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 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
975 BUG_ON(page->mapping != inode->i_mapping);
976 e4b->bd_buddy_page = page;
980 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
982 if (e4b->bd_bitmap_page) {
983 unlock_page(e4b->bd_bitmap_page);
984 page_cache_release(e4b->bd_bitmap_page);
986 if (e4b->bd_buddy_page) {
987 unlock_page(e4b->bd_buddy_page);
988 page_cache_release(e4b->bd_buddy_page);
993 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
994 * block group lock of all groups for this page; do not hold the BG lock when
995 * calling this routine!
997 static noinline_for_stack
998 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1001 struct ext4_group_info *this_grp;
1002 struct ext4_buddy e4b;
1006 mb_debug(1, "init group %u\n", group);
1007 this_grp = ext4_get_group_info(sb, group);
1009 * This ensures that we don't reinit the buddy cache
1010 * page which map to the group from which we are already
1011 * allocating. If we are looking at the buddy cache we would
1012 * have taken a reference using ext4_mb_load_buddy and that
1013 * would have pinned buddy page to page cache.
1015 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1016 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1018 * somebody initialized the group
1019 * return without doing anything
1024 page = e4b.bd_bitmap_page;
1025 ret = ext4_mb_init_cache(page, NULL);
1028 if (!PageUptodate(page)) {
1032 mark_page_accessed(page);
1034 if (e4b.bd_buddy_page == NULL) {
1036 * If both the bitmap and buddy are in
1037 * the same page we don't need to force
1043 /* init buddy cache */
1044 page = e4b.bd_buddy_page;
1045 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1048 if (!PageUptodate(page)) {
1052 mark_page_accessed(page);
1054 ext4_mb_put_buddy_page_lock(&e4b);
1059 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1060 * block group lock of all groups for this page; do not hold the BG lock when
1061 * calling this routine!
1063 static noinline_for_stack int
1064 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1065 struct ext4_buddy *e4b)
1067 int blocks_per_page;
1073 struct ext4_group_info *grp;
1074 struct ext4_sb_info *sbi = EXT4_SB(sb);
1075 struct inode *inode = sbi->s_buddy_cache;
1077 mb_debug(1, "load group %u\n", group);
1079 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1080 grp = ext4_get_group_info(sb, group);
1082 e4b->bd_blkbits = sb->s_blocksize_bits;
1085 e4b->bd_group = group;
1086 e4b->bd_buddy_page = NULL;
1087 e4b->bd_bitmap_page = NULL;
1089 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1091 * we need full data about the group
1092 * to make a good selection
1094 ret = ext4_mb_init_group(sb, group);
1100 * the buddy cache inode stores the block bitmap
1101 * and buddy information in consecutive blocks.
1102 * So for each group we need two blocks.
1105 pnum = block / blocks_per_page;
1106 poff = block % blocks_per_page;
1108 /* we could use find_or_create_page(), but it locks page
1109 * what we'd like to avoid in fast path ... */
1110 page = find_get_page(inode->i_mapping, pnum);
1111 if (page == NULL || !PageUptodate(page)) {
1114 * drop the page reference and try
1115 * to get the page with lock. If we
1116 * are not uptodate that implies
1117 * somebody just created the page but
1118 * is yet to initialize the same. So
1119 * wait for it to initialize.
1121 page_cache_release(page);
1122 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1124 BUG_ON(page->mapping != inode->i_mapping);
1125 if (!PageUptodate(page)) {
1126 ret = ext4_mb_init_cache(page, NULL);
1131 mb_cmp_bitmaps(e4b, page_address(page) +
1132 (poff * sb->s_blocksize));
1137 if (page == NULL || !PageUptodate(page)) {
1141 e4b->bd_bitmap_page = page;
1142 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1143 mark_page_accessed(page);
1146 pnum = block / blocks_per_page;
1147 poff = block % blocks_per_page;
1149 page = find_get_page(inode->i_mapping, pnum);
1150 if (page == NULL || !PageUptodate(page)) {
1152 page_cache_release(page);
1153 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1155 BUG_ON(page->mapping != inode->i_mapping);
1156 if (!PageUptodate(page)) {
1157 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1166 if (page == NULL || !PageUptodate(page)) {
1170 e4b->bd_buddy_page = page;
1171 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1172 mark_page_accessed(page);
1174 BUG_ON(e4b->bd_bitmap_page == NULL);
1175 BUG_ON(e4b->bd_buddy_page == NULL);
1181 page_cache_release(page);
1182 if (e4b->bd_bitmap_page)
1183 page_cache_release(e4b->bd_bitmap_page);
1184 if (e4b->bd_buddy_page)
1185 page_cache_release(e4b->bd_buddy_page);
1186 e4b->bd_buddy = NULL;
1187 e4b->bd_bitmap = NULL;
1191 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1193 if (e4b->bd_bitmap_page)
1194 page_cache_release(e4b->bd_bitmap_page);
1195 if (e4b->bd_buddy_page)
1196 page_cache_release(e4b->bd_buddy_page);
1200 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1205 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1206 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1209 while (order <= e4b->bd_blkbits + 1) {
1211 if (!mb_test_bit(block, bb)) {
1212 /* this block is part of buddy of order 'order' */
1215 bb += 1 << (e4b->bd_blkbits - order);
1221 static void mb_clear_bits(void *bm, int cur, int len)
1227 if ((cur & 31) == 0 && (len - cur) >= 32) {
1228 /* fast path: clear whole word at once */
1229 addr = bm + (cur >> 3);
1234 mb_clear_bit(cur, bm);
1239 void ext4_set_bits(void *bm, int cur, int len)
1245 if ((cur & 31) == 0 && (len - cur) >= 32) {
1246 /* fast path: set whole word at once */
1247 addr = bm + (cur >> 3);
1252 mb_set_bit(cur, bm);
1257 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1258 int first, int count)
1265 struct super_block *sb = e4b->bd_sb;
1267 BUG_ON(first + count > (sb->s_blocksize << 3));
1268 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1269 mb_check_buddy(e4b);
1270 mb_free_blocks_double(inode, e4b, first, count);
1272 e4b->bd_info->bb_free += count;
1273 if (first < e4b->bd_info->bb_first_free)
1274 e4b->bd_info->bb_first_free = first;
1276 /* let's maintain fragments counter */
1278 block = !mb_test_bit(first - 1, e4b->bd_bitmap);
1279 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1280 max = !mb_test_bit(first + count, e4b->bd_bitmap);
1282 e4b->bd_info->bb_fragments--;
1283 else if (!block && !max)
1284 e4b->bd_info->bb_fragments++;
1286 /* let's maintain buddy itself */
1287 while (count-- > 0) {
1291 if (!mb_test_bit(block, e4b->bd_bitmap)) {
1292 ext4_fsblk_t blocknr;
1294 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1295 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1296 ext4_grp_locked_error(sb, e4b->bd_group,
1297 inode ? inode->i_ino : 0,
1299 "freeing already freed block "
1302 mb_clear_bit(block, e4b->bd_bitmap);
1303 e4b->bd_info->bb_counters[order]++;
1305 /* start of the buddy */
1306 buddy = mb_find_buddy(e4b, order, &max);
1310 if (mb_test_bit(block, buddy) ||
1311 mb_test_bit(block + 1, buddy))
1314 /* both the buddies are free, try to coalesce them */
1315 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1321 /* for special purposes, we don't set
1322 * free bits in bitmap */
1323 mb_set_bit(block, buddy);
1324 mb_set_bit(block + 1, buddy);
1326 e4b->bd_info->bb_counters[order]--;
1327 e4b->bd_info->bb_counters[order]--;
1331 e4b->bd_info->bb_counters[order]++;
1333 mb_clear_bit(block, buddy2);
1337 mb_set_largest_free_order(sb, e4b->bd_info);
1338 mb_check_buddy(e4b);
1341 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1342 int needed, struct ext4_free_extent *ex)
1348 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1351 buddy = mb_find_buddy(e4b, order, &max);
1352 BUG_ON(buddy == NULL);
1353 BUG_ON(block >= max);
1354 if (mb_test_bit(block, buddy)) {
1361 /* FIXME dorp order completely ? */
1362 if (likely(order == 0)) {
1363 /* find actual order */
1364 order = mb_find_order_for_block(e4b, block);
1365 block = block >> order;
1368 ex->fe_len = 1 << order;
1369 ex->fe_start = block << order;
1370 ex->fe_group = e4b->bd_group;
1372 /* calc difference from given start */
1373 next = next - ex->fe_start;
1375 ex->fe_start += next;
1377 while (needed > ex->fe_len &&
1378 (buddy = mb_find_buddy(e4b, order, &max))) {
1380 if (block + 1 >= max)
1383 next = (block + 1) * (1 << order);
1384 if (mb_test_bit(next, e4b->bd_bitmap))
1387 order = mb_find_order_for_block(e4b, next);
1389 block = next >> order;
1390 ex->fe_len += 1 << order;
1393 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1397 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1403 int start = ex->fe_start;
1404 int len = ex->fe_len;
1409 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1410 BUG_ON(e4b->bd_group != ex->fe_group);
1411 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1412 mb_check_buddy(e4b);
1413 mb_mark_used_double(e4b, start, len);
1415 e4b->bd_info->bb_free -= len;
1416 if (e4b->bd_info->bb_first_free == start)
1417 e4b->bd_info->bb_first_free += len;
1419 /* let's maintain fragments counter */
1421 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1422 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1423 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1425 e4b->bd_info->bb_fragments++;
1426 else if (!mlen && !max)
1427 e4b->bd_info->bb_fragments--;
1429 /* let's maintain buddy itself */
1431 ord = mb_find_order_for_block(e4b, start);
1433 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1434 /* the whole chunk may be allocated at once! */
1436 buddy = mb_find_buddy(e4b, ord, &max);
1437 BUG_ON((start >> ord) >= max);
1438 mb_set_bit(start >> ord, buddy);
1439 e4b->bd_info->bb_counters[ord]--;
1446 /* store for history */
1448 ret = len | (ord << 16);
1450 /* we have to split large buddy */
1452 buddy = mb_find_buddy(e4b, ord, &max);
1453 mb_set_bit(start >> ord, buddy);
1454 e4b->bd_info->bb_counters[ord]--;
1457 cur = (start >> ord) & ~1U;
1458 buddy = mb_find_buddy(e4b, ord, &max);
1459 mb_clear_bit(cur, buddy);
1460 mb_clear_bit(cur + 1, buddy);
1461 e4b->bd_info->bb_counters[ord]++;
1462 e4b->bd_info->bb_counters[ord]++;
1464 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1466 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1467 mb_check_buddy(e4b);
1473 * Must be called under group lock!
1475 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1476 struct ext4_buddy *e4b)
1478 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1481 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1482 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1484 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1485 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1486 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1488 /* preallocation can change ac_b_ex, thus we store actually
1489 * allocated blocks for history */
1490 ac->ac_f_ex = ac->ac_b_ex;
1492 ac->ac_status = AC_STATUS_FOUND;
1493 ac->ac_tail = ret & 0xffff;
1494 ac->ac_buddy = ret >> 16;
1497 * take the page reference. We want the page to be pinned
1498 * so that we don't get a ext4_mb_init_cache_call for this
1499 * group until we update the bitmap. That would mean we
1500 * double allocate blocks. The reference is dropped
1501 * in ext4_mb_release_context
1503 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1504 get_page(ac->ac_bitmap_page);
1505 ac->ac_buddy_page = e4b->bd_buddy_page;
1506 get_page(ac->ac_buddy_page);
1507 /* store last allocated for subsequent stream allocation */
1508 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1509 spin_lock(&sbi->s_md_lock);
1510 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1511 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1512 spin_unlock(&sbi->s_md_lock);
1517 * regular allocator, for general purposes allocation
1520 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1521 struct ext4_buddy *e4b,
1524 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1525 struct ext4_free_extent *bex = &ac->ac_b_ex;
1526 struct ext4_free_extent *gex = &ac->ac_g_ex;
1527 struct ext4_free_extent ex;
1530 if (ac->ac_status == AC_STATUS_FOUND)
1533 * We don't want to scan for a whole year
1535 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1536 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1537 ac->ac_status = AC_STATUS_BREAK;
1542 * Haven't found good chunk so far, let's continue
1544 if (bex->fe_len < gex->fe_len)
1547 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1548 && bex->fe_group == e4b->bd_group) {
1549 /* recheck chunk's availability - we don't know
1550 * when it was found (within this lock-unlock
1552 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1553 if (max >= gex->fe_len) {
1554 ext4_mb_use_best_found(ac, e4b);
1561 * The routine checks whether found extent is good enough. If it is,
1562 * then the extent gets marked used and flag is set to the context
1563 * to stop scanning. Otherwise, the extent is compared with the
1564 * previous found extent and if new one is better, then it's stored
1565 * in the context. Later, the best found extent will be used, if
1566 * mballoc can't find good enough extent.
1568 * FIXME: real allocation policy is to be designed yet!
1570 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1571 struct ext4_free_extent *ex,
1572 struct ext4_buddy *e4b)
1574 struct ext4_free_extent *bex = &ac->ac_b_ex;
1575 struct ext4_free_extent *gex = &ac->ac_g_ex;
1577 BUG_ON(ex->fe_len <= 0);
1578 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1579 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1580 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1585 * The special case - take what you catch first
1587 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1589 ext4_mb_use_best_found(ac, e4b);
1594 * Let's check whether the chuck is good enough
1596 if (ex->fe_len == gex->fe_len) {
1598 ext4_mb_use_best_found(ac, e4b);
1603 * If this is first found extent, just store it in the context
1605 if (bex->fe_len == 0) {
1611 * If new found extent is better, store it in the context
1613 if (bex->fe_len < gex->fe_len) {
1614 /* if the request isn't satisfied, any found extent
1615 * larger than previous best one is better */
1616 if (ex->fe_len > bex->fe_len)
1618 } else if (ex->fe_len > gex->fe_len) {
1619 /* if the request is satisfied, then we try to find
1620 * an extent that still satisfy the request, but is
1621 * smaller than previous one */
1622 if (ex->fe_len < bex->fe_len)
1626 ext4_mb_check_limits(ac, e4b, 0);
1629 static noinline_for_stack
1630 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1631 struct ext4_buddy *e4b)
1633 struct ext4_free_extent ex = ac->ac_b_ex;
1634 ext4_group_t group = ex.fe_group;
1638 BUG_ON(ex.fe_len <= 0);
1639 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1643 ext4_lock_group(ac->ac_sb, group);
1644 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1648 ext4_mb_use_best_found(ac, e4b);
1651 ext4_unlock_group(ac->ac_sb, group);
1652 ext4_mb_unload_buddy(e4b);
1657 static noinline_for_stack
1658 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1659 struct ext4_buddy *e4b)
1661 ext4_group_t group = ac->ac_g_ex.fe_group;
1664 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1665 struct ext4_free_extent ex;
1667 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1670 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1674 ext4_lock_group(ac->ac_sb, group);
1675 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1676 ac->ac_g_ex.fe_len, &ex);
1678 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1681 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1683 /* use do_div to get remainder (would be 64-bit modulo) */
1684 if (do_div(start, sbi->s_stripe) == 0) {
1687 ext4_mb_use_best_found(ac, e4b);
1689 } else if (max >= ac->ac_g_ex.fe_len) {
1690 BUG_ON(ex.fe_len <= 0);
1691 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1692 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1695 ext4_mb_use_best_found(ac, e4b);
1696 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1697 /* Sometimes, caller may want to merge even small
1698 * number of blocks to an existing extent */
1699 BUG_ON(ex.fe_len <= 0);
1700 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1701 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1704 ext4_mb_use_best_found(ac, e4b);
1706 ext4_unlock_group(ac->ac_sb, group);
1707 ext4_mb_unload_buddy(e4b);
1713 * The routine scans buddy structures (not bitmap!) from given order
1714 * to max order and tries to find big enough chunk to satisfy the req
1716 static noinline_for_stack
1717 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1718 struct ext4_buddy *e4b)
1720 struct super_block *sb = ac->ac_sb;
1721 struct ext4_group_info *grp = e4b->bd_info;
1727 BUG_ON(ac->ac_2order <= 0);
1728 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1729 if (grp->bb_counters[i] == 0)
1732 buddy = mb_find_buddy(e4b, i, &max);
1733 BUG_ON(buddy == NULL);
1735 k = mb_find_next_zero_bit(buddy, max, 0);
1740 ac->ac_b_ex.fe_len = 1 << i;
1741 ac->ac_b_ex.fe_start = k << i;
1742 ac->ac_b_ex.fe_group = e4b->bd_group;
1744 ext4_mb_use_best_found(ac, e4b);
1746 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1748 if (EXT4_SB(sb)->s_mb_stats)
1749 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1756 * The routine scans the group and measures all found extents.
1757 * In order to optimize scanning, caller must pass number of
1758 * free blocks in the group, so the routine can know upper limit.
1760 static noinline_for_stack
1761 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1762 struct ext4_buddy *e4b)
1764 struct super_block *sb = ac->ac_sb;
1765 void *bitmap = e4b->bd_bitmap;
1766 struct ext4_free_extent ex;
1770 free = e4b->bd_info->bb_free;
1773 i = e4b->bd_info->bb_first_free;
1775 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1776 i = mb_find_next_zero_bit(bitmap,
1777 EXT4_CLUSTERS_PER_GROUP(sb), i);
1778 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1780 * IF we have corrupt bitmap, we won't find any
1781 * free blocks even though group info says we
1782 * we have free blocks
1784 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1785 "%d free clusters as per "
1786 "group info. But bitmap says 0",
1791 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1792 BUG_ON(ex.fe_len <= 0);
1793 if (free < ex.fe_len) {
1794 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1795 "%d free clusters as per "
1796 "group info. But got %d blocks",
1799 * The number of free blocks differs. This mostly
1800 * indicate that the bitmap is corrupt. So exit
1801 * without claiming the space.
1806 ext4_mb_measure_extent(ac, &ex, e4b);
1812 ext4_mb_check_limits(ac, e4b, 1);
1816 * This is a special case for storages like raid5
1817 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1819 static noinline_for_stack
1820 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1821 struct ext4_buddy *e4b)
1823 struct super_block *sb = ac->ac_sb;
1824 struct ext4_sb_info *sbi = EXT4_SB(sb);
1825 void *bitmap = e4b->bd_bitmap;
1826 struct ext4_free_extent ex;
1827 ext4_fsblk_t first_group_block;
1832 BUG_ON(sbi->s_stripe == 0);
1834 /* find first stripe-aligned block in group */
1835 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1837 a = first_group_block + sbi->s_stripe - 1;
1838 do_div(a, sbi->s_stripe);
1839 i = (a * sbi->s_stripe) - first_group_block;
1841 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1842 if (!mb_test_bit(i, bitmap)) {
1843 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1844 if (max >= sbi->s_stripe) {
1847 ext4_mb_use_best_found(ac, e4b);
1855 /* This is now called BEFORE we load the buddy bitmap. */
1856 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1857 ext4_group_t group, int cr)
1859 unsigned free, fragments;
1860 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1861 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1863 BUG_ON(cr < 0 || cr >= 4);
1865 /* We only do this if the grp has never been initialized */
1866 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1867 int ret = ext4_mb_init_group(ac->ac_sb, group);
1872 free = grp->bb_free;
1873 fragments = grp->bb_fragments;
1881 BUG_ON(ac->ac_2order == 0);
1883 if (grp->bb_largest_free_order < ac->ac_2order)
1886 /* Avoid using the first bg of a flexgroup for data files */
1887 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1888 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1889 ((group % flex_size) == 0))
1894 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1898 if (free >= ac->ac_g_ex.fe_len)
1910 static noinline_for_stack int
1911 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1913 ext4_group_t ngroups, group, i;
1916 struct ext4_sb_info *sbi;
1917 struct super_block *sb;
1918 struct ext4_buddy e4b;
1922 ngroups = ext4_get_groups_count(sb);
1923 /* non-extent files are limited to low blocks/groups */
1924 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
1925 ngroups = sbi->s_blockfile_groups;
1927 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1929 /* first, try the goal */
1930 err = ext4_mb_find_by_goal(ac, &e4b);
1931 if (err || ac->ac_status == AC_STATUS_FOUND)
1934 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1938 * ac->ac2_order is set only if the fe_len is a power of 2
1939 * if ac2_order is set we also set criteria to 0 so that we
1940 * try exact allocation using buddy.
1942 i = fls(ac->ac_g_ex.fe_len);
1945 * We search using buddy data only if the order of the request
1946 * is greater than equal to the sbi_s_mb_order2_reqs
1947 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1949 if (i >= sbi->s_mb_order2_reqs) {
1951 * This should tell if fe_len is exactly power of 2
1953 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1954 ac->ac_2order = i - 1;
1957 /* if stream allocation is enabled, use global goal */
1958 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1959 /* TBD: may be hot point */
1960 spin_lock(&sbi->s_md_lock);
1961 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1962 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1963 spin_unlock(&sbi->s_md_lock);
1966 /* Let's just scan groups to find more-less suitable blocks */
1967 cr = ac->ac_2order ? 0 : 1;
1969 * cr == 0 try to get exact allocation,
1970 * cr == 3 try to get anything
1973 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1974 ac->ac_criteria = cr;
1976 * searching for the right group start
1977 * from the goal value specified
1979 group = ac->ac_g_ex.fe_group;
1981 for (i = 0; i < ngroups; group++, i++) {
1982 if (group == ngroups)
1985 /* This now checks without needing the buddy page */
1986 if (!ext4_mb_good_group(ac, group, cr))
1989 err = ext4_mb_load_buddy(sb, group, &e4b);
1993 ext4_lock_group(sb, group);
1996 * We need to check again after locking the
1999 if (!ext4_mb_good_group(ac, group, cr)) {
2000 ext4_unlock_group(sb, group);
2001 ext4_mb_unload_buddy(&e4b);
2005 ac->ac_groups_scanned++;
2007 ext4_mb_simple_scan_group(ac, &e4b);
2008 else if (cr == 1 && sbi->s_stripe &&
2009 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2010 ext4_mb_scan_aligned(ac, &e4b);
2012 ext4_mb_complex_scan_group(ac, &e4b);
2014 ext4_unlock_group(sb, group);
2015 ext4_mb_unload_buddy(&e4b);
2017 if (ac->ac_status != AC_STATUS_CONTINUE)
2022 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2023 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2025 * We've been searching too long. Let's try to allocate
2026 * the best chunk we've found so far
2029 ext4_mb_try_best_found(ac, &e4b);
2030 if (ac->ac_status != AC_STATUS_FOUND) {
2032 * Someone more lucky has already allocated it.
2033 * The only thing we can do is just take first
2035 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2037 ac->ac_b_ex.fe_group = 0;
2038 ac->ac_b_ex.fe_start = 0;
2039 ac->ac_b_ex.fe_len = 0;
2040 ac->ac_status = AC_STATUS_CONTINUE;
2041 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2043 atomic_inc(&sbi->s_mb_lost_chunks);
2051 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2053 struct super_block *sb = seq->private;
2056 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2059 return (void *) ((unsigned long) group);
2062 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2064 struct super_block *sb = seq->private;
2068 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2071 return (void *) ((unsigned long) group);
2074 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2076 struct super_block *sb = seq->private;
2077 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2079 int err, buddy_loaded = 0;
2080 struct ext4_buddy e4b;
2081 struct ext4_group_info *grinfo;
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 grinfo = ext4_get_group_info(sb, group);
2099 /* Load the group info in memory only if not already loaded. */
2100 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2101 err = ext4_mb_load_buddy(sb, group, &e4b);
2103 seq_printf(seq, "#%-5u: I/O error\n", group);
2109 memcpy(&sg, ext4_get_group_info(sb, group), i);
2112 ext4_mb_unload_buddy(&e4b);
2114 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2115 sg.info.bb_fragments, sg.info.bb_first_free);
2116 for (i = 0; i <= 13; i++)
2117 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2118 sg.info.bb_counters[i] : 0);
2119 seq_printf(seq, " ]\n");
2124 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2128 static const struct seq_operations ext4_mb_seq_groups_ops = {
2129 .start = ext4_mb_seq_groups_start,
2130 .next = ext4_mb_seq_groups_next,
2131 .stop = ext4_mb_seq_groups_stop,
2132 .show = ext4_mb_seq_groups_show,
2135 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2137 struct super_block *sb = PDE(inode)->data;
2140 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2142 struct seq_file *m = file->private_data;
2149 static const struct file_operations ext4_mb_seq_groups_fops = {
2150 .owner = THIS_MODULE,
2151 .open = ext4_mb_seq_groups_open,
2153 .llseek = seq_lseek,
2154 .release = seq_release,
2157 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2159 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2160 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2166 /* Create and initialize ext4_group_info data for the given group. */
2167 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2168 struct ext4_group_desc *desc)
2172 struct ext4_sb_info *sbi = EXT4_SB(sb);
2173 struct ext4_group_info **meta_group_info;
2174 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2177 * First check if this group is the first of a reserved block.
2178 * If it's true, we have to allocate a new table of pointers
2179 * to ext4_group_info structures
2181 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2182 metalen = sizeof(*meta_group_info) <<
2183 EXT4_DESC_PER_BLOCK_BITS(sb);
2184 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2185 if (meta_group_info == NULL) {
2186 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2187 "for a buddy group");
2188 goto exit_meta_group_info;
2190 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2195 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2196 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2198 meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL);
2199 if (meta_group_info[i] == NULL) {
2200 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2201 goto exit_group_info;
2203 memset(meta_group_info[i], 0, kmem_cache_size(cachep));
2204 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2205 &(meta_group_info[i]->bb_state));
2208 * initialize bb_free to be able to skip
2209 * empty groups without initialization
2211 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2212 meta_group_info[i]->bb_free =
2213 ext4_free_clusters_after_init(sb, group, desc);
2215 meta_group_info[i]->bb_free =
2216 ext4_free_group_clusters(sb, desc);
2219 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2220 init_rwsem(&meta_group_info[i]->alloc_sem);
2221 meta_group_info[i]->bb_free_root = RB_ROOT;
2222 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2226 struct buffer_head *bh;
2227 meta_group_info[i]->bb_bitmap =
2228 kmalloc(sb->s_blocksize, GFP_KERNEL);
2229 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2230 bh = ext4_read_block_bitmap(sb, group);
2232 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2241 /* If a meta_group_info table has been allocated, release it now */
2242 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2243 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2244 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2246 exit_meta_group_info:
2248 } /* ext4_mb_add_groupinfo */
2250 static int ext4_mb_init_backend(struct super_block *sb)
2252 ext4_group_t ngroups = ext4_get_groups_count(sb);
2254 struct ext4_sb_info *sbi = EXT4_SB(sb);
2255 struct ext4_super_block *es = sbi->s_es;
2256 int num_meta_group_infos;
2257 int num_meta_group_infos_max;
2259 struct ext4_group_desc *desc;
2260 struct kmem_cache *cachep;
2262 /* This is the number of blocks used by GDT */
2263 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2264 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2267 * This is the total number of blocks used by GDT including
2268 * the number of reserved blocks for GDT.
2269 * The s_group_info array is allocated with this value
2270 * to allow a clean online resize without a complex
2271 * manipulation of pointer.
2272 * The drawback is the unused memory when no resize
2273 * occurs but it's very low in terms of pages
2274 * (see comments below)
2275 * Need to handle this properly when META_BG resizing is allowed
2277 num_meta_group_infos_max = num_meta_group_infos +
2278 le16_to_cpu(es->s_reserved_gdt_blocks);
2281 * array_size is the size of s_group_info array. We round it
2282 * to the next power of two because this approximation is done
2283 * internally by kmalloc so we can have some more memory
2284 * for free here (e.g. may be used for META_BG resize).
2287 while (array_size < sizeof(*sbi->s_group_info) *
2288 num_meta_group_infos_max)
2289 array_size = array_size << 1;
2290 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2291 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2292 * So a two level scheme suffices for now. */
2293 sbi->s_group_info = ext4_kvzalloc(array_size, GFP_KERNEL);
2294 if (sbi->s_group_info == NULL) {
2295 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2298 sbi->s_buddy_cache = new_inode(sb);
2299 if (sbi->s_buddy_cache == NULL) {
2300 ext4_msg(sb, KERN_ERR, "can't get new inode");
2303 /* To avoid potentially colliding with an valid on-disk inode number,
2304 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2305 * not in the inode hash, so it should never be found by iget(), but
2306 * this will avoid confusion if it ever shows up during debugging. */
2307 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2308 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2309 for (i = 0; i < ngroups; i++) {
2310 desc = ext4_get_group_desc(sb, i, NULL);
2312 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2315 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2322 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2324 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2325 i = num_meta_group_infos;
2327 kfree(sbi->s_group_info[i]);
2328 iput(sbi->s_buddy_cache);
2330 ext4_kvfree(sbi->s_group_info);
2334 static void ext4_groupinfo_destroy_slabs(void)
2338 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2339 if (ext4_groupinfo_caches[i])
2340 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2341 ext4_groupinfo_caches[i] = NULL;
2345 static int ext4_groupinfo_create_slab(size_t size)
2347 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2349 int blocksize_bits = order_base_2(size);
2350 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2351 struct kmem_cache *cachep;
2353 if (cache_index >= NR_GRPINFO_CACHES)
2356 if (unlikely(cache_index < 0))
2359 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2360 if (ext4_groupinfo_caches[cache_index]) {
2361 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2362 return 0; /* Already created */
2365 slab_size = offsetof(struct ext4_group_info,
2366 bb_counters[blocksize_bits + 2]);
2368 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2369 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2372 ext4_groupinfo_caches[cache_index] = cachep;
2374 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2377 "EXT4-fs: no memory for groupinfo slab cache\n");
2384 int ext4_mb_init(struct super_block *sb)
2386 struct ext4_sb_info *sbi = EXT4_SB(sb);
2392 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2394 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2395 if (sbi->s_mb_offsets == NULL) {
2400 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2401 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2402 if (sbi->s_mb_maxs == NULL) {
2407 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2411 /* order 0 is regular bitmap */
2412 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2413 sbi->s_mb_offsets[0] = 0;
2417 max = sb->s_blocksize << 2;
2419 sbi->s_mb_offsets[i] = offset;
2420 sbi->s_mb_maxs[i] = max;
2421 offset += 1 << (sb->s_blocksize_bits - i);
2424 } while (i <= sb->s_blocksize_bits + 1);
2426 spin_lock_init(&sbi->s_md_lock);
2427 spin_lock_init(&sbi->s_bal_lock);
2429 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2430 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2431 sbi->s_mb_stats = MB_DEFAULT_STATS;
2432 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2433 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2435 * The default group preallocation is 512, which for 4k block
2436 * sizes translates to 2 megabytes. However for bigalloc file
2437 * systems, this is probably too big (i.e, if the cluster size
2438 * is 1 megabyte, then group preallocation size becomes half a
2439 * gigabyte!). As a default, we will keep a two megabyte
2440 * group pralloc size for cluster sizes up to 64k, and after
2441 * that, we will force a minimum group preallocation size of
2442 * 32 clusters. This translates to 8 megs when the cluster
2443 * size is 256k, and 32 megs when the cluster size is 1 meg,
2444 * which seems reasonable as a default.
2446 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2447 sbi->s_cluster_bits, 32);
2449 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2450 * to the lowest multiple of s_stripe which is bigger than
2451 * the s_mb_group_prealloc as determined above. We want
2452 * the preallocation size to be an exact multiple of the
2453 * RAID stripe size so that preallocations don't fragment
2456 if (sbi->s_stripe > 1) {
2457 sbi->s_mb_group_prealloc = roundup(
2458 sbi->s_mb_group_prealloc, sbi->s_stripe);
2461 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2462 if (sbi->s_locality_groups == NULL) {
2464 goto out_free_groupinfo_slab;
2466 for_each_possible_cpu(i) {
2467 struct ext4_locality_group *lg;
2468 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2469 mutex_init(&lg->lg_mutex);
2470 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2471 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2472 spin_lock_init(&lg->lg_prealloc_lock);
2475 /* init file for buddy data */
2476 ret = ext4_mb_init_backend(sb);
2478 goto out_free_locality_groups;
2481 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2482 &ext4_mb_seq_groups_fops, sb);
2486 out_free_locality_groups:
2487 free_percpu(sbi->s_locality_groups);
2488 sbi->s_locality_groups = NULL;
2489 out_free_groupinfo_slab:
2490 ext4_groupinfo_destroy_slabs();
2492 kfree(sbi->s_mb_offsets);
2493 sbi->s_mb_offsets = NULL;
2494 kfree(sbi->s_mb_maxs);
2495 sbi->s_mb_maxs = NULL;
2499 /* need to called with the ext4 group lock held */
2500 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2502 struct ext4_prealloc_space *pa;
2503 struct list_head *cur, *tmp;
2506 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2507 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2508 list_del(&pa->pa_group_list);
2510 kmem_cache_free(ext4_pspace_cachep, pa);
2513 mb_debug(1, "mballoc: %u PAs left\n", count);
2517 int ext4_mb_release(struct super_block *sb)
2519 ext4_group_t ngroups = ext4_get_groups_count(sb);
2521 int num_meta_group_infos;
2522 struct ext4_group_info *grinfo;
2523 struct ext4_sb_info *sbi = EXT4_SB(sb);
2524 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2527 remove_proc_entry("mb_groups", sbi->s_proc);
2529 if (sbi->s_group_info) {
2530 for (i = 0; i < ngroups; i++) {
2531 grinfo = ext4_get_group_info(sb, i);
2533 kfree(grinfo->bb_bitmap);
2535 ext4_lock_group(sb, i);
2536 ext4_mb_cleanup_pa(grinfo);
2537 ext4_unlock_group(sb, i);
2538 kmem_cache_free(cachep, grinfo);
2540 num_meta_group_infos = (ngroups +
2541 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2542 EXT4_DESC_PER_BLOCK_BITS(sb);
2543 for (i = 0; i < num_meta_group_infos; i++)
2544 kfree(sbi->s_group_info[i]);
2545 ext4_kvfree(sbi->s_group_info);
2547 kfree(sbi->s_mb_offsets);
2548 kfree(sbi->s_mb_maxs);
2549 if (sbi->s_buddy_cache)
2550 iput(sbi->s_buddy_cache);
2551 if (sbi->s_mb_stats) {
2552 ext4_msg(sb, KERN_INFO,
2553 "mballoc: %u blocks %u reqs (%u success)",
2554 atomic_read(&sbi->s_bal_allocated),
2555 atomic_read(&sbi->s_bal_reqs),
2556 atomic_read(&sbi->s_bal_success));
2557 ext4_msg(sb, KERN_INFO,
2558 "mballoc: %u extents scanned, %u goal hits, "
2559 "%u 2^N hits, %u breaks, %u lost",
2560 atomic_read(&sbi->s_bal_ex_scanned),
2561 atomic_read(&sbi->s_bal_goals),
2562 atomic_read(&sbi->s_bal_2orders),
2563 atomic_read(&sbi->s_bal_breaks),
2564 atomic_read(&sbi->s_mb_lost_chunks));
2565 ext4_msg(sb, KERN_INFO,
2566 "mballoc: %lu generated and it took %Lu",
2567 sbi->s_mb_buddies_generated,
2568 sbi->s_mb_generation_time);
2569 ext4_msg(sb, KERN_INFO,
2570 "mballoc: %u preallocated, %u discarded",
2571 atomic_read(&sbi->s_mb_preallocated),
2572 atomic_read(&sbi->s_mb_discarded));
2575 free_percpu(sbi->s_locality_groups);
2580 static inline int ext4_issue_discard(struct super_block *sb,
2581 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2583 ext4_fsblk_t discard_block;
2585 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2586 ext4_group_first_block_no(sb, block_group));
2587 count = EXT4_C2B(EXT4_SB(sb), count);
2588 trace_ext4_discard_blocks(sb,
2589 (unsigned long long) discard_block, count);
2590 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2594 * This function is called by the jbd2 layer once the commit has finished,
2595 * so we know we can free the blocks that were released with that commit.
2597 static void ext4_free_data_callback(struct super_block *sb,
2598 struct ext4_journal_cb_entry *jce,
2601 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2602 struct ext4_buddy e4b;
2603 struct ext4_group_info *db;
2604 int err, count = 0, count2 = 0;
2606 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2607 entry->efd_count, entry->efd_group, entry);
2609 if (test_opt(sb, DISCARD))
2610 ext4_issue_discard(sb, entry->efd_group,
2611 entry->efd_start_cluster, entry->efd_count);
2613 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2614 /* we expect to find existing buddy because it's pinned */
2619 /* there are blocks to put in buddy to make them really free */
2620 count += entry->efd_count;
2622 ext4_lock_group(sb, entry->efd_group);
2623 /* Take it out of per group rb tree */
2624 rb_erase(&entry->efd_node, &(db->bb_free_root));
2625 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2628 * Clear the trimmed flag for the group so that the next
2629 * ext4_trim_fs can trim it.
2630 * If the volume is mounted with -o discard, online discard
2631 * is supported and the free blocks will be trimmed online.
2633 if (!test_opt(sb, DISCARD))
2634 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2636 if (!db->bb_free_root.rb_node) {
2637 /* No more items in the per group rb tree
2638 * balance refcounts from ext4_mb_free_metadata()
2640 page_cache_release(e4b.bd_buddy_page);
2641 page_cache_release(e4b.bd_bitmap_page);
2643 ext4_unlock_group(sb, entry->efd_group);
2644 kmem_cache_free(ext4_free_data_cachep, entry);
2645 ext4_mb_unload_buddy(&e4b);
2647 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2650 #ifdef CONFIG_EXT4_DEBUG
2651 u8 mb_enable_debug __read_mostly;
2653 static struct dentry *debugfs_dir;
2654 static struct dentry *debugfs_debug;
2656 static void __init ext4_create_debugfs_entry(void)
2658 debugfs_dir = debugfs_create_dir("ext4", NULL);
2660 debugfs_debug = debugfs_create_u8("mballoc-debug",
2666 static void ext4_remove_debugfs_entry(void)
2668 debugfs_remove(debugfs_debug);
2669 debugfs_remove(debugfs_dir);
2674 static void __init ext4_create_debugfs_entry(void)
2678 static void ext4_remove_debugfs_entry(void)
2684 int __init ext4_init_mballoc(void)
2686 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2687 SLAB_RECLAIM_ACCOUNT);
2688 if (ext4_pspace_cachep == NULL)
2691 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2692 SLAB_RECLAIM_ACCOUNT);
2693 if (ext4_ac_cachep == NULL) {
2694 kmem_cache_destroy(ext4_pspace_cachep);
2698 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2699 SLAB_RECLAIM_ACCOUNT);
2700 if (ext4_free_data_cachep == NULL) {
2701 kmem_cache_destroy(ext4_pspace_cachep);
2702 kmem_cache_destroy(ext4_ac_cachep);
2705 ext4_create_debugfs_entry();
2709 void ext4_exit_mballoc(void)
2712 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2713 * before destroying the slab cache.
2716 kmem_cache_destroy(ext4_pspace_cachep);
2717 kmem_cache_destroy(ext4_ac_cachep);
2718 kmem_cache_destroy(ext4_free_data_cachep);
2719 ext4_groupinfo_destroy_slabs();
2720 ext4_remove_debugfs_entry();
2725 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2726 * Returns 0 if success or error code
2728 static noinline_for_stack int
2729 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2730 handle_t *handle, unsigned int reserv_clstrs)
2732 struct buffer_head *bitmap_bh = NULL;
2733 struct ext4_group_desc *gdp;
2734 struct buffer_head *gdp_bh;
2735 struct ext4_sb_info *sbi;
2736 struct super_block *sb;
2740 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2741 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2747 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2751 err = ext4_journal_get_write_access(handle, bitmap_bh);
2756 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2760 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2761 ext4_free_group_clusters(sb, gdp));
2763 err = ext4_journal_get_write_access(handle, gdp_bh);
2767 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2769 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2770 if (!ext4_data_block_valid(sbi, block, len)) {
2771 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2772 "fs metadata", block, block+len);
2773 /* File system mounted not to panic on error
2774 * Fix the bitmap and repeat the block allocation
2775 * We leak some of the blocks here.
2777 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2778 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2779 ac->ac_b_ex.fe_len);
2780 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2781 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2787 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2788 #ifdef AGGRESSIVE_CHECK
2791 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2792 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2793 bitmap_bh->b_data));
2797 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2798 ac->ac_b_ex.fe_len);
2799 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2800 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2801 ext4_free_group_clusters_set(sb, gdp,
2802 ext4_free_clusters_after_init(sb,
2803 ac->ac_b_ex.fe_group, gdp));
2805 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2806 ext4_free_group_clusters_set(sb, gdp, len);
2807 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh,
2808 EXT4_BLOCKS_PER_GROUP(sb) / 8);
2809 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2811 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2812 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2814 * Now reduce the dirty block count also. Should not go negative
2816 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2817 /* release all the reserved blocks if non delalloc */
2818 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2821 if (sbi->s_log_groups_per_flex) {
2822 ext4_group_t flex_group = ext4_flex_group(sbi,
2823 ac->ac_b_ex.fe_group);
2824 atomic_sub(ac->ac_b_ex.fe_len,
2825 &sbi->s_flex_groups[flex_group].free_clusters);
2828 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2831 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2834 ext4_mark_super_dirty(sb);
2840 * here we normalize request for locality group
2841 * Group request are normalized to s_mb_group_prealloc, which goes to
2842 * s_strip if we set the same via mount option.
2843 * s_mb_group_prealloc can be configured via
2844 * /sys/fs/ext4/<partition>/mb_group_prealloc
2846 * XXX: should we try to preallocate more than the group has now?
2848 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2850 struct super_block *sb = ac->ac_sb;
2851 struct ext4_locality_group *lg = ac->ac_lg;
2854 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2855 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2856 current->pid, ac->ac_g_ex.fe_len);
2860 * Normalization means making request better in terms of
2861 * size and alignment
2863 static noinline_for_stack void
2864 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2865 struct ext4_allocation_request *ar)
2867 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2870 loff_t size, start_off;
2871 loff_t orig_size __maybe_unused;
2873 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2874 struct ext4_prealloc_space *pa;
2876 /* do normalize only data requests, metadata requests
2877 do not need preallocation */
2878 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2881 /* sometime caller may want exact blocks */
2882 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2885 /* caller may indicate that preallocation isn't
2886 * required (it's a tail, for example) */
2887 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2890 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2891 ext4_mb_normalize_group_request(ac);
2895 bsbits = ac->ac_sb->s_blocksize_bits;
2897 /* first, let's learn actual file size
2898 * given current request is allocated */
2899 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
2900 size = size << bsbits;
2901 if (size < i_size_read(ac->ac_inode))
2902 size = i_size_read(ac->ac_inode);
2905 /* max size of free chunks */
2908 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2909 (req <= (size) || max <= (chunk_size))
2911 /* first, try to predict filesize */
2912 /* XXX: should this table be tunable? */
2914 if (size <= 16 * 1024) {
2916 } else if (size <= 32 * 1024) {
2918 } else if (size <= 64 * 1024) {
2920 } else if (size <= 128 * 1024) {
2922 } else if (size <= 256 * 1024) {
2924 } else if (size <= 512 * 1024) {
2926 } else if (size <= 1024 * 1024) {
2928 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2929 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2930 (21 - bsbits)) << 21;
2931 size = 2 * 1024 * 1024;
2932 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2933 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2934 (22 - bsbits)) << 22;
2935 size = 4 * 1024 * 1024;
2936 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2937 (8<<20)>>bsbits, max, 8 * 1024)) {
2938 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2939 (23 - bsbits)) << 23;
2940 size = 8 * 1024 * 1024;
2942 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2943 size = ac->ac_o_ex.fe_len << bsbits;
2945 size = size >> bsbits;
2946 start = start_off >> bsbits;
2948 /* don't cover already allocated blocks in selected range */
2949 if (ar->pleft && start <= ar->lleft) {
2950 size -= ar->lleft + 1 - start;
2951 start = ar->lleft + 1;
2953 if (ar->pright && start + size - 1 >= ar->lright)
2954 size -= start + size - ar->lright;
2958 /* check we don't cross already preallocated blocks */
2960 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2965 spin_lock(&pa->pa_lock);
2966 if (pa->pa_deleted) {
2967 spin_unlock(&pa->pa_lock);
2971 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
2974 /* PA must not overlap original request */
2975 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2976 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2978 /* skip PAs this normalized request doesn't overlap with */
2979 if (pa->pa_lstart >= end || pa_end <= start) {
2980 spin_unlock(&pa->pa_lock);
2983 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2985 /* adjust start or end to be adjacent to this pa */
2986 if (pa_end <= ac->ac_o_ex.fe_logical) {
2987 BUG_ON(pa_end < start);
2989 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
2990 BUG_ON(pa->pa_lstart > end);
2991 end = pa->pa_lstart;
2993 spin_unlock(&pa->pa_lock);
2998 /* XXX: extra loop to check we really don't overlap preallocations */
3000 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3003 spin_lock(&pa->pa_lock);
3004 if (pa->pa_deleted == 0) {
3005 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3007 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3009 spin_unlock(&pa->pa_lock);
3013 if (start + size <= ac->ac_o_ex.fe_logical &&
3014 start > ac->ac_o_ex.fe_logical) {
3015 ext4_msg(ac->ac_sb, KERN_ERR,
3016 "start %lu, size %lu, fe_logical %lu",
3017 (unsigned long) start, (unsigned long) size,
3018 (unsigned long) ac->ac_o_ex.fe_logical);
3020 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3021 start > ac->ac_o_ex.fe_logical);
3022 BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
3024 /* now prepare goal request */
3026 /* XXX: is it better to align blocks WRT to logical
3027 * placement or satisfy big request as is */
3028 ac->ac_g_ex.fe_logical = start;
3029 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3031 /* define goal start in order to merge */
3032 if (ar->pright && (ar->lright == (start + size))) {
3033 /* merge to the right */
3034 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3035 &ac->ac_f_ex.fe_group,
3036 &ac->ac_f_ex.fe_start);
3037 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3039 if (ar->pleft && (ar->lleft + 1 == start)) {
3040 /* merge to the left */
3041 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3042 &ac->ac_f_ex.fe_group,
3043 &ac->ac_f_ex.fe_start);
3044 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3047 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3048 (unsigned) orig_size, (unsigned) start);
3051 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3053 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3055 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3056 atomic_inc(&sbi->s_bal_reqs);
3057 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3058 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3059 atomic_inc(&sbi->s_bal_success);
3060 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3061 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3062 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3063 atomic_inc(&sbi->s_bal_goals);
3064 if (ac->ac_found > sbi->s_mb_max_to_scan)
3065 atomic_inc(&sbi->s_bal_breaks);
3068 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3069 trace_ext4_mballoc_alloc(ac);
3071 trace_ext4_mballoc_prealloc(ac);
3075 * Called on failure; free up any blocks from the inode PA for this
3076 * context. We don't need this for MB_GROUP_PA because we only change
3077 * pa_free in ext4_mb_release_context(), but on failure, we've already
3078 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3080 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3082 struct ext4_prealloc_space *pa = ac->ac_pa;
3084 if (pa && pa->pa_type == MB_INODE_PA)
3085 pa->pa_free += ac->ac_b_ex.fe_len;
3089 * use blocks preallocated to inode
3091 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3092 struct ext4_prealloc_space *pa)
3094 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3099 /* found preallocated blocks, use them */
3100 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3101 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3102 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3103 len = EXT4_NUM_B2C(sbi, end - start);
3104 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3105 &ac->ac_b_ex.fe_start);
3106 ac->ac_b_ex.fe_len = len;
3107 ac->ac_status = AC_STATUS_FOUND;
3110 BUG_ON(start < pa->pa_pstart);
3111 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3112 BUG_ON(pa->pa_free < len);
3115 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3119 * use blocks preallocated to locality group
3121 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3122 struct ext4_prealloc_space *pa)
3124 unsigned int len = ac->ac_o_ex.fe_len;
3126 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3127 &ac->ac_b_ex.fe_group,
3128 &ac->ac_b_ex.fe_start);
3129 ac->ac_b_ex.fe_len = len;
3130 ac->ac_status = AC_STATUS_FOUND;
3133 /* we don't correct pa_pstart or pa_plen here to avoid
3134 * possible race when the group is being loaded concurrently
3135 * instead we correct pa later, after blocks are marked
3136 * in on-disk bitmap -- see ext4_mb_release_context()
3137 * Other CPUs are prevented from allocating from this pa by lg_mutex
3139 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3143 * Return the prealloc space that have minimal distance
3144 * from the goal block. @cpa is the prealloc
3145 * space that is having currently known minimal distance
3146 * from the goal block.
3148 static struct ext4_prealloc_space *
3149 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3150 struct ext4_prealloc_space *pa,
3151 struct ext4_prealloc_space *cpa)
3153 ext4_fsblk_t cur_distance, new_distance;
3156 atomic_inc(&pa->pa_count);
3159 cur_distance = abs(goal_block - cpa->pa_pstart);
3160 new_distance = abs(goal_block - pa->pa_pstart);
3162 if (cur_distance <= new_distance)
3165 /* drop the previous reference */
3166 atomic_dec(&cpa->pa_count);
3167 atomic_inc(&pa->pa_count);
3172 * search goal blocks in preallocated space
3174 static noinline_for_stack int
3175 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3177 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3179 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3180 struct ext4_locality_group *lg;
3181 struct ext4_prealloc_space *pa, *cpa = NULL;
3182 ext4_fsblk_t goal_block;
3184 /* only data can be preallocated */
3185 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3188 /* first, try per-file preallocation */
3190 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3192 /* all fields in this condition don't change,
3193 * so we can skip locking for them */
3194 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3195 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3196 EXT4_C2B(sbi, pa->pa_len)))
3199 /* non-extent files can't have physical blocks past 2^32 */
3200 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3201 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3202 EXT4_MAX_BLOCK_FILE_PHYS))
3205 /* found preallocated blocks, use them */
3206 spin_lock(&pa->pa_lock);
3207 if (pa->pa_deleted == 0 && pa->pa_free) {
3208 atomic_inc(&pa->pa_count);
3209 ext4_mb_use_inode_pa(ac, pa);
3210 spin_unlock(&pa->pa_lock);
3211 ac->ac_criteria = 10;
3215 spin_unlock(&pa->pa_lock);
3219 /* can we use group allocation? */
3220 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3223 /* inode may have no locality group for some reason */
3227 order = fls(ac->ac_o_ex.fe_len) - 1;
3228 if (order > PREALLOC_TB_SIZE - 1)
3229 /* The max size of hash table is PREALLOC_TB_SIZE */
3230 order = PREALLOC_TB_SIZE - 1;
3232 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3234 * search for the prealloc space that is having
3235 * minimal distance from the goal block.
3237 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3239 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3241 spin_lock(&pa->pa_lock);
3242 if (pa->pa_deleted == 0 &&
3243 pa->pa_free >= ac->ac_o_ex.fe_len) {
3245 cpa = ext4_mb_check_group_pa(goal_block,
3248 spin_unlock(&pa->pa_lock);
3253 ext4_mb_use_group_pa(ac, cpa);
3254 ac->ac_criteria = 20;
3261 * the function goes through all block freed in the group
3262 * but not yet committed and marks them used in in-core bitmap.
3263 * buddy must be generated from this bitmap
3264 * Need to be called with the ext4 group lock held
3266 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3270 struct ext4_group_info *grp;
3271 struct ext4_free_data *entry;
3273 grp = ext4_get_group_info(sb, group);
3274 n = rb_first(&(grp->bb_free_root));
3277 entry = rb_entry(n, struct ext4_free_data, efd_node);
3278 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3285 * the function goes through all preallocation in this group and marks them
3286 * used in in-core bitmap. buddy must be generated from this bitmap
3287 * Need to be called with ext4 group lock held
3289 static noinline_for_stack
3290 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3293 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3294 struct ext4_prealloc_space *pa;
3295 struct list_head *cur;
3296 ext4_group_t groupnr;
3297 ext4_grpblk_t start;
3298 int preallocated = 0;
3301 /* all form of preallocation discards first load group,
3302 * so the only competing code is preallocation use.
3303 * we don't need any locking here
3304 * notice we do NOT ignore preallocations with pa_deleted
3305 * otherwise we could leave used blocks available for
3306 * allocation in buddy when concurrent ext4_mb_put_pa()
3307 * is dropping preallocation
3309 list_for_each(cur, &grp->bb_prealloc_list) {
3310 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3311 spin_lock(&pa->pa_lock);
3312 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3315 spin_unlock(&pa->pa_lock);
3316 if (unlikely(len == 0))
3318 BUG_ON(groupnr != group);
3319 ext4_set_bits(bitmap, start, len);
3320 preallocated += len;
3322 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3325 static void ext4_mb_pa_callback(struct rcu_head *head)
3327 struct ext4_prealloc_space *pa;
3328 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3329 kmem_cache_free(ext4_pspace_cachep, pa);
3333 * drops a reference to preallocated space descriptor
3334 * if this was the last reference and the space is consumed
3336 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3337 struct super_block *sb, struct ext4_prealloc_space *pa)
3340 ext4_fsblk_t grp_blk;
3342 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3345 /* in this short window concurrent discard can set pa_deleted */
3346 spin_lock(&pa->pa_lock);
3347 if (pa->pa_deleted == 1) {
3348 spin_unlock(&pa->pa_lock);
3353 spin_unlock(&pa->pa_lock);
3355 grp_blk = pa->pa_pstart;
3357 * If doing group-based preallocation, pa_pstart may be in the
3358 * next group when pa is used up
3360 if (pa->pa_type == MB_GROUP_PA)
3363 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3368 * P1 (buddy init) P2 (regular allocation)
3369 * find block B in PA
3370 * copy on-disk bitmap to buddy
3371 * mark B in on-disk bitmap
3372 * drop PA from group
3373 * mark all PAs in buddy
3375 * thus, P1 initializes buddy with B available. to prevent this
3376 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3379 ext4_lock_group(sb, grp);
3380 list_del(&pa->pa_group_list);
3381 ext4_unlock_group(sb, grp);
3383 spin_lock(pa->pa_obj_lock);
3384 list_del_rcu(&pa->pa_inode_list);
3385 spin_unlock(pa->pa_obj_lock);
3387 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3391 * creates new preallocated space for given inode
3393 static noinline_for_stack int
3394 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3396 struct super_block *sb = ac->ac_sb;
3397 struct ext4_sb_info *sbi = EXT4_SB(sb);
3398 struct ext4_prealloc_space *pa;
3399 struct ext4_group_info *grp;
3400 struct ext4_inode_info *ei;
3402 /* preallocate only when found space is larger then requested */
3403 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3404 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3405 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3407 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3411 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3417 /* we can't allocate as much as normalizer wants.
3418 * so, found space must get proper lstart
3419 * to cover original request */
3420 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3421 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3423 /* we're limited by original request in that
3424 * logical block must be covered any way
3425 * winl is window we can move our chunk within */
3426 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3428 /* also, we should cover whole original request */
3429 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3431 /* the smallest one defines real window */
3432 win = min(winl, wins);
3434 offs = ac->ac_o_ex.fe_logical %
3435 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3436 if (offs && offs < win)
3439 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3441 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3442 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3445 /* preallocation can change ac_b_ex, thus we store actually
3446 * allocated blocks for history */
3447 ac->ac_f_ex = ac->ac_b_ex;
3449 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3450 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3451 pa->pa_len = ac->ac_b_ex.fe_len;
3452 pa->pa_free = pa->pa_len;
3453 atomic_set(&pa->pa_count, 1);
3454 spin_lock_init(&pa->pa_lock);
3455 INIT_LIST_HEAD(&pa->pa_inode_list);
3456 INIT_LIST_HEAD(&pa->pa_group_list);
3458 pa->pa_type = MB_INODE_PA;
3460 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3461 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3462 trace_ext4_mb_new_inode_pa(ac, pa);
3464 ext4_mb_use_inode_pa(ac, pa);
3465 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3467 ei = EXT4_I(ac->ac_inode);
3468 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3470 pa->pa_obj_lock = &ei->i_prealloc_lock;
3471 pa->pa_inode = ac->ac_inode;
3473 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3474 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3475 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3477 spin_lock(pa->pa_obj_lock);
3478 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3479 spin_unlock(pa->pa_obj_lock);
3485 * creates new preallocated space for locality group inodes belongs to
3487 static noinline_for_stack int
3488 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3490 struct super_block *sb = ac->ac_sb;
3491 struct ext4_locality_group *lg;
3492 struct ext4_prealloc_space *pa;
3493 struct ext4_group_info *grp;
3495 /* preallocate only when found space is larger then requested */
3496 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3497 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3498 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3500 BUG_ON(ext4_pspace_cachep == NULL);
3501 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3505 /* preallocation can change ac_b_ex, thus we store actually
3506 * allocated blocks for history */
3507 ac->ac_f_ex = ac->ac_b_ex;
3509 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3510 pa->pa_lstart = pa->pa_pstart;
3511 pa->pa_len = ac->ac_b_ex.fe_len;
3512 pa->pa_free = pa->pa_len;
3513 atomic_set(&pa->pa_count, 1);
3514 spin_lock_init(&pa->pa_lock);
3515 INIT_LIST_HEAD(&pa->pa_inode_list);
3516 INIT_LIST_HEAD(&pa->pa_group_list);
3518 pa->pa_type = MB_GROUP_PA;
3520 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3521 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3522 trace_ext4_mb_new_group_pa(ac, pa);
3524 ext4_mb_use_group_pa(ac, pa);
3525 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3527 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3531 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3532 pa->pa_inode = NULL;
3534 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3535 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3536 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3539 * We will later add the new pa to the right bucket
3540 * after updating the pa_free in ext4_mb_release_context
3545 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3549 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3550 err = ext4_mb_new_group_pa(ac);
3552 err = ext4_mb_new_inode_pa(ac);
3557 * finds all unused blocks in on-disk bitmap, frees them in
3558 * in-core bitmap and buddy.
3559 * @pa must be unlinked from inode and group lists, so that
3560 * nobody else can find/use it.
3561 * the caller MUST hold group/inode locks.
3562 * TODO: optimize the case when there are no in-core structures yet
3564 static noinline_for_stack int
3565 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3566 struct ext4_prealloc_space *pa)
3568 struct super_block *sb = e4b->bd_sb;
3569 struct ext4_sb_info *sbi = EXT4_SB(sb);
3574 unsigned long long grp_blk_start;
3578 BUG_ON(pa->pa_deleted == 0);
3579 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3580 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3581 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3582 end = bit + pa->pa_len;
3585 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3588 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3589 mb_debug(1, " free preallocated %u/%u in group %u\n",
3590 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3591 (unsigned) next - bit, (unsigned) group);
3594 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3595 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3596 EXT4_C2B(sbi, bit)),
3598 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3601 if (free != pa->pa_free) {
3602 ext4_msg(e4b->bd_sb, KERN_CRIT,
3603 "pa %p: logic %lu, phys. %lu, len %lu",
3604 pa, (unsigned long) pa->pa_lstart,
3605 (unsigned long) pa->pa_pstart,
3606 (unsigned long) pa->pa_len);
3607 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3610 * pa is already deleted so we use the value obtained
3611 * from the bitmap and continue.
3614 atomic_add(free, &sbi->s_mb_discarded);
3619 static noinline_for_stack int
3620 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3621 struct ext4_prealloc_space *pa)
3623 struct super_block *sb = e4b->bd_sb;
3627 trace_ext4_mb_release_group_pa(sb, pa);
3628 BUG_ON(pa->pa_deleted == 0);
3629 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3630 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3631 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3632 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3633 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3639 * releases all preallocations in given group
3641 * first, we need to decide discard policy:
3642 * - when do we discard
3644 * - how many do we discard
3645 * 1) how many requested
3647 static noinline_for_stack int
3648 ext4_mb_discard_group_preallocations(struct super_block *sb,
3649 ext4_group_t group, int needed)
3651 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3652 struct buffer_head *bitmap_bh = NULL;
3653 struct ext4_prealloc_space *pa, *tmp;
3654 struct list_head list;
3655 struct ext4_buddy e4b;
3660 mb_debug(1, "discard preallocation for group %u\n", group);
3662 if (list_empty(&grp->bb_prealloc_list))
3665 bitmap_bh = ext4_read_block_bitmap(sb, group);
3666 if (bitmap_bh == NULL) {
3667 ext4_error(sb, "Error reading block bitmap for %u", group);
3671 err = ext4_mb_load_buddy(sb, group, &e4b);
3673 ext4_error(sb, "Error loading buddy information for %u", group);
3679 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3681 INIT_LIST_HEAD(&list);
3683 ext4_lock_group(sb, group);
3684 list_for_each_entry_safe(pa, tmp,
3685 &grp->bb_prealloc_list, pa_group_list) {
3686 spin_lock(&pa->pa_lock);
3687 if (atomic_read(&pa->pa_count)) {
3688 spin_unlock(&pa->pa_lock);
3692 if (pa->pa_deleted) {
3693 spin_unlock(&pa->pa_lock);
3697 /* seems this one can be freed ... */
3700 /* we can trust pa_free ... */
3701 free += pa->pa_free;
3703 spin_unlock(&pa->pa_lock);
3705 list_del(&pa->pa_group_list);
3706 list_add(&pa->u.pa_tmp_list, &list);
3709 /* if we still need more blocks and some PAs were used, try again */
3710 if (free < needed && busy) {
3712 ext4_unlock_group(sb, group);
3714 * Yield the CPU here so that we don't get soft lockup
3715 * in non preempt case.
3721 /* found anything to free? */
3722 if (list_empty(&list)) {
3727 /* now free all selected PAs */
3728 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3730 /* remove from object (inode or locality group) */
3731 spin_lock(pa->pa_obj_lock);
3732 list_del_rcu(&pa->pa_inode_list);
3733 spin_unlock(pa->pa_obj_lock);
3735 if (pa->pa_type == MB_GROUP_PA)
3736 ext4_mb_release_group_pa(&e4b, pa);
3738 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3740 list_del(&pa->u.pa_tmp_list);
3741 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3745 ext4_unlock_group(sb, group);
3746 ext4_mb_unload_buddy(&e4b);
3752 * releases all non-used preallocated blocks for given inode
3754 * It's important to discard preallocations under i_data_sem
3755 * We don't want another block to be served from the prealloc
3756 * space when we are discarding the inode prealloc space.
3758 * FIXME!! Make sure it is valid at all the call sites
3760 void ext4_discard_preallocations(struct inode *inode)
3762 struct ext4_inode_info *ei = EXT4_I(inode);
3763 struct super_block *sb = inode->i_sb;
3764 struct buffer_head *bitmap_bh = NULL;
3765 struct ext4_prealloc_space *pa, *tmp;
3766 ext4_group_t group = 0;
3767 struct list_head list;
3768 struct ext4_buddy e4b;
3771 if (!S_ISREG(inode->i_mode)) {
3772 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3776 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3777 trace_ext4_discard_preallocations(inode);
3779 INIT_LIST_HEAD(&list);
3782 /* first, collect all pa's in the inode */
3783 spin_lock(&ei->i_prealloc_lock);
3784 while (!list_empty(&ei->i_prealloc_list)) {
3785 pa = list_entry(ei->i_prealloc_list.next,
3786 struct ext4_prealloc_space, pa_inode_list);
3787 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3788 spin_lock(&pa->pa_lock);
3789 if (atomic_read(&pa->pa_count)) {
3790 /* this shouldn't happen often - nobody should
3791 * use preallocation while we're discarding it */
3792 spin_unlock(&pa->pa_lock);
3793 spin_unlock(&ei->i_prealloc_lock);
3794 ext4_msg(sb, KERN_ERR,
3795 "uh-oh! used pa while discarding");
3797 schedule_timeout_uninterruptible(HZ);
3801 if (pa->pa_deleted == 0) {
3803 spin_unlock(&pa->pa_lock);
3804 list_del_rcu(&pa->pa_inode_list);
3805 list_add(&pa->u.pa_tmp_list, &list);
3809 /* someone is deleting pa right now */
3810 spin_unlock(&pa->pa_lock);
3811 spin_unlock(&ei->i_prealloc_lock);
3813 /* we have to wait here because pa_deleted
3814 * doesn't mean pa is already unlinked from
3815 * the list. as we might be called from
3816 * ->clear_inode() the inode will get freed
3817 * and concurrent thread which is unlinking
3818 * pa from inode's list may access already
3819 * freed memory, bad-bad-bad */
3821 /* XXX: if this happens too often, we can
3822 * add a flag to force wait only in case
3823 * of ->clear_inode(), but not in case of
3824 * regular truncate */
3825 schedule_timeout_uninterruptible(HZ);
3828 spin_unlock(&ei->i_prealloc_lock);
3830 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3831 BUG_ON(pa->pa_type != MB_INODE_PA);
3832 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3834 err = ext4_mb_load_buddy(sb, group, &e4b);
3836 ext4_error(sb, "Error loading buddy information for %u",
3841 bitmap_bh = ext4_read_block_bitmap(sb, group);
3842 if (bitmap_bh == NULL) {
3843 ext4_error(sb, "Error reading block bitmap for %u",
3845 ext4_mb_unload_buddy(&e4b);
3849 ext4_lock_group(sb, group);
3850 list_del(&pa->pa_group_list);
3851 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3852 ext4_unlock_group(sb, group);
3854 ext4_mb_unload_buddy(&e4b);
3857 list_del(&pa->u.pa_tmp_list);
3858 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3862 #ifdef CONFIG_EXT4_DEBUG
3863 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3865 struct super_block *sb = ac->ac_sb;
3866 ext4_group_t ngroups, i;
3868 if (!mb_enable_debug ||
3869 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3872 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
3873 " Allocation context details:");
3874 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
3875 ac->ac_status, ac->ac_flags);
3876 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
3877 "goal %lu/%lu/%lu@%lu, "
3878 "best %lu/%lu/%lu@%lu cr %d",
3879 (unsigned long)ac->ac_o_ex.fe_group,
3880 (unsigned long)ac->ac_o_ex.fe_start,
3881 (unsigned long)ac->ac_o_ex.fe_len,
3882 (unsigned long)ac->ac_o_ex.fe_logical,
3883 (unsigned long)ac->ac_g_ex.fe_group,
3884 (unsigned long)ac->ac_g_ex.fe_start,
3885 (unsigned long)ac->ac_g_ex.fe_len,
3886 (unsigned long)ac->ac_g_ex.fe_logical,
3887 (unsigned long)ac->ac_b_ex.fe_group,
3888 (unsigned long)ac->ac_b_ex.fe_start,
3889 (unsigned long)ac->ac_b_ex.fe_len,
3890 (unsigned long)ac->ac_b_ex.fe_logical,
3891 (int)ac->ac_criteria);
3892 ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found",
3893 ac->ac_ex_scanned, ac->ac_found);
3894 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
3895 ngroups = ext4_get_groups_count(sb);
3896 for (i = 0; i < ngroups; i++) {
3897 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3898 struct ext4_prealloc_space *pa;
3899 ext4_grpblk_t start;
3900 struct list_head *cur;
3901 ext4_lock_group(sb, i);
3902 list_for_each(cur, &grp->bb_prealloc_list) {
3903 pa = list_entry(cur, struct ext4_prealloc_space,
3905 spin_lock(&pa->pa_lock);
3906 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3908 spin_unlock(&pa->pa_lock);
3909 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3912 ext4_unlock_group(sb, i);
3914 if (grp->bb_free == 0)
3916 printk(KERN_ERR "%u: %d/%d \n",
3917 i, grp->bb_free, grp->bb_fragments);
3919 printk(KERN_ERR "\n");
3922 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3929 * We use locality group preallocation for small size file. The size of the
3930 * file is determined by the current size or the resulting size after
3931 * allocation which ever is larger
3933 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3935 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3937 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3938 int bsbits = ac->ac_sb->s_blocksize_bits;
3941 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3944 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3947 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3948 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3951 if ((size == isize) &&
3952 !ext4_fs_is_busy(sbi) &&
3953 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3954 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3958 if (sbi->s_mb_group_prealloc <= 0) {
3959 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3963 /* don't use group allocation for large files */
3964 size = max(size, isize);
3965 if (size > sbi->s_mb_stream_request) {
3966 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3970 BUG_ON(ac->ac_lg != NULL);
3972 * locality group prealloc space are per cpu. The reason for having
3973 * per cpu locality group is to reduce the contention between block
3974 * request from multiple CPUs.
3976 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
3978 /* we're going to use group allocation */
3979 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3981 /* serialize all allocations in the group */
3982 mutex_lock(&ac->ac_lg->lg_mutex);
3985 static noinline_for_stack int
3986 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
3987 struct ext4_allocation_request *ar)
3989 struct super_block *sb = ar->inode->i_sb;
3990 struct ext4_sb_info *sbi = EXT4_SB(sb);
3991 struct ext4_super_block *es = sbi->s_es;
3995 ext4_grpblk_t block;
3997 /* we can't allocate > group size */
4000 /* just a dirty hack to filter too big requests */
4001 if (len >= EXT4_CLUSTERS_PER_GROUP(sb) - 10)
4002 len = EXT4_CLUSTERS_PER_GROUP(sb) - 10;
4004 /* start searching from the goal */
4006 if (goal < le32_to_cpu(es->s_first_data_block) ||
4007 goal >= ext4_blocks_count(es))
4008 goal = le32_to_cpu(es->s_first_data_block);
4009 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4011 /* set up allocation goals */
4012 memset(ac, 0, sizeof(struct ext4_allocation_context));
4013 ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1);
4014 ac->ac_status = AC_STATUS_CONTINUE;
4016 ac->ac_inode = ar->inode;
4017 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4018 ac->ac_o_ex.fe_group = group;
4019 ac->ac_o_ex.fe_start = block;
4020 ac->ac_o_ex.fe_len = len;
4021 ac->ac_g_ex = ac->ac_o_ex;
4022 ac->ac_flags = ar->flags;
4024 /* we have to define context: we'll we work with a file or
4025 * locality group. this is a policy, actually */
4026 ext4_mb_group_or_file(ac);
4028 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4029 "left: %u/%u, right %u/%u to %swritable\n",
4030 (unsigned) ar->len, (unsigned) ar->logical,
4031 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4032 (unsigned) ar->lleft, (unsigned) ar->pleft,
4033 (unsigned) ar->lright, (unsigned) ar->pright,
4034 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4039 static noinline_for_stack void
4040 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4041 struct ext4_locality_group *lg,
4042 int order, int total_entries)
4044 ext4_group_t group = 0;
4045 struct ext4_buddy e4b;
4046 struct list_head discard_list;
4047 struct ext4_prealloc_space *pa, *tmp;
4049 mb_debug(1, "discard locality group preallocation\n");
4051 INIT_LIST_HEAD(&discard_list);
4053 spin_lock(&lg->lg_prealloc_lock);
4054 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4056 spin_lock(&pa->pa_lock);
4057 if (atomic_read(&pa->pa_count)) {
4059 * This is the pa that we just used
4060 * for block allocation. So don't
4063 spin_unlock(&pa->pa_lock);
4066 if (pa->pa_deleted) {
4067 spin_unlock(&pa->pa_lock);
4070 /* only lg prealloc space */
4071 BUG_ON(pa->pa_type != MB_GROUP_PA);
4073 /* seems this one can be freed ... */
4075 spin_unlock(&pa->pa_lock);
4077 list_del_rcu(&pa->pa_inode_list);
4078 list_add(&pa->u.pa_tmp_list, &discard_list);
4081 if (total_entries <= 5) {
4083 * we want to keep only 5 entries
4084 * allowing it to grow to 8. This
4085 * mak sure we don't call discard
4086 * soon for this list.
4091 spin_unlock(&lg->lg_prealloc_lock);
4093 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4095 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4096 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4097 ext4_error(sb, "Error loading buddy information for %u",
4101 ext4_lock_group(sb, group);
4102 list_del(&pa->pa_group_list);
4103 ext4_mb_release_group_pa(&e4b, pa);
4104 ext4_unlock_group(sb, group);
4106 ext4_mb_unload_buddy(&e4b);
4107 list_del(&pa->u.pa_tmp_list);
4108 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4113 * We have incremented pa_count. So it cannot be freed at this
4114 * point. Also we hold lg_mutex. So no parallel allocation is
4115 * possible from this lg. That means pa_free cannot be updated.
4117 * A parallel ext4_mb_discard_group_preallocations is possible.
4118 * which can cause the lg_prealloc_list to be updated.
4121 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4123 int order, added = 0, lg_prealloc_count = 1;
4124 struct super_block *sb = ac->ac_sb;
4125 struct ext4_locality_group *lg = ac->ac_lg;
4126 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4128 order = fls(pa->pa_free) - 1;
4129 if (order > PREALLOC_TB_SIZE - 1)
4130 /* The max size of hash table is PREALLOC_TB_SIZE */
4131 order = PREALLOC_TB_SIZE - 1;
4132 /* Add the prealloc space to lg */
4134 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4136 spin_lock(&tmp_pa->pa_lock);
4137 if (tmp_pa->pa_deleted) {
4138 spin_unlock(&tmp_pa->pa_lock);
4141 if (!added && pa->pa_free < tmp_pa->pa_free) {
4142 /* Add to the tail of the previous entry */
4143 list_add_tail_rcu(&pa->pa_inode_list,
4144 &tmp_pa->pa_inode_list);
4147 * we want to count the total
4148 * number of entries in the list
4151 spin_unlock(&tmp_pa->pa_lock);
4152 lg_prealloc_count++;
4155 list_add_tail_rcu(&pa->pa_inode_list,
4156 &lg->lg_prealloc_list[order]);
4159 /* Now trim the list to be not more than 8 elements */
4160 if (lg_prealloc_count > 8) {
4161 ext4_mb_discard_lg_preallocations(sb, lg,
4162 order, lg_prealloc_count);
4169 * release all resource we used in allocation
4171 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4173 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4174 struct ext4_prealloc_space *pa = ac->ac_pa;
4176 if (pa->pa_type == MB_GROUP_PA) {
4177 /* see comment in ext4_mb_use_group_pa() */
4178 spin_lock(&pa->pa_lock);
4179 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4180 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4181 pa->pa_free -= ac->ac_b_ex.fe_len;
4182 pa->pa_len -= ac->ac_b_ex.fe_len;
4183 spin_unlock(&pa->pa_lock);
4188 * We want to add the pa to the right bucket.
4189 * Remove it from the list and while adding
4190 * make sure the list to which we are adding
4193 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4194 spin_lock(pa->pa_obj_lock);
4195 list_del_rcu(&pa->pa_inode_list);
4196 spin_unlock(pa->pa_obj_lock);
4197 ext4_mb_add_n_trim(ac);
4199 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4201 if (ac->ac_bitmap_page)
4202 page_cache_release(ac->ac_bitmap_page);
4203 if (ac->ac_buddy_page)
4204 page_cache_release(ac->ac_buddy_page);
4205 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4206 mutex_unlock(&ac->ac_lg->lg_mutex);
4207 ext4_mb_collect_stats(ac);
4211 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4213 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4217 trace_ext4_mb_discard_preallocations(sb, needed);
4218 for (i = 0; i < ngroups && needed > 0; i++) {
4219 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4228 * Main entry point into mballoc to allocate blocks
4229 * it tries to use preallocation first, then falls back
4230 * to usual allocation
4232 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4233 struct ext4_allocation_request *ar, int *errp)
4236 struct ext4_allocation_context *ac = NULL;
4237 struct ext4_sb_info *sbi;
4238 struct super_block *sb;
4239 ext4_fsblk_t block = 0;
4240 unsigned int inquota = 0;
4241 unsigned int reserv_clstrs = 0;
4243 sb = ar->inode->i_sb;
4246 trace_ext4_request_blocks(ar);
4248 /* Allow to use superuser reservation for quota file */
4249 if (IS_NOQUOTA(ar->inode))
4250 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4253 * For delayed allocation, we could skip the ENOSPC and
4254 * EDQUOT check, as blocks and quotas have been already
4255 * reserved when data being copied into pagecache.
4257 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4258 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4260 /* Without delayed allocation we need to verify
4261 * there is enough free blocks to do block allocation
4262 * and verify allocation doesn't exceed the quota limits.
4265 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4267 /* let others to free the space */
4269 ar->len = ar->len >> 1;
4275 reserv_clstrs = ar->len;
4276 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4277 dquot_alloc_block_nofail(ar->inode,
4278 EXT4_C2B(sbi, ar->len));
4281 dquot_alloc_block(ar->inode,
4282 EXT4_C2B(sbi, ar->len))) {
4284 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4295 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4302 *errp = ext4_mb_initialize_context(ac, ar);
4308 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4309 if (!ext4_mb_use_preallocated(ac)) {
4310 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4311 ext4_mb_normalize_request(ac, ar);
4313 /* allocate space in core */
4314 *errp = ext4_mb_regular_allocator(ac);
4318 /* as we've just preallocated more space than
4319 * user requested orinally, we store allocated
4320 * space in a special descriptor */
4321 if (ac->ac_status == AC_STATUS_FOUND &&
4322 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4323 ext4_mb_new_preallocation(ac);
4325 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4326 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4327 if (*errp == -EAGAIN) {
4329 * drop the reference that we took
4330 * in ext4_mb_use_best_found
4332 ext4_mb_release_context(ac);
4333 ac->ac_b_ex.fe_group = 0;
4334 ac->ac_b_ex.fe_start = 0;
4335 ac->ac_b_ex.fe_len = 0;
4336 ac->ac_status = AC_STATUS_CONTINUE;
4340 ext4_discard_allocated_blocks(ac);
4342 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4343 ar->len = ac->ac_b_ex.fe_len;
4346 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4353 ac->ac_b_ex.fe_len = 0;
4355 ext4_mb_show_ac(ac);
4357 ext4_mb_release_context(ac);
4360 kmem_cache_free(ext4_ac_cachep, ac);
4361 if (inquota && ar->len < inquota)
4362 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4364 if (!ext4_test_inode_state(ar->inode,
4365 EXT4_STATE_DELALLOC_RESERVED))
4366 /* release all the reserved blocks if non delalloc */
4367 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4371 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4377 * We can merge two free data extents only if the physical blocks
4378 * are contiguous, AND the extents were freed by the same transaction,
4379 * AND the blocks are associated with the same group.
4381 static int can_merge(struct ext4_free_data *entry1,
4382 struct ext4_free_data *entry2)
4384 if ((entry1->efd_tid == entry2->efd_tid) &&
4385 (entry1->efd_group == entry2->efd_group) &&
4386 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4391 static noinline_for_stack int
4392 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4393 struct ext4_free_data *new_entry)
4395 ext4_group_t group = e4b->bd_group;
4396 ext4_grpblk_t cluster;
4397 struct ext4_free_data *entry;
4398 struct ext4_group_info *db = e4b->bd_info;
4399 struct super_block *sb = e4b->bd_sb;
4400 struct ext4_sb_info *sbi = EXT4_SB(sb);
4401 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4402 struct rb_node *parent = NULL, *new_node;
4404 BUG_ON(!ext4_handle_valid(handle));
4405 BUG_ON(e4b->bd_bitmap_page == NULL);
4406 BUG_ON(e4b->bd_buddy_page == NULL);
4408 new_node = &new_entry->efd_node;
4409 cluster = new_entry->efd_start_cluster;
4412 /* first free block exent. We need to
4413 protect buddy cache from being freed,
4414 * otherwise we'll refresh it from
4415 * on-disk bitmap and lose not-yet-available
4417 page_cache_get(e4b->bd_buddy_page);
4418 page_cache_get(e4b->bd_bitmap_page);
4422 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4423 if (cluster < entry->efd_start_cluster)
4425 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4426 n = &(*n)->rb_right;
4428 ext4_grp_locked_error(sb, group, 0,
4429 ext4_group_first_block_no(sb, group) +
4430 EXT4_C2B(sbi, cluster),
4431 "Block already on to-be-freed list");
4436 rb_link_node(new_node, parent, n);
4437 rb_insert_color(new_node, &db->bb_free_root);
4439 /* Now try to see the extent can be merged to left and right */
4440 node = rb_prev(new_node);
4442 entry = rb_entry(node, struct ext4_free_data, efd_node);
4443 if (can_merge(entry, new_entry)) {
4444 new_entry->efd_start_cluster = entry->efd_start_cluster;
4445 new_entry->efd_count += entry->efd_count;
4446 rb_erase(node, &(db->bb_free_root));
4447 ext4_journal_callback_del(handle, &entry->efd_jce);
4448 kmem_cache_free(ext4_free_data_cachep, entry);
4452 node = rb_next(new_node);
4454 entry = rb_entry(node, struct ext4_free_data, efd_node);
4455 if (can_merge(new_entry, entry)) {
4456 new_entry->efd_count += entry->efd_count;
4457 rb_erase(node, &(db->bb_free_root));
4458 ext4_journal_callback_del(handle, &entry->efd_jce);
4459 kmem_cache_free(ext4_free_data_cachep, entry);
4462 /* Add the extent to transaction's private list */
4463 ext4_journal_callback_add(handle, ext4_free_data_callback,
4464 &new_entry->efd_jce);
4469 * ext4_free_blocks() -- Free given blocks and update quota
4470 * @handle: handle for this transaction
4472 * @block: start physical block to free
4473 * @count: number of blocks to count
4474 * @flags: flags used by ext4_free_blocks
4476 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4477 struct buffer_head *bh, ext4_fsblk_t block,
4478 unsigned long count, int flags)
4480 struct buffer_head *bitmap_bh = NULL;
4481 struct super_block *sb = inode->i_sb;
4482 struct ext4_group_desc *gdp;
4483 unsigned long freed = 0;
4484 unsigned int overflow;
4486 struct buffer_head *gd_bh;
4487 ext4_group_t block_group;
4488 struct ext4_sb_info *sbi;
4489 struct ext4_buddy e4b;
4490 unsigned int count_clusters;
4496 BUG_ON(block != bh->b_blocknr);
4498 block = bh->b_blocknr;
4502 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4503 !ext4_data_block_valid(sbi, block, count)) {
4504 ext4_error(sb, "Freeing blocks not in datazone - "
4505 "block = %llu, count = %lu", block, count);
4509 ext4_debug("freeing block %llu\n", block);
4510 trace_ext4_free_blocks(inode, block, count, flags);
4512 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4513 struct buffer_head *tbh = bh;
4516 BUG_ON(bh && (count > 1));
4518 for (i = 0; i < count; i++) {
4520 tbh = sb_find_get_block(inode->i_sb,
4524 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4525 inode, tbh, block + i);
4530 * We need to make sure we don't reuse the freed block until
4531 * after the transaction is committed, which we can do by
4532 * treating the block as metadata, below. We make an
4533 * exception if the inode is to be written in writeback mode
4534 * since writeback mode has weak data consistency guarantees.
4536 if (!ext4_should_writeback_data(inode))
4537 flags |= EXT4_FREE_BLOCKS_METADATA;
4540 * If the extent to be freed does not begin on a cluster
4541 * boundary, we need to deal with partial clusters at the
4542 * beginning and end of the extent. Normally we will free
4543 * blocks at the beginning or the end unless we are explicitly
4544 * requested to avoid doing so.
4546 overflow = block & (sbi->s_cluster_ratio - 1);
4548 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4549 overflow = sbi->s_cluster_ratio - overflow;
4551 if (count > overflow)
4560 overflow = count & (sbi->s_cluster_ratio - 1);
4562 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4563 if (count > overflow)
4568 count += sbi->s_cluster_ratio - overflow;
4573 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4576 * Check to see if we are freeing blocks across a group
4579 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4580 overflow = EXT4_C2B(sbi, bit) + count -
4581 EXT4_BLOCKS_PER_GROUP(sb);
4584 count_clusters = EXT4_B2C(sbi, count);
4585 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4590 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4596 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4597 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4598 in_range(block, ext4_inode_table(sb, gdp),
4599 EXT4_SB(sb)->s_itb_per_group) ||
4600 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4601 EXT4_SB(sb)->s_itb_per_group)) {
4603 ext4_error(sb, "Freeing blocks in system zone - "
4604 "Block = %llu, count = %lu", block, count);
4605 /* err = 0. ext4_std_error should be a no op */
4609 BUFFER_TRACE(bitmap_bh, "getting write access");
4610 err = ext4_journal_get_write_access(handle, bitmap_bh);
4615 * We are about to modify some metadata. Call the journal APIs
4616 * to unshare ->b_data if a currently-committing transaction is
4619 BUFFER_TRACE(gd_bh, "get_write_access");
4620 err = ext4_journal_get_write_access(handle, gd_bh);
4623 #ifdef AGGRESSIVE_CHECK
4626 for (i = 0; i < count_clusters; i++)
4627 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4630 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4632 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4636 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4637 struct ext4_free_data *new_entry;
4639 * blocks being freed are metadata. these blocks shouldn't
4640 * be used until this transaction is committed
4642 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4644 ext4_mb_unload_buddy(&e4b);
4648 new_entry->efd_start_cluster = bit;
4649 new_entry->efd_group = block_group;
4650 new_entry->efd_count = count_clusters;
4651 new_entry->efd_tid = handle->h_transaction->t_tid;
4653 ext4_lock_group(sb, block_group);
4654 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4655 ext4_mb_free_metadata(handle, &e4b, new_entry);
4657 /* need to update group_info->bb_free and bitmap
4658 * with group lock held. generate_buddy look at
4659 * them with group lock_held
4661 ext4_lock_group(sb, block_group);
4662 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4663 mb_free_blocks(inode, &e4b, bit, count_clusters);
4666 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4667 ext4_free_group_clusters_set(sb, gdp, ret);
4668 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh,
4669 EXT4_BLOCKS_PER_GROUP(sb) / 8);
4670 ext4_group_desc_csum_set(sb, block_group, gdp);
4671 ext4_unlock_group(sb, block_group);
4672 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4674 if (sbi->s_log_groups_per_flex) {
4675 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4676 atomic_add(count_clusters,
4677 &sbi->s_flex_groups[flex_group].free_clusters);
4680 ext4_mb_unload_buddy(&e4b);
4684 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4685 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4687 /* We dirtied the bitmap block */
4688 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4689 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4691 /* And the group descriptor block */
4692 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4693 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4697 if (overflow && !err) {
4703 ext4_mark_super_dirty(sb);
4706 ext4_std_error(sb, err);
4711 * ext4_group_add_blocks() -- Add given blocks to an existing group
4712 * @handle: handle to this transaction
4714 * @block: start physcial block to add to the block group
4715 * @count: number of blocks to free
4717 * This marks the blocks as free in the bitmap and buddy.
4719 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4720 ext4_fsblk_t block, unsigned long count)
4722 struct buffer_head *bitmap_bh = NULL;
4723 struct buffer_head *gd_bh;
4724 ext4_group_t block_group;
4727 struct ext4_group_desc *desc;
4728 struct ext4_sb_info *sbi = EXT4_SB(sb);
4729 struct ext4_buddy e4b;
4730 int err = 0, ret, blk_free_count;
4731 ext4_grpblk_t blocks_freed;
4733 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4738 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4740 * Check to see if we are freeing blocks across a group
4743 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4744 ext4_warning(sb, "too much blocks added to group %u\n",
4750 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4756 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4762 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4763 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4764 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4765 in_range(block + count - 1, ext4_inode_table(sb, desc),
4766 sbi->s_itb_per_group)) {
4767 ext4_error(sb, "Adding blocks in system zones - "
4768 "Block = %llu, count = %lu",
4774 BUFFER_TRACE(bitmap_bh, "getting write access");
4775 err = ext4_journal_get_write_access(handle, bitmap_bh);
4780 * We are about to modify some metadata. Call the journal APIs
4781 * to unshare ->b_data if a currently-committing transaction is
4784 BUFFER_TRACE(gd_bh, "get_write_access");
4785 err = ext4_journal_get_write_access(handle, gd_bh);
4789 for (i = 0, blocks_freed = 0; i < count; i++) {
4790 BUFFER_TRACE(bitmap_bh, "clear bit");
4791 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4792 ext4_error(sb, "bit already cleared for block %llu",
4793 (ext4_fsblk_t)(block + i));
4794 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4800 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4805 * need to update group_info->bb_free and bitmap
4806 * with group lock held. generate_buddy look at
4807 * them with group lock_held
4809 ext4_lock_group(sb, block_group);
4810 mb_clear_bits(bitmap_bh->b_data, bit, count);
4811 mb_free_blocks(NULL, &e4b, bit, count);
4812 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4813 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4814 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh,
4815 EXT4_BLOCKS_PER_GROUP(sb) / 8);
4816 ext4_group_desc_csum_set(sb, block_group, desc);
4817 ext4_unlock_group(sb, block_group);
4818 percpu_counter_add(&sbi->s_freeclusters_counter,
4819 EXT4_B2C(sbi, blocks_freed));
4821 if (sbi->s_log_groups_per_flex) {
4822 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4823 atomic_add(EXT4_B2C(sbi, blocks_freed),
4824 &sbi->s_flex_groups[flex_group].free_clusters);
4827 ext4_mb_unload_buddy(&e4b);
4829 /* We dirtied the bitmap block */
4830 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4831 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4833 /* And the group descriptor block */
4834 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4835 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4841 ext4_std_error(sb, err);
4846 * ext4_trim_extent -- function to TRIM one single free extent in the group
4847 * @sb: super block for the file system
4848 * @start: starting block of the free extent in the alloc. group
4849 * @count: number of blocks to TRIM
4850 * @group: alloc. group we are working with
4851 * @e4b: ext4 buddy for the group
4853 * Trim "count" blocks starting at "start" in the "group". To assure that no
4854 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4855 * be called with under the group lock.
4857 static void ext4_trim_extent(struct super_block *sb, int start, int count,
4858 ext4_group_t group, struct ext4_buddy *e4b)
4860 struct ext4_free_extent ex;
4862 trace_ext4_trim_extent(sb, group, start, count);
4864 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4866 ex.fe_start = start;
4867 ex.fe_group = group;
4871 * Mark blocks used, so no one can reuse them while
4874 mb_mark_used(e4b, &ex);
4875 ext4_unlock_group(sb, group);
4876 ext4_issue_discard(sb, group, start, count);
4877 ext4_lock_group(sb, group);
4878 mb_free_blocks(NULL, e4b, start, ex.fe_len);
4882 * ext4_trim_all_free -- function to trim all free space in alloc. group
4883 * @sb: super block for file system
4884 * @group: group to be trimmed
4885 * @start: first group block to examine
4886 * @max: last group block to examine
4887 * @minblocks: minimum extent block count
4889 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4890 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4894 * ext4_trim_all_free walks through group's block bitmap searching for free
4895 * extents. When the free extent is found, mark it as used in group buddy
4896 * bitmap. Then issue a TRIM command on this extent and free the extent in
4897 * the group buddy bitmap. This is done until whole group is scanned.
4899 static ext4_grpblk_t
4900 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4901 ext4_grpblk_t start, ext4_grpblk_t max,
4902 ext4_grpblk_t minblocks)
4905 ext4_grpblk_t next, count = 0, free_count = 0;
4906 struct ext4_buddy e4b;
4909 trace_ext4_trim_all_free(sb, group, start, max);
4911 ret = ext4_mb_load_buddy(sb, group, &e4b);
4913 ext4_error(sb, "Error in loading buddy "
4914 "information for %u", group);
4917 bitmap = e4b.bd_bitmap;
4919 ext4_lock_group(sb, group);
4920 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4921 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4924 start = (e4b.bd_info->bb_first_free > start) ?
4925 e4b.bd_info->bb_first_free : start;
4927 while (start <= max) {
4928 start = mb_find_next_zero_bit(bitmap, max + 1, start);
4931 next = mb_find_next_bit(bitmap, max + 1, start);
4933 if ((next - start) >= minblocks) {
4934 ext4_trim_extent(sb, start,
4935 next - start, group, &e4b);
4936 count += next - start;
4938 free_count += next - start;
4941 if (fatal_signal_pending(current)) {
4942 count = -ERESTARTSYS;
4946 if (need_resched()) {
4947 ext4_unlock_group(sb, group);
4949 ext4_lock_group(sb, group);
4952 if ((e4b.bd_info->bb_free - free_count) < minblocks)
4957 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
4959 ext4_unlock_group(sb, group);
4960 ext4_mb_unload_buddy(&e4b);
4962 ext4_debug("trimmed %d blocks in the group %d\n",
4969 * ext4_trim_fs() -- trim ioctl handle function
4970 * @sb: superblock for filesystem
4971 * @range: fstrim_range structure
4973 * start: First Byte to trim
4974 * len: number of Bytes to trim from start
4975 * minlen: minimum extent length in Bytes
4976 * ext4_trim_fs goes through all allocation groups containing Bytes from
4977 * start to start+len. For each such a group ext4_trim_all_free function
4978 * is invoked to trim all free space.
4980 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
4982 struct ext4_group_info *grp;
4983 ext4_group_t group, first_group, last_group;
4984 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
4985 uint64_t start, end, minlen, trimmed = 0;
4986 ext4_fsblk_t first_data_blk =
4987 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
4988 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
4991 start = range->start >> sb->s_blocksize_bits;
4992 end = start + (range->len >> sb->s_blocksize_bits) - 1;
4993 minlen = range->minlen >> sb->s_blocksize_bits;
4995 if (unlikely(minlen > EXT4_CLUSTERS_PER_GROUP(sb)) ||
4996 unlikely(start >= max_blks))
4998 if (end >= max_blks)
5000 if (end <= first_data_blk)
5002 if (start < first_data_blk)
5003 start = first_data_blk;
5005 /* Determine first and last group to examine based on start and end */
5006 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5007 &first_group, &first_cluster);
5008 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5009 &last_group, &last_cluster);
5011 /* end now represents the last cluster to discard in this group */
5012 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5014 for (group = first_group; group <= last_group; group++) {
5015 grp = ext4_get_group_info(sb, group);
5016 /* We only do this if the grp has never been initialized */
5017 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5018 ret = ext4_mb_init_group(sb, group);
5024 * For all the groups except the last one, last cluster will
5025 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5026 * change it for the last group, note that last_cluster is
5027 * already computed earlier by ext4_get_group_no_and_offset()
5029 if (group == last_group)
5032 if (grp->bb_free >= minlen) {
5033 cnt = ext4_trim_all_free(sb, group, first_cluster,
5043 * For every group except the first one, we are sure
5044 * that the first cluster to discard will be cluster #0.
5050 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5053 range->len = trimmed * sb->s_blocksize;