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 /* Avoid using the first bg of a flexgroup for data files */
1888 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1889 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1890 ((group % flex_size) == 0))
1893 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
1894 (free / fragments) >= ac->ac_g_ex.fe_len)
1897 if (grp->bb_largest_free_order < ac->ac_2order)
1902 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1906 if (free >= ac->ac_g_ex.fe_len)
1918 static noinline_for_stack int
1919 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1921 ext4_group_t ngroups, group, i;
1924 struct ext4_sb_info *sbi;
1925 struct super_block *sb;
1926 struct ext4_buddy e4b;
1930 ngroups = ext4_get_groups_count(sb);
1931 /* non-extent files are limited to low blocks/groups */
1932 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
1933 ngroups = sbi->s_blockfile_groups;
1935 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1937 /* first, try the goal */
1938 err = ext4_mb_find_by_goal(ac, &e4b);
1939 if (err || ac->ac_status == AC_STATUS_FOUND)
1942 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1946 * ac->ac2_order is set only if the fe_len is a power of 2
1947 * if ac2_order is set we also set criteria to 0 so that we
1948 * try exact allocation using buddy.
1950 i = fls(ac->ac_g_ex.fe_len);
1953 * We search using buddy data only if the order of the request
1954 * is greater than equal to the sbi_s_mb_order2_reqs
1955 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1957 if (i >= sbi->s_mb_order2_reqs) {
1959 * This should tell if fe_len is exactly power of 2
1961 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1962 ac->ac_2order = i - 1;
1965 /* if stream allocation is enabled, use global goal */
1966 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1967 /* TBD: may be hot point */
1968 spin_lock(&sbi->s_md_lock);
1969 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1970 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1971 spin_unlock(&sbi->s_md_lock);
1974 /* Let's just scan groups to find more-less suitable blocks */
1975 cr = ac->ac_2order ? 0 : 1;
1977 * cr == 0 try to get exact allocation,
1978 * cr == 3 try to get anything
1981 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1982 ac->ac_criteria = cr;
1984 * searching for the right group start
1985 * from the goal value specified
1987 group = ac->ac_g_ex.fe_group;
1989 for (i = 0; i < ngroups; group++, i++) {
1990 if (group == ngroups)
1993 /* This now checks without needing the buddy page */
1994 if (!ext4_mb_good_group(ac, group, cr))
1997 err = ext4_mb_load_buddy(sb, group, &e4b);
2001 ext4_lock_group(sb, group);
2004 * We need to check again after locking the
2007 if (!ext4_mb_good_group(ac, group, cr)) {
2008 ext4_unlock_group(sb, group);
2009 ext4_mb_unload_buddy(&e4b);
2013 ac->ac_groups_scanned++;
2014 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2015 ext4_mb_simple_scan_group(ac, &e4b);
2016 else if (cr == 1 && sbi->s_stripe &&
2017 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2018 ext4_mb_scan_aligned(ac, &e4b);
2020 ext4_mb_complex_scan_group(ac, &e4b);
2022 ext4_unlock_group(sb, group);
2023 ext4_mb_unload_buddy(&e4b);
2025 if (ac->ac_status != AC_STATUS_CONTINUE)
2030 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2031 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2033 * We've been searching too long. Let's try to allocate
2034 * the best chunk we've found so far
2037 ext4_mb_try_best_found(ac, &e4b);
2038 if (ac->ac_status != AC_STATUS_FOUND) {
2040 * Someone more lucky has already allocated it.
2041 * The only thing we can do is just take first
2043 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2045 ac->ac_b_ex.fe_group = 0;
2046 ac->ac_b_ex.fe_start = 0;
2047 ac->ac_b_ex.fe_len = 0;
2048 ac->ac_status = AC_STATUS_CONTINUE;
2049 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2051 atomic_inc(&sbi->s_mb_lost_chunks);
2059 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2061 struct super_block *sb = seq->private;
2064 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2067 return (void *) ((unsigned long) group);
2070 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2072 struct super_block *sb = seq->private;
2076 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2079 return (void *) ((unsigned long) group);
2082 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2084 struct super_block *sb = seq->private;
2085 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2087 int err, buddy_loaded = 0;
2088 struct ext4_buddy e4b;
2089 struct ext4_group_info *grinfo;
2091 struct ext4_group_info info;
2092 ext4_grpblk_t counters[16];
2097 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2098 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2099 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2100 "group", "free", "frags", "first",
2101 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2102 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2104 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2105 sizeof(struct ext4_group_info);
2106 grinfo = ext4_get_group_info(sb, group);
2107 /* Load the group info in memory only if not already loaded. */
2108 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2109 err = ext4_mb_load_buddy(sb, group, &e4b);
2111 seq_printf(seq, "#%-5u: I/O error\n", group);
2117 memcpy(&sg, ext4_get_group_info(sb, group), i);
2120 ext4_mb_unload_buddy(&e4b);
2122 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2123 sg.info.bb_fragments, sg.info.bb_first_free);
2124 for (i = 0; i <= 13; i++)
2125 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2126 sg.info.bb_counters[i] : 0);
2127 seq_printf(seq, " ]\n");
2132 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2136 static const struct seq_operations ext4_mb_seq_groups_ops = {
2137 .start = ext4_mb_seq_groups_start,
2138 .next = ext4_mb_seq_groups_next,
2139 .stop = ext4_mb_seq_groups_stop,
2140 .show = ext4_mb_seq_groups_show,
2143 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2145 struct super_block *sb = PDE(inode)->data;
2148 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2150 struct seq_file *m = file->private_data;
2157 static const struct file_operations ext4_mb_seq_groups_fops = {
2158 .owner = THIS_MODULE,
2159 .open = ext4_mb_seq_groups_open,
2161 .llseek = seq_lseek,
2162 .release = seq_release,
2165 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2167 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2168 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2175 * Allocate the top-level s_group_info array for the specified number
2178 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2180 struct ext4_sb_info *sbi = EXT4_SB(sb);
2182 struct ext4_group_info ***new_groupinfo;
2184 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2185 EXT4_DESC_PER_BLOCK_BITS(sb);
2186 if (size <= sbi->s_group_info_size)
2189 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2190 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2191 if (!new_groupinfo) {
2192 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2195 if (sbi->s_group_info) {
2196 memcpy(new_groupinfo, sbi->s_group_info,
2197 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2198 ext4_kvfree(sbi->s_group_info);
2200 sbi->s_group_info = new_groupinfo;
2201 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2202 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2203 sbi->s_group_info_size);
2207 /* Create and initialize ext4_group_info data for the given group. */
2208 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2209 struct ext4_group_desc *desc)
2213 struct ext4_sb_info *sbi = EXT4_SB(sb);
2214 struct ext4_group_info **meta_group_info;
2215 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2218 * First check if this group is the first of a reserved block.
2219 * If it's true, we have to allocate a new table of pointers
2220 * to ext4_group_info structures
2222 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2223 metalen = sizeof(*meta_group_info) <<
2224 EXT4_DESC_PER_BLOCK_BITS(sb);
2225 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2226 if (meta_group_info == NULL) {
2227 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2228 "for a buddy group");
2229 goto exit_meta_group_info;
2231 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2236 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2237 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2239 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
2240 if (meta_group_info[i] == NULL) {
2241 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2242 goto exit_group_info;
2244 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2245 &(meta_group_info[i]->bb_state));
2248 * initialize bb_free to be able to skip
2249 * empty groups without initialization
2251 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2252 meta_group_info[i]->bb_free =
2253 ext4_free_clusters_after_init(sb, group, desc);
2255 meta_group_info[i]->bb_free =
2256 ext4_free_group_clusters(sb, desc);
2259 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2260 init_rwsem(&meta_group_info[i]->alloc_sem);
2261 meta_group_info[i]->bb_free_root = RB_ROOT;
2262 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2266 struct buffer_head *bh;
2267 meta_group_info[i]->bb_bitmap =
2268 kmalloc(sb->s_blocksize, GFP_KERNEL);
2269 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2270 bh = ext4_read_block_bitmap(sb, group);
2272 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2281 /* If a meta_group_info table has been allocated, release it now */
2282 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2283 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2284 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2286 exit_meta_group_info:
2288 } /* ext4_mb_add_groupinfo */
2290 static int ext4_mb_init_backend(struct super_block *sb)
2292 ext4_group_t ngroups = ext4_get_groups_count(sb);
2294 struct ext4_sb_info *sbi = EXT4_SB(sb);
2296 struct ext4_group_desc *desc;
2297 struct kmem_cache *cachep;
2299 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2303 sbi->s_buddy_cache = new_inode(sb);
2304 if (sbi->s_buddy_cache == NULL) {
2305 ext4_msg(sb, KERN_ERR, "can't get new inode");
2308 /* To avoid potentially colliding with an valid on-disk inode number,
2309 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2310 * not in the inode hash, so it should never be found by iget(), but
2311 * this will avoid confusion if it ever shows up during debugging. */
2312 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2313 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2314 for (i = 0; i < ngroups; i++) {
2315 desc = ext4_get_group_desc(sb, i, NULL);
2317 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2320 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2327 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2329 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2330 i = sbi->s_group_info_size;
2332 kfree(sbi->s_group_info[i]);
2333 iput(sbi->s_buddy_cache);
2335 ext4_kvfree(sbi->s_group_info);
2339 static void ext4_groupinfo_destroy_slabs(void)
2343 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2344 if (ext4_groupinfo_caches[i])
2345 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2346 ext4_groupinfo_caches[i] = NULL;
2350 static int ext4_groupinfo_create_slab(size_t size)
2352 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2354 int blocksize_bits = order_base_2(size);
2355 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2356 struct kmem_cache *cachep;
2358 if (cache_index >= NR_GRPINFO_CACHES)
2361 if (unlikely(cache_index < 0))
2364 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2365 if (ext4_groupinfo_caches[cache_index]) {
2366 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2367 return 0; /* Already created */
2370 slab_size = offsetof(struct ext4_group_info,
2371 bb_counters[blocksize_bits + 2]);
2373 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2374 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2377 ext4_groupinfo_caches[cache_index] = cachep;
2379 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2382 "EXT4-fs: no memory for groupinfo slab cache\n");
2389 int ext4_mb_init(struct super_block *sb)
2391 struct ext4_sb_info *sbi = EXT4_SB(sb);
2397 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2399 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2400 if (sbi->s_mb_offsets == NULL) {
2405 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2406 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2407 if (sbi->s_mb_maxs == NULL) {
2412 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2416 /* order 0 is regular bitmap */
2417 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2418 sbi->s_mb_offsets[0] = 0;
2422 max = sb->s_blocksize << 2;
2424 sbi->s_mb_offsets[i] = offset;
2425 sbi->s_mb_maxs[i] = max;
2426 offset += 1 << (sb->s_blocksize_bits - i);
2429 } while (i <= sb->s_blocksize_bits + 1);
2431 spin_lock_init(&sbi->s_md_lock);
2432 spin_lock_init(&sbi->s_bal_lock);
2434 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2435 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2436 sbi->s_mb_stats = MB_DEFAULT_STATS;
2437 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2438 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2440 * The default group preallocation is 512, which for 4k block
2441 * sizes translates to 2 megabytes. However for bigalloc file
2442 * systems, this is probably too big (i.e, if the cluster size
2443 * is 1 megabyte, then group preallocation size becomes half a
2444 * gigabyte!). As a default, we will keep a two megabyte
2445 * group pralloc size for cluster sizes up to 64k, and after
2446 * that, we will force a minimum group preallocation size of
2447 * 32 clusters. This translates to 8 megs when the cluster
2448 * size is 256k, and 32 megs when the cluster size is 1 meg,
2449 * which seems reasonable as a default.
2451 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2452 sbi->s_cluster_bits, 32);
2454 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2455 * to the lowest multiple of s_stripe which is bigger than
2456 * the s_mb_group_prealloc as determined above. We want
2457 * the preallocation size to be an exact multiple of the
2458 * RAID stripe size so that preallocations don't fragment
2461 if (sbi->s_stripe > 1) {
2462 sbi->s_mb_group_prealloc = roundup(
2463 sbi->s_mb_group_prealloc, sbi->s_stripe);
2466 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2467 if (sbi->s_locality_groups == NULL) {
2469 goto out_free_groupinfo_slab;
2471 for_each_possible_cpu(i) {
2472 struct ext4_locality_group *lg;
2473 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2474 mutex_init(&lg->lg_mutex);
2475 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2476 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2477 spin_lock_init(&lg->lg_prealloc_lock);
2480 /* init file for buddy data */
2481 ret = ext4_mb_init_backend(sb);
2483 goto out_free_locality_groups;
2486 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2487 &ext4_mb_seq_groups_fops, sb);
2491 out_free_locality_groups:
2492 free_percpu(sbi->s_locality_groups);
2493 sbi->s_locality_groups = NULL;
2494 out_free_groupinfo_slab:
2495 ext4_groupinfo_destroy_slabs();
2497 kfree(sbi->s_mb_offsets);
2498 sbi->s_mb_offsets = NULL;
2499 kfree(sbi->s_mb_maxs);
2500 sbi->s_mb_maxs = NULL;
2504 /* need to called with the ext4 group lock held */
2505 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2507 struct ext4_prealloc_space *pa;
2508 struct list_head *cur, *tmp;
2511 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2512 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2513 list_del(&pa->pa_group_list);
2515 kmem_cache_free(ext4_pspace_cachep, pa);
2518 mb_debug(1, "mballoc: %u PAs left\n", count);
2522 int ext4_mb_release(struct super_block *sb)
2524 ext4_group_t ngroups = ext4_get_groups_count(sb);
2526 int num_meta_group_infos;
2527 struct ext4_group_info *grinfo;
2528 struct ext4_sb_info *sbi = EXT4_SB(sb);
2529 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2532 remove_proc_entry("mb_groups", sbi->s_proc);
2534 if (sbi->s_group_info) {
2535 for (i = 0; i < ngroups; i++) {
2536 grinfo = ext4_get_group_info(sb, i);
2538 kfree(grinfo->bb_bitmap);
2540 ext4_lock_group(sb, i);
2541 ext4_mb_cleanup_pa(grinfo);
2542 ext4_unlock_group(sb, i);
2543 kmem_cache_free(cachep, grinfo);
2545 num_meta_group_infos = (ngroups +
2546 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2547 EXT4_DESC_PER_BLOCK_BITS(sb);
2548 for (i = 0; i < num_meta_group_infos; i++)
2549 kfree(sbi->s_group_info[i]);
2550 ext4_kvfree(sbi->s_group_info);
2552 kfree(sbi->s_mb_offsets);
2553 kfree(sbi->s_mb_maxs);
2554 if (sbi->s_buddy_cache)
2555 iput(sbi->s_buddy_cache);
2556 if (sbi->s_mb_stats) {
2557 ext4_msg(sb, KERN_INFO,
2558 "mballoc: %u blocks %u reqs (%u success)",
2559 atomic_read(&sbi->s_bal_allocated),
2560 atomic_read(&sbi->s_bal_reqs),
2561 atomic_read(&sbi->s_bal_success));
2562 ext4_msg(sb, KERN_INFO,
2563 "mballoc: %u extents scanned, %u goal hits, "
2564 "%u 2^N hits, %u breaks, %u lost",
2565 atomic_read(&sbi->s_bal_ex_scanned),
2566 atomic_read(&sbi->s_bal_goals),
2567 atomic_read(&sbi->s_bal_2orders),
2568 atomic_read(&sbi->s_bal_breaks),
2569 atomic_read(&sbi->s_mb_lost_chunks));
2570 ext4_msg(sb, KERN_INFO,
2571 "mballoc: %lu generated and it took %Lu",
2572 sbi->s_mb_buddies_generated,
2573 sbi->s_mb_generation_time);
2574 ext4_msg(sb, KERN_INFO,
2575 "mballoc: %u preallocated, %u discarded",
2576 atomic_read(&sbi->s_mb_preallocated),
2577 atomic_read(&sbi->s_mb_discarded));
2580 free_percpu(sbi->s_locality_groups);
2585 static inline int ext4_issue_discard(struct super_block *sb,
2586 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2588 ext4_fsblk_t discard_block;
2590 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2591 ext4_group_first_block_no(sb, block_group));
2592 count = EXT4_C2B(EXT4_SB(sb), count);
2593 trace_ext4_discard_blocks(sb,
2594 (unsigned long long) discard_block, count);
2595 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2599 * This function is called by the jbd2 layer once the commit has finished,
2600 * so we know we can free the blocks that were released with that commit.
2602 static void ext4_free_data_callback(struct super_block *sb,
2603 struct ext4_journal_cb_entry *jce,
2606 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2607 struct ext4_buddy e4b;
2608 struct ext4_group_info *db;
2609 int err, count = 0, count2 = 0;
2611 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2612 entry->efd_count, entry->efd_group, entry);
2614 if (test_opt(sb, DISCARD)) {
2615 err = ext4_issue_discard(sb, entry->efd_group,
2616 entry->efd_start_cluster,
2618 if (err && err != -EOPNOTSUPP)
2619 ext4_msg(sb, KERN_WARNING, "discard request in"
2620 " group:%d block:%d count:%d failed"
2621 " with %d", entry->efd_group,
2622 entry->efd_start_cluster,
2623 entry->efd_count, err);
2626 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2627 /* we expect to find existing buddy because it's pinned */
2632 /* there are blocks to put in buddy to make them really free */
2633 count += entry->efd_count;
2635 ext4_lock_group(sb, entry->efd_group);
2636 /* Take it out of per group rb tree */
2637 rb_erase(&entry->efd_node, &(db->bb_free_root));
2638 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2641 * Clear the trimmed flag for the group so that the next
2642 * ext4_trim_fs can trim it.
2643 * If the volume is mounted with -o discard, online discard
2644 * is supported and the free blocks will be trimmed online.
2646 if (!test_opt(sb, DISCARD))
2647 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2649 if (!db->bb_free_root.rb_node) {
2650 /* No more items in the per group rb tree
2651 * balance refcounts from ext4_mb_free_metadata()
2653 page_cache_release(e4b.bd_buddy_page);
2654 page_cache_release(e4b.bd_bitmap_page);
2656 ext4_unlock_group(sb, entry->efd_group);
2657 kmem_cache_free(ext4_free_data_cachep, entry);
2658 ext4_mb_unload_buddy(&e4b);
2660 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2663 #ifdef CONFIG_EXT4_DEBUG
2664 u8 mb_enable_debug __read_mostly;
2666 static struct dentry *debugfs_dir;
2667 static struct dentry *debugfs_debug;
2669 static void __init ext4_create_debugfs_entry(void)
2671 debugfs_dir = debugfs_create_dir("ext4", NULL);
2673 debugfs_debug = debugfs_create_u8("mballoc-debug",
2679 static void ext4_remove_debugfs_entry(void)
2681 debugfs_remove(debugfs_debug);
2682 debugfs_remove(debugfs_dir);
2687 static void __init ext4_create_debugfs_entry(void)
2691 static void ext4_remove_debugfs_entry(void)
2697 int __init ext4_init_mballoc(void)
2699 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2700 SLAB_RECLAIM_ACCOUNT);
2701 if (ext4_pspace_cachep == NULL)
2704 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2705 SLAB_RECLAIM_ACCOUNT);
2706 if (ext4_ac_cachep == NULL) {
2707 kmem_cache_destroy(ext4_pspace_cachep);
2711 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2712 SLAB_RECLAIM_ACCOUNT);
2713 if (ext4_free_data_cachep == NULL) {
2714 kmem_cache_destroy(ext4_pspace_cachep);
2715 kmem_cache_destroy(ext4_ac_cachep);
2718 ext4_create_debugfs_entry();
2722 void ext4_exit_mballoc(void)
2725 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2726 * before destroying the slab cache.
2729 kmem_cache_destroy(ext4_pspace_cachep);
2730 kmem_cache_destroy(ext4_ac_cachep);
2731 kmem_cache_destroy(ext4_free_data_cachep);
2732 ext4_groupinfo_destroy_slabs();
2733 ext4_remove_debugfs_entry();
2738 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2739 * Returns 0 if success or error code
2741 static noinline_for_stack int
2742 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2743 handle_t *handle, unsigned int reserv_clstrs)
2745 struct buffer_head *bitmap_bh = NULL;
2746 struct ext4_group_desc *gdp;
2747 struct buffer_head *gdp_bh;
2748 struct ext4_sb_info *sbi;
2749 struct super_block *sb;
2753 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2754 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2760 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2764 err = ext4_journal_get_write_access(handle, bitmap_bh);
2769 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2773 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2774 ext4_free_group_clusters(sb, gdp));
2776 err = ext4_journal_get_write_access(handle, gdp_bh);
2780 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2782 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2783 if (!ext4_data_block_valid(sbi, block, len)) {
2784 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2785 "fs metadata", block, block+len);
2786 /* File system mounted not to panic on error
2787 * Fix the bitmap and repeat the block allocation
2788 * We leak some of the blocks here.
2790 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2791 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2792 ac->ac_b_ex.fe_len);
2793 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2794 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2800 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2801 #ifdef AGGRESSIVE_CHECK
2804 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2805 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2806 bitmap_bh->b_data));
2810 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2811 ac->ac_b_ex.fe_len);
2812 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2813 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2814 ext4_free_group_clusters_set(sb, gdp,
2815 ext4_free_clusters_after_init(sb,
2816 ac->ac_b_ex.fe_group, gdp));
2818 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2819 ext4_free_group_clusters_set(sb, gdp, len);
2820 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2821 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2823 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2824 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2826 * Now reduce the dirty block count also. Should not go negative
2828 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2829 /* release all the reserved blocks if non delalloc */
2830 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2833 if (sbi->s_log_groups_per_flex) {
2834 ext4_group_t flex_group = ext4_flex_group(sbi,
2835 ac->ac_b_ex.fe_group);
2836 atomic_sub(ac->ac_b_ex.fe_len,
2837 &sbi->s_flex_groups[flex_group].free_clusters);
2840 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2843 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2851 * here we normalize request for locality group
2852 * Group request are normalized to s_mb_group_prealloc, which goes to
2853 * s_strip if we set the same via mount option.
2854 * s_mb_group_prealloc can be configured via
2855 * /sys/fs/ext4/<partition>/mb_group_prealloc
2857 * XXX: should we try to preallocate more than the group has now?
2859 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2861 struct super_block *sb = ac->ac_sb;
2862 struct ext4_locality_group *lg = ac->ac_lg;
2865 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2866 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2867 current->pid, ac->ac_g_ex.fe_len);
2871 * Normalization means making request better in terms of
2872 * size and alignment
2874 static noinline_for_stack void
2875 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2876 struct ext4_allocation_request *ar)
2878 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2881 loff_t size, start_off;
2882 loff_t orig_size __maybe_unused;
2884 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2885 struct ext4_prealloc_space *pa;
2887 /* do normalize only data requests, metadata requests
2888 do not need preallocation */
2889 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2892 /* sometime caller may want exact blocks */
2893 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2896 /* caller may indicate that preallocation isn't
2897 * required (it's a tail, for example) */
2898 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2901 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2902 ext4_mb_normalize_group_request(ac);
2906 bsbits = ac->ac_sb->s_blocksize_bits;
2908 /* first, let's learn actual file size
2909 * given current request is allocated */
2910 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
2911 size = size << bsbits;
2912 if (size < i_size_read(ac->ac_inode))
2913 size = i_size_read(ac->ac_inode);
2916 /* max size of free chunks */
2919 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2920 (req <= (size) || max <= (chunk_size))
2922 /* first, try to predict filesize */
2923 /* XXX: should this table be tunable? */
2925 if (size <= 16 * 1024) {
2927 } else if (size <= 32 * 1024) {
2929 } else if (size <= 64 * 1024) {
2931 } else if (size <= 128 * 1024) {
2933 } else if (size <= 256 * 1024) {
2935 } else if (size <= 512 * 1024) {
2937 } else if (size <= 1024 * 1024) {
2939 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2940 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2941 (21 - bsbits)) << 21;
2942 size = 2 * 1024 * 1024;
2943 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2944 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2945 (22 - bsbits)) << 22;
2946 size = 4 * 1024 * 1024;
2947 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2948 (8<<20)>>bsbits, max, 8 * 1024)) {
2949 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2950 (23 - bsbits)) << 23;
2951 size = 8 * 1024 * 1024;
2953 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2954 size = ac->ac_o_ex.fe_len << bsbits;
2956 size = size >> bsbits;
2957 start = start_off >> bsbits;
2959 /* don't cover already allocated blocks in selected range */
2960 if (ar->pleft && start <= ar->lleft) {
2961 size -= ar->lleft + 1 - start;
2962 start = ar->lleft + 1;
2964 if (ar->pright && start + size - 1 >= ar->lright)
2965 size -= start + size - ar->lright;
2969 /* check we don't cross already preallocated blocks */
2971 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2976 spin_lock(&pa->pa_lock);
2977 if (pa->pa_deleted) {
2978 spin_unlock(&pa->pa_lock);
2982 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
2985 /* PA must not overlap original request */
2986 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2987 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2989 /* skip PAs this normalized request doesn't overlap with */
2990 if (pa->pa_lstart >= end || pa_end <= start) {
2991 spin_unlock(&pa->pa_lock);
2994 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2996 /* adjust start or end to be adjacent to this pa */
2997 if (pa_end <= ac->ac_o_ex.fe_logical) {
2998 BUG_ON(pa_end < start);
3000 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3001 BUG_ON(pa->pa_lstart > end);
3002 end = pa->pa_lstart;
3004 spin_unlock(&pa->pa_lock);
3009 /* XXX: extra loop to check we really don't overlap preallocations */
3011 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3014 spin_lock(&pa->pa_lock);
3015 if (pa->pa_deleted == 0) {
3016 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3018 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3020 spin_unlock(&pa->pa_lock);
3024 if (start + size <= ac->ac_o_ex.fe_logical &&
3025 start > ac->ac_o_ex.fe_logical) {
3026 ext4_msg(ac->ac_sb, KERN_ERR,
3027 "start %lu, size %lu, fe_logical %lu",
3028 (unsigned long) start, (unsigned long) size,
3029 (unsigned long) ac->ac_o_ex.fe_logical);
3031 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3032 start > ac->ac_o_ex.fe_logical);
3033 BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
3035 /* now prepare goal request */
3037 /* XXX: is it better to align blocks WRT to logical
3038 * placement or satisfy big request as is */
3039 ac->ac_g_ex.fe_logical = start;
3040 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3042 /* define goal start in order to merge */
3043 if (ar->pright && (ar->lright == (start + size))) {
3044 /* merge to the right */
3045 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3046 &ac->ac_f_ex.fe_group,
3047 &ac->ac_f_ex.fe_start);
3048 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3050 if (ar->pleft && (ar->lleft + 1 == start)) {
3051 /* merge to the left */
3052 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3053 &ac->ac_f_ex.fe_group,
3054 &ac->ac_f_ex.fe_start);
3055 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3058 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3059 (unsigned) orig_size, (unsigned) start);
3062 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3064 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3066 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3067 atomic_inc(&sbi->s_bal_reqs);
3068 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3069 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3070 atomic_inc(&sbi->s_bal_success);
3071 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3072 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3073 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3074 atomic_inc(&sbi->s_bal_goals);
3075 if (ac->ac_found > sbi->s_mb_max_to_scan)
3076 atomic_inc(&sbi->s_bal_breaks);
3079 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3080 trace_ext4_mballoc_alloc(ac);
3082 trace_ext4_mballoc_prealloc(ac);
3086 * Called on failure; free up any blocks from the inode PA for this
3087 * context. We don't need this for MB_GROUP_PA because we only change
3088 * pa_free in ext4_mb_release_context(), but on failure, we've already
3089 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3091 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3093 struct ext4_prealloc_space *pa = ac->ac_pa;
3095 if (pa && pa->pa_type == MB_INODE_PA)
3096 pa->pa_free += ac->ac_b_ex.fe_len;
3100 * use blocks preallocated to inode
3102 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3103 struct ext4_prealloc_space *pa)
3105 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3110 /* found preallocated blocks, use them */
3111 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3112 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3113 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3114 len = EXT4_NUM_B2C(sbi, end - start);
3115 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3116 &ac->ac_b_ex.fe_start);
3117 ac->ac_b_ex.fe_len = len;
3118 ac->ac_status = AC_STATUS_FOUND;
3121 BUG_ON(start < pa->pa_pstart);
3122 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3123 BUG_ON(pa->pa_free < len);
3126 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3130 * use blocks preallocated to locality group
3132 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3133 struct ext4_prealloc_space *pa)
3135 unsigned int len = ac->ac_o_ex.fe_len;
3137 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3138 &ac->ac_b_ex.fe_group,
3139 &ac->ac_b_ex.fe_start);
3140 ac->ac_b_ex.fe_len = len;
3141 ac->ac_status = AC_STATUS_FOUND;
3144 /* we don't correct pa_pstart or pa_plen here to avoid
3145 * possible race when the group is being loaded concurrently
3146 * instead we correct pa later, after blocks are marked
3147 * in on-disk bitmap -- see ext4_mb_release_context()
3148 * Other CPUs are prevented from allocating from this pa by lg_mutex
3150 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3154 * Return the prealloc space that have minimal distance
3155 * from the goal block. @cpa is the prealloc
3156 * space that is having currently known minimal distance
3157 * from the goal block.
3159 static struct ext4_prealloc_space *
3160 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3161 struct ext4_prealloc_space *pa,
3162 struct ext4_prealloc_space *cpa)
3164 ext4_fsblk_t cur_distance, new_distance;
3167 atomic_inc(&pa->pa_count);
3170 cur_distance = abs(goal_block - cpa->pa_pstart);
3171 new_distance = abs(goal_block - pa->pa_pstart);
3173 if (cur_distance <= new_distance)
3176 /* drop the previous reference */
3177 atomic_dec(&cpa->pa_count);
3178 atomic_inc(&pa->pa_count);
3183 * search goal blocks in preallocated space
3185 static noinline_for_stack int
3186 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3188 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3190 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3191 struct ext4_locality_group *lg;
3192 struct ext4_prealloc_space *pa, *cpa = NULL;
3193 ext4_fsblk_t goal_block;
3195 /* only data can be preallocated */
3196 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3199 /* first, try per-file preallocation */
3201 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3203 /* all fields in this condition don't change,
3204 * so we can skip locking for them */
3205 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3206 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3207 EXT4_C2B(sbi, pa->pa_len)))
3210 /* non-extent files can't have physical blocks past 2^32 */
3211 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3212 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3213 EXT4_MAX_BLOCK_FILE_PHYS))
3216 /* found preallocated blocks, use them */
3217 spin_lock(&pa->pa_lock);
3218 if (pa->pa_deleted == 0 && pa->pa_free) {
3219 atomic_inc(&pa->pa_count);
3220 ext4_mb_use_inode_pa(ac, pa);
3221 spin_unlock(&pa->pa_lock);
3222 ac->ac_criteria = 10;
3226 spin_unlock(&pa->pa_lock);
3230 /* can we use group allocation? */
3231 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3234 /* inode may have no locality group for some reason */
3238 order = fls(ac->ac_o_ex.fe_len) - 1;
3239 if (order > PREALLOC_TB_SIZE - 1)
3240 /* The max size of hash table is PREALLOC_TB_SIZE */
3241 order = PREALLOC_TB_SIZE - 1;
3243 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3245 * search for the prealloc space that is having
3246 * minimal distance from the goal block.
3248 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3250 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3252 spin_lock(&pa->pa_lock);
3253 if (pa->pa_deleted == 0 &&
3254 pa->pa_free >= ac->ac_o_ex.fe_len) {
3256 cpa = ext4_mb_check_group_pa(goal_block,
3259 spin_unlock(&pa->pa_lock);
3264 ext4_mb_use_group_pa(ac, cpa);
3265 ac->ac_criteria = 20;
3272 * the function goes through all block freed in the group
3273 * but not yet committed and marks them used in in-core bitmap.
3274 * buddy must be generated from this bitmap
3275 * Need to be called with the ext4 group lock held
3277 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3281 struct ext4_group_info *grp;
3282 struct ext4_free_data *entry;
3284 grp = ext4_get_group_info(sb, group);
3285 n = rb_first(&(grp->bb_free_root));
3288 entry = rb_entry(n, struct ext4_free_data, efd_node);
3289 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3296 * the function goes through all preallocation in this group and marks them
3297 * used in in-core bitmap. buddy must be generated from this bitmap
3298 * Need to be called with ext4 group lock held
3300 static noinline_for_stack
3301 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3304 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3305 struct ext4_prealloc_space *pa;
3306 struct list_head *cur;
3307 ext4_group_t groupnr;
3308 ext4_grpblk_t start;
3309 int preallocated = 0;
3312 /* all form of preallocation discards first load group,
3313 * so the only competing code is preallocation use.
3314 * we don't need any locking here
3315 * notice we do NOT ignore preallocations with pa_deleted
3316 * otherwise we could leave used blocks available for
3317 * allocation in buddy when concurrent ext4_mb_put_pa()
3318 * is dropping preallocation
3320 list_for_each(cur, &grp->bb_prealloc_list) {
3321 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3322 spin_lock(&pa->pa_lock);
3323 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3326 spin_unlock(&pa->pa_lock);
3327 if (unlikely(len == 0))
3329 BUG_ON(groupnr != group);
3330 ext4_set_bits(bitmap, start, len);
3331 preallocated += len;
3333 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3336 static void ext4_mb_pa_callback(struct rcu_head *head)
3338 struct ext4_prealloc_space *pa;
3339 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3340 kmem_cache_free(ext4_pspace_cachep, pa);
3344 * drops a reference to preallocated space descriptor
3345 * if this was the last reference and the space is consumed
3347 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3348 struct super_block *sb, struct ext4_prealloc_space *pa)
3351 ext4_fsblk_t grp_blk;
3353 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3356 /* in this short window concurrent discard can set pa_deleted */
3357 spin_lock(&pa->pa_lock);
3358 if (pa->pa_deleted == 1) {
3359 spin_unlock(&pa->pa_lock);
3364 spin_unlock(&pa->pa_lock);
3366 grp_blk = pa->pa_pstart;
3368 * If doing group-based preallocation, pa_pstart may be in the
3369 * next group when pa is used up
3371 if (pa->pa_type == MB_GROUP_PA)
3374 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3379 * P1 (buddy init) P2 (regular allocation)
3380 * find block B in PA
3381 * copy on-disk bitmap to buddy
3382 * mark B in on-disk bitmap
3383 * drop PA from group
3384 * mark all PAs in buddy
3386 * thus, P1 initializes buddy with B available. to prevent this
3387 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3390 ext4_lock_group(sb, grp);
3391 list_del(&pa->pa_group_list);
3392 ext4_unlock_group(sb, grp);
3394 spin_lock(pa->pa_obj_lock);
3395 list_del_rcu(&pa->pa_inode_list);
3396 spin_unlock(pa->pa_obj_lock);
3398 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3402 * creates new preallocated space for given inode
3404 static noinline_for_stack int
3405 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3407 struct super_block *sb = ac->ac_sb;
3408 struct ext4_sb_info *sbi = EXT4_SB(sb);
3409 struct ext4_prealloc_space *pa;
3410 struct ext4_group_info *grp;
3411 struct ext4_inode_info *ei;
3413 /* preallocate only when found space is larger then requested */
3414 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3415 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3416 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3418 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3422 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3428 /* we can't allocate as much as normalizer wants.
3429 * so, found space must get proper lstart
3430 * to cover original request */
3431 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3432 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3434 /* we're limited by original request in that
3435 * logical block must be covered any way
3436 * winl is window we can move our chunk within */
3437 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3439 /* also, we should cover whole original request */
3440 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3442 /* the smallest one defines real window */
3443 win = min(winl, wins);
3445 offs = ac->ac_o_ex.fe_logical %
3446 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3447 if (offs && offs < win)
3450 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3452 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3453 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3456 /* preallocation can change ac_b_ex, thus we store actually
3457 * allocated blocks for history */
3458 ac->ac_f_ex = ac->ac_b_ex;
3460 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3461 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3462 pa->pa_len = ac->ac_b_ex.fe_len;
3463 pa->pa_free = pa->pa_len;
3464 atomic_set(&pa->pa_count, 1);
3465 spin_lock_init(&pa->pa_lock);
3466 INIT_LIST_HEAD(&pa->pa_inode_list);
3467 INIT_LIST_HEAD(&pa->pa_group_list);
3469 pa->pa_type = MB_INODE_PA;
3471 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3472 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3473 trace_ext4_mb_new_inode_pa(ac, pa);
3475 ext4_mb_use_inode_pa(ac, pa);
3476 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3478 ei = EXT4_I(ac->ac_inode);
3479 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3481 pa->pa_obj_lock = &ei->i_prealloc_lock;
3482 pa->pa_inode = ac->ac_inode;
3484 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3485 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3486 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3488 spin_lock(pa->pa_obj_lock);
3489 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3490 spin_unlock(pa->pa_obj_lock);
3496 * creates new preallocated space for locality group inodes belongs to
3498 static noinline_for_stack int
3499 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3501 struct super_block *sb = ac->ac_sb;
3502 struct ext4_locality_group *lg;
3503 struct ext4_prealloc_space *pa;
3504 struct ext4_group_info *grp;
3506 /* preallocate only when found space is larger then requested */
3507 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3508 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3509 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3511 BUG_ON(ext4_pspace_cachep == NULL);
3512 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3516 /* preallocation can change ac_b_ex, thus we store actually
3517 * allocated blocks for history */
3518 ac->ac_f_ex = ac->ac_b_ex;
3520 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3521 pa->pa_lstart = pa->pa_pstart;
3522 pa->pa_len = ac->ac_b_ex.fe_len;
3523 pa->pa_free = pa->pa_len;
3524 atomic_set(&pa->pa_count, 1);
3525 spin_lock_init(&pa->pa_lock);
3526 INIT_LIST_HEAD(&pa->pa_inode_list);
3527 INIT_LIST_HEAD(&pa->pa_group_list);
3529 pa->pa_type = MB_GROUP_PA;
3531 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3532 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3533 trace_ext4_mb_new_group_pa(ac, pa);
3535 ext4_mb_use_group_pa(ac, pa);
3536 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3538 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3542 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3543 pa->pa_inode = NULL;
3545 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3546 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3547 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3550 * We will later add the new pa to the right bucket
3551 * after updating the pa_free in ext4_mb_release_context
3556 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3560 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3561 err = ext4_mb_new_group_pa(ac);
3563 err = ext4_mb_new_inode_pa(ac);
3568 * finds all unused blocks in on-disk bitmap, frees them in
3569 * in-core bitmap and buddy.
3570 * @pa must be unlinked from inode and group lists, so that
3571 * nobody else can find/use it.
3572 * the caller MUST hold group/inode locks.
3573 * TODO: optimize the case when there are no in-core structures yet
3575 static noinline_for_stack int
3576 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3577 struct ext4_prealloc_space *pa)
3579 struct super_block *sb = e4b->bd_sb;
3580 struct ext4_sb_info *sbi = EXT4_SB(sb);
3585 unsigned long long grp_blk_start;
3589 BUG_ON(pa->pa_deleted == 0);
3590 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3591 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3592 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3593 end = bit + pa->pa_len;
3596 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3599 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3600 mb_debug(1, " free preallocated %u/%u in group %u\n",
3601 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3602 (unsigned) next - bit, (unsigned) group);
3605 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3606 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3607 EXT4_C2B(sbi, bit)),
3609 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3612 if (free != pa->pa_free) {
3613 ext4_msg(e4b->bd_sb, KERN_CRIT,
3614 "pa %p: logic %lu, phys. %lu, len %lu",
3615 pa, (unsigned long) pa->pa_lstart,
3616 (unsigned long) pa->pa_pstart,
3617 (unsigned long) pa->pa_len);
3618 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3621 * pa is already deleted so we use the value obtained
3622 * from the bitmap and continue.
3625 atomic_add(free, &sbi->s_mb_discarded);
3630 static noinline_for_stack int
3631 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3632 struct ext4_prealloc_space *pa)
3634 struct super_block *sb = e4b->bd_sb;
3638 trace_ext4_mb_release_group_pa(sb, pa);
3639 BUG_ON(pa->pa_deleted == 0);
3640 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3641 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3642 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3643 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3644 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3650 * releases all preallocations in given group
3652 * first, we need to decide discard policy:
3653 * - when do we discard
3655 * - how many do we discard
3656 * 1) how many requested
3658 static noinline_for_stack int
3659 ext4_mb_discard_group_preallocations(struct super_block *sb,
3660 ext4_group_t group, int needed)
3662 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3663 struct buffer_head *bitmap_bh = NULL;
3664 struct ext4_prealloc_space *pa, *tmp;
3665 struct list_head list;
3666 struct ext4_buddy e4b;
3671 mb_debug(1, "discard preallocation for group %u\n", group);
3673 if (list_empty(&grp->bb_prealloc_list))
3676 bitmap_bh = ext4_read_block_bitmap(sb, group);
3677 if (bitmap_bh == NULL) {
3678 ext4_error(sb, "Error reading block bitmap for %u", group);
3682 err = ext4_mb_load_buddy(sb, group, &e4b);
3684 ext4_error(sb, "Error loading buddy information for %u", group);
3690 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3692 INIT_LIST_HEAD(&list);
3694 ext4_lock_group(sb, group);
3695 list_for_each_entry_safe(pa, tmp,
3696 &grp->bb_prealloc_list, pa_group_list) {
3697 spin_lock(&pa->pa_lock);
3698 if (atomic_read(&pa->pa_count)) {
3699 spin_unlock(&pa->pa_lock);
3703 if (pa->pa_deleted) {
3704 spin_unlock(&pa->pa_lock);
3708 /* seems this one can be freed ... */
3711 /* we can trust pa_free ... */
3712 free += pa->pa_free;
3714 spin_unlock(&pa->pa_lock);
3716 list_del(&pa->pa_group_list);
3717 list_add(&pa->u.pa_tmp_list, &list);
3720 /* if we still need more blocks and some PAs were used, try again */
3721 if (free < needed && busy) {
3723 ext4_unlock_group(sb, group);
3725 * Yield the CPU here so that we don't get soft lockup
3726 * in non preempt case.
3732 /* found anything to free? */
3733 if (list_empty(&list)) {
3738 /* now free all selected PAs */
3739 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3741 /* remove from object (inode or locality group) */
3742 spin_lock(pa->pa_obj_lock);
3743 list_del_rcu(&pa->pa_inode_list);
3744 spin_unlock(pa->pa_obj_lock);
3746 if (pa->pa_type == MB_GROUP_PA)
3747 ext4_mb_release_group_pa(&e4b, pa);
3749 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3751 list_del(&pa->u.pa_tmp_list);
3752 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3756 ext4_unlock_group(sb, group);
3757 ext4_mb_unload_buddy(&e4b);
3763 * releases all non-used preallocated blocks for given inode
3765 * It's important to discard preallocations under i_data_sem
3766 * We don't want another block to be served from the prealloc
3767 * space when we are discarding the inode prealloc space.
3769 * FIXME!! Make sure it is valid at all the call sites
3771 void ext4_discard_preallocations(struct inode *inode)
3773 struct ext4_inode_info *ei = EXT4_I(inode);
3774 struct super_block *sb = inode->i_sb;
3775 struct buffer_head *bitmap_bh = NULL;
3776 struct ext4_prealloc_space *pa, *tmp;
3777 ext4_group_t group = 0;
3778 struct list_head list;
3779 struct ext4_buddy e4b;
3782 if (!S_ISREG(inode->i_mode)) {
3783 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3787 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3788 trace_ext4_discard_preallocations(inode);
3790 INIT_LIST_HEAD(&list);
3793 /* first, collect all pa's in the inode */
3794 spin_lock(&ei->i_prealloc_lock);
3795 while (!list_empty(&ei->i_prealloc_list)) {
3796 pa = list_entry(ei->i_prealloc_list.next,
3797 struct ext4_prealloc_space, pa_inode_list);
3798 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3799 spin_lock(&pa->pa_lock);
3800 if (atomic_read(&pa->pa_count)) {
3801 /* this shouldn't happen often - nobody should
3802 * use preallocation while we're discarding it */
3803 spin_unlock(&pa->pa_lock);
3804 spin_unlock(&ei->i_prealloc_lock);
3805 ext4_msg(sb, KERN_ERR,
3806 "uh-oh! used pa while discarding");
3808 schedule_timeout_uninterruptible(HZ);
3812 if (pa->pa_deleted == 0) {
3814 spin_unlock(&pa->pa_lock);
3815 list_del_rcu(&pa->pa_inode_list);
3816 list_add(&pa->u.pa_tmp_list, &list);
3820 /* someone is deleting pa right now */
3821 spin_unlock(&pa->pa_lock);
3822 spin_unlock(&ei->i_prealloc_lock);
3824 /* we have to wait here because pa_deleted
3825 * doesn't mean pa is already unlinked from
3826 * the list. as we might be called from
3827 * ->clear_inode() the inode will get freed
3828 * and concurrent thread which is unlinking
3829 * pa from inode's list may access already
3830 * freed memory, bad-bad-bad */
3832 /* XXX: if this happens too often, we can
3833 * add a flag to force wait only in case
3834 * of ->clear_inode(), but not in case of
3835 * regular truncate */
3836 schedule_timeout_uninterruptible(HZ);
3839 spin_unlock(&ei->i_prealloc_lock);
3841 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3842 BUG_ON(pa->pa_type != MB_INODE_PA);
3843 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3845 err = ext4_mb_load_buddy(sb, group, &e4b);
3847 ext4_error(sb, "Error loading buddy information for %u",
3852 bitmap_bh = ext4_read_block_bitmap(sb, group);
3853 if (bitmap_bh == NULL) {
3854 ext4_error(sb, "Error reading block bitmap for %u",
3856 ext4_mb_unload_buddy(&e4b);
3860 ext4_lock_group(sb, group);
3861 list_del(&pa->pa_group_list);
3862 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3863 ext4_unlock_group(sb, group);
3865 ext4_mb_unload_buddy(&e4b);
3868 list_del(&pa->u.pa_tmp_list);
3869 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3873 #ifdef CONFIG_EXT4_DEBUG
3874 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3876 struct super_block *sb = ac->ac_sb;
3877 ext4_group_t ngroups, i;
3879 if (!mb_enable_debug ||
3880 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3883 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
3884 " Allocation context details:");
3885 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
3886 ac->ac_status, ac->ac_flags);
3887 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
3888 "goal %lu/%lu/%lu@%lu, "
3889 "best %lu/%lu/%lu@%lu cr %d",
3890 (unsigned long)ac->ac_o_ex.fe_group,
3891 (unsigned long)ac->ac_o_ex.fe_start,
3892 (unsigned long)ac->ac_o_ex.fe_len,
3893 (unsigned long)ac->ac_o_ex.fe_logical,
3894 (unsigned long)ac->ac_g_ex.fe_group,
3895 (unsigned long)ac->ac_g_ex.fe_start,
3896 (unsigned long)ac->ac_g_ex.fe_len,
3897 (unsigned long)ac->ac_g_ex.fe_logical,
3898 (unsigned long)ac->ac_b_ex.fe_group,
3899 (unsigned long)ac->ac_b_ex.fe_start,
3900 (unsigned long)ac->ac_b_ex.fe_len,
3901 (unsigned long)ac->ac_b_ex.fe_logical,
3902 (int)ac->ac_criteria);
3903 ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found",
3904 ac->ac_ex_scanned, ac->ac_found);
3905 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
3906 ngroups = ext4_get_groups_count(sb);
3907 for (i = 0; i < ngroups; i++) {
3908 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3909 struct ext4_prealloc_space *pa;
3910 ext4_grpblk_t start;
3911 struct list_head *cur;
3912 ext4_lock_group(sb, i);
3913 list_for_each(cur, &grp->bb_prealloc_list) {
3914 pa = list_entry(cur, struct ext4_prealloc_space,
3916 spin_lock(&pa->pa_lock);
3917 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3919 spin_unlock(&pa->pa_lock);
3920 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3923 ext4_unlock_group(sb, i);
3925 if (grp->bb_free == 0)
3927 printk(KERN_ERR "%u: %d/%d \n",
3928 i, grp->bb_free, grp->bb_fragments);
3930 printk(KERN_ERR "\n");
3933 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3940 * We use locality group preallocation for small size file. The size of the
3941 * file is determined by the current size or the resulting size after
3942 * allocation which ever is larger
3944 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3946 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3948 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3949 int bsbits = ac->ac_sb->s_blocksize_bits;
3952 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3955 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3958 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3959 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3962 if ((size == isize) &&
3963 !ext4_fs_is_busy(sbi) &&
3964 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3965 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3969 if (sbi->s_mb_group_prealloc <= 0) {
3970 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3974 /* don't use group allocation for large files */
3975 size = max(size, isize);
3976 if (size > sbi->s_mb_stream_request) {
3977 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3981 BUG_ON(ac->ac_lg != NULL);
3983 * locality group prealloc space are per cpu. The reason for having
3984 * per cpu locality group is to reduce the contention between block
3985 * request from multiple CPUs.
3987 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
3989 /* we're going to use group allocation */
3990 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3992 /* serialize all allocations in the group */
3993 mutex_lock(&ac->ac_lg->lg_mutex);
3996 static noinline_for_stack int
3997 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
3998 struct ext4_allocation_request *ar)
4000 struct super_block *sb = ar->inode->i_sb;
4001 struct ext4_sb_info *sbi = EXT4_SB(sb);
4002 struct ext4_super_block *es = sbi->s_es;
4006 ext4_grpblk_t block;
4008 /* we can't allocate > group size */
4011 /* just a dirty hack to filter too big requests */
4012 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4013 len = EXT4_CLUSTERS_PER_GROUP(sb);
4015 /* start searching from the goal */
4017 if (goal < le32_to_cpu(es->s_first_data_block) ||
4018 goal >= ext4_blocks_count(es))
4019 goal = le32_to_cpu(es->s_first_data_block);
4020 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4022 /* set up allocation goals */
4023 ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1);
4024 ac->ac_status = AC_STATUS_CONTINUE;
4026 ac->ac_inode = ar->inode;
4027 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4028 ac->ac_o_ex.fe_group = group;
4029 ac->ac_o_ex.fe_start = block;
4030 ac->ac_o_ex.fe_len = len;
4031 ac->ac_g_ex = ac->ac_o_ex;
4032 ac->ac_flags = ar->flags;
4034 /* we have to define context: we'll we work with a file or
4035 * locality group. this is a policy, actually */
4036 ext4_mb_group_or_file(ac);
4038 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4039 "left: %u/%u, right %u/%u to %swritable\n",
4040 (unsigned) ar->len, (unsigned) ar->logical,
4041 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4042 (unsigned) ar->lleft, (unsigned) ar->pleft,
4043 (unsigned) ar->lright, (unsigned) ar->pright,
4044 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4049 static noinline_for_stack void
4050 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4051 struct ext4_locality_group *lg,
4052 int order, int total_entries)
4054 ext4_group_t group = 0;
4055 struct ext4_buddy e4b;
4056 struct list_head discard_list;
4057 struct ext4_prealloc_space *pa, *tmp;
4059 mb_debug(1, "discard locality group preallocation\n");
4061 INIT_LIST_HEAD(&discard_list);
4063 spin_lock(&lg->lg_prealloc_lock);
4064 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4066 spin_lock(&pa->pa_lock);
4067 if (atomic_read(&pa->pa_count)) {
4069 * This is the pa that we just used
4070 * for block allocation. So don't
4073 spin_unlock(&pa->pa_lock);
4076 if (pa->pa_deleted) {
4077 spin_unlock(&pa->pa_lock);
4080 /* only lg prealloc space */
4081 BUG_ON(pa->pa_type != MB_GROUP_PA);
4083 /* seems this one can be freed ... */
4085 spin_unlock(&pa->pa_lock);
4087 list_del_rcu(&pa->pa_inode_list);
4088 list_add(&pa->u.pa_tmp_list, &discard_list);
4091 if (total_entries <= 5) {
4093 * we want to keep only 5 entries
4094 * allowing it to grow to 8. This
4095 * mak sure we don't call discard
4096 * soon for this list.
4101 spin_unlock(&lg->lg_prealloc_lock);
4103 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4105 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4106 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4107 ext4_error(sb, "Error loading buddy information for %u",
4111 ext4_lock_group(sb, group);
4112 list_del(&pa->pa_group_list);
4113 ext4_mb_release_group_pa(&e4b, pa);
4114 ext4_unlock_group(sb, group);
4116 ext4_mb_unload_buddy(&e4b);
4117 list_del(&pa->u.pa_tmp_list);
4118 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4123 * We have incremented pa_count. So it cannot be freed at this
4124 * point. Also we hold lg_mutex. So no parallel allocation is
4125 * possible from this lg. That means pa_free cannot be updated.
4127 * A parallel ext4_mb_discard_group_preallocations is possible.
4128 * which can cause the lg_prealloc_list to be updated.
4131 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4133 int order, added = 0, lg_prealloc_count = 1;
4134 struct super_block *sb = ac->ac_sb;
4135 struct ext4_locality_group *lg = ac->ac_lg;
4136 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4138 order = fls(pa->pa_free) - 1;
4139 if (order > PREALLOC_TB_SIZE - 1)
4140 /* The max size of hash table is PREALLOC_TB_SIZE */
4141 order = PREALLOC_TB_SIZE - 1;
4142 /* Add the prealloc space to lg */
4143 spin_lock(&lg->lg_prealloc_lock);
4144 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4146 spin_lock(&tmp_pa->pa_lock);
4147 if (tmp_pa->pa_deleted) {
4148 spin_unlock(&tmp_pa->pa_lock);
4151 if (!added && pa->pa_free < tmp_pa->pa_free) {
4152 /* Add to the tail of the previous entry */
4153 list_add_tail_rcu(&pa->pa_inode_list,
4154 &tmp_pa->pa_inode_list);
4157 * we want to count the total
4158 * number of entries in the list
4161 spin_unlock(&tmp_pa->pa_lock);
4162 lg_prealloc_count++;
4165 list_add_tail_rcu(&pa->pa_inode_list,
4166 &lg->lg_prealloc_list[order]);
4167 spin_unlock(&lg->lg_prealloc_lock);
4169 /* Now trim the list to be not more than 8 elements */
4170 if (lg_prealloc_count > 8) {
4171 ext4_mb_discard_lg_preallocations(sb, lg,
4172 order, lg_prealloc_count);
4179 * release all resource we used in allocation
4181 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4183 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4184 struct ext4_prealloc_space *pa = ac->ac_pa;
4186 if (pa->pa_type == MB_GROUP_PA) {
4187 /* see comment in ext4_mb_use_group_pa() */
4188 spin_lock(&pa->pa_lock);
4189 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4190 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4191 pa->pa_free -= ac->ac_b_ex.fe_len;
4192 pa->pa_len -= ac->ac_b_ex.fe_len;
4193 spin_unlock(&pa->pa_lock);
4198 * We want to add the pa to the right bucket.
4199 * Remove it from the list and while adding
4200 * make sure the list to which we are adding
4203 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4204 spin_lock(pa->pa_obj_lock);
4205 list_del_rcu(&pa->pa_inode_list);
4206 spin_unlock(pa->pa_obj_lock);
4207 ext4_mb_add_n_trim(ac);
4209 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4211 if (ac->ac_bitmap_page)
4212 page_cache_release(ac->ac_bitmap_page);
4213 if (ac->ac_buddy_page)
4214 page_cache_release(ac->ac_buddy_page);
4215 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4216 mutex_unlock(&ac->ac_lg->lg_mutex);
4217 ext4_mb_collect_stats(ac);
4221 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4223 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4227 trace_ext4_mb_discard_preallocations(sb, needed);
4228 for (i = 0; i < ngroups && needed > 0; i++) {
4229 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4238 * Main entry point into mballoc to allocate blocks
4239 * it tries to use preallocation first, then falls back
4240 * to usual allocation
4242 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4243 struct ext4_allocation_request *ar, int *errp)
4246 struct ext4_allocation_context *ac = NULL;
4247 struct ext4_sb_info *sbi;
4248 struct super_block *sb;
4249 ext4_fsblk_t block = 0;
4250 unsigned int inquota = 0;
4251 unsigned int reserv_clstrs = 0;
4253 sb = ar->inode->i_sb;
4256 trace_ext4_request_blocks(ar);
4258 /* Allow to use superuser reservation for quota file */
4259 if (IS_NOQUOTA(ar->inode))
4260 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4263 * For delayed allocation, we could skip the ENOSPC and
4264 * EDQUOT check, as blocks and quotas have been already
4265 * reserved when data being copied into pagecache.
4267 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4268 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4270 /* Without delayed allocation we need to verify
4271 * there is enough free blocks to do block allocation
4272 * and verify allocation doesn't exceed the quota limits.
4275 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4277 /* let others to free the space */
4279 ar->len = ar->len >> 1;
4285 reserv_clstrs = ar->len;
4286 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4287 dquot_alloc_block_nofail(ar->inode,
4288 EXT4_C2B(sbi, ar->len));
4291 dquot_alloc_block(ar->inode,
4292 EXT4_C2B(sbi, ar->len))) {
4294 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4305 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4312 *errp = ext4_mb_initialize_context(ac, ar);
4318 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4319 if (!ext4_mb_use_preallocated(ac)) {
4320 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4321 ext4_mb_normalize_request(ac, ar);
4323 /* allocate space in core */
4324 *errp = ext4_mb_regular_allocator(ac);
4326 ext4_discard_allocated_blocks(ac);
4330 /* as we've just preallocated more space than
4331 * user requested orinally, we store allocated
4332 * space in a special descriptor */
4333 if (ac->ac_status == AC_STATUS_FOUND &&
4334 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4335 ext4_mb_new_preallocation(ac);
4337 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4338 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4339 if (*errp == -EAGAIN) {
4341 * drop the reference that we took
4342 * in ext4_mb_use_best_found
4344 ext4_mb_release_context(ac);
4345 ac->ac_b_ex.fe_group = 0;
4346 ac->ac_b_ex.fe_start = 0;
4347 ac->ac_b_ex.fe_len = 0;
4348 ac->ac_status = AC_STATUS_CONTINUE;
4351 ext4_discard_allocated_blocks(ac);
4354 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4355 ar->len = ac->ac_b_ex.fe_len;
4358 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4366 ac->ac_b_ex.fe_len = 0;
4368 ext4_mb_show_ac(ac);
4370 ext4_mb_release_context(ac);
4373 kmem_cache_free(ext4_ac_cachep, ac);
4374 if (inquota && ar->len < inquota)
4375 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4377 if (!ext4_test_inode_state(ar->inode,
4378 EXT4_STATE_DELALLOC_RESERVED))
4379 /* release all the reserved blocks if non delalloc */
4380 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4384 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4390 * We can merge two free data extents only if the physical blocks
4391 * are contiguous, AND the extents were freed by the same transaction,
4392 * AND the blocks are associated with the same group.
4394 static int can_merge(struct ext4_free_data *entry1,
4395 struct ext4_free_data *entry2)
4397 if ((entry1->efd_tid == entry2->efd_tid) &&
4398 (entry1->efd_group == entry2->efd_group) &&
4399 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4404 static noinline_for_stack int
4405 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4406 struct ext4_free_data *new_entry)
4408 ext4_group_t group = e4b->bd_group;
4409 ext4_grpblk_t cluster;
4410 struct ext4_free_data *entry;
4411 struct ext4_group_info *db = e4b->bd_info;
4412 struct super_block *sb = e4b->bd_sb;
4413 struct ext4_sb_info *sbi = EXT4_SB(sb);
4414 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4415 struct rb_node *parent = NULL, *new_node;
4417 BUG_ON(!ext4_handle_valid(handle));
4418 BUG_ON(e4b->bd_bitmap_page == NULL);
4419 BUG_ON(e4b->bd_buddy_page == NULL);
4421 new_node = &new_entry->efd_node;
4422 cluster = new_entry->efd_start_cluster;
4425 /* first free block exent. We need to
4426 protect buddy cache from being freed,
4427 * otherwise we'll refresh it from
4428 * on-disk bitmap and lose not-yet-available
4430 page_cache_get(e4b->bd_buddy_page);
4431 page_cache_get(e4b->bd_bitmap_page);
4435 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4436 if (cluster < entry->efd_start_cluster)
4438 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4439 n = &(*n)->rb_right;
4441 ext4_grp_locked_error(sb, group, 0,
4442 ext4_group_first_block_no(sb, group) +
4443 EXT4_C2B(sbi, cluster),
4444 "Block already on to-be-freed list");
4449 rb_link_node(new_node, parent, n);
4450 rb_insert_color(new_node, &db->bb_free_root);
4452 /* Now try to see the extent can be merged to left and right */
4453 node = rb_prev(new_node);
4455 entry = rb_entry(node, struct ext4_free_data, efd_node);
4456 if (can_merge(entry, new_entry)) {
4457 new_entry->efd_start_cluster = entry->efd_start_cluster;
4458 new_entry->efd_count += entry->efd_count;
4459 rb_erase(node, &(db->bb_free_root));
4460 ext4_journal_callback_del(handle, &entry->efd_jce);
4461 kmem_cache_free(ext4_free_data_cachep, entry);
4465 node = rb_next(new_node);
4467 entry = rb_entry(node, struct ext4_free_data, efd_node);
4468 if (can_merge(new_entry, entry)) {
4469 new_entry->efd_count += entry->efd_count;
4470 rb_erase(node, &(db->bb_free_root));
4471 ext4_journal_callback_del(handle, &entry->efd_jce);
4472 kmem_cache_free(ext4_free_data_cachep, entry);
4475 /* Add the extent to transaction's private list */
4476 ext4_journal_callback_add(handle, ext4_free_data_callback,
4477 &new_entry->efd_jce);
4482 * ext4_free_blocks() -- Free given blocks and update quota
4483 * @handle: handle for this transaction
4485 * @block: start physical block to free
4486 * @count: number of blocks to count
4487 * @flags: flags used by ext4_free_blocks
4489 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4490 struct buffer_head *bh, ext4_fsblk_t block,
4491 unsigned long count, int flags)
4493 struct buffer_head *bitmap_bh = NULL;
4494 struct super_block *sb = inode->i_sb;
4495 struct ext4_group_desc *gdp;
4496 unsigned long freed = 0;
4497 unsigned int overflow;
4499 struct buffer_head *gd_bh;
4500 ext4_group_t block_group;
4501 struct ext4_sb_info *sbi;
4502 struct ext4_buddy e4b;
4503 unsigned int count_clusters;
4509 BUG_ON(block != bh->b_blocknr);
4511 block = bh->b_blocknr;
4515 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4516 !ext4_data_block_valid(sbi, block, count)) {
4517 ext4_error(sb, "Freeing blocks not in datazone - "
4518 "block = %llu, count = %lu", block, count);
4522 ext4_debug("freeing block %llu\n", block);
4523 trace_ext4_free_blocks(inode, block, count, flags);
4525 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4526 struct buffer_head *tbh = bh;
4529 BUG_ON(bh && (count > 1));
4531 for (i = 0; i < count; i++) {
4533 tbh = sb_find_get_block(inode->i_sb,
4537 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4538 inode, tbh, block + i);
4543 * We need to make sure we don't reuse the freed block until
4544 * after the transaction is committed, which we can do by
4545 * treating the block as metadata, below. We make an
4546 * exception if the inode is to be written in writeback mode
4547 * since writeback mode has weak data consistency guarantees.
4549 if (!ext4_should_writeback_data(inode))
4550 flags |= EXT4_FREE_BLOCKS_METADATA;
4553 * If the extent to be freed does not begin on a cluster
4554 * boundary, we need to deal with partial clusters at the
4555 * beginning and end of the extent. Normally we will free
4556 * blocks at the beginning or the end unless we are explicitly
4557 * requested to avoid doing so.
4559 overflow = block & (sbi->s_cluster_ratio - 1);
4561 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4562 overflow = sbi->s_cluster_ratio - overflow;
4564 if (count > overflow)
4573 overflow = count & (sbi->s_cluster_ratio - 1);
4575 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4576 if (count > overflow)
4581 count += sbi->s_cluster_ratio - overflow;
4586 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4589 * Check to see if we are freeing blocks across a group
4592 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4593 overflow = EXT4_C2B(sbi, bit) + count -
4594 EXT4_BLOCKS_PER_GROUP(sb);
4597 count_clusters = EXT4_B2C(sbi, count);
4598 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4603 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4609 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4610 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4611 in_range(block, ext4_inode_table(sb, gdp),
4612 EXT4_SB(sb)->s_itb_per_group) ||
4613 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4614 EXT4_SB(sb)->s_itb_per_group)) {
4616 ext4_error(sb, "Freeing blocks in system zone - "
4617 "Block = %llu, count = %lu", block, count);
4618 /* err = 0. ext4_std_error should be a no op */
4622 BUFFER_TRACE(bitmap_bh, "getting write access");
4623 err = ext4_journal_get_write_access(handle, bitmap_bh);
4628 * We are about to modify some metadata. Call the journal APIs
4629 * to unshare ->b_data if a currently-committing transaction is
4632 BUFFER_TRACE(gd_bh, "get_write_access");
4633 err = ext4_journal_get_write_access(handle, gd_bh);
4636 #ifdef AGGRESSIVE_CHECK
4639 for (i = 0; i < count_clusters; i++)
4640 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4643 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4645 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4649 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4650 struct ext4_free_data *new_entry;
4652 * blocks being freed are metadata. these blocks shouldn't
4653 * be used until this transaction is committed
4655 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4657 ext4_mb_unload_buddy(&e4b);
4661 new_entry->efd_start_cluster = bit;
4662 new_entry->efd_group = block_group;
4663 new_entry->efd_count = count_clusters;
4664 new_entry->efd_tid = handle->h_transaction->t_tid;
4666 ext4_lock_group(sb, block_group);
4667 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4668 ext4_mb_free_metadata(handle, &e4b, new_entry);
4670 /* need to update group_info->bb_free and bitmap
4671 * with group lock held. generate_buddy look at
4672 * them with group lock_held
4674 if (test_opt(sb, DISCARD)) {
4675 err = ext4_issue_discard(sb, block_group, bit, count);
4676 if (err && err != -EOPNOTSUPP)
4677 ext4_msg(sb, KERN_WARNING, "discard request in"
4678 " group:%d block:%d count:%lu failed"
4679 " with %d", block_group, bit, count,
4684 ext4_lock_group(sb, block_group);
4685 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4686 mb_free_blocks(inode, &e4b, bit, count_clusters);
4689 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4690 ext4_free_group_clusters_set(sb, gdp, ret);
4691 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4692 ext4_group_desc_csum_set(sb, block_group, gdp);
4693 ext4_unlock_group(sb, block_group);
4694 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4696 if (sbi->s_log_groups_per_flex) {
4697 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4698 atomic_add(count_clusters,
4699 &sbi->s_flex_groups[flex_group].free_clusters);
4702 ext4_mb_unload_buddy(&e4b);
4706 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4707 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4709 /* We dirtied the bitmap block */
4710 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4711 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4713 /* And the group descriptor block */
4714 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4715 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4719 if (overflow && !err) {
4727 ext4_std_error(sb, err);
4732 * ext4_group_add_blocks() -- Add given blocks to an existing group
4733 * @handle: handle to this transaction
4735 * @block: start physical block to add to the block group
4736 * @count: number of blocks to free
4738 * This marks the blocks as free in the bitmap and buddy.
4740 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4741 ext4_fsblk_t block, unsigned long count)
4743 struct buffer_head *bitmap_bh = NULL;
4744 struct buffer_head *gd_bh;
4745 ext4_group_t block_group;
4748 struct ext4_group_desc *desc;
4749 struct ext4_sb_info *sbi = EXT4_SB(sb);
4750 struct ext4_buddy e4b;
4751 int err = 0, ret, blk_free_count;
4752 ext4_grpblk_t blocks_freed;
4754 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4759 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4761 * Check to see if we are freeing blocks across a group
4764 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4765 ext4_warning(sb, "too much blocks added to group %u\n",
4771 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4777 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4783 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4784 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4785 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4786 in_range(block + count - 1, ext4_inode_table(sb, desc),
4787 sbi->s_itb_per_group)) {
4788 ext4_error(sb, "Adding blocks in system zones - "
4789 "Block = %llu, count = %lu",
4795 BUFFER_TRACE(bitmap_bh, "getting write access");
4796 err = ext4_journal_get_write_access(handle, bitmap_bh);
4801 * We are about to modify some metadata. Call the journal APIs
4802 * to unshare ->b_data if a currently-committing transaction is
4805 BUFFER_TRACE(gd_bh, "get_write_access");
4806 err = ext4_journal_get_write_access(handle, gd_bh);
4810 for (i = 0, blocks_freed = 0; i < count; i++) {
4811 BUFFER_TRACE(bitmap_bh, "clear bit");
4812 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4813 ext4_error(sb, "bit already cleared for block %llu",
4814 (ext4_fsblk_t)(block + i));
4815 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4821 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4826 * need to update group_info->bb_free and bitmap
4827 * with group lock held. generate_buddy look at
4828 * them with group lock_held
4830 ext4_lock_group(sb, block_group);
4831 mb_clear_bits(bitmap_bh->b_data, bit, count);
4832 mb_free_blocks(NULL, &e4b, bit, count);
4833 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4834 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4835 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
4836 ext4_group_desc_csum_set(sb, block_group, desc);
4837 ext4_unlock_group(sb, block_group);
4838 percpu_counter_add(&sbi->s_freeclusters_counter,
4839 EXT4_B2C(sbi, blocks_freed));
4841 if (sbi->s_log_groups_per_flex) {
4842 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4843 atomic_add(EXT4_B2C(sbi, blocks_freed),
4844 &sbi->s_flex_groups[flex_group].free_clusters);
4847 ext4_mb_unload_buddy(&e4b);
4849 /* We dirtied the bitmap block */
4850 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4851 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4853 /* And the group descriptor block */
4854 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4855 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4861 ext4_std_error(sb, err);
4866 * ext4_trim_extent -- function to TRIM one single free extent in the group
4867 * @sb: super block for the file system
4868 * @start: starting block of the free extent in the alloc. group
4869 * @count: number of blocks to TRIM
4870 * @group: alloc. group we are working with
4871 * @e4b: ext4 buddy for the group
4873 * Trim "count" blocks starting at "start" in the "group". To assure that no
4874 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4875 * be called with under the group lock.
4877 static int ext4_trim_extent(struct super_block *sb, int start, int count,
4878 ext4_group_t group, struct ext4_buddy *e4b)
4880 struct ext4_free_extent ex;
4883 trace_ext4_trim_extent(sb, group, start, count);
4885 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4887 ex.fe_start = start;
4888 ex.fe_group = group;
4892 * Mark blocks used, so no one can reuse them while
4895 mb_mark_used(e4b, &ex);
4896 ext4_unlock_group(sb, group);
4897 ret = ext4_issue_discard(sb, group, start, count);
4898 ext4_lock_group(sb, group);
4899 mb_free_blocks(NULL, e4b, start, ex.fe_len);
4904 * ext4_trim_all_free -- function to trim all free space in alloc. group
4905 * @sb: super block for file system
4906 * @group: group to be trimmed
4907 * @start: first group block to examine
4908 * @max: last group block to examine
4909 * @minblocks: minimum extent block count
4911 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4912 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4916 * ext4_trim_all_free walks through group's block bitmap searching for free
4917 * extents. When the free extent is found, mark it as used in group buddy
4918 * bitmap. Then issue a TRIM command on this extent and free the extent in
4919 * the group buddy bitmap. This is done until whole group is scanned.
4921 static ext4_grpblk_t
4922 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4923 ext4_grpblk_t start, ext4_grpblk_t max,
4924 ext4_grpblk_t minblocks)
4927 ext4_grpblk_t next, count = 0, free_count = 0;
4928 struct ext4_buddy e4b;
4931 trace_ext4_trim_all_free(sb, group, start, max);
4933 ret = ext4_mb_load_buddy(sb, group, &e4b);
4935 ext4_error(sb, "Error in loading buddy "
4936 "information for %u", group);
4939 bitmap = e4b.bd_bitmap;
4941 ext4_lock_group(sb, group);
4942 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4943 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4946 start = (e4b.bd_info->bb_first_free > start) ?
4947 e4b.bd_info->bb_first_free : start;
4949 while (start <= max) {
4950 start = mb_find_next_zero_bit(bitmap, max + 1, start);
4953 next = mb_find_next_bit(bitmap, max + 1, start);
4955 if ((next - start) >= minblocks) {
4956 ret = ext4_trim_extent(sb, start,
4957 next - start, group, &e4b);
4958 if (ret && ret != -EOPNOTSUPP)
4961 count += next - start;
4963 free_count += next - start;
4966 if (fatal_signal_pending(current)) {
4967 count = -ERESTARTSYS;
4971 if (need_resched()) {
4972 ext4_unlock_group(sb, group);
4974 ext4_lock_group(sb, group);
4977 if ((e4b.bd_info->bb_free - free_count) < minblocks)
4983 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
4986 ext4_unlock_group(sb, group);
4987 ext4_mb_unload_buddy(&e4b);
4989 ext4_debug("trimmed %d blocks in the group %d\n",
4996 * ext4_trim_fs() -- trim ioctl handle function
4997 * @sb: superblock for filesystem
4998 * @range: fstrim_range structure
5000 * start: First Byte to trim
5001 * len: number of Bytes to trim from start
5002 * minlen: minimum extent length in Bytes
5003 * ext4_trim_fs goes through all allocation groups containing Bytes from
5004 * start to start+len. For each such a group ext4_trim_all_free function
5005 * is invoked to trim all free space.
5007 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5009 struct ext4_group_info *grp;
5010 ext4_group_t group, first_group, last_group;
5011 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5012 uint64_t start, end, minlen, trimmed = 0;
5013 ext4_fsblk_t first_data_blk =
5014 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5015 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5018 start = range->start >> sb->s_blocksize_bits;
5019 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5020 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5021 range->minlen >> sb->s_blocksize_bits);
5023 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5024 start >= max_blks ||
5025 range->len < sb->s_blocksize)
5027 if (end >= max_blks)
5029 if (end <= first_data_blk)
5031 if (start < first_data_blk)
5032 start = first_data_blk;
5034 /* Determine first and last group to examine based on start and end */
5035 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5036 &first_group, &first_cluster);
5037 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5038 &last_group, &last_cluster);
5040 /* end now represents the last cluster to discard in this group */
5041 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5043 for (group = first_group; group <= last_group; group++) {
5044 grp = ext4_get_group_info(sb, group);
5045 /* We only do this if the grp has never been initialized */
5046 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5047 ret = ext4_mb_init_group(sb, group);
5053 * For all the groups except the last one, last cluster will
5054 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5055 * change it for the last group, note that last_cluster is
5056 * already computed earlier by ext4_get_group_no_and_offset()
5058 if (group == last_group)
5061 if (grp->bb_free >= minlen) {
5062 cnt = ext4_trim_all_free(sb, group, first_cluster,
5072 * For every group except the first one, we are sure
5073 * that the first cluster to discard will be cluster #0.
5079 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5082 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;