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])) {
891 first_block = page->index * blocks_per_page;
892 for (i = 0; i < blocks_per_page; i++) {
893 group = (first_block + i) >> 1;
894 if (group >= ngroups)
897 if (!bh[group - first_group])
898 /* skip initialized uptodate buddy */
902 * data carry information regarding this
903 * particular group in the format specified
907 data = page_address(page) + (i * blocksize);
908 bitmap = bh[group - first_group]->b_data;
911 * We place the buddy block and bitmap block
914 if ((first_block + i) & 1) {
915 /* this is block of buddy */
916 BUG_ON(incore == NULL);
917 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
918 group, page->index, i * blocksize);
919 trace_ext4_mb_buddy_bitmap_load(sb, group);
920 grinfo = ext4_get_group_info(sb, group);
921 grinfo->bb_fragments = 0;
922 memset(grinfo->bb_counters, 0,
923 sizeof(*grinfo->bb_counters) *
924 (sb->s_blocksize_bits+2));
926 * incore got set to the group block bitmap below
928 ext4_lock_group(sb, group);
930 memset(data, 0xff, blocksize);
931 ext4_mb_generate_buddy(sb, data, incore, group);
932 ext4_unlock_group(sb, group);
935 /* this is block of bitmap */
936 BUG_ON(incore != NULL);
937 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
938 group, page->index, i * blocksize);
939 trace_ext4_mb_bitmap_load(sb, group);
941 /* see comments in ext4_mb_put_pa() */
942 ext4_lock_group(sb, group);
943 memcpy(data, bitmap, blocksize);
945 /* mark all preallocated blks used in in-core bitmap */
946 ext4_mb_generate_from_pa(sb, data, group);
947 ext4_mb_generate_from_freelist(sb, data, group);
948 ext4_unlock_group(sb, group);
950 /* set incore so that the buddy information can be
951 * generated using this
956 SetPageUptodate(page);
960 for (i = 0; i < groups_per_page; i++)
969 * Lock the buddy and bitmap pages. This make sure other parallel init_group
970 * on the same buddy page doesn't happen whild holding the buddy page lock.
971 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
972 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
974 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
975 ext4_group_t group, struct ext4_buddy *e4b)
977 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
978 int block, pnum, poff;
982 e4b->bd_buddy_page = NULL;
983 e4b->bd_bitmap_page = NULL;
985 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
987 * the buddy cache inode stores the block bitmap
988 * and buddy information in consecutive blocks.
989 * So for each group we need two blocks.
992 pnum = block / blocks_per_page;
993 poff = block % blocks_per_page;
994 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
997 BUG_ON(page->mapping != inode->i_mapping);
998 e4b->bd_bitmap_page = page;
999 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1001 if (blocks_per_page >= 2) {
1002 /* buddy and bitmap are on the same page */
1007 pnum = block / blocks_per_page;
1008 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1011 BUG_ON(page->mapping != inode->i_mapping);
1012 e4b->bd_buddy_page = page;
1016 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1018 if (e4b->bd_bitmap_page) {
1019 unlock_page(e4b->bd_bitmap_page);
1020 page_cache_release(e4b->bd_bitmap_page);
1022 if (e4b->bd_buddy_page) {
1023 unlock_page(e4b->bd_buddy_page);
1024 page_cache_release(e4b->bd_buddy_page);
1029 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1030 * block group lock of all groups for this page; do not hold the BG lock when
1031 * calling this routine!
1033 static noinline_for_stack
1034 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1037 struct ext4_group_info *this_grp;
1038 struct ext4_buddy e4b;
1043 mb_debug(1, "init group %u\n", group);
1044 this_grp = ext4_get_group_info(sb, group);
1046 * This ensures that we don't reinit the buddy cache
1047 * page which map to the group from which we are already
1048 * allocating. If we are looking at the buddy cache we would
1049 * have taken a reference using ext4_mb_load_buddy and that
1050 * would have pinned buddy page to page cache.
1051 * The call to ext4_mb_get_buddy_page_lock will mark the
1054 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1055 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1057 * somebody initialized the group
1058 * return without doing anything
1063 page = e4b.bd_bitmap_page;
1064 ret = ext4_mb_init_cache(page, NULL);
1067 if (!PageUptodate(page)) {
1072 if (e4b.bd_buddy_page == NULL) {
1074 * If both the bitmap and buddy are in
1075 * the same page we don't need to force
1081 /* init buddy cache */
1082 page = e4b.bd_buddy_page;
1083 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1086 if (!PageUptodate(page)) {
1091 ext4_mb_put_buddy_page_lock(&e4b);
1096 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1097 * block group lock of all groups for this page; do not hold the BG lock when
1098 * calling this routine!
1100 static noinline_for_stack int
1101 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1102 struct ext4_buddy *e4b)
1104 int blocks_per_page;
1110 struct ext4_group_info *grp;
1111 struct ext4_sb_info *sbi = EXT4_SB(sb);
1112 struct inode *inode = sbi->s_buddy_cache;
1115 mb_debug(1, "load group %u\n", group);
1117 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1118 grp = ext4_get_group_info(sb, group);
1120 e4b->bd_blkbits = sb->s_blocksize_bits;
1123 e4b->bd_group = group;
1124 e4b->bd_buddy_page = NULL;
1125 e4b->bd_bitmap_page = NULL;
1127 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1129 * we need full data about the group
1130 * to make a good selection
1132 ret = ext4_mb_init_group(sb, group);
1138 * the buddy cache inode stores the block bitmap
1139 * and buddy information in consecutive blocks.
1140 * So for each group we need two blocks.
1143 pnum = block / blocks_per_page;
1144 poff = block % blocks_per_page;
1146 /* we could use find_or_create_page(), but it locks page
1147 * what we'd like to avoid in fast path ... */
1148 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1149 if (page == NULL || !PageUptodate(page)) {
1152 * drop the page reference and try
1153 * to get the page with lock. If we
1154 * are not uptodate that implies
1155 * somebody just created the page but
1156 * is yet to initialize the same. So
1157 * wait for it to initialize.
1159 page_cache_release(page);
1160 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1162 BUG_ON(page->mapping != inode->i_mapping);
1163 if (!PageUptodate(page)) {
1164 ret = ext4_mb_init_cache(page, NULL);
1169 mb_cmp_bitmaps(e4b, page_address(page) +
1170 (poff * sb->s_blocksize));
1179 if (!PageUptodate(page)) {
1184 /* Pages marked accessed already */
1185 e4b->bd_bitmap_page = page;
1186 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1189 pnum = block / blocks_per_page;
1190 poff = block % blocks_per_page;
1192 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1193 if (page == NULL || !PageUptodate(page)) {
1195 page_cache_release(page);
1196 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1198 BUG_ON(page->mapping != inode->i_mapping);
1199 if (!PageUptodate(page)) {
1200 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1213 if (!PageUptodate(page)) {
1218 /* Pages marked accessed already */
1219 e4b->bd_buddy_page = page;
1220 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1222 BUG_ON(e4b->bd_bitmap_page == NULL);
1223 BUG_ON(e4b->bd_buddy_page == NULL);
1229 page_cache_release(page);
1230 if (e4b->bd_bitmap_page)
1231 page_cache_release(e4b->bd_bitmap_page);
1232 if (e4b->bd_buddy_page)
1233 page_cache_release(e4b->bd_buddy_page);
1234 e4b->bd_buddy = NULL;
1235 e4b->bd_bitmap = NULL;
1239 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1241 if (e4b->bd_bitmap_page)
1242 page_cache_release(e4b->bd_bitmap_page);
1243 if (e4b->bd_buddy_page)
1244 page_cache_release(e4b->bd_buddy_page);
1248 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1253 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1254 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1257 while (order <= e4b->bd_blkbits + 1) {
1259 if (!mb_test_bit(block, bb)) {
1260 /* this block is part of buddy of order 'order' */
1263 bb += 1 << (e4b->bd_blkbits - order);
1269 static void mb_clear_bits(void *bm, int cur, int len)
1275 if ((cur & 31) == 0 && (len - cur) >= 32) {
1276 /* fast path: clear whole word at once */
1277 addr = bm + (cur >> 3);
1282 mb_clear_bit(cur, bm);
1287 /* clear bits in given range
1288 * will return first found zero bit if any, -1 otherwise
1290 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1297 if ((cur & 31) == 0 && (len - cur) >= 32) {
1298 /* fast path: clear whole word at once */
1299 addr = bm + (cur >> 3);
1300 if (*addr != (__u32)(-1) && zero_bit == -1)
1301 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1306 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1314 void ext4_set_bits(void *bm, int cur, int len)
1320 if ((cur & 31) == 0 && (len - cur) >= 32) {
1321 /* fast path: set whole word at once */
1322 addr = bm + (cur >> 3);
1327 mb_set_bit(cur, bm);
1333 * _________________________________________________________________ */
1335 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1337 if (mb_test_bit(*bit + side, bitmap)) {
1338 mb_clear_bit(*bit, bitmap);
1344 mb_set_bit(*bit, bitmap);
1349 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1353 void *buddy = mb_find_buddy(e4b, order, &max);
1358 /* Bits in range [first; last] are known to be set since
1359 * corresponding blocks were allocated. Bits in range
1360 * (first; last) will stay set because they form buddies on
1361 * upper layer. We just deal with borders if they don't
1362 * align with upper layer and then go up.
1363 * Releasing entire group is all about clearing
1364 * single bit of highest order buddy.
1368 * ---------------------------------
1370 * ---------------------------------
1371 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1372 * ---------------------------------
1374 * \_____________________/
1376 * Neither [1] nor [6] is aligned to above layer.
1377 * Left neighbour [0] is free, so mark it busy,
1378 * decrease bb_counters and extend range to
1380 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1381 * mark [6] free, increase bb_counters and shrink range to
1383 * Then shift range to [0; 2], go up and do the same.
1388 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1390 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1395 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1396 mb_clear_bits(buddy, first, last - first + 1);
1397 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1406 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1407 int first, int count)
1409 int left_is_free = 0;
1410 int right_is_free = 0;
1412 int last = first + count - 1;
1413 struct super_block *sb = e4b->bd_sb;
1415 BUG_ON(last >= (sb->s_blocksize << 3));
1416 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1417 /* Don't bother if the block group is corrupt. */
1418 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1421 mb_check_buddy(e4b);
1422 mb_free_blocks_double(inode, e4b, first, count);
1424 e4b->bd_info->bb_free += count;
1425 if (first < e4b->bd_info->bb_first_free)
1426 e4b->bd_info->bb_first_free = first;
1428 /* access memory sequentially: check left neighbour,
1429 * clear range and then check right neighbour
1432 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1433 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1434 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1435 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1437 if (unlikely(block != -1)) {
1438 struct ext4_sb_info *sbi = EXT4_SB(sb);
1439 ext4_fsblk_t blocknr;
1441 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1442 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1443 ext4_grp_locked_error(sb, e4b->bd_group,
1444 inode ? inode->i_ino : 0,
1446 "freeing already freed block "
1447 "(bit %u); block bitmap corrupt.",
1449 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1450 percpu_counter_sub(&sbi->s_freeclusters_counter,
1451 e4b->bd_info->bb_free);
1452 /* Mark the block group as corrupt. */
1453 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1454 &e4b->bd_info->bb_state);
1455 mb_regenerate_buddy(e4b);
1459 /* let's maintain fragments counter */
1460 if (left_is_free && right_is_free)
1461 e4b->bd_info->bb_fragments--;
1462 else if (!left_is_free && !right_is_free)
1463 e4b->bd_info->bb_fragments++;
1465 /* buddy[0] == bd_bitmap is a special case, so handle
1466 * it right away and let mb_buddy_mark_free stay free of
1467 * zero order checks.
1468 * Check if neighbours are to be coaleasced,
1469 * adjust bitmap bb_counters and borders appropriately.
1472 first += !left_is_free;
1473 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1476 last -= !right_is_free;
1477 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1481 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1484 mb_set_largest_free_order(sb, e4b->bd_info);
1485 mb_check_buddy(e4b);
1488 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1489 int needed, struct ext4_free_extent *ex)
1495 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1498 buddy = mb_find_buddy(e4b, 0, &max);
1499 BUG_ON(buddy == NULL);
1500 BUG_ON(block >= max);
1501 if (mb_test_bit(block, buddy)) {
1508 /* find actual order */
1509 order = mb_find_order_for_block(e4b, block);
1510 block = block >> order;
1512 ex->fe_len = 1 << order;
1513 ex->fe_start = block << order;
1514 ex->fe_group = e4b->bd_group;
1516 /* calc difference from given start */
1517 next = next - ex->fe_start;
1519 ex->fe_start += next;
1521 while (needed > ex->fe_len &&
1522 mb_find_buddy(e4b, order, &max)) {
1524 if (block + 1 >= max)
1527 next = (block + 1) * (1 << order);
1528 if (mb_test_bit(next, e4b->bd_bitmap))
1531 order = mb_find_order_for_block(e4b, next);
1533 block = next >> order;
1534 ex->fe_len += 1 << order;
1537 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1541 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1547 int start = ex->fe_start;
1548 int len = ex->fe_len;
1553 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1554 BUG_ON(e4b->bd_group != ex->fe_group);
1555 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1556 mb_check_buddy(e4b);
1557 mb_mark_used_double(e4b, start, len);
1559 e4b->bd_info->bb_free -= len;
1560 if (e4b->bd_info->bb_first_free == start)
1561 e4b->bd_info->bb_first_free += len;
1563 /* let's maintain fragments counter */
1565 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1566 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1567 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1569 e4b->bd_info->bb_fragments++;
1570 else if (!mlen && !max)
1571 e4b->bd_info->bb_fragments--;
1573 /* let's maintain buddy itself */
1575 ord = mb_find_order_for_block(e4b, start);
1577 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1578 /* the whole chunk may be allocated at once! */
1580 buddy = mb_find_buddy(e4b, ord, &max);
1581 BUG_ON((start >> ord) >= max);
1582 mb_set_bit(start >> ord, buddy);
1583 e4b->bd_info->bb_counters[ord]--;
1590 /* store for history */
1592 ret = len | (ord << 16);
1594 /* we have to split large buddy */
1596 buddy = mb_find_buddy(e4b, ord, &max);
1597 mb_set_bit(start >> ord, buddy);
1598 e4b->bd_info->bb_counters[ord]--;
1601 cur = (start >> ord) & ~1U;
1602 buddy = mb_find_buddy(e4b, ord, &max);
1603 mb_clear_bit(cur, buddy);
1604 mb_clear_bit(cur + 1, buddy);
1605 e4b->bd_info->bb_counters[ord]++;
1606 e4b->bd_info->bb_counters[ord]++;
1608 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1610 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1611 mb_check_buddy(e4b);
1617 * Must be called under group lock!
1619 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1620 struct ext4_buddy *e4b)
1622 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1625 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1626 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1628 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1629 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1630 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1632 /* preallocation can change ac_b_ex, thus we store actually
1633 * allocated blocks for history */
1634 ac->ac_f_ex = ac->ac_b_ex;
1636 ac->ac_status = AC_STATUS_FOUND;
1637 ac->ac_tail = ret & 0xffff;
1638 ac->ac_buddy = ret >> 16;
1641 * take the page reference. We want the page to be pinned
1642 * so that we don't get a ext4_mb_init_cache_call for this
1643 * group until we update the bitmap. That would mean we
1644 * double allocate blocks. The reference is dropped
1645 * in ext4_mb_release_context
1647 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1648 get_page(ac->ac_bitmap_page);
1649 ac->ac_buddy_page = e4b->bd_buddy_page;
1650 get_page(ac->ac_buddy_page);
1651 /* store last allocated for subsequent stream allocation */
1652 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1653 spin_lock(&sbi->s_md_lock);
1654 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1655 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1656 spin_unlock(&sbi->s_md_lock);
1661 * regular allocator, for general purposes allocation
1664 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1665 struct ext4_buddy *e4b,
1668 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1669 struct ext4_free_extent *bex = &ac->ac_b_ex;
1670 struct ext4_free_extent *gex = &ac->ac_g_ex;
1671 struct ext4_free_extent ex;
1674 if (ac->ac_status == AC_STATUS_FOUND)
1677 * We don't want to scan for a whole year
1679 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1680 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1681 ac->ac_status = AC_STATUS_BREAK;
1686 * Haven't found good chunk so far, let's continue
1688 if (bex->fe_len < gex->fe_len)
1691 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1692 && bex->fe_group == e4b->bd_group) {
1693 /* recheck chunk's availability - we don't know
1694 * when it was found (within this lock-unlock
1696 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1697 if (max >= gex->fe_len) {
1698 ext4_mb_use_best_found(ac, e4b);
1705 * The routine checks whether found extent is good enough. If it is,
1706 * then the extent gets marked used and flag is set to the context
1707 * to stop scanning. Otherwise, the extent is compared with the
1708 * previous found extent and if new one is better, then it's stored
1709 * in the context. Later, the best found extent will be used, if
1710 * mballoc can't find good enough extent.
1712 * FIXME: real allocation policy is to be designed yet!
1714 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1715 struct ext4_free_extent *ex,
1716 struct ext4_buddy *e4b)
1718 struct ext4_free_extent *bex = &ac->ac_b_ex;
1719 struct ext4_free_extent *gex = &ac->ac_g_ex;
1721 BUG_ON(ex->fe_len <= 0);
1722 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1723 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1724 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1729 * The special case - take what you catch first
1731 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1733 ext4_mb_use_best_found(ac, e4b);
1738 * Let's check whether the chuck is good enough
1740 if (ex->fe_len == gex->fe_len) {
1742 ext4_mb_use_best_found(ac, e4b);
1747 * If this is first found extent, just store it in the context
1749 if (bex->fe_len == 0) {
1755 * If new found extent is better, store it in the context
1757 if (bex->fe_len < gex->fe_len) {
1758 /* if the request isn't satisfied, any found extent
1759 * larger than previous best one is better */
1760 if (ex->fe_len > bex->fe_len)
1762 } else if (ex->fe_len > gex->fe_len) {
1763 /* if the request is satisfied, then we try to find
1764 * an extent that still satisfy the request, but is
1765 * smaller than previous one */
1766 if (ex->fe_len < bex->fe_len)
1770 ext4_mb_check_limits(ac, e4b, 0);
1773 static noinline_for_stack
1774 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1775 struct ext4_buddy *e4b)
1777 struct ext4_free_extent ex = ac->ac_b_ex;
1778 ext4_group_t group = ex.fe_group;
1782 BUG_ON(ex.fe_len <= 0);
1783 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1787 ext4_lock_group(ac->ac_sb, group);
1788 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1792 ext4_mb_use_best_found(ac, e4b);
1795 ext4_unlock_group(ac->ac_sb, group);
1796 ext4_mb_unload_buddy(e4b);
1801 static noinline_for_stack
1802 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1803 struct ext4_buddy *e4b)
1805 ext4_group_t group = ac->ac_g_ex.fe_group;
1808 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1809 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1810 struct ext4_free_extent ex;
1812 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1814 if (grp->bb_free == 0)
1817 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1821 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1822 ext4_mb_unload_buddy(e4b);
1826 ext4_lock_group(ac->ac_sb, group);
1827 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1828 ac->ac_g_ex.fe_len, &ex);
1829 ex.fe_logical = 0xDEADFA11; /* debug value */
1831 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1834 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1836 /* use do_div to get remainder (would be 64-bit modulo) */
1837 if (do_div(start, sbi->s_stripe) == 0) {
1840 ext4_mb_use_best_found(ac, e4b);
1842 } else if (max >= ac->ac_g_ex.fe_len) {
1843 BUG_ON(ex.fe_len <= 0);
1844 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1845 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1848 ext4_mb_use_best_found(ac, e4b);
1849 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1850 /* Sometimes, caller may want to merge even small
1851 * number of blocks to an existing extent */
1852 BUG_ON(ex.fe_len <= 0);
1853 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1854 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1857 ext4_mb_use_best_found(ac, e4b);
1859 ext4_unlock_group(ac->ac_sb, group);
1860 ext4_mb_unload_buddy(e4b);
1866 * The routine scans buddy structures (not bitmap!) from given order
1867 * to max order and tries to find big enough chunk to satisfy the req
1869 static noinline_for_stack
1870 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1871 struct ext4_buddy *e4b)
1873 struct super_block *sb = ac->ac_sb;
1874 struct ext4_group_info *grp = e4b->bd_info;
1880 BUG_ON(ac->ac_2order <= 0);
1881 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1882 if (grp->bb_counters[i] == 0)
1885 buddy = mb_find_buddy(e4b, i, &max);
1886 BUG_ON(buddy == NULL);
1888 k = mb_find_next_zero_bit(buddy, max, 0);
1893 ac->ac_b_ex.fe_len = 1 << i;
1894 ac->ac_b_ex.fe_start = k << i;
1895 ac->ac_b_ex.fe_group = e4b->bd_group;
1897 ext4_mb_use_best_found(ac, e4b);
1899 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1901 if (EXT4_SB(sb)->s_mb_stats)
1902 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1909 * The routine scans the group and measures all found extents.
1910 * In order to optimize scanning, caller must pass number of
1911 * free blocks in the group, so the routine can know upper limit.
1913 static noinline_for_stack
1914 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1915 struct ext4_buddy *e4b)
1917 struct super_block *sb = ac->ac_sb;
1918 void *bitmap = e4b->bd_bitmap;
1919 struct ext4_free_extent ex;
1923 free = e4b->bd_info->bb_free;
1926 i = e4b->bd_info->bb_first_free;
1928 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1929 i = mb_find_next_zero_bit(bitmap,
1930 EXT4_CLUSTERS_PER_GROUP(sb), i);
1931 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1933 * IF we have corrupt bitmap, we won't find any
1934 * free blocks even though group info says we
1935 * we have free blocks
1937 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1938 "%d free clusters as per "
1939 "group info. But bitmap says 0",
1944 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1945 BUG_ON(ex.fe_len <= 0);
1946 if (free < ex.fe_len) {
1947 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1948 "%d free clusters as per "
1949 "group info. But got %d blocks",
1952 * The number of free blocks differs. This mostly
1953 * indicate that the bitmap is corrupt. So exit
1954 * without claiming the space.
1958 ex.fe_logical = 0xDEADC0DE; /* debug value */
1959 ext4_mb_measure_extent(ac, &ex, e4b);
1965 ext4_mb_check_limits(ac, e4b, 1);
1969 * This is a special case for storages like raid5
1970 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1972 static noinline_for_stack
1973 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1974 struct ext4_buddy *e4b)
1976 struct super_block *sb = ac->ac_sb;
1977 struct ext4_sb_info *sbi = EXT4_SB(sb);
1978 void *bitmap = e4b->bd_bitmap;
1979 struct ext4_free_extent ex;
1980 ext4_fsblk_t first_group_block;
1985 BUG_ON(sbi->s_stripe == 0);
1987 /* find first stripe-aligned block in group */
1988 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1990 a = first_group_block + sbi->s_stripe - 1;
1991 do_div(a, sbi->s_stripe);
1992 i = (a * sbi->s_stripe) - first_group_block;
1994 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1995 if (!mb_test_bit(i, bitmap)) {
1996 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
1997 if (max >= sbi->s_stripe) {
1999 ex.fe_logical = 0xDEADF00D; /* debug value */
2001 ext4_mb_use_best_found(ac, e4b);
2009 /* This is now called BEFORE we load the buddy bitmap. */
2010 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2011 ext4_group_t group, int cr)
2013 unsigned free, fragments;
2014 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2015 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2017 BUG_ON(cr < 0 || cr >= 4);
2019 free = grp->bb_free;
2022 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2025 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2028 /* We only do this if the grp has never been initialized */
2029 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2030 int ret = ext4_mb_init_group(ac->ac_sb, group);
2035 fragments = grp->bb_fragments;
2041 BUG_ON(ac->ac_2order == 0);
2043 /* Avoid using the first bg of a flexgroup for data files */
2044 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2045 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2046 ((group % flex_size) == 0))
2049 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2050 (free / fragments) >= ac->ac_g_ex.fe_len)
2053 if (grp->bb_largest_free_order < ac->ac_2order)
2058 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2062 if (free >= ac->ac_g_ex.fe_len)
2074 static noinline_for_stack int
2075 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2077 ext4_group_t ngroups, group, i;
2080 struct ext4_sb_info *sbi;
2081 struct super_block *sb;
2082 struct ext4_buddy e4b;
2086 ngroups = ext4_get_groups_count(sb);
2087 /* non-extent files are limited to low blocks/groups */
2088 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2089 ngroups = sbi->s_blockfile_groups;
2091 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2093 /* first, try the goal */
2094 err = ext4_mb_find_by_goal(ac, &e4b);
2095 if (err || ac->ac_status == AC_STATUS_FOUND)
2098 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2102 * ac->ac2_order is set only if the fe_len is a power of 2
2103 * if ac2_order is set we also set criteria to 0 so that we
2104 * try exact allocation using buddy.
2106 i = fls(ac->ac_g_ex.fe_len);
2109 * We search using buddy data only if the order of the request
2110 * is greater than equal to the sbi_s_mb_order2_reqs
2111 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2113 if (i >= sbi->s_mb_order2_reqs) {
2115 * This should tell if fe_len is exactly power of 2
2117 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2118 ac->ac_2order = i - 1;
2121 /* if stream allocation is enabled, use global goal */
2122 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2123 /* TBD: may be hot point */
2124 spin_lock(&sbi->s_md_lock);
2125 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2126 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2127 spin_unlock(&sbi->s_md_lock);
2130 /* Let's just scan groups to find more-less suitable blocks */
2131 cr = ac->ac_2order ? 0 : 1;
2133 * cr == 0 try to get exact allocation,
2134 * cr == 3 try to get anything
2137 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2138 ac->ac_criteria = cr;
2140 * searching for the right group start
2141 * from the goal value specified
2143 group = ac->ac_g_ex.fe_group;
2145 for (i = 0; i < ngroups; group++, i++) {
2148 * Artificially restricted ngroups for non-extent
2149 * files makes group > ngroups possible on first loop.
2151 if (group >= ngroups)
2154 /* This now checks without needing the buddy page */
2155 if (!ext4_mb_good_group(ac, group, cr))
2158 err = ext4_mb_load_buddy(sb, group, &e4b);
2162 ext4_lock_group(sb, group);
2165 * We need to check again after locking the
2168 if (!ext4_mb_good_group(ac, group, cr)) {
2169 ext4_unlock_group(sb, group);
2170 ext4_mb_unload_buddy(&e4b);
2174 ac->ac_groups_scanned++;
2175 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2176 ext4_mb_simple_scan_group(ac, &e4b);
2177 else if (cr == 1 && sbi->s_stripe &&
2178 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2179 ext4_mb_scan_aligned(ac, &e4b);
2181 ext4_mb_complex_scan_group(ac, &e4b);
2183 ext4_unlock_group(sb, group);
2184 ext4_mb_unload_buddy(&e4b);
2186 if (ac->ac_status != AC_STATUS_CONTINUE)
2191 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2192 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2194 * We've been searching too long. Let's try to allocate
2195 * the best chunk we've found so far
2198 ext4_mb_try_best_found(ac, &e4b);
2199 if (ac->ac_status != AC_STATUS_FOUND) {
2201 * Someone more lucky has already allocated it.
2202 * The only thing we can do is just take first
2204 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2206 ac->ac_b_ex.fe_group = 0;
2207 ac->ac_b_ex.fe_start = 0;
2208 ac->ac_b_ex.fe_len = 0;
2209 ac->ac_status = AC_STATUS_CONTINUE;
2210 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2212 atomic_inc(&sbi->s_mb_lost_chunks);
2220 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2222 struct super_block *sb = seq->private;
2225 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2228 return (void *) ((unsigned long) group);
2231 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2233 struct super_block *sb = seq->private;
2237 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2240 return (void *) ((unsigned long) group);
2243 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2245 struct super_block *sb = seq->private;
2246 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2248 int err, buddy_loaded = 0;
2249 struct ext4_buddy e4b;
2250 struct ext4_group_info *grinfo;
2252 struct ext4_group_info info;
2253 ext4_grpblk_t counters[16];
2258 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2259 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2260 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2261 "group", "free", "frags", "first",
2262 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2263 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2265 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2266 sizeof(struct ext4_group_info);
2267 grinfo = ext4_get_group_info(sb, group);
2268 /* Load the group info in memory only if not already loaded. */
2269 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2270 err = ext4_mb_load_buddy(sb, group, &e4b);
2272 seq_printf(seq, "#%-5u: I/O error\n", group);
2278 memcpy(&sg, ext4_get_group_info(sb, group), i);
2281 ext4_mb_unload_buddy(&e4b);
2283 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2284 sg.info.bb_fragments, sg.info.bb_first_free);
2285 for (i = 0; i <= 13; i++)
2286 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2287 sg.info.bb_counters[i] : 0);
2288 seq_printf(seq, " ]\n");
2293 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2297 static const struct seq_operations ext4_mb_seq_groups_ops = {
2298 .start = ext4_mb_seq_groups_start,
2299 .next = ext4_mb_seq_groups_next,
2300 .stop = ext4_mb_seq_groups_stop,
2301 .show = ext4_mb_seq_groups_show,
2304 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2306 struct super_block *sb = PDE_DATA(inode);
2309 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2311 struct seq_file *m = file->private_data;
2318 static const struct file_operations ext4_mb_seq_groups_fops = {
2319 .owner = THIS_MODULE,
2320 .open = ext4_mb_seq_groups_open,
2322 .llseek = seq_lseek,
2323 .release = seq_release,
2326 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2328 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2329 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2336 * Allocate the top-level s_group_info array for the specified number
2339 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2341 struct ext4_sb_info *sbi = EXT4_SB(sb);
2343 struct ext4_group_info ***new_groupinfo;
2345 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2346 EXT4_DESC_PER_BLOCK_BITS(sb);
2347 if (size <= sbi->s_group_info_size)
2350 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2351 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2352 if (!new_groupinfo) {
2353 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2356 if (sbi->s_group_info) {
2357 memcpy(new_groupinfo, sbi->s_group_info,
2358 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2359 ext4_kvfree(sbi->s_group_info);
2361 sbi->s_group_info = new_groupinfo;
2362 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2363 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2364 sbi->s_group_info_size);
2368 /* Create and initialize ext4_group_info data for the given group. */
2369 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2370 struct ext4_group_desc *desc)
2374 struct ext4_sb_info *sbi = EXT4_SB(sb);
2375 struct ext4_group_info **meta_group_info;
2376 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2379 * First check if this group is the first of a reserved block.
2380 * If it's true, we have to allocate a new table of pointers
2381 * to ext4_group_info structures
2383 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2384 metalen = sizeof(*meta_group_info) <<
2385 EXT4_DESC_PER_BLOCK_BITS(sb);
2386 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2387 if (meta_group_info == NULL) {
2388 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2389 "for a buddy group");
2390 goto exit_meta_group_info;
2392 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2397 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2398 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2400 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
2401 if (meta_group_info[i] == NULL) {
2402 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2403 goto exit_group_info;
2405 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2406 &(meta_group_info[i]->bb_state));
2409 * initialize bb_free to be able to skip
2410 * empty groups without initialization
2412 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2413 meta_group_info[i]->bb_free =
2414 ext4_free_clusters_after_init(sb, group, desc);
2416 meta_group_info[i]->bb_free =
2417 ext4_free_group_clusters(sb, desc);
2420 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2421 init_rwsem(&meta_group_info[i]->alloc_sem);
2422 meta_group_info[i]->bb_free_root = RB_ROOT;
2423 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2427 struct buffer_head *bh;
2428 meta_group_info[i]->bb_bitmap =
2429 kmalloc(sb->s_blocksize, GFP_KERNEL);
2430 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2431 bh = ext4_read_block_bitmap(sb, group);
2433 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2442 /* If a meta_group_info table has been allocated, release it now */
2443 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2444 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2445 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2447 exit_meta_group_info:
2449 } /* ext4_mb_add_groupinfo */
2451 static int ext4_mb_init_backend(struct super_block *sb)
2453 ext4_group_t ngroups = ext4_get_groups_count(sb);
2455 struct ext4_sb_info *sbi = EXT4_SB(sb);
2457 struct ext4_group_desc *desc;
2458 struct kmem_cache *cachep;
2460 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2464 sbi->s_buddy_cache = new_inode(sb);
2465 if (sbi->s_buddy_cache == NULL) {
2466 ext4_msg(sb, KERN_ERR, "can't get new inode");
2469 /* To avoid potentially colliding with an valid on-disk inode number,
2470 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2471 * not in the inode hash, so it should never be found by iget(), but
2472 * this will avoid confusion if it ever shows up during debugging. */
2473 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2474 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2475 for (i = 0; i < ngroups; i++) {
2476 desc = ext4_get_group_desc(sb, i, NULL);
2478 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2481 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2488 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2490 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2491 i = sbi->s_group_info_size;
2493 kfree(sbi->s_group_info[i]);
2494 iput(sbi->s_buddy_cache);
2496 ext4_kvfree(sbi->s_group_info);
2500 static void ext4_groupinfo_destroy_slabs(void)
2504 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2505 if (ext4_groupinfo_caches[i])
2506 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2507 ext4_groupinfo_caches[i] = NULL;
2511 static int ext4_groupinfo_create_slab(size_t size)
2513 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2515 int blocksize_bits = order_base_2(size);
2516 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2517 struct kmem_cache *cachep;
2519 if (cache_index >= NR_GRPINFO_CACHES)
2522 if (unlikely(cache_index < 0))
2525 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2526 if (ext4_groupinfo_caches[cache_index]) {
2527 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2528 return 0; /* Already created */
2531 slab_size = offsetof(struct ext4_group_info,
2532 bb_counters[blocksize_bits + 2]);
2534 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2535 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2538 ext4_groupinfo_caches[cache_index] = cachep;
2540 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2543 "EXT4-fs: no memory for groupinfo slab cache\n");
2550 int ext4_mb_init(struct super_block *sb)
2552 struct ext4_sb_info *sbi = EXT4_SB(sb);
2558 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2560 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2561 if (sbi->s_mb_offsets == NULL) {
2566 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2567 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2568 if (sbi->s_mb_maxs == NULL) {
2573 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2577 /* order 0 is regular bitmap */
2578 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2579 sbi->s_mb_offsets[0] = 0;
2583 max = sb->s_blocksize << 2;
2585 sbi->s_mb_offsets[i] = offset;
2586 sbi->s_mb_maxs[i] = max;
2587 offset += 1 << (sb->s_blocksize_bits - i);
2590 } while (i <= sb->s_blocksize_bits + 1);
2592 spin_lock_init(&sbi->s_md_lock);
2593 spin_lock_init(&sbi->s_bal_lock);
2595 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2596 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2597 sbi->s_mb_stats = MB_DEFAULT_STATS;
2598 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2599 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2601 * The default group preallocation is 512, which for 4k block
2602 * sizes translates to 2 megabytes. However for bigalloc file
2603 * systems, this is probably too big (i.e, if the cluster size
2604 * is 1 megabyte, then group preallocation size becomes half a
2605 * gigabyte!). As a default, we will keep a two megabyte
2606 * group pralloc size for cluster sizes up to 64k, and after
2607 * that, we will force a minimum group preallocation size of
2608 * 32 clusters. This translates to 8 megs when the cluster
2609 * size is 256k, and 32 megs when the cluster size is 1 meg,
2610 * which seems reasonable as a default.
2612 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2613 sbi->s_cluster_bits, 32);
2615 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2616 * to the lowest multiple of s_stripe which is bigger than
2617 * the s_mb_group_prealloc as determined above. We want
2618 * the preallocation size to be an exact multiple of the
2619 * RAID stripe size so that preallocations don't fragment
2622 if (sbi->s_stripe > 1) {
2623 sbi->s_mb_group_prealloc = roundup(
2624 sbi->s_mb_group_prealloc, sbi->s_stripe);
2627 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2628 if (sbi->s_locality_groups == NULL) {
2632 for_each_possible_cpu(i) {
2633 struct ext4_locality_group *lg;
2634 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2635 mutex_init(&lg->lg_mutex);
2636 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2637 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2638 spin_lock_init(&lg->lg_prealloc_lock);
2641 /* init file for buddy data */
2642 ret = ext4_mb_init_backend(sb);
2644 goto out_free_locality_groups;
2647 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2648 &ext4_mb_seq_groups_fops, sb);
2652 out_free_locality_groups:
2653 free_percpu(sbi->s_locality_groups);
2654 sbi->s_locality_groups = NULL;
2656 kfree(sbi->s_mb_offsets);
2657 sbi->s_mb_offsets = NULL;
2658 kfree(sbi->s_mb_maxs);
2659 sbi->s_mb_maxs = NULL;
2663 /* need to called with the ext4 group lock held */
2664 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2666 struct ext4_prealloc_space *pa;
2667 struct list_head *cur, *tmp;
2670 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2671 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2672 list_del(&pa->pa_group_list);
2674 kmem_cache_free(ext4_pspace_cachep, pa);
2677 mb_debug(1, "mballoc: %u PAs left\n", count);
2681 int ext4_mb_release(struct super_block *sb)
2683 ext4_group_t ngroups = ext4_get_groups_count(sb);
2685 int num_meta_group_infos;
2686 struct ext4_group_info *grinfo;
2687 struct ext4_sb_info *sbi = EXT4_SB(sb);
2688 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2691 remove_proc_entry("mb_groups", sbi->s_proc);
2693 if (sbi->s_group_info) {
2694 for (i = 0; i < ngroups; i++) {
2695 grinfo = ext4_get_group_info(sb, i);
2697 kfree(grinfo->bb_bitmap);
2699 ext4_lock_group(sb, i);
2700 ext4_mb_cleanup_pa(grinfo);
2701 ext4_unlock_group(sb, i);
2702 kmem_cache_free(cachep, grinfo);
2704 num_meta_group_infos = (ngroups +
2705 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2706 EXT4_DESC_PER_BLOCK_BITS(sb);
2707 for (i = 0; i < num_meta_group_infos; i++)
2708 kfree(sbi->s_group_info[i]);
2709 ext4_kvfree(sbi->s_group_info);
2711 kfree(sbi->s_mb_offsets);
2712 kfree(sbi->s_mb_maxs);
2713 if (sbi->s_buddy_cache)
2714 iput(sbi->s_buddy_cache);
2715 if (sbi->s_mb_stats) {
2716 ext4_msg(sb, KERN_INFO,
2717 "mballoc: %u blocks %u reqs (%u success)",
2718 atomic_read(&sbi->s_bal_allocated),
2719 atomic_read(&sbi->s_bal_reqs),
2720 atomic_read(&sbi->s_bal_success));
2721 ext4_msg(sb, KERN_INFO,
2722 "mballoc: %u extents scanned, %u goal hits, "
2723 "%u 2^N hits, %u breaks, %u lost",
2724 atomic_read(&sbi->s_bal_ex_scanned),
2725 atomic_read(&sbi->s_bal_goals),
2726 atomic_read(&sbi->s_bal_2orders),
2727 atomic_read(&sbi->s_bal_breaks),
2728 atomic_read(&sbi->s_mb_lost_chunks));
2729 ext4_msg(sb, KERN_INFO,
2730 "mballoc: %lu generated and it took %Lu",
2731 sbi->s_mb_buddies_generated,
2732 sbi->s_mb_generation_time);
2733 ext4_msg(sb, KERN_INFO,
2734 "mballoc: %u preallocated, %u discarded",
2735 atomic_read(&sbi->s_mb_preallocated),
2736 atomic_read(&sbi->s_mb_discarded));
2739 free_percpu(sbi->s_locality_groups);
2744 static inline int ext4_issue_discard(struct super_block *sb,
2745 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2747 ext4_fsblk_t discard_block;
2749 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2750 ext4_group_first_block_no(sb, block_group));
2751 count = EXT4_C2B(EXT4_SB(sb), count);
2752 trace_ext4_discard_blocks(sb,
2753 (unsigned long long) discard_block, count);
2754 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2758 * This function is called by the jbd2 layer once the commit has finished,
2759 * so we know we can free the blocks that were released with that commit.
2761 static void ext4_free_data_callback(struct super_block *sb,
2762 struct ext4_journal_cb_entry *jce,
2765 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2766 struct ext4_buddy e4b;
2767 struct ext4_group_info *db;
2768 int err, count = 0, count2 = 0;
2770 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2771 entry->efd_count, entry->efd_group, entry);
2773 if (test_opt(sb, DISCARD)) {
2774 err = ext4_issue_discard(sb, entry->efd_group,
2775 entry->efd_start_cluster,
2777 if (err && err != -EOPNOTSUPP)
2778 ext4_msg(sb, KERN_WARNING, "discard request in"
2779 " group:%d block:%d count:%d failed"
2780 " with %d", entry->efd_group,
2781 entry->efd_start_cluster,
2782 entry->efd_count, err);
2785 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2786 /* we expect to find existing buddy because it's pinned */
2791 /* there are blocks to put in buddy to make them really free */
2792 count += entry->efd_count;
2794 ext4_lock_group(sb, entry->efd_group);
2795 /* Take it out of per group rb tree */
2796 rb_erase(&entry->efd_node, &(db->bb_free_root));
2797 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2800 * Clear the trimmed flag for the group so that the next
2801 * ext4_trim_fs can trim it.
2802 * If the volume is mounted with -o discard, online discard
2803 * is supported and the free blocks will be trimmed online.
2805 if (!test_opt(sb, DISCARD))
2806 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2808 if (!db->bb_free_root.rb_node) {
2809 /* No more items in the per group rb tree
2810 * balance refcounts from ext4_mb_free_metadata()
2812 page_cache_release(e4b.bd_buddy_page);
2813 page_cache_release(e4b.bd_bitmap_page);
2815 ext4_unlock_group(sb, entry->efd_group);
2816 kmem_cache_free(ext4_free_data_cachep, entry);
2817 ext4_mb_unload_buddy(&e4b);
2819 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2822 int __init ext4_init_mballoc(void)
2824 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2825 SLAB_RECLAIM_ACCOUNT);
2826 if (ext4_pspace_cachep == NULL)
2829 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2830 SLAB_RECLAIM_ACCOUNT);
2831 if (ext4_ac_cachep == NULL) {
2832 kmem_cache_destroy(ext4_pspace_cachep);
2836 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2837 SLAB_RECLAIM_ACCOUNT);
2838 if (ext4_free_data_cachep == NULL) {
2839 kmem_cache_destroy(ext4_pspace_cachep);
2840 kmem_cache_destroy(ext4_ac_cachep);
2846 void ext4_exit_mballoc(void)
2849 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2850 * before destroying the slab cache.
2853 kmem_cache_destroy(ext4_pspace_cachep);
2854 kmem_cache_destroy(ext4_ac_cachep);
2855 kmem_cache_destroy(ext4_free_data_cachep);
2856 ext4_groupinfo_destroy_slabs();
2861 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2862 * Returns 0 if success or error code
2864 static noinline_for_stack int
2865 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2866 handle_t *handle, unsigned int reserv_clstrs)
2868 struct buffer_head *bitmap_bh = NULL;
2869 struct ext4_group_desc *gdp;
2870 struct buffer_head *gdp_bh;
2871 struct ext4_sb_info *sbi;
2872 struct super_block *sb;
2876 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2877 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2883 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2887 BUFFER_TRACE(bitmap_bh, "getting write access");
2888 err = ext4_journal_get_write_access(handle, bitmap_bh);
2893 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2897 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2898 ext4_free_group_clusters(sb, gdp));
2900 BUFFER_TRACE(gdp_bh, "get_write_access");
2901 err = ext4_journal_get_write_access(handle, gdp_bh);
2905 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2907 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2908 if (!ext4_data_block_valid(sbi, block, len)) {
2909 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2910 "fs metadata", block, block+len);
2911 /* File system mounted not to panic on error
2912 * Fix the bitmap and repeat the block allocation
2913 * We leak some of the blocks here.
2915 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2916 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2917 ac->ac_b_ex.fe_len);
2918 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2919 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2925 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2926 #ifdef AGGRESSIVE_CHECK
2929 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2930 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2931 bitmap_bh->b_data));
2935 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2936 ac->ac_b_ex.fe_len);
2937 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2938 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2939 ext4_free_group_clusters_set(sb, gdp,
2940 ext4_free_clusters_after_init(sb,
2941 ac->ac_b_ex.fe_group, gdp));
2943 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2944 ext4_free_group_clusters_set(sb, gdp, len);
2945 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2946 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2948 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2949 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2951 * Now reduce the dirty block count also. Should not go negative
2953 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2954 /* release all the reserved blocks if non delalloc */
2955 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2958 if (sbi->s_log_groups_per_flex) {
2959 ext4_group_t flex_group = ext4_flex_group(sbi,
2960 ac->ac_b_ex.fe_group);
2961 atomic64_sub(ac->ac_b_ex.fe_len,
2962 &sbi->s_flex_groups[flex_group].free_clusters);
2965 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2968 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2976 * here we normalize request for locality group
2977 * Group request are normalized to s_mb_group_prealloc, which goes to
2978 * s_strip if we set the same via mount option.
2979 * s_mb_group_prealloc can be configured via
2980 * /sys/fs/ext4/<partition>/mb_group_prealloc
2982 * XXX: should we try to preallocate more than the group has now?
2984 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2986 struct super_block *sb = ac->ac_sb;
2987 struct ext4_locality_group *lg = ac->ac_lg;
2990 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2991 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2992 current->pid, ac->ac_g_ex.fe_len);
2996 * Normalization means making request better in terms of
2997 * size and alignment
2999 static noinline_for_stack void
3000 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3001 struct ext4_allocation_request *ar)
3003 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3006 loff_t size, start_off;
3007 loff_t orig_size __maybe_unused;
3009 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3010 struct ext4_prealloc_space *pa;
3012 /* do normalize only data requests, metadata requests
3013 do not need preallocation */
3014 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3017 /* sometime caller may want exact blocks */
3018 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3021 /* caller may indicate that preallocation isn't
3022 * required (it's a tail, for example) */
3023 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3026 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3027 ext4_mb_normalize_group_request(ac);
3031 bsbits = ac->ac_sb->s_blocksize_bits;
3033 /* first, let's learn actual file size
3034 * given current request is allocated */
3035 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3036 size = size << bsbits;
3037 if (size < i_size_read(ac->ac_inode))
3038 size = i_size_read(ac->ac_inode);
3041 /* max size of free chunks */
3044 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3045 (req <= (size) || max <= (chunk_size))
3047 /* first, try to predict filesize */
3048 /* XXX: should this table be tunable? */
3050 if (size <= 16 * 1024) {
3052 } else if (size <= 32 * 1024) {
3054 } else if (size <= 64 * 1024) {
3056 } else if (size <= 128 * 1024) {
3058 } else if (size <= 256 * 1024) {
3060 } else if (size <= 512 * 1024) {
3062 } else if (size <= 1024 * 1024) {
3064 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3065 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3066 (21 - bsbits)) << 21;
3067 size = 2 * 1024 * 1024;
3068 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3069 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3070 (22 - bsbits)) << 22;
3071 size = 4 * 1024 * 1024;
3072 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3073 (8<<20)>>bsbits, max, 8 * 1024)) {
3074 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3075 (23 - bsbits)) << 23;
3076 size = 8 * 1024 * 1024;
3078 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3079 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3080 ac->ac_o_ex.fe_len) << bsbits;
3082 size = size >> bsbits;
3083 start = start_off >> bsbits;
3085 /* don't cover already allocated blocks in selected range */
3086 if (ar->pleft && start <= ar->lleft) {
3087 size -= ar->lleft + 1 - start;
3088 start = ar->lleft + 1;
3090 if (ar->pright && start + size - 1 >= ar->lright)
3091 size -= start + size - ar->lright;
3095 /* check we don't cross already preallocated blocks */
3097 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3102 spin_lock(&pa->pa_lock);
3103 if (pa->pa_deleted) {
3104 spin_unlock(&pa->pa_lock);
3108 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3111 /* PA must not overlap original request */
3112 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3113 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3115 /* skip PAs this normalized request doesn't overlap with */
3116 if (pa->pa_lstart >= end || pa_end <= start) {
3117 spin_unlock(&pa->pa_lock);
3120 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3122 /* adjust start or end to be adjacent to this pa */
3123 if (pa_end <= ac->ac_o_ex.fe_logical) {
3124 BUG_ON(pa_end < start);
3126 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3127 BUG_ON(pa->pa_lstart > end);
3128 end = pa->pa_lstart;
3130 spin_unlock(&pa->pa_lock);
3135 /* XXX: extra loop to check we really don't overlap preallocations */
3137 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3140 spin_lock(&pa->pa_lock);
3141 if (pa->pa_deleted == 0) {
3142 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3144 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3146 spin_unlock(&pa->pa_lock);
3150 if (start + size <= ac->ac_o_ex.fe_logical &&
3151 start > ac->ac_o_ex.fe_logical) {
3152 ext4_msg(ac->ac_sb, KERN_ERR,
3153 "start %lu, size %lu, fe_logical %lu",
3154 (unsigned long) start, (unsigned long) size,
3155 (unsigned long) ac->ac_o_ex.fe_logical);
3157 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3158 start > ac->ac_o_ex.fe_logical);
3159 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3161 /* now prepare goal request */
3163 /* XXX: is it better to align blocks WRT to logical
3164 * placement or satisfy big request as is */
3165 ac->ac_g_ex.fe_logical = start;
3166 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3168 /* define goal start in order to merge */
3169 if (ar->pright && (ar->lright == (start + size))) {
3170 /* merge to the right */
3171 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3172 &ac->ac_f_ex.fe_group,
3173 &ac->ac_f_ex.fe_start);
3174 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3176 if (ar->pleft && (ar->lleft + 1 == start)) {
3177 /* merge to the left */
3178 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3179 &ac->ac_f_ex.fe_group,
3180 &ac->ac_f_ex.fe_start);
3181 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3184 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3185 (unsigned) orig_size, (unsigned) start);
3188 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3190 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3192 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3193 atomic_inc(&sbi->s_bal_reqs);
3194 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3195 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3196 atomic_inc(&sbi->s_bal_success);
3197 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3198 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3199 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3200 atomic_inc(&sbi->s_bal_goals);
3201 if (ac->ac_found > sbi->s_mb_max_to_scan)
3202 atomic_inc(&sbi->s_bal_breaks);
3205 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3206 trace_ext4_mballoc_alloc(ac);
3208 trace_ext4_mballoc_prealloc(ac);
3212 * Called on failure; free up any blocks from the inode PA for this
3213 * context. We don't need this for MB_GROUP_PA because we only change
3214 * pa_free in ext4_mb_release_context(), but on failure, we've already
3215 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3217 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3219 struct ext4_prealloc_space *pa = ac->ac_pa;
3220 struct ext4_buddy e4b;
3224 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3227 * This should never happen since we pin the
3228 * pages in the ext4_allocation_context so
3229 * ext4_mb_load_buddy() should never fail.
3231 WARN(1, "mb_load_buddy failed (%d)", err);
3234 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3235 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3236 ac->ac_f_ex.fe_len);
3237 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3240 if (pa->pa_type == MB_INODE_PA)
3241 pa->pa_free += ac->ac_b_ex.fe_len;
3245 * use blocks preallocated to inode
3247 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3248 struct ext4_prealloc_space *pa)
3250 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3255 /* found preallocated blocks, use them */
3256 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3257 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3258 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3259 len = EXT4_NUM_B2C(sbi, end - start);
3260 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3261 &ac->ac_b_ex.fe_start);
3262 ac->ac_b_ex.fe_len = len;
3263 ac->ac_status = AC_STATUS_FOUND;
3266 BUG_ON(start < pa->pa_pstart);
3267 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3268 BUG_ON(pa->pa_free < len);
3271 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3275 * use blocks preallocated to locality group
3277 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3278 struct ext4_prealloc_space *pa)
3280 unsigned int len = ac->ac_o_ex.fe_len;
3282 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3283 &ac->ac_b_ex.fe_group,
3284 &ac->ac_b_ex.fe_start);
3285 ac->ac_b_ex.fe_len = len;
3286 ac->ac_status = AC_STATUS_FOUND;
3289 /* we don't correct pa_pstart or pa_plen here to avoid
3290 * possible race when the group is being loaded concurrently
3291 * instead we correct pa later, after blocks are marked
3292 * in on-disk bitmap -- see ext4_mb_release_context()
3293 * Other CPUs are prevented from allocating from this pa by lg_mutex
3295 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3299 * Return the prealloc space that have minimal distance
3300 * from the goal block. @cpa is the prealloc
3301 * space that is having currently known minimal distance
3302 * from the goal block.
3304 static struct ext4_prealloc_space *
3305 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3306 struct ext4_prealloc_space *pa,
3307 struct ext4_prealloc_space *cpa)
3309 ext4_fsblk_t cur_distance, new_distance;
3312 atomic_inc(&pa->pa_count);
3315 cur_distance = abs(goal_block - cpa->pa_pstart);
3316 new_distance = abs(goal_block - pa->pa_pstart);
3318 if (cur_distance <= new_distance)
3321 /* drop the previous reference */
3322 atomic_dec(&cpa->pa_count);
3323 atomic_inc(&pa->pa_count);
3328 * search goal blocks in preallocated space
3330 static noinline_for_stack int
3331 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3333 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3335 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3336 struct ext4_locality_group *lg;
3337 struct ext4_prealloc_space *pa, *cpa = NULL;
3338 ext4_fsblk_t goal_block;
3340 /* only data can be preallocated */
3341 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3344 /* first, try per-file preallocation */
3346 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3348 /* all fields in this condition don't change,
3349 * so we can skip locking for them */
3350 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3351 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3352 EXT4_C2B(sbi, pa->pa_len)))
3355 /* non-extent files can't have physical blocks past 2^32 */
3356 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3357 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3358 EXT4_MAX_BLOCK_FILE_PHYS))
3361 /* found preallocated blocks, use them */
3362 spin_lock(&pa->pa_lock);
3363 if (pa->pa_deleted == 0 && pa->pa_free) {
3364 atomic_inc(&pa->pa_count);
3365 ext4_mb_use_inode_pa(ac, pa);
3366 spin_unlock(&pa->pa_lock);
3367 ac->ac_criteria = 10;
3371 spin_unlock(&pa->pa_lock);
3375 /* can we use group allocation? */
3376 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3379 /* inode may have no locality group for some reason */
3383 order = fls(ac->ac_o_ex.fe_len) - 1;
3384 if (order > PREALLOC_TB_SIZE - 1)
3385 /* The max size of hash table is PREALLOC_TB_SIZE */
3386 order = PREALLOC_TB_SIZE - 1;
3388 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3390 * search for the prealloc space that is having
3391 * minimal distance from the goal block.
3393 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3395 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3397 spin_lock(&pa->pa_lock);
3398 if (pa->pa_deleted == 0 &&
3399 pa->pa_free >= ac->ac_o_ex.fe_len) {
3401 cpa = ext4_mb_check_group_pa(goal_block,
3404 spin_unlock(&pa->pa_lock);
3409 ext4_mb_use_group_pa(ac, cpa);
3410 ac->ac_criteria = 20;
3417 * the function goes through all block freed in the group
3418 * but not yet committed and marks them used in in-core bitmap.
3419 * buddy must be generated from this bitmap
3420 * Need to be called with the ext4 group lock held
3422 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3426 struct ext4_group_info *grp;
3427 struct ext4_free_data *entry;
3429 grp = ext4_get_group_info(sb, group);
3430 n = rb_first(&(grp->bb_free_root));
3433 entry = rb_entry(n, struct ext4_free_data, efd_node);
3434 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3441 * the function goes through all preallocation in this group and marks them
3442 * used in in-core bitmap. buddy must be generated from this bitmap
3443 * Need to be called with ext4 group lock held
3445 static noinline_for_stack
3446 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3449 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3450 struct ext4_prealloc_space *pa;
3451 struct list_head *cur;
3452 ext4_group_t groupnr;
3453 ext4_grpblk_t start;
3454 int preallocated = 0;
3457 /* all form of preallocation discards first load group,
3458 * so the only competing code is preallocation use.
3459 * we don't need any locking here
3460 * notice we do NOT ignore preallocations with pa_deleted
3461 * otherwise we could leave used blocks available for
3462 * allocation in buddy when concurrent ext4_mb_put_pa()
3463 * is dropping preallocation
3465 list_for_each(cur, &grp->bb_prealloc_list) {
3466 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3467 spin_lock(&pa->pa_lock);
3468 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3471 spin_unlock(&pa->pa_lock);
3472 if (unlikely(len == 0))
3474 BUG_ON(groupnr != group);
3475 ext4_set_bits(bitmap, start, len);
3476 preallocated += len;
3478 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3481 static void ext4_mb_pa_callback(struct rcu_head *head)
3483 struct ext4_prealloc_space *pa;
3484 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3486 BUG_ON(atomic_read(&pa->pa_count));
3487 BUG_ON(pa->pa_deleted == 0);
3488 kmem_cache_free(ext4_pspace_cachep, pa);
3492 * drops a reference to preallocated space descriptor
3493 * if this was the last reference and the space is consumed
3495 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3496 struct super_block *sb, struct ext4_prealloc_space *pa)
3499 ext4_fsblk_t grp_blk;
3501 /* in this short window concurrent discard can set pa_deleted */
3502 spin_lock(&pa->pa_lock);
3503 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3504 spin_unlock(&pa->pa_lock);
3508 if (pa->pa_deleted == 1) {
3509 spin_unlock(&pa->pa_lock);
3514 spin_unlock(&pa->pa_lock);
3516 grp_blk = pa->pa_pstart;
3518 * If doing group-based preallocation, pa_pstart may be in the
3519 * next group when pa is used up
3521 if (pa->pa_type == MB_GROUP_PA)
3524 grp = ext4_get_group_number(sb, grp_blk);
3529 * P1 (buddy init) P2 (regular allocation)
3530 * find block B in PA
3531 * copy on-disk bitmap to buddy
3532 * mark B in on-disk bitmap
3533 * drop PA from group
3534 * mark all PAs in buddy
3536 * thus, P1 initializes buddy with B available. to prevent this
3537 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3540 ext4_lock_group(sb, grp);
3541 list_del(&pa->pa_group_list);
3542 ext4_unlock_group(sb, grp);
3544 spin_lock(pa->pa_obj_lock);
3545 list_del_rcu(&pa->pa_inode_list);
3546 spin_unlock(pa->pa_obj_lock);
3548 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3552 * creates new preallocated space for given inode
3554 static noinline_for_stack int
3555 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3557 struct super_block *sb = ac->ac_sb;
3558 struct ext4_sb_info *sbi = EXT4_SB(sb);
3559 struct ext4_prealloc_space *pa;
3560 struct ext4_group_info *grp;
3561 struct ext4_inode_info *ei;
3563 /* preallocate only when found space is larger then requested */
3564 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3565 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3566 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3568 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3572 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3578 /* we can't allocate as much as normalizer wants.
3579 * so, found space must get proper lstart
3580 * to cover original request */
3581 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3582 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3584 /* we're limited by original request in that
3585 * logical block must be covered any way
3586 * winl is window we can move our chunk within */
3587 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3589 /* also, we should cover whole original request */
3590 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3592 /* the smallest one defines real window */
3593 win = min(winl, wins);
3595 offs = ac->ac_o_ex.fe_logical %
3596 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3597 if (offs && offs < win)
3600 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3601 EXT4_NUM_B2C(sbi, win);
3602 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3603 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3606 /* preallocation can change ac_b_ex, thus we store actually
3607 * allocated blocks for history */
3608 ac->ac_f_ex = ac->ac_b_ex;
3610 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3611 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3612 pa->pa_len = ac->ac_b_ex.fe_len;
3613 pa->pa_free = pa->pa_len;
3614 atomic_set(&pa->pa_count, 1);
3615 spin_lock_init(&pa->pa_lock);
3616 INIT_LIST_HEAD(&pa->pa_inode_list);
3617 INIT_LIST_HEAD(&pa->pa_group_list);
3619 pa->pa_type = MB_INODE_PA;
3621 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3622 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3623 trace_ext4_mb_new_inode_pa(ac, pa);
3625 ext4_mb_use_inode_pa(ac, pa);
3626 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3628 ei = EXT4_I(ac->ac_inode);
3629 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3631 pa->pa_obj_lock = &ei->i_prealloc_lock;
3632 pa->pa_inode = ac->ac_inode;
3634 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3635 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3636 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3638 spin_lock(pa->pa_obj_lock);
3639 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3640 spin_unlock(pa->pa_obj_lock);
3646 * creates new preallocated space for locality group inodes belongs to
3648 static noinline_for_stack int
3649 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3651 struct super_block *sb = ac->ac_sb;
3652 struct ext4_locality_group *lg;
3653 struct ext4_prealloc_space *pa;
3654 struct ext4_group_info *grp;
3656 /* preallocate only when found space is larger then requested */
3657 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3658 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3659 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3661 BUG_ON(ext4_pspace_cachep == NULL);
3662 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3666 /* preallocation can change ac_b_ex, thus we store actually
3667 * allocated blocks for history */
3668 ac->ac_f_ex = ac->ac_b_ex;
3670 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3671 pa->pa_lstart = pa->pa_pstart;
3672 pa->pa_len = ac->ac_b_ex.fe_len;
3673 pa->pa_free = pa->pa_len;
3674 atomic_set(&pa->pa_count, 1);
3675 spin_lock_init(&pa->pa_lock);
3676 INIT_LIST_HEAD(&pa->pa_inode_list);
3677 INIT_LIST_HEAD(&pa->pa_group_list);
3679 pa->pa_type = MB_GROUP_PA;
3681 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3682 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3683 trace_ext4_mb_new_group_pa(ac, pa);
3685 ext4_mb_use_group_pa(ac, pa);
3686 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3688 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3692 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3693 pa->pa_inode = NULL;
3695 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3696 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3697 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3700 * We will later add the new pa to the right bucket
3701 * after updating the pa_free in ext4_mb_release_context
3706 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3710 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3711 err = ext4_mb_new_group_pa(ac);
3713 err = ext4_mb_new_inode_pa(ac);
3718 * finds all unused blocks in on-disk bitmap, frees them in
3719 * in-core bitmap and buddy.
3720 * @pa must be unlinked from inode and group lists, so that
3721 * nobody else can find/use it.
3722 * the caller MUST hold group/inode locks.
3723 * TODO: optimize the case when there are no in-core structures yet
3725 static noinline_for_stack int
3726 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3727 struct ext4_prealloc_space *pa)
3729 struct super_block *sb = e4b->bd_sb;
3730 struct ext4_sb_info *sbi = EXT4_SB(sb);
3735 unsigned long long grp_blk_start;
3739 BUG_ON(pa->pa_deleted == 0);
3740 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3741 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3742 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3743 end = bit + pa->pa_len;
3746 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3749 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3750 mb_debug(1, " free preallocated %u/%u in group %u\n",
3751 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3752 (unsigned) next - bit, (unsigned) group);
3755 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3756 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3757 EXT4_C2B(sbi, bit)),
3759 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3762 if (free != pa->pa_free) {
3763 ext4_msg(e4b->bd_sb, KERN_CRIT,
3764 "pa %p: logic %lu, phys. %lu, len %lu",
3765 pa, (unsigned long) pa->pa_lstart,
3766 (unsigned long) pa->pa_pstart,
3767 (unsigned long) pa->pa_len);
3768 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3771 * pa is already deleted so we use the value obtained
3772 * from the bitmap and continue.
3775 atomic_add(free, &sbi->s_mb_discarded);
3780 static noinline_for_stack int
3781 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3782 struct ext4_prealloc_space *pa)
3784 struct super_block *sb = e4b->bd_sb;
3788 trace_ext4_mb_release_group_pa(sb, pa);
3789 BUG_ON(pa->pa_deleted == 0);
3790 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3791 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3792 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3793 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3794 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3800 * releases all preallocations in given group
3802 * first, we need to decide discard policy:
3803 * - when do we discard
3805 * - how many do we discard
3806 * 1) how many requested
3808 static noinline_for_stack int
3809 ext4_mb_discard_group_preallocations(struct super_block *sb,
3810 ext4_group_t group, int needed)
3812 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3813 struct buffer_head *bitmap_bh = NULL;
3814 struct ext4_prealloc_space *pa, *tmp;
3815 struct list_head list;
3816 struct ext4_buddy e4b;
3821 mb_debug(1, "discard preallocation for group %u\n", group);
3823 if (list_empty(&grp->bb_prealloc_list))
3826 bitmap_bh = ext4_read_block_bitmap(sb, group);
3827 if (bitmap_bh == NULL) {
3828 ext4_error(sb, "Error reading block bitmap for %u", group);
3832 err = ext4_mb_load_buddy(sb, group, &e4b);
3834 ext4_error(sb, "Error loading buddy information for %u", group);
3840 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3842 INIT_LIST_HEAD(&list);
3844 ext4_lock_group(sb, group);
3845 list_for_each_entry_safe(pa, tmp,
3846 &grp->bb_prealloc_list, pa_group_list) {
3847 spin_lock(&pa->pa_lock);
3848 if (atomic_read(&pa->pa_count)) {
3849 spin_unlock(&pa->pa_lock);
3853 if (pa->pa_deleted) {
3854 spin_unlock(&pa->pa_lock);
3858 /* seems this one can be freed ... */
3861 /* we can trust pa_free ... */
3862 free += pa->pa_free;
3864 spin_unlock(&pa->pa_lock);
3866 list_del(&pa->pa_group_list);
3867 list_add(&pa->u.pa_tmp_list, &list);
3870 /* if we still need more blocks and some PAs were used, try again */
3871 if (free < needed && busy) {
3873 ext4_unlock_group(sb, group);
3878 /* found anything to free? */
3879 if (list_empty(&list)) {
3884 /* now free all selected PAs */
3885 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3887 /* remove from object (inode or locality group) */
3888 spin_lock(pa->pa_obj_lock);
3889 list_del_rcu(&pa->pa_inode_list);
3890 spin_unlock(pa->pa_obj_lock);
3892 if (pa->pa_type == MB_GROUP_PA)
3893 ext4_mb_release_group_pa(&e4b, pa);
3895 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3897 list_del(&pa->u.pa_tmp_list);
3898 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3902 ext4_unlock_group(sb, group);
3903 ext4_mb_unload_buddy(&e4b);
3909 * releases all non-used preallocated blocks for given inode
3911 * It's important to discard preallocations under i_data_sem
3912 * We don't want another block to be served from the prealloc
3913 * space when we are discarding the inode prealloc space.
3915 * FIXME!! Make sure it is valid at all the call sites
3917 void ext4_discard_preallocations(struct inode *inode)
3919 struct ext4_inode_info *ei = EXT4_I(inode);
3920 struct super_block *sb = inode->i_sb;
3921 struct buffer_head *bitmap_bh = NULL;
3922 struct ext4_prealloc_space *pa, *tmp;
3923 ext4_group_t group = 0;
3924 struct list_head list;
3925 struct ext4_buddy e4b;
3928 if (!S_ISREG(inode->i_mode)) {
3929 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3933 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3934 trace_ext4_discard_preallocations(inode);
3936 INIT_LIST_HEAD(&list);
3939 /* first, collect all pa's in the inode */
3940 spin_lock(&ei->i_prealloc_lock);
3941 while (!list_empty(&ei->i_prealloc_list)) {
3942 pa = list_entry(ei->i_prealloc_list.next,
3943 struct ext4_prealloc_space, pa_inode_list);
3944 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3945 spin_lock(&pa->pa_lock);
3946 if (atomic_read(&pa->pa_count)) {
3947 /* this shouldn't happen often - nobody should
3948 * use preallocation while we're discarding it */
3949 spin_unlock(&pa->pa_lock);
3950 spin_unlock(&ei->i_prealloc_lock);
3951 ext4_msg(sb, KERN_ERR,
3952 "uh-oh! used pa while discarding");
3954 schedule_timeout_uninterruptible(HZ);
3958 if (pa->pa_deleted == 0) {
3960 spin_unlock(&pa->pa_lock);
3961 list_del_rcu(&pa->pa_inode_list);
3962 list_add(&pa->u.pa_tmp_list, &list);
3966 /* someone is deleting pa right now */
3967 spin_unlock(&pa->pa_lock);
3968 spin_unlock(&ei->i_prealloc_lock);
3970 /* we have to wait here because pa_deleted
3971 * doesn't mean pa is already unlinked from
3972 * the list. as we might be called from
3973 * ->clear_inode() the inode will get freed
3974 * and concurrent thread which is unlinking
3975 * pa from inode's list may access already
3976 * freed memory, bad-bad-bad */
3978 /* XXX: if this happens too often, we can
3979 * add a flag to force wait only in case
3980 * of ->clear_inode(), but not in case of
3981 * regular truncate */
3982 schedule_timeout_uninterruptible(HZ);
3985 spin_unlock(&ei->i_prealloc_lock);
3987 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3988 BUG_ON(pa->pa_type != MB_INODE_PA);
3989 group = ext4_get_group_number(sb, pa->pa_pstart);
3991 err = ext4_mb_load_buddy(sb, group, &e4b);
3993 ext4_error(sb, "Error loading buddy information for %u",
3998 bitmap_bh = ext4_read_block_bitmap(sb, group);
3999 if (bitmap_bh == NULL) {
4000 ext4_error(sb, "Error reading block bitmap for %u",
4002 ext4_mb_unload_buddy(&e4b);
4006 ext4_lock_group(sb, group);
4007 list_del(&pa->pa_group_list);
4008 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4009 ext4_unlock_group(sb, group);
4011 ext4_mb_unload_buddy(&e4b);
4014 list_del(&pa->u.pa_tmp_list);
4015 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4019 #ifdef CONFIG_EXT4_DEBUG
4020 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4022 struct super_block *sb = ac->ac_sb;
4023 ext4_group_t ngroups, i;
4025 if (!ext4_mballoc_debug ||
4026 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4029 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4030 " Allocation context details:");
4031 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4032 ac->ac_status, ac->ac_flags);
4033 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4034 "goal %lu/%lu/%lu@%lu, "
4035 "best %lu/%lu/%lu@%lu cr %d",
4036 (unsigned long)ac->ac_o_ex.fe_group,
4037 (unsigned long)ac->ac_o_ex.fe_start,
4038 (unsigned long)ac->ac_o_ex.fe_len,
4039 (unsigned long)ac->ac_o_ex.fe_logical,
4040 (unsigned long)ac->ac_g_ex.fe_group,
4041 (unsigned long)ac->ac_g_ex.fe_start,
4042 (unsigned long)ac->ac_g_ex.fe_len,
4043 (unsigned long)ac->ac_g_ex.fe_logical,
4044 (unsigned long)ac->ac_b_ex.fe_group,
4045 (unsigned long)ac->ac_b_ex.fe_start,
4046 (unsigned long)ac->ac_b_ex.fe_len,
4047 (unsigned long)ac->ac_b_ex.fe_logical,
4048 (int)ac->ac_criteria);
4049 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4050 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4051 ngroups = ext4_get_groups_count(sb);
4052 for (i = 0; i < ngroups; i++) {
4053 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4054 struct ext4_prealloc_space *pa;
4055 ext4_grpblk_t start;
4056 struct list_head *cur;
4057 ext4_lock_group(sb, i);
4058 list_for_each(cur, &grp->bb_prealloc_list) {
4059 pa = list_entry(cur, struct ext4_prealloc_space,
4061 spin_lock(&pa->pa_lock);
4062 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4064 spin_unlock(&pa->pa_lock);
4065 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4068 ext4_unlock_group(sb, i);
4070 if (grp->bb_free == 0)
4072 printk(KERN_ERR "%u: %d/%d \n",
4073 i, grp->bb_free, grp->bb_fragments);
4075 printk(KERN_ERR "\n");
4078 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4085 * We use locality group preallocation for small size file. The size of the
4086 * file is determined by the current size or the resulting size after
4087 * allocation which ever is larger
4089 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4091 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4093 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4094 int bsbits = ac->ac_sb->s_blocksize_bits;
4097 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4100 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4103 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4104 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4107 if ((size == isize) &&
4108 !ext4_fs_is_busy(sbi) &&
4109 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4110 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4114 if (sbi->s_mb_group_prealloc <= 0) {
4115 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4119 /* don't use group allocation for large files */
4120 size = max(size, isize);
4121 if (size > sbi->s_mb_stream_request) {
4122 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4126 BUG_ON(ac->ac_lg != NULL);
4128 * locality group prealloc space are per cpu. The reason for having
4129 * per cpu locality group is to reduce the contention between block
4130 * request from multiple CPUs.
4132 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
4134 /* we're going to use group allocation */
4135 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4137 /* serialize all allocations in the group */
4138 mutex_lock(&ac->ac_lg->lg_mutex);
4141 static noinline_for_stack int
4142 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4143 struct ext4_allocation_request *ar)
4145 struct super_block *sb = ar->inode->i_sb;
4146 struct ext4_sb_info *sbi = EXT4_SB(sb);
4147 struct ext4_super_block *es = sbi->s_es;
4151 ext4_grpblk_t block;
4153 /* we can't allocate > group size */
4156 /* just a dirty hack to filter too big requests */
4157 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4158 len = EXT4_CLUSTERS_PER_GROUP(sb);
4160 /* start searching from the goal */
4162 if (goal < le32_to_cpu(es->s_first_data_block) ||
4163 goal >= ext4_blocks_count(es))
4164 goal = le32_to_cpu(es->s_first_data_block);
4165 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4167 /* set up allocation goals */
4168 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4169 ac->ac_status = AC_STATUS_CONTINUE;
4171 ac->ac_inode = ar->inode;
4172 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4173 ac->ac_o_ex.fe_group = group;
4174 ac->ac_o_ex.fe_start = block;
4175 ac->ac_o_ex.fe_len = len;
4176 ac->ac_g_ex = ac->ac_o_ex;
4177 ac->ac_flags = ar->flags;
4179 /* we have to define context: we'll we work with a file or
4180 * locality group. this is a policy, actually */
4181 ext4_mb_group_or_file(ac);
4183 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4184 "left: %u/%u, right %u/%u to %swritable\n",
4185 (unsigned) ar->len, (unsigned) ar->logical,
4186 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4187 (unsigned) ar->lleft, (unsigned) ar->pleft,
4188 (unsigned) ar->lright, (unsigned) ar->pright,
4189 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4194 static noinline_for_stack void
4195 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4196 struct ext4_locality_group *lg,
4197 int order, int total_entries)
4199 ext4_group_t group = 0;
4200 struct ext4_buddy e4b;
4201 struct list_head discard_list;
4202 struct ext4_prealloc_space *pa, *tmp;
4204 mb_debug(1, "discard locality group preallocation\n");
4206 INIT_LIST_HEAD(&discard_list);
4208 spin_lock(&lg->lg_prealloc_lock);
4209 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4211 spin_lock(&pa->pa_lock);
4212 if (atomic_read(&pa->pa_count)) {
4214 * This is the pa that we just used
4215 * for block allocation. So don't
4218 spin_unlock(&pa->pa_lock);
4221 if (pa->pa_deleted) {
4222 spin_unlock(&pa->pa_lock);
4225 /* only lg prealloc space */
4226 BUG_ON(pa->pa_type != MB_GROUP_PA);
4228 /* seems this one can be freed ... */
4230 spin_unlock(&pa->pa_lock);
4232 list_del_rcu(&pa->pa_inode_list);
4233 list_add(&pa->u.pa_tmp_list, &discard_list);
4236 if (total_entries <= 5) {
4238 * we want to keep only 5 entries
4239 * allowing it to grow to 8. This
4240 * mak sure we don't call discard
4241 * soon for this list.
4246 spin_unlock(&lg->lg_prealloc_lock);
4248 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4250 group = ext4_get_group_number(sb, pa->pa_pstart);
4251 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4252 ext4_error(sb, "Error loading buddy information for %u",
4256 ext4_lock_group(sb, group);
4257 list_del(&pa->pa_group_list);
4258 ext4_mb_release_group_pa(&e4b, pa);
4259 ext4_unlock_group(sb, group);
4261 ext4_mb_unload_buddy(&e4b);
4262 list_del(&pa->u.pa_tmp_list);
4263 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4268 * We have incremented pa_count. So it cannot be freed at this
4269 * point. Also we hold lg_mutex. So no parallel allocation is
4270 * possible from this lg. That means pa_free cannot be updated.
4272 * A parallel ext4_mb_discard_group_preallocations is possible.
4273 * which can cause the lg_prealloc_list to be updated.
4276 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4278 int order, added = 0, lg_prealloc_count = 1;
4279 struct super_block *sb = ac->ac_sb;
4280 struct ext4_locality_group *lg = ac->ac_lg;
4281 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4283 order = fls(pa->pa_free) - 1;
4284 if (order > PREALLOC_TB_SIZE - 1)
4285 /* The max size of hash table is PREALLOC_TB_SIZE */
4286 order = PREALLOC_TB_SIZE - 1;
4287 /* Add the prealloc space to lg */
4288 spin_lock(&lg->lg_prealloc_lock);
4289 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4291 spin_lock(&tmp_pa->pa_lock);
4292 if (tmp_pa->pa_deleted) {
4293 spin_unlock(&tmp_pa->pa_lock);
4296 if (!added && pa->pa_free < tmp_pa->pa_free) {
4297 /* Add to the tail of the previous entry */
4298 list_add_tail_rcu(&pa->pa_inode_list,
4299 &tmp_pa->pa_inode_list);
4302 * we want to count the total
4303 * number of entries in the list
4306 spin_unlock(&tmp_pa->pa_lock);
4307 lg_prealloc_count++;
4310 list_add_tail_rcu(&pa->pa_inode_list,
4311 &lg->lg_prealloc_list[order]);
4312 spin_unlock(&lg->lg_prealloc_lock);
4314 /* Now trim the list to be not more than 8 elements */
4315 if (lg_prealloc_count > 8) {
4316 ext4_mb_discard_lg_preallocations(sb, lg,
4317 order, lg_prealloc_count);
4324 * release all resource we used in allocation
4326 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4328 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4329 struct ext4_prealloc_space *pa = ac->ac_pa;
4331 if (pa->pa_type == MB_GROUP_PA) {
4332 /* see comment in ext4_mb_use_group_pa() */
4333 spin_lock(&pa->pa_lock);
4334 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4335 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4336 pa->pa_free -= ac->ac_b_ex.fe_len;
4337 pa->pa_len -= ac->ac_b_ex.fe_len;
4338 spin_unlock(&pa->pa_lock);
4343 * We want to add the pa to the right bucket.
4344 * Remove it from the list and while adding
4345 * make sure the list to which we are adding
4348 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4349 spin_lock(pa->pa_obj_lock);
4350 list_del_rcu(&pa->pa_inode_list);
4351 spin_unlock(pa->pa_obj_lock);
4352 ext4_mb_add_n_trim(ac);
4354 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4356 if (ac->ac_bitmap_page)
4357 page_cache_release(ac->ac_bitmap_page);
4358 if (ac->ac_buddy_page)
4359 page_cache_release(ac->ac_buddy_page);
4360 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4361 mutex_unlock(&ac->ac_lg->lg_mutex);
4362 ext4_mb_collect_stats(ac);
4366 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4368 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4372 trace_ext4_mb_discard_preallocations(sb, needed);
4373 for (i = 0; i < ngroups && needed > 0; i++) {
4374 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4383 * Main entry point into mballoc to allocate blocks
4384 * it tries to use preallocation first, then falls back
4385 * to usual allocation
4387 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4388 struct ext4_allocation_request *ar, int *errp)
4391 struct ext4_allocation_context *ac = NULL;
4392 struct ext4_sb_info *sbi;
4393 struct super_block *sb;
4394 ext4_fsblk_t block = 0;
4395 unsigned int inquota = 0;
4396 unsigned int reserv_clstrs = 0;
4399 sb = ar->inode->i_sb;
4402 trace_ext4_request_blocks(ar);
4404 /* Allow to use superuser reservation for quota file */
4405 if (IS_NOQUOTA(ar->inode))
4406 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4409 * For delayed allocation, we could skip the ENOSPC and
4410 * EDQUOT check, as blocks and quotas have been already
4411 * reserved when data being copied into pagecache.
4413 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4414 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4416 /* Without delayed allocation we need to verify
4417 * there is enough free blocks to do block allocation
4418 * and verify allocation doesn't exceed the quota limits.
4421 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4423 /* let others to free the space */
4425 ar->len = ar->len >> 1;
4431 reserv_clstrs = ar->len;
4432 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4433 dquot_alloc_block_nofail(ar->inode,
4434 EXT4_C2B(sbi, ar->len));
4437 dquot_alloc_block(ar->inode,
4438 EXT4_C2B(sbi, ar->len))) {
4440 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4451 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4458 *errp = ext4_mb_initialize_context(ac, ar);
4464 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4465 if (!ext4_mb_use_preallocated(ac)) {
4466 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4467 ext4_mb_normalize_request(ac, ar);
4469 /* allocate space in core */
4470 *errp = ext4_mb_regular_allocator(ac);
4472 goto discard_and_exit;
4474 /* as we've just preallocated more space than
4475 * user requested originally, we store allocated
4476 * space in a special descriptor */
4477 if (ac->ac_status == AC_STATUS_FOUND &&
4478 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4479 *errp = ext4_mb_new_preallocation(ac);
4482 ext4_discard_allocated_blocks(ac);
4486 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4487 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4488 if (*errp == -EAGAIN) {
4490 * drop the reference that we took
4491 * in ext4_mb_use_best_found
4493 ext4_mb_release_context(ac);
4494 ac->ac_b_ex.fe_group = 0;
4495 ac->ac_b_ex.fe_start = 0;
4496 ac->ac_b_ex.fe_len = 0;
4497 ac->ac_status = AC_STATUS_CONTINUE;
4500 ext4_discard_allocated_blocks(ac);
4503 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4504 ar->len = ac->ac_b_ex.fe_len;
4507 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4515 ac->ac_b_ex.fe_len = 0;
4517 ext4_mb_show_ac(ac);
4519 ext4_mb_release_context(ac);
4522 kmem_cache_free(ext4_ac_cachep, ac);
4523 if (inquota && ar->len < inquota)
4524 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4526 if (!ext4_test_inode_state(ar->inode,
4527 EXT4_STATE_DELALLOC_RESERVED))
4528 /* release all the reserved blocks if non delalloc */
4529 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4533 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4539 * We can merge two free data extents only if the physical blocks
4540 * are contiguous, AND the extents were freed by the same transaction,
4541 * AND the blocks are associated with the same group.
4543 static int can_merge(struct ext4_free_data *entry1,
4544 struct ext4_free_data *entry2)
4546 if ((entry1->efd_tid == entry2->efd_tid) &&
4547 (entry1->efd_group == entry2->efd_group) &&
4548 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4553 static noinline_for_stack int
4554 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4555 struct ext4_free_data *new_entry)
4557 ext4_group_t group = e4b->bd_group;
4558 ext4_grpblk_t cluster;
4559 struct ext4_free_data *entry;
4560 struct ext4_group_info *db = e4b->bd_info;
4561 struct super_block *sb = e4b->bd_sb;
4562 struct ext4_sb_info *sbi = EXT4_SB(sb);
4563 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4564 struct rb_node *parent = NULL, *new_node;
4566 BUG_ON(!ext4_handle_valid(handle));
4567 BUG_ON(e4b->bd_bitmap_page == NULL);
4568 BUG_ON(e4b->bd_buddy_page == NULL);
4570 new_node = &new_entry->efd_node;
4571 cluster = new_entry->efd_start_cluster;
4574 /* first free block exent. We need to
4575 protect buddy cache from being freed,
4576 * otherwise we'll refresh it from
4577 * on-disk bitmap and lose not-yet-available
4579 page_cache_get(e4b->bd_buddy_page);
4580 page_cache_get(e4b->bd_bitmap_page);
4584 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4585 if (cluster < entry->efd_start_cluster)
4587 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4588 n = &(*n)->rb_right;
4590 ext4_grp_locked_error(sb, group, 0,
4591 ext4_group_first_block_no(sb, group) +
4592 EXT4_C2B(sbi, cluster),
4593 "Block already on to-be-freed list");
4598 rb_link_node(new_node, parent, n);
4599 rb_insert_color(new_node, &db->bb_free_root);
4601 /* Now try to see the extent can be merged to left and right */
4602 node = rb_prev(new_node);
4604 entry = rb_entry(node, struct ext4_free_data, efd_node);
4605 if (can_merge(entry, new_entry) &&
4606 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4607 new_entry->efd_start_cluster = entry->efd_start_cluster;
4608 new_entry->efd_count += entry->efd_count;
4609 rb_erase(node, &(db->bb_free_root));
4610 kmem_cache_free(ext4_free_data_cachep, entry);
4614 node = rb_next(new_node);
4616 entry = rb_entry(node, struct ext4_free_data, efd_node);
4617 if (can_merge(new_entry, entry) &&
4618 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4619 new_entry->efd_count += entry->efd_count;
4620 rb_erase(node, &(db->bb_free_root));
4621 kmem_cache_free(ext4_free_data_cachep, entry);
4624 /* Add the extent to transaction's private list */
4625 ext4_journal_callback_add(handle, ext4_free_data_callback,
4626 &new_entry->efd_jce);
4631 * ext4_free_blocks() -- Free given blocks and update quota
4632 * @handle: handle for this transaction
4634 * @block: start physical block to free
4635 * @count: number of blocks to count
4636 * @flags: flags used by ext4_free_blocks
4638 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4639 struct buffer_head *bh, ext4_fsblk_t block,
4640 unsigned long count, int flags)
4642 struct buffer_head *bitmap_bh = NULL;
4643 struct super_block *sb = inode->i_sb;
4644 struct ext4_group_desc *gdp;
4645 unsigned int overflow;
4647 struct buffer_head *gd_bh;
4648 ext4_group_t block_group;
4649 struct ext4_sb_info *sbi;
4650 struct ext4_buddy e4b;
4651 unsigned int count_clusters;
4658 BUG_ON(block != bh->b_blocknr);
4660 block = bh->b_blocknr;
4664 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4665 !ext4_data_block_valid(sbi, block, count)) {
4666 ext4_error(sb, "Freeing blocks not in datazone - "
4667 "block = %llu, count = %lu", block, count);
4671 ext4_debug("freeing block %llu\n", block);
4672 trace_ext4_free_blocks(inode, block, count, flags);
4674 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4675 struct buffer_head *tbh = bh;
4678 BUG_ON(bh && (count > 1));
4680 for (i = 0; i < count; i++) {
4683 tbh = sb_find_get_block(inode->i_sb,
4687 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4688 inode, tbh, block + i);
4693 * We need to make sure we don't reuse the freed block until
4694 * after the transaction is committed, which we can do by
4695 * treating the block as metadata, below. We make an
4696 * exception if the inode is to be written in writeback mode
4697 * since writeback mode has weak data consistency guarantees.
4699 if (!ext4_should_writeback_data(inode))
4700 flags |= EXT4_FREE_BLOCKS_METADATA;
4703 * If the extent to be freed does not begin on a cluster
4704 * boundary, we need to deal with partial clusters at the
4705 * beginning and end of the extent. Normally we will free
4706 * blocks at the beginning or the end unless we are explicitly
4707 * requested to avoid doing so.
4709 overflow = EXT4_PBLK_COFF(sbi, block);
4711 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4712 overflow = sbi->s_cluster_ratio - overflow;
4714 if (count > overflow)
4723 overflow = EXT4_LBLK_COFF(sbi, count);
4725 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4726 if (count > overflow)
4731 count += sbi->s_cluster_ratio - overflow;
4736 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4738 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4739 ext4_get_group_info(sb, block_group))))
4743 * Check to see if we are freeing blocks across a group
4746 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4747 overflow = EXT4_C2B(sbi, bit) + count -
4748 EXT4_BLOCKS_PER_GROUP(sb);
4751 count_clusters = EXT4_NUM_B2C(sbi, count);
4752 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4757 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4763 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4764 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4765 in_range(block, ext4_inode_table(sb, gdp),
4766 EXT4_SB(sb)->s_itb_per_group) ||
4767 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4768 EXT4_SB(sb)->s_itb_per_group)) {
4770 ext4_error(sb, "Freeing blocks in system zone - "
4771 "Block = %llu, count = %lu", block, count);
4772 /* err = 0. ext4_std_error should be a no op */
4776 BUFFER_TRACE(bitmap_bh, "getting write access");
4777 err = ext4_journal_get_write_access(handle, bitmap_bh);
4782 * We are about to modify some metadata. Call the journal APIs
4783 * to unshare ->b_data if a currently-committing transaction is
4786 BUFFER_TRACE(gd_bh, "get_write_access");
4787 err = ext4_journal_get_write_access(handle, gd_bh);
4790 #ifdef AGGRESSIVE_CHECK
4793 for (i = 0; i < count_clusters; i++)
4794 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4797 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4799 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4803 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4804 struct ext4_free_data *new_entry;
4806 * blocks being freed are metadata. these blocks shouldn't
4807 * be used until this transaction is committed
4810 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4813 * We use a retry loop because
4814 * ext4_free_blocks() is not allowed to fail.
4817 congestion_wait(BLK_RW_ASYNC, HZ/50);
4820 new_entry->efd_start_cluster = bit;
4821 new_entry->efd_group = block_group;
4822 new_entry->efd_count = count_clusters;
4823 new_entry->efd_tid = handle->h_transaction->t_tid;
4825 ext4_lock_group(sb, block_group);
4826 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4827 ext4_mb_free_metadata(handle, &e4b, new_entry);
4829 /* need to update group_info->bb_free and bitmap
4830 * with group lock held. generate_buddy look at
4831 * them with group lock_held
4833 if (test_opt(sb, DISCARD)) {
4834 err = ext4_issue_discard(sb, block_group, bit, count);
4835 if (err && err != -EOPNOTSUPP)
4836 ext4_msg(sb, KERN_WARNING, "discard request in"
4837 " group:%d block:%d count:%lu failed"
4838 " with %d", block_group, bit, count,
4841 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4843 ext4_lock_group(sb, block_group);
4844 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4845 mb_free_blocks(inode, &e4b, bit, count_clusters);
4848 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4849 ext4_free_group_clusters_set(sb, gdp, ret);
4850 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4851 ext4_group_desc_csum_set(sb, block_group, gdp);
4852 ext4_unlock_group(sb, block_group);
4854 if (sbi->s_log_groups_per_flex) {
4855 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4856 atomic64_add(count_clusters,
4857 &sbi->s_flex_groups[flex_group].free_clusters);
4860 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4861 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4862 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4864 ext4_mb_unload_buddy(&e4b);
4866 /* We dirtied the bitmap block */
4867 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4868 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4870 /* And the group descriptor block */
4871 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4872 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4876 if (overflow && !err) {
4884 ext4_std_error(sb, err);
4889 * ext4_group_add_blocks() -- Add given blocks to an existing group
4890 * @handle: handle to this transaction
4892 * @block: start physical block to add to the block group
4893 * @count: number of blocks to free
4895 * This marks the blocks as free in the bitmap and buddy.
4897 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4898 ext4_fsblk_t block, unsigned long count)
4900 struct buffer_head *bitmap_bh = NULL;
4901 struct buffer_head *gd_bh;
4902 ext4_group_t block_group;
4905 struct ext4_group_desc *desc;
4906 struct ext4_sb_info *sbi = EXT4_SB(sb);
4907 struct ext4_buddy e4b;
4908 int err = 0, ret, blk_free_count;
4909 ext4_grpblk_t blocks_freed;
4911 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4916 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4918 * Check to see if we are freeing blocks across a group
4921 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4922 ext4_warning(sb, "too much blocks added to group %u\n",
4928 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4934 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4940 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4941 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4942 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4943 in_range(block + count - 1, ext4_inode_table(sb, desc),
4944 sbi->s_itb_per_group)) {
4945 ext4_error(sb, "Adding blocks in system zones - "
4946 "Block = %llu, count = %lu",
4952 BUFFER_TRACE(bitmap_bh, "getting write access");
4953 err = ext4_journal_get_write_access(handle, bitmap_bh);
4958 * We are about to modify some metadata. Call the journal APIs
4959 * to unshare ->b_data if a currently-committing transaction is
4962 BUFFER_TRACE(gd_bh, "get_write_access");
4963 err = ext4_journal_get_write_access(handle, gd_bh);
4967 for (i = 0, blocks_freed = 0; i < count; i++) {
4968 BUFFER_TRACE(bitmap_bh, "clear bit");
4969 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4970 ext4_error(sb, "bit already cleared for block %llu",
4971 (ext4_fsblk_t)(block + i));
4972 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4978 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4983 * need to update group_info->bb_free and bitmap
4984 * with group lock held. generate_buddy look at
4985 * them with group lock_held
4987 ext4_lock_group(sb, block_group);
4988 mb_clear_bits(bitmap_bh->b_data, bit, count);
4989 mb_free_blocks(NULL, &e4b, bit, count);
4990 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4991 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4992 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
4993 ext4_group_desc_csum_set(sb, block_group, desc);
4994 ext4_unlock_group(sb, block_group);
4995 percpu_counter_add(&sbi->s_freeclusters_counter,
4996 EXT4_NUM_B2C(sbi, blocks_freed));
4998 if (sbi->s_log_groups_per_flex) {
4999 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5000 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5001 &sbi->s_flex_groups[flex_group].free_clusters);
5004 ext4_mb_unload_buddy(&e4b);
5006 /* We dirtied the bitmap block */
5007 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5008 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5010 /* And the group descriptor block */
5011 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5012 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5018 ext4_std_error(sb, err);
5023 * ext4_trim_extent -- function to TRIM one single free extent in the group
5024 * @sb: super block for the file system
5025 * @start: starting block of the free extent in the alloc. group
5026 * @count: number of blocks to TRIM
5027 * @group: alloc. group we are working with
5028 * @e4b: ext4 buddy for the group
5030 * Trim "count" blocks starting at "start" in the "group". To assure that no
5031 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5032 * be called with under the group lock.
5034 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5035 ext4_group_t group, struct ext4_buddy *e4b)
5039 struct ext4_free_extent ex;
5042 trace_ext4_trim_extent(sb, group, start, count);
5044 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5046 ex.fe_start = start;
5047 ex.fe_group = group;
5051 * Mark blocks used, so no one can reuse them while
5054 mb_mark_used(e4b, &ex);
5055 ext4_unlock_group(sb, group);
5056 ret = ext4_issue_discard(sb, group, start, count);
5057 ext4_lock_group(sb, group);
5058 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5063 * ext4_trim_all_free -- function to trim all free space in alloc. group
5064 * @sb: super block for file system
5065 * @group: group to be trimmed
5066 * @start: first group block to examine
5067 * @max: last group block to examine
5068 * @minblocks: minimum extent block count
5070 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5071 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5075 * ext4_trim_all_free walks through group's block bitmap searching for free
5076 * extents. When the free extent is found, mark it as used in group buddy
5077 * bitmap. Then issue a TRIM command on this extent and free the extent in
5078 * the group buddy bitmap. This is done until whole group is scanned.
5080 static ext4_grpblk_t
5081 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5082 ext4_grpblk_t start, ext4_grpblk_t max,
5083 ext4_grpblk_t minblocks)
5086 ext4_grpblk_t next, count = 0, free_count = 0;
5087 struct ext4_buddy e4b;
5090 trace_ext4_trim_all_free(sb, group, start, max);
5092 ret = ext4_mb_load_buddy(sb, group, &e4b);
5094 ext4_error(sb, "Error in loading buddy "
5095 "information for %u", group);
5098 bitmap = e4b.bd_bitmap;
5100 ext4_lock_group(sb, group);
5101 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5102 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5105 start = (e4b.bd_info->bb_first_free > start) ?
5106 e4b.bd_info->bb_first_free : start;
5108 while (start <= max) {
5109 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5112 next = mb_find_next_bit(bitmap, max + 1, start);
5114 if ((next - start) >= minblocks) {
5115 ret = ext4_trim_extent(sb, start,
5116 next - start, group, &e4b);
5117 if (ret && ret != -EOPNOTSUPP)
5120 count += next - start;
5122 free_count += next - start;
5125 if (fatal_signal_pending(current)) {
5126 count = -ERESTARTSYS;
5130 if (need_resched()) {
5131 ext4_unlock_group(sb, group);
5133 ext4_lock_group(sb, group);
5136 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5142 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5145 ext4_unlock_group(sb, group);
5146 ext4_mb_unload_buddy(&e4b);
5148 ext4_debug("trimmed %d blocks in the group %d\n",
5155 * ext4_trim_fs() -- trim ioctl handle function
5156 * @sb: superblock for filesystem
5157 * @range: fstrim_range structure
5159 * start: First Byte to trim
5160 * len: number of Bytes to trim from start
5161 * minlen: minimum extent length in Bytes
5162 * ext4_trim_fs goes through all allocation groups containing Bytes from
5163 * start to start+len. For each such a group ext4_trim_all_free function
5164 * is invoked to trim all free space.
5166 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5168 struct ext4_group_info *grp;
5169 ext4_group_t group, first_group, last_group;
5170 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5171 uint64_t start, end, minlen, trimmed = 0;
5172 ext4_fsblk_t first_data_blk =
5173 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5174 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5177 start = range->start >> sb->s_blocksize_bits;
5178 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5179 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5180 range->minlen >> sb->s_blocksize_bits);
5182 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5183 start >= max_blks ||
5184 range->len < sb->s_blocksize)
5186 if (end >= max_blks)
5188 if (end <= first_data_blk)
5190 if (start < first_data_blk)
5191 start = first_data_blk;
5193 /* Determine first and last group to examine based on start and end */
5194 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5195 &first_group, &first_cluster);
5196 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5197 &last_group, &last_cluster);
5199 /* end now represents the last cluster to discard in this group */
5200 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5202 for (group = first_group; group <= last_group; group++) {
5203 grp = ext4_get_group_info(sb, group);
5204 /* We only do this if the grp has never been initialized */
5205 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5206 ret = ext4_mb_init_group(sb, group);
5212 * For all the groups except the last one, last cluster will
5213 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5214 * change it for the last group, note that last_cluster is
5215 * already computed earlier by ext4_get_group_no_and_offset()
5217 if (group == last_group)
5220 if (grp->bb_free >= minlen) {
5221 cnt = ext4_trim_all_free(sb, group, first_cluster,
5231 * For every group except the first one, we are sure
5232 * that the first cluster to discard will be cluster #0.
5238 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5241 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;