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 1 */
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);
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 if (WARN_ON(count == 0))
1317 BUG_ON(first + count > (sb->s_blocksize << 3));
1318 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1319 mb_check_buddy(e4b);
1320 mb_free_blocks_double(inode, e4b, first, count);
1322 e4b->bd_info->bb_free += count;
1323 if (first < e4b->bd_info->bb_first_free)
1324 e4b->bd_info->bb_first_free = first;
1326 /* let's maintain fragments counter */
1328 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1329 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1330 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1332 e4b->bd_info->bb_fragments--;
1333 else if (!block && !max)
1334 e4b->bd_info->bb_fragments++;
1336 /* let's maintain buddy itself */
1337 while (count-- > 0) {
1341 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1342 ext4_fsblk_t blocknr;
1344 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1345 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1346 ext4_grp_locked_error(sb, e4b->bd_group,
1347 inode ? inode->i_ino : 0,
1349 "freeing already freed block "
1352 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1353 e4b->bd_info->bb_counters[order]++;
1355 /* start of the buddy */
1356 buddy = mb_find_buddy(e4b, order, &max);
1360 if (mb_test_bit(block, buddy) ||
1361 mb_test_bit(block + 1, buddy))
1364 /* both the buddies are free, try to coalesce them */
1365 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1371 /* for special purposes, we don't set
1372 * free bits in bitmap */
1373 mb_set_bit(block, buddy);
1374 mb_set_bit(block + 1, buddy);
1376 e4b->bd_info->bb_counters[order]--;
1377 e4b->bd_info->bb_counters[order]--;
1381 e4b->bd_info->bb_counters[order]++;
1383 mb_clear_bit(block, buddy2);
1387 mb_set_largest_free_order(sb, e4b->bd_info);
1388 mb_check_buddy(e4b);
1391 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1392 int needed, struct ext4_free_extent *ex)
1398 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1401 buddy = mb_find_buddy(e4b, order, &max);
1402 BUG_ON(buddy == NULL);
1403 BUG_ON(block >= max);
1404 if (mb_test_bit(block, buddy)) {
1411 /* FIXME dorp order completely ? */
1412 if (likely(order == 0)) {
1413 /* find actual order */
1414 order = mb_find_order_for_block(e4b, block);
1415 block = block >> order;
1418 ex->fe_len = 1 << order;
1419 ex->fe_start = block << order;
1420 ex->fe_group = e4b->bd_group;
1422 /* calc difference from given start */
1423 next = next - ex->fe_start;
1425 ex->fe_start += next;
1427 while (needed > ex->fe_len &&
1428 (buddy = mb_find_buddy(e4b, order, &max))) {
1430 if (block + 1 >= max)
1433 next = (block + 1) * (1 << order);
1434 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1437 order = 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++) {
2033 * Artificially restricted ngroups for non-extent
2034 * files makes group > ngroups possible on first loop.
2036 if (group >= ngroups)
2039 /* This now checks without needing the buddy page */
2040 if (!ext4_mb_good_group(ac, group, cr))
2043 err = ext4_mb_load_buddy(sb, group, &e4b);
2047 ext4_lock_group(sb, group);
2050 * We need to check again after locking the
2053 if (!ext4_mb_good_group(ac, group, cr)) {
2054 ext4_unlock_group(sb, group);
2055 ext4_mb_unload_buddy(&e4b);
2059 ac->ac_groups_scanned++;
2061 ext4_mb_simple_scan_group(ac, &e4b);
2062 else if (cr == 1 && sbi->s_stripe &&
2063 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2064 ext4_mb_scan_aligned(ac, &e4b);
2066 ext4_mb_complex_scan_group(ac, &e4b);
2068 ext4_unlock_group(sb, group);
2069 ext4_mb_unload_buddy(&e4b);
2071 if (ac->ac_status != AC_STATUS_CONTINUE)
2076 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2077 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2079 * We've been searching too long. Let's try to allocate
2080 * the best chunk we've found so far
2083 ext4_mb_try_best_found(ac, &e4b);
2084 if (ac->ac_status != AC_STATUS_FOUND) {
2086 * Someone more lucky has already allocated it.
2087 * The only thing we can do is just take first
2089 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2091 ac->ac_b_ex.fe_group = 0;
2092 ac->ac_b_ex.fe_start = 0;
2093 ac->ac_b_ex.fe_len = 0;
2094 ac->ac_status = AC_STATUS_CONTINUE;
2095 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2097 atomic_inc(&sbi->s_mb_lost_chunks);
2105 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2107 struct super_block *sb = seq->private;
2110 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2113 return (void *) ((unsigned long) group);
2116 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2118 struct super_block *sb = seq->private;
2122 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2125 return (void *) ((unsigned long) group);
2128 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2130 struct super_block *sb = seq->private;
2131 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2134 struct ext4_buddy e4b;
2136 struct ext4_group_info info;
2137 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2142 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2143 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2144 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2145 "group", "free", "frags", "first",
2146 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2147 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2149 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2150 sizeof(struct ext4_group_info);
2151 err = ext4_mb_load_buddy(sb, group, &e4b);
2153 seq_printf(seq, "#%-5u: I/O error\n", group);
2156 ext4_lock_group(sb, group);
2157 memcpy(&sg, ext4_get_group_info(sb, group), i);
2158 ext4_unlock_group(sb, group);
2159 ext4_mb_unload_buddy(&e4b);
2161 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2162 sg.info.bb_fragments, sg.info.bb_first_free);
2163 for (i = 0; i <= 13; i++)
2164 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2165 sg.info.bb_counters[i] : 0);
2166 seq_printf(seq, " ]\n");
2171 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2175 static const struct seq_operations ext4_mb_seq_groups_ops = {
2176 .start = ext4_mb_seq_groups_start,
2177 .next = ext4_mb_seq_groups_next,
2178 .stop = ext4_mb_seq_groups_stop,
2179 .show = ext4_mb_seq_groups_show,
2182 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2184 struct super_block *sb = PDE(inode)->data;
2187 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2189 struct seq_file *m = file->private_data;
2196 static const struct file_operations ext4_mb_seq_groups_fops = {
2197 .owner = THIS_MODULE,
2198 .open = ext4_mb_seq_groups_open,
2200 .llseek = seq_lseek,
2201 .release = seq_release,
2204 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2206 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2207 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2213 /* Create and initialize ext4_group_info data for the given group. */
2214 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2215 struct ext4_group_desc *desc)
2219 struct ext4_sb_info *sbi = EXT4_SB(sb);
2220 struct ext4_group_info **meta_group_info;
2221 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2224 * First check if this group is the first of a reserved block.
2225 * If it's true, we have to allocate a new table of pointers
2226 * to ext4_group_info structures
2228 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2229 metalen = sizeof(*meta_group_info) <<
2230 EXT4_DESC_PER_BLOCK_BITS(sb);
2231 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2232 if (meta_group_info == NULL) {
2233 ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate mem "
2234 "for a buddy group");
2235 goto exit_meta_group_info;
2237 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2242 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2243 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2245 meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL);
2246 if (meta_group_info[i] == NULL) {
2247 ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate buddy mem");
2248 goto exit_group_info;
2250 memset(meta_group_info[i], 0, kmem_cache_size(cachep));
2251 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2252 &(meta_group_info[i]->bb_state));
2255 * initialize bb_free to be able to skip
2256 * empty groups without initialization
2258 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2259 meta_group_info[i]->bb_free =
2260 ext4_free_clusters_after_init(sb, group, desc);
2262 meta_group_info[i]->bb_free =
2263 ext4_free_group_clusters(sb, desc);
2266 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2267 init_rwsem(&meta_group_info[i]->alloc_sem);
2268 meta_group_info[i]->bb_free_root = RB_ROOT;
2269 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2273 struct buffer_head *bh;
2274 meta_group_info[i]->bb_bitmap =
2275 kmalloc(sb->s_blocksize, GFP_KERNEL);
2276 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2277 bh = ext4_read_block_bitmap(sb, group);
2279 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2288 /* If a meta_group_info table has been allocated, release it now */
2289 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2290 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2291 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2293 exit_meta_group_info:
2295 } /* ext4_mb_add_groupinfo */
2297 static int ext4_mb_init_backend(struct super_block *sb)
2299 ext4_group_t ngroups = ext4_get_groups_count(sb);
2301 struct ext4_sb_info *sbi = EXT4_SB(sb);
2302 struct ext4_super_block *es = sbi->s_es;
2303 int num_meta_group_infos;
2304 int num_meta_group_infos_max;
2306 struct ext4_group_desc *desc;
2307 struct kmem_cache *cachep;
2309 /* This is the number of blocks used by GDT */
2310 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2311 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2314 * This is the total number of blocks used by GDT including
2315 * the number of reserved blocks for GDT.
2316 * The s_group_info array is allocated with this value
2317 * to allow a clean online resize without a complex
2318 * manipulation of pointer.
2319 * The drawback is the unused memory when no resize
2320 * occurs but it's very low in terms of pages
2321 * (see comments below)
2322 * Need to handle this properly when META_BG resizing is allowed
2324 num_meta_group_infos_max = num_meta_group_infos +
2325 le16_to_cpu(es->s_reserved_gdt_blocks);
2328 * array_size is the size of s_group_info array. We round it
2329 * to the next power of two because this approximation is done
2330 * internally by kmalloc so we can have some more memory
2331 * for free here (e.g. may be used for META_BG resize).
2334 while (array_size < sizeof(*sbi->s_group_info) *
2335 num_meta_group_infos_max)
2336 array_size = array_size << 1;
2337 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2338 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2339 * So a two level scheme suffices for now. */
2340 sbi->s_group_info = ext4_kvzalloc(array_size, GFP_KERNEL);
2341 if (sbi->s_group_info == NULL) {
2342 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2345 sbi->s_buddy_cache = new_inode(sb);
2346 if (sbi->s_buddy_cache == NULL) {
2347 ext4_msg(sb, KERN_ERR, "can't get new inode");
2350 /* To avoid potentially colliding with an valid on-disk inode number,
2351 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2352 * not in the inode hash, so it should never be found by iget(), but
2353 * this will avoid confusion if it ever shows up during debugging. */
2354 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2355 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2356 for (i = 0; i < ngroups; i++) {
2357 desc = ext4_get_group_desc(sb, i, NULL);
2359 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2362 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2369 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2371 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2372 i = num_meta_group_infos;
2374 kfree(sbi->s_group_info[i]);
2375 iput(sbi->s_buddy_cache);
2377 ext4_kvfree(sbi->s_group_info);
2381 static void ext4_groupinfo_destroy_slabs(void)
2385 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2386 if (ext4_groupinfo_caches[i])
2387 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2388 ext4_groupinfo_caches[i] = NULL;
2392 static int ext4_groupinfo_create_slab(size_t size)
2394 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2396 int blocksize_bits = order_base_2(size);
2397 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2398 struct kmem_cache *cachep;
2400 if (cache_index >= NR_GRPINFO_CACHES)
2403 if (unlikely(cache_index < 0))
2406 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2407 if (ext4_groupinfo_caches[cache_index]) {
2408 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2409 return 0; /* Already created */
2412 slab_size = offsetof(struct ext4_group_info,
2413 bb_counters[blocksize_bits + 2]);
2415 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2416 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2419 ext4_groupinfo_caches[cache_index] = cachep;
2421 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2424 "EXT4-fs: no memory for groupinfo slab cache\n");
2431 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2433 struct ext4_sb_info *sbi = EXT4_SB(sb);
2439 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2441 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2442 if (sbi->s_mb_offsets == NULL) {
2447 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2448 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2449 if (sbi->s_mb_maxs == NULL) {
2454 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2458 /* order 0 is regular bitmap */
2459 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2460 sbi->s_mb_offsets[0] = 0;
2464 max = sb->s_blocksize << 2;
2466 sbi->s_mb_offsets[i] = offset;
2467 sbi->s_mb_maxs[i] = max;
2468 offset += 1 << (sb->s_blocksize_bits - i);
2471 } while (i <= sb->s_blocksize_bits + 1);
2473 spin_lock_init(&sbi->s_md_lock);
2474 spin_lock_init(&sbi->s_bal_lock);
2476 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2477 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2478 sbi->s_mb_stats = MB_DEFAULT_STATS;
2479 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2480 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2482 * The default group preallocation is 512, which for 4k block
2483 * sizes translates to 2 megabytes. However for bigalloc file
2484 * systems, this is probably too big (i.e, if the cluster size
2485 * is 1 megabyte, then group preallocation size becomes half a
2486 * gigabyte!). As a default, we will keep a two megabyte
2487 * group pralloc size for cluster sizes up to 64k, and after
2488 * that, we will force a minimum group preallocation size of
2489 * 32 clusters. This translates to 8 megs when the cluster
2490 * size is 256k, and 32 megs when the cluster size is 1 meg,
2491 * which seems reasonable as a default.
2493 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2494 sbi->s_cluster_bits, 32);
2496 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2497 * to the lowest multiple of s_stripe which is bigger than
2498 * the s_mb_group_prealloc as determined above. We want
2499 * the preallocation size to be an exact multiple of the
2500 * RAID stripe size so that preallocations don't fragment
2503 if (sbi->s_stripe > 1) {
2504 sbi->s_mb_group_prealloc = roundup(
2505 sbi->s_mb_group_prealloc, sbi->s_stripe);
2508 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2509 if (sbi->s_locality_groups == NULL) {
2511 goto out_free_groupinfo_slab;
2513 for_each_possible_cpu(i) {
2514 struct ext4_locality_group *lg;
2515 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2516 mutex_init(&lg->lg_mutex);
2517 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2518 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2519 spin_lock_init(&lg->lg_prealloc_lock);
2522 /* init file for buddy data */
2523 ret = ext4_mb_init_backend(sb);
2525 goto out_free_locality_groups;
2528 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2529 &ext4_mb_seq_groups_fops, sb);
2532 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2536 out_free_locality_groups:
2537 free_percpu(sbi->s_locality_groups);
2538 sbi->s_locality_groups = NULL;
2539 out_free_groupinfo_slab:
2540 ext4_groupinfo_destroy_slabs();
2542 kfree(sbi->s_mb_offsets);
2543 sbi->s_mb_offsets = NULL;
2544 kfree(sbi->s_mb_maxs);
2545 sbi->s_mb_maxs = NULL;
2549 /* need to called with the ext4 group lock held */
2550 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2552 struct ext4_prealloc_space *pa;
2553 struct list_head *cur, *tmp;
2556 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2557 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2558 list_del(&pa->pa_group_list);
2560 kmem_cache_free(ext4_pspace_cachep, pa);
2563 mb_debug(1, "mballoc: %u PAs left\n", count);
2567 int ext4_mb_release(struct super_block *sb)
2569 ext4_group_t ngroups = ext4_get_groups_count(sb);
2571 int num_meta_group_infos;
2572 struct ext4_group_info *grinfo;
2573 struct ext4_sb_info *sbi = EXT4_SB(sb);
2574 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2577 remove_proc_entry("mb_groups", sbi->s_proc);
2579 if (sbi->s_group_info) {
2580 for (i = 0; i < ngroups; i++) {
2581 grinfo = ext4_get_group_info(sb, i);
2583 kfree(grinfo->bb_bitmap);
2585 ext4_lock_group(sb, i);
2586 ext4_mb_cleanup_pa(grinfo);
2587 ext4_unlock_group(sb, i);
2588 kmem_cache_free(cachep, grinfo);
2590 num_meta_group_infos = (ngroups +
2591 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2592 EXT4_DESC_PER_BLOCK_BITS(sb);
2593 for (i = 0; i < num_meta_group_infos; i++)
2594 kfree(sbi->s_group_info[i]);
2595 ext4_kvfree(sbi->s_group_info);
2597 kfree(sbi->s_mb_offsets);
2598 kfree(sbi->s_mb_maxs);
2599 if (sbi->s_buddy_cache)
2600 iput(sbi->s_buddy_cache);
2601 if (sbi->s_mb_stats) {
2602 ext4_msg(sb, KERN_INFO,
2603 "mballoc: %u blocks %u reqs (%u success)",
2604 atomic_read(&sbi->s_bal_allocated),
2605 atomic_read(&sbi->s_bal_reqs),
2606 atomic_read(&sbi->s_bal_success));
2607 ext4_msg(sb, KERN_INFO,
2608 "mballoc: %u extents scanned, %u goal hits, "
2609 "%u 2^N hits, %u breaks, %u lost",
2610 atomic_read(&sbi->s_bal_ex_scanned),
2611 atomic_read(&sbi->s_bal_goals),
2612 atomic_read(&sbi->s_bal_2orders),
2613 atomic_read(&sbi->s_bal_breaks),
2614 atomic_read(&sbi->s_mb_lost_chunks));
2615 ext4_msg(sb, KERN_INFO,
2616 "mballoc: %lu generated and it took %Lu",
2617 sbi->s_mb_buddies_generated,
2618 sbi->s_mb_generation_time);
2619 ext4_msg(sb, KERN_INFO,
2620 "mballoc: %u preallocated, %u discarded",
2621 atomic_read(&sbi->s_mb_preallocated),
2622 atomic_read(&sbi->s_mb_discarded));
2625 free_percpu(sbi->s_locality_groups);
2630 static inline int ext4_issue_discard(struct super_block *sb,
2631 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2633 ext4_fsblk_t discard_block;
2635 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2636 ext4_group_first_block_no(sb, block_group));
2637 count = EXT4_C2B(EXT4_SB(sb), count);
2638 trace_ext4_discard_blocks(sb,
2639 (unsigned long long) discard_block, count);
2640 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2644 * This function is called by the jbd2 layer once the commit has finished,
2645 * so we know we can free the blocks that were released with that commit.
2647 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2649 struct super_block *sb = journal->j_private;
2650 struct ext4_buddy e4b;
2651 struct ext4_group_info *db;
2652 int err, count = 0, count2 = 0;
2653 struct ext4_free_data *entry;
2654 struct list_head *l, *ltmp;
2656 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2657 entry = list_entry(l, struct ext4_free_data, list);
2659 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2660 entry->count, entry->group, entry);
2662 if (test_opt(sb, DISCARD))
2663 ext4_issue_discard(sb, entry->group,
2664 entry->start_cluster, entry->count);
2666 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2667 /* we expect to find existing buddy because it's pinned */
2671 /* there are blocks to put in buddy to make them really free */
2672 count += entry->count;
2674 ext4_lock_group(sb, entry->group);
2675 /* Take it out of per group rb tree */
2676 rb_erase(&entry->node, &(db->bb_free_root));
2677 mb_free_blocks(NULL, &e4b, entry->start_cluster, entry->count);
2680 * Clear the trimmed flag for the group so that the next
2681 * ext4_trim_fs can trim it.
2682 * If the volume is mounted with -o discard, online discard
2683 * is supported and the free blocks will be trimmed online.
2685 if (!test_opt(sb, DISCARD))
2686 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2688 if (!db->bb_free_root.rb_node) {
2689 /* No more items in the per group rb tree
2690 * balance refcounts from ext4_mb_free_metadata()
2692 page_cache_release(e4b.bd_buddy_page);
2693 page_cache_release(e4b.bd_bitmap_page);
2695 ext4_unlock_group(sb, entry->group);
2696 kmem_cache_free(ext4_free_ext_cachep, entry);
2697 ext4_mb_unload_buddy(&e4b);
2700 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2703 #ifdef CONFIG_EXT4_DEBUG
2704 u8 mb_enable_debug __read_mostly;
2706 static struct dentry *debugfs_dir;
2707 static struct dentry *debugfs_debug;
2709 static void __init ext4_create_debugfs_entry(void)
2711 debugfs_dir = debugfs_create_dir("ext4", NULL);
2713 debugfs_debug = debugfs_create_u8("mballoc-debug",
2719 static void ext4_remove_debugfs_entry(void)
2721 debugfs_remove(debugfs_debug);
2722 debugfs_remove(debugfs_dir);
2727 static void __init ext4_create_debugfs_entry(void)
2731 static void ext4_remove_debugfs_entry(void)
2737 int __init ext4_init_mballoc(void)
2739 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2740 SLAB_RECLAIM_ACCOUNT);
2741 if (ext4_pspace_cachep == NULL)
2744 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2745 SLAB_RECLAIM_ACCOUNT);
2746 if (ext4_ac_cachep == NULL) {
2747 kmem_cache_destroy(ext4_pspace_cachep);
2751 ext4_free_ext_cachep = KMEM_CACHE(ext4_free_data,
2752 SLAB_RECLAIM_ACCOUNT);
2753 if (ext4_free_ext_cachep == NULL) {
2754 kmem_cache_destroy(ext4_pspace_cachep);
2755 kmem_cache_destroy(ext4_ac_cachep);
2758 ext4_create_debugfs_entry();
2762 void ext4_exit_mballoc(void)
2765 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2766 * before destroying the slab cache.
2769 kmem_cache_destroy(ext4_pspace_cachep);
2770 kmem_cache_destroy(ext4_ac_cachep);
2771 kmem_cache_destroy(ext4_free_ext_cachep);
2772 ext4_groupinfo_destroy_slabs();
2773 ext4_remove_debugfs_entry();
2778 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2779 * Returns 0 if success or error code
2781 static noinline_for_stack int
2782 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2783 handle_t *handle, unsigned int reserv_clstrs)
2785 struct buffer_head *bitmap_bh = NULL;
2786 struct ext4_group_desc *gdp;
2787 struct buffer_head *gdp_bh;
2788 struct ext4_sb_info *sbi;
2789 struct super_block *sb;
2793 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2794 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2800 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2804 err = ext4_journal_get_write_access(handle, bitmap_bh);
2809 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2813 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2814 ext4_free_group_clusters(sb, gdp));
2816 err = ext4_journal_get_write_access(handle, gdp_bh);
2820 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2822 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2823 if (!ext4_data_block_valid(sbi, block, len)) {
2824 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2825 "fs metadata\n", block, block+len);
2826 /* File system mounted not to panic on error
2827 * Fix the bitmap and return EIO
2828 * We leak some of the blocks here.
2830 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2831 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2832 ac->ac_b_ex.fe_len);
2833 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2834 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2840 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2841 #ifdef AGGRESSIVE_CHECK
2844 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2845 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2846 bitmap_bh->b_data));
2850 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2851 ac->ac_b_ex.fe_len);
2852 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2853 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2854 ext4_free_group_clusters_set(sb, gdp,
2855 ext4_free_clusters_after_init(sb,
2856 ac->ac_b_ex.fe_group, gdp));
2858 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2859 ext4_free_group_clusters_set(sb, gdp, len);
2860 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2862 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2863 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2865 * Now reduce the dirty block count also. Should not go negative
2867 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2868 /* release all the reserved blocks if non delalloc */
2869 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2872 if (sbi->s_log_groups_per_flex) {
2873 ext4_group_t flex_group = ext4_flex_group(sbi,
2874 ac->ac_b_ex.fe_group);
2875 atomic64_sub(ac->ac_b_ex.fe_len,
2876 &sbi->s_flex_groups[flex_group].free_clusters);
2879 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2882 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2885 ext4_mark_super_dirty(sb);
2891 * here we normalize request for locality group
2892 * Group request are normalized to s_mb_group_prealloc, which goes to
2893 * s_strip if we set the same via mount option.
2894 * s_mb_group_prealloc can be configured via
2895 * /sys/fs/ext4/<partition>/mb_group_prealloc
2897 * XXX: should we try to preallocate more than the group has now?
2899 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2901 struct super_block *sb = ac->ac_sb;
2902 struct ext4_locality_group *lg = ac->ac_lg;
2905 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2906 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2907 current->pid, ac->ac_g_ex.fe_len);
2911 * Normalization means making request better in terms of
2912 * size and alignment
2914 static noinline_for_stack void
2915 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2916 struct ext4_allocation_request *ar)
2918 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2921 loff_t size, orig_size, start_off;
2923 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2924 struct ext4_prealloc_space *pa;
2926 /* do normalize only data requests, metadata requests
2927 do not need preallocation */
2928 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2931 /* sometime caller may want exact blocks */
2932 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2935 /* caller may indicate that preallocation isn't
2936 * required (it's a tail, for example) */
2937 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2940 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2941 ext4_mb_normalize_group_request(ac);
2945 bsbits = ac->ac_sb->s_blocksize_bits;
2947 /* first, let's learn actual file size
2948 * given current request is allocated */
2949 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
2950 size = size << bsbits;
2951 if (size < i_size_read(ac->ac_inode))
2952 size = i_size_read(ac->ac_inode);
2955 /* max size of free chunks */
2958 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2959 (req <= (size) || max <= (chunk_size))
2961 /* first, try to predict filesize */
2962 /* XXX: should this table be tunable? */
2964 if (size <= 16 * 1024) {
2966 } else if (size <= 32 * 1024) {
2968 } else if (size <= 64 * 1024) {
2970 } else if (size <= 128 * 1024) {
2972 } else if (size <= 256 * 1024) {
2974 } else if (size <= 512 * 1024) {
2976 } else if (size <= 1024 * 1024) {
2978 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2979 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2980 (21 - bsbits)) << 21;
2981 size = 2 * 1024 * 1024;
2982 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2983 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2984 (22 - bsbits)) << 22;
2985 size = 4 * 1024 * 1024;
2986 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2987 (8<<20)>>bsbits, max, 8 * 1024)) {
2988 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2989 (23 - bsbits)) << 23;
2990 size = 8 * 1024 * 1024;
2992 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2993 size = ac->ac_o_ex.fe_len << bsbits;
2995 size = size >> bsbits;
2996 start = start_off >> bsbits;
2998 /* don't cover already allocated blocks in selected range */
2999 if (ar->pleft && start <= ar->lleft) {
3000 size -= ar->lleft + 1 - start;
3001 start = ar->lleft + 1;
3003 if (ar->pright && start + size - 1 >= ar->lright)
3004 size -= start + size - ar->lright;
3007 * Trim allocation request for filesystems with artificially small
3010 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
3011 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
3015 /* check we don't cross already preallocated blocks */
3017 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3022 spin_lock(&pa->pa_lock);
3023 if (pa->pa_deleted) {
3024 spin_unlock(&pa->pa_lock);
3028 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3031 /* PA must not overlap original request */
3032 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3033 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3035 /* skip PAs this normalized request doesn't overlap with */
3036 if (pa->pa_lstart >= end || pa_end <= start) {
3037 spin_unlock(&pa->pa_lock);
3040 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3042 /* adjust start or end to be adjacent to this pa */
3043 if (pa_end <= ac->ac_o_ex.fe_logical) {
3044 BUG_ON(pa_end < start);
3046 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3047 BUG_ON(pa->pa_lstart > end);
3048 end = pa->pa_lstart;
3050 spin_unlock(&pa->pa_lock);
3055 /* XXX: extra loop to check we really don't overlap preallocations */
3057 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3060 spin_lock(&pa->pa_lock);
3061 if (pa->pa_deleted == 0) {
3062 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3064 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3066 spin_unlock(&pa->pa_lock);
3070 if (start + size <= ac->ac_o_ex.fe_logical &&
3071 start > ac->ac_o_ex.fe_logical) {
3072 ext4_msg(ac->ac_sb, KERN_ERR,
3073 "start %lu, size %lu, fe_logical %lu",
3074 (unsigned long) start, (unsigned long) size,
3075 (unsigned long) ac->ac_o_ex.fe_logical);
3077 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3078 start > ac->ac_o_ex.fe_logical);
3079 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3081 /* now prepare goal request */
3083 /* XXX: is it better to align blocks WRT to logical
3084 * placement or satisfy big request as is */
3085 ac->ac_g_ex.fe_logical = start;
3086 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3088 /* define goal start in order to merge */
3089 if (ar->pright && (ar->lright == (start + size))) {
3090 /* merge to the right */
3091 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3092 &ac->ac_f_ex.fe_group,
3093 &ac->ac_f_ex.fe_start);
3094 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3096 if (ar->pleft && (ar->lleft + 1 == start)) {
3097 /* merge to the left */
3098 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3099 &ac->ac_f_ex.fe_group,
3100 &ac->ac_f_ex.fe_start);
3101 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3104 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3105 (unsigned) orig_size, (unsigned) start);
3108 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3110 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3112 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3113 atomic_inc(&sbi->s_bal_reqs);
3114 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3115 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3116 atomic_inc(&sbi->s_bal_success);
3117 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3118 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3119 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3120 atomic_inc(&sbi->s_bal_goals);
3121 if (ac->ac_found > sbi->s_mb_max_to_scan)
3122 atomic_inc(&sbi->s_bal_breaks);
3125 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3126 trace_ext4_mballoc_alloc(ac);
3128 trace_ext4_mballoc_prealloc(ac);
3132 * Called on failure; free up any blocks from the inode PA for this
3133 * context. We don't need this for MB_GROUP_PA because we only change
3134 * pa_free in ext4_mb_release_context(), but on failure, we've already
3135 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3137 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3139 struct ext4_prealloc_space *pa = ac->ac_pa;
3140 struct ext4_buddy e4b;
3144 if (ac->ac_f_ex.fe_len == 0)
3146 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3149 * This should never happen since we pin the
3150 * pages in the ext4_allocation_context so
3151 * ext4_mb_load_buddy() should never fail.
3153 WARN(1, "mb_load_buddy failed (%d)", err);
3156 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3157 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3158 ac->ac_f_ex.fe_len);
3159 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3160 ext4_mb_unload_buddy(&e4b);
3163 if (pa->pa_type == MB_INODE_PA)
3164 pa->pa_free += ac->ac_b_ex.fe_len;
3168 * use blocks preallocated to inode
3170 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3171 struct ext4_prealloc_space *pa)
3173 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3178 /* found preallocated blocks, use them */
3179 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3180 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3181 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3182 len = EXT4_NUM_B2C(sbi, end - start);
3183 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3184 &ac->ac_b_ex.fe_start);
3185 ac->ac_b_ex.fe_len = len;
3186 ac->ac_status = AC_STATUS_FOUND;
3189 BUG_ON(start < pa->pa_pstart);
3190 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3191 BUG_ON(pa->pa_free < len);
3194 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3198 * use blocks preallocated to locality group
3200 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3201 struct ext4_prealloc_space *pa)
3203 unsigned int len = ac->ac_o_ex.fe_len;
3205 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3206 &ac->ac_b_ex.fe_group,
3207 &ac->ac_b_ex.fe_start);
3208 ac->ac_b_ex.fe_len = len;
3209 ac->ac_status = AC_STATUS_FOUND;
3212 /* we don't correct pa_pstart or pa_plen here to avoid
3213 * possible race when the group is being loaded concurrently
3214 * instead we correct pa later, after blocks are marked
3215 * in on-disk bitmap -- see ext4_mb_release_context()
3216 * Other CPUs are prevented from allocating from this pa by lg_mutex
3218 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3222 * Return the prealloc space that have minimal distance
3223 * from the goal block. @cpa is the prealloc
3224 * space that is having currently known minimal distance
3225 * from the goal block.
3227 static struct ext4_prealloc_space *
3228 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3229 struct ext4_prealloc_space *pa,
3230 struct ext4_prealloc_space *cpa)
3232 ext4_fsblk_t cur_distance, new_distance;
3235 atomic_inc(&pa->pa_count);
3238 cur_distance = abs(goal_block - cpa->pa_pstart);
3239 new_distance = abs(goal_block - pa->pa_pstart);
3241 if (cur_distance <= new_distance)
3244 /* drop the previous reference */
3245 atomic_dec(&cpa->pa_count);
3246 atomic_inc(&pa->pa_count);
3251 * search goal blocks in preallocated space
3253 static noinline_for_stack int
3254 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3256 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3258 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3259 struct ext4_locality_group *lg;
3260 struct ext4_prealloc_space *pa, *cpa = NULL;
3261 ext4_fsblk_t goal_block;
3263 /* only data can be preallocated */
3264 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3267 /* first, try per-file preallocation */
3269 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3271 /* all fields in this condition don't change,
3272 * so we can skip locking for them */
3273 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3274 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3275 EXT4_C2B(sbi, pa->pa_len)))
3278 /* non-extent files can't have physical blocks past 2^32 */
3279 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3280 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3281 EXT4_MAX_BLOCK_FILE_PHYS))
3284 /* found preallocated blocks, use them */
3285 spin_lock(&pa->pa_lock);
3286 if (pa->pa_deleted == 0 && pa->pa_free) {
3287 atomic_inc(&pa->pa_count);
3288 ext4_mb_use_inode_pa(ac, pa);
3289 spin_unlock(&pa->pa_lock);
3290 ac->ac_criteria = 10;
3294 spin_unlock(&pa->pa_lock);
3298 /* can we use group allocation? */
3299 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3302 /* inode may have no locality group for some reason */
3306 order = fls(ac->ac_o_ex.fe_len) - 1;
3307 if (order > PREALLOC_TB_SIZE - 1)
3308 /* The max size of hash table is PREALLOC_TB_SIZE */
3309 order = PREALLOC_TB_SIZE - 1;
3311 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3313 * search for the prealloc space that is having
3314 * minimal distance from the goal block.
3316 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3318 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3320 spin_lock(&pa->pa_lock);
3321 if (pa->pa_deleted == 0 &&
3322 pa->pa_free >= ac->ac_o_ex.fe_len) {
3324 cpa = ext4_mb_check_group_pa(goal_block,
3327 spin_unlock(&pa->pa_lock);
3332 ext4_mb_use_group_pa(ac, cpa);
3333 ac->ac_criteria = 20;
3340 * the function goes through all block freed in the group
3341 * but not yet committed and marks them used in in-core bitmap.
3342 * buddy must be generated from this bitmap
3343 * Need to be called with the ext4 group lock held
3345 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3349 struct ext4_group_info *grp;
3350 struct ext4_free_data *entry;
3352 grp = ext4_get_group_info(sb, group);
3353 n = rb_first(&(grp->bb_free_root));
3356 entry = rb_entry(n, struct ext4_free_data, node);
3357 ext4_set_bits(bitmap, entry->start_cluster, entry->count);
3364 * the function goes through all preallocation in this group and marks them
3365 * used in in-core bitmap. buddy must be generated from this bitmap
3366 * Need to be called with ext4 group lock held
3368 static noinline_for_stack
3369 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3372 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3373 struct ext4_prealloc_space *pa;
3374 struct list_head *cur;
3375 ext4_group_t groupnr;
3376 ext4_grpblk_t start;
3377 int preallocated = 0;
3380 /* all form of preallocation discards first load group,
3381 * so the only competing code is preallocation use.
3382 * we don't need any locking here
3383 * notice we do NOT ignore preallocations with pa_deleted
3384 * otherwise we could leave used blocks available for
3385 * allocation in buddy when concurrent ext4_mb_put_pa()
3386 * is dropping preallocation
3388 list_for_each(cur, &grp->bb_prealloc_list) {
3389 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3390 spin_lock(&pa->pa_lock);
3391 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3394 spin_unlock(&pa->pa_lock);
3395 if (unlikely(len == 0))
3397 BUG_ON(groupnr != group);
3398 ext4_set_bits(bitmap, start, len);
3399 preallocated += len;
3401 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3404 static void ext4_mb_pa_callback(struct rcu_head *head)
3406 struct ext4_prealloc_space *pa;
3407 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3409 BUG_ON(atomic_read(&pa->pa_count));
3410 BUG_ON(pa->pa_deleted == 0);
3411 kmem_cache_free(ext4_pspace_cachep, pa);
3415 * drops a reference to preallocated space descriptor
3416 * if this was the last reference and the space is consumed
3418 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3419 struct super_block *sb, struct ext4_prealloc_space *pa)
3422 ext4_fsblk_t grp_blk;
3424 /* in this short window concurrent discard can set pa_deleted */
3425 spin_lock(&pa->pa_lock);
3426 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3427 spin_unlock(&pa->pa_lock);
3431 if (pa->pa_deleted == 1) {
3432 spin_unlock(&pa->pa_lock);
3437 spin_unlock(&pa->pa_lock);
3439 grp_blk = pa->pa_pstart;
3441 * If doing group-based preallocation, pa_pstart may be in the
3442 * next group when pa is used up
3444 if (pa->pa_type == MB_GROUP_PA)
3447 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3452 * P1 (buddy init) P2 (regular allocation)
3453 * find block B in PA
3454 * copy on-disk bitmap to buddy
3455 * mark B in on-disk bitmap
3456 * drop PA from group
3457 * mark all PAs in buddy
3459 * thus, P1 initializes buddy with B available. to prevent this
3460 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3463 ext4_lock_group(sb, grp);
3464 list_del(&pa->pa_group_list);
3465 ext4_unlock_group(sb, grp);
3467 spin_lock(pa->pa_obj_lock);
3468 list_del_rcu(&pa->pa_inode_list);
3469 spin_unlock(pa->pa_obj_lock);
3471 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3475 * creates new preallocated space for given inode
3477 static noinline_for_stack int
3478 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3480 struct super_block *sb = ac->ac_sb;
3481 struct ext4_sb_info *sbi = EXT4_SB(sb);
3482 struct ext4_prealloc_space *pa;
3483 struct ext4_group_info *grp;
3484 struct ext4_inode_info *ei;
3486 /* preallocate only when found space is larger then requested */
3487 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3488 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3489 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3491 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3495 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3501 /* we can't allocate as much as normalizer wants.
3502 * so, found space must get proper lstart
3503 * to cover original request */
3504 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3505 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3507 /* we're limited by original request in that
3508 * logical block must be covered any way
3509 * winl is window we can move our chunk within */
3510 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3512 /* also, we should cover whole original request */
3513 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3515 /* the smallest one defines real window */
3516 win = min(winl, wins);
3518 offs = ac->ac_o_ex.fe_logical %
3519 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3520 if (offs && offs < win)
3523 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3524 EXT4_NUM_B2C(sbi, win);
3525 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3526 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3529 /* preallocation can change ac_b_ex, thus we store actually
3530 * allocated blocks for history */
3531 ac->ac_f_ex = ac->ac_b_ex;
3533 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3534 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3535 pa->pa_len = ac->ac_b_ex.fe_len;
3536 pa->pa_free = pa->pa_len;
3537 atomic_set(&pa->pa_count, 1);
3538 spin_lock_init(&pa->pa_lock);
3539 INIT_LIST_HEAD(&pa->pa_inode_list);
3540 INIT_LIST_HEAD(&pa->pa_group_list);
3542 pa->pa_type = MB_INODE_PA;
3544 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3545 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3546 trace_ext4_mb_new_inode_pa(ac, pa);
3548 ext4_mb_use_inode_pa(ac, pa);
3549 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3551 ei = EXT4_I(ac->ac_inode);
3552 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3554 pa->pa_obj_lock = &ei->i_prealloc_lock;
3555 pa->pa_inode = ac->ac_inode;
3557 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3558 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3559 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3561 spin_lock(pa->pa_obj_lock);
3562 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3563 spin_unlock(pa->pa_obj_lock);
3569 * creates new preallocated space for locality group inodes belongs to
3571 static noinline_for_stack int
3572 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3574 struct super_block *sb = ac->ac_sb;
3575 struct ext4_locality_group *lg;
3576 struct ext4_prealloc_space *pa;
3577 struct ext4_group_info *grp;
3579 /* preallocate only when found space is larger then requested */
3580 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3581 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3582 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3584 BUG_ON(ext4_pspace_cachep == NULL);
3585 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3589 /* preallocation can change ac_b_ex, thus we store actually
3590 * allocated blocks for history */
3591 ac->ac_f_ex = ac->ac_b_ex;
3593 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3594 pa->pa_lstart = pa->pa_pstart;
3595 pa->pa_len = ac->ac_b_ex.fe_len;
3596 pa->pa_free = pa->pa_len;
3597 atomic_set(&pa->pa_count, 1);
3598 spin_lock_init(&pa->pa_lock);
3599 INIT_LIST_HEAD(&pa->pa_inode_list);
3600 INIT_LIST_HEAD(&pa->pa_group_list);
3602 pa->pa_type = MB_GROUP_PA;
3604 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3605 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3606 trace_ext4_mb_new_group_pa(ac, pa);
3608 ext4_mb_use_group_pa(ac, pa);
3609 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3611 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3615 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3616 pa->pa_inode = NULL;
3618 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3619 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3620 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3623 * We will later add the new pa to the right bucket
3624 * after updating the pa_free in ext4_mb_release_context
3629 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3633 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3634 err = ext4_mb_new_group_pa(ac);
3636 err = ext4_mb_new_inode_pa(ac);
3641 * finds all unused blocks in on-disk bitmap, frees them in
3642 * in-core bitmap and buddy.
3643 * @pa must be unlinked from inode and group lists, so that
3644 * nobody else can find/use it.
3645 * the caller MUST hold group/inode locks.
3646 * TODO: optimize the case when there are no in-core structures yet
3648 static noinline_for_stack int
3649 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3650 struct ext4_prealloc_space *pa)
3652 struct super_block *sb = e4b->bd_sb;
3653 struct ext4_sb_info *sbi = EXT4_SB(sb);
3658 unsigned long long grp_blk_start;
3662 BUG_ON(pa->pa_deleted == 0);
3663 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3664 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3665 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3666 end = bit + pa->pa_len;
3669 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3672 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3673 mb_debug(1, " free preallocated %u/%u in group %u\n",
3674 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3675 (unsigned) next - bit, (unsigned) group);
3678 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3679 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3680 EXT4_C2B(sbi, bit)),
3682 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3685 if (free != pa->pa_free) {
3686 ext4_msg(e4b->bd_sb, KERN_CRIT,
3687 "pa %p: logic %lu, phys. %lu, len %lu",
3688 pa, (unsigned long) pa->pa_lstart,
3689 (unsigned long) pa->pa_pstart,
3690 (unsigned long) pa->pa_len);
3691 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3694 * pa is already deleted so we use the value obtained
3695 * from the bitmap and continue.
3698 atomic_add(free, &sbi->s_mb_discarded);
3703 static noinline_for_stack int
3704 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3705 struct ext4_prealloc_space *pa)
3707 struct super_block *sb = e4b->bd_sb;
3711 trace_ext4_mb_release_group_pa(pa);
3712 BUG_ON(pa->pa_deleted == 0);
3713 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3714 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3715 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3716 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3717 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3723 * releases all preallocations in given group
3725 * first, we need to decide discard policy:
3726 * - when do we discard
3728 * - how many do we discard
3729 * 1) how many requested
3731 static noinline_for_stack int
3732 ext4_mb_discard_group_preallocations(struct super_block *sb,
3733 ext4_group_t group, int needed)
3735 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3736 struct buffer_head *bitmap_bh = NULL;
3737 struct ext4_prealloc_space *pa, *tmp;
3738 struct list_head list;
3739 struct ext4_buddy e4b;
3744 mb_debug(1, "discard preallocation for group %u\n", group);
3746 if (list_empty(&grp->bb_prealloc_list))
3749 bitmap_bh = ext4_read_block_bitmap(sb, group);
3750 if (bitmap_bh == NULL) {
3751 ext4_error(sb, "Error reading block bitmap for %u", group);
3755 err = ext4_mb_load_buddy(sb, group, &e4b);
3757 ext4_error(sb, "Error loading buddy information for %u", group);
3763 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3765 INIT_LIST_HEAD(&list);
3767 ext4_lock_group(sb, group);
3768 list_for_each_entry_safe(pa, tmp,
3769 &grp->bb_prealloc_list, pa_group_list) {
3770 spin_lock(&pa->pa_lock);
3771 if (atomic_read(&pa->pa_count)) {
3772 spin_unlock(&pa->pa_lock);
3776 if (pa->pa_deleted) {
3777 spin_unlock(&pa->pa_lock);
3781 /* seems this one can be freed ... */
3784 /* we can trust pa_free ... */
3785 free += pa->pa_free;
3787 spin_unlock(&pa->pa_lock);
3789 list_del(&pa->pa_group_list);
3790 list_add(&pa->u.pa_tmp_list, &list);
3793 /* if we still need more blocks and some PAs were used, try again */
3794 if (free < needed && busy) {
3796 ext4_unlock_group(sb, group);
3798 * Yield the CPU here so that we don't get soft lockup
3799 * in non preempt case.
3805 /* found anything to free? */
3806 if (list_empty(&list)) {
3811 /* now free all selected PAs */
3812 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3814 /* remove from object (inode or locality group) */
3815 spin_lock(pa->pa_obj_lock);
3816 list_del_rcu(&pa->pa_inode_list);
3817 spin_unlock(pa->pa_obj_lock);
3819 if (pa->pa_type == MB_GROUP_PA)
3820 ext4_mb_release_group_pa(&e4b, pa);
3822 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3824 list_del(&pa->u.pa_tmp_list);
3825 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3829 ext4_unlock_group(sb, group);
3830 ext4_mb_unload_buddy(&e4b);
3836 * releases all non-used preallocated blocks for given inode
3838 * It's important to discard preallocations under i_data_sem
3839 * We don't want another block to be served from the prealloc
3840 * space when we are discarding the inode prealloc space.
3842 * FIXME!! Make sure it is valid at all the call sites
3844 void ext4_discard_preallocations(struct inode *inode)
3846 struct ext4_inode_info *ei = EXT4_I(inode);
3847 struct super_block *sb = inode->i_sb;
3848 struct buffer_head *bitmap_bh = NULL;
3849 struct ext4_prealloc_space *pa, *tmp;
3850 ext4_group_t group = 0;
3851 struct list_head list;
3852 struct ext4_buddy e4b;
3855 if (!S_ISREG(inode->i_mode)) {
3856 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3860 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3861 trace_ext4_discard_preallocations(inode);
3863 INIT_LIST_HEAD(&list);
3866 /* first, collect all pa's in the inode */
3867 spin_lock(&ei->i_prealloc_lock);
3868 while (!list_empty(&ei->i_prealloc_list)) {
3869 pa = list_entry(ei->i_prealloc_list.next,
3870 struct ext4_prealloc_space, pa_inode_list);
3871 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3872 spin_lock(&pa->pa_lock);
3873 if (atomic_read(&pa->pa_count)) {
3874 /* this shouldn't happen often - nobody should
3875 * use preallocation while we're discarding it */
3876 spin_unlock(&pa->pa_lock);
3877 spin_unlock(&ei->i_prealloc_lock);
3878 ext4_msg(sb, KERN_ERR,
3879 "uh-oh! used pa while discarding");
3881 schedule_timeout_uninterruptible(HZ);
3885 if (pa->pa_deleted == 0) {
3887 spin_unlock(&pa->pa_lock);
3888 list_del_rcu(&pa->pa_inode_list);
3889 list_add(&pa->u.pa_tmp_list, &list);
3893 /* someone is deleting pa right now */
3894 spin_unlock(&pa->pa_lock);
3895 spin_unlock(&ei->i_prealloc_lock);
3897 /* we have to wait here because pa_deleted
3898 * doesn't mean pa is already unlinked from
3899 * the list. as we might be called from
3900 * ->clear_inode() the inode will get freed
3901 * and concurrent thread which is unlinking
3902 * pa from inode's list may access already
3903 * freed memory, bad-bad-bad */
3905 /* XXX: if this happens too often, we can
3906 * add a flag to force wait only in case
3907 * of ->clear_inode(), but not in case of
3908 * regular truncate */
3909 schedule_timeout_uninterruptible(HZ);
3912 spin_unlock(&ei->i_prealloc_lock);
3914 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3915 BUG_ON(pa->pa_type != MB_INODE_PA);
3916 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3918 err = ext4_mb_load_buddy(sb, group, &e4b);
3920 ext4_error(sb, "Error loading buddy information for %u",
3925 bitmap_bh = ext4_read_block_bitmap(sb, group);
3926 if (bitmap_bh == NULL) {
3927 ext4_error(sb, "Error reading block bitmap for %u",
3929 ext4_mb_unload_buddy(&e4b);
3933 ext4_lock_group(sb, group);
3934 list_del(&pa->pa_group_list);
3935 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3936 ext4_unlock_group(sb, group);
3938 ext4_mb_unload_buddy(&e4b);
3941 list_del(&pa->u.pa_tmp_list);
3942 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3946 #ifdef CONFIG_EXT4_DEBUG
3947 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3949 struct super_block *sb = ac->ac_sb;
3950 ext4_group_t ngroups, i;
3952 if (!mb_enable_debug ||
3953 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3956 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: Can't allocate:"
3957 " Allocation context details:");
3958 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: status %d flags %d",
3959 ac->ac_status, ac->ac_flags);
3960 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: orig %lu/%lu/%lu@%lu, "
3961 "goal %lu/%lu/%lu@%lu, "
3962 "best %lu/%lu/%lu@%lu cr %d",
3963 (unsigned long)ac->ac_o_ex.fe_group,
3964 (unsigned long)ac->ac_o_ex.fe_start,
3965 (unsigned long)ac->ac_o_ex.fe_len,
3966 (unsigned long)ac->ac_o_ex.fe_logical,
3967 (unsigned long)ac->ac_g_ex.fe_group,
3968 (unsigned long)ac->ac_g_ex.fe_start,
3969 (unsigned long)ac->ac_g_ex.fe_len,
3970 (unsigned long)ac->ac_g_ex.fe_logical,
3971 (unsigned long)ac->ac_b_ex.fe_group,
3972 (unsigned long)ac->ac_b_ex.fe_start,
3973 (unsigned long)ac->ac_b_ex.fe_len,
3974 (unsigned long)ac->ac_b_ex.fe_logical,
3975 (int)ac->ac_criteria);
3976 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: %lu scanned, %d found",
3977 ac->ac_ex_scanned, ac->ac_found);
3978 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: groups: ");
3979 ngroups = ext4_get_groups_count(sb);
3980 for (i = 0; i < ngroups; i++) {
3981 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3982 struct ext4_prealloc_space *pa;
3983 ext4_grpblk_t start;
3984 struct list_head *cur;
3985 ext4_lock_group(sb, i);
3986 list_for_each(cur, &grp->bb_prealloc_list) {
3987 pa = list_entry(cur, struct ext4_prealloc_space,
3989 spin_lock(&pa->pa_lock);
3990 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3992 spin_unlock(&pa->pa_lock);
3993 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3996 ext4_unlock_group(sb, i);
3998 if (grp->bb_free == 0)
4000 printk(KERN_ERR "%u: %d/%d \n",
4001 i, grp->bb_free, grp->bb_fragments);
4003 printk(KERN_ERR "\n");
4006 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4013 * We use locality group preallocation for small size file. The size of the
4014 * file is determined by the current size or the resulting size after
4015 * allocation which ever is larger
4017 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4019 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4021 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4022 int bsbits = ac->ac_sb->s_blocksize_bits;
4025 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4028 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4031 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4032 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4035 if ((size == isize) &&
4036 !ext4_fs_is_busy(sbi) &&
4037 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4038 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4042 if (sbi->s_mb_group_prealloc <= 0) {
4043 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4047 /* don't use group allocation for large files */
4048 size = max(size, isize);
4049 if (size > sbi->s_mb_stream_request) {
4050 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4054 BUG_ON(ac->ac_lg != NULL);
4056 * locality group prealloc space are per cpu. The reason for having
4057 * per cpu locality group is to reduce the contention between block
4058 * request from multiple CPUs.
4060 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
4062 /* we're going to use group allocation */
4063 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4065 /* serialize all allocations in the group */
4066 mutex_lock(&ac->ac_lg->lg_mutex);
4069 static noinline_for_stack int
4070 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4071 struct ext4_allocation_request *ar)
4073 struct super_block *sb = ar->inode->i_sb;
4074 struct ext4_sb_info *sbi = EXT4_SB(sb);
4075 struct ext4_super_block *es = sbi->s_es;
4079 ext4_grpblk_t block;
4081 /* we can't allocate > group size */
4084 /* just a dirty hack to filter too big requests */
4085 if (len >= EXT4_CLUSTERS_PER_GROUP(sb) - 10)
4086 len = EXT4_CLUSTERS_PER_GROUP(sb) - 10;
4088 /* start searching from the goal */
4090 if (goal < le32_to_cpu(es->s_first_data_block) ||
4091 goal >= ext4_blocks_count(es))
4092 goal = le32_to_cpu(es->s_first_data_block);
4093 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4095 /* set up allocation goals */
4096 memset(ac, 0, sizeof(struct ext4_allocation_context));
4097 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4098 ac->ac_status = AC_STATUS_CONTINUE;
4100 ac->ac_inode = ar->inode;
4101 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4102 ac->ac_o_ex.fe_group = group;
4103 ac->ac_o_ex.fe_start = block;
4104 ac->ac_o_ex.fe_len = len;
4105 ac->ac_g_ex = ac->ac_o_ex;
4106 ac->ac_flags = ar->flags;
4108 /* we have to define context: we'll we work with a file or
4109 * locality group. this is a policy, actually */
4110 ext4_mb_group_or_file(ac);
4112 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4113 "left: %u/%u, right %u/%u to %swritable\n",
4114 (unsigned) ar->len, (unsigned) ar->logical,
4115 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4116 (unsigned) ar->lleft, (unsigned) ar->pleft,
4117 (unsigned) ar->lright, (unsigned) ar->pright,
4118 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4123 static noinline_for_stack void
4124 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4125 struct ext4_locality_group *lg,
4126 int order, int total_entries)
4128 ext4_group_t group = 0;
4129 struct ext4_buddy e4b;
4130 struct list_head discard_list;
4131 struct ext4_prealloc_space *pa, *tmp;
4133 mb_debug(1, "discard locality group preallocation\n");
4135 INIT_LIST_HEAD(&discard_list);
4137 spin_lock(&lg->lg_prealloc_lock);
4138 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4140 spin_lock(&pa->pa_lock);
4141 if (atomic_read(&pa->pa_count)) {
4143 * This is the pa that we just used
4144 * for block allocation. So don't
4147 spin_unlock(&pa->pa_lock);
4150 if (pa->pa_deleted) {
4151 spin_unlock(&pa->pa_lock);
4154 /* only lg prealloc space */
4155 BUG_ON(pa->pa_type != MB_GROUP_PA);
4157 /* seems this one can be freed ... */
4159 spin_unlock(&pa->pa_lock);
4161 list_del_rcu(&pa->pa_inode_list);
4162 list_add(&pa->u.pa_tmp_list, &discard_list);
4165 if (total_entries <= 5) {
4167 * we want to keep only 5 entries
4168 * allowing it to grow to 8. This
4169 * mak sure we don't call discard
4170 * soon for this list.
4175 spin_unlock(&lg->lg_prealloc_lock);
4177 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4179 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4180 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4181 ext4_error(sb, "Error loading buddy information for %u",
4185 ext4_lock_group(sb, group);
4186 list_del(&pa->pa_group_list);
4187 ext4_mb_release_group_pa(&e4b, pa);
4188 ext4_unlock_group(sb, group);
4190 ext4_mb_unload_buddy(&e4b);
4191 list_del(&pa->u.pa_tmp_list);
4192 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4197 * We have incremented pa_count. So it cannot be freed at this
4198 * point. Also we hold lg_mutex. So no parallel allocation is
4199 * possible from this lg. That means pa_free cannot be updated.
4201 * A parallel ext4_mb_discard_group_preallocations is possible.
4202 * which can cause the lg_prealloc_list to be updated.
4205 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4207 int order, added = 0, lg_prealloc_count = 1;
4208 struct super_block *sb = ac->ac_sb;
4209 struct ext4_locality_group *lg = ac->ac_lg;
4210 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4212 order = fls(pa->pa_free) - 1;
4213 if (order > PREALLOC_TB_SIZE - 1)
4214 /* The max size of hash table is PREALLOC_TB_SIZE */
4215 order = PREALLOC_TB_SIZE - 1;
4216 /* Add the prealloc space to lg */
4217 spin_lock(&lg->lg_prealloc_lock);
4218 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4220 spin_lock(&tmp_pa->pa_lock);
4221 if (tmp_pa->pa_deleted) {
4222 spin_unlock(&tmp_pa->pa_lock);
4225 if (!added && pa->pa_free < tmp_pa->pa_free) {
4226 /* Add to the tail of the previous entry */
4227 list_add_tail_rcu(&pa->pa_inode_list,
4228 &tmp_pa->pa_inode_list);
4231 * we want to count the total
4232 * number of entries in the list
4235 spin_unlock(&tmp_pa->pa_lock);
4236 lg_prealloc_count++;
4239 list_add_tail_rcu(&pa->pa_inode_list,
4240 &lg->lg_prealloc_list[order]);
4241 spin_unlock(&lg->lg_prealloc_lock);
4243 /* Now trim the list to be not more than 8 elements */
4244 if (lg_prealloc_count > 8) {
4245 ext4_mb_discard_lg_preallocations(sb, lg,
4246 order, lg_prealloc_count);
4253 * release all resource we used in allocation
4255 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4257 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4258 struct ext4_prealloc_space *pa = ac->ac_pa;
4260 if (pa->pa_type == MB_GROUP_PA) {
4261 /* see comment in ext4_mb_use_group_pa() */
4262 spin_lock(&pa->pa_lock);
4263 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4264 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4265 pa->pa_free -= ac->ac_b_ex.fe_len;
4266 pa->pa_len -= ac->ac_b_ex.fe_len;
4267 spin_unlock(&pa->pa_lock);
4272 * We want to add the pa to the right bucket.
4273 * Remove it from the list and while adding
4274 * make sure the list to which we are adding
4277 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4278 spin_lock(pa->pa_obj_lock);
4279 list_del_rcu(&pa->pa_inode_list);
4280 spin_unlock(pa->pa_obj_lock);
4281 ext4_mb_add_n_trim(ac);
4283 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4285 if (ac->ac_bitmap_page)
4286 page_cache_release(ac->ac_bitmap_page);
4287 if (ac->ac_buddy_page)
4288 page_cache_release(ac->ac_buddy_page);
4289 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4290 mutex_unlock(&ac->ac_lg->lg_mutex);
4291 ext4_mb_collect_stats(ac);
4295 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4297 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4301 trace_ext4_mb_discard_preallocations(sb, needed);
4302 for (i = 0; i < ngroups && needed > 0; i++) {
4303 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4312 * Main entry point into mballoc to allocate blocks
4313 * it tries to use preallocation first, then falls back
4314 * to usual allocation
4316 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4317 struct ext4_allocation_request *ar, int *errp)
4320 struct ext4_allocation_context *ac = NULL;
4321 struct ext4_sb_info *sbi;
4322 struct super_block *sb;
4323 ext4_fsblk_t block = 0;
4324 unsigned int inquota = 0;
4325 unsigned int reserv_clstrs = 0;
4327 sb = ar->inode->i_sb;
4330 trace_ext4_request_blocks(ar);
4332 /* Allow to use superuser reservation for quota file */
4333 if (IS_NOQUOTA(ar->inode))
4334 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4337 * For delayed allocation, we could skip the ENOSPC and
4338 * EDQUOT check, as blocks and quotas have been already
4339 * reserved when data being copied into pagecache.
4341 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4342 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4344 /* Without delayed allocation we need to verify
4345 * there is enough free blocks to do block allocation
4346 * and verify allocation doesn't exceed the quota limits.
4349 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4351 /* let others to free the space */
4353 ar->len = ar->len >> 1;
4359 reserv_clstrs = ar->len;
4360 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4361 dquot_alloc_block_nofail(ar->inode,
4362 EXT4_C2B(sbi, ar->len));
4365 dquot_alloc_block(ar->inode,
4366 EXT4_C2B(sbi, ar->len))) {
4368 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4379 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4386 *errp = ext4_mb_initialize_context(ac, ar);
4392 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4393 if (!ext4_mb_use_preallocated(ac)) {
4394 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4395 ext4_mb_normalize_request(ac, ar);
4397 /* allocate space in core */
4398 *errp = ext4_mb_regular_allocator(ac);
4402 /* as we've just preallocated more space than
4403 * user requested orinally, we store allocated
4404 * space in a special descriptor */
4405 if (ac->ac_status == AC_STATUS_FOUND &&
4406 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4407 ext4_mb_new_preallocation(ac);
4409 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4410 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4413 ext4_discard_allocated_blocks(ac);
4415 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4416 ar->len = ac->ac_b_ex.fe_len;
4419 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4426 ac->ac_b_ex.fe_len = 0;
4428 ext4_mb_show_ac(ac);
4430 ext4_mb_release_context(ac);
4433 kmem_cache_free(ext4_ac_cachep, ac);
4434 if (inquota && ar->len < inquota)
4435 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4437 if (!ext4_test_inode_state(ar->inode,
4438 EXT4_STATE_DELALLOC_RESERVED))
4439 /* release all the reserved blocks if non delalloc */
4440 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4444 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4450 * We can merge two free data extents only if the physical blocks
4451 * are contiguous, AND the extents were freed by the same transaction,
4452 * AND the blocks are associated with the same group.
4454 static int can_merge(struct ext4_free_data *entry1,
4455 struct ext4_free_data *entry2)
4457 if ((entry1->t_tid == entry2->t_tid) &&
4458 (entry1->group == entry2->group) &&
4459 ((entry1->start_cluster + entry1->count) == entry2->start_cluster))
4464 static noinline_for_stack int
4465 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4466 struct ext4_free_data *new_entry)
4468 ext4_group_t group = e4b->bd_group;
4469 ext4_grpblk_t cluster;
4470 struct ext4_free_data *entry;
4471 struct ext4_group_info *db = e4b->bd_info;
4472 struct super_block *sb = e4b->bd_sb;
4473 struct ext4_sb_info *sbi = EXT4_SB(sb);
4474 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4475 struct rb_node *parent = NULL, *new_node;
4477 BUG_ON(!ext4_handle_valid(handle));
4478 BUG_ON(e4b->bd_bitmap_page == NULL);
4479 BUG_ON(e4b->bd_buddy_page == NULL);
4481 new_node = &new_entry->node;
4482 cluster = new_entry->start_cluster;
4485 /* first free block exent. We need to
4486 protect buddy cache from being freed,
4487 * otherwise we'll refresh it from
4488 * on-disk bitmap and lose not-yet-available
4490 page_cache_get(e4b->bd_buddy_page);
4491 page_cache_get(e4b->bd_bitmap_page);
4495 entry = rb_entry(parent, struct ext4_free_data, node);
4496 if (cluster < entry->start_cluster)
4498 else if (cluster >= (entry->start_cluster + entry->count))
4499 n = &(*n)->rb_right;
4501 ext4_grp_locked_error(sb, group, 0,
4502 ext4_group_first_block_no(sb, group) +
4503 EXT4_C2B(sbi, cluster),
4504 "Block already on to-be-freed list");
4509 rb_link_node(new_node, parent, n);
4510 rb_insert_color(new_node, &db->bb_free_root);
4512 /* Now try to see the extent can be merged to left and right */
4513 node = rb_prev(new_node);
4515 entry = rb_entry(node, struct ext4_free_data, node);
4516 if (can_merge(entry, new_entry)) {
4517 new_entry->start_cluster = entry->start_cluster;
4518 new_entry->count += entry->count;
4519 rb_erase(node, &(db->bb_free_root));
4520 spin_lock(&sbi->s_md_lock);
4521 list_del(&entry->list);
4522 spin_unlock(&sbi->s_md_lock);
4523 kmem_cache_free(ext4_free_ext_cachep, entry);
4527 node = rb_next(new_node);
4529 entry = rb_entry(node, struct ext4_free_data, node);
4530 if (can_merge(new_entry, entry)) {
4531 new_entry->count += entry->count;
4532 rb_erase(node, &(db->bb_free_root));
4533 spin_lock(&sbi->s_md_lock);
4534 list_del(&entry->list);
4535 spin_unlock(&sbi->s_md_lock);
4536 kmem_cache_free(ext4_free_ext_cachep, entry);
4539 /* Add the extent to transaction's private list */
4540 spin_lock(&sbi->s_md_lock);
4541 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4542 spin_unlock(&sbi->s_md_lock);
4547 * ext4_free_blocks() -- Free given blocks and update quota
4548 * @handle: handle for this transaction
4550 * @block: start physical block to free
4551 * @count: number of blocks to count
4552 * @flags: flags used by ext4_free_blocks
4554 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4555 struct buffer_head *bh, ext4_fsblk_t block,
4556 unsigned long count, int flags)
4558 struct buffer_head *bitmap_bh = NULL;
4559 struct super_block *sb = inode->i_sb;
4560 struct ext4_group_desc *gdp;
4561 unsigned long freed = 0;
4562 unsigned int overflow;
4564 struct buffer_head *gd_bh;
4565 ext4_group_t block_group;
4566 struct ext4_sb_info *sbi;
4567 struct ext4_buddy e4b;
4568 unsigned int count_clusters;
4574 BUG_ON(block != bh->b_blocknr);
4576 block = bh->b_blocknr;
4580 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4581 !ext4_data_block_valid(sbi, block, count)) {
4582 ext4_error(sb, "Freeing blocks not in datazone - "
4583 "block = %llu, count = %lu", block, count);
4587 ext4_debug("freeing block %llu\n", block);
4588 trace_ext4_free_blocks(inode, block, count, flags);
4590 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4591 struct buffer_head *tbh = bh;
4594 BUG_ON(bh && (count > 1));
4596 for (i = 0; i < count; i++) {
4598 tbh = sb_find_get_block(inode->i_sb,
4602 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4603 inode, tbh, block + i);
4608 * We need to make sure we don't reuse the freed block until
4609 * after the transaction is committed, which we can do by
4610 * treating the block as metadata, below. We make an
4611 * exception if the inode is to be written in writeback mode
4612 * since writeback mode has weak data consistency guarantees.
4614 if (!ext4_should_writeback_data(inode))
4615 flags |= EXT4_FREE_BLOCKS_METADATA;
4618 * If the extent to be freed does not begin on a cluster
4619 * boundary, we need to deal with partial clusters at the
4620 * beginning and end of the extent. Normally we will free
4621 * blocks at the beginning or the end unless we are explicitly
4622 * requested to avoid doing so.
4624 overflow = EXT4_PBLK_COFF(sbi, block);
4626 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4627 overflow = sbi->s_cluster_ratio - overflow;
4629 if (count > overflow)
4638 overflow = EXT4_LBLK_COFF(sbi, count);
4640 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4641 if (count > overflow)
4646 count += sbi->s_cluster_ratio - overflow;
4651 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4654 * Check to see if we are freeing blocks across a group
4657 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4658 overflow = EXT4_C2B(sbi, bit) + count -
4659 EXT4_BLOCKS_PER_GROUP(sb);
4662 count_clusters = EXT4_NUM_B2C(sbi, count);
4663 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4668 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4674 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4675 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4676 in_range(block, ext4_inode_table(sb, gdp),
4677 EXT4_SB(sb)->s_itb_per_group) ||
4678 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4679 EXT4_SB(sb)->s_itb_per_group)) {
4681 ext4_error(sb, "Freeing blocks in system zone - "
4682 "Block = %llu, count = %lu", block, count);
4683 /* err = 0. ext4_std_error should be a no op */
4687 BUFFER_TRACE(bitmap_bh, "getting write access");
4688 err = ext4_journal_get_write_access(handle, bitmap_bh);
4693 * We are about to modify some metadata. Call the journal APIs
4694 * to unshare ->b_data if a currently-committing transaction is
4697 BUFFER_TRACE(gd_bh, "get_write_access");
4698 err = ext4_journal_get_write_access(handle, gd_bh);
4701 #ifdef AGGRESSIVE_CHECK
4704 for (i = 0; i < count_clusters; i++)
4705 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4708 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4710 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4714 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4715 struct ext4_free_data *new_entry;
4717 * blocks being freed are metadata. these blocks shouldn't
4718 * be used until this transaction is committed
4720 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4723 new_entry = kmem_cache_alloc(ext4_free_ext_cachep,
4724 GFP_NOFS|__GFP_NOFAIL);
4725 new_entry->start_cluster = bit;
4726 new_entry->group = block_group;
4727 new_entry->count = count_clusters;
4728 new_entry->t_tid = handle->h_transaction->t_tid;
4730 ext4_lock_group(sb, block_group);
4731 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4732 ext4_mb_free_metadata(handle, &e4b, new_entry);
4734 /* need to update group_info->bb_free and bitmap
4735 * with group lock held. generate_buddy look at
4736 * them with group lock_held
4738 ext4_lock_group(sb, block_group);
4739 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4740 mb_free_blocks(inode, &e4b, bit, count_clusters);
4743 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4744 ext4_free_group_clusters_set(sb, gdp, ret);
4745 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4746 ext4_unlock_group(sb, block_group);
4747 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4749 if (sbi->s_log_groups_per_flex) {
4750 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4751 atomic64_add(count_clusters,
4752 &sbi->s_flex_groups[flex_group].free_clusters);
4755 ext4_mb_unload_buddy(&e4b);
4759 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4760 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4762 /* We dirtied the bitmap block */
4763 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4764 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4766 /* And the group descriptor block */
4767 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4768 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4772 if (overflow && !err) {
4778 ext4_mark_super_dirty(sb);
4781 ext4_std_error(sb, err);
4786 * ext4_group_add_blocks() -- Add given blocks to an existing group
4787 * @handle: handle to this transaction
4789 * @block: start physcial block to add to the block group
4790 * @count: number of blocks to free
4792 * This marks the blocks as free in the bitmap and buddy.
4794 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4795 ext4_fsblk_t block, unsigned long count)
4797 struct buffer_head *bitmap_bh = NULL;
4798 struct buffer_head *gd_bh;
4799 ext4_group_t block_group;
4802 struct ext4_group_desc *desc;
4803 struct ext4_sb_info *sbi = EXT4_SB(sb);
4804 struct ext4_buddy e4b;
4805 int err = 0, ret, blk_free_count;
4806 ext4_grpblk_t blocks_freed;
4808 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4813 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4815 * Check to see if we are freeing blocks across a group
4818 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4819 ext4_warning(sb, "too much blocks added to group %u\n",
4825 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4831 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4837 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4838 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4839 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4840 in_range(block + count - 1, ext4_inode_table(sb, desc),
4841 sbi->s_itb_per_group)) {
4842 ext4_error(sb, "Adding blocks in system zones - "
4843 "Block = %llu, count = %lu",
4849 BUFFER_TRACE(bitmap_bh, "getting write access");
4850 err = ext4_journal_get_write_access(handle, bitmap_bh);
4855 * We are about to modify some metadata. Call the journal APIs
4856 * to unshare ->b_data if a currently-committing transaction is
4859 BUFFER_TRACE(gd_bh, "get_write_access");
4860 err = ext4_journal_get_write_access(handle, gd_bh);
4864 for (i = 0, blocks_freed = 0; i < count; i++) {
4865 BUFFER_TRACE(bitmap_bh, "clear bit");
4866 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4867 ext4_error(sb, "bit already cleared for block %llu",
4868 (ext4_fsblk_t)(block + i));
4869 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4875 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4880 * need to update group_info->bb_free and bitmap
4881 * with group lock held. generate_buddy look at
4882 * them with group lock_held
4884 ext4_lock_group(sb, block_group);
4885 mb_clear_bits(bitmap_bh->b_data, bit, count);
4886 mb_free_blocks(NULL, &e4b, bit, count);
4887 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4888 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4889 desc->bg_checksum = ext4_group_desc_csum(sbi, block_group, desc);
4890 ext4_unlock_group(sb, block_group);
4891 percpu_counter_add(&sbi->s_freeclusters_counter,
4892 EXT4_NUM_B2C(sbi, blocks_freed));
4894 if (sbi->s_log_groups_per_flex) {
4895 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4896 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
4897 &sbi->s_flex_groups[flex_group].free_clusters);
4900 ext4_mb_unload_buddy(&e4b);
4902 /* We dirtied the bitmap block */
4903 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4904 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4906 /* And the group descriptor block */
4907 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4908 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4914 ext4_std_error(sb, err);
4919 * ext4_trim_extent -- function to TRIM one single free extent in the group
4920 * @sb: super block for the file system
4921 * @start: starting block of the free extent in the alloc. group
4922 * @count: number of blocks to TRIM
4923 * @group: alloc. group we are working with
4924 * @e4b: ext4 buddy for the group
4926 * Trim "count" blocks starting at "start" in the "group". To assure that no
4927 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4928 * be called with under the group lock.
4930 static void ext4_trim_extent(struct super_block *sb, int start, int count,
4931 ext4_group_t group, struct ext4_buddy *e4b)
4933 struct ext4_free_extent ex;
4935 trace_ext4_trim_extent(sb, group, start, count);
4937 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4939 ex.fe_start = start;
4940 ex.fe_group = group;
4944 * Mark blocks used, so no one can reuse them while
4947 mb_mark_used(e4b, &ex);
4948 ext4_unlock_group(sb, group);
4949 ext4_issue_discard(sb, group, start, count);
4950 ext4_lock_group(sb, group);
4951 mb_free_blocks(NULL, e4b, start, ex.fe_len);
4955 * ext4_trim_all_free -- function to trim all free space in alloc. group
4956 * @sb: super block for file system
4957 * @group: group to be trimmed
4958 * @start: first group block to examine
4959 * @max: last group block to examine
4960 * @minblocks: minimum extent block count
4962 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4963 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4967 * ext4_trim_all_free walks through group's block bitmap searching for free
4968 * extents. When the free extent is found, mark it as used in group buddy
4969 * bitmap. Then issue a TRIM command on this extent and free the extent in
4970 * the group buddy bitmap. This is done until whole group is scanned.
4972 static ext4_grpblk_t
4973 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4974 ext4_grpblk_t start, ext4_grpblk_t max,
4975 ext4_grpblk_t minblocks)
4978 ext4_grpblk_t next, count = 0, free_count = 0;
4979 struct ext4_buddy e4b;
4982 trace_ext4_trim_all_free(sb, group, start, max);
4984 ret = ext4_mb_load_buddy(sb, group, &e4b);
4986 ext4_error(sb, "Error in loading buddy "
4987 "information for %u", group);
4990 bitmap = e4b.bd_bitmap;
4992 ext4_lock_group(sb, group);
4993 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4994 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4997 start = (e4b.bd_info->bb_first_free > start) ?
4998 e4b.bd_info->bb_first_free : start;
5000 while (start < max) {
5001 start = mb_find_next_zero_bit(bitmap, max, start);
5004 next = mb_find_next_bit(bitmap, max, start);
5006 if ((next - start) >= minblocks) {
5007 ext4_trim_extent(sb, start,
5008 next - start, group, &e4b);
5009 count += next - start;
5011 free_count += next - start;
5014 if (fatal_signal_pending(current)) {
5015 count = -ERESTARTSYS;
5019 if (need_resched()) {
5020 ext4_unlock_group(sb, group);
5022 ext4_lock_group(sb, group);
5025 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5030 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5032 ext4_unlock_group(sb, group);
5033 ext4_mb_unload_buddy(&e4b);
5035 ext4_debug("trimmed %d blocks in the group %d\n",
5042 * ext4_trim_fs() -- trim ioctl handle function
5043 * @sb: superblock for filesystem
5044 * @range: fstrim_range structure
5046 * start: First Byte to trim
5047 * len: number of Bytes to trim from start
5048 * minlen: minimum extent length in Bytes
5049 * ext4_trim_fs goes through all allocation groups containing Bytes from
5050 * start to start+len. For each such a group ext4_trim_all_free function
5051 * is invoked to trim all free space.
5053 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5055 struct ext4_group_info *grp;
5056 ext4_group_t first_group, last_group;
5057 ext4_group_t group, ngroups = ext4_get_groups_count(sb);
5058 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5059 uint64_t start, len, minlen, trimmed = 0;
5060 ext4_fsblk_t first_data_blk =
5061 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5064 start = range->start >> sb->s_blocksize_bits;
5065 len = range->len >> sb->s_blocksize_bits;
5066 minlen = range->minlen >> sb->s_blocksize_bits;
5068 if (unlikely(minlen > EXT4_CLUSTERS_PER_GROUP(sb)))
5070 if (start + len <= first_data_blk)
5072 if (start < first_data_blk) {
5073 len -= first_data_blk - start;
5074 start = first_data_blk;
5077 /* Determine first and last group to examine based on start and len */
5078 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5079 &first_group, &first_cluster);
5080 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) (start + len),
5081 &last_group, &last_cluster);
5082 last_group = (last_group > ngroups - 1) ? ngroups - 1 : last_group;
5083 last_cluster = EXT4_CLUSTERS_PER_GROUP(sb);
5085 if (first_group > last_group)
5088 for (group = first_group; group <= last_group; group++) {
5089 grp = ext4_get_group_info(sb, group);
5090 /* We only do this if the grp has never been initialized */
5091 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5092 ret = ext4_mb_init_group(sb, group);
5098 * For all the groups except the last one, last block will
5099 * always be EXT4_BLOCKS_PER_GROUP(sb), so we only need to
5100 * change it for the last group in which case start +
5101 * len < EXT4_BLOCKS_PER_GROUP(sb).
5103 if (first_cluster + len < EXT4_CLUSTERS_PER_GROUP(sb))
5104 last_cluster = first_cluster + len;
5105 len -= last_cluster - first_cluster;
5107 if (grp->bb_free >= minlen) {
5108 cnt = ext4_trim_all_free(sb, group, first_cluster,
5109 last_cluster, minlen);
5118 range->len = trimmed * sb->s_blocksize;
5121 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);