2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
24 #include "ext4_jbd2.h"
26 #include <linux/log2.h>
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <trace/events/ext4.h>
31 #ifdef CONFIG_EXT4_DEBUG
32 ushort ext4_mballoc_debug __read_mostly;
34 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
35 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
40 * - test ext4_ext_search_left() and ext4_ext_search_right()
41 * - search for metadata in few groups
44 * - normalization should take into account whether file is still open
45 * - discard preallocations if no free space left (policy?)
46 * - don't normalize tails
48 * - reservation for superuser
51 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
52 * - track min/max extents in each group for better group selection
53 * - mb_mark_used() may allocate chunk right after splitting buddy
54 * - tree of groups sorted by number of free blocks
59 * The allocation request involve request for multiple number of blocks
60 * near to the goal(block) value specified.
62 * During initialization phase of the allocator we decide to use the
63 * group preallocation or inode preallocation depending on the size of
64 * the file. The size of the file could be the resulting file size we
65 * would have after allocation, or the current file size, which ever
66 * is larger. If the size is less than sbi->s_mb_stream_request we
67 * select to use the group preallocation. The default value of
68 * s_mb_stream_request is 16 blocks. This can also be tuned via
69 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
70 * terms of number of blocks.
72 * The main motivation for having small file use group preallocation is to
73 * ensure that we have small files closer together on the disk.
75 * First stage the allocator looks at the inode prealloc list,
76 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
77 * spaces for this particular inode. The inode prealloc space is
80 * pa_lstart -> the logical start block for this prealloc space
81 * pa_pstart -> the physical start block for this prealloc space
82 * pa_len -> length for this prealloc space (in clusters)
83 * pa_free -> free space available in this prealloc space (in clusters)
85 * The inode preallocation space is used looking at the _logical_ start
86 * block. If only the logical file block falls within the range of prealloc
87 * space we will consume the particular prealloc space. This makes sure that
88 * we have contiguous physical blocks representing the file blocks
90 * The important thing to be noted in case of inode prealloc space is that
91 * we don't modify the values associated to inode prealloc space except
94 * If we are not able to find blocks in the inode prealloc space and if we
95 * have the group allocation flag set then we look at the locality group
96 * prealloc space. These are per CPU prealloc list represented as
98 * ext4_sb_info.s_locality_groups[smp_processor_id()]
100 * The reason for having a per cpu locality group is to reduce the contention
101 * between CPUs. It is possible to get scheduled at this point.
103 * The locality group prealloc space is used looking at whether we have
104 * enough free space (pa_free) within the prealloc space.
106 * If we can't allocate blocks via inode prealloc or/and locality group
107 * prealloc then we look at the buddy cache. The buddy cache is represented
108 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
109 * mapped to the buddy and bitmap information regarding different
110 * groups. The buddy information is attached to buddy cache inode so that
111 * we can access them through the page cache. The information regarding
112 * each group is loaded via ext4_mb_load_buddy. The information involve
113 * block bitmap and buddy information. The information are stored in the
117 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
120 * one block each for bitmap and buddy information. So for each group we
121 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
122 * blocksize) blocks. So it can have information regarding groups_per_page
123 * which is blocks_per_page/2
125 * The buddy cache inode is not stored on disk. The inode is thrown
126 * away when the filesystem is unmounted.
128 * We look for count number of blocks in the buddy cache. If we were able
129 * to locate that many free blocks we return with additional information
130 * regarding rest of the contiguous physical block available
132 * Before allocating blocks via buddy cache we normalize the request
133 * blocks. This ensure we ask for more blocks that we needed. The extra
134 * blocks that we get after allocation is added to the respective prealloc
135 * list. In case of inode preallocation we follow a list of heuristics
136 * based on file size. This can be found in ext4_mb_normalize_request. If
137 * we are doing a group prealloc we try to normalize the request to
138 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
139 * dependent on the cluster size; for non-bigalloc file systems, it is
140 * 512 blocks. This can be tuned via
141 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
142 * terms of number of blocks. If we have mounted the file system with -O
143 * stripe=<value> option the group prealloc request is normalized to the
144 * the smallest multiple of the stripe value (sbi->s_stripe) which is
145 * greater than the default mb_group_prealloc.
147 * The regular allocator (using the buddy cache) supports a few tunables.
149 * /sys/fs/ext4/<partition>/mb_min_to_scan
150 * /sys/fs/ext4/<partition>/mb_max_to_scan
151 * /sys/fs/ext4/<partition>/mb_order2_req
153 * The regular allocator uses buddy scan only if the request len is power of
154 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
155 * value of s_mb_order2_reqs can be tuned via
156 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
157 * stripe size (sbi->s_stripe), we try to search for contiguous block in
158 * stripe size. This should result in better allocation on RAID setups. If
159 * not, we search in the specific group using bitmap for best extents. The
160 * tunable min_to_scan and max_to_scan control the behaviour here.
161 * min_to_scan indicate how long the mballoc __must__ look for a best
162 * extent and max_to_scan indicates how long the mballoc __can__ look for a
163 * best extent in the found extents. Searching for the blocks starts with
164 * the group specified as the goal value in allocation context via
165 * ac_g_ex. Each group is first checked based on the criteria whether it
166 * can be used for allocation. ext4_mb_good_group explains how the groups are
169 * Both the prealloc space are getting populated as above. So for the first
170 * request we will hit the buddy cache which will result in this prealloc
171 * space getting filled. The prealloc space is then later used for the
172 * subsequent request.
176 * mballoc operates on the following data:
178 * - in-core buddy (actually includes buddy and bitmap)
179 * - preallocation descriptors (PAs)
181 * there are two types of preallocations:
183 * assiged to specific inode and can be used for this inode only.
184 * it describes part of inode's space preallocated to specific
185 * physical blocks. any block from that preallocated can be used
186 * independent. the descriptor just tracks number of blocks left
187 * unused. so, before taking some block from descriptor, one must
188 * make sure corresponded logical block isn't allocated yet. this
189 * also means that freeing any block within descriptor's range
190 * must discard all preallocated blocks.
192 * assigned to specific locality group which does not translate to
193 * permanent set of inodes: inode can join and leave group. space
194 * from this type of preallocation can be used for any inode. thus
195 * it's consumed from the beginning to the end.
197 * relation between them can be expressed as:
198 * in-core buddy = on-disk bitmap + preallocation descriptors
200 * this mean blocks mballoc considers used are:
201 * - allocated blocks (persistent)
202 * - preallocated blocks (non-persistent)
204 * consistency in mballoc world means that at any time a block is either
205 * free or used in ALL structures. notice: "any time" should not be read
206 * literally -- time is discrete and delimited by locks.
208 * to keep it simple, we don't use block numbers, instead we count number of
209 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
211 * all operations can be expressed as:
212 * - init buddy: buddy = on-disk + PAs
213 * - new PA: buddy += N; PA = N
214 * - use inode PA: on-disk += N; PA -= N
215 * - discard inode PA buddy -= on-disk - PA; PA = 0
216 * - use locality group PA on-disk += N; PA -= N
217 * - discard locality group PA buddy -= PA; PA = 0
218 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
219 * is used in real operation because we can't know actual used
220 * bits from PA, only from on-disk bitmap
222 * if we follow this strict logic, then all operations above should be atomic.
223 * given some of them can block, we'd have to use something like semaphores
224 * killing performance on high-end SMP hardware. let's try to relax it using
225 * the following knowledge:
226 * 1) if buddy is referenced, it's already initialized
227 * 2) while block is used in buddy and the buddy is referenced,
228 * nobody can re-allocate that block
229 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
230 * bit set and PA claims same block, it's OK. IOW, one can set bit in
231 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
234 * so, now we're building a concurrency table:
237 * blocks for PA are allocated in the buddy, buddy must be referenced
238 * until PA is linked to allocation group to avoid concurrent buddy init
240 * we need to make sure that either on-disk bitmap or PA has uptodate data
241 * given (3) we care that PA-=N operation doesn't interfere with init
243 * the simplest way would be to have buddy initialized by the discard
244 * - use locality group PA
245 * again PA-=N must be serialized with init
246 * - discard locality group PA
247 * the simplest way would be to have buddy initialized by the discard
250 * i_data_sem serializes them
252 * discard process must wait until PA isn't used by another process
253 * - use locality group PA
254 * some mutex should serialize them
255 * - discard locality group PA
256 * discard process must wait until PA isn't used by another process
259 * i_data_sem or another mutex should serializes them
261 * discard process must wait until PA isn't used by another process
262 * - use locality group PA
263 * nothing wrong here -- they're different PAs covering different blocks
264 * - discard locality group PA
265 * discard process must wait until PA isn't used by another process
267 * now we're ready to make few consequences:
268 * - PA is referenced and while it is no discard is possible
269 * - PA is referenced until block isn't marked in on-disk bitmap
270 * - PA changes only after on-disk bitmap
271 * - discard must not compete with init. either init is done before
272 * any discard or they're serialized somehow
273 * - buddy init as sum of on-disk bitmap and PAs is done atomically
275 * a special case when we've used PA to emptiness. no need to modify buddy
276 * in this case, but we should care about concurrent init
281 * Logic in few words:
286 * mark bits in on-disk bitmap
289 * - use preallocation:
290 * find proper PA (per-inode or group)
292 * mark bits in on-disk bitmap
298 * mark bits in on-disk bitmap
301 * - discard preallocations in group:
303 * move them onto local list
304 * load on-disk bitmap
306 * remove PA from object (inode or locality group)
307 * mark free blocks in-core
309 * - discard inode's preallocations:
316 * - bitlock on a group (group)
317 * - object (inode/locality) (object)
328 * - release consumed pa:
333 * - generate in-core bitmap:
337 * - discard all for given object (inode, locality group):
342 * - discard all for given group:
349 static struct kmem_cache *ext4_pspace_cachep;
350 static struct kmem_cache *ext4_ac_cachep;
351 static struct kmem_cache *ext4_free_data_cachep;
353 /* We create slab caches for groupinfo data structures based on the
354 * superblock block size. There will be one per mounted filesystem for
355 * each unique s_blocksize_bits */
356 #define NR_GRPINFO_CACHES 8
357 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
359 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
360 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
361 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
362 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
365 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
367 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
369 static void ext4_free_data_callback(struct super_block *sb,
370 struct ext4_journal_cb_entry *jce, int rc);
372 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
374 #if BITS_PER_LONG == 64
375 *bit += ((unsigned long) addr & 7UL) << 3;
376 addr = (void *) ((unsigned long) addr & ~7UL);
377 #elif BITS_PER_LONG == 32
378 *bit += ((unsigned long) addr & 3UL) << 3;
379 addr = (void *) ((unsigned long) addr & ~3UL);
381 #error "how many bits you are?!"
386 static inline int mb_test_bit(int bit, void *addr)
389 * ext4_test_bit on architecture like powerpc
390 * needs unsigned long aligned address
392 addr = mb_correct_addr_and_bit(&bit, addr);
393 return ext4_test_bit(bit, addr);
396 static inline void mb_set_bit(int bit, void *addr)
398 addr = mb_correct_addr_and_bit(&bit, addr);
399 ext4_set_bit(bit, addr);
402 static inline void mb_clear_bit(int bit, void *addr)
404 addr = mb_correct_addr_and_bit(&bit, addr);
405 ext4_clear_bit(bit, addr);
408 static inline int mb_test_and_clear_bit(int bit, void *addr)
410 addr = mb_correct_addr_and_bit(&bit, addr);
411 return ext4_test_and_clear_bit(bit, addr);
414 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
416 int fix = 0, ret, tmpmax;
417 addr = mb_correct_addr_and_bit(&fix, addr);
421 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
427 static inline int mb_find_next_bit(void *addr, int max, int start)
429 int fix = 0, ret, tmpmax;
430 addr = mb_correct_addr_and_bit(&fix, addr);
434 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
440 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
444 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
447 if (order > e4b->bd_blkbits + 1) {
452 /* at order 0 we see each particular block */
454 *max = 1 << (e4b->bd_blkbits + 3);
455 return e4b->bd_bitmap;
458 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
459 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
465 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
466 int first, int count)
469 struct super_block *sb = e4b->bd_sb;
471 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
473 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
474 for (i = 0; i < count; i++) {
475 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
476 ext4_fsblk_t blocknr;
478 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
479 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
480 ext4_grp_locked_error(sb, e4b->bd_group,
481 inode ? inode->i_ino : 0,
483 "freeing block already freed "
487 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
491 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
495 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
497 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
498 for (i = 0; i < count; i++) {
499 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
500 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
504 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
506 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
507 unsigned char *b1, *b2;
509 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
510 b2 = (unsigned char *) bitmap;
511 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
512 if (b1[i] != b2[i]) {
513 ext4_msg(e4b->bd_sb, KERN_ERR,
514 "corruption in group %u "
515 "at byte %u(%u): %x in copy != %x "
517 e4b->bd_group, i, i * 8, b1[i], b2[i]);
525 static inline void mb_free_blocks_double(struct inode *inode,
526 struct ext4_buddy *e4b, int first, int count)
530 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
531 int first, int count)
535 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
541 #ifdef AGGRESSIVE_CHECK
543 #define MB_CHECK_ASSERT(assert) \
547 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
548 function, file, line, # assert); \
553 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
554 const char *function, int line)
556 struct super_block *sb = e4b->bd_sb;
557 int order = e4b->bd_blkbits + 1;
564 struct ext4_group_info *grp;
567 struct list_head *cur;
572 static int mb_check_counter;
573 if (mb_check_counter++ % 100 != 0)
578 buddy = mb_find_buddy(e4b, order, &max);
579 MB_CHECK_ASSERT(buddy);
580 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
581 MB_CHECK_ASSERT(buddy2);
582 MB_CHECK_ASSERT(buddy != buddy2);
583 MB_CHECK_ASSERT(max * 2 == max2);
586 for (i = 0; i < max; i++) {
588 if (mb_test_bit(i, buddy)) {
589 /* only single bit in buddy2 may be 1 */
590 if (!mb_test_bit(i << 1, buddy2)) {
592 mb_test_bit((i<<1)+1, buddy2));
593 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
595 mb_test_bit(i << 1, buddy2));
600 /* both bits in buddy2 must be 1 */
601 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
602 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
604 for (j = 0; j < (1 << order); j++) {
605 k = (i * (1 << order)) + j;
607 !mb_test_bit(k, e4b->bd_bitmap));
611 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
616 buddy = mb_find_buddy(e4b, 0, &max);
617 for (i = 0; i < max; i++) {
618 if (!mb_test_bit(i, buddy)) {
619 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
627 /* check used bits only */
628 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
629 buddy2 = mb_find_buddy(e4b, j, &max2);
631 MB_CHECK_ASSERT(k < max2);
632 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
635 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
636 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
638 grp = ext4_get_group_info(sb, e4b->bd_group);
639 list_for_each(cur, &grp->bb_prealloc_list) {
640 ext4_group_t groupnr;
641 struct ext4_prealloc_space *pa;
642 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
643 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
644 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
645 for (i = 0; i < pa->pa_len; i++)
646 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
650 #undef MB_CHECK_ASSERT
651 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
652 __FILE__, __func__, __LINE__)
654 #define mb_check_buddy(e4b)
658 * Divide blocks started from @first with length @len into
659 * smaller chunks with power of 2 blocks.
660 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
661 * then increase bb_counters[] for corresponded chunk size.
663 static void ext4_mb_mark_free_simple(struct super_block *sb,
664 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
665 struct ext4_group_info *grp)
667 struct ext4_sb_info *sbi = EXT4_SB(sb);
671 unsigned short border;
673 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
675 border = 2 << sb->s_blocksize_bits;
678 /* find how many blocks can be covered since this position */
679 max = ffs(first | border) - 1;
681 /* find how many blocks of power 2 we need to mark */
688 /* mark multiblock chunks only */
689 grp->bb_counters[min]++;
691 mb_clear_bit(first >> min,
692 buddy + sbi->s_mb_offsets[min]);
700 * Cache the order of the largest free extent we have available in this block
704 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
709 grp->bb_largest_free_order = -1; /* uninit */
711 bits = sb->s_blocksize_bits + 1;
712 for (i = bits; i >= 0; i--) {
713 if (grp->bb_counters[i] > 0) {
714 grp->bb_largest_free_order = i;
720 static noinline_for_stack
721 void ext4_mb_generate_buddy(struct super_block *sb,
722 void *buddy, void *bitmap, ext4_group_t group)
724 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
725 struct ext4_sb_info *sbi = EXT4_SB(sb);
726 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
731 unsigned fragments = 0;
732 unsigned long long period = get_cycles();
734 /* initialize buddy from bitmap which is aggregation
735 * of on-disk bitmap and preallocations */
736 i = mb_find_next_zero_bit(bitmap, max, 0);
737 grp->bb_first_free = i;
741 i = mb_find_next_bit(bitmap, max, i);
745 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
747 grp->bb_counters[0]++;
749 i = mb_find_next_zero_bit(bitmap, max, i);
751 grp->bb_fragments = fragments;
753 if (free != grp->bb_free) {
754 ext4_grp_locked_error(sb, group, 0, 0,
755 "block bitmap and bg descriptor "
756 "inconsistent: %u vs %u free clusters",
759 * If we intend to continue, we consider group descriptor
760 * corrupt and update bb_free using bitmap value
763 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
764 percpu_counter_sub(&sbi->s_freeclusters_counter,
766 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
768 mb_set_largest_free_order(sb, grp);
770 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
772 period = get_cycles() - period;
773 spin_lock(&EXT4_SB(sb)->s_bal_lock);
774 EXT4_SB(sb)->s_mb_buddies_generated++;
775 EXT4_SB(sb)->s_mb_generation_time += period;
776 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
779 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
785 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
786 ext4_set_bits(buddy, 0, count);
788 e4b->bd_info->bb_fragments = 0;
789 memset(e4b->bd_info->bb_counters, 0,
790 sizeof(*e4b->bd_info->bb_counters) *
791 (e4b->bd_sb->s_blocksize_bits + 2));
793 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
794 e4b->bd_bitmap, e4b->bd_group);
797 /* The buddy information is attached the buddy cache inode
798 * for convenience. The information regarding each group
799 * is loaded via ext4_mb_load_buddy. The information involve
800 * block bitmap and buddy information. The information are
801 * stored in the inode as
804 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
807 * one block each for bitmap and buddy information.
808 * So for each group we take up 2 blocks. A page can
809 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
810 * So it can have information regarding groups_per_page which
811 * is blocks_per_page/2
813 * Locking note: This routine takes the block group lock of all groups
814 * for this page; do not hold this lock when calling this routine!
817 static int ext4_mb_init_cache(struct page *page, char *incore)
819 ext4_group_t ngroups;
825 ext4_group_t first_group, group;
827 struct super_block *sb;
828 struct buffer_head *bhs;
829 struct buffer_head **bh = NULL;
833 struct ext4_group_info *grinfo;
835 mb_debug(1, "init page %lu\n", page->index);
837 inode = page->mapping->host;
839 ngroups = ext4_get_groups_count(sb);
840 blocksize = 1 << inode->i_blkbits;
841 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
843 groups_per_page = blocks_per_page >> 1;
844 if (groups_per_page == 0)
847 /* allocate buffer_heads to read bitmaps */
848 if (groups_per_page > 1) {
849 i = sizeof(struct buffer_head *) * groups_per_page;
850 bh = kzalloc(i, GFP_NOFS);
858 first_group = page->index * blocks_per_page / 2;
860 /* read all groups the page covers into the cache */
861 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
862 if (group >= ngroups)
865 grinfo = ext4_get_group_info(sb, group);
867 * If page is uptodate then we came here after online resize
868 * which added some new uninitialized group info structs, so
869 * we must skip all initialized uptodate buddies on the page,
870 * which may be currently in use by an allocating task.
872 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
876 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
880 mb_debug(1, "read bitmap for group %u\n", group);
883 /* wait for I/O completion */
884 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
885 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
891 first_block = page->index * blocks_per_page;
892 for (i = 0; i < blocks_per_page; i++) {
893 group = (first_block + i) >> 1;
894 if (group >= ngroups)
897 if (!bh[group - first_group])
898 /* skip initialized uptodate buddy */
902 * data carry information regarding this
903 * particular group in the format specified
907 data = page_address(page) + (i * blocksize);
908 bitmap = bh[group - first_group]->b_data;
911 * We place the buddy block and bitmap block
914 if ((first_block + i) & 1) {
915 /* this is block of buddy */
916 BUG_ON(incore == NULL);
917 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
918 group, page->index, i * blocksize);
919 trace_ext4_mb_buddy_bitmap_load(sb, group);
920 grinfo = ext4_get_group_info(sb, group);
921 grinfo->bb_fragments = 0;
922 memset(grinfo->bb_counters, 0,
923 sizeof(*grinfo->bb_counters) *
924 (sb->s_blocksize_bits+2));
926 * incore got set to the group block bitmap below
928 ext4_lock_group(sb, group);
930 memset(data, 0xff, blocksize);
931 ext4_mb_generate_buddy(sb, data, incore, group);
932 ext4_unlock_group(sb, group);
935 /* this is block of bitmap */
936 BUG_ON(incore != NULL);
937 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
938 group, page->index, i * blocksize);
939 trace_ext4_mb_bitmap_load(sb, group);
941 /* see comments in ext4_mb_put_pa() */
942 ext4_lock_group(sb, group);
943 memcpy(data, bitmap, blocksize);
945 /* mark all preallocated blks used in in-core bitmap */
946 ext4_mb_generate_from_pa(sb, data, group);
947 ext4_mb_generate_from_freelist(sb, data, group);
948 ext4_unlock_group(sb, group);
950 /* set incore so that the buddy information can be
951 * generated using this
956 SetPageUptodate(page);
960 for (i = 0; i < groups_per_page; i++)
969 * Lock the buddy and bitmap pages. This make sure other parallel init_group
970 * on the same buddy page doesn't happen whild holding the buddy page lock.
971 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
972 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
974 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
975 ext4_group_t group, struct ext4_buddy *e4b)
977 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
978 int block, pnum, poff;
982 e4b->bd_buddy_page = NULL;
983 e4b->bd_bitmap_page = NULL;
985 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
987 * the buddy cache inode stores the block bitmap
988 * and buddy information in consecutive blocks.
989 * So for each group we need two blocks.
992 pnum = block / blocks_per_page;
993 poff = block % blocks_per_page;
994 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
997 BUG_ON(page->mapping != inode->i_mapping);
998 e4b->bd_bitmap_page = page;
999 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1001 if (blocks_per_page >= 2) {
1002 /* buddy and bitmap are on the same page */
1007 pnum = block / blocks_per_page;
1008 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1011 BUG_ON(page->mapping != inode->i_mapping);
1012 e4b->bd_buddy_page = page;
1016 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1018 if (e4b->bd_bitmap_page) {
1019 unlock_page(e4b->bd_bitmap_page);
1020 page_cache_release(e4b->bd_bitmap_page);
1022 if (e4b->bd_buddy_page) {
1023 unlock_page(e4b->bd_buddy_page);
1024 page_cache_release(e4b->bd_buddy_page);
1029 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1030 * block group lock of all groups for this page; do not hold the BG lock when
1031 * calling this routine!
1033 static noinline_for_stack
1034 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1037 struct ext4_group_info *this_grp;
1038 struct ext4_buddy e4b;
1043 mb_debug(1, "init group %u\n", group);
1044 this_grp = ext4_get_group_info(sb, group);
1046 * This ensures that we don't reinit the buddy cache
1047 * page which map to the group from which we are already
1048 * allocating. If we are looking at the buddy cache we would
1049 * have taken a reference using ext4_mb_load_buddy and that
1050 * would have pinned buddy page to page cache.
1051 * The call to ext4_mb_get_buddy_page_lock will mark the
1054 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1055 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1057 * somebody initialized the group
1058 * return without doing anything
1063 page = e4b.bd_bitmap_page;
1064 ret = ext4_mb_init_cache(page, NULL);
1067 if (!PageUptodate(page)) {
1072 if (e4b.bd_buddy_page == NULL) {
1074 * If both the bitmap and buddy are in
1075 * the same page we don't need to force
1081 /* init buddy cache */
1082 page = e4b.bd_buddy_page;
1083 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1086 if (!PageUptodate(page)) {
1091 ext4_mb_put_buddy_page_lock(&e4b);
1096 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1097 * block group lock of all groups for this page; do not hold the BG lock when
1098 * calling this routine!
1100 static noinline_for_stack int
1101 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1102 struct ext4_buddy *e4b)
1104 int blocks_per_page;
1110 struct ext4_group_info *grp;
1111 struct ext4_sb_info *sbi = EXT4_SB(sb);
1112 struct inode *inode = sbi->s_buddy_cache;
1115 mb_debug(1, "load group %u\n", group);
1117 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1118 grp = ext4_get_group_info(sb, group);
1120 e4b->bd_blkbits = sb->s_blocksize_bits;
1123 e4b->bd_group = group;
1124 e4b->bd_buddy_page = NULL;
1125 e4b->bd_bitmap_page = NULL;
1127 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1129 * we need full data about the group
1130 * to make a good selection
1132 ret = ext4_mb_init_group(sb, group);
1138 * the buddy cache inode stores the block bitmap
1139 * and buddy information in consecutive blocks.
1140 * So for each group we need two blocks.
1143 pnum = block / blocks_per_page;
1144 poff = block % blocks_per_page;
1146 /* we could use find_or_create_page(), but it locks page
1147 * what we'd like to avoid in fast path ... */
1148 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1149 if (page == NULL || !PageUptodate(page)) {
1152 * drop the page reference and try
1153 * to get the page with lock. If we
1154 * are not uptodate that implies
1155 * somebody just created the page but
1156 * is yet to initialize the same. So
1157 * wait for it to initialize.
1159 page_cache_release(page);
1160 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1162 BUG_ON(page->mapping != inode->i_mapping);
1163 if (!PageUptodate(page)) {
1164 ret = ext4_mb_init_cache(page, NULL);
1169 mb_cmp_bitmaps(e4b, page_address(page) +
1170 (poff * sb->s_blocksize));
1179 if (!PageUptodate(page)) {
1184 /* Pages marked accessed already */
1185 e4b->bd_bitmap_page = page;
1186 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1189 pnum = block / blocks_per_page;
1190 poff = block % blocks_per_page;
1192 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1193 if (page == NULL || !PageUptodate(page)) {
1195 page_cache_release(page);
1196 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1198 BUG_ON(page->mapping != inode->i_mapping);
1199 if (!PageUptodate(page)) {
1200 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1213 if (!PageUptodate(page)) {
1218 /* Pages marked accessed already */
1219 e4b->bd_buddy_page = page;
1220 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1222 BUG_ON(e4b->bd_bitmap_page == NULL);
1223 BUG_ON(e4b->bd_buddy_page == NULL);
1229 page_cache_release(page);
1230 if (e4b->bd_bitmap_page)
1231 page_cache_release(e4b->bd_bitmap_page);
1232 if (e4b->bd_buddy_page)
1233 page_cache_release(e4b->bd_buddy_page);
1234 e4b->bd_buddy = NULL;
1235 e4b->bd_bitmap = NULL;
1239 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1241 if (e4b->bd_bitmap_page)
1242 page_cache_release(e4b->bd_bitmap_page);
1243 if (e4b->bd_buddy_page)
1244 page_cache_release(e4b->bd_buddy_page);
1248 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1253 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1254 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1257 while (order <= e4b->bd_blkbits + 1) {
1259 if (!mb_test_bit(block, bb)) {
1260 /* this block is part of buddy of order 'order' */
1263 bb += 1 << (e4b->bd_blkbits - order);
1269 static void mb_clear_bits(void *bm, int cur, int len)
1275 if ((cur & 31) == 0 && (len - cur) >= 32) {
1276 /* fast path: clear whole word at once */
1277 addr = bm + (cur >> 3);
1282 mb_clear_bit(cur, bm);
1287 /* clear bits in given range
1288 * will return first found zero bit if any, -1 otherwise
1290 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1297 if ((cur & 31) == 0 && (len - cur) >= 32) {
1298 /* fast path: clear whole word at once */
1299 addr = bm + (cur >> 3);
1300 if (*addr != (__u32)(-1) && zero_bit == -1)
1301 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1306 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1314 void ext4_set_bits(void *bm, int cur, int len)
1320 if ((cur & 31) == 0 && (len - cur) >= 32) {
1321 /* fast path: set whole word at once */
1322 addr = bm + (cur >> 3);
1327 mb_set_bit(cur, bm);
1333 * _________________________________________________________________ */
1335 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1337 if (mb_test_bit(*bit + side, bitmap)) {
1338 mb_clear_bit(*bit, bitmap);
1344 mb_set_bit(*bit, bitmap);
1349 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1353 void *buddy = mb_find_buddy(e4b, order, &max);
1358 /* Bits in range [first; last] are known to be set since
1359 * corresponding blocks were allocated. Bits in range
1360 * (first; last) will stay set because they form buddies on
1361 * upper layer. We just deal with borders if they don't
1362 * align with upper layer and then go up.
1363 * Releasing entire group is all about clearing
1364 * single bit of highest order buddy.
1368 * ---------------------------------
1370 * ---------------------------------
1371 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1372 * ---------------------------------
1374 * \_____________________/
1376 * Neither [1] nor [6] is aligned to above layer.
1377 * Left neighbour [0] is free, so mark it busy,
1378 * decrease bb_counters and extend range to
1380 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1381 * mark [6] free, increase bb_counters and shrink range to
1383 * Then shift range to [0; 2], go up and do the same.
1388 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1390 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1395 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1396 mb_clear_bits(buddy, first, last - first + 1);
1397 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1406 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1407 int first, int count)
1409 int left_is_free = 0;
1410 int right_is_free = 0;
1412 int last = first + count - 1;
1413 struct super_block *sb = e4b->bd_sb;
1415 BUG_ON(last >= (sb->s_blocksize << 3));
1416 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1417 /* Don't bother if the block group is corrupt. */
1418 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1421 mb_check_buddy(e4b);
1422 mb_free_blocks_double(inode, e4b, first, count);
1424 e4b->bd_info->bb_free += count;
1425 if (first < e4b->bd_info->bb_first_free)
1426 e4b->bd_info->bb_first_free = first;
1428 /* access memory sequentially: check left neighbour,
1429 * clear range and then check right neighbour
1432 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1433 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1434 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1435 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1437 if (unlikely(block != -1)) {
1438 struct ext4_sb_info *sbi = EXT4_SB(sb);
1439 ext4_fsblk_t blocknr;
1441 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1442 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1443 ext4_grp_locked_error(sb, e4b->bd_group,
1444 inode ? inode->i_ino : 0,
1446 "freeing already freed block "
1447 "(bit %u); block bitmap corrupt.",
1449 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1450 percpu_counter_sub(&sbi->s_freeclusters_counter,
1451 e4b->bd_info->bb_free);
1452 /* Mark the block group as corrupt. */
1453 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1454 &e4b->bd_info->bb_state);
1455 mb_regenerate_buddy(e4b);
1459 /* let's maintain fragments counter */
1460 if (left_is_free && right_is_free)
1461 e4b->bd_info->bb_fragments--;
1462 else if (!left_is_free && !right_is_free)
1463 e4b->bd_info->bb_fragments++;
1465 /* buddy[0] == bd_bitmap is a special case, so handle
1466 * it right away and let mb_buddy_mark_free stay free of
1467 * zero order checks.
1468 * Check if neighbours are to be coaleasced,
1469 * adjust bitmap bb_counters and borders appropriately.
1472 first += !left_is_free;
1473 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1476 last -= !right_is_free;
1477 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1481 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1484 mb_set_largest_free_order(sb, e4b->bd_info);
1485 mb_check_buddy(e4b);
1488 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1489 int needed, struct ext4_free_extent *ex)
1495 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1498 buddy = mb_find_buddy(e4b, 0, &max);
1499 BUG_ON(buddy == NULL);
1500 BUG_ON(block >= max);
1501 if (mb_test_bit(block, buddy)) {
1508 /* find actual order */
1509 order = mb_find_order_for_block(e4b, block);
1510 block = block >> order;
1512 ex->fe_len = 1 << order;
1513 ex->fe_start = block << order;
1514 ex->fe_group = e4b->bd_group;
1516 /* calc difference from given start */
1517 next = next - ex->fe_start;
1519 ex->fe_start += next;
1521 while (needed > ex->fe_len &&
1522 mb_find_buddy(e4b, order, &max)) {
1524 if (block + 1 >= max)
1527 next = (block + 1) * (1 << order);
1528 if (mb_test_bit(next, e4b->bd_bitmap))
1531 order = mb_find_order_for_block(e4b, next);
1533 block = next >> order;
1534 ex->fe_len += 1 << order;
1537 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1541 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1547 int start = ex->fe_start;
1548 int len = ex->fe_len;
1553 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1554 BUG_ON(e4b->bd_group != ex->fe_group);
1555 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1556 mb_check_buddy(e4b);
1557 mb_mark_used_double(e4b, start, len);
1559 e4b->bd_info->bb_free -= len;
1560 if (e4b->bd_info->bb_first_free == start)
1561 e4b->bd_info->bb_first_free += len;
1563 /* let's maintain fragments counter */
1565 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1566 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1567 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1569 e4b->bd_info->bb_fragments++;
1570 else if (!mlen && !max)
1571 e4b->bd_info->bb_fragments--;
1573 /* let's maintain buddy itself */
1575 ord = mb_find_order_for_block(e4b, start);
1577 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1578 /* the whole chunk may be allocated at once! */
1580 buddy = mb_find_buddy(e4b, ord, &max);
1581 BUG_ON((start >> ord) >= max);
1582 mb_set_bit(start >> ord, buddy);
1583 e4b->bd_info->bb_counters[ord]--;
1590 /* store for history */
1592 ret = len | (ord << 16);
1594 /* we have to split large buddy */
1596 buddy = mb_find_buddy(e4b, ord, &max);
1597 mb_set_bit(start >> ord, buddy);
1598 e4b->bd_info->bb_counters[ord]--;
1601 cur = (start >> ord) & ~1U;
1602 buddy = mb_find_buddy(e4b, ord, &max);
1603 mb_clear_bit(cur, buddy);
1604 mb_clear_bit(cur + 1, buddy);
1605 e4b->bd_info->bb_counters[ord]++;
1606 e4b->bd_info->bb_counters[ord]++;
1608 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1610 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1611 mb_check_buddy(e4b);
1617 * Must be called under group lock!
1619 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1620 struct ext4_buddy *e4b)
1622 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1625 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1626 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1628 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1629 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1630 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1632 /* preallocation can change ac_b_ex, thus we store actually
1633 * allocated blocks for history */
1634 ac->ac_f_ex = ac->ac_b_ex;
1636 ac->ac_status = AC_STATUS_FOUND;
1637 ac->ac_tail = ret & 0xffff;
1638 ac->ac_buddy = ret >> 16;
1641 * take the page reference. We want the page to be pinned
1642 * so that we don't get a ext4_mb_init_cache_call for this
1643 * group until we update the bitmap. That would mean we
1644 * double allocate blocks. The reference is dropped
1645 * in ext4_mb_release_context
1647 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1648 get_page(ac->ac_bitmap_page);
1649 ac->ac_buddy_page = e4b->bd_buddy_page;
1650 get_page(ac->ac_buddy_page);
1651 /* store last allocated for subsequent stream allocation */
1652 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1653 spin_lock(&sbi->s_md_lock);
1654 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1655 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1656 spin_unlock(&sbi->s_md_lock);
1661 * regular allocator, for general purposes allocation
1664 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1665 struct ext4_buddy *e4b,
1668 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1669 struct ext4_free_extent *bex = &ac->ac_b_ex;
1670 struct ext4_free_extent *gex = &ac->ac_g_ex;
1671 struct ext4_free_extent ex;
1674 if (ac->ac_status == AC_STATUS_FOUND)
1677 * We don't want to scan for a whole year
1679 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1680 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1681 ac->ac_status = AC_STATUS_BREAK;
1686 * Haven't found good chunk so far, let's continue
1688 if (bex->fe_len < gex->fe_len)
1691 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1692 && bex->fe_group == e4b->bd_group) {
1693 /* recheck chunk's availability - we don't know
1694 * when it was found (within this lock-unlock
1696 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1697 if (max >= gex->fe_len) {
1698 ext4_mb_use_best_found(ac, e4b);
1705 * The routine checks whether found extent is good enough. If it is,
1706 * then the extent gets marked used and flag is set to the context
1707 * to stop scanning. Otherwise, the extent is compared with the
1708 * previous found extent and if new one is better, then it's stored
1709 * in the context. Later, the best found extent will be used, if
1710 * mballoc can't find good enough extent.
1712 * FIXME: real allocation policy is to be designed yet!
1714 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1715 struct ext4_free_extent *ex,
1716 struct ext4_buddy *e4b)
1718 struct ext4_free_extent *bex = &ac->ac_b_ex;
1719 struct ext4_free_extent *gex = &ac->ac_g_ex;
1721 BUG_ON(ex->fe_len <= 0);
1722 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1723 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1724 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1729 * The special case - take what you catch first
1731 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1733 ext4_mb_use_best_found(ac, e4b);
1738 * Let's check whether the chuck is good enough
1740 if (ex->fe_len == gex->fe_len) {
1742 ext4_mb_use_best_found(ac, e4b);
1747 * If this is first found extent, just store it in the context
1749 if (bex->fe_len == 0) {
1755 * If new found extent is better, store it in the context
1757 if (bex->fe_len < gex->fe_len) {
1758 /* if the request isn't satisfied, any found extent
1759 * larger than previous best one is better */
1760 if (ex->fe_len > bex->fe_len)
1762 } else if (ex->fe_len > gex->fe_len) {
1763 /* if the request is satisfied, then we try to find
1764 * an extent that still satisfy the request, but is
1765 * smaller than previous one */
1766 if (ex->fe_len < bex->fe_len)
1770 ext4_mb_check_limits(ac, e4b, 0);
1773 static noinline_for_stack
1774 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1775 struct ext4_buddy *e4b)
1777 struct ext4_free_extent ex = ac->ac_b_ex;
1778 ext4_group_t group = ex.fe_group;
1782 BUG_ON(ex.fe_len <= 0);
1783 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1787 ext4_lock_group(ac->ac_sb, group);
1788 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1792 ext4_mb_use_best_found(ac, e4b);
1795 ext4_unlock_group(ac->ac_sb, group);
1796 ext4_mb_unload_buddy(e4b);
1801 static noinline_for_stack
1802 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1803 struct ext4_buddy *e4b)
1805 ext4_group_t group = ac->ac_g_ex.fe_group;
1808 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1809 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1810 struct ext4_free_extent ex;
1812 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1814 if (grp->bb_free == 0)
1817 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1821 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1822 ext4_mb_unload_buddy(e4b);
1826 ext4_lock_group(ac->ac_sb, group);
1827 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1828 ac->ac_g_ex.fe_len, &ex);
1829 ex.fe_logical = 0xDEADFA11; /* debug value */
1831 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1834 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1836 /* use do_div to get remainder (would be 64-bit modulo) */
1837 if (do_div(start, sbi->s_stripe) == 0) {
1840 ext4_mb_use_best_found(ac, e4b);
1842 } else if (max >= ac->ac_g_ex.fe_len) {
1843 BUG_ON(ex.fe_len <= 0);
1844 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1845 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1848 ext4_mb_use_best_found(ac, e4b);
1849 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1850 /* Sometimes, caller may want to merge even small
1851 * number of blocks to an existing extent */
1852 BUG_ON(ex.fe_len <= 0);
1853 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1854 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1857 ext4_mb_use_best_found(ac, e4b);
1859 ext4_unlock_group(ac->ac_sb, group);
1860 ext4_mb_unload_buddy(e4b);
1866 * The routine scans buddy structures (not bitmap!) from given order
1867 * to max order and tries to find big enough chunk to satisfy the req
1869 static noinline_for_stack
1870 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1871 struct ext4_buddy *e4b)
1873 struct super_block *sb = ac->ac_sb;
1874 struct ext4_group_info *grp = e4b->bd_info;
1880 BUG_ON(ac->ac_2order <= 0);
1881 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1882 if (grp->bb_counters[i] == 0)
1885 buddy = mb_find_buddy(e4b, i, &max);
1886 BUG_ON(buddy == NULL);
1888 k = mb_find_next_zero_bit(buddy, max, 0);
1893 ac->ac_b_ex.fe_len = 1 << i;
1894 ac->ac_b_ex.fe_start = k << i;
1895 ac->ac_b_ex.fe_group = e4b->bd_group;
1897 ext4_mb_use_best_found(ac, e4b);
1899 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1901 if (EXT4_SB(sb)->s_mb_stats)
1902 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1909 * The routine scans the group and measures all found extents.
1910 * In order to optimize scanning, caller must pass number of
1911 * free blocks in the group, so the routine can know upper limit.
1913 static noinline_for_stack
1914 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1915 struct ext4_buddy *e4b)
1917 struct super_block *sb = ac->ac_sb;
1918 void *bitmap = e4b->bd_bitmap;
1919 struct ext4_free_extent ex;
1923 free = e4b->bd_info->bb_free;
1926 i = e4b->bd_info->bb_first_free;
1928 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1929 i = mb_find_next_zero_bit(bitmap,
1930 EXT4_CLUSTERS_PER_GROUP(sb), i);
1931 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1933 * IF we have corrupt bitmap, we won't find any
1934 * free blocks even though group info says we
1935 * we have free blocks
1937 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1938 "%d free clusters as per "
1939 "group info. But bitmap says 0",
1944 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1945 BUG_ON(ex.fe_len <= 0);
1946 if (free < ex.fe_len) {
1947 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1948 "%d free clusters as per "
1949 "group info. But got %d blocks",
1952 * The number of free blocks differs. This mostly
1953 * indicate that the bitmap is corrupt. So exit
1954 * without claiming the space.
1958 ex.fe_logical = 0xDEADC0DE; /* debug value */
1959 ext4_mb_measure_extent(ac, &ex, e4b);
1965 ext4_mb_check_limits(ac, e4b, 1);
1969 * This is a special case for storages like raid5
1970 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1972 static noinline_for_stack
1973 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1974 struct ext4_buddy *e4b)
1976 struct super_block *sb = ac->ac_sb;
1977 struct ext4_sb_info *sbi = EXT4_SB(sb);
1978 void *bitmap = e4b->bd_bitmap;
1979 struct ext4_free_extent ex;
1980 ext4_fsblk_t first_group_block;
1985 BUG_ON(sbi->s_stripe == 0);
1987 /* find first stripe-aligned block in group */
1988 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1990 a = first_group_block + sbi->s_stripe - 1;
1991 do_div(a, sbi->s_stripe);
1992 i = (a * sbi->s_stripe) - first_group_block;
1994 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1995 if (!mb_test_bit(i, bitmap)) {
1996 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
1997 if (max >= sbi->s_stripe) {
1999 ex.fe_logical = 0xDEADF00D; /* debug value */
2001 ext4_mb_use_best_found(ac, e4b);
2009 /* This is now called BEFORE we load the buddy bitmap. */
2010 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2011 ext4_group_t group, int cr)
2013 unsigned free, fragments;
2014 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2015 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2017 BUG_ON(cr < 0 || cr >= 4);
2019 free = grp->bb_free;
2022 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2025 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2028 /* We only do this if the grp has never been initialized */
2029 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2030 int ret = ext4_mb_init_group(ac->ac_sb, group);
2035 fragments = grp->bb_fragments;
2041 BUG_ON(ac->ac_2order == 0);
2043 /* Avoid using the first bg of a flexgroup for data files */
2044 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2045 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2046 ((group % flex_size) == 0))
2049 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2050 (free / fragments) >= ac->ac_g_ex.fe_len)
2053 if (grp->bb_largest_free_order < ac->ac_2order)
2058 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2062 if (free >= ac->ac_g_ex.fe_len)
2074 static noinline_for_stack int
2075 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2077 ext4_group_t ngroups, group, i;
2080 struct ext4_sb_info *sbi;
2081 struct super_block *sb;
2082 struct ext4_buddy e4b;
2086 ngroups = ext4_get_groups_count(sb);
2087 /* non-extent files are limited to low blocks/groups */
2088 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2089 ngroups = sbi->s_blockfile_groups;
2091 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2093 /* first, try the goal */
2094 err = ext4_mb_find_by_goal(ac, &e4b);
2095 if (err || ac->ac_status == AC_STATUS_FOUND)
2098 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2102 * ac->ac2_order is set only if the fe_len is a power of 2
2103 * if ac2_order is set we also set criteria to 0 so that we
2104 * try exact allocation using buddy.
2106 i = fls(ac->ac_g_ex.fe_len);
2109 * We search using buddy data only if the order of the request
2110 * is greater than equal to the sbi_s_mb_order2_reqs
2111 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2113 if (i >= sbi->s_mb_order2_reqs) {
2115 * This should tell if fe_len is exactly power of 2
2117 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2118 ac->ac_2order = i - 1;
2121 /* if stream allocation is enabled, use global goal */
2122 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2123 /* TBD: may be hot point */
2124 spin_lock(&sbi->s_md_lock);
2125 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2126 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2127 spin_unlock(&sbi->s_md_lock);
2130 /* Let's just scan groups to find more-less suitable blocks */
2131 cr = ac->ac_2order ? 0 : 1;
2133 * cr == 0 try to get exact allocation,
2134 * cr == 3 try to get anything
2137 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2138 ac->ac_criteria = cr;
2140 * searching for the right group start
2141 * from the goal value specified
2143 group = ac->ac_g_ex.fe_group;
2145 for (i = 0; i < ngroups; group++, i++) {
2148 * Artificially restricted ngroups for non-extent
2149 * files makes group > ngroups possible on first loop.
2151 if (group >= ngroups)
2154 /* This now checks without needing the buddy page */
2155 if (!ext4_mb_good_group(ac, group, cr))
2158 err = ext4_mb_load_buddy(sb, group, &e4b);
2162 ext4_lock_group(sb, group);
2165 * We need to check again after locking the
2168 if (!ext4_mb_good_group(ac, group, cr)) {
2169 ext4_unlock_group(sb, group);
2170 ext4_mb_unload_buddy(&e4b);
2174 ac->ac_groups_scanned++;
2175 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2176 ext4_mb_simple_scan_group(ac, &e4b);
2177 else if (cr == 1 && sbi->s_stripe &&
2178 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2179 ext4_mb_scan_aligned(ac, &e4b);
2181 ext4_mb_complex_scan_group(ac, &e4b);
2183 ext4_unlock_group(sb, group);
2184 ext4_mb_unload_buddy(&e4b);
2186 if (ac->ac_status != AC_STATUS_CONTINUE)
2191 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2192 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2194 * We've been searching too long. Let's try to allocate
2195 * the best chunk we've found so far
2198 ext4_mb_try_best_found(ac, &e4b);
2199 if (ac->ac_status != AC_STATUS_FOUND) {
2201 * Someone more lucky has already allocated it.
2202 * The only thing we can do is just take first
2204 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2206 ac->ac_b_ex.fe_group = 0;
2207 ac->ac_b_ex.fe_start = 0;
2208 ac->ac_b_ex.fe_len = 0;
2209 ac->ac_status = AC_STATUS_CONTINUE;
2210 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2212 atomic_inc(&sbi->s_mb_lost_chunks);
2220 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2222 struct super_block *sb = seq->private;
2225 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2228 return (void *) ((unsigned long) group);
2231 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2233 struct super_block *sb = seq->private;
2237 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2240 return (void *) ((unsigned long) group);
2243 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2245 struct super_block *sb = seq->private;
2246 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2248 int err, buddy_loaded = 0;
2249 struct ext4_buddy e4b;
2250 struct ext4_group_info *grinfo;
2252 struct ext4_group_info info;
2253 ext4_grpblk_t counters[16];
2258 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2259 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2260 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2261 "group", "free", "frags", "first",
2262 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2263 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2265 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2266 sizeof(struct ext4_group_info);
2267 grinfo = ext4_get_group_info(sb, group);
2268 /* Load the group info in memory only if not already loaded. */
2269 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2270 err = ext4_mb_load_buddy(sb, group, &e4b);
2272 seq_printf(seq, "#%-5u: I/O error\n", group);
2278 memcpy(&sg, ext4_get_group_info(sb, group), i);
2281 ext4_mb_unload_buddy(&e4b);
2283 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2284 sg.info.bb_fragments, sg.info.bb_first_free);
2285 for (i = 0; i <= 13; i++)
2286 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2287 sg.info.bb_counters[i] : 0);
2288 seq_printf(seq, " ]\n");
2293 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2297 static const struct seq_operations ext4_mb_seq_groups_ops = {
2298 .start = ext4_mb_seq_groups_start,
2299 .next = ext4_mb_seq_groups_next,
2300 .stop = ext4_mb_seq_groups_stop,
2301 .show = ext4_mb_seq_groups_show,
2304 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2306 struct super_block *sb = PDE_DATA(inode);
2309 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2311 struct seq_file *m = file->private_data;
2318 static const struct file_operations ext4_mb_seq_groups_fops = {
2319 .owner = THIS_MODULE,
2320 .open = ext4_mb_seq_groups_open,
2322 .llseek = seq_lseek,
2323 .release = seq_release,
2326 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2328 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2329 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2336 * Allocate the top-level s_group_info array for the specified number
2339 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2341 struct ext4_sb_info *sbi = EXT4_SB(sb);
2343 struct ext4_group_info ***new_groupinfo;
2345 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2346 EXT4_DESC_PER_BLOCK_BITS(sb);
2347 if (size <= sbi->s_group_info_size)
2350 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2351 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2352 if (!new_groupinfo) {
2353 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2356 if (sbi->s_group_info) {
2357 memcpy(new_groupinfo, sbi->s_group_info,
2358 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2359 ext4_kvfree(sbi->s_group_info);
2361 sbi->s_group_info = new_groupinfo;
2362 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2363 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2364 sbi->s_group_info_size);
2368 /* Create and initialize ext4_group_info data for the given group. */
2369 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2370 struct ext4_group_desc *desc)
2374 struct ext4_sb_info *sbi = EXT4_SB(sb);
2375 struct ext4_group_info **meta_group_info;
2376 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2379 * First check if this group is the first of a reserved block.
2380 * If it's true, we have to allocate a new table of pointers
2381 * to ext4_group_info structures
2383 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2384 metalen = sizeof(*meta_group_info) <<
2385 EXT4_DESC_PER_BLOCK_BITS(sb);
2386 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2387 if (meta_group_info == NULL) {
2388 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2389 "for a buddy group");
2390 goto exit_meta_group_info;
2392 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2397 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2398 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2400 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
2401 if (meta_group_info[i] == NULL) {
2402 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2403 goto exit_group_info;
2405 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2406 &(meta_group_info[i]->bb_state));
2409 * initialize bb_free to be able to skip
2410 * empty groups without initialization
2412 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2413 meta_group_info[i]->bb_free =
2414 ext4_free_clusters_after_init(sb, group, desc);
2416 meta_group_info[i]->bb_free =
2417 ext4_free_group_clusters(sb, desc);
2420 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2421 init_rwsem(&meta_group_info[i]->alloc_sem);
2422 meta_group_info[i]->bb_free_root = RB_ROOT;
2423 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2427 struct buffer_head *bh;
2428 meta_group_info[i]->bb_bitmap =
2429 kmalloc(sb->s_blocksize, GFP_KERNEL);
2430 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2431 bh = ext4_read_block_bitmap(sb, group);
2433 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2442 /* If a meta_group_info table has been allocated, release it now */
2443 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2444 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2445 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2447 exit_meta_group_info:
2449 } /* ext4_mb_add_groupinfo */
2451 static int ext4_mb_init_backend(struct super_block *sb)
2453 ext4_group_t ngroups = ext4_get_groups_count(sb);
2455 struct ext4_sb_info *sbi = EXT4_SB(sb);
2457 struct ext4_group_desc *desc;
2458 struct kmem_cache *cachep;
2460 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2464 sbi->s_buddy_cache = new_inode(sb);
2465 if (sbi->s_buddy_cache == NULL) {
2466 ext4_msg(sb, KERN_ERR, "can't get new inode");
2469 /* To avoid potentially colliding with an valid on-disk inode number,
2470 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2471 * not in the inode hash, so it should never be found by iget(), but
2472 * this will avoid confusion if it ever shows up during debugging. */
2473 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2474 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2475 for (i = 0; i < ngroups; i++) {
2476 desc = ext4_get_group_desc(sb, i, NULL);
2478 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2481 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2488 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2490 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2491 i = sbi->s_group_info_size;
2493 kfree(sbi->s_group_info[i]);
2494 iput(sbi->s_buddy_cache);
2496 ext4_kvfree(sbi->s_group_info);
2500 static void ext4_groupinfo_destroy_slabs(void)
2504 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2505 if (ext4_groupinfo_caches[i])
2506 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2507 ext4_groupinfo_caches[i] = NULL;
2511 static int ext4_groupinfo_create_slab(size_t size)
2513 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2515 int blocksize_bits = order_base_2(size);
2516 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2517 struct kmem_cache *cachep;
2519 if (cache_index >= NR_GRPINFO_CACHES)
2522 if (unlikely(cache_index < 0))
2525 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2526 if (ext4_groupinfo_caches[cache_index]) {
2527 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2528 return 0; /* Already created */
2531 slab_size = offsetof(struct ext4_group_info,
2532 bb_counters[blocksize_bits + 2]);
2534 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2535 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2538 ext4_groupinfo_caches[cache_index] = cachep;
2540 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2543 "EXT4-fs: no memory for groupinfo slab cache\n");
2550 int ext4_mb_init(struct super_block *sb)
2552 struct ext4_sb_info *sbi = EXT4_SB(sb);
2558 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2560 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2561 if (sbi->s_mb_offsets == NULL) {
2566 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2567 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2568 if (sbi->s_mb_maxs == NULL) {
2573 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2577 /* order 0 is regular bitmap */
2578 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2579 sbi->s_mb_offsets[0] = 0;
2583 max = sb->s_blocksize << 2;
2585 sbi->s_mb_offsets[i] = offset;
2586 sbi->s_mb_maxs[i] = max;
2587 offset += 1 << (sb->s_blocksize_bits - i);
2590 } while (i <= sb->s_blocksize_bits + 1);
2592 spin_lock_init(&sbi->s_md_lock);
2593 spin_lock_init(&sbi->s_bal_lock);
2595 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2596 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2597 sbi->s_mb_stats = MB_DEFAULT_STATS;
2598 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2599 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2601 * The default group preallocation is 512, which for 4k block
2602 * sizes translates to 2 megabytes. However for bigalloc file
2603 * systems, this is probably too big (i.e, if the cluster size
2604 * is 1 megabyte, then group preallocation size becomes half a
2605 * gigabyte!). As a default, we will keep a two megabyte
2606 * group pralloc size for cluster sizes up to 64k, and after
2607 * that, we will force a minimum group preallocation size of
2608 * 32 clusters. This translates to 8 megs when the cluster
2609 * size is 256k, and 32 megs when the cluster size is 1 meg,
2610 * which seems reasonable as a default.
2612 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2613 sbi->s_cluster_bits, 32);
2615 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2616 * to the lowest multiple of s_stripe which is bigger than
2617 * the s_mb_group_prealloc as determined above. We want
2618 * the preallocation size to be an exact multiple of the
2619 * RAID stripe size so that preallocations don't fragment
2622 if (sbi->s_stripe > 1) {
2623 sbi->s_mb_group_prealloc = roundup(
2624 sbi->s_mb_group_prealloc, sbi->s_stripe);
2627 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2628 if (sbi->s_locality_groups == NULL) {
2632 for_each_possible_cpu(i) {
2633 struct ext4_locality_group *lg;
2634 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2635 mutex_init(&lg->lg_mutex);
2636 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2637 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2638 spin_lock_init(&lg->lg_prealloc_lock);
2641 /* init file for buddy data */
2642 ret = ext4_mb_init_backend(sb);
2644 goto out_free_locality_groups;
2647 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2648 &ext4_mb_seq_groups_fops, sb);
2652 out_free_locality_groups:
2653 free_percpu(sbi->s_locality_groups);
2654 sbi->s_locality_groups = NULL;
2656 kfree(sbi->s_mb_offsets);
2657 sbi->s_mb_offsets = NULL;
2658 kfree(sbi->s_mb_maxs);
2659 sbi->s_mb_maxs = NULL;
2663 /* need to called with the ext4 group lock held */
2664 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2666 struct ext4_prealloc_space *pa;
2667 struct list_head *cur, *tmp;
2670 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2671 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2672 list_del(&pa->pa_group_list);
2674 kmem_cache_free(ext4_pspace_cachep, pa);
2677 mb_debug(1, "mballoc: %u PAs left\n", count);
2681 int ext4_mb_release(struct super_block *sb)
2683 ext4_group_t ngroups = ext4_get_groups_count(sb);
2685 int num_meta_group_infos;
2686 struct ext4_group_info *grinfo;
2687 struct ext4_sb_info *sbi = EXT4_SB(sb);
2688 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2691 remove_proc_entry("mb_groups", sbi->s_proc);
2693 if (sbi->s_group_info) {
2694 for (i = 0; i < ngroups; i++) {
2695 grinfo = ext4_get_group_info(sb, i);
2697 kfree(grinfo->bb_bitmap);
2699 ext4_lock_group(sb, i);
2700 ext4_mb_cleanup_pa(grinfo);
2701 ext4_unlock_group(sb, i);
2702 kmem_cache_free(cachep, grinfo);
2704 num_meta_group_infos = (ngroups +
2705 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2706 EXT4_DESC_PER_BLOCK_BITS(sb);
2707 for (i = 0; i < num_meta_group_infos; i++)
2708 kfree(sbi->s_group_info[i]);
2709 ext4_kvfree(sbi->s_group_info);
2711 kfree(sbi->s_mb_offsets);
2712 kfree(sbi->s_mb_maxs);
2713 if (sbi->s_buddy_cache)
2714 iput(sbi->s_buddy_cache);
2715 if (sbi->s_mb_stats) {
2716 ext4_msg(sb, KERN_INFO,
2717 "mballoc: %u blocks %u reqs (%u success)",
2718 atomic_read(&sbi->s_bal_allocated),
2719 atomic_read(&sbi->s_bal_reqs),
2720 atomic_read(&sbi->s_bal_success));
2721 ext4_msg(sb, KERN_INFO,
2722 "mballoc: %u extents scanned, %u goal hits, "
2723 "%u 2^N hits, %u breaks, %u lost",
2724 atomic_read(&sbi->s_bal_ex_scanned),
2725 atomic_read(&sbi->s_bal_goals),
2726 atomic_read(&sbi->s_bal_2orders),
2727 atomic_read(&sbi->s_bal_breaks),
2728 atomic_read(&sbi->s_mb_lost_chunks));
2729 ext4_msg(sb, KERN_INFO,
2730 "mballoc: %lu generated and it took %Lu",
2731 sbi->s_mb_buddies_generated,
2732 sbi->s_mb_generation_time);
2733 ext4_msg(sb, KERN_INFO,
2734 "mballoc: %u preallocated, %u discarded",
2735 atomic_read(&sbi->s_mb_preallocated),
2736 atomic_read(&sbi->s_mb_discarded));
2739 free_percpu(sbi->s_locality_groups);
2744 static inline int ext4_issue_discard(struct super_block *sb,
2745 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2747 ext4_fsblk_t discard_block;
2749 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2750 ext4_group_first_block_no(sb, block_group));
2751 count = EXT4_C2B(EXT4_SB(sb), count);
2752 trace_ext4_discard_blocks(sb,
2753 (unsigned long long) discard_block, count);
2754 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2758 * This function is called by the jbd2 layer once the commit has finished,
2759 * so we know we can free the blocks that were released with that commit.
2761 static void ext4_free_data_callback(struct super_block *sb,
2762 struct ext4_journal_cb_entry *jce,
2765 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2766 struct ext4_buddy e4b;
2767 struct ext4_group_info *db;
2768 int err, count = 0, count2 = 0;
2770 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2771 entry->efd_count, entry->efd_group, entry);
2773 if (test_opt(sb, DISCARD)) {
2774 err = ext4_issue_discard(sb, entry->efd_group,
2775 entry->efd_start_cluster,
2777 if (err && err != -EOPNOTSUPP)
2778 ext4_msg(sb, KERN_WARNING, "discard request in"
2779 " group:%d block:%d count:%d failed"
2780 " with %d", entry->efd_group,
2781 entry->efd_start_cluster,
2782 entry->efd_count, err);
2785 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2786 /* we expect to find existing buddy because it's pinned */
2791 /* there are blocks to put in buddy to make them really free */
2792 count += entry->efd_count;
2794 ext4_lock_group(sb, entry->efd_group);
2795 /* Take it out of per group rb tree */
2796 rb_erase(&entry->efd_node, &(db->bb_free_root));
2797 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2800 * Clear the trimmed flag for the group so that the next
2801 * ext4_trim_fs can trim it.
2802 * If the volume is mounted with -o discard, online discard
2803 * is supported and the free blocks will be trimmed online.
2805 if (!test_opt(sb, DISCARD))
2806 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2808 if (!db->bb_free_root.rb_node) {
2809 /* No more items in the per group rb tree
2810 * balance refcounts from ext4_mb_free_metadata()
2812 page_cache_release(e4b.bd_buddy_page);
2813 page_cache_release(e4b.bd_bitmap_page);
2815 ext4_unlock_group(sb, entry->efd_group);
2816 kmem_cache_free(ext4_free_data_cachep, entry);
2817 ext4_mb_unload_buddy(&e4b);
2819 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2822 int __init ext4_init_mballoc(void)
2824 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2825 SLAB_RECLAIM_ACCOUNT);
2826 if (ext4_pspace_cachep == NULL)
2829 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2830 SLAB_RECLAIM_ACCOUNT);
2831 if (ext4_ac_cachep == NULL) {
2832 kmem_cache_destroy(ext4_pspace_cachep);
2836 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2837 SLAB_RECLAIM_ACCOUNT);
2838 if (ext4_free_data_cachep == NULL) {
2839 kmem_cache_destroy(ext4_pspace_cachep);
2840 kmem_cache_destroy(ext4_ac_cachep);
2846 void ext4_exit_mballoc(void)
2849 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2850 * before destroying the slab cache.
2853 kmem_cache_destroy(ext4_pspace_cachep);
2854 kmem_cache_destroy(ext4_ac_cachep);
2855 kmem_cache_destroy(ext4_free_data_cachep);
2856 ext4_groupinfo_destroy_slabs();
2861 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2862 * Returns 0 if success or error code
2864 static noinline_for_stack int
2865 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2866 handle_t *handle, unsigned int reserv_clstrs)
2868 struct buffer_head *bitmap_bh = NULL;
2869 struct ext4_group_desc *gdp;
2870 struct buffer_head *gdp_bh;
2871 struct ext4_sb_info *sbi;
2872 struct super_block *sb;
2876 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2877 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2883 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2887 BUFFER_TRACE(bitmap_bh, "getting write access");
2888 err = ext4_journal_get_write_access(handle, bitmap_bh);
2893 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2897 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2898 ext4_free_group_clusters(sb, gdp));
2900 BUFFER_TRACE(gdp_bh, "get_write_access");
2901 err = ext4_journal_get_write_access(handle, gdp_bh);
2905 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2907 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2908 if (!ext4_data_block_valid(sbi, block, len)) {
2909 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2910 "fs metadata", block, block+len);
2911 /* File system mounted not to panic on error
2912 * Fix the bitmap and repeat the block allocation
2913 * We leak some of the blocks here.
2915 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2916 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2917 ac->ac_b_ex.fe_len);
2918 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2919 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2925 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2926 #ifdef AGGRESSIVE_CHECK
2929 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2930 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2931 bitmap_bh->b_data));
2935 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2936 ac->ac_b_ex.fe_len);
2937 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2938 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2939 ext4_free_group_clusters_set(sb, gdp,
2940 ext4_free_clusters_after_init(sb,
2941 ac->ac_b_ex.fe_group, gdp));
2943 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2944 ext4_free_group_clusters_set(sb, gdp, len);
2945 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2946 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2948 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2949 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2951 * Now reduce the dirty block count also. Should not go negative
2953 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2954 /* release all the reserved blocks if non delalloc */
2955 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2958 if (sbi->s_log_groups_per_flex) {
2959 ext4_group_t flex_group = ext4_flex_group(sbi,
2960 ac->ac_b_ex.fe_group);
2961 atomic64_sub(ac->ac_b_ex.fe_len,
2962 &sbi->s_flex_groups[flex_group].free_clusters);
2965 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2968 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2976 * here we normalize request for locality group
2977 * Group request are normalized to s_mb_group_prealloc, which goes to
2978 * s_strip if we set the same via mount option.
2979 * s_mb_group_prealloc can be configured via
2980 * /sys/fs/ext4/<partition>/mb_group_prealloc
2982 * XXX: should we try to preallocate more than the group has now?
2984 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2986 struct super_block *sb = ac->ac_sb;
2987 struct ext4_locality_group *lg = ac->ac_lg;
2990 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2991 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2992 current->pid, ac->ac_g_ex.fe_len);
2996 * Normalization means making request better in terms of
2997 * size and alignment
2999 static noinline_for_stack void
3000 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3001 struct ext4_allocation_request *ar)
3003 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3006 loff_t size, start_off;
3007 loff_t orig_size __maybe_unused;
3009 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3010 struct ext4_prealloc_space *pa;
3012 /* do normalize only data requests, metadata requests
3013 do not need preallocation */
3014 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3017 /* sometime caller may want exact blocks */
3018 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3021 /* caller may indicate that preallocation isn't
3022 * required (it's a tail, for example) */
3023 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3026 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3027 ext4_mb_normalize_group_request(ac);
3031 bsbits = ac->ac_sb->s_blocksize_bits;
3033 /* first, let's learn actual file size
3034 * given current request is allocated */
3035 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3036 size = size << bsbits;
3037 if (size < i_size_read(ac->ac_inode))
3038 size = i_size_read(ac->ac_inode);
3041 /* max size of free chunks */
3044 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3045 (req <= (size) || max <= (chunk_size))
3047 /* first, try to predict filesize */
3048 /* XXX: should this table be tunable? */
3050 if (size <= 16 * 1024) {
3052 } else if (size <= 32 * 1024) {
3054 } else if (size <= 64 * 1024) {
3056 } else if (size <= 128 * 1024) {
3058 } else if (size <= 256 * 1024) {
3060 } else if (size <= 512 * 1024) {
3062 } else if (size <= 1024 * 1024) {
3064 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3065 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3066 (21 - bsbits)) << 21;
3067 size = 2 * 1024 * 1024;
3068 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3069 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3070 (22 - bsbits)) << 22;
3071 size = 4 * 1024 * 1024;
3072 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3073 (8<<20)>>bsbits, max, 8 * 1024)) {
3074 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3075 (23 - bsbits)) << 23;
3076 size = 8 * 1024 * 1024;
3078 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3079 size = ac->ac_o_ex.fe_len << bsbits;
3081 size = size >> bsbits;
3082 start = start_off >> bsbits;
3084 /* don't cover already allocated blocks in selected range */
3085 if (ar->pleft && start <= ar->lleft) {
3086 size -= ar->lleft + 1 - start;
3087 start = ar->lleft + 1;
3089 if (ar->pright && start + size - 1 >= ar->lright)
3090 size -= start + size - ar->lright;
3094 /* check we don't cross already preallocated blocks */
3096 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3101 spin_lock(&pa->pa_lock);
3102 if (pa->pa_deleted) {
3103 spin_unlock(&pa->pa_lock);
3107 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3110 /* PA must not overlap original request */
3111 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3112 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3114 /* skip PAs this normalized request doesn't overlap with */
3115 if (pa->pa_lstart >= end || pa_end <= start) {
3116 spin_unlock(&pa->pa_lock);
3119 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3121 /* adjust start or end to be adjacent to this pa */
3122 if (pa_end <= ac->ac_o_ex.fe_logical) {
3123 BUG_ON(pa_end < start);
3125 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3126 BUG_ON(pa->pa_lstart > end);
3127 end = pa->pa_lstart;
3129 spin_unlock(&pa->pa_lock);
3134 /* XXX: extra loop to check we really don't overlap preallocations */
3136 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3139 spin_lock(&pa->pa_lock);
3140 if (pa->pa_deleted == 0) {
3141 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3143 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3145 spin_unlock(&pa->pa_lock);
3149 if (start + size <= ac->ac_o_ex.fe_logical &&
3150 start > ac->ac_o_ex.fe_logical) {
3151 ext4_msg(ac->ac_sb, KERN_ERR,
3152 "start %lu, size %lu, fe_logical %lu",
3153 (unsigned long) start, (unsigned long) size,
3154 (unsigned long) ac->ac_o_ex.fe_logical);
3156 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3157 start > ac->ac_o_ex.fe_logical);
3158 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3160 /* now prepare goal request */
3162 /* XXX: is it better to align blocks WRT to logical
3163 * placement or satisfy big request as is */
3164 ac->ac_g_ex.fe_logical = start;
3165 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3167 /* define goal start in order to merge */
3168 if (ar->pright && (ar->lright == (start + size))) {
3169 /* merge to the right */
3170 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3171 &ac->ac_f_ex.fe_group,
3172 &ac->ac_f_ex.fe_start);
3173 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3175 if (ar->pleft && (ar->lleft + 1 == start)) {
3176 /* merge to the left */
3177 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3178 &ac->ac_f_ex.fe_group,
3179 &ac->ac_f_ex.fe_start);
3180 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3183 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3184 (unsigned) orig_size, (unsigned) start);
3187 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3189 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3191 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3192 atomic_inc(&sbi->s_bal_reqs);
3193 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3194 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3195 atomic_inc(&sbi->s_bal_success);
3196 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3197 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3198 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3199 atomic_inc(&sbi->s_bal_goals);
3200 if (ac->ac_found > sbi->s_mb_max_to_scan)
3201 atomic_inc(&sbi->s_bal_breaks);
3204 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3205 trace_ext4_mballoc_alloc(ac);
3207 trace_ext4_mballoc_prealloc(ac);
3211 * Called on failure; free up any blocks from the inode PA for this
3212 * context. We don't need this for MB_GROUP_PA because we only change
3213 * pa_free in ext4_mb_release_context(), but on failure, we've already
3214 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3216 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3218 struct ext4_prealloc_space *pa = ac->ac_pa;
3220 if (pa && pa->pa_type == MB_INODE_PA)
3221 pa->pa_free += ac->ac_b_ex.fe_len;
3225 * use blocks preallocated to inode
3227 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3228 struct ext4_prealloc_space *pa)
3230 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3235 /* found preallocated blocks, use them */
3236 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3237 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3238 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3239 len = EXT4_NUM_B2C(sbi, end - start);
3240 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3241 &ac->ac_b_ex.fe_start);
3242 ac->ac_b_ex.fe_len = len;
3243 ac->ac_status = AC_STATUS_FOUND;
3246 BUG_ON(start < pa->pa_pstart);
3247 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3248 BUG_ON(pa->pa_free < len);
3251 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3255 * use blocks preallocated to locality group
3257 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3258 struct ext4_prealloc_space *pa)
3260 unsigned int len = ac->ac_o_ex.fe_len;
3262 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3263 &ac->ac_b_ex.fe_group,
3264 &ac->ac_b_ex.fe_start);
3265 ac->ac_b_ex.fe_len = len;
3266 ac->ac_status = AC_STATUS_FOUND;
3269 /* we don't correct pa_pstart or pa_plen here to avoid
3270 * possible race when the group is being loaded concurrently
3271 * instead we correct pa later, after blocks are marked
3272 * in on-disk bitmap -- see ext4_mb_release_context()
3273 * Other CPUs are prevented from allocating from this pa by lg_mutex
3275 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3279 * Return the prealloc space that have minimal distance
3280 * from the goal block. @cpa is the prealloc
3281 * space that is having currently known minimal distance
3282 * from the goal block.
3284 static struct ext4_prealloc_space *
3285 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3286 struct ext4_prealloc_space *pa,
3287 struct ext4_prealloc_space *cpa)
3289 ext4_fsblk_t cur_distance, new_distance;
3292 atomic_inc(&pa->pa_count);
3295 cur_distance = abs(goal_block - cpa->pa_pstart);
3296 new_distance = abs(goal_block - pa->pa_pstart);
3298 if (cur_distance <= new_distance)
3301 /* drop the previous reference */
3302 atomic_dec(&cpa->pa_count);
3303 atomic_inc(&pa->pa_count);
3308 * search goal blocks in preallocated space
3310 static noinline_for_stack int
3311 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3313 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3315 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3316 struct ext4_locality_group *lg;
3317 struct ext4_prealloc_space *pa, *cpa = NULL;
3318 ext4_fsblk_t goal_block;
3320 /* only data can be preallocated */
3321 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3324 /* first, try per-file preallocation */
3326 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3328 /* all fields in this condition don't change,
3329 * so we can skip locking for them */
3330 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3331 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3332 EXT4_C2B(sbi, pa->pa_len)))
3335 /* non-extent files can't have physical blocks past 2^32 */
3336 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3337 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3338 EXT4_MAX_BLOCK_FILE_PHYS))
3341 /* found preallocated blocks, use them */
3342 spin_lock(&pa->pa_lock);
3343 if (pa->pa_deleted == 0 && pa->pa_free) {
3344 atomic_inc(&pa->pa_count);
3345 ext4_mb_use_inode_pa(ac, pa);
3346 spin_unlock(&pa->pa_lock);
3347 ac->ac_criteria = 10;
3351 spin_unlock(&pa->pa_lock);
3355 /* can we use group allocation? */
3356 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3359 /* inode may have no locality group for some reason */
3363 order = fls(ac->ac_o_ex.fe_len) - 1;
3364 if (order > PREALLOC_TB_SIZE - 1)
3365 /* The max size of hash table is PREALLOC_TB_SIZE */
3366 order = PREALLOC_TB_SIZE - 1;
3368 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3370 * search for the prealloc space that is having
3371 * minimal distance from the goal block.
3373 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3375 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3377 spin_lock(&pa->pa_lock);
3378 if (pa->pa_deleted == 0 &&
3379 pa->pa_free >= ac->ac_o_ex.fe_len) {
3381 cpa = ext4_mb_check_group_pa(goal_block,
3384 spin_unlock(&pa->pa_lock);
3389 ext4_mb_use_group_pa(ac, cpa);
3390 ac->ac_criteria = 20;
3397 * the function goes through all block freed in the group
3398 * but not yet committed and marks them used in in-core bitmap.
3399 * buddy must be generated from this bitmap
3400 * Need to be called with the ext4 group lock held
3402 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3406 struct ext4_group_info *grp;
3407 struct ext4_free_data *entry;
3409 grp = ext4_get_group_info(sb, group);
3410 n = rb_first(&(grp->bb_free_root));
3413 entry = rb_entry(n, struct ext4_free_data, efd_node);
3414 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3421 * the function goes through all preallocation in this group and marks them
3422 * used in in-core bitmap. buddy must be generated from this bitmap
3423 * Need to be called with ext4 group lock held
3425 static noinline_for_stack
3426 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3429 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3430 struct ext4_prealloc_space *pa;
3431 struct list_head *cur;
3432 ext4_group_t groupnr;
3433 ext4_grpblk_t start;
3434 int preallocated = 0;
3437 /* all form of preallocation discards first load group,
3438 * so the only competing code is preallocation use.
3439 * we don't need any locking here
3440 * notice we do NOT ignore preallocations with pa_deleted
3441 * otherwise we could leave used blocks available for
3442 * allocation in buddy when concurrent ext4_mb_put_pa()
3443 * is dropping preallocation
3445 list_for_each(cur, &grp->bb_prealloc_list) {
3446 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3447 spin_lock(&pa->pa_lock);
3448 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3451 spin_unlock(&pa->pa_lock);
3452 if (unlikely(len == 0))
3454 BUG_ON(groupnr != group);
3455 ext4_set_bits(bitmap, start, len);
3456 preallocated += len;
3458 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3461 static void ext4_mb_pa_callback(struct rcu_head *head)
3463 struct ext4_prealloc_space *pa;
3464 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3466 BUG_ON(atomic_read(&pa->pa_count));
3467 BUG_ON(pa->pa_deleted == 0);
3468 kmem_cache_free(ext4_pspace_cachep, pa);
3472 * drops a reference to preallocated space descriptor
3473 * if this was the last reference and the space is consumed
3475 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3476 struct super_block *sb, struct ext4_prealloc_space *pa)
3479 ext4_fsblk_t grp_blk;
3481 /* in this short window concurrent discard can set pa_deleted */
3482 spin_lock(&pa->pa_lock);
3483 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3484 spin_unlock(&pa->pa_lock);
3488 if (pa->pa_deleted == 1) {
3489 spin_unlock(&pa->pa_lock);
3494 spin_unlock(&pa->pa_lock);
3496 grp_blk = pa->pa_pstart;
3498 * If doing group-based preallocation, pa_pstart may be in the
3499 * next group when pa is used up
3501 if (pa->pa_type == MB_GROUP_PA)
3504 grp = ext4_get_group_number(sb, grp_blk);
3509 * P1 (buddy init) P2 (regular allocation)
3510 * find block B in PA
3511 * copy on-disk bitmap to buddy
3512 * mark B in on-disk bitmap
3513 * drop PA from group
3514 * mark all PAs in buddy
3516 * thus, P1 initializes buddy with B available. to prevent this
3517 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3520 ext4_lock_group(sb, grp);
3521 list_del(&pa->pa_group_list);
3522 ext4_unlock_group(sb, grp);
3524 spin_lock(pa->pa_obj_lock);
3525 list_del_rcu(&pa->pa_inode_list);
3526 spin_unlock(pa->pa_obj_lock);
3528 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3532 * creates new preallocated space for given inode
3534 static noinline_for_stack int
3535 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3537 struct super_block *sb = ac->ac_sb;
3538 struct ext4_sb_info *sbi = EXT4_SB(sb);
3539 struct ext4_prealloc_space *pa;
3540 struct ext4_group_info *grp;
3541 struct ext4_inode_info *ei;
3543 /* preallocate only when found space is larger then requested */
3544 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3545 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3546 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3548 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3552 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3558 /* we can't allocate as much as normalizer wants.
3559 * so, found space must get proper lstart
3560 * to cover original request */
3561 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3562 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3564 /* we're limited by original request in that
3565 * logical block must be covered any way
3566 * winl is window we can move our chunk within */
3567 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3569 /* also, we should cover whole original request */
3570 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3572 /* the smallest one defines real window */
3573 win = min(winl, wins);
3575 offs = ac->ac_o_ex.fe_logical %
3576 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3577 if (offs && offs < win)
3580 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3581 EXT4_NUM_B2C(sbi, win);
3582 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3583 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3586 /* preallocation can change ac_b_ex, thus we store actually
3587 * allocated blocks for history */
3588 ac->ac_f_ex = ac->ac_b_ex;
3590 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3591 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3592 pa->pa_len = ac->ac_b_ex.fe_len;
3593 pa->pa_free = pa->pa_len;
3594 atomic_set(&pa->pa_count, 1);
3595 spin_lock_init(&pa->pa_lock);
3596 INIT_LIST_HEAD(&pa->pa_inode_list);
3597 INIT_LIST_HEAD(&pa->pa_group_list);
3599 pa->pa_type = MB_INODE_PA;
3601 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3602 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3603 trace_ext4_mb_new_inode_pa(ac, pa);
3605 ext4_mb_use_inode_pa(ac, pa);
3606 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3608 ei = EXT4_I(ac->ac_inode);
3609 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3611 pa->pa_obj_lock = &ei->i_prealloc_lock;
3612 pa->pa_inode = ac->ac_inode;
3614 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3615 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3616 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3618 spin_lock(pa->pa_obj_lock);
3619 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3620 spin_unlock(pa->pa_obj_lock);
3626 * creates new preallocated space for locality group inodes belongs to
3628 static noinline_for_stack int
3629 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3631 struct super_block *sb = ac->ac_sb;
3632 struct ext4_locality_group *lg;
3633 struct ext4_prealloc_space *pa;
3634 struct ext4_group_info *grp;
3636 /* preallocate only when found space is larger then requested */
3637 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3638 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3639 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3641 BUG_ON(ext4_pspace_cachep == NULL);
3642 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3646 /* preallocation can change ac_b_ex, thus we store actually
3647 * allocated blocks for history */
3648 ac->ac_f_ex = ac->ac_b_ex;
3650 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3651 pa->pa_lstart = pa->pa_pstart;
3652 pa->pa_len = ac->ac_b_ex.fe_len;
3653 pa->pa_free = pa->pa_len;
3654 atomic_set(&pa->pa_count, 1);
3655 spin_lock_init(&pa->pa_lock);
3656 INIT_LIST_HEAD(&pa->pa_inode_list);
3657 INIT_LIST_HEAD(&pa->pa_group_list);
3659 pa->pa_type = MB_GROUP_PA;
3661 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3662 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3663 trace_ext4_mb_new_group_pa(ac, pa);
3665 ext4_mb_use_group_pa(ac, pa);
3666 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3668 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3672 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3673 pa->pa_inode = NULL;
3675 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3676 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3677 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3680 * We will later add the new pa to the right bucket
3681 * after updating the pa_free in ext4_mb_release_context
3686 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3690 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3691 err = ext4_mb_new_group_pa(ac);
3693 err = ext4_mb_new_inode_pa(ac);
3698 * finds all unused blocks in on-disk bitmap, frees them in
3699 * in-core bitmap and buddy.
3700 * @pa must be unlinked from inode and group lists, so that
3701 * nobody else can find/use it.
3702 * the caller MUST hold group/inode locks.
3703 * TODO: optimize the case when there are no in-core structures yet
3705 static noinline_for_stack int
3706 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3707 struct ext4_prealloc_space *pa)
3709 struct super_block *sb = e4b->bd_sb;
3710 struct ext4_sb_info *sbi = EXT4_SB(sb);
3715 unsigned long long grp_blk_start;
3719 BUG_ON(pa->pa_deleted == 0);
3720 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3721 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3722 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3723 end = bit + pa->pa_len;
3726 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3729 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3730 mb_debug(1, " free preallocated %u/%u in group %u\n",
3731 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3732 (unsigned) next - bit, (unsigned) group);
3735 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3736 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3737 EXT4_C2B(sbi, bit)),
3739 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3742 if (free != pa->pa_free) {
3743 ext4_msg(e4b->bd_sb, KERN_CRIT,
3744 "pa %p: logic %lu, phys. %lu, len %lu",
3745 pa, (unsigned long) pa->pa_lstart,
3746 (unsigned long) pa->pa_pstart,
3747 (unsigned long) pa->pa_len);
3748 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3751 * pa is already deleted so we use the value obtained
3752 * from the bitmap and continue.
3755 atomic_add(free, &sbi->s_mb_discarded);
3760 static noinline_for_stack int
3761 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3762 struct ext4_prealloc_space *pa)
3764 struct super_block *sb = e4b->bd_sb;
3768 trace_ext4_mb_release_group_pa(sb, pa);
3769 BUG_ON(pa->pa_deleted == 0);
3770 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3771 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3772 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3773 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3774 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3780 * releases all preallocations in given group
3782 * first, we need to decide discard policy:
3783 * - when do we discard
3785 * - how many do we discard
3786 * 1) how many requested
3788 static noinline_for_stack int
3789 ext4_mb_discard_group_preallocations(struct super_block *sb,
3790 ext4_group_t group, int needed)
3792 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3793 struct buffer_head *bitmap_bh = NULL;
3794 struct ext4_prealloc_space *pa, *tmp;
3795 struct list_head list;
3796 struct ext4_buddy e4b;
3801 mb_debug(1, "discard preallocation for group %u\n", group);
3803 if (list_empty(&grp->bb_prealloc_list))
3806 bitmap_bh = ext4_read_block_bitmap(sb, group);
3807 if (bitmap_bh == NULL) {
3808 ext4_error(sb, "Error reading block bitmap for %u", group);
3812 err = ext4_mb_load_buddy(sb, group, &e4b);
3814 ext4_error(sb, "Error loading buddy information for %u", group);
3820 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3822 INIT_LIST_HEAD(&list);
3824 ext4_lock_group(sb, group);
3825 list_for_each_entry_safe(pa, tmp,
3826 &grp->bb_prealloc_list, pa_group_list) {
3827 spin_lock(&pa->pa_lock);
3828 if (atomic_read(&pa->pa_count)) {
3829 spin_unlock(&pa->pa_lock);
3833 if (pa->pa_deleted) {
3834 spin_unlock(&pa->pa_lock);
3838 /* seems this one can be freed ... */
3841 /* we can trust pa_free ... */
3842 free += pa->pa_free;
3844 spin_unlock(&pa->pa_lock);
3846 list_del(&pa->pa_group_list);
3847 list_add(&pa->u.pa_tmp_list, &list);
3850 /* if we still need more blocks and some PAs were used, try again */
3851 if (free < needed && busy) {
3853 ext4_unlock_group(sb, group);
3858 /* found anything to free? */
3859 if (list_empty(&list)) {
3864 /* now free all selected PAs */
3865 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3867 /* remove from object (inode or locality group) */
3868 spin_lock(pa->pa_obj_lock);
3869 list_del_rcu(&pa->pa_inode_list);
3870 spin_unlock(pa->pa_obj_lock);
3872 if (pa->pa_type == MB_GROUP_PA)
3873 ext4_mb_release_group_pa(&e4b, pa);
3875 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3877 list_del(&pa->u.pa_tmp_list);
3878 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3882 ext4_unlock_group(sb, group);
3883 ext4_mb_unload_buddy(&e4b);
3889 * releases all non-used preallocated blocks for given inode
3891 * It's important to discard preallocations under i_data_sem
3892 * We don't want another block to be served from the prealloc
3893 * space when we are discarding the inode prealloc space.
3895 * FIXME!! Make sure it is valid at all the call sites
3897 void ext4_discard_preallocations(struct inode *inode)
3899 struct ext4_inode_info *ei = EXT4_I(inode);
3900 struct super_block *sb = inode->i_sb;
3901 struct buffer_head *bitmap_bh = NULL;
3902 struct ext4_prealloc_space *pa, *tmp;
3903 ext4_group_t group = 0;
3904 struct list_head list;
3905 struct ext4_buddy e4b;
3908 if (!S_ISREG(inode->i_mode)) {
3909 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3913 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3914 trace_ext4_discard_preallocations(inode);
3916 INIT_LIST_HEAD(&list);
3919 /* first, collect all pa's in the inode */
3920 spin_lock(&ei->i_prealloc_lock);
3921 while (!list_empty(&ei->i_prealloc_list)) {
3922 pa = list_entry(ei->i_prealloc_list.next,
3923 struct ext4_prealloc_space, pa_inode_list);
3924 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3925 spin_lock(&pa->pa_lock);
3926 if (atomic_read(&pa->pa_count)) {
3927 /* this shouldn't happen often - nobody should
3928 * use preallocation while we're discarding it */
3929 spin_unlock(&pa->pa_lock);
3930 spin_unlock(&ei->i_prealloc_lock);
3931 ext4_msg(sb, KERN_ERR,
3932 "uh-oh! used pa while discarding");
3934 schedule_timeout_uninterruptible(HZ);
3938 if (pa->pa_deleted == 0) {
3940 spin_unlock(&pa->pa_lock);
3941 list_del_rcu(&pa->pa_inode_list);
3942 list_add(&pa->u.pa_tmp_list, &list);
3946 /* someone is deleting pa right now */
3947 spin_unlock(&pa->pa_lock);
3948 spin_unlock(&ei->i_prealloc_lock);
3950 /* we have to wait here because pa_deleted
3951 * doesn't mean pa is already unlinked from
3952 * the list. as we might be called from
3953 * ->clear_inode() the inode will get freed
3954 * and concurrent thread which is unlinking
3955 * pa from inode's list may access already
3956 * freed memory, bad-bad-bad */
3958 /* XXX: if this happens too often, we can
3959 * add a flag to force wait only in case
3960 * of ->clear_inode(), but not in case of
3961 * regular truncate */
3962 schedule_timeout_uninterruptible(HZ);
3965 spin_unlock(&ei->i_prealloc_lock);
3967 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3968 BUG_ON(pa->pa_type != MB_INODE_PA);
3969 group = ext4_get_group_number(sb, pa->pa_pstart);
3971 err = ext4_mb_load_buddy(sb, group, &e4b);
3973 ext4_error(sb, "Error loading buddy information for %u",
3978 bitmap_bh = ext4_read_block_bitmap(sb, group);
3979 if (bitmap_bh == NULL) {
3980 ext4_error(sb, "Error reading block bitmap for %u",
3982 ext4_mb_unload_buddy(&e4b);
3986 ext4_lock_group(sb, group);
3987 list_del(&pa->pa_group_list);
3988 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3989 ext4_unlock_group(sb, group);
3991 ext4_mb_unload_buddy(&e4b);
3994 list_del(&pa->u.pa_tmp_list);
3995 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3999 #ifdef CONFIG_EXT4_DEBUG
4000 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4002 struct super_block *sb = ac->ac_sb;
4003 ext4_group_t ngroups, i;
4005 if (!ext4_mballoc_debug ||
4006 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4009 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4010 " Allocation context details:");
4011 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4012 ac->ac_status, ac->ac_flags);
4013 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4014 "goal %lu/%lu/%lu@%lu, "
4015 "best %lu/%lu/%lu@%lu cr %d",
4016 (unsigned long)ac->ac_o_ex.fe_group,
4017 (unsigned long)ac->ac_o_ex.fe_start,
4018 (unsigned long)ac->ac_o_ex.fe_len,
4019 (unsigned long)ac->ac_o_ex.fe_logical,
4020 (unsigned long)ac->ac_g_ex.fe_group,
4021 (unsigned long)ac->ac_g_ex.fe_start,
4022 (unsigned long)ac->ac_g_ex.fe_len,
4023 (unsigned long)ac->ac_g_ex.fe_logical,
4024 (unsigned long)ac->ac_b_ex.fe_group,
4025 (unsigned long)ac->ac_b_ex.fe_start,
4026 (unsigned long)ac->ac_b_ex.fe_len,
4027 (unsigned long)ac->ac_b_ex.fe_logical,
4028 (int)ac->ac_criteria);
4029 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4030 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4031 ngroups = ext4_get_groups_count(sb);
4032 for (i = 0; i < ngroups; i++) {
4033 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4034 struct ext4_prealloc_space *pa;
4035 ext4_grpblk_t start;
4036 struct list_head *cur;
4037 ext4_lock_group(sb, i);
4038 list_for_each(cur, &grp->bb_prealloc_list) {
4039 pa = list_entry(cur, struct ext4_prealloc_space,
4041 spin_lock(&pa->pa_lock);
4042 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4044 spin_unlock(&pa->pa_lock);
4045 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4048 ext4_unlock_group(sb, i);
4050 if (grp->bb_free == 0)
4052 printk(KERN_ERR "%u: %d/%d \n",
4053 i, grp->bb_free, grp->bb_fragments);
4055 printk(KERN_ERR "\n");
4058 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4065 * We use locality group preallocation for small size file. The size of the
4066 * file is determined by the current size or the resulting size after
4067 * allocation which ever is larger
4069 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4071 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4073 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4074 int bsbits = ac->ac_sb->s_blocksize_bits;
4077 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4080 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4083 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4084 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4087 if ((size == isize) &&
4088 !ext4_fs_is_busy(sbi) &&
4089 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4090 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4094 if (sbi->s_mb_group_prealloc <= 0) {
4095 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4099 /* don't use group allocation for large files */
4100 size = max(size, isize);
4101 if (size > sbi->s_mb_stream_request) {
4102 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4106 BUG_ON(ac->ac_lg != NULL);
4108 * locality group prealloc space are per cpu. The reason for having
4109 * per cpu locality group is to reduce the contention between block
4110 * request from multiple CPUs.
4112 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
4114 /* we're going to use group allocation */
4115 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4117 /* serialize all allocations in the group */
4118 mutex_lock(&ac->ac_lg->lg_mutex);
4121 static noinline_for_stack int
4122 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4123 struct ext4_allocation_request *ar)
4125 struct super_block *sb = ar->inode->i_sb;
4126 struct ext4_sb_info *sbi = EXT4_SB(sb);
4127 struct ext4_super_block *es = sbi->s_es;
4131 ext4_grpblk_t block;
4133 /* we can't allocate > group size */
4136 /* just a dirty hack to filter too big requests */
4137 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4138 len = EXT4_CLUSTERS_PER_GROUP(sb);
4140 /* start searching from the goal */
4142 if (goal < le32_to_cpu(es->s_first_data_block) ||
4143 goal >= ext4_blocks_count(es))
4144 goal = le32_to_cpu(es->s_first_data_block);
4145 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4147 /* set up allocation goals */
4148 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4149 ac->ac_status = AC_STATUS_CONTINUE;
4151 ac->ac_inode = ar->inode;
4152 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4153 ac->ac_o_ex.fe_group = group;
4154 ac->ac_o_ex.fe_start = block;
4155 ac->ac_o_ex.fe_len = len;
4156 ac->ac_g_ex = ac->ac_o_ex;
4157 ac->ac_flags = ar->flags;
4159 /* we have to define context: we'll we work with a file or
4160 * locality group. this is a policy, actually */
4161 ext4_mb_group_or_file(ac);
4163 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4164 "left: %u/%u, right %u/%u to %swritable\n",
4165 (unsigned) ar->len, (unsigned) ar->logical,
4166 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4167 (unsigned) ar->lleft, (unsigned) ar->pleft,
4168 (unsigned) ar->lright, (unsigned) ar->pright,
4169 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4174 static noinline_for_stack void
4175 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4176 struct ext4_locality_group *lg,
4177 int order, int total_entries)
4179 ext4_group_t group = 0;
4180 struct ext4_buddy e4b;
4181 struct list_head discard_list;
4182 struct ext4_prealloc_space *pa, *tmp;
4184 mb_debug(1, "discard locality group preallocation\n");
4186 INIT_LIST_HEAD(&discard_list);
4188 spin_lock(&lg->lg_prealloc_lock);
4189 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4191 spin_lock(&pa->pa_lock);
4192 if (atomic_read(&pa->pa_count)) {
4194 * This is the pa that we just used
4195 * for block allocation. So don't
4198 spin_unlock(&pa->pa_lock);
4201 if (pa->pa_deleted) {
4202 spin_unlock(&pa->pa_lock);
4205 /* only lg prealloc space */
4206 BUG_ON(pa->pa_type != MB_GROUP_PA);
4208 /* seems this one can be freed ... */
4210 spin_unlock(&pa->pa_lock);
4212 list_del_rcu(&pa->pa_inode_list);
4213 list_add(&pa->u.pa_tmp_list, &discard_list);
4216 if (total_entries <= 5) {
4218 * we want to keep only 5 entries
4219 * allowing it to grow to 8. This
4220 * mak sure we don't call discard
4221 * soon for this list.
4226 spin_unlock(&lg->lg_prealloc_lock);
4228 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4230 group = ext4_get_group_number(sb, pa->pa_pstart);
4231 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4232 ext4_error(sb, "Error loading buddy information for %u",
4236 ext4_lock_group(sb, group);
4237 list_del(&pa->pa_group_list);
4238 ext4_mb_release_group_pa(&e4b, pa);
4239 ext4_unlock_group(sb, group);
4241 ext4_mb_unload_buddy(&e4b);
4242 list_del(&pa->u.pa_tmp_list);
4243 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4248 * We have incremented pa_count. So it cannot be freed at this
4249 * point. Also we hold lg_mutex. So no parallel allocation is
4250 * possible from this lg. That means pa_free cannot be updated.
4252 * A parallel ext4_mb_discard_group_preallocations is possible.
4253 * which can cause the lg_prealloc_list to be updated.
4256 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4258 int order, added = 0, lg_prealloc_count = 1;
4259 struct super_block *sb = ac->ac_sb;
4260 struct ext4_locality_group *lg = ac->ac_lg;
4261 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4263 order = fls(pa->pa_free) - 1;
4264 if (order > PREALLOC_TB_SIZE - 1)
4265 /* The max size of hash table is PREALLOC_TB_SIZE */
4266 order = PREALLOC_TB_SIZE - 1;
4267 /* Add the prealloc space to lg */
4268 spin_lock(&lg->lg_prealloc_lock);
4269 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4271 spin_lock(&tmp_pa->pa_lock);
4272 if (tmp_pa->pa_deleted) {
4273 spin_unlock(&tmp_pa->pa_lock);
4276 if (!added && pa->pa_free < tmp_pa->pa_free) {
4277 /* Add to the tail of the previous entry */
4278 list_add_tail_rcu(&pa->pa_inode_list,
4279 &tmp_pa->pa_inode_list);
4282 * we want to count the total
4283 * number of entries in the list
4286 spin_unlock(&tmp_pa->pa_lock);
4287 lg_prealloc_count++;
4290 list_add_tail_rcu(&pa->pa_inode_list,
4291 &lg->lg_prealloc_list[order]);
4292 spin_unlock(&lg->lg_prealloc_lock);
4294 /* Now trim the list to be not more than 8 elements */
4295 if (lg_prealloc_count > 8) {
4296 ext4_mb_discard_lg_preallocations(sb, lg,
4297 order, lg_prealloc_count);
4304 * release all resource we used in allocation
4306 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4308 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4309 struct ext4_prealloc_space *pa = ac->ac_pa;
4311 if (pa->pa_type == MB_GROUP_PA) {
4312 /* see comment in ext4_mb_use_group_pa() */
4313 spin_lock(&pa->pa_lock);
4314 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4315 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4316 pa->pa_free -= ac->ac_b_ex.fe_len;
4317 pa->pa_len -= ac->ac_b_ex.fe_len;
4318 spin_unlock(&pa->pa_lock);
4323 * We want to add the pa to the right bucket.
4324 * Remove it from the list and while adding
4325 * make sure the list to which we are adding
4328 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4329 spin_lock(pa->pa_obj_lock);
4330 list_del_rcu(&pa->pa_inode_list);
4331 spin_unlock(pa->pa_obj_lock);
4332 ext4_mb_add_n_trim(ac);
4334 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4336 if (ac->ac_bitmap_page)
4337 page_cache_release(ac->ac_bitmap_page);
4338 if (ac->ac_buddy_page)
4339 page_cache_release(ac->ac_buddy_page);
4340 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4341 mutex_unlock(&ac->ac_lg->lg_mutex);
4342 ext4_mb_collect_stats(ac);
4346 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4348 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4352 trace_ext4_mb_discard_preallocations(sb, needed);
4353 for (i = 0; i < ngroups && needed > 0; i++) {
4354 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4363 * Main entry point into mballoc to allocate blocks
4364 * it tries to use preallocation first, then falls back
4365 * to usual allocation
4367 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4368 struct ext4_allocation_request *ar, int *errp)
4371 struct ext4_allocation_context *ac = NULL;
4372 struct ext4_sb_info *sbi;
4373 struct super_block *sb;
4374 ext4_fsblk_t block = 0;
4375 unsigned int inquota = 0;
4376 unsigned int reserv_clstrs = 0;
4379 sb = ar->inode->i_sb;
4382 trace_ext4_request_blocks(ar);
4384 /* Allow to use superuser reservation for quota file */
4385 if (IS_NOQUOTA(ar->inode))
4386 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4389 * For delayed allocation, we could skip the ENOSPC and
4390 * EDQUOT check, as blocks and quotas have been already
4391 * reserved when data being copied into pagecache.
4393 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4394 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4396 /* Without delayed allocation we need to verify
4397 * there is enough free blocks to do block allocation
4398 * and verify allocation doesn't exceed the quota limits.
4401 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4403 /* let others to free the space */
4405 ar->len = ar->len >> 1;
4411 reserv_clstrs = ar->len;
4412 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4413 dquot_alloc_block_nofail(ar->inode,
4414 EXT4_C2B(sbi, ar->len));
4417 dquot_alloc_block(ar->inode,
4418 EXT4_C2B(sbi, ar->len))) {
4420 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4431 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4438 *errp = ext4_mb_initialize_context(ac, ar);
4444 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4445 if (!ext4_mb_use_preallocated(ac)) {
4446 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4447 ext4_mb_normalize_request(ac, ar);
4449 /* allocate space in core */
4450 *errp = ext4_mb_regular_allocator(ac);
4452 goto discard_and_exit;
4454 /* as we've just preallocated more space than
4455 * user requested originally, we store allocated
4456 * space in a special descriptor */
4457 if (ac->ac_status == AC_STATUS_FOUND &&
4458 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4459 *errp = ext4_mb_new_preallocation(ac);
4462 ext4_discard_allocated_blocks(ac);
4466 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4467 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4468 if (*errp == -EAGAIN) {
4470 * drop the reference that we took
4471 * in ext4_mb_use_best_found
4473 ext4_mb_release_context(ac);
4474 ac->ac_b_ex.fe_group = 0;
4475 ac->ac_b_ex.fe_start = 0;
4476 ac->ac_b_ex.fe_len = 0;
4477 ac->ac_status = AC_STATUS_CONTINUE;
4480 ext4_discard_allocated_blocks(ac);
4483 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4484 ar->len = ac->ac_b_ex.fe_len;
4487 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4495 ac->ac_b_ex.fe_len = 0;
4497 ext4_mb_show_ac(ac);
4499 ext4_mb_release_context(ac);
4502 kmem_cache_free(ext4_ac_cachep, ac);
4503 if (inquota && ar->len < inquota)
4504 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4506 if (!ext4_test_inode_state(ar->inode,
4507 EXT4_STATE_DELALLOC_RESERVED))
4508 /* release all the reserved blocks if non delalloc */
4509 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4513 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4519 * We can merge two free data extents only if the physical blocks
4520 * are contiguous, AND the extents were freed by the same transaction,
4521 * AND the blocks are associated with the same group.
4523 static int can_merge(struct ext4_free_data *entry1,
4524 struct ext4_free_data *entry2)
4526 if ((entry1->efd_tid == entry2->efd_tid) &&
4527 (entry1->efd_group == entry2->efd_group) &&
4528 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4533 static noinline_for_stack int
4534 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4535 struct ext4_free_data *new_entry)
4537 ext4_group_t group = e4b->bd_group;
4538 ext4_grpblk_t cluster;
4539 struct ext4_free_data *entry;
4540 struct ext4_group_info *db = e4b->bd_info;
4541 struct super_block *sb = e4b->bd_sb;
4542 struct ext4_sb_info *sbi = EXT4_SB(sb);
4543 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4544 struct rb_node *parent = NULL, *new_node;
4546 BUG_ON(!ext4_handle_valid(handle));
4547 BUG_ON(e4b->bd_bitmap_page == NULL);
4548 BUG_ON(e4b->bd_buddy_page == NULL);
4550 new_node = &new_entry->efd_node;
4551 cluster = new_entry->efd_start_cluster;
4554 /* first free block exent. We need to
4555 protect buddy cache from being freed,
4556 * otherwise we'll refresh it from
4557 * on-disk bitmap and lose not-yet-available
4559 page_cache_get(e4b->bd_buddy_page);
4560 page_cache_get(e4b->bd_bitmap_page);
4564 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4565 if (cluster < entry->efd_start_cluster)
4567 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4568 n = &(*n)->rb_right;
4570 ext4_grp_locked_error(sb, group, 0,
4571 ext4_group_first_block_no(sb, group) +
4572 EXT4_C2B(sbi, cluster),
4573 "Block already on to-be-freed list");
4578 rb_link_node(new_node, parent, n);
4579 rb_insert_color(new_node, &db->bb_free_root);
4581 /* Now try to see the extent can be merged to left and right */
4582 node = rb_prev(new_node);
4584 entry = rb_entry(node, struct ext4_free_data, efd_node);
4585 if (can_merge(entry, new_entry) &&
4586 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4587 new_entry->efd_start_cluster = entry->efd_start_cluster;
4588 new_entry->efd_count += entry->efd_count;
4589 rb_erase(node, &(db->bb_free_root));
4590 kmem_cache_free(ext4_free_data_cachep, entry);
4594 node = rb_next(new_node);
4596 entry = rb_entry(node, struct ext4_free_data, efd_node);
4597 if (can_merge(new_entry, entry) &&
4598 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4599 new_entry->efd_count += entry->efd_count;
4600 rb_erase(node, &(db->bb_free_root));
4601 kmem_cache_free(ext4_free_data_cachep, entry);
4604 /* Add the extent to transaction's private list */
4605 ext4_journal_callback_add(handle, ext4_free_data_callback,
4606 &new_entry->efd_jce);
4611 * ext4_free_blocks() -- Free given blocks and update quota
4612 * @handle: handle for this transaction
4614 * @block: start physical block to free
4615 * @count: number of blocks to count
4616 * @flags: flags used by ext4_free_blocks
4618 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4619 struct buffer_head *bh, ext4_fsblk_t block,
4620 unsigned long count, int flags)
4622 struct buffer_head *bitmap_bh = NULL;
4623 struct super_block *sb = inode->i_sb;
4624 struct ext4_group_desc *gdp;
4625 unsigned int overflow;
4627 struct buffer_head *gd_bh;
4628 ext4_group_t block_group;
4629 struct ext4_sb_info *sbi;
4630 struct ext4_inode_info *ei = EXT4_I(inode);
4631 struct ext4_buddy e4b;
4632 unsigned int count_clusters;
4639 BUG_ON(block != bh->b_blocknr);
4641 block = bh->b_blocknr;
4645 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4646 !ext4_data_block_valid(sbi, block, count)) {
4647 ext4_error(sb, "Freeing blocks not in datazone - "
4648 "block = %llu, count = %lu", block, count);
4652 ext4_debug("freeing block %llu\n", block);
4653 trace_ext4_free_blocks(inode, block, count, flags);
4655 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4656 struct buffer_head *tbh = bh;
4659 BUG_ON(bh && (count > 1));
4661 for (i = 0; i < count; i++) {
4664 tbh = sb_find_get_block(inode->i_sb,
4668 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4669 inode, tbh, block + i);
4674 * We need to make sure we don't reuse the freed block until
4675 * after the transaction is committed, which we can do by
4676 * treating the block as metadata, below. We make an
4677 * exception if the inode is to be written in writeback mode
4678 * since writeback mode has weak data consistency guarantees.
4680 if (!ext4_should_writeback_data(inode))
4681 flags |= EXT4_FREE_BLOCKS_METADATA;
4684 * If the extent to be freed does not begin on a cluster
4685 * boundary, we need to deal with partial clusters at the
4686 * beginning and end of the extent. Normally we will free
4687 * blocks at the beginning or the end unless we are explicitly
4688 * requested to avoid doing so.
4690 overflow = EXT4_PBLK_COFF(sbi, block);
4692 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4693 overflow = sbi->s_cluster_ratio - overflow;
4695 if (count > overflow)
4704 overflow = EXT4_LBLK_COFF(sbi, count);
4706 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4707 if (count > overflow)
4712 count += sbi->s_cluster_ratio - overflow;
4717 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4719 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4720 ext4_get_group_info(sb, block_group))))
4724 * Check to see if we are freeing blocks across a group
4727 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4728 overflow = EXT4_C2B(sbi, bit) + count -
4729 EXT4_BLOCKS_PER_GROUP(sb);
4732 count_clusters = EXT4_NUM_B2C(sbi, count);
4733 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4738 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4744 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4745 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4746 in_range(block, ext4_inode_table(sb, gdp),
4747 EXT4_SB(sb)->s_itb_per_group) ||
4748 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4749 EXT4_SB(sb)->s_itb_per_group)) {
4751 ext4_error(sb, "Freeing blocks in system zone - "
4752 "Block = %llu, count = %lu", block, count);
4753 /* err = 0. ext4_std_error should be a no op */
4757 BUFFER_TRACE(bitmap_bh, "getting write access");
4758 err = ext4_journal_get_write_access(handle, bitmap_bh);
4763 * We are about to modify some metadata. Call the journal APIs
4764 * to unshare ->b_data if a currently-committing transaction is
4767 BUFFER_TRACE(gd_bh, "get_write_access");
4768 err = ext4_journal_get_write_access(handle, gd_bh);
4771 #ifdef AGGRESSIVE_CHECK
4774 for (i = 0; i < count_clusters; i++)
4775 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4778 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4780 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4784 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4785 struct ext4_free_data *new_entry;
4787 * blocks being freed are metadata. these blocks shouldn't
4788 * be used until this transaction is committed
4791 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4794 * We use a retry loop because
4795 * ext4_free_blocks() is not allowed to fail.
4798 congestion_wait(BLK_RW_ASYNC, HZ/50);
4801 new_entry->efd_start_cluster = bit;
4802 new_entry->efd_group = block_group;
4803 new_entry->efd_count = count_clusters;
4804 new_entry->efd_tid = handle->h_transaction->t_tid;
4806 ext4_lock_group(sb, block_group);
4807 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4808 ext4_mb_free_metadata(handle, &e4b, new_entry);
4810 /* need to update group_info->bb_free and bitmap
4811 * with group lock held. generate_buddy look at
4812 * them with group lock_held
4814 if (test_opt(sb, DISCARD)) {
4815 err = ext4_issue_discard(sb, block_group, bit, count);
4816 if (err && err != -EOPNOTSUPP)
4817 ext4_msg(sb, KERN_WARNING, "discard request in"
4818 " group:%d block:%d count:%lu failed"
4819 " with %d", block_group, bit, count,
4822 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4824 ext4_lock_group(sb, block_group);
4825 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4826 mb_free_blocks(inode, &e4b, bit, count_clusters);
4829 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4830 ext4_free_group_clusters_set(sb, gdp, ret);
4831 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4832 ext4_group_desc_csum_set(sb, block_group, gdp);
4833 ext4_unlock_group(sb, block_group);
4835 if (sbi->s_log_groups_per_flex) {
4836 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4837 atomic64_add(count_clusters,
4838 &sbi->s_flex_groups[flex_group].free_clusters);
4841 if (flags & EXT4_FREE_BLOCKS_RESERVE && ei->i_reserved_data_blocks) {
4842 percpu_counter_add(&sbi->s_dirtyclusters_counter,
4844 spin_lock(&ei->i_block_reservation_lock);
4845 if (flags & EXT4_FREE_BLOCKS_METADATA)
4846 ei->i_reserved_meta_blocks += count_clusters;
4848 ei->i_reserved_data_blocks += count_clusters;
4849 spin_unlock(&ei->i_block_reservation_lock);
4850 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4851 dquot_reclaim_block(inode,
4852 EXT4_C2B(sbi, count_clusters));
4853 } else if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4854 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4855 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4857 ext4_mb_unload_buddy(&e4b);
4859 /* We dirtied the bitmap block */
4860 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4861 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4863 /* And the group descriptor block */
4864 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4865 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4869 if (overflow && !err) {
4877 ext4_std_error(sb, err);
4882 * ext4_group_add_blocks() -- Add given blocks to an existing group
4883 * @handle: handle to this transaction
4885 * @block: start physical block to add to the block group
4886 * @count: number of blocks to free
4888 * This marks the blocks as free in the bitmap and buddy.
4890 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4891 ext4_fsblk_t block, unsigned long count)
4893 struct buffer_head *bitmap_bh = NULL;
4894 struct buffer_head *gd_bh;
4895 ext4_group_t block_group;
4898 struct ext4_group_desc *desc;
4899 struct ext4_sb_info *sbi = EXT4_SB(sb);
4900 struct ext4_buddy e4b;
4901 int err = 0, ret, blk_free_count;
4902 ext4_grpblk_t blocks_freed;
4904 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4909 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4911 * Check to see if we are freeing blocks across a group
4914 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4915 ext4_warning(sb, "too much blocks added to group %u\n",
4921 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4927 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4933 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4934 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4935 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4936 in_range(block + count - 1, ext4_inode_table(sb, desc),
4937 sbi->s_itb_per_group)) {
4938 ext4_error(sb, "Adding blocks in system zones - "
4939 "Block = %llu, count = %lu",
4945 BUFFER_TRACE(bitmap_bh, "getting write access");
4946 err = ext4_journal_get_write_access(handle, bitmap_bh);
4951 * We are about to modify some metadata. Call the journal APIs
4952 * to unshare ->b_data if a currently-committing transaction is
4955 BUFFER_TRACE(gd_bh, "get_write_access");
4956 err = ext4_journal_get_write_access(handle, gd_bh);
4960 for (i = 0, blocks_freed = 0; i < count; i++) {
4961 BUFFER_TRACE(bitmap_bh, "clear bit");
4962 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4963 ext4_error(sb, "bit already cleared for block %llu",
4964 (ext4_fsblk_t)(block + i));
4965 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4971 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4976 * need to update group_info->bb_free and bitmap
4977 * with group lock held. generate_buddy look at
4978 * them with group lock_held
4980 ext4_lock_group(sb, block_group);
4981 mb_clear_bits(bitmap_bh->b_data, bit, count);
4982 mb_free_blocks(NULL, &e4b, bit, count);
4983 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4984 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4985 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
4986 ext4_group_desc_csum_set(sb, block_group, desc);
4987 ext4_unlock_group(sb, block_group);
4988 percpu_counter_add(&sbi->s_freeclusters_counter,
4989 EXT4_NUM_B2C(sbi, blocks_freed));
4991 if (sbi->s_log_groups_per_flex) {
4992 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4993 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
4994 &sbi->s_flex_groups[flex_group].free_clusters);
4997 ext4_mb_unload_buddy(&e4b);
4999 /* We dirtied the bitmap block */
5000 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5001 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5003 /* And the group descriptor block */
5004 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5005 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5011 ext4_std_error(sb, err);
5016 * ext4_trim_extent -- function to TRIM one single free extent in the group
5017 * @sb: super block for the file system
5018 * @start: starting block of the free extent in the alloc. group
5019 * @count: number of blocks to TRIM
5020 * @group: alloc. group we are working with
5021 * @e4b: ext4 buddy for the group
5023 * Trim "count" blocks starting at "start" in the "group". To assure that no
5024 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5025 * be called with under the group lock.
5027 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5028 ext4_group_t group, struct ext4_buddy *e4b)
5032 struct ext4_free_extent ex;
5035 trace_ext4_trim_extent(sb, group, start, count);
5037 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5039 ex.fe_start = start;
5040 ex.fe_group = group;
5044 * Mark blocks used, so no one can reuse them while
5047 mb_mark_used(e4b, &ex);
5048 ext4_unlock_group(sb, group);
5049 ret = ext4_issue_discard(sb, group, start, count);
5050 ext4_lock_group(sb, group);
5051 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5056 * ext4_trim_all_free -- function to trim all free space in alloc. group
5057 * @sb: super block for file system
5058 * @group: group to be trimmed
5059 * @start: first group block to examine
5060 * @max: last group block to examine
5061 * @minblocks: minimum extent block count
5063 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5064 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5068 * ext4_trim_all_free walks through group's block bitmap searching for free
5069 * extents. When the free extent is found, mark it as used in group buddy
5070 * bitmap. Then issue a TRIM command on this extent and free the extent in
5071 * the group buddy bitmap. This is done until whole group is scanned.
5073 static ext4_grpblk_t
5074 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5075 ext4_grpblk_t start, ext4_grpblk_t max,
5076 ext4_grpblk_t minblocks)
5079 ext4_grpblk_t next, count = 0, free_count = 0;
5080 struct ext4_buddy e4b;
5083 trace_ext4_trim_all_free(sb, group, start, max);
5085 ret = ext4_mb_load_buddy(sb, group, &e4b);
5087 ext4_error(sb, "Error in loading buddy "
5088 "information for %u", group);
5091 bitmap = e4b.bd_bitmap;
5093 ext4_lock_group(sb, group);
5094 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5095 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5098 start = (e4b.bd_info->bb_first_free > start) ?
5099 e4b.bd_info->bb_first_free : start;
5101 while (start <= max) {
5102 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5105 next = mb_find_next_bit(bitmap, max + 1, start);
5107 if ((next - start) >= minblocks) {
5108 ret = ext4_trim_extent(sb, start,
5109 next - start, group, &e4b);
5110 if (ret && ret != -EOPNOTSUPP)
5113 count += next - start;
5115 free_count += next - start;
5118 if (fatal_signal_pending(current)) {
5119 count = -ERESTARTSYS;
5123 if (need_resched()) {
5124 ext4_unlock_group(sb, group);
5126 ext4_lock_group(sb, group);
5129 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5135 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5138 ext4_unlock_group(sb, group);
5139 ext4_mb_unload_buddy(&e4b);
5141 ext4_debug("trimmed %d blocks in the group %d\n",
5148 * ext4_trim_fs() -- trim ioctl handle function
5149 * @sb: superblock for filesystem
5150 * @range: fstrim_range structure
5152 * start: First Byte to trim
5153 * len: number of Bytes to trim from start
5154 * minlen: minimum extent length in Bytes
5155 * ext4_trim_fs goes through all allocation groups containing Bytes from
5156 * start to start+len. For each such a group ext4_trim_all_free function
5157 * is invoked to trim all free space.
5159 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5161 struct ext4_group_info *grp;
5162 ext4_group_t group, first_group, last_group;
5163 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5164 uint64_t start, end, minlen, trimmed = 0;
5165 ext4_fsblk_t first_data_blk =
5166 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5167 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5170 start = range->start >> sb->s_blocksize_bits;
5171 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5172 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5173 range->minlen >> sb->s_blocksize_bits);
5175 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5176 start >= max_blks ||
5177 range->len < sb->s_blocksize)
5179 if (end >= max_blks)
5181 if (end <= first_data_blk)
5183 if (start < first_data_blk)
5184 start = first_data_blk;
5186 /* Determine first and last group to examine based on start and end */
5187 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5188 &first_group, &first_cluster);
5189 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5190 &last_group, &last_cluster);
5192 /* end now represents the last cluster to discard in this group */
5193 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5195 for (group = first_group; group <= last_group; group++) {
5196 grp = ext4_get_group_info(sb, group);
5197 /* We only do this if the grp has never been initialized */
5198 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5199 ret = ext4_mb_init_group(sb, group);
5205 * For all the groups except the last one, last cluster will
5206 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5207 * change it for the last group, note that last_cluster is
5208 * already computed earlier by ext4_get_group_no_and_offset()
5210 if (group == last_group)
5213 if (grp->bb_free >= minlen) {
5214 cnt = ext4_trim_all_free(sb, group, first_cluster,
5224 * For every group except the first one, we are sure
5225 * that the first cluster to discard will be cluster #0.
5231 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5234 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;