2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
27 * - test ext4_ext_search_left() and ext4_ext_search_right()
28 * - search for metadata in few groups
31 * - normalization should take into account whether file is still open
32 * - discard preallocations if no free space left (policy?)
33 * - don't normalize tails
35 * - reservation for superuser
38 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
39 * - track min/max extents in each group for better group selection
40 * - mb_mark_used() may allocate chunk right after splitting buddy
41 * - tree of groups sorted by number of free blocks
46 * The allocation request involve request for multiple number of blocks
47 * near to the goal(block) value specified.
49 * During initialization phase of the allocator we decide to use the group
50 * preallocation or inode preallocation depending on the size file. The
51 * size of the file could be the resulting file size we would have after
52 * allocation or the current file size which ever is larger. If the size is
53 * less that sbi->s_mb_stream_request we select the group
54 * preallocation. The default value of s_mb_stream_request is 16
55 * blocks. This can also be tuned via
56 * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms
57 * of number of blocks.
59 * The main motivation for having small file use group preallocation is to
60 * ensure that we have small file closer in the disk.
62 * First stage the allocator looks at the inode prealloc list
63 * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for
64 * this particular inode. The inode prealloc space is represented as:
66 * pa_lstart -> the logical start block for this prealloc space
67 * pa_pstart -> the physical start block for this prealloc space
68 * pa_len -> lenght for this prealloc space
69 * pa_free -> free space available in this prealloc space
71 * The inode preallocation space is used looking at the _logical_ start
72 * block. If only the logical file block falls within the range of prealloc
73 * space we will consume the particular prealloc space. This make sure that
74 * that the we have contiguous physical blocks representing the file blocks
76 * The important thing to be noted in case of inode prealloc space is that
77 * we don't modify the values associated to inode prealloc space except
80 * If we are not able to find blocks in the inode prealloc space and if we
81 * have the group allocation flag set then we look at the locality group
82 * prealloc space. These are per CPU prealloc list repreasented as
84 * ext4_sb_info.s_locality_groups[smp_processor_id()]
86 * The reason for having a per cpu locality group is to reduce the contention
87 * between CPUs. It is possible to get scheduled at this point.
89 * The locality group prealloc space is used looking at whether we have
90 * enough free space (pa_free) withing the prealloc space.
92 * If we can't allocate blocks via inode prealloc or/and locality group
93 * prealloc then we look at the buddy cache. The buddy cache is represented
94 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
95 * mapped to the buddy and bitmap information regarding different
96 * groups. The buddy information is attached to buddy cache inode so that
97 * we can access them through the page cache. The information regarding
98 * each group is loaded via ext4_mb_load_buddy. The information involve
99 * block bitmap and buddy information. The information are stored in the
103 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
106 * one block each for bitmap and buddy information. So for each group we
107 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
108 * blocksize) blocks. So it can have information regarding groups_per_page
109 * which is blocks_per_page/2
111 * The buddy cache inode is not stored on disk. The inode is thrown
112 * away when the filesystem is unmounted.
114 * We look for count number of blocks in the buddy cache. If we were able
115 * to locate that many free blocks we return with additional information
116 * regarding rest of the contiguous physical block available
118 * Before allocating blocks via buddy cache we normalize the request
119 * blocks. This ensure we ask for more blocks that we needed. The extra
120 * blocks that we get after allocation is added to the respective prealloc
121 * list. In case of inode preallocation we follow a list of heuristics
122 * based on file size. This can be found in ext4_mb_normalize_request. If
123 * we are doing a group prealloc we try to normalize the request to
124 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is set to
125 * 512 blocks. This can be tuned via
126 * /proc/fs/ext4/<partition/group_prealloc. The value is represented in
127 * terms of number of blocks. If we have mounted the file system with -O
128 * stripe=<value> option the group prealloc request is normalized to the
129 * stripe value (sbi->s_stripe)
131 * The regular allocator(using the buddy cache) support few tunables.
133 * /proc/fs/ext4/<partition>/min_to_scan
134 * /proc/fs/ext4/<partition>/max_to_scan
135 * /proc/fs/ext4/<partition>/order2_req
137 * The regular allocator use buddy scan only if the request len is power of
138 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
139 * value of s_mb_order2_reqs can be tuned via
140 * /proc/fs/ext4/<partition>/order2_req. If the request len is equal to
141 * stripe size (sbi->s_stripe), we try to search for contigous block in
142 * stripe size. This should result in better allocation on RAID setup. If
143 * not we search in the specific group using bitmap for best extents. The
144 * tunable min_to_scan and max_to_scan controll the behaviour here.
145 * min_to_scan indicate how long the mballoc __must__ look for a best
146 * extent and max_to_scanindicate how long the mballoc __can__ look for a
147 * best extent in the found extents. Searching for the blocks starts with
148 * the group specified as the goal value in allocation context via
149 * ac_g_ex. Each group is first checked based on the criteria whether it
150 * can used for allocation. ext4_mb_good_group explains how the groups are
153 * Both the prealloc space are getting populated as above. So for the first
154 * request we will hit the buddy cache which will result in this prealloc
155 * space getting filled. The prealloc space is then later used for the
156 * subsequent request.
160 * mballoc operates on the following data:
162 * - in-core buddy (actually includes buddy and bitmap)
163 * - preallocation descriptors (PAs)
165 * there are two types of preallocations:
167 * assiged to specific inode and can be used for this inode only.
168 * it describes part of inode's space preallocated to specific
169 * physical blocks. any block from that preallocated can be used
170 * independent. the descriptor just tracks number of blocks left
171 * unused. so, before taking some block from descriptor, one must
172 * make sure corresponded logical block isn't allocated yet. this
173 * also means that freeing any block within descriptor's range
174 * must discard all preallocated blocks.
176 * assigned to specific locality group which does not translate to
177 * permanent set of inodes: inode can join and leave group. space
178 * from this type of preallocation can be used for any inode. thus
179 * it's consumed from the beginning to the end.
181 * relation between them can be expressed as:
182 * in-core buddy = on-disk bitmap + preallocation descriptors
184 * this mean blocks mballoc considers used are:
185 * - allocated blocks (persistent)
186 * - preallocated blocks (non-persistent)
188 * consistency in mballoc world means that at any time a block is either
189 * free or used in ALL structures. notice: "any time" should not be read
190 * literally -- time is discrete and delimited by locks.
192 * to keep it simple, we don't use block numbers, instead we count number of
193 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
195 * all operations can be expressed as:
196 * - init buddy: buddy = on-disk + PAs
197 * - new PA: buddy += N; PA = N
198 * - use inode PA: on-disk += N; PA -= N
199 * - discard inode PA buddy -= on-disk - PA; PA = 0
200 * - use locality group PA on-disk += N; PA -= N
201 * - discard locality group PA buddy -= PA; PA = 0
202 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
203 * is used in real operation because we can't know actual used
204 * bits from PA, only from on-disk bitmap
206 * if we follow this strict logic, then all operations above should be atomic.
207 * given some of them can block, we'd have to use something like semaphores
208 * killing performance on high-end SMP hardware. let's try to relax it using
209 * the following knowledge:
210 * 1) if buddy is referenced, it's already initialized
211 * 2) while block is used in buddy and the buddy is referenced,
212 * nobody can re-allocate that block
213 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
214 * bit set and PA claims same block, it's OK. IOW, one can set bit in
215 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
218 * so, now we're building a concurrency table:
221 * blocks for PA are allocated in the buddy, buddy must be referenced
222 * until PA is linked to allocation group to avoid concurrent buddy init
224 * we need to make sure that either on-disk bitmap or PA has uptodate data
225 * given (3) we care that PA-=N operation doesn't interfere with init
227 * the simplest way would be to have buddy initialized by the discard
228 * - use locality group PA
229 * again PA-=N must be serialized with init
230 * - discard locality group PA
231 * the simplest way would be to have buddy initialized by the discard
234 * i_data_sem serializes them
236 * discard process must wait until PA isn't used by another process
237 * - use locality group PA
238 * some mutex should serialize them
239 * - discard locality group PA
240 * discard process must wait until PA isn't used by another process
243 * i_data_sem or another mutex should serializes them
245 * discard process must wait until PA isn't used by another process
246 * - use locality group PA
247 * nothing wrong here -- they're different PAs covering different blocks
248 * - discard locality group PA
249 * discard process must wait until PA isn't used by another process
251 * now we're ready to make few consequences:
252 * - PA is referenced and while it is no discard is possible
253 * - PA is referenced until block isn't marked in on-disk bitmap
254 * - PA changes only after on-disk bitmap
255 * - discard must not compete with init. either init is done before
256 * any discard or they're serialized somehow
257 * - buddy init as sum of on-disk bitmap and PAs is done atomically
259 * a special case when we've used PA to emptiness. no need to modify buddy
260 * in this case, but we should care about concurrent init
265 * Logic in few words:
270 * mark bits in on-disk bitmap
273 * - use preallocation:
274 * find proper PA (per-inode or group)
276 * mark bits in on-disk bitmap
282 * mark bits in on-disk bitmap
285 * - discard preallocations in group:
287 * move them onto local list
288 * load on-disk bitmap
290 * remove PA from object (inode or locality group)
291 * mark free blocks in-core
293 * - discard inode's preallocations:
300 * - bitlock on a group (group)
301 * - object (inode/locality) (object)
312 * - release consumed pa:
317 * - generate in-core bitmap:
321 * - discard all for given object (inode, locality group):
326 * - discard all for given group:
334 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
336 #if BITS_PER_LONG == 64
337 *bit += ((unsigned long) addr & 7UL) << 3;
338 addr = (void *) ((unsigned long) addr & ~7UL);
339 #elif BITS_PER_LONG == 32
340 *bit += ((unsigned long) addr & 3UL) << 3;
341 addr = (void *) ((unsigned long) addr & ~3UL);
343 #error "how many bits you are?!"
348 static inline int mb_test_bit(int bit, void *addr)
351 * ext4_test_bit on architecture like powerpc
352 * needs unsigned long aligned address
354 addr = mb_correct_addr_and_bit(&bit, addr);
355 return ext4_test_bit(bit, addr);
358 static inline void mb_set_bit(int bit, void *addr)
360 addr = mb_correct_addr_and_bit(&bit, addr);
361 ext4_set_bit(bit, addr);
364 static inline void mb_set_bit_atomic(spinlock_t *lock, int bit, void *addr)
366 addr = mb_correct_addr_and_bit(&bit, addr);
367 ext4_set_bit_atomic(lock, bit, addr);
370 static inline void mb_clear_bit(int bit, void *addr)
372 addr = mb_correct_addr_and_bit(&bit, addr);
373 ext4_clear_bit(bit, addr);
376 static inline void mb_clear_bit_atomic(spinlock_t *lock, int bit, void *addr)
378 addr = mb_correct_addr_and_bit(&bit, addr);
379 ext4_clear_bit_atomic(lock, bit, addr);
382 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
384 int fix = 0, ret, tmpmax;
385 addr = mb_correct_addr_and_bit(&fix, addr);
389 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
395 static inline int mb_find_next_bit(void *addr, int max, int start)
397 int fix = 0, ret, tmpmax;
398 addr = mb_correct_addr_and_bit(&fix, addr);
402 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
408 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
412 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
415 if (order > e4b->bd_blkbits + 1) {
420 /* at order 0 we see each particular block */
421 *max = 1 << (e4b->bd_blkbits + 3);
423 return EXT4_MB_BITMAP(e4b);
425 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
426 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
432 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
433 int first, int count)
436 struct super_block *sb = e4b->bd_sb;
438 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
440 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
441 for (i = 0; i < count; i++) {
442 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
443 ext4_fsblk_t blocknr;
444 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
445 blocknr += first + i;
447 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
449 ext4_error(sb, __func__, "double-free of inode"
450 " %lu's block %llu(bit %u in group %lu)\n",
451 inode ? inode->i_ino : 0, blocknr,
452 first + i, e4b->bd_group);
454 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
458 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
462 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
464 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
465 for (i = 0; i < count; i++) {
466 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
467 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
471 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
473 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
474 unsigned char *b1, *b2;
476 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
477 b2 = (unsigned char *) bitmap;
478 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
479 if (b1[i] != b2[i]) {
480 printk("corruption in group %lu at byte %u(%u):"
481 " %x in copy != %x on disk/prealloc\n",
482 e4b->bd_group, i, i * 8, b1[i], b2[i]);
490 static inline void mb_free_blocks_double(struct inode *inode,
491 struct ext4_buddy *e4b, int first, int count)
495 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
496 int first, int count)
500 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
506 #ifdef AGGRESSIVE_CHECK
508 #define MB_CHECK_ASSERT(assert) \
512 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
513 function, file, line, # assert); \
518 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
519 const char *function, int line)
521 struct super_block *sb = e4b->bd_sb;
522 int order = e4b->bd_blkbits + 1;
529 struct ext4_group_info *grp;
532 struct list_head *cur;
536 if (!test_opt(sb, MBALLOC))
540 static int mb_check_counter;
541 if (mb_check_counter++ % 100 != 0)
546 buddy = mb_find_buddy(e4b, order, &max);
547 MB_CHECK_ASSERT(buddy);
548 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
549 MB_CHECK_ASSERT(buddy2);
550 MB_CHECK_ASSERT(buddy != buddy2);
551 MB_CHECK_ASSERT(max * 2 == max2);
554 for (i = 0; i < max; i++) {
556 if (mb_test_bit(i, buddy)) {
557 /* only single bit in buddy2 may be 1 */
558 if (!mb_test_bit(i << 1, buddy2)) {
560 mb_test_bit((i<<1)+1, buddy2));
561 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
563 mb_test_bit(i << 1, buddy2));
568 /* both bits in buddy2 must be 0 */
569 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
570 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
572 for (j = 0; j < (1 << order); j++) {
573 k = (i * (1 << order)) + j;
575 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
579 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
584 buddy = mb_find_buddy(e4b, 0, &max);
585 for (i = 0; i < max; i++) {
586 if (!mb_test_bit(i, buddy)) {
587 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
595 /* check used bits only */
596 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
597 buddy2 = mb_find_buddy(e4b, j, &max2);
599 MB_CHECK_ASSERT(k < max2);
600 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
603 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
604 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
606 grp = ext4_get_group_info(sb, e4b->bd_group);
607 buddy = mb_find_buddy(e4b, 0, &max);
608 list_for_each(cur, &grp->bb_prealloc_list) {
609 ext4_group_t groupnr;
610 struct ext4_prealloc_space *pa;
611 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
612 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
613 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
614 for (i = 0; i < pa->pa_len; i++)
615 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
619 #undef MB_CHECK_ASSERT
620 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
621 __FILE__, __func__, __LINE__)
623 #define mb_check_buddy(e4b)
626 /* FIXME!! need more doc */
627 static void ext4_mb_mark_free_simple(struct super_block *sb,
628 void *buddy, unsigned first, int len,
629 struct ext4_group_info *grp)
631 struct ext4_sb_info *sbi = EXT4_SB(sb);
634 unsigned short chunk;
635 unsigned short border;
637 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
639 border = 2 << sb->s_blocksize_bits;
642 /* find how many blocks can be covered since this position */
643 max = ffs(first | border) - 1;
645 /* find how many blocks of power 2 we need to mark */
652 /* mark multiblock chunks only */
653 grp->bb_counters[min]++;
655 mb_clear_bit(first >> min,
656 buddy + sbi->s_mb_offsets[min]);
663 static void ext4_mb_generate_buddy(struct super_block *sb,
664 void *buddy, void *bitmap, ext4_group_t group)
666 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
667 unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
668 unsigned short i = 0;
669 unsigned short first;
672 unsigned fragments = 0;
673 unsigned long long period = get_cycles();
675 /* initialize buddy from bitmap which is aggregation
676 * of on-disk bitmap and preallocations */
677 i = mb_find_next_zero_bit(bitmap, max, 0);
678 grp->bb_first_free = i;
682 i = mb_find_next_bit(bitmap, max, i);
686 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
688 grp->bb_counters[0]++;
690 i = mb_find_next_zero_bit(bitmap, max, i);
692 grp->bb_fragments = fragments;
694 if (free != grp->bb_free) {
695 ext4_error(sb, __func__,
696 "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
697 group, free, grp->bb_free);
699 * If we intent to continue, we consider group descritor
700 * corrupt and update bb_free using bitmap value
705 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
707 period = get_cycles() - period;
708 spin_lock(&EXT4_SB(sb)->s_bal_lock);
709 EXT4_SB(sb)->s_mb_buddies_generated++;
710 EXT4_SB(sb)->s_mb_generation_time += period;
711 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
714 /* The buddy information is attached the buddy cache inode
715 * for convenience. The information regarding each group
716 * is loaded via ext4_mb_load_buddy. The information involve
717 * block bitmap and buddy information. The information are
718 * stored in the inode as
721 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
724 * one block each for bitmap and buddy information.
725 * So for each group we take up 2 blocks. A page can
726 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
727 * So it can have information regarding groups_per_page which
728 * is blocks_per_page/2
731 static int ext4_mb_init_cache(struct page *page, char *incore)
738 ext4_group_t first_group;
740 struct super_block *sb;
741 struct buffer_head *bhs;
742 struct buffer_head **bh;
747 mb_debug("init page %lu\n", page->index);
749 inode = page->mapping->host;
751 blocksize = 1 << inode->i_blkbits;
752 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
754 groups_per_page = blocks_per_page >> 1;
755 if (groups_per_page == 0)
758 /* allocate buffer_heads to read bitmaps */
759 if (groups_per_page > 1) {
761 i = sizeof(struct buffer_head *) * groups_per_page;
762 bh = kzalloc(i, GFP_NOFS);
768 first_group = page->index * blocks_per_page / 2;
770 /* read all groups the page covers into the cache */
771 for (i = 0; i < groups_per_page; i++) {
772 struct ext4_group_desc *desc;
774 if (first_group + i >= EXT4_SB(sb)->s_groups_count)
778 desc = ext4_get_group_desc(sb, first_group + i, NULL);
783 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
787 if (bh_uptodate_or_lock(bh[i]))
790 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
791 ext4_init_block_bitmap(sb, bh[i],
792 first_group + i, desc);
793 set_buffer_uptodate(bh[i]);
794 unlock_buffer(bh[i]);
798 bh[i]->b_end_io = end_buffer_read_sync;
799 submit_bh(READ, bh[i]);
800 mb_debug("read bitmap for group %lu\n", first_group + i);
803 /* wait for I/O completion */
804 for (i = 0; i < groups_per_page && bh[i]; i++)
805 wait_on_buffer(bh[i]);
808 for (i = 0; i < groups_per_page && bh[i]; i++)
809 if (!buffer_uptodate(bh[i]))
813 first_block = page->index * blocks_per_page;
814 for (i = 0; i < blocks_per_page; i++) {
816 struct ext4_group_info *grinfo;
818 group = (first_block + i) >> 1;
819 if (group >= EXT4_SB(sb)->s_groups_count)
823 * data carry information regarding this
824 * particular group in the format specified
828 data = page_address(page) + (i * blocksize);
829 bitmap = bh[group - first_group]->b_data;
832 * We place the buddy block and bitmap block
835 if ((first_block + i) & 1) {
836 /* this is block of buddy */
837 BUG_ON(incore == NULL);
838 mb_debug("put buddy for group %u in page %lu/%x\n",
839 group, page->index, i * blocksize);
840 memset(data, 0xff, blocksize);
841 grinfo = ext4_get_group_info(sb, group);
842 grinfo->bb_fragments = 0;
843 memset(grinfo->bb_counters, 0,
844 sizeof(unsigned short)*(sb->s_blocksize_bits+2));
846 * incore got set to the group block bitmap below
848 ext4_mb_generate_buddy(sb, data, incore, group);
851 /* this is block of bitmap */
852 BUG_ON(incore != NULL);
853 mb_debug("put bitmap for group %u in page %lu/%x\n",
854 group, page->index, i * blocksize);
856 /* see comments in ext4_mb_put_pa() */
857 ext4_lock_group(sb, group);
858 memcpy(data, bitmap, blocksize);
860 /* mark all preallocated blks used in in-core bitmap */
861 ext4_mb_generate_from_pa(sb, data, group);
862 ext4_unlock_group(sb, group);
864 /* set incore so that the buddy information can be
865 * generated using this
870 SetPageUptodate(page);
874 for (i = 0; i < groups_per_page && bh[i]; i++)
882 static noinline_for_stack int
883 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
884 struct ext4_buddy *e4b)
886 struct ext4_sb_info *sbi = EXT4_SB(sb);
887 struct inode *inode = sbi->s_buddy_cache;
894 mb_debug("load group %lu\n", group);
896 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
898 e4b->bd_blkbits = sb->s_blocksize_bits;
899 e4b->bd_info = ext4_get_group_info(sb, group);
901 e4b->bd_group = group;
902 e4b->bd_buddy_page = NULL;
903 e4b->bd_bitmap_page = NULL;
906 * the buddy cache inode stores the block bitmap
907 * and buddy information in consecutive blocks.
908 * So for each group we need two blocks.
911 pnum = block / blocks_per_page;
912 poff = block % blocks_per_page;
914 /* we could use find_or_create_page(), but it locks page
915 * what we'd like to avoid in fast path ... */
916 page = find_get_page(inode->i_mapping, pnum);
917 if (page == NULL || !PageUptodate(page)) {
919 page_cache_release(page);
920 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
922 BUG_ON(page->mapping != inode->i_mapping);
923 if (!PageUptodate(page)) {
924 ext4_mb_init_cache(page, NULL);
925 mb_cmp_bitmaps(e4b, page_address(page) +
926 (poff * sb->s_blocksize));
931 if (page == NULL || !PageUptodate(page))
933 e4b->bd_bitmap_page = page;
934 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
935 mark_page_accessed(page);
938 pnum = block / blocks_per_page;
939 poff = block % blocks_per_page;
941 page = find_get_page(inode->i_mapping, pnum);
942 if (page == NULL || !PageUptodate(page)) {
944 page_cache_release(page);
945 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
947 BUG_ON(page->mapping != inode->i_mapping);
948 if (!PageUptodate(page))
949 ext4_mb_init_cache(page, e4b->bd_bitmap);
954 if (page == NULL || !PageUptodate(page))
956 e4b->bd_buddy_page = page;
957 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
958 mark_page_accessed(page);
960 BUG_ON(e4b->bd_bitmap_page == NULL);
961 BUG_ON(e4b->bd_buddy_page == NULL);
966 if (e4b->bd_bitmap_page)
967 page_cache_release(e4b->bd_bitmap_page);
968 if (e4b->bd_buddy_page)
969 page_cache_release(e4b->bd_buddy_page);
970 e4b->bd_buddy = NULL;
971 e4b->bd_bitmap = NULL;
975 static void ext4_mb_release_desc(struct ext4_buddy *e4b)
977 if (e4b->bd_bitmap_page)
978 page_cache_release(e4b->bd_bitmap_page);
979 if (e4b->bd_buddy_page)
980 page_cache_release(e4b->bd_buddy_page);
984 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
989 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
990 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
992 bb = EXT4_MB_BUDDY(e4b);
993 while (order <= e4b->bd_blkbits + 1) {
995 if (!mb_test_bit(block, bb)) {
996 /* this block is part of buddy of order 'order' */
999 bb += 1 << (e4b->bd_blkbits - order);
1005 static void mb_clear_bits(spinlock_t *lock, void *bm, int cur, int len)
1011 if ((cur & 31) == 0 && (len - cur) >= 32) {
1012 /* fast path: clear whole word at once */
1013 addr = bm + (cur >> 3);
1018 mb_clear_bit_atomic(lock, cur, bm);
1023 static void mb_set_bits(spinlock_t *lock, void *bm, int cur, int len)
1029 if ((cur & 31) == 0 && (len - cur) >= 32) {
1030 /* fast path: set whole word at once */
1031 addr = bm + (cur >> 3);
1036 mb_set_bit_atomic(lock, cur, bm);
1041 static int mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1042 int first, int count)
1049 struct super_block *sb = e4b->bd_sb;
1051 BUG_ON(first + count > (sb->s_blocksize << 3));
1052 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
1053 mb_check_buddy(e4b);
1054 mb_free_blocks_double(inode, e4b, first, count);
1056 e4b->bd_info->bb_free += count;
1057 if (first < e4b->bd_info->bb_first_free)
1058 e4b->bd_info->bb_first_free = first;
1060 /* let's maintain fragments counter */
1062 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1063 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1064 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1066 e4b->bd_info->bb_fragments--;
1067 else if (!block && !max)
1068 e4b->bd_info->bb_fragments++;
1070 /* let's maintain buddy itself */
1071 while (count-- > 0) {
1075 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1076 ext4_fsblk_t blocknr;
1077 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1080 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
1082 ext4_error(sb, __func__, "double-free of inode"
1083 " %lu's block %llu(bit %u in group %lu)\n",
1084 inode ? inode->i_ino : 0, blocknr, block,
1087 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1088 e4b->bd_info->bb_counters[order]++;
1090 /* start of the buddy */
1091 buddy = mb_find_buddy(e4b, order, &max);
1095 if (mb_test_bit(block, buddy) ||
1096 mb_test_bit(block + 1, buddy))
1099 /* both the buddies are free, try to coalesce them */
1100 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1106 /* for special purposes, we don't set
1107 * free bits in bitmap */
1108 mb_set_bit(block, buddy);
1109 mb_set_bit(block + 1, buddy);
1111 e4b->bd_info->bb_counters[order]--;
1112 e4b->bd_info->bb_counters[order]--;
1116 e4b->bd_info->bb_counters[order]++;
1118 mb_clear_bit(block, buddy2);
1122 mb_check_buddy(e4b);
1127 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1128 int needed, struct ext4_free_extent *ex)
1135 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1138 buddy = mb_find_buddy(e4b, order, &max);
1139 BUG_ON(buddy == NULL);
1140 BUG_ON(block >= max);
1141 if (mb_test_bit(block, buddy)) {
1148 /* FIXME dorp order completely ? */
1149 if (likely(order == 0)) {
1150 /* find actual order */
1151 order = mb_find_order_for_block(e4b, block);
1152 block = block >> order;
1155 ex->fe_len = 1 << order;
1156 ex->fe_start = block << order;
1157 ex->fe_group = e4b->bd_group;
1159 /* calc difference from given start */
1160 next = next - ex->fe_start;
1162 ex->fe_start += next;
1164 while (needed > ex->fe_len &&
1165 (buddy = mb_find_buddy(e4b, order, &max))) {
1167 if (block + 1 >= max)
1170 next = (block + 1) * (1 << order);
1171 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1174 ord = mb_find_order_for_block(e4b, next);
1177 block = next >> order;
1178 ex->fe_len += 1 << order;
1181 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1185 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1191 int start = ex->fe_start;
1192 int len = ex->fe_len;
1197 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1198 BUG_ON(e4b->bd_group != ex->fe_group);
1199 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1200 mb_check_buddy(e4b);
1201 mb_mark_used_double(e4b, start, len);
1203 e4b->bd_info->bb_free -= len;
1204 if (e4b->bd_info->bb_first_free == start)
1205 e4b->bd_info->bb_first_free += len;
1207 /* let's maintain fragments counter */
1209 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1210 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1211 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1213 e4b->bd_info->bb_fragments++;
1214 else if (!mlen && !max)
1215 e4b->bd_info->bb_fragments--;
1217 /* let's maintain buddy itself */
1219 ord = mb_find_order_for_block(e4b, start);
1221 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1222 /* the whole chunk may be allocated at once! */
1224 buddy = mb_find_buddy(e4b, ord, &max);
1225 BUG_ON((start >> ord) >= max);
1226 mb_set_bit(start >> ord, buddy);
1227 e4b->bd_info->bb_counters[ord]--;
1234 /* store for history */
1236 ret = len | (ord << 16);
1238 /* we have to split large buddy */
1240 buddy = mb_find_buddy(e4b, ord, &max);
1241 mb_set_bit(start >> ord, buddy);
1242 e4b->bd_info->bb_counters[ord]--;
1245 cur = (start >> ord) & ~1U;
1246 buddy = mb_find_buddy(e4b, ord, &max);
1247 mb_clear_bit(cur, buddy);
1248 mb_clear_bit(cur + 1, buddy);
1249 e4b->bd_info->bb_counters[ord]++;
1250 e4b->bd_info->bb_counters[ord]++;
1253 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b->bd_sb), ex->fe_group),
1254 EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1255 mb_check_buddy(e4b);
1261 * Must be called under group lock!
1263 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1264 struct ext4_buddy *e4b)
1266 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1269 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1270 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1272 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1273 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1274 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1276 /* preallocation can change ac_b_ex, thus we store actually
1277 * allocated blocks for history */
1278 ac->ac_f_ex = ac->ac_b_ex;
1280 ac->ac_status = AC_STATUS_FOUND;
1281 ac->ac_tail = ret & 0xffff;
1282 ac->ac_buddy = ret >> 16;
1284 /* XXXXXXX: SUCH A HORRIBLE **CK */
1286 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1287 get_page(ac->ac_bitmap_page);
1288 ac->ac_buddy_page = e4b->bd_buddy_page;
1289 get_page(ac->ac_buddy_page);
1291 /* store last allocated for subsequent stream allocation */
1292 if ((ac->ac_flags & EXT4_MB_HINT_DATA)) {
1293 spin_lock(&sbi->s_md_lock);
1294 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1295 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1296 spin_unlock(&sbi->s_md_lock);
1301 * regular allocator, for general purposes allocation
1304 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1305 struct ext4_buddy *e4b,
1308 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1309 struct ext4_free_extent *bex = &ac->ac_b_ex;
1310 struct ext4_free_extent *gex = &ac->ac_g_ex;
1311 struct ext4_free_extent ex;
1315 * We don't want to scan for a whole year
1317 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1318 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1319 ac->ac_status = AC_STATUS_BREAK;
1324 * Haven't found good chunk so far, let's continue
1326 if (bex->fe_len < gex->fe_len)
1329 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1330 && bex->fe_group == e4b->bd_group) {
1331 /* recheck chunk's availability - we don't know
1332 * when it was found (within this lock-unlock
1334 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1335 if (max >= gex->fe_len) {
1336 ext4_mb_use_best_found(ac, e4b);
1343 * The routine checks whether found extent is good enough. If it is,
1344 * then the extent gets marked used and flag is set to the context
1345 * to stop scanning. Otherwise, the extent is compared with the
1346 * previous found extent and if new one is better, then it's stored
1347 * in the context. Later, the best found extent will be used, if
1348 * mballoc can't find good enough extent.
1350 * FIXME: real allocation policy is to be designed yet!
1352 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1353 struct ext4_free_extent *ex,
1354 struct ext4_buddy *e4b)
1356 struct ext4_free_extent *bex = &ac->ac_b_ex;
1357 struct ext4_free_extent *gex = &ac->ac_g_ex;
1359 BUG_ON(ex->fe_len <= 0);
1360 BUG_ON(ex->fe_len >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1361 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1362 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1367 * The special case - take what you catch first
1369 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1371 ext4_mb_use_best_found(ac, e4b);
1376 * Let's check whether the chuck is good enough
1378 if (ex->fe_len == gex->fe_len) {
1380 ext4_mb_use_best_found(ac, e4b);
1385 * If this is first found extent, just store it in the context
1387 if (bex->fe_len == 0) {
1393 * If new found extent is better, store it in the context
1395 if (bex->fe_len < gex->fe_len) {
1396 /* if the request isn't satisfied, any found extent
1397 * larger than previous best one is better */
1398 if (ex->fe_len > bex->fe_len)
1400 } else if (ex->fe_len > gex->fe_len) {
1401 /* if the request is satisfied, then we try to find
1402 * an extent that still satisfy the request, but is
1403 * smaller than previous one */
1404 if (ex->fe_len < bex->fe_len)
1408 ext4_mb_check_limits(ac, e4b, 0);
1411 static int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1412 struct ext4_buddy *e4b)
1414 struct ext4_free_extent ex = ac->ac_b_ex;
1415 ext4_group_t group = ex.fe_group;
1419 BUG_ON(ex.fe_len <= 0);
1420 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1424 ext4_lock_group(ac->ac_sb, group);
1425 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1429 ext4_mb_use_best_found(ac, e4b);
1432 ext4_unlock_group(ac->ac_sb, group);
1433 ext4_mb_release_desc(e4b);
1438 static int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1439 struct ext4_buddy *e4b)
1441 ext4_group_t group = ac->ac_g_ex.fe_group;
1444 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1445 struct ext4_super_block *es = sbi->s_es;
1446 struct ext4_free_extent ex;
1448 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1451 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1455 ext4_lock_group(ac->ac_sb, group);
1456 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1457 ac->ac_g_ex.fe_len, &ex);
1459 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1462 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1463 ex.fe_start + le32_to_cpu(es->s_first_data_block);
1464 /* use do_div to get remainder (would be 64-bit modulo) */
1465 if (do_div(start, sbi->s_stripe) == 0) {
1468 ext4_mb_use_best_found(ac, e4b);
1470 } else if (max >= ac->ac_g_ex.fe_len) {
1471 BUG_ON(ex.fe_len <= 0);
1472 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1473 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1476 ext4_mb_use_best_found(ac, e4b);
1477 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1478 /* Sometimes, caller may want to merge even small
1479 * number of blocks to an existing extent */
1480 BUG_ON(ex.fe_len <= 0);
1481 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1482 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1485 ext4_mb_use_best_found(ac, e4b);
1487 ext4_unlock_group(ac->ac_sb, group);
1488 ext4_mb_release_desc(e4b);
1494 * The routine scans buddy structures (not bitmap!) from given order
1495 * to max order and tries to find big enough chunk to satisfy the req
1497 static void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1498 struct ext4_buddy *e4b)
1500 struct super_block *sb = ac->ac_sb;
1501 struct ext4_group_info *grp = e4b->bd_info;
1507 BUG_ON(ac->ac_2order <= 0);
1508 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1509 if (grp->bb_counters[i] == 0)
1512 buddy = mb_find_buddy(e4b, i, &max);
1513 BUG_ON(buddy == NULL);
1515 k = mb_find_next_zero_bit(buddy, max, 0);
1520 ac->ac_b_ex.fe_len = 1 << i;
1521 ac->ac_b_ex.fe_start = k << i;
1522 ac->ac_b_ex.fe_group = e4b->bd_group;
1524 ext4_mb_use_best_found(ac, e4b);
1526 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1528 if (EXT4_SB(sb)->s_mb_stats)
1529 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1536 * The routine scans the group and measures all found extents.
1537 * In order to optimize scanning, caller must pass number of
1538 * free blocks in the group, so the routine can know upper limit.
1540 static void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1541 struct ext4_buddy *e4b)
1543 struct super_block *sb = ac->ac_sb;
1544 void *bitmap = EXT4_MB_BITMAP(e4b);
1545 struct ext4_free_extent ex;
1549 free = e4b->bd_info->bb_free;
1552 i = e4b->bd_info->bb_first_free;
1554 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1555 i = mb_find_next_zero_bit(bitmap,
1556 EXT4_BLOCKS_PER_GROUP(sb), i);
1557 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1559 * IF we have corrupt bitmap, we won't find any
1560 * free blocks even though group info says we
1561 * we have free blocks
1563 ext4_error(sb, __func__, "%d free blocks as per "
1564 "group info. But bitmap says 0\n",
1569 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1570 BUG_ON(ex.fe_len <= 0);
1571 if (free < ex.fe_len) {
1572 ext4_error(sb, __func__, "%d free blocks as per "
1573 "group info. But got %d blocks\n",
1576 * The number of free blocks differs. This mostly
1577 * indicate that the bitmap is corrupt. So exit
1578 * without claiming the space.
1583 ext4_mb_measure_extent(ac, &ex, e4b);
1589 ext4_mb_check_limits(ac, e4b, 1);
1593 * This is a special case for storages like raid5
1594 * we try to find stripe-aligned chunks for stripe-size requests
1595 * XXX should do so at least for multiples of stripe size as well
1597 static void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1598 struct ext4_buddy *e4b)
1600 struct super_block *sb = ac->ac_sb;
1601 struct ext4_sb_info *sbi = EXT4_SB(sb);
1602 void *bitmap = EXT4_MB_BITMAP(e4b);
1603 struct ext4_free_extent ex;
1604 ext4_fsblk_t first_group_block;
1609 BUG_ON(sbi->s_stripe == 0);
1611 /* find first stripe-aligned block in group */
1612 first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1613 + le32_to_cpu(sbi->s_es->s_first_data_block);
1614 a = first_group_block + sbi->s_stripe - 1;
1615 do_div(a, sbi->s_stripe);
1616 i = (a * sbi->s_stripe) - first_group_block;
1618 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1619 if (!mb_test_bit(i, bitmap)) {
1620 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1621 if (max >= sbi->s_stripe) {
1624 ext4_mb_use_best_found(ac, e4b);
1632 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1633 ext4_group_t group, int cr)
1635 unsigned free, fragments;
1637 struct ext4_group_desc *desc;
1638 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1640 BUG_ON(cr < 0 || cr >= 4);
1641 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1643 free = grp->bb_free;
1644 fragments = grp->bb_fragments;
1652 BUG_ON(ac->ac_2order == 0);
1653 /* If this group is uninitialized, skip it initially */
1654 desc = ext4_get_group_desc(ac->ac_sb, group, NULL);
1655 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
1658 bits = ac->ac_sb->s_blocksize_bits + 1;
1659 for (i = ac->ac_2order; i <= bits; i++)
1660 if (grp->bb_counters[i] > 0)
1664 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1668 if (free >= ac->ac_g_ex.fe_len)
1680 static noinline_for_stack int
1681 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1688 struct ext4_sb_info *sbi;
1689 struct super_block *sb;
1690 struct ext4_buddy e4b;
1695 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1697 /* first, try the goal */
1698 err = ext4_mb_find_by_goal(ac, &e4b);
1699 if (err || ac->ac_status == AC_STATUS_FOUND)
1702 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1706 * ac->ac2_order is set only if the fe_len is a power of 2
1707 * if ac2_order is set we also set criteria to 0 so that we
1708 * try exact allocation using buddy.
1710 i = fls(ac->ac_g_ex.fe_len);
1713 * We search using buddy data only if the order of the request
1714 * is greater than equal to the sbi_s_mb_order2_reqs
1715 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1717 if (i >= sbi->s_mb_order2_reqs) {
1719 * This should tell if fe_len is exactly power of 2
1721 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1722 ac->ac_2order = i - 1;
1725 bsbits = ac->ac_sb->s_blocksize_bits;
1726 /* if stream allocation is enabled, use global goal */
1727 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
1728 isize = i_size_read(ac->ac_inode) >> bsbits;
1732 if (size < sbi->s_mb_stream_request &&
1733 (ac->ac_flags & EXT4_MB_HINT_DATA)) {
1734 /* TBD: may be hot point */
1735 spin_lock(&sbi->s_md_lock);
1736 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1737 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1738 spin_unlock(&sbi->s_md_lock);
1740 /* Let's just scan groups to find more-less suitable blocks */
1741 cr = ac->ac_2order ? 0 : 1;
1743 * cr == 0 try to get exact allocation,
1744 * cr == 3 try to get anything
1747 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1748 ac->ac_criteria = cr;
1750 * searching for the right group start
1751 * from the goal value specified
1753 group = ac->ac_g_ex.fe_group;
1755 for (i = 0; i < EXT4_SB(sb)->s_groups_count; group++, i++) {
1756 struct ext4_group_info *grp;
1757 struct ext4_group_desc *desc;
1759 if (group == EXT4_SB(sb)->s_groups_count)
1762 /* quick check to skip empty groups */
1763 grp = ext4_get_group_info(ac->ac_sb, group);
1764 if (grp->bb_free == 0)
1768 * if the group is already init we check whether it is
1769 * a good group and if not we don't load the buddy
1771 if (EXT4_MB_GRP_NEED_INIT(grp)) {
1773 * we need full data about the group
1774 * to make a good selection
1776 err = ext4_mb_load_buddy(sb, group, &e4b);
1779 ext4_mb_release_desc(&e4b);
1783 * If the particular group doesn't satisfy our
1784 * criteria we continue with the next group
1786 if (!ext4_mb_good_group(ac, group, cr))
1789 err = ext4_mb_load_buddy(sb, group, &e4b);
1793 ext4_lock_group(sb, group);
1794 if (!ext4_mb_good_group(ac, group, cr)) {
1795 /* someone did allocation from this group */
1796 ext4_unlock_group(sb, group);
1797 ext4_mb_release_desc(&e4b);
1801 ac->ac_groups_scanned++;
1802 desc = ext4_get_group_desc(sb, group, NULL);
1803 if (cr == 0 || (desc->bg_flags &
1804 cpu_to_le16(EXT4_BG_BLOCK_UNINIT) &&
1805 ac->ac_2order != 0))
1806 ext4_mb_simple_scan_group(ac, &e4b);
1808 ac->ac_g_ex.fe_len == sbi->s_stripe)
1809 ext4_mb_scan_aligned(ac, &e4b);
1811 ext4_mb_complex_scan_group(ac, &e4b);
1813 ext4_unlock_group(sb, group);
1814 ext4_mb_release_desc(&e4b);
1816 if (ac->ac_status != AC_STATUS_CONTINUE)
1821 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
1822 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1824 * We've been searching too long. Let's try to allocate
1825 * the best chunk we've found so far
1828 ext4_mb_try_best_found(ac, &e4b);
1829 if (ac->ac_status != AC_STATUS_FOUND) {
1831 * Someone more lucky has already allocated it.
1832 * The only thing we can do is just take first
1834 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
1836 ac->ac_b_ex.fe_group = 0;
1837 ac->ac_b_ex.fe_start = 0;
1838 ac->ac_b_ex.fe_len = 0;
1839 ac->ac_status = AC_STATUS_CONTINUE;
1840 ac->ac_flags |= EXT4_MB_HINT_FIRST;
1842 atomic_inc(&sbi->s_mb_lost_chunks);
1850 #ifdef EXT4_MB_HISTORY
1851 struct ext4_mb_proc_session {
1852 struct ext4_mb_history *history;
1853 struct super_block *sb;
1858 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
1859 struct ext4_mb_history *hs,
1862 if (hs == s->history + s->max)
1864 if (!first && hs == s->history + s->start)
1866 while (hs->orig.fe_len == 0) {
1868 if (hs == s->history + s->max)
1870 if (hs == s->history + s->start)
1876 static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
1878 struct ext4_mb_proc_session *s = seq->private;
1879 struct ext4_mb_history *hs;
1883 return SEQ_START_TOKEN;
1884 hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
1887 while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
1891 static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
1894 struct ext4_mb_proc_session *s = seq->private;
1895 struct ext4_mb_history *hs = v;
1898 if (v == SEQ_START_TOKEN)
1899 return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
1901 return ext4_mb_history_skip_empty(s, ++hs, 0);
1904 static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
1906 char buf[25], buf2[25], buf3[25], *fmt;
1907 struct ext4_mb_history *hs = v;
1909 if (v == SEQ_START_TOKEN) {
1910 seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
1911 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
1912 "pid", "inode", "original", "goal", "result", "found",
1913 "grps", "cr", "flags", "merge", "tail", "broken");
1917 if (hs->op == EXT4_MB_HISTORY_ALLOC) {
1918 fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
1919 "%-5u %-5s %-5u %-6u\n";
1920 sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
1921 hs->result.fe_start, hs->result.fe_len,
1922 hs->result.fe_logical);
1923 sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
1924 hs->orig.fe_start, hs->orig.fe_len,
1925 hs->orig.fe_logical);
1926 sprintf(buf3, "%lu/%d/%u@%u", hs->goal.fe_group,
1927 hs->goal.fe_start, hs->goal.fe_len,
1928 hs->goal.fe_logical);
1929 seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
1930 hs->found, hs->groups, hs->cr, hs->flags,
1931 hs->merged ? "M" : "", hs->tail,
1932 hs->buddy ? 1 << hs->buddy : 0);
1933 } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
1934 fmt = "%-5u %-8u %-23s %-23s %-23s\n";
1935 sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
1936 hs->result.fe_start, hs->result.fe_len,
1937 hs->result.fe_logical);
1938 sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
1939 hs->orig.fe_start, hs->orig.fe_len,
1940 hs->orig.fe_logical);
1941 seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
1942 } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
1943 sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
1944 hs->result.fe_start, hs->result.fe_len);
1945 seq_printf(seq, "%-5u %-8u %-23s discard\n",
1946 hs->pid, hs->ino, buf2);
1947 } else if (hs->op == EXT4_MB_HISTORY_FREE) {
1948 sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
1949 hs->result.fe_start, hs->result.fe_len);
1950 seq_printf(seq, "%-5u %-8u %-23s free\n",
1951 hs->pid, hs->ino, buf2);
1956 static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
1960 static struct seq_operations ext4_mb_seq_history_ops = {
1961 .start = ext4_mb_seq_history_start,
1962 .next = ext4_mb_seq_history_next,
1963 .stop = ext4_mb_seq_history_stop,
1964 .show = ext4_mb_seq_history_show,
1967 static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
1969 struct super_block *sb = PDE(inode)->data;
1970 struct ext4_sb_info *sbi = EXT4_SB(sb);
1971 struct ext4_mb_proc_session *s;
1975 s = kmalloc(sizeof(*s), GFP_KERNEL);
1979 size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
1980 s->history = kmalloc(size, GFP_KERNEL);
1981 if (s->history == NULL) {
1986 spin_lock(&sbi->s_mb_history_lock);
1987 memcpy(s->history, sbi->s_mb_history, size);
1988 s->max = sbi->s_mb_history_max;
1989 s->start = sbi->s_mb_history_cur % s->max;
1990 spin_unlock(&sbi->s_mb_history_lock);
1992 rc = seq_open(file, &ext4_mb_seq_history_ops);
1994 struct seq_file *m = (struct seq_file *)file->private_data;
2004 static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
2006 struct seq_file *seq = (struct seq_file *)file->private_data;
2007 struct ext4_mb_proc_session *s = seq->private;
2010 return seq_release(inode, file);
2013 static ssize_t ext4_mb_seq_history_write(struct file *file,
2014 const char __user *buffer,
2015 size_t count, loff_t *ppos)
2017 struct seq_file *seq = (struct seq_file *)file->private_data;
2018 struct ext4_mb_proc_session *s = seq->private;
2019 struct super_block *sb = s->sb;
2023 if (count >= sizeof(str)) {
2024 printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
2025 "mb_history", (int)sizeof(str));
2029 if (copy_from_user(str, buffer, count))
2032 value = simple_strtol(str, NULL, 0);
2035 EXT4_SB(sb)->s_mb_history_filter = value;
2040 static struct file_operations ext4_mb_seq_history_fops = {
2041 .owner = THIS_MODULE,
2042 .open = ext4_mb_seq_history_open,
2044 .write = ext4_mb_seq_history_write,
2045 .llseek = seq_lseek,
2046 .release = ext4_mb_seq_history_release,
2049 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2051 struct super_block *sb = seq->private;
2052 struct ext4_sb_info *sbi = EXT4_SB(sb);
2055 if (*pos < 0 || *pos >= sbi->s_groups_count)
2059 return (void *) group;
2062 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2064 struct super_block *sb = seq->private;
2065 struct ext4_sb_info *sbi = EXT4_SB(sb);
2069 if (*pos < 0 || *pos >= sbi->s_groups_count)
2072 return (void *) group;;
2075 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2077 struct super_block *sb = seq->private;
2078 long group = (long) v;
2081 struct ext4_buddy e4b;
2083 struct ext4_group_info info;
2084 unsigned short counters[16];
2089 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2090 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2091 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2092 "group", "free", "frags", "first",
2093 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2094 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2096 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2097 sizeof(struct ext4_group_info);
2098 err = ext4_mb_load_buddy(sb, group, &e4b);
2100 seq_printf(seq, "#%-5lu: I/O error\n", group);
2103 ext4_lock_group(sb, group);
2104 memcpy(&sg, ext4_get_group_info(sb, group), i);
2105 ext4_unlock_group(sb, group);
2106 ext4_mb_release_desc(&e4b);
2108 seq_printf(seq, "#%-5lu: %-5u %-5u %-5u [", group, sg.info.bb_free,
2109 sg.info.bb_fragments, sg.info.bb_first_free);
2110 for (i = 0; i <= 13; i++)
2111 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2112 sg.info.bb_counters[i] : 0);
2113 seq_printf(seq, " ]\n");
2118 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2122 static struct seq_operations ext4_mb_seq_groups_ops = {
2123 .start = ext4_mb_seq_groups_start,
2124 .next = ext4_mb_seq_groups_next,
2125 .stop = ext4_mb_seq_groups_stop,
2126 .show = ext4_mb_seq_groups_show,
2129 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2131 struct super_block *sb = PDE(inode)->data;
2134 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2136 struct seq_file *m = (struct seq_file *)file->private_data;
2143 static struct file_operations ext4_mb_seq_groups_fops = {
2144 .owner = THIS_MODULE,
2145 .open = ext4_mb_seq_groups_open,
2147 .llseek = seq_lseek,
2148 .release = seq_release,
2151 static void ext4_mb_history_release(struct super_block *sb)
2153 struct ext4_sb_info *sbi = EXT4_SB(sb);
2155 remove_proc_entry("mb_groups", sbi->s_mb_proc);
2156 remove_proc_entry("mb_history", sbi->s_mb_proc);
2158 kfree(sbi->s_mb_history);
2161 static void ext4_mb_history_init(struct super_block *sb)
2163 struct ext4_sb_info *sbi = EXT4_SB(sb);
2166 if (sbi->s_mb_proc != NULL) {
2167 proc_create_data("mb_history", S_IRUGO, sbi->s_mb_proc,
2168 &ext4_mb_seq_history_fops, sb);
2169 proc_create_data("mb_groups", S_IRUGO, sbi->s_mb_proc,
2170 &ext4_mb_seq_groups_fops, sb);
2173 sbi->s_mb_history_max = 1000;
2174 sbi->s_mb_history_cur = 0;
2175 spin_lock_init(&sbi->s_mb_history_lock);
2176 i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
2177 sbi->s_mb_history = kmalloc(i, GFP_KERNEL);
2178 if (likely(sbi->s_mb_history != NULL))
2179 memset(sbi->s_mb_history, 0, i);
2180 /* if we can't allocate history, then we simple won't use it */
2183 static noinline_for_stack void
2184 ext4_mb_store_history(struct ext4_allocation_context *ac)
2186 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2187 struct ext4_mb_history h;
2189 if (unlikely(sbi->s_mb_history == NULL))
2192 if (!(ac->ac_op & sbi->s_mb_history_filter))
2196 h.pid = current->pid;
2197 h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2198 h.orig = ac->ac_o_ex;
2199 h.result = ac->ac_b_ex;
2200 h.flags = ac->ac_flags;
2201 h.found = ac->ac_found;
2202 h.groups = ac->ac_groups_scanned;
2203 h.cr = ac->ac_criteria;
2204 h.tail = ac->ac_tail;
2205 h.buddy = ac->ac_buddy;
2207 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
2208 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2209 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2211 h.goal = ac->ac_g_ex;
2212 h.result = ac->ac_f_ex;
2215 spin_lock(&sbi->s_mb_history_lock);
2216 memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2217 if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2218 sbi->s_mb_history_cur = 0;
2219 spin_unlock(&sbi->s_mb_history_lock);
2223 #define ext4_mb_history_release(sb)
2224 #define ext4_mb_history_init(sb)
2227 static int ext4_mb_init_backend(struct super_block *sb)
2230 int j, len, metalen;
2231 struct ext4_sb_info *sbi = EXT4_SB(sb);
2232 int num_meta_group_infos =
2233 (sbi->s_groups_count + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2234 EXT4_DESC_PER_BLOCK_BITS(sb);
2235 struct ext4_group_info **meta_group_info;
2237 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2238 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2239 * So a two level scheme suffices for now. */
2240 sbi->s_group_info = kmalloc(sizeof(*sbi->s_group_info) *
2241 num_meta_group_infos, GFP_KERNEL);
2242 if (sbi->s_group_info == NULL) {
2243 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2246 sbi->s_buddy_cache = new_inode(sb);
2247 if (sbi->s_buddy_cache == NULL) {
2248 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2251 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2253 metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb);
2254 for (i = 0; i < num_meta_group_infos; i++) {
2255 if ((i + 1) == num_meta_group_infos)
2256 metalen = sizeof(*meta_group_info) *
2257 (sbi->s_groups_count -
2258 (i << EXT4_DESC_PER_BLOCK_BITS(sb)));
2259 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2260 if (meta_group_info == NULL) {
2261 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2265 sbi->s_group_info[i] = meta_group_info;
2269 * calculate needed size. if change bb_counters size,
2270 * don't forget about ext4_mb_generate_buddy()
2272 len = sizeof(struct ext4_group_info);
2273 len += sizeof(unsigned short) * (sb->s_blocksize_bits + 2);
2274 for (i = 0; i < sbi->s_groups_count; i++) {
2275 struct ext4_group_desc *desc;
2278 sbi->s_group_info[i >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2279 j = i & (EXT4_DESC_PER_BLOCK(sb) - 1);
2281 meta_group_info[j] = kzalloc(len, GFP_KERNEL);
2282 if (meta_group_info[j] == NULL) {
2283 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2286 desc = ext4_get_group_desc(sb, i, NULL);
2289 "EXT4-fs: can't read descriptor %lu\n", i);
2293 memset(meta_group_info[j], 0, len);
2294 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2295 &(meta_group_info[j]->bb_state));
2298 * initialize bb_free to be able to skip
2299 * empty groups without initialization
2301 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2302 meta_group_info[j]->bb_free =
2303 ext4_free_blocks_after_init(sb, i, desc);
2305 meta_group_info[j]->bb_free =
2306 le16_to_cpu(desc->bg_free_blocks_count);
2309 INIT_LIST_HEAD(&meta_group_info[j]->bb_prealloc_list);
2313 struct buffer_head *bh;
2314 meta_group_info[j]->bb_bitmap =
2315 kmalloc(sb->s_blocksize, GFP_KERNEL);
2316 BUG_ON(meta_group_info[j]->bb_bitmap == NULL);
2317 bh = read_block_bitmap(sb, i);
2319 memcpy(meta_group_info[j]->bb_bitmap, bh->b_data,
2331 kfree(ext4_get_group_info(sb, i));
2332 i = num_meta_group_infos;
2335 kfree(sbi->s_group_info[i]);
2336 iput(sbi->s_buddy_cache);
2338 kfree(sbi->s_group_info);
2342 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2344 struct ext4_sb_info *sbi = EXT4_SB(sb);
2349 if (!test_opt(sb, MBALLOC))
2352 i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
2354 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2355 if (sbi->s_mb_offsets == NULL) {
2356 clear_opt(sbi->s_mount_opt, MBALLOC);
2359 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2360 if (sbi->s_mb_maxs == NULL) {
2361 clear_opt(sbi->s_mount_opt, MBALLOC);
2362 kfree(sbi->s_mb_maxs);
2366 /* order 0 is regular bitmap */
2367 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2368 sbi->s_mb_offsets[0] = 0;
2372 max = sb->s_blocksize << 2;
2374 sbi->s_mb_offsets[i] = offset;
2375 sbi->s_mb_maxs[i] = max;
2376 offset += 1 << (sb->s_blocksize_bits - i);
2379 } while (i <= sb->s_blocksize_bits + 1);
2381 /* init file for buddy data */
2382 i = ext4_mb_init_backend(sb);
2384 clear_opt(sbi->s_mount_opt, MBALLOC);
2385 kfree(sbi->s_mb_offsets);
2386 kfree(sbi->s_mb_maxs);
2390 spin_lock_init(&sbi->s_md_lock);
2391 INIT_LIST_HEAD(&sbi->s_active_transaction);
2392 INIT_LIST_HEAD(&sbi->s_closed_transaction);
2393 INIT_LIST_HEAD(&sbi->s_committed_transaction);
2394 spin_lock_init(&sbi->s_bal_lock);
2396 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2397 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2398 sbi->s_mb_stats = MB_DEFAULT_STATS;
2399 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2400 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2401 sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
2402 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2404 i = sizeof(struct ext4_locality_group) * NR_CPUS;
2405 sbi->s_locality_groups = kmalloc(i, GFP_KERNEL);
2406 if (sbi->s_locality_groups == NULL) {
2407 clear_opt(sbi->s_mount_opt, MBALLOC);
2408 kfree(sbi->s_mb_offsets);
2409 kfree(sbi->s_mb_maxs);
2412 for (i = 0; i < NR_CPUS; i++) {
2413 struct ext4_locality_group *lg;
2414 lg = &sbi->s_locality_groups[i];
2415 mutex_init(&lg->lg_mutex);
2416 INIT_LIST_HEAD(&lg->lg_prealloc_list);
2417 spin_lock_init(&lg->lg_prealloc_lock);
2420 ext4_mb_init_per_dev_proc(sb);
2421 ext4_mb_history_init(sb);
2423 printk("EXT4-fs: mballoc enabled\n");
2427 /* need to called with ext4 group lock (ext4_lock_group) */
2428 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2430 struct ext4_prealloc_space *pa;
2431 struct list_head *cur, *tmp;
2434 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2435 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2436 list_del(&pa->pa_group_list);
2441 mb_debug("mballoc: %u PAs left\n", count);
2445 int ext4_mb_release(struct super_block *sb)
2448 int num_meta_group_infos;
2449 struct ext4_group_info *grinfo;
2450 struct ext4_sb_info *sbi = EXT4_SB(sb);
2452 if (!test_opt(sb, MBALLOC))
2455 /* release freed, non-committed blocks */
2456 spin_lock(&sbi->s_md_lock);
2457 list_splice_init(&sbi->s_closed_transaction,
2458 &sbi->s_committed_transaction);
2459 list_splice_init(&sbi->s_active_transaction,
2460 &sbi->s_committed_transaction);
2461 spin_unlock(&sbi->s_md_lock);
2462 ext4_mb_free_committed_blocks(sb);
2464 if (sbi->s_group_info) {
2465 for (i = 0; i < sbi->s_groups_count; i++) {
2466 grinfo = ext4_get_group_info(sb, i);
2468 kfree(grinfo->bb_bitmap);
2470 ext4_lock_group(sb, i);
2471 ext4_mb_cleanup_pa(grinfo);
2472 ext4_unlock_group(sb, i);
2475 num_meta_group_infos = (sbi->s_groups_count +
2476 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2477 EXT4_DESC_PER_BLOCK_BITS(sb);
2478 for (i = 0; i < num_meta_group_infos; i++)
2479 kfree(sbi->s_group_info[i]);
2480 kfree(sbi->s_group_info);
2482 kfree(sbi->s_mb_offsets);
2483 kfree(sbi->s_mb_maxs);
2484 if (sbi->s_buddy_cache)
2485 iput(sbi->s_buddy_cache);
2486 if (sbi->s_mb_stats) {
2488 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2489 atomic_read(&sbi->s_bal_allocated),
2490 atomic_read(&sbi->s_bal_reqs),
2491 atomic_read(&sbi->s_bal_success));
2493 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2494 "%u 2^N hits, %u breaks, %u lost\n",
2495 atomic_read(&sbi->s_bal_ex_scanned),
2496 atomic_read(&sbi->s_bal_goals),
2497 atomic_read(&sbi->s_bal_2orders),
2498 atomic_read(&sbi->s_bal_breaks),
2499 atomic_read(&sbi->s_mb_lost_chunks));
2501 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2502 sbi->s_mb_buddies_generated++,
2503 sbi->s_mb_generation_time);
2505 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2506 atomic_read(&sbi->s_mb_preallocated),
2507 atomic_read(&sbi->s_mb_discarded));
2510 kfree(sbi->s_locality_groups);
2512 ext4_mb_history_release(sb);
2513 ext4_mb_destroy_per_dev_proc(sb);
2518 static noinline_for_stack void
2519 ext4_mb_free_committed_blocks(struct super_block *sb)
2521 struct ext4_sb_info *sbi = EXT4_SB(sb);
2526 struct ext4_free_metadata *md;
2527 struct ext4_buddy e4b;
2529 if (list_empty(&sbi->s_committed_transaction))
2532 /* there is committed blocks to be freed yet */
2534 /* get next array of blocks */
2536 spin_lock(&sbi->s_md_lock);
2537 if (!list_empty(&sbi->s_committed_transaction)) {
2538 md = list_entry(sbi->s_committed_transaction.next,
2539 struct ext4_free_metadata, list);
2540 list_del(&md->list);
2542 spin_unlock(&sbi->s_md_lock);
2547 mb_debug("gonna free %u blocks in group %lu (0x%p):",
2548 md->num, md->group, md);
2550 err = ext4_mb_load_buddy(sb, md->group, &e4b);
2551 /* we expect to find existing buddy because it's pinned */
2554 /* there are blocks to put in buddy to make them really free */
2557 ext4_lock_group(sb, md->group);
2558 for (i = 0; i < md->num; i++) {
2559 mb_debug(" %u", md->blocks[i]);
2560 err = mb_free_blocks(NULL, &e4b, md->blocks[i], 1);
2564 ext4_unlock_group(sb, md->group);
2566 /* balance refcounts from ext4_mb_free_metadata() */
2567 page_cache_release(e4b.bd_buddy_page);
2568 page_cache_release(e4b.bd_bitmap_page);
2571 ext4_mb_release_desc(&e4b);
2575 mb_debug("freed %u blocks in %u structures\n", count, count2);
2578 #define EXT4_MB_STATS_NAME "stats"
2579 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2580 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2581 #define EXT4_MB_ORDER2_REQ "order2_req"
2582 #define EXT4_MB_STREAM_REQ "stream_req"
2583 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2587 #define MB_PROC_FOPS(name) \
2588 static int ext4_mb_##name##_proc_show(struct seq_file *m, void *v) \
2590 struct ext4_sb_info *sbi = m->private; \
2592 seq_printf(m, "%ld\n", sbi->s_mb_##name); \
2596 static int ext4_mb_##name##_proc_open(struct inode *inode, struct file *file)\
2598 return single_open(file, ext4_mb_##name##_proc_show, PDE(inode)->data);\
2601 static ssize_t ext4_mb_##name##_proc_write(struct file *file, \
2602 const char __user *buf, size_t cnt, loff_t *ppos) \
2604 struct ext4_sb_info *sbi = PDE(file->f_path.dentry->d_inode)->data;\
2607 if (cnt >= sizeof(str)) \
2609 if (copy_from_user(str, buf, cnt)) \
2611 value = simple_strtol(str, NULL, 0); \
2614 sbi->s_mb_##name = value; \
2618 static const struct file_operations ext4_mb_##name##_proc_fops = { \
2619 .owner = THIS_MODULE, \
2620 .open = ext4_mb_##name##_proc_open, \
2622 .llseek = seq_lseek, \
2623 .release = single_release, \
2624 .write = ext4_mb_##name##_proc_write, \
2627 MB_PROC_FOPS(stats);
2628 MB_PROC_FOPS(max_to_scan);
2629 MB_PROC_FOPS(min_to_scan);
2630 MB_PROC_FOPS(order2_reqs);
2631 MB_PROC_FOPS(stream_request);
2632 MB_PROC_FOPS(group_prealloc);
2634 #define MB_PROC_HANDLER(name, var) \
2636 proc = proc_create_data(name, mode, sbi->s_mb_proc, \
2637 &ext4_mb_##var##_proc_fops, sbi); \
2638 if (proc == NULL) { \
2639 printk(KERN_ERR "EXT4-fs: can't to create %s\n", name); \
2644 static int ext4_mb_init_per_dev_proc(struct super_block *sb)
2646 mode_t mode = S_IFREG | S_IRUGO | S_IWUSR;
2647 struct ext4_sb_info *sbi = EXT4_SB(sb);
2648 struct proc_dir_entry *proc;
2651 bdevname(sb->s_bdev, devname);
2652 sbi->s_mb_proc = proc_mkdir(devname, proc_root_ext4);
2654 MB_PROC_HANDLER(EXT4_MB_STATS_NAME, stats);
2655 MB_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME, max_to_scan);
2656 MB_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME, min_to_scan);
2657 MB_PROC_HANDLER(EXT4_MB_ORDER2_REQ, order2_reqs);
2658 MB_PROC_HANDLER(EXT4_MB_STREAM_REQ, stream_request);
2659 MB_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC, group_prealloc);
2664 printk(KERN_ERR "EXT4-fs: Unable to create %s\n", devname);
2665 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_mb_proc);
2666 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_mb_proc);
2667 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_mb_proc);
2668 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
2669 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
2670 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_mb_proc);
2671 remove_proc_entry(devname, proc_root_ext4);
2672 sbi->s_mb_proc = NULL;
2677 static int ext4_mb_destroy_per_dev_proc(struct super_block *sb)
2679 struct ext4_sb_info *sbi = EXT4_SB(sb);
2682 if (sbi->s_mb_proc == NULL)
2685 bdevname(sb->s_bdev, devname);
2686 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_mb_proc);
2687 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_mb_proc);
2688 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_mb_proc);
2689 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
2690 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
2691 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_mb_proc);
2692 remove_proc_entry(devname, proc_root_ext4);
2697 int __init init_ext4_mballoc(void)
2699 ext4_pspace_cachep =
2700 kmem_cache_create("ext4_prealloc_space",
2701 sizeof(struct ext4_prealloc_space),
2702 0, SLAB_RECLAIM_ACCOUNT, NULL);
2703 if (ext4_pspace_cachep == NULL)
2707 kmem_cache_create("ext4_alloc_context",
2708 sizeof(struct ext4_allocation_context),
2709 0, SLAB_RECLAIM_ACCOUNT, NULL);
2710 if (ext4_ac_cachep == NULL) {
2711 kmem_cache_destroy(ext4_pspace_cachep);
2714 #ifdef CONFIG_PROC_FS
2715 proc_root_ext4 = proc_mkdir("fs/ext4", NULL);
2716 if (proc_root_ext4 == NULL)
2717 printk(KERN_ERR "EXT4-fs: Unable to create fs/ext4\n");
2722 void exit_ext4_mballoc(void)
2724 /* XXX: synchronize_rcu(); */
2725 kmem_cache_destroy(ext4_pspace_cachep);
2726 kmem_cache_destroy(ext4_ac_cachep);
2727 #ifdef CONFIG_PROC_FS
2728 remove_proc_entry("fs/ext4", NULL);
2734 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2735 * Returns 0 if success or error code
2737 static noinline_for_stack int
2738 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2741 struct buffer_head *bitmap_bh = NULL;
2742 struct ext4_super_block *es;
2743 struct ext4_group_desc *gdp;
2744 struct buffer_head *gdp_bh;
2745 struct ext4_sb_info *sbi;
2746 struct super_block *sb;
2750 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2751 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2759 bitmap_bh = read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2763 err = ext4_journal_get_write_access(handle, bitmap_bh);
2768 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2772 ext4_debug("using block group %lu(%d)\n", ac->ac_b_ex.fe_group,
2773 gdp->bg_free_blocks_count);
2775 err = ext4_journal_get_write_access(handle, gdp_bh);
2779 block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
2780 + ac->ac_b_ex.fe_start
2781 + le32_to_cpu(es->s_first_data_block);
2783 len = ac->ac_b_ex.fe_len;
2784 if (in_range(ext4_block_bitmap(sb, gdp), block, len) ||
2785 in_range(ext4_inode_bitmap(sb, gdp), block, len) ||
2786 in_range(block, ext4_inode_table(sb, gdp),
2787 EXT4_SB(sb)->s_itb_per_group) ||
2788 in_range(block + len - 1, ext4_inode_table(sb, gdp),
2789 EXT4_SB(sb)->s_itb_per_group)) {
2790 ext4_error(sb, __func__,
2791 "Allocating block in system zone - block = %llu",
2793 /* File system mounted not to panic on error
2794 * Fix the bitmap and repeat the block allocation
2795 * We leak some of the blocks here.
2797 mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group),
2798 bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2799 ac->ac_b_ex.fe_len);
2800 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
2805 #ifdef AGGRESSIVE_CHECK
2808 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2809 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2810 bitmap_bh->b_data));
2814 mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), bitmap_bh->b_data,
2815 ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
2817 spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
2818 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2819 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2820 gdp->bg_free_blocks_count =
2821 cpu_to_le16(ext4_free_blocks_after_init(sb,
2822 ac->ac_b_ex.fe_group,
2825 le16_add_cpu(&gdp->bg_free_blocks_count, -ac->ac_b_ex.fe_len);
2826 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2827 spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
2828 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
2830 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
2833 err = ext4_journal_dirty_metadata(handle, gdp_bh);
2842 * here we normalize request for locality group
2843 * Group request are normalized to s_strip size if we set the same via mount
2844 * option. If not we set it to s_mb_group_prealloc which can be configured via
2845 * /proc/fs/ext4/<partition>/group_prealloc
2847 * XXX: should we try to preallocate more than the group has now?
2849 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2851 struct super_block *sb = ac->ac_sb;
2852 struct ext4_locality_group *lg = ac->ac_lg;
2855 if (EXT4_SB(sb)->s_stripe)
2856 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
2858 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2859 mb_debug("#%u: goal %u blocks for locality group\n",
2860 current->pid, ac->ac_g_ex.fe_len);
2864 * Normalization means making request better in terms of
2865 * size and alignment
2867 static noinline_for_stack void
2868 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2869 struct ext4_allocation_request *ar)
2873 loff_t size, orig_size, start_off;
2874 ext4_lblk_t start, orig_start;
2875 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2876 struct ext4_prealloc_space *pa;
2878 /* do normalize only data requests, metadata requests
2879 do not need preallocation */
2880 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2883 /* sometime caller may want exact blocks */
2884 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2887 /* caller may indicate that preallocation isn't
2888 * required (it's a tail, for example) */
2889 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2892 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2893 ext4_mb_normalize_group_request(ac);
2897 bsbits = ac->ac_sb->s_blocksize_bits;
2899 /* first, let's learn actual file size
2900 * given current request is allocated */
2901 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2902 size = size << bsbits;
2903 if (size < i_size_read(ac->ac_inode))
2904 size = i_size_read(ac->ac_inode);
2906 /* max size of free chunks */
2909 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2910 (req <= (size) || max <= (chunk_size))
2912 /* first, try to predict filesize */
2913 /* XXX: should this table be tunable? */
2915 if (size <= 16 * 1024) {
2917 } else if (size <= 32 * 1024) {
2919 } else if (size <= 64 * 1024) {
2921 } else if (size <= 128 * 1024) {
2923 } else if (size <= 256 * 1024) {
2925 } else if (size <= 512 * 1024) {
2927 } else if (size <= 1024 * 1024) {
2929 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2930 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2931 (21 - bsbits)) << 21;
2932 size = 2 * 1024 * 1024;
2933 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2934 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2935 (22 - bsbits)) << 22;
2936 size = 4 * 1024 * 1024;
2937 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2938 (8<<20)>>bsbits, max, 8 * 1024)) {
2939 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2940 (23 - bsbits)) << 23;
2941 size = 8 * 1024 * 1024;
2943 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2944 size = ac->ac_o_ex.fe_len << bsbits;
2946 orig_size = size = size >> bsbits;
2947 orig_start = start = start_off >> bsbits;
2949 /* don't cover already allocated blocks in selected range */
2950 if (ar->pleft && start <= ar->lleft) {
2951 size -= ar->lleft + 1 - start;
2952 start = ar->lleft + 1;
2954 if (ar->pright && start + size - 1 >= ar->lright)
2955 size -= start + size - ar->lright;
2959 /* check we don't cross already preallocated blocks */
2961 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2962 unsigned long pa_end;
2966 spin_lock(&pa->pa_lock);
2967 if (pa->pa_deleted) {
2968 spin_unlock(&pa->pa_lock);
2972 pa_end = pa->pa_lstart + pa->pa_len;
2974 /* PA must not overlap original request */
2975 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2976 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2978 /* skip PA normalized request doesn't overlap with */
2979 if (pa->pa_lstart >= end) {
2980 spin_unlock(&pa->pa_lock);
2983 if (pa_end <= start) {
2984 spin_unlock(&pa->pa_lock);
2987 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2989 if (pa_end <= ac->ac_o_ex.fe_logical) {
2990 BUG_ON(pa_end < start);
2994 if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
2995 BUG_ON(pa->pa_lstart > end);
2996 end = pa->pa_lstart;
2998 spin_unlock(&pa->pa_lock);
3003 /* XXX: extra loop to check we really don't overlap preallocations */
3005 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3006 unsigned long pa_end;
3007 spin_lock(&pa->pa_lock);
3008 if (pa->pa_deleted == 0) {
3009 pa_end = pa->pa_lstart + pa->pa_len;
3010 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3012 spin_unlock(&pa->pa_lock);
3016 if (start + size <= ac->ac_o_ex.fe_logical &&
3017 start > ac->ac_o_ex.fe_logical) {
3018 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3019 (unsigned long) start, (unsigned long) size,
3020 (unsigned long) ac->ac_o_ex.fe_logical);
3022 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3023 start > ac->ac_o_ex.fe_logical);
3024 BUG_ON(size <= 0 || size >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3026 /* now prepare goal request */
3028 /* XXX: is it better to align blocks WRT to logical
3029 * placement or satisfy big request as is */
3030 ac->ac_g_ex.fe_logical = start;
3031 ac->ac_g_ex.fe_len = size;
3033 /* define goal start in order to merge */
3034 if (ar->pright && (ar->lright == (start + size))) {
3035 /* merge to the right */
3036 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3037 &ac->ac_f_ex.fe_group,
3038 &ac->ac_f_ex.fe_start);
3039 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3041 if (ar->pleft && (ar->lleft + 1 == start)) {
3042 /* merge to the left */
3043 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3044 &ac->ac_f_ex.fe_group,
3045 &ac->ac_f_ex.fe_start);
3046 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3049 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
3050 (unsigned) orig_size, (unsigned) start);
3053 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3055 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3057 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3058 atomic_inc(&sbi->s_bal_reqs);
3059 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3060 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3061 atomic_inc(&sbi->s_bal_success);
3062 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3063 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3064 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3065 atomic_inc(&sbi->s_bal_goals);
3066 if (ac->ac_found > sbi->s_mb_max_to_scan)
3067 atomic_inc(&sbi->s_bal_breaks);
3070 ext4_mb_store_history(ac);
3074 * use blocks preallocated to inode
3076 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3077 struct ext4_prealloc_space *pa)
3083 /* found preallocated blocks, use them */
3084 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3085 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3087 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3088 &ac->ac_b_ex.fe_start);
3089 ac->ac_b_ex.fe_len = len;
3090 ac->ac_status = AC_STATUS_FOUND;
3093 BUG_ON(start < pa->pa_pstart);
3094 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3095 BUG_ON(pa->pa_free < len);
3098 mb_debug("use %llu/%u from inode pa %p\n", start, len, pa);
3102 * use blocks preallocated to locality group
3104 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3105 struct ext4_prealloc_space *pa)
3107 unsigned int len = ac->ac_o_ex.fe_len;
3108 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3109 &ac->ac_b_ex.fe_group,
3110 &ac->ac_b_ex.fe_start);
3111 ac->ac_b_ex.fe_len = len;
3112 ac->ac_status = AC_STATUS_FOUND;
3115 /* we don't correct pa_pstart or pa_plen here to avoid
3116 * possible race when the group is being loaded concurrently
3117 * instead we correct pa later, after blocks are marked
3118 * in on-disk bitmap -- see ext4_mb_release_context()
3119 * Other CPUs are prevented from allocating from this pa by lg_mutex
3121 mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3125 * search goal blocks in preallocated space
3127 static noinline_for_stack int
3128 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3130 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3131 struct ext4_locality_group *lg;
3132 struct ext4_prealloc_space *pa;
3134 /* only data can be preallocated */
3135 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3138 /* first, try per-file preallocation */
3140 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3142 /* all fields in this condition don't change,
3143 * so we can skip locking for them */
3144 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3145 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3148 /* found preallocated blocks, use them */
3149 spin_lock(&pa->pa_lock);
3150 if (pa->pa_deleted == 0 && pa->pa_free) {
3151 atomic_inc(&pa->pa_count);
3152 ext4_mb_use_inode_pa(ac, pa);
3153 spin_unlock(&pa->pa_lock);
3154 ac->ac_criteria = 10;
3158 spin_unlock(&pa->pa_lock);
3162 /* can we use group allocation? */
3163 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3166 /* inode may have no locality group for some reason */
3172 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list, pa_inode_list) {
3173 spin_lock(&pa->pa_lock);
3174 if (pa->pa_deleted == 0 && pa->pa_free >= ac->ac_o_ex.fe_len) {
3175 atomic_inc(&pa->pa_count);
3176 ext4_mb_use_group_pa(ac, pa);
3177 spin_unlock(&pa->pa_lock);
3178 ac->ac_criteria = 20;
3182 spin_unlock(&pa->pa_lock);
3190 * the function goes through all preallocation in this group and marks them
3191 * used in in-core bitmap. buddy must be generated from this bitmap
3192 * Need to be called with ext4 group lock (ext4_lock_group)
3194 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3197 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3198 struct ext4_prealloc_space *pa;
3199 struct list_head *cur;
3200 ext4_group_t groupnr;
3201 ext4_grpblk_t start;
3202 int preallocated = 0;
3206 /* all form of preallocation discards first load group,
3207 * so the only competing code is preallocation use.
3208 * we don't need any locking here
3209 * notice we do NOT ignore preallocations with pa_deleted
3210 * otherwise we could leave used blocks available for
3211 * allocation in buddy when concurrent ext4_mb_put_pa()
3212 * is dropping preallocation
3214 list_for_each(cur, &grp->bb_prealloc_list) {
3215 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3216 spin_lock(&pa->pa_lock);
3217 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3220 spin_unlock(&pa->pa_lock);
3221 if (unlikely(len == 0))
3223 BUG_ON(groupnr != group);
3224 mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
3225 bitmap, start, len);
3226 preallocated += len;
3229 mb_debug("prellocated %u for group %lu\n", preallocated, group);
3232 static void ext4_mb_pa_callback(struct rcu_head *head)
3234 struct ext4_prealloc_space *pa;
3235 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3236 kmem_cache_free(ext4_pspace_cachep, pa);
3240 * drops a reference to preallocated space descriptor
3241 * if this was the last reference and the space is consumed
3243 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3244 struct super_block *sb, struct ext4_prealloc_space *pa)
3248 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3251 /* in this short window concurrent discard can set pa_deleted */
3252 spin_lock(&pa->pa_lock);
3253 if (pa->pa_deleted == 1) {
3254 spin_unlock(&pa->pa_lock);
3259 spin_unlock(&pa->pa_lock);
3261 /* -1 is to protect from crossing allocation group */
3262 ext4_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL);
3267 * P1 (buddy init) P2 (regular allocation)
3268 * find block B in PA
3269 * copy on-disk bitmap to buddy
3270 * mark B in on-disk bitmap
3271 * drop PA from group
3272 * mark all PAs in buddy
3274 * thus, P1 initializes buddy with B available. to prevent this
3275 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3278 ext4_lock_group(sb, grp);
3279 list_del(&pa->pa_group_list);
3280 ext4_unlock_group(sb, grp);
3282 spin_lock(pa->pa_obj_lock);
3283 list_del_rcu(&pa->pa_inode_list);
3284 spin_unlock(pa->pa_obj_lock);
3286 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3290 * creates new preallocated space for given inode
3292 static noinline_for_stack int
3293 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3295 struct super_block *sb = ac->ac_sb;
3296 struct ext4_prealloc_space *pa;
3297 struct ext4_group_info *grp;
3298 struct ext4_inode_info *ei;
3300 /* preallocate only when found space is larger then requested */
3301 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3302 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3303 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3305 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3309 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3315 /* we can't allocate as much as normalizer wants.
3316 * so, found space must get proper lstart
3317 * to cover original request */
3318 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3319 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3321 /* we're limited by original request in that
3322 * logical block must be covered any way
3323 * winl is window we can move our chunk within */
3324 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3326 /* also, we should cover whole original request */
3327 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3329 /* the smallest one defines real window */
3330 win = min(winl, wins);
3332 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3333 if (offs && offs < win)
3336 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3337 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3338 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3341 /* preallocation can change ac_b_ex, thus we store actually
3342 * allocated blocks for history */
3343 ac->ac_f_ex = ac->ac_b_ex;
3345 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3346 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3347 pa->pa_len = ac->ac_b_ex.fe_len;
3348 pa->pa_free = pa->pa_len;
3349 atomic_set(&pa->pa_count, 1);
3350 spin_lock_init(&pa->pa_lock);
3354 mb_debug("new inode pa %p: %llu/%u for %u\n", pa,
3355 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3357 ext4_mb_use_inode_pa(ac, pa);
3358 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3360 ei = EXT4_I(ac->ac_inode);
3361 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3363 pa->pa_obj_lock = &ei->i_prealloc_lock;
3364 pa->pa_inode = ac->ac_inode;
3366 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3367 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3368 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3370 spin_lock(pa->pa_obj_lock);
3371 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3372 spin_unlock(pa->pa_obj_lock);
3378 * creates new preallocated space for locality group inodes belongs to
3380 static noinline_for_stack int
3381 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3383 struct super_block *sb = ac->ac_sb;
3384 struct ext4_locality_group *lg;
3385 struct ext4_prealloc_space *pa;
3386 struct ext4_group_info *grp;
3388 /* preallocate only when found space is larger then requested */
3389 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3390 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3391 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3393 BUG_ON(ext4_pspace_cachep == NULL);
3394 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3398 /* preallocation can change ac_b_ex, thus we store actually
3399 * allocated blocks for history */
3400 ac->ac_f_ex = ac->ac_b_ex;
3402 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3403 pa->pa_lstart = pa->pa_pstart;
3404 pa->pa_len = ac->ac_b_ex.fe_len;
3405 pa->pa_free = pa->pa_len;
3406 atomic_set(&pa->pa_count, 1);
3407 spin_lock_init(&pa->pa_lock);
3411 mb_debug("new group pa %p: %llu/%u for %u\n", pa,
3412 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3414 ext4_mb_use_group_pa(ac, pa);
3415 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3417 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3421 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3422 pa->pa_inode = NULL;
3424 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3425 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3426 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3428 spin_lock(pa->pa_obj_lock);
3429 list_add_tail_rcu(&pa->pa_inode_list, &lg->lg_prealloc_list);
3430 spin_unlock(pa->pa_obj_lock);
3435 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3439 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3440 err = ext4_mb_new_group_pa(ac);
3442 err = ext4_mb_new_inode_pa(ac);
3447 * finds all unused blocks in on-disk bitmap, frees them in
3448 * in-core bitmap and buddy.
3449 * @pa must be unlinked from inode and group lists, so that
3450 * nobody else can find/use it.
3451 * the caller MUST hold group/inode locks.
3452 * TODO: optimize the case when there are no in-core structures yet
3454 static noinline_for_stack int
3455 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3456 struct ext4_prealloc_space *pa,
3457 struct ext4_allocation_context *ac)
3459 struct super_block *sb = e4b->bd_sb;
3460 struct ext4_sb_info *sbi = EXT4_SB(sb);
3469 BUG_ON(pa->pa_deleted == 0);
3470 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3471 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3472 end = bit + pa->pa_len;
3476 ac->ac_inode = pa->pa_inode;
3477 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3481 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3484 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3485 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3486 le32_to_cpu(sbi->s_es->s_first_data_block);
3487 mb_debug(" free preallocated %u/%u in group %u\n",
3488 (unsigned) start, (unsigned) next - bit,
3493 ac->ac_b_ex.fe_group = group;
3494 ac->ac_b_ex.fe_start = bit;
3495 ac->ac_b_ex.fe_len = next - bit;
3496 ac->ac_b_ex.fe_logical = 0;
3497 ext4_mb_store_history(ac);
3500 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3503 if (free != pa->pa_free) {
3504 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3505 pa, (unsigned long) pa->pa_lstart,
3506 (unsigned long) pa->pa_pstart,
3507 (unsigned long) pa->pa_len);
3508 ext4_error(sb, __func__, "free %u, pa_free %u\n",
3511 * pa is already deleted so we use the value obtained
3512 * from the bitmap and continue.
3515 atomic_add(free, &sbi->s_mb_discarded);
3520 static noinline_for_stack int
3521 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3522 struct ext4_prealloc_space *pa,
3523 struct ext4_allocation_context *ac)
3525 struct super_block *sb = e4b->bd_sb;
3530 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3532 BUG_ON(pa->pa_deleted == 0);
3533 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3534 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3535 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3536 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3540 ac->ac_inode = NULL;
3541 ac->ac_b_ex.fe_group = group;
3542 ac->ac_b_ex.fe_start = bit;
3543 ac->ac_b_ex.fe_len = pa->pa_len;
3544 ac->ac_b_ex.fe_logical = 0;
3545 ext4_mb_store_history(ac);
3552 * releases all preallocations in given group
3554 * first, we need to decide discard policy:
3555 * - when do we discard
3557 * - how many do we discard
3558 * 1) how many requested
3560 static noinline_for_stack int
3561 ext4_mb_discard_group_preallocations(struct super_block *sb,
3562 ext4_group_t group, int needed)
3564 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3565 struct buffer_head *bitmap_bh = NULL;
3566 struct ext4_prealloc_space *pa, *tmp;
3567 struct ext4_allocation_context *ac;
3568 struct list_head list;
3569 struct ext4_buddy e4b;
3574 mb_debug("discard preallocation for group %lu\n", group);
3576 if (list_empty(&grp->bb_prealloc_list))
3579 bitmap_bh = read_block_bitmap(sb, group);
3580 if (bitmap_bh == NULL) {
3581 /* error handling here */
3582 ext4_mb_release_desc(&e4b);
3583 BUG_ON(bitmap_bh == NULL);
3586 err = ext4_mb_load_buddy(sb, group, &e4b);
3587 BUG_ON(err != 0); /* error handling here */
3590 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3592 grp = ext4_get_group_info(sb, group);
3593 INIT_LIST_HEAD(&list);
3595 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3597 ext4_lock_group(sb, group);
3598 list_for_each_entry_safe(pa, tmp,
3599 &grp->bb_prealloc_list, pa_group_list) {
3600 spin_lock(&pa->pa_lock);
3601 if (atomic_read(&pa->pa_count)) {
3602 spin_unlock(&pa->pa_lock);
3606 if (pa->pa_deleted) {
3607 spin_unlock(&pa->pa_lock);
3611 /* seems this one can be freed ... */
3614 /* we can trust pa_free ... */
3615 free += pa->pa_free;
3617 spin_unlock(&pa->pa_lock);
3619 list_del(&pa->pa_group_list);
3620 list_add(&pa->u.pa_tmp_list, &list);
3623 /* if we still need more blocks and some PAs were used, try again */
3624 if (free < needed && busy) {
3626 ext4_unlock_group(sb, group);
3628 * Yield the CPU here so that we don't get soft lockup
3629 * in non preempt case.
3635 /* found anything to free? */
3636 if (list_empty(&list)) {
3641 /* now free all selected PAs */
3642 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3644 /* remove from object (inode or locality group) */
3645 spin_lock(pa->pa_obj_lock);
3646 list_del_rcu(&pa->pa_inode_list);
3647 spin_unlock(pa->pa_obj_lock);
3650 ext4_mb_release_group_pa(&e4b, pa, ac);
3652 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3654 list_del(&pa->u.pa_tmp_list);
3655 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3659 ext4_unlock_group(sb, group);
3661 kmem_cache_free(ext4_ac_cachep, ac);
3662 ext4_mb_release_desc(&e4b);
3668 * releases all non-used preallocated blocks for given inode
3670 * It's important to discard preallocations under i_data_sem
3671 * We don't want another block to be served from the prealloc
3672 * space when we are discarding the inode prealloc space.
3674 * FIXME!! Make sure it is valid at all the call sites
3676 void ext4_mb_discard_inode_preallocations(struct inode *inode)
3678 struct ext4_inode_info *ei = EXT4_I(inode);
3679 struct super_block *sb = inode->i_sb;
3680 struct buffer_head *bitmap_bh = NULL;
3681 struct ext4_prealloc_space *pa, *tmp;
3682 struct ext4_allocation_context *ac;
3683 ext4_group_t group = 0;
3684 struct list_head list;
3685 struct ext4_buddy e4b;
3688 if (!test_opt(sb, MBALLOC) || !S_ISREG(inode->i_mode)) {
3689 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3693 mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
3695 INIT_LIST_HEAD(&list);
3697 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3699 /* first, collect all pa's in the inode */
3700 spin_lock(&ei->i_prealloc_lock);
3701 while (!list_empty(&ei->i_prealloc_list)) {
3702 pa = list_entry(ei->i_prealloc_list.next,
3703 struct ext4_prealloc_space, pa_inode_list);
3704 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3705 spin_lock(&pa->pa_lock);
3706 if (atomic_read(&pa->pa_count)) {
3707 /* this shouldn't happen often - nobody should
3708 * use preallocation while we're discarding it */
3709 spin_unlock(&pa->pa_lock);
3710 spin_unlock(&ei->i_prealloc_lock);
3711 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3713 schedule_timeout_uninterruptible(HZ);
3717 if (pa->pa_deleted == 0) {
3719 spin_unlock(&pa->pa_lock);
3720 list_del_rcu(&pa->pa_inode_list);
3721 list_add(&pa->u.pa_tmp_list, &list);
3725 /* someone is deleting pa right now */
3726 spin_unlock(&pa->pa_lock);
3727 spin_unlock(&ei->i_prealloc_lock);
3729 /* we have to wait here because pa_deleted
3730 * doesn't mean pa is already unlinked from
3731 * the list. as we might be called from
3732 * ->clear_inode() the inode will get freed
3733 * and concurrent thread which is unlinking
3734 * pa from inode's list may access already
3735 * freed memory, bad-bad-bad */
3737 /* XXX: if this happens too often, we can
3738 * add a flag to force wait only in case
3739 * of ->clear_inode(), but not in case of
3740 * regular truncate */
3741 schedule_timeout_uninterruptible(HZ);
3744 spin_unlock(&ei->i_prealloc_lock);
3746 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3747 BUG_ON(pa->pa_linear != 0);
3748 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3750 err = ext4_mb_load_buddy(sb, group, &e4b);
3751 BUG_ON(err != 0); /* error handling here */
3753 bitmap_bh = read_block_bitmap(sb, group);
3754 if (bitmap_bh == NULL) {
3755 /* error handling here */
3756 ext4_mb_release_desc(&e4b);
3757 BUG_ON(bitmap_bh == NULL);
3760 ext4_lock_group(sb, group);
3761 list_del(&pa->pa_group_list);
3762 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3763 ext4_unlock_group(sb, group);
3765 ext4_mb_release_desc(&e4b);
3768 list_del(&pa->u.pa_tmp_list);
3769 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3772 kmem_cache_free(ext4_ac_cachep, ac);
3776 * finds all preallocated spaces and return blocks being freed to them
3777 * if preallocated space becomes full (no block is used from the space)
3778 * then the function frees space in buddy
3779 * XXX: at the moment, truncate (which is the only way to free blocks)
3780 * discards all preallocations
3782 static void ext4_mb_return_to_preallocation(struct inode *inode,
3783 struct ext4_buddy *e4b,
3784 sector_t block, int count)
3786 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
3789 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3791 struct super_block *sb = ac->ac_sb;
3794 printk(KERN_ERR "EXT4-fs: Can't allocate:"
3795 " Allocation context details:\n");
3796 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
3797 ac->ac_status, ac->ac_flags);
3798 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3799 "best %lu/%lu/%lu@%lu cr %d\n",
3800 (unsigned long)ac->ac_o_ex.fe_group,
3801 (unsigned long)ac->ac_o_ex.fe_start,
3802 (unsigned long)ac->ac_o_ex.fe_len,
3803 (unsigned long)ac->ac_o_ex.fe_logical,
3804 (unsigned long)ac->ac_g_ex.fe_group,
3805 (unsigned long)ac->ac_g_ex.fe_start,
3806 (unsigned long)ac->ac_g_ex.fe_len,
3807 (unsigned long)ac->ac_g_ex.fe_logical,
3808 (unsigned long)ac->ac_b_ex.fe_group,
3809 (unsigned long)ac->ac_b_ex.fe_start,
3810 (unsigned long)ac->ac_b_ex.fe_len,
3811 (unsigned long)ac->ac_b_ex.fe_logical,
3812 (int)ac->ac_criteria);
3813 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
3815 printk(KERN_ERR "EXT4-fs: groups: \n");
3816 for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
3817 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3818 struct ext4_prealloc_space *pa;
3819 ext4_grpblk_t start;
3820 struct list_head *cur;
3821 ext4_lock_group(sb, i);
3822 list_for_each(cur, &grp->bb_prealloc_list) {
3823 pa = list_entry(cur, struct ext4_prealloc_space,
3825 spin_lock(&pa->pa_lock);
3826 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3828 spin_unlock(&pa->pa_lock);
3829 printk(KERN_ERR "PA:%lu:%d:%u \n", i,
3832 ext4_unlock_group(sb, i);
3834 if (grp->bb_free == 0)
3836 printk(KERN_ERR "%lu: %d/%d \n",
3837 i, grp->bb_free, grp->bb_fragments);
3839 printk(KERN_ERR "\n");
3842 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3849 * We use locality group preallocation for small size file. The size of the
3850 * file is determined by the current size or the resulting size after
3851 * allocation which ever is larger
3853 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
3855 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3857 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3858 int bsbits = ac->ac_sb->s_blocksize_bits;
3861 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3864 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3865 isize = i_size_read(ac->ac_inode) >> bsbits;
3866 size = max(size, isize);
3868 /* don't use group allocation for large files */
3869 if (size >= sbi->s_mb_stream_request)
3872 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3875 BUG_ON(ac->ac_lg != NULL);
3877 * locality group prealloc space are per cpu. The reason for having
3878 * per cpu locality group is to reduce the contention between block
3879 * request from multiple CPUs.
3881 ac->ac_lg = &sbi->s_locality_groups[get_cpu()];
3884 /* we're going to use group allocation */
3885 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3887 /* serialize all allocations in the group */
3888 mutex_lock(&ac->ac_lg->lg_mutex);
3891 static noinline_for_stack int
3892 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
3893 struct ext4_allocation_request *ar)
3895 struct super_block *sb = ar->inode->i_sb;
3896 struct ext4_sb_info *sbi = EXT4_SB(sb);
3897 struct ext4_super_block *es = sbi->s_es;
3901 ext4_grpblk_t block;
3903 /* we can't allocate > group size */
3906 /* just a dirty hack to filter too big requests */
3907 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
3908 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
3910 /* start searching from the goal */
3912 if (goal < le32_to_cpu(es->s_first_data_block) ||
3913 goal >= ext4_blocks_count(es))
3914 goal = le32_to_cpu(es->s_first_data_block);
3915 ext4_get_group_no_and_offset(sb, goal, &group, &block);
3917 /* set up allocation goals */
3918 ac->ac_b_ex.fe_logical = ar->logical;
3919 ac->ac_b_ex.fe_group = 0;
3920 ac->ac_b_ex.fe_start = 0;
3921 ac->ac_b_ex.fe_len = 0;
3922 ac->ac_status = AC_STATUS_CONTINUE;
3923 ac->ac_groups_scanned = 0;
3924 ac->ac_ex_scanned = 0;
3927 ac->ac_inode = ar->inode;
3928 ac->ac_o_ex.fe_logical = ar->logical;
3929 ac->ac_o_ex.fe_group = group;
3930 ac->ac_o_ex.fe_start = block;
3931 ac->ac_o_ex.fe_len = len;
3932 ac->ac_g_ex.fe_logical = ar->logical;
3933 ac->ac_g_ex.fe_group = group;
3934 ac->ac_g_ex.fe_start = block;
3935 ac->ac_g_ex.fe_len = len;
3936 ac->ac_f_ex.fe_len = 0;
3937 ac->ac_flags = ar->flags;
3939 ac->ac_criteria = 0;
3941 ac->ac_bitmap_page = NULL;
3942 ac->ac_buddy_page = NULL;
3945 /* we have to define context: we'll we work with a file or
3946 * locality group. this is a policy, actually */
3947 ext4_mb_group_or_file(ac);
3949 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
3950 "left: %u/%u, right %u/%u to %swritable\n",
3951 (unsigned) ar->len, (unsigned) ar->logical,
3952 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
3953 (unsigned) ar->lleft, (unsigned) ar->pleft,
3954 (unsigned) ar->lright, (unsigned) ar->pright,
3955 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
3961 * release all resource we used in allocation
3963 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
3966 if (ac->ac_pa->pa_linear) {
3967 /* see comment in ext4_mb_use_group_pa() */
3968 spin_lock(&ac->ac_pa->pa_lock);
3969 ac->ac_pa->pa_pstart += ac->ac_b_ex.fe_len;
3970 ac->ac_pa->pa_lstart += ac->ac_b_ex.fe_len;
3971 ac->ac_pa->pa_free -= ac->ac_b_ex.fe_len;
3972 ac->ac_pa->pa_len -= ac->ac_b_ex.fe_len;
3973 spin_unlock(&ac->ac_pa->pa_lock);
3975 ext4_mb_put_pa(ac, ac->ac_sb, ac->ac_pa);
3977 if (ac->ac_bitmap_page)
3978 page_cache_release(ac->ac_bitmap_page);
3979 if (ac->ac_buddy_page)
3980 page_cache_release(ac->ac_buddy_page);
3981 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3982 mutex_unlock(&ac->ac_lg->lg_mutex);
3983 ext4_mb_collect_stats(ac);
3987 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
3993 for (i = 0; i < EXT4_SB(sb)->s_groups_count && needed > 0; i++) {
3994 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4003 * Main entry point into mballoc to allocate blocks
4004 * it tries to use preallocation first, then falls back
4005 * to usual allocation
4007 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4008 struct ext4_allocation_request *ar, int *errp)
4010 struct ext4_allocation_context *ac = NULL;
4011 struct ext4_sb_info *sbi;
4012 struct super_block *sb;
4013 ext4_fsblk_t block = 0;
4017 sb = ar->inode->i_sb;
4020 if (!test_opt(sb, MBALLOC)) {
4021 block = ext4_new_blocks_old(handle, ar->inode, ar->goal,
4026 while (ar->len && DQUOT_ALLOC_BLOCK(ar->inode, ar->len)) {
4027 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4036 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4042 ext4_mb_poll_new_transaction(sb, handle);
4044 *errp = ext4_mb_initialize_context(ac, ar);
4050 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4051 if (!ext4_mb_use_preallocated(ac)) {
4053 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4054 ext4_mb_normalize_request(ac, ar);
4056 /* allocate space in core */
4057 ext4_mb_regular_allocator(ac);
4059 /* as we've just preallocated more space than
4060 * user requested orinally, we store allocated
4061 * space in a special descriptor */
4062 if (ac->ac_status == AC_STATUS_FOUND &&
4063 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4064 ext4_mb_new_preallocation(ac);
4067 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4068 *errp = ext4_mb_mark_diskspace_used(ac, handle);
4069 if (*errp == -EAGAIN) {
4070 ac->ac_b_ex.fe_group = 0;
4071 ac->ac_b_ex.fe_start = 0;
4072 ac->ac_b_ex.fe_len = 0;
4073 ac->ac_status = AC_STATUS_CONTINUE;
4076 ac->ac_b_ex.fe_len = 0;
4078 ext4_mb_show_ac(ac);
4080 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4081 ar->len = ac->ac_b_ex.fe_len;
4084 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4088 ac->ac_b_ex.fe_len = 0;
4090 ext4_mb_show_ac(ac);
4093 ext4_mb_release_context(ac);
4096 if (ar->len < inquota)
4097 DQUOT_FREE_BLOCK(ar->inode, inquota - ar->len);
4099 kmem_cache_free(ext4_ac_cachep, ac);
4102 static void ext4_mb_poll_new_transaction(struct super_block *sb,
4105 struct ext4_sb_info *sbi = EXT4_SB(sb);
4107 if (sbi->s_last_transaction == handle->h_transaction->t_tid)
4110 /* new transaction! time to close last one and free blocks for
4111 * committed transaction. we know that only transaction can be
4112 * active, so previos transaction can be being logged and we
4113 * know that transaction before previous is known to be already
4114 * logged. this means that now we may free blocks freed in all
4115 * transactions before previous one. hope I'm clear enough ... */
4117 spin_lock(&sbi->s_md_lock);
4118 if (sbi->s_last_transaction != handle->h_transaction->t_tid) {
4119 mb_debug("new transaction %lu, old %lu\n",
4120 (unsigned long) handle->h_transaction->t_tid,
4121 (unsigned long) sbi->s_last_transaction);
4122 list_splice_init(&sbi->s_closed_transaction,
4123 &sbi->s_committed_transaction);
4124 list_splice_init(&sbi->s_active_transaction,
4125 &sbi->s_closed_transaction);
4126 sbi->s_last_transaction = handle->h_transaction->t_tid;
4128 spin_unlock(&sbi->s_md_lock);
4130 ext4_mb_free_committed_blocks(sb);
4133 static noinline_for_stack int
4134 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4135 ext4_group_t group, ext4_grpblk_t block, int count)
4137 struct ext4_group_info *db = e4b->bd_info;
4138 struct super_block *sb = e4b->bd_sb;
4139 struct ext4_sb_info *sbi = EXT4_SB(sb);
4140 struct ext4_free_metadata *md;
4143 BUG_ON(e4b->bd_bitmap_page == NULL);
4144 BUG_ON(e4b->bd_buddy_page == NULL);
4146 ext4_lock_group(sb, group);
4147 for (i = 0; i < count; i++) {
4149 if (md && db->bb_tid != handle->h_transaction->t_tid) {
4150 db->bb_md_cur = NULL;
4155 ext4_unlock_group(sb, group);
4156 md = kmalloc(sizeof(*md), GFP_NOFS);
4162 ext4_lock_group(sb, group);
4163 if (db->bb_md_cur == NULL) {
4164 spin_lock(&sbi->s_md_lock);
4165 list_add(&md->list, &sbi->s_active_transaction);
4166 spin_unlock(&sbi->s_md_lock);
4167 /* protect buddy cache from being freed,
4168 * otherwise we'll refresh it from
4169 * on-disk bitmap and lose not-yet-available
4171 page_cache_get(e4b->bd_buddy_page);
4172 page_cache_get(e4b->bd_bitmap_page);
4174 db->bb_tid = handle->h_transaction->t_tid;
4175 mb_debug("new md 0x%p for group %lu\n",
4183 BUG_ON(md->num >= EXT4_BB_MAX_BLOCKS);
4184 md->blocks[md->num] = block + i;
4186 if (md->num == EXT4_BB_MAX_BLOCKS) {
4187 /* no more space, put full container on a sb's list */
4188 db->bb_md_cur = NULL;
4191 ext4_unlock_group(sb, group);
4196 * Main entry point into mballoc to free blocks
4198 void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
4199 unsigned long block, unsigned long count,
4200 int metadata, unsigned long *freed)
4202 struct buffer_head *bitmap_bh = NULL;
4203 struct super_block *sb = inode->i_sb;
4204 struct ext4_allocation_context *ac = NULL;
4205 struct ext4_group_desc *gdp;
4206 struct ext4_super_block *es;
4207 unsigned long overflow;
4209 struct buffer_head *gd_bh;
4210 ext4_group_t block_group;
4211 struct ext4_sb_info *sbi;
4212 struct ext4_buddy e4b;
4218 ext4_mb_poll_new_transaction(sb, handle);
4221 es = EXT4_SB(sb)->s_es;
4222 if (block < le32_to_cpu(es->s_first_data_block) ||
4223 block + count < block ||
4224 block + count > ext4_blocks_count(es)) {
4225 ext4_error(sb, __func__,
4226 "Freeing blocks not in datazone - "
4227 "block = %lu, count = %lu", block, count);
4231 ext4_debug("freeing block %lu\n", block);
4233 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4235 ac->ac_op = EXT4_MB_HISTORY_FREE;
4236 ac->ac_inode = inode;
4242 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4245 * Check to see if we are freeing blocks across a group
4248 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4249 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4252 bitmap_bh = read_block_bitmap(sb, block_group);
4255 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4259 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4260 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4261 in_range(block, ext4_inode_table(sb, gdp),
4262 EXT4_SB(sb)->s_itb_per_group) ||
4263 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4264 EXT4_SB(sb)->s_itb_per_group)) {
4266 ext4_error(sb, __func__,
4267 "Freeing blocks in system zone - "
4268 "Block = %lu, count = %lu", block, count);
4269 /* err = 0. ext4_std_error should be a no op */
4273 BUFFER_TRACE(bitmap_bh, "getting write access");
4274 err = ext4_journal_get_write_access(handle, bitmap_bh);
4279 * We are about to modify some metadata. Call the journal APIs
4280 * to unshare ->b_data if a currently-committing transaction is
4283 BUFFER_TRACE(gd_bh, "get_write_access");
4284 err = ext4_journal_get_write_access(handle, gd_bh);
4288 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4292 #ifdef AGGRESSIVE_CHECK
4295 for (i = 0; i < count; i++)
4296 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4299 mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data,
4302 /* We dirtied the bitmap block */
4303 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4304 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
4307 ac->ac_b_ex.fe_group = block_group;
4308 ac->ac_b_ex.fe_start = bit;
4309 ac->ac_b_ex.fe_len = count;
4310 ext4_mb_store_history(ac);
4314 /* blocks being freed are metadata. these blocks shouldn't
4315 * be used until this transaction is committed */
4316 ext4_mb_free_metadata(handle, &e4b, block_group, bit, count);
4318 ext4_lock_group(sb, block_group);
4319 err = mb_free_blocks(inode, &e4b, bit, count);
4320 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4321 ext4_unlock_group(sb, block_group);
4325 spin_lock(sb_bgl_lock(sbi, block_group));
4326 le16_add_cpu(&gdp->bg_free_blocks_count, count);
4327 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4328 spin_unlock(sb_bgl_lock(sbi, block_group));
4329 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4331 ext4_mb_release_desc(&e4b);
4335 /* And the group descriptor block */
4336 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4337 ret = ext4_journal_dirty_metadata(handle, gd_bh);
4341 if (overflow && !err) {
4350 ext4_std_error(sb, err);
4352 kmem_cache_free(ext4_ac_cachep, ac);