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 License
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_SIZE /
123 * blocksize) blocks. So it can have information regarding groups_per_page
124 * which is blocks_per_page/2
126 * The buddy cache inode is not stored on disk. The inode is thrown
127 * away when the filesystem is unmounted.
129 * We look for count number of blocks in the buddy cache. If we were able
130 * to locate that many free blocks we return with additional information
131 * regarding rest of the contiguous physical block available
133 * Before allocating blocks via buddy cache we normalize the request
134 * blocks. This ensure we ask for more blocks that we needed. The extra
135 * blocks that we get after allocation is added to the respective prealloc
136 * list. In case of inode preallocation we follow a list of heuristics
137 * based on file size. This can be found in ext4_mb_normalize_request. If
138 * we are doing a group prealloc we try to normalize the request to
139 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
140 * dependent on the cluster size; for non-bigalloc file systems, it is
141 * 512 blocks. This can be tuned via
142 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
143 * terms of number of blocks. If we have mounted the file system with -O
144 * stripe=<value> option the group prealloc request is normalized to the
145 * the smallest multiple of the stripe value (sbi->s_stripe) which is
146 * greater than the default mb_group_prealloc.
148 * The regular allocator (using the buddy cache) supports a few tunables.
150 * /sys/fs/ext4/<partition>/mb_min_to_scan
151 * /sys/fs/ext4/<partition>/mb_max_to_scan
152 * /sys/fs/ext4/<partition>/mb_order2_req
154 * The regular allocator uses buddy scan only if the request len is power of
155 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
156 * value of s_mb_order2_reqs can be tuned via
157 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
158 * stripe size (sbi->s_stripe), we try to search for contiguous block in
159 * stripe size. This should result in better allocation on RAID setups. If
160 * not, we search in the specific group using bitmap for best extents. The
161 * tunable min_to_scan and max_to_scan control the behaviour here.
162 * min_to_scan indicate how long the mballoc __must__ look for a best
163 * extent and max_to_scan indicates how long the mballoc __can__ look for a
164 * best extent in the found extents. Searching for the blocks starts with
165 * the group specified as the goal value in allocation context via
166 * ac_g_ex. Each group is first checked based on the criteria whether it
167 * can be used for allocation. ext4_mb_good_group explains how the groups are
170 * Both the prealloc space are getting populated as above. So for the first
171 * request we will hit the buddy cache which will result in this prealloc
172 * space getting filled. The prealloc space is then later used for the
173 * subsequent request.
177 * mballoc operates on the following data:
179 * - in-core buddy (actually includes buddy and bitmap)
180 * - preallocation descriptors (PAs)
182 * there are two types of preallocations:
184 * assiged to specific inode and can be used for this inode only.
185 * it describes part of inode's space preallocated to specific
186 * physical blocks. any block from that preallocated can be used
187 * independent. the descriptor just tracks number of blocks left
188 * unused. so, before taking some block from descriptor, one must
189 * make sure corresponded logical block isn't allocated yet. this
190 * also means that freeing any block within descriptor's range
191 * must discard all preallocated blocks.
193 * assigned to specific locality group which does not translate to
194 * permanent set of inodes: inode can join and leave group. space
195 * from this type of preallocation can be used for any inode. thus
196 * it's consumed from the beginning to the end.
198 * relation between them can be expressed as:
199 * in-core buddy = on-disk bitmap + preallocation descriptors
201 * this mean blocks mballoc considers used are:
202 * - allocated blocks (persistent)
203 * - preallocated blocks (non-persistent)
205 * consistency in mballoc world means that at any time a block is either
206 * free or used in ALL structures. notice: "any time" should not be read
207 * literally -- time is discrete and delimited by locks.
209 * to keep it simple, we don't use block numbers, instead we count number of
210 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
212 * all operations can be expressed as:
213 * - init buddy: buddy = on-disk + PAs
214 * - new PA: buddy += N; PA = N
215 * - use inode PA: on-disk += N; PA -= N
216 * - discard inode PA buddy -= on-disk - PA; PA = 0
217 * - use locality group PA on-disk += N; PA -= N
218 * - discard locality group PA buddy -= PA; PA = 0
219 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
220 * is used in real operation because we can't know actual used
221 * bits from PA, only from on-disk bitmap
223 * if we follow this strict logic, then all operations above should be atomic.
224 * given some of them can block, we'd have to use something like semaphores
225 * killing performance on high-end SMP hardware. let's try to relax it using
226 * the following knowledge:
227 * 1) if buddy is referenced, it's already initialized
228 * 2) while block is used in buddy and the buddy is referenced,
229 * nobody can re-allocate that block
230 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
231 * bit set and PA claims same block, it's OK. IOW, one can set bit in
232 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
235 * so, now we're building a concurrency table:
238 * blocks for PA are allocated in the buddy, buddy must be referenced
239 * until PA is linked to allocation group to avoid concurrent buddy init
241 * we need to make sure that either on-disk bitmap or PA has uptodate data
242 * given (3) we care that PA-=N operation doesn't interfere with init
244 * the simplest way would be to have buddy initialized by the discard
245 * - use locality group PA
246 * again PA-=N must be serialized with init
247 * - discard locality group PA
248 * the simplest way would be to have buddy initialized by the discard
251 * i_data_sem serializes them
253 * discard process must wait until PA isn't used by another process
254 * - use locality group PA
255 * some mutex should serialize them
256 * - discard locality group PA
257 * discard process must wait until PA isn't used by another process
260 * i_data_sem or another mutex should serializes them
262 * discard process must wait until PA isn't used by another process
263 * - use locality group PA
264 * nothing wrong here -- they're different PAs covering different blocks
265 * - discard locality group PA
266 * discard process must wait until PA isn't used by another process
268 * now we're ready to make few consequences:
269 * - PA is referenced and while it is no discard is possible
270 * - PA is referenced until block isn't marked in on-disk bitmap
271 * - PA changes only after on-disk bitmap
272 * - discard must not compete with init. either init is done before
273 * any discard or they're serialized somehow
274 * - buddy init as sum of on-disk bitmap and PAs is done atomically
276 * a special case when we've used PA to emptiness. no need to modify buddy
277 * in this case, but we should care about concurrent init
282 * Logic in few words:
287 * mark bits in on-disk bitmap
290 * - use preallocation:
291 * find proper PA (per-inode or group)
293 * mark bits in on-disk bitmap
299 * mark bits in on-disk bitmap
302 * - discard preallocations in group:
304 * move them onto local list
305 * load on-disk bitmap
307 * remove PA from object (inode or locality group)
308 * mark free blocks in-core
310 * - discard inode's preallocations:
317 * - bitlock on a group (group)
318 * - object (inode/locality) (object)
329 * - release consumed pa:
334 * - generate in-core bitmap:
338 * - discard all for given object (inode, locality group):
343 * - discard all for given group:
350 static struct kmem_cache *ext4_pspace_cachep;
351 static struct kmem_cache *ext4_ac_cachep;
352 static struct kmem_cache *ext4_free_data_cachep;
354 /* We create slab caches for groupinfo data structures based on the
355 * superblock block size. There will be one per mounted filesystem for
356 * each unique s_blocksize_bits */
357 #define NR_GRPINFO_CACHES 8
358 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
360 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
361 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
362 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
363 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
366 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
368 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
370 static void ext4_free_data_callback(struct super_block *sb,
371 struct ext4_journal_cb_entry *jce, int rc);
373 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
375 #if BITS_PER_LONG == 64
376 *bit += ((unsigned long) addr & 7UL) << 3;
377 addr = (void *) ((unsigned long) addr & ~7UL);
378 #elif BITS_PER_LONG == 32
379 *bit += ((unsigned long) addr & 3UL) << 3;
380 addr = (void *) ((unsigned long) addr & ~3UL);
382 #error "how many bits you are?!"
387 static inline int mb_test_bit(int bit, void *addr)
390 * ext4_test_bit on architecture like powerpc
391 * needs unsigned long aligned address
393 addr = mb_correct_addr_and_bit(&bit, addr);
394 return ext4_test_bit(bit, addr);
397 static inline void mb_set_bit(int bit, void *addr)
399 addr = mb_correct_addr_and_bit(&bit, addr);
400 ext4_set_bit(bit, addr);
403 static inline void mb_clear_bit(int bit, void *addr)
405 addr = mb_correct_addr_and_bit(&bit, addr);
406 ext4_clear_bit(bit, addr);
409 static inline int mb_test_and_clear_bit(int bit, void *addr)
411 addr = mb_correct_addr_and_bit(&bit, addr);
412 return ext4_test_and_clear_bit(bit, addr);
415 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
417 int fix = 0, ret, tmpmax;
418 addr = mb_correct_addr_and_bit(&fix, addr);
422 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
428 static inline int mb_find_next_bit(void *addr, int max, int start)
430 int fix = 0, ret, tmpmax;
431 addr = mb_correct_addr_and_bit(&fix, addr);
435 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
441 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
445 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
448 if (order > e4b->bd_blkbits + 1) {
453 /* at order 0 we see each particular block */
455 *max = 1 << (e4b->bd_blkbits + 3);
456 return e4b->bd_bitmap;
459 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
460 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
466 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
467 int first, int count)
470 struct super_block *sb = e4b->bd_sb;
472 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
474 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
475 for (i = 0; i < count; i++) {
476 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
477 ext4_fsblk_t blocknr;
479 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
480 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
481 ext4_grp_locked_error(sb, e4b->bd_group,
482 inode ? inode->i_ino : 0,
484 "freeing block already freed "
488 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
492 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
496 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
498 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
499 for (i = 0; i < count; i++) {
500 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
501 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
505 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
507 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
508 unsigned char *b1, *b2;
510 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
511 b2 = (unsigned char *) bitmap;
512 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
513 if (b1[i] != b2[i]) {
514 ext4_msg(e4b->bd_sb, KERN_ERR,
515 "corruption in group %u "
516 "at byte %u(%u): %x in copy != %x "
518 e4b->bd_group, i, i * 8, b1[i], b2[i]);
526 static inline void mb_free_blocks_double(struct inode *inode,
527 struct ext4_buddy *e4b, int first, int count)
531 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
532 int first, int count)
536 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
542 #ifdef AGGRESSIVE_CHECK
544 #define MB_CHECK_ASSERT(assert) \
548 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
549 function, file, line, # assert); \
554 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
555 const char *function, int line)
557 struct super_block *sb = e4b->bd_sb;
558 int order = e4b->bd_blkbits + 1;
565 struct ext4_group_info *grp;
568 struct list_head *cur;
573 static int mb_check_counter;
574 if (mb_check_counter++ % 100 != 0)
579 buddy = mb_find_buddy(e4b, order, &max);
580 MB_CHECK_ASSERT(buddy);
581 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
582 MB_CHECK_ASSERT(buddy2);
583 MB_CHECK_ASSERT(buddy != buddy2);
584 MB_CHECK_ASSERT(max * 2 == max2);
587 for (i = 0; i < max; i++) {
589 if (mb_test_bit(i, buddy)) {
590 /* only single bit in buddy2 may be 1 */
591 if (!mb_test_bit(i << 1, buddy2)) {
593 mb_test_bit((i<<1)+1, buddy2));
594 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
596 mb_test_bit(i << 1, buddy2));
601 /* both bits in buddy2 must be 1 */
602 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
603 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
605 for (j = 0; j < (1 << order); j++) {
606 k = (i * (1 << order)) + j;
608 !mb_test_bit(k, e4b->bd_bitmap));
612 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
617 buddy = mb_find_buddy(e4b, 0, &max);
618 for (i = 0; i < max; i++) {
619 if (!mb_test_bit(i, buddy)) {
620 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
628 /* check used bits only */
629 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
630 buddy2 = mb_find_buddy(e4b, j, &max2);
632 MB_CHECK_ASSERT(k < max2);
633 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
636 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
637 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
639 grp = ext4_get_group_info(sb, e4b->bd_group);
640 list_for_each(cur, &grp->bb_prealloc_list) {
641 ext4_group_t groupnr;
642 struct ext4_prealloc_space *pa;
643 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
644 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
645 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
646 for (i = 0; i < pa->pa_len; i++)
647 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
651 #undef MB_CHECK_ASSERT
652 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
653 __FILE__, __func__, __LINE__)
655 #define mb_check_buddy(e4b)
659 * Divide blocks started from @first with length @len into
660 * smaller chunks with power of 2 blocks.
661 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
662 * then increase bb_counters[] for corresponded chunk size.
664 static void ext4_mb_mark_free_simple(struct super_block *sb,
665 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
666 struct ext4_group_info *grp)
668 struct ext4_sb_info *sbi = EXT4_SB(sb);
672 unsigned short border;
674 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
676 border = 2 << sb->s_blocksize_bits;
679 /* find how many blocks can be covered since this position */
680 max = ffs(first | border) - 1;
682 /* find how many blocks of power 2 we need to mark */
689 /* mark multiblock chunks only */
690 grp->bb_counters[min]++;
692 mb_clear_bit(first >> min,
693 buddy + sbi->s_mb_offsets[min]);
701 * Cache the order of the largest free extent we have available in this block
705 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
710 grp->bb_largest_free_order = -1; /* uninit */
712 bits = sb->s_blocksize_bits + 1;
713 for (i = bits; i >= 0; i--) {
714 if (grp->bb_counters[i] > 0) {
715 grp->bb_largest_free_order = i;
721 static noinline_for_stack
722 void ext4_mb_generate_buddy(struct super_block *sb,
723 void *buddy, void *bitmap, ext4_group_t group)
725 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
726 struct ext4_sb_info *sbi = EXT4_SB(sb);
727 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
732 unsigned fragments = 0;
733 unsigned long long period = get_cycles();
735 /* initialize buddy from bitmap which is aggregation
736 * of on-disk bitmap and preallocations */
737 i = mb_find_next_zero_bit(bitmap, max, 0);
738 grp->bb_first_free = i;
742 i = mb_find_next_bit(bitmap, max, i);
746 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
748 grp->bb_counters[0]++;
750 i = mb_find_next_zero_bit(bitmap, max, i);
752 grp->bb_fragments = fragments;
754 if (free != grp->bb_free) {
755 ext4_grp_locked_error(sb, group, 0, 0,
756 "block bitmap and bg descriptor "
757 "inconsistent: %u vs %u free clusters",
760 * If we intend to continue, we consider group descriptor
761 * corrupt and update bb_free using bitmap value
764 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
765 percpu_counter_sub(&sbi->s_freeclusters_counter,
767 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
769 mb_set_largest_free_order(sb, grp);
771 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
773 period = get_cycles() - period;
774 spin_lock(&EXT4_SB(sb)->s_bal_lock);
775 EXT4_SB(sb)->s_mb_buddies_generated++;
776 EXT4_SB(sb)->s_mb_generation_time += period;
777 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
780 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
786 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
787 ext4_set_bits(buddy, 0, count);
789 e4b->bd_info->bb_fragments = 0;
790 memset(e4b->bd_info->bb_counters, 0,
791 sizeof(*e4b->bd_info->bb_counters) *
792 (e4b->bd_sb->s_blocksize_bits + 2));
794 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
795 e4b->bd_bitmap, e4b->bd_group);
798 /* The buddy information is attached the buddy cache inode
799 * for convenience. The information regarding each group
800 * is loaded via ext4_mb_load_buddy. The information involve
801 * block bitmap and buddy information. The information are
802 * stored in the inode as
805 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
808 * one block each for bitmap and buddy information.
809 * So for each group we take up 2 blocks. A page can
810 * contain blocks_per_page (PAGE_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_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_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 put_page(e4b->bd_bitmap_page);
1033 if (e4b->bd_buddy_page) {
1034 unlock_page(e4b->bd_buddy_page);
1035 put_page(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_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.
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)) {
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);
1242 if (e4b->bd_bitmap_page)
1243 put_page(e4b->bd_bitmap_page);
1244 if (e4b->bd_buddy_page)
1245 put_page(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 put_page(e4b->bd_bitmap_page);
1261 if (e4b->bd_buddy_page)
1262 put_page(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[16];
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 ]\n");
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 .open = ext4_mb_seq_groups_open,
2355 .llseek = seq_lseek,
2356 .release = seq_release,
2359 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2361 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2362 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2369 * Allocate the top-level s_group_info array for the specified number
2372 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2374 struct ext4_sb_info *sbi = EXT4_SB(sb);
2376 struct ext4_group_info ***new_groupinfo;
2378 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2379 EXT4_DESC_PER_BLOCK_BITS(sb);
2380 if (size <= sbi->s_group_info_size)
2383 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2384 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2385 if (!new_groupinfo) {
2386 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2389 if (sbi->s_group_info) {
2390 memcpy(new_groupinfo, sbi->s_group_info,
2391 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2392 kvfree(sbi->s_group_info);
2394 sbi->s_group_info = new_groupinfo;
2395 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2396 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2397 sbi->s_group_info_size);
2401 /* Create and initialize ext4_group_info data for the given group. */
2402 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2403 struct ext4_group_desc *desc)
2407 struct ext4_sb_info *sbi = EXT4_SB(sb);
2408 struct ext4_group_info **meta_group_info;
2409 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2412 * First check if this group is the first of a reserved block.
2413 * If it's true, we have to allocate a new table of pointers
2414 * to ext4_group_info structures
2416 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2417 metalen = sizeof(*meta_group_info) <<
2418 EXT4_DESC_PER_BLOCK_BITS(sb);
2419 meta_group_info = kmalloc(metalen, GFP_NOFS);
2420 if (meta_group_info == NULL) {
2421 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2422 "for a buddy group");
2423 goto exit_meta_group_info;
2425 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2430 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2431 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2433 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2434 if (meta_group_info[i] == NULL) {
2435 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2436 goto exit_group_info;
2438 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2439 &(meta_group_info[i]->bb_state));
2442 * initialize bb_free to be able to skip
2443 * empty groups without initialization
2445 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2446 meta_group_info[i]->bb_free =
2447 ext4_free_clusters_after_init(sb, group, desc);
2449 meta_group_info[i]->bb_free =
2450 ext4_free_group_clusters(sb, desc);
2453 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2454 init_rwsem(&meta_group_info[i]->alloc_sem);
2455 meta_group_info[i]->bb_free_root = RB_ROOT;
2456 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2460 struct buffer_head *bh;
2461 meta_group_info[i]->bb_bitmap =
2462 kmalloc(sb->s_blocksize, GFP_NOFS);
2463 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2464 bh = ext4_read_block_bitmap(sb, group);
2465 BUG_ON(IS_ERR_OR_NULL(bh));
2466 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2475 /* If a meta_group_info table has been allocated, release it now */
2476 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2477 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2478 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2480 exit_meta_group_info:
2482 } /* ext4_mb_add_groupinfo */
2484 static int ext4_mb_init_backend(struct super_block *sb)
2486 ext4_group_t ngroups = ext4_get_groups_count(sb);
2488 struct ext4_sb_info *sbi = EXT4_SB(sb);
2490 struct ext4_group_desc *desc;
2491 struct kmem_cache *cachep;
2493 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2497 sbi->s_buddy_cache = new_inode(sb);
2498 if (sbi->s_buddy_cache == NULL) {
2499 ext4_msg(sb, KERN_ERR, "can't get new inode");
2502 /* To avoid potentially colliding with an valid on-disk inode number,
2503 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2504 * not in the inode hash, so it should never be found by iget(), but
2505 * this will avoid confusion if it ever shows up during debugging. */
2506 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2507 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2508 for (i = 0; i < ngroups; i++) {
2509 desc = ext4_get_group_desc(sb, i, NULL);
2511 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2514 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2521 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2523 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2524 i = sbi->s_group_info_size;
2526 kfree(sbi->s_group_info[i]);
2527 iput(sbi->s_buddy_cache);
2529 kvfree(sbi->s_group_info);
2533 static void ext4_groupinfo_destroy_slabs(void)
2537 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2538 if (ext4_groupinfo_caches[i])
2539 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2540 ext4_groupinfo_caches[i] = NULL;
2544 static int ext4_groupinfo_create_slab(size_t size)
2546 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2548 int blocksize_bits = order_base_2(size);
2549 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2550 struct kmem_cache *cachep;
2552 if (cache_index >= NR_GRPINFO_CACHES)
2555 if (unlikely(cache_index < 0))
2558 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2559 if (ext4_groupinfo_caches[cache_index]) {
2560 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2561 return 0; /* Already created */
2564 slab_size = offsetof(struct ext4_group_info,
2565 bb_counters[blocksize_bits + 2]);
2567 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2568 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2571 ext4_groupinfo_caches[cache_index] = cachep;
2573 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2576 "EXT4-fs: no memory for groupinfo slab cache\n");
2583 int ext4_mb_init(struct super_block *sb)
2585 struct ext4_sb_info *sbi = EXT4_SB(sb);
2587 unsigned offset, offset_incr;
2591 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2593 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2594 if (sbi->s_mb_offsets == NULL) {
2599 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2600 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2601 if (sbi->s_mb_maxs == NULL) {
2606 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2610 /* order 0 is regular bitmap */
2611 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2612 sbi->s_mb_offsets[0] = 0;
2616 offset_incr = 1 << (sb->s_blocksize_bits - 1);
2617 max = sb->s_blocksize << 2;
2619 sbi->s_mb_offsets[i] = offset;
2620 sbi->s_mb_maxs[i] = max;
2621 offset += offset_incr;
2622 offset_incr = offset_incr >> 1;
2625 } while (i <= sb->s_blocksize_bits + 1);
2627 spin_lock_init(&sbi->s_md_lock);
2628 spin_lock_init(&sbi->s_bal_lock);
2630 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2631 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2632 sbi->s_mb_stats = MB_DEFAULT_STATS;
2633 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2634 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2636 * The default group preallocation is 512, which for 4k block
2637 * sizes translates to 2 megabytes. However for bigalloc file
2638 * systems, this is probably too big (i.e, if the cluster size
2639 * is 1 megabyte, then group preallocation size becomes half a
2640 * gigabyte!). As a default, we will keep a two megabyte
2641 * group pralloc size for cluster sizes up to 64k, and after
2642 * that, we will force a minimum group preallocation size of
2643 * 32 clusters. This translates to 8 megs when the cluster
2644 * size is 256k, and 32 megs when the cluster size is 1 meg,
2645 * which seems reasonable as a default.
2647 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2648 sbi->s_cluster_bits, 32);
2650 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2651 * to the lowest multiple of s_stripe which is bigger than
2652 * the s_mb_group_prealloc as determined above. We want
2653 * the preallocation size to be an exact multiple of the
2654 * RAID stripe size so that preallocations don't fragment
2657 if (sbi->s_stripe > 1) {
2658 sbi->s_mb_group_prealloc = roundup(
2659 sbi->s_mb_group_prealloc, sbi->s_stripe);
2662 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2663 if (sbi->s_locality_groups == NULL) {
2667 for_each_possible_cpu(i) {
2668 struct ext4_locality_group *lg;
2669 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2670 mutex_init(&lg->lg_mutex);
2671 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2672 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2673 spin_lock_init(&lg->lg_prealloc_lock);
2676 /* init file for buddy data */
2677 ret = ext4_mb_init_backend(sb);
2679 goto out_free_locality_groups;
2683 out_free_locality_groups:
2684 free_percpu(sbi->s_locality_groups);
2685 sbi->s_locality_groups = NULL;
2687 kfree(sbi->s_mb_offsets);
2688 sbi->s_mb_offsets = NULL;
2689 kfree(sbi->s_mb_maxs);
2690 sbi->s_mb_maxs = NULL;
2694 /* need to called with the ext4 group lock held */
2695 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2697 struct ext4_prealloc_space *pa;
2698 struct list_head *cur, *tmp;
2701 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2702 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2703 list_del(&pa->pa_group_list);
2705 kmem_cache_free(ext4_pspace_cachep, pa);
2708 mb_debug(1, "mballoc: %u PAs left\n", count);
2712 int ext4_mb_release(struct super_block *sb)
2714 ext4_group_t ngroups = ext4_get_groups_count(sb);
2716 int num_meta_group_infos;
2717 struct ext4_group_info *grinfo;
2718 struct ext4_sb_info *sbi = EXT4_SB(sb);
2719 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2721 if (sbi->s_group_info) {
2722 for (i = 0; i < ngroups; i++) {
2723 grinfo = ext4_get_group_info(sb, i);
2725 kfree(grinfo->bb_bitmap);
2727 ext4_lock_group(sb, i);
2728 ext4_mb_cleanup_pa(grinfo);
2729 ext4_unlock_group(sb, i);
2730 kmem_cache_free(cachep, grinfo);
2732 num_meta_group_infos = (ngroups +
2733 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2734 EXT4_DESC_PER_BLOCK_BITS(sb);
2735 for (i = 0; i < num_meta_group_infos; i++)
2736 kfree(sbi->s_group_info[i]);
2737 kvfree(sbi->s_group_info);
2739 kfree(sbi->s_mb_offsets);
2740 kfree(sbi->s_mb_maxs);
2741 iput(sbi->s_buddy_cache);
2742 if (sbi->s_mb_stats) {
2743 ext4_msg(sb, KERN_INFO,
2744 "mballoc: %u blocks %u reqs (%u success)",
2745 atomic_read(&sbi->s_bal_allocated),
2746 atomic_read(&sbi->s_bal_reqs),
2747 atomic_read(&sbi->s_bal_success));
2748 ext4_msg(sb, KERN_INFO,
2749 "mballoc: %u extents scanned, %u goal hits, "
2750 "%u 2^N hits, %u breaks, %u lost",
2751 atomic_read(&sbi->s_bal_ex_scanned),
2752 atomic_read(&sbi->s_bal_goals),
2753 atomic_read(&sbi->s_bal_2orders),
2754 atomic_read(&sbi->s_bal_breaks),
2755 atomic_read(&sbi->s_mb_lost_chunks));
2756 ext4_msg(sb, KERN_INFO,
2757 "mballoc: %lu generated and it took %Lu",
2758 sbi->s_mb_buddies_generated,
2759 sbi->s_mb_generation_time);
2760 ext4_msg(sb, KERN_INFO,
2761 "mballoc: %u preallocated, %u discarded",
2762 atomic_read(&sbi->s_mb_preallocated),
2763 atomic_read(&sbi->s_mb_discarded));
2766 free_percpu(sbi->s_locality_groups);
2771 static inline int ext4_issue_discard(struct super_block *sb,
2772 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2774 ext4_fsblk_t discard_block;
2776 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2777 ext4_group_first_block_no(sb, block_group));
2778 count = EXT4_C2B(EXT4_SB(sb), count);
2779 trace_ext4_discard_blocks(sb,
2780 (unsigned long long) discard_block, count);
2781 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2785 * This function is called by the jbd2 layer once the commit has finished,
2786 * so we know we can free the blocks that were released with that commit.
2788 static void ext4_free_data_callback(struct super_block *sb,
2789 struct ext4_journal_cb_entry *jce,
2792 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2793 struct ext4_buddy e4b;
2794 struct ext4_group_info *db;
2795 int err, count = 0, count2 = 0;
2797 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2798 entry->efd_count, entry->efd_group, entry);
2800 if (test_opt(sb, DISCARD)) {
2801 err = ext4_issue_discard(sb, entry->efd_group,
2802 entry->efd_start_cluster,
2804 if (err && err != -EOPNOTSUPP)
2805 ext4_msg(sb, KERN_WARNING, "discard request in"
2806 " group:%d block:%d count:%d failed"
2807 " with %d", entry->efd_group,
2808 entry->efd_start_cluster,
2809 entry->efd_count, err);
2812 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2813 /* we expect to find existing buddy because it's pinned */
2818 /* there are blocks to put in buddy to make them really free */
2819 count += entry->efd_count;
2821 ext4_lock_group(sb, entry->efd_group);
2822 /* Take it out of per group rb tree */
2823 rb_erase(&entry->efd_node, &(db->bb_free_root));
2824 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2827 * Clear the trimmed flag for the group so that the next
2828 * ext4_trim_fs can trim it.
2829 * If the volume is mounted with -o discard, online discard
2830 * is supported and the free blocks will be trimmed online.
2832 if (!test_opt(sb, DISCARD))
2833 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2835 if (!db->bb_free_root.rb_node) {
2836 /* No more items in the per group rb tree
2837 * balance refcounts from ext4_mb_free_metadata()
2839 put_page(e4b.bd_buddy_page);
2840 put_page(e4b.bd_bitmap_page);
2842 ext4_unlock_group(sb, entry->efd_group);
2843 kmem_cache_free(ext4_free_data_cachep, entry);
2844 ext4_mb_unload_buddy(&e4b);
2846 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2849 int __init ext4_init_mballoc(void)
2851 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2852 SLAB_RECLAIM_ACCOUNT);
2853 if (ext4_pspace_cachep == NULL)
2856 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2857 SLAB_RECLAIM_ACCOUNT);
2858 if (ext4_ac_cachep == NULL) {
2859 kmem_cache_destroy(ext4_pspace_cachep);
2863 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2864 SLAB_RECLAIM_ACCOUNT);
2865 if (ext4_free_data_cachep == NULL) {
2866 kmem_cache_destroy(ext4_pspace_cachep);
2867 kmem_cache_destroy(ext4_ac_cachep);
2873 void ext4_exit_mballoc(void)
2876 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2877 * before destroying the slab cache.
2880 kmem_cache_destroy(ext4_pspace_cachep);
2881 kmem_cache_destroy(ext4_ac_cachep);
2882 kmem_cache_destroy(ext4_free_data_cachep);
2883 ext4_groupinfo_destroy_slabs();
2888 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2889 * Returns 0 if success or error code
2891 static noinline_for_stack int
2892 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2893 handle_t *handle, unsigned int reserv_clstrs)
2895 struct buffer_head *bitmap_bh = NULL;
2896 struct ext4_group_desc *gdp;
2897 struct buffer_head *gdp_bh;
2898 struct ext4_sb_info *sbi;
2899 struct super_block *sb;
2903 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2904 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2909 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2910 if (IS_ERR(bitmap_bh)) {
2911 err = PTR_ERR(bitmap_bh);
2916 BUFFER_TRACE(bitmap_bh, "getting write access");
2917 err = ext4_journal_get_write_access(handle, bitmap_bh);
2922 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2926 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2927 ext4_free_group_clusters(sb, gdp));
2929 BUFFER_TRACE(gdp_bh, "get_write_access");
2930 err = ext4_journal_get_write_access(handle, gdp_bh);
2934 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2936 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2937 if (!ext4_data_block_valid(sbi, block, len)) {
2938 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2939 "fs metadata", block, block+len);
2940 /* File system mounted not to panic on error
2941 * Fix the bitmap and repeat the block allocation
2942 * We leak some of the blocks here.
2944 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2945 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2946 ac->ac_b_ex.fe_len);
2947 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2948 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2954 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2955 #ifdef AGGRESSIVE_CHECK
2958 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2959 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2960 bitmap_bh->b_data));
2964 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2965 ac->ac_b_ex.fe_len);
2966 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2967 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2968 ext4_free_group_clusters_set(sb, gdp,
2969 ext4_free_clusters_after_init(sb,
2970 ac->ac_b_ex.fe_group, gdp));
2972 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2973 ext4_free_group_clusters_set(sb, gdp, len);
2974 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2975 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2977 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2978 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2980 * Now reduce the dirty block count also. Should not go negative
2982 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2983 /* release all the reserved blocks if non delalloc */
2984 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2987 if (sbi->s_log_groups_per_flex) {
2988 ext4_group_t flex_group = ext4_flex_group(sbi,
2989 ac->ac_b_ex.fe_group);
2990 atomic64_sub(ac->ac_b_ex.fe_len,
2991 &sbi->s_flex_groups[flex_group].free_clusters);
2994 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2997 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3005 * here we normalize request for locality group
3006 * Group request are normalized to s_mb_group_prealloc, which goes to
3007 * s_strip if we set the same via mount option.
3008 * s_mb_group_prealloc can be configured via
3009 * /sys/fs/ext4/<partition>/mb_group_prealloc
3011 * XXX: should we try to preallocate more than the group has now?
3013 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3015 struct super_block *sb = ac->ac_sb;
3016 struct ext4_locality_group *lg = ac->ac_lg;
3019 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3020 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3021 current->pid, ac->ac_g_ex.fe_len);
3025 * Normalization means making request better in terms of
3026 * size and alignment
3028 static noinline_for_stack void
3029 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3030 struct ext4_allocation_request *ar)
3032 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3035 loff_t size, start_off;
3036 loff_t orig_size __maybe_unused;
3038 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3039 struct ext4_prealloc_space *pa;
3041 /* do normalize only data requests, metadata requests
3042 do not need preallocation */
3043 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3046 /* sometime caller may want exact blocks */
3047 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3050 /* caller may indicate that preallocation isn't
3051 * required (it's a tail, for example) */
3052 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3055 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3056 ext4_mb_normalize_group_request(ac);
3060 bsbits = ac->ac_sb->s_blocksize_bits;
3062 /* first, let's learn actual file size
3063 * given current request is allocated */
3064 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3065 size = size << bsbits;
3066 if (size < i_size_read(ac->ac_inode))
3067 size = i_size_read(ac->ac_inode);
3070 /* max size of free chunks */
3073 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3074 (req <= (size) || max <= (chunk_size))
3076 /* first, try to predict filesize */
3077 /* XXX: should this table be tunable? */
3079 if (size <= 16 * 1024) {
3081 } else if (size <= 32 * 1024) {
3083 } else if (size <= 64 * 1024) {
3085 } else if (size <= 128 * 1024) {
3087 } else if (size <= 256 * 1024) {
3089 } else if (size <= 512 * 1024) {
3091 } else if (size <= 1024 * 1024) {
3093 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3094 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3095 (21 - bsbits)) << 21;
3096 size = 2 * 1024 * 1024;
3097 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3098 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3099 (22 - bsbits)) << 22;
3100 size = 4 * 1024 * 1024;
3101 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3102 (8<<20)>>bsbits, max, 8 * 1024)) {
3103 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3104 (23 - bsbits)) << 23;
3105 size = 8 * 1024 * 1024;
3107 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3108 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3109 ac->ac_o_ex.fe_len) << bsbits;
3111 size = size >> bsbits;
3112 start = start_off >> bsbits;
3114 /* don't cover already allocated blocks in selected range */
3115 if (ar->pleft && start <= ar->lleft) {
3116 size -= ar->lleft + 1 - start;
3117 start = ar->lleft + 1;
3119 if (ar->pright && start + size - 1 >= ar->lright)
3120 size -= start + size - ar->lright;
3124 /* check we don't cross already preallocated blocks */
3126 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3131 spin_lock(&pa->pa_lock);
3132 if (pa->pa_deleted) {
3133 spin_unlock(&pa->pa_lock);
3137 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3140 /* PA must not overlap original request */
3141 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3142 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3144 /* skip PAs this normalized request doesn't overlap with */
3145 if (pa->pa_lstart >= end || pa_end <= start) {
3146 spin_unlock(&pa->pa_lock);
3149 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3151 /* adjust start or end to be adjacent to this pa */
3152 if (pa_end <= ac->ac_o_ex.fe_logical) {
3153 BUG_ON(pa_end < start);
3155 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3156 BUG_ON(pa->pa_lstart > end);
3157 end = pa->pa_lstart;
3159 spin_unlock(&pa->pa_lock);
3164 /* XXX: extra loop to check we really don't overlap preallocations */
3166 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3169 spin_lock(&pa->pa_lock);
3170 if (pa->pa_deleted == 0) {
3171 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3173 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3175 spin_unlock(&pa->pa_lock);
3179 if (start + size <= ac->ac_o_ex.fe_logical &&
3180 start > ac->ac_o_ex.fe_logical) {
3181 ext4_msg(ac->ac_sb, KERN_ERR,
3182 "start %lu, size %lu, fe_logical %lu",
3183 (unsigned long) start, (unsigned long) size,
3184 (unsigned long) ac->ac_o_ex.fe_logical);
3187 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3189 /* now prepare goal request */
3191 /* XXX: is it better to align blocks WRT to logical
3192 * placement or satisfy big request as is */
3193 ac->ac_g_ex.fe_logical = start;
3194 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3196 /* define goal start in order to merge */
3197 if (ar->pright && (ar->lright == (start + size))) {
3198 /* merge to the right */
3199 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3200 &ac->ac_f_ex.fe_group,
3201 &ac->ac_f_ex.fe_start);
3202 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3204 if (ar->pleft && (ar->lleft + 1 == start)) {
3205 /* merge to the left */
3206 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3207 &ac->ac_f_ex.fe_group,
3208 &ac->ac_f_ex.fe_start);
3209 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3212 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3213 (unsigned) orig_size, (unsigned) start);
3216 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3218 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3220 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3221 atomic_inc(&sbi->s_bal_reqs);
3222 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3223 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3224 atomic_inc(&sbi->s_bal_success);
3225 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3226 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3227 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3228 atomic_inc(&sbi->s_bal_goals);
3229 if (ac->ac_found > sbi->s_mb_max_to_scan)
3230 atomic_inc(&sbi->s_bal_breaks);
3233 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3234 trace_ext4_mballoc_alloc(ac);
3236 trace_ext4_mballoc_prealloc(ac);
3240 * Called on failure; free up any blocks from the inode PA for this
3241 * context. We don't need this for MB_GROUP_PA because we only change
3242 * pa_free in ext4_mb_release_context(), but on failure, we've already
3243 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3245 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3247 struct ext4_prealloc_space *pa = ac->ac_pa;
3248 struct ext4_buddy e4b;
3252 if (ac->ac_f_ex.fe_len == 0)
3254 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3257 * This should never happen since we pin the
3258 * pages in the ext4_allocation_context so
3259 * ext4_mb_load_buddy() should never fail.
3261 WARN(1, "mb_load_buddy failed (%d)", err);
3264 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3265 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3266 ac->ac_f_ex.fe_len);
3267 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3268 ext4_mb_unload_buddy(&e4b);
3271 if (pa->pa_type == MB_INODE_PA)
3272 pa->pa_free += ac->ac_b_ex.fe_len;
3276 * use blocks preallocated to inode
3278 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3279 struct ext4_prealloc_space *pa)
3281 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3286 /* found preallocated blocks, use them */
3287 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3288 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3289 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3290 len = EXT4_NUM_B2C(sbi, end - start);
3291 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3292 &ac->ac_b_ex.fe_start);
3293 ac->ac_b_ex.fe_len = len;
3294 ac->ac_status = AC_STATUS_FOUND;
3297 BUG_ON(start < pa->pa_pstart);
3298 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3299 BUG_ON(pa->pa_free < len);
3302 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3306 * use blocks preallocated to locality group
3308 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3309 struct ext4_prealloc_space *pa)
3311 unsigned int len = ac->ac_o_ex.fe_len;
3313 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3314 &ac->ac_b_ex.fe_group,
3315 &ac->ac_b_ex.fe_start);
3316 ac->ac_b_ex.fe_len = len;
3317 ac->ac_status = AC_STATUS_FOUND;
3320 /* we don't correct pa_pstart or pa_plen here to avoid
3321 * possible race when the group is being loaded concurrently
3322 * instead we correct pa later, after blocks are marked
3323 * in on-disk bitmap -- see ext4_mb_release_context()
3324 * Other CPUs are prevented from allocating from this pa by lg_mutex
3326 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3330 * Return the prealloc space that have minimal distance
3331 * from the goal block. @cpa is the prealloc
3332 * space that is having currently known minimal distance
3333 * from the goal block.
3335 static struct ext4_prealloc_space *
3336 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3337 struct ext4_prealloc_space *pa,
3338 struct ext4_prealloc_space *cpa)
3340 ext4_fsblk_t cur_distance, new_distance;
3343 atomic_inc(&pa->pa_count);
3346 cur_distance = abs(goal_block - cpa->pa_pstart);
3347 new_distance = abs(goal_block - pa->pa_pstart);
3349 if (cur_distance <= new_distance)
3352 /* drop the previous reference */
3353 atomic_dec(&cpa->pa_count);
3354 atomic_inc(&pa->pa_count);
3359 * search goal blocks in preallocated space
3361 static noinline_for_stack int
3362 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3364 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3366 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3367 struct ext4_locality_group *lg;
3368 struct ext4_prealloc_space *pa, *cpa = NULL;
3369 ext4_fsblk_t goal_block;
3371 /* only data can be preallocated */
3372 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3375 /* first, try per-file preallocation */
3377 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3379 /* all fields in this condition don't change,
3380 * so we can skip locking for them */
3381 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3382 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3383 EXT4_C2B(sbi, pa->pa_len)))
3386 /* non-extent files can't have physical blocks past 2^32 */
3387 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3388 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3389 EXT4_MAX_BLOCK_FILE_PHYS))
3392 /* found preallocated blocks, use them */
3393 spin_lock(&pa->pa_lock);
3394 if (pa->pa_deleted == 0 && pa->pa_free) {
3395 atomic_inc(&pa->pa_count);
3396 ext4_mb_use_inode_pa(ac, pa);
3397 spin_unlock(&pa->pa_lock);
3398 ac->ac_criteria = 10;
3402 spin_unlock(&pa->pa_lock);
3406 /* can we use group allocation? */
3407 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3410 /* inode may have no locality group for some reason */
3414 order = fls(ac->ac_o_ex.fe_len) - 1;
3415 if (order > PREALLOC_TB_SIZE - 1)
3416 /* The max size of hash table is PREALLOC_TB_SIZE */
3417 order = PREALLOC_TB_SIZE - 1;
3419 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3421 * search for the prealloc space that is having
3422 * minimal distance from the goal block.
3424 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3426 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3428 spin_lock(&pa->pa_lock);
3429 if (pa->pa_deleted == 0 &&
3430 pa->pa_free >= ac->ac_o_ex.fe_len) {
3432 cpa = ext4_mb_check_group_pa(goal_block,
3435 spin_unlock(&pa->pa_lock);
3440 ext4_mb_use_group_pa(ac, cpa);
3441 ac->ac_criteria = 20;
3448 * the function goes through all block freed in the group
3449 * but not yet committed and marks them used in in-core bitmap.
3450 * buddy must be generated from this bitmap
3451 * Need to be called with the ext4 group lock held
3453 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3457 struct ext4_group_info *grp;
3458 struct ext4_free_data *entry;
3460 grp = ext4_get_group_info(sb, group);
3461 n = rb_first(&(grp->bb_free_root));
3464 entry = rb_entry(n, struct ext4_free_data, efd_node);
3465 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3472 * the function goes through all preallocation in this group and marks them
3473 * used in in-core bitmap. buddy must be generated from this bitmap
3474 * Need to be called with ext4 group lock held
3476 static noinline_for_stack
3477 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3480 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3481 struct ext4_prealloc_space *pa;
3482 struct list_head *cur;
3483 ext4_group_t groupnr;
3484 ext4_grpblk_t start;
3485 int preallocated = 0;
3488 /* all form of preallocation discards first load group,
3489 * so the only competing code is preallocation use.
3490 * we don't need any locking here
3491 * notice we do NOT ignore preallocations with pa_deleted
3492 * otherwise we could leave used blocks available for
3493 * allocation in buddy when concurrent ext4_mb_put_pa()
3494 * is dropping preallocation
3496 list_for_each(cur, &grp->bb_prealloc_list) {
3497 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3498 spin_lock(&pa->pa_lock);
3499 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3502 spin_unlock(&pa->pa_lock);
3503 if (unlikely(len == 0))
3505 BUG_ON(groupnr != group);
3506 ext4_set_bits(bitmap, start, len);
3507 preallocated += len;
3509 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3512 static void ext4_mb_pa_callback(struct rcu_head *head)
3514 struct ext4_prealloc_space *pa;
3515 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3517 BUG_ON(atomic_read(&pa->pa_count));
3518 BUG_ON(pa->pa_deleted == 0);
3519 kmem_cache_free(ext4_pspace_cachep, pa);
3523 * drops a reference to preallocated space descriptor
3524 * if this was the last reference and the space is consumed
3526 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3527 struct super_block *sb, struct ext4_prealloc_space *pa)
3530 ext4_fsblk_t grp_blk;
3532 /* in this short window concurrent discard can set pa_deleted */
3533 spin_lock(&pa->pa_lock);
3534 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3535 spin_unlock(&pa->pa_lock);
3539 if (pa->pa_deleted == 1) {
3540 spin_unlock(&pa->pa_lock);
3545 spin_unlock(&pa->pa_lock);
3547 grp_blk = pa->pa_pstart;
3549 * If doing group-based preallocation, pa_pstart may be in the
3550 * next group when pa is used up
3552 if (pa->pa_type == MB_GROUP_PA)
3555 grp = ext4_get_group_number(sb, grp_blk);
3560 * P1 (buddy init) P2 (regular allocation)
3561 * find block B in PA
3562 * copy on-disk bitmap to buddy
3563 * mark B in on-disk bitmap
3564 * drop PA from group
3565 * mark all PAs in buddy
3567 * thus, P1 initializes buddy with B available. to prevent this
3568 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3571 ext4_lock_group(sb, grp);
3572 list_del(&pa->pa_group_list);
3573 ext4_unlock_group(sb, grp);
3575 spin_lock(pa->pa_obj_lock);
3576 list_del_rcu(&pa->pa_inode_list);
3577 spin_unlock(pa->pa_obj_lock);
3579 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3583 * creates new preallocated space for given inode
3585 static noinline_for_stack int
3586 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3588 struct super_block *sb = ac->ac_sb;
3589 struct ext4_sb_info *sbi = EXT4_SB(sb);
3590 struct ext4_prealloc_space *pa;
3591 struct ext4_group_info *grp;
3592 struct ext4_inode_info *ei;
3594 /* preallocate only when found space is larger then requested */
3595 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3596 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3597 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3599 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3603 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3609 /* we can't allocate as much as normalizer wants.
3610 * so, found space must get proper lstart
3611 * to cover original request */
3612 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3613 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3615 /* we're limited by original request in that
3616 * logical block must be covered any way
3617 * winl is window we can move our chunk within */
3618 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3620 /* also, we should cover whole original request */
3621 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3623 /* the smallest one defines real window */
3624 win = min(winl, wins);
3626 offs = ac->ac_o_ex.fe_logical %
3627 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3628 if (offs && offs < win)
3631 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3632 EXT4_NUM_B2C(sbi, win);
3633 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3634 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3637 /* preallocation can change ac_b_ex, thus we store actually
3638 * allocated blocks for history */
3639 ac->ac_f_ex = ac->ac_b_ex;
3641 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3642 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3643 pa->pa_len = ac->ac_b_ex.fe_len;
3644 pa->pa_free = pa->pa_len;
3645 atomic_set(&pa->pa_count, 1);
3646 spin_lock_init(&pa->pa_lock);
3647 INIT_LIST_HEAD(&pa->pa_inode_list);
3648 INIT_LIST_HEAD(&pa->pa_group_list);
3650 pa->pa_type = MB_INODE_PA;
3652 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3653 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3654 trace_ext4_mb_new_inode_pa(ac, pa);
3656 ext4_mb_use_inode_pa(ac, pa);
3657 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3659 ei = EXT4_I(ac->ac_inode);
3660 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3662 pa->pa_obj_lock = &ei->i_prealloc_lock;
3663 pa->pa_inode = ac->ac_inode;
3665 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3666 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3667 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3669 spin_lock(pa->pa_obj_lock);
3670 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3671 spin_unlock(pa->pa_obj_lock);
3677 * creates new preallocated space for locality group inodes belongs to
3679 static noinline_for_stack int
3680 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3682 struct super_block *sb = ac->ac_sb;
3683 struct ext4_locality_group *lg;
3684 struct ext4_prealloc_space *pa;
3685 struct ext4_group_info *grp;
3687 /* preallocate only when found space is larger then requested */
3688 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3689 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3690 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3692 BUG_ON(ext4_pspace_cachep == NULL);
3693 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3697 /* preallocation can change ac_b_ex, thus we store actually
3698 * allocated blocks for history */
3699 ac->ac_f_ex = ac->ac_b_ex;
3701 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3702 pa->pa_lstart = pa->pa_pstart;
3703 pa->pa_len = ac->ac_b_ex.fe_len;
3704 pa->pa_free = pa->pa_len;
3705 atomic_set(&pa->pa_count, 1);
3706 spin_lock_init(&pa->pa_lock);
3707 INIT_LIST_HEAD(&pa->pa_inode_list);
3708 INIT_LIST_HEAD(&pa->pa_group_list);
3710 pa->pa_type = MB_GROUP_PA;
3712 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3713 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3714 trace_ext4_mb_new_group_pa(ac, pa);
3716 ext4_mb_use_group_pa(ac, pa);
3717 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3719 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3723 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3724 pa->pa_inode = NULL;
3726 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3727 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3728 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3731 * We will later add the new pa to the right bucket
3732 * after updating the pa_free in ext4_mb_release_context
3737 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3741 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3742 err = ext4_mb_new_group_pa(ac);
3744 err = ext4_mb_new_inode_pa(ac);
3749 * finds all unused blocks in on-disk bitmap, frees them in
3750 * in-core bitmap and buddy.
3751 * @pa must be unlinked from inode and group lists, so that
3752 * nobody else can find/use it.
3753 * the caller MUST hold group/inode locks.
3754 * TODO: optimize the case when there are no in-core structures yet
3756 static noinline_for_stack int
3757 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3758 struct ext4_prealloc_space *pa)
3760 struct super_block *sb = e4b->bd_sb;
3761 struct ext4_sb_info *sbi = EXT4_SB(sb);
3766 unsigned long long grp_blk_start;
3770 BUG_ON(pa->pa_deleted == 0);
3771 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3772 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3773 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3774 end = bit + pa->pa_len;
3777 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3780 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3781 mb_debug(1, " free preallocated %u/%u in group %u\n",
3782 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3783 (unsigned) next - bit, (unsigned) group);
3786 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3787 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3788 EXT4_C2B(sbi, bit)),
3790 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3793 if (free != pa->pa_free) {
3794 ext4_msg(e4b->bd_sb, KERN_CRIT,
3795 "pa %p: logic %lu, phys. %lu, len %lu",
3796 pa, (unsigned long) pa->pa_lstart,
3797 (unsigned long) pa->pa_pstart,
3798 (unsigned long) pa->pa_len);
3799 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3802 * pa is already deleted so we use the value obtained
3803 * from the bitmap and continue.
3806 atomic_add(free, &sbi->s_mb_discarded);
3811 static noinline_for_stack int
3812 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3813 struct ext4_prealloc_space *pa)
3815 struct super_block *sb = e4b->bd_sb;
3819 trace_ext4_mb_release_group_pa(sb, pa);
3820 BUG_ON(pa->pa_deleted == 0);
3821 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3822 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3823 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3824 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3825 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3831 * releases all preallocations in given group
3833 * first, we need to decide discard policy:
3834 * - when do we discard
3836 * - how many do we discard
3837 * 1) how many requested
3839 static noinline_for_stack int
3840 ext4_mb_discard_group_preallocations(struct super_block *sb,
3841 ext4_group_t group, int needed)
3843 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3844 struct buffer_head *bitmap_bh = NULL;
3845 struct ext4_prealloc_space *pa, *tmp;
3846 struct list_head list;
3847 struct ext4_buddy e4b;
3852 mb_debug(1, "discard preallocation for group %u\n", group);
3854 if (list_empty(&grp->bb_prealloc_list))
3857 bitmap_bh = ext4_read_block_bitmap(sb, group);
3858 if (IS_ERR(bitmap_bh)) {
3859 err = PTR_ERR(bitmap_bh);
3860 ext4_error(sb, "Error %d reading block bitmap for %u",
3865 err = ext4_mb_load_buddy(sb, group, &e4b);
3867 ext4_error(sb, "Error loading buddy information for %u", group);
3873 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3875 INIT_LIST_HEAD(&list);
3877 ext4_lock_group(sb, group);
3878 list_for_each_entry_safe(pa, tmp,
3879 &grp->bb_prealloc_list, pa_group_list) {
3880 spin_lock(&pa->pa_lock);
3881 if (atomic_read(&pa->pa_count)) {
3882 spin_unlock(&pa->pa_lock);
3886 if (pa->pa_deleted) {
3887 spin_unlock(&pa->pa_lock);
3891 /* seems this one can be freed ... */
3894 /* we can trust pa_free ... */
3895 free += pa->pa_free;
3897 spin_unlock(&pa->pa_lock);
3899 list_del(&pa->pa_group_list);
3900 list_add(&pa->u.pa_tmp_list, &list);
3903 /* if we still need more blocks and some PAs were used, try again */
3904 if (free < needed && busy) {
3906 ext4_unlock_group(sb, group);
3911 /* found anything to free? */
3912 if (list_empty(&list)) {
3917 /* now free all selected PAs */
3918 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3920 /* remove from object (inode or locality group) */
3921 spin_lock(pa->pa_obj_lock);
3922 list_del_rcu(&pa->pa_inode_list);
3923 spin_unlock(pa->pa_obj_lock);
3925 if (pa->pa_type == MB_GROUP_PA)
3926 ext4_mb_release_group_pa(&e4b, pa);
3928 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3930 list_del(&pa->u.pa_tmp_list);
3931 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3935 ext4_unlock_group(sb, group);
3936 ext4_mb_unload_buddy(&e4b);
3942 * releases all non-used preallocated blocks for given inode
3944 * It's important to discard preallocations under i_data_sem
3945 * We don't want another block to be served from the prealloc
3946 * space when we are discarding the inode prealloc space.
3948 * FIXME!! Make sure it is valid at all the call sites
3950 void ext4_discard_preallocations(struct inode *inode)
3952 struct ext4_inode_info *ei = EXT4_I(inode);
3953 struct super_block *sb = inode->i_sb;
3954 struct buffer_head *bitmap_bh = NULL;
3955 struct ext4_prealloc_space *pa, *tmp;
3956 ext4_group_t group = 0;
3957 struct list_head list;
3958 struct ext4_buddy e4b;
3961 if (!S_ISREG(inode->i_mode)) {
3962 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3966 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3967 trace_ext4_discard_preallocations(inode);
3969 INIT_LIST_HEAD(&list);
3972 /* first, collect all pa's in the inode */
3973 spin_lock(&ei->i_prealloc_lock);
3974 while (!list_empty(&ei->i_prealloc_list)) {
3975 pa = list_entry(ei->i_prealloc_list.next,
3976 struct ext4_prealloc_space, pa_inode_list);
3977 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3978 spin_lock(&pa->pa_lock);
3979 if (atomic_read(&pa->pa_count)) {
3980 /* this shouldn't happen often - nobody should
3981 * use preallocation while we're discarding it */
3982 spin_unlock(&pa->pa_lock);
3983 spin_unlock(&ei->i_prealloc_lock);
3984 ext4_msg(sb, KERN_ERR,
3985 "uh-oh! used pa while discarding");
3987 schedule_timeout_uninterruptible(HZ);
3991 if (pa->pa_deleted == 0) {
3993 spin_unlock(&pa->pa_lock);
3994 list_del_rcu(&pa->pa_inode_list);
3995 list_add(&pa->u.pa_tmp_list, &list);
3999 /* someone is deleting pa right now */
4000 spin_unlock(&pa->pa_lock);
4001 spin_unlock(&ei->i_prealloc_lock);
4003 /* we have to wait here because pa_deleted
4004 * doesn't mean pa is already unlinked from
4005 * the list. as we might be called from
4006 * ->clear_inode() the inode will get freed
4007 * and concurrent thread which is unlinking
4008 * pa from inode's list may access already
4009 * freed memory, bad-bad-bad */
4011 /* XXX: if this happens too often, we can
4012 * add a flag to force wait only in case
4013 * of ->clear_inode(), but not in case of
4014 * regular truncate */
4015 schedule_timeout_uninterruptible(HZ);
4018 spin_unlock(&ei->i_prealloc_lock);
4020 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4021 BUG_ON(pa->pa_type != MB_INODE_PA);
4022 group = ext4_get_group_number(sb, pa->pa_pstart);
4024 err = ext4_mb_load_buddy(sb, group, &e4b);
4026 ext4_error(sb, "Error loading buddy information for %u",
4031 bitmap_bh = ext4_read_block_bitmap(sb, group);
4032 if (IS_ERR(bitmap_bh)) {
4033 err = PTR_ERR(bitmap_bh);
4034 ext4_error(sb, "Error %d reading block bitmap for %u",
4036 ext4_mb_unload_buddy(&e4b);
4040 ext4_lock_group(sb, group);
4041 list_del(&pa->pa_group_list);
4042 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4043 ext4_unlock_group(sb, group);
4045 ext4_mb_unload_buddy(&e4b);
4048 list_del(&pa->u.pa_tmp_list);
4049 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4053 #ifdef CONFIG_EXT4_DEBUG
4054 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4056 struct super_block *sb = ac->ac_sb;
4057 ext4_group_t ngroups, i;
4059 if (!ext4_mballoc_debug ||
4060 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4063 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4064 " Allocation context details:");
4065 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4066 ac->ac_status, ac->ac_flags);
4067 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4068 "goal %lu/%lu/%lu@%lu, "
4069 "best %lu/%lu/%lu@%lu cr %d",
4070 (unsigned long)ac->ac_o_ex.fe_group,
4071 (unsigned long)ac->ac_o_ex.fe_start,
4072 (unsigned long)ac->ac_o_ex.fe_len,
4073 (unsigned long)ac->ac_o_ex.fe_logical,
4074 (unsigned long)ac->ac_g_ex.fe_group,
4075 (unsigned long)ac->ac_g_ex.fe_start,
4076 (unsigned long)ac->ac_g_ex.fe_len,
4077 (unsigned long)ac->ac_g_ex.fe_logical,
4078 (unsigned long)ac->ac_b_ex.fe_group,
4079 (unsigned long)ac->ac_b_ex.fe_start,
4080 (unsigned long)ac->ac_b_ex.fe_len,
4081 (unsigned long)ac->ac_b_ex.fe_logical,
4082 (int)ac->ac_criteria);
4083 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4084 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4085 ngroups = ext4_get_groups_count(sb);
4086 for (i = 0; i < ngroups; i++) {
4087 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4088 struct ext4_prealloc_space *pa;
4089 ext4_grpblk_t start;
4090 struct list_head *cur;
4091 ext4_lock_group(sb, i);
4092 list_for_each(cur, &grp->bb_prealloc_list) {
4093 pa = list_entry(cur, struct ext4_prealloc_space,
4095 spin_lock(&pa->pa_lock);
4096 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4098 spin_unlock(&pa->pa_lock);
4099 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4102 ext4_unlock_group(sb, i);
4104 if (grp->bb_free == 0)
4106 printk(KERN_ERR "%u: %d/%d \n",
4107 i, grp->bb_free, grp->bb_fragments);
4109 printk(KERN_ERR "\n");
4112 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4119 * We use locality group preallocation for small size file. The size of the
4120 * file is determined by the current size or the resulting size after
4121 * allocation which ever is larger
4123 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4125 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4127 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4128 int bsbits = ac->ac_sb->s_blocksize_bits;
4131 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4134 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4137 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4138 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4141 if ((size == isize) &&
4142 !ext4_fs_is_busy(sbi) &&
4143 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4144 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4148 if (sbi->s_mb_group_prealloc <= 0) {
4149 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4153 /* don't use group allocation for large files */
4154 size = max(size, isize);
4155 if (size > sbi->s_mb_stream_request) {
4156 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4160 BUG_ON(ac->ac_lg != NULL);
4162 * locality group prealloc space are per cpu. The reason for having
4163 * per cpu locality group is to reduce the contention between block
4164 * request from multiple CPUs.
4166 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4168 /* we're going to use group allocation */
4169 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4171 /* serialize all allocations in the group */
4172 mutex_lock(&ac->ac_lg->lg_mutex);
4175 static noinline_for_stack int
4176 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4177 struct ext4_allocation_request *ar)
4179 struct super_block *sb = ar->inode->i_sb;
4180 struct ext4_sb_info *sbi = EXT4_SB(sb);
4181 struct ext4_super_block *es = sbi->s_es;
4185 ext4_grpblk_t block;
4187 /* we can't allocate > group size */
4190 /* just a dirty hack to filter too big requests */
4191 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4192 len = EXT4_CLUSTERS_PER_GROUP(sb);
4194 /* start searching from the goal */
4196 if (goal < le32_to_cpu(es->s_first_data_block) ||
4197 goal >= ext4_blocks_count(es))
4198 goal = le32_to_cpu(es->s_first_data_block);
4199 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4201 /* set up allocation goals */
4202 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4203 ac->ac_status = AC_STATUS_CONTINUE;
4205 ac->ac_inode = ar->inode;
4206 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4207 ac->ac_o_ex.fe_group = group;
4208 ac->ac_o_ex.fe_start = block;
4209 ac->ac_o_ex.fe_len = len;
4210 ac->ac_g_ex = ac->ac_o_ex;
4211 ac->ac_flags = ar->flags;
4213 /* we have to define context: we'll we work with a file or
4214 * locality group. this is a policy, actually */
4215 ext4_mb_group_or_file(ac);
4217 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4218 "left: %u/%u, right %u/%u to %swritable\n",
4219 (unsigned) ar->len, (unsigned) ar->logical,
4220 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4221 (unsigned) ar->lleft, (unsigned) ar->pleft,
4222 (unsigned) ar->lright, (unsigned) ar->pright,
4223 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4228 static noinline_for_stack void
4229 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4230 struct ext4_locality_group *lg,
4231 int order, int total_entries)
4233 ext4_group_t group = 0;
4234 struct ext4_buddy e4b;
4235 struct list_head discard_list;
4236 struct ext4_prealloc_space *pa, *tmp;
4238 mb_debug(1, "discard locality group preallocation\n");
4240 INIT_LIST_HEAD(&discard_list);
4242 spin_lock(&lg->lg_prealloc_lock);
4243 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4245 spin_lock(&pa->pa_lock);
4246 if (atomic_read(&pa->pa_count)) {
4248 * This is the pa that we just used
4249 * for block allocation. So don't
4252 spin_unlock(&pa->pa_lock);
4255 if (pa->pa_deleted) {
4256 spin_unlock(&pa->pa_lock);
4259 /* only lg prealloc space */
4260 BUG_ON(pa->pa_type != MB_GROUP_PA);
4262 /* seems this one can be freed ... */
4264 spin_unlock(&pa->pa_lock);
4266 list_del_rcu(&pa->pa_inode_list);
4267 list_add(&pa->u.pa_tmp_list, &discard_list);
4270 if (total_entries <= 5) {
4272 * we want to keep only 5 entries
4273 * allowing it to grow to 8. This
4274 * mak sure we don't call discard
4275 * soon for this list.
4280 spin_unlock(&lg->lg_prealloc_lock);
4282 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4284 group = ext4_get_group_number(sb, pa->pa_pstart);
4285 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4286 ext4_error(sb, "Error loading buddy information for %u",
4290 ext4_lock_group(sb, group);
4291 list_del(&pa->pa_group_list);
4292 ext4_mb_release_group_pa(&e4b, pa);
4293 ext4_unlock_group(sb, group);
4295 ext4_mb_unload_buddy(&e4b);
4296 list_del(&pa->u.pa_tmp_list);
4297 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4302 * We have incremented pa_count. So it cannot be freed at this
4303 * point. Also we hold lg_mutex. So no parallel allocation is
4304 * possible from this lg. That means pa_free cannot be updated.
4306 * A parallel ext4_mb_discard_group_preallocations is possible.
4307 * which can cause the lg_prealloc_list to be updated.
4310 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4312 int order, added = 0, lg_prealloc_count = 1;
4313 struct super_block *sb = ac->ac_sb;
4314 struct ext4_locality_group *lg = ac->ac_lg;
4315 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4317 order = fls(pa->pa_free) - 1;
4318 if (order > PREALLOC_TB_SIZE - 1)
4319 /* The max size of hash table is PREALLOC_TB_SIZE */
4320 order = PREALLOC_TB_SIZE - 1;
4321 /* Add the prealloc space to lg */
4322 spin_lock(&lg->lg_prealloc_lock);
4323 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4325 spin_lock(&tmp_pa->pa_lock);
4326 if (tmp_pa->pa_deleted) {
4327 spin_unlock(&tmp_pa->pa_lock);
4330 if (!added && pa->pa_free < tmp_pa->pa_free) {
4331 /* Add to the tail of the previous entry */
4332 list_add_tail_rcu(&pa->pa_inode_list,
4333 &tmp_pa->pa_inode_list);
4336 * we want to count the total
4337 * number of entries in the list
4340 spin_unlock(&tmp_pa->pa_lock);
4341 lg_prealloc_count++;
4344 list_add_tail_rcu(&pa->pa_inode_list,
4345 &lg->lg_prealloc_list[order]);
4346 spin_unlock(&lg->lg_prealloc_lock);
4348 /* Now trim the list to be not more than 8 elements */
4349 if (lg_prealloc_count > 8) {
4350 ext4_mb_discard_lg_preallocations(sb, lg,
4351 order, lg_prealloc_count);
4358 * release all resource we used in allocation
4360 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4362 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4363 struct ext4_prealloc_space *pa = ac->ac_pa;
4365 if (pa->pa_type == MB_GROUP_PA) {
4366 /* see comment in ext4_mb_use_group_pa() */
4367 spin_lock(&pa->pa_lock);
4368 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4369 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4370 pa->pa_free -= ac->ac_b_ex.fe_len;
4371 pa->pa_len -= ac->ac_b_ex.fe_len;
4372 spin_unlock(&pa->pa_lock);
4377 * We want to add the pa to the right bucket.
4378 * Remove it from the list and while adding
4379 * make sure the list to which we are adding
4382 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4383 spin_lock(pa->pa_obj_lock);
4384 list_del_rcu(&pa->pa_inode_list);
4385 spin_unlock(pa->pa_obj_lock);
4386 ext4_mb_add_n_trim(ac);
4388 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4390 if (ac->ac_bitmap_page)
4391 put_page(ac->ac_bitmap_page);
4392 if (ac->ac_buddy_page)
4393 put_page(ac->ac_buddy_page);
4394 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4395 mutex_unlock(&ac->ac_lg->lg_mutex);
4396 ext4_mb_collect_stats(ac);
4400 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4402 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4406 trace_ext4_mb_discard_preallocations(sb, needed);
4407 for (i = 0; i < ngroups && needed > 0; i++) {
4408 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4417 * Main entry point into mballoc to allocate blocks
4418 * it tries to use preallocation first, then falls back
4419 * to usual allocation
4421 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4422 struct ext4_allocation_request *ar, int *errp)
4425 struct ext4_allocation_context *ac = NULL;
4426 struct ext4_sb_info *sbi;
4427 struct super_block *sb;
4428 ext4_fsblk_t block = 0;
4429 unsigned int inquota = 0;
4430 unsigned int reserv_clstrs = 0;
4433 sb = ar->inode->i_sb;
4436 trace_ext4_request_blocks(ar);
4438 /* Allow to use superuser reservation for quota file */
4439 if (IS_NOQUOTA(ar->inode))
4440 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4442 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4443 /* Without delayed allocation we need to verify
4444 * there is enough free blocks to do block allocation
4445 * and verify allocation doesn't exceed the quota limits.
4448 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4450 /* let others to free the space */
4452 ar->len = ar->len >> 1;
4458 reserv_clstrs = ar->len;
4459 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4460 dquot_alloc_block_nofail(ar->inode,
4461 EXT4_C2B(sbi, ar->len));
4464 dquot_alloc_block(ar->inode,
4465 EXT4_C2B(sbi, ar->len))) {
4467 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4478 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4485 *errp = ext4_mb_initialize_context(ac, ar);
4491 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4492 if (!ext4_mb_use_preallocated(ac)) {
4493 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4494 ext4_mb_normalize_request(ac, ar);
4496 /* allocate space in core */
4497 *errp = ext4_mb_regular_allocator(ac);
4499 goto discard_and_exit;
4501 /* as we've just preallocated more space than
4502 * user requested originally, we store allocated
4503 * space in a special descriptor */
4504 if (ac->ac_status == AC_STATUS_FOUND &&
4505 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4506 *errp = ext4_mb_new_preallocation(ac);
4509 ext4_discard_allocated_blocks(ac);
4513 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4514 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4515 if (*errp == -EAGAIN) {
4517 * drop the reference that we took
4518 * in ext4_mb_use_best_found
4520 ext4_mb_release_context(ac);
4521 ac->ac_b_ex.fe_group = 0;
4522 ac->ac_b_ex.fe_start = 0;
4523 ac->ac_b_ex.fe_len = 0;
4524 ac->ac_status = AC_STATUS_CONTINUE;
4527 ext4_discard_allocated_blocks(ac);
4530 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4531 ar->len = ac->ac_b_ex.fe_len;
4534 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4542 ac->ac_b_ex.fe_len = 0;
4544 ext4_mb_show_ac(ac);
4546 ext4_mb_release_context(ac);
4549 kmem_cache_free(ext4_ac_cachep, ac);
4550 if (inquota && ar->len < inquota)
4551 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4553 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4554 /* release all the reserved blocks if non delalloc */
4555 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4559 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4565 * We can merge two free data extents only if the physical blocks
4566 * are contiguous, AND the extents were freed by the same transaction,
4567 * AND the blocks are associated with the same group.
4569 static int can_merge(struct ext4_free_data *entry1,
4570 struct ext4_free_data *entry2)
4572 if ((entry1->efd_tid == entry2->efd_tid) &&
4573 (entry1->efd_group == entry2->efd_group) &&
4574 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4579 static noinline_for_stack int
4580 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4581 struct ext4_free_data *new_entry)
4583 ext4_group_t group = e4b->bd_group;
4584 ext4_grpblk_t cluster;
4585 struct ext4_free_data *entry;
4586 struct ext4_group_info *db = e4b->bd_info;
4587 struct super_block *sb = e4b->bd_sb;
4588 struct ext4_sb_info *sbi = EXT4_SB(sb);
4589 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4590 struct rb_node *parent = NULL, *new_node;
4592 BUG_ON(!ext4_handle_valid(handle));
4593 BUG_ON(e4b->bd_bitmap_page == NULL);
4594 BUG_ON(e4b->bd_buddy_page == NULL);
4596 new_node = &new_entry->efd_node;
4597 cluster = new_entry->efd_start_cluster;
4600 /* first free block exent. We need to
4601 protect buddy cache from being freed,
4602 * otherwise we'll refresh it from
4603 * on-disk bitmap and lose not-yet-available
4605 get_page(e4b->bd_buddy_page);
4606 get_page(e4b->bd_bitmap_page);
4610 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4611 if (cluster < entry->efd_start_cluster)
4613 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4614 n = &(*n)->rb_right;
4616 ext4_grp_locked_error(sb, group, 0,
4617 ext4_group_first_block_no(sb, group) +
4618 EXT4_C2B(sbi, cluster),
4619 "Block already on to-be-freed list");
4624 rb_link_node(new_node, parent, n);
4625 rb_insert_color(new_node, &db->bb_free_root);
4627 /* Now try to see the extent can be merged to left and right */
4628 node = rb_prev(new_node);
4630 entry = rb_entry(node, struct ext4_free_data, efd_node);
4631 if (can_merge(entry, new_entry) &&
4632 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4633 new_entry->efd_start_cluster = entry->efd_start_cluster;
4634 new_entry->efd_count += entry->efd_count;
4635 rb_erase(node, &(db->bb_free_root));
4636 kmem_cache_free(ext4_free_data_cachep, entry);
4640 node = rb_next(new_node);
4642 entry = rb_entry(node, struct ext4_free_data, efd_node);
4643 if (can_merge(new_entry, entry) &&
4644 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4645 new_entry->efd_count += entry->efd_count;
4646 rb_erase(node, &(db->bb_free_root));
4647 kmem_cache_free(ext4_free_data_cachep, entry);
4650 /* Add the extent to transaction's private list */
4651 ext4_journal_callback_add(handle, ext4_free_data_callback,
4652 &new_entry->efd_jce);
4657 * ext4_free_blocks() -- Free given blocks and update quota
4658 * @handle: handle for this transaction
4660 * @block: start physical block to free
4661 * @count: number of blocks to count
4662 * @flags: flags used by ext4_free_blocks
4664 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4665 struct buffer_head *bh, ext4_fsblk_t block,
4666 unsigned long count, int flags)
4668 struct buffer_head *bitmap_bh = NULL;
4669 struct super_block *sb = inode->i_sb;
4670 struct ext4_group_desc *gdp;
4671 unsigned int overflow;
4673 struct buffer_head *gd_bh;
4674 ext4_group_t block_group;
4675 struct ext4_sb_info *sbi;
4676 struct ext4_buddy e4b;
4677 unsigned int count_clusters;
4684 BUG_ON(block != bh->b_blocknr);
4686 block = bh->b_blocknr;
4690 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4691 !ext4_data_block_valid(sbi, block, count)) {
4692 ext4_error(sb, "Freeing blocks not in datazone - "
4693 "block = %llu, count = %lu", block, count);
4697 ext4_debug("freeing block %llu\n", block);
4698 trace_ext4_free_blocks(inode, block, count, flags);
4700 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4703 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4708 * If the extent to be freed does not begin on a cluster
4709 * boundary, we need to deal with partial clusters at the
4710 * beginning and end of the extent. Normally we will free
4711 * blocks at the beginning or the end unless we are explicitly
4712 * requested to avoid doing so.
4714 overflow = EXT4_PBLK_COFF(sbi, block);
4716 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4717 overflow = sbi->s_cluster_ratio - overflow;
4719 if (count > overflow)
4728 overflow = EXT4_LBLK_COFF(sbi, count);
4730 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4731 if (count > overflow)
4736 count += sbi->s_cluster_ratio - overflow;
4739 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4741 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
4743 for (i = 0; i < count; i++) {
4746 bh = sb_find_get_block(inode->i_sb, block + i);
4747 ext4_forget(handle, is_metadata, inode, bh, block + i);
4753 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4755 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4756 ext4_get_group_info(sb, block_group))))
4760 * Check to see if we are freeing blocks across a group
4763 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4764 overflow = EXT4_C2B(sbi, bit) + count -
4765 EXT4_BLOCKS_PER_GROUP(sb);
4768 count_clusters = EXT4_NUM_B2C(sbi, count);
4769 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4770 if (IS_ERR(bitmap_bh)) {
4771 err = PTR_ERR(bitmap_bh);
4775 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4781 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4782 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4783 in_range(block, ext4_inode_table(sb, gdp),
4784 EXT4_SB(sb)->s_itb_per_group) ||
4785 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4786 EXT4_SB(sb)->s_itb_per_group)) {
4788 ext4_error(sb, "Freeing blocks in system zone - "
4789 "Block = %llu, count = %lu", block, count);
4790 /* err = 0. ext4_std_error should be a no op */
4794 BUFFER_TRACE(bitmap_bh, "getting write access");
4795 err = ext4_journal_get_write_access(handle, bitmap_bh);
4800 * We are about to modify some metadata. Call the journal APIs
4801 * to unshare ->b_data if a currently-committing transaction is
4804 BUFFER_TRACE(gd_bh, "get_write_access");
4805 err = ext4_journal_get_write_access(handle, gd_bh);
4808 #ifdef AGGRESSIVE_CHECK
4811 for (i = 0; i < count_clusters; i++)
4812 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4815 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4817 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4818 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
4819 GFP_NOFS|__GFP_NOFAIL);
4824 * We need to make sure we don't reuse the freed block until after the
4825 * transaction is committed. We make an exception if the inode is to be
4826 * written in writeback mode since writeback mode has weak data
4827 * consistency guarantees.
4829 if (ext4_handle_valid(handle) &&
4830 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
4831 !ext4_should_writeback_data(inode))) {
4832 struct ext4_free_data *new_entry;
4834 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4837 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4838 GFP_NOFS|__GFP_NOFAIL);
4839 new_entry->efd_start_cluster = bit;
4840 new_entry->efd_group = block_group;
4841 new_entry->efd_count = count_clusters;
4842 new_entry->efd_tid = handle->h_transaction->t_tid;
4844 ext4_lock_group(sb, block_group);
4845 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4846 ext4_mb_free_metadata(handle, &e4b, new_entry);
4848 /* need to update group_info->bb_free and bitmap
4849 * with group lock held. generate_buddy look at
4850 * them with group lock_held
4852 if (test_opt(sb, DISCARD)) {
4853 err = ext4_issue_discard(sb, block_group, bit, count);
4854 if (err && err != -EOPNOTSUPP)
4855 ext4_msg(sb, KERN_WARNING, "discard request in"
4856 " group:%d block:%d count:%lu failed"
4857 " with %d", block_group, bit, count,
4860 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4862 ext4_lock_group(sb, block_group);
4863 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4864 mb_free_blocks(inode, &e4b, bit, count_clusters);
4867 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4868 ext4_free_group_clusters_set(sb, gdp, ret);
4869 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4870 ext4_group_desc_csum_set(sb, block_group, gdp);
4871 ext4_unlock_group(sb, block_group);
4873 if (sbi->s_log_groups_per_flex) {
4874 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4875 atomic64_add(count_clusters,
4876 &sbi->s_flex_groups[flex_group].free_clusters);
4879 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4880 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4881 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4883 ext4_mb_unload_buddy(&e4b);
4885 /* We dirtied the bitmap block */
4886 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4887 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4889 /* And the group descriptor block */
4890 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4891 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4895 if (overflow && !err) {
4903 ext4_std_error(sb, err);
4908 * ext4_group_add_blocks() -- Add given blocks to an existing group
4909 * @handle: handle to this transaction
4911 * @block: start physical block to add to the block group
4912 * @count: number of blocks to free
4914 * This marks the blocks as free in the bitmap and buddy.
4916 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4917 ext4_fsblk_t block, unsigned long count)
4919 struct buffer_head *bitmap_bh = NULL;
4920 struct buffer_head *gd_bh;
4921 ext4_group_t block_group;
4924 struct ext4_group_desc *desc;
4925 struct ext4_sb_info *sbi = EXT4_SB(sb);
4926 struct ext4_buddy e4b;
4927 int err = 0, ret, blk_free_count;
4928 ext4_grpblk_t blocks_freed;
4930 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4935 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4937 * Check to see if we are freeing blocks across a group
4940 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4941 ext4_warning(sb, "too much blocks added to group %u",
4947 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4948 if (IS_ERR(bitmap_bh)) {
4949 err = PTR_ERR(bitmap_bh);
4954 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4960 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4961 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4962 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4963 in_range(block + count - 1, ext4_inode_table(sb, desc),
4964 sbi->s_itb_per_group)) {
4965 ext4_error(sb, "Adding blocks in system zones - "
4966 "Block = %llu, count = %lu",
4972 BUFFER_TRACE(bitmap_bh, "getting write access");
4973 err = ext4_journal_get_write_access(handle, bitmap_bh);
4978 * We are about to modify some metadata. Call the journal APIs
4979 * to unshare ->b_data if a currently-committing transaction is
4982 BUFFER_TRACE(gd_bh, "get_write_access");
4983 err = ext4_journal_get_write_access(handle, gd_bh);
4987 for (i = 0, blocks_freed = 0; i < count; i++) {
4988 BUFFER_TRACE(bitmap_bh, "clear bit");
4989 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4990 ext4_error(sb, "bit already cleared for block %llu",
4991 (ext4_fsblk_t)(block + i));
4992 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4998 err = ext4_mb_load_buddy(sb, block_group, &e4b);
5003 * need to update group_info->bb_free and bitmap
5004 * with group lock held. generate_buddy look at
5005 * them with group lock_held
5007 ext4_lock_group(sb, block_group);
5008 mb_clear_bits(bitmap_bh->b_data, bit, count);
5009 mb_free_blocks(NULL, &e4b, bit, count);
5010 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
5011 ext4_free_group_clusters_set(sb, desc, blk_free_count);
5012 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5013 ext4_group_desc_csum_set(sb, block_group, desc);
5014 ext4_unlock_group(sb, block_group);
5015 percpu_counter_add(&sbi->s_freeclusters_counter,
5016 EXT4_NUM_B2C(sbi, blocks_freed));
5018 if (sbi->s_log_groups_per_flex) {
5019 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5020 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5021 &sbi->s_flex_groups[flex_group].free_clusters);
5024 ext4_mb_unload_buddy(&e4b);
5026 /* We dirtied the bitmap block */
5027 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5028 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5030 /* And the group descriptor block */
5031 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5032 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5038 ext4_std_error(sb, err);
5043 * ext4_trim_extent -- function to TRIM one single free extent in the group
5044 * @sb: super block for the file system
5045 * @start: starting block of the free extent in the alloc. group
5046 * @count: number of blocks to TRIM
5047 * @group: alloc. group we are working with
5048 * @e4b: ext4 buddy for the group
5050 * Trim "count" blocks starting at "start" in the "group". To assure that no
5051 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5052 * be called with under the group lock.
5054 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5055 ext4_group_t group, struct ext4_buddy *e4b)
5059 struct ext4_free_extent ex;
5062 trace_ext4_trim_extent(sb, group, start, count);
5064 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5066 ex.fe_start = start;
5067 ex.fe_group = group;
5071 * Mark blocks used, so no one can reuse them while
5074 mb_mark_used(e4b, &ex);
5075 ext4_unlock_group(sb, group);
5076 ret = ext4_issue_discard(sb, group, start, count);
5077 ext4_lock_group(sb, group);
5078 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5083 * ext4_trim_all_free -- function to trim all free space in alloc. group
5084 * @sb: super block for file system
5085 * @group: group to be trimmed
5086 * @start: first group block to examine
5087 * @max: last group block to examine
5088 * @minblocks: minimum extent block count
5090 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5091 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5095 * ext4_trim_all_free walks through group's block bitmap searching for free
5096 * extents. When the free extent is found, mark it as used in group buddy
5097 * bitmap. Then issue a TRIM command on this extent and free the extent in
5098 * the group buddy bitmap. This is done until whole group is scanned.
5100 static ext4_grpblk_t
5101 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5102 ext4_grpblk_t start, ext4_grpblk_t max,
5103 ext4_grpblk_t minblocks)
5106 ext4_grpblk_t next, count = 0, free_count = 0;
5107 struct ext4_buddy e4b;
5110 trace_ext4_trim_all_free(sb, group, start, max);
5112 ret = ext4_mb_load_buddy(sb, group, &e4b);
5114 ext4_error(sb, "Error in loading buddy "
5115 "information for %u", group);
5118 bitmap = e4b.bd_bitmap;
5120 ext4_lock_group(sb, group);
5121 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5122 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5125 start = (e4b.bd_info->bb_first_free > start) ?
5126 e4b.bd_info->bb_first_free : start;
5128 while (start <= max) {
5129 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5132 next = mb_find_next_bit(bitmap, max + 1, start);
5134 if ((next - start) >= minblocks) {
5135 ret = ext4_trim_extent(sb, start,
5136 next - start, group, &e4b);
5137 if (ret && ret != -EOPNOTSUPP)
5140 count += next - start;
5142 free_count += next - start;
5145 if (fatal_signal_pending(current)) {
5146 count = -ERESTARTSYS;
5150 if (need_resched()) {
5151 ext4_unlock_group(sb, group);
5153 ext4_lock_group(sb, group);
5156 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5162 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5165 ext4_unlock_group(sb, group);
5166 ext4_mb_unload_buddy(&e4b);
5168 ext4_debug("trimmed %d blocks in the group %d\n",
5175 * ext4_trim_fs() -- trim ioctl handle function
5176 * @sb: superblock for filesystem
5177 * @range: fstrim_range structure
5179 * start: First Byte to trim
5180 * len: number of Bytes to trim from start
5181 * minlen: minimum extent length in Bytes
5182 * ext4_trim_fs goes through all allocation groups containing Bytes from
5183 * start to start+len. For each such a group ext4_trim_all_free function
5184 * is invoked to trim all free space.
5186 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5188 struct ext4_group_info *grp;
5189 ext4_group_t group, first_group, last_group;
5190 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5191 uint64_t start, end, minlen, trimmed = 0;
5192 ext4_fsblk_t first_data_blk =
5193 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5194 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5197 start = range->start >> sb->s_blocksize_bits;
5198 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5199 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5200 range->minlen >> sb->s_blocksize_bits);
5202 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5203 start >= max_blks ||
5204 range->len < sb->s_blocksize)
5206 if (end >= max_blks)
5208 if (end <= first_data_blk)
5210 if (start < first_data_blk)
5211 start = first_data_blk;
5213 /* Determine first and last group to examine based on start and end */
5214 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5215 &first_group, &first_cluster);
5216 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5217 &last_group, &last_cluster);
5219 /* end now represents the last cluster to discard in this group */
5220 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5222 for (group = first_group; group <= last_group; group++) {
5223 grp = ext4_get_group_info(sb, group);
5224 /* We only do this if the grp has never been initialized */
5225 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5226 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5232 * For all the groups except the last one, last cluster will
5233 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5234 * change it for the last group, note that last_cluster is
5235 * already computed earlier by ext4_get_group_no_and_offset()
5237 if (group == last_group)
5240 if (grp->bb_free >= minlen) {
5241 cnt = ext4_trim_all_free(sb, group, first_cluster,
5251 * For every group except the first one, we are sure
5252 * that the first cluster to discard will be cluster #0.
5258 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5261 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;