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);
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, gfp_t gfp)
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);
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, gfp_t gfp)
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);
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);
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, gfp_t gfp)
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, gfp);
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, gfp);
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, gfp);
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_gfp(struct super_block *sb, ext4_group_t group,
1113 struct ext4_buddy *e4b, gfp_t gfp)
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, gfp);
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);
1173 BUG_ON(page->mapping != inode->i_mapping);
1174 if (!PageUptodate(page)) {
1175 ret = ext4_mb_init_cache(page, NULL, gfp);
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);
1209 BUG_ON(page->mapping != inode->i_mapping);
1210 if (!PageUptodate(page)) {
1211 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1225 if (!PageUptodate(page)) {
1230 /* Pages marked accessed already */
1231 e4b->bd_buddy_page = page;
1232 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1234 BUG_ON(e4b->bd_bitmap_page == NULL);
1235 BUG_ON(e4b->bd_buddy_page == NULL);
1241 page_cache_release(page);
1242 if (e4b->bd_bitmap_page)
1243 page_cache_release(e4b->bd_bitmap_page);
1244 if (e4b->bd_buddy_page)
1245 page_cache_release(e4b->bd_buddy_page);
1246 e4b->bd_buddy = NULL;
1247 e4b->bd_bitmap = NULL;
1251 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1252 struct ext4_buddy *e4b)
1254 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1257 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1259 if (e4b->bd_bitmap_page)
1260 page_cache_release(e4b->bd_bitmap_page);
1261 if (e4b->bd_buddy_page)
1262 page_cache_release(e4b->bd_buddy_page);
1266 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1269 int bb_incr = 1 << (e4b->bd_blkbits - 1);
1272 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1273 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1276 while (order <= e4b->bd_blkbits + 1) {
1278 if (!mb_test_bit(block, bb)) {
1279 /* this block is part of buddy of order 'order' */
1289 static void mb_clear_bits(void *bm, int cur, int len)
1295 if ((cur & 31) == 0 && (len - cur) >= 32) {
1296 /* fast path: clear whole word at once */
1297 addr = bm + (cur >> 3);
1302 mb_clear_bit(cur, bm);
1307 /* clear bits in given range
1308 * will return first found zero bit if any, -1 otherwise
1310 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1317 if ((cur & 31) == 0 && (len - cur) >= 32) {
1318 /* fast path: clear whole word at once */
1319 addr = bm + (cur >> 3);
1320 if (*addr != (__u32)(-1) && zero_bit == -1)
1321 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1326 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1334 void ext4_set_bits(void *bm, int cur, int len)
1340 if ((cur & 31) == 0 && (len - cur) >= 32) {
1341 /* fast path: set whole word at once */
1342 addr = bm + (cur >> 3);
1347 mb_set_bit(cur, bm);
1353 * _________________________________________________________________ */
1355 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1357 if (mb_test_bit(*bit + side, bitmap)) {
1358 mb_clear_bit(*bit, bitmap);
1364 mb_set_bit(*bit, bitmap);
1369 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1373 void *buddy = mb_find_buddy(e4b, order, &max);
1378 /* Bits in range [first; last] are known to be set since
1379 * corresponding blocks were allocated. Bits in range
1380 * (first; last) will stay set because they form buddies on
1381 * upper layer. We just deal with borders if they don't
1382 * align with upper layer and then go up.
1383 * Releasing entire group is all about clearing
1384 * single bit of highest order buddy.
1388 * ---------------------------------
1390 * ---------------------------------
1391 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1392 * ---------------------------------
1394 * \_____________________/
1396 * Neither [1] nor [6] is aligned to above layer.
1397 * Left neighbour [0] is free, so mark it busy,
1398 * decrease bb_counters and extend range to
1400 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1401 * mark [6] free, increase bb_counters and shrink range to
1403 * Then shift range to [0; 2], go up and do the same.
1408 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1410 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1415 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1416 mb_clear_bits(buddy, first, last - first + 1);
1417 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1426 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1427 int first, int count)
1429 int left_is_free = 0;
1430 int right_is_free = 0;
1432 int last = first + count - 1;
1433 struct super_block *sb = e4b->bd_sb;
1435 if (WARN_ON(count == 0))
1437 BUG_ON(last >= (sb->s_blocksize << 3));
1438 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1439 /* Don't bother if the block group is corrupt. */
1440 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1443 mb_check_buddy(e4b);
1444 mb_free_blocks_double(inode, e4b, first, count);
1446 e4b->bd_info->bb_free += count;
1447 if (first < e4b->bd_info->bb_first_free)
1448 e4b->bd_info->bb_first_free = first;
1450 /* access memory sequentially: check left neighbour,
1451 * clear range and then check right neighbour
1454 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1455 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1456 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1457 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1459 if (unlikely(block != -1)) {
1460 struct ext4_sb_info *sbi = EXT4_SB(sb);
1461 ext4_fsblk_t blocknr;
1463 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1464 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1465 ext4_grp_locked_error(sb, e4b->bd_group,
1466 inode ? inode->i_ino : 0,
1468 "freeing already freed block "
1469 "(bit %u); block bitmap corrupt.",
1471 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1472 percpu_counter_sub(&sbi->s_freeclusters_counter,
1473 e4b->bd_info->bb_free);
1474 /* Mark the block group as corrupt. */
1475 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1476 &e4b->bd_info->bb_state);
1477 mb_regenerate_buddy(e4b);
1481 /* let's maintain fragments counter */
1482 if (left_is_free && right_is_free)
1483 e4b->bd_info->bb_fragments--;
1484 else if (!left_is_free && !right_is_free)
1485 e4b->bd_info->bb_fragments++;
1487 /* buddy[0] == bd_bitmap is a special case, so handle
1488 * it right away and let mb_buddy_mark_free stay free of
1489 * zero order checks.
1490 * Check if neighbours are to be coaleasced,
1491 * adjust bitmap bb_counters and borders appropriately.
1494 first += !left_is_free;
1495 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1498 last -= !right_is_free;
1499 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1503 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1506 mb_set_largest_free_order(sb, e4b->bd_info);
1507 mb_check_buddy(e4b);
1510 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1511 int needed, struct ext4_free_extent *ex)
1517 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1520 buddy = mb_find_buddy(e4b, 0, &max);
1521 BUG_ON(buddy == NULL);
1522 BUG_ON(block >= max);
1523 if (mb_test_bit(block, buddy)) {
1530 /* find actual order */
1531 order = mb_find_order_for_block(e4b, block);
1532 block = block >> order;
1534 ex->fe_len = 1 << order;
1535 ex->fe_start = block << order;
1536 ex->fe_group = e4b->bd_group;
1538 /* calc difference from given start */
1539 next = next - ex->fe_start;
1541 ex->fe_start += next;
1543 while (needed > ex->fe_len &&
1544 mb_find_buddy(e4b, order, &max)) {
1546 if (block + 1 >= max)
1549 next = (block + 1) * (1 << order);
1550 if (mb_test_bit(next, e4b->bd_bitmap))
1553 order = mb_find_order_for_block(e4b, next);
1555 block = next >> order;
1556 ex->fe_len += 1 << order;
1559 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1563 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1569 int start = ex->fe_start;
1570 int len = ex->fe_len;
1575 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1576 BUG_ON(e4b->bd_group != ex->fe_group);
1577 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1578 mb_check_buddy(e4b);
1579 mb_mark_used_double(e4b, start, len);
1581 e4b->bd_info->bb_free -= len;
1582 if (e4b->bd_info->bb_first_free == start)
1583 e4b->bd_info->bb_first_free += len;
1585 /* let's maintain fragments counter */
1587 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1588 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1589 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1591 e4b->bd_info->bb_fragments++;
1592 else if (!mlen && !max)
1593 e4b->bd_info->bb_fragments--;
1595 /* let's maintain buddy itself */
1597 ord = mb_find_order_for_block(e4b, start);
1599 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1600 /* the whole chunk may be allocated at once! */
1602 buddy = mb_find_buddy(e4b, ord, &max);
1603 BUG_ON((start >> ord) >= max);
1604 mb_set_bit(start >> ord, buddy);
1605 e4b->bd_info->bb_counters[ord]--;
1612 /* store for history */
1614 ret = len | (ord << 16);
1616 /* we have to split large buddy */
1618 buddy = mb_find_buddy(e4b, ord, &max);
1619 mb_set_bit(start >> ord, buddy);
1620 e4b->bd_info->bb_counters[ord]--;
1623 cur = (start >> ord) & ~1U;
1624 buddy = mb_find_buddy(e4b, ord, &max);
1625 mb_clear_bit(cur, buddy);
1626 mb_clear_bit(cur + 1, buddy);
1627 e4b->bd_info->bb_counters[ord]++;
1628 e4b->bd_info->bb_counters[ord]++;
1630 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1632 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1633 mb_check_buddy(e4b);
1639 * Must be called under group lock!
1641 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1642 struct ext4_buddy *e4b)
1644 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1647 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1648 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1650 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1651 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1652 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1654 /* preallocation can change ac_b_ex, thus we store actually
1655 * allocated blocks for history */
1656 ac->ac_f_ex = ac->ac_b_ex;
1658 ac->ac_status = AC_STATUS_FOUND;
1659 ac->ac_tail = ret & 0xffff;
1660 ac->ac_buddy = ret >> 16;
1663 * take the page reference. We want the page to be pinned
1664 * so that we don't get a ext4_mb_init_cache_call for this
1665 * group until we update the bitmap. That would mean we
1666 * double allocate blocks. The reference is dropped
1667 * in ext4_mb_release_context
1669 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1670 get_page(ac->ac_bitmap_page);
1671 ac->ac_buddy_page = e4b->bd_buddy_page;
1672 get_page(ac->ac_buddy_page);
1673 /* store last allocated for subsequent stream allocation */
1674 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1675 spin_lock(&sbi->s_md_lock);
1676 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1677 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1678 spin_unlock(&sbi->s_md_lock);
1683 * regular allocator, for general purposes allocation
1686 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1687 struct ext4_buddy *e4b,
1690 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1691 struct ext4_free_extent *bex = &ac->ac_b_ex;
1692 struct ext4_free_extent *gex = &ac->ac_g_ex;
1693 struct ext4_free_extent ex;
1696 if (ac->ac_status == AC_STATUS_FOUND)
1699 * We don't want to scan for a whole year
1701 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1702 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1703 ac->ac_status = AC_STATUS_BREAK;
1708 * Haven't found good chunk so far, let's continue
1710 if (bex->fe_len < gex->fe_len)
1713 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1714 && bex->fe_group == e4b->bd_group) {
1715 /* recheck chunk's availability - we don't know
1716 * when it was found (within this lock-unlock
1718 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1719 if (max >= gex->fe_len) {
1720 ext4_mb_use_best_found(ac, e4b);
1727 * The routine checks whether found extent is good enough. If it is,
1728 * then the extent gets marked used and flag is set to the context
1729 * to stop scanning. Otherwise, the extent is compared with the
1730 * previous found extent and if new one is better, then it's stored
1731 * in the context. Later, the best found extent will be used, if
1732 * mballoc can't find good enough extent.
1734 * FIXME: real allocation policy is to be designed yet!
1736 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1737 struct ext4_free_extent *ex,
1738 struct ext4_buddy *e4b)
1740 struct ext4_free_extent *bex = &ac->ac_b_ex;
1741 struct ext4_free_extent *gex = &ac->ac_g_ex;
1743 BUG_ON(ex->fe_len <= 0);
1744 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1745 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1746 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1751 * The special case - take what you catch first
1753 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1755 ext4_mb_use_best_found(ac, e4b);
1760 * Let's check whether the chuck is good enough
1762 if (ex->fe_len == gex->fe_len) {
1764 ext4_mb_use_best_found(ac, e4b);
1769 * If this is first found extent, just store it in the context
1771 if (bex->fe_len == 0) {
1777 * If new found extent is better, store it in the context
1779 if (bex->fe_len < gex->fe_len) {
1780 /* if the request isn't satisfied, any found extent
1781 * larger than previous best one is better */
1782 if (ex->fe_len > bex->fe_len)
1784 } else if (ex->fe_len > gex->fe_len) {
1785 /* if the request is satisfied, then we try to find
1786 * an extent that still satisfy the request, but is
1787 * smaller than previous one */
1788 if (ex->fe_len < bex->fe_len)
1792 ext4_mb_check_limits(ac, e4b, 0);
1795 static noinline_for_stack
1796 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1797 struct ext4_buddy *e4b)
1799 struct ext4_free_extent ex = ac->ac_b_ex;
1800 ext4_group_t group = ex.fe_group;
1804 BUG_ON(ex.fe_len <= 0);
1805 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1809 ext4_lock_group(ac->ac_sb, group);
1810 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1814 ext4_mb_use_best_found(ac, e4b);
1817 ext4_unlock_group(ac->ac_sb, group);
1818 ext4_mb_unload_buddy(e4b);
1823 static noinline_for_stack
1824 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1825 struct ext4_buddy *e4b)
1827 ext4_group_t group = ac->ac_g_ex.fe_group;
1830 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1831 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1832 struct ext4_free_extent ex;
1834 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1836 if (grp->bb_free == 0)
1839 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1843 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1844 ext4_mb_unload_buddy(e4b);
1848 ext4_lock_group(ac->ac_sb, group);
1849 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1850 ac->ac_g_ex.fe_len, &ex);
1851 ex.fe_logical = 0xDEADFA11; /* debug value */
1853 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1856 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1858 /* use do_div to get remainder (would be 64-bit modulo) */
1859 if (do_div(start, sbi->s_stripe) == 0) {
1862 ext4_mb_use_best_found(ac, e4b);
1864 } else if (max >= ac->ac_g_ex.fe_len) {
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);
1871 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1872 /* Sometimes, caller may want to merge even small
1873 * number of blocks to an existing extent */
1874 BUG_ON(ex.fe_len <= 0);
1875 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1876 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1879 ext4_mb_use_best_found(ac, e4b);
1881 ext4_unlock_group(ac->ac_sb, group);
1882 ext4_mb_unload_buddy(e4b);
1888 * The routine scans buddy structures (not bitmap!) from given order
1889 * to max order and tries to find big enough chunk to satisfy the req
1891 static noinline_for_stack
1892 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1893 struct ext4_buddy *e4b)
1895 struct super_block *sb = ac->ac_sb;
1896 struct ext4_group_info *grp = e4b->bd_info;
1902 BUG_ON(ac->ac_2order <= 0);
1903 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1904 if (grp->bb_counters[i] == 0)
1907 buddy = mb_find_buddy(e4b, i, &max);
1908 BUG_ON(buddy == NULL);
1910 k = mb_find_next_zero_bit(buddy, max, 0);
1915 ac->ac_b_ex.fe_len = 1 << i;
1916 ac->ac_b_ex.fe_start = k << i;
1917 ac->ac_b_ex.fe_group = e4b->bd_group;
1919 ext4_mb_use_best_found(ac, e4b);
1921 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1923 if (EXT4_SB(sb)->s_mb_stats)
1924 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1931 * The routine scans the group and measures all found extents.
1932 * In order to optimize scanning, caller must pass number of
1933 * free blocks in the group, so the routine can know upper limit.
1935 static noinline_for_stack
1936 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1937 struct ext4_buddy *e4b)
1939 struct super_block *sb = ac->ac_sb;
1940 void *bitmap = e4b->bd_bitmap;
1941 struct ext4_free_extent ex;
1945 free = e4b->bd_info->bb_free;
1948 i = e4b->bd_info->bb_first_free;
1950 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1951 i = mb_find_next_zero_bit(bitmap,
1952 EXT4_CLUSTERS_PER_GROUP(sb), i);
1953 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1955 * IF we have corrupt bitmap, we won't find any
1956 * free blocks even though group info says we
1957 * we have free blocks
1959 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1960 "%d free clusters as per "
1961 "group info. But bitmap says 0",
1966 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1967 BUG_ON(ex.fe_len <= 0);
1968 if (free < ex.fe_len) {
1969 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1970 "%d free clusters as per "
1971 "group info. But got %d blocks",
1974 * The number of free blocks differs. This mostly
1975 * indicate that the bitmap is corrupt. So exit
1976 * without claiming the space.
1980 ex.fe_logical = 0xDEADC0DE; /* debug value */
1981 ext4_mb_measure_extent(ac, &ex, e4b);
1987 ext4_mb_check_limits(ac, e4b, 1);
1991 * This is a special case for storages like raid5
1992 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1994 static noinline_for_stack
1995 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1996 struct ext4_buddy *e4b)
1998 struct super_block *sb = ac->ac_sb;
1999 struct ext4_sb_info *sbi = EXT4_SB(sb);
2000 void *bitmap = e4b->bd_bitmap;
2001 struct ext4_free_extent ex;
2002 ext4_fsblk_t first_group_block;
2007 BUG_ON(sbi->s_stripe == 0);
2009 /* find first stripe-aligned block in group */
2010 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2012 a = first_group_block + sbi->s_stripe - 1;
2013 do_div(a, sbi->s_stripe);
2014 i = (a * sbi->s_stripe) - first_group_block;
2016 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2017 if (!mb_test_bit(i, bitmap)) {
2018 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2019 if (max >= sbi->s_stripe) {
2021 ex.fe_logical = 0xDEADF00D; /* debug value */
2023 ext4_mb_use_best_found(ac, e4b);
2032 * This is now called BEFORE we load the buddy bitmap.
2033 * Returns either 1 or 0 indicating that the group is either suitable
2034 * for the allocation or not. In addition it can also return negative
2035 * error code when something goes wrong.
2037 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2038 ext4_group_t group, int cr)
2040 unsigned free, fragments;
2041 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2042 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2044 BUG_ON(cr < 0 || cr >= 4);
2046 free = grp->bb_free;
2049 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2052 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2055 /* We only do this if the grp has never been initialized */
2056 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2057 int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS);
2062 fragments = grp->bb_fragments;
2068 BUG_ON(ac->ac_2order == 0);
2070 /* Avoid using the first bg of a flexgroup for data files */
2071 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2072 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2073 ((group % flex_size) == 0))
2076 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2077 (free / fragments) >= ac->ac_g_ex.fe_len)
2080 if (grp->bb_largest_free_order < ac->ac_2order)
2085 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2089 if (free >= ac->ac_g_ex.fe_len)
2101 static noinline_for_stack int
2102 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2104 ext4_group_t ngroups, group, i;
2106 int err = 0, first_err = 0;
2107 struct ext4_sb_info *sbi;
2108 struct super_block *sb;
2109 struct ext4_buddy e4b;
2113 ngroups = ext4_get_groups_count(sb);
2114 /* non-extent files are limited to low blocks/groups */
2115 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2116 ngroups = sbi->s_blockfile_groups;
2118 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2120 /* first, try the goal */
2121 err = ext4_mb_find_by_goal(ac, &e4b);
2122 if (err || ac->ac_status == AC_STATUS_FOUND)
2125 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2129 * ac->ac2_order is set only if the fe_len is a power of 2
2130 * if ac2_order is set we also set criteria to 0 so that we
2131 * try exact allocation using buddy.
2133 i = fls(ac->ac_g_ex.fe_len);
2136 * We search using buddy data only if the order of the request
2137 * is greater than equal to the sbi_s_mb_order2_reqs
2138 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2140 if (i >= sbi->s_mb_order2_reqs) {
2142 * This should tell if fe_len is exactly power of 2
2144 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2145 ac->ac_2order = i - 1;
2148 /* if stream allocation is enabled, use global goal */
2149 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2150 /* TBD: may be hot point */
2151 spin_lock(&sbi->s_md_lock);
2152 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2153 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2154 spin_unlock(&sbi->s_md_lock);
2157 /* Let's just scan groups to find more-less suitable blocks */
2158 cr = ac->ac_2order ? 0 : 1;
2160 * cr == 0 try to get exact allocation,
2161 * cr == 3 try to get anything
2164 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2165 ac->ac_criteria = cr;
2167 * searching for the right group start
2168 * from the goal value specified
2170 group = ac->ac_g_ex.fe_group;
2172 for (i = 0; i < ngroups; group++, i++) {
2176 * Artificially restricted ngroups for non-extent
2177 * files makes group > ngroups possible on first loop.
2179 if (group >= ngroups)
2182 /* This now checks without needing the buddy page */
2183 ret = ext4_mb_good_group(ac, group, cr);
2190 err = ext4_mb_load_buddy(sb, group, &e4b);
2194 ext4_lock_group(sb, group);
2197 * We need to check again after locking the
2200 ret = ext4_mb_good_group(ac, group, cr);
2202 ext4_unlock_group(sb, group);
2203 ext4_mb_unload_buddy(&e4b);
2209 ac->ac_groups_scanned++;
2210 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2211 ext4_mb_simple_scan_group(ac, &e4b);
2212 else if (cr == 1 && sbi->s_stripe &&
2213 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2214 ext4_mb_scan_aligned(ac, &e4b);
2216 ext4_mb_complex_scan_group(ac, &e4b);
2218 ext4_unlock_group(sb, group);
2219 ext4_mb_unload_buddy(&e4b);
2221 if (ac->ac_status != AC_STATUS_CONTINUE)
2226 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2227 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2229 * We've been searching too long. Let's try to allocate
2230 * the best chunk we've found so far
2233 ext4_mb_try_best_found(ac, &e4b);
2234 if (ac->ac_status != AC_STATUS_FOUND) {
2236 * Someone more lucky has already allocated it.
2237 * The only thing we can do is just take first
2239 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2241 ac->ac_b_ex.fe_group = 0;
2242 ac->ac_b_ex.fe_start = 0;
2243 ac->ac_b_ex.fe_len = 0;
2244 ac->ac_status = AC_STATUS_CONTINUE;
2245 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2247 atomic_inc(&sbi->s_mb_lost_chunks);
2252 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2257 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2259 struct super_block *sb = seq->private;
2262 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2265 return (void *) ((unsigned long) group);
2268 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2270 struct super_block *sb = seq->private;
2274 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2277 return (void *) ((unsigned long) group);
2280 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2282 struct super_block *sb = seq->private;
2283 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2285 int err, buddy_loaded = 0;
2286 struct ext4_buddy e4b;
2287 struct ext4_group_info *grinfo;
2289 struct ext4_group_info info;
2290 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2295 seq_puts(seq, "#group: free frags first ["
2296 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2297 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]");
2299 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2300 sizeof(struct ext4_group_info);
2301 grinfo = ext4_get_group_info(sb, group);
2302 /* Load the group info in memory only if not already loaded. */
2303 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2304 err = ext4_mb_load_buddy(sb, group, &e4b);
2306 seq_printf(seq, "#%-5u: I/O error\n", group);
2312 memcpy(&sg, ext4_get_group_info(sb, group), i);
2315 ext4_mb_unload_buddy(&e4b);
2317 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2318 sg.info.bb_fragments, sg.info.bb_first_free);
2319 for (i = 0; i <= 13; i++)
2320 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2321 sg.info.bb_counters[i] : 0);
2322 seq_printf(seq, " ]\n");
2327 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2331 static const struct seq_operations ext4_mb_seq_groups_ops = {
2332 .start = ext4_mb_seq_groups_start,
2333 .next = ext4_mb_seq_groups_next,
2334 .stop = ext4_mb_seq_groups_stop,
2335 .show = ext4_mb_seq_groups_show,
2338 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2340 struct super_block *sb = PDE_DATA(inode);
2343 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2345 struct seq_file *m = file->private_data;
2352 const struct file_operations ext4_seq_mb_groups_fops = {
2353 .owner = THIS_MODULE,
2354 .open = ext4_mb_seq_groups_open,
2356 .llseek = seq_lseek,
2357 .release = seq_release,
2360 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2362 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2363 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2370 * Allocate the top-level s_group_info array for the specified number
2373 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2375 struct ext4_sb_info *sbi = EXT4_SB(sb);
2377 struct ext4_group_info ***new_groupinfo;
2379 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2380 EXT4_DESC_PER_BLOCK_BITS(sb);
2381 if (size <= sbi->s_group_info_size)
2384 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2385 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2386 if (!new_groupinfo) {
2387 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2390 if (sbi->s_group_info) {
2391 memcpy(new_groupinfo, sbi->s_group_info,
2392 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2393 kvfree(sbi->s_group_info);
2395 sbi->s_group_info = new_groupinfo;
2396 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2397 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2398 sbi->s_group_info_size);
2402 /* Create and initialize ext4_group_info data for the given group. */
2403 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2404 struct ext4_group_desc *desc)
2408 struct ext4_sb_info *sbi = EXT4_SB(sb);
2409 struct ext4_group_info **meta_group_info;
2410 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2413 * First check if this group is the first of a reserved block.
2414 * If it's true, we have to allocate a new table of pointers
2415 * to ext4_group_info structures
2417 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2418 metalen = sizeof(*meta_group_info) <<
2419 EXT4_DESC_PER_BLOCK_BITS(sb);
2420 meta_group_info = kmalloc(metalen, GFP_NOFS);
2421 if (meta_group_info == NULL) {
2422 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2423 "for a buddy group");
2424 goto exit_meta_group_info;
2426 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2431 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2432 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2434 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2435 if (meta_group_info[i] == NULL) {
2436 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2437 goto exit_group_info;
2439 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2440 &(meta_group_info[i]->bb_state));
2443 * initialize bb_free to be able to skip
2444 * empty groups without initialization
2446 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2447 meta_group_info[i]->bb_free =
2448 ext4_free_clusters_after_init(sb, group, desc);
2450 meta_group_info[i]->bb_free =
2451 ext4_free_group_clusters(sb, desc);
2454 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2455 init_rwsem(&meta_group_info[i]->alloc_sem);
2456 meta_group_info[i]->bb_free_root = RB_ROOT;
2457 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2461 struct buffer_head *bh;
2462 meta_group_info[i]->bb_bitmap =
2463 kmalloc(sb->s_blocksize, GFP_NOFS);
2464 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2465 bh = ext4_read_block_bitmap(sb, group);
2466 BUG_ON(IS_ERR_OR_NULL(bh));
2467 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2476 /* If a meta_group_info table has been allocated, release it now */
2477 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2478 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2479 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2481 exit_meta_group_info:
2483 } /* ext4_mb_add_groupinfo */
2485 static int ext4_mb_init_backend(struct super_block *sb)
2487 ext4_group_t ngroups = ext4_get_groups_count(sb);
2489 struct ext4_sb_info *sbi = EXT4_SB(sb);
2491 struct ext4_group_desc *desc;
2492 struct kmem_cache *cachep;
2494 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2498 sbi->s_buddy_cache = new_inode(sb);
2499 if (sbi->s_buddy_cache == NULL) {
2500 ext4_msg(sb, KERN_ERR, "can't get new inode");
2503 /* To avoid potentially colliding with an valid on-disk inode number,
2504 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2505 * not in the inode hash, so it should never be found by iget(), but
2506 * this will avoid confusion if it ever shows up during debugging. */
2507 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2508 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2509 for (i = 0; i < ngroups; i++) {
2510 desc = ext4_get_group_desc(sb, i, NULL);
2512 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2515 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2522 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2524 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2525 i = sbi->s_group_info_size;
2527 kfree(sbi->s_group_info[i]);
2528 iput(sbi->s_buddy_cache);
2530 kvfree(sbi->s_group_info);
2534 static void ext4_groupinfo_destroy_slabs(void)
2538 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2539 if (ext4_groupinfo_caches[i])
2540 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2541 ext4_groupinfo_caches[i] = NULL;
2545 static int ext4_groupinfo_create_slab(size_t size)
2547 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2549 int blocksize_bits = order_base_2(size);
2550 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2551 struct kmem_cache *cachep;
2553 if (cache_index >= NR_GRPINFO_CACHES)
2556 if (unlikely(cache_index < 0))
2559 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2560 if (ext4_groupinfo_caches[cache_index]) {
2561 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2562 return 0; /* Already created */
2565 slab_size = offsetof(struct ext4_group_info,
2566 bb_counters[blocksize_bits + 2]);
2568 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2569 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2572 ext4_groupinfo_caches[cache_index] = cachep;
2574 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2577 "EXT4-fs: no memory for groupinfo slab cache\n");
2584 int ext4_mb_init(struct super_block *sb)
2586 struct ext4_sb_info *sbi = EXT4_SB(sb);
2588 unsigned offset, offset_incr;
2592 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2594 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2595 if (sbi->s_mb_offsets == NULL) {
2600 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2601 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2602 if (sbi->s_mb_maxs == NULL) {
2607 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2611 /* order 0 is regular bitmap */
2612 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2613 sbi->s_mb_offsets[0] = 0;
2617 offset_incr = 1 << (sb->s_blocksize_bits - 1);
2618 max = sb->s_blocksize << 2;
2620 sbi->s_mb_offsets[i] = offset;
2621 sbi->s_mb_maxs[i] = max;
2622 offset += offset_incr;
2623 offset_incr = offset_incr >> 1;
2626 } while (i <= sb->s_blocksize_bits + 1);
2628 spin_lock_init(&sbi->s_md_lock);
2629 spin_lock_init(&sbi->s_bal_lock);
2631 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2632 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2633 sbi->s_mb_stats = MB_DEFAULT_STATS;
2634 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2635 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2637 * The default group preallocation is 512, which for 4k block
2638 * sizes translates to 2 megabytes. However for bigalloc file
2639 * systems, this is probably too big (i.e, if the cluster size
2640 * is 1 megabyte, then group preallocation size becomes half a
2641 * gigabyte!). As a default, we will keep a two megabyte
2642 * group pralloc size for cluster sizes up to 64k, and after
2643 * that, we will force a minimum group preallocation size of
2644 * 32 clusters. This translates to 8 megs when the cluster
2645 * size is 256k, and 32 megs when the cluster size is 1 meg,
2646 * which seems reasonable as a default.
2648 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2649 sbi->s_cluster_bits, 32);
2651 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2652 * to the lowest multiple of s_stripe which is bigger than
2653 * the s_mb_group_prealloc as determined above. We want
2654 * the preallocation size to be an exact multiple of the
2655 * RAID stripe size so that preallocations don't fragment
2658 if (sbi->s_stripe > 1) {
2659 sbi->s_mb_group_prealloc = roundup(
2660 sbi->s_mb_group_prealloc, sbi->s_stripe);
2663 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2664 if (sbi->s_locality_groups == NULL) {
2668 for_each_possible_cpu(i) {
2669 struct ext4_locality_group *lg;
2670 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2671 mutex_init(&lg->lg_mutex);
2672 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2673 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2674 spin_lock_init(&lg->lg_prealloc_lock);
2677 /* init file for buddy data */
2678 ret = ext4_mb_init_backend(sb);
2680 goto out_free_locality_groups;
2684 out_free_locality_groups:
2685 free_percpu(sbi->s_locality_groups);
2686 sbi->s_locality_groups = NULL;
2688 kfree(sbi->s_mb_offsets);
2689 sbi->s_mb_offsets = NULL;
2690 kfree(sbi->s_mb_maxs);
2691 sbi->s_mb_maxs = NULL;
2695 /* need to called with the ext4 group lock held */
2696 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2698 struct ext4_prealloc_space *pa;
2699 struct list_head *cur, *tmp;
2702 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2703 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2704 list_del(&pa->pa_group_list);
2706 kmem_cache_free(ext4_pspace_cachep, pa);
2709 mb_debug(1, "mballoc: %u PAs left\n", count);
2713 int ext4_mb_release(struct super_block *sb)
2715 ext4_group_t ngroups = ext4_get_groups_count(sb);
2717 int num_meta_group_infos;
2718 struct ext4_group_info *grinfo;
2719 struct ext4_sb_info *sbi = EXT4_SB(sb);
2720 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2722 if (sbi->s_group_info) {
2723 for (i = 0; i < ngroups; i++) {
2724 grinfo = ext4_get_group_info(sb, i);
2726 kfree(grinfo->bb_bitmap);
2728 ext4_lock_group(sb, i);
2729 ext4_mb_cleanup_pa(grinfo);
2730 ext4_unlock_group(sb, i);
2731 kmem_cache_free(cachep, grinfo);
2733 num_meta_group_infos = (ngroups +
2734 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2735 EXT4_DESC_PER_BLOCK_BITS(sb);
2736 for (i = 0; i < num_meta_group_infos; i++)
2737 kfree(sbi->s_group_info[i]);
2738 kvfree(sbi->s_group_info);
2740 kfree(sbi->s_mb_offsets);
2741 kfree(sbi->s_mb_maxs);
2742 iput(sbi->s_buddy_cache);
2743 if (sbi->s_mb_stats) {
2744 ext4_msg(sb, KERN_INFO,
2745 "mballoc: %u blocks %u reqs (%u success)",
2746 atomic_read(&sbi->s_bal_allocated),
2747 atomic_read(&sbi->s_bal_reqs),
2748 atomic_read(&sbi->s_bal_success));
2749 ext4_msg(sb, KERN_INFO,
2750 "mballoc: %u extents scanned, %u goal hits, "
2751 "%u 2^N hits, %u breaks, %u lost",
2752 atomic_read(&sbi->s_bal_ex_scanned),
2753 atomic_read(&sbi->s_bal_goals),
2754 atomic_read(&sbi->s_bal_2orders),
2755 atomic_read(&sbi->s_bal_breaks),
2756 atomic_read(&sbi->s_mb_lost_chunks));
2757 ext4_msg(sb, KERN_INFO,
2758 "mballoc: %lu generated and it took %Lu",
2759 sbi->s_mb_buddies_generated,
2760 sbi->s_mb_generation_time);
2761 ext4_msg(sb, KERN_INFO,
2762 "mballoc: %u preallocated, %u discarded",
2763 atomic_read(&sbi->s_mb_preallocated),
2764 atomic_read(&sbi->s_mb_discarded));
2767 free_percpu(sbi->s_locality_groups);
2772 static inline int ext4_issue_discard(struct super_block *sb,
2773 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
2774 unsigned long flags)
2776 ext4_fsblk_t discard_block;
2778 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2779 ext4_group_first_block_no(sb, block_group));
2780 count = EXT4_C2B(EXT4_SB(sb), count);
2781 trace_ext4_discard_blocks(sb,
2782 (unsigned long long) discard_block, count);
2783 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, flags);
2787 * This function is called by the jbd2 layer once the commit has finished,
2788 * so we know we can free the blocks that were released with that commit.
2790 static void ext4_free_data_callback(struct super_block *sb,
2791 struct ext4_journal_cb_entry *jce,
2794 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2795 struct ext4_buddy e4b;
2796 struct ext4_group_info *db;
2797 int err, count = 0, count2 = 0;
2799 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2800 entry->efd_count, entry->efd_group, entry);
2802 if (test_opt(sb, DISCARD)) {
2803 err = ext4_issue_discard(sb, entry->efd_group,
2804 entry->efd_start_cluster,
2805 entry->efd_count, 0);
2806 if (err && err != -EOPNOTSUPP)
2807 ext4_msg(sb, KERN_WARNING, "discard request in"
2808 " group:%d block:%d count:%d failed"
2809 " with %d", entry->efd_group,
2810 entry->efd_start_cluster,
2811 entry->efd_count, err);
2814 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2815 /* we expect to find existing buddy because it's pinned */
2820 /* there are blocks to put in buddy to make them really free */
2821 count += entry->efd_count;
2823 ext4_lock_group(sb, entry->efd_group);
2824 /* Take it out of per group rb tree */
2825 rb_erase(&entry->efd_node, &(db->bb_free_root));
2826 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2829 * Clear the trimmed flag for the group so that the next
2830 * ext4_trim_fs can trim it.
2831 * If the volume is mounted with -o discard, online discard
2832 * is supported and the free blocks will be trimmed online.
2834 if (!test_opt(sb, DISCARD))
2835 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2837 if (!db->bb_free_root.rb_node) {
2838 /* No more items in the per group rb tree
2839 * balance refcounts from ext4_mb_free_metadata()
2841 page_cache_release(e4b.bd_buddy_page);
2842 page_cache_release(e4b.bd_bitmap_page);
2844 ext4_unlock_group(sb, entry->efd_group);
2845 kmem_cache_free(ext4_free_data_cachep, entry);
2846 ext4_mb_unload_buddy(&e4b);
2848 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2851 int __init ext4_init_mballoc(void)
2853 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2854 SLAB_RECLAIM_ACCOUNT);
2855 if (ext4_pspace_cachep == NULL)
2858 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2859 SLAB_RECLAIM_ACCOUNT);
2860 if (ext4_ac_cachep == NULL) {
2861 kmem_cache_destroy(ext4_pspace_cachep);
2865 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2866 SLAB_RECLAIM_ACCOUNT);
2867 if (ext4_free_data_cachep == NULL) {
2868 kmem_cache_destroy(ext4_pspace_cachep);
2869 kmem_cache_destroy(ext4_ac_cachep);
2875 void ext4_exit_mballoc(void)
2878 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2879 * before destroying the slab cache.
2882 kmem_cache_destroy(ext4_pspace_cachep);
2883 kmem_cache_destroy(ext4_ac_cachep);
2884 kmem_cache_destroy(ext4_free_data_cachep);
2885 ext4_groupinfo_destroy_slabs();
2890 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2891 * Returns 0 if success or error code
2893 static noinline_for_stack int
2894 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2895 handle_t *handle, unsigned int reserv_clstrs)
2897 struct buffer_head *bitmap_bh = NULL;
2898 struct ext4_group_desc *gdp;
2899 struct buffer_head *gdp_bh;
2900 struct ext4_sb_info *sbi;
2901 struct super_block *sb;
2905 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2906 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2911 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2912 if (IS_ERR(bitmap_bh)) {
2913 err = PTR_ERR(bitmap_bh);
2918 BUFFER_TRACE(bitmap_bh, "getting write access");
2919 err = ext4_journal_get_write_access(handle, bitmap_bh);
2924 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2928 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2929 ext4_free_group_clusters(sb, gdp));
2931 BUFFER_TRACE(gdp_bh, "get_write_access");
2932 err = ext4_journal_get_write_access(handle, gdp_bh);
2936 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2938 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2939 if (!ext4_data_block_valid(sbi, block, len)) {
2940 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2941 "fs metadata", block, block+len);
2942 /* File system mounted not to panic on error
2943 * Fix the bitmap and return EFSCORRUPTED
2944 * We leak some of the blocks here.
2946 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2947 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2948 ac->ac_b_ex.fe_len);
2949 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2950 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2952 err = -EFSCORRUPTED;
2956 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2957 #ifdef AGGRESSIVE_CHECK
2960 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2961 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2962 bitmap_bh->b_data));
2966 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2967 ac->ac_b_ex.fe_len);
2968 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2969 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2970 ext4_free_group_clusters_set(sb, gdp,
2971 ext4_free_clusters_after_init(sb,
2972 ac->ac_b_ex.fe_group, gdp));
2974 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2975 ext4_free_group_clusters_set(sb, gdp, len);
2976 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2977 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2979 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2980 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2982 * Now reduce the dirty block count also. Should not go negative
2984 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2985 /* release all the reserved blocks if non delalloc */
2986 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2989 if (sbi->s_log_groups_per_flex) {
2990 ext4_group_t flex_group = ext4_flex_group(sbi,
2991 ac->ac_b_ex.fe_group);
2992 atomic64_sub(ac->ac_b_ex.fe_len,
2993 &sbi->s_flex_groups[flex_group].free_clusters);
2996 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2999 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3007 * here we normalize request for locality group
3008 * Group request are normalized to s_mb_group_prealloc, which goes to
3009 * s_strip if we set the same via mount option.
3010 * s_mb_group_prealloc can be configured via
3011 * /sys/fs/ext4/<partition>/mb_group_prealloc
3013 * XXX: should we try to preallocate more than the group has now?
3015 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3017 struct super_block *sb = ac->ac_sb;
3018 struct ext4_locality_group *lg = ac->ac_lg;
3021 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3022 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3023 current->pid, ac->ac_g_ex.fe_len);
3027 * Normalization means making request better in terms of
3028 * size and alignment
3030 static noinline_for_stack void
3031 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3032 struct ext4_allocation_request *ar)
3034 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3037 loff_t size, start_off;
3038 loff_t orig_size __maybe_unused;
3040 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3041 struct ext4_prealloc_space *pa;
3043 /* do normalize only data requests, metadata requests
3044 do not need preallocation */
3045 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3048 /* sometime caller may want exact blocks */
3049 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3052 /* caller may indicate that preallocation isn't
3053 * required (it's a tail, for example) */
3054 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3057 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3058 ext4_mb_normalize_group_request(ac);
3062 bsbits = ac->ac_sb->s_blocksize_bits;
3064 /* first, let's learn actual file size
3065 * given current request is allocated */
3066 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3067 size = size << bsbits;
3068 if (size < i_size_read(ac->ac_inode))
3069 size = i_size_read(ac->ac_inode);
3072 /* max size of free chunks */
3075 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3076 (req <= (size) || max <= (chunk_size))
3078 /* first, try to predict filesize */
3079 /* XXX: should this table be tunable? */
3081 if (size <= 16 * 1024) {
3083 } else if (size <= 32 * 1024) {
3085 } else if (size <= 64 * 1024) {
3087 } else if (size <= 128 * 1024) {
3089 } else if (size <= 256 * 1024) {
3091 } else if (size <= 512 * 1024) {
3093 } else if (size <= 1024 * 1024) {
3095 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3096 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3097 (21 - bsbits)) << 21;
3098 size = 2 * 1024 * 1024;
3099 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3100 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3101 (22 - bsbits)) << 22;
3102 size = 4 * 1024 * 1024;
3103 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3104 (8<<20)>>bsbits, max, 8 * 1024)) {
3105 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3106 (23 - bsbits)) << 23;
3107 size = 8 * 1024 * 1024;
3109 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3110 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3111 ac->ac_o_ex.fe_len) << bsbits;
3113 size = size >> bsbits;
3114 start = start_off >> bsbits;
3116 /* don't cover already allocated blocks in selected range */
3117 if (ar->pleft && start <= ar->lleft) {
3118 size -= ar->lleft + 1 - start;
3119 start = ar->lleft + 1;
3121 if (ar->pright && start + size - 1 >= ar->lright)
3122 size -= start + size - ar->lright;
3125 * Trim allocation request for filesystems with artificially small
3128 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
3129 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
3133 /* check we don't cross already preallocated blocks */
3135 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3140 spin_lock(&pa->pa_lock);
3141 if (pa->pa_deleted) {
3142 spin_unlock(&pa->pa_lock);
3146 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3149 /* PA must not overlap original request */
3150 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3151 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3153 /* skip PAs this normalized request doesn't overlap with */
3154 if (pa->pa_lstart >= end || pa_end <= start) {
3155 spin_unlock(&pa->pa_lock);
3158 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3160 /* adjust start or end to be adjacent to this pa */
3161 if (pa_end <= ac->ac_o_ex.fe_logical) {
3162 BUG_ON(pa_end < start);
3164 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3165 BUG_ON(pa->pa_lstart > end);
3166 end = pa->pa_lstart;
3168 spin_unlock(&pa->pa_lock);
3173 /* XXX: extra loop to check we really don't overlap preallocations */
3175 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3178 spin_lock(&pa->pa_lock);
3179 if (pa->pa_deleted == 0) {
3180 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3182 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3184 spin_unlock(&pa->pa_lock);
3188 if (start + size <= ac->ac_o_ex.fe_logical &&
3189 start > ac->ac_o_ex.fe_logical) {
3190 ext4_msg(ac->ac_sb, KERN_ERR,
3191 "start %lu, size %lu, fe_logical %lu",
3192 (unsigned long) start, (unsigned long) size,
3193 (unsigned long) ac->ac_o_ex.fe_logical);
3196 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3198 /* now prepare goal request */
3200 /* XXX: is it better to align blocks WRT to logical
3201 * placement or satisfy big request as is */
3202 ac->ac_g_ex.fe_logical = start;
3203 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3205 /* define goal start in order to merge */
3206 if (ar->pright && (ar->lright == (start + size))) {
3207 /* merge to the right */
3208 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3209 &ac->ac_f_ex.fe_group,
3210 &ac->ac_f_ex.fe_start);
3211 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3213 if (ar->pleft && (ar->lleft + 1 == start)) {
3214 /* merge to the left */
3215 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3216 &ac->ac_f_ex.fe_group,
3217 &ac->ac_f_ex.fe_start);
3218 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3221 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3222 (unsigned) orig_size, (unsigned) start);
3225 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3227 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3229 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3230 atomic_inc(&sbi->s_bal_reqs);
3231 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3232 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3233 atomic_inc(&sbi->s_bal_success);
3234 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3235 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3236 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3237 atomic_inc(&sbi->s_bal_goals);
3238 if (ac->ac_found > sbi->s_mb_max_to_scan)
3239 atomic_inc(&sbi->s_bal_breaks);
3242 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3243 trace_ext4_mballoc_alloc(ac);
3245 trace_ext4_mballoc_prealloc(ac);
3249 * Called on failure; free up any blocks from the inode PA for this
3250 * context. We don't need this for MB_GROUP_PA because we only change
3251 * pa_free in ext4_mb_release_context(), but on failure, we've already
3252 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3254 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3256 struct ext4_prealloc_space *pa = ac->ac_pa;
3257 struct ext4_buddy e4b;
3261 if (ac->ac_f_ex.fe_len == 0)
3263 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3266 * This should never happen since we pin the
3267 * pages in the ext4_allocation_context so
3268 * ext4_mb_load_buddy() should never fail.
3270 WARN(1, "mb_load_buddy failed (%d)", err);
3273 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3274 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3275 ac->ac_f_ex.fe_len);
3276 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3277 ext4_mb_unload_buddy(&e4b);
3280 if (pa->pa_type == MB_INODE_PA)
3281 pa->pa_free += ac->ac_b_ex.fe_len;
3285 * use blocks preallocated to inode
3287 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3288 struct ext4_prealloc_space *pa)
3290 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3295 /* found preallocated blocks, use them */
3296 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3297 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3298 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3299 len = EXT4_NUM_B2C(sbi, end - start);
3300 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3301 &ac->ac_b_ex.fe_start);
3302 ac->ac_b_ex.fe_len = len;
3303 ac->ac_status = AC_STATUS_FOUND;
3306 BUG_ON(start < pa->pa_pstart);
3307 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3308 BUG_ON(pa->pa_free < len);
3311 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3315 * use blocks preallocated to locality group
3317 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3318 struct ext4_prealloc_space *pa)
3320 unsigned int len = ac->ac_o_ex.fe_len;
3322 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3323 &ac->ac_b_ex.fe_group,
3324 &ac->ac_b_ex.fe_start);
3325 ac->ac_b_ex.fe_len = len;
3326 ac->ac_status = AC_STATUS_FOUND;
3329 /* we don't correct pa_pstart or pa_plen here to avoid
3330 * possible race when the group is being loaded concurrently
3331 * instead we correct pa later, after blocks are marked
3332 * in on-disk bitmap -- see ext4_mb_release_context()
3333 * Other CPUs are prevented from allocating from this pa by lg_mutex
3335 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3339 * Return the prealloc space that have minimal distance
3340 * from the goal block. @cpa is the prealloc
3341 * space that is having currently known minimal distance
3342 * from the goal block.
3344 static struct ext4_prealloc_space *
3345 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3346 struct ext4_prealloc_space *pa,
3347 struct ext4_prealloc_space *cpa)
3349 ext4_fsblk_t cur_distance, new_distance;
3352 atomic_inc(&pa->pa_count);
3355 cur_distance = abs(goal_block - cpa->pa_pstart);
3356 new_distance = abs(goal_block - pa->pa_pstart);
3358 if (cur_distance <= new_distance)
3361 /* drop the previous reference */
3362 atomic_dec(&cpa->pa_count);
3363 atomic_inc(&pa->pa_count);
3368 * search goal blocks in preallocated space
3370 static noinline_for_stack int
3371 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3373 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3375 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3376 struct ext4_locality_group *lg;
3377 struct ext4_prealloc_space *pa, *cpa = NULL;
3378 ext4_fsblk_t goal_block;
3380 /* only data can be preallocated */
3381 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3384 /* first, try per-file preallocation */
3386 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3388 /* all fields in this condition don't change,
3389 * so we can skip locking for them */
3390 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3391 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3392 EXT4_C2B(sbi, pa->pa_len)))
3395 /* non-extent files can't have physical blocks past 2^32 */
3396 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3397 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3398 EXT4_MAX_BLOCK_FILE_PHYS))
3401 /* found preallocated blocks, use them */
3402 spin_lock(&pa->pa_lock);
3403 if (pa->pa_deleted == 0 && pa->pa_free) {
3404 atomic_inc(&pa->pa_count);
3405 ext4_mb_use_inode_pa(ac, pa);
3406 spin_unlock(&pa->pa_lock);
3407 ac->ac_criteria = 10;
3411 spin_unlock(&pa->pa_lock);
3415 /* can we use group allocation? */
3416 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3419 /* inode may have no locality group for some reason */
3423 order = fls(ac->ac_o_ex.fe_len) - 1;
3424 if (order > PREALLOC_TB_SIZE - 1)
3425 /* The max size of hash table is PREALLOC_TB_SIZE */
3426 order = PREALLOC_TB_SIZE - 1;
3428 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3430 * search for the prealloc space that is having
3431 * minimal distance from the goal block.
3433 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3435 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3437 spin_lock(&pa->pa_lock);
3438 if (pa->pa_deleted == 0 &&
3439 pa->pa_free >= ac->ac_o_ex.fe_len) {
3441 cpa = ext4_mb_check_group_pa(goal_block,
3444 spin_unlock(&pa->pa_lock);
3449 ext4_mb_use_group_pa(ac, cpa);
3450 ac->ac_criteria = 20;
3457 * the function goes through all block freed in the group
3458 * but not yet committed and marks them used in in-core bitmap.
3459 * buddy must be generated from this bitmap
3460 * Need to be called with the ext4 group lock held
3462 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3466 struct ext4_group_info *grp;
3467 struct ext4_free_data *entry;
3469 grp = ext4_get_group_info(sb, group);
3470 n = rb_first(&(grp->bb_free_root));
3473 entry = rb_entry(n, struct ext4_free_data, efd_node);
3474 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3481 * the function goes through all preallocation in this group and marks them
3482 * used in in-core bitmap. buddy must be generated from this bitmap
3483 * Need to be called with ext4 group lock held
3485 static noinline_for_stack
3486 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3489 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3490 struct ext4_prealloc_space *pa;
3491 struct list_head *cur;
3492 ext4_group_t groupnr;
3493 ext4_grpblk_t start;
3494 int preallocated = 0;
3497 /* all form of preallocation discards first load group,
3498 * so the only competing code is preallocation use.
3499 * we don't need any locking here
3500 * notice we do NOT ignore preallocations with pa_deleted
3501 * otherwise we could leave used blocks available for
3502 * allocation in buddy when concurrent ext4_mb_put_pa()
3503 * is dropping preallocation
3505 list_for_each(cur, &grp->bb_prealloc_list) {
3506 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3507 spin_lock(&pa->pa_lock);
3508 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3511 spin_unlock(&pa->pa_lock);
3512 if (unlikely(len == 0))
3514 BUG_ON(groupnr != group);
3515 ext4_set_bits(bitmap, start, len);
3516 preallocated += len;
3518 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3521 static void ext4_mb_pa_callback(struct rcu_head *head)
3523 struct ext4_prealloc_space *pa;
3524 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3526 BUG_ON(atomic_read(&pa->pa_count));
3527 BUG_ON(pa->pa_deleted == 0);
3528 kmem_cache_free(ext4_pspace_cachep, pa);
3532 * drops a reference to preallocated space descriptor
3533 * if this was the last reference and the space is consumed
3535 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3536 struct super_block *sb, struct ext4_prealloc_space *pa)
3539 ext4_fsblk_t grp_blk;
3541 /* in this short window concurrent discard can set pa_deleted */
3542 spin_lock(&pa->pa_lock);
3543 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3544 spin_unlock(&pa->pa_lock);
3548 if (pa->pa_deleted == 1) {
3549 spin_unlock(&pa->pa_lock);
3554 spin_unlock(&pa->pa_lock);
3556 grp_blk = pa->pa_pstart;
3558 * If doing group-based preallocation, pa_pstart may be in the
3559 * next group when pa is used up
3561 if (pa->pa_type == MB_GROUP_PA)
3564 grp = ext4_get_group_number(sb, grp_blk);
3569 * P1 (buddy init) P2 (regular allocation)
3570 * find block B in PA
3571 * copy on-disk bitmap to buddy
3572 * mark B in on-disk bitmap
3573 * drop PA from group
3574 * mark all PAs in buddy
3576 * thus, P1 initializes buddy with B available. to prevent this
3577 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3580 ext4_lock_group(sb, grp);
3581 list_del(&pa->pa_group_list);
3582 ext4_unlock_group(sb, grp);
3584 spin_lock(pa->pa_obj_lock);
3585 list_del_rcu(&pa->pa_inode_list);
3586 spin_unlock(pa->pa_obj_lock);
3588 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3592 * creates new preallocated space for given inode
3594 static noinline_for_stack int
3595 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3597 struct super_block *sb = ac->ac_sb;
3598 struct ext4_sb_info *sbi = EXT4_SB(sb);
3599 struct ext4_prealloc_space *pa;
3600 struct ext4_group_info *grp;
3601 struct ext4_inode_info *ei;
3603 /* preallocate only when found space is larger then requested */
3604 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3605 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3606 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3608 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3612 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3618 /* we can't allocate as much as normalizer wants.
3619 * so, found space must get proper lstart
3620 * to cover original request */
3621 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3622 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3624 /* we're limited by original request in that
3625 * logical block must be covered any way
3626 * winl is window we can move our chunk within */
3627 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3629 /* also, we should cover whole original request */
3630 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3632 /* the smallest one defines real window */
3633 win = min(winl, wins);
3635 offs = ac->ac_o_ex.fe_logical %
3636 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3637 if (offs && offs < win)
3640 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3641 EXT4_NUM_B2C(sbi, win);
3642 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3643 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3646 /* preallocation can change ac_b_ex, thus we store actually
3647 * allocated blocks for history */
3648 ac->ac_f_ex = ac->ac_b_ex;
3650 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3651 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3652 pa->pa_len = ac->ac_b_ex.fe_len;
3653 pa->pa_free = pa->pa_len;
3654 atomic_set(&pa->pa_count, 1);
3655 spin_lock_init(&pa->pa_lock);
3656 INIT_LIST_HEAD(&pa->pa_inode_list);
3657 INIT_LIST_HEAD(&pa->pa_group_list);
3659 pa->pa_type = MB_INODE_PA;
3661 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3662 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3663 trace_ext4_mb_new_inode_pa(ac, pa);
3665 ext4_mb_use_inode_pa(ac, pa);
3666 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3668 ei = EXT4_I(ac->ac_inode);
3669 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3671 pa->pa_obj_lock = &ei->i_prealloc_lock;
3672 pa->pa_inode = ac->ac_inode;
3674 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3675 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3676 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3678 spin_lock(pa->pa_obj_lock);
3679 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3680 spin_unlock(pa->pa_obj_lock);
3686 * creates new preallocated space for locality group inodes belongs to
3688 static noinline_for_stack int
3689 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3691 struct super_block *sb = ac->ac_sb;
3692 struct ext4_locality_group *lg;
3693 struct ext4_prealloc_space *pa;
3694 struct ext4_group_info *grp;
3696 /* preallocate only when found space is larger then requested */
3697 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3698 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3699 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3701 BUG_ON(ext4_pspace_cachep == NULL);
3702 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3706 /* preallocation can change ac_b_ex, thus we store actually
3707 * allocated blocks for history */
3708 ac->ac_f_ex = ac->ac_b_ex;
3710 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3711 pa->pa_lstart = pa->pa_pstart;
3712 pa->pa_len = ac->ac_b_ex.fe_len;
3713 pa->pa_free = pa->pa_len;
3714 atomic_set(&pa->pa_count, 1);
3715 spin_lock_init(&pa->pa_lock);
3716 INIT_LIST_HEAD(&pa->pa_inode_list);
3717 INIT_LIST_HEAD(&pa->pa_group_list);
3719 pa->pa_type = MB_GROUP_PA;
3721 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3722 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3723 trace_ext4_mb_new_group_pa(ac, pa);
3725 ext4_mb_use_group_pa(ac, pa);
3726 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3728 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3732 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3733 pa->pa_inode = NULL;
3735 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3736 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3737 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3740 * We will later add the new pa to the right bucket
3741 * after updating the pa_free in ext4_mb_release_context
3746 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3750 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3751 err = ext4_mb_new_group_pa(ac);
3753 err = ext4_mb_new_inode_pa(ac);
3758 * finds all unused blocks in on-disk bitmap, frees them in
3759 * in-core bitmap and buddy.
3760 * @pa must be unlinked from inode and group lists, so that
3761 * nobody else can find/use it.
3762 * the caller MUST hold group/inode locks.
3763 * TODO: optimize the case when there are no in-core structures yet
3765 static noinline_for_stack int
3766 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3767 struct ext4_prealloc_space *pa)
3769 struct super_block *sb = e4b->bd_sb;
3770 struct ext4_sb_info *sbi = EXT4_SB(sb);
3775 unsigned long long grp_blk_start;
3779 BUG_ON(pa->pa_deleted == 0);
3780 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3781 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3782 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3783 end = bit + pa->pa_len;
3786 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3789 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3790 mb_debug(1, " free preallocated %u/%u in group %u\n",
3791 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3792 (unsigned) next - bit, (unsigned) group);
3795 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3796 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3797 EXT4_C2B(sbi, bit)),
3799 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3802 if (free != pa->pa_free) {
3803 ext4_msg(e4b->bd_sb, KERN_CRIT,
3804 "pa %p: logic %lu, phys. %lu, len %lu",
3805 pa, (unsigned long) pa->pa_lstart,
3806 (unsigned long) pa->pa_pstart,
3807 (unsigned long) pa->pa_len);
3808 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3811 * pa is already deleted so we use the value obtained
3812 * from the bitmap and continue.
3815 atomic_add(free, &sbi->s_mb_discarded);
3820 static noinline_for_stack int
3821 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3822 struct ext4_prealloc_space *pa)
3824 struct super_block *sb = e4b->bd_sb;
3828 trace_ext4_mb_release_group_pa(sb, pa);
3829 BUG_ON(pa->pa_deleted == 0);
3830 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3831 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3832 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3833 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3834 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3840 * releases all preallocations in given group
3842 * first, we need to decide discard policy:
3843 * - when do we discard
3845 * - how many do we discard
3846 * 1) how many requested
3848 static noinline_for_stack int
3849 ext4_mb_discard_group_preallocations(struct super_block *sb,
3850 ext4_group_t group, int needed)
3852 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3853 struct buffer_head *bitmap_bh = NULL;
3854 struct ext4_prealloc_space *pa, *tmp;
3855 struct list_head list;
3856 struct ext4_buddy e4b;
3861 mb_debug(1, "discard preallocation for group %u\n", group);
3863 if (list_empty(&grp->bb_prealloc_list))
3866 bitmap_bh = ext4_read_block_bitmap(sb, group);
3867 if (IS_ERR(bitmap_bh)) {
3868 err = PTR_ERR(bitmap_bh);
3869 ext4_error(sb, "Error %d reading block bitmap for %u",
3874 err = ext4_mb_load_buddy(sb, group, &e4b);
3876 ext4_error(sb, "Error loading buddy information for %u", group);
3882 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3884 INIT_LIST_HEAD(&list);
3886 ext4_lock_group(sb, group);
3887 list_for_each_entry_safe(pa, tmp,
3888 &grp->bb_prealloc_list, pa_group_list) {
3889 spin_lock(&pa->pa_lock);
3890 if (atomic_read(&pa->pa_count)) {
3891 spin_unlock(&pa->pa_lock);
3895 if (pa->pa_deleted) {
3896 spin_unlock(&pa->pa_lock);
3900 /* seems this one can be freed ... */
3903 /* we can trust pa_free ... */
3904 free += pa->pa_free;
3906 spin_unlock(&pa->pa_lock);
3908 list_del(&pa->pa_group_list);
3909 list_add(&pa->u.pa_tmp_list, &list);
3912 /* if we still need more blocks and some PAs were used, try again */
3913 if (free < needed && busy) {
3915 ext4_unlock_group(sb, group);
3920 /* found anything to free? */
3921 if (list_empty(&list)) {
3926 /* now free all selected PAs */
3927 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3929 /* remove from object (inode or locality group) */
3930 spin_lock(pa->pa_obj_lock);
3931 list_del_rcu(&pa->pa_inode_list);
3932 spin_unlock(pa->pa_obj_lock);
3934 if (pa->pa_type == MB_GROUP_PA)
3935 ext4_mb_release_group_pa(&e4b, pa);
3937 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3939 list_del(&pa->u.pa_tmp_list);
3940 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3944 ext4_unlock_group(sb, group);
3945 ext4_mb_unload_buddy(&e4b);
3951 * releases all non-used preallocated blocks for given inode
3953 * It's important to discard preallocations under i_data_sem
3954 * We don't want another block to be served from the prealloc
3955 * space when we are discarding the inode prealloc space.
3957 * FIXME!! Make sure it is valid at all the call sites
3959 void ext4_discard_preallocations(struct inode *inode)
3961 struct ext4_inode_info *ei = EXT4_I(inode);
3962 struct super_block *sb = inode->i_sb;
3963 struct buffer_head *bitmap_bh = NULL;
3964 struct ext4_prealloc_space *pa, *tmp;
3965 ext4_group_t group = 0;
3966 struct list_head list;
3967 struct ext4_buddy e4b;
3970 if (!S_ISREG(inode->i_mode)) {
3971 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3975 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3976 trace_ext4_discard_preallocations(inode);
3978 INIT_LIST_HEAD(&list);
3981 /* first, collect all pa's in the inode */
3982 spin_lock(&ei->i_prealloc_lock);
3983 while (!list_empty(&ei->i_prealloc_list)) {
3984 pa = list_entry(ei->i_prealloc_list.next,
3985 struct ext4_prealloc_space, pa_inode_list);
3986 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3987 spin_lock(&pa->pa_lock);
3988 if (atomic_read(&pa->pa_count)) {
3989 /* this shouldn't happen often - nobody should
3990 * use preallocation while we're discarding it */
3991 spin_unlock(&pa->pa_lock);
3992 spin_unlock(&ei->i_prealloc_lock);
3993 ext4_msg(sb, KERN_ERR,
3994 "uh-oh! used pa while discarding");
3996 schedule_timeout_uninterruptible(HZ);
4000 if (pa->pa_deleted == 0) {
4002 spin_unlock(&pa->pa_lock);
4003 list_del_rcu(&pa->pa_inode_list);
4004 list_add(&pa->u.pa_tmp_list, &list);
4008 /* someone is deleting pa right now */
4009 spin_unlock(&pa->pa_lock);
4010 spin_unlock(&ei->i_prealloc_lock);
4012 /* we have to wait here because pa_deleted
4013 * doesn't mean pa is already unlinked from
4014 * the list. as we might be called from
4015 * ->clear_inode() the inode will get freed
4016 * and concurrent thread which is unlinking
4017 * pa from inode's list may access already
4018 * freed memory, bad-bad-bad */
4020 /* XXX: if this happens too often, we can
4021 * add a flag to force wait only in case
4022 * of ->clear_inode(), but not in case of
4023 * regular truncate */
4024 schedule_timeout_uninterruptible(HZ);
4027 spin_unlock(&ei->i_prealloc_lock);
4029 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4030 BUG_ON(pa->pa_type != MB_INODE_PA);
4031 group = ext4_get_group_number(sb, pa->pa_pstart);
4033 err = ext4_mb_load_buddy(sb, group, &e4b);
4035 ext4_error(sb, "Error loading buddy information for %u",
4040 bitmap_bh = ext4_read_block_bitmap(sb, group);
4041 if (IS_ERR(bitmap_bh)) {
4042 err = PTR_ERR(bitmap_bh);
4043 ext4_error(sb, "Error %d reading block bitmap for %u",
4045 ext4_mb_unload_buddy(&e4b);
4049 ext4_lock_group(sb, group);
4050 list_del(&pa->pa_group_list);
4051 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4052 ext4_unlock_group(sb, group);
4054 ext4_mb_unload_buddy(&e4b);
4057 list_del(&pa->u.pa_tmp_list);
4058 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4062 #ifdef CONFIG_EXT4_DEBUG
4063 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4065 struct super_block *sb = ac->ac_sb;
4066 ext4_group_t ngroups, i;
4068 if (!ext4_mballoc_debug ||
4069 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4072 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4073 " Allocation context details:");
4074 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4075 ac->ac_status, ac->ac_flags);
4076 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4077 "goal %lu/%lu/%lu@%lu, "
4078 "best %lu/%lu/%lu@%lu cr %d",
4079 (unsigned long)ac->ac_o_ex.fe_group,
4080 (unsigned long)ac->ac_o_ex.fe_start,
4081 (unsigned long)ac->ac_o_ex.fe_len,
4082 (unsigned long)ac->ac_o_ex.fe_logical,
4083 (unsigned long)ac->ac_g_ex.fe_group,
4084 (unsigned long)ac->ac_g_ex.fe_start,
4085 (unsigned long)ac->ac_g_ex.fe_len,
4086 (unsigned long)ac->ac_g_ex.fe_logical,
4087 (unsigned long)ac->ac_b_ex.fe_group,
4088 (unsigned long)ac->ac_b_ex.fe_start,
4089 (unsigned long)ac->ac_b_ex.fe_len,
4090 (unsigned long)ac->ac_b_ex.fe_logical,
4091 (int)ac->ac_criteria);
4092 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4093 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4094 ngroups = ext4_get_groups_count(sb);
4095 for (i = 0; i < ngroups; i++) {
4096 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4097 struct ext4_prealloc_space *pa;
4098 ext4_grpblk_t start;
4099 struct list_head *cur;
4100 ext4_lock_group(sb, i);
4101 list_for_each(cur, &grp->bb_prealloc_list) {
4102 pa = list_entry(cur, struct ext4_prealloc_space,
4104 spin_lock(&pa->pa_lock);
4105 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4107 spin_unlock(&pa->pa_lock);
4108 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4111 ext4_unlock_group(sb, i);
4113 if (grp->bb_free == 0)
4115 printk(KERN_ERR "%u: %d/%d \n",
4116 i, grp->bb_free, grp->bb_fragments);
4118 printk(KERN_ERR "\n");
4121 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4128 * We use locality group preallocation for small size file. The size of the
4129 * file is determined by the current size or the resulting size after
4130 * allocation which ever is larger
4132 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4134 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4136 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4137 int bsbits = ac->ac_sb->s_blocksize_bits;
4140 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4143 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4146 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4147 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4150 if ((size == isize) &&
4151 !ext4_fs_is_busy(sbi) &&
4152 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4153 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4157 if (sbi->s_mb_group_prealloc <= 0) {
4158 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4162 /* don't use group allocation for large files */
4163 size = max(size, isize);
4164 if (size > sbi->s_mb_stream_request) {
4165 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4169 BUG_ON(ac->ac_lg != NULL);
4171 * locality group prealloc space are per cpu. The reason for having
4172 * per cpu locality group is to reduce the contention between block
4173 * request from multiple CPUs.
4175 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4177 /* we're going to use group allocation */
4178 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4180 /* serialize all allocations in the group */
4181 mutex_lock(&ac->ac_lg->lg_mutex);
4184 static noinline_for_stack int
4185 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4186 struct ext4_allocation_request *ar)
4188 struct super_block *sb = ar->inode->i_sb;
4189 struct ext4_sb_info *sbi = EXT4_SB(sb);
4190 struct ext4_super_block *es = sbi->s_es;
4194 ext4_grpblk_t block;
4196 /* we can't allocate > group size */
4199 /* just a dirty hack to filter too big requests */
4200 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4201 len = EXT4_CLUSTERS_PER_GROUP(sb);
4203 /* start searching from the goal */
4205 if (goal < le32_to_cpu(es->s_first_data_block) ||
4206 goal >= ext4_blocks_count(es))
4207 goal = le32_to_cpu(es->s_first_data_block);
4208 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4210 /* set up allocation goals */
4211 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4212 ac->ac_status = AC_STATUS_CONTINUE;
4214 ac->ac_inode = ar->inode;
4215 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4216 ac->ac_o_ex.fe_group = group;
4217 ac->ac_o_ex.fe_start = block;
4218 ac->ac_o_ex.fe_len = len;
4219 ac->ac_g_ex = ac->ac_o_ex;
4220 ac->ac_flags = ar->flags;
4222 /* we have to define context: we'll we work with a file or
4223 * locality group. this is a policy, actually */
4224 ext4_mb_group_or_file(ac);
4226 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4227 "left: %u/%u, right %u/%u to %swritable\n",
4228 (unsigned) ar->len, (unsigned) ar->logical,
4229 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4230 (unsigned) ar->lleft, (unsigned) ar->pleft,
4231 (unsigned) ar->lright, (unsigned) ar->pright,
4232 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4237 static noinline_for_stack void
4238 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4239 struct ext4_locality_group *lg,
4240 int order, int total_entries)
4242 ext4_group_t group = 0;
4243 struct ext4_buddy e4b;
4244 struct list_head discard_list;
4245 struct ext4_prealloc_space *pa, *tmp;
4247 mb_debug(1, "discard locality group preallocation\n");
4249 INIT_LIST_HEAD(&discard_list);
4251 spin_lock(&lg->lg_prealloc_lock);
4252 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4254 spin_lock(&pa->pa_lock);
4255 if (atomic_read(&pa->pa_count)) {
4257 * This is the pa that we just used
4258 * for block allocation. So don't
4261 spin_unlock(&pa->pa_lock);
4264 if (pa->pa_deleted) {
4265 spin_unlock(&pa->pa_lock);
4268 /* only lg prealloc space */
4269 BUG_ON(pa->pa_type != MB_GROUP_PA);
4271 /* seems this one can be freed ... */
4273 spin_unlock(&pa->pa_lock);
4275 list_del_rcu(&pa->pa_inode_list);
4276 list_add(&pa->u.pa_tmp_list, &discard_list);
4279 if (total_entries <= 5) {
4281 * we want to keep only 5 entries
4282 * allowing it to grow to 8. This
4283 * mak sure we don't call discard
4284 * soon for this list.
4289 spin_unlock(&lg->lg_prealloc_lock);
4291 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4293 group = ext4_get_group_number(sb, pa->pa_pstart);
4294 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4295 ext4_error(sb, "Error loading buddy information for %u",
4299 ext4_lock_group(sb, group);
4300 list_del(&pa->pa_group_list);
4301 ext4_mb_release_group_pa(&e4b, pa);
4302 ext4_unlock_group(sb, group);
4304 ext4_mb_unload_buddy(&e4b);
4305 list_del(&pa->u.pa_tmp_list);
4306 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4311 * We have incremented pa_count. So it cannot be freed at this
4312 * point. Also we hold lg_mutex. So no parallel allocation is
4313 * possible from this lg. That means pa_free cannot be updated.
4315 * A parallel ext4_mb_discard_group_preallocations is possible.
4316 * which can cause the lg_prealloc_list to be updated.
4319 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4321 int order, added = 0, lg_prealloc_count = 1;
4322 struct super_block *sb = ac->ac_sb;
4323 struct ext4_locality_group *lg = ac->ac_lg;
4324 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4326 order = fls(pa->pa_free) - 1;
4327 if (order > PREALLOC_TB_SIZE - 1)
4328 /* The max size of hash table is PREALLOC_TB_SIZE */
4329 order = PREALLOC_TB_SIZE - 1;
4330 /* Add the prealloc space to lg */
4331 spin_lock(&lg->lg_prealloc_lock);
4332 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4334 spin_lock(&tmp_pa->pa_lock);
4335 if (tmp_pa->pa_deleted) {
4336 spin_unlock(&tmp_pa->pa_lock);
4339 if (!added && pa->pa_free < tmp_pa->pa_free) {
4340 /* Add to the tail of the previous entry */
4341 list_add_tail_rcu(&pa->pa_inode_list,
4342 &tmp_pa->pa_inode_list);
4345 * we want to count the total
4346 * number of entries in the list
4349 spin_unlock(&tmp_pa->pa_lock);
4350 lg_prealloc_count++;
4353 list_add_tail_rcu(&pa->pa_inode_list,
4354 &lg->lg_prealloc_list[order]);
4355 spin_unlock(&lg->lg_prealloc_lock);
4357 /* Now trim the list to be not more than 8 elements */
4358 if (lg_prealloc_count > 8) {
4359 ext4_mb_discard_lg_preallocations(sb, lg,
4360 order, lg_prealloc_count);
4367 * release all resource we used in allocation
4369 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4371 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4372 struct ext4_prealloc_space *pa = ac->ac_pa;
4374 if (pa->pa_type == MB_GROUP_PA) {
4375 /* see comment in ext4_mb_use_group_pa() */
4376 spin_lock(&pa->pa_lock);
4377 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4378 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4379 pa->pa_free -= ac->ac_b_ex.fe_len;
4380 pa->pa_len -= ac->ac_b_ex.fe_len;
4381 spin_unlock(&pa->pa_lock);
4386 * We want to add the pa to the right bucket.
4387 * Remove it from the list and while adding
4388 * make sure the list to which we are adding
4391 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4392 spin_lock(pa->pa_obj_lock);
4393 list_del_rcu(&pa->pa_inode_list);
4394 spin_unlock(pa->pa_obj_lock);
4395 ext4_mb_add_n_trim(ac);
4397 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4399 if (ac->ac_bitmap_page)
4400 page_cache_release(ac->ac_bitmap_page);
4401 if (ac->ac_buddy_page)
4402 page_cache_release(ac->ac_buddy_page);
4403 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4404 mutex_unlock(&ac->ac_lg->lg_mutex);
4405 ext4_mb_collect_stats(ac);
4409 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4411 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4415 trace_ext4_mb_discard_preallocations(sb, needed);
4416 for (i = 0; i < ngroups && needed > 0; i++) {
4417 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4426 * Main entry point into mballoc to allocate blocks
4427 * it tries to use preallocation first, then falls back
4428 * to usual allocation
4430 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4431 struct ext4_allocation_request *ar, int *errp)
4434 struct ext4_allocation_context *ac = NULL;
4435 struct ext4_sb_info *sbi;
4436 struct super_block *sb;
4437 ext4_fsblk_t block = 0;
4438 unsigned int inquota = 0;
4439 unsigned int reserv_clstrs = 0;
4442 sb = ar->inode->i_sb;
4445 trace_ext4_request_blocks(ar);
4447 /* Allow to use superuser reservation for quota file */
4448 if (IS_NOQUOTA(ar->inode))
4449 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4451 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4452 /* Without delayed allocation we need to verify
4453 * there is enough free blocks to do block allocation
4454 * and verify allocation doesn't exceed the quota limits.
4457 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4459 /* let others to free the space */
4461 ar->len = ar->len >> 1;
4467 reserv_clstrs = ar->len;
4468 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4469 dquot_alloc_block_nofail(ar->inode,
4470 EXT4_C2B(sbi, ar->len));
4473 dquot_alloc_block(ar->inode,
4474 EXT4_C2B(sbi, ar->len))) {
4476 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4487 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4494 *errp = ext4_mb_initialize_context(ac, ar);
4500 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4501 if (!ext4_mb_use_preallocated(ac)) {
4502 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4503 ext4_mb_normalize_request(ac, ar);
4505 /* allocate space in core */
4506 *errp = ext4_mb_regular_allocator(ac);
4508 goto discard_and_exit;
4510 /* as we've just preallocated more space than
4511 * user requested originally, we store allocated
4512 * space in a special descriptor */
4513 if (ac->ac_status == AC_STATUS_FOUND &&
4514 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4515 *errp = ext4_mb_new_preallocation(ac);
4518 ext4_discard_allocated_blocks(ac);
4522 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4523 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4525 ext4_discard_allocated_blocks(ac);
4528 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4529 ar->len = ac->ac_b_ex.fe_len;
4532 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4540 ac->ac_b_ex.fe_len = 0;
4542 ext4_mb_show_ac(ac);
4544 ext4_mb_release_context(ac);
4547 kmem_cache_free(ext4_ac_cachep, ac);
4548 if (inquota && ar->len < inquota)
4549 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4551 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4552 /* release all the reserved blocks if non delalloc */
4553 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4557 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4563 * We can merge two free data extents only if the physical blocks
4564 * are contiguous, AND the extents were freed by the same transaction,
4565 * AND the blocks are associated with the same group.
4567 static int can_merge(struct ext4_free_data *entry1,
4568 struct ext4_free_data *entry2)
4570 if ((entry1->efd_tid == entry2->efd_tid) &&
4571 (entry1->efd_group == entry2->efd_group) &&
4572 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4577 static noinline_for_stack int
4578 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4579 struct ext4_free_data *new_entry)
4581 ext4_group_t group = e4b->bd_group;
4582 ext4_grpblk_t cluster;
4583 struct ext4_free_data *entry;
4584 struct ext4_group_info *db = e4b->bd_info;
4585 struct super_block *sb = e4b->bd_sb;
4586 struct ext4_sb_info *sbi = EXT4_SB(sb);
4587 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4588 struct rb_node *parent = NULL, *new_node;
4590 BUG_ON(!ext4_handle_valid(handle));
4591 BUG_ON(e4b->bd_bitmap_page == NULL);
4592 BUG_ON(e4b->bd_buddy_page == NULL);
4594 new_node = &new_entry->efd_node;
4595 cluster = new_entry->efd_start_cluster;
4598 /* first free block exent. We need to
4599 protect buddy cache from being freed,
4600 * otherwise we'll refresh it from
4601 * on-disk bitmap and lose not-yet-available
4603 page_cache_get(e4b->bd_buddy_page);
4604 page_cache_get(e4b->bd_bitmap_page);
4608 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4609 if (cluster < entry->efd_start_cluster)
4611 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4612 n = &(*n)->rb_right;
4614 ext4_grp_locked_error(sb, group, 0,
4615 ext4_group_first_block_no(sb, group) +
4616 EXT4_C2B(sbi, cluster),
4617 "Block already on to-be-freed list");
4622 rb_link_node(new_node, parent, n);
4623 rb_insert_color(new_node, &db->bb_free_root);
4625 /* Now try to see the extent can be merged to left and right */
4626 node = rb_prev(new_node);
4628 entry = rb_entry(node, struct ext4_free_data, efd_node);
4629 if (can_merge(entry, new_entry) &&
4630 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4631 new_entry->efd_start_cluster = entry->efd_start_cluster;
4632 new_entry->efd_count += entry->efd_count;
4633 rb_erase(node, &(db->bb_free_root));
4634 kmem_cache_free(ext4_free_data_cachep, entry);
4638 node = rb_next(new_node);
4640 entry = rb_entry(node, struct ext4_free_data, efd_node);
4641 if (can_merge(new_entry, entry) &&
4642 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4643 new_entry->efd_count += entry->efd_count;
4644 rb_erase(node, &(db->bb_free_root));
4645 kmem_cache_free(ext4_free_data_cachep, entry);
4648 /* Add the extent to transaction's private list */
4649 ext4_journal_callback_add(handle, ext4_free_data_callback,
4650 &new_entry->efd_jce);
4655 * ext4_free_blocks() -- Free given blocks and update quota
4656 * @handle: handle for this transaction
4658 * @block: start physical block to free
4659 * @count: number of blocks to count
4660 * @flags: flags used by ext4_free_blocks
4662 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4663 struct buffer_head *bh, ext4_fsblk_t block,
4664 unsigned long count, int flags)
4666 struct buffer_head *bitmap_bh = NULL;
4667 struct super_block *sb = inode->i_sb;
4668 struct ext4_group_desc *gdp;
4669 unsigned int overflow;
4671 struct buffer_head *gd_bh;
4672 ext4_group_t block_group;
4673 struct ext4_sb_info *sbi;
4674 struct ext4_buddy e4b;
4675 unsigned int count_clusters;
4682 BUG_ON(block != bh->b_blocknr);
4684 block = bh->b_blocknr;
4688 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4689 !ext4_data_block_valid(sbi, block, count)) {
4690 ext4_error(sb, "Freeing blocks not in datazone - "
4691 "block = %llu, count = %lu", block, count);
4695 ext4_debug("freeing block %llu\n", block);
4696 trace_ext4_free_blocks(inode, block, count, flags);
4698 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4701 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4706 * We need to make sure we don't reuse the freed block until
4707 * after the transaction is committed, which we can do by
4708 * treating the block as metadata, below. We make an
4709 * exception if the inode is to be written in writeback mode
4710 * since writeback mode has weak data consistency guarantees.
4712 if (!ext4_should_writeback_data(inode))
4713 flags |= EXT4_FREE_BLOCKS_METADATA;
4716 * If the extent to be freed does not begin on a cluster
4717 * boundary, we need to deal with partial clusters at the
4718 * beginning and end of the extent. Normally we will free
4719 * blocks at the beginning or the end unless we are explicitly
4720 * requested to avoid doing so.
4722 overflow = EXT4_PBLK_COFF(sbi, block);
4724 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4725 overflow = sbi->s_cluster_ratio - overflow;
4727 if (count > overflow)
4736 overflow = EXT4_LBLK_COFF(sbi, count);
4738 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4739 if (count > overflow)
4744 count += sbi->s_cluster_ratio - overflow;
4747 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4750 for (i = 0; i < count; i++) {
4752 bh = sb_find_get_block(inode->i_sb, block + i);
4755 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4756 inode, bh, block + i);
4762 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4764 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4765 ext4_get_group_info(sb, block_group))))
4769 * Check to see if we are freeing blocks across a group
4772 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4773 overflow = EXT4_C2B(sbi, bit) + count -
4774 EXT4_BLOCKS_PER_GROUP(sb);
4777 count_clusters = EXT4_NUM_B2C(sbi, count);
4778 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4779 if (IS_ERR(bitmap_bh)) {
4780 err = PTR_ERR(bitmap_bh);
4784 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4790 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4791 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4792 in_range(block, ext4_inode_table(sb, gdp),
4793 EXT4_SB(sb)->s_itb_per_group) ||
4794 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4795 EXT4_SB(sb)->s_itb_per_group)) {
4797 ext4_error(sb, "Freeing blocks in system zone - "
4798 "Block = %llu, count = %lu", block, count);
4799 /* err = 0. ext4_std_error should be a no op */
4803 BUFFER_TRACE(bitmap_bh, "getting write access");
4804 err = ext4_journal_get_write_access(handle, bitmap_bh);
4809 * We are about to modify some metadata. Call the journal APIs
4810 * to unshare ->b_data if a currently-committing transaction is
4813 BUFFER_TRACE(gd_bh, "get_write_access");
4814 err = ext4_journal_get_write_access(handle, gd_bh);
4817 #ifdef AGGRESSIVE_CHECK
4820 for (i = 0; i < count_clusters; i++)
4821 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4824 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4826 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4827 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
4828 GFP_NOFS|__GFP_NOFAIL);
4832 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4833 struct ext4_free_data *new_entry;
4835 * blocks being freed are metadata. these blocks shouldn't
4836 * be used until this transaction is committed
4838 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4841 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4842 GFP_NOFS|__GFP_NOFAIL);
4843 new_entry->efd_start_cluster = bit;
4844 new_entry->efd_group = block_group;
4845 new_entry->efd_count = count_clusters;
4846 new_entry->efd_tid = handle->h_transaction->t_tid;
4848 ext4_lock_group(sb, block_group);
4849 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4850 ext4_mb_free_metadata(handle, &e4b, new_entry);
4852 /* need to update group_info->bb_free and bitmap
4853 * with group lock held. generate_buddy look at
4854 * them with group lock_held
4856 if (test_opt(sb, DISCARD)) {
4857 err = ext4_issue_discard(sb, block_group, bit, count,
4859 if (err && err != -EOPNOTSUPP)
4860 ext4_msg(sb, KERN_WARNING, "discard request in"
4861 " group:%d block:%d count:%lu failed"
4862 " with %d", block_group, bit, count,
4865 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4867 ext4_lock_group(sb, block_group);
4868 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4869 mb_free_blocks(inode, &e4b, bit, count_clusters);
4872 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4873 ext4_free_group_clusters_set(sb, gdp, ret);
4874 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4875 ext4_group_desc_csum_set(sb, block_group, gdp);
4876 ext4_unlock_group(sb, block_group);
4878 if (sbi->s_log_groups_per_flex) {
4879 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4880 atomic64_add(count_clusters,
4881 &sbi->s_flex_groups[flex_group].free_clusters);
4884 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4885 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4886 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4888 ext4_mb_unload_buddy(&e4b);
4890 /* We dirtied the bitmap block */
4891 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4892 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4894 /* And the group descriptor block */
4895 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4896 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4900 if (overflow && !err) {
4908 ext4_std_error(sb, err);
4913 * ext4_group_add_blocks() -- Add given blocks to an existing group
4914 * @handle: handle to this transaction
4916 * @block: start physical block to add to the block group
4917 * @count: number of blocks to free
4919 * This marks the blocks as free in the bitmap and buddy.
4921 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4922 ext4_fsblk_t block, unsigned long count)
4924 struct buffer_head *bitmap_bh = NULL;
4925 struct buffer_head *gd_bh;
4926 ext4_group_t block_group;
4929 struct ext4_group_desc *desc;
4930 struct ext4_sb_info *sbi = EXT4_SB(sb);
4931 struct ext4_buddy e4b;
4932 int err = 0, ret, blk_free_count;
4933 ext4_grpblk_t blocks_freed;
4935 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4940 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4942 * Check to see if we are freeing blocks across a group
4945 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4946 ext4_warning(sb, "too much blocks added to group %u\n",
4952 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4953 if (IS_ERR(bitmap_bh)) {
4954 err = PTR_ERR(bitmap_bh);
4959 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4965 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4966 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4967 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4968 in_range(block + count - 1, ext4_inode_table(sb, desc),
4969 sbi->s_itb_per_group)) {
4970 ext4_error(sb, "Adding blocks in system zones - "
4971 "Block = %llu, count = %lu",
4977 BUFFER_TRACE(bitmap_bh, "getting write access");
4978 err = ext4_journal_get_write_access(handle, bitmap_bh);
4983 * We are about to modify some metadata. Call the journal APIs
4984 * to unshare ->b_data if a currently-committing transaction is
4987 BUFFER_TRACE(gd_bh, "get_write_access");
4988 err = ext4_journal_get_write_access(handle, gd_bh);
4992 for (i = 0, blocks_freed = 0; i < count; i++) {
4993 BUFFER_TRACE(bitmap_bh, "clear bit");
4994 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4995 ext4_error(sb, "bit already cleared for block %llu",
4996 (ext4_fsblk_t)(block + i));
4997 BUFFER_TRACE(bitmap_bh, "bit already cleared");
5003 err = ext4_mb_load_buddy(sb, block_group, &e4b);
5008 * need to update group_info->bb_free and bitmap
5009 * with group lock held. generate_buddy look at
5010 * them with group lock_held
5012 ext4_lock_group(sb, block_group);
5013 mb_clear_bits(bitmap_bh->b_data, bit, count);
5014 mb_free_blocks(NULL, &e4b, bit, count);
5015 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
5016 ext4_free_group_clusters_set(sb, desc, blk_free_count);
5017 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5018 ext4_group_desc_csum_set(sb, block_group, desc);
5019 ext4_unlock_group(sb, block_group);
5020 percpu_counter_add(&sbi->s_freeclusters_counter,
5021 EXT4_NUM_B2C(sbi, blocks_freed));
5023 if (sbi->s_log_groups_per_flex) {
5024 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5025 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5026 &sbi->s_flex_groups[flex_group].free_clusters);
5029 ext4_mb_unload_buddy(&e4b);
5031 /* We dirtied the bitmap block */
5032 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5033 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5035 /* And the group descriptor block */
5036 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5037 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5043 ext4_std_error(sb, err);
5048 * ext4_trim_extent -- function to TRIM one single free extent in the group
5049 * @sb: super block for the file system
5050 * @start: starting block of the free extent in the alloc. group
5051 * @count: number of blocks to TRIM
5052 * @group: alloc. group we are working with
5053 * @e4b: ext4 buddy for the group
5054 * @blkdev_flags: flags for the block device
5056 * Trim "count" blocks starting at "start" in the "group". To assure that no
5057 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5058 * be called with under the group lock.
5060 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5061 ext4_group_t group, struct ext4_buddy *e4b,
5062 unsigned long blkdev_flags)
5066 struct ext4_free_extent ex;
5069 trace_ext4_trim_extent(sb, group, start, count);
5071 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5073 ex.fe_start = start;
5074 ex.fe_group = group;
5078 * Mark blocks used, so no one can reuse them while
5081 mb_mark_used(e4b, &ex);
5082 ext4_unlock_group(sb, group);
5083 ret = ext4_issue_discard(sb, group, start, count, blkdev_flags);
5084 ext4_lock_group(sb, group);
5085 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5090 * ext4_trim_all_free -- function to trim all free space in alloc. group
5091 * @sb: super block for file system
5092 * @group: group to be trimmed
5093 * @start: first group block to examine
5094 * @max: last group block to examine
5095 * @minblocks: minimum extent block count
5096 * @blkdev_flags: flags for the block device
5098 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5099 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5103 * ext4_trim_all_free walks through group's block bitmap searching for free
5104 * extents. When the free extent is found, mark it as used in group buddy
5105 * bitmap. Then issue a TRIM command on this extent and free the extent in
5106 * the group buddy bitmap. This is done until whole group is scanned.
5108 static ext4_grpblk_t
5109 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5110 ext4_grpblk_t start, ext4_grpblk_t max,
5111 ext4_grpblk_t minblocks, unsigned long blkdev_flags)
5114 ext4_grpblk_t next, count = 0, free_count = 0;
5115 struct ext4_buddy e4b;
5118 trace_ext4_trim_all_free(sb, group, start, max);
5120 ret = ext4_mb_load_buddy(sb, group, &e4b);
5122 ext4_error(sb, "Error in loading buddy "
5123 "information for %u", group);
5126 bitmap = e4b.bd_bitmap;
5128 ext4_lock_group(sb, group);
5129 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5130 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5133 start = (e4b.bd_info->bb_first_free > start) ?
5134 e4b.bd_info->bb_first_free : start;
5136 while (start <= max) {
5137 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5140 next = mb_find_next_bit(bitmap, max + 1, start);
5142 if ((next - start) >= minblocks) {
5143 ret = ext4_trim_extent(sb, start,
5144 next - start, group, &e4b,
5146 if (ret && ret != -EOPNOTSUPP)
5149 count += next - start;
5151 free_count += next - start;
5154 if (fatal_signal_pending(current)) {
5155 count = -ERESTARTSYS;
5159 if (need_resched()) {
5160 ext4_unlock_group(sb, group);
5162 ext4_lock_group(sb, group);
5165 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5171 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5174 ext4_unlock_group(sb, group);
5175 ext4_mb_unload_buddy(&e4b);
5177 ext4_debug("trimmed %d blocks in the group %d\n",
5184 * ext4_trim_fs() -- trim ioctl handle function
5185 * @sb: superblock for filesystem
5186 * @range: fstrim_range structure
5187 * @blkdev_flags: flags for the block device
5189 * start: First Byte to trim
5190 * len: number of Bytes to trim from start
5191 * minlen: minimum extent length in Bytes
5192 * ext4_trim_fs goes through all allocation groups containing Bytes from
5193 * start to start+len. For each such a group ext4_trim_all_free function
5194 * is invoked to trim all free space.
5196 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range,
5197 unsigned long blkdev_flags)
5199 struct ext4_group_info *grp;
5200 ext4_group_t group, first_group, last_group;
5201 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5202 uint64_t start, end, minlen, trimmed = 0;
5203 ext4_fsblk_t first_data_blk =
5204 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5205 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5208 start = range->start >> sb->s_blocksize_bits;
5209 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5210 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5211 range->minlen >> sb->s_blocksize_bits);
5213 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5214 start >= max_blks ||
5215 range->len < sb->s_blocksize)
5217 if (end >= max_blks)
5219 if (end <= first_data_blk)
5221 if (start < first_data_blk)
5222 start = first_data_blk;
5224 /* Determine first and last group to examine based on start and end */
5225 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5226 &first_group, &first_cluster);
5227 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5228 &last_group, &last_cluster);
5230 /* end now represents the last cluster to discard in this group */
5231 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5233 for (group = first_group; group <= last_group; group++) {
5234 grp = ext4_get_group_info(sb, group);
5235 /* We only do this if the grp has never been initialized */
5236 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5237 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5243 * For all the groups except the last one, last cluster will
5244 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5245 * change it for the last group, note that last_cluster is
5246 * already computed earlier by ext4_get_group_no_and_offset()
5248 if (group == last_group)
5251 if (grp->bb_free >= minlen) {
5252 cnt = ext4_trim_all_free(sb, group, first_cluster,
5253 end, minlen, blkdev_flags);
5262 * For every group except the first one, we are sure
5263 * that the first cluster to discard will be cluster #0.
5269 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5272 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;