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
24 #include "ext4_jbd2.h"
26 #include <linux/debugfs.h>
27 #include <linux/log2.h>
28 #include <linux/slab.h>
29 #include <trace/events/ext4.h>
33 * - test ext4_ext_search_left() and ext4_ext_search_right()
34 * - search for metadata in few groups
37 * - normalization should take into account whether file is still open
38 * - discard preallocations if no free space left (policy?)
39 * - don't normalize tails
41 * - reservation for superuser
44 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
45 * - track min/max extents in each group for better group selection
46 * - mb_mark_used() may allocate chunk right after splitting buddy
47 * - tree of groups sorted by number of free blocks
52 * The allocation request involve request for multiple number of blocks
53 * near to the goal(block) value specified.
55 * During initialization phase of the allocator we decide to use the
56 * group preallocation or inode preallocation depending on the size of
57 * the file. The size of the file could be the resulting file size we
58 * would have after allocation, or the current file size, which ever
59 * is larger. If the size is less than sbi->s_mb_stream_request we
60 * select to use the group preallocation. The default value of
61 * s_mb_stream_request is 16 blocks. This can also be tuned via
62 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
63 * terms of number of blocks.
65 * The main motivation for having small file use group preallocation is to
66 * ensure that we have small files closer together on the disk.
68 * First stage the allocator looks at the inode prealloc list,
69 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
70 * spaces for this particular inode. The inode prealloc space is
73 * pa_lstart -> the logical start block for this prealloc space
74 * pa_pstart -> the physical start block for this prealloc space
75 * pa_len -> length for this prealloc space (in clusters)
76 * pa_free -> free space available in this prealloc space (in clusters)
78 * The inode preallocation space is used looking at the _logical_ start
79 * block. If only the logical file block falls within the range of prealloc
80 * space we will consume the particular prealloc space. This makes sure that
81 * we have contiguous physical blocks representing the file blocks
83 * The important thing to be noted in case of inode prealloc space is that
84 * we don't modify the values associated to inode prealloc space except
87 * If we are not able to find blocks in the inode prealloc space and if we
88 * have the group allocation flag set then we look at the locality group
89 * prealloc space. These are per CPU prealloc list represented as
91 * ext4_sb_info.s_locality_groups[smp_processor_id()]
93 * The reason for having a per cpu locality group is to reduce the contention
94 * between CPUs. It is possible to get scheduled at this point.
96 * The locality group prealloc space is used looking at whether we have
97 * enough free space (pa_free) within the prealloc space.
99 * If we can't allocate blocks via inode prealloc or/and locality group
100 * prealloc then we look at the buddy cache. The buddy cache is represented
101 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
102 * mapped to the buddy and bitmap information regarding different
103 * groups. The buddy information is attached to buddy cache inode so that
104 * we can access them through the page cache. The information regarding
105 * each group is loaded via ext4_mb_load_buddy. The information involve
106 * block bitmap and buddy information. The information are stored in the
110 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
113 * one block each for bitmap and buddy information. So for each group we
114 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
115 * blocksize) blocks. So it can have information regarding groups_per_page
116 * which is blocks_per_page/2
118 * The buddy cache inode is not stored on disk. The inode is thrown
119 * away when the filesystem is unmounted.
121 * We look for count number of blocks in the buddy cache. If we were able
122 * to locate that many free blocks we return with additional information
123 * regarding rest of the contiguous physical block available
125 * Before allocating blocks via buddy cache we normalize the request
126 * blocks. This ensure we ask for more blocks that we needed. The extra
127 * blocks that we get after allocation is added to the respective prealloc
128 * list. In case of inode preallocation we follow a list of heuristics
129 * based on file size. This can be found in ext4_mb_normalize_request. If
130 * we are doing a group prealloc we try to normalize the request to
131 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
132 * dependent on the cluster size; for non-bigalloc file systems, it is
133 * 512 blocks. This can be tuned via
134 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
135 * terms of number of blocks. If we have mounted the file system with -O
136 * stripe=<value> option the group prealloc request is normalized to the
137 * the smallest multiple of the stripe value (sbi->s_stripe) which is
138 * greater than the default mb_group_prealloc.
140 * The regular allocator (using the buddy cache) supports a few tunables.
142 * /sys/fs/ext4/<partition>/mb_min_to_scan
143 * /sys/fs/ext4/<partition>/mb_max_to_scan
144 * /sys/fs/ext4/<partition>/mb_order2_req
146 * The regular allocator uses buddy scan only if the request len is power of
147 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
148 * value of s_mb_order2_reqs can be tuned via
149 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
150 * stripe size (sbi->s_stripe), we try to search for contiguous block in
151 * stripe size. This should result in better allocation on RAID setups. If
152 * not, we search in the specific group using bitmap for best extents. The
153 * tunable min_to_scan and max_to_scan control the behaviour here.
154 * min_to_scan indicate how long the mballoc __must__ look for a best
155 * extent and max_to_scan indicates how long the mballoc __can__ look for a
156 * best extent in the found extents. Searching for the blocks starts with
157 * the group specified as the goal value in allocation context via
158 * ac_g_ex. Each group is first checked based on the criteria whether it
159 * can be used for allocation. ext4_mb_good_group explains how the groups are
162 * Both the prealloc space are getting populated as above. So for the first
163 * request we will hit the buddy cache which will result in this prealloc
164 * space getting filled. The prealloc space is then later used for the
165 * subsequent request.
169 * mballoc operates on the following data:
171 * - in-core buddy (actually includes buddy and bitmap)
172 * - preallocation descriptors (PAs)
174 * there are two types of preallocations:
176 * assiged to specific inode and can be used for this inode only.
177 * it describes part of inode's space preallocated to specific
178 * physical blocks. any block from that preallocated can be used
179 * independent. the descriptor just tracks number of blocks left
180 * unused. so, before taking some block from descriptor, one must
181 * make sure corresponded logical block isn't allocated yet. this
182 * also means that freeing any block within descriptor's range
183 * must discard all preallocated blocks.
185 * assigned to specific locality group which does not translate to
186 * permanent set of inodes: inode can join and leave group. space
187 * from this type of preallocation can be used for any inode. thus
188 * it's consumed from the beginning to the end.
190 * relation between them can be expressed as:
191 * in-core buddy = on-disk bitmap + preallocation descriptors
193 * this mean blocks mballoc considers used are:
194 * - allocated blocks (persistent)
195 * - preallocated blocks (non-persistent)
197 * consistency in mballoc world means that at any time a block is either
198 * free or used in ALL structures. notice: "any time" should not be read
199 * literally -- time is discrete and delimited by locks.
201 * to keep it simple, we don't use block numbers, instead we count number of
202 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
204 * all operations can be expressed as:
205 * - init buddy: buddy = on-disk + PAs
206 * - new PA: buddy += N; PA = N
207 * - use inode PA: on-disk += N; PA -= N
208 * - discard inode PA buddy -= on-disk - PA; PA = 0
209 * - use locality group PA on-disk += N; PA -= N
210 * - discard locality group PA buddy -= PA; PA = 0
211 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
212 * is used in real operation because we can't know actual used
213 * bits from PA, only from on-disk bitmap
215 * if we follow this strict logic, then all operations above should be atomic.
216 * given some of them can block, we'd have to use something like semaphores
217 * killing performance on high-end SMP hardware. let's try to relax it using
218 * the following knowledge:
219 * 1) if buddy is referenced, it's already initialized
220 * 2) while block is used in buddy and the buddy is referenced,
221 * nobody can re-allocate that block
222 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
223 * bit set and PA claims same block, it's OK. IOW, one can set bit in
224 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
227 * so, now we're building a concurrency table:
230 * blocks for PA are allocated in the buddy, buddy must be referenced
231 * until PA is linked to allocation group to avoid concurrent buddy init
233 * we need to make sure that either on-disk bitmap or PA has uptodate data
234 * given (3) we care that PA-=N operation doesn't interfere with init
236 * the simplest way would be to have buddy initialized by the discard
237 * - use locality group PA
238 * again PA-=N must be serialized with init
239 * - discard locality group PA
240 * the simplest way would be to have buddy initialized by the discard
243 * i_data_sem serializes them
245 * discard process must wait until PA isn't used by another process
246 * - use locality group PA
247 * some mutex should serialize them
248 * - discard locality group PA
249 * discard process must wait until PA isn't used by another process
252 * i_data_sem or another mutex should serializes them
254 * discard process must wait until PA isn't used by another process
255 * - use locality group PA
256 * nothing wrong here -- they're different PAs covering different blocks
257 * - discard locality group PA
258 * discard process must wait until PA isn't used by another process
260 * now we're ready to make few consequences:
261 * - PA is referenced and while it is no discard is possible
262 * - PA is referenced until block isn't marked in on-disk bitmap
263 * - PA changes only after on-disk bitmap
264 * - discard must not compete with init. either init is done before
265 * any discard or they're serialized somehow
266 * - buddy init as sum of on-disk bitmap and PAs is done atomically
268 * a special case when we've used PA to emptiness. no need to modify buddy
269 * in this case, but we should care about concurrent init
274 * Logic in few words:
279 * mark bits in on-disk bitmap
282 * - use preallocation:
283 * find proper PA (per-inode or group)
285 * mark bits in on-disk bitmap
291 * mark bits in on-disk bitmap
294 * - discard preallocations in group:
296 * move them onto local list
297 * load on-disk bitmap
299 * remove PA from object (inode or locality group)
300 * mark free blocks in-core
302 * - discard inode's preallocations:
309 * - bitlock on a group (group)
310 * - object (inode/locality) (object)
321 * - release consumed pa:
326 * - generate in-core bitmap:
330 * - discard all for given object (inode, locality group):
335 * - discard all for given group:
342 static struct kmem_cache *ext4_pspace_cachep;
343 static struct kmem_cache *ext4_ac_cachep;
344 static struct kmem_cache *ext4_free_data_cachep;
346 /* We create slab caches for groupinfo data structures based on the
347 * superblock block size. There will be one per mounted filesystem for
348 * each unique s_blocksize_bits */
349 #define NR_GRPINFO_CACHES 8
350 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
352 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
353 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
354 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
355 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
358 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
360 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
362 static void ext4_free_data_callback(struct super_block *sb,
363 struct ext4_journal_cb_entry *jce, int rc);
365 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
367 #if BITS_PER_LONG == 64
368 *bit += ((unsigned long) addr & 7UL) << 3;
369 addr = (void *) ((unsigned long) addr & ~7UL);
370 #elif BITS_PER_LONG == 32
371 *bit += ((unsigned long) addr & 3UL) << 3;
372 addr = (void *) ((unsigned long) addr & ~3UL);
374 #error "how many bits you are?!"
379 static inline int mb_test_bit(int bit, void *addr)
382 * ext4_test_bit on architecture like powerpc
383 * needs unsigned long aligned address
385 addr = mb_correct_addr_and_bit(&bit, addr);
386 return ext4_test_bit(bit, addr);
389 static inline void mb_set_bit(int bit, void *addr)
391 addr = mb_correct_addr_and_bit(&bit, addr);
392 ext4_set_bit(bit, addr);
395 static inline void mb_clear_bit(int bit, void *addr)
397 addr = mb_correct_addr_and_bit(&bit, addr);
398 ext4_clear_bit(bit, addr);
401 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
403 int fix = 0, ret, tmpmax;
404 addr = mb_correct_addr_and_bit(&fix, addr);
408 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
414 static inline int mb_find_next_bit(void *addr, int max, int start)
416 int fix = 0, ret, tmpmax;
417 addr = mb_correct_addr_and_bit(&fix, addr);
421 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
427 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
431 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
434 if (order > e4b->bd_blkbits + 1) {
439 /* at order 0 we see each particular block */
441 *max = 1 << (e4b->bd_blkbits + 3);
442 return e4b->bd_bitmap;
445 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
446 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
452 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
453 int first, int count)
456 struct super_block *sb = e4b->bd_sb;
458 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
460 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
461 for (i = 0; i < count; i++) {
462 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
463 ext4_fsblk_t blocknr;
465 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
466 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
467 ext4_grp_locked_error(sb, e4b->bd_group,
468 inode ? inode->i_ino : 0,
470 "freeing block already freed "
474 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
478 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
482 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
484 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
485 for (i = 0; i < count; i++) {
486 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
487 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
491 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
493 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
494 unsigned char *b1, *b2;
496 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
497 b2 = (unsigned char *) bitmap;
498 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
499 if (b1[i] != b2[i]) {
500 ext4_msg(e4b->bd_sb, KERN_ERR,
501 "corruption in group %u "
502 "at byte %u(%u): %x in copy != %x "
504 e4b->bd_group, i, i * 8, b1[i], b2[i]);
512 static inline void mb_free_blocks_double(struct inode *inode,
513 struct ext4_buddy *e4b, int first, int count)
517 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
518 int first, int count)
522 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
528 #ifdef AGGRESSIVE_CHECK
530 #define MB_CHECK_ASSERT(assert) \
534 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
535 function, file, line, # assert); \
540 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
541 const char *function, int line)
543 struct super_block *sb = e4b->bd_sb;
544 int order = e4b->bd_blkbits + 1;
551 struct ext4_group_info *grp;
554 struct list_head *cur;
559 static int mb_check_counter;
560 if (mb_check_counter++ % 100 != 0)
565 buddy = mb_find_buddy(e4b, order, &max);
566 MB_CHECK_ASSERT(buddy);
567 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
568 MB_CHECK_ASSERT(buddy2);
569 MB_CHECK_ASSERT(buddy != buddy2);
570 MB_CHECK_ASSERT(max * 2 == max2);
573 for (i = 0; i < max; i++) {
575 if (mb_test_bit(i, buddy)) {
576 /* only single bit in buddy2 may be 1 */
577 if (!mb_test_bit(i << 1, buddy2)) {
579 mb_test_bit((i<<1)+1, buddy2));
580 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
582 mb_test_bit(i << 1, buddy2));
587 /* both bits in buddy2 must be 1 */
588 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
589 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
591 for (j = 0; j < (1 << order); j++) {
592 k = (i * (1 << order)) + j;
594 !mb_test_bit(k, e4b->bd_bitmap));
598 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
603 buddy = mb_find_buddy(e4b, 0, &max);
604 for (i = 0; i < max; i++) {
605 if (!mb_test_bit(i, buddy)) {
606 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
614 /* check used bits only */
615 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
616 buddy2 = mb_find_buddy(e4b, j, &max2);
618 MB_CHECK_ASSERT(k < max2);
619 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
622 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
623 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
625 grp = ext4_get_group_info(sb, e4b->bd_group);
626 list_for_each(cur, &grp->bb_prealloc_list) {
627 ext4_group_t groupnr;
628 struct ext4_prealloc_space *pa;
629 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
630 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
631 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
632 for (i = 0; i < pa->pa_len; i++)
633 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
637 #undef MB_CHECK_ASSERT
638 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
639 __FILE__, __func__, __LINE__)
641 #define mb_check_buddy(e4b)
645 * Divide blocks started from @first with length @len into
646 * smaller chunks with power of 2 blocks.
647 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
648 * then increase bb_counters[] for corresponded chunk size.
650 static void ext4_mb_mark_free_simple(struct super_block *sb,
651 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
652 struct ext4_group_info *grp)
654 struct ext4_sb_info *sbi = EXT4_SB(sb);
658 unsigned short border;
660 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
662 border = 2 << sb->s_blocksize_bits;
665 /* find how many blocks can be covered since this position */
666 max = ffs(first | border) - 1;
668 /* find how many blocks of power 2 we need to mark */
675 /* mark multiblock chunks only */
676 grp->bb_counters[min]++;
678 mb_clear_bit(first >> min,
679 buddy + sbi->s_mb_offsets[min]);
687 * Cache the order of the largest free extent we have available in this block
691 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
696 grp->bb_largest_free_order = -1; /* uninit */
698 bits = sb->s_blocksize_bits + 1;
699 for (i = bits; i >= 0; i--) {
700 if (grp->bb_counters[i] > 0) {
701 grp->bb_largest_free_order = i;
707 static noinline_for_stack
708 void ext4_mb_generate_buddy(struct super_block *sb,
709 void *buddy, void *bitmap, ext4_group_t group)
711 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
712 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
717 unsigned fragments = 0;
718 unsigned long long period = get_cycles();
720 /* initialize buddy from bitmap which is aggregation
721 * of on-disk bitmap and preallocations */
722 i = mb_find_next_zero_bit(bitmap, max, 0);
723 grp->bb_first_free = i;
727 i = mb_find_next_bit(bitmap, max, i);
731 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
733 grp->bb_counters[0]++;
735 i = mb_find_next_zero_bit(bitmap, max, i);
737 grp->bb_fragments = fragments;
739 if (free != grp->bb_free) {
740 ext4_grp_locked_error(sb, group, 0, 0,
741 "%u clusters in bitmap, %u in gd",
744 * If we intent to continue, we consider group descritor
745 * corrupt and update bb_free using bitmap value
749 mb_set_largest_free_order(sb, grp);
751 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
753 period = get_cycles() - period;
754 spin_lock(&EXT4_SB(sb)->s_bal_lock);
755 EXT4_SB(sb)->s_mb_buddies_generated++;
756 EXT4_SB(sb)->s_mb_generation_time += period;
757 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
760 /* The buddy information is attached the buddy cache inode
761 * for convenience. The information regarding each group
762 * is loaded via ext4_mb_load_buddy. The information involve
763 * block bitmap and buddy information. The information are
764 * stored in the inode as
767 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
770 * one block each for bitmap and buddy information.
771 * So for each group we take up 2 blocks. A page can
772 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
773 * So it can have information regarding groups_per_page which
774 * is blocks_per_page/2
776 * Locking note: This routine takes the block group lock of all groups
777 * for this page; do not hold this lock when calling this routine!
780 static int ext4_mb_init_cache(struct page *page, char *incore)
782 ext4_group_t ngroups;
788 ext4_group_t first_group, group;
790 struct super_block *sb;
791 struct buffer_head *bhs;
792 struct buffer_head **bh = NULL;
796 struct ext4_group_info *grinfo;
798 mb_debug(1, "init page %lu\n", page->index);
800 inode = page->mapping->host;
802 ngroups = ext4_get_groups_count(sb);
803 blocksize = 1 << inode->i_blkbits;
804 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
806 groups_per_page = blocks_per_page >> 1;
807 if (groups_per_page == 0)
810 /* allocate buffer_heads to read bitmaps */
811 if (groups_per_page > 1) {
812 i = sizeof(struct buffer_head *) * groups_per_page;
813 bh = kzalloc(i, GFP_NOFS);
821 first_group = page->index * blocks_per_page / 2;
823 /* read all groups the page covers into the cache */
824 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
825 if (group >= ngroups)
828 grinfo = ext4_get_group_info(sb, group);
830 * If page is uptodate then we came here after online resize
831 * which added some new uninitialized group info structs, so
832 * we must skip all initialized uptodate buddies on the page,
833 * which may be currently in use by an allocating task.
835 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
839 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
843 mb_debug(1, "read bitmap for group %u\n", group);
846 /* wait for I/O completion */
847 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
848 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
854 first_block = page->index * blocks_per_page;
855 for (i = 0; i < blocks_per_page; i++) {
858 group = (first_block + i) >> 1;
859 if (group >= ngroups)
862 if (!bh[group - first_group])
863 /* skip initialized uptodate buddy */
867 * data carry information regarding this
868 * particular group in the format specified
872 data = page_address(page) + (i * blocksize);
873 bitmap = bh[group - first_group]->b_data;
876 * We place the buddy block and bitmap block
879 if ((first_block + i) & 1) {
880 /* this is block of buddy */
881 BUG_ON(incore == NULL);
882 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
883 group, page->index, i * blocksize);
884 trace_ext4_mb_buddy_bitmap_load(sb, group);
885 grinfo = ext4_get_group_info(sb, group);
886 grinfo->bb_fragments = 0;
887 memset(grinfo->bb_counters, 0,
888 sizeof(*grinfo->bb_counters) *
889 (sb->s_blocksize_bits+2));
891 * incore got set to the group block bitmap below
893 ext4_lock_group(sb, group);
895 memset(data, 0xff, blocksize);
896 ext4_mb_generate_buddy(sb, data, incore, group);
897 ext4_unlock_group(sb, group);
900 /* this is block of bitmap */
901 BUG_ON(incore != NULL);
902 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
903 group, page->index, i * blocksize);
904 trace_ext4_mb_bitmap_load(sb, group);
906 /* see comments in ext4_mb_put_pa() */
907 ext4_lock_group(sb, group);
908 memcpy(data, bitmap, blocksize);
910 /* mark all preallocated blks used in in-core bitmap */
911 ext4_mb_generate_from_pa(sb, data, group);
912 ext4_mb_generate_from_freelist(sb, data, group);
913 ext4_unlock_group(sb, group);
915 /* set incore so that the buddy information can be
916 * generated using this
921 SetPageUptodate(page);
925 for (i = 0; i < groups_per_page; i++)
934 * Lock the buddy and bitmap pages. This make sure other parallel init_group
935 * on the same buddy page doesn't happen whild holding the buddy page lock.
936 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
937 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
939 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
940 ext4_group_t group, struct ext4_buddy *e4b)
942 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
943 int block, pnum, poff;
947 e4b->bd_buddy_page = NULL;
948 e4b->bd_bitmap_page = NULL;
950 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
952 * the buddy cache inode stores the block bitmap
953 * and buddy information in consecutive blocks.
954 * So for each group we need two blocks.
957 pnum = block / blocks_per_page;
958 poff = block % blocks_per_page;
959 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
962 BUG_ON(page->mapping != inode->i_mapping);
963 e4b->bd_bitmap_page = page;
964 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
966 if (blocks_per_page >= 2) {
967 /* buddy and bitmap are on the same page */
972 pnum = block / blocks_per_page;
973 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
976 BUG_ON(page->mapping != inode->i_mapping);
977 e4b->bd_buddy_page = page;
981 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
983 if (e4b->bd_bitmap_page) {
984 unlock_page(e4b->bd_bitmap_page);
985 page_cache_release(e4b->bd_bitmap_page);
987 if (e4b->bd_buddy_page) {
988 unlock_page(e4b->bd_buddy_page);
989 page_cache_release(e4b->bd_buddy_page);
994 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
995 * block group lock of all groups for this page; do not hold the BG lock when
996 * calling this routine!
998 static noinline_for_stack
999 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1002 struct ext4_group_info *this_grp;
1003 struct ext4_buddy e4b;
1007 mb_debug(1, "init group %u\n", group);
1008 this_grp = ext4_get_group_info(sb, group);
1010 * This ensures that we don't reinit the buddy cache
1011 * page which map to the group from which we are already
1012 * allocating. If we are looking at the buddy cache we would
1013 * have taken a reference using ext4_mb_load_buddy and that
1014 * would have pinned buddy page to page cache.
1016 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1017 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1019 * somebody initialized the group
1020 * return without doing anything
1025 page = e4b.bd_bitmap_page;
1026 ret = ext4_mb_init_cache(page, NULL);
1029 if (!PageUptodate(page)) {
1033 mark_page_accessed(page);
1035 if (e4b.bd_buddy_page == NULL) {
1037 * If both the bitmap and buddy are in
1038 * the same page we don't need to force
1044 /* init buddy cache */
1045 page = e4b.bd_buddy_page;
1046 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1049 if (!PageUptodate(page)) {
1053 mark_page_accessed(page);
1055 ext4_mb_put_buddy_page_lock(&e4b);
1060 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1061 * block group lock of all groups for this page; do not hold the BG lock when
1062 * calling this routine!
1064 static noinline_for_stack int
1065 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1066 struct ext4_buddy *e4b)
1068 int blocks_per_page;
1074 struct ext4_group_info *grp;
1075 struct ext4_sb_info *sbi = EXT4_SB(sb);
1076 struct inode *inode = sbi->s_buddy_cache;
1078 mb_debug(1, "load group %u\n", group);
1080 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1081 grp = ext4_get_group_info(sb, group);
1083 e4b->bd_blkbits = sb->s_blocksize_bits;
1086 e4b->bd_group = group;
1087 e4b->bd_buddy_page = NULL;
1088 e4b->bd_bitmap_page = NULL;
1090 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1092 * we need full data about the group
1093 * to make a good selection
1095 ret = ext4_mb_init_group(sb, group);
1101 * the buddy cache inode stores the block bitmap
1102 * and buddy information in consecutive blocks.
1103 * So for each group we need two blocks.
1106 pnum = block / blocks_per_page;
1107 poff = block % blocks_per_page;
1109 /* we could use find_or_create_page(), but it locks page
1110 * what we'd like to avoid in fast path ... */
1111 page = find_get_page(inode->i_mapping, pnum);
1112 if (page == NULL || !PageUptodate(page)) {
1115 * drop the page reference and try
1116 * to get the page with lock. If we
1117 * are not uptodate that implies
1118 * somebody just created the page but
1119 * is yet to initialize the same. So
1120 * wait for it to initialize.
1122 page_cache_release(page);
1123 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1125 BUG_ON(page->mapping != inode->i_mapping);
1126 if (!PageUptodate(page)) {
1127 ret = ext4_mb_init_cache(page, NULL);
1132 mb_cmp_bitmaps(e4b, page_address(page) +
1133 (poff * sb->s_blocksize));
1138 if (page == NULL || !PageUptodate(page)) {
1142 e4b->bd_bitmap_page = page;
1143 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1144 mark_page_accessed(page);
1147 pnum = block / blocks_per_page;
1148 poff = block % blocks_per_page;
1150 page = find_get_page(inode->i_mapping, pnum);
1151 if (page == NULL || !PageUptodate(page)) {
1153 page_cache_release(page);
1154 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1156 BUG_ON(page->mapping != inode->i_mapping);
1157 if (!PageUptodate(page)) {
1158 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1167 if (page == NULL || !PageUptodate(page)) {
1171 e4b->bd_buddy_page = page;
1172 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1173 mark_page_accessed(page);
1175 BUG_ON(e4b->bd_bitmap_page == NULL);
1176 BUG_ON(e4b->bd_buddy_page == NULL);
1182 page_cache_release(page);
1183 if (e4b->bd_bitmap_page)
1184 page_cache_release(e4b->bd_bitmap_page);
1185 if (e4b->bd_buddy_page)
1186 page_cache_release(e4b->bd_buddy_page);
1187 e4b->bd_buddy = NULL;
1188 e4b->bd_bitmap = NULL;
1192 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1194 if (e4b->bd_bitmap_page)
1195 page_cache_release(e4b->bd_bitmap_page);
1196 if (e4b->bd_buddy_page)
1197 page_cache_release(e4b->bd_buddy_page);
1201 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1206 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1207 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1210 while (order <= e4b->bd_blkbits + 1) {
1212 if (!mb_test_bit(block, bb)) {
1213 /* this block is part of buddy of order 'order' */
1216 bb += 1 << (e4b->bd_blkbits - order);
1222 static void mb_clear_bits(void *bm, int cur, int len)
1228 if ((cur & 31) == 0 && (len - cur) >= 32) {
1229 /* fast path: clear whole word at once */
1230 addr = bm + (cur >> 3);
1235 mb_clear_bit(cur, bm);
1240 void ext4_set_bits(void *bm, int cur, int len)
1246 if ((cur & 31) == 0 && (len - cur) >= 32) {
1247 /* fast path: set whole word at once */
1248 addr = bm + (cur >> 3);
1253 mb_set_bit(cur, bm);
1258 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1259 int first, int count)
1266 struct super_block *sb = e4b->bd_sb;
1268 BUG_ON(first + count > (sb->s_blocksize << 3));
1269 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1270 mb_check_buddy(e4b);
1271 mb_free_blocks_double(inode, e4b, first, count);
1273 e4b->bd_info->bb_free += count;
1274 if (first < e4b->bd_info->bb_first_free)
1275 e4b->bd_info->bb_first_free = first;
1277 /* let's maintain fragments counter */
1279 block = !mb_test_bit(first - 1, e4b->bd_bitmap);
1280 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1281 max = !mb_test_bit(first + count, e4b->bd_bitmap);
1283 e4b->bd_info->bb_fragments--;
1284 else if (!block && !max)
1285 e4b->bd_info->bb_fragments++;
1287 /* let's maintain buddy itself */
1288 while (count-- > 0) {
1292 if (!mb_test_bit(block, e4b->bd_bitmap)) {
1293 ext4_fsblk_t blocknr;
1295 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1296 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1297 ext4_grp_locked_error(sb, e4b->bd_group,
1298 inode ? inode->i_ino : 0,
1300 "freeing already freed block "
1303 mb_clear_bit(block, e4b->bd_bitmap);
1304 e4b->bd_info->bb_counters[order]++;
1306 /* start of the buddy */
1307 buddy = mb_find_buddy(e4b, order, &max);
1311 if (mb_test_bit(block, buddy) ||
1312 mb_test_bit(block + 1, buddy))
1315 /* both the buddies are free, try to coalesce them */
1316 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1322 /* for special purposes, we don't set
1323 * free bits in bitmap */
1324 mb_set_bit(block, buddy);
1325 mb_set_bit(block + 1, buddy);
1327 e4b->bd_info->bb_counters[order]--;
1328 e4b->bd_info->bb_counters[order]--;
1332 e4b->bd_info->bb_counters[order]++;
1334 mb_clear_bit(block, buddy2);
1338 mb_set_largest_free_order(sb, e4b->bd_info);
1339 mb_check_buddy(e4b);
1342 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1343 int needed, struct ext4_free_extent *ex)
1349 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1352 buddy = mb_find_buddy(e4b, 0, &max);
1353 BUG_ON(buddy == NULL);
1354 BUG_ON(block >= max);
1355 if (mb_test_bit(block, buddy)) {
1362 /* find actual order */
1363 order = mb_find_order_for_block(e4b, block);
1364 block = block >> order;
1366 ex->fe_len = 1 << order;
1367 ex->fe_start = block << order;
1368 ex->fe_group = e4b->bd_group;
1370 /* calc difference from given start */
1371 next = next - ex->fe_start;
1373 ex->fe_start += next;
1375 while (needed > ex->fe_len &&
1376 mb_find_buddy(e4b, order, &max)) {
1378 if (block + 1 >= max)
1381 next = (block + 1) * (1 << order);
1382 if (mb_test_bit(next, e4b->bd_bitmap))
1385 order = mb_find_order_for_block(e4b, next);
1387 block = next >> order;
1388 ex->fe_len += 1 << order;
1391 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1395 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1401 int start = ex->fe_start;
1402 int len = ex->fe_len;
1407 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1408 BUG_ON(e4b->bd_group != ex->fe_group);
1409 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1410 mb_check_buddy(e4b);
1411 mb_mark_used_double(e4b, start, len);
1413 e4b->bd_info->bb_free -= len;
1414 if (e4b->bd_info->bb_first_free == start)
1415 e4b->bd_info->bb_first_free += len;
1417 /* let's maintain fragments counter */
1419 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1420 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1421 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1423 e4b->bd_info->bb_fragments++;
1424 else if (!mlen && !max)
1425 e4b->bd_info->bb_fragments--;
1427 /* let's maintain buddy itself */
1429 ord = mb_find_order_for_block(e4b, start);
1431 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1432 /* the whole chunk may be allocated at once! */
1434 buddy = mb_find_buddy(e4b, ord, &max);
1435 BUG_ON((start >> ord) >= max);
1436 mb_set_bit(start >> ord, buddy);
1437 e4b->bd_info->bb_counters[ord]--;
1444 /* store for history */
1446 ret = len | (ord << 16);
1448 /* we have to split large buddy */
1450 buddy = mb_find_buddy(e4b, ord, &max);
1451 mb_set_bit(start >> ord, buddy);
1452 e4b->bd_info->bb_counters[ord]--;
1455 cur = (start >> ord) & ~1U;
1456 buddy = mb_find_buddy(e4b, ord, &max);
1457 mb_clear_bit(cur, buddy);
1458 mb_clear_bit(cur + 1, buddy);
1459 e4b->bd_info->bb_counters[ord]++;
1460 e4b->bd_info->bb_counters[ord]++;
1462 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1464 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1465 mb_check_buddy(e4b);
1471 * Must be called under group lock!
1473 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1474 struct ext4_buddy *e4b)
1476 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1479 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1480 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1482 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1483 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1484 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1486 /* preallocation can change ac_b_ex, thus we store actually
1487 * allocated blocks for history */
1488 ac->ac_f_ex = ac->ac_b_ex;
1490 ac->ac_status = AC_STATUS_FOUND;
1491 ac->ac_tail = ret & 0xffff;
1492 ac->ac_buddy = ret >> 16;
1495 * take the page reference. We want the page to be pinned
1496 * so that we don't get a ext4_mb_init_cache_call for this
1497 * group until we update the bitmap. That would mean we
1498 * double allocate blocks. The reference is dropped
1499 * in ext4_mb_release_context
1501 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1502 get_page(ac->ac_bitmap_page);
1503 ac->ac_buddy_page = e4b->bd_buddy_page;
1504 get_page(ac->ac_buddy_page);
1505 /* store last allocated for subsequent stream allocation */
1506 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1507 spin_lock(&sbi->s_md_lock);
1508 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1509 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1510 spin_unlock(&sbi->s_md_lock);
1515 * regular allocator, for general purposes allocation
1518 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1519 struct ext4_buddy *e4b,
1522 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1523 struct ext4_free_extent *bex = &ac->ac_b_ex;
1524 struct ext4_free_extent *gex = &ac->ac_g_ex;
1525 struct ext4_free_extent ex;
1528 if (ac->ac_status == AC_STATUS_FOUND)
1531 * We don't want to scan for a whole year
1533 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1534 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1535 ac->ac_status = AC_STATUS_BREAK;
1540 * Haven't found good chunk so far, let's continue
1542 if (bex->fe_len < gex->fe_len)
1545 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1546 && bex->fe_group == e4b->bd_group) {
1547 /* recheck chunk's availability - we don't know
1548 * when it was found (within this lock-unlock
1550 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1551 if (max >= gex->fe_len) {
1552 ext4_mb_use_best_found(ac, e4b);
1559 * The routine checks whether found extent is good enough. If it is,
1560 * then the extent gets marked used and flag is set to the context
1561 * to stop scanning. Otherwise, the extent is compared with the
1562 * previous found extent and if new one is better, then it's stored
1563 * in the context. Later, the best found extent will be used, if
1564 * mballoc can't find good enough extent.
1566 * FIXME: real allocation policy is to be designed yet!
1568 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1569 struct ext4_free_extent *ex,
1570 struct ext4_buddy *e4b)
1572 struct ext4_free_extent *bex = &ac->ac_b_ex;
1573 struct ext4_free_extent *gex = &ac->ac_g_ex;
1575 BUG_ON(ex->fe_len <= 0);
1576 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1577 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1578 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1583 * The special case - take what you catch first
1585 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1587 ext4_mb_use_best_found(ac, e4b);
1592 * Let's check whether the chuck is good enough
1594 if (ex->fe_len == gex->fe_len) {
1596 ext4_mb_use_best_found(ac, e4b);
1601 * If this is first found extent, just store it in the context
1603 if (bex->fe_len == 0) {
1609 * If new found extent is better, store it in the context
1611 if (bex->fe_len < gex->fe_len) {
1612 /* if the request isn't satisfied, any found extent
1613 * larger than previous best one is better */
1614 if (ex->fe_len > bex->fe_len)
1616 } else if (ex->fe_len > gex->fe_len) {
1617 /* if the request is satisfied, then we try to find
1618 * an extent that still satisfy the request, but is
1619 * smaller than previous one */
1620 if (ex->fe_len < bex->fe_len)
1624 ext4_mb_check_limits(ac, e4b, 0);
1627 static noinline_for_stack
1628 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1629 struct ext4_buddy *e4b)
1631 struct ext4_free_extent ex = ac->ac_b_ex;
1632 ext4_group_t group = ex.fe_group;
1636 BUG_ON(ex.fe_len <= 0);
1637 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1641 ext4_lock_group(ac->ac_sb, group);
1642 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1646 ext4_mb_use_best_found(ac, e4b);
1649 ext4_unlock_group(ac->ac_sb, group);
1650 ext4_mb_unload_buddy(e4b);
1655 static noinline_for_stack
1656 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1657 struct ext4_buddy *e4b)
1659 ext4_group_t group = ac->ac_g_ex.fe_group;
1662 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1663 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1664 struct ext4_free_extent ex;
1666 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1668 if (grp->bb_free == 0)
1671 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1675 ext4_lock_group(ac->ac_sb, group);
1676 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1677 ac->ac_g_ex.fe_len, &ex);
1679 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1682 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1684 /* use do_div to get remainder (would be 64-bit modulo) */
1685 if (do_div(start, sbi->s_stripe) == 0) {
1688 ext4_mb_use_best_found(ac, e4b);
1690 } else if (max >= ac->ac_g_ex.fe_len) {
1691 BUG_ON(ex.fe_len <= 0);
1692 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1693 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1696 ext4_mb_use_best_found(ac, e4b);
1697 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1698 /* Sometimes, caller may want to merge even small
1699 * number of blocks to an existing extent */
1700 BUG_ON(ex.fe_len <= 0);
1701 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1702 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1705 ext4_mb_use_best_found(ac, e4b);
1707 ext4_unlock_group(ac->ac_sb, group);
1708 ext4_mb_unload_buddy(e4b);
1714 * The routine scans buddy structures (not bitmap!) from given order
1715 * to max order and tries to find big enough chunk to satisfy the req
1717 static noinline_for_stack
1718 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1719 struct ext4_buddy *e4b)
1721 struct super_block *sb = ac->ac_sb;
1722 struct ext4_group_info *grp = e4b->bd_info;
1728 BUG_ON(ac->ac_2order <= 0);
1729 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1730 if (grp->bb_counters[i] == 0)
1733 buddy = mb_find_buddy(e4b, i, &max);
1734 BUG_ON(buddy == NULL);
1736 k = mb_find_next_zero_bit(buddy, max, 0);
1741 ac->ac_b_ex.fe_len = 1 << i;
1742 ac->ac_b_ex.fe_start = k << i;
1743 ac->ac_b_ex.fe_group = e4b->bd_group;
1745 ext4_mb_use_best_found(ac, e4b);
1747 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1749 if (EXT4_SB(sb)->s_mb_stats)
1750 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1757 * The routine scans the group and measures all found extents.
1758 * In order to optimize scanning, caller must pass number of
1759 * free blocks in the group, so the routine can know upper limit.
1761 static noinline_for_stack
1762 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1763 struct ext4_buddy *e4b)
1765 struct super_block *sb = ac->ac_sb;
1766 void *bitmap = e4b->bd_bitmap;
1767 struct ext4_free_extent ex;
1771 free = e4b->bd_info->bb_free;
1774 i = e4b->bd_info->bb_first_free;
1776 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1777 i = mb_find_next_zero_bit(bitmap,
1778 EXT4_CLUSTERS_PER_GROUP(sb), i);
1779 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1781 * IF we have corrupt bitmap, we won't find any
1782 * free blocks even though group info says we
1783 * we have free blocks
1785 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1786 "%d free clusters as per "
1787 "group info. But bitmap says 0",
1792 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1793 BUG_ON(ex.fe_len <= 0);
1794 if (free < ex.fe_len) {
1795 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1796 "%d free clusters as per "
1797 "group info. But got %d blocks",
1800 * The number of free blocks differs. This mostly
1801 * indicate that the bitmap is corrupt. So exit
1802 * without claiming the space.
1807 ext4_mb_measure_extent(ac, &ex, e4b);
1813 ext4_mb_check_limits(ac, e4b, 1);
1817 * This is a special case for storages like raid5
1818 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1820 static noinline_for_stack
1821 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1822 struct ext4_buddy *e4b)
1824 struct super_block *sb = ac->ac_sb;
1825 struct ext4_sb_info *sbi = EXT4_SB(sb);
1826 void *bitmap = e4b->bd_bitmap;
1827 struct ext4_free_extent ex;
1828 ext4_fsblk_t first_group_block;
1833 BUG_ON(sbi->s_stripe == 0);
1835 /* find first stripe-aligned block in group */
1836 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1838 a = first_group_block + sbi->s_stripe - 1;
1839 do_div(a, sbi->s_stripe);
1840 i = (a * sbi->s_stripe) - first_group_block;
1842 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1843 if (!mb_test_bit(i, bitmap)) {
1844 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
1845 if (max >= sbi->s_stripe) {
1848 ext4_mb_use_best_found(ac, e4b);
1856 /* This is now called BEFORE we load the buddy bitmap. */
1857 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1858 ext4_group_t group, int cr)
1860 unsigned free, fragments;
1861 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1862 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1864 BUG_ON(cr < 0 || cr >= 4);
1866 free = grp->bb_free;
1869 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
1872 /* We only do this if the grp has never been initialized */
1873 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1874 int ret = ext4_mb_init_group(ac->ac_sb, group);
1879 fragments = grp->bb_fragments;
1885 BUG_ON(ac->ac_2order == 0);
1887 if (grp->bb_largest_free_order < ac->ac_2order)
1890 /* Avoid using the first bg of a flexgroup for data files */
1891 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1892 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1893 ((group % flex_size) == 0))
1898 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1902 if (free >= ac->ac_g_ex.fe_len)
1914 static noinline_for_stack int
1915 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1917 ext4_group_t ngroups, group, i;
1920 struct ext4_sb_info *sbi;
1921 struct super_block *sb;
1922 struct ext4_buddy e4b;
1926 ngroups = ext4_get_groups_count(sb);
1927 /* non-extent files are limited to low blocks/groups */
1928 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
1929 ngroups = sbi->s_blockfile_groups;
1931 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1933 /* first, try the goal */
1934 err = ext4_mb_find_by_goal(ac, &e4b);
1935 if (err || ac->ac_status == AC_STATUS_FOUND)
1938 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1942 * ac->ac2_order is set only if the fe_len is a power of 2
1943 * if ac2_order is set we also set criteria to 0 so that we
1944 * try exact allocation using buddy.
1946 i = fls(ac->ac_g_ex.fe_len);
1949 * We search using buddy data only if the order of the request
1950 * is greater than equal to the sbi_s_mb_order2_reqs
1951 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1953 if (i >= sbi->s_mb_order2_reqs) {
1955 * This should tell if fe_len is exactly power of 2
1957 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1958 ac->ac_2order = i - 1;
1961 /* if stream allocation is enabled, use global goal */
1962 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1963 /* TBD: may be hot point */
1964 spin_lock(&sbi->s_md_lock);
1965 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1966 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1967 spin_unlock(&sbi->s_md_lock);
1970 /* Let's just scan groups to find more-less suitable blocks */
1971 cr = ac->ac_2order ? 0 : 1;
1973 * cr == 0 try to get exact allocation,
1974 * cr == 3 try to get anything
1977 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1978 ac->ac_criteria = cr;
1980 * searching for the right group start
1981 * from the goal value specified
1983 group = ac->ac_g_ex.fe_group;
1985 for (i = 0; i < ngroups; group++, i++) {
1986 if (group == ngroups)
1989 /* This now checks without needing the buddy page */
1990 if (!ext4_mb_good_group(ac, group, cr))
1993 err = ext4_mb_load_buddy(sb, group, &e4b);
1997 ext4_lock_group(sb, group);
2000 * We need to check again after locking the
2003 if (!ext4_mb_good_group(ac, group, cr)) {
2004 ext4_unlock_group(sb, group);
2005 ext4_mb_unload_buddy(&e4b);
2009 ac->ac_groups_scanned++;
2011 ext4_mb_simple_scan_group(ac, &e4b);
2012 else if (cr == 1 && sbi->s_stripe &&
2013 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2014 ext4_mb_scan_aligned(ac, &e4b);
2016 ext4_mb_complex_scan_group(ac, &e4b);
2018 ext4_unlock_group(sb, group);
2019 ext4_mb_unload_buddy(&e4b);
2021 if (ac->ac_status != AC_STATUS_CONTINUE)
2026 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2027 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2029 * We've been searching too long. Let's try to allocate
2030 * the best chunk we've found so far
2033 ext4_mb_try_best_found(ac, &e4b);
2034 if (ac->ac_status != AC_STATUS_FOUND) {
2036 * Someone more lucky has already allocated it.
2037 * The only thing we can do is just take first
2039 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2041 ac->ac_b_ex.fe_group = 0;
2042 ac->ac_b_ex.fe_start = 0;
2043 ac->ac_b_ex.fe_len = 0;
2044 ac->ac_status = AC_STATUS_CONTINUE;
2045 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2047 atomic_inc(&sbi->s_mb_lost_chunks);
2055 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2057 struct super_block *sb = seq->private;
2060 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2063 return (void *) ((unsigned long) group);
2066 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2068 struct super_block *sb = seq->private;
2072 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2075 return (void *) ((unsigned long) group);
2078 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2080 struct super_block *sb = seq->private;
2081 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2083 int err, buddy_loaded = 0;
2084 struct ext4_buddy e4b;
2085 struct ext4_group_info *grinfo;
2087 struct ext4_group_info info;
2088 ext4_grpblk_t counters[16];
2093 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2094 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2095 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2096 "group", "free", "frags", "first",
2097 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2098 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2100 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2101 sizeof(struct ext4_group_info);
2102 grinfo = ext4_get_group_info(sb, group);
2103 /* Load the group info in memory only if not already loaded. */
2104 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2105 err = ext4_mb_load_buddy(sb, group, &e4b);
2107 seq_printf(seq, "#%-5u: I/O error\n", group);
2113 memcpy(&sg, ext4_get_group_info(sb, group), i);
2116 ext4_mb_unload_buddy(&e4b);
2118 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2119 sg.info.bb_fragments, sg.info.bb_first_free);
2120 for (i = 0; i <= 13; i++)
2121 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2122 sg.info.bb_counters[i] : 0);
2123 seq_printf(seq, " ]\n");
2128 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2132 static const struct seq_operations ext4_mb_seq_groups_ops = {
2133 .start = ext4_mb_seq_groups_start,
2134 .next = ext4_mb_seq_groups_next,
2135 .stop = ext4_mb_seq_groups_stop,
2136 .show = ext4_mb_seq_groups_show,
2139 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2141 struct super_block *sb = PDE(inode)->data;
2144 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2146 struct seq_file *m = file->private_data;
2153 static const struct file_operations ext4_mb_seq_groups_fops = {
2154 .owner = THIS_MODULE,
2155 .open = ext4_mb_seq_groups_open,
2157 .llseek = seq_lseek,
2158 .release = seq_release,
2161 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2163 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2164 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2171 * Allocate the top-level s_group_info array for the specified number
2174 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2176 struct ext4_sb_info *sbi = EXT4_SB(sb);
2178 struct ext4_group_info ***new_groupinfo;
2180 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2181 EXT4_DESC_PER_BLOCK_BITS(sb);
2182 if (size <= sbi->s_group_info_size)
2185 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2186 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2187 if (!new_groupinfo) {
2188 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2191 if (sbi->s_group_info) {
2192 memcpy(new_groupinfo, sbi->s_group_info,
2193 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2194 ext4_kvfree(sbi->s_group_info);
2196 sbi->s_group_info = new_groupinfo;
2197 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2198 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2199 sbi->s_group_info_size);
2203 /* Create and initialize ext4_group_info data for the given group. */
2204 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2205 struct ext4_group_desc *desc)
2209 struct ext4_sb_info *sbi = EXT4_SB(sb);
2210 struct ext4_group_info **meta_group_info;
2211 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2214 * First check if this group is the first of a reserved block.
2215 * If it's true, we have to allocate a new table of pointers
2216 * to ext4_group_info structures
2218 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2219 metalen = sizeof(*meta_group_info) <<
2220 EXT4_DESC_PER_BLOCK_BITS(sb);
2221 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2222 if (meta_group_info == NULL) {
2223 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2224 "for a buddy group");
2225 goto exit_meta_group_info;
2227 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2232 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2233 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2235 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
2236 if (meta_group_info[i] == NULL) {
2237 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2238 goto exit_group_info;
2240 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2241 &(meta_group_info[i]->bb_state));
2244 * initialize bb_free to be able to skip
2245 * empty groups without initialization
2247 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2248 meta_group_info[i]->bb_free =
2249 ext4_free_clusters_after_init(sb, group, desc);
2251 meta_group_info[i]->bb_free =
2252 ext4_free_group_clusters(sb, desc);
2255 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2256 init_rwsem(&meta_group_info[i]->alloc_sem);
2257 meta_group_info[i]->bb_free_root = RB_ROOT;
2258 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2262 struct buffer_head *bh;
2263 meta_group_info[i]->bb_bitmap =
2264 kmalloc(sb->s_blocksize, GFP_KERNEL);
2265 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2266 bh = ext4_read_block_bitmap(sb, group);
2268 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2277 /* If a meta_group_info table has been allocated, release it now */
2278 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2279 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2280 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2282 exit_meta_group_info:
2284 } /* ext4_mb_add_groupinfo */
2286 static int ext4_mb_init_backend(struct super_block *sb)
2288 ext4_group_t ngroups = ext4_get_groups_count(sb);
2290 struct ext4_sb_info *sbi = EXT4_SB(sb);
2292 struct ext4_group_desc *desc;
2293 struct kmem_cache *cachep;
2295 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2299 sbi->s_buddy_cache = new_inode(sb);
2300 if (sbi->s_buddy_cache == NULL) {
2301 ext4_msg(sb, KERN_ERR, "can't get new inode");
2304 /* To avoid potentially colliding with an valid on-disk inode number,
2305 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2306 * not in the inode hash, so it should never be found by iget(), but
2307 * this will avoid confusion if it ever shows up during debugging. */
2308 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2309 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2310 for (i = 0; i < ngroups; i++) {
2311 desc = ext4_get_group_desc(sb, i, NULL);
2313 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2316 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2323 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2325 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2326 i = sbi->s_group_info_size;
2328 kfree(sbi->s_group_info[i]);
2329 iput(sbi->s_buddy_cache);
2331 ext4_kvfree(sbi->s_group_info);
2335 static void ext4_groupinfo_destroy_slabs(void)
2339 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2340 if (ext4_groupinfo_caches[i])
2341 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2342 ext4_groupinfo_caches[i] = NULL;
2346 static int ext4_groupinfo_create_slab(size_t size)
2348 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2350 int blocksize_bits = order_base_2(size);
2351 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2352 struct kmem_cache *cachep;
2354 if (cache_index >= NR_GRPINFO_CACHES)
2357 if (unlikely(cache_index < 0))
2360 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2361 if (ext4_groupinfo_caches[cache_index]) {
2362 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2363 return 0; /* Already created */
2366 slab_size = offsetof(struct ext4_group_info,
2367 bb_counters[blocksize_bits + 2]);
2369 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2370 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2373 ext4_groupinfo_caches[cache_index] = cachep;
2375 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2378 "EXT4-fs: no memory for groupinfo slab cache\n");
2385 int ext4_mb_init(struct super_block *sb)
2387 struct ext4_sb_info *sbi = EXT4_SB(sb);
2393 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2395 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2396 if (sbi->s_mb_offsets == NULL) {
2401 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2402 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2403 if (sbi->s_mb_maxs == NULL) {
2408 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2412 /* order 0 is regular bitmap */
2413 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2414 sbi->s_mb_offsets[0] = 0;
2418 max = sb->s_blocksize << 2;
2420 sbi->s_mb_offsets[i] = offset;
2421 sbi->s_mb_maxs[i] = max;
2422 offset += 1 << (sb->s_blocksize_bits - i);
2425 } while (i <= sb->s_blocksize_bits + 1);
2427 spin_lock_init(&sbi->s_md_lock);
2428 spin_lock_init(&sbi->s_bal_lock);
2430 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2431 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2432 sbi->s_mb_stats = MB_DEFAULT_STATS;
2433 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2434 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2436 * The default group preallocation is 512, which for 4k block
2437 * sizes translates to 2 megabytes. However for bigalloc file
2438 * systems, this is probably too big (i.e, if the cluster size
2439 * is 1 megabyte, then group preallocation size becomes half a
2440 * gigabyte!). As a default, we will keep a two megabyte
2441 * group pralloc size for cluster sizes up to 64k, and after
2442 * that, we will force a minimum group preallocation size of
2443 * 32 clusters. This translates to 8 megs when the cluster
2444 * size is 256k, and 32 megs when the cluster size is 1 meg,
2445 * which seems reasonable as a default.
2447 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2448 sbi->s_cluster_bits, 32);
2450 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2451 * to the lowest multiple of s_stripe which is bigger than
2452 * the s_mb_group_prealloc as determined above. We want
2453 * the preallocation size to be an exact multiple of the
2454 * RAID stripe size so that preallocations don't fragment
2457 if (sbi->s_stripe > 1) {
2458 sbi->s_mb_group_prealloc = roundup(
2459 sbi->s_mb_group_prealloc, sbi->s_stripe);
2462 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2463 if (sbi->s_locality_groups == NULL) {
2465 goto out_free_groupinfo_slab;
2467 for_each_possible_cpu(i) {
2468 struct ext4_locality_group *lg;
2469 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2470 mutex_init(&lg->lg_mutex);
2471 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2472 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2473 spin_lock_init(&lg->lg_prealloc_lock);
2476 /* init file for buddy data */
2477 ret = ext4_mb_init_backend(sb);
2479 goto out_free_locality_groups;
2482 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2483 &ext4_mb_seq_groups_fops, sb);
2487 out_free_locality_groups:
2488 free_percpu(sbi->s_locality_groups);
2489 sbi->s_locality_groups = NULL;
2490 out_free_groupinfo_slab:
2491 ext4_groupinfo_destroy_slabs();
2493 kfree(sbi->s_mb_offsets);
2494 sbi->s_mb_offsets = NULL;
2495 kfree(sbi->s_mb_maxs);
2496 sbi->s_mb_maxs = NULL;
2500 /* need to called with the ext4 group lock held */
2501 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2503 struct ext4_prealloc_space *pa;
2504 struct list_head *cur, *tmp;
2507 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2508 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2509 list_del(&pa->pa_group_list);
2511 kmem_cache_free(ext4_pspace_cachep, pa);
2514 mb_debug(1, "mballoc: %u PAs left\n", count);
2518 int ext4_mb_release(struct super_block *sb)
2520 ext4_group_t ngroups = ext4_get_groups_count(sb);
2522 int num_meta_group_infos;
2523 struct ext4_group_info *grinfo;
2524 struct ext4_sb_info *sbi = EXT4_SB(sb);
2525 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2528 remove_proc_entry("mb_groups", sbi->s_proc);
2530 if (sbi->s_group_info) {
2531 for (i = 0; i < ngroups; i++) {
2532 grinfo = ext4_get_group_info(sb, i);
2534 kfree(grinfo->bb_bitmap);
2536 ext4_lock_group(sb, i);
2537 ext4_mb_cleanup_pa(grinfo);
2538 ext4_unlock_group(sb, i);
2539 kmem_cache_free(cachep, grinfo);
2541 num_meta_group_infos = (ngroups +
2542 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2543 EXT4_DESC_PER_BLOCK_BITS(sb);
2544 for (i = 0; i < num_meta_group_infos; i++)
2545 kfree(sbi->s_group_info[i]);
2546 ext4_kvfree(sbi->s_group_info);
2548 kfree(sbi->s_mb_offsets);
2549 kfree(sbi->s_mb_maxs);
2550 if (sbi->s_buddy_cache)
2551 iput(sbi->s_buddy_cache);
2552 if (sbi->s_mb_stats) {
2553 ext4_msg(sb, KERN_INFO,
2554 "mballoc: %u blocks %u reqs (%u success)",
2555 atomic_read(&sbi->s_bal_allocated),
2556 atomic_read(&sbi->s_bal_reqs),
2557 atomic_read(&sbi->s_bal_success));
2558 ext4_msg(sb, KERN_INFO,
2559 "mballoc: %u extents scanned, %u goal hits, "
2560 "%u 2^N hits, %u breaks, %u lost",
2561 atomic_read(&sbi->s_bal_ex_scanned),
2562 atomic_read(&sbi->s_bal_goals),
2563 atomic_read(&sbi->s_bal_2orders),
2564 atomic_read(&sbi->s_bal_breaks),
2565 atomic_read(&sbi->s_mb_lost_chunks));
2566 ext4_msg(sb, KERN_INFO,
2567 "mballoc: %lu generated and it took %Lu",
2568 sbi->s_mb_buddies_generated,
2569 sbi->s_mb_generation_time);
2570 ext4_msg(sb, KERN_INFO,
2571 "mballoc: %u preallocated, %u discarded",
2572 atomic_read(&sbi->s_mb_preallocated),
2573 atomic_read(&sbi->s_mb_discarded));
2576 free_percpu(sbi->s_locality_groups);
2581 static inline int ext4_issue_discard(struct super_block *sb,
2582 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2584 ext4_fsblk_t discard_block;
2586 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2587 ext4_group_first_block_no(sb, block_group));
2588 count = EXT4_C2B(EXT4_SB(sb), count);
2589 trace_ext4_discard_blocks(sb,
2590 (unsigned long long) discard_block, count);
2591 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2595 * This function is called by the jbd2 layer once the commit has finished,
2596 * so we know we can free the blocks that were released with that commit.
2598 static void ext4_free_data_callback(struct super_block *sb,
2599 struct ext4_journal_cb_entry *jce,
2602 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2603 struct ext4_buddy e4b;
2604 struct ext4_group_info *db;
2605 int err, count = 0, count2 = 0;
2607 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2608 entry->efd_count, entry->efd_group, entry);
2610 if (test_opt(sb, DISCARD)) {
2611 err = ext4_issue_discard(sb, entry->efd_group,
2612 entry->efd_start_cluster,
2614 if (err && err != -EOPNOTSUPP)
2615 ext4_msg(sb, KERN_WARNING, "discard request in"
2616 " group:%d block:%d count:%d failed"
2617 " with %d", entry->efd_group,
2618 entry->efd_start_cluster,
2619 entry->efd_count, err);
2622 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2623 /* we expect to find existing buddy because it's pinned */
2628 /* there are blocks to put in buddy to make them really free */
2629 count += entry->efd_count;
2631 ext4_lock_group(sb, entry->efd_group);
2632 /* Take it out of per group rb tree */
2633 rb_erase(&entry->efd_node, &(db->bb_free_root));
2634 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2637 * Clear the trimmed flag for the group so that the next
2638 * ext4_trim_fs can trim it.
2639 * If the volume is mounted with -o discard, online discard
2640 * is supported and the free blocks will be trimmed online.
2642 if (!test_opt(sb, DISCARD))
2643 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2645 if (!db->bb_free_root.rb_node) {
2646 /* No more items in the per group rb tree
2647 * balance refcounts from ext4_mb_free_metadata()
2649 page_cache_release(e4b.bd_buddy_page);
2650 page_cache_release(e4b.bd_bitmap_page);
2652 ext4_unlock_group(sb, entry->efd_group);
2653 kmem_cache_free(ext4_free_data_cachep, entry);
2654 ext4_mb_unload_buddy(&e4b);
2656 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2659 #ifdef CONFIG_EXT4_DEBUG
2660 u8 mb_enable_debug __read_mostly;
2662 static struct dentry *debugfs_dir;
2663 static struct dentry *debugfs_debug;
2665 static void __init ext4_create_debugfs_entry(void)
2667 debugfs_dir = debugfs_create_dir("ext4", NULL);
2669 debugfs_debug = debugfs_create_u8("mballoc-debug",
2675 static void ext4_remove_debugfs_entry(void)
2677 debugfs_remove(debugfs_debug);
2678 debugfs_remove(debugfs_dir);
2683 static void __init ext4_create_debugfs_entry(void)
2687 static void ext4_remove_debugfs_entry(void)
2693 int __init ext4_init_mballoc(void)
2695 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2696 SLAB_RECLAIM_ACCOUNT);
2697 if (ext4_pspace_cachep == NULL)
2700 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2701 SLAB_RECLAIM_ACCOUNT);
2702 if (ext4_ac_cachep == NULL) {
2703 kmem_cache_destroy(ext4_pspace_cachep);
2707 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2708 SLAB_RECLAIM_ACCOUNT);
2709 if (ext4_free_data_cachep == NULL) {
2710 kmem_cache_destroy(ext4_pspace_cachep);
2711 kmem_cache_destroy(ext4_ac_cachep);
2714 ext4_create_debugfs_entry();
2718 void ext4_exit_mballoc(void)
2721 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2722 * before destroying the slab cache.
2725 kmem_cache_destroy(ext4_pspace_cachep);
2726 kmem_cache_destroy(ext4_ac_cachep);
2727 kmem_cache_destroy(ext4_free_data_cachep);
2728 ext4_groupinfo_destroy_slabs();
2729 ext4_remove_debugfs_entry();
2734 * Check quota and mark chosen 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,
2739 handle_t *handle, unsigned int reserv_clstrs)
2741 struct buffer_head *bitmap_bh = NULL;
2742 struct ext4_group_desc *gdp;
2743 struct buffer_head *gdp_bh;
2744 struct ext4_sb_info *sbi;
2745 struct super_block *sb;
2749 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2750 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2756 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2760 err = ext4_journal_get_write_access(handle, bitmap_bh);
2765 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2769 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2770 ext4_free_group_clusters(sb, gdp));
2772 err = ext4_journal_get_write_access(handle, gdp_bh);
2776 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2778 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2779 if (!ext4_data_block_valid(sbi, block, len)) {
2780 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2781 "fs metadata", block, block+len);
2782 /* File system mounted not to panic on error
2783 * Fix the bitmap and repeat the block allocation
2784 * We leak some of the blocks here.
2786 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2787 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2788 ac->ac_b_ex.fe_len);
2789 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2790 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2796 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2797 #ifdef AGGRESSIVE_CHECK
2800 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2801 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2802 bitmap_bh->b_data));
2806 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2807 ac->ac_b_ex.fe_len);
2808 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2809 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2810 ext4_free_group_clusters_set(sb, gdp,
2811 ext4_free_clusters_after_init(sb,
2812 ac->ac_b_ex.fe_group, gdp));
2814 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2815 ext4_free_group_clusters_set(sb, gdp, len);
2816 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2817 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2819 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2820 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2822 * Now reduce the dirty block count also. Should not go negative
2824 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2825 /* release all the reserved blocks if non delalloc */
2826 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2829 if (sbi->s_log_groups_per_flex) {
2830 ext4_group_t flex_group = ext4_flex_group(sbi,
2831 ac->ac_b_ex.fe_group);
2832 atomic_sub(ac->ac_b_ex.fe_len,
2833 &sbi->s_flex_groups[flex_group].free_clusters);
2836 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2839 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2847 * here we normalize request for locality group
2848 * Group request are normalized to s_mb_group_prealloc, which goes to
2849 * s_strip if we set the same via mount option.
2850 * s_mb_group_prealloc can be configured via
2851 * /sys/fs/ext4/<partition>/mb_group_prealloc
2853 * XXX: should we try to preallocate more than the group has now?
2855 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2857 struct super_block *sb = ac->ac_sb;
2858 struct ext4_locality_group *lg = ac->ac_lg;
2861 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2862 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2863 current->pid, ac->ac_g_ex.fe_len);
2867 * Normalization means making request better in terms of
2868 * size and alignment
2870 static noinline_for_stack void
2871 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2872 struct ext4_allocation_request *ar)
2874 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2877 loff_t size, start_off;
2878 loff_t orig_size __maybe_unused;
2880 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2881 struct ext4_prealloc_space *pa;
2883 /* do normalize only data requests, metadata requests
2884 do not need preallocation */
2885 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2888 /* sometime caller may want exact blocks */
2889 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2892 /* caller may indicate that preallocation isn't
2893 * required (it's a tail, for example) */
2894 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2897 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2898 ext4_mb_normalize_group_request(ac);
2902 bsbits = ac->ac_sb->s_blocksize_bits;
2904 /* first, let's learn actual file size
2905 * given current request is allocated */
2906 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
2907 size = size << bsbits;
2908 if (size < i_size_read(ac->ac_inode))
2909 size = i_size_read(ac->ac_inode);
2912 /* max size of free chunks */
2915 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2916 (req <= (size) || max <= (chunk_size))
2918 /* first, try to predict filesize */
2919 /* XXX: should this table be tunable? */
2921 if (size <= 16 * 1024) {
2923 } else if (size <= 32 * 1024) {
2925 } else if (size <= 64 * 1024) {
2927 } else if (size <= 128 * 1024) {
2929 } else if (size <= 256 * 1024) {
2931 } else if (size <= 512 * 1024) {
2933 } else if (size <= 1024 * 1024) {
2935 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2936 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2937 (21 - bsbits)) << 21;
2938 size = 2 * 1024 * 1024;
2939 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2940 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2941 (22 - bsbits)) << 22;
2942 size = 4 * 1024 * 1024;
2943 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2944 (8<<20)>>bsbits, max, 8 * 1024)) {
2945 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2946 (23 - bsbits)) << 23;
2947 size = 8 * 1024 * 1024;
2949 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2950 size = ac->ac_o_ex.fe_len << bsbits;
2952 size = size >> bsbits;
2953 start = start_off >> bsbits;
2955 /* don't cover already allocated blocks in selected range */
2956 if (ar->pleft && start <= ar->lleft) {
2957 size -= ar->lleft + 1 - start;
2958 start = ar->lleft + 1;
2960 if (ar->pright && start + size - 1 >= ar->lright)
2961 size -= start + size - ar->lright;
2965 /* check we don't cross already preallocated blocks */
2967 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2972 spin_lock(&pa->pa_lock);
2973 if (pa->pa_deleted) {
2974 spin_unlock(&pa->pa_lock);
2978 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
2981 /* PA must not overlap original request */
2982 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2983 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2985 /* skip PAs this normalized request doesn't overlap with */
2986 if (pa->pa_lstart >= end || pa_end <= start) {
2987 spin_unlock(&pa->pa_lock);
2990 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2992 /* adjust start or end to be adjacent to this pa */
2993 if (pa_end <= ac->ac_o_ex.fe_logical) {
2994 BUG_ON(pa_end < start);
2996 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
2997 BUG_ON(pa->pa_lstart > end);
2998 end = pa->pa_lstart;
3000 spin_unlock(&pa->pa_lock);
3005 /* XXX: extra loop to check we really don't overlap preallocations */
3007 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3010 spin_lock(&pa->pa_lock);
3011 if (pa->pa_deleted == 0) {
3012 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3014 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3016 spin_unlock(&pa->pa_lock);
3020 if (start + size <= ac->ac_o_ex.fe_logical &&
3021 start > ac->ac_o_ex.fe_logical) {
3022 ext4_msg(ac->ac_sb, KERN_ERR,
3023 "start %lu, size %lu, fe_logical %lu",
3024 (unsigned long) start, (unsigned long) size,
3025 (unsigned long) ac->ac_o_ex.fe_logical);
3027 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3028 start > ac->ac_o_ex.fe_logical);
3029 BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
3031 /* now prepare goal request */
3033 /* XXX: is it better to align blocks WRT to logical
3034 * placement or satisfy big request as is */
3035 ac->ac_g_ex.fe_logical = start;
3036 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3038 /* define goal start in order to merge */
3039 if (ar->pright && (ar->lright == (start + size))) {
3040 /* merge to the right */
3041 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3042 &ac->ac_f_ex.fe_group,
3043 &ac->ac_f_ex.fe_start);
3044 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3046 if (ar->pleft && (ar->lleft + 1 == start)) {
3047 /* merge to the left */
3048 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3049 &ac->ac_f_ex.fe_group,
3050 &ac->ac_f_ex.fe_start);
3051 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3054 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3055 (unsigned) orig_size, (unsigned) start);
3058 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3060 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3062 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3063 atomic_inc(&sbi->s_bal_reqs);
3064 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3065 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3066 atomic_inc(&sbi->s_bal_success);
3067 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3068 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3069 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3070 atomic_inc(&sbi->s_bal_goals);
3071 if (ac->ac_found > sbi->s_mb_max_to_scan)
3072 atomic_inc(&sbi->s_bal_breaks);
3075 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3076 trace_ext4_mballoc_alloc(ac);
3078 trace_ext4_mballoc_prealloc(ac);
3082 * Called on failure; free up any blocks from the inode PA for this
3083 * context. We don't need this for MB_GROUP_PA because we only change
3084 * pa_free in ext4_mb_release_context(), but on failure, we've already
3085 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3087 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3089 struct ext4_prealloc_space *pa = ac->ac_pa;
3091 if (pa && pa->pa_type == MB_INODE_PA)
3092 pa->pa_free += ac->ac_b_ex.fe_len;
3096 * use blocks preallocated to inode
3098 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3099 struct ext4_prealloc_space *pa)
3101 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3106 /* found preallocated blocks, use them */
3107 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3108 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3109 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3110 len = EXT4_NUM_B2C(sbi, end - start);
3111 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3112 &ac->ac_b_ex.fe_start);
3113 ac->ac_b_ex.fe_len = len;
3114 ac->ac_status = AC_STATUS_FOUND;
3117 BUG_ON(start < pa->pa_pstart);
3118 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3119 BUG_ON(pa->pa_free < len);
3122 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3126 * use blocks preallocated to locality group
3128 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3129 struct ext4_prealloc_space *pa)
3131 unsigned int len = ac->ac_o_ex.fe_len;
3133 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3134 &ac->ac_b_ex.fe_group,
3135 &ac->ac_b_ex.fe_start);
3136 ac->ac_b_ex.fe_len = len;
3137 ac->ac_status = AC_STATUS_FOUND;
3140 /* we don't correct pa_pstart or pa_plen here to avoid
3141 * possible race when the group is being loaded concurrently
3142 * instead we correct pa later, after blocks are marked
3143 * in on-disk bitmap -- see ext4_mb_release_context()
3144 * Other CPUs are prevented from allocating from this pa by lg_mutex
3146 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3150 * Return the prealloc space that have minimal distance
3151 * from the goal block. @cpa is the prealloc
3152 * space that is having currently known minimal distance
3153 * from the goal block.
3155 static struct ext4_prealloc_space *
3156 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3157 struct ext4_prealloc_space *pa,
3158 struct ext4_prealloc_space *cpa)
3160 ext4_fsblk_t cur_distance, new_distance;
3163 atomic_inc(&pa->pa_count);
3166 cur_distance = abs(goal_block - cpa->pa_pstart);
3167 new_distance = abs(goal_block - pa->pa_pstart);
3169 if (cur_distance <= new_distance)
3172 /* drop the previous reference */
3173 atomic_dec(&cpa->pa_count);
3174 atomic_inc(&pa->pa_count);
3179 * search goal blocks in preallocated space
3181 static noinline_for_stack int
3182 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3184 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3186 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3187 struct ext4_locality_group *lg;
3188 struct ext4_prealloc_space *pa, *cpa = NULL;
3189 ext4_fsblk_t goal_block;
3191 /* only data can be preallocated */
3192 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3195 /* first, try per-file preallocation */
3197 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3199 /* all fields in this condition don't change,
3200 * so we can skip locking for them */
3201 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3202 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3203 EXT4_C2B(sbi, pa->pa_len)))
3206 /* non-extent files can't have physical blocks past 2^32 */
3207 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3208 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3209 EXT4_MAX_BLOCK_FILE_PHYS))
3212 /* found preallocated blocks, use them */
3213 spin_lock(&pa->pa_lock);
3214 if (pa->pa_deleted == 0 && pa->pa_free) {
3215 atomic_inc(&pa->pa_count);
3216 ext4_mb_use_inode_pa(ac, pa);
3217 spin_unlock(&pa->pa_lock);
3218 ac->ac_criteria = 10;
3222 spin_unlock(&pa->pa_lock);
3226 /* can we use group allocation? */
3227 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3230 /* inode may have no locality group for some reason */
3234 order = fls(ac->ac_o_ex.fe_len) - 1;
3235 if (order > PREALLOC_TB_SIZE - 1)
3236 /* The max size of hash table is PREALLOC_TB_SIZE */
3237 order = PREALLOC_TB_SIZE - 1;
3239 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3241 * search for the prealloc space that is having
3242 * minimal distance from the goal block.
3244 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3246 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3248 spin_lock(&pa->pa_lock);
3249 if (pa->pa_deleted == 0 &&
3250 pa->pa_free >= ac->ac_o_ex.fe_len) {
3252 cpa = ext4_mb_check_group_pa(goal_block,
3255 spin_unlock(&pa->pa_lock);
3260 ext4_mb_use_group_pa(ac, cpa);
3261 ac->ac_criteria = 20;
3268 * the function goes through all block freed in the group
3269 * but not yet committed and marks them used in in-core bitmap.
3270 * buddy must be generated from this bitmap
3271 * Need to be called with the ext4 group lock held
3273 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3277 struct ext4_group_info *grp;
3278 struct ext4_free_data *entry;
3280 grp = ext4_get_group_info(sb, group);
3281 n = rb_first(&(grp->bb_free_root));
3284 entry = rb_entry(n, struct ext4_free_data, efd_node);
3285 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3292 * the function goes through all preallocation in this group and marks them
3293 * used in in-core bitmap. buddy must be generated from this bitmap
3294 * Need to be called with ext4 group lock held
3296 static noinline_for_stack
3297 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3300 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3301 struct ext4_prealloc_space *pa;
3302 struct list_head *cur;
3303 ext4_group_t groupnr;
3304 ext4_grpblk_t start;
3305 int preallocated = 0;
3308 /* all form of preallocation discards first load group,
3309 * so the only competing code is preallocation use.
3310 * we don't need any locking here
3311 * notice we do NOT ignore preallocations with pa_deleted
3312 * otherwise we could leave used blocks available for
3313 * allocation in buddy when concurrent ext4_mb_put_pa()
3314 * is dropping preallocation
3316 list_for_each(cur, &grp->bb_prealloc_list) {
3317 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3318 spin_lock(&pa->pa_lock);
3319 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3322 spin_unlock(&pa->pa_lock);
3323 if (unlikely(len == 0))
3325 BUG_ON(groupnr != group);
3326 ext4_set_bits(bitmap, start, len);
3327 preallocated += len;
3329 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3332 static void ext4_mb_pa_callback(struct rcu_head *head)
3334 struct ext4_prealloc_space *pa;
3335 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3336 kmem_cache_free(ext4_pspace_cachep, pa);
3340 * drops a reference to preallocated space descriptor
3341 * if this was the last reference and the space is consumed
3343 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3344 struct super_block *sb, struct ext4_prealloc_space *pa)
3347 ext4_fsblk_t grp_blk;
3349 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3352 /* in this short window concurrent discard can set pa_deleted */
3353 spin_lock(&pa->pa_lock);
3354 if (pa->pa_deleted == 1) {
3355 spin_unlock(&pa->pa_lock);
3360 spin_unlock(&pa->pa_lock);
3362 grp_blk = pa->pa_pstart;
3364 * If doing group-based preallocation, pa_pstart may be in the
3365 * next group when pa is used up
3367 if (pa->pa_type == MB_GROUP_PA)
3370 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3375 * P1 (buddy init) P2 (regular allocation)
3376 * find block B in PA
3377 * copy on-disk bitmap to buddy
3378 * mark B in on-disk bitmap
3379 * drop PA from group
3380 * mark all PAs in buddy
3382 * thus, P1 initializes buddy with B available. to prevent this
3383 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3386 ext4_lock_group(sb, grp);
3387 list_del(&pa->pa_group_list);
3388 ext4_unlock_group(sb, grp);
3390 spin_lock(pa->pa_obj_lock);
3391 list_del_rcu(&pa->pa_inode_list);
3392 spin_unlock(pa->pa_obj_lock);
3394 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3398 * creates new preallocated space for given inode
3400 static noinline_for_stack int
3401 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3403 struct super_block *sb = ac->ac_sb;
3404 struct ext4_sb_info *sbi = EXT4_SB(sb);
3405 struct ext4_prealloc_space *pa;
3406 struct ext4_group_info *grp;
3407 struct ext4_inode_info *ei;
3409 /* preallocate only when found space is larger then requested */
3410 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3411 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3412 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3414 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3418 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3424 /* we can't allocate as much as normalizer wants.
3425 * so, found space must get proper lstart
3426 * to cover original request */
3427 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3428 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3430 /* we're limited by original request in that
3431 * logical block must be covered any way
3432 * winl is window we can move our chunk within */
3433 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3435 /* also, we should cover whole original request */
3436 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3438 /* the smallest one defines real window */
3439 win = min(winl, wins);
3441 offs = ac->ac_o_ex.fe_logical %
3442 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3443 if (offs && offs < win)
3446 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3448 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3449 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3452 /* preallocation can change ac_b_ex, thus we store actually
3453 * allocated blocks for history */
3454 ac->ac_f_ex = ac->ac_b_ex;
3456 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3457 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3458 pa->pa_len = ac->ac_b_ex.fe_len;
3459 pa->pa_free = pa->pa_len;
3460 atomic_set(&pa->pa_count, 1);
3461 spin_lock_init(&pa->pa_lock);
3462 INIT_LIST_HEAD(&pa->pa_inode_list);
3463 INIT_LIST_HEAD(&pa->pa_group_list);
3465 pa->pa_type = MB_INODE_PA;
3467 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3468 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3469 trace_ext4_mb_new_inode_pa(ac, pa);
3471 ext4_mb_use_inode_pa(ac, pa);
3472 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3474 ei = EXT4_I(ac->ac_inode);
3475 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3477 pa->pa_obj_lock = &ei->i_prealloc_lock;
3478 pa->pa_inode = ac->ac_inode;
3480 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3481 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3482 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3484 spin_lock(pa->pa_obj_lock);
3485 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3486 spin_unlock(pa->pa_obj_lock);
3492 * creates new preallocated space for locality group inodes belongs to
3494 static noinline_for_stack int
3495 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3497 struct super_block *sb = ac->ac_sb;
3498 struct ext4_locality_group *lg;
3499 struct ext4_prealloc_space *pa;
3500 struct ext4_group_info *grp;
3502 /* preallocate only when found space is larger then requested */
3503 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3504 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3505 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3507 BUG_ON(ext4_pspace_cachep == NULL);
3508 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3512 /* preallocation can change ac_b_ex, thus we store actually
3513 * allocated blocks for history */
3514 ac->ac_f_ex = ac->ac_b_ex;
3516 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3517 pa->pa_lstart = pa->pa_pstart;
3518 pa->pa_len = ac->ac_b_ex.fe_len;
3519 pa->pa_free = pa->pa_len;
3520 atomic_set(&pa->pa_count, 1);
3521 spin_lock_init(&pa->pa_lock);
3522 INIT_LIST_HEAD(&pa->pa_inode_list);
3523 INIT_LIST_HEAD(&pa->pa_group_list);
3525 pa->pa_type = MB_GROUP_PA;
3527 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3528 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3529 trace_ext4_mb_new_group_pa(ac, pa);
3531 ext4_mb_use_group_pa(ac, pa);
3532 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3534 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3538 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3539 pa->pa_inode = NULL;
3541 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3542 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3543 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3546 * We will later add the new pa to the right bucket
3547 * after updating the pa_free in ext4_mb_release_context
3552 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3556 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3557 err = ext4_mb_new_group_pa(ac);
3559 err = ext4_mb_new_inode_pa(ac);
3564 * finds all unused blocks in on-disk bitmap, frees them in
3565 * in-core bitmap and buddy.
3566 * @pa must be unlinked from inode and group lists, so that
3567 * nobody else can find/use it.
3568 * the caller MUST hold group/inode locks.
3569 * TODO: optimize the case when there are no in-core structures yet
3571 static noinline_for_stack int
3572 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3573 struct ext4_prealloc_space *pa)
3575 struct super_block *sb = e4b->bd_sb;
3576 struct ext4_sb_info *sbi = EXT4_SB(sb);
3581 unsigned long long grp_blk_start;
3585 BUG_ON(pa->pa_deleted == 0);
3586 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3587 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3588 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3589 end = bit + pa->pa_len;
3592 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3595 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3596 mb_debug(1, " free preallocated %u/%u in group %u\n",
3597 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3598 (unsigned) next - bit, (unsigned) group);
3601 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3602 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3603 EXT4_C2B(sbi, bit)),
3605 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3608 if (free != pa->pa_free) {
3609 ext4_msg(e4b->bd_sb, KERN_CRIT,
3610 "pa %p: logic %lu, phys. %lu, len %lu",
3611 pa, (unsigned long) pa->pa_lstart,
3612 (unsigned long) pa->pa_pstart,
3613 (unsigned long) pa->pa_len);
3614 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3617 * pa is already deleted so we use the value obtained
3618 * from the bitmap and continue.
3621 atomic_add(free, &sbi->s_mb_discarded);
3626 static noinline_for_stack int
3627 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3628 struct ext4_prealloc_space *pa)
3630 struct super_block *sb = e4b->bd_sb;
3634 trace_ext4_mb_release_group_pa(sb, pa);
3635 BUG_ON(pa->pa_deleted == 0);
3636 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3637 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3638 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3639 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3640 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3646 * releases all preallocations in given group
3648 * first, we need to decide discard policy:
3649 * - when do we discard
3651 * - how many do we discard
3652 * 1) how many requested
3654 static noinline_for_stack int
3655 ext4_mb_discard_group_preallocations(struct super_block *sb,
3656 ext4_group_t group, int needed)
3658 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3659 struct buffer_head *bitmap_bh = NULL;
3660 struct ext4_prealloc_space *pa, *tmp;
3661 struct list_head list;
3662 struct ext4_buddy e4b;
3667 mb_debug(1, "discard preallocation for group %u\n", group);
3669 if (list_empty(&grp->bb_prealloc_list))
3672 bitmap_bh = ext4_read_block_bitmap(sb, group);
3673 if (bitmap_bh == NULL) {
3674 ext4_error(sb, "Error reading block bitmap for %u", group);
3678 err = ext4_mb_load_buddy(sb, group, &e4b);
3680 ext4_error(sb, "Error loading buddy information for %u", group);
3686 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3688 INIT_LIST_HEAD(&list);
3690 ext4_lock_group(sb, group);
3691 list_for_each_entry_safe(pa, tmp,
3692 &grp->bb_prealloc_list, pa_group_list) {
3693 spin_lock(&pa->pa_lock);
3694 if (atomic_read(&pa->pa_count)) {
3695 spin_unlock(&pa->pa_lock);
3699 if (pa->pa_deleted) {
3700 spin_unlock(&pa->pa_lock);
3704 /* seems this one can be freed ... */
3707 /* we can trust pa_free ... */
3708 free += pa->pa_free;
3710 spin_unlock(&pa->pa_lock);
3712 list_del(&pa->pa_group_list);
3713 list_add(&pa->u.pa_tmp_list, &list);
3716 /* if we still need more blocks and some PAs were used, try again */
3717 if (free < needed && busy) {
3719 ext4_unlock_group(sb, group);
3721 * Yield the CPU here so that we don't get soft lockup
3722 * in non preempt case.
3728 /* found anything to free? */
3729 if (list_empty(&list)) {
3734 /* now free all selected PAs */
3735 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3737 /* remove from object (inode or locality group) */
3738 spin_lock(pa->pa_obj_lock);
3739 list_del_rcu(&pa->pa_inode_list);
3740 spin_unlock(pa->pa_obj_lock);
3742 if (pa->pa_type == MB_GROUP_PA)
3743 ext4_mb_release_group_pa(&e4b, pa);
3745 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3747 list_del(&pa->u.pa_tmp_list);
3748 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3752 ext4_unlock_group(sb, group);
3753 ext4_mb_unload_buddy(&e4b);
3759 * releases all non-used preallocated blocks for given inode
3761 * It's important to discard preallocations under i_data_sem
3762 * We don't want another block to be served from the prealloc
3763 * space when we are discarding the inode prealloc space.
3765 * FIXME!! Make sure it is valid at all the call sites
3767 void ext4_discard_preallocations(struct inode *inode)
3769 struct ext4_inode_info *ei = EXT4_I(inode);
3770 struct super_block *sb = inode->i_sb;
3771 struct buffer_head *bitmap_bh = NULL;
3772 struct ext4_prealloc_space *pa, *tmp;
3773 ext4_group_t group = 0;
3774 struct list_head list;
3775 struct ext4_buddy e4b;
3778 if (!S_ISREG(inode->i_mode)) {
3779 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3783 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3784 trace_ext4_discard_preallocations(inode);
3786 INIT_LIST_HEAD(&list);
3789 /* first, collect all pa's in the inode */
3790 spin_lock(&ei->i_prealloc_lock);
3791 while (!list_empty(&ei->i_prealloc_list)) {
3792 pa = list_entry(ei->i_prealloc_list.next,
3793 struct ext4_prealloc_space, pa_inode_list);
3794 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3795 spin_lock(&pa->pa_lock);
3796 if (atomic_read(&pa->pa_count)) {
3797 /* this shouldn't happen often - nobody should
3798 * use preallocation while we're discarding it */
3799 spin_unlock(&pa->pa_lock);
3800 spin_unlock(&ei->i_prealloc_lock);
3801 ext4_msg(sb, KERN_ERR,
3802 "uh-oh! used pa while discarding");
3804 schedule_timeout_uninterruptible(HZ);
3808 if (pa->pa_deleted == 0) {
3810 spin_unlock(&pa->pa_lock);
3811 list_del_rcu(&pa->pa_inode_list);
3812 list_add(&pa->u.pa_tmp_list, &list);
3816 /* someone is deleting pa right now */
3817 spin_unlock(&pa->pa_lock);
3818 spin_unlock(&ei->i_prealloc_lock);
3820 /* we have to wait here because pa_deleted
3821 * doesn't mean pa is already unlinked from
3822 * the list. as we might be called from
3823 * ->clear_inode() the inode will get freed
3824 * and concurrent thread which is unlinking
3825 * pa from inode's list may access already
3826 * freed memory, bad-bad-bad */
3828 /* XXX: if this happens too often, we can
3829 * add a flag to force wait only in case
3830 * of ->clear_inode(), but not in case of
3831 * regular truncate */
3832 schedule_timeout_uninterruptible(HZ);
3835 spin_unlock(&ei->i_prealloc_lock);
3837 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3838 BUG_ON(pa->pa_type != MB_INODE_PA);
3839 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3841 err = ext4_mb_load_buddy(sb, group, &e4b);
3843 ext4_error(sb, "Error loading buddy information for %u",
3848 bitmap_bh = ext4_read_block_bitmap(sb, group);
3849 if (bitmap_bh == NULL) {
3850 ext4_error(sb, "Error reading block bitmap for %u",
3852 ext4_mb_unload_buddy(&e4b);
3856 ext4_lock_group(sb, group);
3857 list_del(&pa->pa_group_list);
3858 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3859 ext4_unlock_group(sb, group);
3861 ext4_mb_unload_buddy(&e4b);
3864 list_del(&pa->u.pa_tmp_list);
3865 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3869 #ifdef CONFIG_EXT4_DEBUG
3870 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3872 struct super_block *sb = ac->ac_sb;
3873 ext4_group_t ngroups, i;
3875 if (!mb_enable_debug ||
3876 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3879 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
3880 " Allocation context details:");
3881 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
3882 ac->ac_status, ac->ac_flags);
3883 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
3884 "goal %lu/%lu/%lu@%lu, "
3885 "best %lu/%lu/%lu@%lu cr %d",
3886 (unsigned long)ac->ac_o_ex.fe_group,
3887 (unsigned long)ac->ac_o_ex.fe_start,
3888 (unsigned long)ac->ac_o_ex.fe_len,
3889 (unsigned long)ac->ac_o_ex.fe_logical,
3890 (unsigned long)ac->ac_g_ex.fe_group,
3891 (unsigned long)ac->ac_g_ex.fe_start,
3892 (unsigned long)ac->ac_g_ex.fe_len,
3893 (unsigned long)ac->ac_g_ex.fe_logical,
3894 (unsigned long)ac->ac_b_ex.fe_group,
3895 (unsigned long)ac->ac_b_ex.fe_start,
3896 (unsigned long)ac->ac_b_ex.fe_len,
3897 (unsigned long)ac->ac_b_ex.fe_logical,
3898 (int)ac->ac_criteria);
3899 ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found",
3900 ac->ac_ex_scanned, ac->ac_found);
3901 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
3902 ngroups = ext4_get_groups_count(sb);
3903 for (i = 0; i < ngroups; i++) {
3904 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3905 struct ext4_prealloc_space *pa;
3906 ext4_grpblk_t start;
3907 struct list_head *cur;
3908 ext4_lock_group(sb, i);
3909 list_for_each(cur, &grp->bb_prealloc_list) {
3910 pa = list_entry(cur, struct ext4_prealloc_space,
3912 spin_lock(&pa->pa_lock);
3913 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3915 spin_unlock(&pa->pa_lock);
3916 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3919 ext4_unlock_group(sb, i);
3921 if (grp->bb_free == 0)
3923 printk(KERN_ERR "%u: %d/%d \n",
3924 i, grp->bb_free, grp->bb_fragments);
3926 printk(KERN_ERR "\n");
3929 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3936 * We use locality group preallocation for small size file. The size of the
3937 * file is determined by the current size or the resulting size after
3938 * allocation which ever is larger
3940 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3942 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3944 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3945 int bsbits = ac->ac_sb->s_blocksize_bits;
3948 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3951 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3954 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3955 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3958 if ((size == isize) &&
3959 !ext4_fs_is_busy(sbi) &&
3960 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3961 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3965 if (sbi->s_mb_group_prealloc <= 0) {
3966 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3970 /* don't use group allocation for large files */
3971 size = max(size, isize);
3972 if (size > sbi->s_mb_stream_request) {
3973 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3977 BUG_ON(ac->ac_lg != NULL);
3979 * locality group prealloc space are per cpu. The reason for having
3980 * per cpu locality group is to reduce the contention between block
3981 * request from multiple CPUs.
3983 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
3985 /* we're going to use group allocation */
3986 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3988 /* serialize all allocations in the group */
3989 mutex_lock(&ac->ac_lg->lg_mutex);
3992 static noinline_for_stack int
3993 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
3994 struct ext4_allocation_request *ar)
3996 struct super_block *sb = ar->inode->i_sb;
3997 struct ext4_sb_info *sbi = EXT4_SB(sb);
3998 struct ext4_super_block *es = sbi->s_es;
4002 ext4_grpblk_t block;
4004 /* we can't allocate > group size */
4007 /* just a dirty hack to filter too big requests */
4008 if (len >= EXT4_CLUSTERS_PER_GROUP(sb) - 10)
4009 len = EXT4_CLUSTERS_PER_GROUP(sb) - 10;
4011 /* start searching from the goal */
4013 if (goal < le32_to_cpu(es->s_first_data_block) ||
4014 goal >= ext4_blocks_count(es))
4015 goal = le32_to_cpu(es->s_first_data_block);
4016 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4018 /* set up allocation goals */
4019 ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1);
4020 ac->ac_status = AC_STATUS_CONTINUE;
4022 ac->ac_inode = ar->inode;
4023 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4024 ac->ac_o_ex.fe_group = group;
4025 ac->ac_o_ex.fe_start = block;
4026 ac->ac_o_ex.fe_len = len;
4027 ac->ac_g_ex = ac->ac_o_ex;
4028 ac->ac_flags = ar->flags;
4030 /* we have to define context: we'll we work with a file or
4031 * locality group. this is a policy, actually */
4032 ext4_mb_group_or_file(ac);
4034 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4035 "left: %u/%u, right %u/%u to %swritable\n",
4036 (unsigned) ar->len, (unsigned) ar->logical,
4037 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4038 (unsigned) ar->lleft, (unsigned) ar->pleft,
4039 (unsigned) ar->lright, (unsigned) ar->pright,
4040 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4045 static noinline_for_stack void
4046 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4047 struct ext4_locality_group *lg,
4048 int order, int total_entries)
4050 ext4_group_t group = 0;
4051 struct ext4_buddy e4b;
4052 struct list_head discard_list;
4053 struct ext4_prealloc_space *pa, *tmp;
4055 mb_debug(1, "discard locality group preallocation\n");
4057 INIT_LIST_HEAD(&discard_list);
4059 spin_lock(&lg->lg_prealloc_lock);
4060 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4062 spin_lock(&pa->pa_lock);
4063 if (atomic_read(&pa->pa_count)) {
4065 * This is the pa that we just used
4066 * for block allocation. So don't
4069 spin_unlock(&pa->pa_lock);
4072 if (pa->pa_deleted) {
4073 spin_unlock(&pa->pa_lock);
4076 /* only lg prealloc space */
4077 BUG_ON(pa->pa_type != MB_GROUP_PA);
4079 /* seems this one can be freed ... */
4081 spin_unlock(&pa->pa_lock);
4083 list_del_rcu(&pa->pa_inode_list);
4084 list_add(&pa->u.pa_tmp_list, &discard_list);
4087 if (total_entries <= 5) {
4089 * we want to keep only 5 entries
4090 * allowing it to grow to 8. This
4091 * mak sure we don't call discard
4092 * soon for this list.
4097 spin_unlock(&lg->lg_prealloc_lock);
4099 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4101 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4102 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4103 ext4_error(sb, "Error loading buddy information for %u",
4107 ext4_lock_group(sb, group);
4108 list_del(&pa->pa_group_list);
4109 ext4_mb_release_group_pa(&e4b, pa);
4110 ext4_unlock_group(sb, group);
4112 ext4_mb_unload_buddy(&e4b);
4113 list_del(&pa->u.pa_tmp_list);
4114 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4119 * We have incremented pa_count. So it cannot be freed at this
4120 * point. Also we hold lg_mutex. So no parallel allocation is
4121 * possible from this lg. That means pa_free cannot be updated.
4123 * A parallel ext4_mb_discard_group_preallocations is possible.
4124 * which can cause the lg_prealloc_list to be updated.
4127 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4129 int order, added = 0, lg_prealloc_count = 1;
4130 struct super_block *sb = ac->ac_sb;
4131 struct ext4_locality_group *lg = ac->ac_lg;
4132 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4134 order = fls(pa->pa_free) - 1;
4135 if (order > PREALLOC_TB_SIZE - 1)
4136 /* The max size of hash table is PREALLOC_TB_SIZE */
4137 order = PREALLOC_TB_SIZE - 1;
4138 /* Add the prealloc space to lg */
4140 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4142 spin_lock(&tmp_pa->pa_lock);
4143 if (tmp_pa->pa_deleted) {
4144 spin_unlock(&tmp_pa->pa_lock);
4147 if (!added && pa->pa_free < tmp_pa->pa_free) {
4148 /* Add to the tail of the previous entry */
4149 list_add_tail_rcu(&pa->pa_inode_list,
4150 &tmp_pa->pa_inode_list);
4153 * we want to count the total
4154 * number of entries in the list
4157 spin_unlock(&tmp_pa->pa_lock);
4158 lg_prealloc_count++;
4161 list_add_tail_rcu(&pa->pa_inode_list,
4162 &lg->lg_prealloc_list[order]);
4165 /* Now trim the list to be not more than 8 elements */
4166 if (lg_prealloc_count > 8) {
4167 ext4_mb_discard_lg_preallocations(sb, lg,
4168 order, lg_prealloc_count);
4175 * release all resource we used in allocation
4177 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4179 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4180 struct ext4_prealloc_space *pa = ac->ac_pa;
4182 if (pa->pa_type == MB_GROUP_PA) {
4183 /* see comment in ext4_mb_use_group_pa() */
4184 spin_lock(&pa->pa_lock);
4185 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4186 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4187 pa->pa_free -= ac->ac_b_ex.fe_len;
4188 pa->pa_len -= ac->ac_b_ex.fe_len;
4189 spin_unlock(&pa->pa_lock);
4194 * We want to add the pa to the right bucket.
4195 * Remove it from the list and while adding
4196 * make sure the list to which we are adding
4199 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4200 spin_lock(pa->pa_obj_lock);
4201 list_del_rcu(&pa->pa_inode_list);
4202 spin_unlock(pa->pa_obj_lock);
4203 ext4_mb_add_n_trim(ac);
4205 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4207 if (ac->ac_bitmap_page)
4208 page_cache_release(ac->ac_bitmap_page);
4209 if (ac->ac_buddy_page)
4210 page_cache_release(ac->ac_buddy_page);
4211 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4212 mutex_unlock(&ac->ac_lg->lg_mutex);
4213 ext4_mb_collect_stats(ac);
4217 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4219 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4223 trace_ext4_mb_discard_preallocations(sb, needed);
4224 for (i = 0; i < ngroups && needed > 0; i++) {
4225 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4234 * Main entry point into mballoc to allocate blocks
4235 * it tries to use preallocation first, then falls back
4236 * to usual allocation
4238 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4239 struct ext4_allocation_request *ar, int *errp)
4242 struct ext4_allocation_context *ac = NULL;
4243 struct ext4_sb_info *sbi;
4244 struct super_block *sb;
4245 ext4_fsblk_t block = 0;
4246 unsigned int inquota = 0;
4247 unsigned int reserv_clstrs = 0;
4249 sb = ar->inode->i_sb;
4252 trace_ext4_request_blocks(ar);
4254 /* Allow to use superuser reservation for quota file */
4255 if (IS_NOQUOTA(ar->inode))
4256 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4259 * For delayed allocation, we could skip the ENOSPC and
4260 * EDQUOT check, as blocks and quotas have been already
4261 * reserved when data being copied into pagecache.
4263 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4264 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4266 /* Without delayed allocation we need to verify
4267 * there is enough free blocks to do block allocation
4268 * and verify allocation doesn't exceed the quota limits.
4271 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4273 /* let others to free the space */
4275 ar->len = ar->len >> 1;
4281 reserv_clstrs = ar->len;
4282 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4283 dquot_alloc_block_nofail(ar->inode,
4284 EXT4_C2B(sbi, ar->len));
4287 dquot_alloc_block(ar->inode,
4288 EXT4_C2B(sbi, ar->len))) {
4290 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4301 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4308 *errp = ext4_mb_initialize_context(ac, ar);
4314 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4315 if (!ext4_mb_use_preallocated(ac)) {
4316 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4317 ext4_mb_normalize_request(ac, ar);
4319 /* allocate space in core */
4320 *errp = ext4_mb_regular_allocator(ac);
4322 ext4_discard_allocated_blocks(ac);
4326 /* as we've just preallocated more space than
4327 * user requested orinally, we store allocated
4328 * space in a special descriptor */
4329 if (ac->ac_status == AC_STATUS_FOUND &&
4330 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4331 ext4_mb_new_preallocation(ac);
4333 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4334 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4335 if (*errp == -EAGAIN) {
4337 * drop the reference that we took
4338 * in ext4_mb_use_best_found
4340 ext4_mb_release_context(ac);
4341 ac->ac_b_ex.fe_group = 0;
4342 ac->ac_b_ex.fe_start = 0;
4343 ac->ac_b_ex.fe_len = 0;
4344 ac->ac_status = AC_STATUS_CONTINUE;
4347 ext4_discard_allocated_blocks(ac);
4350 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4351 ar->len = ac->ac_b_ex.fe_len;
4354 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4362 ac->ac_b_ex.fe_len = 0;
4364 ext4_mb_show_ac(ac);
4366 ext4_mb_release_context(ac);
4369 kmem_cache_free(ext4_ac_cachep, ac);
4370 if (inquota && ar->len < inquota)
4371 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4373 if (!ext4_test_inode_state(ar->inode,
4374 EXT4_STATE_DELALLOC_RESERVED))
4375 /* release all the reserved blocks if non delalloc */
4376 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4380 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4386 * We can merge two free data extents only if the physical blocks
4387 * are contiguous, AND the extents were freed by the same transaction,
4388 * AND the blocks are associated with the same group.
4390 static int can_merge(struct ext4_free_data *entry1,
4391 struct ext4_free_data *entry2)
4393 if ((entry1->efd_tid == entry2->efd_tid) &&
4394 (entry1->efd_group == entry2->efd_group) &&
4395 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4400 static noinline_for_stack int
4401 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4402 struct ext4_free_data *new_entry)
4404 ext4_group_t group = e4b->bd_group;
4405 ext4_grpblk_t cluster;
4406 struct ext4_free_data *entry;
4407 struct ext4_group_info *db = e4b->bd_info;
4408 struct super_block *sb = e4b->bd_sb;
4409 struct ext4_sb_info *sbi = EXT4_SB(sb);
4410 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4411 struct rb_node *parent = NULL, *new_node;
4413 BUG_ON(!ext4_handle_valid(handle));
4414 BUG_ON(e4b->bd_bitmap_page == NULL);
4415 BUG_ON(e4b->bd_buddy_page == NULL);
4417 new_node = &new_entry->efd_node;
4418 cluster = new_entry->efd_start_cluster;
4421 /* first free block exent. We need to
4422 protect buddy cache from being freed,
4423 * otherwise we'll refresh it from
4424 * on-disk bitmap and lose not-yet-available
4426 page_cache_get(e4b->bd_buddy_page);
4427 page_cache_get(e4b->bd_bitmap_page);
4431 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4432 if (cluster < entry->efd_start_cluster)
4434 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4435 n = &(*n)->rb_right;
4437 ext4_grp_locked_error(sb, group, 0,
4438 ext4_group_first_block_no(sb, group) +
4439 EXT4_C2B(sbi, cluster),
4440 "Block already on to-be-freed list");
4445 rb_link_node(new_node, parent, n);
4446 rb_insert_color(new_node, &db->bb_free_root);
4448 /* Now try to see the extent can be merged to left and right */
4449 node = rb_prev(new_node);
4451 entry = rb_entry(node, struct ext4_free_data, efd_node);
4452 if (can_merge(entry, new_entry)) {
4453 new_entry->efd_start_cluster = entry->efd_start_cluster;
4454 new_entry->efd_count += entry->efd_count;
4455 rb_erase(node, &(db->bb_free_root));
4456 ext4_journal_callback_del(handle, &entry->efd_jce);
4457 kmem_cache_free(ext4_free_data_cachep, entry);
4461 node = rb_next(new_node);
4463 entry = rb_entry(node, struct ext4_free_data, efd_node);
4464 if (can_merge(new_entry, entry)) {
4465 new_entry->efd_count += entry->efd_count;
4466 rb_erase(node, &(db->bb_free_root));
4467 ext4_journal_callback_del(handle, &entry->efd_jce);
4468 kmem_cache_free(ext4_free_data_cachep, entry);
4471 /* Add the extent to transaction's private list */
4472 ext4_journal_callback_add(handle, ext4_free_data_callback,
4473 &new_entry->efd_jce);
4478 * ext4_free_blocks() -- Free given blocks and update quota
4479 * @handle: handle for this transaction
4481 * @block: start physical block to free
4482 * @count: number of blocks to count
4483 * @flags: flags used by ext4_free_blocks
4485 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4486 struct buffer_head *bh, ext4_fsblk_t block,
4487 unsigned long count, int flags)
4489 struct buffer_head *bitmap_bh = NULL;
4490 struct super_block *sb = inode->i_sb;
4491 struct ext4_group_desc *gdp;
4492 unsigned long freed = 0;
4493 unsigned int overflow;
4495 struct buffer_head *gd_bh;
4496 ext4_group_t block_group;
4497 struct ext4_sb_info *sbi;
4498 struct ext4_buddy e4b;
4499 unsigned int count_clusters;
4505 BUG_ON(block != bh->b_blocknr);
4507 block = bh->b_blocknr;
4511 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4512 !ext4_data_block_valid(sbi, block, count)) {
4513 ext4_error(sb, "Freeing blocks not in datazone - "
4514 "block = %llu, count = %lu", block, count);
4518 ext4_debug("freeing block %llu\n", block);
4519 trace_ext4_free_blocks(inode, block, count, flags);
4521 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4522 struct buffer_head *tbh = bh;
4525 BUG_ON(bh && (count > 1));
4527 for (i = 0; i < count; i++) {
4529 tbh = sb_find_get_block(inode->i_sb,
4533 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4534 inode, tbh, block + i);
4539 * We need to make sure we don't reuse the freed block until
4540 * after the transaction is committed, which we can do by
4541 * treating the block as metadata, below. We make an
4542 * exception if the inode is to be written in writeback mode
4543 * since writeback mode has weak data consistency guarantees.
4545 if (!ext4_should_writeback_data(inode))
4546 flags |= EXT4_FREE_BLOCKS_METADATA;
4549 * If the extent to be freed does not begin on a cluster
4550 * boundary, we need to deal with partial clusters at the
4551 * beginning and end of the extent. Normally we will free
4552 * blocks at the beginning or the end unless we are explicitly
4553 * requested to avoid doing so.
4555 overflow = block & (sbi->s_cluster_ratio - 1);
4557 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4558 overflow = sbi->s_cluster_ratio - overflow;
4560 if (count > overflow)
4569 overflow = count & (sbi->s_cluster_ratio - 1);
4571 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4572 if (count > overflow)
4577 count += sbi->s_cluster_ratio - overflow;
4582 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4585 * Check to see if we are freeing blocks across a group
4588 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4589 overflow = EXT4_C2B(sbi, bit) + count -
4590 EXT4_BLOCKS_PER_GROUP(sb);
4593 count_clusters = EXT4_B2C(sbi, count);
4594 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4599 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4605 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4606 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4607 in_range(block, ext4_inode_table(sb, gdp),
4608 EXT4_SB(sb)->s_itb_per_group) ||
4609 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4610 EXT4_SB(sb)->s_itb_per_group)) {
4612 ext4_error(sb, "Freeing blocks in system zone - "
4613 "Block = %llu, count = %lu", block, count);
4614 /* err = 0. ext4_std_error should be a no op */
4618 BUFFER_TRACE(bitmap_bh, "getting write access");
4619 err = ext4_journal_get_write_access(handle, bitmap_bh);
4624 * We are about to modify some metadata. Call the journal APIs
4625 * to unshare ->b_data if a currently-committing transaction is
4628 BUFFER_TRACE(gd_bh, "get_write_access");
4629 err = ext4_journal_get_write_access(handle, gd_bh);
4632 #ifdef AGGRESSIVE_CHECK
4635 for (i = 0; i < count_clusters; i++)
4636 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4639 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4641 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4645 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4646 struct ext4_free_data *new_entry;
4648 * blocks being freed are metadata. these blocks shouldn't
4649 * be used until this transaction is committed
4651 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4653 ext4_mb_unload_buddy(&e4b);
4657 new_entry->efd_start_cluster = bit;
4658 new_entry->efd_group = block_group;
4659 new_entry->efd_count = count_clusters;
4660 new_entry->efd_tid = handle->h_transaction->t_tid;
4662 ext4_lock_group(sb, block_group);
4663 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4664 ext4_mb_free_metadata(handle, &e4b, new_entry);
4666 /* need to update group_info->bb_free and bitmap
4667 * with group lock held. generate_buddy look at
4668 * them with group lock_held
4670 if (test_opt(sb, DISCARD)) {
4671 err = ext4_issue_discard(sb, block_group, bit, count);
4672 if (err && err != -EOPNOTSUPP)
4673 ext4_msg(sb, KERN_WARNING, "discard request in"
4674 " group:%d block:%d count:%lu failed"
4675 " with %d", block_group, bit, count,
4680 ext4_lock_group(sb, block_group);
4681 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4682 mb_free_blocks(inode, &e4b, bit, count_clusters);
4685 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4686 ext4_free_group_clusters_set(sb, gdp, ret);
4687 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4688 ext4_group_desc_csum_set(sb, block_group, gdp);
4689 ext4_unlock_group(sb, block_group);
4690 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4692 if (sbi->s_log_groups_per_flex) {
4693 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4694 atomic_add(count_clusters,
4695 &sbi->s_flex_groups[flex_group].free_clusters);
4698 ext4_mb_unload_buddy(&e4b);
4702 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4703 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4705 /* We dirtied the bitmap block */
4706 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4707 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4709 /* And the group descriptor block */
4710 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4711 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4715 if (overflow && !err) {
4723 ext4_std_error(sb, err);
4728 * ext4_group_add_blocks() -- Add given blocks to an existing group
4729 * @handle: handle to this transaction
4731 * @block: start physical block to add to the block group
4732 * @count: number of blocks to free
4734 * This marks the blocks as free in the bitmap and buddy.
4736 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4737 ext4_fsblk_t block, unsigned long count)
4739 struct buffer_head *bitmap_bh = NULL;
4740 struct buffer_head *gd_bh;
4741 ext4_group_t block_group;
4744 struct ext4_group_desc *desc;
4745 struct ext4_sb_info *sbi = EXT4_SB(sb);
4746 struct ext4_buddy e4b;
4747 int err = 0, ret, blk_free_count;
4748 ext4_grpblk_t blocks_freed;
4750 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4755 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4757 * Check to see if we are freeing blocks across a group
4760 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4761 ext4_warning(sb, "too much blocks added to group %u\n",
4767 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4773 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4779 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4780 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4781 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4782 in_range(block + count - 1, ext4_inode_table(sb, desc),
4783 sbi->s_itb_per_group)) {
4784 ext4_error(sb, "Adding blocks in system zones - "
4785 "Block = %llu, count = %lu",
4791 BUFFER_TRACE(bitmap_bh, "getting write access");
4792 err = ext4_journal_get_write_access(handle, bitmap_bh);
4797 * We are about to modify some metadata. Call the journal APIs
4798 * to unshare ->b_data if a currently-committing transaction is
4801 BUFFER_TRACE(gd_bh, "get_write_access");
4802 err = ext4_journal_get_write_access(handle, gd_bh);
4806 for (i = 0, blocks_freed = 0; i < count; i++) {
4807 BUFFER_TRACE(bitmap_bh, "clear bit");
4808 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4809 ext4_error(sb, "bit already cleared for block %llu",
4810 (ext4_fsblk_t)(block + i));
4811 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4817 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4822 * need to update group_info->bb_free and bitmap
4823 * with group lock held. generate_buddy look at
4824 * them with group lock_held
4826 ext4_lock_group(sb, block_group);
4827 mb_clear_bits(bitmap_bh->b_data, bit, count);
4828 mb_free_blocks(NULL, &e4b, bit, count);
4829 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4830 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4831 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
4832 ext4_group_desc_csum_set(sb, block_group, desc);
4833 ext4_unlock_group(sb, block_group);
4834 percpu_counter_add(&sbi->s_freeclusters_counter,
4835 EXT4_B2C(sbi, blocks_freed));
4837 if (sbi->s_log_groups_per_flex) {
4838 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4839 atomic_add(EXT4_B2C(sbi, blocks_freed),
4840 &sbi->s_flex_groups[flex_group].free_clusters);
4843 ext4_mb_unload_buddy(&e4b);
4845 /* We dirtied the bitmap block */
4846 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4847 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4849 /* And the group descriptor block */
4850 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4851 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4857 ext4_std_error(sb, err);
4862 * ext4_trim_extent -- function to TRIM one single free extent in the group
4863 * @sb: super block for the file system
4864 * @start: starting block of the free extent in the alloc. group
4865 * @count: number of blocks to TRIM
4866 * @group: alloc. group we are working with
4867 * @e4b: ext4 buddy for the group
4869 * Trim "count" blocks starting at "start" in the "group". To assure that no
4870 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4871 * be called with under the group lock.
4873 static int ext4_trim_extent(struct super_block *sb, int start, int count,
4874 ext4_group_t group, struct ext4_buddy *e4b)
4876 struct ext4_free_extent ex;
4879 trace_ext4_trim_extent(sb, group, start, count);
4881 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4883 ex.fe_start = start;
4884 ex.fe_group = group;
4888 * Mark blocks used, so no one can reuse them while
4891 mb_mark_used(e4b, &ex);
4892 ext4_unlock_group(sb, group);
4893 ret = ext4_issue_discard(sb, group, start, count);
4894 ext4_lock_group(sb, group);
4895 mb_free_blocks(NULL, e4b, start, ex.fe_len);
4900 * ext4_trim_all_free -- function to trim all free space in alloc. group
4901 * @sb: super block for file system
4902 * @group: group to be trimmed
4903 * @start: first group block to examine
4904 * @max: last group block to examine
4905 * @minblocks: minimum extent block count
4907 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4908 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4912 * ext4_trim_all_free walks through group's block bitmap searching for free
4913 * extents. When the free extent is found, mark it as used in group buddy
4914 * bitmap. Then issue a TRIM command on this extent and free the extent in
4915 * the group buddy bitmap. This is done until whole group is scanned.
4917 static ext4_grpblk_t
4918 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4919 ext4_grpblk_t start, ext4_grpblk_t max,
4920 ext4_grpblk_t minblocks)
4923 ext4_grpblk_t next, count = 0, free_count = 0;
4924 struct ext4_buddy e4b;
4927 trace_ext4_trim_all_free(sb, group, start, max);
4929 ret = ext4_mb_load_buddy(sb, group, &e4b);
4931 ext4_error(sb, "Error in loading buddy "
4932 "information for %u", group);
4935 bitmap = e4b.bd_bitmap;
4937 ext4_lock_group(sb, group);
4938 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4939 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4942 start = (e4b.bd_info->bb_first_free > start) ?
4943 e4b.bd_info->bb_first_free : start;
4945 while (start <= max) {
4946 start = mb_find_next_zero_bit(bitmap, max + 1, start);
4949 next = mb_find_next_bit(bitmap, max + 1, start);
4951 if ((next - start) >= minblocks) {
4952 ret = ext4_trim_extent(sb, start,
4953 next - start, group, &e4b);
4954 if (ret && ret != -EOPNOTSUPP)
4957 count += next - start;
4959 free_count += next - start;
4962 if (fatal_signal_pending(current)) {
4963 count = -ERESTARTSYS;
4967 if (need_resched()) {
4968 ext4_unlock_group(sb, group);
4970 ext4_lock_group(sb, group);
4973 if ((e4b.bd_info->bb_free - free_count) < minblocks)
4979 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
4982 ext4_unlock_group(sb, group);
4983 ext4_mb_unload_buddy(&e4b);
4985 ext4_debug("trimmed %d blocks in the group %d\n",
4992 * ext4_trim_fs() -- trim ioctl handle function
4993 * @sb: superblock for filesystem
4994 * @range: fstrim_range structure
4996 * start: First Byte to trim
4997 * len: number of Bytes to trim from start
4998 * minlen: minimum extent length in Bytes
4999 * ext4_trim_fs goes through all allocation groups containing Bytes from
5000 * start to start+len. For each such a group ext4_trim_all_free function
5001 * is invoked to trim all free space.
5003 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5005 struct ext4_group_info *grp;
5006 ext4_group_t group, first_group, last_group;
5007 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5008 uint64_t start, end, minlen, trimmed = 0;
5009 ext4_fsblk_t first_data_blk =
5010 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5011 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5014 start = range->start >> sb->s_blocksize_bits;
5015 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5016 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5017 range->minlen >> sb->s_blocksize_bits);
5019 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5020 start >= max_blks ||
5021 range->len < sb->s_blocksize)
5023 if (end >= max_blks)
5025 if (end <= first_data_blk)
5027 if (start < first_data_blk)
5028 start = first_data_blk;
5030 /* Determine first and last group to examine based on start and end */
5031 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5032 &first_group, &first_cluster);
5033 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5034 &last_group, &last_cluster);
5036 /* end now represents the last cluster to discard in this group */
5037 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5039 for (group = first_group; group <= last_group; group++) {
5040 grp = ext4_get_group_info(sb, group);
5041 /* We only do this if the grp has never been initialized */
5042 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5043 ret = ext4_mb_init_group(sb, group);
5049 * For all the groups except the last one, last cluster will
5050 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5051 * change it for the last group, note that last_cluster is
5052 * already computed earlier by ext4_get_group_no_and_offset()
5054 if (group == last_group)
5057 if (grp->bb_free >= minlen) {
5058 cnt = ext4_trim_all_free(sb, group, first_cluster,
5068 * For every group except the first one, we are sure
5069 * that the first cluster to discard will be cluster #0.
5075 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5078 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;