->pa_lock);
3500 ext4_mb_use_group_pa(ac, cpa);
3501 ac->ac_criteria = 20;
3508 * the function goes through all block freed in the group
3509 * but not yet committed and marks them used in in-core bitmap.
3510 * buddy must be generated from this bitmap
3511 * Need to be called with ext4 group lock (ext4_lock_group)
3513 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3517 struct ext4_group_info *grp;
3518 struct ext4_free_data *entry;
3520 grp = ext4_get_group_info(sb, group);
3521 n = rb_first(&(grp->bb_free_root));
3524 entry = rb_entry(n, struct ext4_free_data, node);
3525 mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
3526 bitmap, entry->start_blk,
3534 * the function goes through all preallocation in this group and marks them
3535 * used in in-core bitmap. buddy must be generated from this bitmap
3536 * Need to be called with ext4 group lock (ext4_lock_group)
3538 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3541 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3542 struct ext4_prealloc_space *pa;
3543 struct list_head *cur;
3544 ext4_group_t groupnr;
3545 ext4_grpblk_t start;
3546 int preallocated = 0;
3550 /* all form of preallocation discards first load group,
3551 * so the only competing code is preallocation use.
3552 * we don't need any locking here
3553 * notice we do NOT ignore preallocations with pa_deleted
3554 * otherwise we could leave used blocks available for
3555 * allocation in buddy when concurrent ext4_mb_put_pa()
3556 * is dropping preallocation
3558 list_for_each(cur, &grp->bb_prealloc_list) {
3559 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3560 spin_lock(&pa->pa_lock);
3561 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3564 spin_unlock(&pa->pa_lock);
3565 if (unlikely(len == 0))
3567 BUG_ON(groupnr != group);
3568 mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
3569 bitmap, start, len);
3570 preallocated += len;
3573 mb_debug("prellocated %u for group %u\n", preallocated, group);
3576 static void ext4_mb_pa_callback(struct rcu_head *head)
3578 struct ext4_prealloc_space *pa;
3579 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3580 kmem_cache_free(ext4_pspace_cachep, pa);
3584 * drops a reference to preallocated space descriptor
3585 * if this was the last reference and the space is consumed
3587 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3588 struct super_block *sb, struct ext4_prealloc_space *pa)
3592 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3595 /* in this short window concurrent discard can set pa_deleted */
3596 spin_lock(&pa->pa_lock);
3597 if (pa->pa_deleted == 1) {
3598 spin_unlock(&pa->pa_lock);
3603 spin_unlock(&pa->pa_lock);
3605 /* -1 is to protect from crossing allocation group */
3606 ext4_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL);
3611 * P1 (buddy init) P2 (regular allocation)
3612 * find block B in PA
3613 * copy on-disk bitmap to buddy
3614 * mark B in on-disk bitmap
3615 * drop PA from group
3616 * mark all PAs in buddy
3618 * thus, P1 initializes buddy with B available. to prevent this
3619 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3622 ext4_lock_group(sb, grp);
3623 list_del(&pa->pa_group_list);
3624 ext4_unlock_group(sb, grp);
3626 spin_lock(pa->pa_obj_lock);
3627 list_del_rcu(&pa->pa_inode_list);
3628 spin_unlock(pa->pa_obj_lock);
3630 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3634 * creates new preallocated space for given inode
3636 static noinline_for_stack int
3637 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3639 struct super_block *sb = ac->ac_sb;
3640 struct ext4_prealloc_space *pa;
3641 struct ext4_group_info *grp;
3642 struct ext4_inode_info *ei;
3644 /* preallocate only when found space is larger then requested */
3645 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3646 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3647 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3649 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3653 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3659 /* we can't allocate as much as normalizer wants.
3660 * so, found space must get proper lstart
3661 * to cover original request */
3662 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3663 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3665 /* we're limited by original request in that
3666 * logical block must be covered any way
3667 * winl is window we can move our chunk within */
3668 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3670 /* also, we should cover whole original request */
3671 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3673 /* the smallest one defines real window */
3674 win = min(winl, wins);
3676 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3677 if (offs && offs < win)
3680 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3681 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3682 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3685 /* preallocation can change ac_b_ex, thus we store actually
3686 * allocated blocks for history */
3687 ac->ac_f_ex = ac->ac_b_ex;
3689 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3690 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3691 pa->pa_len = ac->ac_b_ex.fe_len;
3692 pa->pa_free = pa->pa_len;
3693 atomic_set(&pa->pa_count, 1);
3694 spin_lock_init(&pa->pa_lock);
3698 mb_debug("new inode pa %p: %llu/%u for %u\n", pa,
3699 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3701 ext4_mb_use_inode_pa(ac, pa);
3702 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3704 ei = EXT4_I(ac->ac_inode);
3705 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3707 pa->pa_obj_lock = &ei->i_prealloc_lock;
3708 pa->pa_inode = ac->ac_inode;
3710 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3711 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3712 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3714 spin_lock(pa->pa_obj_lock);
3715 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3716 spin_unlock(pa->pa_obj_lock);
3722 * creates new preallocated space for locality group inodes belongs to
3724 static noinline_for_stack int
3725 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3727 struct super_block *sb = ac->ac_sb;
3728 struct ext4_locality_group *lg;
3729 struct ext4_prealloc_space *pa;
3730 struct ext4_group_info *grp;
3732 /* preallocate only when found space is larger then requested */
3733 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3734 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3735 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3737 BUG_ON(ext4_pspace_cachep == NULL);
3738 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3742 /* preallocation can change ac_b_ex, thus we store actually
3743 * allocated blocks for history */
3744 ac->ac_f_ex = ac->ac_b_ex;
3746 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3747 pa->pa_lstart = pa->pa_pstart;
3748 pa->pa_len = ac->ac_b_ex.fe_len;
3749 pa->pa_free = pa->pa_len;
3750 atomic_set(&pa->pa_count, 1);
3751 spin_lock_init(&pa->pa_lock);
3752 INIT_LIST_HEAD(&pa->pa_inode_list);
3756 mb_debug("new group pa %p: %llu/%u for %u\n", pa,
3757 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3759 ext4_mb_use_group_pa(ac, pa);
3760 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3762 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3766 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3767 pa->pa_inode = NULL;
3769 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3770 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3771 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3774 * We will later add the new pa to the right bucket
3775 * after updating the pa_free in ext4_mb_release_context
3780 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3784 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3785 err = ext4_mb_new_group_pa(ac);
3787 err = ext4_mb_new_inode_pa(ac);
3792 * finds all unused blocks in on-disk bitmap, frees them in
3793 * in-core bitmap and buddy.
3794 * @pa must be unlinked from inode and group lists, so that
3795 * nobody else can find/use it.
3796 * the caller MUST hold group/inode locks.
3797 * TODO: optimize the case when there are no in-core structures yet
3799 static noinline_for_stack int
3800 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3801 struct ext4_prealloc_space *pa,
3802 struct ext4_allocation_context *ac)
3804 struct super_block *sb = e4b->bd_sb;
3805 struct ext4_sb_info *sbi = EXT4_SB(sb);
3814 BUG_ON(pa->pa_deleted == 0);
3815 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3816 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3817 end = bit + pa->pa_len;
3821 ac->ac_inode = pa->pa_inode;
3822 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3826 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3829 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3830 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3831 le32_to_cpu(sbi->s_es->s_first_data_block);
3832 mb_debug(" free preallocated %u/%u in group %u\n",
3833 (unsigned) start, (unsigned) next - bit,
3838 ac->ac_b_ex.fe_group = group;
3839 ac->ac_b_ex.fe_start = bit;
3840 ac->ac_b_ex.fe_len = next - bit;
3841 ac->ac_b_ex.fe_logical = 0;
3842 ext4_mb_store_history(ac);
3845 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3848 if (free != pa->pa_free) {
3849 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3850 pa, (unsigned long) pa->pa_lstart,
3851 (unsigned long) pa->pa_pstart,
3852 (unsigned long) pa->pa_len);
3853 ext4_grp_locked_error(sb, group,
3854 __func__, "free %u, pa_free %u",
3857 * pa is already deleted so we use the value obtained
3858 * from the bitmap and continue.
3861 atomic_add(free, &sbi->s_mb_discarded);
3866 static noinline_for_stack int
3867 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3868 struct ext4_prealloc_space *pa,
3869 struct ext4_allocation_context *ac)
3871 struct super_block *sb = e4b->bd_sb;
3876 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3878 BUG_ON(pa->pa_deleted == 0);
3879 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3880 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3881 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3882 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3886 ac->ac_inode = NULL;
3887 ac->ac_b_ex.fe_group = group;
3888 ac->ac_b_ex.fe_start = bit;
3889 ac->ac_b_ex.fe_len = pa->pa_len;
3890 ac->ac_b_ex.fe_logical = 0;
3891 ext4_mb_store_history(ac);
3898 * releases all preallocations in given group
3900 * first, we need to decide discard policy:
3901 * - when do we discard
3903 * - how many do we discard
3904 * 1) how many requested
3906 static noinline_for_stack int
3907 ext4_mb_discard_group_preallocations(struct super_block *sb,
3908 ext4_group_t group, int needed)
3910 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3911 struct buffer_head *bitmap_bh = NULL;
3912 struct ext4_prealloc_space *pa, *tmp;
3913 struct ext4_allocation_context *ac;
3914 struct list_head list;
3915 struct ext4_buddy e4b;
3920 mb_debug("discard preallocation for group %u\n", group);
3922 if (list_empty(&grp->bb_prealloc_list))
3925 bitmap_bh = ext4_read_block_bitmap(sb, group);
3926 if (bitmap_bh == NULL) {
3927 ext4_error(sb, __func__, "Error in reading block "
3928 "bitmap for %u", group);
3932 err = ext4_mb_load_buddy(sb, group, &e4b);
3934 ext4_error(sb, __func__, "Error in loading buddy "
3935 "information for %u", group);
3941 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3943 INIT_LIST_HEAD(&list);
3944 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3946 ext4_lock_group(sb, group);
3947 list_for_each_entry_safe(pa, tmp,
3948 &grp->bb_prealloc_list, pa_group_list) {
3949 spin_lock(&pa->pa_lock);
3950 if (atomic_read(&pa->pa_count)) {
3951 spin_unlock(&pa->pa_lock);
3955 if (pa->pa_deleted) {
3956 spin_unlock(&pa->pa_lock);
3960 /* seems this one can be freed ... */
3963 /* we can trust pa_free ... */
3964 free += pa->pa_free;
3966 spin_unlock(&pa->pa_lock);
3968 list_del(&pa->pa_group_list);
3969 list_add(&pa->u.pa_tmp_list, &list);
3972 /* if we still need more blocks and some PAs were used, try again */
3973 if (free < needed && busy) {
3975 ext4_unlock_group(sb, group);
3977 * Yield the CPU here so that we don't get soft lockup
3978 * in non preempt case.
3984 /* found anything to free? */
3985 if (list_empty(&list)) {
3990 /* now free all selected PAs */
3991 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3993 /* remove from object (inode or locality group) */
3994 spin_lock(pa->pa_obj_lock);
3995 list_del_rcu(&pa->pa_inode_list);
3996 spin_unlock(pa->pa_obj_lock);
3999 ext4_mb_release_group_pa(&e4b, pa, ac);
4001 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
4003 list_del(&pa->u.pa_tmp_list);
4004 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4008 ext4_unlock_group(sb, group);
4010 kmem_cache_free(ext4_ac_cachep, ac);
4011 ext4_mb_release_desc(&e4b);
4017 * releases all non-used preallocated blocks for given inode
4019 * It's important to discard preallocations under i_data_sem
4020 * We don't want another block to be served from the prealloc
4021 * space when we are discarding the inode prealloc space.
4023 * FIXME!! Make sure it is valid at all the call sites
4025 void ext4_discard_preallocations(struct inode *inode)
4027 struct ext4_inode_info *ei = EXT4_I(inode);
4028 struct super_block *sb = inode->i_sb;
4029 struct buffer_head *bitmap_bh = NULL;
4030 struct ext4_prealloc_space *pa, *tmp;
4031 struct ext4_allocation_context *ac;
4032 ext4_group_t group = 0;
4033 struct list_head list;
4034 struct ext4_buddy e4b;
4037 if (!S_ISREG(inode->i_mode)) {
4038 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4042 mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
4044 INIT_LIST_HEAD(&list);
4046 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4048 /* first, collect all pa's in the inode */
4049 spin_lock(&ei->i_prealloc_lock);
4050 while (!list_empty(&ei->i_prealloc_list)) {
4051 pa = list_entry(ei->i_prealloc_list.next,
4052 struct ext4_prealloc_space, pa_inode_list);
4053 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4054 spin_lock(&pa->pa_lock);
4055 if (atomic_read(&pa->pa_count)) {
4056 /* this shouldn't happen often - nobody should
4057 * use preallocation while we're discarding it */
4058 spin_unlock(&pa->pa_lock);
4059 spin_unlock(&ei->i_prealloc_lock);
4060 printk(KERN_ERR "uh-oh! used pa while discarding\n");
4062 schedule_timeout_uninterruptible(HZ);
4066 if (pa->pa_deleted == 0) {
4068 spin_unlock(&pa->pa_lock);
4069 list_del_rcu(&pa->pa_inode_list);
4070 list_add(&pa->u.pa_tmp_list, &list);
4074 /* someone is deleting pa right now */
4075 spin_unlock(&pa->pa_lock);
4076 spin_unlock(&ei->i_prealloc_lock);
4078 /* we have to wait here because pa_deleted
4079 * doesn't mean pa is already unlinked from
4080 * the list. as we might be called from
4081 * ->clear_inode() the inode will get freed
4082 * and concurrent thread which is unlinking
4083 * pa from inode's list may access already
4084 * freed memory, bad-bad-bad */
4086 /* XXX: if this happens too often, we can
4087 * add a flag to force wait only in case
4088 * of ->clear_inode(), but not in case of
4089 * regular truncate */
4090 schedule_timeout_uninterruptible(HZ);
4093 spin_unlock(&ei->i_prealloc_lock);
4095 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4096 BUG_ON(pa->pa_linear != 0);
4097 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4099 err = ext4_mb_load_buddy(sb, group, &e4b);
4101 ext4_error(sb, __func__, "Error in loading buddy "
4102 "information for %u", group);
4106 bitmap_bh = ext4_read_block_bitmap(sb, group);
4107 if (bitmap_bh == NULL) {
4108 ext4_error(sb, __func__, "Error in reading block "
4109 "bitmap for %u", group);
4110 ext4_mb_release_desc(&e4b);
4114 ext4_lock_group(sb, group);
4115 list_del(&pa->pa_group_list);
4116 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
4117 ext4_unlock_group(sb, group);
4119 ext4_mb_release_desc(&e4b);
4122 list_del(&pa->u.pa_tmp_list);
4123 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4126 kmem_cache_free(ext4_ac_cachep, ac);
4130 * finds all preallocated spaces and return blocks being freed to them
4131 * if preallocated space becomes full (no block is used from the space)
4132 * then the function frees space in buddy
4133 * XXX: at the moment, truncate (which is the only way to free blocks)
4134 * discards all preallocations
4136 static void ext4_mb_return_to_preallocation(struct inode *inode,
4137 struct ext4_buddy *e4b,
4138 sector_t block, int count)
4140 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
4143 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4145 struct super_block *sb = ac->ac_sb;
4148 printk(KERN_ERR "EXT4-fs: Can't allocate:"
4149 " Allocation context details:\n");
4150 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
4151 ac->ac_status, ac->ac_flags);
4152 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4153 "best %lu/%lu/%lu@%lu cr %d\n",
4154 (unsigned long)ac->ac_o_ex.fe_group,
4155 (unsigned long)ac->ac_o_ex.fe_start,
4156 (unsigned long)ac->ac_o_ex.fe_len,
4157 (unsigned long)ac->ac_o_ex.fe_logical,
4158 (unsigned long)ac->ac_g_ex.fe_group,
4159 (unsigned long)ac->ac_g_ex.fe_start,
4160 (unsigned long)ac->ac_g_ex.fe_len,
4161 (unsigned long)ac->ac_g_ex.fe_logical,
4162 (unsigned long)ac->ac_b_ex.fe_group,
4163 (unsigned long)ac->ac_b_ex.fe_start,
4164 (unsigned long)ac->ac_b_ex.fe_len,
4165 (unsigned long)ac->ac_b_ex.fe_logical,
4166 (int)ac->ac_criteria);
4167 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
4169 printk(KERN_ERR "EXT4-fs: groups: \n");
4170 for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
4171 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4172 struct ext4_prealloc_space *pa;
4173 ext4_grpblk_t start;
4174 struct list_head *cur;
4175 ext4_lock_group(sb, i);
4176 list_for_each(cur, &grp->bb_prealloc_list) {
4177 pa = list_entry(cur, struct ext4_prealloc_space,
4179 spin_lock(&pa->pa_lock);
4180 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4182 spin_unlock(&pa->pa_lock);
4183 printk(KERN_ERR "PA:%lu:%d:%u \n", i,
4186 ext4_unlock_group(sb, i);
4188 if (grp->bb_free == 0)
4190 printk(KERN_ERR "%lu: %d/%d \n",
4191 i, grp->bb_free, grp->bb_fragments);
4193 printk(KERN_ERR "\n");
4196 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4203 * We use locality group preallocation for small size file. The size of the
4204 * file is determined by the current size or the resulting size after
4205 * allocation which ever is larger
4207 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
4209 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4211 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4212 int bsbits = ac->ac_sb->s_blocksize_bits;
4215 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4218 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4219 isize = i_size_read(ac->ac_inode) >> bsbits;
4220 size = max(size, isize);
4222 /* don't use group allocation for large files */
4223 if (size >= sbi->s_mb_stream_request)
4226 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4229 BUG_ON(ac->ac_lg != NULL);
4231 * locality group prealloc space are per cpu. The reason for having
4232 * per cpu locality group is to reduce the contention between block
4233 * request from multiple CPUs.
4235 ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id());
4237 /* we're going to use group allocation */
4238 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4240 /* serialize all allocations in the group */
4241 mutex_lock(&ac->ac_lg->lg_mutex);
4244 static noinline_for_stack int
4245 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4246 struct ext4_allocation_request *ar)
4248 struct super_block *sb = ar->inode->i_sb;
4249 struct ext4_sb_info *sbi = EXT4_SB(sb);
4250 struct ext4_super_block *es = sbi->s_es;
4254 ext4_grpblk_t block;
4256 /* we can't allocate > group size */
4259 /* just a dirty hack to filter too big requests */
4260 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4261 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4263 /* start searching from the goal */
4265 if (goal < le32_to_cpu(es->s_first_data_block) ||
4266 goal >= ext4_blocks_count(es))
4267 goal = le32_to_cpu(es->s_first_data_block);
4268 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4270 /* set up allocation goals */
4271 ac->ac_b_ex.fe_logical = ar->logical;
4272 ac->ac_b_ex.fe_group = 0;
4273 ac->ac_b_ex.fe_start = 0;
4274 ac->ac_b_ex.fe_len = 0;
4275 ac->ac_status = AC_STATUS_CONTINUE;
4276 ac->ac_groups_scanned = 0;
4277 ac->ac_ex_scanned = 0;
4280 ac->ac_inode = ar->inode;
4281 ac->ac_o_ex.fe_logical = ar->logical;
4282 ac->ac_o_ex.fe_group = group;
4283 ac->ac_o_ex.fe_start = block;
4284 ac->ac_o_ex.fe_len = len;
4285 ac->ac_g_ex.fe_logical = ar->logical;
4286 ac->ac_g_ex.fe_group = group;
4287 ac->ac_g_ex.fe_start = block;
4288 ac->ac_g_ex.fe_len = len;
4289 ac->ac_f_ex.fe_len = 0;
4290 ac->ac_flags = ar->flags;
4292 ac->ac_criteria = 0;
4294 ac->ac_bitmap_page = NULL;
4295 ac->ac_buddy_page = NULL;
4296 ac->alloc_semp = NULL;
4299 /* we have to define context: we'll we work with a file or
4300 * locality group. this is a policy, actually */
4301 ext4_mb_group_or_file(ac);
4303 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4304 "left: %u/%u, right %u/%u to %swritable\n",
4305 (unsigned) ar->len, (unsigned) ar->logical,
4306 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4307 (unsigned) ar->lleft, (unsigned) ar->pleft,
4308 (unsigned) ar->lright, (unsigned) ar->pright,
4309 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4314 static noinline_for_stack void
4315 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4316 struct ext4_locality_group *lg,
4317 int order, int total_entries)
4319 ext4_group_t group = 0;
4320 struct ext4_buddy e4b;
4321 struct list_head discard_list;
4322 struct ext4_prealloc_space *pa, *tmp;
4323 struct ext4_allocation_context *ac;
4325 mb_debug("discard locality group preallocation\n");
4327 INIT_LIST_HEAD(&discard_list);
4328 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4330 spin_lock(&lg->lg_prealloc_lock);
4331 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4333 spin_lock(&pa->pa_lock);
4334 if (atomic_read(&pa->pa_count)) {
4336 * This is the pa that we just used
4337 * for block allocation. So don't
4340 spin_unlock(&pa->pa_lock);
4343 if (pa->pa_deleted) {
4344 spin_unlock(&pa->pa_lock);
4347 /* only lg prealloc space */
4348 BUG_ON(!pa->pa_linear);
4350 /* seems this one can be freed ... */
4352 spin_unlock(&pa->pa_lock);
4354 list_del_rcu(&pa->pa_inode_list);
4355 list_add(&pa->u.pa_tmp_list, &discard_list);
4358 if (total_entries <= 5) {
4360 * we want to keep only 5 entries
4361 * allowing it to grow to 8. This
4362 * mak sure we don't call discard
4363 * soon for this list.
4368 spin_unlock(&lg->lg_prealloc_lock);
4370 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4372 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4373 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4374 ext4_error(sb, __func__, "Error in loading buddy "
4375 "information for %u", group);
4378 ext4_lock_group(sb, group);
4379 list_del(&pa->pa_group_list);
4380 ext4_mb_release_group_pa(&e4b, pa, ac);
4381 ext4_unlock_group(sb, group);
4383 ext4_mb_release_desc(&e4b);
4384 list_del(&pa->u.pa_tmp_list);
4385 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4388 kmem_cache_free(ext4_ac_cachep, ac);
4392 * We have incremented pa_count. So it cannot be freed at this
4393 * point. Also we hold lg_mutex. So no parallel allocation is
4394 * possible from this lg. That means pa_free cannot be updated.
4396 * A parallel ext4_mb_discard_group_preallocations is possible.
4397 * which can cause the lg_prealloc_list to be updated.
4400 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4402 int order, added = 0, lg_prealloc_count = 1;
4403 struct super_block *sb = ac->ac_sb;
4404 struct ext4_locality_group *lg = ac->ac_lg;
4405 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4407 order = fls(pa->pa_free) - 1;
4408 if (order > PREALLOC_TB_SIZE - 1)
4409 /* The max size of hash table is PREALLOC_TB_SIZE */
4410 order = PREALLOC_TB_SIZE - 1;
4411 /* Add the prealloc space to lg */
4413 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4415 spin_lock(&tmp_pa->pa_lock);
4416 if (tmp_pa->pa_deleted) {
4417 spin_unlock(&pa->pa_lock);
4420 if (!added && pa->pa_free < tmp_pa->pa_free) {
4421 /* Add to the tail of the previous entry */
4422 list_add_tail_rcu(&pa->pa_inode_list,
4423 &tmp_pa->pa_inode_list);
4426 * we want to count the total
4427 * number of entries in the list
4430 spin_unlock(&tmp_pa->pa_lock);
4431 lg_prealloc_count++;
4434 list_add_tail_rcu(&pa->pa_inode_list,
4435 &lg->lg_prealloc_list[order]);
4438 /* Now trim the list to be not more than 8 elements */
4439 if (lg_prealloc_count > 8) {
4440 ext4_mb_discard_lg_preallocations(sb, lg,
4441 order, lg_prealloc_count);
4448 * release all resource we used in allocation
4450 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4452 struct ext4_prealloc_space *pa = ac->ac_pa;
4454 if (pa->pa_linear) {
4455 /* see comment in ext4_mb_use_group_pa() */
4456 spin_lock(&pa->pa_lock);
4457 pa->pa_pstart += ac->ac_b_ex.fe_len;
4458 pa->pa_lstart += ac->ac_b_ex.fe_len;
4459 pa->pa_free -= ac->ac_b_ex.fe_len;
4460 pa->pa_len -= ac->ac_b_ex.fe_len;
4461 spin_unlock(&pa->pa_lock);
4463 * We want to add the pa to the right bucket.
4464 * Remove it from the list and while adding
4465 * make sure the list to which we are adding
4468 if (likely(pa->pa_free)) {
4469 spin_lock(pa->pa_obj_lock);
4470 list_del_rcu(&pa->pa_inode_list);
4471 spin_unlock(pa->pa_obj_lock);
4472 ext4_mb_add_n_trim(ac);
4475 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4478 up_read(ac->alloc_semp);
4479 if (ac->ac_bitmap_page)
4480 page_cache_release(ac->ac_bitmap_page);
4481 if (ac->ac_buddy_page)
4482 page_cache_release(ac->ac_buddy_page);
4483 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4484 mutex_unlock(&ac->ac_lg->lg_mutex);
4485 ext4_mb_collect_stats(ac);
4489 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4495 for (i = 0; i < EXT4_SB(sb)->s_groups_count && needed > 0; i++) {
4496 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4505 * Main entry point into mballoc to allocate blocks
4506 * it tries to use preallocation first, then falls back
4507 * to usual allocation
4509 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4510 struct ext4_allocation_request *ar, int *errp)
4513 struct ext4_allocation_context *ac = NULL;
4514 struct ext4_sb_info *sbi;
4515 struct super_block *sb;
4516 ext4_fsblk_t block = 0;
4517 unsigned int inquota;
4518 unsigned int reserv_blks = 0;
4520 sb = ar->inode->i_sb;
4523 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) {
4525 * With delalloc we already reserved the blocks
4527 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4528 /* let others to free the space */
4530 ar->len = ar->len >> 1;
4536 reserv_blks = ar->len;
4538 while (ar->len && DQUOT_ALLOC_BLOCK(ar->inode, ar->len)) {
4539 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4548 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4549 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4551 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4558 *errp = ext4_mb_initialize_context(ac, ar);
4564 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4565 if (!ext4_mb_use_preallocated(ac)) {
4566 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4567 ext4_mb_normalize_request(ac, ar);
4569 /* allocate space in core */
4570 ext4_mb_regular_allocator(ac);
4572 /* as we've just preallocated more space than
4573 * user requested orinally, we store allocated
4574 * space in a special descriptor */
4575 if (ac->ac_status == AC_STATUS_FOUND &&
4576 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4577 ext4_mb_new_preallocation(ac);
4579 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4580 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4581 if (*errp == -EAGAIN) {
4583 * drop the reference that we took
4584 * in ext4_mb_use_best_found
4586 ext4_mb_release_context(ac);
4587 ac->ac_b_ex.fe_group = 0;
4588 ac->ac_b_ex.fe_start = 0;
4589 ac->ac_b_ex.fe_len = 0;
4590 ac->ac_status = AC_STATUS_CONTINUE;
4593 ac->ac_b_ex.fe_len = 0;
4595 ext4_mb_show_ac(ac);
4597 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4598 ar->len = ac->ac_b_ex.fe_len;
4601 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4605 ac->ac_b_ex.fe_len = 0;
4607 ext4_mb_show_ac(ac);
4610 ext4_mb_release_context(ac);
4613 kmem_cache_free(ext4_ac_cachep, ac);
4615 if (ar->len < inquota)
4616 DQUOT_FREE_BLOCK(ar->inode, inquota - ar->len);
4622 * We can merge two free data extents only if the physical blocks
4623 * are contiguous, AND the extents were freed by the same transaction,
4624 * AND the blocks are associated with the same group.
4626 static int can_merge(struct ext4_free_data *entry1,
4627 struct ext4_free_data *entry2)
4629 if ((entry1->t_tid == entry2->t_tid) &&
4630 (entry1->group == entry2->group) &&
4631 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4636 static noinline_for_stack int
4637 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4638 struct ext4_free_data *new_entry)
4640 ext4_grpblk_t block;
4641 struct ext4_free_data *entry;
4642 struct ext4_group_info *db = e4b->bd_info;
4643 struct super_block *sb = e4b->bd_sb;
4644 struct ext4_sb_info *sbi = EXT4_SB(sb);
4645 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4646 struct rb_node *parent = NULL, *new_node;
4648 BUG_ON(!ext4_handle_valid(handle));
4649 BUG_ON(e4b->bd_bitmap_page == NULL);
4650 BUG_ON(e4b->bd_buddy_page == NULL);
4652 new_node = &new_entry->node;
4653 block = new_entry->start_blk;
4656 /* first free block exent. We need to
4657 protect buddy cache from being freed,
4658 * otherwise we'll refresh it from
4659 * on-disk bitmap and lose not-yet-available
4661 page_cache_get(e4b->bd_buddy_page);
4662 page_cache_get(e4b->bd_bitmap_page);
4666 entry = rb_entry(parent, struct ext4_free_data, node);
4667 if (block < entry->start_blk)
4669 else if (block >= (entry->start_blk + entry->count))
4670 n = &(*n)->rb_right;
4672 ext4_grp_locked_error(sb, e4b->bd_group, __func__,
4673 "Double free of blocks %d (%d %d)",
4674 block, entry->start_blk, entry->count);
4679 rb_link_node(new_node, parent, n);
4680 rb_insert_color(new_node, &db->bb_free_root);
4682 /* Now try to see the extent can be merged to left and right */
4683 node = rb_prev(new_node);
4685 entry = rb_entry(node, struct ext4_free_data, node);
4686 if (can_merge(entry, new_entry)) {
4687 new_entry->start_blk = entry->start_blk;
4688 new_entry->count += entry->count;
4689 rb_erase(node, &(db->bb_free_root));
4690 spin_lock(&sbi->s_md_lock);
4691 list_del(&entry->list);
4692 spin_unlock(&sbi->s_md_lock);
4693 kmem_cache_free(ext4_free_ext_cachep, entry);
4697 node = rb_next(new_node);
4699 entry = rb_entry(node, struct ext4_free_data, node);
4700 if (can_merge(new_entry, entry)) {
4701 new_entry->count += entry->count;
4702 rb_erase(node, &(db->bb_free_root));
4703 spin_lock(&sbi->s_md_lock);
4704 list_del(&entry->list);
4705 spin_unlock(&sbi->s_md_lock);
4706 kmem_cache_free(ext4_free_ext_cachep, entry);
4709 /* Add the extent to transaction's private list */
4710 spin_lock(&sbi->s_md_lock);
4711 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4712 spin_unlock(&sbi->s_md_lock);
4717 * Main entry point into mballoc to free blocks
4719 void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
4720 unsigned long block, unsigned long count,
4721 int metadata, unsigned long *freed)
4723 struct buffer_head *bitmap_bh = NULL;
4724 struct super_block *sb = inode->i_sb;
4725 struct ext4_allocation_context *ac = NULL;
4726 struct ext4_group_desc *gdp;
4727 struct ext4_super_block *es;
4728 unsigned int overflow;
4730 struct buffer_head *gd_bh;
4731 ext4_group_t block_group;
4732 struct ext4_sb_info *sbi;
4733 struct ext4_buddy e4b;
4740 es = EXT4_SB(sb)->s_es;
4741 if (block < le32_to_cpu(es->s_first_data_block) ||
4742 block + count < block ||
4743 block + count > ext4_blocks_count(es)) {
4744 ext4_error(sb, __func__,
4745 "Freeing blocks not in datazone - "
4746 "block = %lu, count = %lu", block, count);
4750 ext4_debug("freeing block %lu\n", block);
4752 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4754 ac->ac_op = EXT4_MB_HISTORY_FREE;
4755 ac->ac_inode = inode;
4761 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4764 * Check to see if we are freeing blocks across a group
4767 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4768 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4771 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4776 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4782 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4783 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4784 in_range(block, ext4_inode_table(sb, gdp),
4785 EXT4_SB(sb)->s_itb_per_group) ||
4786 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4787 EXT4_SB(sb)->s_itb_per_group)) {
4789 ext4_error(sb, __func__,
4790 "Freeing blocks in system zone - "
4791 "Block = %lu, count = %lu", block, count);
4792 /* err = 0. ext4_std_error should be a no op */
4796 BUFFER_TRACE(bitmap_bh, "getting write access");
4797 err = ext4_journal_get_write_access(handle, bitmap_bh);
4802 * We are about to modify some metadata. Call the journal APIs
4803 * to unshare ->b_data if a currently-committing transaction is
4806 BUFFER_TRACE(gd_bh, "get_write_access");
4807 err = ext4_journal_get_write_access(handle, gd_bh);
4810 #ifdef AGGRESSIVE_CHECK
4813 for (i = 0; i < count; i++)
4814 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4818 ac->ac_b_ex.fe_group = block_group;
4819 ac->ac_b_ex.fe_start = bit;
4820 ac->ac_b_ex.fe_len = count;
4821 ext4_mb_store_history(ac);
4824 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4827 if (metadata && ext4_handle_valid(handle)) {
4828 struct ext4_free_data *new_entry;
4830 * blocks being freed are metadata. these blocks shouldn't
4831 * be used until this transaction is committed
4833 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4834 new_entry->start_blk = bit;
4835 new_entry->group = block_group;
4836 new_entry->count = count;
4837 new_entry->t_tid = handle->h_transaction->t_tid;
4838 ext4_lock_group(sb, block_group);
4839 mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data,
4841 ext4_mb_free_metadata(handle, &e4b, new_entry);
4842 ext4_unlock_group(sb, block_group);
4844 ext4_lock_group(sb, block_group);
4845 /* need to update group_info->bb_free and bitmap
4846 * with group lock held. generate_buddy look at
4847 * them with group lock_held
4849 mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data,
4851 mb_free_blocks(inode, &e4b, bit, count);
4852 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4853 ext4_unlock_group(sb, block_group);
4856 spin_lock(sb_bgl_lock(sbi, block_group));
4857 ret = ext4_free_blks_count(sb, gdp) + count;
4858 ext4_free_blks_set(sb, gdp, ret);
4859 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4860 spin_unlock(sb_bgl_lock(sbi, block_group));
4861 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4863 if (sbi->s_log_groups_per_flex) {
4864 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4865 spin_lock(sb_bgl_lock(sbi, flex_group));
4866 sbi->s_flex_groups[flex_group].free_blocks += count;
4867 spin_unlock(sb_bgl_lock(sbi, flex_group));
4870 ext4_mb_release_desc(&e4b);
4874 /* We dirtied the bitmap block */
4875 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4876 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4878 /* And the group descriptor block */
4879 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4880 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4884 if (overflow && !err) {
4893 ext4_std_error(sb, err);
4895 kmem_cache_free(ext4_ac_cachep, ac);