2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
25 #include <linux/debugfs.h>
26 #include <linux/slab.h>
27 #include <trace/events/ext4.h>
31 * - test ext4_ext_search_left() and ext4_ext_search_right()
32 * - search for metadata in few groups
35 * - normalization should take into account whether file is still open
36 * - discard preallocations if no free space left (policy?)
37 * - don't normalize tails
39 * - reservation for superuser
42 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
43 * - track min/max extents in each group for better group selection
44 * - mb_mark_used() may allocate chunk right after splitting buddy
45 * - tree of groups sorted by number of free blocks
50 * The allocation request involve request for multiple number of blocks
51 * near to the goal(block) value specified.
53 * During initialization phase of the allocator we decide to use the
54 * group preallocation or inode preallocation depending on the size of
55 * the file. The size of the file could be the resulting file size we
56 * would have after allocation, or the current file size, which ever
57 * is larger. If the size is less than sbi->s_mb_stream_request we
58 * select to use the group preallocation. The default value of
59 * s_mb_stream_request is 16 blocks. This can also be tuned via
60 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
61 * terms of number of blocks.
63 * The main motivation for having small file use group preallocation is to
64 * ensure that we have small files closer together on the disk.
66 * First stage the allocator looks at the inode prealloc list,
67 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
68 * spaces for this particular inode. The inode prealloc space is
71 * pa_lstart -> the logical start block for this prealloc space
72 * pa_pstart -> the physical start block for this prealloc space
73 * pa_len -> length for this prealloc space
74 * pa_free -> free space available in this prealloc space
76 * The inode preallocation space is used looking at the _logical_ start
77 * block. If only the logical file block falls within the range of prealloc
78 * space we will consume the particular prealloc space. This make sure that
79 * that the we have contiguous physical blocks representing the file blocks
81 * The important thing to be noted in case of inode prealloc space is that
82 * we don't modify the values associated to inode prealloc space except
85 * If we are not able to find blocks in the inode prealloc space and if we
86 * have the group allocation flag set then we look at the locality group
87 * prealloc space. These are per CPU prealloc list repreasented as
89 * ext4_sb_info.s_locality_groups[smp_processor_id()]
91 * The reason for having a per cpu locality group is to reduce the contention
92 * between CPUs. It is possible to get scheduled at this point.
94 * The locality group prealloc space is used looking at whether we have
95 * enough free space (pa_free) withing the prealloc space.
97 * If we can't allocate blocks via inode prealloc or/and locality group
98 * prealloc then we look at the buddy cache. The buddy cache is represented
99 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
100 * mapped to the buddy and bitmap information regarding different
101 * groups. The buddy information is attached to buddy cache inode so that
102 * we can access them through the page cache. The information regarding
103 * each group is loaded via ext4_mb_load_buddy. The information involve
104 * block bitmap and buddy information. The information are stored in the
108 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
111 * one block each for bitmap and buddy information. So for each group we
112 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
113 * blocksize) blocks. So it can have information regarding groups_per_page
114 * which is blocks_per_page/2
116 * The buddy cache inode is not stored on disk. The inode is thrown
117 * away when the filesystem is unmounted.
119 * We look for count number of blocks in the buddy cache. If we were able
120 * to locate that many free blocks we return with additional information
121 * regarding rest of the contiguous physical block available
123 * Before allocating blocks via buddy cache we normalize the request
124 * blocks. This ensure we ask for more blocks that we needed. The extra
125 * blocks that we get after allocation is added to the respective prealloc
126 * list. In case of inode preallocation we follow a list of heuristics
127 * based on file size. This can be found in ext4_mb_normalize_request. If
128 * we are doing a group prealloc we try to normalize the request to
129 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
130 * 512 blocks. This can be tuned via
131 * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
132 * terms of number of blocks. If we have mounted the file system with -O
133 * stripe=<value> option the group prealloc request is normalized to the
134 * stripe value (sbi->s_stripe)
136 * The regular allocator(using the buddy cache) supports few tunables.
138 * /sys/fs/ext4/<partition>/mb_min_to_scan
139 * /sys/fs/ext4/<partition>/mb_max_to_scan
140 * /sys/fs/ext4/<partition>/mb_order2_req
142 * The regular allocator uses buddy scan only if the request len is power of
143 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
144 * value of s_mb_order2_reqs can be tuned via
145 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
146 * stripe size (sbi->s_stripe), we try to search for contiguous block in
147 * stripe size. This should result in better allocation on RAID setups. If
148 * not, we search in the specific group using bitmap for best extents. The
149 * tunable min_to_scan and max_to_scan control the behaviour here.
150 * min_to_scan indicate how long the mballoc __must__ look for a best
151 * extent and max_to_scan indicates how long the mballoc __can__ look for a
152 * best extent in the found extents. Searching for the blocks starts with
153 * the group specified as the goal value in allocation context via
154 * ac_g_ex. Each group is first checked based on the criteria whether it
155 * can used for allocation. ext4_mb_good_group explains how the groups are
158 * Both the prealloc space are getting populated as above. So for the first
159 * request we will hit the buddy cache which will result in this prealloc
160 * space getting filled. The prealloc space is then later used for the
161 * subsequent request.
165 * mballoc operates on the following data:
167 * - in-core buddy (actually includes buddy and bitmap)
168 * - preallocation descriptors (PAs)
170 * there are two types of preallocations:
172 * assiged to specific inode and can be used for this inode only.
173 * it describes part of inode's space preallocated to specific
174 * physical blocks. any block from that preallocated can be used
175 * independent. the descriptor just tracks number of blocks left
176 * unused. so, before taking some block from descriptor, one must
177 * make sure corresponded logical block isn't allocated yet. this
178 * also means that freeing any block within descriptor's range
179 * must discard all preallocated blocks.
181 * assigned to specific locality group which does not translate to
182 * permanent set of inodes: inode can join and leave group. space
183 * from this type of preallocation can be used for any inode. thus
184 * it's consumed from the beginning to the end.
186 * relation between them can be expressed as:
187 * in-core buddy = on-disk bitmap + preallocation descriptors
189 * this mean blocks mballoc considers used are:
190 * - allocated blocks (persistent)
191 * - preallocated blocks (non-persistent)
193 * consistency in mballoc world means that at any time a block is either
194 * free or used in ALL structures. notice: "any time" should not be read
195 * literally -- time is discrete and delimited by locks.
197 * to keep it simple, we don't use block numbers, instead we count number of
198 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
200 * all operations can be expressed as:
201 * - init buddy: buddy = on-disk + PAs
202 * - new PA: buddy += N; PA = N
203 * - use inode PA: on-disk += N; PA -= N
204 * - discard inode PA buddy -= on-disk - PA; PA = 0
205 * - use locality group PA on-disk += N; PA -= N
206 * - discard locality group PA buddy -= PA; PA = 0
207 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
208 * is used in real operation because we can't know actual used
209 * bits from PA, only from on-disk bitmap
211 * if we follow this strict logic, then all operations above should be atomic.
212 * given some of them can block, we'd have to use something like semaphores
213 * killing performance on high-end SMP hardware. let's try to relax it using
214 * the following knowledge:
215 * 1) if buddy is referenced, it's already initialized
216 * 2) while block is used in buddy and the buddy is referenced,
217 * nobody can re-allocate that block
218 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
219 * bit set and PA claims same block, it's OK. IOW, one can set bit in
220 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
223 * so, now we're building a concurrency table:
226 * blocks for PA are allocated in the buddy, buddy must be referenced
227 * until PA is linked to allocation group to avoid concurrent buddy init
229 * we need to make sure that either on-disk bitmap or PA has uptodate data
230 * given (3) we care that PA-=N operation doesn't interfere with init
232 * the simplest way would be to have buddy initialized by the discard
233 * - use locality group PA
234 * again PA-=N must be serialized with init
235 * - discard locality group PA
236 * the simplest way would be to have buddy initialized by the discard
239 * i_data_sem serializes them
241 * discard process must wait until PA isn't used by another process
242 * - use locality group PA
243 * some mutex should serialize them
244 * - discard locality group PA
245 * discard process must wait until PA isn't used by another process
248 * i_data_sem or another mutex should serializes them
250 * discard process must wait until PA isn't used by another process
251 * - use locality group PA
252 * nothing wrong here -- they're different PAs covering different blocks
253 * - discard locality group PA
254 * discard process must wait until PA isn't used by another process
256 * now we're ready to make few consequences:
257 * - PA is referenced and while it is no discard is possible
258 * - PA is referenced until block isn't marked in on-disk bitmap
259 * - PA changes only after on-disk bitmap
260 * - discard must not compete with init. either init is done before
261 * any discard or they're serialized somehow
262 * - buddy init as sum of on-disk bitmap and PAs is done atomically
264 * a special case when we've used PA to emptiness. no need to modify buddy
265 * in this case, but we should care about concurrent init
270 * Logic in few words:
275 * mark bits in on-disk bitmap
278 * - use preallocation:
279 * find proper PA (per-inode or group)
281 * mark bits in on-disk bitmap
287 * mark bits in on-disk bitmap
290 * - discard preallocations in group:
292 * move them onto local list
293 * load on-disk bitmap
295 * remove PA from object (inode or locality group)
296 * mark free blocks in-core
298 * - discard inode's preallocations:
305 * - bitlock on a group (group)
306 * - object (inode/locality) (object)
317 * - release consumed pa:
322 * - generate in-core bitmap:
326 * - discard all for given object (inode, locality group):
331 * - discard all for given group:
338 static struct kmem_cache *ext4_pspace_cachep;
339 static struct kmem_cache *ext4_ac_cachep;
340 static struct kmem_cache *ext4_free_ext_cachep;
342 /* We create slab caches for groupinfo data structures based on the
343 * superblock block size. There will be one per mounted filesystem for
344 * each unique s_blocksize_bits */
345 #define NR_GRPINFO_CACHES \
346 (EXT4_MAX_BLOCK_LOG_SIZE - EXT4_MIN_BLOCK_LOG_SIZE + 1)
347 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
349 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
351 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
353 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
355 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
357 #if BITS_PER_LONG == 64
358 *bit += ((unsigned long) addr & 7UL) << 3;
359 addr = (void *) ((unsigned long) addr & ~7UL);
360 #elif BITS_PER_LONG == 32
361 *bit += ((unsigned long) addr & 3UL) << 3;
362 addr = (void *) ((unsigned long) addr & ~3UL);
364 #error "how many bits you are?!"
369 static inline int mb_test_bit(int bit, void *addr)
372 * ext4_test_bit on architecture like powerpc
373 * needs unsigned long aligned address
375 addr = mb_correct_addr_and_bit(&bit, addr);
376 return ext4_test_bit(bit, addr);
379 static inline void mb_set_bit(int bit, void *addr)
381 addr = mb_correct_addr_and_bit(&bit, addr);
382 ext4_set_bit(bit, addr);
385 static inline void mb_clear_bit(int bit, void *addr)
387 addr = mb_correct_addr_and_bit(&bit, addr);
388 ext4_clear_bit(bit, addr);
391 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
393 int fix = 0, ret, tmpmax;
394 addr = mb_correct_addr_and_bit(&fix, addr);
398 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
404 static inline int mb_find_next_bit(void *addr, int max, int start)
406 int fix = 0, ret, tmpmax;
407 addr = mb_correct_addr_and_bit(&fix, addr);
411 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
417 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
421 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
424 if (order > e4b->bd_blkbits + 1) {
429 /* at order 0 we see each particular block */
430 *max = 1 << (e4b->bd_blkbits + 3);
432 return EXT4_MB_BITMAP(e4b);
434 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
435 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
441 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
442 int first, int count)
445 struct super_block *sb = e4b->bd_sb;
447 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
449 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
450 for (i = 0; i < count; i++) {
451 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
452 ext4_fsblk_t blocknr;
454 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
455 blocknr += first + i;
456 ext4_grp_locked_error(sb, e4b->bd_group,
457 inode ? inode->i_ino : 0,
459 "freeing block already freed "
463 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
467 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
471 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
473 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
474 for (i = 0; i < count; i++) {
475 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
476 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
480 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
482 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
483 unsigned char *b1, *b2;
485 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
486 b2 = (unsigned char *) bitmap;
487 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
488 if (b1[i] != b2[i]) {
489 printk(KERN_ERR "corruption in group %u "
490 "at byte %u(%u): %x in copy != %x "
491 "on disk/prealloc\n",
492 e4b->bd_group, i, i * 8, b1[i], b2[i]);
500 static inline void mb_free_blocks_double(struct inode *inode,
501 struct ext4_buddy *e4b, int first, int count)
505 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
506 int first, int count)
510 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
516 #ifdef AGGRESSIVE_CHECK
518 #define MB_CHECK_ASSERT(assert) \
522 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
523 function, file, line, # assert); \
528 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
529 const char *function, int line)
531 struct super_block *sb = e4b->bd_sb;
532 int order = e4b->bd_blkbits + 1;
539 struct ext4_group_info *grp;
542 struct list_head *cur;
547 static int mb_check_counter;
548 if (mb_check_counter++ % 100 != 0)
553 buddy = mb_find_buddy(e4b, order, &max);
554 MB_CHECK_ASSERT(buddy);
555 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
556 MB_CHECK_ASSERT(buddy2);
557 MB_CHECK_ASSERT(buddy != buddy2);
558 MB_CHECK_ASSERT(max * 2 == max2);
561 for (i = 0; i < max; i++) {
563 if (mb_test_bit(i, buddy)) {
564 /* only single bit in buddy2 may be 1 */
565 if (!mb_test_bit(i << 1, buddy2)) {
567 mb_test_bit((i<<1)+1, buddy2));
568 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
570 mb_test_bit(i << 1, buddy2));
575 /* both bits in buddy2 must be 0 */
576 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
577 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
579 for (j = 0; j < (1 << order); j++) {
580 k = (i * (1 << order)) + j;
582 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
586 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
591 buddy = mb_find_buddy(e4b, 0, &max);
592 for (i = 0; i < max; i++) {
593 if (!mb_test_bit(i, buddy)) {
594 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
602 /* check used bits only */
603 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
604 buddy2 = mb_find_buddy(e4b, j, &max2);
606 MB_CHECK_ASSERT(k < max2);
607 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
610 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
611 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
613 grp = ext4_get_group_info(sb, e4b->bd_group);
614 buddy = mb_find_buddy(e4b, 0, &max);
615 list_for_each(cur, &grp->bb_prealloc_list) {
616 ext4_group_t groupnr;
617 struct ext4_prealloc_space *pa;
618 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
619 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
620 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
621 for (i = 0; i < pa->pa_len; i++)
622 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
626 #undef MB_CHECK_ASSERT
627 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
628 __FILE__, __func__, __LINE__)
630 #define mb_check_buddy(e4b)
633 /* FIXME!! need more doc */
634 static void ext4_mb_mark_free_simple(struct super_block *sb,
635 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
636 struct ext4_group_info *grp)
638 struct ext4_sb_info *sbi = EXT4_SB(sb);
642 unsigned short border;
644 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
646 border = 2 << sb->s_blocksize_bits;
649 /* find how many blocks can be covered since this position */
650 max = ffs(first | border) - 1;
652 /* find how many blocks of power 2 we need to mark */
659 /* mark multiblock chunks only */
660 grp->bb_counters[min]++;
662 mb_clear_bit(first >> min,
663 buddy + sbi->s_mb_offsets[min]);
671 * Cache the order of the largest free extent we have available in this block
675 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
680 grp->bb_largest_free_order = -1; /* uninit */
682 bits = sb->s_blocksize_bits + 1;
683 for (i = bits; i >= 0; i--) {
684 if (grp->bb_counters[i] > 0) {
685 grp->bb_largest_free_order = i;
691 static noinline_for_stack
692 void ext4_mb_generate_buddy(struct super_block *sb,
693 void *buddy, void *bitmap, ext4_group_t group)
695 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
696 ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb);
701 unsigned fragments = 0;
702 unsigned long long period = get_cycles();
704 /* initialize buddy from bitmap which is aggregation
705 * of on-disk bitmap and preallocations */
706 i = mb_find_next_zero_bit(bitmap, max, 0);
707 grp->bb_first_free = i;
711 i = mb_find_next_bit(bitmap, max, i);
715 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
717 grp->bb_counters[0]++;
719 i = mb_find_next_zero_bit(bitmap, max, i);
721 grp->bb_fragments = fragments;
723 if (free != grp->bb_free) {
724 ext4_grp_locked_error(sb, group, 0, 0,
725 "%u blocks in bitmap, %u in gd",
728 * If we intent to continue, we consider group descritor
729 * corrupt and update bb_free using bitmap value
733 mb_set_largest_free_order(sb, grp);
735 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
737 period = get_cycles() - period;
738 spin_lock(&EXT4_SB(sb)->s_bal_lock);
739 EXT4_SB(sb)->s_mb_buddies_generated++;
740 EXT4_SB(sb)->s_mb_generation_time += period;
741 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
744 /* The buddy information is attached the buddy cache inode
745 * for convenience. The information regarding each group
746 * is loaded via ext4_mb_load_buddy. The information involve
747 * block bitmap and buddy information. The information are
748 * stored in the inode as
751 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
754 * one block each for bitmap and buddy information.
755 * So for each group we take up 2 blocks. A page can
756 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
757 * So it can have information regarding groups_per_page which
758 * is blocks_per_page/2
760 * Locking note: This routine takes the block group lock of all groups
761 * for this page; do not hold this lock when calling this routine!
764 static int ext4_mb_init_cache(struct page *page, char *incore)
766 ext4_group_t ngroups;
772 ext4_group_t first_group;
774 struct super_block *sb;
775 struct buffer_head *bhs;
776 struct buffer_head **bh;
781 mb_debug(1, "init page %lu\n", page->index);
783 inode = page->mapping->host;
785 ngroups = ext4_get_groups_count(sb);
786 blocksize = 1 << inode->i_blkbits;
787 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
789 groups_per_page = blocks_per_page >> 1;
790 if (groups_per_page == 0)
793 /* allocate buffer_heads to read bitmaps */
794 if (groups_per_page > 1) {
796 i = sizeof(struct buffer_head *) * groups_per_page;
797 bh = kzalloc(i, GFP_NOFS);
803 first_group = page->index * blocks_per_page / 2;
805 /* read all groups the page covers into the cache */
806 for (i = 0; i < groups_per_page; i++) {
807 struct ext4_group_desc *desc;
809 if (first_group + i >= ngroups)
813 desc = ext4_get_group_desc(sb, first_group + i, NULL);
818 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
822 if (bitmap_uptodate(bh[i]))
826 if (bitmap_uptodate(bh[i])) {
827 unlock_buffer(bh[i]);
830 ext4_lock_group(sb, first_group + i);
831 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
832 ext4_init_block_bitmap(sb, bh[i],
833 first_group + i, desc);
834 set_bitmap_uptodate(bh[i]);
835 set_buffer_uptodate(bh[i]);
836 ext4_unlock_group(sb, first_group + i);
837 unlock_buffer(bh[i]);
840 ext4_unlock_group(sb, first_group + i);
841 if (buffer_uptodate(bh[i])) {
843 * if not uninit if bh is uptodate,
844 * bitmap is also uptodate
846 set_bitmap_uptodate(bh[i]);
847 unlock_buffer(bh[i]);
852 * submit the buffer_head for read. We can
853 * safely mark the bitmap as uptodate now.
854 * We do it here so the bitmap uptodate bit
855 * get set with buffer lock held.
857 set_bitmap_uptodate(bh[i]);
858 bh[i]->b_end_io = end_buffer_read_sync;
859 submit_bh(READ, bh[i]);
860 mb_debug(1, "read bitmap for group %u\n", first_group + i);
863 /* wait for I/O completion */
864 for (i = 0; i < groups_per_page && bh[i]; i++)
865 wait_on_buffer(bh[i]);
868 for (i = 0; i < groups_per_page && bh[i]; i++)
869 if (!buffer_uptodate(bh[i]))
873 first_block = page->index * blocks_per_page;
875 memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
876 for (i = 0; i < blocks_per_page; i++) {
878 struct ext4_group_info *grinfo;
880 group = (first_block + i) >> 1;
881 if (group >= ngroups)
885 * data carry information regarding this
886 * particular group in the format specified
890 data = page_address(page) + (i * blocksize);
891 bitmap = bh[group - first_group]->b_data;
894 * We place the buddy block and bitmap block
897 if ((first_block + i) & 1) {
898 /* this is block of buddy */
899 BUG_ON(incore == NULL);
900 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
901 group, page->index, i * blocksize);
902 trace_ext4_mb_buddy_bitmap_load(sb, group);
903 grinfo = ext4_get_group_info(sb, group);
904 grinfo->bb_fragments = 0;
905 memset(grinfo->bb_counters, 0,
906 sizeof(*grinfo->bb_counters) *
907 (sb->s_blocksize_bits+2));
909 * incore got set to the group block bitmap below
911 ext4_lock_group(sb, group);
912 ext4_mb_generate_buddy(sb, data, incore, group);
913 ext4_unlock_group(sb, group);
916 /* this is block of bitmap */
917 BUG_ON(incore != NULL);
918 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
919 group, page->index, i * blocksize);
920 trace_ext4_mb_bitmap_load(sb, group);
922 /* see comments in ext4_mb_put_pa() */
923 ext4_lock_group(sb, group);
924 memcpy(data, bitmap, blocksize);
926 /* mark all preallocated blks used in in-core bitmap */
927 ext4_mb_generate_from_pa(sb, data, group);
928 ext4_mb_generate_from_freelist(sb, data, group);
929 ext4_unlock_group(sb, group);
931 /* set incore so that the buddy information can be
932 * generated using this
937 SetPageUptodate(page);
941 for (i = 0; i < groups_per_page && bh[i]; i++)
950 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
951 * block group lock of all groups for this page; do not hold the BG lock when
952 * calling this routine!
954 static noinline_for_stack
955 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
961 int block, pnum, poff;
962 int num_grp_locked = 0;
963 struct ext4_group_info *this_grp;
964 struct ext4_sb_info *sbi = EXT4_SB(sb);
965 struct inode *inode = sbi->s_buddy_cache;
966 struct page *page = NULL, *bitmap_page = NULL;
968 mb_debug(1, "init group %u\n", group);
969 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
970 this_grp = ext4_get_group_info(sb, group);
972 * This ensures that we don't reinit the buddy cache
973 * page which map to the group from which we are already
974 * allocating. If we are looking at the buddy cache we would
975 * have taken a reference using ext4_mb_load_buddy and that
976 * would have taken the alloc_sem lock.
978 num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group);
979 if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
981 * somebody initialized the group
982 * return without doing anything
988 * the buddy cache inode stores the block bitmap
989 * and buddy information in consecutive blocks.
990 * So for each group we need two blocks.
993 pnum = block / blocks_per_page;
994 poff = block % blocks_per_page;
995 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
997 BUG_ON(page->mapping != inode->i_mapping);
998 ret = ext4_mb_init_cache(page, NULL);
1005 if (page == NULL || !PageUptodate(page)) {
1009 mark_page_accessed(page);
1011 bitmap = page_address(page) + (poff * sb->s_blocksize);
1013 /* init buddy cache */
1015 pnum = block / blocks_per_page;
1016 poff = block % blocks_per_page;
1017 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1018 if (page == bitmap_page) {
1020 * If both the bitmap and buddy are in
1021 * the same page we don't need to force
1026 BUG_ON(page->mapping != inode->i_mapping);
1027 ret = ext4_mb_init_cache(page, bitmap);
1034 if (page == NULL || !PageUptodate(page)) {
1038 mark_page_accessed(page);
1040 ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
1042 page_cache_release(bitmap_page);
1044 page_cache_release(page);
1049 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1050 * block group lock of all groups for this page; do not hold the BG lock when
1051 * calling this routine!
1053 static noinline_for_stack int
1054 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1055 struct ext4_buddy *e4b)
1057 int blocks_per_page;
1063 struct ext4_group_info *grp;
1064 struct ext4_sb_info *sbi = EXT4_SB(sb);
1065 struct inode *inode = sbi->s_buddy_cache;
1067 mb_debug(1, "load group %u\n", group);
1069 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1070 grp = ext4_get_group_info(sb, group);
1072 e4b->bd_blkbits = sb->s_blocksize_bits;
1073 e4b->bd_info = ext4_get_group_info(sb, group);
1075 e4b->bd_group = group;
1076 e4b->bd_buddy_page = NULL;
1077 e4b->bd_bitmap_page = NULL;
1078 e4b->alloc_semp = &grp->alloc_sem;
1080 /* Take the read lock on the group alloc
1081 * sem. This would make sure a parallel
1082 * ext4_mb_init_group happening on other
1083 * groups mapped by the page is blocked
1084 * till we are done with allocation
1087 down_read(e4b->alloc_semp);
1089 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1090 /* we need to check for group need init flag
1091 * with alloc_semp held so that we can be sure
1092 * that new blocks didn't get added to the group
1093 * when we are loading the buddy cache
1095 up_read(e4b->alloc_semp);
1097 * we need full data about the group
1098 * to make a good selection
1100 ret = ext4_mb_init_group(sb, group);
1103 goto repeat_load_buddy;
1107 * the buddy cache inode stores the block bitmap
1108 * and buddy information in consecutive blocks.
1109 * So for each group we need two blocks.
1112 pnum = block / blocks_per_page;
1113 poff = block % blocks_per_page;
1115 /* we could use find_or_create_page(), but it locks page
1116 * what we'd like to avoid in fast path ... */
1117 page = find_get_page(inode->i_mapping, pnum);
1118 if (page == NULL || !PageUptodate(page)) {
1121 * drop the page reference and try
1122 * to get the page with lock. If we
1123 * are not uptodate that implies
1124 * somebody just created the page but
1125 * is yet to initialize the same. So
1126 * wait for it to initialize.
1128 page_cache_release(page);
1129 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1131 BUG_ON(page->mapping != inode->i_mapping);
1132 if (!PageUptodate(page)) {
1133 ret = ext4_mb_init_cache(page, NULL);
1138 mb_cmp_bitmaps(e4b, page_address(page) +
1139 (poff * sb->s_blocksize));
1144 if (page == NULL || !PageUptodate(page)) {
1148 e4b->bd_bitmap_page = page;
1149 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1150 mark_page_accessed(page);
1153 pnum = block / blocks_per_page;
1154 poff = block % blocks_per_page;
1156 page = find_get_page(inode->i_mapping, pnum);
1157 if (page == NULL || !PageUptodate(page)) {
1159 page_cache_release(page);
1160 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1162 BUG_ON(page->mapping != inode->i_mapping);
1163 if (!PageUptodate(page)) {
1164 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1173 if (page == NULL || !PageUptodate(page)) {
1177 e4b->bd_buddy_page = page;
1178 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1179 mark_page_accessed(page);
1181 BUG_ON(e4b->bd_bitmap_page == NULL);
1182 BUG_ON(e4b->bd_buddy_page == NULL);
1187 if (e4b->bd_bitmap_page)
1188 page_cache_release(e4b->bd_bitmap_page);
1189 if (e4b->bd_buddy_page)
1190 page_cache_release(e4b->bd_buddy_page);
1191 e4b->bd_buddy = NULL;
1192 e4b->bd_bitmap = NULL;
1194 /* Done with the buddy cache */
1195 up_read(e4b->alloc_semp);
1199 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1201 if (e4b->bd_bitmap_page)
1202 page_cache_release(e4b->bd_bitmap_page);
1203 if (e4b->bd_buddy_page)
1204 page_cache_release(e4b->bd_buddy_page);
1205 /* Done with the buddy cache */
1206 if (e4b->alloc_semp)
1207 up_read(e4b->alloc_semp);
1211 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1216 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1217 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1219 bb = EXT4_MB_BUDDY(e4b);
1220 while (order <= e4b->bd_blkbits + 1) {
1222 if (!mb_test_bit(block, bb)) {
1223 /* this block is part of buddy of order 'order' */
1226 bb += 1 << (e4b->bd_blkbits - order);
1232 static void mb_clear_bits(void *bm, int cur, int len)
1238 if ((cur & 31) == 0 && (len - cur) >= 32) {
1239 /* fast path: clear whole word at once */
1240 addr = bm + (cur >> 3);
1245 mb_clear_bit(cur, bm);
1250 static void mb_set_bits(void *bm, int cur, int len)
1256 if ((cur & 31) == 0 && (len - cur) >= 32) {
1257 /* fast path: set whole word at once */
1258 addr = bm + (cur >> 3);
1263 mb_set_bit(cur, bm);
1268 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1269 int first, int count)
1276 struct super_block *sb = e4b->bd_sb;
1278 BUG_ON(first + count > (sb->s_blocksize << 3));
1279 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1280 mb_check_buddy(e4b);
1281 mb_free_blocks_double(inode, e4b, first, count);
1283 e4b->bd_info->bb_free += count;
1284 if (first < e4b->bd_info->bb_first_free)
1285 e4b->bd_info->bb_first_free = first;
1287 /* let's maintain fragments counter */
1289 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1290 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1291 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1293 e4b->bd_info->bb_fragments--;
1294 else if (!block && !max)
1295 e4b->bd_info->bb_fragments++;
1297 /* let's maintain buddy itself */
1298 while (count-- > 0) {
1302 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1303 ext4_fsblk_t blocknr;
1305 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1307 ext4_grp_locked_error(sb, e4b->bd_group,
1308 inode ? inode->i_ino : 0,
1310 "freeing already freed block "
1313 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1314 e4b->bd_info->bb_counters[order]++;
1316 /* start of the buddy */
1317 buddy = mb_find_buddy(e4b, order, &max);
1321 if (mb_test_bit(block, buddy) ||
1322 mb_test_bit(block + 1, buddy))
1325 /* both the buddies are free, try to coalesce them */
1326 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1332 /* for special purposes, we don't set
1333 * free bits in bitmap */
1334 mb_set_bit(block, buddy);
1335 mb_set_bit(block + 1, buddy);
1337 e4b->bd_info->bb_counters[order]--;
1338 e4b->bd_info->bb_counters[order]--;
1342 e4b->bd_info->bb_counters[order]++;
1344 mb_clear_bit(block, buddy2);
1348 mb_set_largest_free_order(sb, e4b->bd_info);
1349 mb_check_buddy(e4b);
1352 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1353 int needed, struct ext4_free_extent *ex)
1360 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1363 buddy = mb_find_buddy(e4b, order, &max);
1364 BUG_ON(buddy == NULL);
1365 BUG_ON(block >= max);
1366 if (mb_test_bit(block, buddy)) {
1373 /* FIXME dorp order completely ? */
1374 if (likely(order == 0)) {
1375 /* find actual order */
1376 order = mb_find_order_for_block(e4b, block);
1377 block = block >> order;
1380 ex->fe_len = 1 << order;
1381 ex->fe_start = block << order;
1382 ex->fe_group = e4b->bd_group;
1384 /* calc difference from given start */
1385 next = next - ex->fe_start;
1387 ex->fe_start += next;
1389 while (needed > ex->fe_len &&
1390 (buddy = mb_find_buddy(e4b, order, &max))) {
1392 if (block + 1 >= max)
1395 next = (block + 1) * (1 << order);
1396 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1399 ord = mb_find_order_for_block(e4b, next);
1402 block = next >> order;
1403 ex->fe_len += 1 << order;
1406 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1410 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1416 int start = ex->fe_start;
1417 int len = ex->fe_len;
1422 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1423 BUG_ON(e4b->bd_group != ex->fe_group);
1424 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1425 mb_check_buddy(e4b);
1426 mb_mark_used_double(e4b, start, len);
1428 e4b->bd_info->bb_free -= len;
1429 if (e4b->bd_info->bb_first_free == start)
1430 e4b->bd_info->bb_first_free += len;
1432 /* let's maintain fragments counter */
1434 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1435 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1436 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1438 e4b->bd_info->bb_fragments++;
1439 else if (!mlen && !max)
1440 e4b->bd_info->bb_fragments--;
1442 /* let's maintain buddy itself */
1444 ord = mb_find_order_for_block(e4b, start);
1446 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1447 /* the whole chunk may be allocated at once! */
1449 buddy = mb_find_buddy(e4b, ord, &max);
1450 BUG_ON((start >> ord) >= max);
1451 mb_set_bit(start >> ord, buddy);
1452 e4b->bd_info->bb_counters[ord]--;
1459 /* store for history */
1461 ret = len | (ord << 16);
1463 /* we have to split large buddy */
1465 buddy = mb_find_buddy(e4b, ord, &max);
1466 mb_set_bit(start >> ord, buddy);
1467 e4b->bd_info->bb_counters[ord]--;
1470 cur = (start >> ord) & ~1U;
1471 buddy = mb_find_buddy(e4b, ord, &max);
1472 mb_clear_bit(cur, buddy);
1473 mb_clear_bit(cur + 1, buddy);
1474 e4b->bd_info->bb_counters[ord]++;
1475 e4b->bd_info->bb_counters[ord]++;
1477 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1479 mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1480 mb_check_buddy(e4b);
1486 * Must be called under group lock!
1488 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1489 struct ext4_buddy *e4b)
1491 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1494 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1495 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1497 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1498 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1499 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1501 /* preallocation can change ac_b_ex, thus we store actually
1502 * allocated blocks for history */
1503 ac->ac_f_ex = ac->ac_b_ex;
1505 ac->ac_status = AC_STATUS_FOUND;
1506 ac->ac_tail = ret & 0xffff;
1507 ac->ac_buddy = ret >> 16;
1510 * take the page reference. We want the page to be pinned
1511 * so that we don't get a ext4_mb_init_cache_call for this
1512 * group until we update the bitmap. That would mean we
1513 * double allocate blocks. The reference is dropped
1514 * in ext4_mb_release_context
1516 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1517 get_page(ac->ac_bitmap_page);
1518 ac->ac_buddy_page = e4b->bd_buddy_page;
1519 get_page(ac->ac_buddy_page);
1520 /* on allocation we use ac to track the held semaphore */
1521 ac->alloc_semp = e4b->alloc_semp;
1522 e4b->alloc_semp = NULL;
1523 /* store last allocated for subsequent stream allocation */
1524 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1525 spin_lock(&sbi->s_md_lock);
1526 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1527 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1528 spin_unlock(&sbi->s_md_lock);
1533 * regular allocator, for general purposes allocation
1536 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1537 struct ext4_buddy *e4b,
1540 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1541 struct ext4_free_extent *bex = &ac->ac_b_ex;
1542 struct ext4_free_extent *gex = &ac->ac_g_ex;
1543 struct ext4_free_extent ex;
1546 if (ac->ac_status == AC_STATUS_FOUND)
1549 * We don't want to scan for a whole year
1551 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1552 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1553 ac->ac_status = AC_STATUS_BREAK;
1558 * Haven't found good chunk so far, let's continue
1560 if (bex->fe_len < gex->fe_len)
1563 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1564 && bex->fe_group == e4b->bd_group) {
1565 /* recheck chunk's availability - we don't know
1566 * when it was found (within this lock-unlock
1568 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1569 if (max >= gex->fe_len) {
1570 ext4_mb_use_best_found(ac, e4b);
1577 * The routine checks whether found extent is good enough. If it is,
1578 * then the extent gets marked used and flag is set to the context
1579 * to stop scanning. Otherwise, the extent is compared with the
1580 * previous found extent and if new one is better, then it's stored
1581 * in the context. Later, the best found extent will be used, if
1582 * mballoc can't find good enough extent.
1584 * FIXME: real allocation policy is to be designed yet!
1586 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1587 struct ext4_free_extent *ex,
1588 struct ext4_buddy *e4b)
1590 struct ext4_free_extent *bex = &ac->ac_b_ex;
1591 struct ext4_free_extent *gex = &ac->ac_g_ex;
1593 BUG_ON(ex->fe_len <= 0);
1594 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1595 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1596 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1601 * The special case - take what you catch first
1603 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1605 ext4_mb_use_best_found(ac, e4b);
1610 * Let's check whether the chuck is good enough
1612 if (ex->fe_len == gex->fe_len) {
1614 ext4_mb_use_best_found(ac, e4b);
1619 * If this is first found extent, just store it in the context
1621 if (bex->fe_len == 0) {
1627 * If new found extent is better, store it in the context
1629 if (bex->fe_len < gex->fe_len) {
1630 /* if the request isn't satisfied, any found extent
1631 * larger than previous best one is better */
1632 if (ex->fe_len > bex->fe_len)
1634 } else if (ex->fe_len > gex->fe_len) {
1635 /* if the request is satisfied, then we try to find
1636 * an extent that still satisfy the request, but is
1637 * smaller than previous one */
1638 if (ex->fe_len < bex->fe_len)
1642 ext4_mb_check_limits(ac, e4b, 0);
1645 static noinline_for_stack
1646 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1647 struct ext4_buddy *e4b)
1649 struct ext4_free_extent ex = ac->ac_b_ex;
1650 ext4_group_t group = ex.fe_group;
1654 BUG_ON(ex.fe_len <= 0);
1655 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1659 ext4_lock_group(ac->ac_sb, group);
1660 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1664 ext4_mb_use_best_found(ac, e4b);
1667 ext4_unlock_group(ac->ac_sb, group);
1668 ext4_mb_unload_buddy(e4b);
1673 static noinline_for_stack
1674 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1675 struct ext4_buddy *e4b)
1677 ext4_group_t group = ac->ac_g_ex.fe_group;
1680 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1681 struct ext4_free_extent ex;
1683 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1686 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1690 ext4_lock_group(ac->ac_sb, group);
1691 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1692 ac->ac_g_ex.fe_len, &ex);
1694 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1697 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1699 /* use do_div to get remainder (would be 64-bit modulo) */
1700 if (do_div(start, sbi->s_stripe) == 0) {
1703 ext4_mb_use_best_found(ac, e4b);
1705 } else if (max >= ac->ac_g_ex.fe_len) {
1706 BUG_ON(ex.fe_len <= 0);
1707 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1708 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1711 ext4_mb_use_best_found(ac, e4b);
1712 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1713 /* Sometimes, caller may want to merge even small
1714 * number of blocks to an existing extent */
1715 BUG_ON(ex.fe_len <= 0);
1716 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1717 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1720 ext4_mb_use_best_found(ac, e4b);
1722 ext4_unlock_group(ac->ac_sb, group);
1723 ext4_mb_unload_buddy(e4b);
1729 * The routine scans buddy structures (not bitmap!) from given order
1730 * to max order and tries to find big enough chunk to satisfy the req
1732 static noinline_for_stack
1733 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1734 struct ext4_buddy *e4b)
1736 struct super_block *sb = ac->ac_sb;
1737 struct ext4_group_info *grp = e4b->bd_info;
1743 BUG_ON(ac->ac_2order <= 0);
1744 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1745 if (grp->bb_counters[i] == 0)
1748 buddy = mb_find_buddy(e4b, i, &max);
1749 BUG_ON(buddy == NULL);
1751 k = mb_find_next_zero_bit(buddy, max, 0);
1756 ac->ac_b_ex.fe_len = 1 << i;
1757 ac->ac_b_ex.fe_start = k << i;
1758 ac->ac_b_ex.fe_group = e4b->bd_group;
1760 ext4_mb_use_best_found(ac, e4b);
1762 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1764 if (EXT4_SB(sb)->s_mb_stats)
1765 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1772 * The routine scans the group and measures all found extents.
1773 * In order to optimize scanning, caller must pass number of
1774 * free blocks in the group, so the routine can know upper limit.
1776 static noinline_for_stack
1777 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1778 struct ext4_buddy *e4b)
1780 struct super_block *sb = ac->ac_sb;
1781 void *bitmap = EXT4_MB_BITMAP(e4b);
1782 struct ext4_free_extent ex;
1786 free = e4b->bd_info->bb_free;
1789 i = e4b->bd_info->bb_first_free;
1791 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1792 i = mb_find_next_zero_bit(bitmap,
1793 EXT4_BLOCKS_PER_GROUP(sb), i);
1794 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1796 * IF we have corrupt bitmap, we won't find any
1797 * free blocks even though group info says we
1798 * we have free blocks
1800 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1801 "%d free blocks as per "
1802 "group info. But bitmap says 0",
1807 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1808 BUG_ON(ex.fe_len <= 0);
1809 if (free < ex.fe_len) {
1810 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1811 "%d free blocks as per "
1812 "group info. But got %d blocks",
1815 * The number of free blocks differs. This mostly
1816 * indicate that the bitmap is corrupt. So exit
1817 * without claiming the space.
1822 ext4_mb_measure_extent(ac, &ex, e4b);
1828 ext4_mb_check_limits(ac, e4b, 1);
1832 * This is a special case for storages like raid5
1833 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1835 static noinline_for_stack
1836 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1837 struct ext4_buddy *e4b)
1839 struct super_block *sb = ac->ac_sb;
1840 struct ext4_sb_info *sbi = EXT4_SB(sb);
1841 void *bitmap = EXT4_MB_BITMAP(e4b);
1842 struct ext4_free_extent ex;
1843 ext4_fsblk_t first_group_block;
1848 BUG_ON(sbi->s_stripe == 0);
1850 /* find first stripe-aligned block in group */
1851 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1853 a = first_group_block + sbi->s_stripe - 1;
1854 do_div(a, sbi->s_stripe);
1855 i = (a * sbi->s_stripe) - first_group_block;
1857 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1858 if (!mb_test_bit(i, bitmap)) {
1859 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1860 if (max >= sbi->s_stripe) {
1863 ext4_mb_use_best_found(ac, e4b);
1871 /* This is now called BEFORE we load the buddy bitmap. */
1872 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1873 ext4_group_t group, int cr)
1875 unsigned free, fragments;
1876 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1877 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1879 BUG_ON(cr < 0 || cr >= 4);
1881 /* We only do this if the grp has never been initialized */
1882 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1883 int ret = ext4_mb_init_group(ac->ac_sb, group);
1888 free = grp->bb_free;
1889 fragments = grp->bb_fragments;
1897 BUG_ON(ac->ac_2order == 0);
1899 if (grp->bb_largest_free_order < ac->ac_2order)
1902 /* Avoid using the first bg of a flexgroup for data files */
1903 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1904 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1905 ((group % flex_size) == 0))
1910 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1914 if (free >= ac->ac_g_ex.fe_len)
1927 * lock the group_info alloc_sem of all the groups
1928 * belonging to the same buddy cache page. This
1929 * make sure other parallel operation on the buddy
1930 * cache doesn't happen whild holding the buddy cache
1933 int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1937 int blocks_per_page;
1938 int groups_per_page;
1939 ext4_group_t ngroups = ext4_get_groups_count(sb);
1940 ext4_group_t first_group;
1941 struct ext4_group_info *grp;
1943 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1945 * the buddy cache inode stores the block bitmap
1946 * and buddy information in consecutive blocks.
1947 * So for each group we need two blocks.
1950 pnum = block / blocks_per_page;
1951 first_group = pnum * blocks_per_page / 2;
1953 groups_per_page = blocks_per_page >> 1;
1954 if (groups_per_page == 0)
1955 groups_per_page = 1;
1956 /* read all groups the page covers into the cache */
1957 for (i = 0; i < groups_per_page; i++) {
1959 if ((first_group + i) >= ngroups)
1961 grp = ext4_get_group_info(sb, first_group + i);
1962 /* take all groups write allocation
1963 * semaphore. This make sure there is
1964 * no block allocation going on in any
1967 down_write_nested(&grp->alloc_sem, i);
1972 void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1973 ext4_group_t group, int locked_group)
1977 int blocks_per_page;
1978 ext4_group_t first_group;
1979 struct ext4_group_info *grp;
1981 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1983 * the buddy cache inode stores the block bitmap
1984 * and buddy information in consecutive blocks.
1985 * So for each group we need two blocks.
1988 pnum = block / blocks_per_page;
1989 first_group = pnum * blocks_per_page / 2;
1990 /* release locks on all the groups */
1991 for (i = 0; i < locked_group; i++) {
1993 grp = ext4_get_group_info(sb, first_group + i);
1994 /* take all groups write allocation
1995 * semaphore. This make sure there is
1996 * no block allocation going on in any
1999 up_write(&grp->alloc_sem);
2004 static noinline_for_stack int
2005 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2007 ext4_group_t ngroups, group, i;
2010 struct ext4_sb_info *sbi;
2011 struct super_block *sb;
2012 struct ext4_buddy e4b;
2016 ngroups = ext4_get_groups_count(sb);
2017 /* non-extent files are limited to low blocks/groups */
2018 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2019 ngroups = sbi->s_blockfile_groups;
2021 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2023 /* first, try the goal */
2024 err = ext4_mb_find_by_goal(ac, &e4b);
2025 if (err || ac->ac_status == AC_STATUS_FOUND)
2028 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2032 * ac->ac2_order is set only if the fe_len is a power of 2
2033 * if ac2_order is set we also set criteria to 0 so that we
2034 * try exact allocation using buddy.
2036 i = fls(ac->ac_g_ex.fe_len);
2039 * We search using buddy data only if the order of the request
2040 * is greater than equal to the sbi_s_mb_order2_reqs
2041 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2043 if (i >= sbi->s_mb_order2_reqs) {
2045 * This should tell if fe_len is exactly power of 2
2047 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2048 ac->ac_2order = i - 1;
2051 /* if stream allocation is enabled, use global goal */
2052 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2053 /* TBD: may be hot point */
2054 spin_lock(&sbi->s_md_lock);
2055 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2056 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2057 spin_unlock(&sbi->s_md_lock);
2060 /* Let's just scan groups to find more-less suitable blocks */
2061 cr = ac->ac_2order ? 0 : 1;
2063 * cr == 0 try to get exact allocation,
2064 * cr == 3 try to get anything
2067 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2068 ac->ac_criteria = cr;
2070 * searching for the right group start
2071 * from the goal value specified
2073 group = ac->ac_g_ex.fe_group;
2075 for (i = 0; i < ngroups; group++, i++) {
2076 if (group == ngroups)
2079 /* This now checks without needing the buddy page */
2080 if (!ext4_mb_good_group(ac, group, cr))
2083 err = ext4_mb_load_buddy(sb, group, &e4b);
2087 ext4_lock_group(sb, group);
2090 * We need to check again after locking the
2093 if (!ext4_mb_good_group(ac, group, cr)) {
2094 ext4_unlock_group(sb, group);
2095 ext4_mb_unload_buddy(&e4b);
2099 ac->ac_groups_scanned++;
2101 ext4_mb_simple_scan_group(ac, &e4b);
2102 else if (cr == 1 && sbi->s_stripe &&
2103 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2104 ext4_mb_scan_aligned(ac, &e4b);
2106 ext4_mb_complex_scan_group(ac, &e4b);
2108 ext4_unlock_group(sb, group);
2109 ext4_mb_unload_buddy(&e4b);
2111 if (ac->ac_status != AC_STATUS_CONTINUE)
2116 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2117 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2119 * We've been searching too long. Let's try to allocate
2120 * the best chunk we've found so far
2123 ext4_mb_try_best_found(ac, &e4b);
2124 if (ac->ac_status != AC_STATUS_FOUND) {
2126 * Someone more lucky has already allocated it.
2127 * The only thing we can do is just take first
2129 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2131 ac->ac_b_ex.fe_group = 0;
2132 ac->ac_b_ex.fe_start = 0;
2133 ac->ac_b_ex.fe_len = 0;
2134 ac->ac_status = AC_STATUS_CONTINUE;
2135 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2137 atomic_inc(&sbi->s_mb_lost_chunks);
2145 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2147 struct super_block *sb = seq->private;
2150 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2153 return (void *) ((unsigned long) group);
2156 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2158 struct super_block *sb = seq->private;
2162 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2165 return (void *) ((unsigned long) group);
2168 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2170 struct super_block *sb = seq->private;
2171 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2174 struct ext4_buddy e4b;
2176 struct ext4_group_info info;
2177 ext4_grpblk_t counters[16];
2182 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2183 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2184 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2185 "group", "free", "frags", "first",
2186 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2187 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2189 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2190 sizeof(struct ext4_group_info);
2191 err = ext4_mb_load_buddy(sb, group, &e4b);
2193 seq_printf(seq, "#%-5u: I/O error\n", group);
2196 ext4_lock_group(sb, group);
2197 memcpy(&sg, ext4_get_group_info(sb, group), i);
2198 ext4_unlock_group(sb, group);
2199 ext4_mb_unload_buddy(&e4b);
2201 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2202 sg.info.bb_fragments, sg.info.bb_first_free);
2203 for (i = 0; i <= 13; i++)
2204 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2205 sg.info.bb_counters[i] : 0);
2206 seq_printf(seq, " ]\n");
2211 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2215 static const struct seq_operations ext4_mb_seq_groups_ops = {
2216 .start = ext4_mb_seq_groups_start,
2217 .next = ext4_mb_seq_groups_next,
2218 .stop = ext4_mb_seq_groups_stop,
2219 .show = ext4_mb_seq_groups_show,
2222 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2224 struct super_block *sb = PDE(inode)->data;
2227 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2229 struct seq_file *m = file->private_data;
2236 static const struct file_operations ext4_mb_seq_groups_fops = {
2237 .owner = THIS_MODULE,
2238 .open = ext4_mb_seq_groups_open,
2240 .llseek = seq_lseek,
2241 .release = seq_release,
2244 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2246 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2247 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2253 /* Create and initialize ext4_group_info data for the given group. */
2254 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2255 struct ext4_group_desc *desc)
2259 struct ext4_sb_info *sbi = EXT4_SB(sb);
2260 struct ext4_group_info **meta_group_info;
2261 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2264 * First check if this group is the first of a reserved block.
2265 * If it's true, we have to allocate a new table of pointers
2266 * to ext4_group_info structures
2268 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2269 metalen = sizeof(*meta_group_info) <<
2270 EXT4_DESC_PER_BLOCK_BITS(sb);
2271 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2272 if (meta_group_info == NULL) {
2273 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2275 goto exit_meta_group_info;
2277 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2282 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2283 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2285 meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL);
2286 if (meta_group_info[i] == NULL) {
2287 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2288 goto exit_group_info;
2290 memset(meta_group_info[i], 0, kmem_cache_size(cachep));
2291 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2292 &(meta_group_info[i]->bb_state));
2295 * initialize bb_free to be able to skip
2296 * empty groups without initialization
2298 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2299 meta_group_info[i]->bb_free =
2300 ext4_free_blocks_after_init(sb, group, desc);
2302 meta_group_info[i]->bb_free =
2303 ext4_free_blks_count(sb, desc);
2306 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2307 init_rwsem(&meta_group_info[i]->alloc_sem);
2308 meta_group_info[i]->bb_free_root = RB_ROOT;
2309 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2313 struct buffer_head *bh;
2314 meta_group_info[i]->bb_bitmap =
2315 kmalloc(sb->s_blocksize, GFP_KERNEL);
2316 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2317 bh = ext4_read_block_bitmap(sb, group);
2319 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2328 /* If a meta_group_info table has been allocated, release it now */
2329 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2330 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2331 exit_meta_group_info:
2333 } /* ext4_mb_add_groupinfo */
2335 static int ext4_mb_init_backend(struct super_block *sb)
2337 ext4_group_t ngroups = ext4_get_groups_count(sb);
2339 struct ext4_sb_info *sbi = EXT4_SB(sb);
2340 struct ext4_super_block *es = sbi->s_es;
2341 int num_meta_group_infos;
2342 int num_meta_group_infos_max;
2344 struct ext4_group_desc *desc;
2345 struct kmem_cache *cachep;
2347 /* This is the number of blocks used by GDT */
2348 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2349 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2352 * This is the total number of blocks used by GDT including
2353 * the number of reserved blocks for GDT.
2354 * The s_group_info array is allocated with this value
2355 * to allow a clean online resize without a complex
2356 * manipulation of pointer.
2357 * The drawback is the unused memory when no resize
2358 * occurs but it's very low in terms of pages
2359 * (see comments below)
2360 * Need to handle this properly when META_BG resizing is allowed
2362 num_meta_group_infos_max = num_meta_group_infos +
2363 le16_to_cpu(es->s_reserved_gdt_blocks);
2366 * array_size is the size of s_group_info array. We round it
2367 * to the next power of two because this approximation is done
2368 * internally by kmalloc so we can have some more memory
2369 * for free here (e.g. may be used for META_BG resize).
2372 while (array_size < sizeof(*sbi->s_group_info) *
2373 num_meta_group_infos_max)
2374 array_size = array_size << 1;
2375 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2376 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2377 * So a two level scheme suffices for now. */
2378 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2379 if (sbi->s_group_info == NULL) {
2380 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2383 sbi->s_buddy_cache = new_inode(sb);
2384 if (sbi->s_buddy_cache == NULL) {
2385 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2388 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2389 for (i = 0; i < ngroups; i++) {
2390 desc = ext4_get_group_desc(sb, i, NULL);
2393 "EXT4-fs: can't read descriptor %u\n", i);
2396 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2403 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2405 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2406 i = num_meta_group_infos;
2408 kfree(sbi->s_group_info[i]);
2409 iput(sbi->s_buddy_cache);
2411 kfree(sbi->s_group_info);
2415 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2417 struct ext4_sb_info *sbi = EXT4_SB(sb);
2423 struct kmem_cache *cachep;
2426 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2428 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2429 if (sbi->s_mb_offsets == NULL) {
2434 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2435 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2436 if (sbi->s_mb_maxs == NULL) {
2441 cache_index = sb->s_blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2442 cachep = ext4_groupinfo_caches[cache_index];
2445 int len = offsetof(struct ext4_group_info,
2446 bb_counters[sb->s_blocksize_bits + 2]);
2448 sprintf(name, "ext4_groupinfo_%d", sb->s_blocksize_bits);
2449 namep = kstrdup(name, GFP_KERNEL);
2455 /* Need to free the kmem_cache_name() when we
2456 * destroy the slab */
2457 cachep = kmem_cache_create(namep, len, 0,
2458 SLAB_RECLAIM_ACCOUNT, NULL);
2463 ext4_groupinfo_caches[cache_index] = cachep;
2466 /* order 0 is regular bitmap */
2467 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2468 sbi->s_mb_offsets[0] = 0;
2472 max = sb->s_blocksize << 2;
2474 sbi->s_mb_offsets[i] = offset;
2475 sbi->s_mb_maxs[i] = max;
2476 offset += 1 << (sb->s_blocksize_bits - i);
2479 } while (i <= sb->s_blocksize_bits + 1);
2481 /* init file for buddy data */
2482 ret = ext4_mb_init_backend(sb);
2487 spin_lock_init(&sbi->s_md_lock);
2488 spin_lock_init(&sbi->s_bal_lock);
2490 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2491 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2492 sbi->s_mb_stats = MB_DEFAULT_STATS;
2493 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2494 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2495 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2497 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2498 if (sbi->s_locality_groups == NULL) {
2502 for_each_possible_cpu(i) {
2503 struct ext4_locality_group *lg;
2504 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2505 mutex_init(&lg->lg_mutex);
2506 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2507 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2508 spin_lock_init(&lg->lg_prealloc_lock);
2512 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2513 &ext4_mb_seq_groups_fops, sb);
2516 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2519 kfree(sbi->s_mb_offsets);
2520 kfree(sbi->s_mb_maxs);
2526 /* need to called with the ext4 group lock held */
2527 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2529 struct ext4_prealloc_space *pa;
2530 struct list_head *cur, *tmp;
2533 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2534 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2535 list_del(&pa->pa_group_list);
2537 kmem_cache_free(ext4_pspace_cachep, pa);
2540 mb_debug(1, "mballoc: %u PAs left\n", count);
2544 int ext4_mb_release(struct super_block *sb)
2546 ext4_group_t ngroups = ext4_get_groups_count(sb);
2548 int num_meta_group_infos;
2549 struct ext4_group_info *grinfo;
2550 struct ext4_sb_info *sbi = EXT4_SB(sb);
2551 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2553 if (sbi->s_group_info) {
2554 for (i = 0; i < ngroups; i++) {
2555 grinfo = ext4_get_group_info(sb, i);
2557 kfree(grinfo->bb_bitmap);
2559 ext4_lock_group(sb, i);
2560 ext4_mb_cleanup_pa(grinfo);
2561 ext4_unlock_group(sb, i);
2562 kmem_cache_free(cachep, grinfo);
2564 num_meta_group_infos = (ngroups +
2565 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2566 EXT4_DESC_PER_BLOCK_BITS(sb);
2567 for (i = 0; i < num_meta_group_infos; i++)
2568 kfree(sbi->s_group_info[i]);
2569 kfree(sbi->s_group_info);
2571 kfree(sbi->s_mb_offsets);
2572 kfree(sbi->s_mb_maxs);
2573 if (sbi->s_buddy_cache)
2574 iput(sbi->s_buddy_cache);
2575 if (sbi->s_mb_stats) {
2577 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2578 atomic_read(&sbi->s_bal_allocated),
2579 atomic_read(&sbi->s_bal_reqs),
2580 atomic_read(&sbi->s_bal_success));
2582 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2583 "%u 2^N hits, %u breaks, %u lost\n",
2584 atomic_read(&sbi->s_bal_ex_scanned),
2585 atomic_read(&sbi->s_bal_goals),
2586 atomic_read(&sbi->s_bal_2orders),
2587 atomic_read(&sbi->s_bal_breaks),
2588 atomic_read(&sbi->s_mb_lost_chunks));
2590 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2591 sbi->s_mb_buddies_generated++,
2592 sbi->s_mb_generation_time);
2594 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2595 atomic_read(&sbi->s_mb_preallocated),
2596 atomic_read(&sbi->s_mb_discarded));
2599 free_percpu(sbi->s_locality_groups);
2601 remove_proc_entry("mb_groups", sbi->s_proc);
2606 static inline int ext4_issue_discard(struct super_block *sb,
2607 ext4_group_t block_group, ext4_grpblk_t block, int count)
2610 ext4_fsblk_t discard_block;
2612 discard_block = block + ext4_group_first_block_no(sb, block_group);
2613 trace_ext4_discard_blocks(sb,
2614 (unsigned long long) discard_block, count);
2615 ret = sb_issue_discard(sb, discard_block, count);
2616 if (ret == -EOPNOTSUPP) {
2617 ext4_warning(sb, "discard not supported, disabling");
2618 clear_opt(EXT4_SB(sb)->s_mount_opt, DISCARD);
2624 * This function is called by the jbd2 layer once the commit has finished,
2625 * so we know we can free the blocks that were released with that commit.
2627 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2629 struct super_block *sb = journal->j_private;
2630 struct ext4_buddy e4b;
2631 struct ext4_group_info *db;
2632 int err, count = 0, count2 = 0;
2633 struct ext4_free_data *entry;
2634 struct list_head *l, *ltmp;
2636 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2637 entry = list_entry(l, struct ext4_free_data, list);
2639 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2640 entry->count, entry->group, entry);
2642 if (test_opt(sb, DISCARD))
2643 ext4_issue_discard(sb, entry->group,
2644 entry->start_blk, entry->count);
2646 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2647 /* we expect to find existing buddy because it's pinned */
2651 /* there are blocks to put in buddy to make them really free */
2652 count += entry->count;
2654 ext4_lock_group(sb, entry->group);
2655 /* Take it out of per group rb tree */
2656 rb_erase(&entry->node, &(db->bb_free_root));
2657 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2659 if (!db->bb_free_root.rb_node) {
2660 /* No more items in the per group rb tree
2661 * balance refcounts from ext4_mb_free_metadata()
2663 page_cache_release(e4b.bd_buddy_page);
2664 page_cache_release(e4b.bd_bitmap_page);
2666 ext4_unlock_group(sb, entry->group);
2667 kmem_cache_free(ext4_free_ext_cachep, entry);
2668 ext4_mb_unload_buddy(&e4b);
2671 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2674 #ifdef CONFIG_EXT4_DEBUG
2675 u8 mb_enable_debug __read_mostly;
2677 static struct dentry *debugfs_dir;
2678 static struct dentry *debugfs_debug;
2680 static void __init ext4_create_debugfs_entry(void)
2682 debugfs_dir = debugfs_create_dir("ext4", NULL);
2684 debugfs_debug = debugfs_create_u8("mballoc-debug",
2690 static void ext4_remove_debugfs_entry(void)
2692 debugfs_remove(debugfs_debug);
2693 debugfs_remove(debugfs_dir);
2698 static void __init ext4_create_debugfs_entry(void)
2702 static void ext4_remove_debugfs_entry(void)
2708 int __init init_ext4_mballoc(void)
2710 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2711 SLAB_RECLAIM_ACCOUNT);
2712 if (ext4_pspace_cachep == NULL)
2715 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2716 SLAB_RECLAIM_ACCOUNT);
2717 if (ext4_ac_cachep == NULL) {
2718 kmem_cache_destroy(ext4_pspace_cachep);
2722 ext4_free_ext_cachep = KMEM_CACHE(ext4_free_data,
2723 SLAB_RECLAIM_ACCOUNT);
2724 if (ext4_free_ext_cachep == NULL) {
2725 kmem_cache_destroy(ext4_pspace_cachep);
2726 kmem_cache_destroy(ext4_ac_cachep);
2729 ext4_create_debugfs_entry();
2733 void exit_ext4_mballoc(void)
2737 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2738 * before destroying the slab cache.
2741 kmem_cache_destroy(ext4_pspace_cachep);
2742 kmem_cache_destroy(ext4_ac_cachep);
2743 kmem_cache_destroy(ext4_free_ext_cachep);
2745 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2746 struct kmem_cache *cachep = ext4_groupinfo_caches[i];
2748 char *name = (char *)kmem_cache_name(cachep);
2749 kmem_cache_destroy(cachep);
2753 ext4_remove_debugfs_entry();
2758 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2759 * Returns 0 if success or error code
2761 static noinline_for_stack int
2762 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2763 handle_t *handle, unsigned int reserv_blks)
2765 struct buffer_head *bitmap_bh = NULL;
2766 struct ext4_group_desc *gdp;
2767 struct buffer_head *gdp_bh;
2768 struct ext4_sb_info *sbi;
2769 struct super_block *sb;
2773 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2774 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2780 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2784 err = ext4_journal_get_write_access(handle, bitmap_bh);
2789 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2793 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2794 ext4_free_blks_count(sb, gdp));
2796 err = ext4_journal_get_write_access(handle, gdp_bh);
2800 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2802 len = ac->ac_b_ex.fe_len;
2803 if (!ext4_data_block_valid(sbi, block, len)) {
2804 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2805 "fs metadata\n", block, block+len);
2806 /* File system mounted not to panic on error
2807 * Fix the bitmap and repeat the block allocation
2808 * We leak some of the blocks here.
2810 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2811 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2812 ac->ac_b_ex.fe_len);
2813 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2814 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2820 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2821 #ifdef AGGRESSIVE_CHECK
2824 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2825 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2826 bitmap_bh->b_data));
2830 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
2831 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2832 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2833 ext4_free_blks_set(sb, gdp,
2834 ext4_free_blocks_after_init(sb,
2835 ac->ac_b_ex.fe_group, gdp));
2837 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
2838 ext4_free_blks_set(sb, gdp, len);
2839 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2841 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2842 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
2844 * Now reduce the dirty block count also. Should not go negative
2846 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2847 /* release all the reserved blocks if non delalloc */
2848 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
2850 if (sbi->s_log_groups_per_flex) {
2851 ext4_group_t flex_group = ext4_flex_group(sbi,
2852 ac->ac_b_ex.fe_group);
2853 atomic_sub(ac->ac_b_ex.fe_len,
2854 &sbi->s_flex_groups[flex_group].free_blocks);
2857 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2860 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2863 ext4_mark_super_dirty(sb);
2869 * here we normalize request for locality group
2870 * Group request are normalized to s_strip size if we set the same via mount
2871 * option. If not we set it to s_mb_group_prealloc which can be configured via
2872 * /sys/fs/ext4/<partition>/mb_group_prealloc
2874 * XXX: should we try to preallocate more than the group has now?
2876 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2878 struct super_block *sb = ac->ac_sb;
2879 struct ext4_locality_group *lg = ac->ac_lg;
2882 if (EXT4_SB(sb)->s_stripe)
2883 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
2885 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2886 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2887 current->pid, ac->ac_g_ex.fe_len);
2891 * Normalization means making request better in terms of
2892 * size and alignment
2894 static noinline_for_stack void
2895 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2896 struct ext4_allocation_request *ar)
2900 loff_t size, orig_size, start_off;
2902 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2903 struct ext4_prealloc_space *pa;
2905 /* do normalize only data requests, metadata requests
2906 do not need preallocation */
2907 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2910 /* sometime caller may want exact blocks */
2911 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2914 /* caller may indicate that preallocation isn't
2915 * required (it's a tail, for example) */
2916 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2919 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2920 ext4_mb_normalize_group_request(ac);
2924 bsbits = ac->ac_sb->s_blocksize_bits;
2926 /* first, let's learn actual file size
2927 * given current request is allocated */
2928 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2929 size = size << bsbits;
2930 if (size < i_size_read(ac->ac_inode))
2931 size = i_size_read(ac->ac_inode);
2934 /* max size of free chunks */
2937 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2938 (req <= (size) || max <= (chunk_size))
2940 /* first, try to predict filesize */
2941 /* XXX: should this table be tunable? */
2943 if (size <= 16 * 1024) {
2945 } else if (size <= 32 * 1024) {
2947 } else if (size <= 64 * 1024) {
2949 } else if (size <= 128 * 1024) {
2951 } else if (size <= 256 * 1024) {
2953 } else if (size <= 512 * 1024) {
2955 } else if (size <= 1024 * 1024) {
2957 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2958 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2959 (21 - bsbits)) << 21;
2960 size = 2 * 1024 * 1024;
2961 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2962 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2963 (22 - bsbits)) << 22;
2964 size = 4 * 1024 * 1024;
2965 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2966 (8<<20)>>bsbits, max, 8 * 1024)) {
2967 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2968 (23 - bsbits)) << 23;
2969 size = 8 * 1024 * 1024;
2971 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2972 size = ac->ac_o_ex.fe_len << bsbits;
2974 size = size >> bsbits;
2975 start = start_off >> bsbits;
2977 /* don't cover already allocated blocks in selected range */
2978 if (ar->pleft && start <= ar->lleft) {
2979 size -= ar->lleft + 1 - start;
2980 start = ar->lleft + 1;
2982 if (ar->pright && start + size - 1 >= ar->lright)
2983 size -= start + size - ar->lright;
2987 /* check we don't cross already preallocated blocks */
2989 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2994 spin_lock(&pa->pa_lock);
2995 if (pa->pa_deleted) {
2996 spin_unlock(&pa->pa_lock);
3000 pa_end = pa->pa_lstart + pa->pa_len;
3002 /* PA must not overlap original request */
3003 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3004 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3006 /* skip PAs this normalized request doesn't overlap with */
3007 if (pa->pa_lstart >= end || pa_end <= start) {
3008 spin_unlock(&pa->pa_lock);
3011 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3013 /* adjust start or end to be adjacent to this pa */
3014 if (pa_end <= ac->ac_o_ex.fe_logical) {
3015 BUG_ON(pa_end < start);
3017 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3018 BUG_ON(pa->pa_lstart > end);
3019 end = pa->pa_lstart;
3021 spin_unlock(&pa->pa_lock);
3026 /* XXX: extra loop to check we really don't overlap preallocations */
3028 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3030 spin_lock(&pa->pa_lock);
3031 if (pa->pa_deleted == 0) {
3032 pa_end = pa->pa_lstart + pa->pa_len;
3033 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3035 spin_unlock(&pa->pa_lock);
3039 if (start + size <= ac->ac_o_ex.fe_logical &&
3040 start > ac->ac_o_ex.fe_logical) {
3041 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3042 (unsigned long) start, (unsigned long) size,
3043 (unsigned long) ac->ac_o_ex.fe_logical);
3045 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3046 start > ac->ac_o_ex.fe_logical);
3047 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3049 /* now prepare goal request */
3051 /* XXX: is it better to align blocks WRT to logical
3052 * placement or satisfy big request as is */
3053 ac->ac_g_ex.fe_logical = start;
3054 ac->ac_g_ex.fe_len = size;
3056 /* define goal start in order to merge */
3057 if (ar->pright && (ar->lright == (start + size))) {
3058 /* merge to the right */
3059 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3060 &ac->ac_f_ex.fe_group,
3061 &ac->ac_f_ex.fe_start);
3062 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3064 if (ar->pleft && (ar->lleft + 1 == start)) {
3065 /* merge to the left */
3066 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3067 &ac->ac_f_ex.fe_group,
3068 &ac->ac_f_ex.fe_start);
3069 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3072 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3073 (unsigned) orig_size, (unsigned) start);
3076 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3078 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3080 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3081 atomic_inc(&sbi->s_bal_reqs);
3082 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3083 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3084 atomic_inc(&sbi->s_bal_success);
3085 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3086 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3087 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3088 atomic_inc(&sbi->s_bal_goals);
3089 if (ac->ac_found > sbi->s_mb_max_to_scan)
3090 atomic_inc(&sbi->s_bal_breaks);
3093 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3094 trace_ext4_mballoc_alloc(ac);
3096 trace_ext4_mballoc_prealloc(ac);
3100 * Called on failure; free up any blocks from the inode PA for this
3101 * context. We don't need this for MB_GROUP_PA because we only change
3102 * pa_free in ext4_mb_release_context(), but on failure, we've already
3103 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3105 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3107 struct ext4_prealloc_space *pa = ac->ac_pa;
3110 if (pa && pa->pa_type == MB_INODE_PA) {
3111 len = ac->ac_b_ex.fe_len;
3118 * use blocks preallocated to inode
3120 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3121 struct ext4_prealloc_space *pa)
3127 /* found preallocated blocks, use them */
3128 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3129 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3131 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3132 &ac->ac_b_ex.fe_start);
3133 ac->ac_b_ex.fe_len = len;
3134 ac->ac_status = AC_STATUS_FOUND;
3137 BUG_ON(start < pa->pa_pstart);
3138 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3139 BUG_ON(pa->pa_free < len);
3142 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3146 * use blocks preallocated to locality group
3148 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3149 struct ext4_prealloc_space *pa)
3151 unsigned int len = ac->ac_o_ex.fe_len;
3153 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3154 &ac->ac_b_ex.fe_group,
3155 &ac->ac_b_ex.fe_start);
3156 ac->ac_b_ex.fe_len = len;
3157 ac->ac_status = AC_STATUS_FOUND;
3160 /* we don't correct pa_pstart or pa_plen here to avoid
3161 * possible race when the group is being loaded concurrently
3162 * instead we correct pa later, after blocks are marked
3163 * in on-disk bitmap -- see ext4_mb_release_context()
3164 * Other CPUs are prevented from allocating from this pa by lg_mutex
3166 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3170 * Return the prealloc space that have minimal distance
3171 * from the goal block. @cpa is the prealloc
3172 * space that is having currently known minimal distance
3173 * from the goal block.
3175 static struct ext4_prealloc_space *
3176 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3177 struct ext4_prealloc_space *pa,
3178 struct ext4_prealloc_space *cpa)
3180 ext4_fsblk_t cur_distance, new_distance;
3183 atomic_inc(&pa->pa_count);
3186 cur_distance = abs(goal_block - cpa->pa_pstart);
3187 new_distance = abs(goal_block - pa->pa_pstart);
3189 if (cur_distance < new_distance)
3192 /* drop the previous reference */
3193 atomic_dec(&cpa->pa_count);
3194 atomic_inc(&pa->pa_count);
3199 * search goal blocks in preallocated space
3201 static noinline_for_stack int
3202 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3205 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3206 struct ext4_locality_group *lg;
3207 struct ext4_prealloc_space *pa, *cpa = NULL;
3208 ext4_fsblk_t goal_block;
3210 /* only data can be preallocated */
3211 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3214 /* first, try per-file preallocation */
3216 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3218 /* all fields in this condition don't change,
3219 * so we can skip locking for them */
3220 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3221 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3224 /* non-extent files can't have physical blocks past 2^32 */
3225 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3226 pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3229 /* found preallocated blocks, use them */
3230 spin_lock(&pa->pa_lock);
3231 if (pa->pa_deleted == 0 && pa->pa_free) {
3232 atomic_inc(&pa->pa_count);
3233 ext4_mb_use_inode_pa(ac, pa);
3234 spin_unlock(&pa->pa_lock);
3235 ac->ac_criteria = 10;
3239 spin_unlock(&pa->pa_lock);
3243 /* can we use group allocation? */
3244 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3247 /* inode may have no locality group for some reason */
3251 order = fls(ac->ac_o_ex.fe_len) - 1;
3252 if (order > PREALLOC_TB_SIZE - 1)
3253 /* The max size of hash table is PREALLOC_TB_SIZE */
3254 order = PREALLOC_TB_SIZE - 1;
3256 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3258 * search for the prealloc space that is having
3259 * minimal distance from the goal block.
3261 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3263 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3265 spin_lock(&pa->pa_lock);
3266 if (pa->pa_deleted == 0 &&
3267 pa->pa_free >= ac->ac_o_ex.fe_len) {
3269 cpa = ext4_mb_check_group_pa(goal_block,
3272 spin_unlock(&pa->pa_lock);
3277 ext4_mb_use_group_pa(ac, cpa);
3278 ac->ac_criteria = 20;
3285 * the function goes through all block freed in the group
3286 * but not yet committed and marks them used in in-core bitmap.
3287 * buddy must be generated from this bitmap
3288 * Need to be called with the ext4 group lock held
3290 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3294 struct ext4_group_info *grp;
3295 struct ext4_free_data *entry;
3297 grp = ext4_get_group_info(sb, group);
3298 n = rb_first(&(grp->bb_free_root));
3301 entry = rb_entry(n, struct ext4_free_data, node);
3302 mb_set_bits(bitmap, entry->start_blk, entry->count);
3309 * the function goes through all preallocation in this group and marks them
3310 * used in in-core bitmap. buddy must be generated from this bitmap
3311 * Need to be called with ext4 group lock held
3313 static noinline_for_stack
3314 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3317 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3318 struct ext4_prealloc_space *pa;
3319 struct list_head *cur;
3320 ext4_group_t groupnr;
3321 ext4_grpblk_t start;
3322 int preallocated = 0;
3326 /* all form of preallocation discards first load group,
3327 * so the only competing code is preallocation use.
3328 * we don't need any locking here
3329 * notice we do NOT ignore preallocations with pa_deleted
3330 * otherwise we could leave used blocks available for
3331 * allocation in buddy when concurrent ext4_mb_put_pa()
3332 * is dropping preallocation
3334 list_for_each(cur, &grp->bb_prealloc_list) {
3335 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3336 spin_lock(&pa->pa_lock);
3337 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3340 spin_unlock(&pa->pa_lock);
3341 if (unlikely(len == 0))
3343 BUG_ON(groupnr != group);
3344 mb_set_bits(bitmap, start, len);
3345 preallocated += len;
3348 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3351 static void ext4_mb_pa_callback(struct rcu_head *head)
3353 struct ext4_prealloc_space *pa;
3354 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3355 kmem_cache_free(ext4_pspace_cachep, pa);
3359 * drops a reference to preallocated space descriptor
3360 * if this was the last reference and the space is consumed
3362 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3363 struct super_block *sb, struct ext4_prealloc_space *pa)
3366 ext4_fsblk_t grp_blk;
3368 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3371 /* in this short window concurrent discard can set pa_deleted */
3372 spin_lock(&pa->pa_lock);
3373 if (pa->pa_deleted == 1) {
3374 spin_unlock(&pa->pa_lock);
3379 spin_unlock(&pa->pa_lock);
3381 grp_blk = pa->pa_pstart;
3383 * If doing group-based preallocation, pa_pstart may be in the
3384 * next group when pa is used up
3386 if (pa->pa_type == MB_GROUP_PA)
3389 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3394 * P1 (buddy init) P2 (regular allocation)
3395 * find block B in PA
3396 * copy on-disk bitmap to buddy
3397 * mark B in on-disk bitmap
3398 * drop PA from group
3399 * mark all PAs in buddy
3401 * thus, P1 initializes buddy with B available. to prevent this
3402 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3405 ext4_lock_group(sb, grp);
3406 list_del(&pa->pa_group_list);
3407 ext4_unlock_group(sb, grp);
3409 spin_lock(pa->pa_obj_lock);
3410 list_del_rcu(&pa->pa_inode_list);
3411 spin_unlock(pa->pa_obj_lock);
3413 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3417 * creates new preallocated space for given inode
3419 static noinline_for_stack int
3420 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3422 struct super_block *sb = ac->ac_sb;
3423 struct ext4_prealloc_space *pa;
3424 struct ext4_group_info *grp;
3425 struct ext4_inode_info *ei;
3427 /* preallocate only when found space is larger then requested */
3428 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3429 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3430 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3432 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3436 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3442 /* we can't allocate as much as normalizer wants.
3443 * so, found space must get proper lstart
3444 * to cover original request */
3445 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3446 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3448 /* we're limited by original request in that
3449 * logical block must be covered any way
3450 * winl is window we can move our chunk within */
3451 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3453 /* also, we should cover whole original request */
3454 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3456 /* the smallest one defines real window */
3457 win = min(winl, wins);
3459 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3460 if (offs && offs < win)
3463 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3464 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3465 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3468 /* preallocation can change ac_b_ex, thus we store actually
3469 * allocated blocks for history */
3470 ac->ac_f_ex = ac->ac_b_ex;
3472 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3473 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3474 pa->pa_len = ac->ac_b_ex.fe_len;
3475 pa->pa_free = pa->pa_len;
3476 atomic_set(&pa->pa_count, 1);
3477 spin_lock_init(&pa->pa_lock);
3478 INIT_LIST_HEAD(&pa->pa_inode_list);
3479 INIT_LIST_HEAD(&pa->pa_group_list);
3481 pa->pa_type = MB_INODE_PA;
3483 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3484 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3485 trace_ext4_mb_new_inode_pa(ac, pa);
3487 ext4_mb_use_inode_pa(ac, pa);
3488 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3490 ei = EXT4_I(ac->ac_inode);
3491 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3493 pa->pa_obj_lock = &ei->i_prealloc_lock;
3494 pa->pa_inode = ac->ac_inode;
3496 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3497 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3498 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3500 spin_lock(pa->pa_obj_lock);
3501 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3502 spin_unlock(pa->pa_obj_lock);
3508 * creates new preallocated space for locality group inodes belongs to
3510 static noinline_for_stack int
3511 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3513 struct super_block *sb = ac->ac_sb;
3514 struct ext4_locality_group *lg;
3515 struct ext4_prealloc_space *pa;
3516 struct ext4_group_info *grp;
3518 /* preallocate only when found space is larger then requested */
3519 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3520 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3521 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3523 BUG_ON(ext4_pspace_cachep == NULL);
3524 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3528 /* preallocation can change ac_b_ex, thus we store actually
3529 * allocated blocks for history */
3530 ac->ac_f_ex = ac->ac_b_ex;
3532 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3533 pa->pa_lstart = pa->pa_pstart;
3534 pa->pa_len = ac->ac_b_ex.fe_len;
3535 pa->pa_free = pa->pa_len;
3536 atomic_set(&pa->pa_count, 1);
3537 spin_lock_init(&pa->pa_lock);
3538 INIT_LIST_HEAD(&pa->pa_inode_list);
3539 INIT_LIST_HEAD(&pa->pa_group_list);
3541 pa->pa_type = MB_GROUP_PA;
3543 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3544 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3545 trace_ext4_mb_new_group_pa(ac, pa);
3547 ext4_mb_use_group_pa(ac, pa);
3548 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3550 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3554 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3555 pa->pa_inode = NULL;
3557 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3558 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3559 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3562 * We will later add the new pa to the right bucket
3563 * after updating the pa_free in ext4_mb_release_context
3568 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3572 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3573 err = ext4_mb_new_group_pa(ac);
3575 err = ext4_mb_new_inode_pa(ac);
3580 * finds all unused blocks in on-disk bitmap, frees them in
3581 * in-core bitmap and buddy.
3582 * @pa must be unlinked from inode and group lists, so that
3583 * nobody else can find/use it.
3584 * the caller MUST hold group/inode locks.
3585 * TODO: optimize the case when there are no in-core structures yet
3587 static noinline_for_stack int
3588 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3589 struct ext4_prealloc_space *pa)
3591 struct super_block *sb = e4b->bd_sb;
3592 struct ext4_sb_info *sbi = EXT4_SB(sb);
3597 unsigned long long grp_blk_start;
3601 BUG_ON(pa->pa_deleted == 0);
3602 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3603 grp_blk_start = pa->pa_pstart - bit;
3604 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3605 end = bit + pa->pa_len;
3608 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3611 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3612 mb_debug(1, " free preallocated %u/%u in group %u\n",
3613 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3614 (unsigned) next - bit, (unsigned) group);
3617 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3618 trace_ext4_mb_release_inode_pa(sb, pa->pa_inode, pa,
3619 grp_blk_start + bit, next - bit);
3620 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3623 if (free != pa->pa_free) {
3624 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3625 pa, (unsigned long) pa->pa_lstart,
3626 (unsigned long) pa->pa_pstart,
3627 (unsigned long) pa->pa_len);
3628 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3631 * pa is already deleted so we use the value obtained
3632 * from the bitmap and continue.
3635 atomic_add(free, &sbi->s_mb_discarded);
3640 static noinline_for_stack int
3641 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3642 struct ext4_prealloc_space *pa)
3644 struct super_block *sb = e4b->bd_sb;
3648 trace_ext4_mb_release_group_pa(sb, pa);
3649 BUG_ON(pa->pa_deleted == 0);
3650 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3651 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3652 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3653 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3654 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3660 * releases all preallocations in given group
3662 * first, we need to decide discard policy:
3663 * - when do we discard
3665 * - how many do we discard
3666 * 1) how many requested
3668 static noinline_for_stack int
3669 ext4_mb_discard_group_preallocations(struct super_block *sb,
3670 ext4_group_t group, int needed)
3672 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3673 struct buffer_head *bitmap_bh = NULL;
3674 struct ext4_prealloc_space *pa, *tmp;
3675 struct list_head list;
3676 struct ext4_buddy e4b;
3681 mb_debug(1, "discard preallocation for group %u\n", group);
3683 if (list_empty(&grp->bb_prealloc_list))
3686 bitmap_bh = ext4_read_block_bitmap(sb, group);
3687 if (bitmap_bh == NULL) {
3688 ext4_error(sb, "Error reading block bitmap for %u", group);
3692 err = ext4_mb_load_buddy(sb, group, &e4b);
3694 ext4_error(sb, "Error loading buddy information for %u", group);
3700 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3702 INIT_LIST_HEAD(&list);
3704 ext4_lock_group(sb, group);
3705 list_for_each_entry_safe(pa, tmp,
3706 &grp->bb_prealloc_list, pa_group_list) {
3707 spin_lock(&pa->pa_lock);
3708 if (atomic_read(&pa->pa_count)) {
3709 spin_unlock(&pa->pa_lock);
3713 if (pa->pa_deleted) {
3714 spin_unlock(&pa->pa_lock);
3718 /* seems this one can be freed ... */
3721 /* we can trust pa_free ... */
3722 free += pa->pa_free;
3724 spin_unlock(&pa->pa_lock);
3726 list_del(&pa->pa_group_list);
3727 list_add(&pa->u.pa_tmp_list, &list);
3730 /* if we still need more blocks and some PAs were used, try again */
3731 if (free < needed && busy) {
3733 ext4_unlock_group(sb, group);
3735 * Yield the CPU here so that we don't get soft lockup
3736 * in non preempt case.
3742 /* found anything to free? */
3743 if (list_empty(&list)) {
3748 /* now free all selected PAs */
3749 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3751 /* remove from object (inode or locality group) */
3752 spin_lock(pa->pa_obj_lock);
3753 list_del_rcu(&pa->pa_inode_list);
3754 spin_unlock(pa->pa_obj_lock);
3756 if (pa->pa_type == MB_GROUP_PA)
3757 ext4_mb_release_group_pa(&e4b, pa);
3759 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3761 list_del(&pa->u.pa_tmp_list);
3762 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3766 ext4_unlock_group(sb, group);
3767 ext4_mb_unload_buddy(&e4b);
3773 * releases all non-used preallocated blocks for given inode
3775 * It's important to discard preallocations under i_data_sem
3776 * We don't want another block to be served from the prealloc
3777 * space when we are discarding the inode prealloc space.
3779 * FIXME!! Make sure it is valid at all the call sites
3781 void ext4_discard_preallocations(struct inode *inode)
3783 struct ext4_inode_info *ei = EXT4_I(inode);
3784 struct super_block *sb = inode->i_sb;
3785 struct buffer_head *bitmap_bh = NULL;
3786 struct ext4_prealloc_space *pa, *tmp;
3787 ext4_group_t group = 0;
3788 struct list_head list;
3789 struct ext4_buddy e4b;
3792 if (!S_ISREG(inode->i_mode)) {
3793 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3797 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3798 trace_ext4_discard_preallocations(inode);
3800 INIT_LIST_HEAD(&list);
3803 /* first, collect all pa's in the inode */
3804 spin_lock(&ei->i_prealloc_lock);
3805 while (!list_empty(&ei->i_prealloc_list)) {
3806 pa = list_entry(ei->i_prealloc_list.next,
3807 struct ext4_prealloc_space, pa_inode_list);
3808 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3809 spin_lock(&pa->pa_lock);
3810 if (atomic_read(&pa->pa_count)) {
3811 /* this shouldn't happen often - nobody should
3812 * use preallocation while we're discarding it */
3813 spin_unlock(&pa->pa_lock);
3814 spin_unlock(&ei->i_prealloc_lock);
3815 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3817 schedule_timeout_uninterruptible(HZ);
3821 if (pa->pa_deleted == 0) {
3823 spin_unlock(&pa->pa_lock);
3824 list_del_rcu(&pa->pa_inode_list);
3825 list_add(&pa->u.pa_tmp_list, &list);
3829 /* someone is deleting pa right now */
3830 spin_unlock(&pa->pa_lock);
3831 spin_unlock(&ei->i_prealloc_lock);
3833 /* we have to wait here because pa_deleted
3834 * doesn't mean pa is already unlinked from
3835 * the list. as we might be called from
3836 * ->clear_inode() the inode will get freed
3837 * and concurrent thread which is unlinking
3838 * pa from inode's list may access already
3839 * freed memory, bad-bad-bad */
3841 /* XXX: if this happens too often, we can
3842 * add a flag to force wait only in case
3843 * of ->clear_inode(), but not in case of
3844 * regular truncate */
3845 schedule_timeout_uninterruptible(HZ);
3848 spin_unlock(&ei->i_prealloc_lock);
3850 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3851 BUG_ON(pa->pa_type != MB_INODE_PA);
3852 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3854 err = ext4_mb_load_buddy(sb, group, &e4b);
3856 ext4_error(sb, "Error loading buddy information for %u",
3861 bitmap_bh = ext4_read_block_bitmap(sb, group);
3862 if (bitmap_bh == NULL) {
3863 ext4_error(sb, "Error reading block bitmap for %u",
3865 ext4_mb_unload_buddy(&e4b);
3869 ext4_lock_group(sb, group);
3870 list_del(&pa->pa_group_list);
3871 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3872 ext4_unlock_group(sb, group);
3874 ext4_mb_unload_buddy(&e4b);
3877 list_del(&pa->u.pa_tmp_list);
3878 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3883 * finds all preallocated spaces and return blocks being freed to them
3884 * if preallocated space becomes full (no block is used from the space)
3885 * then the function frees space in buddy
3886 * XXX: at the moment, truncate (which is the only way to free blocks)
3887 * discards all preallocations
3889 static void ext4_mb_return_to_preallocation(struct inode *inode,
3890 struct ext4_buddy *e4b,
3891 sector_t block, int count)
3893 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
3895 #ifdef CONFIG_EXT4_DEBUG
3896 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3898 struct super_block *sb = ac->ac_sb;
3899 ext4_group_t ngroups, i;
3901 if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
3904 printk(KERN_ERR "EXT4-fs: Can't allocate:"
3905 " Allocation context details:\n");
3906 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
3907 ac->ac_status, ac->ac_flags);
3908 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3909 "best %lu/%lu/%lu@%lu cr %d\n",
3910 (unsigned long)ac->ac_o_ex.fe_group,
3911 (unsigned long)ac->ac_o_ex.fe_start,
3912 (unsigned long)ac->ac_o_ex.fe_len,
3913 (unsigned long)ac->ac_o_ex.fe_logical,
3914 (unsigned long)ac->ac_g_ex.fe_group,
3915 (unsigned long)ac->ac_g_ex.fe_start,
3916 (unsigned long)ac->ac_g_ex.fe_len,
3917 (unsigned long)ac->ac_g_ex.fe_logical,
3918 (unsigned long)ac->ac_b_ex.fe_group,
3919 (unsigned long)ac->ac_b_ex.fe_start,
3920 (unsigned long)ac->ac_b_ex.fe_len,
3921 (unsigned long)ac->ac_b_ex.fe_logical,
3922 (int)ac->ac_criteria);
3923 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
3925 printk(KERN_ERR "EXT4-fs: groups: \n");
3926 ngroups = ext4_get_groups_count(sb);
3927 for (i = 0; i < ngroups; i++) {
3928 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3929 struct ext4_prealloc_space *pa;
3930 ext4_grpblk_t start;
3931 struct list_head *cur;
3932 ext4_lock_group(sb, i);
3933 list_for_each(cur, &grp->bb_prealloc_list) {
3934 pa = list_entry(cur, struct ext4_prealloc_space,
3936 spin_lock(&pa->pa_lock);
3937 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3939 spin_unlock(&pa->pa_lock);
3940 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3943 ext4_unlock_group(sb, i);
3945 if (grp->bb_free == 0)
3947 printk(KERN_ERR "%u: %d/%d \n",
3948 i, grp->bb_free, grp->bb_fragments);
3950 printk(KERN_ERR "\n");
3953 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3960 * We use locality group preallocation for small size file. The size of the
3961 * file is determined by the current size or the resulting size after
3962 * allocation which ever is larger
3964 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3966 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3968 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3969 int bsbits = ac->ac_sb->s_blocksize_bits;
3972 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3975 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3978 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3979 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3982 if ((size == isize) &&
3983 !ext4_fs_is_busy(sbi) &&
3984 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3985 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3989 /* don't use group allocation for large files */
3990 size = max(size, isize);
3991 if (size > sbi->s_mb_stream_request) {
3992 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3996 BUG_ON(ac->ac_lg != NULL);
3998 * locality group prealloc space are per cpu. The reason for having
3999 * per cpu locality group is to reduce the contention between block
4000 * request from multiple CPUs.
4002 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
4004 /* we're going to use group allocation */
4005 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4007 /* serialize all allocations in the group */
4008 mutex_lock(&ac->ac_lg->lg_mutex);
4011 static noinline_for_stack int
4012 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4013 struct ext4_allocation_request *ar)
4015 struct super_block *sb = ar->inode->i_sb;
4016 struct ext4_sb_info *sbi = EXT4_SB(sb);
4017 struct ext4_super_block *es = sbi->s_es;
4021 ext4_grpblk_t block;
4023 /* we can't allocate > group size */
4026 /* just a dirty hack to filter too big requests */
4027 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4028 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4030 /* start searching from the goal */
4032 if (goal < le32_to_cpu(es->s_first_data_block) ||
4033 goal >= ext4_blocks_count(es))
4034 goal = le32_to_cpu(es->s_first_data_block);
4035 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4037 /* set up allocation goals */
4038 memset(ac, 0, sizeof(struct ext4_allocation_context));
4039 ac->ac_b_ex.fe_logical = ar->logical;
4040 ac->ac_status = AC_STATUS_CONTINUE;
4042 ac->ac_inode = ar->inode;
4043 ac->ac_o_ex.fe_logical = ar->logical;
4044 ac->ac_o_ex.fe_group = group;
4045 ac->ac_o_ex.fe_start = block;
4046 ac->ac_o_ex.fe_len = len;
4047 ac->ac_g_ex.fe_logical = ar->logical;
4048 ac->ac_g_ex.fe_group = group;
4049 ac->ac_g_ex.fe_start = block;
4050 ac->ac_g_ex.fe_len = len;
4051 ac->ac_flags = ar->flags;
4053 /* we have to define context: we'll we work with a file or
4054 * locality group. this is a policy, actually */
4055 ext4_mb_group_or_file(ac);
4057 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4058 "left: %u/%u, right %u/%u to %swritable\n",
4059 (unsigned) ar->len, (unsigned) ar->logical,
4060 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4061 (unsigned) ar->lleft, (unsigned) ar->pleft,
4062 (unsigned) ar->lright, (unsigned) ar->pright,
4063 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4068 static noinline_for_stack void
4069 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4070 struct ext4_locality_group *lg,
4071 int order, int total_entries)
4073 ext4_group_t group = 0;
4074 struct ext4_buddy e4b;
4075 struct list_head discard_list;
4076 struct ext4_prealloc_space *pa, *tmp;
4078 mb_debug(1, "discard locality group preallocation\n");
4080 INIT_LIST_HEAD(&discard_list);
4082 spin_lock(&lg->lg_prealloc_lock);
4083 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4085 spin_lock(&pa->pa_lock);
4086 if (atomic_read(&pa->pa_count)) {
4088 * This is the pa that we just used
4089 * for block allocation. So don't
4092 spin_unlock(&pa->pa_lock);
4095 if (pa->pa_deleted) {
4096 spin_unlock(&pa->pa_lock);
4099 /* only lg prealloc space */
4100 BUG_ON(pa->pa_type != MB_GROUP_PA);
4102 /* seems this one can be freed ... */
4104 spin_unlock(&pa->pa_lock);
4106 list_del_rcu(&pa->pa_inode_list);
4107 list_add(&pa->u.pa_tmp_list, &discard_list);
4110 if (total_entries <= 5) {
4112 * we want to keep only 5 entries
4113 * allowing it to grow to 8. This
4114 * mak sure we don't call discard
4115 * soon for this list.
4120 spin_unlock(&lg->lg_prealloc_lock);
4122 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4124 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4125 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4126 ext4_error(sb, "Error loading buddy information for %u",
4130 ext4_lock_group(sb, group);
4131 list_del(&pa->pa_group_list);
4132 ext4_mb_release_group_pa(&e4b, pa);
4133 ext4_unlock_group(sb, group);
4135 ext4_mb_unload_buddy(&e4b);
4136 list_del(&pa->u.pa_tmp_list);
4137 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4142 * We have incremented pa_count. So it cannot be freed at this
4143 * point. Also we hold lg_mutex. So no parallel allocation is
4144 * possible from this lg. That means pa_free cannot be updated.
4146 * A parallel ext4_mb_discard_group_preallocations is possible.
4147 * which can cause the lg_prealloc_list to be updated.
4150 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4152 int order, added = 0, lg_prealloc_count = 1;
4153 struct super_block *sb = ac->ac_sb;
4154 struct ext4_locality_group *lg = ac->ac_lg;
4155 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4157 order = fls(pa->pa_free) - 1;
4158 if (order > PREALLOC_TB_SIZE - 1)
4159 /* The max size of hash table is PREALLOC_TB_SIZE */
4160 order = PREALLOC_TB_SIZE - 1;
4161 /* Add the prealloc space to lg */
4163 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4165 spin_lock(&tmp_pa->pa_lock);
4166 if (tmp_pa->pa_deleted) {
4167 spin_unlock(&tmp_pa->pa_lock);
4170 if (!added && pa->pa_free < tmp_pa->pa_free) {
4171 /* Add to the tail of the previous entry */
4172 list_add_tail_rcu(&pa->pa_inode_list,
4173 &tmp_pa->pa_inode_list);
4176 * we want to count the total
4177 * number of entries in the list
4180 spin_unlock(&tmp_pa->pa_lock);
4181 lg_prealloc_count++;
4184 list_add_tail_rcu(&pa->pa_inode_list,
4185 &lg->lg_prealloc_list[order]);
4188 /* Now trim the list to be not more than 8 elements */
4189 if (lg_prealloc_count > 8) {
4190 ext4_mb_discard_lg_preallocations(sb, lg,
4191 order, lg_prealloc_count);
4198 * release all resource we used in allocation
4200 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4202 struct ext4_prealloc_space *pa = ac->ac_pa;
4204 if (pa->pa_type == MB_GROUP_PA) {
4205 /* see comment in ext4_mb_use_group_pa() */
4206 spin_lock(&pa->pa_lock);
4207 pa->pa_pstart += ac->ac_b_ex.fe_len;
4208 pa->pa_lstart += ac->ac_b_ex.fe_len;
4209 pa->pa_free -= ac->ac_b_ex.fe_len;
4210 pa->pa_len -= ac->ac_b_ex.fe_len;
4211 spin_unlock(&pa->pa_lock);
4215 up_read(ac->alloc_semp);
4218 * We want to add the pa to the right bucket.
4219 * Remove it from the list and while adding
4220 * make sure the list to which we are adding
4221 * doesn't grow big. We need to release
4222 * alloc_semp before calling ext4_mb_add_n_trim()
4224 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4225 spin_lock(pa->pa_obj_lock);
4226 list_del_rcu(&pa->pa_inode_list);
4227 spin_unlock(pa->pa_obj_lock);
4228 ext4_mb_add_n_trim(ac);
4230 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4232 if (ac->ac_bitmap_page)
4233 page_cache_release(ac->ac_bitmap_page);
4234 if (ac->ac_buddy_page)
4235 page_cache_release(ac->ac_buddy_page);
4236 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4237 mutex_unlock(&ac->ac_lg->lg_mutex);
4238 ext4_mb_collect_stats(ac);
4242 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4244 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4248 trace_ext4_mb_discard_preallocations(sb, needed);
4249 for (i = 0; i < ngroups && needed > 0; i++) {
4250 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4259 * Main entry point into mballoc to allocate blocks
4260 * it tries to use preallocation first, then falls back
4261 * to usual allocation
4263 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4264 struct ext4_allocation_request *ar, int *errp)
4267 struct ext4_allocation_context *ac = NULL;
4268 struct ext4_sb_info *sbi;
4269 struct super_block *sb;
4270 ext4_fsblk_t block = 0;
4271 unsigned int inquota = 0;
4272 unsigned int reserv_blks = 0;
4274 sb = ar->inode->i_sb;
4277 trace_ext4_request_blocks(ar);
4280 * For delayed allocation, we could skip the ENOSPC and
4281 * EDQUOT check, as blocks and quotas have been already
4282 * reserved when data being copied into pagecache.
4284 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4285 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4287 /* Without delayed allocation we need to verify
4288 * there is enough free blocks to do block allocation
4289 * and verify allocation doesn't exceed the quota limits.
4291 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4292 /* let others to free the space */
4294 ar->len = ar->len >> 1;
4300 reserv_blks = ar->len;
4301 while (ar->len && dquot_alloc_block(ar->inode, ar->len)) {
4302 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4312 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4319 *errp = ext4_mb_initialize_context(ac, ar);
4325 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4326 if (!ext4_mb_use_preallocated(ac)) {
4327 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4328 ext4_mb_normalize_request(ac, ar);
4330 /* allocate space in core */
4331 *errp = ext4_mb_regular_allocator(ac);
4335 /* as we've just preallocated more space than
4336 * user requested orinally, we store allocated
4337 * space in a special descriptor */
4338 if (ac->ac_status == AC_STATUS_FOUND &&
4339 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4340 ext4_mb_new_preallocation(ac);
4342 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4343 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4344 if (*errp == -EAGAIN) {
4346 * drop the reference that we took
4347 * in ext4_mb_use_best_found
4349 ext4_mb_release_context(ac);
4350 ac->ac_b_ex.fe_group = 0;
4351 ac->ac_b_ex.fe_start = 0;
4352 ac->ac_b_ex.fe_len = 0;
4353 ac->ac_status = AC_STATUS_CONTINUE;
4357 ext4_discard_allocated_blocks(ac);
4359 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4360 ar->len = ac->ac_b_ex.fe_len;
4363 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4370 ac->ac_b_ex.fe_len = 0;
4372 ext4_mb_show_ac(ac);
4374 ext4_mb_release_context(ac);
4377 kmem_cache_free(ext4_ac_cachep, ac);
4378 if (inquota && ar->len < inquota)
4379 dquot_free_block(ar->inode, inquota - ar->len);
4381 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4382 /* release all the reserved blocks if non delalloc */
4383 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4387 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4393 * We can merge two free data extents only if the physical blocks
4394 * are contiguous, AND the extents were freed by the same transaction,
4395 * AND the blocks are associated with the same group.
4397 static int can_merge(struct ext4_free_data *entry1,
4398 struct ext4_free_data *entry2)
4400 if ((entry1->t_tid == entry2->t_tid) &&
4401 (entry1->group == entry2->group) &&
4402 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4407 static noinline_for_stack int
4408 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4409 struct ext4_free_data *new_entry)
4411 ext4_group_t group = e4b->bd_group;
4412 ext4_grpblk_t block;
4413 struct ext4_free_data *entry;
4414 struct ext4_group_info *db = e4b->bd_info;
4415 struct super_block *sb = e4b->bd_sb;
4416 struct ext4_sb_info *sbi = EXT4_SB(sb);
4417 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4418 struct rb_node *parent = NULL, *new_node;
4420 BUG_ON(!ext4_handle_valid(handle));
4421 BUG_ON(e4b->bd_bitmap_page == NULL);
4422 BUG_ON(e4b->bd_buddy_page == NULL);
4424 new_node = &new_entry->node;
4425 block = new_entry->start_blk;
4428 /* first free block exent. We need to
4429 protect buddy cache from being freed,
4430 * otherwise we'll refresh it from
4431 * on-disk bitmap and lose not-yet-available
4433 page_cache_get(e4b->bd_buddy_page);
4434 page_cache_get(e4b->bd_bitmap_page);
4438 entry = rb_entry(parent, struct ext4_free_data, node);
4439 if (block < entry->start_blk)
4441 else if (block >= (entry->start_blk + entry->count))
4442 n = &(*n)->rb_right;
4444 ext4_grp_locked_error(sb, group, 0,
4445 ext4_group_first_block_no(sb, group) + block,
4446 "Block already on to-be-freed list");
4451 rb_link_node(new_node, parent, n);
4452 rb_insert_color(new_node, &db->bb_free_root);
4454 /* Now try to see the extent can be merged to left and right */
4455 node = rb_prev(new_node);
4457 entry = rb_entry(node, struct ext4_free_data, node);
4458 if (can_merge(entry, new_entry)) {
4459 new_entry->start_blk = entry->start_blk;
4460 new_entry->count += entry->count;
4461 rb_erase(node, &(db->bb_free_root));
4462 spin_lock(&sbi->s_md_lock);
4463 list_del(&entry->list);
4464 spin_unlock(&sbi->s_md_lock);
4465 kmem_cache_free(ext4_free_ext_cachep, entry);
4469 node = rb_next(new_node);
4471 entry = rb_entry(node, struct ext4_free_data, node);
4472 if (can_merge(new_entry, entry)) {
4473 new_entry->count += entry->count;
4474 rb_erase(node, &(db->bb_free_root));
4475 spin_lock(&sbi->s_md_lock);
4476 list_del(&entry->list);
4477 spin_unlock(&sbi->s_md_lock);
4478 kmem_cache_free(ext4_free_ext_cachep, entry);
4481 /* Add the extent to transaction's private list */
4482 spin_lock(&sbi->s_md_lock);
4483 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4484 spin_unlock(&sbi->s_md_lock);
4489 * ext4_free_blocks() -- Free given blocks and update quota
4490 * @handle: handle for this transaction
4492 * @block: start physical block to free
4493 * @count: number of blocks to count
4494 * @metadata: Are these metadata blocks
4496 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4497 struct buffer_head *bh, ext4_fsblk_t block,
4498 unsigned long count, int flags)
4500 struct buffer_head *bitmap_bh = NULL;
4501 struct super_block *sb = inode->i_sb;
4502 struct ext4_group_desc *gdp;
4503 unsigned long freed = 0;
4504 unsigned int overflow;
4506 struct buffer_head *gd_bh;
4507 ext4_group_t block_group;
4508 struct ext4_sb_info *sbi;
4509 struct ext4_buddy e4b;
4515 BUG_ON(block != bh->b_blocknr);
4517 block = bh->b_blocknr;
4521 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4522 !ext4_data_block_valid(sbi, block, count)) {
4523 ext4_error(sb, "Freeing blocks not in datazone - "
4524 "block = %llu, count = %lu", block, count);
4528 ext4_debug("freeing block %llu\n", block);
4529 trace_ext4_free_blocks(inode, block, count, flags);
4531 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4532 struct buffer_head *tbh = bh;
4535 BUG_ON(bh && (count > 1));
4537 for (i = 0; i < count; i++) {
4539 tbh = sb_find_get_block(inode->i_sb,
4543 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4544 inode, tbh, block + i);
4549 * We need to make sure we don't reuse the freed block until
4550 * after the transaction is committed, which we can do by
4551 * treating the block as metadata, below. We make an
4552 * exception if the inode is to be written in writeback mode
4553 * since writeback mode has weak data consistency guarantees.
4555 if (!ext4_should_writeback_data(inode))
4556 flags |= EXT4_FREE_BLOCKS_METADATA;
4560 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4563 * Check to see if we are freeing blocks across a group
4566 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4567 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4570 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4575 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4581 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4582 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4583 in_range(block, ext4_inode_table(sb, gdp),
4584 EXT4_SB(sb)->s_itb_per_group) ||
4585 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4586 EXT4_SB(sb)->s_itb_per_group)) {
4588 ext4_error(sb, "Freeing blocks in system zone - "
4589 "Block = %llu, count = %lu", block, count);
4590 /* err = 0. ext4_std_error should be a no op */
4594 BUFFER_TRACE(bitmap_bh, "getting write access");
4595 err = ext4_journal_get_write_access(handle, bitmap_bh);
4600 * We are about to modify some metadata. Call the journal APIs
4601 * to unshare ->b_data if a currently-committing transaction is
4604 BUFFER_TRACE(gd_bh, "get_write_access");
4605 err = ext4_journal_get_write_access(handle, gd_bh);
4608 #ifdef AGGRESSIVE_CHECK
4611 for (i = 0; i < count; i++)
4612 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4615 trace_ext4_mballoc_free(sb, inode, block_group, bit, count);
4617 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4621 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4622 struct ext4_free_data *new_entry;
4624 * blocks being freed are metadata. these blocks shouldn't
4625 * be used until this transaction is committed
4627 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4628 new_entry->start_blk = bit;
4629 new_entry->group = block_group;
4630 new_entry->count = count;
4631 new_entry->t_tid = handle->h_transaction->t_tid;
4633 ext4_lock_group(sb, block_group);
4634 mb_clear_bits(bitmap_bh->b_data, bit, count);
4635 ext4_mb_free_metadata(handle, &e4b, new_entry);
4637 /* need to update group_info->bb_free and bitmap
4638 * with group lock held. generate_buddy look at
4639 * them with group lock_held
4641 if (test_opt(sb, DISCARD))
4642 ext4_issue_discard(sb, block_group, bit, count);
4643 ext4_lock_group(sb, block_group);
4644 mb_clear_bits(bitmap_bh->b_data, bit, count);
4645 mb_free_blocks(inode, &e4b, bit, count);
4646 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4649 ret = ext4_free_blks_count(sb, gdp) + count;
4650 ext4_free_blks_set(sb, gdp, ret);
4651 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4652 ext4_unlock_group(sb, block_group);
4653 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4655 if (sbi->s_log_groups_per_flex) {
4656 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4657 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4660 ext4_mb_unload_buddy(&e4b);
4664 /* We dirtied the bitmap block */
4665 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4666 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4668 /* And the group descriptor block */
4669 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4670 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4674 if (overflow && !err) {
4680 ext4_mark_super_dirty(sb);
4683 dquot_free_block(inode, freed);
4685 ext4_std_error(sb, err);
4690 * ext4_trim_extent -- function to TRIM one single free extent in the group
4691 * @sb: super block for the file system
4692 * @start: starting block of the free extent in the alloc. group
4693 * @count: number of blocks to TRIM
4694 * @group: alloc. group we are working with
4695 * @e4b: ext4 buddy for the group
4697 * Trim "count" blocks starting at "start" in the "group". To assure that no
4698 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4699 * be called with under the group lock.
4701 static int ext4_trim_extent(struct super_block *sb, int start, int count,
4702 ext4_group_t group, struct ext4_buddy *e4b)
4704 struct ext4_free_extent ex;
4707 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4709 ex.fe_start = start;
4710 ex.fe_group = group;
4714 * Mark blocks used, so no one can reuse them while
4717 mb_mark_used(e4b, &ex);
4718 ext4_unlock_group(sb, group);
4720 ret = ext4_issue_discard(sb, group, start, count);
4722 ext4_std_error(sb, ret);
4724 ext4_lock_group(sb, group);
4725 mb_free_blocks(NULL, e4b, start, ex.fe_len);
4730 * ext4_trim_all_free -- function to trim all free space in alloc. group
4731 * @sb: super block for file system
4733 * @start: first group block to examine
4734 * @max: last group block to examine
4735 * @minblocks: minimum extent block count
4737 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4738 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4742 * ext4_trim_all_free walks through group's block bitmap searching for free
4743 * extents. When the free extent is found, mark it as used in group buddy
4744 * bitmap. Then issue a TRIM command on this extent and free the extent in
4745 * the group buddy bitmap. This is done until whole group is scanned.
4747 ext4_grpblk_t ext4_trim_all_free(struct super_block *sb, struct ext4_buddy *e4b,
4748 ext4_grpblk_t start, ext4_grpblk_t max, ext4_grpblk_t minblocks)
4751 ext4_grpblk_t next, count = 0;
4755 BUG_ON(e4b == NULL);
4757 bitmap = e4b->bd_bitmap;
4758 group = e4b->bd_group;
4759 start = (e4b->bd_info->bb_first_free > start) ?
4760 e4b->bd_info->bb_first_free : start;
4761 ext4_lock_group(sb, group);
4763 while (start < max) {
4764 start = mb_find_next_zero_bit(bitmap, max, start);
4767 next = mb_find_next_bit(bitmap, max, start);
4769 if ((next - start) >= minblocks) {
4770 ret = ext4_trim_extent(sb, start,
4771 next - start, group, e4b);
4774 count += next - start;
4778 if (fatal_signal_pending(current)) {
4779 count = -ERESTARTSYS;
4783 if (need_resched()) {
4784 ext4_unlock_group(sb, group);
4786 ext4_lock_group(sb, group);
4789 if ((e4b->bd_info->bb_free - count) < minblocks)
4792 ext4_unlock_group(sb, group);
4794 ext4_debug("trimmed %d blocks in the group %d\n",
4804 * ext4_trim_fs() -- trim ioctl handle function
4805 * @sb: superblock for filesystem
4806 * @range: fstrim_range structure
4808 * start: First Byte to trim
4809 * len: number of Bytes to trim from start
4810 * minlen: minimum extent length in Bytes
4811 * ext4_trim_fs goes through all allocation groups containing Bytes from
4812 * start to start+len. For each such a group ext4_trim_all_free function
4813 * is invoked to trim all free space.
4815 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
4817 struct ext4_buddy e4b;
4818 ext4_group_t first_group, last_group;
4819 ext4_group_t group, ngroups = ext4_get_groups_count(sb);
4820 ext4_grpblk_t cnt = 0, first_block, last_block;
4821 uint64_t start, len, minlen, trimmed;
4824 start = range->start >> sb->s_blocksize_bits;
4825 len = range->len >> sb->s_blocksize_bits;
4826 minlen = range->minlen >> sb->s_blocksize_bits;
4829 if (unlikely(minlen > EXT4_BLOCKS_PER_GROUP(sb)))
4832 /* Determine first and last group to examine based on start and len */
4833 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
4834 &first_group, &first_block);
4835 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) (start + len),
4836 &last_group, &last_block);
4837 last_group = (last_group > ngroups - 1) ? ngroups - 1 : last_group;
4838 last_block = EXT4_BLOCKS_PER_GROUP(sb);
4840 if (first_group > last_group)
4843 for (group = first_group; group <= last_group; group++) {
4844 ret = ext4_mb_load_buddy(sb, group, &e4b);
4846 ext4_error(sb, "Error in loading buddy "
4847 "information for %u", group);
4851 if (len >= EXT4_BLOCKS_PER_GROUP(sb))
4852 len -= (EXT4_BLOCKS_PER_GROUP(sb) - first_block);
4856 if (e4b.bd_info->bb_free >= minlen) {
4857 cnt = ext4_trim_all_free(sb, &e4b, first_block,
4858 last_block, minlen);
4861 ext4_mb_unload_buddy(&e4b);
4865 ext4_mb_unload_buddy(&e4b);
4869 range->len = trimmed * sb->s_blocksize;