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/slab.h>
28 #include <trace/events/ext4.h>
32 * - test ext4_ext_search_left() and ext4_ext_search_right()
33 * - search for metadata in few groups
36 * - normalization should take into account whether file is still open
37 * - discard preallocations if no free space left (policy?)
38 * - don't normalize tails
40 * - reservation for superuser
43 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
44 * - track min/max extents in each group for better group selection
45 * - mb_mark_used() may allocate chunk right after splitting buddy
46 * - tree of groups sorted by number of free blocks
51 * The allocation request involve request for multiple number of blocks
52 * near to the goal(block) value specified.
54 * During initialization phase of the allocator we decide to use the
55 * group preallocation or inode preallocation depending on the size of
56 * the file. The size of the file could be the resulting file size we
57 * would have after allocation, or the current file size, which ever
58 * is larger. If the size is less than sbi->s_mb_stream_request we
59 * select to use the group preallocation. The default value of
60 * s_mb_stream_request is 16 blocks. This can also be tuned via
61 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
62 * terms of number of blocks.
64 * The main motivation for having small file use group preallocation is to
65 * ensure that we have small files closer together on the disk.
67 * First stage the allocator looks at the inode prealloc list,
68 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
69 * spaces for this particular inode. The inode prealloc space is
72 * pa_lstart -> the logical start block for this prealloc space
73 * pa_pstart -> the physical start block for this prealloc space
74 * pa_len -> length for this prealloc space (in clusters)
75 * pa_free -> free space available in this prealloc space (in clusters)
77 * The inode preallocation space is used looking at the _logical_ start
78 * block. If only the logical file block falls within the range of prealloc
79 * space we will consume the particular prealloc space. This makes sure that
80 * we have contiguous physical blocks representing the file blocks
82 * The important thing to be noted in case of inode prealloc space is that
83 * we don't modify the values associated to inode prealloc space except
86 * If we are not able to find blocks in the inode prealloc space and if we
87 * have the group allocation flag set then we look at the locality group
88 * prealloc space. These are per CPU prealloc list represented as
90 * ext4_sb_info.s_locality_groups[smp_processor_id()]
92 * The reason for having a per cpu locality group is to reduce the contention
93 * between CPUs. It is possible to get scheduled at this point.
95 * The locality group prealloc space is used looking at whether we have
96 * enough free space (pa_free) within the prealloc space.
98 * If we can't allocate blocks via inode prealloc or/and locality group
99 * prealloc then we look at the buddy cache. The buddy cache is represented
100 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
101 * mapped to the buddy and bitmap information regarding different
102 * groups. The buddy information is attached to buddy cache inode so that
103 * we can access them through the page cache. The information regarding
104 * each group is loaded via ext4_mb_load_buddy. The information involve
105 * block bitmap and buddy information. The information are stored in the
109 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
112 * one block each for bitmap and buddy information. So for each group we
113 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
114 * blocksize) blocks. So it can have information regarding groups_per_page
115 * which is blocks_per_page/2
117 * The buddy cache inode is not stored on disk. The inode is thrown
118 * away when the filesystem is unmounted.
120 * We look for count number of blocks in the buddy cache. If we were able
121 * to locate that many free blocks we return with additional information
122 * regarding rest of the contiguous physical block available
124 * Before allocating blocks via buddy cache we normalize the request
125 * blocks. This ensure we ask for more blocks that we needed. The extra
126 * blocks that we get after allocation is added to the respective prealloc
127 * list. In case of inode preallocation we follow a list of heuristics
128 * based on file size. This can be found in ext4_mb_normalize_request. If
129 * we are doing a group prealloc we try to normalize the request to
130 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
131 * dependent on the cluster size; for non-bigalloc file systems, it is
132 * 512 blocks. This can be tuned via
133 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
134 * terms of number of blocks. If we have mounted the file system with -O
135 * stripe=<value> option the group prealloc request is normalized to the
136 * the smallest multiple of the stripe value (sbi->s_stripe) which is
137 * greater than the default mb_group_prealloc.
139 * The regular allocator (using the buddy cache) supports a few tunables.
141 * /sys/fs/ext4/<partition>/mb_min_to_scan
142 * /sys/fs/ext4/<partition>/mb_max_to_scan
143 * /sys/fs/ext4/<partition>/mb_order2_req
145 * The regular allocator uses buddy scan only if the request len is power of
146 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
147 * value of s_mb_order2_reqs can be tuned via
148 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
149 * stripe size (sbi->s_stripe), we try to search for contiguous block in
150 * stripe size. This should result in better allocation on RAID setups. If
151 * not, we search in the specific group using bitmap for best extents. The
152 * tunable min_to_scan and max_to_scan control the behaviour here.
153 * min_to_scan indicate how long the mballoc __must__ look for a best
154 * extent and max_to_scan indicates how long the mballoc __can__ look for a
155 * best extent in the found extents. Searching for the blocks starts with
156 * the group specified as the goal value in allocation context via
157 * ac_g_ex. Each group is first checked based on the criteria whether it
158 * can be used for allocation. ext4_mb_good_group explains how the groups are
161 * Both the prealloc space are getting populated as above. So for the first
162 * request we will hit the buddy cache which will result in this prealloc
163 * space getting filled. The prealloc space is then later used for the
164 * subsequent request.
168 * mballoc operates on the following data:
170 * - in-core buddy (actually includes buddy and bitmap)
171 * - preallocation descriptors (PAs)
173 * there are two types of preallocations:
175 * assiged to specific inode and can be used for this inode only.
176 * it describes part of inode's space preallocated to specific
177 * physical blocks. any block from that preallocated can be used
178 * independent. the descriptor just tracks number of blocks left
179 * unused. so, before taking some block from descriptor, one must
180 * make sure corresponded logical block isn't allocated yet. this
181 * also means that freeing any block within descriptor's range
182 * must discard all preallocated blocks.
184 * assigned to specific locality group which does not translate to
185 * permanent set of inodes: inode can join and leave group. space
186 * from this type of preallocation can be used for any inode. thus
187 * it's consumed from the beginning to the end.
189 * relation between them can be expressed as:
190 * in-core buddy = on-disk bitmap + preallocation descriptors
192 * this mean blocks mballoc considers used are:
193 * - allocated blocks (persistent)
194 * - preallocated blocks (non-persistent)
196 * consistency in mballoc world means that at any time a block is either
197 * free or used in ALL structures. notice: "any time" should not be read
198 * literally -- time is discrete and delimited by locks.
200 * to keep it simple, we don't use block numbers, instead we count number of
201 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
203 * all operations can be expressed as:
204 * - init buddy: buddy = on-disk + PAs
205 * - new PA: buddy += N; PA = N
206 * - use inode PA: on-disk += N; PA -= N
207 * - discard inode PA buddy -= on-disk - PA; PA = 0
208 * - use locality group PA on-disk += N; PA -= N
209 * - discard locality group PA buddy -= PA; PA = 0
210 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
211 * is used in real operation because we can't know actual used
212 * bits from PA, only from on-disk bitmap
214 * if we follow this strict logic, then all operations above should be atomic.
215 * given some of them can block, we'd have to use something like semaphores
216 * killing performance on high-end SMP hardware. let's try to relax it using
217 * the following knowledge:
218 * 1) if buddy is referenced, it's already initialized
219 * 2) while block is used in buddy and the buddy is referenced,
220 * nobody can re-allocate that block
221 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
222 * bit set and PA claims same block, it's OK. IOW, one can set bit in
223 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
226 * so, now we're building a concurrency table:
229 * blocks for PA are allocated in the buddy, buddy must be referenced
230 * until PA is linked to allocation group to avoid concurrent buddy init
232 * we need to make sure that either on-disk bitmap or PA has uptodate data
233 * given (3) we care that PA-=N operation doesn't interfere with init
235 * the simplest way would be to have buddy initialized by the discard
236 * - use locality group PA
237 * again PA-=N must be serialized with init
238 * - discard locality group PA
239 * the simplest way would be to have buddy initialized by the discard
242 * i_data_sem serializes them
244 * discard process must wait until PA isn't used by another process
245 * - use locality group PA
246 * some mutex should serialize them
247 * - discard locality group PA
248 * discard process must wait until PA isn't used by another process
251 * i_data_sem or another mutex should serializes them
253 * discard process must wait until PA isn't used by another process
254 * - use locality group PA
255 * nothing wrong here -- they're different PAs covering different blocks
256 * - discard locality group PA
257 * discard process must wait until PA isn't used by another process
259 * now we're ready to make few consequences:
260 * - PA is referenced and while it is no discard is possible
261 * - PA is referenced until block isn't marked in on-disk bitmap
262 * - PA changes only after on-disk bitmap
263 * - discard must not compete with init. either init is done before
264 * any discard or they're serialized somehow
265 * - buddy init as sum of on-disk bitmap and PAs is done atomically
267 * a special case when we've used PA to emptiness. no need to modify buddy
268 * in this case, but we should care about concurrent init
273 * Logic in few words:
278 * mark bits in on-disk bitmap
281 * - use preallocation:
282 * find proper PA (per-inode or group)
284 * mark bits in on-disk bitmap
290 * mark bits in on-disk bitmap
293 * - discard preallocations in group:
295 * move them onto local list
296 * load on-disk bitmap
298 * remove PA from object (inode or locality group)
299 * mark free blocks in-core
301 * - discard inode's preallocations:
308 * - bitlock on a group (group)
309 * - object (inode/locality) (object)
320 * - release consumed pa:
325 * - generate in-core bitmap:
329 * - discard all for given object (inode, locality group):
334 * - discard all for given group:
341 static struct kmem_cache *ext4_pspace_cachep;
342 static struct kmem_cache *ext4_ac_cachep;
343 static struct kmem_cache *ext4_free_data_cachep;
345 /* We create slab caches for groupinfo data structures based on the
346 * superblock block size. There will be one per mounted filesystem for
347 * each unique s_blocksize_bits */
348 #define NR_GRPINFO_CACHES 8
349 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
351 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
352 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
353 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
354 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
357 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
359 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
361 static void ext4_free_data_callback(struct super_block *sb,
362 struct ext4_journal_cb_entry *jce, int rc);
364 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
366 #if BITS_PER_LONG == 64
367 *bit += ((unsigned long) addr & 7UL) << 3;
368 addr = (void *) ((unsigned long) addr & ~7UL);
369 #elif BITS_PER_LONG == 32
370 *bit += ((unsigned long) addr & 3UL) << 3;
371 addr = (void *) ((unsigned long) addr & ~3UL);
373 #error "how many bits you are?!"
378 static inline int mb_test_bit(int bit, void *addr)
381 * ext4_test_bit on architecture like powerpc
382 * needs unsigned long aligned address
384 addr = mb_correct_addr_and_bit(&bit, addr);
385 return ext4_test_bit(bit, addr);
388 static inline void mb_set_bit(int bit, void *addr)
390 addr = mb_correct_addr_and_bit(&bit, addr);
391 ext4_set_bit(bit, addr);
394 static inline void mb_clear_bit(int bit, void *addr)
396 addr = mb_correct_addr_and_bit(&bit, addr);
397 ext4_clear_bit(bit, addr);
400 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
402 int fix = 0, ret, tmpmax;
403 addr = mb_correct_addr_and_bit(&fix, addr);
407 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
413 static inline int mb_find_next_bit(void *addr, int max, int start)
415 int fix = 0, ret, tmpmax;
416 addr = mb_correct_addr_and_bit(&fix, addr);
420 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
426 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
430 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
433 if (order > e4b->bd_blkbits + 1) {
438 /* at order 0 we see each particular block */
440 *max = 1 << (e4b->bd_blkbits + 3);
441 return e4b->bd_bitmap;
444 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
445 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
451 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
452 int first, int count)
455 struct super_block *sb = e4b->bd_sb;
457 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
459 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
460 for (i = 0; i < count; i++) {
461 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
462 ext4_fsblk_t blocknr;
464 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
465 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
466 ext4_grp_locked_error(sb, e4b->bd_group,
467 inode ? inode->i_ino : 0,
469 "freeing block already freed "
473 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
477 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
481 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
483 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
484 for (i = 0; i < count; i++) {
485 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
486 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
490 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
492 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
493 unsigned char *b1, *b2;
495 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
496 b2 = (unsigned char *) bitmap;
497 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
498 if (b1[i] != b2[i]) {
499 ext4_msg(e4b->bd_sb, KERN_ERR,
500 "corruption in group %u "
501 "at byte %u(%u): %x in copy != %x "
503 e4b->bd_group, i, i * 8, b1[i], b2[i]);
511 static inline void mb_free_blocks_double(struct inode *inode,
512 struct ext4_buddy *e4b, int first, int count)
516 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
517 int first, int count)
521 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
527 #ifdef AGGRESSIVE_CHECK
529 #define MB_CHECK_ASSERT(assert) \
533 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
534 function, file, line, # assert); \
539 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
540 const char *function, int line)
542 struct super_block *sb = e4b->bd_sb;
543 int order = e4b->bd_blkbits + 1;
550 struct ext4_group_info *grp;
553 struct list_head *cur;
558 static int mb_check_counter;
559 if (mb_check_counter++ % 100 != 0)
564 buddy = mb_find_buddy(e4b, order, &max);
565 MB_CHECK_ASSERT(buddy);
566 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
567 MB_CHECK_ASSERT(buddy2);
568 MB_CHECK_ASSERT(buddy != buddy2);
569 MB_CHECK_ASSERT(max * 2 == max2);
572 for (i = 0; i < max; i++) {
574 if (mb_test_bit(i, buddy)) {
575 /* only single bit in buddy2 may be 1 */
576 if (!mb_test_bit(i << 1, buddy2)) {
578 mb_test_bit((i<<1)+1, buddy2));
579 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
581 mb_test_bit(i << 1, buddy2));
586 /* both bits in buddy2 must be 1 */
587 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
588 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
590 for (j = 0; j < (1 << order); j++) {
591 k = (i * (1 << order)) + j;
593 !mb_test_bit(k, e4b->bd_bitmap));
597 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
602 buddy = mb_find_buddy(e4b, 0, &max);
603 for (i = 0; i < max; i++) {
604 if (!mb_test_bit(i, buddy)) {
605 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
613 /* check used bits only */
614 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
615 buddy2 = mb_find_buddy(e4b, j, &max2);
617 MB_CHECK_ASSERT(k < max2);
618 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
621 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
622 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
624 grp = ext4_get_group_info(sb, e4b->bd_group);
625 list_for_each(cur, &grp->bb_prealloc_list) {
626 ext4_group_t groupnr;
627 struct ext4_prealloc_space *pa;
628 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
629 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
630 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
631 for (i = 0; i < pa->pa_len; i++)
632 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
636 #undef MB_CHECK_ASSERT
637 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
638 __FILE__, __func__, __LINE__)
640 #define mb_check_buddy(e4b)
644 * Divide blocks started from @first with length @len into
645 * smaller chunks with power of 2 blocks.
646 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
647 * then increase bb_counters[] for corresponded chunk size.
649 static void ext4_mb_mark_free_simple(struct super_block *sb,
650 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
651 struct ext4_group_info *grp)
653 struct ext4_sb_info *sbi = EXT4_SB(sb);
657 unsigned short border;
659 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
661 border = 2 << sb->s_blocksize_bits;
664 /* find how many blocks can be covered since this position */
665 max = ffs(first | border) - 1;
667 /* find how many blocks of power 2 we need to mark */
674 /* mark multiblock chunks only */
675 grp->bb_counters[min]++;
677 mb_clear_bit(first >> min,
678 buddy + sbi->s_mb_offsets[min]);
686 * Cache the order of the largest free extent we have available in this block
690 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
695 grp->bb_largest_free_order = -1; /* uninit */
697 bits = sb->s_blocksize_bits + 1;
698 for (i = bits; i >= 0; i--) {
699 if (grp->bb_counters[i] > 0) {
700 grp->bb_largest_free_order = i;
706 static noinline_for_stack
707 void ext4_mb_generate_buddy(struct super_block *sb,
708 void *buddy, void *bitmap, ext4_group_t group)
710 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
711 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
716 unsigned fragments = 0;
717 unsigned long long period = get_cycles();
719 /* initialize buddy from bitmap which is aggregation
720 * of on-disk bitmap and preallocations */
721 i = mb_find_next_zero_bit(bitmap, max, 0);
722 grp->bb_first_free = i;
726 i = mb_find_next_bit(bitmap, max, i);
730 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
732 grp->bb_counters[0]++;
734 i = mb_find_next_zero_bit(bitmap, max, i);
736 grp->bb_fragments = fragments;
738 if (free != grp->bb_free) {
739 ext4_grp_locked_error(sb, group, 0, 0,
740 "%u clusters in bitmap, %u in gd",
743 * If we intent to continue, we consider group descritor
744 * corrupt and update bb_free using bitmap value
748 mb_set_largest_free_order(sb, grp);
750 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
752 period = get_cycles() - period;
753 spin_lock(&EXT4_SB(sb)->s_bal_lock);
754 EXT4_SB(sb)->s_mb_buddies_generated++;
755 EXT4_SB(sb)->s_mb_generation_time += period;
756 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
759 /* The buddy information is attached the buddy cache inode
760 * for convenience. The information regarding each group
761 * is loaded via ext4_mb_load_buddy. The information involve
762 * block bitmap and buddy information. The information are
763 * stored in the inode as
766 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
769 * one block each for bitmap and buddy information.
770 * So for each group we take up 2 blocks. A page can
771 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
772 * So it can have information regarding groups_per_page which
773 * is blocks_per_page/2
775 * Locking note: This routine takes the block group lock of all groups
776 * for this page; do not hold this lock when calling this routine!
779 static int ext4_mb_init_cache(struct page *page, char *incore)
781 ext4_group_t ngroups;
787 ext4_group_t first_group, group;
789 struct super_block *sb;
790 struct buffer_head *bhs;
791 struct buffer_head **bh = NULL;
795 struct ext4_group_info *grinfo;
797 mb_debug(1, "init page %lu\n", page->index);
799 inode = page->mapping->host;
801 ngroups = ext4_get_groups_count(sb);
802 blocksize = 1 << inode->i_blkbits;
803 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
805 groups_per_page = blocks_per_page >> 1;
806 if (groups_per_page == 0)
809 /* allocate buffer_heads to read bitmaps */
810 if (groups_per_page > 1) {
811 i = sizeof(struct buffer_head *) * groups_per_page;
812 bh = kzalloc(i, GFP_NOFS);
820 first_group = page->index * blocks_per_page / 2;
822 /* read all groups the page covers into the cache */
823 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
824 if (group >= ngroups)
827 grinfo = ext4_get_group_info(sb, group);
829 * If page is uptodate then we came here after online resize
830 * which added some new uninitialized group info structs, so
831 * we must skip all initialized uptodate buddies on the page,
832 * which may be currently in use by an allocating task.
834 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
838 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
842 mb_debug(1, "read bitmap for group %u\n", group);
845 /* wait for I/O completion */
846 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
847 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
853 first_block = page->index * blocks_per_page;
854 for (i = 0; i < blocks_per_page; i++) {
857 group = (first_block + i) >> 1;
858 if (group >= ngroups)
861 if (!bh[group - first_group])
862 /* skip initialized uptodate buddy */
866 * data carry information regarding this
867 * particular group in the format specified
871 data = page_address(page) + (i * blocksize);
872 bitmap = bh[group - first_group]->b_data;
875 * We place the buddy block and bitmap block
878 if ((first_block + i) & 1) {
879 /* this is block of buddy */
880 BUG_ON(incore == NULL);
881 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
882 group, page->index, i * blocksize);
883 trace_ext4_mb_buddy_bitmap_load(sb, group);
884 grinfo = ext4_get_group_info(sb, group);
885 grinfo->bb_fragments = 0;
886 memset(grinfo->bb_counters, 0,
887 sizeof(*grinfo->bb_counters) *
888 (sb->s_blocksize_bits+2));
890 * incore got set to the group block bitmap below
892 ext4_lock_group(sb, group);
894 memset(data, 0xff, blocksize);
895 ext4_mb_generate_buddy(sb, data, incore, group);
896 ext4_unlock_group(sb, group);
899 /* this is block of bitmap */
900 BUG_ON(incore != NULL);
901 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
902 group, page->index, i * blocksize);
903 trace_ext4_mb_bitmap_load(sb, group);
905 /* see comments in ext4_mb_put_pa() */
906 ext4_lock_group(sb, group);
907 memcpy(data, bitmap, blocksize);
909 /* mark all preallocated blks used in in-core bitmap */
910 ext4_mb_generate_from_pa(sb, data, group);
911 ext4_mb_generate_from_freelist(sb, data, group);
912 ext4_unlock_group(sb, group);
914 /* set incore so that the buddy information can be
915 * generated using this
920 SetPageUptodate(page);
924 for (i = 0; i < groups_per_page; i++)
933 * Lock the buddy and bitmap pages. This make sure other parallel init_group
934 * on the same buddy page doesn't happen whild holding the buddy page lock.
935 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
936 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
938 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
939 ext4_group_t group, struct ext4_buddy *e4b)
941 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
942 int block, pnum, poff;
946 e4b->bd_buddy_page = NULL;
947 e4b->bd_bitmap_page = NULL;
949 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
951 * the buddy cache inode stores the block bitmap
952 * and buddy information in consecutive blocks.
953 * So for each group we need two blocks.
956 pnum = block / blocks_per_page;
957 poff = block % blocks_per_page;
958 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
961 BUG_ON(page->mapping != inode->i_mapping);
962 e4b->bd_bitmap_page = page;
963 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
965 if (blocks_per_page >= 2) {
966 /* buddy and bitmap are on the same page */
971 pnum = block / blocks_per_page;
972 poff = 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 order, 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, order, &max);
1353 BUG_ON(buddy == NULL);
1354 BUG_ON(block >= max);
1355 if (mb_test_bit(block, buddy)) {
1362 /* FIXME dorp order completely ? */
1363 if (likely(order == 0)) {
1364 /* find actual order */
1365 order = mb_find_order_for_block(e4b, block);
1366 block = block >> order;
1369 ex->fe_len = 1 << order;
1370 ex->fe_start = block << order;
1371 ex->fe_group = e4b->bd_group;
1373 /* calc difference from given start */
1374 next = next - ex->fe_start;
1376 ex->fe_start += next;
1378 while (needed > ex->fe_len &&
1379 (buddy = mb_find_buddy(e4b, order, &max))) {
1381 if (block + 1 >= max)
1384 next = (block + 1) * (1 << order);
1385 if (mb_test_bit(next, e4b->bd_bitmap))
1388 order = mb_find_order_for_block(e4b, next);
1390 block = next >> order;
1391 ex->fe_len += 1 << order;
1394 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1398 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1404 int start = ex->fe_start;
1405 int len = ex->fe_len;
1410 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1411 BUG_ON(e4b->bd_group != ex->fe_group);
1412 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1413 mb_check_buddy(e4b);
1414 mb_mark_used_double(e4b, start, len);
1416 e4b->bd_info->bb_free -= len;
1417 if (e4b->bd_info->bb_first_free == start)
1418 e4b->bd_info->bb_first_free += len;
1420 /* let's maintain fragments counter */
1422 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1423 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1424 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1426 e4b->bd_info->bb_fragments++;
1427 else if (!mlen && !max)
1428 e4b->bd_info->bb_fragments--;
1430 /* let's maintain buddy itself */
1432 ord = mb_find_order_for_block(e4b, start);
1434 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1435 /* the whole chunk may be allocated at once! */
1437 buddy = mb_find_buddy(e4b, ord, &max);
1438 BUG_ON((start >> ord) >= max);
1439 mb_set_bit(start >> ord, buddy);
1440 e4b->bd_info->bb_counters[ord]--;
1447 /* store for history */
1449 ret = len | (ord << 16);
1451 /* we have to split large buddy */
1453 buddy = mb_find_buddy(e4b, ord, &max);
1454 mb_set_bit(start >> ord, buddy);
1455 e4b->bd_info->bb_counters[ord]--;
1458 cur = (start >> ord) & ~1U;
1459 buddy = mb_find_buddy(e4b, ord, &max);
1460 mb_clear_bit(cur, buddy);
1461 mb_clear_bit(cur + 1, buddy);
1462 e4b->bd_info->bb_counters[ord]++;
1463 e4b->bd_info->bb_counters[ord]++;
1465 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1467 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1468 mb_check_buddy(e4b);
1474 * Must be called under group lock!
1476 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1477 struct ext4_buddy *e4b)
1479 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1482 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1483 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1485 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1486 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1487 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1489 /* preallocation can change ac_b_ex, thus we store actually
1490 * allocated blocks for history */
1491 ac->ac_f_ex = ac->ac_b_ex;
1493 ac->ac_status = AC_STATUS_FOUND;
1494 ac->ac_tail = ret & 0xffff;
1495 ac->ac_buddy = ret >> 16;
1498 * take the page reference. We want the page to be pinned
1499 * so that we don't get a ext4_mb_init_cache_call for this
1500 * group until we update the bitmap. That would mean we
1501 * double allocate blocks. The reference is dropped
1502 * in ext4_mb_release_context
1504 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1505 get_page(ac->ac_bitmap_page);
1506 ac->ac_buddy_page = e4b->bd_buddy_page;
1507 get_page(ac->ac_buddy_page);
1508 /* store last allocated for subsequent stream allocation */
1509 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1510 spin_lock(&sbi->s_md_lock);
1511 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1512 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1513 spin_unlock(&sbi->s_md_lock);
1518 * regular allocator, for general purposes allocation
1521 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1522 struct ext4_buddy *e4b,
1525 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1526 struct ext4_free_extent *bex = &ac->ac_b_ex;
1527 struct ext4_free_extent *gex = &ac->ac_g_ex;
1528 struct ext4_free_extent ex;
1531 if (ac->ac_status == AC_STATUS_FOUND)
1534 * We don't want to scan for a whole year
1536 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1537 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1538 ac->ac_status = AC_STATUS_BREAK;
1543 * Haven't found good chunk so far, let's continue
1545 if (bex->fe_len < gex->fe_len)
1548 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1549 && bex->fe_group == e4b->bd_group) {
1550 /* recheck chunk's availability - we don't know
1551 * when it was found (within this lock-unlock
1553 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1554 if (max >= gex->fe_len) {
1555 ext4_mb_use_best_found(ac, e4b);
1562 * The routine checks whether found extent is good enough. If it is,
1563 * then the extent gets marked used and flag is set to the context
1564 * to stop scanning. Otherwise, the extent is compared with the
1565 * previous found extent and if new one is better, then it's stored
1566 * in the context. Later, the best found extent will be used, if
1567 * mballoc can't find good enough extent.
1569 * FIXME: real allocation policy is to be designed yet!
1571 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1572 struct ext4_free_extent *ex,
1573 struct ext4_buddy *e4b)
1575 struct ext4_free_extent *bex = &ac->ac_b_ex;
1576 struct ext4_free_extent *gex = &ac->ac_g_ex;
1578 BUG_ON(ex->fe_len <= 0);
1579 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1580 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1581 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1586 * The special case - take what you catch first
1588 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1590 ext4_mb_use_best_found(ac, e4b);
1595 * Let's check whether the chuck is good enough
1597 if (ex->fe_len == gex->fe_len) {
1599 ext4_mb_use_best_found(ac, e4b);
1604 * If this is first found extent, just store it in the context
1606 if (bex->fe_len == 0) {
1612 * If new found extent is better, store it in the context
1614 if (bex->fe_len < gex->fe_len) {
1615 /* if the request isn't satisfied, any found extent
1616 * larger than previous best one is better */
1617 if (ex->fe_len > bex->fe_len)
1619 } else if (ex->fe_len > gex->fe_len) {
1620 /* if the request is satisfied, then we try to find
1621 * an extent that still satisfy the request, but is
1622 * smaller than previous one */
1623 if (ex->fe_len < bex->fe_len)
1627 ext4_mb_check_limits(ac, e4b, 0);
1630 static noinline_for_stack
1631 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1632 struct ext4_buddy *e4b)
1634 struct ext4_free_extent ex = ac->ac_b_ex;
1635 ext4_group_t group = ex.fe_group;
1639 BUG_ON(ex.fe_len <= 0);
1640 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1644 ext4_lock_group(ac->ac_sb, group);
1645 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1649 ext4_mb_use_best_found(ac, e4b);
1652 ext4_unlock_group(ac->ac_sb, group);
1653 ext4_mb_unload_buddy(e4b);
1658 static noinline_for_stack
1659 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1660 struct ext4_buddy *e4b)
1662 ext4_group_t group = ac->ac_g_ex.fe_group;
1665 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1666 struct ext4_free_extent ex;
1668 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
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, 0, 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, 0, 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, 0, 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 /* We only do this if the grp has never been initialized */
1867 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1868 int ret = ext4_mb_init_group(ac->ac_sb, group);
1873 free = grp->bb_free;
1874 fragments = grp->bb_fragments;
1882 BUG_ON(ac->ac_2order == 0);
1884 if (grp->bb_largest_free_order < ac->ac_2order)
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))
1895 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1899 if (free >= ac->ac_g_ex.fe_len)
1911 static noinline_for_stack int
1912 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1914 ext4_group_t ngroups, group, i;
1917 struct ext4_sb_info *sbi;
1918 struct super_block *sb;
1919 struct ext4_buddy e4b;
1923 ngroups = ext4_get_groups_count(sb);
1924 /* non-extent files are limited to low blocks/groups */
1925 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
1926 ngroups = sbi->s_blockfile_groups;
1928 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1930 /* first, try the goal */
1931 err = ext4_mb_find_by_goal(ac, &e4b);
1932 if (err || ac->ac_status == AC_STATUS_FOUND)
1935 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1939 * ac->ac2_order is set only if the fe_len is a power of 2
1940 * if ac2_order is set we also set criteria to 0 so that we
1941 * try exact allocation using buddy.
1943 i = fls(ac->ac_g_ex.fe_len);
1946 * We search using buddy data only if the order of the request
1947 * is greater than equal to the sbi_s_mb_order2_reqs
1948 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1950 if (i >= sbi->s_mb_order2_reqs) {
1952 * This should tell if fe_len is exactly power of 2
1954 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1955 ac->ac_2order = i - 1;
1958 /* if stream allocation is enabled, use global goal */
1959 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1960 /* TBD: may be hot point */
1961 spin_lock(&sbi->s_md_lock);
1962 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1963 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1964 spin_unlock(&sbi->s_md_lock);
1967 /* Let's just scan groups to find more-less suitable blocks */
1968 cr = ac->ac_2order ? 0 : 1;
1970 * cr == 0 try to get exact allocation,
1971 * cr == 3 try to get anything
1974 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1975 ac->ac_criteria = cr;
1977 * searching for the right group start
1978 * from the goal value specified
1980 group = ac->ac_g_ex.fe_group;
1982 for (i = 0; i < ngroups; group++, i++) {
1983 if (group == ngroups)
1986 /* This now checks without needing the buddy page */
1987 if (!ext4_mb_good_group(ac, group, cr))
1990 err = ext4_mb_load_buddy(sb, group, &e4b);
1994 ext4_lock_group(sb, group);
1997 * We need to check again after locking the
2000 if (!ext4_mb_good_group(ac, group, cr)) {
2001 ext4_unlock_group(sb, group);
2002 ext4_mb_unload_buddy(&e4b);
2006 ac->ac_groups_scanned++;
2008 ext4_mb_simple_scan_group(ac, &e4b);
2009 else if (cr == 1 && sbi->s_stripe &&
2010 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2011 ext4_mb_scan_aligned(ac, &e4b);
2013 ext4_mb_complex_scan_group(ac, &e4b);
2015 ext4_unlock_group(sb, group);
2016 ext4_mb_unload_buddy(&e4b);
2018 if (ac->ac_status != AC_STATUS_CONTINUE)
2023 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2024 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2026 * We've been searching too long. Let's try to allocate
2027 * the best chunk we've found so far
2030 ext4_mb_try_best_found(ac, &e4b);
2031 if (ac->ac_status != AC_STATUS_FOUND) {
2033 * Someone more lucky has already allocated it.
2034 * The only thing we can do is just take first
2036 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2038 ac->ac_b_ex.fe_group = 0;
2039 ac->ac_b_ex.fe_start = 0;
2040 ac->ac_b_ex.fe_len = 0;
2041 ac->ac_status = AC_STATUS_CONTINUE;
2042 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2044 atomic_inc(&sbi->s_mb_lost_chunks);
2052 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2054 struct super_block *sb = seq->private;
2057 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2060 return (void *) ((unsigned long) group);
2063 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2065 struct super_block *sb = seq->private;
2069 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2072 return (void *) ((unsigned long) group);
2075 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2077 struct super_block *sb = seq->private;
2078 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2081 struct ext4_buddy e4b;
2083 struct ext4_group_info info;
2084 ext4_grpblk_t counters[16];
2089 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2090 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2091 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2092 "group", "free", "frags", "first",
2093 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2094 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2096 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2097 sizeof(struct ext4_group_info);
2098 err = ext4_mb_load_buddy(sb, group, &e4b);
2100 seq_printf(seq, "#%-5u: I/O error\n", group);
2103 ext4_lock_group(sb, group);
2104 memcpy(&sg, ext4_get_group_info(sb, group), i);
2105 ext4_unlock_group(sb, group);
2106 ext4_mb_unload_buddy(&e4b);
2108 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2109 sg.info.bb_fragments, sg.info.bb_first_free);
2110 for (i = 0; i <= 13; i++)
2111 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2112 sg.info.bb_counters[i] : 0);
2113 seq_printf(seq, " ]\n");
2118 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2122 static const struct seq_operations ext4_mb_seq_groups_ops = {
2123 .start = ext4_mb_seq_groups_start,
2124 .next = ext4_mb_seq_groups_next,
2125 .stop = ext4_mb_seq_groups_stop,
2126 .show = ext4_mb_seq_groups_show,
2129 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2131 struct super_block *sb = PDE(inode)->data;
2134 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2136 struct seq_file *m = file->private_data;
2143 static const struct file_operations ext4_mb_seq_groups_fops = {
2144 .owner = THIS_MODULE,
2145 .open = ext4_mb_seq_groups_open,
2147 .llseek = seq_lseek,
2148 .release = seq_release,
2151 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2153 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2154 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2160 /* Create and initialize ext4_group_info data for the given group. */
2161 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2162 struct ext4_group_desc *desc)
2166 struct ext4_sb_info *sbi = EXT4_SB(sb);
2167 struct ext4_group_info **meta_group_info;
2168 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2171 * First check if this group is the first of a reserved block.
2172 * If it's true, we have to allocate a new table of pointers
2173 * to ext4_group_info structures
2175 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2176 metalen = sizeof(*meta_group_info) <<
2177 EXT4_DESC_PER_BLOCK_BITS(sb);
2178 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2179 if (meta_group_info == NULL) {
2180 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2181 "for a buddy group");
2182 goto exit_meta_group_info;
2184 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2189 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2190 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2192 meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL);
2193 if (meta_group_info[i] == NULL) {
2194 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2195 goto exit_group_info;
2197 memset(meta_group_info[i], 0, kmem_cache_size(cachep));
2198 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2199 &(meta_group_info[i]->bb_state));
2202 * initialize bb_free to be able to skip
2203 * empty groups without initialization
2205 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2206 meta_group_info[i]->bb_free =
2207 ext4_free_clusters_after_init(sb, group, desc);
2209 meta_group_info[i]->bb_free =
2210 ext4_free_group_clusters(sb, desc);
2213 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2214 init_rwsem(&meta_group_info[i]->alloc_sem);
2215 meta_group_info[i]->bb_free_root = RB_ROOT;
2216 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2220 struct buffer_head *bh;
2221 meta_group_info[i]->bb_bitmap =
2222 kmalloc(sb->s_blocksize, GFP_KERNEL);
2223 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2224 bh = ext4_read_block_bitmap(sb, group);
2226 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2235 /* If a meta_group_info table has been allocated, release it now */
2236 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2237 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2238 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2240 exit_meta_group_info:
2242 } /* ext4_mb_add_groupinfo */
2244 static int ext4_mb_init_backend(struct super_block *sb)
2246 ext4_group_t ngroups = ext4_get_groups_count(sb);
2248 struct ext4_sb_info *sbi = EXT4_SB(sb);
2249 struct ext4_super_block *es = sbi->s_es;
2250 int num_meta_group_infos;
2251 int num_meta_group_infos_max;
2253 struct ext4_group_desc *desc;
2254 struct kmem_cache *cachep;
2256 /* This is the number of blocks used by GDT */
2257 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2258 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2261 * This is the total number of blocks used by GDT including
2262 * the number of reserved blocks for GDT.
2263 * The s_group_info array is allocated with this value
2264 * to allow a clean online resize without a complex
2265 * manipulation of pointer.
2266 * The drawback is the unused memory when no resize
2267 * occurs but it's very low in terms of pages
2268 * (see comments below)
2269 * Need to handle this properly when META_BG resizing is allowed
2271 num_meta_group_infos_max = num_meta_group_infos +
2272 le16_to_cpu(es->s_reserved_gdt_blocks);
2275 * array_size is the size of s_group_info array. We round it
2276 * to the next power of two because this approximation is done
2277 * internally by kmalloc so we can have some more memory
2278 * for free here (e.g. may be used for META_BG resize).
2281 while (array_size < sizeof(*sbi->s_group_info) *
2282 num_meta_group_infos_max)
2283 array_size = array_size << 1;
2284 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2285 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2286 * So a two level scheme suffices for now. */
2287 sbi->s_group_info = ext4_kvzalloc(array_size, GFP_KERNEL);
2288 if (sbi->s_group_info == NULL) {
2289 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2292 sbi->s_buddy_cache = new_inode(sb);
2293 if (sbi->s_buddy_cache == NULL) {
2294 ext4_msg(sb, KERN_ERR, "can't get new inode");
2297 /* To avoid potentially colliding with an valid on-disk inode number,
2298 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2299 * not in the inode hash, so it should never be found by iget(), but
2300 * this will avoid confusion if it ever shows up during debugging. */
2301 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2302 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2303 for (i = 0; i < ngroups; i++) {
2304 desc = ext4_get_group_desc(sb, i, NULL);
2306 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2309 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2316 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2318 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2319 i = num_meta_group_infos;
2321 kfree(sbi->s_group_info[i]);
2322 iput(sbi->s_buddy_cache);
2324 ext4_kvfree(sbi->s_group_info);
2328 static void ext4_groupinfo_destroy_slabs(void)
2332 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2333 if (ext4_groupinfo_caches[i])
2334 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2335 ext4_groupinfo_caches[i] = NULL;
2339 static int ext4_groupinfo_create_slab(size_t size)
2341 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2343 int blocksize_bits = order_base_2(size);
2344 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2345 struct kmem_cache *cachep;
2347 if (cache_index >= NR_GRPINFO_CACHES)
2350 if (unlikely(cache_index < 0))
2353 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2354 if (ext4_groupinfo_caches[cache_index]) {
2355 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2356 return 0; /* Already created */
2359 slab_size = offsetof(struct ext4_group_info,
2360 bb_counters[blocksize_bits + 2]);
2362 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2363 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2366 ext4_groupinfo_caches[cache_index] = cachep;
2368 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2371 "EXT4-fs: no memory for groupinfo slab cache\n");
2378 int ext4_mb_init(struct super_block *sb)
2380 struct ext4_sb_info *sbi = EXT4_SB(sb);
2386 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2388 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2389 if (sbi->s_mb_offsets == NULL) {
2394 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2395 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2396 if (sbi->s_mb_maxs == NULL) {
2401 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2405 /* order 0 is regular bitmap */
2406 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2407 sbi->s_mb_offsets[0] = 0;
2411 max = sb->s_blocksize << 2;
2413 sbi->s_mb_offsets[i] = offset;
2414 sbi->s_mb_maxs[i] = max;
2415 offset += 1 << (sb->s_blocksize_bits - i);
2418 } while (i <= sb->s_blocksize_bits + 1);
2420 spin_lock_init(&sbi->s_md_lock);
2421 spin_lock_init(&sbi->s_bal_lock);
2423 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2424 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2425 sbi->s_mb_stats = MB_DEFAULT_STATS;
2426 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2427 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2429 * The default group preallocation is 512, which for 4k block
2430 * sizes translates to 2 megabytes. However for bigalloc file
2431 * systems, this is probably too big (i.e, if the cluster size
2432 * is 1 megabyte, then group preallocation size becomes half a
2433 * gigabyte!). As a default, we will keep a two megabyte
2434 * group pralloc size for cluster sizes up to 64k, and after
2435 * that, we will force a minimum group preallocation size of
2436 * 32 clusters. This translates to 8 megs when the cluster
2437 * size is 256k, and 32 megs when the cluster size is 1 meg,
2438 * which seems reasonable as a default.
2440 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2441 sbi->s_cluster_bits, 32);
2443 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2444 * to the lowest multiple of s_stripe which is bigger than
2445 * the s_mb_group_prealloc as determined above. We want
2446 * the preallocation size to be an exact multiple of the
2447 * RAID stripe size so that preallocations don't fragment
2450 if (sbi->s_stripe > 1) {
2451 sbi->s_mb_group_prealloc = roundup(
2452 sbi->s_mb_group_prealloc, sbi->s_stripe);
2455 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2456 if (sbi->s_locality_groups == NULL) {
2458 goto out_free_groupinfo_slab;
2460 for_each_possible_cpu(i) {
2461 struct ext4_locality_group *lg;
2462 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2463 mutex_init(&lg->lg_mutex);
2464 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2465 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2466 spin_lock_init(&lg->lg_prealloc_lock);
2469 /* init file for buddy data */
2470 ret = ext4_mb_init_backend(sb);
2472 goto out_free_locality_groups;
2475 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2476 &ext4_mb_seq_groups_fops, sb);
2480 out_free_locality_groups:
2481 free_percpu(sbi->s_locality_groups);
2482 sbi->s_locality_groups = NULL;
2483 out_free_groupinfo_slab:
2484 ext4_groupinfo_destroy_slabs();
2486 kfree(sbi->s_mb_offsets);
2487 sbi->s_mb_offsets = NULL;
2488 kfree(sbi->s_mb_maxs);
2489 sbi->s_mb_maxs = NULL;
2493 /* need to called with the ext4 group lock held */
2494 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2496 struct ext4_prealloc_space *pa;
2497 struct list_head *cur, *tmp;
2500 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2501 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2502 list_del(&pa->pa_group_list);
2504 kmem_cache_free(ext4_pspace_cachep, pa);
2507 mb_debug(1, "mballoc: %u PAs left\n", count);
2511 int ext4_mb_release(struct super_block *sb)
2513 ext4_group_t ngroups = ext4_get_groups_count(sb);
2515 int num_meta_group_infos;
2516 struct ext4_group_info *grinfo;
2517 struct ext4_sb_info *sbi = EXT4_SB(sb);
2518 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2520 if (sbi->s_group_info) {
2521 for (i = 0; i < ngroups; i++) {
2522 grinfo = ext4_get_group_info(sb, i);
2524 kfree(grinfo->bb_bitmap);
2526 ext4_lock_group(sb, i);
2527 ext4_mb_cleanup_pa(grinfo);
2528 ext4_unlock_group(sb, i);
2529 kmem_cache_free(cachep, grinfo);
2531 num_meta_group_infos = (ngroups +
2532 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2533 EXT4_DESC_PER_BLOCK_BITS(sb);
2534 for (i = 0; i < num_meta_group_infos; i++)
2535 kfree(sbi->s_group_info[i]);
2536 ext4_kvfree(sbi->s_group_info);
2538 kfree(sbi->s_mb_offsets);
2539 kfree(sbi->s_mb_maxs);
2540 if (sbi->s_buddy_cache)
2541 iput(sbi->s_buddy_cache);
2542 if (sbi->s_mb_stats) {
2543 ext4_msg(sb, KERN_INFO,
2544 "mballoc: %u blocks %u reqs (%u success)",
2545 atomic_read(&sbi->s_bal_allocated),
2546 atomic_read(&sbi->s_bal_reqs),
2547 atomic_read(&sbi->s_bal_success));
2548 ext4_msg(sb, KERN_INFO,
2549 "mballoc: %u extents scanned, %u goal hits, "
2550 "%u 2^N hits, %u breaks, %u lost",
2551 atomic_read(&sbi->s_bal_ex_scanned),
2552 atomic_read(&sbi->s_bal_goals),
2553 atomic_read(&sbi->s_bal_2orders),
2554 atomic_read(&sbi->s_bal_breaks),
2555 atomic_read(&sbi->s_mb_lost_chunks));
2556 ext4_msg(sb, KERN_INFO,
2557 "mballoc: %lu generated and it took %Lu",
2558 sbi->s_mb_buddies_generated,
2559 sbi->s_mb_generation_time);
2560 ext4_msg(sb, KERN_INFO,
2561 "mballoc: %u preallocated, %u discarded",
2562 atomic_read(&sbi->s_mb_preallocated),
2563 atomic_read(&sbi->s_mb_discarded));
2566 free_percpu(sbi->s_locality_groups);
2568 remove_proc_entry("mb_groups", sbi->s_proc);
2573 static inline int ext4_issue_discard(struct super_block *sb,
2574 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2576 ext4_fsblk_t discard_block;
2578 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2579 ext4_group_first_block_no(sb, block_group));
2580 count = EXT4_C2B(EXT4_SB(sb), count);
2581 trace_ext4_discard_blocks(sb,
2582 (unsigned long long) discard_block, count);
2583 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2587 * This function is called by the jbd2 layer once the commit has finished,
2588 * so we know we can free the blocks that were released with that commit.
2590 static void ext4_free_data_callback(struct super_block *sb,
2591 struct ext4_journal_cb_entry *jce,
2594 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2595 struct ext4_buddy e4b;
2596 struct ext4_group_info *db;
2597 int err, count = 0, count2 = 0;
2599 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2600 entry->efd_count, entry->efd_group, entry);
2602 if (test_opt(sb, DISCARD))
2603 ext4_issue_discard(sb, entry->efd_group,
2604 entry->efd_start_cluster, entry->efd_count);
2606 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2607 /* we expect to find existing buddy because it's pinned */
2612 /* there are blocks to put in buddy to make them really free */
2613 count += entry->efd_count;
2615 ext4_lock_group(sb, entry->efd_group);
2616 /* Take it out of per group rb tree */
2617 rb_erase(&entry->efd_node, &(db->bb_free_root));
2618 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2621 * Clear the trimmed flag for the group so that the next
2622 * ext4_trim_fs can trim it.
2623 * If the volume is mounted with -o discard, online discard
2624 * is supported and the free blocks will be trimmed online.
2626 if (!test_opt(sb, DISCARD))
2627 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2629 if (!db->bb_free_root.rb_node) {
2630 /* No more items in the per group rb tree
2631 * balance refcounts from ext4_mb_free_metadata()
2633 page_cache_release(e4b.bd_buddy_page);
2634 page_cache_release(e4b.bd_bitmap_page);
2636 ext4_unlock_group(sb, entry->efd_group);
2637 kmem_cache_free(ext4_free_data_cachep, entry);
2638 ext4_mb_unload_buddy(&e4b);
2640 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2643 #ifdef CONFIG_EXT4_DEBUG
2644 u8 mb_enable_debug __read_mostly;
2646 static struct dentry *debugfs_dir;
2647 static struct dentry *debugfs_debug;
2649 static void __init ext4_create_debugfs_entry(void)
2651 debugfs_dir = debugfs_create_dir("ext4", NULL);
2653 debugfs_debug = debugfs_create_u8("mballoc-debug",
2659 static void ext4_remove_debugfs_entry(void)
2661 debugfs_remove(debugfs_debug);
2662 debugfs_remove(debugfs_dir);
2667 static void __init ext4_create_debugfs_entry(void)
2671 static void ext4_remove_debugfs_entry(void)
2677 int __init ext4_init_mballoc(void)
2679 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2680 SLAB_RECLAIM_ACCOUNT);
2681 if (ext4_pspace_cachep == NULL)
2684 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2685 SLAB_RECLAIM_ACCOUNT);
2686 if (ext4_ac_cachep == NULL) {
2687 kmem_cache_destroy(ext4_pspace_cachep);
2691 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2692 SLAB_RECLAIM_ACCOUNT);
2693 if (ext4_free_data_cachep == NULL) {
2694 kmem_cache_destroy(ext4_pspace_cachep);
2695 kmem_cache_destroy(ext4_ac_cachep);
2698 ext4_create_debugfs_entry();
2702 void ext4_exit_mballoc(void)
2705 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2706 * before destroying the slab cache.
2709 kmem_cache_destroy(ext4_pspace_cachep);
2710 kmem_cache_destroy(ext4_ac_cachep);
2711 kmem_cache_destroy(ext4_free_data_cachep);
2712 ext4_groupinfo_destroy_slabs();
2713 ext4_remove_debugfs_entry();
2718 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2719 * Returns 0 if success or error code
2721 static noinline_for_stack int
2722 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2723 handle_t *handle, unsigned int reserv_clstrs)
2725 struct buffer_head *bitmap_bh = NULL;
2726 struct ext4_group_desc *gdp;
2727 struct buffer_head *gdp_bh;
2728 struct ext4_sb_info *sbi;
2729 struct super_block *sb;
2733 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2734 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2740 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2744 err = ext4_journal_get_write_access(handle, bitmap_bh);
2749 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2753 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2754 ext4_free_group_clusters(sb, gdp));
2756 err = ext4_journal_get_write_access(handle, gdp_bh);
2760 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2762 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2763 if (!ext4_data_block_valid(sbi, block, len)) {
2764 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2765 "fs metadata", block, block+len);
2766 /* File system mounted not to panic on error
2767 * Fix the bitmap and repeat the block allocation
2768 * We leak some of the blocks here.
2770 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2771 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2772 ac->ac_b_ex.fe_len);
2773 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2774 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2780 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2781 #ifdef AGGRESSIVE_CHECK
2784 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2785 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2786 bitmap_bh->b_data));
2790 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2791 ac->ac_b_ex.fe_len);
2792 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2793 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2794 ext4_free_group_clusters_set(sb, gdp,
2795 ext4_free_clusters_after_init(sb,
2796 ac->ac_b_ex.fe_group, gdp));
2798 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2799 ext4_free_group_clusters_set(sb, gdp, len);
2800 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh,
2801 EXT4_BLOCKS_PER_GROUP(sb) / 8);
2802 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2804 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2805 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2807 * Now reduce the dirty block count also. Should not go negative
2809 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2810 /* release all the reserved blocks if non delalloc */
2811 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2814 if (sbi->s_log_groups_per_flex) {
2815 ext4_group_t flex_group = ext4_flex_group(sbi,
2816 ac->ac_b_ex.fe_group);
2817 atomic_sub(ac->ac_b_ex.fe_len,
2818 &sbi->s_flex_groups[flex_group].free_clusters);
2821 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2824 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2827 ext4_mark_super_dirty(sb);
2833 * here we normalize request for locality group
2834 * Group request are normalized to s_mb_group_prealloc, which goes to
2835 * s_strip if we set the same via mount option.
2836 * s_mb_group_prealloc can be configured via
2837 * /sys/fs/ext4/<partition>/mb_group_prealloc
2839 * XXX: should we try to preallocate more than the group has now?
2841 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2843 struct super_block *sb = ac->ac_sb;
2844 struct ext4_locality_group *lg = ac->ac_lg;
2847 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2848 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2849 current->pid, ac->ac_g_ex.fe_len);
2853 * Normalization means making request better in terms of
2854 * size and alignment
2856 static noinline_for_stack void
2857 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2858 struct ext4_allocation_request *ar)
2860 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2863 loff_t size, start_off;
2864 loff_t orig_size __maybe_unused;
2866 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2867 struct ext4_prealloc_space *pa;
2869 /* do normalize only data requests, metadata requests
2870 do not need preallocation */
2871 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2874 /* sometime caller may want exact blocks */
2875 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2878 /* caller may indicate that preallocation isn't
2879 * required (it's a tail, for example) */
2880 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2883 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2884 ext4_mb_normalize_group_request(ac);
2888 bsbits = ac->ac_sb->s_blocksize_bits;
2890 /* first, let's learn actual file size
2891 * given current request is allocated */
2892 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
2893 size = size << bsbits;
2894 if (size < i_size_read(ac->ac_inode))
2895 size = i_size_read(ac->ac_inode);
2898 /* max size of free chunks */
2901 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2902 (req <= (size) || max <= (chunk_size))
2904 /* first, try to predict filesize */
2905 /* XXX: should this table be tunable? */
2907 if (size <= 16 * 1024) {
2909 } else if (size <= 32 * 1024) {
2911 } else if (size <= 64 * 1024) {
2913 } else if (size <= 128 * 1024) {
2915 } else if (size <= 256 * 1024) {
2917 } else if (size <= 512 * 1024) {
2919 } else if (size <= 1024 * 1024) {
2921 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2922 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2923 (21 - bsbits)) << 21;
2924 size = 2 * 1024 * 1024;
2925 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2926 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2927 (22 - bsbits)) << 22;
2928 size = 4 * 1024 * 1024;
2929 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2930 (8<<20)>>bsbits, max, 8 * 1024)) {
2931 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2932 (23 - bsbits)) << 23;
2933 size = 8 * 1024 * 1024;
2935 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2936 size = ac->ac_o_ex.fe_len << bsbits;
2938 size = size >> bsbits;
2939 start = start_off >> bsbits;
2941 /* don't cover already allocated blocks in selected range */
2942 if (ar->pleft && start <= ar->lleft) {
2943 size -= ar->lleft + 1 - start;
2944 start = ar->lleft + 1;
2946 if (ar->pright && start + size - 1 >= ar->lright)
2947 size -= start + size - ar->lright;
2951 /* check we don't cross already preallocated blocks */
2953 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2958 spin_lock(&pa->pa_lock);
2959 if (pa->pa_deleted) {
2960 spin_unlock(&pa->pa_lock);
2964 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
2967 /* PA must not overlap original request */
2968 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2969 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2971 /* skip PAs this normalized request doesn't overlap with */
2972 if (pa->pa_lstart >= end || pa_end <= start) {
2973 spin_unlock(&pa->pa_lock);
2976 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2978 /* adjust start or end to be adjacent to this pa */
2979 if (pa_end <= ac->ac_o_ex.fe_logical) {
2980 BUG_ON(pa_end < start);
2982 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
2983 BUG_ON(pa->pa_lstart > end);
2984 end = pa->pa_lstart;
2986 spin_unlock(&pa->pa_lock);
2991 /* XXX: extra loop to check we really don't overlap preallocations */
2993 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2996 spin_lock(&pa->pa_lock);
2997 if (pa->pa_deleted == 0) {
2998 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3000 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3002 spin_unlock(&pa->pa_lock);
3006 if (start + size <= ac->ac_o_ex.fe_logical &&
3007 start > ac->ac_o_ex.fe_logical) {
3008 ext4_msg(ac->ac_sb, KERN_ERR,
3009 "start %lu, size %lu, fe_logical %lu",
3010 (unsigned long) start, (unsigned long) size,
3011 (unsigned long) ac->ac_o_ex.fe_logical);
3013 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3014 start > ac->ac_o_ex.fe_logical);
3015 BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
3017 /* now prepare goal request */
3019 /* XXX: is it better to align blocks WRT to logical
3020 * placement or satisfy big request as is */
3021 ac->ac_g_ex.fe_logical = start;
3022 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3024 /* define goal start in order to merge */
3025 if (ar->pright && (ar->lright == (start + size))) {
3026 /* merge to the right */
3027 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3028 &ac->ac_f_ex.fe_group,
3029 &ac->ac_f_ex.fe_start);
3030 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3032 if (ar->pleft && (ar->lleft + 1 == start)) {
3033 /* merge to the left */
3034 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3035 &ac->ac_f_ex.fe_group,
3036 &ac->ac_f_ex.fe_start);
3037 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3040 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3041 (unsigned) orig_size, (unsigned) start);
3044 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3046 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3048 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3049 atomic_inc(&sbi->s_bal_reqs);
3050 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3051 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3052 atomic_inc(&sbi->s_bal_success);
3053 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3054 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3055 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3056 atomic_inc(&sbi->s_bal_goals);
3057 if (ac->ac_found > sbi->s_mb_max_to_scan)
3058 atomic_inc(&sbi->s_bal_breaks);
3061 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3062 trace_ext4_mballoc_alloc(ac);
3064 trace_ext4_mballoc_prealloc(ac);
3068 * Called on failure; free up any blocks from the inode PA for this
3069 * context. We don't need this for MB_GROUP_PA because we only change
3070 * pa_free in ext4_mb_release_context(), but on failure, we've already
3071 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3073 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3075 struct ext4_prealloc_space *pa = ac->ac_pa;
3077 if (pa && pa->pa_type == MB_INODE_PA)
3078 pa->pa_free += ac->ac_b_ex.fe_len;
3082 * use blocks preallocated to inode
3084 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3085 struct ext4_prealloc_space *pa)
3087 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3092 /* found preallocated blocks, use them */
3093 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3094 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3095 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3096 len = EXT4_NUM_B2C(sbi, end - start);
3097 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3098 &ac->ac_b_ex.fe_start);
3099 ac->ac_b_ex.fe_len = len;
3100 ac->ac_status = AC_STATUS_FOUND;
3103 BUG_ON(start < pa->pa_pstart);
3104 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3105 BUG_ON(pa->pa_free < len);
3108 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3112 * use blocks preallocated to locality group
3114 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3115 struct ext4_prealloc_space *pa)
3117 unsigned int len = ac->ac_o_ex.fe_len;
3119 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3120 &ac->ac_b_ex.fe_group,
3121 &ac->ac_b_ex.fe_start);
3122 ac->ac_b_ex.fe_len = len;
3123 ac->ac_status = AC_STATUS_FOUND;
3126 /* we don't correct pa_pstart or pa_plen here to avoid
3127 * possible race when the group is being loaded concurrently
3128 * instead we correct pa later, after blocks are marked
3129 * in on-disk bitmap -- see ext4_mb_release_context()
3130 * Other CPUs are prevented from allocating from this pa by lg_mutex
3132 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3136 * Return the prealloc space that have minimal distance
3137 * from the goal block. @cpa is the prealloc
3138 * space that is having currently known minimal distance
3139 * from the goal block.
3141 static struct ext4_prealloc_space *
3142 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3143 struct ext4_prealloc_space *pa,
3144 struct ext4_prealloc_space *cpa)
3146 ext4_fsblk_t cur_distance, new_distance;
3149 atomic_inc(&pa->pa_count);
3152 cur_distance = abs(goal_block - cpa->pa_pstart);
3153 new_distance = abs(goal_block - pa->pa_pstart);
3155 if (cur_distance <= new_distance)
3158 /* drop the previous reference */
3159 atomic_dec(&cpa->pa_count);
3160 atomic_inc(&pa->pa_count);
3165 * search goal blocks in preallocated space
3167 static noinline_for_stack int
3168 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3170 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3172 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3173 struct ext4_locality_group *lg;
3174 struct ext4_prealloc_space *pa, *cpa = NULL;
3175 ext4_fsblk_t goal_block;
3177 /* only data can be preallocated */
3178 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3181 /* first, try per-file preallocation */
3183 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3185 /* all fields in this condition don't change,
3186 * so we can skip locking for them */
3187 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3188 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3189 EXT4_C2B(sbi, pa->pa_len)))
3192 /* non-extent files can't have physical blocks past 2^32 */
3193 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3194 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3195 EXT4_MAX_BLOCK_FILE_PHYS))
3198 /* found preallocated blocks, use them */
3199 spin_lock(&pa->pa_lock);
3200 if (pa->pa_deleted == 0 && pa->pa_free) {
3201 atomic_inc(&pa->pa_count);
3202 ext4_mb_use_inode_pa(ac, pa);
3203 spin_unlock(&pa->pa_lock);
3204 ac->ac_criteria = 10;
3208 spin_unlock(&pa->pa_lock);
3212 /* can we use group allocation? */
3213 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3216 /* inode may have no locality group for some reason */
3220 order = fls(ac->ac_o_ex.fe_len) - 1;
3221 if (order > PREALLOC_TB_SIZE - 1)
3222 /* The max size of hash table is PREALLOC_TB_SIZE */
3223 order = PREALLOC_TB_SIZE - 1;
3225 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3227 * search for the prealloc space that is having
3228 * minimal distance from the goal block.
3230 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3232 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3234 spin_lock(&pa->pa_lock);
3235 if (pa->pa_deleted == 0 &&
3236 pa->pa_free >= ac->ac_o_ex.fe_len) {
3238 cpa = ext4_mb_check_group_pa(goal_block,
3241 spin_unlock(&pa->pa_lock);
3246 ext4_mb_use_group_pa(ac, cpa);
3247 ac->ac_criteria = 20;
3254 * the function goes through all block freed in the group
3255 * but not yet committed and marks them used in in-core bitmap.
3256 * buddy must be generated from this bitmap
3257 * Need to be called with the ext4 group lock held
3259 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3263 struct ext4_group_info *grp;
3264 struct ext4_free_data *entry;
3266 grp = ext4_get_group_info(sb, group);
3267 n = rb_first(&(grp->bb_free_root));
3270 entry = rb_entry(n, struct ext4_free_data, efd_node);
3271 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3278 * the function goes through all preallocation in this group and marks them
3279 * used in in-core bitmap. buddy must be generated from this bitmap
3280 * Need to be called with ext4 group lock held
3282 static noinline_for_stack
3283 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3286 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3287 struct ext4_prealloc_space *pa;
3288 struct list_head *cur;
3289 ext4_group_t groupnr;
3290 ext4_grpblk_t start;
3291 int preallocated = 0;
3294 /* all form of preallocation discards first load group,
3295 * so the only competing code is preallocation use.
3296 * we don't need any locking here
3297 * notice we do NOT ignore preallocations with pa_deleted
3298 * otherwise we could leave used blocks available for
3299 * allocation in buddy when concurrent ext4_mb_put_pa()
3300 * is dropping preallocation
3302 list_for_each(cur, &grp->bb_prealloc_list) {
3303 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3304 spin_lock(&pa->pa_lock);
3305 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3308 spin_unlock(&pa->pa_lock);
3309 if (unlikely(len == 0))
3311 BUG_ON(groupnr != group);
3312 ext4_set_bits(bitmap, start, len);
3313 preallocated += len;
3315 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3318 static void ext4_mb_pa_callback(struct rcu_head *head)
3320 struct ext4_prealloc_space *pa;
3321 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3322 kmem_cache_free(ext4_pspace_cachep, pa);
3326 * drops a reference to preallocated space descriptor
3327 * if this was the last reference and the space is consumed
3329 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3330 struct super_block *sb, struct ext4_prealloc_space *pa)
3333 ext4_fsblk_t grp_blk;
3335 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3338 /* in this short window concurrent discard can set pa_deleted */
3339 spin_lock(&pa->pa_lock);
3340 if (pa->pa_deleted == 1) {
3341 spin_unlock(&pa->pa_lock);
3346 spin_unlock(&pa->pa_lock);
3348 grp_blk = pa->pa_pstart;
3350 * If doing group-based preallocation, pa_pstart may be in the
3351 * next group when pa is used up
3353 if (pa->pa_type == MB_GROUP_PA)
3356 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3361 * P1 (buddy init) P2 (regular allocation)
3362 * find block B in PA
3363 * copy on-disk bitmap to buddy
3364 * mark B in on-disk bitmap
3365 * drop PA from group
3366 * mark all PAs in buddy
3368 * thus, P1 initializes buddy with B available. to prevent this
3369 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3372 ext4_lock_group(sb, grp);
3373 list_del(&pa->pa_group_list);
3374 ext4_unlock_group(sb, grp);
3376 spin_lock(pa->pa_obj_lock);
3377 list_del_rcu(&pa->pa_inode_list);
3378 spin_unlock(pa->pa_obj_lock);
3380 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3384 * creates new preallocated space for given inode
3386 static noinline_for_stack int
3387 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3389 struct super_block *sb = ac->ac_sb;
3390 struct ext4_sb_info *sbi = EXT4_SB(sb);
3391 struct ext4_prealloc_space *pa;
3392 struct ext4_group_info *grp;
3393 struct ext4_inode_info *ei;
3395 /* preallocate only when found space is larger then requested */
3396 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3397 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3398 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3400 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3404 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3410 /* we can't allocate as much as normalizer wants.
3411 * so, found space must get proper lstart
3412 * to cover original request */
3413 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3414 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3416 /* we're limited by original request in that
3417 * logical block must be covered any way
3418 * winl is window we can move our chunk within */
3419 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3421 /* also, we should cover whole original request */
3422 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3424 /* the smallest one defines real window */
3425 win = min(winl, wins);
3427 offs = ac->ac_o_ex.fe_logical %
3428 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3429 if (offs && offs < win)
3432 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3434 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3435 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3438 /* preallocation can change ac_b_ex, thus we store actually
3439 * allocated blocks for history */
3440 ac->ac_f_ex = ac->ac_b_ex;
3442 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3443 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3444 pa->pa_len = ac->ac_b_ex.fe_len;
3445 pa->pa_free = pa->pa_len;
3446 atomic_set(&pa->pa_count, 1);
3447 spin_lock_init(&pa->pa_lock);
3448 INIT_LIST_HEAD(&pa->pa_inode_list);
3449 INIT_LIST_HEAD(&pa->pa_group_list);
3451 pa->pa_type = MB_INODE_PA;
3453 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3454 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3455 trace_ext4_mb_new_inode_pa(ac, pa);
3457 ext4_mb_use_inode_pa(ac, pa);
3458 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3460 ei = EXT4_I(ac->ac_inode);
3461 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3463 pa->pa_obj_lock = &ei->i_prealloc_lock;
3464 pa->pa_inode = ac->ac_inode;
3466 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3467 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3468 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3470 spin_lock(pa->pa_obj_lock);
3471 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3472 spin_unlock(pa->pa_obj_lock);
3478 * creates new preallocated space for locality group inodes belongs to
3480 static noinline_for_stack int
3481 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3483 struct super_block *sb = ac->ac_sb;
3484 struct ext4_locality_group *lg;
3485 struct ext4_prealloc_space *pa;
3486 struct ext4_group_info *grp;
3488 /* preallocate only when found space is larger then requested */
3489 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3490 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3491 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3493 BUG_ON(ext4_pspace_cachep == NULL);
3494 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3498 /* preallocation can change ac_b_ex, thus we store actually
3499 * allocated blocks for history */
3500 ac->ac_f_ex = ac->ac_b_ex;
3502 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3503 pa->pa_lstart = pa->pa_pstart;
3504 pa->pa_len = ac->ac_b_ex.fe_len;
3505 pa->pa_free = pa->pa_len;
3506 atomic_set(&pa->pa_count, 1);
3507 spin_lock_init(&pa->pa_lock);
3508 INIT_LIST_HEAD(&pa->pa_inode_list);
3509 INIT_LIST_HEAD(&pa->pa_group_list);
3511 pa->pa_type = MB_GROUP_PA;
3513 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3514 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3515 trace_ext4_mb_new_group_pa(ac, pa);
3517 ext4_mb_use_group_pa(ac, pa);
3518 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3520 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3524 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3525 pa->pa_inode = NULL;
3527 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3528 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3529 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3532 * We will later add the new pa to the right bucket
3533 * after updating the pa_free in ext4_mb_release_context
3538 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3542 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3543 err = ext4_mb_new_group_pa(ac);
3545 err = ext4_mb_new_inode_pa(ac);
3550 * finds all unused blocks in on-disk bitmap, frees them in
3551 * in-core bitmap and buddy.
3552 * @pa must be unlinked from inode and group lists, so that
3553 * nobody else can find/use it.
3554 * the caller MUST hold group/inode locks.
3555 * TODO: optimize the case when there are no in-core structures yet
3557 static noinline_for_stack int
3558 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3559 struct ext4_prealloc_space *pa)
3561 struct super_block *sb = e4b->bd_sb;
3562 struct ext4_sb_info *sbi = EXT4_SB(sb);
3567 unsigned long long grp_blk_start;
3571 BUG_ON(pa->pa_deleted == 0);
3572 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3573 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3574 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3575 end = bit + pa->pa_len;
3578 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3581 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3582 mb_debug(1, " free preallocated %u/%u in group %u\n",
3583 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3584 (unsigned) next - bit, (unsigned) group);
3587 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3588 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3589 EXT4_C2B(sbi, bit)),
3591 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3594 if (free != pa->pa_free) {
3595 ext4_msg(e4b->bd_sb, KERN_CRIT,
3596 "pa %p: logic %lu, phys. %lu, len %lu",
3597 pa, (unsigned long) pa->pa_lstart,
3598 (unsigned long) pa->pa_pstart,
3599 (unsigned long) pa->pa_len);
3600 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3603 * pa is already deleted so we use the value obtained
3604 * from the bitmap and continue.
3607 atomic_add(free, &sbi->s_mb_discarded);
3612 static noinline_for_stack int
3613 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3614 struct ext4_prealloc_space *pa)
3616 struct super_block *sb = e4b->bd_sb;
3620 trace_ext4_mb_release_group_pa(sb, pa);
3621 BUG_ON(pa->pa_deleted == 0);
3622 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3623 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3624 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3625 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3626 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3632 * releases all preallocations in given group
3634 * first, we need to decide discard policy:
3635 * - when do we discard
3637 * - how many do we discard
3638 * 1) how many requested
3640 static noinline_for_stack int
3641 ext4_mb_discard_group_preallocations(struct super_block *sb,
3642 ext4_group_t group, int needed)
3644 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3645 struct buffer_head *bitmap_bh = NULL;
3646 struct ext4_prealloc_space *pa, *tmp;
3647 struct list_head list;
3648 struct ext4_buddy e4b;
3653 mb_debug(1, "discard preallocation for group %u\n", group);
3655 if (list_empty(&grp->bb_prealloc_list))
3658 bitmap_bh = ext4_read_block_bitmap(sb, group);
3659 if (bitmap_bh == NULL) {
3660 ext4_error(sb, "Error reading block bitmap for %u", group);
3664 err = ext4_mb_load_buddy(sb, group, &e4b);
3666 ext4_error(sb, "Error loading buddy information for %u", group);
3672 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3674 INIT_LIST_HEAD(&list);
3676 ext4_lock_group(sb, group);
3677 list_for_each_entry_safe(pa, tmp,
3678 &grp->bb_prealloc_list, pa_group_list) {
3679 spin_lock(&pa->pa_lock);
3680 if (atomic_read(&pa->pa_count)) {
3681 spin_unlock(&pa->pa_lock);
3685 if (pa->pa_deleted) {
3686 spin_unlock(&pa->pa_lock);
3690 /* seems this one can be freed ... */
3693 /* we can trust pa_free ... */
3694 free += pa->pa_free;
3696 spin_unlock(&pa->pa_lock);
3698 list_del(&pa->pa_group_list);
3699 list_add(&pa->u.pa_tmp_list, &list);
3702 /* if we still need more blocks and some PAs were used, try again */
3703 if (free < needed && busy) {
3705 ext4_unlock_group(sb, group);
3707 * Yield the CPU here so that we don't get soft lockup
3708 * in non preempt case.
3714 /* found anything to free? */
3715 if (list_empty(&list)) {
3720 /* now free all selected PAs */
3721 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3723 /* remove from object (inode or locality group) */
3724 spin_lock(pa->pa_obj_lock);
3725 list_del_rcu(&pa->pa_inode_list);
3726 spin_unlock(pa->pa_obj_lock);
3728 if (pa->pa_type == MB_GROUP_PA)
3729 ext4_mb_release_group_pa(&e4b, pa);
3731 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3733 list_del(&pa->u.pa_tmp_list);
3734 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3738 ext4_unlock_group(sb, group);
3739 ext4_mb_unload_buddy(&e4b);
3745 * releases all non-used preallocated blocks for given inode
3747 * It's important to discard preallocations under i_data_sem
3748 * We don't want another block to be served from the prealloc
3749 * space when we are discarding the inode prealloc space.
3751 * FIXME!! Make sure it is valid at all the call sites
3753 void ext4_discard_preallocations(struct inode *inode)
3755 struct ext4_inode_info *ei = EXT4_I(inode);
3756 struct super_block *sb = inode->i_sb;
3757 struct buffer_head *bitmap_bh = NULL;
3758 struct ext4_prealloc_space *pa, *tmp;
3759 ext4_group_t group = 0;
3760 struct list_head list;
3761 struct ext4_buddy e4b;
3764 if (!S_ISREG(inode->i_mode)) {
3765 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3769 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3770 trace_ext4_discard_preallocations(inode);
3772 INIT_LIST_HEAD(&list);
3775 /* first, collect all pa's in the inode */
3776 spin_lock(&ei->i_prealloc_lock);
3777 while (!list_empty(&ei->i_prealloc_list)) {
3778 pa = list_entry(ei->i_prealloc_list.next,
3779 struct ext4_prealloc_space, pa_inode_list);
3780 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3781 spin_lock(&pa->pa_lock);
3782 if (atomic_read(&pa->pa_count)) {
3783 /* this shouldn't happen often - nobody should
3784 * use preallocation while we're discarding it */
3785 spin_unlock(&pa->pa_lock);
3786 spin_unlock(&ei->i_prealloc_lock);
3787 ext4_msg(sb, KERN_ERR,
3788 "uh-oh! used pa while discarding");
3790 schedule_timeout_uninterruptible(HZ);
3794 if (pa->pa_deleted == 0) {
3796 spin_unlock(&pa->pa_lock);
3797 list_del_rcu(&pa->pa_inode_list);
3798 list_add(&pa->u.pa_tmp_list, &list);
3802 /* someone is deleting pa right now */
3803 spin_unlock(&pa->pa_lock);
3804 spin_unlock(&ei->i_prealloc_lock);
3806 /* we have to wait here because pa_deleted
3807 * doesn't mean pa is already unlinked from
3808 * the list. as we might be called from
3809 * ->clear_inode() the inode will get freed
3810 * and concurrent thread which is unlinking
3811 * pa from inode's list may access already
3812 * freed memory, bad-bad-bad */
3814 /* XXX: if this happens too often, we can
3815 * add a flag to force wait only in case
3816 * of ->clear_inode(), but not in case of
3817 * regular truncate */
3818 schedule_timeout_uninterruptible(HZ);
3821 spin_unlock(&ei->i_prealloc_lock);
3823 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3824 BUG_ON(pa->pa_type != MB_INODE_PA);
3825 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3827 err = ext4_mb_load_buddy(sb, group, &e4b);
3829 ext4_error(sb, "Error loading buddy information for %u",
3834 bitmap_bh = ext4_read_block_bitmap(sb, group);
3835 if (bitmap_bh == NULL) {
3836 ext4_error(sb, "Error reading block bitmap for %u",
3838 ext4_mb_unload_buddy(&e4b);
3842 ext4_lock_group(sb, group);
3843 list_del(&pa->pa_group_list);
3844 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3845 ext4_unlock_group(sb, group);
3847 ext4_mb_unload_buddy(&e4b);
3850 list_del(&pa->u.pa_tmp_list);
3851 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3855 #ifdef CONFIG_EXT4_DEBUG
3856 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3858 struct super_block *sb = ac->ac_sb;
3859 ext4_group_t ngroups, i;
3861 if (!mb_enable_debug ||
3862 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3865 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
3866 " Allocation context details:");
3867 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
3868 ac->ac_status, ac->ac_flags);
3869 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
3870 "goal %lu/%lu/%lu@%lu, "
3871 "best %lu/%lu/%lu@%lu cr %d",
3872 (unsigned long)ac->ac_o_ex.fe_group,
3873 (unsigned long)ac->ac_o_ex.fe_start,
3874 (unsigned long)ac->ac_o_ex.fe_len,
3875 (unsigned long)ac->ac_o_ex.fe_logical,
3876 (unsigned long)ac->ac_g_ex.fe_group,
3877 (unsigned long)ac->ac_g_ex.fe_start,
3878 (unsigned long)ac->ac_g_ex.fe_len,
3879 (unsigned long)ac->ac_g_ex.fe_logical,
3880 (unsigned long)ac->ac_b_ex.fe_group,
3881 (unsigned long)ac->ac_b_ex.fe_start,
3882 (unsigned long)ac->ac_b_ex.fe_len,
3883 (unsigned long)ac->ac_b_ex.fe_logical,
3884 (int)ac->ac_criteria);
3885 ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found",
3886 ac->ac_ex_scanned, ac->ac_found);
3887 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
3888 ngroups = ext4_get_groups_count(sb);
3889 for (i = 0; i < ngroups; i++) {
3890 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3891 struct ext4_prealloc_space *pa;
3892 ext4_grpblk_t start;
3893 struct list_head *cur;
3894 ext4_lock_group(sb, i);
3895 list_for_each(cur, &grp->bb_prealloc_list) {
3896 pa = list_entry(cur, struct ext4_prealloc_space,
3898 spin_lock(&pa->pa_lock);
3899 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3901 spin_unlock(&pa->pa_lock);
3902 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3905 ext4_unlock_group(sb, i);
3907 if (grp->bb_free == 0)
3909 printk(KERN_ERR "%u: %d/%d \n",
3910 i, grp->bb_free, grp->bb_fragments);
3912 printk(KERN_ERR "\n");
3915 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3922 * We use locality group preallocation for small size file. The size of the
3923 * file is determined by the current size or the resulting size after
3924 * allocation which ever is larger
3926 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3928 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3930 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3931 int bsbits = ac->ac_sb->s_blocksize_bits;
3934 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3937 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3940 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3941 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3944 if ((size == isize) &&
3945 !ext4_fs_is_busy(sbi) &&
3946 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3947 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3951 if (sbi->s_mb_group_prealloc <= 0) {
3952 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3956 /* don't use group allocation for large files */
3957 size = max(size, isize);
3958 if (size > sbi->s_mb_stream_request) {
3959 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3963 BUG_ON(ac->ac_lg != NULL);
3965 * locality group prealloc space are per cpu. The reason for having
3966 * per cpu locality group is to reduce the contention between block
3967 * request from multiple CPUs.
3969 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
3971 /* we're going to use group allocation */
3972 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3974 /* serialize all allocations in the group */
3975 mutex_lock(&ac->ac_lg->lg_mutex);
3978 static noinline_for_stack int
3979 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
3980 struct ext4_allocation_request *ar)
3982 struct super_block *sb = ar->inode->i_sb;
3983 struct ext4_sb_info *sbi = EXT4_SB(sb);
3984 struct ext4_super_block *es = sbi->s_es;
3988 ext4_grpblk_t block;
3990 /* we can't allocate > group size */
3993 /* just a dirty hack to filter too big requests */
3994 if (len >= EXT4_CLUSTERS_PER_GROUP(sb) - 10)
3995 len = EXT4_CLUSTERS_PER_GROUP(sb) - 10;
3997 /* start searching from the goal */
3999 if (goal < le32_to_cpu(es->s_first_data_block) ||
4000 goal >= ext4_blocks_count(es))
4001 goal = le32_to_cpu(es->s_first_data_block);
4002 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4004 /* set up allocation goals */
4005 memset(ac, 0, sizeof(struct ext4_allocation_context));
4006 ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1);
4007 ac->ac_status = AC_STATUS_CONTINUE;
4009 ac->ac_inode = ar->inode;
4010 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4011 ac->ac_o_ex.fe_group = group;
4012 ac->ac_o_ex.fe_start = block;
4013 ac->ac_o_ex.fe_len = len;
4014 ac->ac_g_ex = ac->ac_o_ex;
4015 ac->ac_flags = ar->flags;
4017 /* we have to define context: we'll we work with a file or
4018 * locality group. this is a policy, actually */
4019 ext4_mb_group_or_file(ac);
4021 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4022 "left: %u/%u, right %u/%u to %swritable\n",
4023 (unsigned) ar->len, (unsigned) ar->logical,
4024 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4025 (unsigned) ar->lleft, (unsigned) ar->pleft,
4026 (unsigned) ar->lright, (unsigned) ar->pright,
4027 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4032 static noinline_for_stack void
4033 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4034 struct ext4_locality_group *lg,
4035 int order, int total_entries)
4037 ext4_group_t group = 0;
4038 struct ext4_buddy e4b;
4039 struct list_head discard_list;
4040 struct ext4_prealloc_space *pa, *tmp;
4042 mb_debug(1, "discard locality group preallocation\n");
4044 INIT_LIST_HEAD(&discard_list);
4046 spin_lock(&lg->lg_prealloc_lock);
4047 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4049 spin_lock(&pa->pa_lock);
4050 if (atomic_read(&pa->pa_count)) {
4052 * This is the pa that we just used
4053 * for block allocation. So don't
4056 spin_unlock(&pa->pa_lock);
4059 if (pa->pa_deleted) {
4060 spin_unlock(&pa->pa_lock);
4063 /* only lg prealloc space */
4064 BUG_ON(pa->pa_type != MB_GROUP_PA);
4066 /* seems this one can be freed ... */
4068 spin_unlock(&pa->pa_lock);
4070 list_del_rcu(&pa->pa_inode_list);
4071 list_add(&pa->u.pa_tmp_list, &discard_list);
4074 if (total_entries <= 5) {
4076 * we want to keep only 5 entries
4077 * allowing it to grow to 8. This
4078 * mak sure we don't call discard
4079 * soon for this list.
4084 spin_unlock(&lg->lg_prealloc_lock);
4086 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4088 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4089 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4090 ext4_error(sb, "Error loading buddy information for %u",
4094 ext4_lock_group(sb, group);
4095 list_del(&pa->pa_group_list);
4096 ext4_mb_release_group_pa(&e4b, pa);
4097 ext4_unlock_group(sb, group);
4099 ext4_mb_unload_buddy(&e4b);
4100 list_del(&pa->u.pa_tmp_list);
4101 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4106 * We have incremented pa_count. So it cannot be freed at this
4107 * point. Also we hold lg_mutex. So no parallel allocation is
4108 * possible from this lg. That means pa_free cannot be updated.
4110 * A parallel ext4_mb_discard_group_preallocations is possible.
4111 * which can cause the lg_prealloc_list to be updated.
4114 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4116 int order, added = 0, lg_prealloc_count = 1;
4117 struct super_block *sb = ac->ac_sb;
4118 struct ext4_locality_group *lg = ac->ac_lg;
4119 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4121 order = fls(pa->pa_free) - 1;
4122 if (order > PREALLOC_TB_SIZE - 1)
4123 /* The max size of hash table is PREALLOC_TB_SIZE */
4124 order = PREALLOC_TB_SIZE - 1;
4125 /* Add the prealloc space to lg */
4127 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4129 spin_lock(&tmp_pa->pa_lock);
4130 if (tmp_pa->pa_deleted) {
4131 spin_unlock(&tmp_pa->pa_lock);
4134 if (!added && pa->pa_free < tmp_pa->pa_free) {
4135 /* Add to the tail of the previous entry */
4136 list_add_tail_rcu(&pa->pa_inode_list,
4137 &tmp_pa->pa_inode_list);
4140 * we want to count the total
4141 * number of entries in the list
4144 spin_unlock(&tmp_pa->pa_lock);
4145 lg_prealloc_count++;
4148 list_add_tail_rcu(&pa->pa_inode_list,
4149 &lg->lg_prealloc_list[order]);
4152 /* Now trim the list to be not more than 8 elements */
4153 if (lg_prealloc_count > 8) {
4154 ext4_mb_discard_lg_preallocations(sb, lg,
4155 order, lg_prealloc_count);
4162 * release all resource we used in allocation
4164 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4166 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4167 struct ext4_prealloc_space *pa = ac->ac_pa;
4169 if (pa->pa_type == MB_GROUP_PA) {
4170 /* see comment in ext4_mb_use_group_pa() */
4171 spin_lock(&pa->pa_lock);
4172 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4173 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4174 pa->pa_free -= ac->ac_b_ex.fe_len;
4175 pa->pa_len -= ac->ac_b_ex.fe_len;
4176 spin_unlock(&pa->pa_lock);
4181 * We want to add the pa to the right bucket.
4182 * Remove it from the list and while adding
4183 * make sure the list to which we are adding
4186 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4187 spin_lock(pa->pa_obj_lock);
4188 list_del_rcu(&pa->pa_inode_list);
4189 spin_unlock(pa->pa_obj_lock);
4190 ext4_mb_add_n_trim(ac);
4192 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4194 if (ac->ac_bitmap_page)
4195 page_cache_release(ac->ac_bitmap_page);
4196 if (ac->ac_buddy_page)
4197 page_cache_release(ac->ac_buddy_page);
4198 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4199 mutex_unlock(&ac->ac_lg->lg_mutex);
4200 ext4_mb_collect_stats(ac);
4204 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4206 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4210 trace_ext4_mb_discard_preallocations(sb, needed);
4211 for (i = 0; i < ngroups && needed > 0; i++) {
4212 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4221 * Main entry point into mballoc to allocate blocks
4222 * it tries to use preallocation first, then falls back
4223 * to usual allocation
4225 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4226 struct ext4_allocation_request *ar, int *errp)
4229 struct ext4_allocation_context *ac = NULL;
4230 struct ext4_sb_info *sbi;
4231 struct super_block *sb;
4232 ext4_fsblk_t block = 0;
4233 unsigned int inquota = 0;
4234 unsigned int reserv_clstrs = 0;
4236 sb = ar->inode->i_sb;
4239 trace_ext4_request_blocks(ar);
4241 /* Allow to use superuser reservation for quota file */
4242 if (IS_NOQUOTA(ar->inode))
4243 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4246 * For delayed allocation, we could skip the ENOSPC and
4247 * EDQUOT check, as blocks and quotas have been already
4248 * reserved when data being copied into pagecache.
4250 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4251 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4253 /* Without delayed allocation we need to verify
4254 * there is enough free blocks to do block allocation
4255 * and verify allocation doesn't exceed the quota limits.
4258 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4260 /* let others to free the space */
4262 ar->len = ar->len >> 1;
4268 reserv_clstrs = ar->len;
4269 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4270 dquot_alloc_block_nofail(ar->inode,
4271 EXT4_C2B(sbi, ar->len));
4274 dquot_alloc_block(ar->inode,
4275 EXT4_C2B(sbi, ar->len))) {
4277 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4288 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4295 *errp = ext4_mb_initialize_context(ac, ar);
4301 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4302 if (!ext4_mb_use_preallocated(ac)) {
4303 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4304 ext4_mb_normalize_request(ac, ar);
4306 /* allocate space in core */
4307 *errp = ext4_mb_regular_allocator(ac);
4311 /* as we've just preallocated more space than
4312 * user requested orinally, we store allocated
4313 * space in a special descriptor */
4314 if (ac->ac_status == AC_STATUS_FOUND &&
4315 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4316 ext4_mb_new_preallocation(ac);
4318 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4319 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4320 if (*errp == -EAGAIN) {
4322 * drop the reference that we took
4323 * in ext4_mb_use_best_found
4325 ext4_mb_release_context(ac);
4326 ac->ac_b_ex.fe_group = 0;
4327 ac->ac_b_ex.fe_start = 0;
4328 ac->ac_b_ex.fe_len = 0;
4329 ac->ac_status = AC_STATUS_CONTINUE;
4333 ext4_discard_allocated_blocks(ac);
4335 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4336 ar->len = ac->ac_b_ex.fe_len;
4339 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4346 ac->ac_b_ex.fe_len = 0;
4348 ext4_mb_show_ac(ac);
4350 ext4_mb_release_context(ac);
4353 kmem_cache_free(ext4_ac_cachep, ac);
4354 if (inquota && ar->len < inquota)
4355 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4357 if (!ext4_test_inode_state(ar->inode,
4358 EXT4_STATE_DELALLOC_RESERVED))
4359 /* release all the reserved blocks if non delalloc */
4360 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4364 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4370 * We can merge two free data extents only if the physical blocks
4371 * are contiguous, AND the extents were freed by the same transaction,
4372 * AND the blocks are associated with the same group.
4374 static int can_merge(struct ext4_free_data *entry1,
4375 struct ext4_free_data *entry2)
4377 if ((entry1->efd_tid == entry2->efd_tid) &&
4378 (entry1->efd_group == entry2->efd_group) &&
4379 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4384 static noinline_for_stack int
4385 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4386 struct ext4_free_data *new_entry)
4388 ext4_group_t group = e4b->bd_group;
4389 ext4_grpblk_t cluster;
4390 struct ext4_free_data *entry;
4391 struct ext4_group_info *db = e4b->bd_info;
4392 struct super_block *sb = e4b->bd_sb;
4393 struct ext4_sb_info *sbi = EXT4_SB(sb);
4394 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4395 struct rb_node *parent = NULL, *new_node;
4397 BUG_ON(!ext4_handle_valid(handle));
4398 BUG_ON(e4b->bd_bitmap_page == NULL);
4399 BUG_ON(e4b->bd_buddy_page == NULL);
4401 new_node = &new_entry->efd_node;
4402 cluster = new_entry->efd_start_cluster;
4405 /* first free block exent. We need to
4406 protect buddy cache from being freed,
4407 * otherwise we'll refresh it from
4408 * on-disk bitmap and lose not-yet-available
4410 page_cache_get(e4b->bd_buddy_page);
4411 page_cache_get(e4b->bd_bitmap_page);
4415 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4416 if (cluster < entry->efd_start_cluster)
4418 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4419 n = &(*n)->rb_right;
4421 ext4_grp_locked_error(sb, group, 0,
4422 ext4_group_first_block_no(sb, group) +
4423 EXT4_C2B(sbi, cluster),
4424 "Block already on to-be-freed list");
4429 rb_link_node(new_node, parent, n);
4430 rb_insert_color(new_node, &db->bb_free_root);
4432 /* Now try to see the extent can be merged to left and right */
4433 node = rb_prev(new_node);
4435 entry = rb_entry(node, struct ext4_free_data, efd_node);
4436 if (can_merge(entry, new_entry)) {
4437 new_entry->efd_start_cluster = entry->efd_start_cluster;
4438 new_entry->efd_count += entry->efd_count;
4439 rb_erase(node, &(db->bb_free_root));
4440 ext4_journal_callback_del(handle, &entry->efd_jce);
4441 kmem_cache_free(ext4_free_data_cachep, entry);
4445 node = rb_next(new_node);
4447 entry = rb_entry(node, struct ext4_free_data, efd_node);
4448 if (can_merge(new_entry, entry)) {
4449 new_entry->efd_count += entry->efd_count;
4450 rb_erase(node, &(db->bb_free_root));
4451 ext4_journal_callback_del(handle, &entry->efd_jce);
4452 kmem_cache_free(ext4_free_data_cachep, entry);
4455 /* Add the extent to transaction's private list */
4456 ext4_journal_callback_add(handle, ext4_free_data_callback,
4457 &new_entry->efd_jce);
4462 * ext4_free_blocks() -- Free given blocks and update quota
4463 * @handle: handle for this transaction
4465 * @block: start physical block to free
4466 * @count: number of blocks to count
4467 * @flags: flags used by ext4_free_blocks
4469 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4470 struct buffer_head *bh, ext4_fsblk_t block,
4471 unsigned long count, int flags)
4473 struct buffer_head *bitmap_bh = NULL;
4474 struct super_block *sb = inode->i_sb;
4475 struct ext4_group_desc *gdp;
4476 unsigned long freed = 0;
4477 unsigned int overflow;
4479 struct buffer_head *gd_bh;
4480 ext4_group_t block_group;
4481 struct ext4_sb_info *sbi;
4482 struct ext4_buddy e4b;
4483 unsigned int count_clusters;
4489 BUG_ON(block != bh->b_blocknr);
4491 block = bh->b_blocknr;
4495 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4496 !ext4_data_block_valid(sbi, block, count)) {
4497 ext4_error(sb, "Freeing blocks not in datazone - "
4498 "block = %llu, count = %lu", block, count);
4502 ext4_debug("freeing block %llu\n", block);
4503 trace_ext4_free_blocks(inode, block, count, flags);
4505 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4506 struct buffer_head *tbh = bh;
4509 BUG_ON(bh && (count > 1));
4511 for (i = 0; i < count; i++) {
4513 tbh = sb_find_get_block(inode->i_sb,
4517 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4518 inode, tbh, block + i);
4523 * We need to make sure we don't reuse the freed block until
4524 * after the transaction is committed, which we can do by
4525 * treating the block as metadata, below. We make an
4526 * exception if the inode is to be written in writeback mode
4527 * since writeback mode has weak data consistency guarantees.
4529 if (!ext4_should_writeback_data(inode))
4530 flags |= EXT4_FREE_BLOCKS_METADATA;
4533 * If the extent to be freed does not begin on a cluster
4534 * boundary, we need to deal with partial clusters at the
4535 * beginning and end of the extent. Normally we will free
4536 * blocks at the beginning or the end unless we are explicitly
4537 * requested to avoid doing so.
4539 overflow = block & (sbi->s_cluster_ratio - 1);
4541 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4542 overflow = sbi->s_cluster_ratio - overflow;
4544 if (count > overflow)
4553 overflow = count & (sbi->s_cluster_ratio - 1);
4555 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4556 if (count > overflow)
4561 count += sbi->s_cluster_ratio - overflow;
4566 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4569 * Check to see if we are freeing blocks across a group
4572 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4573 overflow = EXT4_C2B(sbi, bit) + count -
4574 EXT4_BLOCKS_PER_GROUP(sb);
4577 count_clusters = EXT4_B2C(sbi, count);
4578 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4583 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4589 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4590 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4591 in_range(block, ext4_inode_table(sb, gdp),
4592 EXT4_SB(sb)->s_itb_per_group) ||
4593 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4594 EXT4_SB(sb)->s_itb_per_group)) {
4596 ext4_error(sb, "Freeing blocks in system zone - "
4597 "Block = %llu, count = %lu", block, count);
4598 /* err = 0. ext4_std_error should be a no op */
4602 BUFFER_TRACE(bitmap_bh, "getting write access");
4603 err = ext4_journal_get_write_access(handle, bitmap_bh);
4608 * We are about to modify some metadata. Call the journal APIs
4609 * to unshare ->b_data if a currently-committing transaction is
4612 BUFFER_TRACE(gd_bh, "get_write_access");
4613 err = ext4_journal_get_write_access(handle, gd_bh);
4616 #ifdef AGGRESSIVE_CHECK
4619 for (i = 0; i < count_clusters; i++)
4620 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4623 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4625 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4629 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4630 struct ext4_free_data *new_entry;
4632 * blocks being freed are metadata. these blocks shouldn't
4633 * be used until this transaction is committed
4635 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4637 ext4_mb_unload_buddy(&e4b);
4641 new_entry->efd_start_cluster = bit;
4642 new_entry->efd_group = block_group;
4643 new_entry->efd_count = count_clusters;
4644 new_entry->efd_tid = handle->h_transaction->t_tid;
4646 ext4_lock_group(sb, block_group);
4647 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4648 ext4_mb_free_metadata(handle, &e4b, new_entry);
4650 /* need to update group_info->bb_free and bitmap
4651 * with group lock held. generate_buddy look at
4652 * them with group lock_held
4654 ext4_lock_group(sb, block_group);
4655 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4656 mb_free_blocks(inode, &e4b, bit, count_clusters);
4659 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4660 ext4_free_group_clusters_set(sb, gdp, ret);
4661 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh,
4662 EXT4_BLOCKS_PER_GROUP(sb) / 8);
4663 ext4_group_desc_csum_set(sb, block_group, gdp);
4664 ext4_unlock_group(sb, block_group);
4665 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4667 if (sbi->s_log_groups_per_flex) {
4668 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4669 atomic_add(count_clusters,
4670 &sbi->s_flex_groups[flex_group].free_clusters);
4673 ext4_mb_unload_buddy(&e4b);
4677 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4678 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4680 /* We dirtied the bitmap block */
4681 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4682 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4684 /* And the group descriptor block */
4685 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4686 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4690 if (overflow && !err) {
4696 ext4_mark_super_dirty(sb);
4699 ext4_std_error(sb, err);
4704 * ext4_group_add_blocks() -- Add given blocks to an existing group
4705 * @handle: handle to this transaction
4707 * @block: start physcial block to add to the block group
4708 * @count: number of blocks to free
4710 * This marks the blocks as free in the bitmap and buddy.
4712 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4713 ext4_fsblk_t block, unsigned long count)
4715 struct buffer_head *bitmap_bh = NULL;
4716 struct buffer_head *gd_bh;
4717 ext4_group_t block_group;
4720 struct ext4_group_desc *desc;
4721 struct ext4_sb_info *sbi = EXT4_SB(sb);
4722 struct ext4_buddy e4b;
4723 int err = 0, ret, blk_free_count;
4724 ext4_grpblk_t blocks_freed;
4726 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4731 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4733 * Check to see if we are freeing blocks across a group
4736 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4737 ext4_warning(sb, "too much blocks added to group %u\n",
4743 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4749 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4755 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4756 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4757 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4758 in_range(block + count - 1, ext4_inode_table(sb, desc),
4759 sbi->s_itb_per_group)) {
4760 ext4_error(sb, "Adding blocks in system zones - "
4761 "Block = %llu, count = %lu",
4767 BUFFER_TRACE(bitmap_bh, "getting write access");
4768 err = ext4_journal_get_write_access(handle, bitmap_bh);
4773 * We are about to modify some metadata. Call the journal APIs
4774 * to unshare ->b_data if a currently-committing transaction is
4777 BUFFER_TRACE(gd_bh, "get_write_access");
4778 err = ext4_journal_get_write_access(handle, gd_bh);
4782 for (i = 0, blocks_freed = 0; i < count; i++) {
4783 BUFFER_TRACE(bitmap_bh, "clear bit");
4784 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4785 ext4_error(sb, "bit already cleared for block %llu",
4786 (ext4_fsblk_t)(block + i));
4787 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4793 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4798 * need to update group_info->bb_free and bitmap
4799 * with group lock held. generate_buddy look at
4800 * them with group lock_held
4802 ext4_lock_group(sb, block_group);
4803 mb_clear_bits(bitmap_bh->b_data, bit, count);
4804 mb_free_blocks(NULL, &e4b, bit, count);
4805 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4806 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4807 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh,
4808 EXT4_BLOCKS_PER_GROUP(sb) / 8);
4809 ext4_group_desc_csum_set(sb, block_group, desc);
4810 ext4_unlock_group(sb, block_group);
4811 percpu_counter_add(&sbi->s_freeclusters_counter,
4812 EXT4_B2C(sbi, blocks_freed));
4814 if (sbi->s_log_groups_per_flex) {
4815 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4816 atomic_add(EXT4_B2C(sbi, blocks_freed),
4817 &sbi->s_flex_groups[flex_group].free_clusters);
4820 ext4_mb_unload_buddy(&e4b);
4822 /* We dirtied the bitmap block */
4823 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4824 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4826 /* And the group descriptor block */
4827 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4828 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4834 ext4_std_error(sb, err);
4839 * ext4_trim_extent -- function to TRIM one single free extent in the group
4840 * @sb: super block for the file system
4841 * @start: starting block of the free extent in the alloc. group
4842 * @count: number of blocks to TRIM
4843 * @group: alloc. group we are working with
4844 * @e4b: ext4 buddy for the group
4846 * Trim "count" blocks starting at "start" in the "group". To assure that no
4847 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4848 * be called with under the group lock.
4850 static void ext4_trim_extent(struct super_block *sb, int start, int count,
4851 ext4_group_t group, struct ext4_buddy *e4b)
4853 struct ext4_free_extent ex;
4855 trace_ext4_trim_extent(sb, group, start, count);
4857 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4859 ex.fe_start = start;
4860 ex.fe_group = group;
4864 * Mark blocks used, so no one can reuse them while
4867 mb_mark_used(e4b, &ex);
4868 ext4_unlock_group(sb, group);
4869 ext4_issue_discard(sb, group, start, count);
4870 ext4_lock_group(sb, group);
4871 mb_free_blocks(NULL, e4b, start, ex.fe_len);
4875 * ext4_trim_all_free -- function to trim all free space in alloc. group
4876 * @sb: super block for file system
4877 * @group: group to be trimmed
4878 * @start: first group block to examine
4879 * @max: last group block to examine
4880 * @minblocks: minimum extent block count
4882 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4883 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4887 * ext4_trim_all_free walks through group's block bitmap searching for free
4888 * extents. When the free extent is found, mark it as used in group buddy
4889 * bitmap. Then issue a TRIM command on this extent and free the extent in
4890 * the group buddy bitmap. This is done until whole group is scanned.
4892 static ext4_grpblk_t
4893 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4894 ext4_grpblk_t start, ext4_grpblk_t max,
4895 ext4_grpblk_t minblocks)
4898 ext4_grpblk_t next, count = 0, free_count = 0;
4899 struct ext4_buddy e4b;
4902 trace_ext4_trim_all_free(sb, group, start, max);
4904 ret = ext4_mb_load_buddy(sb, group, &e4b);
4906 ext4_error(sb, "Error in loading buddy "
4907 "information for %u", group);
4910 bitmap = e4b.bd_bitmap;
4912 ext4_lock_group(sb, group);
4913 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4914 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4917 start = (e4b.bd_info->bb_first_free > start) ?
4918 e4b.bd_info->bb_first_free : start;
4920 while (start <= max) {
4921 start = mb_find_next_zero_bit(bitmap, max + 1, start);
4924 next = mb_find_next_bit(bitmap, max + 1, start);
4926 if ((next - start) >= minblocks) {
4927 ext4_trim_extent(sb, start,
4928 next - start, group, &e4b);
4929 count += next - start;
4931 free_count += next - start;
4934 if (fatal_signal_pending(current)) {
4935 count = -ERESTARTSYS;
4939 if (need_resched()) {
4940 ext4_unlock_group(sb, group);
4942 ext4_lock_group(sb, group);
4945 if ((e4b.bd_info->bb_free - free_count) < minblocks)
4950 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
4952 ext4_unlock_group(sb, group);
4953 ext4_mb_unload_buddy(&e4b);
4955 ext4_debug("trimmed %d blocks in the group %d\n",
4962 * ext4_trim_fs() -- trim ioctl handle function
4963 * @sb: superblock for filesystem
4964 * @range: fstrim_range structure
4966 * start: First Byte to trim
4967 * len: number of Bytes to trim from start
4968 * minlen: minimum extent length in Bytes
4969 * ext4_trim_fs goes through all allocation groups containing Bytes from
4970 * start to start+len. For each such a group ext4_trim_all_free function
4971 * is invoked to trim all free space.
4973 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
4975 struct ext4_group_info *grp;
4976 ext4_group_t group, first_group, last_group;
4977 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
4978 uint64_t start, end, minlen, trimmed = 0;
4979 ext4_fsblk_t first_data_blk =
4980 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
4981 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
4984 start = range->start >> sb->s_blocksize_bits;
4985 end = start + (range->len >> sb->s_blocksize_bits) - 1;
4986 minlen = range->minlen >> sb->s_blocksize_bits;
4988 if (unlikely(minlen > EXT4_CLUSTERS_PER_GROUP(sb)) ||
4989 unlikely(start >= max_blks))
4991 if (end >= max_blks)
4993 if (end <= first_data_blk)
4995 if (start < first_data_blk)
4996 start = first_data_blk;
4998 /* Determine first and last group to examine based on start and end */
4999 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5000 &first_group, &first_cluster);
5001 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5002 &last_group, &last_cluster);
5004 /* end now represents the last cluster to discard in this group */
5005 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5007 for (group = first_group; group <= last_group; group++) {
5008 grp = ext4_get_group_info(sb, group);
5009 /* We only do this if the grp has never been initialized */
5010 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5011 ret = ext4_mb_init_group(sb, group);
5017 * For all the groups except the last one, last cluster will
5018 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5019 * change it for the last group, note that last_cluster is
5020 * already computed earlier by ext4_get_group_no_and_offset()
5022 if (group == last_group)
5025 if (grp->bb_free >= minlen) {
5026 cnt = ext4_trim_all_free(sb, group, first_cluster,
5036 * For every group except the first one, we are sure
5037 * that the first cluster to discard will be cluster #0.
5043 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5046 range->len = trimmed * sb->s_blocksize;