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/log2.h>
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <trace/events/ext4.h>
31 #ifdef CONFIG_EXT4_DEBUG
32 ushort ext4_mballoc_debug __read_mostly;
34 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
35 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
40 * - test ext4_ext_search_left() and ext4_ext_search_right()
41 * - search for metadata in few groups
44 * - normalization should take into account whether file is still open
45 * - discard preallocations if no free space left (policy?)
46 * - don't normalize tails
48 * - reservation for superuser
51 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
52 * - track min/max extents in each group for better group selection
53 * - mb_mark_used() may allocate chunk right after splitting buddy
54 * - tree of groups sorted by number of free blocks
59 * The allocation request involve request for multiple number of blocks
60 * near to the goal(block) value specified.
62 * During initialization phase of the allocator we decide to use the
63 * group preallocation or inode preallocation depending on the size of
64 * the file. The size of the file could be the resulting file size we
65 * would have after allocation, or the current file size, which ever
66 * is larger. If the size is less than sbi->s_mb_stream_request we
67 * select to use the group preallocation. The default value of
68 * s_mb_stream_request is 16 blocks. This can also be tuned via
69 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
70 * terms of number of blocks.
72 * The main motivation for having small file use group preallocation is to
73 * ensure that we have small files closer together on the disk.
75 * First stage the allocator looks at the inode prealloc list,
76 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
77 * spaces for this particular inode. The inode prealloc space is
80 * pa_lstart -> the logical start block for this prealloc space
81 * pa_pstart -> the physical start block for this prealloc space
82 * pa_len -> length for this prealloc space (in clusters)
83 * pa_free -> free space available in this prealloc space (in clusters)
85 * The inode preallocation space is used looking at the _logical_ start
86 * block. If only the logical file block falls within the range of prealloc
87 * space we will consume the particular prealloc space. This makes sure that
88 * we have contiguous physical blocks representing the file blocks
90 * The important thing to be noted in case of inode prealloc space is that
91 * we don't modify the values associated to inode prealloc space except
94 * If we are not able to find blocks in the inode prealloc space and if we
95 * have the group allocation flag set then we look at the locality group
96 * prealloc space. These are per CPU prealloc list represented as
98 * ext4_sb_info.s_locality_groups[smp_processor_id()]
100 * The reason for having a per cpu locality group is to reduce the contention
101 * between CPUs. It is possible to get scheduled at this point.
103 * The locality group prealloc space is used looking at whether we have
104 * enough free space (pa_free) within the prealloc space.
106 * If we can't allocate blocks via inode prealloc or/and locality group
107 * prealloc then we look at the buddy cache. The buddy cache is represented
108 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
109 * mapped to the buddy and bitmap information regarding different
110 * groups. The buddy information is attached to buddy cache inode so that
111 * we can access them through the page cache. The information regarding
112 * each group is loaded via ext4_mb_load_buddy. The information involve
113 * block bitmap and buddy information. The information are stored in the
117 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
120 * one block each for bitmap and buddy information. So for each group we
121 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
122 * blocksize) blocks. So it can have information regarding groups_per_page
123 * which is blocks_per_page/2
125 * The buddy cache inode is not stored on disk. The inode is thrown
126 * away when the filesystem is unmounted.
128 * We look for count number of blocks in the buddy cache. If we were able
129 * to locate that many free blocks we return with additional information
130 * regarding rest of the contiguous physical block available
132 * Before allocating blocks via buddy cache we normalize the request
133 * blocks. This ensure we ask for more blocks that we needed. The extra
134 * blocks that we get after allocation is added to the respective prealloc
135 * list. In case of inode preallocation we follow a list of heuristics
136 * based on file size. This can be found in ext4_mb_normalize_request. If
137 * we are doing a group prealloc we try to normalize the request to
138 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
139 * dependent on the cluster size; for non-bigalloc file systems, it is
140 * 512 blocks. This can be tuned via
141 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
142 * terms of number of blocks. If we have mounted the file system with -O
143 * stripe=<value> option the group prealloc request is normalized to the
144 * the smallest multiple of the stripe value (sbi->s_stripe) which is
145 * greater than the default mb_group_prealloc.
147 * The regular allocator (using the buddy cache) supports a few tunables.
149 * /sys/fs/ext4/<partition>/mb_min_to_scan
150 * /sys/fs/ext4/<partition>/mb_max_to_scan
151 * /sys/fs/ext4/<partition>/mb_order2_req
153 * The regular allocator uses buddy scan only if the request len is power of
154 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
155 * value of s_mb_order2_reqs can be tuned via
156 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
157 * stripe size (sbi->s_stripe), we try to search for contiguous block in
158 * stripe size. This should result in better allocation on RAID setups. If
159 * not, we search in the specific group using bitmap for best extents. The
160 * tunable min_to_scan and max_to_scan control the behaviour here.
161 * min_to_scan indicate how long the mballoc __must__ look for a best
162 * extent and max_to_scan indicates how long the mballoc __can__ look for a
163 * best extent in the found extents. Searching for the blocks starts with
164 * the group specified as the goal value in allocation context via
165 * ac_g_ex. Each group is first checked based on the criteria whether it
166 * can be used for allocation. ext4_mb_good_group explains how the groups are
169 * Both the prealloc space are getting populated as above. So for the first
170 * request we will hit the buddy cache which will result in this prealloc
171 * space getting filled. The prealloc space is then later used for the
172 * subsequent request.
176 * mballoc operates on the following data:
178 * - in-core buddy (actually includes buddy and bitmap)
179 * - preallocation descriptors (PAs)
181 * there are two types of preallocations:
183 * assiged to specific inode and can be used for this inode only.
184 * it describes part of inode's space preallocated to specific
185 * physical blocks. any block from that preallocated can be used
186 * independent. the descriptor just tracks number of blocks left
187 * unused. so, before taking some block from descriptor, one must
188 * make sure corresponded logical block isn't allocated yet. this
189 * also means that freeing any block within descriptor's range
190 * must discard all preallocated blocks.
192 * assigned to specific locality group which does not translate to
193 * permanent set of inodes: inode can join and leave group. space
194 * from this type of preallocation can be used for any inode. thus
195 * it's consumed from the beginning to the end.
197 * relation between them can be expressed as:
198 * in-core buddy = on-disk bitmap + preallocation descriptors
200 * this mean blocks mballoc considers used are:
201 * - allocated blocks (persistent)
202 * - preallocated blocks (non-persistent)
204 * consistency in mballoc world means that at any time a block is either
205 * free or used in ALL structures. notice: "any time" should not be read
206 * literally -- time is discrete and delimited by locks.
208 * to keep it simple, we don't use block numbers, instead we count number of
209 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
211 * all operations can be expressed as:
212 * - init buddy: buddy = on-disk + PAs
213 * - new PA: buddy += N; PA = N
214 * - use inode PA: on-disk += N; PA -= N
215 * - discard inode PA buddy -= on-disk - PA; PA = 0
216 * - use locality group PA on-disk += N; PA -= N
217 * - discard locality group PA buddy -= PA; PA = 0
218 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
219 * is used in real operation because we can't know actual used
220 * bits from PA, only from on-disk bitmap
222 * if we follow this strict logic, then all operations above should be atomic.
223 * given some of them can block, we'd have to use something like semaphores
224 * killing performance on high-end SMP hardware. let's try to relax it using
225 * the following knowledge:
226 * 1) if buddy is referenced, it's already initialized
227 * 2) while block is used in buddy and the buddy is referenced,
228 * nobody can re-allocate that block
229 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
230 * bit set and PA claims same block, it's OK. IOW, one can set bit in
231 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
234 * so, now we're building a concurrency table:
237 * blocks for PA are allocated in the buddy, buddy must be referenced
238 * until PA is linked to allocation group to avoid concurrent buddy init
240 * we need to make sure that either on-disk bitmap or PA has uptodate data
241 * given (3) we care that PA-=N operation doesn't interfere with init
243 * the simplest way would be to have buddy initialized by the discard
244 * - use locality group PA
245 * again PA-=N must be serialized with init
246 * - discard locality group PA
247 * the simplest way would be to have buddy initialized by the discard
250 * i_data_sem serializes them
252 * discard process must wait until PA isn't used by another process
253 * - use locality group PA
254 * some mutex should serialize them
255 * - discard locality group PA
256 * discard process must wait until PA isn't used by another process
259 * i_data_sem or another mutex should serializes them
261 * discard process must wait until PA isn't used by another process
262 * - use locality group PA
263 * nothing wrong here -- they're different PAs covering different blocks
264 * - discard locality group PA
265 * discard process must wait until PA isn't used by another process
267 * now we're ready to make few consequences:
268 * - PA is referenced and while it is no discard is possible
269 * - PA is referenced until block isn't marked in on-disk bitmap
270 * - PA changes only after on-disk bitmap
271 * - discard must not compete with init. either init is done before
272 * any discard or they're serialized somehow
273 * - buddy init as sum of on-disk bitmap and PAs is done atomically
275 * a special case when we've used PA to emptiness. no need to modify buddy
276 * in this case, but we should care about concurrent init
281 * Logic in few words:
286 * mark bits in on-disk bitmap
289 * - use preallocation:
290 * find proper PA (per-inode or group)
292 * mark bits in on-disk bitmap
298 * mark bits in on-disk bitmap
301 * - discard preallocations in group:
303 * move them onto local list
304 * load on-disk bitmap
306 * remove PA from object (inode or locality group)
307 * mark free blocks in-core
309 * - discard inode's preallocations:
316 * - bitlock on a group (group)
317 * - object (inode/locality) (object)
328 * - release consumed pa:
333 * - generate in-core bitmap:
337 * - discard all for given object (inode, locality group):
342 * - discard all for given group:
349 static struct kmem_cache *ext4_pspace_cachep;
350 static struct kmem_cache *ext4_ac_cachep;
351 static struct kmem_cache *ext4_free_data_cachep;
353 /* We create slab caches for groupinfo data structures based on the
354 * superblock block size. There will be one per mounted filesystem for
355 * each unique s_blocksize_bits */
356 #define NR_GRPINFO_CACHES 8
357 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
359 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
360 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
361 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
362 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
365 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
367 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
369 static void ext4_free_data_callback(struct super_block *sb,
370 struct ext4_journal_cb_entry *jce, int rc);
372 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
374 #if BITS_PER_LONG == 64
375 *bit += ((unsigned long) addr & 7UL) << 3;
376 addr = (void *) ((unsigned long) addr & ~7UL);
377 #elif BITS_PER_LONG == 32
378 *bit += ((unsigned long) addr & 3UL) << 3;
379 addr = (void *) ((unsigned long) addr & ~3UL);
381 #error "how many bits you are?!"
386 static inline int mb_test_bit(int bit, void *addr)
389 * ext4_test_bit on architecture like powerpc
390 * needs unsigned long aligned address
392 addr = mb_correct_addr_and_bit(&bit, addr);
393 return ext4_test_bit(bit, addr);
396 static inline void mb_set_bit(int bit, void *addr)
398 addr = mb_correct_addr_and_bit(&bit, addr);
399 ext4_set_bit(bit, addr);
402 static inline void mb_clear_bit(int bit, void *addr)
404 addr = mb_correct_addr_and_bit(&bit, addr);
405 ext4_clear_bit(bit, addr);
408 static inline int mb_test_and_clear_bit(int bit, void *addr)
410 addr = mb_correct_addr_and_bit(&bit, addr);
411 return ext4_test_and_clear_bit(bit, addr);
414 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
416 int fix = 0, ret, tmpmax;
417 addr = mb_correct_addr_and_bit(&fix, addr);
421 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
427 static inline int mb_find_next_bit(void *addr, int max, int start)
429 int fix = 0, ret, tmpmax;
430 addr = mb_correct_addr_and_bit(&fix, addr);
434 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
440 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
444 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
447 if (order > e4b->bd_blkbits + 1) {
452 /* at order 0 we see each particular block */
454 *max = 1 << (e4b->bd_blkbits + 3);
455 return e4b->bd_bitmap;
458 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
459 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
465 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
466 int first, int count)
469 struct super_block *sb = e4b->bd_sb;
471 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
473 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
474 for (i = 0; i < count; i++) {
475 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
476 ext4_fsblk_t blocknr;
478 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
479 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
480 ext4_grp_locked_error(sb, e4b->bd_group,
481 inode ? inode->i_ino : 0,
483 "freeing block already freed "
487 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
491 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
495 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
497 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
498 for (i = 0; i < count; i++) {
499 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
500 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
504 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
506 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
507 unsigned char *b1, *b2;
509 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
510 b2 = (unsigned char *) bitmap;
511 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
512 if (b1[i] != b2[i]) {
513 ext4_msg(e4b->bd_sb, KERN_ERR,
514 "corruption in group %u "
515 "at byte %u(%u): %x in copy != %x "
517 e4b->bd_group, i, i * 8, b1[i], b2[i]);
525 static inline void mb_free_blocks_double(struct inode *inode,
526 struct ext4_buddy *e4b, int first, int count)
530 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
531 int first, int count)
535 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
541 #ifdef AGGRESSIVE_CHECK
543 #define MB_CHECK_ASSERT(assert) \
547 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
548 function, file, line, # assert); \
553 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
554 const char *function, int line)
556 struct super_block *sb = e4b->bd_sb;
557 int order = e4b->bd_blkbits + 1;
564 struct ext4_group_info *grp;
567 struct list_head *cur;
572 static int mb_check_counter;
573 if (mb_check_counter++ % 100 != 0)
578 buddy = mb_find_buddy(e4b, order, &max);
579 MB_CHECK_ASSERT(buddy);
580 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
581 MB_CHECK_ASSERT(buddy2);
582 MB_CHECK_ASSERT(buddy != buddy2);
583 MB_CHECK_ASSERT(max * 2 == max2);
586 for (i = 0; i < max; i++) {
588 if (mb_test_bit(i, buddy)) {
589 /* only single bit in buddy2 may be 1 */
590 if (!mb_test_bit(i << 1, buddy2)) {
592 mb_test_bit((i<<1)+1, buddy2));
593 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
595 mb_test_bit(i << 1, buddy2));
600 /* both bits in buddy2 must be 1 */
601 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
602 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
604 for (j = 0; j < (1 << order); j++) {
605 k = (i * (1 << order)) + j;
607 !mb_test_bit(k, e4b->bd_bitmap));
611 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
616 buddy = mb_find_buddy(e4b, 0, &max);
617 for (i = 0; i < max; i++) {
618 if (!mb_test_bit(i, buddy)) {
619 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
627 /* check used bits only */
628 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
629 buddy2 = mb_find_buddy(e4b, j, &max2);
631 MB_CHECK_ASSERT(k < max2);
632 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
635 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
636 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
638 grp = ext4_get_group_info(sb, e4b->bd_group);
639 list_for_each(cur, &grp->bb_prealloc_list) {
640 ext4_group_t groupnr;
641 struct ext4_prealloc_space *pa;
642 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
643 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
644 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
645 for (i = 0; i < pa->pa_len; i++)
646 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
650 #undef MB_CHECK_ASSERT
651 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
652 __FILE__, __func__, __LINE__)
654 #define mb_check_buddy(e4b)
658 * Divide blocks started from @first with length @len into
659 * smaller chunks with power of 2 blocks.
660 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
661 * then increase bb_counters[] for corresponded chunk size.
663 static void ext4_mb_mark_free_simple(struct super_block *sb,
664 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
665 struct ext4_group_info *grp)
667 struct ext4_sb_info *sbi = EXT4_SB(sb);
671 unsigned short border;
673 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
675 border = 2 << sb->s_blocksize_bits;
678 /* find how many blocks can be covered since this position */
679 max = ffs(first | border) - 1;
681 /* find how many blocks of power 2 we need to mark */
688 /* mark multiblock chunks only */
689 grp->bb_counters[min]++;
691 mb_clear_bit(first >> min,
692 buddy + sbi->s_mb_offsets[min]);
700 * Cache the order of the largest free extent we have available in this block
704 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
709 grp->bb_largest_free_order = -1; /* uninit */
711 bits = sb->s_blocksize_bits + 1;
712 for (i = bits; i >= 0; i--) {
713 if (grp->bb_counters[i] > 0) {
714 grp->bb_largest_free_order = i;
720 static noinline_for_stack
721 void ext4_mb_generate_buddy(struct super_block *sb,
722 void *buddy, void *bitmap, ext4_group_t group)
724 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
725 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
730 unsigned fragments = 0;
731 unsigned long long period = get_cycles();
733 /* initialize buddy from bitmap which is aggregation
734 * of on-disk bitmap and preallocations */
735 i = mb_find_next_zero_bit(bitmap, max, 0);
736 grp->bb_first_free = i;
740 i = mb_find_next_bit(bitmap, max, i);
744 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
746 grp->bb_counters[0]++;
748 i = mb_find_next_zero_bit(bitmap, max, i);
750 grp->bb_fragments = fragments;
752 if (free != grp->bb_free) {
753 ext4_grp_locked_error(sb, group, 0, 0,
754 "%u clusters in bitmap, %u in gd",
757 * If we intent to continue, we consider group descritor
758 * corrupt and update bb_free using bitmap value
762 mb_set_largest_free_order(sb, grp);
764 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
766 period = get_cycles() - period;
767 spin_lock(&EXT4_SB(sb)->s_bal_lock);
768 EXT4_SB(sb)->s_mb_buddies_generated++;
769 EXT4_SB(sb)->s_mb_generation_time += period;
770 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
773 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
779 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
780 ext4_set_bits(buddy, 0, count);
782 e4b->bd_info->bb_fragments = 0;
783 memset(e4b->bd_info->bb_counters, 0,
784 sizeof(*e4b->bd_info->bb_counters) *
785 (e4b->bd_sb->s_blocksize_bits + 2));
787 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
788 e4b->bd_bitmap, e4b->bd_group);
791 /* The buddy information is attached the buddy cache inode
792 * for convenience. The information regarding each group
793 * is loaded via ext4_mb_load_buddy. The information involve
794 * block bitmap and buddy information. The information are
795 * stored in the inode as
798 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
801 * one block each for bitmap and buddy information.
802 * So for each group we take up 2 blocks. A page can
803 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
804 * So it can have information regarding groups_per_page which
805 * is blocks_per_page/2
807 * Locking note: This routine takes the block group lock of all groups
808 * for this page; do not hold this lock when calling this routine!
811 static int ext4_mb_init_cache(struct page *page, char *incore)
813 ext4_group_t ngroups;
819 ext4_group_t first_group, group;
821 struct super_block *sb;
822 struct buffer_head *bhs;
823 struct buffer_head **bh = NULL;
827 struct ext4_group_info *grinfo;
829 mb_debug(1, "init page %lu\n", page->index);
831 inode = page->mapping->host;
833 ngroups = ext4_get_groups_count(sb);
834 blocksize = 1 << inode->i_blkbits;
835 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
837 groups_per_page = blocks_per_page >> 1;
838 if (groups_per_page == 0)
841 /* allocate buffer_heads to read bitmaps */
842 if (groups_per_page > 1) {
843 i = sizeof(struct buffer_head *) * groups_per_page;
844 bh = kzalloc(i, GFP_NOFS);
852 first_group = page->index * blocks_per_page / 2;
854 /* read all groups the page covers into the cache */
855 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
856 if (group >= ngroups)
859 grinfo = ext4_get_group_info(sb, group);
861 * If page is uptodate then we came here after online resize
862 * which added some new uninitialized group info structs, so
863 * we must skip all initialized uptodate buddies on the page,
864 * which may be currently in use by an allocating task.
866 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
870 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
874 mb_debug(1, "read bitmap for group %u\n", group);
877 /* wait for I/O completion */
878 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
879 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
885 first_block = page->index * blocks_per_page;
886 for (i = 0; i < blocks_per_page; i++) {
887 group = (first_block + i) >> 1;
888 if (group >= ngroups)
891 if (!bh[group - first_group])
892 /* skip initialized uptodate buddy */
896 * data carry information regarding this
897 * particular group in the format specified
901 data = page_address(page) + (i * blocksize);
902 bitmap = bh[group - first_group]->b_data;
905 * We place the buddy block and bitmap block
908 if ((first_block + i) & 1) {
909 /* this is block of buddy */
910 BUG_ON(incore == NULL);
911 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
912 group, page->index, i * blocksize);
913 trace_ext4_mb_buddy_bitmap_load(sb, group);
914 grinfo = ext4_get_group_info(sb, group);
915 grinfo->bb_fragments = 0;
916 memset(grinfo->bb_counters, 0,
917 sizeof(*grinfo->bb_counters) *
918 (sb->s_blocksize_bits+2));
920 * incore got set to the group block bitmap below
922 ext4_lock_group(sb, group);
924 memset(data, 0xff, blocksize);
925 ext4_mb_generate_buddy(sb, data, incore, group);
926 ext4_unlock_group(sb, group);
929 /* this is block of bitmap */
930 BUG_ON(incore != NULL);
931 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
932 group, page->index, i * blocksize);
933 trace_ext4_mb_bitmap_load(sb, group);
935 /* see comments in ext4_mb_put_pa() */
936 ext4_lock_group(sb, group);
937 memcpy(data, bitmap, blocksize);
939 /* mark all preallocated blks used in in-core bitmap */
940 ext4_mb_generate_from_pa(sb, data, group);
941 ext4_mb_generate_from_freelist(sb, data, group);
942 ext4_unlock_group(sb, group);
944 /* set incore so that the buddy information can be
945 * generated using this
950 SetPageUptodate(page);
954 for (i = 0; i < groups_per_page; i++)
963 * Lock the buddy and bitmap pages. This make sure other parallel init_group
964 * on the same buddy page doesn't happen whild holding the buddy page lock.
965 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
966 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
968 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
969 ext4_group_t group, struct ext4_buddy *e4b)
971 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
972 int block, pnum, poff;
976 e4b->bd_buddy_page = NULL;
977 e4b->bd_bitmap_page = NULL;
979 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
981 * the buddy cache inode stores the block bitmap
982 * and buddy information in consecutive blocks.
983 * So for each group we need two blocks.
986 pnum = block / blocks_per_page;
987 poff = block % blocks_per_page;
988 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
991 BUG_ON(page->mapping != inode->i_mapping);
992 e4b->bd_bitmap_page = page;
993 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
995 if (blocks_per_page >= 2) {
996 /* buddy and bitmap are on the same page */
1001 pnum = block / blocks_per_page;
1002 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1005 BUG_ON(page->mapping != inode->i_mapping);
1006 e4b->bd_buddy_page = page;
1010 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1012 if (e4b->bd_bitmap_page) {
1013 unlock_page(e4b->bd_bitmap_page);
1014 page_cache_release(e4b->bd_bitmap_page);
1016 if (e4b->bd_buddy_page) {
1017 unlock_page(e4b->bd_buddy_page);
1018 page_cache_release(e4b->bd_buddy_page);
1023 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1024 * block group lock of all groups for this page; do not hold the BG lock when
1025 * calling this routine!
1027 static noinline_for_stack
1028 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1031 struct ext4_group_info *this_grp;
1032 struct ext4_buddy e4b;
1037 mb_debug(1, "init group %u\n", group);
1038 this_grp = ext4_get_group_info(sb, group);
1040 * This ensures that we don't reinit the buddy cache
1041 * page which map to the group from which we are already
1042 * allocating. If we are looking at the buddy cache we would
1043 * have taken a reference using ext4_mb_load_buddy and that
1044 * would have pinned buddy page to page cache.
1046 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1047 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1049 * somebody initialized the group
1050 * return without doing anything
1055 page = e4b.bd_bitmap_page;
1056 ret = ext4_mb_init_cache(page, NULL);
1059 if (!PageUptodate(page)) {
1063 mark_page_accessed(page);
1065 if (e4b.bd_buddy_page == NULL) {
1067 * If both the bitmap and buddy are in
1068 * the same page we don't need to force
1074 /* init buddy cache */
1075 page = e4b.bd_buddy_page;
1076 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1079 if (!PageUptodate(page)) {
1083 mark_page_accessed(page);
1085 ext4_mb_put_buddy_page_lock(&e4b);
1090 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1091 * block group lock of all groups for this page; do not hold the BG lock when
1092 * calling this routine!
1094 static noinline_for_stack int
1095 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1096 struct ext4_buddy *e4b)
1098 int blocks_per_page;
1104 struct ext4_group_info *grp;
1105 struct ext4_sb_info *sbi = EXT4_SB(sb);
1106 struct inode *inode = sbi->s_buddy_cache;
1109 mb_debug(1, "load group %u\n", group);
1111 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1112 grp = ext4_get_group_info(sb, group);
1114 e4b->bd_blkbits = sb->s_blocksize_bits;
1117 e4b->bd_group = group;
1118 e4b->bd_buddy_page = NULL;
1119 e4b->bd_bitmap_page = NULL;
1121 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1123 * we need full data about the group
1124 * to make a good selection
1126 ret = ext4_mb_init_group(sb, group);
1132 * the buddy cache inode stores the block bitmap
1133 * and buddy information in consecutive blocks.
1134 * So for each group we need two blocks.
1137 pnum = block / blocks_per_page;
1138 poff = block % blocks_per_page;
1140 /* we could use find_or_create_page(), but it locks page
1141 * what we'd like to avoid in fast path ... */
1142 page = find_get_page(inode->i_mapping, pnum);
1143 if (page == NULL || !PageUptodate(page)) {
1146 * drop the page reference and try
1147 * to get the page with lock. If we
1148 * are not uptodate that implies
1149 * somebody just created the page but
1150 * is yet to initialize the same. So
1151 * wait for it to initialize.
1153 page_cache_release(page);
1154 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1156 BUG_ON(page->mapping != inode->i_mapping);
1157 if (!PageUptodate(page)) {
1158 ret = ext4_mb_init_cache(page, NULL);
1163 mb_cmp_bitmaps(e4b, page_address(page) +
1164 (poff * sb->s_blocksize));
1169 if (page == NULL || !PageUptodate(page)) {
1173 e4b->bd_bitmap_page = page;
1174 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1175 mark_page_accessed(page);
1178 pnum = block / blocks_per_page;
1179 poff = block % blocks_per_page;
1181 page = find_get_page(inode->i_mapping, pnum);
1182 if (page == NULL || !PageUptodate(page)) {
1184 page_cache_release(page);
1185 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1187 BUG_ON(page->mapping != inode->i_mapping);
1188 if (!PageUptodate(page)) {
1189 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1198 if (page == NULL || !PageUptodate(page)) {
1202 e4b->bd_buddy_page = page;
1203 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1204 mark_page_accessed(page);
1206 BUG_ON(e4b->bd_bitmap_page == NULL);
1207 BUG_ON(e4b->bd_buddy_page == NULL);
1213 page_cache_release(page);
1214 if (e4b->bd_bitmap_page)
1215 page_cache_release(e4b->bd_bitmap_page);
1216 if (e4b->bd_buddy_page)
1217 page_cache_release(e4b->bd_buddy_page);
1218 e4b->bd_buddy = NULL;
1219 e4b->bd_bitmap = NULL;
1223 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1225 if (e4b->bd_bitmap_page)
1226 page_cache_release(e4b->bd_bitmap_page);
1227 if (e4b->bd_buddy_page)
1228 page_cache_release(e4b->bd_buddy_page);
1232 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1237 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1238 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1241 while (order <= e4b->bd_blkbits + 1) {
1243 if (!mb_test_bit(block, bb)) {
1244 /* this block is part of buddy of order 'order' */
1247 bb += 1 << (e4b->bd_blkbits - order);
1253 static void mb_clear_bits(void *bm, int cur, int len)
1259 if ((cur & 31) == 0 && (len - cur) >= 32) {
1260 /* fast path: clear whole word at once */
1261 addr = bm + (cur >> 3);
1266 mb_clear_bit(cur, bm);
1271 /* clear bits in given range
1272 * will return first found zero bit if any, -1 otherwise
1274 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1281 if ((cur & 31) == 0 && (len - cur) >= 32) {
1282 /* fast path: clear whole word at once */
1283 addr = bm + (cur >> 3);
1284 if (*addr != (__u32)(-1) && zero_bit == -1)
1285 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1290 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1298 void ext4_set_bits(void *bm, int cur, int len)
1304 if ((cur & 31) == 0 && (len - cur) >= 32) {
1305 /* fast path: set whole word at once */
1306 addr = bm + (cur >> 3);
1311 mb_set_bit(cur, bm);
1317 * _________________________________________________________________ */
1319 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1321 if (mb_test_bit(*bit + side, bitmap)) {
1322 mb_clear_bit(*bit, bitmap);
1328 mb_set_bit(*bit, bitmap);
1333 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1337 void *buddy = mb_find_buddy(e4b, order, &max);
1342 /* Bits in range [first; last] are known to be set since
1343 * corresponding blocks were allocated. Bits in range
1344 * (first; last) will stay set because they form buddies on
1345 * upper layer. We just deal with borders if they don't
1346 * align with upper layer and then go up.
1347 * Releasing entire group is all about clearing
1348 * single bit of highest order buddy.
1352 * ---------------------------------
1354 * ---------------------------------
1355 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1356 * ---------------------------------
1358 * \_____________________/
1360 * Neither [1] nor [6] is aligned to above layer.
1361 * Left neighbour [0] is free, so mark it busy,
1362 * decrease bb_counters and extend range to
1364 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1365 * mark [6] free, increase bb_counters and shrink range to
1367 * Then shift range to [0; 2], go up and do the same.
1372 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1374 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1379 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1380 mb_clear_bits(buddy, first, last - first + 1);
1381 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1390 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1391 int first, int count)
1393 int left_is_free = 0;
1394 int right_is_free = 0;
1396 int last = first + count - 1;
1397 struct super_block *sb = e4b->bd_sb;
1399 if (WARN_ON(count == 0))
1401 BUG_ON(last >= (sb->s_blocksize << 3));
1402 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1403 mb_check_buddy(e4b);
1404 mb_free_blocks_double(inode, e4b, first, count);
1406 e4b->bd_info->bb_free += count;
1407 if (first < e4b->bd_info->bb_first_free)
1408 e4b->bd_info->bb_first_free = first;
1410 /* access memory sequentially: check left neighbour,
1411 * clear range and then check right neighbour
1414 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1415 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1416 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1417 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1419 if (unlikely(block != -1)) {
1420 ext4_fsblk_t blocknr;
1422 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1423 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1424 ext4_grp_locked_error(sb, e4b->bd_group,
1425 inode ? inode->i_ino : 0,
1427 "freeing already freed block "
1429 mb_regenerate_buddy(e4b);
1433 /* let's maintain fragments counter */
1434 if (left_is_free && right_is_free)
1435 e4b->bd_info->bb_fragments--;
1436 else if (!left_is_free && !right_is_free)
1437 e4b->bd_info->bb_fragments++;
1439 /* buddy[0] == bd_bitmap is a special case, so handle
1440 * it right away and let mb_buddy_mark_free stay free of
1441 * zero order checks.
1442 * Check if neighbours are to be coaleasced,
1443 * adjust bitmap bb_counters and borders appropriately.
1446 first += !left_is_free;
1447 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1450 last -= !right_is_free;
1451 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1455 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1458 mb_set_largest_free_order(sb, e4b->bd_info);
1459 mb_check_buddy(e4b);
1462 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1463 int needed, struct ext4_free_extent *ex)
1469 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1472 buddy = mb_find_buddy(e4b, 0, &max);
1473 BUG_ON(buddy == NULL);
1474 BUG_ON(block >= max);
1475 if (mb_test_bit(block, buddy)) {
1482 /* find actual order */
1483 order = mb_find_order_for_block(e4b, block);
1484 block = block >> order;
1486 ex->fe_len = 1 << order;
1487 ex->fe_start = block << order;
1488 ex->fe_group = e4b->bd_group;
1490 /* calc difference from given start */
1491 next = next - ex->fe_start;
1493 ex->fe_start += next;
1495 while (needed > ex->fe_len &&
1496 mb_find_buddy(e4b, order, &max)) {
1498 if (block + 1 >= max)
1501 next = (block + 1) * (1 << order);
1502 if (mb_test_bit(next, e4b->bd_bitmap))
1505 order = mb_find_order_for_block(e4b, next);
1507 block = next >> order;
1508 ex->fe_len += 1 << order;
1511 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1515 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1521 int start = ex->fe_start;
1522 int len = ex->fe_len;
1527 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1528 BUG_ON(e4b->bd_group != ex->fe_group);
1529 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1530 mb_check_buddy(e4b);
1531 mb_mark_used_double(e4b, start, len);
1533 e4b->bd_info->bb_free -= len;
1534 if (e4b->bd_info->bb_first_free == start)
1535 e4b->bd_info->bb_first_free += len;
1537 /* let's maintain fragments counter */
1539 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1540 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1541 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1543 e4b->bd_info->bb_fragments++;
1544 else if (!mlen && !max)
1545 e4b->bd_info->bb_fragments--;
1547 /* let's maintain buddy itself */
1549 ord = mb_find_order_for_block(e4b, start);
1551 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1552 /* the whole chunk may be allocated at once! */
1554 buddy = mb_find_buddy(e4b, ord, &max);
1555 BUG_ON((start >> ord) >= max);
1556 mb_set_bit(start >> ord, buddy);
1557 e4b->bd_info->bb_counters[ord]--;
1564 /* store for history */
1566 ret = len | (ord << 16);
1568 /* we have to split large buddy */
1570 buddy = mb_find_buddy(e4b, ord, &max);
1571 mb_set_bit(start >> ord, buddy);
1572 e4b->bd_info->bb_counters[ord]--;
1575 cur = (start >> ord) & ~1U;
1576 buddy = mb_find_buddy(e4b, ord, &max);
1577 mb_clear_bit(cur, buddy);
1578 mb_clear_bit(cur + 1, buddy);
1579 e4b->bd_info->bb_counters[ord]++;
1580 e4b->bd_info->bb_counters[ord]++;
1582 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1584 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1585 mb_check_buddy(e4b);
1591 * Must be called under group lock!
1593 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1594 struct ext4_buddy *e4b)
1596 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1599 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1600 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1602 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1603 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1604 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1606 /* preallocation can change ac_b_ex, thus we store actually
1607 * allocated blocks for history */
1608 ac->ac_f_ex = ac->ac_b_ex;
1610 ac->ac_status = AC_STATUS_FOUND;
1611 ac->ac_tail = ret & 0xffff;
1612 ac->ac_buddy = ret >> 16;
1615 * take the page reference. We want the page to be pinned
1616 * so that we don't get a ext4_mb_init_cache_call for this
1617 * group until we update the bitmap. That would mean we
1618 * double allocate blocks. The reference is dropped
1619 * in ext4_mb_release_context
1621 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1622 get_page(ac->ac_bitmap_page);
1623 ac->ac_buddy_page = e4b->bd_buddy_page;
1624 get_page(ac->ac_buddy_page);
1625 /* store last allocated for subsequent stream allocation */
1626 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1627 spin_lock(&sbi->s_md_lock);
1628 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1629 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1630 spin_unlock(&sbi->s_md_lock);
1635 * regular allocator, for general purposes allocation
1638 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1639 struct ext4_buddy *e4b,
1642 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1643 struct ext4_free_extent *bex = &ac->ac_b_ex;
1644 struct ext4_free_extent *gex = &ac->ac_g_ex;
1645 struct ext4_free_extent ex;
1648 if (ac->ac_status == AC_STATUS_FOUND)
1651 * We don't want to scan for a whole year
1653 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1654 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1655 ac->ac_status = AC_STATUS_BREAK;
1660 * Haven't found good chunk so far, let's continue
1662 if (bex->fe_len < gex->fe_len)
1665 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1666 && bex->fe_group == e4b->bd_group) {
1667 /* recheck chunk's availability - we don't know
1668 * when it was found (within this lock-unlock
1670 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1671 if (max >= gex->fe_len) {
1672 ext4_mb_use_best_found(ac, e4b);
1679 * The routine checks whether found extent is good enough. If it is,
1680 * then the extent gets marked used and flag is set to the context
1681 * to stop scanning. Otherwise, the extent is compared with the
1682 * previous found extent and if new one is better, then it's stored
1683 * in the context. Later, the best found extent will be used, if
1684 * mballoc can't find good enough extent.
1686 * FIXME: real allocation policy is to be designed yet!
1688 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1689 struct ext4_free_extent *ex,
1690 struct ext4_buddy *e4b)
1692 struct ext4_free_extent *bex = &ac->ac_b_ex;
1693 struct ext4_free_extent *gex = &ac->ac_g_ex;
1695 BUG_ON(ex->fe_len <= 0);
1696 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1697 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1698 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1703 * The special case - take what you catch first
1705 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1707 ext4_mb_use_best_found(ac, e4b);
1712 * Let's check whether the chuck is good enough
1714 if (ex->fe_len == gex->fe_len) {
1716 ext4_mb_use_best_found(ac, e4b);
1721 * If this is first found extent, just store it in the context
1723 if (bex->fe_len == 0) {
1729 * If new found extent is better, store it in the context
1731 if (bex->fe_len < gex->fe_len) {
1732 /* if the request isn't satisfied, any found extent
1733 * larger than previous best one is better */
1734 if (ex->fe_len > bex->fe_len)
1736 } else if (ex->fe_len > gex->fe_len) {
1737 /* if the request is satisfied, then we try to find
1738 * an extent that still satisfy the request, but is
1739 * smaller than previous one */
1740 if (ex->fe_len < bex->fe_len)
1744 ext4_mb_check_limits(ac, e4b, 0);
1747 static noinline_for_stack
1748 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1749 struct ext4_buddy *e4b)
1751 struct ext4_free_extent ex = ac->ac_b_ex;
1752 ext4_group_t group = ex.fe_group;
1756 BUG_ON(ex.fe_len <= 0);
1757 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1761 ext4_lock_group(ac->ac_sb, group);
1762 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1766 ext4_mb_use_best_found(ac, e4b);
1769 ext4_unlock_group(ac->ac_sb, group);
1770 ext4_mb_unload_buddy(e4b);
1775 static noinline_for_stack
1776 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1777 struct ext4_buddy *e4b)
1779 ext4_group_t group = ac->ac_g_ex.fe_group;
1782 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1783 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1784 struct ext4_free_extent ex;
1786 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1788 if (grp->bb_free == 0)
1791 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1795 ext4_lock_group(ac->ac_sb, group);
1796 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1797 ac->ac_g_ex.fe_len, &ex);
1799 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1802 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1804 /* use do_div to get remainder (would be 64-bit modulo) */
1805 if (do_div(start, sbi->s_stripe) == 0) {
1808 ext4_mb_use_best_found(ac, e4b);
1810 } else if (max >= ac->ac_g_ex.fe_len) {
1811 BUG_ON(ex.fe_len <= 0);
1812 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1813 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1816 ext4_mb_use_best_found(ac, e4b);
1817 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1818 /* Sometimes, caller may want to merge even small
1819 * number of blocks to an existing extent */
1820 BUG_ON(ex.fe_len <= 0);
1821 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1822 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1825 ext4_mb_use_best_found(ac, e4b);
1827 ext4_unlock_group(ac->ac_sb, group);
1828 ext4_mb_unload_buddy(e4b);
1834 * The routine scans buddy structures (not bitmap!) from given order
1835 * to max order and tries to find big enough chunk to satisfy the req
1837 static noinline_for_stack
1838 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1839 struct ext4_buddy *e4b)
1841 struct super_block *sb = ac->ac_sb;
1842 struct ext4_group_info *grp = e4b->bd_info;
1848 BUG_ON(ac->ac_2order <= 0);
1849 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1850 if (grp->bb_counters[i] == 0)
1853 buddy = mb_find_buddy(e4b, i, &max);
1854 BUG_ON(buddy == NULL);
1856 k = mb_find_next_zero_bit(buddy, max, 0);
1861 ac->ac_b_ex.fe_len = 1 << i;
1862 ac->ac_b_ex.fe_start = k << i;
1863 ac->ac_b_ex.fe_group = e4b->bd_group;
1865 ext4_mb_use_best_found(ac, e4b);
1867 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1869 if (EXT4_SB(sb)->s_mb_stats)
1870 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1877 * The routine scans the group and measures all found extents.
1878 * In order to optimize scanning, caller must pass number of
1879 * free blocks in the group, so the routine can know upper limit.
1881 static noinline_for_stack
1882 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1883 struct ext4_buddy *e4b)
1885 struct super_block *sb = ac->ac_sb;
1886 void *bitmap = e4b->bd_bitmap;
1887 struct ext4_free_extent ex;
1891 free = e4b->bd_info->bb_free;
1894 i = e4b->bd_info->bb_first_free;
1896 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1897 i = mb_find_next_zero_bit(bitmap,
1898 EXT4_CLUSTERS_PER_GROUP(sb), i);
1899 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1901 * IF we have corrupt bitmap, we won't find any
1902 * free blocks even though group info says we
1903 * we have free blocks
1905 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1906 "%d free clusters as per "
1907 "group info. But bitmap says 0",
1912 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1913 BUG_ON(ex.fe_len <= 0);
1914 if (free < ex.fe_len) {
1915 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1916 "%d free clusters as per "
1917 "group info. But got %d blocks",
1920 * The number of free blocks differs. This mostly
1921 * indicate that the bitmap is corrupt. So exit
1922 * without claiming the space.
1927 ext4_mb_measure_extent(ac, &ex, e4b);
1933 ext4_mb_check_limits(ac, e4b, 1);
1937 * This is a special case for storages like raid5
1938 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1940 static noinline_for_stack
1941 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1942 struct ext4_buddy *e4b)
1944 struct super_block *sb = ac->ac_sb;
1945 struct ext4_sb_info *sbi = EXT4_SB(sb);
1946 void *bitmap = e4b->bd_bitmap;
1947 struct ext4_free_extent ex;
1948 ext4_fsblk_t first_group_block;
1953 BUG_ON(sbi->s_stripe == 0);
1955 /* find first stripe-aligned block in group */
1956 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1958 a = first_group_block + sbi->s_stripe - 1;
1959 do_div(a, sbi->s_stripe);
1960 i = (a * sbi->s_stripe) - first_group_block;
1962 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1963 if (!mb_test_bit(i, bitmap)) {
1964 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
1965 if (max >= sbi->s_stripe) {
1968 ext4_mb_use_best_found(ac, e4b);
1976 /* This is now called BEFORE we load the buddy bitmap. */
1977 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1978 ext4_group_t group, int cr)
1980 unsigned free, fragments;
1981 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1982 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1984 BUG_ON(cr < 0 || cr >= 4);
1986 free = grp->bb_free;
1989 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
1992 /* We only do this if the grp has never been initialized */
1993 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1994 int ret = ext4_mb_init_group(ac->ac_sb, group);
1999 fragments = grp->bb_fragments;
2005 BUG_ON(ac->ac_2order == 0);
2007 /* Avoid using the first bg of a flexgroup for data files */
2008 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2009 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2010 ((group % flex_size) == 0))
2013 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2014 (free / fragments) >= ac->ac_g_ex.fe_len)
2017 if (grp->bb_largest_free_order < ac->ac_2order)
2022 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2026 if (free >= ac->ac_g_ex.fe_len)
2038 static noinline_for_stack int
2039 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2041 ext4_group_t ngroups, group, i;
2044 struct ext4_sb_info *sbi;
2045 struct super_block *sb;
2046 struct ext4_buddy e4b;
2050 ngroups = ext4_get_groups_count(sb);
2051 /* non-extent files are limited to low blocks/groups */
2052 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2053 ngroups = sbi->s_blockfile_groups;
2055 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2057 /* first, try the goal */
2058 err = ext4_mb_find_by_goal(ac, &e4b);
2059 if (err || ac->ac_status == AC_STATUS_FOUND)
2062 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2066 * ac->ac2_order is set only if the fe_len is a power of 2
2067 * if ac2_order is set we also set criteria to 0 so that we
2068 * try exact allocation using buddy.
2070 i = fls(ac->ac_g_ex.fe_len);
2073 * We search using buddy data only if the order of the request
2074 * is greater than equal to the sbi_s_mb_order2_reqs
2075 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2077 if (i >= sbi->s_mb_order2_reqs) {
2079 * This should tell if fe_len is exactly power of 2
2081 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2082 ac->ac_2order = i - 1;
2085 /* if stream allocation is enabled, use global goal */
2086 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2087 /* TBD: may be hot point */
2088 spin_lock(&sbi->s_md_lock);
2089 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2090 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2091 spin_unlock(&sbi->s_md_lock);
2094 /* Let's just scan groups to find more-less suitable blocks */
2095 cr = ac->ac_2order ? 0 : 1;
2097 * cr == 0 try to get exact allocation,
2098 * cr == 3 try to get anything
2101 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2102 ac->ac_criteria = cr;
2104 * searching for the right group start
2105 * from the goal value specified
2107 group = ac->ac_g_ex.fe_group;
2109 for (i = 0; i < ngroups; group++, i++) {
2111 * Artificially restricted ngroups for non-extent
2112 * files makes group > ngroups possible on first loop.
2114 if (group >= ngroups)
2117 /* This now checks without needing the buddy page */
2118 if (!ext4_mb_good_group(ac, group, cr))
2121 err = ext4_mb_load_buddy(sb, group, &e4b);
2125 ext4_lock_group(sb, group);
2128 * We need to check again after locking the
2131 if (!ext4_mb_good_group(ac, group, cr)) {
2132 ext4_unlock_group(sb, group);
2133 ext4_mb_unload_buddy(&e4b);
2137 ac->ac_groups_scanned++;
2138 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2139 ext4_mb_simple_scan_group(ac, &e4b);
2140 else if (cr == 1 && sbi->s_stripe &&
2141 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2142 ext4_mb_scan_aligned(ac, &e4b);
2144 ext4_mb_complex_scan_group(ac, &e4b);
2146 ext4_unlock_group(sb, group);
2147 ext4_mb_unload_buddy(&e4b);
2149 if (ac->ac_status != AC_STATUS_CONTINUE)
2154 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2155 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2157 * We've been searching too long. Let's try to allocate
2158 * the best chunk we've found so far
2161 ext4_mb_try_best_found(ac, &e4b);
2162 if (ac->ac_status != AC_STATUS_FOUND) {
2164 * Someone more lucky has already allocated it.
2165 * The only thing we can do is just take first
2167 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2169 ac->ac_b_ex.fe_group = 0;
2170 ac->ac_b_ex.fe_start = 0;
2171 ac->ac_b_ex.fe_len = 0;
2172 ac->ac_status = AC_STATUS_CONTINUE;
2173 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2175 atomic_inc(&sbi->s_mb_lost_chunks);
2183 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2185 struct super_block *sb = seq->private;
2188 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2191 return (void *) ((unsigned long) group);
2194 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2196 struct super_block *sb = seq->private;
2200 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2203 return (void *) ((unsigned long) group);
2206 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2208 struct super_block *sb = seq->private;
2209 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2211 int err, buddy_loaded = 0;
2212 struct ext4_buddy e4b;
2213 struct ext4_group_info *grinfo;
2215 struct ext4_group_info info;
2216 ext4_grpblk_t counters[16];
2221 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2222 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2223 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2224 "group", "free", "frags", "first",
2225 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2226 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2228 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2229 sizeof(struct ext4_group_info);
2230 grinfo = ext4_get_group_info(sb, group);
2231 /* Load the group info in memory only if not already loaded. */
2232 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2233 err = ext4_mb_load_buddy(sb, group, &e4b);
2235 seq_printf(seq, "#%-5u: I/O error\n", group);
2241 memcpy(&sg, ext4_get_group_info(sb, group), i);
2244 ext4_mb_unload_buddy(&e4b);
2246 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2247 sg.info.bb_fragments, sg.info.bb_first_free);
2248 for (i = 0; i <= 13; i++)
2249 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2250 sg.info.bb_counters[i] : 0);
2251 seq_printf(seq, " ]\n");
2256 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2260 static const struct seq_operations ext4_mb_seq_groups_ops = {
2261 .start = ext4_mb_seq_groups_start,
2262 .next = ext4_mb_seq_groups_next,
2263 .stop = ext4_mb_seq_groups_stop,
2264 .show = ext4_mb_seq_groups_show,
2267 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2269 struct super_block *sb = PDE_DATA(inode);
2272 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2274 struct seq_file *m = file->private_data;
2281 static const struct file_operations ext4_mb_seq_groups_fops = {
2282 .owner = THIS_MODULE,
2283 .open = ext4_mb_seq_groups_open,
2285 .llseek = seq_lseek,
2286 .release = seq_release,
2289 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2291 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2292 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2299 * Allocate the top-level s_group_info array for the specified number
2302 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2304 struct ext4_sb_info *sbi = EXT4_SB(sb);
2306 struct ext4_group_info ***new_groupinfo;
2308 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2309 EXT4_DESC_PER_BLOCK_BITS(sb);
2310 if (size <= sbi->s_group_info_size)
2313 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2314 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2315 if (!new_groupinfo) {
2316 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2319 if (sbi->s_group_info) {
2320 memcpy(new_groupinfo, sbi->s_group_info,
2321 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2322 ext4_kvfree(sbi->s_group_info);
2324 sbi->s_group_info = new_groupinfo;
2325 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2326 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2327 sbi->s_group_info_size);
2331 /* Create and initialize ext4_group_info data for the given group. */
2332 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2333 struct ext4_group_desc *desc)
2337 struct ext4_sb_info *sbi = EXT4_SB(sb);
2338 struct ext4_group_info **meta_group_info;
2339 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2342 * First check if this group is the first of a reserved block.
2343 * If it's true, we have to allocate a new table of pointers
2344 * to ext4_group_info structures
2346 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2347 metalen = sizeof(*meta_group_info) <<
2348 EXT4_DESC_PER_BLOCK_BITS(sb);
2349 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2350 if (meta_group_info == NULL) {
2351 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2352 "for a buddy group");
2353 goto exit_meta_group_info;
2355 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2360 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2361 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2363 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
2364 if (meta_group_info[i] == NULL) {
2365 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2366 goto exit_group_info;
2368 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2369 &(meta_group_info[i]->bb_state));
2372 * initialize bb_free to be able to skip
2373 * empty groups without initialization
2375 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2376 meta_group_info[i]->bb_free =
2377 ext4_free_clusters_after_init(sb, group, desc);
2379 meta_group_info[i]->bb_free =
2380 ext4_free_group_clusters(sb, desc);
2383 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2384 init_rwsem(&meta_group_info[i]->alloc_sem);
2385 meta_group_info[i]->bb_free_root = RB_ROOT;
2386 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2390 struct buffer_head *bh;
2391 meta_group_info[i]->bb_bitmap =
2392 kmalloc(sb->s_blocksize, GFP_KERNEL);
2393 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2394 bh = ext4_read_block_bitmap(sb, group);
2396 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2405 /* If a meta_group_info table has been allocated, release it now */
2406 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2407 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2408 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2410 exit_meta_group_info:
2412 } /* ext4_mb_add_groupinfo */
2414 static int ext4_mb_init_backend(struct super_block *sb)
2416 ext4_group_t ngroups = ext4_get_groups_count(sb);
2418 struct ext4_sb_info *sbi = EXT4_SB(sb);
2420 struct ext4_group_desc *desc;
2421 struct kmem_cache *cachep;
2423 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2427 sbi->s_buddy_cache = new_inode(sb);
2428 if (sbi->s_buddy_cache == NULL) {
2429 ext4_msg(sb, KERN_ERR, "can't get new inode");
2432 /* To avoid potentially colliding with an valid on-disk inode number,
2433 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2434 * not in the inode hash, so it should never be found by iget(), but
2435 * this will avoid confusion if it ever shows up during debugging. */
2436 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2437 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2438 for (i = 0; i < ngroups; i++) {
2439 desc = ext4_get_group_desc(sb, i, NULL);
2441 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2444 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2451 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2453 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2454 i = sbi->s_group_info_size;
2456 kfree(sbi->s_group_info[i]);
2457 iput(sbi->s_buddy_cache);
2459 ext4_kvfree(sbi->s_group_info);
2463 static void ext4_groupinfo_destroy_slabs(void)
2467 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2468 if (ext4_groupinfo_caches[i])
2469 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2470 ext4_groupinfo_caches[i] = NULL;
2474 static int ext4_groupinfo_create_slab(size_t size)
2476 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2478 int blocksize_bits = order_base_2(size);
2479 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2480 struct kmem_cache *cachep;
2482 if (cache_index >= NR_GRPINFO_CACHES)
2485 if (unlikely(cache_index < 0))
2488 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2489 if (ext4_groupinfo_caches[cache_index]) {
2490 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2491 return 0; /* Already created */
2494 slab_size = offsetof(struct ext4_group_info,
2495 bb_counters[blocksize_bits + 2]);
2497 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2498 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2501 ext4_groupinfo_caches[cache_index] = cachep;
2503 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2506 "EXT4-fs: no memory for groupinfo slab cache\n");
2513 int ext4_mb_init(struct super_block *sb)
2515 struct ext4_sb_info *sbi = EXT4_SB(sb);
2521 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2523 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2524 if (sbi->s_mb_offsets == NULL) {
2529 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2530 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2531 if (sbi->s_mb_maxs == NULL) {
2536 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2540 /* order 0 is regular bitmap */
2541 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2542 sbi->s_mb_offsets[0] = 0;
2546 max = sb->s_blocksize << 2;
2548 sbi->s_mb_offsets[i] = offset;
2549 sbi->s_mb_maxs[i] = max;
2550 offset += 1 << (sb->s_blocksize_bits - i);
2553 } while (i <= sb->s_blocksize_bits + 1);
2555 spin_lock_init(&sbi->s_md_lock);
2556 spin_lock_init(&sbi->s_bal_lock);
2558 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2559 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2560 sbi->s_mb_stats = MB_DEFAULT_STATS;
2561 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2562 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2564 * The default group preallocation is 512, which for 4k block
2565 * sizes translates to 2 megabytes. However for bigalloc file
2566 * systems, this is probably too big (i.e, if the cluster size
2567 * is 1 megabyte, then group preallocation size becomes half a
2568 * gigabyte!). As a default, we will keep a two megabyte
2569 * group pralloc size for cluster sizes up to 64k, and after
2570 * that, we will force a minimum group preallocation size of
2571 * 32 clusters. This translates to 8 megs when the cluster
2572 * size is 256k, and 32 megs when the cluster size is 1 meg,
2573 * which seems reasonable as a default.
2575 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2576 sbi->s_cluster_bits, 32);
2578 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2579 * to the lowest multiple of s_stripe which is bigger than
2580 * the s_mb_group_prealloc as determined above. We want
2581 * the preallocation size to be an exact multiple of the
2582 * RAID stripe size so that preallocations don't fragment
2585 if (sbi->s_stripe > 1) {
2586 sbi->s_mb_group_prealloc = roundup(
2587 sbi->s_mb_group_prealloc, sbi->s_stripe);
2590 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2591 if (sbi->s_locality_groups == NULL) {
2593 goto out_free_groupinfo_slab;
2595 for_each_possible_cpu(i) {
2596 struct ext4_locality_group *lg;
2597 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2598 mutex_init(&lg->lg_mutex);
2599 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2600 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2601 spin_lock_init(&lg->lg_prealloc_lock);
2604 /* init file for buddy data */
2605 ret = ext4_mb_init_backend(sb);
2607 goto out_free_locality_groups;
2610 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2611 &ext4_mb_seq_groups_fops, sb);
2615 out_free_locality_groups:
2616 free_percpu(sbi->s_locality_groups);
2617 sbi->s_locality_groups = NULL;
2618 out_free_groupinfo_slab:
2619 ext4_groupinfo_destroy_slabs();
2621 kfree(sbi->s_mb_offsets);
2622 sbi->s_mb_offsets = NULL;
2623 kfree(sbi->s_mb_maxs);
2624 sbi->s_mb_maxs = NULL;
2628 /* need to called with the ext4 group lock held */
2629 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2631 struct ext4_prealloc_space *pa;
2632 struct list_head *cur, *tmp;
2635 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2636 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2637 list_del(&pa->pa_group_list);
2639 kmem_cache_free(ext4_pspace_cachep, pa);
2642 mb_debug(1, "mballoc: %u PAs left\n", count);
2646 int ext4_mb_release(struct super_block *sb)
2648 ext4_group_t ngroups = ext4_get_groups_count(sb);
2650 int num_meta_group_infos;
2651 struct ext4_group_info *grinfo;
2652 struct ext4_sb_info *sbi = EXT4_SB(sb);
2653 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2656 remove_proc_entry("mb_groups", sbi->s_proc);
2658 if (sbi->s_group_info) {
2659 for (i = 0; i < ngroups; i++) {
2660 grinfo = ext4_get_group_info(sb, i);
2662 kfree(grinfo->bb_bitmap);
2664 ext4_lock_group(sb, i);
2665 ext4_mb_cleanup_pa(grinfo);
2666 ext4_unlock_group(sb, i);
2667 kmem_cache_free(cachep, grinfo);
2669 num_meta_group_infos = (ngroups +
2670 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2671 EXT4_DESC_PER_BLOCK_BITS(sb);
2672 for (i = 0; i < num_meta_group_infos; i++)
2673 kfree(sbi->s_group_info[i]);
2674 ext4_kvfree(sbi->s_group_info);
2676 kfree(sbi->s_mb_offsets);
2677 kfree(sbi->s_mb_maxs);
2678 if (sbi->s_buddy_cache)
2679 iput(sbi->s_buddy_cache);
2680 if (sbi->s_mb_stats) {
2681 ext4_msg(sb, KERN_INFO,
2682 "mballoc: %u blocks %u reqs (%u success)",
2683 atomic_read(&sbi->s_bal_allocated),
2684 atomic_read(&sbi->s_bal_reqs),
2685 atomic_read(&sbi->s_bal_success));
2686 ext4_msg(sb, KERN_INFO,
2687 "mballoc: %u extents scanned, %u goal hits, "
2688 "%u 2^N hits, %u breaks, %u lost",
2689 atomic_read(&sbi->s_bal_ex_scanned),
2690 atomic_read(&sbi->s_bal_goals),
2691 atomic_read(&sbi->s_bal_2orders),
2692 atomic_read(&sbi->s_bal_breaks),
2693 atomic_read(&sbi->s_mb_lost_chunks));
2694 ext4_msg(sb, KERN_INFO,
2695 "mballoc: %lu generated and it took %Lu",
2696 sbi->s_mb_buddies_generated,
2697 sbi->s_mb_generation_time);
2698 ext4_msg(sb, KERN_INFO,
2699 "mballoc: %u preallocated, %u discarded",
2700 atomic_read(&sbi->s_mb_preallocated),
2701 atomic_read(&sbi->s_mb_discarded));
2704 free_percpu(sbi->s_locality_groups);
2709 static inline int ext4_issue_discard(struct super_block *sb,
2710 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
2711 unsigned long flags)
2713 ext4_fsblk_t discard_block;
2715 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2716 ext4_group_first_block_no(sb, block_group));
2717 count = EXT4_C2B(EXT4_SB(sb), count);
2718 trace_ext4_discard_blocks(sb,
2719 (unsigned long long) discard_block, count);
2720 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, flags);
2724 * This function is called by the jbd2 layer once the commit has finished,
2725 * so we know we can free the blocks that were released with that commit.
2727 static void ext4_free_data_callback(struct super_block *sb,
2728 struct ext4_journal_cb_entry *jce,
2731 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2732 struct ext4_buddy e4b;
2733 struct ext4_group_info *db;
2734 int err, count = 0, count2 = 0;
2736 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2737 entry->efd_count, entry->efd_group, entry);
2739 if (test_opt(sb, DISCARD)) {
2740 err = ext4_issue_discard(sb, entry->efd_group,
2741 entry->efd_start_cluster,
2742 entry->efd_count, 0);
2743 if (err && err != -EOPNOTSUPP)
2744 ext4_msg(sb, KERN_WARNING, "discard request in"
2745 " group:%d block:%d count:%d failed"
2746 " with %d", entry->efd_group,
2747 entry->efd_start_cluster,
2748 entry->efd_count, err);
2751 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2752 /* we expect to find existing buddy because it's pinned */
2757 /* there are blocks to put in buddy to make them really free */
2758 count += entry->efd_count;
2760 ext4_lock_group(sb, entry->efd_group);
2761 /* Take it out of per group rb tree */
2762 rb_erase(&entry->efd_node, &(db->bb_free_root));
2763 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2766 * Clear the trimmed flag for the group so that the next
2767 * ext4_trim_fs can trim it.
2768 * If the volume is mounted with -o discard, online discard
2769 * is supported and the free blocks will be trimmed online.
2771 if (!test_opt(sb, DISCARD))
2772 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2774 if (!db->bb_free_root.rb_node) {
2775 /* No more items in the per group rb tree
2776 * balance refcounts from ext4_mb_free_metadata()
2778 page_cache_release(e4b.bd_buddy_page);
2779 page_cache_release(e4b.bd_bitmap_page);
2781 ext4_unlock_group(sb, entry->efd_group);
2782 kmem_cache_free(ext4_free_data_cachep, entry);
2783 ext4_mb_unload_buddy(&e4b);
2785 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2788 int __init ext4_init_mballoc(void)
2790 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2791 SLAB_RECLAIM_ACCOUNT);
2792 if (ext4_pspace_cachep == NULL)
2795 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2796 SLAB_RECLAIM_ACCOUNT);
2797 if (ext4_ac_cachep == NULL) {
2798 kmem_cache_destroy(ext4_pspace_cachep);
2802 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2803 SLAB_RECLAIM_ACCOUNT);
2804 if (ext4_free_data_cachep == NULL) {
2805 kmem_cache_destroy(ext4_pspace_cachep);
2806 kmem_cache_destroy(ext4_ac_cachep);
2812 void ext4_exit_mballoc(void)
2815 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2816 * before destroying the slab cache.
2819 kmem_cache_destroy(ext4_pspace_cachep);
2820 kmem_cache_destroy(ext4_ac_cachep);
2821 kmem_cache_destroy(ext4_free_data_cachep);
2822 ext4_groupinfo_destroy_slabs();
2827 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2828 * Returns 0 if success or error code
2830 static noinline_for_stack int
2831 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2832 handle_t *handle, unsigned int reserv_clstrs)
2834 struct buffer_head *bitmap_bh = NULL;
2835 struct ext4_group_desc *gdp;
2836 struct buffer_head *gdp_bh;
2837 struct ext4_sb_info *sbi;
2838 struct super_block *sb;
2842 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2843 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2849 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2853 err = ext4_journal_get_write_access(handle, bitmap_bh);
2858 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2862 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2863 ext4_free_group_clusters(sb, gdp));
2865 err = ext4_journal_get_write_access(handle, gdp_bh);
2869 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2871 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2872 if (!ext4_data_block_valid(sbi, block, len)) {
2873 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2874 "fs metadata", block, block+len);
2875 /* File system mounted not to panic on error
2876 * Fix the bitmap and repeat the block allocation
2877 * We leak some of the blocks here.
2879 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2880 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2881 ac->ac_b_ex.fe_len);
2882 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2883 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2889 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2890 #ifdef AGGRESSIVE_CHECK
2893 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2894 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2895 bitmap_bh->b_data));
2899 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2900 ac->ac_b_ex.fe_len);
2901 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2902 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2903 ext4_free_group_clusters_set(sb, gdp,
2904 ext4_free_clusters_after_init(sb,
2905 ac->ac_b_ex.fe_group, gdp));
2907 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2908 ext4_free_group_clusters_set(sb, gdp, len);
2909 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2910 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2912 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2913 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2915 * Now reduce the dirty block count also. Should not go negative
2917 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2918 /* release all the reserved blocks if non delalloc */
2919 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2922 if (sbi->s_log_groups_per_flex) {
2923 ext4_group_t flex_group = ext4_flex_group(sbi,
2924 ac->ac_b_ex.fe_group);
2925 atomic64_sub(ac->ac_b_ex.fe_len,
2926 &sbi->s_flex_groups[flex_group].free_clusters);
2929 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2932 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2940 * here we normalize request for locality group
2941 * Group request are normalized to s_mb_group_prealloc, which goes to
2942 * s_strip if we set the same via mount option.
2943 * s_mb_group_prealloc can be configured via
2944 * /sys/fs/ext4/<partition>/mb_group_prealloc
2946 * XXX: should we try to preallocate more than the group has now?
2948 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2950 struct super_block *sb = ac->ac_sb;
2951 struct ext4_locality_group *lg = ac->ac_lg;
2954 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2955 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2956 current->pid, ac->ac_g_ex.fe_len);
2960 * Normalization means making request better in terms of
2961 * size and alignment
2963 static noinline_for_stack void
2964 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2965 struct ext4_allocation_request *ar)
2967 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2970 loff_t size, start_off;
2971 loff_t orig_size __maybe_unused;
2973 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2974 struct ext4_prealloc_space *pa;
2976 /* do normalize only data requests, metadata requests
2977 do not need preallocation */
2978 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2981 /* sometime caller may want exact blocks */
2982 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2985 /* caller may indicate that preallocation isn't
2986 * required (it's a tail, for example) */
2987 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2990 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2991 ext4_mb_normalize_group_request(ac);
2995 bsbits = ac->ac_sb->s_blocksize_bits;
2997 /* first, let's learn actual file size
2998 * given current request is allocated */
2999 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3000 size = size << bsbits;
3001 if (size < i_size_read(ac->ac_inode))
3002 size = i_size_read(ac->ac_inode);
3005 /* max size of free chunks */
3008 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3009 (req <= (size) || max <= (chunk_size))
3011 /* first, try to predict filesize */
3012 /* XXX: should this table be tunable? */
3014 if (size <= 16 * 1024) {
3016 } else if (size <= 32 * 1024) {
3018 } else if (size <= 64 * 1024) {
3020 } else if (size <= 128 * 1024) {
3022 } else if (size <= 256 * 1024) {
3024 } else if (size <= 512 * 1024) {
3026 } else if (size <= 1024 * 1024) {
3028 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3029 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3030 (21 - bsbits)) << 21;
3031 size = 2 * 1024 * 1024;
3032 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3033 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3034 (22 - bsbits)) << 22;
3035 size = 4 * 1024 * 1024;
3036 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3037 (8<<20)>>bsbits, max, 8 * 1024)) {
3038 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3039 (23 - bsbits)) << 23;
3040 size = 8 * 1024 * 1024;
3042 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3043 size = ac->ac_o_ex.fe_len << bsbits;
3045 size = size >> bsbits;
3046 start = start_off >> bsbits;
3048 /* don't cover already allocated blocks in selected range */
3049 if (ar->pleft && start <= ar->lleft) {
3050 size -= ar->lleft + 1 - start;
3051 start = ar->lleft + 1;
3053 if (ar->pright && start + size - 1 >= ar->lright)
3054 size -= start + size - ar->lright;
3058 /* check we don't cross already preallocated blocks */
3060 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3065 spin_lock(&pa->pa_lock);
3066 if (pa->pa_deleted) {
3067 spin_unlock(&pa->pa_lock);
3071 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3074 /* PA must not overlap original request */
3075 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3076 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3078 /* skip PAs this normalized request doesn't overlap with */
3079 if (pa->pa_lstart >= end || pa_end <= start) {
3080 spin_unlock(&pa->pa_lock);
3083 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3085 /* adjust start or end to be adjacent to this pa */
3086 if (pa_end <= ac->ac_o_ex.fe_logical) {
3087 BUG_ON(pa_end < start);
3089 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3090 BUG_ON(pa->pa_lstart > end);
3091 end = pa->pa_lstart;
3093 spin_unlock(&pa->pa_lock);
3098 /* XXX: extra loop to check we really don't overlap preallocations */
3100 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3103 spin_lock(&pa->pa_lock);
3104 if (pa->pa_deleted == 0) {
3105 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3107 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3109 spin_unlock(&pa->pa_lock);
3113 if (start + size <= ac->ac_o_ex.fe_logical &&
3114 start > ac->ac_o_ex.fe_logical) {
3115 ext4_msg(ac->ac_sb, KERN_ERR,
3116 "start %lu, size %lu, fe_logical %lu",
3117 (unsigned long) start, (unsigned long) size,
3118 (unsigned long) ac->ac_o_ex.fe_logical);
3120 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3121 start > ac->ac_o_ex.fe_logical);
3122 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3124 /* now prepare goal request */
3126 /* XXX: is it better to align blocks WRT to logical
3127 * placement or satisfy big request as is */
3128 ac->ac_g_ex.fe_logical = start;
3129 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3131 /* define goal start in order to merge */
3132 if (ar->pright && (ar->lright == (start + size))) {
3133 /* merge to the right */
3134 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3135 &ac->ac_f_ex.fe_group,
3136 &ac->ac_f_ex.fe_start);
3137 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3139 if (ar->pleft && (ar->lleft + 1 == start)) {
3140 /* merge to the left */
3141 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3142 &ac->ac_f_ex.fe_group,
3143 &ac->ac_f_ex.fe_start);
3144 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3147 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3148 (unsigned) orig_size, (unsigned) start);
3151 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3153 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3155 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3156 atomic_inc(&sbi->s_bal_reqs);
3157 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3158 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3159 atomic_inc(&sbi->s_bal_success);
3160 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3161 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3162 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3163 atomic_inc(&sbi->s_bal_goals);
3164 if (ac->ac_found > sbi->s_mb_max_to_scan)
3165 atomic_inc(&sbi->s_bal_breaks);
3168 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3169 trace_ext4_mballoc_alloc(ac);
3171 trace_ext4_mballoc_prealloc(ac);
3175 * Called on failure; free up any blocks from the inode PA for this
3176 * context. We don't need this for MB_GROUP_PA because we only change
3177 * pa_free in ext4_mb_release_context(), but on failure, we've already
3178 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3180 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3182 struct ext4_prealloc_space *pa = ac->ac_pa;
3183 struct ext4_buddy e4b;
3187 if (ac->ac_f_ex.fe_len == 0)
3189 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3192 * This should never happen since we pin the
3193 * pages in the ext4_allocation_context so
3194 * ext4_mb_load_buddy() should never fail.
3196 WARN(1, "mb_load_buddy failed (%d)", err);
3199 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3200 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3201 ac->ac_f_ex.fe_len);
3202 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3203 ext4_mb_unload_buddy(&e4b);
3206 if (pa->pa_type == MB_INODE_PA)
3207 pa->pa_free += ac->ac_b_ex.fe_len;
3211 * use blocks preallocated to inode
3213 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3214 struct ext4_prealloc_space *pa)
3216 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3221 /* found preallocated blocks, use them */
3222 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3223 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3224 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3225 len = EXT4_NUM_B2C(sbi, end - start);
3226 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3227 &ac->ac_b_ex.fe_start);
3228 ac->ac_b_ex.fe_len = len;
3229 ac->ac_status = AC_STATUS_FOUND;
3232 BUG_ON(start < pa->pa_pstart);
3233 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3234 BUG_ON(pa->pa_free < len);
3237 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3241 * use blocks preallocated to locality group
3243 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3244 struct ext4_prealloc_space *pa)
3246 unsigned int len = ac->ac_o_ex.fe_len;
3248 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3249 &ac->ac_b_ex.fe_group,
3250 &ac->ac_b_ex.fe_start);
3251 ac->ac_b_ex.fe_len = len;
3252 ac->ac_status = AC_STATUS_FOUND;
3255 /* we don't correct pa_pstart or pa_plen here to avoid
3256 * possible race when the group is being loaded concurrently
3257 * instead we correct pa later, after blocks are marked
3258 * in on-disk bitmap -- see ext4_mb_release_context()
3259 * Other CPUs are prevented from allocating from this pa by lg_mutex
3261 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3265 * Return the prealloc space that have minimal distance
3266 * from the goal block. @cpa is the prealloc
3267 * space that is having currently known minimal distance
3268 * from the goal block.
3270 static struct ext4_prealloc_space *
3271 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3272 struct ext4_prealloc_space *pa,
3273 struct ext4_prealloc_space *cpa)
3275 ext4_fsblk_t cur_distance, new_distance;
3278 atomic_inc(&pa->pa_count);
3281 cur_distance = abs(goal_block - cpa->pa_pstart);
3282 new_distance = abs(goal_block - pa->pa_pstart);
3284 if (cur_distance <= new_distance)
3287 /* drop the previous reference */
3288 atomic_dec(&cpa->pa_count);
3289 atomic_inc(&pa->pa_count);
3294 * search goal blocks in preallocated space
3296 static noinline_for_stack int
3297 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3299 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3301 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3302 struct ext4_locality_group *lg;
3303 struct ext4_prealloc_space *pa, *cpa = NULL;
3304 ext4_fsblk_t goal_block;
3306 /* only data can be preallocated */
3307 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3310 /* first, try per-file preallocation */
3312 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3314 /* all fields in this condition don't change,
3315 * so we can skip locking for them */
3316 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3317 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3318 EXT4_C2B(sbi, pa->pa_len)))
3321 /* non-extent files can't have physical blocks past 2^32 */
3322 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3323 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3324 EXT4_MAX_BLOCK_FILE_PHYS))
3327 /* found preallocated blocks, use them */
3328 spin_lock(&pa->pa_lock);
3329 if (pa->pa_deleted == 0 && pa->pa_free) {
3330 atomic_inc(&pa->pa_count);
3331 ext4_mb_use_inode_pa(ac, pa);
3332 spin_unlock(&pa->pa_lock);
3333 ac->ac_criteria = 10;
3337 spin_unlock(&pa->pa_lock);
3341 /* can we use group allocation? */
3342 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3345 /* inode may have no locality group for some reason */
3349 order = fls(ac->ac_o_ex.fe_len) - 1;
3350 if (order > PREALLOC_TB_SIZE - 1)
3351 /* The max size of hash table is PREALLOC_TB_SIZE */
3352 order = PREALLOC_TB_SIZE - 1;
3354 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3356 * search for the prealloc space that is having
3357 * minimal distance from the goal block.
3359 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3361 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3363 spin_lock(&pa->pa_lock);
3364 if (pa->pa_deleted == 0 &&
3365 pa->pa_free >= ac->ac_o_ex.fe_len) {
3367 cpa = ext4_mb_check_group_pa(goal_block,
3370 spin_unlock(&pa->pa_lock);
3375 ext4_mb_use_group_pa(ac, cpa);
3376 ac->ac_criteria = 20;
3383 * the function goes through all block freed in the group
3384 * but not yet committed and marks them used in in-core bitmap.
3385 * buddy must be generated from this bitmap
3386 * Need to be called with the ext4 group lock held
3388 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3392 struct ext4_group_info *grp;
3393 struct ext4_free_data *entry;
3395 grp = ext4_get_group_info(sb, group);
3396 n = rb_first(&(grp->bb_free_root));
3399 entry = rb_entry(n, struct ext4_free_data, efd_node);
3400 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3407 * the function goes through all preallocation in this group and marks them
3408 * used in in-core bitmap. buddy must be generated from this bitmap
3409 * Need to be called with ext4 group lock held
3411 static noinline_for_stack
3412 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3415 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3416 struct ext4_prealloc_space *pa;
3417 struct list_head *cur;
3418 ext4_group_t groupnr;
3419 ext4_grpblk_t start;
3420 int preallocated = 0;
3423 /* all form of preallocation discards first load group,
3424 * so the only competing code is preallocation use.
3425 * we don't need any locking here
3426 * notice we do NOT ignore preallocations with pa_deleted
3427 * otherwise we could leave used blocks available for
3428 * allocation in buddy when concurrent ext4_mb_put_pa()
3429 * is dropping preallocation
3431 list_for_each(cur, &grp->bb_prealloc_list) {
3432 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3433 spin_lock(&pa->pa_lock);
3434 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3437 spin_unlock(&pa->pa_lock);
3438 if (unlikely(len == 0))
3440 BUG_ON(groupnr != group);
3441 ext4_set_bits(bitmap, start, len);
3442 preallocated += len;
3444 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3447 static void ext4_mb_pa_callback(struct rcu_head *head)
3449 struct ext4_prealloc_space *pa;
3450 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3452 BUG_ON(atomic_read(&pa->pa_count));
3453 BUG_ON(pa->pa_deleted == 0);
3454 kmem_cache_free(ext4_pspace_cachep, pa);
3458 * drops a reference to preallocated space descriptor
3459 * if this was the last reference and the space is consumed
3461 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3462 struct super_block *sb, struct ext4_prealloc_space *pa)
3465 ext4_fsblk_t grp_blk;
3467 /* in this short window concurrent discard can set pa_deleted */
3468 spin_lock(&pa->pa_lock);
3469 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3470 spin_unlock(&pa->pa_lock);
3474 if (pa->pa_deleted == 1) {
3475 spin_unlock(&pa->pa_lock);
3480 spin_unlock(&pa->pa_lock);
3482 grp_blk = pa->pa_pstart;
3484 * If doing group-based preallocation, pa_pstart may be in the
3485 * next group when pa is used up
3487 if (pa->pa_type == MB_GROUP_PA)
3490 grp = ext4_get_group_number(sb, grp_blk);
3495 * P1 (buddy init) P2 (regular allocation)
3496 * find block B in PA
3497 * copy on-disk bitmap to buddy
3498 * mark B in on-disk bitmap
3499 * drop PA from group
3500 * mark all PAs in buddy
3502 * thus, P1 initializes buddy with B available. to prevent this
3503 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3506 ext4_lock_group(sb, grp);
3507 list_del(&pa->pa_group_list);
3508 ext4_unlock_group(sb, grp);
3510 spin_lock(pa->pa_obj_lock);
3511 list_del_rcu(&pa->pa_inode_list);
3512 spin_unlock(pa->pa_obj_lock);
3514 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3518 * creates new preallocated space for given inode
3520 static noinline_for_stack int
3521 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3523 struct super_block *sb = ac->ac_sb;
3524 struct ext4_sb_info *sbi = EXT4_SB(sb);
3525 struct ext4_prealloc_space *pa;
3526 struct ext4_group_info *grp;
3527 struct ext4_inode_info *ei;
3529 /* preallocate only when found space is larger then requested */
3530 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3531 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3532 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3534 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3538 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3544 /* we can't allocate as much as normalizer wants.
3545 * so, found space must get proper lstart
3546 * to cover original request */
3547 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3548 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3550 /* we're limited by original request in that
3551 * logical block must be covered any way
3552 * winl is window we can move our chunk within */
3553 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3555 /* also, we should cover whole original request */
3556 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3558 /* the smallest one defines real window */
3559 win = min(winl, wins);
3561 offs = ac->ac_o_ex.fe_logical %
3562 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3563 if (offs && offs < win)
3566 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3567 EXT4_NUM_B2C(sbi, win);
3568 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3569 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3572 /* preallocation can change ac_b_ex, thus we store actually
3573 * allocated blocks for history */
3574 ac->ac_f_ex = ac->ac_b_ex;
3576 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3577 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3578 pa->pa_len = ac->ac_b_ex.fe_len;
3579 pa->pa_free = pa->pa_len;
3580 atomic_set(&pa->pa_count, 1);
3581 spin_lock_init(&pa->pa_lock);
3582 INIT_LIST_HEAD(&pa->pa_inode_list);
3583 INIT_LIST_HEAD(&pa->pa_group_list);
3585 pa->pa_type = MB_INODE_PA;
3587 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3588 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3589 trace_ext4_mb_new_inode_pa(ac, pa);
3591 ext4_mb_use_inode_pa(ac, pa);
3592 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3594 ei = EXT4_I(ac->ac_inode);
3595 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3597 pa->pa_obj_lock = &ei->i_prealloc_lock;
3598 pa->pa_inode = ac->ac_inode;
3600 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3601 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3602 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3604 spin_lock(pa->pa_obj_lock);
3605 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3606 spin_unlock(pa->pa_obj_lock);
3612 * creates new preallocated space for locality group inodes belongs to
3614 static noinline_for_stack int
3615 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3617 struct super_block *sb = ac->ac_sb;
3618 struct ext4_locality_group *lg;
3619 struct ext4_prealloc_space *pa;
3620 struct ext4_group_info *grp;
3622 /* preallocate only when found space is larger then requested */
3623 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3624 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3625 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3627 BUG_ON(ext4_pspace_cachep == NULL);
3628 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3632 /* preallocation can change ac_b_ex, thus we store actually
3633 * allocated blocks for history */
3634 ac->ac_f_ex = ac->ac_b_ex;
3636 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3637 pa->pa_lstart = pa->pa_pstart;
3638 pa->pa_len = ac->ac_b_ex.fe_len;
3639 pa->pa_free = pa->pa_len;
3640 atomic_set(&pa->pa_count, 1);
3641 spin_lock_init(&pa->pa_lock);
3642 INIT_LIST_HEAD(&pa->pa_inode_list);
3643 INIT_LIST_HEAD(&pa->pa_group_list);
3645 pa->pa_type = MB_GROUP_PA;
3647 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3648 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3649 trace_ext4_mb_new_group_pa(ac, pa);
3651 ext4_mb_use_group_pa(ac, pa);
3652 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3654 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3658 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3659 pa->pa_inode = NULL;
3661 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3662 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3663 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3666 * We will later add the new pa to the right bucket
3667 * after updating the pa_free in ext4_mb_release_context
3672 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3676 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3677 err = ext4_mb_new_group_pa(ac);
3679 err = ext4_mb_new_inode_pa(ac);
3684 * finds all unused blocks in on-disk bitmap, frees them in
3685 * in-core bitmap and buddy.
3686 * @pa must be unlinked from inode and group lists, so that
3687 * nobody else can find/use it.
3688 * the caller MUST hold group/inode locks.
3689 * TODO: optimize the case when there are no in-core structures yet
3691 static noinline_for_stack int
3692 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3693 struct ext4_prealloc_space *pa)
3695 struct super_block *sb = e4b->bd_sb;
3696 struct ext4_sb_info *sbi = EXT4_SB(sb);
3701 unsigned long long grp_blk_start;
3705 BUG_ON(pa->pa_deleted == 0);
3706 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3707 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3708 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3709 end = bit + pa->pa_len;
3712 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3715 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3716 mb_debug(1, " free preallocated %u/%u in group %u\n",
3717 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3718 (unsigned) next - bit, (unsigned) group);
3721 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3722 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3723 EXT4_C2B(sbi, bit)),
3725 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3728 if (free != pa->pa_free) {
3729 ext4_msg(e4b->bd_sb, KERN_CRIT,
3730 "pa %p: logic %lu, phys. %lu, len %lu",
3731 pa, (unsigned long) pa->pa_lstart,
3732 (unsigned long) pa->pa_pstart,
3733 (unsigned long) pa->pa_len);
3734 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3737 * pa is already deleted so we use the value obtained
3738 * from the bitmap and continue.
3741 atomic_add(free, &sbi->s_mb_discarded);
3746 static noinline_for_stack int
3747 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3748 struct ext4_prealloc_space *pa)
3750 struct super_block *sb = e4b->bd_sb;
3754 trace_ext4_mb_release_group_pa(sb, pa);
3755 BUG_ON(pa->pa_deleted == 0);
3756 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3757 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3758 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3759 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3760 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3766 * releases all preallocations in given group
3768 * first, we need to decide discard policy:
3769 * - when do we discard
3771 * - how many do we discard
3772 * 1) how many requested
3774 static noinline_for_stack int
3775 ext4_mb_discard_group_preallocations(struct super_block *sb,
3776 ext4_group_t group, int needed)
3778 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3779 struct buffer_head *bitmap_bh = NULL;
3780 struct ext4_prealloc_space *pa, *tmp;
3781 struct list_head list;
3782 struct ext4_buddy e4b;
3787 mb_debug(1, "discard preallocation for group %u\n", group);
3789 if (list_empty(&grp->bb_prealloc_list))
3792 bitmap_bh = ext4_read_block_bitmap(sb, group);
3793 if (bitmap_bh == NULL) {
3794 ext4_error(sb, "Error reading block bitmap for %u", group);
3798 err = ext4_mb_load_buddy(sb, group, &e4b);
3800 ext4_error(sb, "Error loading buddy information for %u", group);
3806 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3808 INIT_LIST_HEAD(&list);
3810 ext4_lock_group(sb, group);
3811 list_for_each_entry_safe(pa, tmp,
3812 &grp->bb_prealloc_list, pa_group_list) {
3813 spin_lock(&pa->pa_lock);
3814 if (atomic_read(&pa->pa_count)) {
3815 spin_unlock(&pa->pa_lock);
3819 if (pa->pa_deleted) {
3820 spin_unlock(&pa->pa_lock);
3824 /* seems this one can be freed ... */
3827 /* we can trust pa_free ... */
3828 free += pa->pa_free;
3830 spin_unlock(&pa->pa_lock);
3832 list_del(&pa->pa_group_list);
3833 list_add(&pa->u.pa_tmp_list, &list);
3836 /* if we still need more blocks and some PAs were used, try again */
3837 if (free < needed && busy) {
3839 ext4_unlock_group(sb, group);
3844 /* found anything to free? */
3845 if (list_empty(&list)) {
3850 /* now free all selected PAs */
3851 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3853 /* remove from object (inode or locality group) */
3854 spin_lock(pa->pa_obj_lock);
3855 list_del_rcu(&pa->pa_inode_list);
3856 spin_unlock(pa->pa_obj_lock);
3858 if (pa->pa_type == MB_GROUP_PA)
3859 ext4_mb_release_group_pa(&e4b, pa);
3861 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3863 list_del(&pa->u.pa_tmp_list);
3864 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3868 ext4_unlock_group(sb, group);
3869 ext4_mb_unload_buddy(&e4b);
3875 * releases all non-used preallocated blocks for given inode
3877 * It's important to discard preallocations under i_data_sem
3878 * We don't want another block to be served from the prealloc
3879 * space when we are discarding the inode prealloc space.
3881 * FIXME!! Make sure it is valid at all the call sites
3883 void ext4_discard_preallocations(struct inode *inode)
3885 struct ext4_inode_info *ei = EXT4_I(inode);
3886 struct super_block *sb = inode->i_sb;
3887 struct buffer_head *bitmap_bh = NULL;
3888 struct ext4_prealloc_space *pa, *tmp;
3889 ext4_group_t group = 0;
3890 struct list_head list;
3891 struct ext4_buddy e4b;
3894 if (!S_ISREG(inode->i_mode)) {
3895 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3899 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3900 trace_ext4_discard_preallocations(inode);
3902 INIT_LIST_HEAD(&list);
3905 /* first, collect all pa's in the inode */
3906 spin_lock(&ei->i_prealloc_lock);
3907 while (!list_empty(&ei->i_prealloc_list)) {
3908 pa = list_entry(ei->i_prealloc_list.next,
3909 struct ext4_prealloc_space, pa_inode_list);
3910 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3911 spin_lock(&pa->pa_lock);
3912 if (atomic_read(&pa->pa_count)) {
3913 /* this shouldn't happen often - nobody should
3914 * use preallocation while we're discarding it */
3915 spin_unlock(&pa->pa_lock);
3916 spin_unlock(&ei->i_prealloc_lock);
3917 ext4_msg(sb, KERN_ERR,
3918 "uh-oh! used pa while discarding");
3920 schedule_timeout_uninterruptible(HZ);
3924 if (pa->pa_deleted == 0) {
3926 spin_unlock(&pa->pa_lock);
3927 list_del_rcu(&pa->pa_inode_list);
3928 list_add(&pa->u.pa_tmp_list, &list);
3932 /* someone is deleting pa right now */
3933 spin_unlock(&pa->pa_lock);
3934 spin_unlock(&ei->i_prealloc_lock);
3936 /* we have to wait here because pa_deleted
3937 * doesn't mean pa is already unlinked from
3938 * the list. as we might be called from
3939 * ->clear_inode() the inode will get freed
3940 * and concurrent thread which is unlinking
3941 * pa from inode's list may access already
3942 * freed memory, bad-bad-bad */
3944 /* XXX: if this happens too often, we can
3945 * add a flag to force wait only in case
3946 * of ->clear_inode(), but not in case of
3947 * regular truncate */
3948 schedule_timeout_uninterruptible(HZ);
3951 spin_unlock(&ei->i_prealloc_lock);
3953 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3954 BUG_ON(pa->pa_type != MB_INODE_PA);
3955 group = ext4_get_group_number(sb, pa->pa_pstart);
3957 err = ext4_mb_load_buddy(sb, group, &e4b);
3959 ext4_error(sb, "Error loading buddy information for %u",
3964 bitmap_bh = ext4_read_block_bitmap(sb, group);
3965 if (bitmap_bh == NULL) {
3966 ext4_error(sb, "Error reading block bitmap for %u",
3968 ext4_mb_unload_buddy(&e4b);
3972 ext4_lock_group(sb, group);
3973 list_del(&pa->pa_group_list);
3974 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3975 ext4_unlock_group(sb, group);
3977 ext4_mb_unload_buddy(&e4b);
3980 list_del(&pa->u.pa_tmp_list);
3981 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3985 #ifdef CONFIG_EXT4_DEBUG
3986 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3988 struct super_block *sb = ac->ac_sb;
3989 ext4_group_t ngroups, i;
3991 if (!ext4_mballoc_debug ||
3992 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3995 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
3996 " Allocation context details:");
3997 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
3998 ac->ac_status, ac->ac_flags);
3999 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4000 "goal %lu/%lu/%lu@%lu, "
4001 "best %lu/%lu/%lu@%lu cr %d",
4002 (unsigned long)ac->ac_o_ex.fe_group,
4003 (unsigned long)ac->ac_o_ex.fe_start,
4004 (unsigned long)ac->ac_o_ex.fe_len,
4005 (unsigned long)ac->ac_o_ex.fe_logical,
4006 (unsigned long)ac->ac_g_ex.fe_group,
4007 (unsigned long)ac->ac_g_ex.fe_start,
4008 (unsigned long)ac->ac_g_ex.fe_len,
4009 (unsigned long)ac->ac_g_ex.fe_logical,
4010 (unsigned long)ac->ac_b_ex.fe_group,
4011 (unsigned long)ac->ac_b_ex.fe_start,
4012 (unsigned long)ac->ac_b_ex.fe_len,
4013 (unsigned long)ac->ac_b_ex.fe_logical,
4014 (int)ac->ac_criteria);
4015 ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found",
4016 ac->ac_ex_scanned, ac->ac_found);
4017 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4018 ngroups = ext4_get_groups_count(sb);
4019 for (i = 0; i < ngroups; i++) {
4020 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4021 struct ext4_prealloc_space *pa;
4022 ext4_grpblk_t start;
4023 struct list_head *cur;
4024 ext4_lock_group(sb, i);
4025 list_for_each(cur, &grp->bb_prealloc_list) {
4026 pa = list_entry(cur, struct ext4_prealloc_space,
4028 spin_lock(&pa->pa_lock);
4029 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4031 spin_unlock(&pa->pa_lock);
4032 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4035 ext4_unlock_group(sb, i);
4037 if (grp->bb_free == 0)
4039 printk(KERN_ERR "%u: %d/%d \n",
4040 i, grp->bb_free, grp->bb_fragments);
4042 printk(KERN_ERR "\n");
4045 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4052 * We use locality group preallocation for small size file. The size of the
4053 * file is determined by the current size or the resulting size after
4054 * allocation which ever is larger
4056 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4058 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4060 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4061 int bsbits = ac->ac_sb->s_blocksize_bits;
4064 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4067 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4070 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4071 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4074 if ((size == isize) &&
4075 !ext4_fs_is_busy(sbi) &&
4076 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4077 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4081 if (sbi->s_mb_group_prealloc <= 0) {
4082 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4086 /* don't use group allocation for large files */
4087 size = max(size, isize);
4088 if (size > sbi->s_mb_stream_request) {
4089 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4093 BUG_ON(ac->ac_lg != NULL);
4095 * locality group prealloc space are per cpu. The reason for having
4096 * per cpu locality group is to reduce the contention between block
4097 * request from multiple CPUs.
4099 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
4101 /* we're going to use group allocation */
4102 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4104 /* serialize all allocations in the group */
4105 mutex_lock(&ac->ac_lg->lg_mutex);
4108 static noinline_for_stack int
4109 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4110 struct ext4_allocation_request *ar)
4112 struct super_block *sb = ar->inode->i_sb;
4113 struct ext4_sb_info *sbi = EXT4_SB(sb);
4114 struct ext4_super_block *es = sbi->s_es;
4118 ext4_grpblk_t block;
4120 /* we can't allocate > group size */
4123 /* just a dirty hack to filter too big requests */
4124 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4125 len = EXT4_CLUSTERS_PER_GROUP(sb);
4127 /* start searching from the goal */
4129 if (goal < le32_to_cpu(es->s_first_data_block) ||
4130 goal >= ext4_blocks_count(es))
4131 goal = le32_to_cpu(es->s_first_data_block);
4132 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4134 /* set up allocation goals */
4135 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4136 ac->ac_status = AC_STATUS_CONTINUE;
4138 ac->ac_inode = ar->inode;
4139 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4140 ac->ac_o_ex.fe_group = group;
4141 ac->ac_o_ex.fe_start = block;
4142 ac->ac_o_ex.fe_len = len;
4143 ac->ac_g_ex = ac->ac_o_ex;
4144 ac->ac_flags = ar->flags;
4146 /* we have to define context: we'll we work with a file or
4147 * locality group. this is a policy, actually */
4148 ext4_mb_group_or_file(ac);
4150 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4151 "left: %u/%u, right %u/%u to %swritable\n",
4152 (unsigned) ar->len, (unsigned) ar->logical,
4153 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4154 (unsigned) ar->lleft, (unsigned) ar->pleft,
4155 (unsigned) ar->lright, (unsigned) ar->pright,
4156 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4161 static noinline_for_stack void
4162 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4163 struct ext4_locality_group *lg,
4164 int order, int total_entries)
4166 ext4_group_t group = 0;
4167 struct ext4_buddy e4b;
4168 struct list_head discard_list;
4169 struct ext4_prealloc_space *pa, *tmp;
4171 mb_debug(1, "discard locality group preallocation\n");
4173 INIT_LIST_HEAD(&discard_list);
4175 spin_lock(&lg->lg_prealloc_lock);
4176 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4178 spin_lock(&pa->pa_lock);
4179 if (atomic_read(&pa->pa_count)) {
4181 * This is the pa that we just used
4182 * for block allocation. So don't
4185 spin_unlock(&pa->pa_lock);
4188 if (pa->pa_deleted) {
4189 spin_unlock(&pa->pa_lock);
4192 /* only lg prealloc space */
4193 BUG_ON(pa->pa_type != MB_GROUP_PA);
4195 /* seems this one can be freed ... */
4197 spin_unlock(&pa->pa_lock);
4199 list_del_rcu(&pa->pa_inode_list);
4200 list_add(&pa->u.pa_tmp_list, &discard_list);
4203 if (total_entries <= 5) {
4205 * we want to keep only 5 entries
4206 * allowing it to grow to 8. This
4207 * mak sure we don't call discard
4208 * soon for this list.
4213 spin_unlock(&lg->lg_prealloc_lock);
4215 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4217 group = ext4_get_group_number(sb, pa->pa_pstart);
4218 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4219 ext4_error(sb, "Error loading buddy information for %u",
4223 ext4_lock_group(sb, group);
4224 list_del(&pa->pa_group_list);
4225 ext4_mb_release_group_pa(&e4b, pa);
4226 ext4_unlock_group(sb, group);
4228 ext4_mb_unload_buddy(&e4b);
4229 list_del(&pa->u.pa_tmp_list);
4230 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4235 * We have incremented pa_count. So it cannot be freed at this
4236 * point. Also we hold lg_mutex. So no parallel allocation is
4237 * possible from this lg. That means pa_free cannot be updated.
4239 * A parallel ext4_mb_discard_group_preallocations is possible.
4240 * which can cause the lg_prealloc_list to be updated.
4243 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4245 int order, added = 0, lg_prealloc_count = 1;
4246 struct super_block *sb = ac->ac_sb;
4247 struct ext4_locality_group *lg = ac->ac_lg;
4248 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4250 order = fls(pa->pa_free) - 1;
4251 if (order > PREALLOC_TB_SIZE - 1)
4252 /* The max size of hash table is PREALLOC_TB_SIZE */
4253 order = PREALLOC_TB_SIZE - 1;
4254 /* Add the prealloc space to lg */
4255 spin_lock(&lg->lg_prealloc_lock);
4256 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4258 spin_lock(&tmp_pa->pa_lock);
4259 if (tmp_pa->pa_deleted) {
4260 spin_unlock(&tmp_pa->pa_lock);
4263 if (!added && pa->pa_free < tmp_pa->pa_free) {
4264 /* Add to the tail of the previous entry */
4265 list_add_tail_rcu(&pa->pa_inode_list,
4266 &tmp_pa->pa_inode_list);
4269 * we want to count the total
4270 * number of entries in the list
4273 spin_unlock(&tmp_pa->pa_lock);
4274 lg_prealloc_count++;
4277 list_add_tail_rcu(&pa->pa_inode_list,
4278 &lg->lg_prealloc_list[order]);
4279 spin_unlock(&lg->lg_prealloc_lock);
4281 /* Now trim the list to be not more than 8 elements */
4282 if (lg_prealloc_count > 8) {
4283 ext4_mb_discard_lg_preallocations(sb, lg,
4284 order, lg_prealloc_count);
4291 * release all resource we used in allocation
4293 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4295 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4296 struct ext4_prealloc_space *pa = ac->ac_pa;
4298 if (pa->pa_type == MB_GROUP_PA) {
4299 /* see comment in ext4_mb_use_group_pa() */
4300 spin_lock(&pa->pa_lock);
4301 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4302 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4303 pa->pa_free -= ac->ac_b_ex.fe_len;
4304 pa->pa_len -= ac->ac_b_ex.fe_len;
4305 spin_unlock(&pa->pa_lock);
4310 * We want to add the pa to the right bucket.
4311 * Remove it from the list and while adding
4312 * make sure the list to which we are adding
4315 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4316 spin_lock(pa->pa_obj_lock);
4317 list_del_rcu(&pa->pa_inode_list);
4318 spin_unlock(pa->pa_obj_lock);
4319 ext4_mb_add_n_trim(ac);
4321 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4323 if (ac->ac_bitmap_page)
4324 page_cache_release(ac->ac_bitmap_page);
4325 if (ac->ac_buddy_page)
4326 page_cache_release(ac->ac_buddy_page);
4327 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4328 mutex_unlock(&ac->ac_lg->lg_mutex);
4329 ext4_mb_collect_stats(ac);
4333 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4335 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4339 trace_ext4_mb_discard_preallocations(sb, needed);
4340 for (i = 0; i < ngroups && needed > 0; i++) {
4341 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4350 * Main entry point into mballoc to allocate blocks
4351 * it tries to use preallocation first, then falls back
4352 * to usual allocation
4354 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4355 struct ext4_allocation_request *ar, int *errp)
4358 struct ext4_allocation_context *ac = NULL;
4359 struct ext4_sb_info *sbi;
4360 struct super_block *sb;
4361 ext4_fsblk_t block = 0;
4362 unsigned int inquota = 0;
4363 unsigned int reserv_clstrs = 0;
4366 sb = ar->inode->i_sb;
4369 trace_ext4_request_blocks(ar);
4371 /* Allow to use superuser reservation for quota file */
4372 if (IS_NOQUOTA(ar->inode))
4373 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4376 * For delayed allocation, we could skip the ENOSPC and
4377 * EDQUOT check, as blocks and quotas have been already
4378 * reserved when data being copied into pagecache.
4380 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4381 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4383 /* Without delayed allocation we need to verify
4384 * there is enough free blocks to do block allocation
4385 * and verify allocation doesn't exceed the quota limits.
4388 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4390 /* let others to free the space */
4392 ar->len = ar->len >> 1;
4398 reserv_clstrs = ar->len;
4399 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4400 dquot_alloc_block_nofail(ar->inode,
4401 EXT4_C2B(sbi, ar->len));
4404 dquot_alloc_block(ar->inode,
4405 EXT4_C2B(sbi, ar->len))) {
4407 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4418 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4425 *errp = ext4_mb_initialize_context(ac, ar);
4431 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4432 if (!ext4_mb_use_preallocated(ac)) {
4433 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4434 ext4_mb_normalize_request(ac, ar);
4436 /* allocate space in core */
4437 *errp = ext4_mb_regular_allocator(ac);
4439 ext4_discard_allocated_blocks(ac);
4443 /* as we've just preallocated more space than
4444 * user requested orinally, we store allocated
4445 * space in a special descriptor */
4446 if (ac->ac_status == AC_STATUS_FOUND &&
4447 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4448 ext4_mb_new_preallocation(ac);
4450 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4451 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4452 if (*errp == -EAGAIN) {
4454 * drop the reference that we took
4455 * in ext4_mb_use_best_found
4457 ext4_mb_release_context(ac);
4458 ac->ac_b_ex.fe_group = 0;
4459 ac->ac_b_ex.fe_start = 0;
4460 ac->ac_b_ex.fe_len = 0;
4461 ac->ac_status = AC_STATUS_CONTINUE;
4464 ext4_discard_allocated_blocks(ac);
4467 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4468 ar->len = ac->ac_b_ex.fe_len;
4471 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4479 ac->ac_b_ex.fe_len = 0;
4481 ext4_mb_show_ac(ac);
4483 ext4_mb_release_context(ac);
4486 kmem_cache_free(ext4_ac_cachep, ac);
4487 if (inquota && ar->len < inquota)
4488 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4490 if (!ext4_test_inode_state(ar->inode,
4491 EXT4_STATE_DELALLOC_RESERVED))
4492 /* release all the reserved blocks if non delalloc */
4493 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4497 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4503 * We can merge two free data extents only if the physical blocks
4504 * are contiguous, AND the extents were freed by the same transaction,
4505 * AND the blocks are associated with the same group.
4507 static int can_merge(struct ext4_free_data *entry1,
4508 struct ext4_free_data *entry2)
4510 if ((entry1->efd_tid == entry2->efd_tid) &&
4511 (entry1->efd_group == entry2->efd_group) &&
4512 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4517 static noinline_for_stack int
4518 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4519 struct ext4_free_data *new_entry)
4521 ext4_group_t group = e4b->bd_group;
4522 ext4_grpblk_t cluster;
4523 struct ext4_free_data *entry;
4524 struct ext4_group_info *db = e4b->bd_info;
4525 struct super_block *sb = e4b->bd_sb;
4526 struct ext4_sb_info *sbi = EXT4_SB(sb);
4527 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4528 struct rb_node *parent = NULL, *new_node;
4530 BUG_ON(!ext4_handle_valid(handle));
4531 BUG_ON(e4b->bd_bitmap_page == NULL);
4532 BUG_ON(e4b->bd_buddy_page == NULL);
4534 new_node = &new_entry->efd_node;
4535 cluster = new_entry->efd_start_cluster;
4538 /* first free block exent. We need to
4539 protect buddy cache from being freed,
4540 * otherwise we'll refresh it from
4541 * on-disk bitmap and lose not-yet-available
4543 page_cache_get(e4b->bd_buddy_page);
4544 page_cache_get(e4b->bd_bitmap_page);
4548 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4549 if (cluster < entry->efd_start_cluster)
4551 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4552 n = &(*n)->rb_right;
4554 ext4_grp_locked_error(sb, group, 0,
4555 ext4_group_first_block_no(sb, group) +
4556 EXT4_C2B(sbi, cluster),
4557 "Block already on to-be-freed list");
4562 rb_link_node(new_node, parent, n);
4563 rb_insert_color(new_node, &db->bb_free_root);
4565 /* Now try to see the extent can be merged to left and right */
4566 node = rb_prev(new_node);
4568 entry = rb_entry(node, struct ext4_free_data, efd_node);
4569 if (can_merge(entry, new_entry) &&
4570 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4571 new_entry->efd_start_cluster = entry->efd_start_cluster;
4572 new_entry->efd_count += entry->efd_count;
4573 rb_erase(node, &(db->bb_free_root));
4574 kmem_cache_free(ext4_free_data_cachep, entry);
4578 node = rb_next(new_node);
4580 entry = rb_entry(node, struct ext4_free_data, efd_node);
4581 if (can_merge(new_entry, entry) &&
4582 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4583 new_entry->efd_count += entry->efd_count;
4584 rb_erase(node, &(db->bb_free_root));
4585 kmem_cache_free(ext4_free_data_cachep, entry);
4588 /* Add the extent to transaction's private list */
4589 ext4_journal_callback_add(handle, ext4_free_data_callback,
4590 &new_entry->efd_jce);
4595 * ext4_free_blocks() -- Free given blocks and update quota
4596 * @handle: handle for this transaction
4598 * @block: start physical block to free
4599 * @count: number of blocks to count
4600 * @flags: flags used by ext4_free_blocks
4602 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4603 struct buffer_head *bh, ext4_fsblk_t block,
4604 unsigned long count, int flags)
4606 struct buffer_head *bitmap_bh = NULL;
4607 struct super_block *sb = inode->i_sb;
4608 struct ext4_group_desc *gdp;
4609 unsigned int overflow;
4611 struct buffer_head *gd_bh;
4612 ext4_group_t block_group;
4613 struct ext4_sb_info *sbi;
4614 struct ext4_buddy e4b;
4615 unsigned int count_clusters;
4622 BUG_ON(block != bh->b_blocknr);
4624 block = bh->b_blocknr;
4628 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4629 !ext4_data_block_valid(sbi, block, count)) {
4630 ext4_error(sb, "Freeing blocks not in datazone - "
4631 "block = %llu, count = %lu", block, count);
4635 ext4_debug("freeing block %llu\n", block);
4636 trace_ext4_free_blocks(inode, block, count, flags);
4638 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4639 struct buffer_head *tbh = bh;
4642 BUG_ON(bh && (count > 1));
4644 for (i = 0; i < count; i++) {
4646 tbh = sb_find_get_block(inode->i_sb,
4650 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4651 inode, tbh, block + i);
4656 * We need to make sure we don't reuse the freed block until
4657 * after the transaction is committed, which we can do by
4658 * treating the block as metadata, below. We make an
4659 * exception if the inode is to be written in writeback mode
4660 * since writeback mode has weak data consistency guarantees.
4662 if (!ext4_should_writeback_data(inode))
4663 flags |= EXT4_FREE_BLOCKS_METADATA;
4666 * If the extent to be freed does not begin on a cluster
4667 * boundary, we need to deal with partial clusters at the
4668 * beginning and end of the extent. Normally we will free
4669 * blocks at the beginning or the end unless we are explicitly
4670 * requested to avoid doing so.
4672 overflow = EXT4_PBLK_COFF(sbi, block);
4674 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4675 overflow = sbi->s_cluster_ratio - overflow;
4677 if (count > overflow)
4686 overflow = EXT4_LBLK_COFF(sbi, count);
4688 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4689 if (count > overflow)
4694 count += sbi->s_cluster_ratio - overflow;
4699 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4702 * Check to see if we are freeing blocks across a group
4705 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4706 overflow = EXT4_C2B(sbi, bit) + count -
4707 EXT4_BLOCKS_PER_GROUP(sb);
4710 count_clusters = EXT4_NUM_B2C(sbi, count);
4711 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4716 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4722 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4723 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4724 in_range(block, ext4_inode_table(sb, gdp),
4725 EXT4_SB(sb)->s_itb_per_group) ||
4726 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4727 EXT4_SB(sb)->s_itb_per_group)) {
4729 ext4_error(sb, "Freeing blocks in system zone - "
4730 "Block = %llu, count = %lu", block, count);
4731 /* err = 0. ext4_std_error should be a no op */
4735 BUFFER_TRACE(bitmap_bh, "getting write access");
4736 err = ext4_journal_get_write_access(handle, bitmap_bh);
4741 * We are about to modify some metadata. Call the journal APIs
4742 * to unshare ->b_data if a currently-committing transaction is
4745 BUFFER_TRACE(gd_bh, "get_write_access");
4746 err = ext4_journal_get_write_access(handle, gd_bh);
4749 #ifdef AGGRESSIVE_CHECK
4752 for (i = 0; i < count_clusters; i++)
4753 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4756 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4758 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4762 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4763 struct ext4_free_data *new_entry;
4765 * blocks being freed are metadata. these blocks shouldn't
4766 * be used until this transaction is committed
4769 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4772 * We use a retry loop because
4773 * ext4_free_blocks() is not allowed to fail.
4776 congestion_wait(BLK_RW_ASYNC, HZ/50);
4779 new_entry->efd_start_cluster = bit;
4780 new_entry->efd_group = block_group;
4781 new_entry->efd_count = count_clusters;
4782 new_entry->efd_tid = handle->h_transaction->t_tid;
4784 ext4_lock_group(sb, block_group);
4785 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4786 ext4_mb_free_metadata(handle, &e4b, new_entry);
4788 /* need to update group_info->bb_free and bitmap
4789 * with group lock held. generate_buddy look at
4790 * them with group lock_held
4792 if (test_opt(sb, DISCARD)) {
4793 err = ext4_issue_discard(sb, block_group, bit, count,
4795 if (err && err != -EOPNOTSUPP)
4796 ext4_msg(sb, KERN_WARNING, "discard request in"
4797 " group:%d block:%d count:%lu failed"
4798 " with %d", block_group, bit, count,
4801 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4803 ext4_lock_group(sb, block_group);
4804 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4805 mb_free_blocks(inode, &e4b, bit, count_clusters);
4808 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4809 ext4_free_group_clusters_set(sb, gdp, ret);
4810 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4811 ext4_group_desc_csum_set(sb, block_group, gdp);
4812 ext4_unlock_group(sb, block_group);
4813 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4815 if (sbi->s_log_groups_per_flex) {
4816 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4817 atomic64_add(count_clusters,
4818 &sbi->s_flex_groups[flex_group].free_clusters);
4821 ext4_mb_unload_buddy(&e4b);
4823 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4824 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4826 /* We dirtied the bitmap block */
4827 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4828 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4830 /* And the group descriptor block */
4831 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4832 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4836 if (overflow && !err) {
4844 ext4_std_error(sb, err);
4849 * ext4_group_add_blocks() -- Add given blocks to an existing group
4850 * @handle: handle to this transaction
4852 * @block: start physical block to add to the block group
4853 * @count: number of blocks to free
4855 * This marks the blocks as free in the bitmap and buddy.
4857 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4858 ext4_fsblk_t block, unsigned long count)
4860 struct buffer_head *bitmap_bh = NULL;
4861 struct buffer_head *gd_bh;
4862 ext4_group_t block_group;
4865 struct ext4_group_desc *desc;
4866 struct ext4_sb_info *sbi = EXT4_SB(sb);
4867 struct ext4_buddy e4b;
4868 int err = 0, ret, blk_free_count;
4869 ext4_grpblk_t blocks_freed;
4871 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4876 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4878 * Check to see if we are freeing blocks across a group
4881 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4882 ext4_warning(sb, "too much blocks added to group %u\n",
4888 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4894 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4900 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4901 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4902 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4903 in_range(block + count - 1, ext4_inode_table(sb, desc),
4904 sbi->s_itb_per_group)) {
4905 ext4_error(sb, "Adding blocks in system zones - "
4906 "Block = %llu, count = %lu",
4912 BUFFER_TRACE(bitmap_bh, "getting write access");
4913 err = ext4_journal_get_write_access(handle, bitmap_bh);
4918 * We are about to modify some metadata. Call the journal APIs
4919 * to unshare ->b_data if a currently-committing transaction is
4922 BUFFER_TRACE(gd_bh, "get_write_access");
4923 err = ext4_journal_get_write_access(handle, gd_bh);
4927 for (i = 0, blocks_freed = 0; i < count; i++) {
4928 BUFFER_TRACE(bitmap_bh, "clear bit");
4929 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4930 ext4_error(sb, "bit already cleared for block %llu",
4931 (ext4_fsblk_t)(block + i));
4932 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4938 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4943 * need to update group_info->bb_free and bitmap
4944 * with group lock held. generate_buddy look at
4945 * them with group lock_held
4947 ext4_lock_group(sb, block_group);
4948 mb_clear_bits(bitmap_bh->b_data, bit, count);
4949 mb_free_blocks(NULL, &e4b, bit, count);
4950 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4951 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4952 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
4953 ext4_group_desc_csum_set(sb, block_group, desc);
4954 ext4_unlock_group(sb, block_group);
4955 percpu_counter_add(&sbi->s_freeclusters_counter,
4956 EXT4_NUM_B2C(sbi, blocks_freed));
4958 if (sbi->s_log_groups_per_flex) {
4959 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4960 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
4961 &sbi->s_flex_groups[flex_group].free_clusters);
4964 ext4_mb_unload_buddy(&e4b);
4966 /* We dirtied the bitmap block */
4967 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4968 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4970 /* And the group descriptor block */
4971 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4972 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4978 ext4_std_error(sb, err);
4983 * ext4_trim_extent -- function to TRIM one single free extent in the group
4984 * @sb: super block for the file system
4985 * @start: starting block of the free extent in the alloc. group
4986 * @count: number of blocks to TRIM
4987 * @group: alloc. group we are working with
4988 * @e4b: ext4 buddy for the group
4989 * @blkdev_flags: flags for the block device
4991 * Trim "count" blocks starting at "start" in the "group". To assure that no
4992 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4993 * be called with under the group lock.
4995 static int ext4_trim_extent(struct super_block *sb, int start, int count,
4996 ext4_group_t group, struct ext4_buddy *e4b,
4997 unsigned long blkdev_flags)
4999 struct ext4_free_extent ex;
5002 trace_ext4_trim_extent(sb, group, start, count);
5004 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5006 ex.fe_start = start;
5007 ex.fe_group = group;
5011 * Mark blocks used, so no one can reuse them while
5014 mb_mark_used(e4b, &ex);
5015 ext4_unlock_group(sb, group);
5016 ret = ext4_issue_discard(sb, group, start, count, blkdev_flags);
5017 ext4_lock_group(sb, group);
5018 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5023 * ext4_trim_all_free -- function to trim all free space in alloc. group
5024 * @sb: super block for file system
5025 * @group: group to be trimmed
5026 * @start: first group block to examine
5027 * @max: last group block to examine
5028 * @minblocks: minimum extent block count
5029 * @blkdev_flags: flags for the block device
5031 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5032 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5036 * ext4_trim_all_free walks through group's block bitmap searching for free
5037 * extents. When the free extent is found, mark it as used in group buddy
5038 * bitmap. Then issue a TRIM command on this extent and free the extent in
5039 * the group buddy bitmap. This is done until whole group is scanned.
5041 static ext4_grpblk_t
5042 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5043 ext4_grpblk_t start, ext4_grpblk_t max,
5044 ext4_grpblk_t minblocks, unsigned long blkdev_flags)
5047 ext4_grpblk_t next, count = 0, free_count = 0;
5048 struct ext4_buddy e4b;
5051 trace_ext4_trim_all_free(sb, group, start, max);
5053 ret = ext4_mb_load_buddy(sb, group, &e4b);
5055 ext4_error(sb, "Error in loading buddy "
5056 "information for %u", group);
5059 bitmap = e4b.bd_bitmap;
5061 ext4_lock_group(sb, group);
5062 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5063 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5066 start = (e4b.bd_info->bb_first_free > start) ?
5067 e4b.bd_info->bb_first_free : start;
5069 while (start <= max) {
5070 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5073 next = mb_find_next_bit(bitmap, max + 1, start);
5075 if ((next - start) >= minblocks) {
5076 ret = ext4_trim_extent(sb, start,
5077 next - start, group, &e4b,
5079 if (ret && ret != -EOPNOTSUPP)
5082 count += next - start;
5084 free_count += next - start;
5087 if (fatal_signal_pending(current)) {
5088 count = -ERESTARTSYS;
5092 if (need_resched()) {
5093 ext4_unlock_group(sb, group);
5095 ext4_lock_group(sb, group);
5098 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5104 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5107 ext4_unlock_group(sb, group);
5108 ext4_mb_unload_buddy(&e4b);
5110 ext4_debug("trimmed %d blocks in the group %d\n",
5117 * ext4_trim_fs() -- trim ioctl handle function
5118 * @sb: superblock for filesystem
5119 * @range: fstrim_range structure
5120 * @blkdev_flags: flags for the block device
5122 * start: First Byte to trim
5123 * len: number of Bytes to trim from start
5124 * minlen: minimum extent length in Bytes
5125 * ext4_trim_fs goes through all allocation groups containing Bytes from
5126 * start to start+len. For each such a group ext4_trim_all_free function
5127 * is invoked to trim all free space.
5129 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range,
5130 unsigned long blkdev_flags)
5132 struct ext4_group_info *grp;
5133 ext4_group_t group, first_group, last_group;
5134 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5135 uint64_t start, end, minlen, trimmed = 0;
5136 ext4_fsblk_t first_data_blk =
5137 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5138 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5141 start = range->start >> sb->s_blocksize_bits;
5142 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5143 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5144 range->minlen >> sb->s_blocksize_bits);
5146 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5147 start >= max_blks ||
5148 range->len < sb->s_blocksize)
5150 if (end >= max_blks)
5152 if (end <= first_data_blk)
5154 if (start < first_data_blk)
5155 start = first_data_blk;
5157 /* Determine first and last group to examine based on start and end */
5158 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5159 &first_group, &first_cluster);
5160 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5161 &last_group, &last_cluster);
5163 /* end now represents the last cluster to discard in this group */
5164 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5166 for (group = first_group; group <= last_group; group++) {
5167 grp = ext4_get_group_info(sb, group);
5168 /* We only do this if the grp has never been initialized */
5169 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5170 ret = ext4_mb_init_group(sb, group);
5176 * For all the groups except the last one, last cluster will
5177 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5178 * change it for the last group, note that last_cluster is
5179 * already computed earlier by ext4_get_group_no_and_offset()
5181 if (group == last_group)
5184 if (grp->bb_free >= minlen) {
5185 cnt = ext4_trim_all_free(sb, group, first_cluster,
5186 end, minlen, blkdev_flags);
5195 * For every group except the first one, we are sure
5196 * that the first cluster to discard will be cluster #0.
5202 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5205 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;