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 <linux/backing-dev.h>
30 #include <trace/events/ext4.h>
32 #ifdef CONFIG_EXT4_DEBUG
33 ushort ext4_mballoc_debug __read_mostly;
35 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
36 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
41 * - test ext4_ext_search_left() and ext4_ext_search_right()
42 * - search for metadata in few groups
45 * - normalization should take into account whether file is still open
46 * - discard preallocations if no free space left (policy?)
47 * - don't normalize tails
49 * - reservation for superuser
52 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
53 * - track min/max extents in each group for better group selection
54 * - mb_mark_used() may allocate chunk right after splitting buddy
55 * - tree of groups sorted by number of free blocks
60 * The allocation request involve request for multiple number of blocks
61 * near to the goal(block) value specified.
63 * During initialization phase of the allocator we decide to use the
64 * group preallocation or inode preallocation depending on the size of
65 * the file. The size of the file could be the resulting file size we
66 * would have after allocation, or the current file size, which ever
67 * is larger. If the size is less than sbi->s_mb_stream_request we
68 * select to use the group preallocation. The default value of
69 * s_mb_stream_request is 16 blocks. This can also be tuned via
70 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
71 * terms of number of blocks.
73 * The main motivation for having small file use group preallocation is to
74 * ensure that we have small files closer together on the disk.
76 * First stage the allocator looks at the inode prealloc list,
77 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
78 * spaces for this particular inode. The inode prealloc space is
81 * pa_lstart -> the logical start block for this prealloc space
82 * pa_pstart -> the physical start block for this prealloc space
83 * pa_len -> length for this prealloc space (in clusters)
84 * pa_free -> free space available in this prealloc space (in clusters)
86 * The inode preallocation space is used looking at the _logical_ start
87 * block. If only the logical file block falls within the range of prealloc
88 * space we will consume the particular prealloc space. This makes sure that
89 * we have contiguous physical blocks representing the file blocks
91 * The important thing to be noted in case of inode prealloc space is that
92 * we don't modify the values associated to inode prealloc space except
95 * If we are not able to find blocks in the inode prealloc space and if we
96 * have the group allocation flag set then we look at the locality group
97 * prealloc space. These are per CPU prealloc list represented as
99 * ext4_sb_info.s_locality_groups[smp_processor_id()]
101 * The reason for having a per cpu locality group is to reduce the contention
102 * between CPUs. It is possible to get scheduled at this point.
104 * The locality group prealloc space is used looking at whether we have
105 * enough free space (pa_free) within the prealloc space.
107 * If we can't allocate blocks via inode prealloc or/and locality group
108 * prealloc then we look at the buddy cache. The buddy cache is represented
109 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
110 * mapped to the buddy and bitmap information regarding different
111 * groups. The buddy information is attached to buddy cache inode so that
112 * we can access them through the page cache. The information regarding
113 * each group is loaded via ext4_mb_load_buddy. The information involve
114 * block bitmap and buddy information. The information are stored in the
118 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
121 * one block each for bitmap and buddy information. So for each group we
122 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
123 * blocksize) blocks. So it can have information regarding groups_per_page
124 * which is blocks_per_page/2
126 * The buddy cache inode is not stored on disk. The inode is thrown
127 * away when the filesystem is unmounted.
129 * We look for count number of blocks in the buddy cache. If we were able
130 * to locate that many free blocks we return with additional information
131 * regarding rest of the contiguous physical block available
133 * Before allocating blocks via buddy cache we normalize the request
134 * blocks. This ensure we ask for more blocks that we needed. The extra
135 * blocks that we get after allocation is added to the respective prealloc
136 * list. In case of inode preallocation we follow a list of heuristics
137 * based on file size. This can be found in ext4_mb_normalize_request. If
138 * we are doing a group prealloc we try to normalize the request to
139 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
140 * dependent on the cluster size; for non-bigalloc file systems, it is
141 * 512 blocks. This can be tuned via
142 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
143 * terms of number of blocks. If we have mounted the file system with -O
144 * stripe=<value> option the group prealloc request is normalized to the
145 * the smallest multiple of the stripe value (sbi->s_stripe) which is
146 * greater than the default mb_group_prealloc.
148 * The regular allocator (using the buddy cache) supports a few tunables.
150 * /sys/fs/ext4/<partition>/mb_min_to_scan
151 * /sys/fs/ext4/<partition>/mb_max_to_scan
152 * /sys/fs/ext4/<partition>/mb_order2_req
154 * The regular allocator uses buddy scan only if the request len is power of
155 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
156 * value of s_mb_order2_reqs can be tuned via
157 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
158 * stripe size (sbi->s_stripe), we try to search for contiguous block in
159 * stripe size. This should result in better allocation on RAID setups. If
160 * not, we search in the specific group using bitmap for best extents. The
161 * tunable min_to_scan and max_to_scan control the behaviour here.
162 * min_to_scan indicate how long the mballoc __must__ look for a best
163 * extent and max_to_scan indicates how long the mballoc __can__ look for a
164 * best extent in the found extents. Searching for the blocks starts with
165 * the group specified as the goal value in allocation context via
166 * ac_g_ex. Each group is first checked based on the criteria whether it
167 * can be used for allocation. ext4_mb_good_group explains how the groups are
170 * Both the prealloc space are getting populated as above. So for the first
171 * request we will hit the buddy cache which will result in this prealloc
172 * space getting filled. The prealloc space is then later used for the
173 * subsequent request.
177 * mballoc operates on the following data:
179 * - in-core buddy (actually includes buddy and bitmap)
180 * - preallocation descriptors (PAs)
182 * there are two types of preallocations:
184 * assiged to specific inode and can be used for this inode only.
185 * it describes part of inode's space preallocated to specific
186 * physical blocks. any block from that preallocated can be used
187 * independent. the descriptor just tracks number of blocks left
188 * unused. so, before taking some block from descriptor, one must
189 * make sure corresponded logical block isn't allocated yet. this
190 * also means that freeing any block within descriptor's range
191 * must discard all preallocated blocks.
193 * assigned to specific locality group which does not translate to
194 * permanent set of inodes: inode can join and leave group. space
195 * from this type of preallocation can be used for any inode. thus
196 * it's consumed from the beginning to the end.
198 * relation between them can be expressed as:
199 * in-core buddy = on-disk bitmap + preallocation descriptors
201 * this mean blocks mballoc considers used are:
202 * - allocated blocks (persistent)
203 * - preallocated blocks (non-persistent)
205 * consistency in mballoc world means that at any time a block is either
206 * free or used in ALL structures. notice: "any time" should not be read
207 * literally -- time is discrete and delimited by locks.
209 * to keep it simple, we don't use block numbers, instead we count number of
210 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
212 * all operations can be expressed as:
213 * - init buddy: buddy = on-disk + PAs
214 * - new PA: buddy += N; PA = N
215 * - use inode PA: on-disk += N; PA -= N
216 * - discard inode PA buddy -= on-disk - PA; PA = 0
217 * - use locality group PA on-disk += N; PA -= N
218 * - discard locality group PA buddy -= PA; PA = 0
219 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
220 * is used in real operation because we can't know actual used
221 * bits from PA, only from on-disk bitmap
223 * if we follow this strict logic, then all operations above should be atomic.
224 * given some of them can block, we'd have to use something like semaphores
225 * killing performance on high-end SMP hardware. let's try to relax it using
226 * the following knowledge:
227 * 1) if buddy is referenced, it's already initialized
228 * 2) while block is used in buddy and the buddy is referenced,
229 * nobody can re-allocate that block
230 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
231 * bit set and PA claims same block, it's OK. IOW, one can set bit in
232 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
235 * so, now we're building a concurrency table:
238 * blocks for PA are allocated in the buddy, buddy must be referenced
239 * until PA is linked to allocation group to avoid concurrent buddy init
241 * we need to make sure that either on-disk bitmap or PA has uptodate data
242 * given (3) we care that PA-=N operation doesn't interfere with init
244 * the simplest way would be to have buddy initialized by the discard
245 * - use locality group PA
246 * again PA-=N must be serialized with init
247 * - discard locality group PA
248 * the simplest way would be to have buddy initialized by the discard
251 * i_data_sem serializes them
253 * discard process must wait until PA isn't used by another process
254 * - use locality group PA
255 * some mutex should serialize them
256 * - discard locality group PA
257 * discard process must wait until PA isn't used by another process
260 * i_data_sem or another mutex should serializes them
262 * discard process must wait until PA isn't used by another process
263 * - use locality group PA
264 * nothing wrong here -- they're different PAs covering different blocks
265 * - discard locality group PA
266 * discard process must wait until PA isn't used by another process
268 * now we're ready to make few consequences:
269 * - PA is referenced and while it is no discard is possible
270 * - PA is referenced until block isn't marked in on-disk bitmap
271 * - PA changes only after on-disk bitmap
272 * - discard must not compete with init. either init is done before
273 * any discard or they're serialized somehow
274 * - buddy init as sum of on-disk bitmap and PAs is done atomically
276 * a special case when we've used PA to emptiness. no need to modify buddy
277 * in this case, but we should care about concurrent init
282 * Logic in few words:
287 * mark bits in on-disk bitmap
290 * - use preallocation:
291 * find proper PA (per-inode or group)
293 * mark bits in on-disk bitmap
299 * mark bits in on-disk bitmap
302 * - discard preallocations in group:
304 * move them onto local list
305 * load on-disk bitmap
307 * remove PA from object (inode or locality group)
308 * mark free blocks in-core
310 * - discard inode's preallocations:
317 * - bitlock on a group (group)
318 * - object (inode/locality) (object)
329 * - release consumed pa:
334 * - generate in-core bitmap:
338 * - discard all for given object (inode, locality group):
343 * - discard all for given group:
350 static struct kmem_cache *ext4_pspace_cachep;
351 static struct kmem_cache *ext4_ac_cachep;
352 static struct kmem_cache *ext4_free_data_cachep;
354 /* We create slab caches for groupinfo data structures based on the
355 * superblock block size. There will be one per mounted filesystem for
356 * each unique s_blocksize_bits */
357 #define NR_GRPINFO_CACHES 8
358 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
360 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
361 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
362 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
363 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
366 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
368 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
370 static void ext4_free_data_callback(struct super_block *sb,
371 struct ext4_journal_cb_entry *jce, int rc);
373 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
375 #if BITS_PER_LONG == 64
376 *bit += ((unsigned long) addr & 7UL) << 3;
377 addr = (void *) ((unsigned long) addr & ~7UL);
378 #elif BITS_PER_LONG == 32
379 *bit += ((unsigned long) addr & 3UL) << 3;
380 addr = (void *) ((unsigned long) addr & ~3UL);
382 #error "how many bits you are?!"
387 static inline int mb_test_bit(int bit, void *addr)
390 * ext4_test_bit on architecture like powerpc
391 * needs unsigned long aligned address
393 addr = mb_correct_addr_and_bit(&bit, addr);
394 return ext4_test_bit(bit, addr);
397 static inline void mb_set_bit(int bit, void *addr)
399 addr = mb_correct_addr_and_bit(&bit, addr);
400 ext4_set_bit(bit, addr);
403 static inline void mb_clear_bit(int bit, void *addr)
405 addr = mb_correct_addr_and_bit(&bit, addr);
406 ext4_clear_bit(bit, addr);
409 static inline int mb_test_and_clear_bit(int bit, void *addr)
411 addr = mb_correct_addr_and_bit(&bit, addr);
412 return ext4_test_and_clear_bit(bit, addr);
415 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
417 int fix = 0, ret, tmpmax;
418 addr = mb_correct_addr_and_bit(&fix, addr);
422 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
428 static inline int mb_find_next_bit(void *addr, int max, int start)
430 int fix = 0, ret, tmpmax;
431 addr = mb_correct_addr_and_bit(&fix, addr);
435 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
441 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
445 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
448 if (order > e4b->bd_blkbits + 1) {
453 /* at order 0 we see each particular block */
455 *max = 1 << (e4b->bd_blkbits + 3);
456 return e4b->bd_bitmap;
459 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
460 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
466 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
467 int first, int count)
470 struct super_block *sb = e4b->bd_sb;
472 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
474 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
475 for (i = 0; i < count; i++) {
476 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
477 ext4_fsblk_t blocknr;
479 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
480 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
481 ext4_grp_locked_error(sb, e4b->bd_group,
482 inode ? inode->i_ino : 0,
484 "freeing block already freed "
488 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
492 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
496 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
498 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
499 for (i = 0; i < count; i++) {
500 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
501 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
505 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
507 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
508 unsigned char *b1, *b2;
510 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
511 b2 = (unsigned char *) bitmap;
512 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
513 if (b1[i] != b2[i]) {
514 ext4_msg(e4b->bd_sb, KERN_ERR,
515 "corruption in group %u "
516 "at byte %u(%u): %x in copy != %x "
518 e4b->bd_group, i, i * 8, b1[i], b2[i]);
526 static inline void mb_free_blocks_double(struct inode *inode,
527 struct ext4_buddy *e4b, int first, int count)
531 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
532 int first, int count)
536 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
542 #ifdef AGGRESSIVE_CHECK
544 #define MB_CHECK_ASSERT(assert) \
548 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
549 function, file, line, # assert); \
554 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
555 const char *function, int line)
557 struct super_block *sb = e4b->bd_sb;
558 int order = e4b->bd_blkbits + 1;
565 struct ext4_group_info *grp;
568 struct list_head *cur;
573 static int mb_check_counter;
574 if (mb_check_counter++ % 100 != 0)
579 buddy = mb_find_buddy(e4b, order, &max);
580 MB_CHECK_ASSERT(buddy);
581 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
582 MB_CHECK_ASSERT(buddy2);
583 MB_CHECK_ASSERT(buddy != buddy2);
584 MB_CHECK_ASSERT(max * 2 == max2);
587 for (i = 0; i < max; i++) {
589 if (mb_test_bit(i, buddy)) {
590 /* only single bit in buddy2 may be 1 */
591 if (!mb_test_bit(i << 1, buddy2)) {
593 mb_test_bit((i<<1)+1, buddy2));
594 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
596 mb_test_bit(i << 1, buddy2));
601 /* both bits in buddy2 must be 1 */
602 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
603 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
605 for (j = 0; j < (1 << order); j++) {
606 k = (i * (1 << order)) + j;
608 !mb_test_bit(k, e4b->bd_bitmap));
612 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
617 buddy = mb_find_buddy(e4b, 0, &max);
618 for (i = 0; i < max; i++) {
619 if (!mb_test_bit(i, buddy)) {
620 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
628 /* check used bits only */
629 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
630 buddy2 = mb_find_buddy(e4b, j, &max2);
632 MB_CHECK_ASSERT(k < max2);
633 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
636 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
637 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
639 grp = ext4_get_group_info(sb, e4b->bd_group);
640 list_for_each(cur, &grp->bb_prealloc_list) {
641 ext4_group_t groupnr;
642 struct ext4_prealloc_space *pa;
643 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
644 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
645 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
646 for (i = 0; i < pa->pa_len; i++)
647 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
651 #undef MB_CHECK_ASSERT
652 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
653 __FILE__, __func__, __LINE__)
655 #define mb_check_buddy(e4b)
659 * Divide blocks started from @first with length @len into
660 * smaller chunks with power of 2 blocks.
661 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
662 * then increase bb_counters[] for corresponded chunk size.
664 static void ext4_mb_mark_free_simple(struct super_block *sb,
665 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
666 struct ext4_group_info *grp)
668 struct ext4_sb_info *sbi = EXT4_SB(sb);
672 unsigned short border;
674 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
676 border = 2 << sb->s_blocksize_bits;
679 /* find how many blocks can be covered since this position */
680 max = ffs(first | border) - 1;
682 /* find how many blocks of power 2 we need to mark */
689 /* mark multiblock chunks only */
690 grp->bb_counters[min]++;
692 mb_clear_bit(first >> min,
693 buddy + sbi->s_mb_offsets[min]);
701 * Cache the order of the largest free extent we have available in this block
705 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
710 grp->bb_largest_free_order = -1; /* uninit */
712 bits = sb->s_blocksize_bits + 1;
713 for (i = bits; i >= 0; i--) {
714 if (grp->bb_counters[i] > 0) {
715 grp->bb_largest_free_order = i;
721 static noinline_for_stack
722 void ext4_mb_generate_buddy(struct super_block *sb,
723 void *buddy, void *bitmap, ext4_group_t group)
725 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
726 struct ext4_sb_info *sbi = EXT4_SB(sb);
727 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
732 unsigned fragments = 0;
733 unsigned long long period = get_cycles();
735 /* initialize buddy from bitmap which is aggregation
736 * of on-disk bitmap and preallocations */
737 i = mb_find_next_zero_bit(bitmap, max, 0);
738 grp->bb_first_free = i;
742 i = mb_find_next_bit(bitmap, max, i);
746 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
748 grp->bb_counters[0]++;
750 i = mb_find_next_zero_bit(bitmap, max, i);
752 grp->bb_fragments = fragments;
754 if (free != grp->bb_free) {
755 ext4_grp_locked_error(sb, group, 0, 0,
756 "block bitmap and bg descriptor "
757 "inconsistent: %u vs %u free clusters",
760 * If we intend to continue, we consider group descriptor
761 * corrupt and update bb_free using bitmap value
764 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
765 percpu_counter_sub(&sbi->s_freeclusters_counter,
767 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
769 mb_set_largest_free_order(sb, grp);
771 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
773 period = get_cycles() - period;
774 spin_lock(&EXT4_SB(sb)->s_bal_lock);
775 EXT4_SB(sb)->s_mb_buddies_generated++;
776 EXT4_SB(sb)->s_mb_generation_time += period;
777 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
780 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
786 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
787 ext4_set_bits(buddy, 0, count);
789 e4b->bd_info->bb_fragments = 0;
790 memset(e4b->bd_info->bb_counters, 0,
791 sizeof(*e4b->bd_info->bb_counters) *
792 (e4b->bd_sb->s_blocksize_bits + 2));
794 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
795 e4b->bd_bitmap, e4b->bd_group);
798 /* The buddy information is attached the buddy cache inode
799 * for convenience. The information regarding each group
800 * is loaded via ext4_mb_load_buddy. The information involve
801 * block bitmap and buddy information. The information are
802 * stored in the inode as
805 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
808 * one block each for bitmap and buddy information.
809 * So for each group we take up 2 blocks. A page can
810 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
811 * So it can have information regarding groups_per_page which
812 * is blocks_per_page/2
814 * Locking note: This routine takes the block group lock of all groups
815 * for this page; do not hold this lock when calling this routine!
818 static int ext4_mb_init_cache(struct page *page, char *incore)
820 ext4_group_t ngroups;
826 ext4_group_t first_group, group;
828 struct super_block *sb;
829 struct buffer_head *bhs;
830 struct buffer_head **bh = NULL;
834 struct ext4_group_info *grinfo;
836 mb_debug(1, "init page %lu\n", page->index);
838 inode = page->mapping->host;
840 ngroups = ext4_get_groups_count(sb);
841 blocksize = 1 << inode->i_blkbits;
842 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
844 groups_per_page = blocks_per_page >> 1;
845 if (groups_per_page == 0)
848 /* allocate buffer_heads to read bitmaps */
849 if (groups_per_page > 1) {
850 i = sizeof(struct buffer_head *) * groups_per_page;
851 bh = kzalloc(i, GFP_NOFS);
859 first_group = page->index * blocks_per_page / 2;
861 /* read all groups the page covers into the cache */
862 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
863 if (group >= ngroups)
866 grinfo = ext4_get_group_info(sb, group);
868 * If page is uptodate then we came here after online resize
869 * which added some new uninitialized group info structs, so
870 * we must skip all initialized uptodate buddies on the page,
871 * which may be currently in use by an allocating task.
873 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
877 bh[i] = ext4_read_block_bitmap_nowait(sb, group);
879 err = PTR_ERR(bh[i]);
883 mb_debug(1, "read bitmap for group %u\n", group);
886 /* wait for I/O completion */
887 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
892 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
897 first_block = page->index * blocks_per_page;
898 for (i = 0; i < blocks_per_page; i++) {
899 group = (first_block + i) >> 1;
900 if (group >= ngroups)
903 if (!bh[group - first_group])
904 /* skip initialized uptodate buddy */
907 if (!buffer_verified(bh[group - first_group]))
908 /* Skip faulty bitmaps */
913 * data carry information regarding this
914 * particular group in the format specified
918 data = page_address(page) + (i * blocksize);
919 bitmap = bh[group - first_group]->b_data;
922 * We place the buddy block and bitmap block
925 if ((first_block + i) & 1) {
926 /* this is block of buddy */
927 BUG_ON(incore == NULL);
928 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
929 group, page->index, i * blocksize);
930 trace_ext4_mb_buddy_bitmap_load(sb, group);
931 grinfo = ext4_get_group_info(sb, group);
932 grinfo->bb_fragments = 0;
933 memset(grinfo->bb_counters, 0,
934 sizeof(*grinfo->bb_counters) *
935 (sb->s_blocksize_bits+2));
937 * incore got set to the group block bitmap below
939 ext4_lock_group(sb, group);
941 memset(data, 0xff, blocksize);
942 ext4_mb_generate_buddy(sb, data, incore, group);
943 ext4_unlock_group(sb, group);
946 /* this is block of bitmap */
947 BUG_ON(incore != NULL);
948 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
949 group, page->index, i * blocksize);
950 trace_ext4_mb_bitmap_load(sb, group);
952 /* see comments in ext4_mb_put_pa() */
953 ext4_lock_group(sb, group);
954 memcpy(data, bitmap, blocksize);
956 /* mark all preallocated blks used in in-core bitmap */
957 ext4_mb_generate_from_pa(sb, data, group);
958 ext4_mb_generate_from_freelist(sb, data, group);
959 ext4_unlock_group(sb, group);
961 /* set incore so that the buddy information can be
962 * generated using this
967 SetPageUptodate(page);
971 for (i = 0; i < groups_per_page; i++)
980 * Lock the buddy and bitmap pages. This make sure other parallel init_group
981 * on the same buddy page doesn't happen whild holding the buddy page lock.
982 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
983 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
985 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
986 ext4_group_t group, struct ext4_buddy *e4b)
988 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
989 int block, pnum, poff;
993 e4b->bd_buddy_page = NULL;
994 e4b->bd_bitmap_page = NULL;
996 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
998 * the buddy cache inode stores the block bitmap
999 * and buddy information in consecutive blocks.
1000 * So for each group we need two blocks.
1003 pnum = block / blocks_per_page;
1004 poff = block % blocks_per_page;
1005 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1008 BUG_ON(page->mapping != inode->i_mapping);
1009 e4b->bd_bitmap_page = page;
1010 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1012 if (blocks_per_page >= 2) {
1013 /* buddy and bitmap are on the same page */
1018 pnum = block / blocks_per_page;
1019 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1022 BUG_ON(page->mapping != inode->i_mapping);
1023 e4b->bd_buddy_page = page;
1027 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1029 if (e4b->bd_bitmap_page) {
1030 unlock_page(e4b->bd_bitmap_page);
1031 page_cache_release(e4b->bd_bitmap_page);
1033 if (e4b->bd_buddy_page) {
1034 unlock_page(e4b->bd_buddy_page);
1035 page_cache_release(e4b->bd_buddy_page);
1040 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1041 * block group lock of all groups for this page; do not hold the BG lock when
1042 * calling this routine!
1044 static noinline_for_stack
1045 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1048 struct ext4_group_info *this_grp;
1049 struct ext4_buddy e4b;
1054 mb_debug(1, "init group %u\n", group);
1055 this_grp = ext4_get_group_info(sb, group);
1057 * This ensures that we don't reinit the buddy cache
1058 * page which map to the group from which we are already
1059 * allocating. If we are looking at the buddy cache we would
1060 * have taken a reference using ext4_mb_load_buddy and that
1061 * would have pinned buddy page to page cache.
1062 * The call to ext4_mb_get_buddy_page_lock will mark the
1065 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1066 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1068 * somebody initialized the group
1069 * return without doing anything
1074 page = e4b.bd_bitmap_page;
1075 ret = ext4_mb_init_cache(page, NULL);
1078 if (!PageUptodate(page)) {
1083 if (e4b.bd_buddy_page == NULL) {
1085 * If both the bitmap and buddy are in
1086 * the same page we don't need to force
1092 /* init buddy cache */
1093 page = e4b.bd_buddy_page;
1094 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1097 if (!PageUptodate(page)) {
1102 ext4_mb_put_buddy_page_lock(&e4b);
1107 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1108 * block group lock of all groups for this page; do not hold the BG lock when
1109 * calling this routine!
1111 static noinline_for_stack int
1112 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1113 struct ext4_buddy *e4b)
1115 int blocks_per_page;
1121 struct ext4_group_info *grp;
1122 struct ext4_sb_info *sbi = EXT4_SB(sb);
1123 struct inode *inode = sbi->s_buddy_cache;
1126 mb_debug(1, "load group %u\n", group);
1128 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1129 grp = ext4_get_group_info(sb, group);
1131 e4b->bd_blkbits = sb->s_blocksize_bits;
1134 e4b->bd_group = group;
1135 e4b->bd_buddy_page = NULL;
1136 e4b->bd_bitmap_page = NULL;
1138 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1140 * we need full data about the group
1141 * to make a good selection
1143 ret = ext4_mb_init_group(sb, group);
1149 * the buddy cache inode stores the block bitmap
1150 * and buddy information in consecutive blocks.
1151 * So for each group we need two blocks.
1154 pnum = block / blocks_per_page;
1155 poff = block % blocks_per_page;
1157 /* we could use find_or_create_page(), but it locks page
1158 * what we'd like to avoid in fast path ... */
1159 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1160 if (page == NULL || !PageUptodate(page)) {
1163 * drop the page reference and try
1164 * to get the page with lock. If we
1165 * are not uptodate that implies
1166 * somebody just created the page but
1167 * is yet to initialize the same. So
1168 * wait for it to initialize.
1170 page_cache_release(page);
1171 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1173 BUG_ON(page->mapping != inode->i_mapping);
1174 if (!PageUptodate(page)) {
1175 ret = ext4_mb_init_cache(page, NULL);
1180 mb_cmp_bitmaps(e4b, page_address(page) +
1181 (poff * sb->s_blocksize));
1190 if (!PageUptodate(page)) {
1195 /* Pages marked accessed already */
1196 e4b->bd_bitmap_page = page;
1197 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1200 pnum = block / blocks_per_page;
1201 poff = block % blocks_per_page;
1203 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1204 if (page == NULL || !PageUptodate(page)) {
1206 page_cache_release(page);
1207 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1209 BUG_ON(page->mapping != inode->i_mapping);
1210 if (!PageUptodate(page)) {
1211 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1224 if (!PageUptodate(page)) {
1229 /* Pages marked accessed already */
1230 e4b->bd_buddy_page = page;
1231 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1233 BUG_ON(e4b->bd_bitmap_page == NULL);
1234 BUG_ON(e4b->bd_buddy_page == NULL);
1240 page_cache_release(page);
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);
1245 e4b->bd_buddy = NULL;
1246 e4b->bd_bitmap = NULL;
1250 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1252 if (e4b->bd_bitmap_page)
1253 page_cache_release(e4b->bd_bitmap_page);
1254 if (e4b->bd_buddy_page)
1255 page_cache_release(e4b->bd_buddy_page);
1259 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1264 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1265 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1268 while (order <= e4b->bd_blkbits + 1) {
1270 if (!mb_test_bit(block, bb)) {
1271 /* this block is part of buddy of order 'order' */
1274 bb += 1 << (e4b->bd_blkbits - order);
1280 static void mb_clear_bits(void *bm, int cur, int len)
1286 if ((cur & 31) == 0 && (len - cur) >= 32) {
1287 /* fast path: clear whole word at once */
1288 addr = bm + (cur >> 3);
1293 mb_clear_bit(cur, bm);
1298 /* clear bits in given range
1299 * will return first found zero bit if any, -1 otherwise
1301 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1308 if ((cur & 31) == 0 && (len - cur) >= 32) {
1309 /* fast path: clear whole word at once */
1310 addr = bm + (cur >> 3);
1311 if (*addr != (__u32)(-1) && zero_bit == -1)
1312 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1317 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1325 void ext4_set_bits(void *bm, int cur, int len)
1331 if ((cur & 31) == 0 && (len - cur) >= 32) {
1332 /* fast path: set whole word at once */
1333 addr = bm + (cur >> 3);
1338 mb_set_bit(cur, bm);
1344 * _________________________________________________________________ */
1346 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1348 if (mb_test_bit(*bit + side, bitmap)) {
1349 mb_clear_bit(*bit, bitmap);
1355 mb_set_bit(*bit, bitmap);
1360 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1364 void *buddy = mb_find_buddy(e4b, order, &max);
1369 /* Bits in range [first; last] are known to be set since
1370 * corresponding blocks were allocated. Bits in range
1371 * (first; last) will stay set because they form buddies on
1372 * upper layer. We just deal with borders if they don't
1373 * align with upper layer and then go up.
1374 * Releasing entire group is all about clearing
1375 * single bit of highest order buddy.
1379 * ---------------------------------
1381 * ---------------------------------
1382 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1383 * ---------------------------------
1385 * \_____________________/
1387 * Neither [1] nor [6] is aligned to above layer.
1388 * Left neighbour [0] is free, so mark it busy,
1389 * decrease bb_counters and extend range to
1391 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1392 * mark [6] free, increase bb_counters and shrink range to
1394 * Then shift range to [0; 2], go up and do the same.
1399 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1401 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1406 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1407 mb_clear_bits(buddy, first, last - first + 1);
1408 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1417 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1418 int first, int count)
1420 int left_is_free = 0;
1421 int right_is_free = 0;
1423 int last = first + count - 1;
1424 struct super_block *sb = e4b->bd_sb;
1426 if (WARN_ON(count == 0))
1428 BUG_ON(last >= (sb->s_blocksize << 3));
1429 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1430 /* Don't bother if the block group is corrupt. */
1431 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1434 mb_check_buddy(e4b);
1435 mb_free_blocks_double(inode, e4b, first, count);
1437 e4b->bd_info->bb_free += count;
1438 if (first < e4b->bd_info->bb_first_free)
1439 e4b->bd_info->bb_first_free = first;
1441 /* access memory sequentially: check left neighbour,
1442 * clear range and then check right neighbour
1445 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1446 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1447 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1448 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1450 if (unlikely(block != -1)) {
1451 struct ext4_sb_info *sbi = EXT4_SB(sb);
1452 ext4_fsblk_t blocknr;
1454 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1455 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1456 ext4_grp_locked_error(sb, e4b->bd_group,
1457 inode ? inode->i_ino : 0,
1459 "freeing already freed block "
1460 "(bit %u); block bitmap corrupt.",
1462 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1463 percpu_counter_sub(&sbi->s_freeclusters_counter,
1464 e4b->bd_info->bb_free);
1465 /* Mark the block group as corrupt. */
1466 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1467 &e4b->bd_info->bb_state);
1468 mb_regenerate_buddy(e4b);
1472 /* let's maintain fragments counter */
1473 if (left_is_free && right_is_free)
1474 e4b->bd_info->bb_fragments--;
1475 else if (!left_is_free && !right_is_free)
1476 e4b->bd_info->bb_fragments++;
1478 /* buddy[0] == bd_bitmap is a special case, so handle
1479 * it right away and let mb_buddy_mark_free stay free of
1480 * zero order checks.
1481 * Check if neighbours are to be coaleasced,
1482 * adjust bitmap bb_counters and borders appropriately.
1485 first += !left_is_free;
1486 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1489 last -= !right_is_free;
1490 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1494 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1497 mb_set_largest_free_order(sb, e4b->bd_info);
1498 mb_check_buddy(e4b);
1501 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1502 int needed, struct ext4_free_extent *ex)
1508 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1511 buddy = mb_find_buddy(e4b, 0, &max);
1512 BUG_ON(buddy == NULL);
1513 BUG_ON(block >= max);
1514 if (mb_test_bit(block, buddy)) {
1521 /* find actual order */
1522 order = mb_find_order_for_block(e4b, block);
1523 block = block >> order;
1525 ex->fe_len = 1 << order;
1526 ex->fe_start = block << order;
1527 ex->fe_group = e4b->bd_group;
1529 /* calc difference from given start */
1530 next = next - ex->fe_start;
1532 ex->fe_start += next;
1534 while (needed > ex->fe_len &&
1535 mb_find_buddy(e4b, order, &max)) {
1537 if (block + 1 >= max)
1540 next = (block + 1) * (1 << order);
1541 if (mb_test_bit(next, e4b->bd_bitmap))
1544 order = mb_find_order_for_block(e4b, next);
1546 block = next >> order;
1547 ex->fe_len += 1 << order;
1550 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1554 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1560 int start = ex->fe_start;
1561 int len = ex->fe_len;
1566 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1567 BUG_ON(e4b->bd_group != ex->fe_group);
1568 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1569 mb_check_buddy(e4b);
1570 mb_mark_used_double(e4b, start, len);
1572 e4b->bd_info->bb_free -= len;
1573 if (e4b->bd_info->bb_first_free == start)
1574 e4b->bd_info->bb_first_free += len;
1576 /* let's maintain fragments counter */
1578 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1579 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1580 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1582 e4b->bd_info->bb_fragments++;
1583 else if (!mlen && !max)
1584 e4b->bd_info->bb_fragments--;
1586 /* let's maintain buddy itself */
1588 ord = mb_find_order_for_block(e4b, start);
1590 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1591 /* the whole chunk may be allocated at once! */
1593 buddy = mb_find_buddy(e4b, ord, &max);
1594 BUG_ON((start >> ord) >= max);
1595 mb_set_bit(start >> ord, buddy);
1596 e4b->bd_info->bb_counters[ord]--;
1603 /* store for history */
1605 ret = len | (ord << 16);
1607 /* we have to split large buddy */
1609 buddy = mb_find_buddy(e4b, ord, &max);
1610 mb_set_bit(start >> ord, buddy);
1611 e4b->bd_info->bb_counters[ord]--;
1614 cur = (start >> ord) & ~1U;
1615 buddy = mb_find_buddy(e4b, ord, &max);
1616 mb_clear_bit(cur, buddy);
1617 mb_clear_bit(cur + 1, buddy);
1618 e4b->bd_info->bb_counters[ord]++;
1619 e4b->bd_info->bb_counters[ord]++;
1621 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1623 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1624 mb_check_buddy(e4b);
1630 * Must be called under group lock!
1632 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1633 struct ext4_buddy *e4b)
1635 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1638 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1639 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1641 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1642 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1643 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1645 /* preallocation can change ac_b_ex, thus we store actually
1646 * allocated blocks for history */
1647 ac->ac_f_ex = ac->ac_b_ex;
1649 ac->ac_status = AC_STATUS_FOUND;
1650 ac->ac_tail = ret & 0xffff;
1651 ac->ac_buddy = ret >> 16;
1654 * take the page reference. We want the page to be pinned
1655 * so that we don't get a ext4_mb_init_cache_call for this
1656 * group until we update the bitmap. That would mean we
1657 * double allocate blocks. The reference is dropped
1658 * in ext4_mb_release_context
1660 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1661 get_page(ac->ac_bitmap_page);
1662 ac->ac_buddy_page = e4b->bd_buddy_page;
1663 get_page(ac->ac_buddy_page);
1664 /* store last allocated for subsequent stream allocation */
1665 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1666 spin_lock(&sbi->s_md_lock);
1667 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1668 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1669 spin_unlock(&sbi->s_md_lock);
1674 * regular allocator, for general purposes allocation
1677 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1678 struct ext4_buddy *e4b,
1681 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1682 struct ext4_free_extent *bex = &ac->ac_b_ex;
1683 struct ext4_free_extent *gex = &ac->ac_g_ex;
1684 struct ext4_free_extent ex;
1687 if (ac->ac_status == AC_STATUS_FOUND)
1690 * We don't want to scan for a whole year
1692 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1693 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1694 ac->ac_status = AC_STATUS_BREAK;
1699 * Haven't found good chunk so far, let's continue
1701 if (bex->fe_len < gex->fe_len)
1704 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1705 && bex->fe_group == e4b->bd_group) {
1706 /* recheck chunk's availability - we don't know
1707 * when it was found (within this lock-unlock
1709 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1710 if (max >= gex->fe_len) {
1711 ext4_mb_use_best_found(ac, e4b);
1718 * The routine checks whether found extent is good enough. If it is,
1719 * then the extent gets marked used and flag is set to the context
1720 * to stop scanning. Otherwise, the extent is compared with the
1721 * previous found extent and if new one is better, then it's stored
1722 * in the context. Later, the best found extent will be used, if
1723 * mballoc can't find good enough extent.
1725 * FIXME: real allocation policy is to be designed yet!
1727 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1728 struct ext4_free_extent *ex,
1729 struct ext4_buddy *e4b)
1731 struct ext4_free_extent *bex = &ac->ac_b_ex;
1732 struct ext4_free_extent *gex = &ac->ac_g_ex;
1734 BUG_ON(ex->fe_len <= 0);
1735 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1736 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1737 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1742 * The special case - take what you catch first
1744 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1746 ext4_mb_use_best_found(ac, e4b);
1751 * Let's check whether the chuck is good enough
1753 if (ex->fe_len == gex->fe_len) {
1755 ext4_mb_use_best_found(ac, e4b);
1760 * If this is first found extent, just store it in the context
1762 if (bex->fe_len == 0) {
1768 * If new found extent is better, store it in the context
1770 if (bex->fe_len < gex->fe_len) {
1771 /* if the request isn't satisfied, any found extent
1772 * larger than previous best one is better */
1773 if (ex->fe_len > bex->fe_len)
1775 } else if (ex->fe_len > gex->fe_len) {
1776 /* if the request is satisfied, then we try to find
1777 * an extent that still satisfy the request, but is
1778 * smaller than previous one */
1779 if (ex->fe_len < bex->fe_len)
1783 ext4_mb_check_limits(ac, e4b, 0);
1786 static noinline_for_stack
1787 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1788 struct ext4_buddy *e4b)
1790 struct ext4_free_extent ex = ac->ac_b_ex;
1791 ext4_group_t group = ex.fe_group;
1795 BUG_ON(ex.fe_len <= 0);
1796 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1800 ext4_lock_group(ac->ac_sb, group);
1801 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1805 ext4_mb_use_best_found(ac, e4b);
1808 ext4_unlock_group(ac->ac_sb, group);
1809 ext4_mb_unload_buddy(e4b);
1814 static noinline_for_stack
1815 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1816 struct ext4_buddy *e4b)
1818 ext4_group_t group = ac->ac_g_ex.fe_group;
1821 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1822 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1823 struct ext4_free_extent ex;
1825 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1827 if (grp->bb_free == 0)
1830 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1834 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1835 ext4_mb_unload_buddy(e4b);
1839 ext4_lock_group(ac->ac_sb, group);
1840 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1841 ac->ac_g_ex.fe_len, &ex);
1842 ex.fe_logical = 0xDEADFA11; /* debug value */
1844 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1847 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1849 /* use do_div to get remainder (would be 64-bit modulo) */
1850 if (do_div(start, sbi->s_stripe) == 0) {
1853 ext4_mb_use_best_found(ac, e4b);
1855 } else if (max >= ac->ac_g_ex.fe_len) {
1856 BUG_ON(ex.fe_len <= 0);
1857 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1858 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1861 ext4_mb_use_best_found(ac, e4b);
1862 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1863 /* Sometimes, caller may want to merge even small
1864 * number of blocks to an existing extent */
1865 BUG_ON(ex.fe_len <= 0);
1866 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1867 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1870 ext4_mb_use_best_found(ac, e4b);
1872 ext4_unlock_group(ac->ac_sb, group);
1873 ext4_mb_unload_buddy(e4b);
1879 * The routine scans buddy structures (not bitmap!) from given order
1880 * to max order and tries to find big enough chunk to satisfy the req
1882 static noinline_for_stack
1883 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1884 struct ext4_buddy *e4b)
1886 struct super_block *sb = ac->ac_sb;
1887 struct ext4_group_info *grp = e4b->bd_info;
1893 BUG_ON(ac->ac_2order <= 0);
1894 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1895 if (grp->bb_counters[i] == 0)
1898 buddy = mb_find_buddy(e4b, i, &max);
1899 BUG_ON(buddy == NULL);
1901 k = mb_find_next_zero_bit(buddy, max, 0);
1906 ac->ac_b_ex.fe_len = 1 << i;
1907 ac->ac_b_ex.fe_start = k << i;
1908 ac->ac_b_ex.fe_group = e4b->bd_group;
1910 ext4_mb_use_best_found(ac, e4b);
1912 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1914 if (EXT4_SB(sb)->s_mb_stats)
1915 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1922 * The routine scans the group and measures all found extents.
1923 * In order to optimize scanning, caller must pass number of
1924 * free blocks in the group, so the routine can know upper limit.
1926 static noinline_for_stack
1927 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1928 struct ext4_buddy *e4b)
1930 struct super_block *sb = ac->ac_sb;
1931 void *bitmap = e4b->bd_bitmap;
1932 struct ext4_free_extent ex;
1936 free = e4b->bd_info->bb_free;
1939 i = e4b->bd_info->bb_first_free;
1941 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1942 i = mb_find_next_zero_bit(bitmap,
1943 EXT4_CLUSTERS_PER_GROUP(sb), i);
1944 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1946 * IF we have corrupt bitmap, we won't find any
1947 * free blocks even though group info says we
1948 * we have free blocks
1950 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1951 "%d free clusters as per "
1952 "group info. But bitmap says 0",
1957 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1958 BUG_ON(ex.fe_len <= 0);
1959 if (free < ex.fe_len) {
1960 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1961 "%d free clusters as per "
1962 "group info. But got %d blocks",
1965 * The number of free blocks differs. This mostly
1966 * indicate that the bitmap is corrupt. So exit
1967 * without claiming the space.
1971 ex.fe_logical = 0xDEADC0DE; /* debug value */
1972 ext4_mb_measure_extent(ac, &ex, e4b);
1978 ext4_mb_check_limits(ac, e4b, 1);
1982 * This is a special case for storages like raid5
1983 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1985 static noinline_for_stack
1986 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1987 struct ext4_buddy *e4b)
1989 struct super_block *sb = ac->ac_sb;
1990 struct ext4_sb_info *sbi = EXT4_SB(sb);
1991 void *bitmap = e4b->bd_bitmap;
1992 struct ext4_free_extent ex;
1993 ext4_fsblk_t first_group_block;
1998 BUG_ON(sbi->s_stripe == 0);
2000 /* find first stripe-aligned block in group */
2001 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2003 a = first_group_block + sbi->s_stripe - 1;
2004 do_div(a, sbi->s_stripe);
2005 i = (a * sbi->s_stripe) - first_group_block;
2007 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2008 if (!mb_test_bit(i, bitmap)) {
2009 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2010 if (max >= sbi->s_stripe) {
2012 ex.fe_logical = 0xDEADF00D; /* debug value */
2014 ext4_mb_use_best_found(ac, e4b);
2023 * This is now called BEFORE we load the buddy bitmap.
2024 * Returns either 1 or 0 indicating that the group is either suitable
2025 * for the allocation or not. In addition it can also return negative
2026 * error code when something goes wrong.
2028 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2029 ext4_group_t group, int cr)
2031 unsigned free, fragments;
2032 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2033 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2035 BUG_ON(cr < 0 || cr >= 4);
2037 free = grp->bb_free;
2040 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2043 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2046 /* We only do this if the grp has never been initialized */
2047 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2048 int ret = ext4_mb_init_group(ac->ac_sb, group);
2053 fragments = grp->bb_fragments;
2059 BUG_ON(ac->ac_2order == 0);
2061 /* Avoid using the first bg of a flexgroup for data files */
2062 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2063 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2064 ((group % flex_size) == 0))
2067 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2068 (free / fragments) >= ac->ac_g_ex.fe_len)
2071 if (grp->bb_largest_free_order < ac->ac_2order)
2076 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2080 if (free >= ac->ac_g_ex.fe_len)
2092 static noinline_for_stack int
2093 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2095 ext4_group_t ngroups, group, i;
2097 int err = 0, first_err = 0;
2098 struct ext4_sb_info *sbi;
2099 struct super_block *sb;
2100 struct ext4_buddy e4b;
2104 ngroups = ext4_get_groups_count(sb);
2105 /* non-extent files are limited to low blocks/groups */
2106 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2107 ngroups = sbi->s_blockfile_groups;
2109 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2111 /* first, try the goal */
2112 err = ext4_mb_find_by_goal(ac, &e4b);
2113 if (err || ac->ac_status == AC_STATUS_FOUND)
2116 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2120 * ac->ac2_order is set only if the fe_len is a power of 2
2121 * if ac2_order is set we also set criteria to 0 so that we
2122 * try exact allocation using buddy.
2124 i = fls(ac->ac_g_ex.fe_len);
2127 * We search using buddy data only if the order of the request
2128 * is greater than equal to the sbi_s_mb_order2_reqs
2129 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2131 if (i >= sbi->s_mb_order2_reqs) {
2133 * This should tell if fe_len is exactly power of 2
2135 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2136 ac->ac_2order = i - 1;
2139 /* if stream allocation is enabled, use global goal */
2140 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2141 /* TBD: may be hot point */
2142 spin_lock(&sbi->s_md_lock);
2143 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2144 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2145 spin_unlock(&sbi->s_md_lock);
2148 /* Let's just scan groups to find more-less suitable blocks */
2149 cr = ac->ac_2order ? 0 : 1;
2151 * cr == 0 try to get exact allocation,
2152 * cr == 3 try to get anything
2155 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2156 ac->ac_criteria = cr;
2158 * searching for the right group start
2159 * from the goal value specified
2161 group = ac->ac_g_ex.fe_group;
2163 for (i = 0; i < ngroups; group++, i++) {
2167 * Artificially restricted ngroups for non-extent
2168 * files makes group > ngroups possible on first loop.
2170 if (group >= ngroups)
2173 /* This now checks without needing the buddy page */
2174 ret = ext4_mb_good_group(ac, group, cr);
2181 err = ext4_mb_load_buddy(sb, group, &e4b);
2185 ext4_lock_group(sb, group);
2188 * We need to check again after locking the
2191 ret = ext4_mb_good_group(ac, group, cr);
2193 ext4_unlock_group(sb, group);
2194 ext4_mb_unload_buddy(&e4b);
2200 ac->ac_groups_scanned++;
2201 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2202 ext4_mb_simple_scan_group(ac, &e4b);
2203 else if (cr == 1 && sbi->s_stripe &&
2204 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2205 ext4_mb_scan_aligned(ac, &e4b);
2207 ext4_mb_complex_scan_group(ac, &e4b);
2209 ext4_unlock_group(sb, group);
2210 ext4_mb_unload_buddy(&e4b);
2212 if (ac->ac_status != AC_STATUS_CONTINUE)
2217 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2218 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2220 * We've been searching too long. Let's try to allocate
2221 * the best chunk we've found so far
2224 ext4_mb_try_best_found(ac, &e4b);
2225 if (ac->ac_status != AC_STATUS_FOUND) {
2227 * Someone more lucky has already allocated it.
2228 * The only thing we can do is just take first
2230 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2232 ac->ac_b_ex.fe_group = 0;
2233 ac->ac_b_ex.fe_start = 0;
2234 ac->ac_b_ex.fe_len = 0;
2235 ac->ac_status = AC_STATUS_CONTINUE;
2236 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2238 atomic_inc(&sbi->s_mb_lost_chunks);
2243 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2248 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2250 struct super_block *sb = seq->private;
2253 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2256 return (void *) ((unsigned long) group);
2259 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2261 struct super_block *sb = seq->private;
2265 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2268 return (void *) ((unsigned long) group);
2271 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2273 struct super_block *sb = seq->private;
2274 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2276 int err, buddy_loaded = 0;
2277 struct ext4_buddy e4b;
2278 struct ext4_group_info *grinfo;
2280 struct ext4_group_info info;
2281 ext4_grpblk_t counters[16];
2286 seq_puts(seq, "#group: free frags first ["
2287 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2288 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]");
2290 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2291 sizeof(struct ext4_group_info);
2292 grinfo = ext4_get_group_info(sb, group);
2293 /* Load the group info in memory only if not already loaded. */
2294 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2295 err = ext4_mb_load_buddy(sb, group, &e4b);
2297 seq_printf(seq, "#%-5u: I/O error\n", group);
2303 memcpy(&sg, ext4_get_group_info(sb, group), i);
2306 ext4_mb_unload_buddy(&e4b);
2308 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2309 sg.info.bb_fragments, sg.info.bb_first_free);
2310 for (i = 0; i <= 13; i++)
2311 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2312 sg.info.bb_counters[i] : 0);
2313 seq_printf(seq, " ]\n");
2318 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2322 static const struct seq_operations ext4_mb_seq_groups_ops = {
2323 .start = ext4_mb_seq_groups_start,
2324 .next = ext4_mb_seq_groups_next,
2325 .stop = ext4_mb_seq_groups_stop,
2326 .show = ext4_mb_seq_groups_show,
2329 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2331 struct super_block *sb = PDE_DATA(inode);
2334 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2336 struct seq_file *m = file->private_data;
2343 const struct file_operations ext4_seq_mb_groups_fops = {
2344 .owner = THIS_MODULE,
2345 .open = ext4_mb_seq_groups_open,
2347 .llseek = seq_lseek,
2348 .release = seq_release,
2351 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2353 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2354 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2361 * Allocate the top-level s_group_info array for the specified number
2364 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2366 struct ext4_sb_info *sbi = EXT4_SB(sb);
2368 struct ext4_group_info ***new_groupinfo;
2370 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2371 EXT4_DESC_PER_BLOCK_BITS(sb);
2372 if (size <= sbi->s_group_info_size)
2375 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2376 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2377 if (!new_groupinfo) {
2378 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2381 if (sbi->s_group_info) {
2382 memcpy(new_groupinfo, sbi->s_group_info,
2383 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2384 kvfree(sbi->s_group_info);
2386 sbi->s_group_info = new_groupinfo;
2387 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2388 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2389 sbi->s_group_info_size);
2393 /* Create and initialize ext4_group_info data for the given group. */
2394 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2395 struct ext4_group_desc *desc)
2399 struct ext4_sb_info *sbi = EXT4_SB(sb);
2400 struct ext4_group_info **meta_group_info;
2401 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2404 * First check if this group is the first of a reserved block.
2405 * If it's true, we have to allocate a new table of pointers
2406 * to ext4_group_info structures
2408 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2409 metalen = sizeof(*meta_group_info) <<
2410 EXT4_DESC_PER_BLOCK_BITS(sb);
2411 meta_group_info = kmalloc(metalen, GFP_NOFS);
2412 if (meta_group_info == NULL) {
2413 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2414 "for a buddy group");
2415 goto exit_meta_group_info;
2417 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2422 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2423 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2425 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2426 if (meta_group_info[i] == NULL) {
2427 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2428 goto exit_group_info;
2430 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2431 &(meta_group_info[i]->bb_state));
2434 * initialize bb_free to be able to skip
2435 * empty groups without initialization
2437 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2438 meta_group_info[i]->bb_free =
2439 ext4_free_clusters_after_init(sb, group, desc);
2441 meta_group_info[i]->bb_free =
2442 ext4_free_group_clusters(sb, desc);
2445 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2446 init_rwsem(&meta_group_info[i]->alloc_sem);
2447 meta_group_info[i]->bb_free_root = RB_ROOT;
2448 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2452 struct buffer_head *bh;
2453 meta_group_info[i]->bb_bitmap =
2454 kmalloc(sb->s_blocksize, GFP_NOFS);
2455 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2456 bh = ext4_read_block_bitmap(sb, group);
2457 BUG_ON(IS_ERR_OR_NULL(bh));
2458 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2467 /* If a meta_group_info table has been allocated, release it now */
2468 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2469 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2470 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2472 exit_meta_group_info:
2474 } /* ext4_mb_add_groupinfo */
2476 static int ext4_mb_init_backend(struct super_block *sb)
2478 ext4_group_t ngroups = ext4_get_groups_count(sb);
2480 struct ext4_sb_info *sbi = EXT4_SB(sb);
2482 struct ext4_group_desc *desc;
2483 struct kmem_cache *cachep;
2485 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2489 sbi->s_buddy_cache = new_inode(sb);
2490 if (sbi->s_buddy_cache == NULL) {
2491 ext4_msg(sb, KERN_ERR, "can't get new inode");
2494 /* To avoid potentially colliding with an valid on-disk inode number,
2495 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2496 * not in the inode hash, so it should never be found by iget(), but
2497 * this will avoid confusion if it ever shows up during debugging. */
2498 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2499 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2500 for (i = 0; i < ngroups; i++) {
2501 desc = ext4_get_group_desc(sb, i, NULL);
2503 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2506 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2513 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2515 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2516 i = sbi->s_group_info_size;
2518 kfree(sbi->s_group_info[i]);
2519 iput(sbi->s_buddy_cache);
2521 kvfree(sbi->s_group_info);
2525 static void ext4_groupinfo_destroy_slabs(void)
2529 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2530 if (ext4_groupinfo_caches[i])
2531 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2532 ext4_groupinfo_caches[i] = NULL;
2536 static int ext4_groupinfo_create_slab(size_t size)
2538 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2540 int blocksize_bits = order_base_2(size);
2541 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2542 struct kmem_cache *cachep;
2544 if (cache_index >= NR_GRPINFO_CACHES)
2547 if (unlikely(cache_index < 0))
2550 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2551 if (ext4_groupinfo_caches[cache_index]) {
2552 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2553 return 0; /* Already created */
2556 slab_size = offsetof(struct ext4_group_info,
2557 bb_counters[blocksize_bits + 2]);
2559 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2560 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2563 ext4_groupinfo_caches[cache_index] = cachep;
2565 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2568 "EXT4-fs: no memory for groupinfo slab cache\n");
2575 int ext4_mb_init(struct super_block *sb)
2577 struct ext4_sb_info *sbi = EXT4_SB(sb);
2583 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2585 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2586 if (sbi->s_mb_offsets == NULL) {
2591 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2592 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2593 if (sbi->s_mb_maxs == NULL) {
2598 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2602 /* order 0 is regular bitmap */
2603 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2604 sbi->s_mb_offsets[0] = 0;
2608 max = sb->s_blocksize << 2;
2610 sbi->s_mb_offsets[i] = offset;
2611 sbi->s_mb_maxs[i] = max;
2612 offset += 1 << (sb->s_blocksize_bits - i);
2615 } while (i <= sb->s_blocksize_bits + 1);
2617 spin_lock_init(&sbi->s_md_lock);
2618 spin_lock_init(&sbi->s_bal_lock);
2620 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2621 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2622 sbi->s_mb_stats = MB_DEFAULT_STATS;
2623 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2624 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2626 * The default group preallocation is 512, which for 4k block
2627 * sizes translates to 2 megabytes. However for bigalloc file
2628 * systems, this is probably too big (i.e, if the cluster size
2629 * is 1 megabyte, then group preallocation size becomes half a
2630 * gigabyte!). As a default, we will keep a two megabyte
2631 * group pralloc size for cluster sizes up to 64k, and after
2632 * that, we will force a minimum group preallocation size of
2633 * 32 clusters. This translates to 8 megs when the cluster
2634 * size is 256k, and 32 megs when the cluster size is 1 meg,
2635 * which seems reasonable as a default.
2637 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2638 sbi->s_cluster_bits, 32);
2640 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2641 * to the lowest multiple of s_stripe which is bigger than
2642 * the s_mb_group_prealloc as determined above. We want
2643 * the preallocation size to be an exact multiple of the
2644 * RAID stripe size so that preallocations don't fragment
2647 if (sbi->s_stripe > 1) {
2648 sbi->s_mb_group_prealloc = roundup(
2649 sbi->s_mb_group_prealloc, sbi->s_stripe);
2652 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2653 if (sbi->s_locality_groups == NULL) {
2657 for_each_possible_cpu(i) {
2658 struct ext4_locality_group *lg;
2659 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2660 mutex_init(&lg->lg_mutex);
2661 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2662 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2663 spin_lock_init(&lg->lg_prealloc_lock);
2666 /* init file for buddy data */
2667 ret = ext4_mb_init_backend(sb);
2669 goto out_free_locality_groups;
2673 out_free_locality_groups:
2674 free_percpu(sbi->s_locality_groups);
2675 sbi->s_locality_groups = NULL;
2677 kfree(sbi->s_mb_offsets);
2678 sbi->s_mb_offsets = NULL;
2679 kfree(sbi->s_mb_maxs);
2680 sbi->s_mb_maxs = NULL;
2684 /* need to called with the ext4 group lock held */
2685 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2687 struct ext4_prealloc_space *pa;
2688 struct list_head *cur, *tmp;
2691 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2692 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2693 list_del(&pa->pa_group_list);
2695 kmem_cache_free(ext4_pspace_cachep, pa);
2698 mb_debug(1, "mballoc: %u PAs left\n", count);
2702 int ext4_mb_release(struct super_block *sb)
2704 ext4_group_t ngroups = ext4_get_groups_count(sb);
2706 int num_meta_group_infos;
2707 struct ext4_group_info *grinfo;
2708 struct ext4_sb_info *sbi = EXT4_SB(sb);
2709 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2711 if (sbi->s_group_info) {
2712 for (i = 0; i < ngroups; i++) {
2713 grinfo = ext4_get_group_info(sb, i);
2715 kfree(grinfo->bb_bitmap);
2717 ext4_lock_group(sb, i);
2718 ext4_mb_cleanup_pa(grinfo);
2719 ext4_unlock_group(sb, i);
2720 kmem_cache_free(cachep, grinfo);
2722 num_meta_group_infos = (ngroups +
2723 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2724 EXT4_DESC_PER_BLOCK_BITS(sb);
2725 for (i = 0; i < num_meta_group_infos; i++)
2726 kfree(sbi->s_group_info[i]);
2727 kvfree(sbi->s_group_info);
2729 kfree(sbi->s_mb_offsets);
2730 kfree(sbi->s_mb_maxs);
2731 iput(sbi->s_buddy_cache);
2732 if (sbi->s_mb_stats) {
2733 ext4_msg(sb, KERN_INFO,
2734 "mballoc: %u blocks %u reqs (%u success)",
2735 atomic_read(&sbi->s_bal_allocated),
2736 atomic_read(&sbi->s_bal_reqs),
2737 atomic_read(&sbi->s_bal_success));
2738 ext4_msg(sb, KERN_INFO,
2739 "mballoc: %u extents scanned, %u goal hits, "
2740 "%u 2^N hits, %u breaks, %u lost",
2741 atomic_read(&sbi->s_bal_ex_scanned),
2742 atomic_read(&sbi->s_bal_goals),
2743 atomic_read(&sbi->s_bal_2orders),
2744 atomic_read(&sbi->s_bal_breaks),
2745 atomic_read(&sbi->s_mb_lost_chunks));
2746 ext4_msg(sb, KERN_INFO,
2747 "mballoc: %lu generated and it took %Lu",
2748 sbi->s_mb_buddies_generated,
2749 sbi->s_mb_generation_time);
2750 ext4_msg(sb, KERN_INFO,
2751 "mballoc: %u preallocated, %u discarded",
2752 atomic_read(&sbi->s_mb_preallocated),
2753 atomic_read(&sbi->s_mb_discarded));
2756 free_percpu(sbi->s_locality_groups);
2761 static inline int ext4_issue_discard(struct super_block *sb,
2762 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
2763 unsigned long flags)
2765 ext4_fsblk_t discard_block;
2767 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2768 ext4_group_first_block_no(sb, block_group));
2769 count = EXT4_C2B(EXT4_SB(sb), count);
2770 trace_ext4_discard_blocks(sb,
2771 (unsigned long long) discard_block, count);
2772 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, flags);
2776 * This function is called by the jbd2 layer once the commit has finished,
2777 * so we know we can free the blocks that were released with that commit.
2779 static void ext4_free_data_callback(struct super_block *sb,
2780 struct ext4_journal_cb_entry *jce,
2783 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2784 struct ext4_buddy e4b;
2785 struct ext4_group_info *db;
2786 int err, count = 0, count2 = 0;
2788 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2789 entry->efd_count, entry->efd_group, entry);
2791 if (test_opt(sb, DISCARD)) {
2792 err = ext4_issue_discard(sb, entry->efd_group,
2793 entry->efd_start_cluster,
2794 entry->efd_count, 0);
2795 if (err && err != -EOPNOTSUPP)
2796 ext4_msg(sb, KERN_WARNING, "discard request in"
2797 " group:%d block:%d count:%d failed"
2798 " with %d", entry->efd_group,
2799 entry->efd_start_cluster,
2800 entry->efd_count, err);
2803 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2804 /* we expect to find existing buddy because it's pinned */
2809 /* there are blocks to put in buddy to make them really free */
2810 count += entry->efd_count;
2812 ext4_lock_group(sb, entry->efd_group);
2813 /* Take it out of per group rb tree */
2814 rb_erase(&entry->efd_node, &(db->bb_free_root));
2815 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2818 * Clear the trimmed flag for the group so that the next
2819 * ext4_trim_fs can trim it.
2820 * If the volume is mounted with -o discard, online discard
2821 * is supported and the free blocks will be trimmed online.
2823 if (!test_opt(sb, DISCARD))
2824 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2826 if (!db->bb_free_root.rb_node) {
2827 /* No more items in the per group rb tree
2828 * balance refcounts from ext4_mb_free_metadata()
2830 page_cache_release(e4b.bd_buddy_page);
2831 page_cache_release(e4b.bd_bitmap_page);
2833 ext4_unlock_group(sb, entry->efd_group);
2834 kmem_cache_free(ext4_free_data_cachep, entry);
2835 ext4_mb_unload_buddy(&e4b);
2837 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2840 int __init ext4_init_mballoc(void)
2842 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2843 SLAB_RECLAIM_ACCOUNT);
2844 if (ext4_pspace_cachep == NULL)
2847 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2848 SLAB_RECLAIM_ACCOUNT);
2849 if (ext4_ac_cachep == NULL) {
2850 kmem_cache_destroy(ext4_pspace_cachep);
2854 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2855 SLAB_RECLAIM_ACCOUNT);
2856 if (ext4_free_data_cachep == NULL) {
2857 kmem_cache_destroy(ext4_pspace_cachep);
2858 kmem_cache_destroy(ext4_ac_cachep);
2864 void ext4_exit_mballoc(void)
2867 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2868 * before destroying the slab cache.
2871 kmem_cache_destroy(ext4_pspace_cachep);
2872 kmem_cache_destroy(ext4_ac_cachep);
2873 kmem_cache_destroy(ext4_free_data_cachep);
2874 ext4_groupinfo_destroy_slabs();
2879 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2880 * Returns 0 if success or error code
2882 static noinline_for_stack int
2883 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2884 handle_t *handle, unsigned int reserv_clstrs)
2886 struct buffer_head *bitmap_bh = NULL;
2887 struct ext4_group_desc *gdp;
2888 struct buffer_head *gdp_bh;
2889 struct ext4_sb_info *sbi;
2890 struct super_block *sb;
2894 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2895 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2900 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2901 if (IS_ERR(bitmap_bh)) {
2902 err = PTR_ERR(bitmap_bh);
2907 BUFFER_TRACE(bitmap_bh, "getting write access");
2908 err = ext4_journal_get_write_access(handle, bitmap_bh);
2913 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2917 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2918 ext4_free_group_clusters(sb, gdp));
2920 BUFFER_TRACE(gdp_bh, "get_write_access");
2921 err = ext4_journal_get_write_access(handle, gdp_bh);
2925 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2927 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2928 if (!ext4_data_block_valid(sbi, block, len)) {
2929 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2930 "fs metadata", block, block+len);
2931 /* File system mounted not to panic on error
2932 * Fix the bitmap and repeat the block allocation
2933 * We leak some of the blocks here.
2935 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2936 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2937 ac->ac_b_ex.fe_len);
2938 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2939 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2945 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2946 #ifdef AGGRESSIVE_CHECK
2949 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2950 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2951 bitmap_bh->b_data));
2955 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2956 ac->ac_b_ex.fe_len);
2957 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2958 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2959 ext4_free_group_clusters_set(sb, gdp,
2960 ext4_free_clusters_after_init(sb,
2961 ac->ac_b_ex.fe_group, gdp));
2963 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2964 ext4_free_group_clusters_set(sb, gdp, len);
2965 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2966 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2968 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2969 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2971 * Now reduce the dirty block count also. Should not go negative
2973 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2974 /* release all the reserved blocks if non delalloc */
2975 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2978 if (sbi->s_log_groups_per_flex) {
2979 ext4_group_t flex_group = ext4_flex_group(sbi,
2980 ac->ac_b_ex.fe_group);
2981 atomic64_sub(ac->ac_b_ex.fe_len,
2982 &sbi->s_flex_groups[flex_group].free_clusters);
2985 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2988 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2996 * here we normalize request for locality group
2997 * Group request are normalized to s_mb_group_prealloc, which goes to
2998 * s_strip if we set the same via mount option.
2999 * s_mb_group_prealloc can be configured via
3000 * /sys/fs/ext4/<partition>/mb_group_prealloc
3002 * XXX: should we try to preallocate more than the group has now?
3004 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3006 struct super_block *sb = ac->ac_sb;
3007 struct ext4_locality_group *lg = ac->ac_lg;
3010 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3011 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3012 current->pid, ac->ac_g_ex.fe_len);
3016 * Normalization means making request better in terms of
3017 * size and alignment
3019 static noinline_for_stack void
3020 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3021 struct ext4_allocation_request *ar)
3023 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3026 loff_t size, start_off;
3027 loff_t orig_size __maybe_unused;
3029 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3030 struct ext4_prealloc_space *pa;
3032 /* do normalize only data requests, metadata requests
3033 do not need preallocation */
3034 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3037 /* sometime caller may want exact blocks */
3038 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3041 /* caller may indicate that preallocation isn't
3042 * required (it's a tail, for example) */
3043 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3046 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3047 ext4_mb_normalize_group_request(ac);
3051 bsbits = ac->ac_sb->s_blocksize_bits;
3053 /* first, let's learn actual file size
3054 * given current request is allocated */
3055 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3056 size = size << bsbits;
3057 if (size < i_size_read(ac->ac_inode))
3058 size = i_size_read(ac->ac_inode);
3061 /* max size of free chunks */
3064 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3065 (req <= (size) || max <= (chunk_size))
3067 /* first, try to predict filesize */
3068 /* XXX: should this table be tunable? */
3070 if (size <= 16 * 1024) {
3072 } else if (size <= 32 * 1024) {
3074 } else if (size <= 64 * 1024) {
3076 } else if (size <= 128 * 1024) {
3078 } else if (size <= 256 * 1024) {
3080 } else if (size <= 512 * 1024) {
3082 } else if (size <= 1024 * 1024) {
3084 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3085 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3086 (21 - bsbits)) << 21;
3087 size = 2 * 1024 * 1024;
3088 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3089 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3090 (22 - bsbits)) << 22;
3091 size = 4 * 1024 * 1024;
3092 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3093 (8<<20)>>bsbits, max, 8 * 1024)) {
3094 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3095 (23 - bsbits)) << 23;
3096 size = 8 * 1024 * 1024;
3098 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3099 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3100 ac->ac_o_ex.fe_len) << bsbits;
3102 size = size >> bsbits;
3103 start = start_off >> bsbits;
3105 /* don't cover already allocated blocks in selected range */
3106 if (ar->pleft && start <= ar->lleft) {
3107 size -= ar->lleft + 1 - start;
3108 start = ar->lleft + 1;
3110 if (ar->pright && start + size - 1 >= ar->lright)
3111 size -= start + size - ar->lright;
3115 /* check we don't cross already preallocated blocks */
3117 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3122 spin_lock(&pa->pa_lock);
3123 if (pa->pa_deleted) {
3124 spin_unlock(&pa->pa_lock);
3128 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3131 /* PA must not overlap original request */
3132 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3133 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3135 /* skip PAs this normalized request doesn't overlap with */
3136 if (pa->pa_lstart >= end || pa_end <= start) {
3137 spin_unlock(&pa->pa_lock);
3140 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3142 /* adjust start or end to be adjacent to this pa */
3143 if (pa_end <= ac->ac_o_ex.fe_logical) {
3144 BUG_ON(pa_end < start);
3146 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3147 BUG_ON(pa->pa_lstart > end);
3148 end = pa->pa_lstart;
3150 spin_unlock(&pa->pa_lock);
3155 /* XXX: extra loop to check we really don't overlap preallocations */
3157 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3160 spin_lock(&pa->pa_lock);
3161 if (pa->pa_deleted == 0) {
3162 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3164 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3166 spin_unlock(&pa->pa_lock);
3170 if (start + size <= ac->ac_o_ex.fe_logical &&
3171 start > ac->ac_o_ex.fe_logical) {
3172 ext4_msg(ac->ac_sb, KERN_ERR,
3173 "start %lu, size %lu, fe_logical %lu",
3174 (unsigned long) start, (unsigned long) size,
3175 (unsigned long) ac->ac_o_ex.fe_logical);
3178 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3180 /* now prepare goal request */
3182 /* XXX: is it better to align blocks WRT to logical
3183 * placement or satisfy big request as is */
3184 ac->ac_g_ex.fe_logical = start;
3185 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3187 /* define goal start in order to merge */
3188 if (ar->pright && (ar->lright == (start + size))) {
3189 /* merge to the right */
3190 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3191 &ac->ac_f_ex.fe_group,
3192 &ac->ac_f_ex.fe_start);
3193 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3195 if (ar->pleft && (ar->lleft + 1 == start)) {
3196 /* merge to the left */
3197 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3198 &ac->ac_f_ex.fe_group,
3199 &ac->ac_f_ex.fe_start);
3200 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3203 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3204 (unsigned) orig_size, (unsigned) start);
3207 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3209 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3211 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3212 atomic_inc(&sbi->s_bal_reqs);
3213 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3214 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3215 atomic_inc(&sbi->s_bal_success);
3216 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3217 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3218 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3219 atomic_inc(&sbi->s_bal_goals);
3220 if (ac->ac_found > sbi->s_mb_max_to_scan)
3221 atomic_inc(&sbi->s_bal_breaks);
3224 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3225 trace_ext4_mballoc_alloc(ac);
3227 trace_ext4_mballoc_prealloc(ac);
3231 * Called on failure; free up any blocks from the inode PA for this
3232 * context. We don't need this for MB_GROUP_PA because we only change
3233 * pa_free in ext4_mb_release_context(), but on failure, we've already
3234 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3236 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3238 struct ext4_prealloc_space *pa = ac->ac_pa;
3239 struct ext4_buddy e4b;
3243 if (ac->ac_f_ex.fe_len == 0)
3245 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3248 * This should never happen since we pin the
3249 * pages in the ext4_allocation_context so
3250 * ext4_mb_load_buddy() should never fail.
3252 WARN(1, "mb_load_buddy failed (%d)", err);
3255 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3256 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3257 ac->ac_f_ex.fe_len);
3258 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3259 ext4_mb_unload_buddy(&e4b);
3262 if (pa->pa_type == MB_INODE_PA)
3263 pa->pa_free += ac->ac_b_ex.fe_len;
3267 * use blocks preallocated to inode
3269 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3270 struct ext4_prealloc_space *pa)
3272 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3277 /* found preallocated blocks, use them */
3278 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3279 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3280 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3281 len = EXT4_NUM_B2C(sbi, end - start);
3282 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3283 &ac->ac_b_ex.fe_start);
3284 ac->ac_b_ex.fe_len = len;
3285 ac->ac_status = AC_STATUS_FOUND;
3288 BUG_ON(start < pa->pa_pstart);
3289 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3290 BUG_ON(pa->pa_free < len);
3293 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3297 * use blocks preallocated to locality group
3299 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3300 struct ext4_prealloc_space *pa)
3302 unsigned int len = ac->ac_o_ex.fe_len;
3304 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3305 &ac->ac_b_ex.fe_group,
3306 &ac->ac_b_ex.fe_start);
3307 ac->ac_b_ex.fe_len = len;
3308 ac->ac_status = AC_STATUS_FOUND;
3311 /* we don't correct pa_pstart or pa_plen here to avoid
3312 * possible race when the group is being loaded concurrently
3313 * instead we correct pa later, after blocks are marked
3314 * in on-disk bitmap -- see ext4_mb_release_context()
3315 * Other CPUs are prevented from allocating from this pa by lg_mutex
3317 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3321 * Return the prealloc space that have minimal distance
3322 * from the goal block. @cpa is the prealloc
3323 * space that is having currently known minimal distance
3324 * from the goal block.
3326 static struct ext4_prealloc_space *
3327 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3328 struct ext4_prealloc_space *pa,
3329 struct ext4_prealloc_space *cpa)
3331 ext4_fsblk_t cur_distance, new_distance;
3334 atomic_inc(&pa->pa_count);
3337 cur_distance = abs(goal_block - cpa->pa_pstart);
3338 new_distance = abs(goal_block - pa->pa_pstart);
3340 if (cur_distance <= new_distance)
3343 /* drop the previous reference */
3344 atomic_dec(&cpa->pa_count);
3345 atomic_inc(&pa->pa_count);
3350 * search goal blocks in preallocated space
3352 static noinline_for_stack int
3353 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3355 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3357 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3358 struct ext4_locality_group *lg;
3359 struct ext4_prealloc_space *pa, *cpa = NULL;
3360 ext4_fsblk_t goal_block;
3362 /* only data can be preallocated */
3363 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3366 /* first, try per-file preallocation */
3368 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3370 /* all fields in this condition don't change,
3371 * so we can skip locking for them */
3372 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3373 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3374 EXT4_C2B(sbi, pa->pa_len)))
3377 /* non-extent files can't have physical blocks past 2^32 */
3378 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3379 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3380 EXT4_MAX_BLOCK_FILE_PHYS))
3383 /* found preallocated blocks, use them */
3384 spin_lock(&pa->pa_lock);
3385 if (pa->pa_deleted == 0 && pa->pa_free) {
3386 atomic_inc(&pa->pa_count);
3387 ext4_mb_use_inode_pa(ac, pa);
3388 spin_unlock(&pa->pa_lock);
3389 ac->ac_criteria = 10;
3393 spin_unlock(&pa->pa_lock);
3397 /* can we use group allocation? */
3398 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3401 /* inode may have no locality group for some reason */
3405 order = fls(ac->ac_o_ex.fe_len) - 1;
3406 if (order > PREALLOC_TB_SIZE - 1)
3407 /* The max size of hash table is PREALLOC_TB_SIZE */
3408 order = PREALLOC_TB_SIZE - 1;
3410 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3412 * search for the prealloc space that is having
3413 * minimal distance from the goal block.
3415 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3417 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3419 spin_lock(&pa->pa_lock);
3420 if (pa->pa_deleted == 0 &&
3421 pa->pa_free >= ac->ac_o_ex.fe_len) {
3423 cpa = ext4_mb_check_group_pa(goal_block,
3426 spin_unlock(&pa->pa_lock);
3431 ext4_mb_use_group_pa(ac, cpa);
3432 ac->ac_criteria = 20;
3439 * the function goes through all block freed in the group
3440 * but not yet committed and marks them used in in-core bitmap.
3441 * buddy must be generated from this bitmap
3442 * Need to be called with the ext4 group lock held
3444 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3448 struct ext4_group_info *grp;
3449 struct ext4_free_data *entry;
3451 grp = ext4_get_group_info(sb, group);
3452 n = rb_first(&(grp->bb_free_root));
3455 entry = rb_entry(n, struct ext4_free_data, efd_node);
3456 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3463 * the function goes through all preallocation in this group and marks them
3464 * used in in-core bitmap. buddy must be generated from this bitmap
3465 * Need to be called with ext4 group lock held
3467 static noinline_for_stack
3468 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3471 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3472 struct ext4_prealloc_space *pa;
3473 struct list_head *cur;
3474 ext4_group_t groupnr;
3475 ext4_grpblk_t start;
3476 int preallocated = 0;
3479 /* all form of preallocation discards first load group,
3480 * so the only competing code is preallocation use.
3481 * we don't need any locking here
3482 * notice we do NOT ignore preallocations with pa_deleted
3483 * otherwise we could leave used blocks available for
3484 * allocation in buddy when concurrent ext4_mb_put_pa()
3485 * is dropping preallocation
3487 list_for_each(cur, &grp->bb_prealloc_list) {
3488 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3489 spin_lock(&pa->pa_lock);
3490 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3493 spin_unlock(&pa->pa_lock);
3494 if (unlikely(len == 0))
3496 BUG_ON(groupnr != group);
3497 ext4_set_bits(bitmap, start, len);
3498 preallocated += len;
3500 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3503 static void ext4_mb_pa_callback(struct rcu_head *head)
3505 struct ext4_prealloc_space *pa;
3506 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3508 BUG_ON(atomic_read(&pa->pa_count));
3509 BUG_ON(pa->pa_deleted == 0);
3510 kmem_cache_free(ext4_pspace_cachep, pa);
3514 * drops a reference to preallocated space descriptor
3515 * if this was the last reference and the space is consumed
3517 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3518 struct super_block *sb, struct ext4_prealloc_space *pa)
3521 ext4_fsblk_t grp_blk;
3523 /* in this short window concurrent discard can set pa_deleted */
3524 spin_lock(&pa->pa_lock);
3525 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3526 spin_unlock(&pa->pa_lock);
3530 if (pa->pa_deleted == 1) {
3531 spin_unlock(&pa->pa_lock);
3536 spin_unlock(&pa->pa_lock);
3538 grp_blk = pa->pa_pstart;
3540 * If doing group-based preallocation, pa_pstart may be in the
3541 * next group when pa is used up
3543 if (pa->pa_type == MB_GROUP_PA)
3546 grp = ext4_get_group_number(sb, grp_blk);
3551 * P1 (buddy init) P2 (regular allocation)
3552 * find block B in PA
3553 * copy on-disk bitmap to buddy
3554 * mark B in on-disk bitmap
3555 * drop PA from group
3556 * mark all PAs in buddy
3558 * thus, P1 initializes buddy with B available. to prevent this
3559 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3562 ext4_lock_group(sb, grp);
3563 list_del(&pa->pa_group_list);
3564 ext4_unlock_group(sb, grp);
3566 spin_lock(pa->pa_obj_lock);
3567 list_del_rcu(&pa->pa_inode_list);
3568 spin_unlock(pa->pa_obj_lock);
3570 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3574 * creates new preallocated space for given inode
3576 static noinline_for_stack int
3577 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3579 struct super_block *sb = ac->ac_sb;
3580 struct ext4_sb_info *sbi = EXT4_SB(sb);
3581 struct ext4_prealloc_space *pa;
3582 struct ext4_group_info *grp;
3583 struct ext4_inode_info *ei;
3585 /* preallocate only when found space is larger then requested */
3586 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3587 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3588 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3590 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3594 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3600 /* we can't allocate as much as normalizer wants.
3601 * so, found space must get proper lstart
3602 * to cover original request */
3603 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3604 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3606 /* we're limited by original request in that
3607 * logical block must be covered any way
3608 * winl is window we can move our chunk within */
3609 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3611 /* also, we should cover whole original request */
3612 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3614 /* the smallest one defines real window */
3615 win = min(winl, wins);
3617 offs = ac->ac_o_ex.fe_logical %
3618 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3619 if (offs && offs < win)
3622 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3623 EXT4_NUM_B2C(sbi, win);
3624 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3625 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3628 /* preallocation can change ac_b_ex, thus we store actually
3629 * allocated blocks for history */
3630 ac->ac_f_ex = ac->ac_b_ex;
3632 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3633 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3634 pa->pa_len = ac->ac_b_ex.fe_len;
3635 pa->pa_free = pa->pa_len;
3636 atomic_set(&pa->pa_count, 1);
3637 spin_lock_init(&pa->pa_lock);
3638 INIT_LIST_HEAD(&pa->pa_inode_list);
3639 INIT_LIST_HEAD(&pa->pa_group_list);
3641 pa->pa_type = MB_INODE_PA;
3643 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3644 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3645 trace_ext4_mb_new_inode_pa(ac, pa);
3647 ext4_mb_use_inode_pa(ac, pa);
3648 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3650 ei = EXT4_I(ac->ac_inode);
3651 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3653 pa->pa_obj_lock = &ei->i_prealloc_lock;
3654 pa->pa_inode = ac->ac_inode;
3656 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3657 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3658 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3660 spin_lock(pa->pa_obj_lock);
3661 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3662 spin_unlock(pa->pa_obj_lock);
3668 * creates new preallocated space for locality group inodes belongs to
3670 static noinline_for_stack int
3671 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3673 struct super_block *sb = ac->ac_sb;
3674 struct ext4_locality_group *lg;
3675 struct ext4_prealloc_space *pa;
3676 struct ext4_group_info *grp;
3678 /* preallocate only when found space is larger then requested */
3679 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3680 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3681 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3683 BUG_ON(ext4_pspace_cachep == NULL);
3684 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3688 /* preallocation can change ac_b_ex, thus we store actually
3689 * allocated blocks for history */
3690 ac->ac_f_ex = ac->ac_b_ex;
3692 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3693 pa->pa_lstart = pa->pa_pstart;
3694 pa->pa_len = ac->ac_b_ex.fe_len;
3695 pa->pa_free = pa->pa_len;
3696 atomic_set(&pa->pa_count, 1);
3697 spin_lock_init(&pa->pa_lock);
3698 INIT_LIST_HEAD(&pa->pa_inode_list);
3699 INIT_LIST_HEAD(&pa->pa_group_list);
3701 pa->pa_type = MB_GROUP_PA;
3703 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3704 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3705 trace_ext4_mb_new_group_pa(ac, pa);
3707 ext4_mb_use_group_pa(ac, pa);
3708 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3710 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3714 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3715 pa->pa_inode = NULL;
3717 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3718 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3719 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3722 * We will later add the new pa to the right bucket
3723 * after updating the pa_free in ext4_mb_release_context
3728 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3732 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3733 err = ext4_mb_new_group_pa(ac);
3735 err = ext4_mb_new_inode_pa(ac);
3740 * finds all unused blocks in on-disk bitmap, frees them in
3741 * in-core bitmap and buddy.
3742 * @pa must be unlinked from inode and group lists, so that
3743 * nobody else can find/use it.
3744 * the caller MUST hold group/inode locks.
3745 * TODO: optimize the case when there are no in-core structures yet
3747 static noinline_for_stack int
3748 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3749 struct ext4_prealloc_space *pa)
3751 struct super_block *sb = e4b->bd_sb;
3752 struct ext4_sb_info *sbi = EXT4_SB(sb);
3757 unsigned long long grp_blk_start;
3761 BUG_ON(pa->pa_deleted == 0);
3762 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3763 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3764 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3765 end = bit + pa->pa_len;
3768 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3771 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3772 mb_debug(1, " free preallocated %u/%u in group %u\n",
3773 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3774 (unsigned) next - bit, (unsigned) group);
3777 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3778 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3779 EXT4_C2B(sbi, bit)),
3781 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3784 if (free != pa->pa_free) {
3785 ext4_msg(e4b->bd_sb, KERN_CRIT,
3786 "pa %p: logic %lu, phys. %lu, len %lu",
3787 pa, (unsigned long) pa->pa_lstart,
3788 (unsigned long) pa->pa_pstart,
3789 (unsigned long) pa->pa_len);
3790 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3793 * pa is already deleted so we use the value obtained
3794 * from the bitmap and continue.
3797 atomic_add(free, &sbi->s_mb_discarded);
3802 static noinline_for_stack int
3803 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3804 struct ext4_prealloc_space *pa)
3806 struct super_block *sb = e4b->bd_sb;
3810 trace_ext4_mb_release_group_pa(sb, pa);
3811 BUG_ON(pa->pa_deleted == 0);
3812 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3813 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3814 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3815 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3816 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3822 * releases all preallocations in given group
3824 * first, we need to decide discard policy:
3825 * - when do we discard
3827 * - how many do we discard
3828 * 1) how many requested
3830 static noinline_for_stack int
3831 ext4_mb_discard_group_preallocations(struct super_block *sb,
3832 ext4_group_t group, int needed)
3834 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3835 struct buffer_head *bitmap_bh = NULL;
3836 struct ext4_prealloc_space *pa, *tmp;
3837 struct list_head list;
3838 struct ext4_buddy e4b;
3843 mb_debug(1, "discard preallocation for group %u\n", group);
3845 if (list_empty(&grp->bb_prealloc_list))
3848 bitmap_bh = ext4_read_block_bitmap(sb, group);
3849 if (IS_ERR(bitmap_bh)) {
3850 err = PTR_ERR(bitmap_bh);
3851 ext4_error(sb, "Error %d reading block bitmap for %u",
3856 err = ext4_mb_load_buddy(sb, group, &e4b);
3858 ext4_error(sb, "Error loading buddy information for %u", group);
3864 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3866 INIT_LIST_HEAD(&list);
3868 ext4_lock_group(sb, group);
3869 list_for_each_entry_safe(pa, tmp,
3870 &grp->bb_prealloc_list, pa_group_list) {
3871 spin_lock(&pa->pa_lock);
3872 if (atomic_read(&pa->pa_count)) {
3873 spin_unlock(&pa->pa_lock);
3877 if (pa->pa_deleted) {
3878 spin_unlock(&pa->pa_lock);
3882 /* seems this one can be freed ... */
3885 /* we can trust pa_free ... */
3886 free += pa->pa_free;
3888 spin_unlock(&pa->pa_lock);
3890 list_del(&pa->pa_group_list);
3891 list_add(&pa->u.pa_tmp_list, &list);
3894 /* if we still need more blocks and some PAs were used, try again */
3895 if (free < needed && busy) {
3897 ext4_unlock_group(sb, group);
3902 /* found anything to free? */
3903 if (list_empty(&list)) {
3908 /* now free all selected PAs */
3909 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3911 /* remove from object (inode or locality group) */
3912 spin_lock(pa->pa_obj_lock);
3913 list_del_rcu(&pa->pa_inode_list);
3914 spin_unlock(pa->pa_obj_lock);
3916 if (pa->pa_type == MB_GROUP_PA)
3917 ext4_mb_release_group_pa(&e4b, pa);
3919 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3921 list_del(&pa->u.pa_tmp_list);
3922 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3926 ext4_unlock_group(sb, group);
3927 ext4_mb_unload_buddy(&e4b);
3933 * releases all non-used preallocated blocks for given inode
3935 * It's important to discard preallocations under i_data_sem
3936 * We don't want another block to be served from the prealloc
3937 * space when we are discarding the inode prealloc space.
3939 * FIXME!! Make sure it is valid at all the call sites
3941 void ext4_discard_preallocations(struct inode *inode)
3943 struct ext4_inode_info *ei = EXT4_I(inode);
3944 struct super_block *sb = inode->i_sb;
3945 struct buffer_head *bitmap_bh = NULL;
3946 struct ext4_prealloc_space *pa, *tmp;
3947 ext4_group_t group = 0;
3948 struct list_head list;
3949 struct ext4_buddy e4b;
3952 if (!S_ISREG(inode->i_mode)) {
3953 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3957 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3958 trace_ext4_discard_preallocations(inode);
3960 INIT_LIST_HEAD(&list);
3963 /* first, collect all pa's in the inode */
3964 spin_lock(&ei->i_prealloc_lock);
3965 while (!list_empty(&ei->i_prealloc_list)) {
3966 pa = list_entry(ei->i_prealloc_list.next,
3967 struct ext4_prealloc_space, pa_inode_list);
3968 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3969 spin_lock(&pa->pa_lock);
3970 if (atomic_read(&pa->pa_count)) {
3971 /* this shouldn't happen often - nobody should
3972 * use preallocation while we're discarding it */
3973 spin_unlock(&pa->pa_lock);
3974 spin_unlock(&ei->i_prealloc_lock);
3975 ext4_msg(sb, KERN_ERR,
3976 "uh-oh! used pa while discarding");
3978 schedule_timeout_uninterruptible(HZ);
3982 if (pa->pa_deleted == 0) {
3984 spin_unlock(&pa->pa_lock);
3985 list_del_rcu(&pa->pa_inode_list);
3986 list_add(&pa->u.pa_tmp_list, &list);
3990 /* someone is deleting pa right now */
3991 spin_unlock(&pa->pa_lock);
3992 spin_unlock(&ei->i_prealloc_lock);
3994 /* we have to wait here because pa_deleted
3995 * doesn't mean pa is already unlinked from
3996 * the list. as we might be called from
3997 * ->clear_inode() the inode will get freed
3998 * and concurrent thread which is unlinking
3999 * pa from inode's list may access already
4000 * freed memory, bad-bad-bad */
4002 /* XXX: if this happens too often, we can
4003 * add a flag to force wait only in case
4004 * of ->clear_inode(), but not in case of
4005 * regular truncate */
4006 schedule_timeout_uninterruptible(HZ);
4009 spin_unlock(&ei->i_prealloc_lock);
4011 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4012 BUG_ON(pa->pa_type != MB_INODE_PA);
4013 group = ext4_get_group_number(sb, pa->pa_pstart);
4015 err = ext4_mb_load_buddy(sb, group, &e4b);
4017 ext4_error(sb, "Error loading buddy information for %u",
4022 bitmap_bh = ext4_read_block_bitmap(sb, group);
4023 if (IS_ERR(bitmap_bh)) {
4024 err = PTR_ERR(bitmap_bh);
4025 ext4_error(sb, "Error %d reading block bitmap for %u",
4027 ext4_mb_unload_buddy(&e4b);
4031 ext4_lock_group(sb, group);
4032 list_del(&pa->pa_group_list);
4033 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4034 ext4_unlock_group(sb, group);
4036 ext4_mb_unload_buddy(&e4b);
4039 list_del(&pa->u.pa_tmp_list);
4040 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4044 #ifdef CONFIG_EXT4_DEBUG
4045 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4047 struct super_block *sb = ac->ac_sb;
4048 ext4_group_t ngroups, i;
4050 if (!ext4_mballoc_debug ||
4051 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4054 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4055 " Allocation context details:");
4056 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4057 ac->ac_status, ac->ac_flags);
4058 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4059 "goal %lu/%lu/%lu@%lu, "
4060 "best %lu/%lu/%lu@%lu cr %d",
4061 (unsigned long)ac->ac_o_ex.fe_group,
4062 (unsigned long)ac->ac_o_ex.fe_start,
4063 (unsigned long)ac->ac_o_ex.fe_len,
4064 (unsigned long)ac->ac_o_ex.fe_logical,
4065 (unsigned long)ac->ac_g_ex.fe_group,
4066 (unsigned long)ac->ac_g_ex.fe_start,
4067 (unsigned long)ac->ac_g_ex.fe_len,
4068 (unsigned long)ac->ac_g_ex.fe_logical,
4069 (unsigned long)ac->ac_b_ex.fe_group,
4070 (unsigned long)ac->ac_b_ex.fe_start,
4071 (unsigned long)ac->ac_b_ex.fe_len,
4072 (unsigned long)ac->ac_b_ex.fe_logical,
4073 (int)ac->ac_criteria);
4074 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4075 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4076 ngroups = ext4_get_groups_count(sb);
4077 for (i = 0; i < ngroups; i++) {
4078 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4079 struct ext4_prealloc_space *pa;
4080 ext4_grpblk_t start;
4081 struct list_head *cur;
4082 ext4_lock_group(sb, i);
4083 list_for_each(cur, &grp->bb_prealloc_list) {
4084 pa = list_entry(cur, struct ext4_prealloc_space,
4086 spin_lock(&pa->pa_lock);
4087 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4089 spin_unlock(&pa->pa_lock);
4090 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4093 ext4_unlock_group(sb, i);
4095 if (grp->bb_free == 0)
4097 printk(KERN_ERR "%u: %d/%d \n",
4098 i, grp->bb_free, grp->bb_fragments);
4100 printk(KERN_ERR "\n");
4103 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4110 * We use locality group preallocation for small size file. The size of the
4111 * file is determined by the current size or the resulting size after
4112 * allocation which ever is larger
4114 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4116 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4118 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4119 int bsbits = ac->ac_sb->s_blocksize_bits;
4122 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4125 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4128 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4129 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4132 if ((size == isize) &&
4133 !ext4_fs_is_busy(sbi) &&
4134 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4135 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4139 if (sbi->s_mb_group_prealloc <= 0) {
4140 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4144 /* don't use group allocation for large files */
4145 size = max(size, isize);
4146 if (size > sbi->s_mb_stream_request) {
4147 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4151 BUG_ON(ac->ac_lg != NULL);
4153 * locality group prealloc space are per cpu. The reason for having
4154 * per cpu locality group is to reduce the contention between block
4155 * request from multiple CPUs.
4157 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4159 /* we're going to use group allocation */
4160 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4162 /* serialize all allocations in the group */
4163 mutex_lock(&ac->ac_lg->lg_mutex);
4166 static noinline_for_stack int
4167 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4168 struct ext4_allocation_request *ar)
4170 struct super_block *sb = ar->inode->i_sb;
4171 struct ext4_sb_info *sbi = EXT4_SB(sb);
4172 struct ext4_super_block *es = sbi->s_es;
4176 ext4_grpblk_t block;
4178 /* we can't allocate > group size */
4181 /* just a dirty hack to filter too big requests */
4182 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4183 len = EXT4_CLUSTERS_PER_GROUP(sb);
4185 /* start searching from the goal */
4187 if (goal < le32_to_cpu(es->s_first_data_block) ||
4188 goal >= ext4_blocks_count(es))
4189 goal = le32_to_cpu(es->s_first_data_block);
4190 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4192 /* set up allocation goals */
4193 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4194 ac->ac_status = AC_STATUS_CONTINUE;
4196 ac->ac_inode = ar->inode;
4197 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4198 ac->ac_o_ex.fe_group = group;
4199 ac->ac_o_ex.fe_start = block;
4200 ac->ac_o_ex.fe_len = len;
4201 ac->ac_g_ex = ac->ac_o_ex;
4202 ac->ac_flags = ar->flags;
4204 /* we have to define context: we'll we work with a file or
4205 * locality group. this is a policy, actually */
4206 ext4_mb_group_or_file(ac);
4208 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4209 "left: %u/%u, right %u/%u to %swritable\n",
4210 (unsigned) ar->len, (unsigned) ar->logical,
4211 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4212 (unsigned) ar->lleft, (unsigned) ar->pleft,
4213 (unsigned) ar->lright, (unsigned) ar->pright,
4214 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4219 static noinline_for_stack void
4220 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4221 struct ext4_locality_group *lg,
4222 int order, int total_entries)
4224 ext4_group_t group = 0;
4225 struct ext4_buddy e4b;
4226 struct list_head discard_list;
4227 struct ext4_prealloc_space *pa, *tmp;
4229 mb_debug(1, "discard locality group preallocation\n");
4231 INIT_LIST_HEAD(&discard_list);
4233 spin_lock(&lg->lg_prealloc_lock);
4234 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4236 spin_lock(&pa->pa_lock);
4237 if (atomic_read(&pa->pa_count)) {
4239 * This is the pa that we just used
4240 * for block allocation. So don't
4243 spin_unlock(&pa->pa_lock);
4246 if (pa->pa_deleted) {
4247 spin_unlock(&pa->pa_lock);
4250 /* only lg prealloc space */
4251 BUG_ON(pa->pa_type != MB_GROUP_PA);
4253 /* seems this one can be freed ... */
4255 spin_unlock(&pa->pa_lock);
4257 list_del_rcu(&pa->pa_inode_list);
4258 list_add(&pa->u.pa_tmp_list, &discard_list);
4261 if (total_entries <= 5) {
4263 * we want to keep only 5 entries
4264 * allowing it to grow to 8. This
4265 * mak sure we don't call discard
4266 * soon for this list.
4271 spin_unlock(&lg->lg_prealloc_lock);
4273 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4275 group = ext4_get_group_number(sb, pa->pa_pstart);
4276 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4277 ext4_error(sb, "Error loading buddy information for %u",
4281 ext4_lock_group(sb, group);
4282 list_del(&pa->pa_group_list);
4283 ext4_mb_release_group_pa(&e4b, pa);
4284 ext4_unlock_group(sb, group);
4286 ext4_mb_unload_buddy(&e4b);
4287 list_del(&pa->u.pa_tmp_list);
4288 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4293 * We have incremented pa_count. So it cannot be freed at this
4294 * point. Also we hold lg_mutex. So no parallel allocation is
4295 * possible from this lg. That means pa_free cannot be updated.
4297 * A parallel ext4_mb_discard_group_preallocations is possible.
4298 * which can cause the lg_prealloc_list to be updated.
4301 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4303 int order, added = 0, lg_prealloc_count = 1;
4304 struct super_block *sb = ac->ac_sb;
4305 struct ext4_locality_group *lg = ac->ac_lg;
4306 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4308 order = fls(pa->pa_free) - 1;
4309 if (order > PREALLOC_TB_SIZE - 1)
4310 /* The max size of hash table is PREALLOC_TB_SIZE */
4311 order = PREALLOC_TB_SIZE - 1;
4312 /* Add the prealloc space to lg */
4313 spin_lock(&lg->lg_prealloc_lock);
4314 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4316 spin_lock(&tmp_pa->pa_lock);
4317 if (tmp_pa->pa_deleted) {
4318 spin_unlock(&tmp_pa->pa_lock);
4321 if (!added && pa->pa_free < tmp_pa->pa_free) {
4322 /* Add to the tail of the previous entry */
4323 list_add_tail_rcu(&pa->pa_inode_list,
4324 &tmp_pa->pa_inode_list);
4327 * we want to count the total
4328 * number of entries in the list
4331 spin_unlock(&tmp_pa->pa_lock);
4332 lg_prealloc_count++;
4335 list_add_tail_rcu(&pa->pa_inode_list,
4336 &lg->lg_prealloc_list[order]);
4337 spin_unlock(&lg->lg_prealloc_lock);
4339 /* Now trim the list to be not more than 8 elements */
4340 if (lg_prealloc_count > 8) {
4341 ext4_mb_discard_lg_preallocations(sb, lg,
4342 order, lg_prealloc_count);
4349 * release all resource we used in allocation
4351 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4353 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4354 struct ext4_prealloc_space *pa = ac->ac_pa;
4356 if (pa->pa_type == MB_GROUP_PA) {
4357 /* see comment in ext4_mb_use_group_pa() */
4358 spin_lock(&pa->pa_lock);
4359 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4360 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4361 pa->pa_free -= ac->ac_b_ex.fe_len;
4362 pa->pa_len -= ac->ac_b_ex.fe_len;
4363 spin_unlock(&pa->pa_lock);
4368 * We want to add the pa to the right bucket.
4369 * Remove it from the list and while adding
4370 * make sure the list to which we are adding
4373 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4374 spin_lock(pa->pa_obj_lock);
4375 list_del_rcu(&pa->pa_inode_list);
4376 spin_unlock(pa->pa_obj_lock);
4377 ext4_mb_add_n_trim(ac);
4379 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4381 if (ac->ac_bitmap_page)
4382 page_cache_release(ac->ac_bitmap_page);
4383 if (ac->ac_buddy_page)
4384 page_cache_release(ac->ac_buddy_page);
4385 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4386 mutex_unlock(&ac->ac_lg->lg_mutex);
4387 ext4_mb_collect_stats(ac);
4391 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4393 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4397 trace_ext4_mb_discard_preallocations(sb, needed);
4398 for (i = 0; i < ngroups && needed > 0; i++) {
4399 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4408 * Main entry point into mballoc to allocate blocks
4409 * it tries to use preallocation first, then falls back
4410 * to usual allocation
4412 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4413 struct ext4_allocation_request *ar, int *errp)
4416 struct ext4_allocation_context *ac = NULL;
4417 struct ext4_sb_info *sbi;
4418 struct super_block *sb;
4419 ext4_fsblk_t block = 0;
4420 unsigned int inquota = 0;
4421 unsigned int reserv_clstrs = 0;
4424 sb = ar->inode->i_sb;
4427 trace_ext4_request_blocks(ar);
4429 /* Allow to use superuser reservation for quota file */
4430 if (IS_NOQUOTA(ar->inode))
4431 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4433 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4434 /* Without delayed allocation we need to verify
4435 * there is enough free blocks to do block allocation
4436 * and verify allocation doesn't exceed the quota limits.
4439 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4441 /* let others to free the space */
4443 ar->len = ar->len >> 1;
4449 reserv_clstrs = ar->len;
4450 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4451 dquot_alloc_block_nofail(ar->inode,
4452 EXT4_C2B(sbi, ar->len));
4455 dquot_alloc_block(ar->inode,
4456 EXT4_C2B(sbi, ar->len))) {
4458 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4469 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4476 *errp = ext4_mb_initialize_context(ac, ar);
4482 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4483 if (!ext4_mb_use_preallocated(ac)) {
4484 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4485 ext4_mb_normalize_request(ac, ar);
4487 /* allocate space in core */
4488 *errp = ext4_mb_regular_allocator(ac);
4490 goto discard_and_exit;
4492 /* as we've just preallocated more space than
4493 * user requested originally, we store allocated
4494 * space in a special descriptor */
4495 if (ac->ac_status == AC_STATUS_FOUND &&
4496 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4497 *errp = ext4_mb_new_preallocation(ac);
4500 ext4_discard_allocated_blocks(ac);
4504 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4505 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4506 if (*errp == -EAGAIN) {
4508 * drop the reference that we took
4509 * in ext4_mb_use_best_found
4511 ext4_mb_release_context(ac);
4512 ac->ac_b_ex.fe_group = 0;
4513 ac->ac_b_ex.fe_start = 0;
4514 ac->ac_b_ex.fe_len = 0;
4515 ac->ac_status = AC_STATUS_CONTINUE;
4518 ext4_discard_allocated_blocks(ac);
4521 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4522 ar->len = ac->ac_b_ex.fe_len;
4525 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4533 ac->ac_b_ex.fe_len = 0;
4535 ext4_mb_show_ac(ac);
4537 ext4_mb_release_context(ac);
4540 kmem_cache_free(ext4_ac_cachep, ac);
4541 if (inquota && ar->len < inquota)
4542 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4544 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4545 /* release all the reserved blocks if non delalloc */
4546 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4550 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4556 * We can merge two free data extents only if the physical blocks
4557 * are contiguous, AND the extents were freed by the same transaction,
4558 * AND the blocks are associated with the same group.
4560 static int can_merge(struct ext4_free_data *entry1,
4561 struct ext4_free_data *entry2)
4563 if ((entry1->efd_tid == entry2->efd_tid) &&
4564 (entry1->efd_group == entry2->efd_group) &&
4565 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4570 static noinline_for_stack int
4571 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4572 struct ext4_free_data *new_entry)
4574 ext4_group_t group = e4b->bd_group;
4575 ext4_grpblk_t cluster;
4576 struct ext4_free_data *entry;
4577 struct ext4_group_info *db = e4b->bd_info;
4578 struct super_block *sb = e4b->bd_sb;
4579 struct ext4_sb_info *sbi = EXT4_SB(sb);
4580 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4581 struct rb_node *parent = NULL, *new_node;
4583 BUG_ON(!ext4_handle_valid(handle));
4584 BUG_ON(e4b->bd_bitmap_page == NULL);
4585 BUG_ON(e4b->bd_buddy_page == NULL);
4587 new_node = &new_entry->efd_node;
4588 cluster = new_entry->efd_start_cluster;
4591 /* first free block exent. We need to
4592 protect buddy cache from being freed,
4593 * otherwise we'll refresh it from
4594 * on-disk bitmap and lose not-yet-available
4596 page_cache_get(e4b->bd_buddy_page);
4597 page_cache_get(e4b->bd_bitmap_page);
4601 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4602 if (cluster < entry->efd_start_cluster)
4604 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4605 n = &(*n)->rb_right;
4607 ext4_grp_locked_error(sb, group, 0,
4608 ext4_group_first_block_no(sb, group) +
4609 EXT4_C2B(sbi, cluster),
4610 "Block already on to-be-freed list");
4615 rb_link_node(new_node, parent, n);
4616 rb_insert_color(new_node, &db->bb_free_root);
4618 /* Now try to see the extent can be merged to left and right */
4619 node = rb_prev(new_node);
4621 entry = rb_entry(node, struct ext4_free_data, efd_node);
4622 if (can_merge(entry, new_entry) &&
4623 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4624 new_entry->efd_start_cluster = entry->efd_start_cluster;
4625 new_entry->efd_count += entry->efd_count;
4626 rb_erase(node, &(db->bb_free_root));
4627 kmem_cache_free(ext4_free_data_cachep, entry);
4631 node = rb_next(new_node);
4633 entry = rb_entry(node, struct ext4_free_data, efd_node);
4634 if (can_merge(new_entry, entry) &&
4635 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4636 new_entry->efd_count += entry->efd_count;
4637 rb_erase(node, &(db->bb_free_root));
4638 kmem_cache_free(ext4_free_data_cachep, entry);
4641 /* Add the extent to transaction's private list */
4642 ext4_journal_callback_add(handle, ext4_free_data_callback,
4643 &new_entry->efd_jce);
4648 * ext4_free_blocks() -- Free given blocks and update quota
4649 * @handle: handle for this transaction
4651 * @block: start physical block to free
4652 * @count: number of blocks to count
4653 * @flags: flags used by ext4_free_blocks
4655 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4656 struct buffer_head *bh, ext4_fsblk_t block,
4657 unsigned long count, int flags)
4659 struct buffer_head *bitmap_bh = NULL;
4660 struct super_block *sb = inode->i_sb;
4661 struct ext4_group_desc *gdp;
4662 unsigned int overflow;
4664 struct buffer_head *gd_bh;
4665 ext4_group_t block_group;
4666 struct ext4_sb_info *sbi;
4667 struct ext4_buddy e4b;
4668 unsigned int count_clusters;
4675 BUG_ON(block != bh->b_blocknr);
4677 block = bh->b_blocknr;
4681 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4682 !ext4_data_block_valid(sbi, block, count)) {
4683 ext4_error(sb, "Freeing blocks not in datazone - "
4684 "block = %llu, count = %lu", block, count);
4688 ext4_debug("freeing block %llu\n", block);
4689 trace_ext4_free_blocks(inode, block, count, flags);
4691 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4694 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4699 * We need to make sure we don't reuse the freed block until
4700 * after the transaction is committed, which we can do by
4701 * treating the block as metadata, below. We make an
4702 * exception if the inode is to be written in writeback mode
4703 * since writeback mode has weak data consistency guarantees.
4705 if (!ext4_should_writeback_data(inode))
4706 flags |= EXT4_FREE_BLOCKS_METADATA;
4709 * If the extent to be freed does not begin on a cluster
4710 * boundary, we need to deal with partial clusters at the
4711 * beginning and end of the extent. Normally we will free
4712 * blocks at the beginning or the end unless we are explicitly
4713 * requested to avoid doing so.
4715 overflow = EXT4_PBLK_COFF(sbi, block);
4717 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4718 overflow = sbi->s_cluster_ratio - overflow;
4720 if (count > overflow)
4729 overflow = EXT4_LBLK_COFF(sbi, count);
4731 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4732 if (count > overflow)
4737 count += sbi->s_cluster_ratio - overflow;
4740 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4743 for (i = 0; i < count; i++) {
4745 bh = sb_find_get_block(inode->i_sb, block + i);
4748 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4749 inode, bh, block + i);
4755 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4757 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4758 ext4_get_group_info(sb, block_group))))
4762 * Check to see if we are freeing blocks across a group
4765 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4766 overflow = EXT4_C2B(sbi, bit) + count -
4767 EXT4_BLOCKS_PER_GROUP(sb);
4770 count_clusters = EXT4_NUM_B2C(sbi, count);
4771 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4772 if (IS_ERR(bitmap_bh)) {
4773 err = PTR_ERR(bitmap_bh);
4777 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4783 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4784 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4785 in_range(block, ext4_inode_table(sb, gdp),
4786 EXT4_SB(sb)->s_itb_per_group) ||
4787 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4788 EXT4_SB(sb)->s_itb_per_group)) {
4790 ext4_error(sb, "Freeing blocks in system zone - "
4791 "Block = %llu, count = %lu", block, count);
4792 /* err = 0. ext4_std_error should be a no op */
4796 BUFFER_TRACE(bitmap_bh, "getting write access");
4797 err = ext4_journal_get_write_access(handle, bitmap_bh);
4802 * We are about to modify some metadata. Call the journal APIs
4803 * to unshare ->b_data if a currently-committing transaction is
4806 BUFFER_TRACE(gd_bh, "get_write_access");
4807 err = ext4_journal_get_write_access(handle, gd_bh);
4810 #ifdef AGGRESSIVE_CHECK
4813 for (i = 0; i < count_clusters; i++)
4814 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4817 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4819 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4823 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4824 struct ext4_free_data *new_entry;
4826 * blocks being freed are metadata. these blocks shouldn't
4827 * be used until this transaction is committed
4829 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4832 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4833 GFP_NOFS|__GFP_NOFAIL);
4834 new_entry->efd_start_cluster = bit;
4835 new_entry->efd_group = block_group;
4836 new_entry->efd_count = count_clusters;
4837 new_entry->efd_tid = handle->h_transaction->t_tid;
4839 ext4_lock_group(sb, block_group);
4840 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4841 ext4_mb_free_metadata(handle, &e4b, new_entry);
4843 /* need to update group_info->bb_free and bitmap
4844 * with group lock held. generate_buddy look at
4845 * them with group lock_held
4847 if (test_opt(sb, DISCARD)) {
4848 err = ext4_issue_discard(sb, block_group, bit, count,
4850 if (err && err != -EOPNOTSUPP)
4851 ext4_msg(sb, KERN_WARNING, "discard request in"
4852 " group:%d block:%d count:%lu failed"
4853 " with %d", block_group, bit, count,
4856 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4858 ext4_lock_group(sb, block_group);
4859 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4860 mb_free_blocks(inode, &e4b, bit, count_clusters);
4863 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4864 ext4_free_group_clusters_set(sb, gdp, ret);
4865 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4866 ext4_group_desc_csum_set(sb, block_group, gdp);
4867 ext4_unlock_group(sb, block_group);
4869 if (sbi->s_log_groups_per_flex) {
4870 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4871 atomic64_add(count_clusters,
4872 &sbi->s_flex_groups[flex_group].free_clusters);
4875 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4876 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4877 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4879 ext4_mb_unload_buddy(&e4b);
4881 /* We dirtied the bitmap block */
4882 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4883 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4885 /* And the group descriptor block */
4886 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4887 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4891 if (overflow && !err) {
4899 ext4_std_error(sb, err);
4904 * ext4_group_add_blocks() -- Add given blocks to an existing group
4905 * @handle: handle to this transaction
4907 * @block: start physical block to add to the block group
4908 * @count: number of blocks to free
4910 * This marks the blocks as free in the bitmap and buddy.
4912 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4913 ext4_fsblk_t block, unsigned long count)
4915 struct buffer_head *bitmap_bh = NULL;
4916 struct buffer_head *gd_bh;
4917 ext4_group_t block_group;
4920 struct ext4_group_desc *desc;
4921 struct ext4_sb_info *sbi = EXT4_SB(sb);
4922 struct ext4_buddy e4b;
4923 int err = 0, ret, blk_free_count;
4924 ext4_grpblk_t blocks_freed;
4926 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4931 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4933 * Check to see if we are freeing blocks across a group
4936 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4937 ext4_warning(sb, "too much blocks added to group %u\n",
4943 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4944 if (IS_ERR(bitmap_bh)) {
4945 err = PTR_ERR(bitmap_bh);
4950 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4956 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4957 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4958 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4959 in_range(block + count - 1, ext4_inode_table(sb, desc),
4960 sbi->s_itb_per_group)) {
4961 ext4_error(sb, "Adding blocks in system zones - "
4962 "Block = %llu, count = %lu",
4968 BUFFER_TRACE(bitmap_bh, "getting write access");
4969 err = ext4_journal_get_write_access(handle, bitmap_bh);
4974 * We are about to modify some metadata. Call the journal APIs
4975 * to unshare ->b_data if a currently-committing transaction is
4978 BUFFER_TRACE(gd_bh, "get_write_access");
4979 err = ext4_journal_get_write_access(handle, gd_bh);
4983 for (i = 0, blocks_freed = 0; i < count; i++) {
4984 BUFFER_TRACE(bitmap_bh, "clear bit");
4985 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4986 ext4_error(sb, "bit already cleared for block %llu",
4987 (ext4_fsblk_t)(block + i));
4988 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4994 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4999 * need to update group_info->bb_free and bitmap
5000 * with group lock held. generate_buddy look at
5001 * them with group lock_held
5003 ext4_lock_group(sb, block_group);
5004 mb_clear_bits(bitmap_bh->b_data, bit, count);
5005 mb_free_blocks(NULL, &e4b, bit, count);
5006 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
5007 ext4_free_group_clusters_set(sb, desc, blk_free_count);
5008 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5009 ext4_group_desc_csum_set(sb, block_group, desc);
5010 ext4_unlock_group(sb, block_group);
5011 percpu_counter_add(&sbi->s_freeclusters_counter,
5012 EXT4_NUM_B2C(sbi, blocks_freed));
5014 if (sbi->s_log_groups_per_flex) {
5015 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5016 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5017 &sbi->s_flex_groups[flex_group].free_clusters);
5020 ext4_mb_unload_buddy(&e4b);
5022 /* We dirtied the bitmap block */
5023 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5024 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5026 /* And the group descriptor block */
5027 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5028 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5034 ext4_std_error(sb, err);
5039 * ext4_trim_extent -- function to TRIM one single free extent in the group
5040 * @sb: super block for the file system
5041 * @start: starting block of the free extent in the alloc. group
5042 * @count: number of blocks to TRIM
5043 * @group: alloc. group we are working with
5044 * @e4b: ext4 buddy for the group
5045 * @blkdev_flags: flags for the block device
5047 * Trim "count" blocks starting at "start" in the "group". To assure that no
5048 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5049 * be called with under the group lock.
5051 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5052 ext4_group_t group, struct ext4_buddy *e4b,
5053 unsigned long blkdev_flags)
5057 struct ext4_free_extent ex;
5060 trace_ext4_trim_extent(sb, group, start, count);
5062 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5064 ex.fe_start = start;
5065 ex.fe_group = group;
5069 * Mark blocks used, so no one can reuse them while
5072 mb_mark_used(e4b, &ex);
5073 ext4_unlock_group(sb, group);
5074 ret = ext4_issue_discard(sb, group, start, count, blkdev_flags);
5075 ext4_lock_group(sb, group);
5076 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5081 * ext4_trim_all_free -- function to trim all free space in alloc. group
5082 * @sb: super block for file system
5083 * @group: group to be trimmed
5084 * @start: first group block to examine
5085 * @max: last group block to examine
5086 * @minblocks: minimum extent block count
5087 * @blkdev_flags: flags for the block device
5089 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5090 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5094 * ext4_trim_all_free walks through group's block bitmap searching for free
5095 * extents. When the free extent is found, mark it as used in group buddy
5096 * bitmap. Then issue a TRIM command on this extent and free the extent in
5097 * the group buddy bitmap. This is done until whole group is scanned.
5099 static ext4_grpblk_t
5100 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5101 ext4_grpblk_t start, ext4_grpblk_t max,
5102 ext4_grpblk_t minblocks, unsigned long blkdev_flags)
5105 ext4_grpblk_t next, count = 0, free_count = 0;
5106 struct ext4_buddy e4b;
5109 trace_ext4_trim_all_free(sb, group, start, max);
5111 ret = ext4_mb_load_buddy(sb, group, &e4b);
5113 ext4_error(sb, "Error in loading buddy "
5114 "information for %u", group);
5117 bitmap = e4b.bd_bitmap;
5119 ext4_lock_group(sb, group);
5120 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5121 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5124 start = (e4b.bd_info->bb_first_free > start) ?
5125 e4b.bd_info->bb_first_free : start;
5127 while (start <= max) {
5128 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5131 next = mb_find_next_bit(bitmap, max + 1, start);
5133 if ((next - start) >= minblocks) {
5134 ret = ext4_trim_extent(sb, start,
5135 next - start, group, &e4b,
5137 if (ret && ret != -EOPNOTSUPP)
5140 count += next - start;
5142 free_count += next - start;
5145 if (fatal_signal_pending(current)) {
5146 count = -ERESTARTSYS;
5150 if (need_resched()) {
5151 ext4_unlock_group(sb, group);
5153 ext4_lock_group(sb, group);
5156 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5162 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5165 ext4_unlock_group(sb, group);
5166 ext4_mb_unload_buddy(&e4b);
5168 ext4_debug("trimmed %d blocks in the group %d\n",
5175 * ext4_trim_fs() -- trim ioctl handle function
5176 * @sb: superblock for filesystem
5177 * @range: fstrim_range structure
5178 * @blkdev_flags: flags for the block device
5180 * start: First Byte to trim
5181 * len: number of Bytes to trim from start
5182 * minlen: minimum extent length in Bytes
5183 * ext4_trim_fs goes through all allocation groups containing Bytes from
5184 * start to start+len. For each such a group ext4_trim_all_free function
5185 * is invoked to trim all free space.
5187 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range,
5188 unsigned long blkdev_flags)
5190 struct ext4_group_info *grp;
5191 ext4_group_t group, first_group, last_group;
5192 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5193 uint64_t start, end, minlen, trimmed = 0;
5194 ext4_fsblk_t first_data_blk =
5195 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5196 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5199 start = range->start >> sb->s_blocksize_bits;
5200 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5201 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5202 range->minlen >> sb->s_blocksize_bits);
5204 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5205 start >= max_blks ||
5206 range->len < sb->s_blocksize)
5208 if (end >= max_blks)
5210 if (end <= first_data_blk)
5212 if (start < first_data_blk)
5213 start = first_data_blk;
5215 /* Determine first and last group to examine based on start and end */
5216 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5217 &first_group, &first_cluster);
5218 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5219 &last_group, &last_cluster);
5221 /* end now represents the last cluster to discard in this group */
5222 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5224 for (group = first_group; group <= last_group; group++) {
5225 grp = ext4_get_group_info(sb, group);
5226 /* We only do this if the grp has never been initialized */
5227 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5228 ret = ext4_mb_init_group(sb, group);
5234 * For all the groups except the last one, last cluster will
5235 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5236 * change it for the last group, note that last_cluster is
5237 * already computed earlier by ext4_get_group_no_and_offset()
5239 if (group == last_group)
5242 if (grp->bb_free >= minlen) {
5243 cnt = ext4_trim_all_free(sb, group, first_cluster,
5244 end, minlen, blkdev_flags);
5253 * For every group except the first one, we are sure
5254 * that the first cluster to discard will be cluster #0.
5260 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5263 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;