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 (in clusters)
74 * pa_free -> free space available in this prealloc space (in clusters)
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 makes sure that
79 * 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 represented 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) within 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. The default value of s_mb_group_prealloc is
130 * dependent on the cluster size; for non-bigalloc file systems, it is
131 * 512 blocks. This can be tuned via
132 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
133 * terms of number of blocks. If we have mounted the file system with -O
134 * stripe=<value> option the group prealloc request is normalized to the
135 * the smallest multiple of the stripe value (sbi->s_stripe) which is
136 * greater than the default mb_group_prealloc.
138 * The regular allocator (using the buddy cache) supports a few tunables.
140 * /sys/fs/ext4/<partition>/mb_min_to_scan
141 * /sys/fs/ext4/<partition>/mb_max_to_scan
142 * /sys/fs/ext4/<partition>/mb_order2_req
144 * The regular allocator uses buddy scan only if the request len is power of
145 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
146 * value of s_mb_order2_reqs can be tuned via
147 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
148 * stripe size (sbi->s_stripe), we try to search for contiguous block in
149 * stripe size. This should result in better allocation on RAID setups. If
150 * not, we search in the specific group using bitmap for best extents. The
151 * tunable min_to_scan and max_to_scan control the behaviour here.
152 * min_to_scan indicate how long the mballoc __must__ look for a best
153 * extent and max_to_scan indicates how long the mballoc __can__ look for a
154 * best extent in the found extents. Searching for the blocks starts with
155 * the group specified as the goal value in allocation context via
156 * ac_g_ex. Each group is first checked based on the criteria whether it
157 * can be used for allocation. ext4_mb_good_group explains how the groups are
160 * Both the prealloc space are getting populated as above. So for the first
161 * request we will hit the buddy cache which will result in this prealloc
162 * space getting filled. The prealloc space is then later used for the
163 * subsequent request.
167 * mballoc operates on the following data:
169 * - in-core buddy (actually includes buddy and bitmap)
170 * - preallocation descriptors (PAs)
172 * there are two types of preallocations:
174 * assiged to specific inode and can be used for this inode only.
175 * it describes part of inode's space preallocated to specific
176 * physical blocks. any block from that preallocated can be used
177 * independent. the descriptor just tracks number of blocks left
178 * unused. so, before taking some block from descriptor, one must
179 * make sure corresponded logical block isn't allocated yet. this
180 * also means that freeing any block within descriptor's range
181 * must discard all preallocated blocks.
183 * assigned to specific locality group which does not translate to
184 * permanent set of inodes: inode can join and leave group. space
185 * from this type of preallocation can be used for any inode. thus
186 * it's consumed from the beginning to the end.
188 * relation between them can be expressed as:
189 * in-core buddy = on-disk bitmap + preallocation descriptors
191 * this mean blocks mballoc considers used are:
192 * - allocated blocks (persistent)
193 * - preallocated blocks (non-persistent)
195 * consistency in mballoc world means that at any time a block is either
196 * free or used in ALL structures. notice: "any time" should not be read
197 * literally -- time is discrete and delimited by locks.
199 * to keep it simple, we don't use block numbers, instead we count number of
200 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
202 * all operations can be expressed as:
203 * - init buddy: buddy = on-disk + PAs
204 * - new PA: buddy += N; PA = N
205 * - use inode PA: on-disk += N; PA -= N
206 * - discard inode PA buddy -= on-disk - PA; PA = 0
207 * - use locality group PA on-disk += N; PA -= N
208 * - discard locality group PA buddy -= PA; PA = 0
209 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
210 * is used in real operation because we can't know actual used
211 * bits from PA, only from on-disk bitmap
213 * if we follow this strict logic, then all operations above should be atomic.
214 * given some of them can block, we'd have to use something like semaphores
215 * killing performance on high-end SMP hardware. let's try to relax it using
216 * the following knowledge:
217 * 1) if buddy is referenced, it's already initialized
218 * 2) while block is used in buddy and the buddy is referenced,
219 * nobody can re-allocate that block
220 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
221 * bit set and PA claims same block, it's OK. IOW, one can set bit in
222 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
225 * so, now we're building a concurrency table:
228 * blocks for PA are allocated in the buddy, buddy must be referenced
229 * until PA is linked to allocation group to avoid concurrent buddy init
231 * we need to make sure that either on-disk bitmap or PA has uptodate data
232 * given (3) we care that PA-=N operation doesn't interfere with init
234 * the simplest way would be to have buddy initialized by the discard
235 * - use locality group PA
236 * again PA-=N must be serialized with init
237 * - discard locality group PA
238 * the simplest way would be to have buddy initialized by the discard
241 * i_data_sem serializes them
243 * discard process must wait until PA isn't used by another process
244 * - use locality group PA
245 * some mutex should serialize them
246 * - discard locality group PA
247 * discard process must wait until PA isn't used by another process
250 * i_data_sem or another mutex should serializes them
252 * discard process must wait until PA isn't used by another process
253 * - use locality group PA
254 * nothing wrong here -- they're different PAs covering different blocks
255 * - discard locality group PA
256 * discard process must wait until PA isn't used by another process
258 * now we're ready to make few consequences:
259 * - PA is referenced and while it is no discard is possible
260 * - PA is referenced until block isn't marked in on-disk bitmap
261 * - PA changes only after on-disk bitmap
262 * - discard must not compete with init. either init is done before
263 * any discard or they're serialized somehow
264 * - buddy init as sum of on-disk bitmap and PAs is done atomically
266 * a special case when we've used PA to emptiness. no need to modify buddy
267 * in this case, but we should care about concurrent init
272 * Logic in few words:
277 * mark bits in on-disk bitmap
280 * - use preallocation:
281 * find proper PA (per-inode or group)
283 * mark bits in on-disk bitmap
289 * mark bits in on-disk bitmap
292 * - discard preallocations in group:
294 * move them onto local list
295 * load on-disk bitmap
297 * remove PA from object (inode or locality group)
298 * mark free blocks in-core
300 * - discard inode's preallocations:
307 * - bitlock on a group (group)
308 * - object (inode/locality) (object)
319 * - release consumed pa:
324 * - generate in-core bitmap:
328 * - discard all for given object (inode, locality group):
333 * - discard all for given group:
340 static struct kmem_cache *ext4_pspace_cachep;
341 static struct kmem_cache *ext4_ac_cachep;
342 static struct kmem_cache *ext4_free_ext_cachep;
344 /* We create slab caches for groupinfo data structures based on the
345 * superblock block size. There will be one per mounted filesystem for
346 * each unique s_blocksize_bits */
347 #define NR_GRPINFO_CACHES 8
348 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
350 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
351 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
352 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
353 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
356 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
358 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
360 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
362 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
364 #if BITS_PER_LONG == 64
365 *bit += ((unsigned long) addr & 7UL) << 3;
366 addr = (void *) ((unsigned long) addr & ~7UL);
367 #elif BITS_PER_LONG == 32
368 *bit += ((unsigned long) addr & 3UL) << 3;
369 addr = (void *) ((unsigned long) addr & ~3UL);
371 #error "how many bits you are?!"
376 static inline int mb_test_bit(int bit, void *addr)
379 * ext4_test_bit on architecture like powerpc
380 * needs unsigned long aligned address
382 addr = mb_correct_addr_and_bit(&bit, addr);
383 return ext4_test_bit(bit, addr);
386 static inline void mb_set_bit(int bit, void *addr)
388 addr = mb_correct_addr_and_bit(&bit, addr);
389 ext4_set_bit(bit, addr);
392 static inline void mb_clear_bit(int bit, void *addr)
394 addr = mb_correct_addr_and_bit(&bit, addr);
395 ext4_clear_bit(bit, addr);
398 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
400 int fix = 0, ret, tmpmax;
401 addr = mb_correct_addr_and_bit(&fix, addr);
405 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
411 static inline int mb_find_next_bit(void *addr, int max, int start)
413 int fix = 0, ret, tmpmax;
414 addr = mb_correct_addr_and_bit(&fix, addr);
418 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
424 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
428 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
431 if (order > e4b->bd_blkbits + 1) {
436 /* at order 0 we see each particular block */
438 *max = 1 << (e4b->bd_blkbits + 3);
439 return EXT4_MB_BITMAP(e4b);
442 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
443 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
449 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
450 int first, int count)
453 struct super_block *sb = e4b->bd_sb;
455 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
457 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
458 for (i = 0; i < count; i++) {
459 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
460 ext4_fsblk_t blocknr;
462 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
463 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
464 ext4_grp_locked_error(sb, e4b->bd_group,
465 inode ? inode->i_ino : 0,
467 "freeing block already freed "
471 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
475 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
479 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
481 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
482 for (i = 0; i < count; i++) {
483 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
484 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
488 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
490 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
491 unsigned char *b1, *b2;
493 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
494 b2 = (unsigned char *) bitmap;
495 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
496 if (b1[i] != b2[i]) {
497 ext4_msg(e4b->bd_sb, KERN_ERR,
498 "corruption in group %u "
499 "at byte %u(%u): %x in copy != %x "
501 e4b->bd_group, i, i * 8, b1[i], b2[i]);
509 static inline void mb_free_blocks_double(struct inode *inode,
510 struct ext4_buddy *e4b, int first, int count)
514 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
515 int first, int count)
519 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
525 #ifdef AGGRESSIVE_CHECK
527 #define MB_CHECK_ASSERT(assert) \
531 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
532 function, file, line, # assert); \
537 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
538 const char *function, int line)
540 struct super_block *sb = e4b->bd_sb;
541 int order = e4b->bd_blkbits + 1;
548 struct ext4_group_info *grp;
551 struct list_head *cur;
556 static int mb_check_counter;
557 if (mb_check_counter++ % 100 != 0)
562 buddy = mb_find_buddy(e4b, order, &max);
563 MB_CHECK_ASSERT(buddy);
564 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
565 MB_CHECK_ASSERT(buddy2);
566 MB_CHECK_ASSERT(buddy != buddy2);
567 MB_CHECK_ASSERT(max * 2 == max2);
570 for (i = 0; i < max; i++) {
572 if (mb_test_bit(i, buddy)) {
573 /* only single bit in buddy2 may be 1 */
574 if (!mb_test_bit(i << 1, buddy2)) {
576 mb_test_bit((i<<1)+1, buddy2));
577 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
579 mb_test_bit(i << 1, buddy2));
584 /* both bits in buddy2 must be 0 */
585 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
586 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
588 for (j = 0; j < (1 << order); j++) {
589 k = (i * (1 << order)) + j;
591 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
595 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
600 buddy = mb_find_buddy(e4b, 0, &max);
601 for (i = 0; i < max; i++) {
602 if (!mb_test_bit(i, buddy)) {
603 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
611 /* check used bits only */
612 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
613 buddy2 = mb_find_buddy(e4b, j, &max2);
615 MB_CHECK_ASSERT(k < max2);
616 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
619 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
620 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
622 grp = ext4_get_group_info(sb, e4b->bd_group);
623 list_for_each(cur, &grp->bb_prealloc_list) {
624 ext4_group_t groupnr;
625 struct ext4_prealloc_space *pa;
626 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
627 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
628 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
629 for (i = 0; i < pa->pa_len; i++)
630 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
634 #undef MB_CHECK_ASSERT
635 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
636 __FILE__, __func__, __LINE__)
638 #define mb_check_buddy(e4b)
642 * Divide blocks started from @first with length @len into
643 * smaller chunks with power of 2 blocks.
644 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
645 * then increase bb_counters[] for corresponded chunk size.
647 static void ext4_mb_mark_free_simple(struct super_block *sb,
648 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
649 struct ext4_group_info *grp)
651 struct ext4_sb_info *sbi = EXT4_SB(sb);
655 unsigned short border;
657 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
659 border = 2 << sb->s_blocksize_bits;
662 /* find how many blocks can be covered since this position */
663 max = ffs(first | border) - 1;
665 /* find how many blocks of power 2 we need to mark */
672 /* mark multiblock chunks only */
673 grp->bb_counters[min]++;
675 mb_clear_bit(first >> min,
676 buddy + sbi->s_mb_offsets[min]);
684 * Cache the order of the largest free extent we have available in this block
688 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
693 grp->bb_largest_free_order = -1; /* uninit */
695 bits = sb->s_blocksize_bits + 1;
696 for (i = bits; i >= 0; i--) {
697 if (grp->bb_counters[i] > 0) {
698 grp->bb_largest_free_order = i;
704 static noinline_for_stack
705 void ext4_mb_generate_buddy(struct super_block *sb,
706 void *buddy, void *bitmap, ext4_group_t group)
708 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
709 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
714 unsigned fragments = 0;
715 unsigned long long period = get_cycles();
717 /* initialize buddy from bitmap which is aggregation
718 * of on-disk bitmap and preallocations */
719 i = mb_find_next_zero_bit(bitmap, max, 0);
720 grp->bb_first_free = i;
724 i = mb_find_next_bit(bitmap, max, i);
728 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
730 grp->bb_counters[0]++;
732 i = mb_find_next_zero_bit(bitmap, max, i);
734 grp->bb_fragments = fragments;
736 if (free != grp->bb_free) {
737 ext4_grp_locked_error(sb, group, 0, 0,
738 "%u clusters in bitmap, %u in gd",
741 * If we intent to continue, we consider group descritor
742 * corrupt and update bb_free using bitmap value
746 mb_set_largest_free_order(sb, grp);
748 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
750 period = get_cycles() - period;
751 spin_lock(&EXT4_SB(sb)->s_bal_lock);
752 EXT4_SB(sb)->s_mb_buddies_generated++;
753 EXT4_SB(sb)->s_mb_generation_time += period;
754 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
757 /* The buddy information is attached the buddy cache inode
758 * for convenience. The information regarding each group
759 * is loaded via ext4_mb_load_buddy. The information involve
760 * block bitmap and buddy information. The information are
761 * stored in the inode as
764 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
767 * one block each for bitmap and buddy information.
768 * So for each group we take up 2 blocks. A page can
769 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
770 * So it can have information regarding groups_per_page which
771 * is blocks_per_page/2
773 * Locking note: This routine takes the block group lock of all groups
774 * for this page; do not hold this lock when calling this routine!
777 static int ext4_mb_init_cache(struct page *page, char *incore)
779 ext4_group_t ngroups;
785 ext4_group_t first_group;
787 struct super_block *sb;
788 struct buffer_head *bhs;
789 struct buffer_head **bh;
793 struct ext4_group_info *grinfo;
795 mb_debug(1, "init page %lu\n", page->index);
797 inode = page->mapping->host;
799 ngroups = ext4_get_groups_count(sb);
800 blocksize = 1 << inode->i_blkbits;
801 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
803 groups_per_page = blocks_per_page >> 1;
804 if (groups_per_page == 0)
807 /* allocate buffer_heads to read bitmaps */
808 if (groups_per_page > 1) {
810 i = sizeof(struct buffer_head *) * groups_per_page;
811 bh = kzalloc(i, GFP_NOFS);
817 first_group = page->index * blocks_per_page / 2;
819 /* read all groups the page covers into the cache */
820 for (i = 0; i < groups_per_page; i++) {
821 struct ext4_group_desc *desc;
823 if (first_group + i >= ngroups)
826 grinfo = ext4_get_group_info(sb, first_group + i);
828 * If page is uptodate then we came here after online resize
829 * which added some new uninitialized group info structs, so
830 * we must skip all initialized uptodate buddies on the page,
831 * which may be currently in use by an allocating task.
833 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
839 desc = ext4_get_group_desc(sb, first_group + i, NULL);
844 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
848 if (bitmap_uptodate(bh[i]))
852 if (bitmap_uptodate(bh[i])) {
853 unlock_buffer(bh[i]);
856 ext4_lock_group(sb, first_group + i);
857 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
858 ext4_init_block_bitmap(sb, bh[i],
859 first_group + i, desc);
860 set_bitmap_uptodate(bh[i]);
861 set_buffer_uptodate(bh[i]);
862 ext4_unlock_group(sb, first_group + i);
863 unlock_buffer(bh[i]);
866 ext4_unlock_group(sb, first_group + i);
867 if (buffer_uptodate(bh[i])) {
869 * if not uninit if bh is uptodate,
870 * bitmap is also uptodate
872 set_bitmap_uptodate(bh[i]);
873 unlock_buffer(bh[i]);
878 * submit the buffer_head for read. We can
879 * safely mark the bitmap as uptodate now.
880 * We do it here so the bitmap uptodate bit
881 * get set with buffer lock held.
883 set_bitmap_uptodate(bh[i]);
884 bh[i]->b_end_io = end_buffer_read_sync;
885 submit_bh(READ, bh[i]);
886 mb_debug(1, "read bitmap for group %u\n", first_group + i);
889 /* wait for I/O completion */
890 for (i = 0; i < groups_per_page; i++)
892 wait_on_buffer(bh[i]);
895 for (i = 0; i < groups_per_page; i++)
896 if (bh[i] && !buffer_uptodate(bh[i]))
900 first_block = page->index * blocks_per_page;
901 for (i = 0; i < blocks_per_page; i++) {
904 group = (first_block + i) >> 1;
905 if (group >= ngroups)
908 if (!bh[group - first_group])
909 /* skip initialized uptodate buddy */
913 * data carry information regarding this
914 * particular group in the format specified
918 data = page_address(page) + (i * blocksize);
919 bitmap = bh[group - first_group]->b_data;
922 * We place the buddy block and bitmap block
925 if ((first_block + i) & 1) {
926 /* this is block of buddy */
927 BUG_ON(incore == NULL);
928 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
929 group, page->index, i * blocksize);
930 trace_ext4_mb_buddy_bitmap_load(sb, group);
931 grinfo = ext4_get_group_info(sb, group);
932 grinfo->bb_fragments = 0;
933 memset(grinfo->bb_counters, 0,
934 sizeof(*grinfo->bb_counters) *
935 (sb->s_blocksize_bits+2));
937 * incore got set to the group block bitmap below
939 ext4_lock_group(sb, group);
941 memset(data, 0xff, blocksize);
942 ext4_mb_generate_buddy(sb, data, incore, group);
943 ext4_unlock_group(sb, group);
946 /* this is block of bitmap */
947 BUG_ON(incore != NULL);
948 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
949 group, page->index, i * blocksize);
950 trace_ext4_mb_bitmap_load(sb, group);
952 /* see comments in ext4_mb_put_pa() */
953 ext4_lock_group(sb, group);
954 memcpy(data, bitmap, blocksize);
956 /* mark all preallocated blks used in in-core bitmap */
957 ext4_mb_generate_from_pa(sb, data, group);
958 ext4_mb_generate_from_freelist(sb, data, group);
959 ext4_unlock_group(sb, group);
961 /* set incore so that the buddy information can be
962 * generated using this
967 SetPageUptodate(page);
971 for (i = 0; i < groups_per_page; i++)
980 * Lock the buddy and bitmap pages. This make sure other parallel init_group
981 * on the same buddy page doesn't happen whild holding the buddy page lock.
982 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
983 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
985 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
986 ext4_group_t group, struct ext4_buddy *e4b)
988 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
989 int block, pnum, poff;
993 e4b->bd_buddy_page = NULL;
994 e4b->bd_bitmap_page = NULL;
996 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
998 * the buddy cache inode stores the block bitmap
999 * and buddy information in consecutive blocks.
1000 * So for each group we need two blocks.
1003 pnum = block / blocks_per_page;
1004 poff = block % blocks_per_page;
1005 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1008 BUG_ON(page->mapping != inode->i_mapping);
1009 e4b->bd_bitmap_page = page;
1010 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1012 if (blocks_per_page >= 2) {
1013 /* buddy and bitmap are on the same page */
1018 pnum = block / blocks_per_page;
1019 poff = block % blocks_per_page;
1020 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1023 BUG_ON(page->mapping != inode->i_mapping);
1024 e4b->bd_buddy_page = page;
1028 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1030 if (e4b->bd_bitmap_page) {
1031 unlock_page(e4b->bd_bitmap_page);
1032 page_cache_release(e4b->bd_bitmap_page);
1034 if (e4b->bd_buddy_page) {
1035 unlock_page(e4b->bd_buddy_page);
1036 page_cache_release(e4b->bd_buddy_page);
1041 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1042 * block group lock of all groups for this page; do not hold the BG lock when
1043 * calling this routine!
1045 static noinline_for_stack
1046 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1049 struct ext4_group_info *this_grp;
1050 struct ext4_buddy e4b;
1054 mb_debug(1, "init group %u\n", group);
1055 this_grp = ext4_get_group_info(sb, group);
1057 * This ensures that we don't reinit the buddy cache
1058 * page which map to the group from which we are already
1059 * allocating. If we are looking at the buddy cache we would
1060 * have taken a reference using ext4_mb_load_buddy and that
1061 * would have pinned buddy page to page cache.
1063 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1064 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1066 * somebody initialized the group
1067 * return without doing anything
1072 page = e4b.bd_bitmap_page;
1073 ret = ext4_mb_init_cache(page, NULL);
1076 if (!PageUptodate(page)) {
1080 mark_page_accessed(page);
1082 if (e4b.bd_buddy_page == NULL) {
1084 * If both the bitmap and buddy are in
1085 * the same page we don't need to force
1091 /* init buddy cache */
1092 page = e4b.bd_buddy_page;
1093 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1096 if (!PageUptodate(page)) {
1100 mark_page_accessed(page);
1102 ext4_mb_put_buddy_page_lock(&e4b);
1107 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1108 * block group lock of all groups for this page; do not hold the BG lock when
1109 * calling this routine!
1111 static noinline_for_stack int
1112 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1113 struct ext4_buddy *e4b)
1115 int blocks_per_page;
1121 struct ext4_group_info *grp;
1122 struct ext4_sb_info *sbi = EXT4_SB(sb);
1123 struct inode *inode = sbi->s_buddy_cache;
1125 mb_debug(1, "load group %u\n", group);
1127 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1128 grp = ext4_get_group_info(sb, group);
1130 e4b->bd_blkbits = sb->s_blocksize_bits;
1133 e4b->bd_group = group;
1134 e4b->bd_buddy_page = NULL;
1135 e4b->bd_bitmap_page = NULL;
1137 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1139 * we need full data about the group
1140 * to make a good selection
1142 ret = ext4_mb_init_group(sb, group);
1148 * the buddy cache inode stores the block bitmap
1149 * and buddy information in consecutive blocks.
1150 * So for each group we need two blocks.
1153 pnum = block / blocks_per_page;
1154 poff = block % blocks_per_page;
1156 /* we could use find_or_create_page(), but it locks page
1157 * what we'd like to avoid in fast path ... */
1158 page = find_get_page(inode->i_mapping, pnum);
1159 if (page == NULL || !PageUptodate(page)) {
1162 * drop the page reference and try
1163 * to get the page with lock. If we
1164 * are not uptodate that implies
1165 * somebody just created the page but
1166 * is yet to initialize the same. So
1167 * wait for it to initialize.
1169 page_cache_release(page);
1170 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1172 BUG_ON(page->mapping != inode->i_mapping);
1173 if (!PageUptodate(page)) {
1174 ret = ext4_mb_init_cache(page, NULL);
1179 mb_cmp_bitmaps(e4b, page_address(page) +
1180 (poff * sb->s_blocksize));
1185 if (page == NULL || !PageUptodate(page)) {
1189 e4b->bd_bitmap_page = page;
1190 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1191 mark_page_accessed(page);
1194 pnum = block / blocks_per_page;
1195 poff = block % blocks_per_page;
1197 page = find_get_page(inode->i_mapping, pnum);
1198 if (page == NULL || !PageUptodate(page)) {
1200 page_cache_release(page);
1201 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1203 BUG_ON(page->mapping != inode->i_mapping);
1204 if (!PageUptodate(page)) {
1205 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1214 if (page == NULL || !PageUptodate(page)) {
1218 e4b->bd_buddy_page = page;
1219 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1220 mark_page_accessed(page);
1222 BUG_ON(e4b->bd_bitmap_page == NULL);
1223 BUG_ON(e4b->bd_buddy_page == NULL);
1229 page_cache_release(page);
1230 if (e4b->bd_bitmap_page)
1231 page_cache_release(e4b->bd_bitmap_page);
1232 if (e4b->bd_buddy_page)
1233 page_cache_release(e4b->bd_buddy_page);
1234 e4b->bd_buddy = NULL;
1235 e4b->bd_bitmap = NULL;
1239 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1241 if (e4b->bd_bitmap_page)
1242 page_cache_release(e4b->bd_bitmap_page);
1243 if (e4b->bd_buddy_page)
1244 page_cache_release(e4b->bd_buddy_page);
1248 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1253 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1254 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1256 bb = EXT4_MB_BUDDY(e4b);
1257 while (order <= e4b->bd_blkbits + 1) {
1259 if (!mb_test_bit(block, bb)) {
1260 /* this block is part of buddy of order 'order' */
1263 bb += 1 << (e4b->bd_blkbits - order);
1269 static void mb_clear_bits(void *bm, int cur, int len)
1275 if ((cur & 31) == 0 && (len - cur) >= 32) {
1276 /* fast path: clear whole word at once */
1277 addr = bm + (cur >> 3);
1282 mb_clear_bit(cur, bm);
1287 void ext4_set_bits(void *bm, int cur, int len)
1293 if ((cur & 31) == 0 && (len - cur) >= 32) {
1294 /* fast path: set whole word at once */
1295 addr = bm + (cur >> 3);
1300 mb_set_bit(cur, bm);
1305 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1306 int first, int count)
1313 struct super_block *sb = e4b->bd_sb;
1315 BUG_ON(first + count > (sb->s_blocksize << 3));
1316 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1317 mb_check_buddy(e4b);
1318 mb_free_blocks_double(inode, e4b, first, count);
1320 e4b->bd_info->bb_free += count;
1321 if (first < e4b->bd_info->bb_first_free)
1322 e4b->bd_info->bb_first_free = first;
1324 /* let's maintain fragments counter */
1326 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1327 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1328 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1330 e4b->bd_info->bb_fragments--;
1331 else if (!block && !max)
1332 e4b->bd_info->bb_fragments++;
1334 /* let's maintain buddy itself */
1335 while (count-- > 0) {
1339 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1340 ext4_fsblk_t blocknr;
1342 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1343 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1344 ext4_grp_locked_error(sb, e4b->bd_group,
1345 inode ? inode->i_ino : 0,
1347 "freeing already freed block "
1350 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1351 e4b->bd_info->bb_counters[order]++;
1353 /* start of the buddy */
1354 buddy = mb_find_buddy(e4b, order, &max);
1358 if (mb_test_bit(block, buddy) ||
1359 mb_test_bit(block + 1, buddy))
1362 /* both the buddies are free, try to coalesce them */
1363 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1369 /* for special purposes, we don't set
1370 * free bits in bitmap */
1371 mb_set_bit(block, buddy);
1372 mb_set_bit(block + 1, buddy);
1374 e4b->bd_info->bb_counters[order]--;
1375 e4b->bd_info->bb_counters[order]--;
1379 e4b->bd_info->bb_counters[order]++;
1381 mb_clear_bit(block, buddy2);
1385 mb_set_largest_free_order(sb, e4b->bd_info);
1386 mb_check_buddy(e4b);
1389 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1390 int needed, struct ext4_free_extent *ex)
1397 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1400 buddy = mb_find_buddy(e4b, order, &max);
1401 BUG_ON(buddy == NULL);
1402 BUG_ON(block >= max);
1403 if (mb_test_bit(block, buddy)) {
1410 /* FIXME dorp order completely ? */
1411 if (likely(order == 0)) {
1412 /* find actual order */
1413 order = mb_find_order_for_block(e4b, block);
1414 block = block >> order;
1417 ex->fe_len = 1 << order;
1418 ex->fe_start = block << order;
1419 ex->fe_group = e4b->bd_group;
1421 /* calc difference from given start */
1422 next = next - ex->fe_start;
1424 ex->fe_start += next;
1426 while (needed > ex->fe_len &&
1427 (buddy = mb_find_buddy(e4b, order, &max))) {
1429 if (block + 1 >= max)
1432 next = (block + 1) * (1 << order);
1433 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1436 ord = mb_find_order_for_block(e4b, next);
1439 block = next >> order;
1440 ex->fe_len += 1 << order;
1443 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1447 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1453 int start = ex->fe_start;
1454 int len = ex->fe_len;
1459 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1460 BUG_ON(e4b->bd_group != ex->fe_group);
1461 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1462 mb_check_buddy(e4b);
1463 mb_mark_used_double(e4b, start, len);
1465 e4b->bd_info->bb_free -= len;
1466 if (e4b->bd_info->bb_first_free == start)
1467 e4b->bd_info->bb_first_free += len;
1469 /* let's maintain fragments counter */
1471 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1472 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1473 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1475 e4b->bd_info->bb_fragments++;
1476 else if (!mlen && !max)
1477 e4b->bd_info->bb_fragments--;
1479 /* let's maintain buddy itself */
1481 ord = mb_find_order_for_block(e4b, start);
1483 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1484 /* the whole chunk may be allocated at once! */
1486 buddy = mb_find_buddy(e4b, ord, &max);
1487 BUG_ON((start >> ord) >= max);
1488 mb_set_bit(start >> ord, buddy);
1489 e4b->bd_info->bb_counters[ord]--;
1496 /* store for history */
1498 ret = len | (ord << 16);
1500 /* we have to split large buddy */
1502 buddy = mb_find_buddy(e4b, ord, &max);
1503 mb_set_bit(start >> ord, buddy);
1504 e4b->bd_info->bb_counters[ord]--;
1507 cur = (start >> ord) & ~1U;
1508 buddy = mb_find_buddy(e4b, ord, &max);
1509 mb_clear_bit(cur, buddy);
1510 mb_clear_bit(cur + 1, buddy);
1511 e4b->bd_info->bb_counters[ord]++;
1512 e4b->bd_info->bb_counters[ord]++;
1514 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1516 ext4_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1517 mb_check_buddy(e4b);
1523 * Must be called under group lock!
1525 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1526 struct ext4_buddy *e4b)
1528 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1531 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1532 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1534 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1535 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1536 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1538 /* preallocation can change ac_b_ex, thus we store actually
1539 * allocated blocks for history */
1540 ac->ac_f_ex = ac->ac_b_ex;
1542 ac->ac_status = AC_STATUS_FOUND;
1543 ac->ac_tail = ret & 0xffff;
1544 ac->ac_buddy = ret >> 16;
1547 * take the page reference. We want the page to be pinned
1548 * so that we don't get a ext4_mb_init_cache_call for this
1549 * group until we update the bitmap. That would mean we
1550 * double allocate blocks. The reference is dropped
1551 * in ext4_mb_release_context
1553 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1554 get_page(ac->ac_bitmap_page);
1555 ac->ac_buddy_page = e4b->bd_buddy_page;
1556 get_page(ac->ac_buddy_page);
1557 /* store last allocated for subsequent stream allocation */
1558 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1559 spin_lock(&sbi->s_md_lock);
1560 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1561 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1562 spin_unlock(&sbi->s_md_lock);
1567 * regular allocator, for general purposes allocation
1570 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1571 struct ext4_buddy *e4b,
1574 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1575 struct ext4_free_extent *bex = &ac->ac_b_ex;
1576 struct ext4_free_extent *gex = &ac->ac_g_ex;
1577 struct ext4_free_extent ex;
1580 if (ac->ac_status == AC_STATUS_FOUND)
1583 * We don't want to scan for a whole year
1585 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1586 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1587 ac->ac_status = AC_STATUS_BREAK;
1592 * Haven't found good chunk so far, let's continue
1594 if (bex->fe_len < gex->fe_len)
1597 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1598 && bex->fe_group == e4b->bd_group) {
1599 /* recheck chunk's availability - we don't know
1600 * when it was found (within this lock-unlock
1602 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1603 if (max >= gex->fe_len) {
1604 ext4_mb_use_best_found(ac, e4b);
1611 * The routine checks whether found extent is good enough. If it is,
1612 * then the extent gets marked used and flag is set to the context
1613 * to stop scanning. Otherwise, the extent is compared with the
1614 * previous found extent and if new one is better, then it's stored
1615 * in the context. Later, the best found extent will be used, if
1616 * mballoc can't find good enough extent.
1618 * FIXME: real allocation policy is to be designed yet!
1620 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1621 struct ext4_free_extent *ex,
1622 struct ext4_buddy *e4b)
1624 struct ext4_free_extent *bex = &ac->ac_b_ex;
1625 struct ext4_free_extent *gex = &ac->ac_g_ex;
1627 BUG_ON(ex->fe_len <= 0);
1628 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1629 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1630 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1635 * The special case - take what you catch first
1637 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1639 ext4_mb_use_best_found(ac, e4b);
1644 * Let's check whether the chuck is good enough
1646 if (ex->fe_len == gex->fe_len) {
1648 ext4_mb_use_best_found(ac, e4b);
1653 * If this is first found extent, just store it in the context
1655 if (bex->fe_len == 0) {
1661 * If new found extent is better, store it in the context
1663 if (bex->fe_len < gex->fe_len) {
1664 /* if the request isn't satisfied, any found extent
1665 * larger than previous best one is better */
1666 if (ex->fe_len > bex->fe_len)
1668 } else if (ex->fe_len > gex->fe_len) {
1669 /* if the request is satisfied, then we try to find
1670 * an extent that still satisfy the request, but is
1671 * smaller than previous one */
1672 if (ex->fe_len < bex->fe_len)
1676 ext4_mb_check_limits(ac, e4b, 0);
1679 static noinline_for_stack
1680 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1681 struct ext4_buddy *e4b)
1683 struct ext4_free_extent ex = ac->ac_b_ex;
1684 ext4_group_t group = ex.fe_group;
1688 BUG_ON(ex.fe_len <= 0);
1689 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1693 ext4_lock_group(ac->ac_sb, group);
1694 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1698 ext4_mb_use_best_found(ac, e4b);
1701 ext4_unlock_group(ac->ac_sb, group);
1702 ext4_mb_unload_buddy(e4b);
1707 static noinline_for_stack
1708 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1709 struct ext4_buddy *e4b)
1711 ext4_group_t group = ac->ac_g_ex.fe_group;
1714 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1715 struct ext4_free_extent ex;
1717 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1720 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1724 ext4_lock_group(ac->ac_sb, group);
1725 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1726 ac->ac_g_ex.fe_len, &ex);
1728 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1731 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1733 /* use do_div to get remainder (would be 64-bit modulo) */
1734 if (do_div(start, sbi->s_stripe) == 0) {
1737 ext4_mb_use_best_found(ac, e4b);
1739 } else if (max >= ac->ac_g_ex.fe_len) {
1740 BUG_ON(ex.fe_len <= 0);
1741 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1742 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1745 ext4_mb_use_best_found(ac, e4b);
1746 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1747 /* Sometimes, caller may want to merge even small
1748 * number of blocks to an existing extent */
1749 BUG_ON(ex.fe_len <= 0);
1750 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1751 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1754 ext4_mb_use_best_found(ac, e4b);
1756 ext4_unlock_group(ac->ac_sb, group);
1757 ext4_mb_unload_buddy(e4b);
1763 * The routine scans buddy structures (not bitmap!) from given order
1764 * to max order and tries to find big enough chunk to satisfy the req
1766 static noinline_for_stack
1767 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1768 struct ext4_buddy *e4b)
1770 struct super_block *sb = ac->ac_sb;
1771 struct ext4_group_info *grp = e4b->bd_info;
1777 BUG_ON(ac->ac_2order <= 0);
1778 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1779 if (grp->bb_counters[i] == 0)
1782 buddy = mb_find_buddy(e4b, i, &max);
1783 BUG_ON(buddy == NULL);
1785 k = mb_find_next_zero_bit(buddy, max, 0);
1790 ac->ac_b_ex.fe_len = 1 << i;
1791 ac->ac_b_ex.fe_start = k << i;
1792 ac->ac_b_ex.fe_group = e4b->bd_group;
1794 ext4_mb_use_best_found(ac, e4b);
1796 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1798 if (EXT4_SB(sb)->s_mb_stats)
1799 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1806 * The routine scans the group and measures all found extents.
1807 * In order to optimize scanning, caller must pass number of
1808 * free blocks in the group, so the routine can know upper limit.
1810 static noinline_for_stack
1811 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1812 struct ext4_buddy *e4b)
1814 struct super_block *sb = ac->ac_sb;
1815 void *bitmap = EXT4_MB_BITMAP(e4b);
1816 struct ext4_free_extent ex;
1820 free = e4b->bd_info->bb_free;
1823 i = e4b->bd_info->bb_first_free;
1825 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1826 i = mb_find_next_zero_bit(bitmap,
1827 EXT4_CLUSTERS_PER_GROUP(sb), i);
1828 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1830 * IF we have corrupt bitmap, we won't find any
1831 * free blocks even though group info says we
1832 * we have free blocks
1834 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1835 "%d free clusters as per "
1836 "group info. But bitmap says 0",
1841 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1842 BUG_ON(ex.fe_len <= 0);
1843 if (free < ex.fe_len) {
1844 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1845 "%d free clusters as per "
1846 "group info. But got %d blocks",
1849 * The number of free blocks differs. This mostly
1850 * indicate that the bitmap is corrupt. So exit
1851 * without claiming the space.
1856 ext4_mb_measure_extent(ac, &ex, e4b);
1862 ext4_mb_check_limits(ac, e4b, 1);
1866 * This is a special case for storages like raid5
1867 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1869 static noinline_for_stack
1870 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1871 struct ext4_buddy *e4b)
1873 struct super_block *sb = ac->ac_sb;
1874 struct ext4_sb_info *sbi = EXT4_SB(sb);
1875 void *bitmap = EXT4_MB_BITMAP(e4b);
1876 struct ext4_free_extent ex;
1877 ext4_fsblk_t first_group_block;
1882 BUG_ON(sbi->s_stripe == 0);
1884 /* find first stripe-aligned block in group */
1885 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1887 a = first_group_block + sbi->s_stripe - 1;
1888 do_div(a, sbi->s_stripe);
1889 i = (a * sbi->s_stripe) - first_group_block;
1891 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1892 if (!mb_test_bit(i, bitmap)) {
1893 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1894 if (max >= sbi->s_stripe) {
1897 ext4_mb_use_best_found(ac, e4b);
1905 /* This is now called BEFORE we load the buddy bitmap. */
1906 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1907 ext4_group_t group, int cr)
1909 unsigned free, fragments;
1910 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1911 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1913 BUG_ON(cr < 0 || cr >= 4);
1915 /* We only do this if the grp has never been initialized */
1916 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1917 int ret = ext4_mb_init_group(ac->ac_sb, group);
1922 free = grp->bb_free;
1923 fragments = grp->bb_fragments;
1931 BUG_ON(ac->ac_2order == 0);
1933 if (grp->bb_largest_free_order < ac->ac_2order)
1936 /* Avoid using the first bg of a flexgroup for data files */
1937 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1938 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1939 ((group % flex_size) == 0))
1944 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1948 if (free >= ac->ac_g_ex.fe_len)
1960 static noinline_for_stack int
1961 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1963 ext4_group_t ngroups, group, i;
1966 struct ext4_sb_info *sbi;
1967 struct super_block *sb;
1968 struct ext4_buddy e4b;
1972 ngroups = ext4_get_groups_count(sb);
1973 /* non-extent files are limited to low blocks/groups */
1974 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
1975 ngroups = sbi->s_blockfile_groups;
1977 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1979 /* first, try the goal */
1980 err = ext4_mb_find_by_goal(ac, &e4b);
1981 if (err || ac->ac_status == AC_STATUS_FOUND)
1984 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1988 * ac->ac2_order is set only if the fe_len is a power of 2
1989 * if ac2_order is set we also set criteria to 0 so that we
1990 * try exact allocation using buddy.
1992 i = fls(ac->ac_g_ex.fe_len);
1995 * We search using buddy data only if the order of the request
1996 * is greater than equal to the sbi_s_mb_order2_reqs
1997 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1999 if (i >= sbi->s_mb_order2_reqs) {
2001 * This should tell if fe_len is exactly power of 2
2003 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2004 ac->ac_2order = i - 1;
2007 /* if stream allocation is enabled, use global goal */
2008 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2009 /* TBD: may be hot point */
2010 spin_lock(&sbi->s_md_lock);
2011 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2012 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2013 spin_unlock(&sbi->s_md_lock);
2016 /* Let's just scan groups to find more-less suitable blocks */
2017 cr = ac->ac_2order ? 0 : 1;
2019 * cr == 0 try to get exact allocation,
2020 * cr == 3 try to get anything
2023 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2024 ac->ac_criteria = cr;
2026 * searching for the right group start
2027 * from the goal value specified
2029 group = ac->ac_g_ex.fe_group;
2031 for (i = 0; i < ngroups; group++, i++) {
2032 if (group == ngroups)
2035 /* This now checks without needing the buddy page */
2036 if (!ext4_mb_good_group(ac, group, cr))
2039 err = ext4_mb_load_buddy(sb, group, &e4b);
2043 ext4_lock_group(sb, group);
2046 * We need to check again after locking the
2049 if (!ext4_mb_good_group(ac, group, cr)) {
2050 ext4_unlock_group(sb, group);
2051 ext4_mb_unload_buddy(&e4b);
2055 ac->ac_groups_scanned++;
2057 ext4_mb_simple_scan_group(ac, &e4b);
2058 else if (cr == 1 && sbi->s_stripe &&
2059 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2060 ext4_mb_scan_aligned(ac, &e4b);
2062 ext4_mb_complex_scan_group(ac, &e4b);
2064 ext4_unlock_group(sb, group);
2065 ext4_mb_unload_buddy(&e4b);
2067 if (ac->ac_status != AC_STATUS_CONTINUE)
2072 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2073 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2075 * We've been searching too long. Let's try to allocate
2076 * the best chunk we've found so far
2079 ext4_mb_try_best_found(ac, &e4b);
2080 if (ac->ac_status != AC_STATUS_FOUND) {
2082 * Someone more lucky has already allocated it.
2083 * The only thing we can do is just take first
2085 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2087 ac->ac_b_ex.fe_group = 0;
2088 ac->ac_b_ex.fe_start = 0;
2089 ac->ac_b_ex.fe_len = 0;
2090 ac->ac_status = AC_STATUS_CONTINUE;
2091 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2093 atomic_inc(&sbi->s_mb_lost_chunks);
2101 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2103 struct super_block *sb = seq->private;
2106 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2109 return (void *) ((unsigned long) group);
2112 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2114 struct super_block *sb = seq->private;
2118 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2121 return (void *) ((unsigned long) group);
2124 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2126 struct super_block *sb = seq->private;
2127 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2130 struct ext4_buddy e4b;
2132 struct ext4_group_info info;
2133 ext4_grpblk_t counters[16];
2138 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2139 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2140 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2141 "group", "free", "frags", "first",
2142 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2143 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2145 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2146 sizeof(struct ext4_group_info);
2147 err = ext4_mb_load_buddy(sb, group, &e4b);
2149 seq_printf(seq, "#%-5u: I/O error\n", group);
2152 ext4_lock_group(sb, group);
2153 memcpy(&sg, ext4_get_group_info(sb, group), i);
2154 ext4_unlock_group(sb, group);
2155 ext4_mb_unload_buddy(&e4b);
2157 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2158 sg.info.bb_fragments, sg.info.bb_first_free);
2159 for (i = 0; i <= 13; i++)
2160 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2161 sg.info.bb_counters[i] : 0);
2162 seq_printf(seq, " ]\n");
2167 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2171 static const struct seq_operations ext4_mb_seq_groups_ops = {
2172 .start = ext4_mb_seq_groups_start,
2173 .next = ext4_mb_seq_groups_next,
2174 .stop = ext4_mb_seq_groups_stop,
2175 .show = ext4_mb_seq_groups_show,
2178 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2180 struct super_block *sb = PDE(inode)->data;
2183 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2185 struct seq_file *m = file->private_data;
2192 static const struct file_operations ext4_mb_seq_groups_fops = {
2193 .owner = THIS_MODULE,
2194 .open = ext4_mb_seq_groups_open,
2196 .llseek = seq_lseek,
2197 .release = seq_release,
2200 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2202 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2203 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2209 /* Create and initialize ext4_group_info data for the given group. */
2210 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2211 struct ext4_group_desc *desc)
2215 struct ext4_sb_info *sbi = EXT4_SB(sb);
2216 struct ext4_group_info **meta_group_info;
2217 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2220 * First check if this group is the first of a reserved block.
2221 * If it's true, we have to allocate a new table of pointers
2222 * to ext4_group_info structures
2224 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2225 metalen = sizeof(*meta_group_info) <<
2226 EXT4_DESC_PER_BLOCK_BITS(sb);
2227 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2228 if (meta_group_info == NULL) {
2229 ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate mem "
2230 "for a buddy group");
2231 goto exit_meta_group_info;
2233 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2238 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2239 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2241 meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL);
2242 if (meta_group_info[i] == NULL) {
2243 ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate buddy mem");
2244 goto exit_group_info;
2246 memset(meta_group_info[i], 0, kmem_cache_size(cachep));
2247 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2248 &(meta_group_info[i]->bb_state));
2251 * initialize bb_free to be able to skip
2252 * empty groups without initialization
2254 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2255 meta_group_info[i]->bb_free =
2256 ext4_free_clusters_after_init(sb, group, desc);
2258 meta_group_info[i]->bb_free =
2259 ext4_free_group_clusters(sb, desc);
2262 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2263 init_rwsem(&meta_group_info[i]->alloc_sem);
2264 meta_group_info[i]->bb_free_root = RB_ROOT;
2265 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2269 struct buffer_head *bh;
2270 meta_group_info[i]->bb_bitmap =
2271 kmalloc(sb->s_blocksize, GFP_KERNEL);
2272 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2273 bh = ext4_read_block_bitmap(sb, group);
2275 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2284 /* If a meta_group_info table has been allocated, release it now */
2285 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2286 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2287 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2289 exit_meta_group_info:
2291 } /* ext4_mb_add_groupinfo */
2293 static int ext4_mb_init_backend(struct super_block *sb)
2295 ext4_group_t ngroups = ext4_get_groups_count(sb);
2297 struct ext4_sb_info *sbi = EXT4_SB(sb);
2298 struct ext4_super_block *es = sbi->s_es;
2299 int num_meta_group_infos;
2300 int num_meta_group_infos_max;
2302 struct ext4_group_desc *desc;
2303 struct kmem_cache *cachep;
2305 /* This is the number of blocks used by GDT */
2306 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2307 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2310 * This is the total number of blocks used by GDT including
2311 * the number of reserved blocks for GDT.
2312 * The s_group_info array is allocated with this value
2313 * to allow a clean online resize without a complex
2314 * manipulation of pointer.
2315 * The drawback is the unused memory when no resize
2316 * occurs but it's very low in terms of pages
2317 * (see comments below)
2318 * Need to handle this properly when META_BG resizing is allowed
2320 num_meta_group_infos_max = num_meta_group_infos +
2321 le16_to_cpu(es->s_reserved_gdt_blocks);
2324 * array_size is the size of s_group_info array. We round it
2325 * to the next power of two because this approximation is done
2326 * internally by kmalloc so we can have some more memory
2327 * for free here (e.g. may be used for META_BG resize).
2330 while (array_size < sizeof(*sbi->s_group_info) *
2331 num_meta_group_infos_max)
2332 array_size = array_size << 1;
2333 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2334 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2335 * So a two level scheme suffices for now. */
2336 sbi->s_group_info = ext4_kvzalloc(array_size, GFP_KERNEL);
2337 if (sbi->s_group_info == NULL) {
2338 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2341 sbi->s_buddy_cache = new_inode(sb);
2342 if (sbi->s_buddy_cache == NULL) {
2343 ext4_msg(sb, KERN_ERR, "can't get new inode");
2346 /* To avoid potentially colliding with an valid on-disk inode number,
2347 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2348 * not in the inode hash, so it should never be found by iget(), but
2349 * this will avoid confusion if it ever shows up during debugging. */
2350 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2351 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2352 for (i = 0; i < ngroups; i++) {
2353 desc = ext4_get_group_desc(sb, i, NULL);
2355 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2358 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2365 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2367 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2368 i = num_meta_group_infos;
2370 kfree(sbi->s_group_info[i]);
2371 iput(sbi->s_buddy_cache);
2373 ext4_kvfree(sbi->s_group_info);
2377 static void ext4_groupinfo_destroy_slabs(void)
2381 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2382 if (ext4_groupinfo_caches[i])
2383 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2384 ext4_groupinfo_caches[i] = NULL;
2388 static int ext4_groupinfo_create_slab(size_t size)
2390 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2392 int blocksize_bits = order_base_2(size);
2393 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2394 struct kmem_cache *cachep;
2396 if (cache_index >= NR_GRPINFO_CACHES)
2399 if (unlikely(cache_index < 0))
2402 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2403 if (ext4_groupinfo_caches[cache_index]) {
2404 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2405 return 0; /* Already created */
2408 slab_size = offsetof(struct ext4_group_info,
2409 bb_counters[blocksize_bits + 2]);
2411 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2412 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2415 ext4_groupinfo_caches[cache_index] = cachep;
2417 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2420 "EXT4-fs: no memory for groupinfo slab cache\n");
2427 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2429 struct ext4_sb_info *sbi = EXT4_SB(sb);
2435 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2437 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2438 if (sbi->s_mb_offsets == NULL) {
2443 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2444 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2445 if (sbi->s_mb_maxs == NULL) {
2450 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2454 /* order 0 is regular bitmap */
2455 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2456 sbi->s_mb_offsets[0] = 0;
2460 max = sb->s_blocksize << 2;
2462 sbi->s_mb_offsets[i] = offset;
2463 sbi->s_mb_maxs[i] = max;
2464 offset += 1 << (sb->s_blocksize_bits - i);
2467 } while (i <= sb->s_blocksize_bits + 1);
2469 spin_lock_init(&sbi->s_md_lock);
2470 spin_lock_init(&sbi->s_bal_lock);
2472 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2473 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2474 sbi->s_mb_stats = MB_DEFAULT_STATS;
2475 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2476 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2478 * The default group preallocation is 512, which for 4k block
2479 * sizes translates to 2 megabytes. However for bigalloc file
2480 * systems, this is probably too big (i.e, if the cluster size
2481 * is 1 megabyte, then group preallocation size becomes half a
2482 * gigabyte!). As a default, we will keep a two megabyte
2483 * group pralloc size for cluster sizes up to 64k, and after
2484 * that, we will force a minimum group preallocation size of
2485 * 32 clusters. This translates to 8 megs when the cluster
2486 * size is 256k, and 32 megs when the cluster size is 1 meg,
2487 * which seems reasonable as a default.
2489 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2490 sbi->s_cluster_bits, 32);
2492 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2493 * to the lowest multiple of s_stripe which is bigger than
2494 * the s_mb_group_prealloc as determined above. We want
2495 * the preallocation size to be an exact multiple of the
2496 * RAID stripe size so that preallocations don't fragment
2499 if (sbi->s_stripe > 1) {
2500 sbi->s_mb_group_prealloc = roundup(
2501 sbi->s_mb_group_prealloc, sbi->s_stripe);
2504 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2505 if (sbi->s_locality_groups == NULL) {
2507 goto out_free_groupinfo_slab;
2509 for_each_possible_cpu(i) {
2510 struct ext4_locality_group *lg;
2511 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2512 mutex_init(&lg->lg_mutex);
2513 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2514 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2515 spin_lock_init(&lg->lg_prealloc_lock);
2518 /* init file for buddy data */
2519 ret = ext4_mb_init_backend(sb);
2521 goto out_free_locality_groups;
2524 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2525 &ext4_mb_seq_groups_fops, sb);
2528 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2532 out_free_locality_groups:
2533 free_percpu(sbi->s_locality_groups);
2534 sbi->s_locality_groups = NULL;
2535 out_free_groupinfo_slab:
2536 ext4_groupinfo_destroy_slabs();
2538 kfree(sbi->s_mb_offsets);
2539 sbi->s_mb_offsets = NULL;
2540 kfree(sbi->s_mb_maxs);
2541 sbi->s_mb_maxs = NULL;
2545 /* need to called with the ext4 group lock held */
2546 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2548 struct ext4_prealloc_space *pa;
2549 struct list_head *cur, *tmp;
2552 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2553 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2554 list_del(&pa->pa_group_list);
2556 kmem_cache_free(ext4_pspace_cachep, pa);
2559 mb_debug(1, "mballoc: %u PAs left\n", count);
2563 int ext4_mb_release(struct super_block *sb)
2565 ext4_group_t ngroups = ext4_get_groups_count(sb);
2567 int num_meta_group_infos;
2568 struct ext4_group_info *grinfo;
2569 struct ext4_sb_info *sbi = EXT4_SB(sb);
2570 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2572 if (sbi->s_group_info) {
2573 for (i = 0; i < ngroups; i++) {
2574 grinfo = ext4_get_group_info(sb, i);
2576 kfree(grinfo->bb_bitmap);
2578 ext4_lock_group(sb, i);
2579 ext4_mb_cleanup_pa(grinfo);
2580 ext4_unlock_group(sb, i);
2581 kmem_cache_free(cachep, grinfo);
2583 num_meta_group_infos = (ngroups +
2584 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2585 EXT4_DESC_PER_BLOCK_BITS(sb);
2586 for (i = 0; i < num_meta_group_infos; i++)
2587 kfree(sbi->s_group_info[i]);
2588 ext4_kvfree(sbi->s_group_info);
2590 kfree(sbi->s_mb_offsets);
2591 kfree(sbi->s_mb_maxs);
2592 if (sbi->s_buddy_cache)
2593 iput(sbi->s_buddy_cache);
2594 if (sbi->s_mb_stats) {
2595 ext4_msg(sb, KERN_INFO,
2596 "mballoc: %u blocks %u reqs (%u success)",
2597 atomic_read(&sbi->s_bal_allocated),
2598 atomic_read(&sbi->s_bal_reqs),
2599 atomic_read(&sbi->s_bal_success));
2600 ext4_msg(sb, KERN_INFO,
2601 "mballoc: %u extents scanned, %u goal hits, "
2602 "%u 2^N hits, %u breaks, %u lost",
2603 atomic_read(&sbi->s_bal_ex_scanned),
2604 atomic_read(&sbi->s_bal_goals),
2605 atomic_read(&sbi->s_bal_2orders),
2606 atomic_read(&sbi->s_bal_breaks),
2607 atomic_read(&sbi->s_mb_lost_chunks));
2608 ext4_msg(sb, KERN_INFO,
2609 "mballoc: %lu generated and it took %Lu",
2610 sbi->s_mb_buddies_generated,
2611 sbi->s_mb_generation_time);
2612 ext4_msg(sb, KERN_INFO,
2613 "mballoc: %u preallocated, %u discarded",
2614 atomic_read(&sbi->s_mb_preallocated),
2615 atomic_read(&sbi->s_mb_discarded));
2618 free_percpu(sbi->s_locality_groups);
2620 remove_proc_entry("mb_groups", sbi->s_proc);
2625 static inline int ext4_issue_discard(struct super_block *sb,
2626 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2628 ext4_fsblk_t discard_block;
2630 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2631 ext4_group_first_block_no(sb, block_group));
2632 count = EXT4_C2B(EXT4_SB(sb), count);
2633 trace_ext4_discard_blocks(sb,
2634 (unsigned long long) discard_block, count);
2635 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2639 * This function is called by the jbd2 layer once the commit has finished,
2640 * so we know we can free the blocks that were released with that commit.
2642 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2644 struct super_block *sb = journal->j_private;
2645 struct ext4_buddy e4b;
2646 struct ext4_group_info *db;
2647 int err, count = 0, count2 = 0;
2648 struct ext4_free_data *entry;
2649 struct list_head *l, *ltmp;
2651 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2652 entry = list_entry(l, struct ext4_free_data, list);
2654 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2655 entry->count, entry->group, entry);
2657 if (test_opt(sb, DISCARD))
2658 ext4_issue_discard(sb, entry->group,
2659 entry->start_cluster, entry->count);
2661 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2662 /* we expect to find existing buddy because it's pinned */
2666 /* there are blocks to put in buddy to make them really free */
2667 count += entry->count;
2669 ext4_lock_group(sb, entry->group);
2670 /* Take it out of per group rb tree */
2671 rb_erase(&entry->node, &(db->bb_free_root));
2672 mb_free_blocks(NULL, &e4b, entry->start_cluster, entry->count);
2675 * Clear the trimmed flag for the group so that the next
2676 * ext4_trim_fs can trim it.
2677 * If the volume is mounted with -o discard, online discard
2678 * is supported and the free blocks will be trimmed online.
2680 if (!test_opt(sb, DISCARD))
2681 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2683 if (!db->bb_free_root.rb_node) {
2684 /* No more items in the per group rb tree
2685 * balance refcounts from ext4_mb_free_metadata()
2687 page_cache_release(e4b.bd_buddy_page);
2688 page_cache_release(e4b.bd_bitmap_page);
2690 ext4_unlock_group(sb, entry->group);
2691 kmem_cache_free(ext4_free_ext_cachep, entry);
2692 ext4_mb_unload_buddy(&e4b);
2695 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2698 #ifdef CONFIG_EXT4_DEBUG
2699 u8 mb_enable_debug __read_mostly;
2701 static struct dentry *debugfs_dir;
2702 static struct dentry *debugfs_debug;
2704 static void __init ext4_create_debugfs_entry(void)
2706 debugfs_dir = debugfs_create_dir("ext4", NULL);
2708 debugfs_debug = debugfs_create_u8("mballoc-debug",
2714 static void ext4_remove_debugfs_entry(void)
2716 debugfs_remove(debugfs_debug);
2717 debugfs_remove(debugfs_dir);
2722 static void __init ext4_create_debugfs_entry(void)
2726 static void ext4_remove_debugfs_entry(void)
2732 int __init ext4_init_mballoc(void)
2734 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2735 SLAB_RECLAIM_ACCOUNT);
2736 if (ext4_pspace_cachep == NULL)
2739 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2740 SLAB_RECLAIM_ACCOUNT);
2741 if (ext4_ac_cachep == NULL) {
2742 kmem_cache_destroy(ext4_pspace_cachep);
2746 ext4_free_ext_cachep = KMEM_CACHE(ext4_free_data,
2747 SLAB_RECLAIM_ACCOUNT);
2748 if (ext4_free_ext_cachep == NULL) {
2749 kmem_cache_destroy(ext4_pspace_cachep);
2750 kmem_cache_destroy(ext4_ac_cachep);
2753 ext4_create_debugfs_entry();
2757 void ext4_exit_mballoc(void)
2760 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2761 * before destroying the slab cache.
2764 kmem_cache_destroy(ext4_pspace_cachep);
2765 kmem_cache_destroy(ext4_ac_cachep);
2766 kmem_cache_destroy(ext4_free_ext_cachep);
2767 ext4_groupinfo_destroy_slabs();
2768 ext4_remove_debugfs_entry();
2773 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2774 * Returns 0 if success or error code
2776 static noinline_for_stack int
2777 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2778 handle_t *handle, unsigned int reserv_clstrs)
2780 struct buffer_head *bitmap_bh = NULL;
2781 struct ext4_group_desc *gdp;
2782 struct buffer_head *gdp_bh;
2783 struct ext4_sb_info *sbi;
2784 struct super_block *sb;
2788 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2789 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2795 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2799 err = ext4_journal_get_write_access(handle, bitmap_bh);
2804 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2808 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2809 ext4_free_group_clusters(sb, gdp));
2811 err = ext4_journal_get_write_access(handle, gdp_bh);
2815 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2817 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2818 if (!ext4_data_block_valid(sbi, block, len)) {
2819 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2820 "fs metadata\n", block, block+len);
2821 /* File system mounted not to panic on error
2822 * Fix the bitmap and repeat the block allocation
2823 * We leak some of the blocks here.
2825 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2826 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2827 ac->ac_b_ex.fe_len);
2828 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2829 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2835 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2836 #ifdef AGGRESSIVE_CHECK
2839 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2840 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2841 bitmap_bh->b_data));
2845 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2846 ac->ac_b_ex.fe_len);
2847 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2848 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2849 ext4_free_group_clusters_set(sb, gdp,
2850 ext4_free_clusters_after_init(sb,
2851 ac->ac_b_ex.fe_group, gdp));
2853 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2854 ext4_free_group_clusters_set(sb, gdp, len);
2855 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2857 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2858 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2860 * Now reduce the dirty block count also. Should not go negative
2862 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2863 /* release all the reserved blocks if non delalloc */
2864 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2867 if (sbi->s_log_groups_per_flex) {
2868 ext4_group_t flex_group = ext4_flex_group(sbi,
2869 ac->ac_b_ex.fe_group);
2870 atomic_sub(ac->ac_b_ex.fe_len,
2871 &sbi->s_flex_groups[flex_group].free_clusters);
2874 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2877 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2880 ext4_mark_super_dirty(sb);
2886 * here we normalize request for locality group
2887 * Group request are normalized to s_mb_group_prealloc, which goes to
2888 * s_strip if we set the same via mount option.
2889 * s_mb_group_prealloc can be configured via
2890 * /sys/fs/ext4/<partition>/mb_group_prealloc
2892 * XXX: should we try to preallocate more than the group has now?
2894 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2896 struct super_block *sb = ac->ac_sb;
2897 struct ext4_locality_group *lg = ac->ac_lg;
2900 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2901 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2902 current->pid, ac->ac_g_ex.fe_len);
2906 * Normalization means making request better in terms of
2907 * size and alignment
2909 static noinline_for_stack void
2910 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2911 struct ext4_allocation_request *ar)
2913 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2916 loff_t size, orig_size, start_off;
2918 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2919 struct ext4_prealloc_space *pa;
2921 /* do normalize only data requests, metadata requests
2922 do not need preallocation */
2923 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2926 /* sometime caller may want exact blocks */
2927 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2930 /* caller may indicate that preallocation isn't
2931 * required (it's a tail, for example) */
2932 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2935 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2936 ext4_mb_normalize_group_request(ac);
2940 bsbits = ac->ac_sb->s_blocksize_bits;
2942 /* first, let's learn actual file size
2943 * given current request is allocated */
2944 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
2945 size = size << bsbits;
2946 if (size < i_size_read(ac->ac_inode))
2947 size = i_size_read(ac->ac_inode);
2950 /* max size of free chunks */
2953 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2954 (req <= (size) || max <= (chunk_size))
2956 /* first, try to predict filesize */
2957 /* XXX: should this table be tunable? */
2959 if (size <= 16 * 1024) {
2961 } else if (size <= 32 * 1024) {
2963 } else if (size <= 64 * 1024) {
2965 } else if (size <= 128 * 1024) {
2967 } else if (size <= 256 * 1024) {
2969 } else if (size <= 512 * 1024) {
2971 } else if (size <= 1024 * 1024) {
2973 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2974 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2975 (21 - bsbits)) << 21;
2976 size = 2 * 1024 * 1024;
2977 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2978 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2979 (22 - bsbits)) << 22;
2980 size = 4 * 1024 * 1024;
2981 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2982 (8<<20)>>bsbits, max, 8 * 1024)) {
2983 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2984 (23 - bsbits)) << 23;
2985 size = 8 * 1024 * 1024;
2987 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2988 size = ac->ac_o_ex.fe_len << bsbits;
2990 size = size >> bsbits;
2991 start = start_off >> bsbits;
2993 /* don't cover already allocated blocks in selected range */
2994 if (ar->pleft && start <= ar->lleft) {
2995 size -= ar->lleft + 1 - start;
2996 start = ar->lleft + 1;
2998 if (ar->pright && start + size - 1 >= ar->lright)
2999 size -= start + size - ar->lright;
3003 /* check we don't cross already preallocated blocks */
3005 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3010 spin_lock(&pa->pa_lock);
3011 if (pa->pa_deleted) {
3012 spin_unlock(&pa->pa_lock);
3016 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3019 /* PA must not overlap original request */
3020 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3021 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3023 /* skip PAs this normalized request doesn't overlap with */
3024 if (pa->pa_lstart >= end || pa_end <= start) {
3025 spin_unlock(&pa->pa_lock);
3028 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3030 /* adjust start or end to be adjacent to this pa */
3031 if (pa_end <= ac->ac_o_ex.fe_logical) {
3032 BUG_ON(pa_end < start);
3034 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3035 BUG_ON(pa->pa_lstart > end);
3036 end = pa->pa_lstart;
3038 spin_unlock(&pa->pa_lock);
3043 /* XXX: extra loop to check we really don't overlap preallocations */
3045 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3048 spin_lock(&pa->pa_lock);
3049 if (pa->pa_deleted == 0) {
3050 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3052 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3054 spin_unlock(&pa->pa_lock);
3058 if (start + size <= ac->ac_o_ex.fe_logical &&
3059 start > ac->ac_o_ex.fe_logical) {
3060 ext4_msg(ac->ac_sb, KERN_ERR,
3061 "start %lu, size %lu, fe_logical %lu",
3062 (unsigned long) start, (unsigned long) size,
3063 (unsigned long) ac->ac_o_ex.fe_logical);
3065 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3066 start > ac->ac_o_ex.fe_logical);
3067 BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
3069 /* now prepare goal request */
3071 /* XXX: is it better to align blocks WRT to logical
3072 * placement or satisfy big request as is */
3073 ac->ac_g_ex.fe_logical = start;
3074 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3076 /* define goal start in order to merge */
3077 if (ar->pright && (ar->lright == (start + size))) {
3078 /* merge to the right */
3079 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3080 &ac->ac_f_ex.fe_group,
3081 &ac->ac_f_ex.fe_start);
3082 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3084 if (ar->pleft && (ar->lleft + 1 == start)) {
3085 /* merge to the left */
3086 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3087 &ac->ac_f_ex.fe_group,
3088 &ac->ac_f_ex.fe_start);
3089 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3092 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3093 (unsigned) orig_size, (unsigned) start);
3096 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3098 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3100 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3101 atomic_inc(&sbi->s_bal_reqs);
3102 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3103 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3104 atomic_inc(&sbi->s_bal_success);
3105 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3106 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3107 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3108 atomic_inc(&sbi->s_bal_goals);
3109 if (ac->ac_found > sbi->s_mb_max_to_scan)
3110 atomic_inc(&sbi->s_bal_breaks);
3113 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3114 trace_ext4_mballoc_alloc(ac);
3116 trace_ext4_mballoc_prealloc(ac);
3120 * Called on failure; free up any blocks from the inode PA for this
3121 * context. We don't need this for MB_GROUP_PA because we only change
3122 * pa_free in ext4_mb_release_context(), but on failure, we've already
3123 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3125 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3127 struct ext4_prealloc_space *pa = ac->ac_pa;
3130 if (pa && pa->pa_type == MB_INODE_PA) {
3131 len = ac->ac_b_ex.fe_len;
3138 * use blocks preallocated to inode
3140 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3141 struct ext4_prealloc_space *pa)
3143 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3148 /* found preallocated blocks, use them */
3149 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3150 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3151 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3152 len = EXT4_NUM_B2C(sbi, end - start);
3153 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3154 &ac->ac_b_ex.fe_start);
3155 ac->ac_b_ex.fe_len = len;
3156 ac->ac_status = AC_STATUS_FOUND;
3159 BUG_ON(start < pa->pa_pstart);
3160 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3161 BUG_ON(pa->pa_free < len);
3164 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3168 * use blocks preallocated to locality group
3170 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3171 struct ext4_prealloc_space *pa)
3173 unsigned int len = ac->ac_o_ex.fe_len;
3175 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3176 &ac->ac_b_ex.fe_group,
3177 &ac->ac_b_ex.fe_start);
3178 ac->ac_b_ex.fe_len = len;
3179 ac->ac_status = AC_STATUS_FOUND;
3182 /* we don't correct pa_pstart or pa_plen here to avoid
3183 * possible race when the group is being loaded concurrently
3184 * instead we correct pa later, after blocks are marked
3185 * in on-disk bitmap -- see ext4_mb_release_context()
3186 * Other CPUs are prevented from allocating from this pa by lg_mutex
3188 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3192 * Return the prealloc space that have minimal distance
3193 * from the goal block. @cpa is the prealloc
3194 * space that is having currently known minimal distance
3195 * from the goal block.
3197 static struct ext4_prealloc_space *
3198 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3199 struct ext4_prealloc_space *pa,
3200 struct ext4_prealloc_space *cpa)
3202 ext4_fsblk_t cur_distance, new_distance;
3205 atomic_inc(&pa->pa_count);
3208 cur_distance = abs(goal_block - cpa->pa_pstart);
3209 new_distance = abs(goal_block - pa->pa_pstart);
3211 if (cur_distance <= new_distance)
3214 /* drop the previous reference */
3215 atomic_dec(&cpa->pa_count);
3216 atomic_inc(&pa->pa_count);
3221 * search goal blocks in preallocated space
3223 static noinline_for_stack int
3224 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3226 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3228 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3229 struct ext4_locality_group *lg;
3230 struct ext4_prealloc_space *pa, *cpa = NULL;
3231 ext4_fsblk_t goal_block;
3233 /* only data can be preallocated */
3234 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3237 /* first, try per-file preallocation */
3239 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3241 /* all fields in this condition don't change,
3242 * so we can skip locking for them */
3243 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3244 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3245 EXT4_C2B(sbi, pa->pa_len)))
3248 /* non-extent files can't have physical blocks past 2^32 */
3249 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3250 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3251 EXT4_MAX_BLOCK_FILE_PHYS))
3254 /* found preallocated blocks, use them */
3255 spin_lock(&pa->pa_lock);
3256 if (pa->pa_deleted == 0 && pa->pa_free) {
3257 atomic_inc(&pa->pa_count);
3258 ext4_mb_use_inode_pa(ac, pa);
3259 spin_unlock(&pa->pa_lock);
3260 ac->ac_criteria = 10;
3264 spin_unlock(&pa->pa_lock);
3268 /* can we use group allocation? */
3269 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3272 /* inode may have no locality group for some reason */
3276 order = fls(ac->ac_o_ex.fe_len) - 1;
3277 if (order > PREALLOC_TB_SIZE - 1)
3278 /* The max size of hash table is PREALLOC_TB_SIZE */
3279 order = PREALLOC_TB_SIZE - 1;
3281 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3283 * search for the prealloc space that is having
3284 * minimal distance from the goal block.
3286 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3288 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3290 spin_lock(&pa->pa_lock);
3291 if (pa->pa_deleted == 0 &&
3292 pa->pa_free >= ac->ac_o_ex.fe_len) {
3294 cpa = ext4_mb_check_group_pa(goal_block,
3297 spin_unlock(&pa->pa_lock);
3302 ext4_mb_use_group_pa(ac, cpa);
3303 ac->ac_criteria = 20;
3310 * the function goes through all block freed in the group
3311 * but not yet committed and marks them used in in-core bitmap.
3312 * buddy must be generated from this bitmap
3313 * Need to be called with the ext4 group lock held
3315 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3319 struct ext4_group_info *grp;
3320 struct ext4_free_data *entry;
3322 grp = ext4_get_group_info(sb, group);
3323 n = rb_first(&(grp->bb_free_root));
3326 entry = rb_entry(n, struct ext4_free_data, node);
3327 ext4_set_bits(bitmap, entry->start_cluster, entry->count);
3334 * the function goes through all preallocation in this group and marks them
3335 * used in in-core bitmap. buddy must be generated from this bitmap
3336 * Need to be called with ext4 group lock held
3338 static noinline_for_stack
3339 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3342 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3343 struct ext4_prealloc_space *pa;
3344 struct list_head *cur;
3345 ext4_group_t groupnr;
3346 ext4_grpblk_t start;
3347 int preallocated = 0;
3350 /* all form of preallocation discards first load group,
3351 * so the only competing code is preallocation use.
3352 * we don't need any locking here
3353 * notice we do NOT ignore preallocations with pa_deleted
3354 * otherwise we could leave used blocks available for
3355 * allocation in buddy when concurrent ext4_mb_put_pa()
3356 * is dropping preallocation
3358 list_for_each(cur, &grp->bb_prealloc_list) {
3359 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3360 spin_lock(&pa->pa_lock);
3361 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3364 spin_unlock(&pa->pa_lock);
3365 if (unlikely(len == 0))
3367 BUG_ON(groupnr != group);
3368 ext4_set_bits(bitmap, start, len);
3369 preallocated += len;
3371 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3374 static void ext4_mb_pa_callback(struct rcu_head *head)
3376 struct ext4_prealloc_space *pa;
3377 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3378 kmem_cache_free(ext4_pspace_cachep, pa);
3382 * drops a reference to preallocated space descriptor
3383 * if this was the last reference and the space is consumed
3385 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3386 struct super_block *sb, struct ext4_prealloc_space *pa)
3389 ext4_fsblk_t grp_blk;
3391 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3394 /* in this short window concurrent discard can set pa_deleted */
3395 spin_lock(&pa->pa_lock);
3396 if (pa->pa_deleted == 1) {
3397 spin_unlock(&pa->pa_lock);
3402 spin_unlock(&pa->pa_lock);
3404 grp_blk = pa->pa_pstart;
3406 * If doing group-based preallocation, pa_pstart may be in the
3407 * next group when pa is used up
3409 if (pa->pa_type == MB_GROUP_PA)
3412 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3417 * P1 (buddy init) P2 (regular allocation)
3418 * find block B in PA
3419 * copy on-disk bitmap to buddy
3420 * mark B in on-disk bitmap
3421 * drop PA from group
3422 * mark all PAs in buddy
3424 * thus, P1 initializes buddy with B available. to prevent this
3425 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3428 ext4_lock_group(sb, grp);
3429 list_del(&pa->pa_group_list);
3430 ext4_unlock_group(sb, grp);
3432 spin_lock(pa->pa_obj_lock);
3433 list_del_rcu(&pa->pa_inode_list);
3434 spin_unlock(pa->pa_obj_lock);
3436 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3440 * creates new preallocated space for given inode
3442 static noinline_for_stack int
3443 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3445 struct super_block *sb = ac->ac_sb;
3446 struct ext4_sb_info *sbi = EXT4_SB(sb);
3447 struct ext4_prealloc_space *pa;
3448 struct ext4_group_info *grp;
3449 struct ext4_inode_info *ei;
3451 /* preallocate only when found space is larger then requested */
3452 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3453 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3454 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3456 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3460 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3466 /* we can't allocate as much as normalizer wants.
3467 * so, found space must get proper lstart
3468 * to cover original request */
3469 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3470 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3472 /* we're limited by original request in that
3473 * logical block must be covered any way
3474 * winl is window we can move our chunk within */
3475 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3477 /* also, we should cover whole original request */
3478 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3480 /* the smallest one defines real window */
3481 win = min(winl, wins);
3483 offs = ac->ac_o_ex.fe_logical %
3484 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3485 if (offs && offs < win)
3488 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3490 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3491 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3494 /* preallocation can change ac_b_ex, thus we store actually
3495 * allocated blocks for history */
3496 ac->ac_f_ex = ac->ac_b_ex;
3498 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3499 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3500 pa->pa_len = ac->ac_b_ex.fe_len;
3501 pa->pa_free = pa->pa_len;
3502 atomic_set(&pa->pa_count, 1);
3503 spin_lock_init(&pa->pa_lock);
3504 INIT_LIST_HEAD(&pa->pa_inode_list);
3505 INIT_LIST_HEAD(&pa->pa_group_list);
3507 pa->pa_type = MB_INODE_PA;
3509 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3510 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3511 trace_ext4_mb_new_inode_pa(ac, pa);
3513 ext4_mb_use_inode_pa(ac, pa);
3514 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3516 ei = EXT4_I(ac->ac_inode);
3517 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3519 pa->pa_obj_lock = &ei->i_prealloc_lock;
3520 pa->pa_inode = ac->ac_inode;
3522 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3523 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3524 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3526 spin_lock(pa->pa_obj_lock);
3527 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3528 spin_unlock(pa->pa_obj_lock);
3534 * creates new preallocated space for locality group inodes belongs to
3536 static noinline_for_stack int
3537 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3539 struct super_block *sb = ac->ac_sb;
3540 struct ext4_locality_group *lg;
3541 struct ext4_prealloc_space *pa;
3542 struct ext4_group_info *grp;
3544 /* preallocate only when found space is larger then requested */
3545 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3546 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3547 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3549 BUG_ON(ext4_pspace_cachep == NULL);
3550 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3554 /* preallocation can change ac_b_ex, thus we store actually
3555 * allocated blocks for history */
3556 ac->ac_f_ex = ac->ac_b_ex;
3558 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3559 pa->pa_lstart = pa->pa_pstart;
3560 pa->pa_len = ac->ac_b_ex.fe_len;
3561 pa->pa_free = pa->pa_len;
3562 atomic_set(&pa->pa_count, 1);
3563 spin_lock_init(&pa->pa_lock);
3564 INIT_LIST_HEAD(&pa->pa_inode_list);
3565 INIT_LIST_HEAD(&pa->pa_group_list);
3567 pa->pa_type = MB_GROUP_PA;
3569 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3570 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3571 trace_ext4_mb_new_group_pa(ac, pa);
3573 ext4_mb_use_group_pa(ac, pa);
3574 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3576 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3580 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3581 pa->pa_inode = NULL;
3583 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3584 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3585 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3588 * We will later add the new pa to the right bucket
3589 * after updating the pa_free in ext4_mb_release_context
3594 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3598 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3599 err = ext4_mb_new_group_pa(ac);
3601 err = ext4_mb_new_inode_pa(ac);
3606 * finds all unused blocks in on-disk bitmap, frees them in
3607 * in-core bitmap and buddy.
3608 * @pa must be unlinked from inode and group lists, so that
3609 * nobody else can find/use it.
3610 * the caller MUST hold group/inode locks.
3611 * TODO: optimize the case when there are no in-core structures yet
3613 static noinline_for_stack int
3614 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3615 struct ext4_prealloc_space *pa)
3617 struct super_block *sb = e4b->bd_sb;
3618 struct ext4_sb_info *sbi = EXT4_SB(sb);
3623 unsigned long long grp_blk_start;
3627 BUG_ON(pa->pa_deleted == 0);
3628 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3629 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3630 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3631 end = bit + pa->pa_len;
3634 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3637 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3638 mb_debug(1, " free preallocated %u/%u in group %u\n",
3639 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3640 (unsigned) next - bit, (unsigned) group);
3643 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3644 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3645 EXT4_C2B(sbi, bit)),
3647 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3650 if (free != pa->pa_free) {
3651 ext4_msg(e4b->bd_sb, KERN_CRIT,
3652 "pa %p: logic %lu, phys. %lu, len %lu",
3653 pa, (unsigned long) pa->pa_lstart,
3654 (unsigned long) pa->pa_pstart,
3655 (unsigned long) pa->pa_len);
3656 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3659 * pa is already deleted so we use the value obtained
3660 * from the bitmap and continue.
3663 atomic_add(free, &sbi->s_mb_discarded);
3668 static noinline_for_stack int
3669 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3670 struct ext4_prealloc_space *pa)
3672 struct super_block *sb = e4b->bd_sb;
3676 trace_ext4_mb_release_group_pa(pa);
3677 BUG_ON(pa->pa_deleted == 0);
3678 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3679 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3680 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3681 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3682 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3688 * releases all preallocations in given group
3690 * first, we need to decide discard policy:
3691 * - when do we discard
3693 * - how many do we discard
3694 * 1) how many requested
3696 static noinline_for_stack int
3697 ext4_mb_discard_group_preallocations(struct super_block *sb,
3698 ext4_group_t group, int needed)
3700 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3701 struct buffer_head *bitmap_bh = NULL;
3702 struct ext4_prealloc_space *pa, *tmp;
3703 struct list_head list;
3704 struct ext4_buddy e4b;
3709 mb_debug(1, "discard preallocation for group %u\n", group);
3711 if (list_empty(&grp->bb_prealloc_list))
3714 bitmap_bh = ext4_read_block_bitmap(sb, group);
3715 if (bitmap_bh == NULL) {
3716 ext4_error(sb, "Error reading block bitmap for %u", group);
3720 err = ext4_mb_load_buddy(sb, group, &e4b);
3722 ext4_error(sb, "Error loading buddy information for %u", group);
3728 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3730 INIT_LIST_HEAD(&list);
3732 ext4_lock_group(sb, group);
3733 list_for_each_entry_safe(pa, tmp,
3734 &grp->bb_prealloc_list, pa_group_list) {
3735 spin_lock(&pa->pa_lock);
3736 if (atomic_read(&pa->pa_count)) {
3737 spin_unlock(&pa->pa_lock);
3741 if (pa->pa_deleted) {
3742 spin_unlock(&pa->pa_lock);
3746 /* seems this one can be freed ... */
3749 /* we can trust pa_free ... */
3750 free += pa->pa_free;
3752 spin_unlock(&pa->pa_lock);
3754 list_del(&pa->pa_group_list);
3755 list_add(&pa->u.pa_tmp_list, &list);
3758 /* if we still need more blocks and some PAs were used, try again */
3759 if (free < needed && busy) {
3761 ext4_unlock_group(sb, group);
3763 * Yield the CPU here so that we don't get soft lockup
3764 * in non preempt case.
3770 /* found anything to free? */
3771 if (list_empty(&list)) {
3776 /* now free all selected PAs */
3777 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3779 /* remove from object (inode or locality group) */
3780 spin_lock(pa->pa_obj_lock);
3781 list_del_rcu(&pa->pa_inode_list);
3782 spin_unlock(pa->pa_obj_lock);
3784 if (pa->pa_type == MB_GROUP_PA)
3785 ext4_mb_release_group_pa(&e4b, pa);
3787 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3789 list_del(&pa->u.pa_tmp_list);
3790 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3794 ext4_unlock_group(sb, group);
3795 ext4_mb_unload_buddy(&e4b);
3801 * releases all non-used preallocated blocks for given inode
3803 * It's important to discard preallocations under i_data_sem
3804 * We don't want another block to be served from the prealloc
3805 * space when we are discarding the inode prealloc space.
3807 * FIXME!! Make sure it is valid at all the call sites
3809 void ext4_discard_preallocations(struct inode *inode)
3811 struct ext4_inode_info *ei = EXT4_I(inode);
3812 struct super_block *sb = inode->i_sb;
3813 struct buffer_head *bitmap_bh = NULL;
3814 struct ext4_prealloc_space *pa, *tmp;
3815 ext4_group_t group = 0;
3816 struct list_head list;
3817 struct ext4_buddy e4b;
3820 if (!S_ISREG(inode->i_mode)) {
3821 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3825 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3826 trace_ext4_discard_preallocations(inode);
3828 INIT_LIST_HEAD(&list);
3831 /* first, collect all pa's in the inode */
3832 spin_lock(&ei->i_prealloc_lock);
3833 while (!list_empty(&ei->i_prealloc_list)) {
3834 pa = list_entry(ei->i_prealloc_list.next,
3835 struct ext4_prealloc_space, pa_inode_list);
3836 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3837 spin_lock(&pa->pa_lock);
3838 if (atomic_read(&pa->pa_count)) {
3839 /* this shouldn't happen often - nobody should
3840 * use preallocation while we're discarding it */
3841 spin_unlock(&pa->pa_lock);
3842 spin_unlock(&ei->i_prealloc_lock);
3843 ext4_msg(sb, KERN_ERR,
3844 "uh-oh! used pa while discarding");
3846 schedule_timeout_uninterruptible(HZ);
3850 if (pa->pa_deleted == 0) {
3852 spin_unlock(&pa->pa_lock);
3853 list_del_rcu(&pa->pa_inode_list);
3854 list_add(&pa->u.pa_tmp_list, &list);
3858 /* someone is deleting pa right now */
3859 spin_unlock(&pa->pa_lock);
3860 spin_unlock(&ei->i_prealloc_lock);
3862 /* we have to wait here because pa_deleted
3863 * doesn't mean pa is already unlinked from
3864 * the list. as we might be called from
3865 * ->clear_inode() the inode will get freed
3866 * and concurrent thread which is unlinking
3867 * pa from inode's list may access already
3868 * freed memory, bad-bad-bad */
3870 /* XXX: if this happens too often, we can
3871 * add a flag to force wait only in case
3872 * of ->clear_inode(), but not in case of
3873 * regular truncate */
3874 schedule_timeout_uninterruptible(HZ);
3877 spin_unlock(&ei->i_prealloc_lock);
3879 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3880 BUG_ON(pa->pa_type != MB_INODE_PA);
3881 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3883 err = ext4_mb_load_buddy(sb, group, &e4b);
3885 ext4_error(sb, "Error loading buddy information for %u",
3890 bitmap_bh = ext4_read_block_bitmap(sb, group);
3891 if (bitmap_bh == NULL) {
3892 ext4_error(sb, "Error reading block bitmap for %u",
3894 ext4_mb_unload_buddy(&e4b);
3898 ext4_lock_group(sb, group);
3899 list_del(&pa->pa_group_list);
3900 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3901 ext4_unlock_group(sb, group);
3903 ext4_mb_unload_buddy(&e4b);
3906 list_del(&pa->u.pa_tmp_list);
3907 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3911 #ifdef CONFIG_EXT4_DEBUG
3912 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3914 struct super_block *sb = ac->ac_sb;
3915 ext4_group_t ngroups, i;
3917 if (!mb_enable_debug ||
3918 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3921 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: Can't allocate:"
3922 " Allocation context details:");
3923 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: status %d flags %d",
3924 ac->ac_status, ac->ac_flags);
3925 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: orig %lu/%lu/%lu@%lu, "
3926 "goal %lu/%lu/%lu@%lu, "
3927 "best %lu/%lu/%lu@%lu cr %d",
3928 (unsigned long)ac->ac_o_ex.fe_group,
3929 (unsigned long)ac->ac_o_ex.fe_start,
3930 (unsigned long)ac->ac_o_ex.fe_len,
3931 (unsigned long)ac->ac_o_ex.fe_logical,
3932 (unsigned long)ac->ac_g_ex.fe_group,
3933 (unsigned long)ac->ac_g_ex.fe_start,
3934 (unsigned long)ac->ac_g_ex.fe_len,
3935 (unsigned long)ac->ac_g_ex.fe_logical,
3936 (unsigned long)ac->ac_b_ex.fe_group,
3937 (unsigned long)ac->ac_b_ex.fe_start,
3938 (unsigned long)ac->ac_b_ex.fe_len,
3939 (unsigned long)ac->ac_b_ex.fe_logical,
3940 (int)ac->ac_criteria);
3941 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: %lu scanned, %d found",
3942 ac->ac_ex_scanned, ac->ac_found);
3943 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: groups: ");
3944 ngroups = ext4_get_groups_count(sb);
3945 for (i = 0; i < ngroups; i++) {
3946 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3947 struct ext4_prealloc_space *pa;
3948 ext4_grpblk_t start;
3949 struct list_head *cur;
3950 ext4_lock_group(sb, i);
3951 list_for_each(cur, &grp->bb_prealloc_list) {
3952 pa = list_entry(cur, struct ext4_prealloc_space,
3954 spin_lock(&pa->pa_lock);
3955 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3957 spin_unlock(&pa->pa_lock);
3958 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3961 ext4_unlock_group(sb, i);
3963 if (grp->bb_free == 0)
3965 printk(KERN_ERR "%u: %d/%d \n",
3966 i, grp->bb_free, grp->bb_fragments);
3968 printk(KERN_ERR "\n");
3971 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3978 * We use locality group preallocation for small size file. The size of the
3979 * file is determined by the current size or the resulting size after
3980 * allocation which ever is larger
3982 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3984 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3986 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3987 int bsbits = ac->ac_sb->s_blocksize_bits;
3990 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3993 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3996 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3997 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4000 if ((size == isize) &&
4001 !ext4_fs_is_busy(sbi) &&
4002 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4003 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4007 if (sbi->s_mb_group_prealloc <= 0) {
4008 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4012 /* don't use group allocation for large files */
4013 size = max(size, isize);
4014 if (size > sbi->s_mb_stream_request) {
4015 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4019 BUG_ON(ac->ac_lg != NULL);
4021 * locality group prealloc space are per cpu. The reason for having
4022 * per cpu locality group is to reduce the contention between block
4023 * request from multiple CPUs.
4025 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
4027 /* we're going to use group allocation */
4028 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4030 /* serialize all allocations in the group */
4031 mutex_lock(&ac->ac_lg->lg_mutex);
4034 static noinline_for_stack int
4035 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4036 struct ext4_allocation_request *ar)
4038 struct super_block *sb = ar->inode->i_sb;
4039 struct ext4_sb_info *sbi = EXT4_SB(sb);
4040 struct ext4_super_block *es = sbi->s_es;
4044 ext4_grpblk_t block;
4046 /* we can't allocate > group size */
4049 /* just a dirty hack to filter too big requests */
4050 if (len >= EXT4_CLUSTERS_PER_GROUP(sb) - 10)
4051 len = EXT4_CLUSTERS_PER_GROUP(sb) - 10;
4053 /* start searching from the goal */
4055 if (goal < le32_to_cpu(es->s_first_data_block) ||
4056 goal >= ext4_blocks_count(es))
4057 goal = le32_to_cpu(es->s_first_data_block);
4058 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4060 /* set up allocation goals */
4061 memset(ac, 0, sizeof(struct ext4_allocation_context));
4062 ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1);
4063 ac->ac_status = AC_STATUS_CONTINUE;
4065 ac->ac_inode = ar->inode;
4066 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4067 ac->ac_o_ex.fe_group = group;
4068 ac->ac_o_ex.fe_start = block;
4069 ac->ac_o_ex.fe_len = len;
4070 ac->ac_g_ex = ac->ac_o_ex;
4071 ac->ac_flags = ar->flags;
4073 /* we have to define context: we'll we work with a file or
4074 * locality group. this is a policy, actually */
4075 ext4_mb_group_or_file(ac);
4077 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4078 "left: %u/%u, right %u/%u to %swritable\n",
4079 (unsigned) ar->len, (unsigned) ar->logical,
4080 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4081 (unsigned) ar->lleft, (unsigned) ar->pleft,
4082 (unsigned) ar->lright, (unsigned) ar->pright,
4083 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4088 static noinline_for_stack void
4089 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4090 struct ext4_locality_group *lg,
4091 int order, int total_entries)
4093 ext4_group_t group = 0;
4094 struct ext4_buddy e4b;
4095 struct list_head discard_list;
4096 struct ext4_prealloc_space *pa, *tmp;
4098 mb_debug(1, "discard locality group preallocation\n");
4100 INIT_LIST_HEAD(&discard_list);
4102 spin_lock(&lg->lg_prealloc_lock);
4103 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4105 spin_lock(&pa->pa_lock);
4106 if (atomic_read(&pa->pa_count)) {
4108 * This is the pa that we just used
4109 * for block allocation. So don't
4112 spin_unlock(&pa->pa_lock);
4115 if (pa->pa_deleted) {
4116 spin_unlock(&pa->pa_lock);
4119 /* only lg prealloc space */
4120 BUG_ON(pa->pa_type != MB_GROUP_PA);
4122 /* seems this one can be freed ... */
4124 spin_unlock(&pa->pa_lock);
4126 list_del_rcu(&pa->pa_inode_list);
4127 list_add(&pa->u.pa_tmp_list, &discard_list);
4130 if (total_entries <= 5) {
4132 * we want to keep only 5 entries
4133 * allowing it to grow to 8. This
4134 * mak sure we don't call discard
4135 * soon for this list.
4140 spin_unlock(&lg->lg_prealloc_lock);
4142 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4144 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4145 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4146 ext4_error(sb, "Error loading buddy information for %u",
4150 ext4_lock_group(sb, group);
4151 list_del(&pa->pa_group_list);
4152 ext4_mb_release_group_pa(&e4b, pa);
4153 ext4_unlock_group(sb, group);
4155 ext4_mb_unload_buddy(&e4b);
4156 list_del(&pa->u.pa_tmp_list);
4157 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4162 * We have incremented pa_count. So it cannot be freed at this
4163 * point. Also we hold lg_mutex. So no parallel allocation is
4164 * possible from this lg. That means pa_free cannot be updated.
4166 * A parallel ext4_mb_discard_group_preallocations is possible.
4167 * which can cause the lg_prealloc_list to be updated.
4170 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4172 int order, added = 0, lg_prealloc_count = 1;
4173 struct super_block *sb = ac->ac_sb;
4174 struct ext4_locality_group *lg = ac->ac_lg;
4175 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4177 order = fls(pa->pa_free) - 1;
4178 if (order > PREALLOC_TB_SIZE - 1)
4179 /* The max size of hash table is PREALLOC_TB_SIZE */
4180 order = PREALLOC_TB_SIZE - 1;
4181 /* Add the prealloc space to lg */
4183 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4185 spin_lock(&tmp_pa->pa_lock);
4186 if (tmp_pa->pa_deleted) {
4187 spin_unlock(&tmp_pa->pa_lock);
4190 if (!added && pa->pa_free < tmp_pa->pa_free) {
4191 /* Add to the tail of the previous entry */
4192 list_add_tail_rcu(&pa->pa_inode_list,
4193 &tmp_pa->pa_inode_list);
4196 * we want to count the total
4197 * number of entries in the list
4200 spin_unlock(&tmp_pa->pa_lock);
4201 lg_prealloc_count++;
4204 list_add_tail_rcu(&pa->pa_inode_list,
4205 &lg->lg_prealloc_list[order]);
4208 /* Now trim the list to be not more than 8 elements */
4209 if (lg_prealloc_count > 8) {
4210 ext4_mb_discard_lg_preallocations(sb, lg,
4211 order, lg_prealloc_count);
4218 * release all resource we used in allocation
4220 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4222 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4223 struct ext4_prealloc_space *pa = ac->ac_pa;
4225 if (pa->pa_type == MB_GROUP_PA) {
4226 /* see comment in ext4_mb_use_group_pa() */
4227 spin_lock(&pa->pa_lock);
4228 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4229 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4230 pa->pa_free -= ac->ac_b_ex.fe_len;
4231 pa->pa_len -= ac->ac_b_ex.fe_len;
4232 spin_unlock(&pa->pa_lock);
4237 * We want to add the pa to the right bucket.
4238 * Remove it from the list and while adding
4239 * make sure the list to which we are adding
4242 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4243 spin_lock(pa->pa_obj_lock);
4244 list_del_rcu(&pa->pa_inode_list);
4245 spin_unlock(pa->pa_obj_lock);
4246 ext4_mb_add_n_trim(ac);
4248 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4250 if (ac->ac_bitmap_page)
4251 page_cache_release(ac->ac_bitmap_page);
4252 if (ac->ac_buddy_page)
4253 page_cache_release(ac->ac_buddy_page);
4254 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4255 mutex_unlock(&ac->ac_lg->lg_mutex);
4256 ext4_mb_collect_stats(ac);
4260 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4262 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4266 trace_ext4_mb_discard_preallocations(sb, needed);
4267 for (i = 0; i < ngroups && needed > 0; i++) {
4268 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4277 * Main entry point into mballoc to allocate blocks
4278 * it tries to use preallocation first, then falls back
4279 * to usual allocation
4281 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4282 struct ext4_allocation_request *ar, int *errp)
4285 struct ext4_allocation_context *ac = NULL;
4286 struct ext4_sb_info *sbi;
4287 struct super_block *sb;
4288 ext4_fsblk_t block = 0;
4289 unsigned int inquota = 0;
4290 unsigned int reserv_clstrs = 0;
4292 sb = ar->inode->i_sb;
4295 trace_ext4_request_blocks(ar);
4297 /* Allow to use superuser reservation for quota file */
4298 if (IS_NOQUOTA(ar->inode))
4299 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4302 * For delayed allocation, we could skip the ENOSPC and
4303 * EDQUOT check, as blocks and quotas have been already
4304 * reserved when data being copied into pagecache.
4306 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4307 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4309 /* Without delayed allocation we need to verify
4310 * there is enough free blocks to do block allocation
4311 * and verify allocation doesn't exceed the quota limits.
4314 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4316 /* let others to free the space */
4318 ar->len = ar->len >> 1;
4324 reserv_clstrs = ar->len;
4325 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4326 dquot_alloc_block_nofail(ar->inode,
4327 EXT4_C2B(sbi, ar->len));
4330 dquot_alloc_block(ar->inode,
4331 EXT4_C2B(sbi, ar->len))) {
4333 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4344 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4351 *errp = ext4_mb_initialize_context(ac, ar);
4357 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4358 if (!ext4_mb_use_preallocated(ac)) {
4359 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4360 ext4_mb_normalize_request(ac, ar);
4362 /* allocate space in core */
4363 *errp = ext4_mb_regular_allocator(ac);
4367 /* as we've just preallocated more space than
4368 * user requested orinally, we store allocated
4369 * space in a special descriptor */
4370 if (ac->ac_status == AC_STATUS_FOUND &&
4371 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4372 ext4_mb_new_preallocation(ac);
4374 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4375 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4376 if (*errp == -EAGAIN) {
4378 * drop the reference that we took
4379 * in ext4_mb_use_best_found
4381 ext4_mb_release_context(ac);
4382 ac->ac_b_ex.fe_group = 0;
4383 ac->ac_b_ex.fe_start = 0;
4384 ac->ac_b_ex.fe_len = 0;
4385 ac->ac_status = AC_STATUS_CONTINUE;
4389 ext4_discard_allocated_blocks(ac);
4391 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4392 ar->len = ac->ac_b_ex.fe_len;
4395 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4402 ac->ac_b_ex.fe_len = 0;
4404 ext4_mb_show_ac(ac);
4406 ext4_mb_release_context(ac);
4409 kmem_cache_free(ext4_ac_cachep, ac);
4410 if (inquota && ar->len < inquota)
4411 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4413 if (!ext4_test_inode_state(ar->inode,
4414 EXT4_STATE_DELALLOC_RESERVED))
4415 /* release all the reserved blocks if non delalloc */
4416 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4420 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4426 * We can merge two free data extents only if the physical blocks
4427 * are contiguous, AND the extents were freed by the same transaction,
4428 * AND the blocks are associated with the same group.
4430 static int can_merge(struct ext4_free_data *entry1,
4431 struct ext4_free_data *entry2)
4433 if ((entry1->t_tid == entry2->t_tid) &&
4434 (entry1->group == entry2->group) &&
4435 ((entry1->start_cluster + entry1->count) == entry2->start_cluster))
4440 static noinline_for_stack int
4441 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4442 struct ext4_free_data *new_entry)
4444 ext4_group_t group = e4b->bd_group;
4445 ext4_grpblk_t cluster;
4446 struct ext4_free_data *entry;
4447 struct ext4_group_info *db = e4b->bd_info;
4448 struct super_block *sb = e4b->bd_sb;
4449 struct ext4_sb_info *sbi = EXT4_SB(sb);
4450 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4451 struct rb_node *parent = NULL, *new_node;
4453 BUG_ON(!ext4_handle_valid(handle));
4454 BUG_ON(e4b->bd_bitmap_page == NULL);
4455 BUG_ON(e4b->bd_buddy_page == NULL);
4457 new_node = &new_entry->node;
4458 cluster = new_entry->start_cluster;
4461 /* first free block exent. We need to
4462 protect buddy cache from being freed,
4463 * otherwise we'll refresh it from
4464 * on-disk bitmap and lose not-yet-available
4466 page_cache_get(e4b->bd_buddy_page);
4467 page_cache_get(e4b->bd_bitmap_page);
4471 entry = rb_entry(parent, struct ext4_free_data, node);
4472 if (cluster < entry->start_cluster)
4474 else if (cluster >= (entry->start_cluster + entry->count))
4475 n = &(*n)->rb_right;
4477 ext4_grp_locked_error(sb, group, 0,
4478 ext4_group_first_block_no(sb, group) +
4479 EXT4_C2B(sbi, cluster),
4480 "Block already on to-be-freed list");
4485 rb_link_node(new_node, parent, n);
4486 rb_insert_color(new_node, &db->bb_free_root);
4488 /* Now try to see the extent can be merged to left and right */
4489 node = rb_prev(new_node);
4491 entry = rb_entry(node, struct ext4_free_data, node);
4492 if (can_merge(entry, new_entry)) {
4493 new_entry->start_cluster = entry->start_cluster;
4494 new_entry->count += entry->count;
4495 rb_erase(node, &(db->bb_free_root));
4496 spin_lock(&sbi->s_md_lock);
4497 list_del(&entry->list);
4498 spin_unlock(&sbi->s_md_lock);
4499 kmem_cache_free(ext4_free_ext_cachep, entry);
4503 node = rb_next(new_node);
4505 entry = rb_entry(node, struct ext4_free_data, node);
4506 if (can_merge(new_entry, entry)) {
4507 new_entry->count += entry->count;
4508 rb_erase(node, &(db->bb_free_root));
4509 spin_lock(&sbi->s_md_lock);
4510 list_del(&entry->list);
4511 spin_unlock(&sbi->s_md_lock);
4512 kmem_cache_free(ext4_free_ext_cachep, entry);
4515 /* Add the extent to transaction's private list */
4516 spin_lock(&sbi->s_md_lock);
4517 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4518 spin_unlock(&sbi->s_md_lock);
4523 * ext4_free_blocks() -- Free given blocks and update quota
4524 * @handle: handle for this transaction
4526 * @block: start physical block to free
4527 * @count: number of blocks to count
4528 * @flags: flags used by ext4_free_blocks
4530 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4531 struct buffer_head *bh, ext4_fsblk_t block,
4532 unsigned long count, int flags)
4534 struct buffer_head *bitmap_bh = NULL;
4535 struct super_block *sb = inode->i_sb;
4536 struct ext4_group_desc *gdp;
4537 unsigned long freed = 0;
4538 unsigned int overflow;
4540 struct buffer_head *gd_bh;
4541 ext4_group_t block_group;
4542 struct ext4_sb_info *sbi;
4543 struct ext4_buddy e4b;
4544 unsigned int count_clusters;
4550 BUG_ON(block != bh->b_blocknr);
4552 block = bh->b_blocknr;
4556 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4557 !ext4_data_block_valid(sbi, block, count)) {
4558 ext4_error(sb, "Freeing blocks not in datazone - "
4559 "block = %llu, count = %lu", block, count);
4563 ext4_debug("freeing block %llu\n", block);
4564 trace_ext4_free_blocks(inode, block, count, flags);
4566 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4567 struct buffer_head *tbh = bh;
4570 BUG_ON(bh && (count > 1));
4572 for (i = 0; i < count; i++) {
4574 tbh = sb_find_get_block(inode->i_sb,
4578 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4579 inode, tbh, block + i);
4584 * We need to make sure we don't reuse the freed block until
4585 * after the transaction is committed, which we can do by
4586 * treating the block as metadata, below. We make an
4587 * exception if the inode is to be written in writeback mode
4588 * since writeback mode has weak data consistency guarantees.
4590 if (!ext4_should_writeback_data(inode))
4591 flags |= EXT4_FREE_BLOCKS_METADATA;
4594 * If the extent to be freed does not begin on a cluster
4595 * boundary, we need to deal with partial clusters at the
4596 * beginning and end of the extent. Normally we will free
4597 * blocks at the beginning or the end unless we are explicitly
4598 * requested to avoid doing so.
4600 overflow = block & (sbi->s_cluster_ratio - 1);
4602 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4603 overflow = sbi->s_cluster_ratio - overflow;
4605 if (count > overflow)
4614 overflow = count & (sbi->s_cluster_ratio - 1);
4616 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4617 if (count > overflow)
4622 count += sbi->s_cluster_ratio - overflow;
4627 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4630 * Check to see if we are freeing blocks across a group
4633 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4634 overflow = EXT4_C2B(sbi, bit) + count -
4635 EXT4_BLOCKS_PER_GROUP(sb);
4638 count_clusters = EXT4_B2C(sbi, count);
4639 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4644 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4650 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4651 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4652 in_range(block, ext4_inode_table(sb, gdp),
4653 EXT4_SB(sb)->s_itb_per_group) ||
4654 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4655 EXT4_SB(sb)->s_itb_per_group)) {
4657 ext4_error(sb, "Freeing blocks in system zone - "
4658 "Block = %llu, count = %lu", block, count);
4659 /* err = 0. ext4_std_error should be a no op */
4663 BUFFER_TRACE(bitmap_bh, "getting write access");
4664 err = ext4_journal_get_write_access(handle, bitmap_bh);
4669 * We are about to modify some metadata. Call the journal APIs
4670 * to unshare ->b_data if a currently-committing transaction is
4673 BUFFER_TRACE(gd_bh, "get_write_access");
4674 err = ext4_journal_get_write_access(handle, gd_bh);
4677 #ifdef AGGRESSIVE_CHECK
4680 for (i = 0; i < count_clusters; i++)
4681 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4684 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4686 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4690 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4691 struct ext4_free_data *new_entry;
4693 * blocks being freed are metadata. these blocks shouldn't
4694 * be used until this transaction is committed
4696 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4701 new_entry->start_cluster = bit;
4702 new_entry->group = block_group;
4703 new_entry->count = count_clusters;
4704 new_entry->t_tid = handle->h_transaction->t_tid;
4706 ext4_lock_group(sb, block_group);
4707 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4708 ext4_mb_free_metadata(handle, &e4b, new_entry);
4710 /* need to update group_info->bb_free and bitmap
4711 * with group lock held. generate_buddy look at
4712 * them with group lock_held
4714 ext4_lock_group(sb, block_group);
4715 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4716 mb_free_blocks(inode, &e4b, bit, count_clusters);
4719 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4720 ext4_free_group_clusters_set(sb, gdp, ret);
4721 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4722 ext4_unlock_group(sb, block_group);
4723 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4725 if (sbi->s_log_groups_per_flex) {
4726 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4727 atomic_add(count_clusters,
4728 &sbi->s_flex_groups[flex_group].free_clusters);
4731 ext4_mb_unload_buddy(&e4b);
4735 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4736 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4738 /* We dirtied the bitmap block */
4739 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4740 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4742 /* And the group descriptor block */
4743 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4744 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4748 if (overflow && !err) {
4754 ext4_mark_super_dirty(sb);
4757 ext4_std_error(sb, err);
4762 * ext4_group_add_blocks() -- Add given blocks to an existing group
4763 * @handle: handle to this transaction
4765 * @block: start physcial block to add to the block group
4766 * @count: number of blocks to free
4768 * This marks the blocks as free in the bitmap and buddy.
4770 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4771 ext4_fsblk_t block, unsigned long count)
4773 struct buffer_head *bitmap_bh = NULL;
4774 struct buffer_head *gd_bh;
4775 ext4_group_t block_group;
4778 struct ext4_group_desc *desc;
4779 struct ext4_sb_info *sbi = EXT4_SB(sb);
4780 struct ext4_buddy e4b;
4781 int err = 0, ret, blk_free_count;
4782 ext4_grpblk_t blocks_freed;
4784 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4789 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4791 * Check to see if we are freeing blocks across a group
4794 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4795 ext4_warning(sb, "too much blocks added to group %u\n",
4801 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4807 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4813 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4814 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4815 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4816 in_range(block + count - 1, ext4_inode_table(sb, desc),
4817 sbi->s_itb_per_group)) {
4818 ext4_error(sb, "Adding blocks in system zones - "
4819 "Block = %llu, count = %lu",
4825 BUFFER_TRACE(bitmap_bh, "getting write access");
4826 err = ext4_journal_get_write_access(handle, bitmap_bh);
4831 * We are about to modify some metadata. Call the journal APIs
4832 * to unshare ->b_data if a currently-committing transaction is
4835 BUFFER_TRACE(gd_bh, "get_write_access");
4836 err = ext4_journal_get_write_access(handle, gd_bh);
4840 for (i = 0, blocks_freed = 0; i < count; i++) {
4841 BUFFER_TRACE(bitmap_bh, "clear bit");
4842 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4843 ext4_error(sb, "bit already cleared for block %llu",
4844 (ext4_fsblk_t)(block + i));
4845 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4851 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4856 * need to update group_info->bb_free and bitmap
4857 * with group lock held. generate_buddy look at
4858 * them with group lock_held
4860 ext4_lock_group(sb, block_group);
4861 mb_clear_bits(bitmap_bh->b_data, bit, count);
4862 mb_free_blocks(NULL, &e4b, bit, count);
4863 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4864 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4865 desc->bg_checksum = ext4_group_desc_csum(sbi, block_group, desc);
4866 ext4_unlock_group(sb, block_group);
4867 percpu_counter_add(&sbi->s_freeclusters_counter,
4868 EXT4_B2C(sbi, blocks_freed));
4870 if (sbi->s_log_groups_per_flex) {
4871 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4872 atomic_add(EXT4_B2C(sbi, blocks_freed),
4873 &sbi->s_flex_groups[flex_group].free_clusters);
4876 ext4_mb_unload_buddy(&e4b);
4878 /* We dirtied the bitmap block */
4879 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4880 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4882 /* And the group descriptor block */
4883 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4884 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4890 ext4_std_error(sb, err);
4895 * ext4_trim_extent -- function to TRIM one single free extent in the group
4896 * @sb: super block for the file system
4897 * @start: starting block of the free extent in the alloc. group
4898 * @count: number of blocks to TRIM
4899 * @group: alloc. group we are working with
4900 * @e4b: ext4 buddy for the group
4902 * Trim "count" blocks starting at "start" in the "group". To assure that no
4903 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4904 * be called with under the group lock.
4906 static void ext4_trim_extent(struct super_block *sb, int start, int count,
4907 ext4_group_t group, struct ext4_buddy *e4b)
4909 struct ext4_free_extent ex;
4911 trace_ext4_trim_extent(sb, group, start, count);
4913 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4915 ex.fe_start = start;
4916 ex.fe_group = group;
4920 * Mark blocks used, so no one can reuse them while
4923 mb_mark_used(e4b, &ex);
4924 ext4_unlock_group(sb, group);
4925 ext4_issue_discard(sb, group, start, count);
4926 ext4_lock_group(sb, group);
4927 mb_free_blocks(NULL, e4b, start, ex.fe_len);
4931 * ext4_trim_all_free -- function to trim all free space in alloc. group
4932 * @sb: super block for file system
4933 * @group: group to be trimmed
4934 * @start: first group block to examine
4935 * @max: last group block to examine
4936 * @minblocks: minimum extent block count
4938 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4939 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4943 * ext4_trim_all_free walks through group's block bitmap searching for free
4944 * extents. When the free extent is found, mark it as used in group buddy
4945 * bitmap. Then issue a TRIM command on this extent and free the extent in
4946 * the group buddy bitmap. This is done until whole group is scanned.
4948 static ext4_grpblk_t
4949 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4950 ext4_grpblk_t start, ext4_grpblk_t max,
4951 ext4_grpblk_t minblocks)
4954 ext4_grpblk_t next, count = 0, free_count = 0;
4955 struct ext4_buddy e4b;
4958 trace_ext4_trim_all_free(sb, group, start, max);
4960 ret = ext4_mb_load_buddy(sb, group, &e4b);
4962 ext4_error(sb, "Error in loading buddy "
4963 "information for %u", group);
4966 bitmap = e4b.bd_bitmap;
4968 ext4_lock_group(sb, group);
4969 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4970 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4973 start = (e4b.bd_info->bb_first_free > start) ?
4974 e4b.bd_info->bb_first_free : start;
4976 while (start < max) {
4977 start = mb_find_next_zero_bit(bitmap, max, start);
4980 next = mb_find_next_bit(bitmap, max, start);
4982 if ((next - start) >= minblocks) {
4983 ext4_trim_extent(sb, start,
4984 next - start, group, &e4b);
4985 count += next - start;
4987 free_count += next - start;
4990 if (fatal_signal_pending(current)) {
4991 count = -ERESTARTSYS;
4995 if (need_resched()) {
4996 ext4_unlock_group(sb, group);
4998 ext4_lock_group(sb, group);
5001 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5006 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5008 ext4_unlock_group(sb, group);
5009 ext4_mb_unload_buddy(&e4b);
5011 ext4_debug("trimmed %d blocks in the group %d\n",
5018 * ext4_trim_fs() -- trim ioctl handle function
5019 * @sb: superblock for filesystem
5020 * @range: fstrim_range structure
5022 * start: First Byte to trim
5023 * len: number of Bytes to trim from start
5024 * minlen: minimum extent length in Bytes
5025 * ext4_trim_fs goes through all allocation groups containing Bytes from
5026 * start to start+len. For each such a group ext4_trim_all_free function
5027 * is invoked to trim all free space.
5029 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5031 struct ext4_group_info *grp;
5032 ext4_group_t first_group, last_group;
5033 ext4_group_t group, ngroups = ext4_get_groups_count(sb);
5034 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5035 uint64_t start, len, minlen, trimmed = 0;
5036 ext4_fsblk_t first_data_blk =
5037 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5040 start = range->start >> sb->s_blocksize_bits;
5041 len = range->len >> sb->s_blocksize_bits;
5042 minlen = range->minlen >> sb->s_blocksize_bits;
5044 if (unlikely(minlen > EXT4_CLUSTERS_PER_GROUP(sb)))
5046 if (start + len <= first_data_blk)
5048 if (start < first_data_blk) {
5049 len -= first_data_blk - start;
5050 start = first_data_blk;
5053 /* Determine first and last group to examine based on start and len */
5054 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5055 &first_group, &first_cluster);
5056 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) (start + len),
5057 &last_group, &last_cluster);
5058 last_group = (last_group > ngroups - 1) ? ngroups - 1 : last_group;
5059 last_cluster = EXT4_CLUSTERS_PER_GROUP(sb);
5061 if (first_group > last_group)
5064 for (group = first_group; group <= last_group; group++) {
5065 grp = ext4_get_group_info(sb, group);
5066 /* We only do this if the grp has never been initialized */
5067 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5068 ret = ext4_mb_init_group(sb, group);
5074 * For all the groups except the last one, last block will
5075 * always be EXT4_BLOCKS_PER_GROUP(sb), so we only need to
5076 * change it for the last group in which case start +
5077 * len < EXT4_BLOCKS_PER_GROUP(sb).
5079 if (first_cluster + len < EXT4_CLUSTERS_PER_GROUP(sb))
5080 last_cluster = first_cluster + len;
5081 len -= last_cluster - first_cluster;
5083 if (grp->bb_free >= minlen) {
5084 cnt = ext4_trim_all_free(sb, group, first_cluster,
5085 last_cluster, minlen);
5094 range->len = trimmed * sb->s_blocksize;
5097 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);