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 struct ext4_sb_info *sbi = EXT4_SB(sb);
726 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
731 unsigned fragments = 0;
732 unsigned long long period = get_cycles();
734 /* initialize buddy from bitmap which is aggregation
735 * of on-disk bitmap and preallocations */
736 i = mb_find_next_zero_bit(bitmap, max, 0);
737 grp->bb_first_free = i;
741 i = mb_find_next_bit(bitmap, max, i);
745 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
747 grp->bb_counters[0]++;
749 i = mb_find_next_zero_bit(bitmap, max, i);
751 grp->bb_fragments = fragments;
753 if (free != grp->bb_free) {
754 ext4_grp_locked_error(sb, group, 0, 0,
755 "block bitmap and bg descriptor "
756 "inconsistent: %u vs %u free clusters",
759 * If we intend to continue, we consider group descriptor
760 * corrupt and update bb_free using bitmap value
763 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
764 percpu_counter_sub(&sbi->s_freeclusters_counter,
766 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
768 mb_set_largest_free_order(sb, grp);
770 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
772 period = get_cycles() - period;
773 spin_lock(&EXT4_SB(sb)->s_bal_lock);
774 EXT4_SB(sb)->s_mb_buddies_generated++;
775 EXT4_SB(sb)->s_mb_generation_time += period;
776 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
779 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
785 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
786 ext4_set_bits(buddy, 0, count);
788 e4b->bd_info->bb_fragments = 0;
789 memset(e4b->bd_info->bb_counters, 0,
790 sizeof(*e4b->bd_info->bb_counters) *
791 (e4b->bd_sb->s_blocksize_bits + 2));
793 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
794 e4b->bd_bitmap, e4b->bd_group);
797 /* The buddy information is attached the buddy cache inode
798 * for convenience. The information regarding each group
799 * is loaded via ext4_mb_load_buddy. The information involve
800 * block bitmap and buddy information. The information are
801 * stored in the inode as
804 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
807 * one block each for bitmap and buddy information.
808 * So for each group we take up 2 blocks. A page can
809 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
810 * So it can have information regarding groups_per_page which
811 * is blocks_per_page/2
813 * Locking note: This routine takes the block group lock of all groups
814 * for this page; do not hold this lock when calling this routine!
817 static int ext4_mb_init_cache(struct page *page, char *incore)
819 ext4_group_t ngroups;
825 ext4_group_t first_group, group;
827 struct super_block *sb;
828 struct buffer_head *bhs;
829 struct buffer_head **bh = NULL;
833 struct ext4_group_info *grinfo;
835 mb_debug(1, "init page %lu\n", page->index);
837 inode = page->mapping->host;
839 ngroups = ext4_get_groups_count(sb);
840 blocksize = 1 << inode->i_blkbits;
841 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
843 groups_per_page = blocks_per_page >> 1;
844 if (groups_per_page == 0)
847 /* allocate buffer_heads to read bitmaps */
848 if (groups_per_page > 1) {
849 i = sizeof(struct buffer_head *) * groups_per_page;
850 bh = kzalloc(i, GFP_NOFS);
858 first_group = page->index * blocks_per_page / 2;
860 /* read all groups the page covers into the cache */
861 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
862 if (group >= ngroups)
865 grinfo = ext4_get_group_info(sb, group);
867 * If page is uptodate then we came here after online resize
868 * which added some new uninitialized group info structs, so
869 * we must skip all initialized uptodate buddies on the page,
870 * which may be currently in use by an allocating task.
872 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
876 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
880 mb_debug(1, "read bitmap for group %u\n", group);
883 /* wait for I/O completion */
884 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
885 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i]))
889 first_block = page->index * blocks_per_page;
890 for (i = 0; i < blocks_per_page; i++) {
891 group = (first_block + i) >> 1;
892 if (group >= ngroups)
895 if (!bh[group - first_group])
896 /* skip initialized uptodate buddy */
899 if (!buffer_verified(bh[group - first_group]))
900 /* Skip faulty bitmaps */
905 * data carry information regarding this
906 * particular group in the format specified
910 data = page_address(page) + (i * blocksize);
911 bitmap = bh[group - first_group]->b_data;
914 * We place the buddy block and bitmap block
917 if ((first_block + i) & 1) {
918 /* this is block of buddy */
919 BUG_ON(incore == NULL);
920 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
921 group, page->index, i * blocksize);
922 trace_ext4_mb_buddy_bitmap_load(sb, group);
923 grinfo = ext4_get_group_info(sb, group);
924 grinfo->bb_fragments = 0;
925 memset(grinfo->bb_counters, 0,
926 sizeof(*grinfo->bb_counters) *
927 (sb->s_blocksize_bits+2));
929 * incore got set to the group block bitmap below
931 ext4_lock_group(sb, group);
933 memset(data, 0xff, blocksize);
934 ext4_mb_generate_buddy(sb, data, incore, group);
935 ext4_unlock_group(sb, group);
938 /* this is block of bitmap */
939 BUG_ON(incore != NULL);
940 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
941 group, page->index, i * blocksize);
942 trace_ext4_mb_bitmap_load(sb, group);
944 /* see comments in ext4_mb_put_pa() */
945 ext4_lock_group(sb, group);
946 memcpy(data, bitmap, blocksize);
948 /* mark all preallocated blks used in in-core bitmap */
949 ext4_mb_generate_from_pa(sb, data, group);
950 ext4_mb_generate_from_freelist(sb, data, group);
951 ext4_unlock_group(sb, group);
953 /* set incore so that the buddy information can be
954 * generated using this
959 SetPageUptodate(page);
963 for (i = 0; i < groups_per_page; i++)
972 * Lock the buddy and bitmap pages. This make sure other parallel init_group
973 * on the same buddy page doesn't happen whild holding the buddy page lock.
974 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
975 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
977 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
978 ext4_group_t group, struct ext4_buddy *e4b)
980 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
981 int block, pnum, poff;
985 e4b->bd_buddy_page = NULL;
986 e4b->bd_bitmap_page = NULL;
988 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
990 * the buddy cache inode stores the block bitmap
991 * and buddy information in consecutive blocks.
992 * So for each group we need two blocks.
995 pnum = block / blocks_per_page;
996 poff = block % blocks_per_page;
997 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1000 BUG_ON(page->mapping != inode->i_mapping);
1001 e4b->bd_bitmap_page = page;
1002 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1004 if (blocks_per_page >= 2) {
1005 /* buddy and bitmap are on the same page */
1010 pnum = block / blocks_per_page;
1011 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1014 BUG_ON(page->mapping != inode->i_mapping);
1015 e4b->bd_buddy_page = page;
1019 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1021 if (e4b->bd_bitmap_page) {
1022 unlock_page(e4b->bd_bitmap_page);
1023 page_cache_release(e4b->bd_bitmap_page);
1025 if (e4b->bd_buddy_page) {
1026 unlock_page(e4b->bd_buddy_page);
1027 page_cache_release(e4b->bd_buddy_page);
1032 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1033 * block group lock of all groups for this page; do not hold the BG lock when
1034 * calling this routine!
1036 static noinline_for_stack
1037 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1040 struct ext4_group_info *this_grp;
1041 struct ext4_buddy e4b;
1046 mb_debug(1, "init group %u\n", group);
1047 this_grp = ext4_get_group_info(sb, group);
1049 * This ensures that we don't reinit the buddy cache
1050 * page which map to the group from which we are already
1051 * allocating. If we are looking at the buddy cache we would
1052 * have taken a reference using ext4_mb_load_buddy and that
1053 * would have pinned buddy page to page cache.
1054 * The call to ext4_mb_get_buddy_page_lock will mark the
1057 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1058 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1060 * somebody initialized the group
1061 * return without doing anything
1066 page = e4b.bd_bitmap_page;
1067 ret = ext4_mb_init_cache(page, NULL);
1070 if (!PageUptodate(page)) {
1075 if (e4b.bd_buddy_page == NULL) {
1077 * If both the bitmap and buddy are in
1078 * the same page we don't need to force
1084 /* init buddy cache */
1085 page = e4b.bd_buddy_page;
1086 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1089 if (!PageUptodate(page)) {
1094 ext4_mb_put_buddy_page_lock(&e4b);
1099 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1100 * block group lock of all groups for this page; do not hold the BG lock when
1101 * calling this routine!
1103 static noinline_for_stack int
1104 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1105 struct ext4_buddy *e4b)
1107 int blocks_per_page;
1113 struct ext4_group_info *grp;
1114 struct ext4_sb_info *sbi = EXT4_SB(sb);
1115 struct inode *inode = sbi->s_buddy_cache;
1118 mb_debug(1, "load group %u\n", group);
1120 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1121 grp = ext4_get_group_info(sb, group);
1123 e4b->bd_blkbits = sb->s_blocksize_bits;
1126 e4b->bd_group = group;
1127 e4b->bd_buddy_page = NULL;
1128 e4b->bd_bitmap_page = NULL;
1130 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1132 * we need full data about the group
1133 * to make a good selection
1135 ret = ext4_mb_init_group(sb, group);
1141 * the buddy cache inode stores the block bitmap
1142 * and buddy information in consecutive blocks.
1143 * So for each group we need two blocks.
1146 pnum = block / blocks_per_page;
1147 poff = block % blocks_per_page;
1149 /* we could use find_or_create_page(), but it locks page
1150 * what we'd like to avoid in fast path ... */
1151 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1152 if (page == NULL || !PageUptodate(page)) {
1155 * drop the page reference and try
1156 * to get the page with lock. If we
1157 * are not uptodate that implies
1158 * somebody just created the page but
1159 * is yet to initialize the same. So
1160 * wait for it to initialize.
1162 page_cache_release(page);
1163 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1165 BUG_ON(page->mapping != inode->i_mapping);
1166 if (!PageUptodate(page)) {
1167 ret = ext4_mb_init_cache(page, NULL);
1172 mb_cmp_bitmaps(e4b, page_address(page) +
1173 (poff * sb->s_blocksize));
1182 if (!PageUptodate(page)) {
1187 /* Pages marked accessed already */
1188 e4b->bd_bitmap_page = page;
1189 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1192 pnum = block / blocks_per_page;
1193 poff = block % blocks_per_page;
1195 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1196 if (page == NULL || !PageUptodate(page)) {
1198 page_cache_release(page);
1199 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1201 BUG_ON(page->mapping != inode->i_mapping);
1202 if (!PageUptodate(page)) {
1203 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1216 if (!PageUptodate(page)) {
1221 /* Pages marked accessed already */
1222 e4b->bd_buddy_page = page;
1223 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1225 BUG_ON(e4b->bd_bitmap_page == NULL);
1226 BUG_ON(e4b->bd_buddy_page == NULL);
1232 page_cache_release(page);
1233 if (e4b->bd_bitmap_page)
1234 page_cache_release(e4b->bd_bitmap_page);
1235 if (e4b->bd_buddy_page)
1236 page_cache_release(e4b->bd_buddy_page);
1237 e4b->bd_buddy = NULL;
1238 e4b->bd_bitmap = NULL;
1242 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1244 if (e4b->bd_bitmap_page)
1245 page_cache_release(e4b->bd_bitmap_page);
1246 if (e4b->bd_buddy_page)
1247 page_cache_release(e4b->bd_buddy_page);
1251 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1256 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1257 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1260 while (order <= e4b->bd_blkbits + 1) {
1262 if (!mb_test_bit(block, bb)) {
1263 /* this block is part of buddy of order 'order' */
1266 bb += 1 << (e4b->bd_blkbits - order);
1272 static void mb_clear_bits(void *bm, int cur, int len)
1278 if ((cur & 31) == 0 && (len - cur) >= 32) {
1279 /* fast path: clear whole word at once */
1280 addr = bm + (cur >> 3);
1285 mb_clear_bit(cur, bm);
1290 /* clear bits in given range
1291 * will return first found zero bit if any, -1 otherwise
1293 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1300 if ((cur & 31) == 0 && (len - cur) >= 32) {
1301 /* fast path: clear whole word at once */
1302 addr = bm + (cur >> 3);
1303 if (*addr != (__u32)(-1) && zero_bit == -1)
1304 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1309 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1317 void ext4_set_bits(void *bm, int cur, int len)
1323 if ((cur & 31) == 0 && (len - cur) >= 32) {
1324 /* fast path: set whole word at once */
1325 addr = bm + (cur >> 3);
1330 mb_set_bit(cur, bm);
1336 * _________________________________________________________________ */
1338 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1340 if (mb_test_bit(*bit + side, bitmap)) {
1341 mb_clear_bit(*bit, bitmap);
1347 mb_set_bit(*bit, bitmap);
1352 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1356 void *buddy = mb_find_buddy(e4b, order, &max);
1361 /* Bits in range [first; last] are known to be set since
1362 * corresponding blocks were allocated. Bits in range
1363 * (first; last) will stay set because they form buddies on
1364 * upper layer. We just deal with borders if they don't
1365 * align with upper layer and then go up.
1366 * Releasing entire group is all about clearing
1367 * single bit of highest order buddy.
1371 * ---------------------------------
1373 * ---------------------------------
1374 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1375 * ---------------------------------
1377 * \_____________________/
1379 * Neither [1] nor [6] is aligned to above layer.
1380 * Left neighbour [0] is free, so mark it busy,
1381 * decrease bb_counters and extend range to
1383 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1384 * mark [6] free, increase bb_counters and shrink range to
1386 * Then shift range to [0; 2], go up and do the same.
1391 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1393 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1398 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1399 mb_clear_bits(buddy, first, last - first + 1);
1400 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1409 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1410 int first, int count)
1412 int left_is_free = 0;
1413 int right_is_free = 0;
1415 int last = first + count - 1;
1416 struct super_block *sb = e4b->bd_sb;
1418 if (WARN_ON(count == 0))
1420 BUG_ON(last >= (sb->s_blocksize << 3));
1421 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1422 /* Don't bother if the block group is corrupt. */
1423 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1426 mb_check_buddy(e4b);
1427 mb_free_blocks_double(inode, e4b, first, count);
1429 e4b->bd_info->bb_free += count;
1430 if (first < e4b->bd_info->bb_first_free)
1431 e4b->bd_info->bb_first_free = first;
1433 /* access memory sequentially: check left neighbour,
1434 * clear range and then check right neighbour
1437 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1438 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1439 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1440 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1442 if (unlikely(block != -1)) {
1443 struct ext4_sb_info *sbi = EXT4_SB(sb);
1444 ext4_fsblk_t blocknr;
1446 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1447 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1448 ext4_grp_locked_error(sb, e4b->bd_group,
1449 inode ? inode->i_ino : 0,
1451 "freeing already freed block "
1452 "(bit %u); block bitmap corrupt.",
1454 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1455 percpu_counter_sub(&sbi->s_freeclusters_counter,
1456 e4b->bd_info->bb_free);
1457 /* Mark the block group as corrupt. */
1458 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1459 &e4b->bd_info->bb_state);
1460 mb_regenerate_buddy(e4b);
1464 /* let's maintain fragments counter */
1465 if (left_is_free && right_is_free)
1466 e4b->bd_info->bb_fragments--;
1467 else if (!left_is_free && !right_is_free)
1468 e4b->bd_info->bb_fragments++;
1470 /* buddy[0] == bd_bitmap is a special case, so handle
1471 * it right away and let mb_buddy_mark_free stay free of
1472 * zero order checks.
1473 * Check if neighbours are to be coaleasced,
1474 * adjust bitmap bb_counters and borders appropriately.
1477 first += !left_is_free;
1478 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1481 last -= !right_is_free;
1482 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1486 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1489 mb_set_largest_free_order(sb, e4b->bd_info);
1490 mb_check_buddy(e4b);
1493 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1494 int needed, struct ext4_free_extent *ex)
1500 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1503 buddy = mb_find_buddy(e4b, 0, &max);
1504 BUG_ON(buddy == NULL);
1505 BUG_ON(block >= max);
1506 if (mb_test_bit(block, buddy)) {
1513 /* find actual order */
1514 order = mb_find_order_for_block(e4b, block);
1515 block = block >> order;
1517 ex->fe_len = 1 << order;
1518 ex->fe_start = block << order;
1519 ex->fe_group = e4b->bd_group;
1521 /* calc difference from given start */
1522 next = next - ex->fe_start;
1524 ex->fe_start += next;
1526 while (needed > ex->fe_len &&
1527 mb_find_buddy(e4b, order, &max)) {
1529 if (block + 1 >= max)
1532 next = (block + 1) * (1 << order);
1533 if (mb_test_bit(next, e4b->bd_bitmap))
1536 order = mb_find_order_for_block(e4b, next);
1538 block = next >> order;
1539 ex->fe_len += 1 << order;
1542 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1546 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1552 int start = ex->fe_start;
1553 int len = ex->fe_len;
1558 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1559 BUG_ON(e4b->bd_group != ex->fe_group);
1560 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1561 mb_check_buddy(e4b);
1562 mb_mark_used_double(e4b, start, len);
1564 e4b->bd_info->bb_free -= len;
1565 if (e4b->bd_info->bb_first_free == start)
1566 e4b->bd_info->bb_first_free += len;
1568 /* let's maintain fragments counter */
1570 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1571 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1572 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1574 e4b->bd_info->bb_fragments++;
1575 else if (!mlen && !max)
1576 e4b->bd_info->bb_fragments--;
1578 /* let's maintain buddy itself */
1580 ord = mb_find_order_for_block(e4b, start);
1582 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1583 /* the whole chunk may be allocated at once! */
1585 buddy = mb_find_buddy(e4b, ord, &max);
1586 BUG_ON((start >> ord) >= max);
1587 mb_set_bit(start >> ord, buddy);
1588 e4b->bd_info->bb_counters[ord]--;
1595 /* store for history */
1597 ret = len | (ord << 16);
1599 /* we have to split large buddy */
1601 buddy = mb_find_buddy(e4b, ord, &max);
1602 mb_set_bit(start >> ord, buddy);
1603 e4b->bd_info->bb_counters[ord]--;
1606 cur = (start >> ord) & ~1U;
1607 buddy = mb_find_buddy(e4b, ord, &max);
1608 mb_clear_bit(cur, buddy);
1609 mb_clear_bit(cur + 1, buddy);
1610 e4b->bd_info->bb_counters[ord]++;
1611 e4b->bd_info->bb_counters[ord]++;
1613 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1615 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1616 mb_check_buddy(e4b);
1622 * Must be called under group lock!
1624 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1625 struct ext4_buddy *e4b)
1627 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1630 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1631 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1633 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1634 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1635 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1637 /* preallocation can change ac_b_ex, thus we store actually
1638 * allocated blocks for history */
1639 ac->ac_f_ex = ac->ac_b_ex;
1641 ac->ac_status = AC_STATUS_FOUND;
1642 ac->ac_tail = ret & 0xffff;
1643 ac->ac_buddy = ret >> 16;
1646 * take the page reference. We want the page to be pinned
1647 * so that we don't get a ext4_mb_init_cache_call for this
1648 * group until we update the bitmap. That would mean we
1649 * double allocate blocks. The reference is dropped
1650 * in ext4_mb_release_context
1652 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1653 get_page(ac->ac_bitmap_page);
1654 ac->ac_buddy_page = e4b->bd_buddy_page;
1655 get_page(ac->ac_buddy_page);
1656 /* store last allocated for subsequent stream allocation */
1657 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1658 spin_lock(&sbi->s_md_lock);
1659 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1660 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1661 spin_unlock(&sbi->s_md_lock);
1666 * regular allocator, for general purposes allocation
1669 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1670 struct ext4_buddy *e4b,
1673 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1674 struct ext4_free_extent *bex = &ac->ac_b_ex;
1675 struct ext4_free_extent *gex = &ac->ac_g_ex;
1676 struct ext4_free_extent ex;
1679 if (ac->ac_status == AC_STATUS_FOUND)
1682 * We don't want to scan for a whole year
1684 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1685 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1686 ac->ac_status = AC_STATUS_BREAK;
1691 * Haven't found good chunk so far, let's continue
1693 if (bex->fe_len < gex->fe_len)
1696 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1697 && bex->fe_group == e4b->bd_group) {
1698 /* recheck chunk's availability - we don't know
1699 * when it was found (within this lock-unlock
1701 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1702 if (max >= gex->fe_len) {
1703 ext4_mb_use_best_found(ac, e4b);
1710 * The routine checks whether found extent is good enough. If it is,
1711 * then the extent gets marked used and flag is set to the context
1712 * to stop scanning. Otherwise, the extent is compared with the
1713 * previous found extent and if new one is better, then it's stored
1714 * in the context. Later, the best found extent will be used, if
1715 * mballoc can't find good enough extent.
1717 * FIXME: real allocation policy is to be designed yet!
1719 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1720 struct ext4_free_extent *ex,
1721 struct ext4_buddy *e4b)
1723 struct ext4_free_extent *bex = &ac->ac_b_ex;
1724 struct ext4_free_extent *gex = &ac->ac_g_ex;
1726 BUG_ON(ex->fe_len <= 0);
1727 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1728 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1729 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1734 * The special case - take what you catch first
1736 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1738 ext4_mb_use_best_found(ac, e4b);
1743 * Let's check whether the chuck is good enough
1745 if (ex->fe_len == gex->fe_len) {
1747 ext4_mb_use_best_found(ac, e4b);
1752 * If this is first found extent, just store it in the context
1754 if (bex->fe_len == 0) {
1760 * If new found extent is better, store it in the context
1762 if (bex->fe_len < gex->fe_len) {
1763 /* if the request isn't satisfied, any found extent
1764 * larger than previous best one is better */
1765 if (ex->fe_len > bex->fe_len)
1767 } else if (ex->fe_len > gex->fe_len) {
1768 /* if the request is satisfied, then we try to find
1769 * an extent that still satisfy the request, but is
1770 * smaller than previous one */
1771 if (ex->fe_len < bex->fe_len)
1775 ext4_mb_check_limits(ac, e4b, 0);
1778 static noinline_for_stack
1779 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1780 struct ext4_buddy *e4b)
1782 struct ext4_free_extent ex = ac->ac_b_ex;
1783 ext4_group_t group = ex.fe_group;
1787 BUG_ON(ex.fe_len <= 0);
1788 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1792 ext4_lock_group(ac->ac_sb, group);
1793 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1797 ext4_mb_use_best_found(ac, e4b);
1800 ext4_unlock_group(ac->ac_sb, group);
1801 ext4_mb_unload_buddy(e4b);
1806 static noinline_for_stack
1807 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1808 struct ext4_buddy *e4b)
1810 ext4_group_t group = ac->ac_g_ex.fe_group;
1813 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1814 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1815 struct ext4_free_extent ex;
1817 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1819 if (grp->bb_free == 0)
1822 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1826 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1827 ext4_mb_unload_buddy(e4b);
1831 ext4_lock_group(ac->ac_sb, group);
1832 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1833 ac->ac_g_ex.fe_len, &ex);
1834 ex.fe_logical = 0xDEADFA11; /* debug value */
1836 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1839 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1841 /* use do_div to get remainder (would be 64-bit modulo) */
1842 if (do_div(start, sbi->s_stripe) == 0) {
1845 ext4_mb_use_best_found(ac, e4b);
1847 } else if (max >= ac->ac_g_ex.fe_len) {
1848 BUG_ON(ex.fe_len <= 0);
1849 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1850 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1853 ext4_mb_use_best_found(ac, e4b);
1854 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1855 /* Sometimes, caller may want to merge even small
1856 * number of blocks to an existing extent */
1857 BUG_ON(ex.fe_len <= 0);
1858 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1859 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1862 ext4_mb_use_best_found(ac, e4b);
1864 ext4_unlock_group(ac->ac_sb, group);
1865 ext4_mb_unload_buddy(e4b);
1871 * The routine scans buddy structures (not bitmap!) from given order
1872 * to max order and tries to find big enough chunk to satisfy the req
1874 static noinline_for_stack
1875 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1876 struct ext4_buddy *e4b)
1878 struct super_block *sb = ac->ac_sb;
1879 struct ext4_group_info *grp = e4b->bd_info;
1885 BUG_ON(ac->ac_2order <= 0);
1886 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1887 if (grp->bb_counters[i] == 0)
1890 buddy = mb_find_buddy(e4b, i, &max);
1891 BUG_ON(buddy == NULL);
1893 k = mb_find_next_zero_bit(buddy, max, 0);
1898 ac->ac_b_ex.fe_len = 1 << i;
1899 ac->ac_b_ex.fe_start = k << i;
1900 ac->ac_b_ex.fe_group = e4b->bd_group;
1902 ext4_mb_use_best_found(ac, e4b);
1904 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1906 if (EXT4_SB(sb)->s_mb_stats)
1907 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1914 * The routine scans the group and measures all found extents.
1915 * In order to optimize scanning, caller must pass number of
1916 * free blocks in the group, so the routine can know upper limit.
1918 static noinline_for_stack
1919 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1920 struct ext4_buddy *e4b)
1922 struct super_block *sb = ac->ac_sb;
1923 void *bitmap = e4b->bd_bitmap;
1924 struct ext4_free_extent ex;
1928 free = e4b->bd_info->bb_free;
1931 i = e4b->bd_info->bb_first_free;
1933 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1934 i = mb_find_next_zero_bit(bitmap,
1935 EXT4_CLUSTERS_PER_GROUP(sb), i);
1936 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1938 * IF we have corrupt bitmap, we won't find any
1939 * free blocks even though group info says we
1940 * we have free blocks
1942 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1943 "%d free clusters as per "
1944 "group info. But bitmap says 0",
1949 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1950 BUG_ON(ex.fe_len <= 0);
1951 if (free < ex.fe_len) {
1952 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1953 "%d free clusters as per "
1954 "group info. But got %d blocks",
1957 * The number of free blocks differs. This mostly
1958 * indicate that the bitmap is corrupt. So exit
1959 * without claiming the space.
1963 ex.fe_logical = 0xDEADC0DE; /* debug value */
1964 ext4_mb_measure_extent(ac, &ex, e4b);
1970 ext4_mb_check_limits(ac, e4b, 1);
1974 * This is a special case for storages like raid5
1975 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1977 static noinline_for_stack
1978 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1979 struct ext4_buddy *e4b)
1981 struct super_block *sb = ac->ac_sb;
1982 struct ext4_sb_info *sbi = EXT4_SB(sb);
1983 void *bitmap = e4b->bd_bitmap;
1984 struct ext4_free_extent ex;
1985 ext4_fsblk_t first_group_block;
1990 BUG_ON(sbi->s_stripe == 0);
1992 /* find first stripe-aligned block in group */
1993 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1995 a = first_group_block + sbi->s_stripe - 1;
1996 do_div(a, sbi->s_stripe);
1997 i = (a * sbi->s_stripe) - first_group_block;
1999 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2000 if (!mb_test_bit(i, bitmap)) {
2001 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2002 if (max >= sbi->s_stripe) {
2004 ex.fe_logical = 0xDEADF00D; /* debug value */
2006 ext4_mb_use_best_found(ac, e4b);
2015 * This is now called BEFORE we load the buddy bitmap.
2016 * Returns either 1 or 0 indicating that the group is either suitable
2017 * for the allocation or not. In addition it can also return negative
2018 * error code when something goes wrong.
2020 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2021 ext4_group_t group, int cr)
2023 unsigned free, fragments;
2024 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2025 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2027 BUG_ON(cr < 0 || cr >= 4);
2029 free = grp->bb_free;
2032 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2035 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2038 /* We only do this if the grp has never been initialized */
2039 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2040 int ret = ext4_mb_init_group(ac->ac_sb, group);
2045 fragments = grp->bb_fragments;
2051 BUG_ON(ac->ac_2order == 0);
2053 /* Avoid using the first bg of a flexgroup for data files */
2054 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2055 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2056 ((group % flex_size) == 0))
2059 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2060 (free / fragments) >= ac->ac_g_ex.fe_len)
2063 if (grp->bb_largest_free_order < ac->ac_2order)
2068 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2072 if (free >= ac->ac_g_ex.fe_len)
2084 static noinline_for_stack int
2085 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2087 ext4_group_t ngroups, group, i;
2089 int err = 0, first_err = 0;
2090 struct ext4_sb_info *sbi;
2091 struct super_block *sb;
2092 struct ext4_buddy e4b;
2096 ngroups = ext4_get_groups_count(sb);
2097 /* non-extent files are limited to low blocks/groups */
2098 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2099 ngroups = sbi->s_blockfile_groups;
2101 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2103 /* first, try the goal */
2104 err = ext4_mb_find_by_goal(ac, &e4b);
2105 if (err || ac->ac_status == AC_STATUS_FOUND)
2108 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2112 * ac->ac2_order is set only if the fe_len is a power of 2
2113 * if ac2_order is set we also set criteria to 0 so that we
2114 * try exact allocation using buddy.
2116 i = fls(ac->ac_g_ex.fe_len);
2119 * We search using buddy data only if the order of the request
2120 * is greater than equal to the sbi_s_mb_order2_reqs
2121 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2123 if (i >= sbi->s_mb_order2_reqs) {
2125 * This should tell if fe_len is exactly power of 2
2127 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2128 ac->ac_2order = i - 1;
2131 /* if stream allocation is enabled, use global goal */
2132 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2133 /* TBD: may be hot point */
2134 spin_lock(&sbi->s_md_lock);
2135 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2136 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2137 spin_unlock(&sbi->s_md_lock);
2140 /* Let's just scan groups to find more-less suitable blocks */
2141 cr = ac->ac_2order ? 0 : 1;
2143 * cr == 0 try to get exact allocation,
2144 * cr == 3 try to get anything
2147 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2148 ac->ac_criteria = cr;
2150 * searching for the right group start
2151 * from the goal value specified
2153 group = ac->ac_g_ex.fe_group;
2155 for (i = 0; i < ngroups; group++, i++) {
2159 * Artificially restricted ngroups for non-extent
2160 * files makes group > ngroups possible on first loop.
2162 if (group >= ngroups)
2165 /* This now checks without needing the buddy page */
2166 ret = ext4_mb_good_group(ac, group, cr);
2173 err = ext4_mb_load_buddy(sb, group, &e4b);
2177 ext4_lock_group(sb, group);
2180 * We need to check again after locking the
2183 ret = ext4_mb_good_group(ac, group, cr);
2185 ext4_unlock_group(sb, group);
2186 ext4_mb_unload_buddy(&e4b);
2192 ac->ac_groups_scanned++;
2193 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2194 ext4_mb_simple_scan_group(ac, &e4b);
2195 else if (cr == 1 && sbi->s_stripe &&
2196 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2197 ext4_mb_scan_aligned(ac, &e4b);
2199 ext4_mb_complex_scan_group(ac, &e4b);
2201 ext4_unlock_group(sb, group);
2202 ext4_mb_unload_buddy(&e4b);
2204 if (ac->ac_status != AC_STATUS_CONTINUE)
2209 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2210 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2212 * We've been searching too long. Let's try to allocate
2213 * the best chunk we've found so far
2216 ext4_mb_try_best_found(ac, &e4b);
2217 if (ac->ac_status != AC_STATUS_FOUND) {
2219 * Someone more lucky has already allocated it.
2220 * The only thing we can do is just take first
2222 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2224 ac->ac_b_ex.fe_group = 0;
2225 ac->ac_b_ex.fe_start = 0;
2226 ac->ac_b_ex.fe_len = 0;
2227 ac->ac_status = AC_STATUS_CONTINUE;
2228 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2230 atomic_inc(&sbi->s_mb_lost_chunks);
2235 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2240 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2242 struct super_block *sb = seq->private;
2245 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2248 return (void *) ((unsigned long) group);
2251 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2253 struct super_block *sb = seq->private;
2257 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2260 return (void *) ((unsigned long) group);
2263 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2265 struct super_block *sb = seq->private;
2266 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2268 int err, buddy_loaded = 0;
2269 struct ext4_buddy e4b;
2270 struct ext4_group_info *grinfo;
2272 struct ext4_group_info info;
2273 ext4_grpblk_t counters[16];
2278 seq_puts(seq, "#group: free frags first ["
2279 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2280 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]");
2282 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2283 sizeof(struct ext4_group_info);
2284 grinfo = ext4_get_group_info(sb, group);
2285 /* Load the group info in memory only if not already loaded. */
2286 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2287 err = ext4_mb_load_buddy(sb, group, &e4b);
2289 seq_printf(seq, "#%-5u: I/O error\n", group);
2295 memcpy(&sg, ext4_get_group_info(sb, group), i);
2298 ext4_mb_unload_buddy(&e4b);
2300 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2301 sg.info.bb_fragments, sg.info.bb_first_free);
2302 for (i = 0; i <= 13; i++)
2303 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2304 sg.info.bb_counters[i] : 0);
2305 seq_printf(seq, " ]\n");
2310 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2314 static const struct seq_operations ext4_mb_seq_groups_ops = {
2315 .start = ext4_mb_seq_groups_start,
2316 .next = ext4_mb_seq_groups_next,
2317 .stop = ext4_mb_seq_groups_stop,
2318 .show = ext4_mb_seq_groups_show,
2321 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2323 struct super_block *sb = PDE_DATA(inode);
2326 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2328 struct seq_file *m = file->private_data;
2335 static const struct file_operations ext4_mb_seq_groups_fops = {
2336 .owner = THIS_MODULE,
2337 .open = ext4_mb_seq_groups_open,
2339 .llseek = seq_lseek,
2340 .release = seq_release,
2343 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2345 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2346 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2353 * Allocate the top-level s_group_info array for the specified number
2356 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2358 struct ext4_sb_info *sbi = EXT4_SB(sb);
2360 struct ext4_group_info ***new_groupinfo;
2362 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2363 EXT4_DESC_PER_BLOCK_BITS(sb);
2364 if (size <= sbi->s_group_info_size)
2367 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2368 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2369 if (!new_groupinfo) {
2370 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2373 if (sbi->s_group_info) {
2374 memcpy(new_groupinfo, sbi->s_group_info,
2375 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2376 kvfree(sbi->s_group_info);
2378 sbi->s_group_info = new_groupinfo;
2379 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2380 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2381 sbi->s_group_info_size);
2385 /* Create and initialize ext4_group_info data for the given group. */
2386 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2387 struct ext4_group_desc *desc)
2391 struct ext4_sb_info *sbi = EXT4_SB(sb);
2392 struct ext4_group_info **meta_group_info;
2393 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2396 * First check if this group is the first of a reserved block.
2397 * If it's true, we have to allocate a new table of pointers
2398 * to ext4_group_info structures
2400 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2401 metalen = sizeof(*meta_group_info) <<
2402 EXT4_DESC_PER_BLOCK_BITS(sb);
2403 meta_group_info = kmalloc(metalen, GFP_NOFS);
2404 if (meta_group_info == NULL) {
2405 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2406 "for a buddy group");
2407 goto exit_meta_group_info;
2409 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2414 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2415 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2417 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2418 if (meta_group_info[i] == NULL) {
2419 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2420 goto exit_group_info;
2422 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2423 &(meta_group_info[i]->bb_state));
2426 * initialize bb_free to be able to skip
2427 * empty groups without initialization
2429 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2430 meta_group_info[i]->bb_free =
2431 ext4_free_clusters_after_init(sb, group, desc);
2433 meta_group_info[i]->bb_free =
2434 ext4_free_group_clusters(sb, desc);
2437 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2438 init_rwsem(&meta_group_info[i]->alloc_sem);
2439 meta_group_info[i]->bb_free_root = RB_ROOT;
2440 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2444 struct buffer_head *bh;
2445 meta_group_info[i]->bb_bitmap =
2446 kmalloc(sb->s_blocksize, GFP_NOFS);
2447 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2448 bh = ext4_read_block_bitmap(sb, group);
2450 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2459 /* If a meta_group_info table has been allocated, release it now */
2460 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2461 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2462 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2464 exit_meta_group_info:
2466 } /* ext4_mb_add_groupinfo */
2468 static int ext4_mb_init_backend(struct super_block *sb)
2470 ext4_group_t ngroups = ext4_get_groups_count(sb);
2472 struct ext4_sb_info *sbi = EXT4_SB(sb);
2474 struct ext4_group_desc *desc;
2475 struct kmem_cache *cachep;
2477 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2481 sbi->s_buddy_cache = new_inode(sb);
2482 if (sbi->s_buddy_cache == NULL) {
2483 ext4_msg(sb, KERN_ERR, "can't get new inode");
2486 /* To avoid potentially colliding with an valid on-disk inode number,
2487 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2488 * not in the inode hash, so it should never be found by iget(), but
2489 * this will avoid confusion if it ever shows up during debugging. */
2490 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2491 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2492 for (i = 0; i < ngroups; i++) {
2493 desc = ext4_get_group_desc(sb, i, NULL);
2495 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2498 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2505 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2507 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2508 i = sbi->s_group_info_size;
2510 kfree(sbi->s_group_info[i]);
2511 iput(sbi->s_buddy_cache);
2513 kvfree(sbi->s_group_info);
2517 static void ext4_groupinfo_destroy_slabs(void)
2521 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2522 if (ext4_groupinfo_caches[i])
2523 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2524 ext4_groupinfo_caches[i] = NULL;
2528 static int ext4_groupinfo_create_slab(size_t size)
2530 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2532 int blocksize_bits = order_base_2(size);
2533 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2534 struct kmem_cache *cachep;
2536 if (cache_index >= NR_GRPINFO_CACHES)
2539 if (unlikely(cache_index < 0))
2542 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2543 if (ext4_groupinfo_caches[cache_index]) {
2544 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2545 return 0; /* Already created */
2548 slab_size = offsetof(struct ext4_group_info,
2549 bb_counters[blocksize_bits + 2]);
2551 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2552 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2555 ext4_groupinfo_caches[cache_index] = cachep;
2557 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2560 "EXT4-fs: no memory for groupinfo slab cache\n");
2567 int ext4_mb_init(struct super_block *sb)
2569 struct ext4_sb_info *sbi = EXT4_SB(sb);
2575 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2577 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2578 if (sbi->s_mb_offsets == NULL) {
2583 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2584 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2585 if (sbi->s_mb_maxs == NULL) {
2590 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2594 /* order 0 is regular bitmap */
2595 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2596 sbi->s_mb_offsets[0] = 0;
2600 max = sb->s_blocksize << 2;
2602 sbi->s_mb_offsets[i] = offset;
2603 sbi->s_mb_maxs[i] = max;
2604 offset += 1 << (sb->s_blocksize_bits - i);
2607 } while (i <= sb->s_blocksize_bits + 1);
2609 spin_lock_init(&sbi->s_md_lock);
2610 spin_lock_init(&sbi->s_bal_lock);
2612 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2613 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2614 sbi->s_mb_stats = MB_DEFAULT_STATS;
2615 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2616 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2618 * The default group preallocation is 512, which for 4k block
2619 * sizes translates to 2 megabytes. However for bigalloc file
2620 * systems, this is probably too big (i.e, if the cluster size
2621 * is 1 megabyte, then group preallocation size becomes half a
2622 * gigabyte!). As a default, we will keep a two megabyte
2623 * group pralloc size for cluster sizes up to 64k, and after
2624 * that, we will force a minimum group preallocation size of
2625 * 32 clusters. This translates to 8 megs when the cluster
2626 * size is 256k, and 32 megs when the cluster size is 1 meg,
2627 * which seems reasonable as a default.
2629 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2630 sbi->s_cluster_bits, 32);
2632 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2633 * to the lowest multiple of s_stripe which is bigger than
2634 * the s_mb_group_prealloc as determined above. We want
2635 * the preallocation size to be an exact multiple of the
2636 * RAID stripe size so that preallocations don't fragment
2639 if (sbi->s_stripe > 1) {
2640 sbi->s_mb_group_prealloc = roundup(
2641 sbi->s_mb_group_prealloc, sbi->s_stripe);
2644 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2645 if (sbi->s_locality_groups == NULL) {
2649 for_each_possible_cpu(i) {
2650 struct ext4_locality_group *lg;
2651 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2652 mutex_init(&lg->lg_mutex);
2653 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2654 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2655 spin_lock_init(&lg->lg_prealloc_lock);
2658 /* init file for buddy data */
2659 ret = ext4_mb_init_backend(sb);
2661 goto out_free_locality_groups;
2664 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2665 &ext4_mb_seq_groups_fops, sb);
2669 out_free_locality_groups:
2670 free_percpu(sbi->s_locality_groups);
2671 sbi->s_locality_groups = NULL;
2673 kfree(sbi->s_mb_offsets);
2674 sbi->s_mb_offsets = NULL;
2675 kfree(sbi->s_mb_maxs);
2676 sbi->s_mb_maxs = NULL;
2680 /* need to called with the ext4 group lock held */
2681 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2683 struct ext4_prealloc_space *pa;
2684 struct list_head *cur, *tmp;
2687 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2688 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2689 list_del(&pa->pa_group_list);
2691 kmem_cache_free(ext4_pspace_cachep, pa);
2694 mb_debug(1, "mballoc: %u PAs left\n", count);
2698 int ext4_mb_release(struct super_block *sb)
2700 ext4_group_t ngroups = ext4_get_groups_count(sb);
2702 int num_meta_group_infos;
2703 struct ext4_group_info *grinfo;
2704 struct ext4_sb_info *sbi = EXT4_SB(sb);
2705 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2708 remove_proc_entry("mb_groups", sbi->s_proc);
2710 if (sbi->s_group_info) {
2711 for (i = 0; i < ngroups; i++) {
2712 grinfo = ext4_get_group_info(sb, i);
2714 kfree(grinfo->bb_bitmap);
2716 ext4_lock_group(sb, i);
2717 ext4_mb_cleanup_pa(grinfo);
2718 ext4_unlock_group(sb, i);
2719 kmem_cache_free(cachep, grinfo);
2721 num_meta_group_infos = (ngroups +
2722 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2723 EXT4_DESC_PER_BLOCK_BITS(sb);
2724 for (i = 0; i < num_meta_group_infos; i++)
2725 kfree(sbi->s_group_info[i]);
2726 kvfree(sbi->s_group_info);
2728 kfree(sbi->s_mb_offsets);
2729 kfree(sbi->s_mb_maxs);
2730 iput(sbi->s_buddy_cache);
2731 if (sbi->s_mb_stats) {
2732 ext4_msg(sb, KERN_INFO,
2733 "mballoc: %u blocks %u reqs (%u success)",
2734 atomic_read(&sbi->s_bal_allocated),
2735 atomic_read(&sbi->s_bal_reqs),
2736 atomic_read(&sbi->s_bal_success));
2737 ext4_msg(sb, KERN_INFO,
2738 "mballoc: %u extents scanned, %u goal hits, "
2739 "%u 2^N hits, %u breaks, %u lost",
2740 atomic_read(&sbi->s_bal_ex_scanned),
2741 atomic_read(&sbi->s_bal_goals),
2742 atomic_read(&sbi->s_bal_2orders),
2743 atomic_read(&sbi->s_bal_breaks),
2744 atomic_read(&sbi->s_mb_lost_chunks));
2745 ext4_msg(sb, KERN_INFO,
2746 "mballoc: %lu generated and it took %Lu",
2747 sbi->s_mb_buddies_generated,
2748 sbi->s_mb_generation_time);
2749 ext4_msg(sb, KERN_INFO,
2750 "mballoc: %u preallocated, %u discarded",
2751 atomic_read(&sbi->s_mb_preallocated),
2752 atomic_read(&sbi->s_mb_discarded));
2755 free_percpu(sbi->s_locality_groups);
2760 static inline int ext4_issue_discard(struct super_block *sb,
2761 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2763 ext4_fsblk_t discard_block;
2765 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2766 ext4_group_first_block_no(sb, block_group));
2767 count = EXT4_C2B(EXT4_SB(sb), count);
2768 trace_ext4_discard_blocks(sb,
2769 (unsigned long long) discard_block, count);
2770 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2774 * This function is called by the jbd2 layer once the commit has finished,
2775 * so we know we can free the blocks that were released with that commit.
2777 static void ext4_free_data_callback(struct super_block *sb,
2778 struct ext4_journal_cb_entry *jce,
2781 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2782 struct ext4_buddy e4b;
2783 struct ext4_group_info *db;
2784 int err, count = 0, count2 = 0;
2786 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2787 entry->efd_count, entry->efd_group, entry);
2789 if (test_opt(sb, DISCARD)) {
2790 err = ext4_issue_discard(sb, entry->efd_group,
2791 entry->efd_start_cluster,
2793 if (err && err != -EOPNOTSUPP)
2794 ext4_msg(sb, KERN_WARNING, "discard request in"
2795 " group:%d block:%d count:%d failed"
2796 " with %d", entry->efd_group,
2797 entry->efd_start_cluster,
2798 entry->efd_count, err);
2801 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2802 /* we expect to find existing buddy because it's pinned */
2807 /* there are blocks to put in buddy to make them really free */
2808 count += entry->efd_count;
2810 ext4_lock_group(sb, entry->efd_group);
2811 /* Take it out of per group rb tree */
2812 rb_erase(&entry->efd_node, &(db->bb_free_root));
2813 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2816 * Clear the trimmed flag for the group so that the next
2817 * ext4_trim_fs can trim it.
2818 * If the volume is mounted with -o discard, online discard
2819 * is supported and the free blocks will be trimmed online.
2821 if (!test_opt(sb, DISCARD))
2822 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2824 if (!db->bb_free_root.rb_node) {
2825 /* No more items in the per group rb tree
2826 * balance refcounts from ext4_mb_free_metadata()
2828 page_cache_release(e4b.bd_buddy_page);
2829 page_cache_release(e4b.bd_bitmap_page);
2831 ext4_unlock_group(sb, entry->efd_group);
2832 kmem_cache_free(ext4_free_data_cachep, entry);
2833 ext4_mb_unload_buddy(&e4b);
2835 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2838 int __init ext4_init_mballoc(void)
2840 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2841 SLAB_RECLAIM_ACCOUNT);
2842 if (ext4_pspace_cachep == NULL)
2845 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2846 SLAB_RECLAIM_ACCOUNT);
2847 if (ext4_ac_cachep == NULL) {
2848 kmem_cache_destroy(ext4_pspace_cachep);
2852 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2853 SLAB_RECLAIM_ACCOUNT);
2854 if (ext4_free_data_cachep == NULL) {
2855 kmem_cache_destroy(ext4_pspace_cachep);
2856 kmem_cache_destroy(ext4_ac_cachep);
2862 void ext4_exit_mballoc(void)
2865 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2866 * before destroying the slab cache.
2869 kmem_cache_destroy(ext4_pspace_cachep);
2870 kmem_cache_destroy(ext4_ac_cachep);
2871 kmem_cache_destroy(ext4_free_data_cachep);
2872 ext4_groupinfo_destroy_slabs();
2877 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2878 * Returns 0 if success or error code
2880 static noinline_for_stack int
2881 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2882 handle_t *handle, unsigned int reserv_clstrs)
2884 struct buffer_head *bitmap_bh = NULL;
2885 struct ext4_group_desc *gdp;
2886 struct buffer_head *gdp_bh;
2887 struct ext4_sb_info *sbi;
2888 struct super_block *sb;
2892 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2893 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2899 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2903 BUFFER_TRACE(bitmap_bh, "getting write access");
2904 err = ext4_journal_get_write_access(handle, bitmap_bh);
2909 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2913 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2914 ext4_free_group_clusters(sb, gdp));
2916 BUFFER_TRACE(gdp_bh, "get_write_access");
2917 err = ext4_journal_get_write_access(handle, gdp_bh);
2921 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2923 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2924 if (!ext4_data_block_valid(sbi, block, len)) {
2925 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2926 "fs metadata", block, block+len);
2927 /* File system mounted not to panic on error
2928 * Fix the bitmap and repeat the block allocation
2929 * We leak some of the blocks here.
2931 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2932 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2933 ac->ac_b_ex.fe_len);
2934 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2935 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2941 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2942 #ifdef AGGRESSIVE_CHECK
2945 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2946 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2947 bitmap_bh->b_data));
2951 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2952 ac->ac_b_ex.fe_len);
2953 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2954 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2955 ext4_free_group_clusters_set(sb, gdp,
2956 ext4_free_clusters_after_init(sb,
2957 ac->ac_b_ex.fe_group, gdp));
2959 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2960 ext4_free_group_clusters_set(sb, gdp, len);
2961 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2962 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2964 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2965 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2967 * Now reduce the dirty block count also. Should not go negative
2969 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2970 /* release all the reserved blocks if non delalloc */
2971 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2974 if (sbi->s_log_groups_per_flex) {
2975 ext4_group_t flex_group = ext4_flex_group(sbi,
2976 ac->ac_b_ex.fe_group);
2977 atomic64_sub(ac->ac_b_ex.fe_len,
2978 &sbi->s_flex_groups[flex_group].free_clusters);
2981 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2984 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2992 * here we normalize request for locality group
2993 * Group request are normalized to s_mb_group_prealloc, which goes to
2994 * s_strip if we set the same via mount option.
2995 * s_mb_group_prealloc can be configured via
2996 * /sys/fs/ext4/<partition>/mb_group_prealloc
2998 * XXX: should we try to preallocate more than the group has now?
3000 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3002 struct super_block *sb = ac->ac_sb;
3003 struct ext4_locality_group *lg = ac->ac_lg;
3006 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3007 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3008 current->pid, ac->ac_g_ex.fe_len);
3012 * Normalization means making request better in terms of
3013 * size and alignment
3015 static noinline_for_stack void
3016 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3017 struct ext4_allocation_request *ar)
3019 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3022 loff_t size, start_off;
3023 loff_t orig_size __maybe_unused;
3025 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3026 struct ext4_prealloc_space *pa;
3028 /* do normalize only data requests, metadata requests
3029 do not need preallocation */
3030 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3033 /* sometime caller may want exact blocks */
3034 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3037 /* caller may indicate that preallocation isn't
3038 * required (it's a tail, for example) */
3039 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3042 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3043 ext4_mb_normalize_group_request(ac);
3047 bsbits = ac->ac_sb->s_blocksize_bits;
3049 /* first, let's learn actual file size
3050 * given current request is allocated */
3051 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3052 size = size << bsbits;
3053 if (size < i_size_read(ac->ac_inode))
3054 size = i_size_read(ac->ac_inode);
3057 /* max size of free chunks */
3060 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3061 (req <= (size) || max <= (chunk_size))
3063 /* first, try to predict filesize */
3064 /* XXX: should this table be tunable? */
3066 if (size <= 16 * 1024) {
3068 } else if (size <= 32 * 1024) {
3070 } else if (size <= 64 * 1024) {
3072 } else if (size <= 128 * 1024) {
3074 } else if (size <= 256 * 1024) {
3076 } else if (size <= 512 * 1024) {
3078 } else if (size <= 1024 * 1024) {
3080 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3081 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3082 (21 - bsbits)) << 21;
3083 size = 2 * 1024 * 1024;
3084 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3085 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3086 (22 - bsbits)) << 22;
3087 size = 4 * 1024 * 1024;
3088 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3089 (8<<20)>>bsbits, max, 8 * 1024)) {
3090 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3091 (23 - bsbits)) << 23;
3092 size = 8 * 1024 * 1024;
3094 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3095 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3096 ac->ac_o_ex.fe_len) << bsbits;
3098 size = size >> bsbits;
3099 start = start_off >> bsbits;
3101 /* don't cover already allocated blocks in selected range */
3102 if (ar->pleft && start <= ar->lleft) {
3103 size -= ar->lleft + 1 - start;
3104 start = ar->lleft + 1;
3106 if (ar->pright && start + size - 1 >= ar->lright)
3107 size -= start + size - ar->lright;
3111 /* check we don't cross already preallocated blocks */
3113 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3118 spin_lock(&pa->pa_lock);
3119 if (pa->pa_deleted) {
3120 spin_unlock(&pa->pa_lock);
3124 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3127 /* PA must not overlap original request */
3128 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3129 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3131 /* skip PAs this normalized request doesn't overlap with */
3132 if (pa->pa_lstart >= end || pa_end <= start) {
3133 spin_unlock(&pa->pa_lock);
3136 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3138 /* adjust start or end to be adjacent to this pa */
3139 if (pa_end <= ac->ac_o_ex.fe_logical) {
3140 BUG_ON(pa_end < start);
3142 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3143 BUG_ON(pa->pa_lstart > end);
3144 end = pa->pa_lstart;
3146 spin_unlock(&pa->pa_lock);
3151 /* XXX: extra loop to check we really don't overlap preallocations */
3153 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3156 spin_lock(&pa->pa_lock);
3157 if (pa->pa_deleted == 0) {
3158 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3160 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3162 spin_unlock(&pa->pa_lock);
3166 if (start + size <= ac->ac_o_ex.fe_logical &&
3167 start > ac->ac_o_ex.fe_logical) {
3168 ext4_msg(ac->ac_sb, KERN_ERR,
3169 "start %lu, size %lu, fe_logical %lu",
3170 (unsigned long) start, (unsigned long) size,
3171 (unsigned long) ac->ac_o_ex.fe_logical);
3174 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3176 /* now prepare goal request */
3178 /* XXX: is it better to align blocks WRT to logical
3179 * placement or satisfy big request as is */
3180 ac->ac_g_ex.fe_logical = start;
3181 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3183 /* define goal start in order to merge */
3184 if (ar->pright && (ar->lright == (start + size))) {
3185 /* merge to the right */
3186 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3187 &ac->ac_f_ex.fe_group,
3188 &ac->ac_f_ex.fe_start);
3189 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3191 if (ar->pleft && (ar->lleft + 1 == start)) {
3192 /* merge to the left */
3193 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3194 &ac->ac_f_ex.fe_group,
3195 &ac->ac_f_ex.fe_start);
3196 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3199 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3200 (unsigned) orig_size, (unsigned) start);
3203 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3205 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3207 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3208 atomic_inc(&sbi->s_bal_reqs);
3209 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3210 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3211 atomic_inc(&sbi->s_bal_success);
3212 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3213 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3214 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3215 atomic_inc(&sbi->s_bal_goals);
3216 if (ac->ac_found > sbi->s_mb_max_to_scan)
3217 atomic_inc(&sbi->s_bal_breaks);
3220 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3221 trace_ext4_mballoc_alloc(ac);
3223 trace_ext4_mballoc_prealloc(ac);
3227 * Called on failure; free up any blocks from the inode PA for this
3228 * context. We don't need this for MB_GROUP_PA because we only change
3229 * pa_free in ext4_mb_release_context(), but on failure, we've already
3230 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3232 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3234 struct ext4_prealloc_space *pa = ac->ac_pa;
3235 struct ext4_buddy e4b;
3239 if (ac->ac_f_ex.fe_len == 0)
3241 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3244 * This should never happen since we pin the
3245 * pages in the ext4_allocation_context so
3246 * ext4_mb_load_buddy() should never fail.
3248 WARN(1, "mb_load_buddy failed (%d)", err);
3251 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3252 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3253 ac->ac_f_ex.fe_len);
3254 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3255 ext4_mb_unload_buddy(&e4b);
3258 if (pa->pa_type == MB_INODE_PA)
3259 pa->pa_free += ac->ac_b_ex.fe_len;
3263 * use blocks preallocated to inode
3265 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3266 struct ext4_prealloc_space *pa)
3268 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3273 /* found preallocated blocks, use them */
3274 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3275 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3276 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3277 len = EXT4_NUM_B2C(sbi, end - start);
3278 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3279 &ac->ac_b_ex.fe_start);
3280 ac->ac_b_ex.fe_len = len;
3281 ac->ac_status = AC_STATUS_FOUND;
3284 BUG_ON(start < pa->pa_pstart);
3285 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3286 BUG_ON(pa->pa_free < len);
3289 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3293 * use blocks preallocated to locality group
3295 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3296 struct ext4_prealloc_space *pa)
3298 unsigned int len = ac->ac_o_ex.fe_len;
3300 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3301 &ac->ac_b_ex.fe_group,
3302 &ac->ac_b_ex.fe_start);
3303 ac->ac_b_ex.fe_len = len;
3304 ac->ac_status = AC_STATUS_FOUND;
3307 /* we don't correct pa_pstart or pa_plen here to avoid
3308 * possible race when the group is being loaded concurrently
3309 * instead we correct pa later, after blocks are marked
3310 * in on-disk bitmap -- see ext4_mb_release_context()
3311 * Other CPUs are prevented from allocating from this pa by lg_mutex
3313 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3317 * Return the prealloc space that have minimal distance
3318 * from the goal block. @cpa is the prealloc
3319 * space that is having currently known minimal distance
3320 * from the goal block.
3322 static struct ext4_prealloc_space *
3323 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3324 struct ext4_prealloc_space *pa,
3325 struct ext4_prealloc_space *cpa)
3327 ext4_fsblk_t cur_distance, new_distance;
3330 atomic_inc(&pa->pa_count);
3333 cur_distance = abs(goal_block - cpa->pa_pstart);
3334 new_distance = abs(goal_block - pa->pa_pstart);
3336 if (cur_distance <= new_distance)
3339 /* drop the previous reference */
3340 atomic_dec(&cpa->pa_count);
3341 atomic_inc(&pa->pa_count);
3346 * search goal blocks in preallocated space
3348 static noinline_for_stack int
3349 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3351 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3353 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3354 struct ext4_locality_group *lg;
3355 struct ext4_prealloc_space *pa, *cpa = NULL;
3356 ext4_fsblk_t goal_block;
3358 /* only data can be preallocated */
3359 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3362 /* first, try per-file preallocation */
3364 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3366 /* all fields in this condition don't change,
3367 * so we can skip locking for them */
3368 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3369 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3370 EXT4_C2B(sbi, pa->pa_len)))
3373 /* non-extent files can't have physical blocks past 2^32 */
3374 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3375 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3376 EXT4_MAX_BLOCK_FILE_PHYS))
3379 /* found preallocated blocks, use them */
3380 spin_lock(&pa->pa_lock);
3381 if (pa->pa_deleted == 0 && pa->pa_free) {
3382 atomic_inc(&pa->pa_count);
3383 ext4_mb_use_inode_pa(ac, pa);
3384 spin_unlock(&pa->pa_lock);
3385 ac->ac_criteria = 10;
3389 spin_unlock(&pa->pa_lock);
3393 /* can we use group allocation? */
3394 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3397 /* inode may have no locality group for some reason */
3401 order = fls(ac->ac_o_ex.fe_len) - 1;
3402 if (order > PREALLOC_TB_SIZE - 1)
3403 /* The max size of hash table is PREALLOC_TB_SIZE */
3404 order = PREALLOC_TB_SIZE - 1;
3406 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3408 * search for the prealloc space that is having
3409 * minimal distance from the goal block.
3411 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3413 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3415 spin_lock(&pa->pa_lock);
3416 if (pa->pa_deleted == 0 &&
3417 pa->pa_free >= ac->ac_o_ex.fe_len) {
3419 cpa = ext4_mb_check_group_pa(goal_block,
3422 spin_unlock(&pa->pa_lock);
3427 ext4_mb_use_group_pa(ac, cpa);
3428 ac->ac_criteria = 20;
3435 * the function goes through all block freed in the group
3436 * but not yet committed and marks them used in in-core bitmap.
3437 * buddy must be generated from this bitmap
3438 * Need to be called with the ext4 group lock held
3440 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3444 struct ext4_group_info *grp;
3445 struct ext4_free_data *entry;
3447 grp = ext4_get_group_info(sb, group);
3448 n = rb_first(&(grp->bb_free_root));
3451 entry = rb_entry(n, struct ext4_free_data, efd_node);
3452 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3459 * the function goes through all preallocation in this group and marks them
3460 * used in in-core bitmap. buddy must be generated from this bitmap
3461 * Need to be called with ext4 group lock held
3463 static noinline_for_stack
3464 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3467 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3468 struct ext4_prealloc_space *pa;
3469 struct list_head *cur;
3470 ext4_group_t groupnr;
3471 ext4_grpblk_t start;
3472 int preallocated = 0;
3475 /* all form of preallocation discards first load group,
3476 * so the only competing code is preallocation use.
3477 * we don't need any locking here
3478 * notice we do NOT ignore preallocations with pa_deleted
3479 * otherwise we could leave used blocks available for
3480 * allocation in buddy when concurrent ext4_mb_put_pa()
3481 * is dropping preallocation
3483 list_for_each(cur, &grp->bb_prealloc_list) {
3484 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3485 spin_lock(&pa->pa_lock);
3486 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3489 spin_unlock(&pa->pa_lock);
3490 if (unlikely(len == 0))
3492 BUG_ON(groupnr != group);
3493 ext4_set_bits(bitmap, start, len);
3494 preallocated += len;
3496 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3499 static void ext4_mb_pa_callback(struct rcu_head *head)
3501 struct ext4_prealloc_space *pa;
3502 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3504 BUG_ON(atomic_read(&pa->pa_count));
3505 BUG_ON(pa->pa_deleted == 0);
3506 kmem_cache_free(ext4_pspace_cachep, pa);
3510 * drops a reference to preallocated space descriptor
3511 * if this was the last reference and the space is consumed
3513 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3514 struct super_block *sb, struct ext4_prealloc_space *pa)
3517 ext4_fsblk_t grp_blk;
3519 /* in this short window concurrent discard can set pa_deleted */
3520 spin_lock(&pa->pa_lock);
3521 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3522 spin_unlock(&pa->pa_lock);
3526 if (pa->pa_deleted == 1) {
3527 spin_unlock(&pa->pa_lock);
3532 spin_unlock(&pa->pa_lock);
3534 grp_blk = pa->pa_pstart;
3536 * If doing group-based preallocation, pa_pstart may be in the
3537 * next group when pa is used up
3539 if (pa->pa_type == MB_GROUP_PA)
3542 grp = ext4_get_group_number(sb, grp_blk);
3547 * P1 (buddy init) P2 (regular allocation)
3548 * find block B in PA
3549 * copy on-disk bitmap to buddy
3550 * mark B in on-disk bitmap
3551 * drop PA from group
3552 * mark all PAs in buddy
3554 * thus, P1 initializes buddy with B available. to prevent this
3555 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3558 ext4_lock_group(sb, grp);
3559 list_del(&pa->pa_group_list);
3560 ext4_unlock_group(sb, grp);
3562 spin_lock(pa->pa_obj_lock);
3563 list_del_rcu(&pa->pa_inode_list);
3564 spin_unlock(pa->pa_obj_lock);
3566 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3570 * creates new preallocated space for given inode
3572 static noinline_for_stack int
3573 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3575 struct super_block *sb = ac->ac_sb;
3576 struct ext4_sb_info *sbi = EXT4_SB(sb);
3577 struct ext4_prealloc_space *pa;
3578 struct ext4_group_info *grp;
3579 struct ext4_inode_info *ei;
3581 /* preallocate only when found space is larger then requested */
3582 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3583 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3584 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3586 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3590 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3596 /* we can't allocate as much as normalizer wants.
3597 * so, found space must get proper lstart
3598 * to cover original request */
3599 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3600 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3602 /* we're limited by original request in that
3603 * logical block must be covered any way
3604 * winl is window we can move our chunk within */
3605 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3607 /* also, we should cover whole original request */
3608 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3610 /* the smallest one defines real window */
3611 win = min(winl, wins);
3613 offs = ac->ac_o_ex.fe_logical %
3614 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3615 if (offs && offs < win)
3618 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3619 EXT4_NUM_B2C(sbi, win);
3620 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3621 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3624 /* preallocation can change ac_b_ex, thus we store actually
3625 * allocated blocks for history */
3626 ac->ac_f_ex = ac->ac_b_ex;
3628 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3629 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3630 pa->pa_len = ac->ac_b_ex.fe_len;
3631 pa->pa_free = pa->pa_len;
3632 atomic_set(&pa->pa_count, 1);
3633 spin_lock_init(&pa->pa_lock);
3634 INIT_LIST_HEAD(&pa->pa_inode_list);
3635 INIT_LIST_HEAD(&pa->pa_group_list);
3637 pa->pa_type = MB_INODE_PA;
3639 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3640 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3641 trace_ext4_mb_new_inode_pa(ac, pa);
3643 ext4_mb_use_inode_pa(ac, pa);
3644 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3646 ei = EXT4_I(ac->ac_inode);
3647 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3649 pa->pa_obj_lock = &ei->i_prealloc_lock;
3650 pa->pa_inode = ac->ac_inode;
3652 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3653 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3654 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3656 spin_lock(pa->pa_obj_lock);
3657 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3658 spin_unlock(pa->pa_obj_lock);
3664 * creates new preallocated space for locality group inodes belongs to
3666 static noinline_for_stack int
3667 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3669 struct super_block *sb = ac->ac_sb;
3670 struct ext4_locality_group *lg;
3671 struct ext4_prealloc_space *pa;
3672 struct ext4_group_info *grp;
3674 /* preallocate only when found space is larger then requested */
3675 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3676 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3677 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3679 BUG_ON(ext4_pspace_cachep == NULL);
3680 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3684 /* preallocation can change ac_b_ex, thus we store actually
3685 * allocated blocks for history */
3686 ac->ac_f_ex = ac->ac_b_ex;
3688 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3689 pa->pa_lstart = pa->pa_pstart;
3690 pa->pa_len = ac->ac_b_ex.fe_len;
3691 pa->pa_free = pa->pa_len;
3692 atomic_set(&pa->pa_count, 1);
3693 spin_lock_init(&pa->pa_lock);
3694 INIT_LIST_HEAD(&pa->pa_inode_list);
3695 INIT_LIST_HEAD(&pa->pa_group_list);
3697 pa->pa_type = MB_GROUP_PA;
3699 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3700 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3701 trace_ext4_mb_new_group_pa(ac, pa);
3703 ext4_mb_use_group_pa(ac, pa);
3704 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3706 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3710 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3711 pa->pa_inode = NULL;
3713 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3714 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3715 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3718 * We will later add the new pa to the right bucket
3719 * after updating the pa_free in ext4_mb_release_context
3724 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3728 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3729 err = ext4_mb_new_group_pa(ac);
3731 err = ext4_mb_new_inode_pa(ac);
3736 * finds all unused blocks in on-disk bitmap, frees them in
3737 * in-core bitmap and buddy.
3738 * @pa must be unlinked from inode and group lists, so that
3739 * nobody else can find/use it.
3740 * the caller MUST hold group/inode locks.
3741 * TODO: optimize the case when there are no in-core structures yet
3743 static noinline_for_stack int
3744 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3745 struct ext4_prealloc_space *pa)
3747 struct super_block *sb = e4b->bd_sb;
3748 struct ext4_sb_info *sbi = EXT4_SB(sb);
3753 unsigned long long grp_blk_start;
3757 BUG_ON(pa->pa_deleted == 0);
3758 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3759 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3760 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3761 end = bit + pa->pa_len;
3764 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3767 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3768 mb_debug(1, " free preallocated %u/%u in group %u\n",
3769 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3770 (unsigned) next - bit, (unsigned) group);
3773 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3774 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3775 EXT4_C2B(sbi, bit)),
3777 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3780 if (free != pa->pa_free) {
3781 ext4_msg(e4b->bd_sb, KERN_CRIT,
3782 "pa %p: logic %lu, phys. %lu, len %lu",
3783 pa, (unsigned long) pa->pa_lstart,
3784 (unsigned long) pa->pa_pstart,
3785 (unsigned long) pa->pa_len);
3786 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3789 * pa is already deleted so we use the value obtained
3790 * from the bitmap and continue.
3793 atomic_add(free, &sbi->s_mb_discarded);
3798 static noinline_for_stack int
3799 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3800 struct ext4_prealloc_space *pa)
3802 struct super_block *sb = e4b->bd_sb;
3806 trace_ext4_mb_release_group_pa(sb, pa);
3807 BUG_ON(pa->pa_deleted == 0);
3808 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3809 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3810 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3811 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3812 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3818 * releases all preallocations in given group
3820 * first, we need to decide discard policy:
3821 * - when do we discard
3823 * - how many do we discard
3824 * 1) how many requested
3826 static noinline_for_stack int
3827 ext4_mb_discard_group_preallocations(struct super_block *sb,
3828 ext4_group_t group, int needed)
3830 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3831 struct buffer_head *bitmap_bh = NULL;
3832 struct ext4_prealloc_space *pa, *tmp;
3833 struct list_head list;
3834 struct ext4_buddy e4b;
3839 mb_debug(1, "discard preallocation for group %u\n", group);
3841 if (list_empty(&grp->bb_prealloc_list))
3844 bitmap_bh = ext4_read_block_bitmap(sb, group);
3845 if (bitmap_bh == NULL) {
3846 ext4_error(sb, "Error reading block bitmap for %u", group);
3850 err = ext4_mb_load_buddy(sb, group, &e4b);
3852 ext4_error(sb, "Error loading buddy information for %u", group);
3858 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3860 INIT_LIST_HEAD(&list);
3862 ext4_lock_group(sb, group);
3863 list_for_each_entry_safe(pa, tmp,
3864 &grp->bb_prealloc_list, pa_group_list) {
3865 spin_lock(&pa->pa_lock);
3866 if (atomic_read(&pa->pa_count)) {
3867 spin_unlock(&pa->pa_lock);
3871 if (pa->pa_deleted) {
3872 spin_unlock(&pa->pa_lock);
3876 /* seems this one can be freed ... */
3879 /* we can trust pa_free ... */
3880 free += pa->pa_free;
3882 spin_unlock(&pa->pa_lock);
3884 list_del(&pa->pa_group_list);
3885 list_add(&pa->u.pa_tmp_list, &list);
3888 /* if we still need more blocks and some PAs were used, try again */
3889 if (free < needed && busy) {
3891 ext4_unlock_group(sb, group);
3896 /* found anything to free? */
3897 if (list_empty(&list)) {
3902 /* now free all selected PAs */
3903 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3905 /* remove from object (inode or locality group) */
3906 spin_lock(pa->pa_obj_lock);
3907 list_del_rcu(&pa->pa_inode_list);
3908 spin_unlock(pa->pa_obj_lock);
3910 if (pa->pa_type == MB_GROUP_PA)
3911 ext4_mb_release_group_pa(&e4b, pa);
3913 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3915 list_del(&pa->u.pa_tmp_list);
3916 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3920 ext4_unlock_group(sb, group);
3921 ext4_mb_unload_buddy(&e4b);
3927 * releases all non-used preallocated blocks for given inode
3929 * It's important to discard preallocations under i_data_sem
3930 * We don't want another block to be served from the prealloc
3931 * space when we are discarding the inode prealloc space.
3933 * FIXME!! Make sure it is valid at all the call sites
3935 void ext4_discard_preallocations(struct inode *inode)
3937 struct ext4_inode_info *ei = EXT4_I(inode);
3938 struct super_block *sb = inode->i_sb;
3939 struct buffer_head *bitmap_bh = NULL;
3940 struct ext4_prealloc_space *pa, *tmp;
3941 ext4_group_t group = 0;
3942 struct list_head list;
3943 struct ext4_buddy e4b;
3946 if (!S_ISREG(inode->i_mode)) {
3947 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3951 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3952 trace_ext4_discard_preallocations(inode);
3954 INIT_LIST_HEAD(&list);
3957 /* first, collect all pa's in the inode */
3958 spin_lock(&ei->i_prealloc_lock);
3959 while (!list_empty(&ei->i_prealloc_list)) {
3960 pa = list_entry(ei->i_prealloc_list.next,
3961 struct ext4_prealloc_space, pa_inode_list);
3962 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3963 spin_lock(&pa->pa_lock);
3964 if (atomic_read(&pa->pa_count)) {
3965 /* this shouldn't happen often - nobody should
3966 * use preallocation while we're discarding it */
3967 spin_unlock(&pa->pa_lock);
3968 spin_unlock(&ei->i_prealloc_lock);
3969 ext4_msg(sb, KERN_ERR,
3970 "uh-oh! used pa while discarding");
3972 schedule_timeout_uninterruptible(HZ);
3976 if (pa->pa_deleted == 0) {
3978 spin_unlock(&pa->pa_lock);
3979 list_del_rcu(&pa->pa_inode_list);
3980 list_add(&pa->u.pa_tmp_list, &list);
3984 /* someone is deleting pa right now */
3985 spin_unlock(&pa->pa_lock);
3986 spin_unlock(&ei->i_prealloc_lock);
3988 /* we have to wait here because pa_deleted
3989 * doesn't mean pa is already unlinked from
3990 * the list. as we might be called from
3991 * ->clear_inode() the inode will get freed
3992 * and concurrent thread which is unlinking
3993 * pa from inode's list may access already
3994 * freed memory, bad-bad-bad */
3996 /* XXX: if this happens too often, we can
3997 * add a flag to force wait only in case
3998 * of ->clear_inode(), but not in case of
3999 * regular truncate */
4000 schedule_timeout_uninterruptible(HZ);
4003 spin_unlock(&ei->i_prealloc_lock);
4005 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4006 BUG_ON(pa->pa_type != MB_INODE_PA);
4007 group = ext4_get_group_number(sb, pa->pa_pstart);
4009 err = ext4_mb_load_buddy(sb, group, &e4b);
4011 ext4_error(sb, "Error loading buddy information for %u",
4016 bitmap_bh = ext4_read_block_bitmap(sb, group);
4017 if (bitmap_bh == NULL) {
4018 ext4_error(sb, "Error reading block bitmap for %u",
4020 ext4_mb_unload_buddy(&e4b);
4024 ext4_lock_group(sb, group);
4025 list_del(&pa->pa_group_list);
4026 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4027 ext4_unlock_group(sb, group);
4029 ext4_mb_unload_buddy(&e4b);
4032 list_del(&pa->u.pa_tmp_list);
4033 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4037 #ifdef CONFIG_EXT4_DEBUG
4038 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4040 struct super_block *sb = ac->ac_sb;
4041 ext4_group_t ngroups, i;
4043 if (!ext4_mballoc_debug ||
4044 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4047 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4048 " Allocation context details:");
4049 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4050 ac->ac_status, ac->ac_flags);
4051 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4052 "goal %lu/%lu/%lu@%lu, "
4053 "best %lu/%lu/%lu@%lu cr %d",
4054 (unsigned long)ac->ac_o_ex.fe_group,
4055 (unsigned long)ac->ac_o_ex.fe_start,
4056 (unsigned long)ac->ac_o_ex.fe_len,
4057 (unsigned long)ac->ac_o_ex.fe_logical,
4058 (unsigned long)ac->ac_g_ex.fe_group,
4059 (unsigned long)ac->ac_g_ex.fe_start,
4060 (unsigned long)ac->ac_g_ex.fe_len,
4061 (unsigned long)ac->ac_g_ex.fe_logical,
4062 (unsigned long)ac->ac_b_ex.fe_group,
4063 (unsigned long)ac->ac_b_ex.fe_start,
4064 (unsigned long)ac->ac_b_ex.fe_len,
4065 (unsigned long)ac->ac_b_ex.fe_logical,
4066 (int)ac->ac_criteria);
4067 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4068 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4069 ngroups = ext4_get_groups_count(sb);
4070 for (i = 0; i < ngroups; i++) {
4071 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4072 struct ext4_prealloc_space *pa;
4073 ext4_grpblk_t start;
4074 struct list_head *cur;
4075 ext4_lock_group(sb, i);
4076 list_for_each(cur, &grp->bb_prealloc_list) {
4077 pa = list_entry(cur, struct ext4_prealloc_space,
4079 spin_lock(&pa->pa_lock);
4080 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4082 spin_unlock(&pa->pa_lock);
4083 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4086 ext4_unlock_group(sb, i);
4088 if (grp->bb_free == 0)
4090 printk(KERN_ERR "%u: %d/%d \n",
4091 i, grp->bb_free, grp->bb_fragments);
4093 printk(KERN_ERR "\n");
4096 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4103 * We use locality group preallocation for small size file. The size of the
4104 * file is determined by the current size or the resulting size after
4105 * allocation which ever is larger
4107 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4109 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4111 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4112 int bsbits = ac->ac_sb->s_blocksize_bits;
4115 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4118 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4121 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4122 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4125 if ((size == isize) &&
4126 !ext4_fs_is_busy(sbi) &&
4127 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4128 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4132 if (sbi->s_mb_group_prealloc <= 0) {
4133 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4137 /* don't use group allocation for large files */
4138 size = max(size, isize);
4139 if (size > sbi->s_mb_stream_request) {
4140 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4144 BUG_ON(ac->ac_lg != NULL);
4146 * locality group prealloc space are per cpu. The reason for having
4147 * per cpu locality group is to reduce the contention between block
4148 * request from multiple CPUs.
4150 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4152 /* we're going to use group allocation */
4153 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4155 /* serialize all allocations in the group */
4156 mutex_lock(&ac->ac_lg->lg_mutex);
4159 static noinline_for_stack int
4160 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4161 struct ext4_allocation_request *ar)
4163 struct super_block *sb = ar->inode->i_sb;
4164 struct ext4_sb_info *sbi = EXT4_SB(sb);
4165 struct ext4_super_block *es = sbi->s_es;
4169 ext4_grpblk_t block;
4171 /* we can't allocate > group size */
4174 /* just a dirty hack to filter too big requests */
4175 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4176 len = EXT4_CLUSTERS_PER_GROUP(sb);
4178 /* start searching from the goal */
4180 if (goal < le32_to_cpu(es->s_first_data_block) ||
4181 goal >= ext4_blocks_count(es))
4182 goal = le32_to_cpu(es->s_first_data_block);
4183 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4185 /* set up allocation goals */
4186 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4187 ac->ac_status = AC_STATUS_CONTINUE;
4189 ac->ac_inode = ar->inode;
4190 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4191 ac->ac_o_ex.fe_group = group;
4192 ac->ac_o_ex.fe_start = block;
4193 ac->ac_o_ex.fe_len = len;
4194 ac->ac_g_ex = ac->ac_o_ex;
4195 ac->ac_flags = ar->flags;
4197 /* we have to define context: we'll we work with a file or
4198 * locality group. this is a policy, actually */
4199 ext4_mb_group_or_file(ac);
4201 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4202 "left: %u/%u, right %u/%u to %swritable\n",
4203 (unsigned) ar->len, (unsigned) ar->logical,
4204 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4205 (unsigned) ar->lleft, (unsigned) ar->pleft,
4206 (unsigned) ar->lright, (unsigned) ar->pright,
4207 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4212 static noinline_for_stack void
4213 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4214 struct ext4_locality_group *lg,
4215 int order, int total_entries)
4217 ext4_group_t group = 0;
4218 struct ext4_buddy e4b;
4219 struct list_head discard_list;
4220 struct ext4_prealloc_space *pa, *tmp;
4222 mb_debug(1, "discard locality group preallocation\n");
4224 INIT_LIST_HEAD(&discard_list);
4226 spin_lock(&lg->lg_prealloc_lock);
4227 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4229 spin_lock(&pa->pa_lock);
4230 if (atomic_read(&pa->pa_count)) {
4232 * This is the pa that we just used
4233 * for block allocation. So don't
4236 spin_unlock(&pa->pa_lock);
4239 if (pa->pa_deleted) {
4240 spin_unlock(&pa->pa_lock);
4243 /* only lg prealloc space */
4244 BUG_ON(pa->pa_type != MB_GROUP_PA);
4246 /* seems this one can be freed ... */
4248 spin_unlock(&pa->pa_lock);
4250 list_del_rcu(&pa->pa_inode_list);
4251 list_add(&pa->u.pa_tmp_list, &discard_list);
4254 if (total_entries <= 5) {
4256 * we want to keep only 5 entries
4257 * allowing it to grow to 8. This
4258 * mak sure we don't call discard
4259 * soon for this list.
4264 spin_unlock(&lg->lg_prealloc_lock);
4266 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4268 group = ext4_get_group_number(sb, pa->pa_pstart);
4269 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4270 ext4_error(sb, "Error loading buddy information for %u",
4274 ext4_lock_group(sb, group);
4275 list_del(&pa->pa_group_list);
4276 ext4_mb_release_group_pa(&e4b, pa);
4277 ext4_unlock_group(sb, group);
4279 ext4_mb_unload_buddy(&e4b);
4280 list_del(&pa->u.pa_tmp_list);
4281 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4286 * We have incremented pa_count. So it cannot be freed at this
4287 * point. Also we hold lg_mutex. So no parallel allocation is
4288 * possible from this lg. That means pa_free cannot be updated.
4290 * A parallel ext4_mb_discard_group_preallocations is possible.
4291 * which can cause the lg_prealloc_list to be updated.
4294 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4296 int order, added = 0, lg_prealloc_count = 1;
4297 struct super_block *sb = ac->ac_sb;
4298 struct ext4_locality_group *lg = ac->ac_lg;
4299 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4301 order = fls(pa->pa_free) - 1;
4302 if (order > PREALLOC_TB_SIZE - 1)
4303 /* The max size of hash table is PREALLOC_TB_SIZE */
4304 order = PREALLOC_TB_SIZE - 1;
4305 /* Add the prealloc space to lg */
4306 spin_lock(&lg->lg_prealloc_lock);
4307 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4309 spin_lock(&tmp_pa->pa_lock);
4310 if (tmp_pa->pa_deleted) {
4311 spin_unlock(&tmp_pa->pa_lock);
4314 if (!added && pa->pa_free < tmp_pa->pa_free) {
4315 /* Add to the tail of the previous entry */
4316 list_add_tail_rcu(&pa->pa_inode_list,
4317 &tmp_pa->pa_inode_list);
4320 * we want to count the total
4321 * number of entries in the list
4324 spin_unlock(&tmp_pa->pa_lock);
4325 lg_prealloc_count++;
4328 list_add_tail_rcu(&pa->pa_inode_list,
4329 &lg->lg_prealloc_list[order]);
4330 spin_unlock(&lg->lg_prealloc_lock);
4332 /* Now trim the list to be not more than 8 elements */
4333 if (lg_prealloc_count > 8) {
4334 ext4_mb_discard_lg_preallocations(sb, lg,
4335 order, lg_prealloc_count);
4342 * release all resource we used in allocation
4344 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4346 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4347 struct ext4_prealloc_space *pa = ac->ac_pa;
4349 if (pa->pa_type == MB_GROUP_PA) {
4350 /* see comment in ext4_mb_use_group_pa() */
4351 spin_lock(&pa->pa_lock);
4352 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4353 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4354 pa->pa_free -= ac->ac_b_ex.fe_len;
4355 pa->pa_len -= ac->ac_b_ex.fe_len;
4356 spin_unlock(&pa->pa_lock);
4361 * We want to add the pa to the right bucket.
4362 * Remove it from the list and while adding
4363 * make sure the list to which we are adding
4366 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4367 spin_lock(pa->pa_obj_lock);
4368 list_del_rcu(&pa->pa_inode_list);
4369 spin_unlock(pa->pa_obj_lock);
4370 ext4_mb_add_n_trim(ac);
4372 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4374 if (ac->ac_bitmap_page)
4375 page_cache_release(ac->ac_bitmap_page);
4376 if (ac->ac_buddy_page)
4377 page_cache_release(ac->ac_buddy_page);
4378 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4379 mutex_unlock(&ac->ac_lg->lg_mutex);
4380 ext4_mb_collect_stats(ac);
4384 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4386 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4390 trace_ext4_mb_discard_preallocations(sb, needed);
4391 for (i = 0; i < ngroups && needed > 0; i++) {
4392 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4401 * Main entry point into mballoc to allocate blocks
4402 * it tries to use preallocation first, then falls back
4403 * to usual allocation
4405 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4406 struct ext4_allocation_request *ar, int *errp)
4409 struct ext4_allocation_context *ac = NULL;
4410 struct ext4_sb_info *sbi;
4411 struct super_block *sb;
4412 ext4_fsblk_t block = 0;
4413 unsigned int inquota = 0;
4414 unsigned int reserv_clstrs = 0;
4417 sb = ar->inode->i_sb;
4420 trace_ext4_request_blocks(ar);
4422 /* Allow to use superuser reservation for quota file */
4423 if (IS_NOQUOTA(ar->inode))
4424 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4426 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4427 /* Without delayed allocation we need to verify
4428 * there is enough free blocks to do block allocation
4429 * and verify allocation doesn't exceed the quota limits.
4432 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4434 /* let others to free the space */
4436 ar->len = ar->len >> 1;
4442 reserv_clstrs = ar->len;
4443 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4444 dquot_alloc_block_nofail(ar->inode,
4445 EXT4_C2B(sbi, ar->len));
4448 dquot_alloc_block(ar->inode,
4449 EXT4_C2B(sbi, ar->len))) {
4451 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4462 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4469 *errp = ext4_mb_initialize_context(ac, ar);
4475 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4476 if (!ext4_mb_use_preallocated(ac)) {
4477 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4478 ext4_mb_normalize_request(ac, ar);
4480 /* allocate space in core */
4481 *errp = ext4_mb_regular_allocator(ac);
4483 goto discard_and_exit;
4485 /* as we've just preallocated more space than
4486 * user requested originally, we store allocated
4487 * space in a special descriptor */
4488 if (ac->ac_status == AC_STATUS_FOUND &&
4489 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4490 *errp = ext4_mb_new_preallocation(ac);
4493 ext4_discard_allocated_blocks(ac);
4497 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4498 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4499 if (*errp == -EAGAIN) {
4501 * drop the reference that we took
4502 * in ext4_mb_use_best_found
4504 ext4_mb_release_context(ac);
4505 ac->ac_b_ex.fe_group = 0;
4506 ac->ac_b_ex.fe_start = 0;
4507 ac->ac_b_ex.fe_len = 0;
4508 ac->ac_status = AC_STATUS_CONTINUE;
4511 ext4_discard_allocated_blocks(ac);
4514 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4515 ar->len = ac->ac_b_ex.fe_len;
4518 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4526 ac->ac_b_ex.fe_len = 0;
4528 ext4_mb_show_ac(ac);
4530 ext4_mb_release_context(ac);
4533 kmem_cache_free(ext4_ac_cachep, ac);
4534 if (inquota && ar->len < inquota)
4535 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4537 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4538 /* release all the reserved blocks if non delalloc */
4539 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4543 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4549 * We can merge two free data extents only if the physical blocks
4550 * are contiguous, AND the extents were freed by the same transaction,
4551 * AND the blocks are associated with the same group.
4553 static int can_merge(struct ext4_free_data *entry1,
4554 struct ext4_free_data *entry2)
4556 if ((entry1->efd_tid == entry2->efd_tid) &&
4557 (entry1->efd_group == entry2->efd_group) &&
4558 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4563 static noinline_for_stack int
4564 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4565 struct ext4_free_data *new_entry)
4567 ext4_group_t group = e4b->bd_group;
4568 ext4_grpblk_t cluster;
4569 struct ext4_free_data *entry;
4570 struct ext4_group_info *db = e4b->bd_info;
4571 struct super_block *sb = e4b->bd_sb;
4572 struct ext4_sb_info *sbi = EXT4_SB(sb);
4573 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4574 struct rb_node *parent = NULL, *new_node;
4576 BUG_ON(!ext4_handle_valid(handle));
4577 BUG_ON(e4b->bd_bitmap_page == NULL);
4578 BUG_ON(e4b->bd_buddy_page == NULL);
4580 new_node = &new_entry->efd_node;
4581 cluster = new_entry->efd_start_cluster;
4584 /* first free block exent. We need to
4585 protect buddy cache from being freed,
4586 * otherwise we'll refresh it from
4587 * on-disk bitmap and lose not-yet-available
4589 page_cache_get(e4b->bd_buddy_page);
4590 page_cache_get(e4b->bd_bitmap_page);
4594 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4595 if (cluster < entry->efd_start_cluster)
4597 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4598 n = &(*n)->rb_right;
4600 ext4_grp_locked_error(sb, group, 0,
4601 ext4_group_first_block_no(sb, group) +
4602 EXT4_C2B(sbi, cluster),
4603 "Block already on to-be-freed list");
4608 rb_link_node(new_node, parent, n);
4609 rb_insert_color(new_node, &db->bb_free_root);
4611 /* Now try to see the extent can be merged to left and right */
4612 node = rb_prev(new_node);
4614 entry = rb_entry(node, struct ext4_free_data, efd_node);
4615 if (can_merge(entry, new_entry) &&
4616 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4617 new_entry->efd_start_cluster = entry->efd_start_cluster;
4618 new_entry->efd_count += entry->efd_count;
4619 rb_erase(node, &(db->bb_free_root));
4620 kmem_cache_free(ext4_free_data_cachep, entry);
4624 node = rb_next(new_node);
4626 entry = rb_entry(node, struct ext4_free_data, efd_node);
4627 if (can_merge(new_entry, entry) &&
4628 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4629 new_entry->efd_count += entry->efd_count;
4630 rb_erase(node, &(db->bb_free_root));
4631 kmem_cache_free(ext4_free_data_cachep, entry);
4634 /* Add the extent to transaction's private list */
4635 ext4_journal_callback_add(handle, ext4_free_data_callback,
4636 &new_entry->efd_jce);
4641 * ext4_free_blocks() -- Free given blocks and update quota
4642 * @handle: handle for this transaction
4644 * @block: start physical block to free
4645 * @count: number of blocks to count
4646 * @flags: flags used by ext4_free_blocks
4648 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4649 struct buffer_head *bh, ext4_fsblk_t block,
4650 unsigned long count, int flags)
4652 struct buffer_head *bitmap_bh = NULL;
4653 struct super_block *sb = inode->i_sb;
4654 struct ext4_group_desc *gdp;
4655 unsigned int overflow;
4657 struct buffer_head *gd_bh;
4658 ext4_group_t block_group;
4659 struct ext4_sb_info *sbi;
4660 struct ext4_buddy e4b;
4661 unsigned int count_clusters;
4668 BUG_ON(block != bh->b_blocknr);
4670 block = bh->b_blocknr;
4674 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4675 !ext4_data_block_valid(sbi, block, count)) {
4676 ext4_error(sb, "Freeing blocks not in datazone - "
4677 "block = %llu, count = %lu", block, count);
4681 ext4_debug("freeing block %llu\n", block);
4682 trace_ext4_free_blocks(inode, block, count, flags);
4684 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4685 struct buffer_head *tbh = bh;
4688 BUG_ON(bh && (count > 1));
4690 for (i = 0; i < count; i++) {
4693 tbh = sb_find_get_block(inode->i_sb,
4697 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4698 inode, tbh, block + i);
4703 * We need to make sure we don't reuse the freed block until
4704 * after the transaction is committed, which we can do by
4705 * treating the block as metadata, below. We make an
4706 * exception if the inode is to be written in writeback mode
4707 * since writeback mode has weak data consistency guarantees.
4709 if (!ext4_should_writeback_data(inode))
4710 flags |= EXT4_FREE_BLOCKS_METADATA;
4713 * If the extent to be freed does not begin on a cluster
4714 * boundary, we need to deal with partial clusters at the
4715 * beginning and end of the extent. Normally we will free
4716 * blocks at the beginning or the end unless we are explicitly
4717 * requested to avoid doing so.
4719 overflow = EXT4_PBLK_COFF(sbi, block);
4721 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4722 overflow = sbi->s_cluster_ratio - overflow;
4724 if (count > overflow)
4733 overflow = EXT4_LBLK_COFF(sbi, count);
4735 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4736 if (count > overflow)
4741 count += sbi->s_cluster_ratio - overflow;
4746 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4748 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4749 ext4_get_group_info(sb, block_group))))
4753 * Check to see if we are freeing blocks across a group
4756 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4757 overflow = EXT4_C2B(sbi, bit) + count -
4758 EXT4_BLOCKS_PER_GROUP(sb);
4761 count_clusters = EXT4_NUM_B2C(sbi, count);
4762 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4767 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4773 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4774 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4775 in_range(block, ext4_inode_table(sb, gdp),
4776 EXT4_SB(sb)->s_itb_per_group) ||
4777 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4778 EXT4_SB(sb)->s_itb_per_group)) {
4780 ext4_error(sb, "Freeing blocks in system zone - "
4781 "Block = %llu, count = %lu", block, count);
4782 /* err = 0. ext4_std_error should be a no op */
4786 BUFFER_TRACE(bitmap_bh, "getting write access");
4787 err = ext4_journal_get_write_access(handle, bitmap_bh);
4792 * We are about to modify some metadata. Call the journal APIs
4793 * to unshare ->b_data if a currently-committing transaction is
4796 BUFFER_TRACE(gd_bh, "get_write_access");
4797 err = ext4_journal_get_write_access(handle, gd_bh);
4800 #ifdef AGGRESSIVE_CHECK
4803 for (i = 0; i < count_clusters; i++)
4804 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4807 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4809 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4813 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4814 struct ext4_free_data *new_entry;
4816 * blocks being freed are metadata. these blocks shouldn't
4817 * be used until this transaction is committed
4819 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4822 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4823 GFP_NOFS|__GFP_NOFAIL);
4824 new_entry->efd_start_cluster = bit;
4825 new_entry->efd_group = block_group;
4826 new_entry->efd_count = count_clusters;
4827 new_entry->efd_tid = handle->h_transaction->t_tid;
4829 ext4_lock_group(sb, block_group);
4830 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4831 ext4_mb_free_metadata(handle, &e4b, new_entry);
4833 /* need to update group_info->bb_free and bitmap
4834 * with group lock held. generate_buddy look at
4835 * them with group lock_held
4837 if (test_opt(sb, DISCARD)) {
4838 err = ext4_issue_discard(sb, block_group, bit, count);
4839 if (err && err != -EOPNOTSUPP)
4840 ext4_msg(sb, KERN_WARNING, "discard request in"
4841 " group:%d block:%d count:%lu failed"
4842 " with %d", block_group, bit, count,
4845 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4847 ext4_lock_group(sb, block_group);
4848 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4849 mb_free_blocks(inode, &e4b, bit, count_clusters);
4852 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4853 ext4_free_group_clusters_set(sb, gdp, ret);
4854 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4855 ext4_group_desc_csum_set(sb, block_group, gdp);
4856 ext4_unlock_group(sb, block_group);
4858 if (sbi->s_log_groups_per_flex) {
4859 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4860 atomic64_add(count_clusters,
4861 &sbi->s_flex_groups[flex_group].free_clusters);
4864 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4865 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4866 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4868 ext4_mb_unload_buddy(&e4b);
4870 /* We dirtied the bitmap block */
4871 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4872 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4874 /* And the group descriptor block */
4875 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4876 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4880 if (overflow && !err) {
4888 ext4_std_error(sb, err);
4893 * ext4_group_add_blocks() -- Add given blocks to an existing group
4894 * @handle: handle to this transaction
4896 * @block: start physical block to add to the block group
4897 * @count: number of blocks to free
4899 * This marks the blocks as free in the bitmap and buddy.
4901 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4902 ext4_fsblk_t block, unsigned long count)
4904 struct buffer_head *bitmap_bh = NULL;
4905 struct buffer_head *gd_bh;
4906 ext4_group_t block_group;
4909 struct ext4_group_desc *desc;
4910 struct ext4_sb_info *sbi = EXT4_SB(sb);
4911 struct ext4_buddy e4b;
4912 int err = 0, ret, blk_free_count;
4913 ext4_grpblk_t blocks_freed;
4915 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4920 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4922 * Check to see if we are freeing blocks across a group
4925 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4926 ext4_warning(sb, "too much blocks added to group %u\n",
4932 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4938 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4944 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4945 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4946 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4947 in_range(block + count - 1, ext4_inode_table(sb, desc),
4948 sbi->s_itb_per_group)) {
4949 ext4_error(sb, "Adding blocks in system zones - "
4950 "Block = %llu, count = %lu",
4956 BUFFER_TRACE(bitmap_bh, "getting write access");
4957 err = ext4_journal_get_write_access(handle, bitmap_bh);
4962 * We are about to modify some metadata. Call the journal APIs
4963 * to unshare ->b_data if a currently-committing transaction is
4966 BUFFER_TRACE(gd_bh, "get_write_access");
4967 err = ext4_journal_get_write_access(handle, gd_bh);
4971 for (i = 0, blocks_freed = 0; i < count; i++) {
4972 BUFFER_TRACE(bitmap_bh, "clear bit");
4973 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4974 ext4_error(sb, "bit already cleared for block %llu",
4975 (ext4_fsblk_t)(block + i));
4976 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4982 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4987 * need to update group_info->bb_free and bitmap
4988 * with group lock held. generate_buddy look at
4989 * them with group lock_held
4991 ext4_lock_group(sb, block_group);
4992 mb_clear_bits(bitmap_bh->b_data, bit, count);
4993 mb_free_blocks(NULL, &e4b, bit, count);
4994 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4995 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4996 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
4997 ext4_group_desc_csum_set(sb, block_group, desc);
4998 ext4_unlock_group(sb, block_group);
4999 percpu_counter_add(&sbi->s_freeclusters_counter,
5000 EXT4_NUM_B2C(sbi, blocks_freed));
5002 if (sbi->s_log_groups_per_flex) {
5003 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5004 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5005 &sbi->s_flex_groups[flex_group].free_clusters);
5008 ext4_mb_unload_buddy(&e4b);
5010 /* We dirtied the bitmap block */
5011 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5012 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5014 /* And the group descriptor block */
5015 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5016 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5022 ext4_std_error(sb, err);
5027 * ext4_trim_extent -- function to TRIM one single free extent in the group
5028 * @sb: super block for the file system
5029 * @start: starting block of the free extent in the alloc. group
5030 * @count: number of blocks to TRIM
5031 * @group: alloc. group we are working with
5032 * @e4b: ext4 buddy for the group
5034 * Trim "count" blocks starting at "start" in the "group". To assure that no
5035 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5036 * be called with under the group lock.
5038 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5039 ext4_group_t group, struct ext4_buddy *e4b)
5043 struct ext4_free_extent ex;
5046 trace_ext4_trim_extent(sb, group, start, count);
5048 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5050 ex.fe_start = start;
5051 ex.fe_group = group;
5055 * Mark blocks used, so no one can reuse them while
5058 mb_mark_used(e4b, &ex);
5059 ext4_unlock_group(sb, group);
5060 ret = ext4_issue_discard(sb, group, start, count);
5061 ext4_lock_group(sb, group);
5062 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5067 * ext4_trim_all_free -- function to trim all free space in alloc. group
5068 * @sb: super block for file system
5069 * @group: group to be trimmed
5070 * @start: first group block to examine
5071 * @max: last group block to examine
5072 * @minblocks: minimum extent block count
5074 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5075 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5079 * ext4_trim_all_free walks through group's block bitmap searching for free
5080 * extents. When the free extent is found, mark it as used in group buddy
5081 * bitmap. Then issue a TRIM command on this extent and free the extent in
5082 * the group buddy bitmap. This is done until whole group is scanned.
5084 static ext4_grpblk_t
5085 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5086 ext4_grpblk_t start, ext4_grpblk_t max,
5087 ext4_grpblk_t minblocks)
5090 ext4_grpblk_t next, count = 0, free_count = 0;
5091 struct ext4_buddy e4b;
5094 trace_ext4_trim_all_free(sb, group, start, max);
5096 ret = ext4_mb_load_buddy(sb, group, &e4b);
5098 ext4_error(sb, "Error in loading buddy "
5099 "information for %u", group);
5102 bitmap = e4b.bd_bitmap;
5104 ext4_lock_group(sb, group);
5105 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5106 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5109 start = (e4b.bd_info->bb_first_free > start) ?
5110 e4b.bd_info->bb_first_free : start;
5112 while (start <= max) {
5113 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5116 next = mb_find_next_bit(bitmap, max + 1, start);
5118 if ((next - start) >= minblocks) {
5119 ret = ext4_trim_extent(sb, start,
5120 next - start, group, &e4b);
5121 if (ret && ret != -EOPNOTSUPP)
5124 count += next - start;
5126 free_count += next - start;
5129 if (fatal_signal_pending(current)) {
5130 count = -ERESTARTSYS;
5134 if (need_resched()) {
5135 ext4_unlock_group(sb, group);
5137 ext4_lock_group(sb, group);
5140 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5146 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5149 ext4_unlock_group(sb, group);
5150 ext4_mb_unload_buddy(&e4b);
5152 ext4_debug("trimmed %d blocks in the group %d\n",
5159 * ext4_trim_fs() -- trim ioctl handle function
5160 * @sb: superblock for filesystem
5161 * @range: fstrim_range structure
5163 * start: First Byte to trim
5164 * len: number of Bytes to trim from start
5165 * minlen: minimum extent length in Bytes
5166 * ext4_trim_fs goes through all allocation groups containing Bytes from
5167 * start to start+len. For each such a group ext4_trim_all_free function
5168 * is invoked to trim all free space.
5170 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5172 struct ext4_group_info *grp;
5173 ext4_group_t group, first_group, last_group;
5174 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5175 uint64_t start, end, minlen, trimmed = 0;
5176 ext4_fsblk_t first_data_blk =
5177 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5178 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5181 start = range->start >> sb->s_blocksize_bits;
5182 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5183 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5184 range->minlen >> sb->s_blocksize_bits);
5186 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5187 start >= max_blks ||
5188 range->len < sb->s_blocksize)
5190 if (end >= max_blks)
5192 if (end <= first_data_blk)
5194 if (start < first_data_blk)
5195 start = first_data_blk;
5197 /* Determine first and last group to examine based on start and end */
5198 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5199 &first_group, &first_cluster);
5200 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5201 &last_group, &last_cluster);
5203 /* end now represents the last cluster to discard in this group */
5204 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5206 for (group = first_group; group <= last_group; group++) {
5207 grp = ext4_get_group_info(sb, group);
5208 /* We only do this if the grp has never been initialized */
5209 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5210 ret = ext4_mb_init_group(sb, group);
5216 * For all the groups except the last one, last cluster will
5217 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5218 * change it for the last group, note that last_cluster is
5219 * already computed earlier by ext4_get_group_no_and_offset()
5221 if (group == last_group)
5224 if (grp->bb_free >= minlen) {
5225 cnt = ext4_trim_all_free(sb, group, first_cluster,
5235 * For every group except the first one, we are sure
5236 * that the first cluster to discard will be cluster #0.
5242 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5245 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;