1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
4 * Written by Alex Tomas <alex@clusterfs.com>
9 * mballoc.c contains the multiblocks allocation routines
12 #include "ext4_jbd2.h"
14 #include <linux/log2.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/backing-dev.h>
18 #include <trace/events/ext4.h>
20 #ifdef CONFIG_EXT4_DEBUG
21 ushort ext4_mballoc_debug __read_mostly;
23 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
24 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
29 * - test ext4_ext_search_left() and ext4_ext_search_right()
30 * - search for metadata in few groups
33 * - normalization should take into account whether file is still open
34 * - discard preallocations if no free space left (policy?)
35 * - don't normalize tails
37 * - reservation for superuser
40 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
41 * - track min/max extents in each group for better group selection
42 * - mb_mark_used() may allocate chunk right after splitting buddy
43 * - tree of groups sorted by number of free blocks
48 * The allocation request involve request for multiple number of blocks
49 * near to the goal(block) value specified.
51 * During initialization phase of the allocator we decide to use the
52 * group preallocation or inode preallocation depending on the size of
53 * the file. The size of the file could be the resulting file size we
54 * would have after allocation, or the current file size, which ever
55 * is larger. If the size is less than sbi->s_mb_stream_request we
56 * select to use the group preallocation. The default value of
57 * s_mb_stream_request is 16 blocks. This can also be tuned via
58 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
59 * terms of number of blocks.
61 * The main motivation for having small file use group preallocation is to
62 * ensure that we have small files closer together on the disk.
64 * First stage the allocator looks at the inode prealloc list,
65 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
66 * spaces for this particular inode. The inode prealloc space is
69 * pa_lstart -> the logical start block for this prealloc space
70 * pa_pstart -> the physical start block for this prealloc space
71 * pa_len -> length for this prealloc space (in clusters)
72 * pa_free -> free space available in this prealloc space (in clusters)
74 * The inode preallocation space is used looking at the _logical_ start
75 * block. If only the logical file block falls within the range of prealloc
76 * space we will consume the particular prealloc space. This makes sure that
77 * we have contiguous physical blocks representing the file blocks
79 * The important thing to be noted in case of inode prealloc space is that
80 * we don't modify the values associated to inode prealloc space except
83 * If we are not able to find blocks in the inode prealloc space and if we
84 * have the group allocation flag set then we look at the locality group
85 * prealloc space. These are per CPU prealloc list represented as
87 * ext4_sb_info.s_locality_groups[smp_processor_id()]
89 * The reason for having a per cpu locality group is to reduce the contention
90 * between CPUs. It is possible to get scheduled at this point.
92 * The locality group prealloc space is used looking at whether we have
93 * enough free space (pa_free) within the prealloc space.
95 * If we can't allocate blocks via inode prealloc or/and locality group
96 * prealloc then we look at the buddy cache. The buddy cache is represented
97 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
98 * mapped to the buddy and bitmap information regarding different
99 * groups. The buddy information is attached to buddy cache inode so that
100 * we can access them through the page cache. The information regarding
101 * each group is loaded via ext4_mb_load_buddy. The information involve
102 * block bitmap and buddy information. The information are stored in the
106 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
109 * one block each for bitmap and buddy information. So for each group we
110 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
111 * blocksize) blocks. So it can have information regarding groups_per_page
112 * which is blocks_per_page/2
114 * The buddy cache inode is not stored on disk. The inode is thrown
115 * away when the filesystem is unmounted.
117 * We look for count number of blocks in the buddy cache. If we were able
118 * to locate that many free blocks we return with additional information
119 * regarding rest of the contiguous physical block available
121 * Before allocating blocks via buddy cache we normalize the request
122 * blocks. This ensure we ask for more blocks that we needed. The extra
123 * blocks that we get after allocation is added to the respective prealloc
124 * list. In case of inode preallocation we follow a list of heuristics
125 * based on file size. This can be found in ext4_mb_normalize_request. If
126 * we are doing a group prealloc we try to normalize the request to
127 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
128 * dependent on the cluster size; for non-bigalloc file systems, it is
129 * 512 blocks. This can be tuned via
130 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
131 * terms of number of blocks. If we have mounted the file system with -O
132 * stripe=<value> option the group prealloc request is normalized to the
133 * the smallest multiple of the stripe value (sbi->s_stripe) which is
134 * greater than the default mb_group_prealloc.
136 * The regular allocator (using the buddy cache) supports a few tunables.
138 * /sys/fs/ext4/<partition>/mb_min_to_scan
139 * /sys/fs/ext4/<partition>/mb_max_to_scan
140 * /sys/fs/ext4/<partition>/mb_order2_req
142 * The regular allocator uses buddy scan only if the request len is power of
143 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
144 * value of s_mb_order2_reqs can be tuned via
145 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
146 * stripe size (sbi->s_stripe), we try to search for contiguous block in
147 * stripe size. This should result in better allocation on RAID setups. If
148 * not, we search in the specific group using bitmap for best extents. The
149 * tunable min_to_scan and max_to_scan control the behaviour here.
150 * min_to_scan indicate how long the mballoc __must__ look for a best
151 * extent and max_to_scan indicates how long the mballoc __can__ look for a
152 * best extent in the found extents. Searching for the blocks starts with
153 * the group specified as the goal value in allocation context via
154 * ac_g_ex. Each group is first checked based on the criteria whether it
155 * can be used for allocation. ext4_mb_good_group explains how the groups are
158 * Both the prealloc space are getting populated as above. So for the first
159 * request we will hit the buddy cache which will result in this prealloc
160 * space getting filled. The prealloc space is then later used for the
161 * subsequent request.
165 * mballoc operates on the following data:
167 * - in-core buddy (actually includes buddy and bitmap)
168 * - preallocation descriptors (PAs)
170 * there are two types of preallocations:
172 * assiged to specific inode and can be used for this inode only.
173 * it describes part of inode's space preallocated to specific
174 * physical blocks. any block from that preallocated can be used
175 * independent. the descriptor just tracks number of blocks left
176 * unused. so, before taking some block from descriptor, one must
177 * make sure corresponded logical block isn't allocated yet. this
178 * also means that freeing any block within descriptor's range
179 * must discard all preallocated blocks.
181 * assigned to specific locality group which does not translate to
182 * permanent set of inodes: inode can join and leave group. space
183 * from this type of preallocation can be used for any inode. thus
184 * it's consumed from the beginning to the end.
186 * relation between them can be expressed as:
187 * in-core buddy = on-disk bitmap + preallocation descriptors
189 * this mean blocks mballoc considers used are:
190 * - allocated blocks (persistent)
191 * - preallocated blocks (non-persistent)
193 * consistency in mballoc world means that at any time a block is either
194 * free or used in ALL structures. notice: "any time" should not be read
195 * literally -- time is discrete and delimited by locks.
197 * to keep it simple, we don't use block numbers, instead we count number of
198 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
200 * all operations can be expressed as:
201 * - init buddy: buddy = on-disk + PAs
202 * - new PA: buddy += N; PA = N
203 * - use inode PA: on-disk += N; PA -= N
204 * - discard inode PA buddy -= on-disk - PA; PA = 0
205 * - use locality group PA on-disk += N; PA -= N
206 * - discard locality group PA buddy -= PA; PA = 0
207 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
208 * is used in real operation because we can't know actual used
209 * bits from PA, only from on-disk bitmap
211 * if we follow this strict logic, then all operations above should be atomic.
212 * given some of them can block, we'd have to use something like semaphores
213 * killing performance on high-end SMP hardware. let's try to relax it using
214 * the following knowledge:
215 * 1) if buddy is referenced, it's already initialized
216 * 2) while block is used in buddy and the buddy is referenced,
217 * nobody can re-allocate that block
218 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
219 * bit set and PA claims same block, it's OK. IOW, one can set bit in
220 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
223 * so, now we're building a concurrency table:
226 * blocks for PA are allocated in the buddy, buddy must be referenced
227 * until PA is linked to allocation group to avoid concurrent buddy init
229 * we need to make sure that either on-disk bitmap or PA has uptodate data
230 * given (3) we care that PA-=N operation doesn't interfere with init
232 * the simplest way would be to have buddy initialized by the discard
233 * - use locality group PA
234 * again PA-=N must be serialized with init
235 * - discard locality group PA
236 * the simplest way would be to have buddy initialized by the discard
239 * i_data_sem serializes them
241 * discard process must wait until PA isn't used by another process
242 * - use locality group PA
243 * some mutex should serialize them
244 * - discard locality group PA
245 * discard process must wait until PA isn't used by another process
248 * i_data_sem or another mutex should serializes them
250 * discard process must wait until PA isn't used by another process
251 * - use locality group PA
252 * nothing wrong here -- they're different PAs covering different blocks
253 * - discard locality group PA
254 * discard process must wait until PA isn't used by another process
256 * now we're ready to make few consequences:
257 * - PA is referenced and while it is no discard is possible
258 * - PA is referenced until block isn't marked in on-disk bitmap
259 * - PA changes only after on-disk bitmap
260 * - discard must not compete with init. either init is done before
261 * any discard or they're serialized somehow
262 * - buddy init as sum of on-disk bitmap and PAs is done atomically
264 * a special case when we've used PA to emptiness. no need to modify buddy
265 * in this case, but we should care about concurrent init
270 * Logic in few words:
275 * mark bits in on-disk bitmap
278 * - use preallocation:
279 * find proper PA (per-inode or group)
281 * mark bits in on-disk bitmap
287 * mark bits in on-disk bitmap
290 * - discard preallocations in group:
292 * move them onto local list
293 * load on-disk bitmap
295 * remove PA from object (inode or locality group)
296 * mark free blocks in-core
298 * - discard inode's preallocations:
305 * - bitlock on a group (group)
306 * - object (inode/locality) (object)
317 * - release consumed pa:
322 * - generate in-core bitmap:
326 * - discard all for given object (inode, locality group):
331 * - discard all for given group:
338 static struct kmem_cache *ext4_pspace_cachep;
339 static struct kmem_cache *ext4_ac_cachep;
340 static struct kmem_cache *ext4_free_data_cachep;
342 /* We create slab caches for groupinfo data structures based on the
343 * superblock block size. There will be one per mounted filesystem for
344 * each unique s_blocksize_bits */
345 #define NR_GRPINFO_CACHES 8
346 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
348 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
349 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
350 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
351 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
354 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
356 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
359 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
361 #if BITS_PER_LONG == 64
362 *bit += ((unsigned long) addr & 7UL) << 3;
363 addr = (void *) ((unsigned long) addr & ~7UL);
364 #elif BITS_PER_LONG == 32
365 *bit += ((unsigned long) addr & 3UL) << 3;
366 addr = (void *) ((unsigned long) addr & ~3UL);
368 #error "how many bits you are?!"
373 static inline int mb_test_bit(int bit, void *addr)
376 * ext4_test_bit on architecture like powerpc
377 * needs unsigned long aligned address
379 addr = mb_correct_addr_and_bit(&bit, addr);
380 return ext4_test_bit(bit, addr);
383 static inline void mb_set_bit(int bit, void *addr)
385 addr = mb_correct_addr_and_bit(&bit, addr);
386 ext4_set_bit(bit, addr);
389 static inline void mb_clear_bit(int bit, void *addr)
391 addr = mb_correct_addr_and_bit(&bit, addr);
392 ext4_clear_bit(bit, addr);
395 static inline int mb_test_and_clear_bit(int bit, void *addr)
397 addr = mb_correct_addr_and_bit(&bit, addr);
398 return ext4_test_and_clear_bit(bit, addr);
401 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
403 int fix = 0, ret, tmpmax;
404 addr = mb_correct_addr_and_bit(&fix, addr);
408 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
414 static inline int mb_find_next_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_bit(addr, tmpmax, start) - fix;
427 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
431 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
434 if (order > e4b->bd_blkbits + 1) {
439 /* at order 0 we see each particular block */
441 *max = 1 << (e4b->bd_blkbits + 3);
442 return e4b->bd_bitmap;
445 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
446 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
452 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
453 int first, int count)
456 struct super_block *sb = e4b->bd_sb;
458 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
460 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
461 for (i = 0; i < count; i++) {
462 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
463 ext4_fsblk_t blocknr;
465 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
466 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
467 ext4_grp_locked_error(sb, e4b->bd_group,
468 inode ? inode->i_ino : 0,
470 "freeing block already freed "
473 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
474 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
476 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
480 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
484 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
486 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
487 for (i = 0; i < count; i++) {
488 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
489 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
493 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
495 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
496 unsigned char *b1, *b2;
498 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
499 b2 = (unsigned char *) bitmap;
500 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
501 if (b1[i] != b2[i]) {
502 ext4_msg(e4b->bd_sb, KERN_ERR,
503 "corruption in group %u "
504 "at byte %u(%u): %x in copy != %x "
506 e4b->bd_group, i, i * 8, b1[i], b2[i]);
514 static inline void mb_free_blocks_double(struct inode *inode,
515 struct ext4_buddy *e4b, int first, int count)
519 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
520 int first, int count)
524 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
530 #ifdef AGGRESSIVE_CHECK
532 #define MB_CHECK_ASSERT(assert) \
536 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
537 function, file, line, # assert); \
542 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
543 const char *function, int line)
545 struct super_block *sb = e4b->bd_sb;
546 int order = e4b->bd_blkbits + 1;
553 struct ext4_group_info *grp;
556 struct list_head *cur;
561 static int mb_check_counter;
562 if (mb_check_counter++ % 100 != 0)
567 buddy = mb_find_buddy(e4b, order, &max);
568 MB_CHECK_ASSERT(buddy);
569 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
570 MB_CHECK_ASSERT(buddy2);
571 MB_CHECK_ASSERT(buddy != buddy2);
572 MB_CHECK_ASSERT(max * 2 == max2);
575 for (i = 0; i < max; i++) {
577 if (mb_test_bit(i, buddy)) {
578 /* only single bit in buddy2 may be 1 */
579 if (!mb_test_bit(i << 1, buddy2)) {
581 mb_test_bit((i<<1)+1, buddy2));
582 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
584 mb_test_bit(i << 1, buddy2));
589 /* both bits in buddy2 must be 1 */
590 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
591 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
593 for (j = 0; j < (1 << order); j++) {
594 k = (i * (1 << order)) + j;
596 !mb_test_bit(k, e4b->bd_bitmap));
600 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
605 buddy = mb_find_buddy(e4b, 0, &max);
606 for (i = 0; i < max; i++) {
607 if (!mb_test_bit(i, buddy)) {
608 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
616 /* check used bits only */
617 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
618 buddy2 = mb_find_buddy(e4b, j, &max2);
620 MB_CHECK_ASSERT(k < max2);
621 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
624 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
625 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
627 grp = ext4_get_group_info(sb, e4b->bd_group);
628 list_for_each(cur, &grp->bb_prealloc_list) {
629 ext4_group_t groupnr;
630 struct ext4_prealloc_space *pa;
631 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
632 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
633 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
634 for (i = 0; i < pa->pa_len; i++)
635 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
639 #undef MB_CHECK_ASSERT
640 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
641 __FILE__, __func__, __LINE__)
643 #define mb_check_buddy(e4b)
647 * Divide blocks started from @first with length @len into
648 * smaller chunks with power of 2 blocks.
649 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
650 * then increase bb_counters[] for corresponded chunk size.
652 static void ext4_mb_mark_free_simple(struct super_block *sb,
653 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
654 struct ext4_group_info *grp)
656 struct ext4_sb_info *sbi = EXT4_SB(sb);
662 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
664 border = 2 << sb->s_blocksize_bits;
667 /* find how many blocks can be covered since this position */
668 max = ffs(first | border) - 1;
670 /* find how many blocks of power 2 we need to mark */
677 /* mark multiblock chunks only */
678 grp->bb_counters[min]++;
680 mb_clear_bit(first >> min,
681 buddy + sbi->s_mb_offsets[min]);
689 * Cache the order of the largest free extent we have available in this block
693 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
698 grp->bb_largest_free_order = -1; /* uninit */
700 bits = sb->s_blocksize_bits + 1;
701 for (i = bits; i >= 0; i--) {
702 if (grp->bb_counters[i] > 0) {
703 grp->bb_largest_free_order = i;
709 static noinline_for_stack
710 void ext4_mb_generate_buddy(struct super_block *sb,
711 void *buddy, void *bitmap, ext4_group_t group)
713 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
714 struct ext4_sb_info *sbi = EXT4_SB(sb);
715 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
720 unsigned fragments = 0;
721 unsigned long long period = get_cycles();
723 /* initialize buddy from bitmap which is aggregation
724 * of on-disk bitmap and preallocations */
725 i = mb_find_next_zero_bit(bitmap, max, 0);
726 grp->bb_first_free = i;
730 i = mb_find_next_bit(bitmap, max, i);
734 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
736 grp->bb_counters[0]++;
738 i = mb_find_next_zero_bit(bitmap, max, i);
740 grp->bb_fragments = fragments;
742 if (free != grp->bb_free) {
743 ext4_grp_locked_error(sb, group, 0, 0,
744 "block bitmap and bg descriptor "
745 "inconsistent: %u vs %u free clusters",
748 * If we intend to continue, we consider group descriptor
749 * corrupt and update bb_free using bitmap value
752 ext4_mark_group_bitmap_corrupted(sb, group,
753 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
755 mb_set_largest_free_order(sb, grp);
757 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
759 period = get_cycles() - period;
760 spin_lock(&sbi->s_bal_lock);
761 sbi->s_mb_buddies_generated++;
762 sbi->s_mb_generation_time += period;
763 spin_unlock(&sbi->s_bal_lock);
766 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
772 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
773 ext4_set_bits(buddy, 0, count);
775 e4b->bd_info->bb_fragments = 0;
776 memset(e4b->bd_info->bb_counters, 0,
777 sizeof(*e4b->bd_info->bb_counters) *
778 (e4b->bd_sb->s_blocksize_bits + 2));
780 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
781 e4b->bd_bitmap, e4b->bd_group);
784 /* The buddy information is attached the buddy cache inode
785 * for convenience. The information regarding each group
786 * is loaded via ext4_mb_load_buddy. The information involve
787 * block bitmap and buddy information. The information are
788 * stored in the inode as
791 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
794 * one block each for bitmap and buddy information.
795 * So for each group we take up 2 blocks. A page can
796 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
797 * So it can have information regarding groups_per_page which
798 * is blocks_per_page/2
800 * Locking note: This routine takes the block group lock of all groups
801 * for this page; do not hold this lock when calling this routine!
804 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
806 ext4_group_t ngroups;
812 ext4_group_t first_group, group;
814 struct super_block *sb;
815 struct buffer_head *bhs;
816 struct buffer_head **bh = NULL;
820 struct ext4_group_info *grinfo;
822 mb_debug(1, "init page %lu\n", page->index);
824 inode = page->mapping->host;
826 ngroups = ext4_get_groups_count(sb);
827 blocksize = i_blocksize(inode);
828 blocks_per_page = PAGE_SIZE / blocksize;
830 groups_per_page = blocks_per_page >> 1;
831 if (groups_per_page == 0)
834 /* allocate buffer_heads to read bitmaps */
835 if (groups_per_page > 1) {
836 i = sizeof(struct buffer_head *) * groups_per_page;
837 bh = kzalloc(i, gfp);
845 first_group = page->index * blocks_per_page / 2;
847 /* read all groups the page covers into the cache */
848 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
849 if (group >= ngroups)
852 grinfo = ext4_get_group_info(sb, group);
854 * If page is uptodate then we came here after online resize
855 * which added some new uninitialized group info structs, so
856 * we must skip all initialized uptodate buddies on the page,
857 * which may be currently in use by an allocating task.
859 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
863 bh[i] = ext4_read_block_bitmap_nowait(sb, group);
865 err = PTR_ERR(bh[i]);
869 mb_debug(1, "read bitmap for group %u\n", group);
872 /* wait for I/O completion */
873 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
878 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
883 first_block = page->index * blocks_per_page;
884 for (i = 0; i < blocks_per_page; i++) {
885 group = (first_block + i) >> 1;
886 if (group >= ngroups)
889 if (!bh[group - first_group])
890 /* skip initialized uptodate buddy */
893 if (!buffer_verified(bh[group - first_group]))
894 /* Skip faulty bitmaps */
899 * data carry information regarding this
900 * particular group in the format specified
904 data = page_address(page) + (i * blocksize);
905 bitmap = bh[group - first_group]->b_data;
908 * We place the buddy block and bitmap block
911 if ((first_block + i) & 1) {
912 /* this is block of buddy */
913 BUG_ON(incore == NULL);
914 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
915 group, page->index, i * blocksize);
916 trace_ext4_mb_buddy_bitmap_load(sb, group);
917 grinfo = ext4_get_group_info(sb, group);
918 grinfo->bb_fragments = 0;
919 memset(grinfo->bb_counters, 0,
920 sizeof(*grinfo->bb_counters) *
921 (sb->s_blocksize_bits+2));
923 * incore got set to the group block bitmap below
925 ext4_lock_group(sb, group);
927 memset(data, 0xff, blocksize);
928 ext4_mb_generate_buddy(sb, data, incore, group);
929 ext4_unlock_group(sb, group);
932 /* this is block of bitmap */
933 BUG_ON(incore != NULL);
934 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
935 group, page->index, i * blocksize);
936 trace_ext4_mb_bitmap_load(sb, group);
938 /* see comments in ext4_mb_put_pa() */
939 ext4_lock_group(sb, group);
940 memcpy(data, bitmap, blocksize);
942 /* mark all preallocated blks used in in-core bitmap */
943 ext4_mb_generate_from_pa(sb, data, group);
944 ext4_mb_generate_from_freelist(sb, data, group);
945 ext4_unlock_group(sb, group);
947 /* set incore so that the buddy information can be
948 * generated using this
953 SetPageUptodate(page);
957 for (i = 0; i < groups_per_page; i++)
966 * Lock the buddy and bitmap pages. This make sure other parallel init_group
967 * on the same buddy page doesn't happen whild holding the buddy page lock.
968 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
969 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
971 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
972 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
974 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
975 int block, pnum, poff;
979 e4b->bd_buddy_page = NULL;
980 e4b->bd_bitmap_page = NULL;
982 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
984 * the buddy cache inode stores the block bitmap
985 * and buddy information in consecutive blocks.
986 * So for each group we need two blocks.
989 pnum = block / blocks_per_page;
990 poff = block % blocks_per_page;
991 page = find_or_create_page(inode->i_mapping, pnum, gfp);
994 BUG_ON(page->mapping != inode->i_mapping);
995 e4b->bd_bitmap_page = page;
996 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
998 if (blocks_per_page >= 2) {
999 /* buddy and bitmap are on the same page */
1004 pnum = block / blocks_per_page;
1005 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1008 BUG_ON(page->mapping != inode->i_mapping);
1009 e4b->bd_buddy_page = page;
1013 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1015 if (e4b->bd_bitmap_page) {
1016 unlock_page(e4b->bd_bitmap_page);
1017 put_page(e4b->bd_bitmap_page);
1019 if (e4b->bd_buddy_page) {
1020 unlock_page(e4b->bd_buddy_page);
1021 put_page(e4b->bd_buddy_page);
1026 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1027 * block group lock of all groups for this page; do not hold the BG lock when
1028 * calling this routine!
1030 static noinline_for_stack
1031 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1034 struct ext4_group_info *this_grp;
1035 struct ext4_buddy e4b;
1040 mb_debug(1, "init group %u\n", group);
1041 this_grp = ext4_get_group_info(sb, group);
1043 * This ensures that we don't reinit the buddy cache
1044 * page which map to the group from which we are already
1045 * allocating. If we are looking at the buddy cache we would
1046 * have taken a reference using ext4_mb_load_buddy and that
1047 * would have pinned buddy page to page cache.
1048 * The call to ext4_mb_get_buddy_page_lock will mark the
1051 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1052 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1054 * somebody initialized the group
1055 * return without doing anything
1060 page = e4b.bd_bitmap_page;
1061 ret = ext4_mb_init_cache(page, NULL, gfp);
1064 if (!PageUptodate(page)) {
1069 if (e4b.bd_buddy_page == NULL) {
1071 * If both the bitmap and buddy are in
1072 * the same page we don't need to force
1078 /* init buddy cache */
1079 page = e4b.bd_buddy_page;
1080 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1083 if (!PageUptodate(page)) {
1088 ext4_mb_put_buddy_page_lock(&e4b);
1093 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1094 * block group lock of all groups for this page; do not hold the BG lock when
1095 * calling this routine!
1097 static noinline_for_stack int
1098 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1099 struct ext4_buddy *e4b, gfp_t gfp)
1101 int blocks_per_page;
1107 struct ext4_group_info *grp;
1108 struct ext4_sb_info *sbi = EXT4_SB(sb);
1109 struct inode *inode = sbi->s_buddy_cache;
1112 mb_debug(1, "load group %u\n", group);
1114 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1115 grp = ext4_get_group_info(sb, group);
1117 e4b->bd_blkbits = sb->s_blocksize_bits;
1120 e4b->bd_group = group;
1121 e4b->bd_buddy_page = NULL;
1122 e4b->bd_bitmap_page = NULL;
1124 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1126 * we need full data about the group
1127 * to make a good selection
1129 ret = ext4_mb_init_group(sb, group, gfp);
1135 * the buddy cache inode stores the block bitmap
1136 * and buddy information in consecutive blocks.
1137 * So for each group we need two blocks.
1140 pnum = block / blocks_per_page;
1141 poff = block % blocks_per_page;
1143 /* we could use find_or_create_page(), but it locks page
1144 * what we'd like to avoid in fast path ... */
1145 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1146 if (page == NULL || !PageUptodate(page)) {
1149 * drop the page reference and try
1150 * to get the page with lock. If we
1151 * are not uptodate that implies
1152 * somebody just created the page but
1153 * is yet to initialize the same. So
1154 * wait for it to initialize.
1157 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1159 BUG_ON(page->mapping != inode->i_mapping);
1160 if (!PageUptodate(page)) {
1161 ret = ext4_mb_init_cache(page, NULL, gfp);
1166 mb_cmp_bitmaps(e4b, page_address(page) +
1167 (poff * sb->s_blocksize));
1176 if (!PageUptodate(page)) {
1181 /* Pages marked accessed already */
1182 e4b->bd_bitmap_page = page;
1183 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1186 pnum = block / blocks_per_page;
1187 poff = block % blocks_per_page;
1189 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1190 if (page == NULL || !PageUptodate(page)) {
1193 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1195 BUG_ON(page->mapping != inode->i_mapping);
1196 if (!PageUptodate(page)) {
1197 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1211 if (!PageUptodate(page)) {
1216 /* Pages marked accessed already */
1217 e4b->bd_buddy_page = page;
1218 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1220 BUG_ON(e4b->bd_bitmap_page == NULL);
1221 BUG_ON(e4b->bd_buddy_page == NULL);
1228 if (e4b->bd_bitmap_page)
1229 put_page(e4b->bd_bitmap_page);
1230 if (e4b->bd_buddy_page)
1231 put_page(e4b->bd_buddy_page);
1232 e4b->bd_buddy = NULL;
1233 e4b->bd_bitmap = NULL;
1237 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1238 struct ext4_buddy *e4b)
1240 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1243 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1245 if (e4b->bd_bitmap_page)
1246 put_page(e4b->bd_bitmap_page);
1247 if (e4b->bd_buddy_page)
1248 put_page(e4b->bd_buddy_page);
1252 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1255 int bb_incr = 1 << (e4b->bd_blkbits - 1);
1258 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1259 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1262 while (order <= e4b->bd_blkbits + 1) {
1264 if (!mb_test_bit(block, bb)) {
1265 /* this block is part of buddy of order 'order' */
1275 static void mb_clear_bits(void *bm, int cur, int len)
1281 if ((cur & 31) == 0 && (len - cur) >= 32) {
1282 /* fast path: clear whole word at once */
1283 addr = bm + (cur >> 3);
1288 mb_clear_bit(cur, bm);
1293 /* clear bits in given range
1294 * will return first found zero bit if any, -1 otherwise
1296 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1303 if ((cur & 31) == 0 && (len - cur) >= 32) {
1304 /* fast path: clear whole word at once */
1305 addr = bm + (cur >> 3);
1306 if (*addr != (__u32)(-1) && zero_bit == -1)
1307 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1312 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1320 void ext4_set_bits(void *bm, int cur, int len)
1326 if ((cur & 31) == 0 && (len - cur) >= 32) {
1327 /* fast path: set whole word at once */
1328 addr = bm + (cur >> 3);
1333 mb_set_bit(cur, bm);
1339 * _________________________________________________________________ */
1341 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1343 if (mb_test_bit(*bit + side, bitmap)) {
1344 mb_clear_bit(*bit, bitmap);
1350 mb_set_bit(*bit, bitmap);
1355 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1359 void *buddy = mb_find_buddy(e4b, order, &max);
1364 /* Bits in range [first; last] are known to be set since
1365 * corresponding blocks were allocated. Bits in range
1366 * (first; last) will stay set because they form buddies on
1367 * upper layer. We just deal with borders if they don't
1368 * align with upper layer and then go up.
1369 * Releasing entire group is all about clearing
1370 * single bit of highest order buddy.
1374 * ---------------------------------
1376 * ---------------------------------
1377 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1378 * ---------------------------------
1380 * \_____________________/
1382 * Neither [1] nor [6] is aligned to above layer.
1383 * Left neighbour [0] is free, so mark it busy,
1384 * decrease bb_counters and extend range to
1386 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1387 * mark [6] free, increase bb_counters and shrink range to
1389 * Then shift range to [0; 2], go up and do the same.
1394 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1396 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1401 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1402 mb_clear_bits(buddy, first, last - first + 1);
1403 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1412 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1413 int first, int count)
1415 int left_is_free = 0;
1416 int right_is_free = 0;
1418 int last = first + count - 1;
1419 struct super_block *sb = e4b->bd_sb;
1421 if (WARN_ON(count == 0))
1423 BUG_ON(last >= (sb->s_blocksize << 3));
1424 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1425 /* Don't bother if the block group is corrupt. */
1426 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1429 mb_check_buddy(e4b);
1430 mb_free_blocks_double(inode, e4b, first, count);
1432 e4b->bd_info->bb_free += count;
1433 if (first < e4b->bd_info->bb_first_free)
1434 e4b->bd_info->bb_first_free = first;
1436 /* access memory sequentially: check left neighbour,
1437 * clear range and then check right neighbour
1440 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1441 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1442 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1443 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1445 if (unlikely(block != -1)) {
1446 struct ext4_sb_info *sbi = EXT4_SB(sb);
1447 ext4_fsblk_t blocknr;
1449 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1450 blocknr += EXT4_C2B(sbi, block);
1451 ext4_grp_locked_error(sb, e4b->bd_group,
1452 inode ? inode->i_ino : 0,
1454 "freeing already freed block "
1455 "(bit %u); block bitmap corrupt.",
1457 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1458 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1459 mb_regenerate_buddy(e4b);
1463 /* let's maintain fragments counter */
1464 if (left_is_free && right_is_free)
1465 e4b->bd_info->bb_fragments--;
1466 else if (!left_is_free && !right_is_free)
1467 e4b->bd_info->bb_fragments++;
1469 /* buddy[0] == bd_bitmap is a special case, so handle
1470 * it right away and let mb_buddy_mark_free stay free of
1471 * zero order checks.
1472 * Check if neighbours are to be coaleasced,
1473 * adjust bitmap bb_counters and borders appropriately.
1476 first += !left_is_free;
1477 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1480 last -= !right_is_free;
1481 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1485 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1488 mb_set_largest_free_order(sb, e4b->bd_info);
1489 mb_check_buddy(e4b);
1492 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1493 int needed, struct ext4_free_extent *ex)
1499 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1502 buddy = mb_find_buddy(e4b, 0, &max);
1503 BUG_ON(buddy == NULL);
1504 BUG_ON(block >= max);
1505 if (mb_test_bit(block, buddy)) {
1512 /* find actual order */
1513 order = mb_find_order_for_block(e4b, block);
1514 block = block >> order;
1516 ex->fe_len = 1 << order;
1517 ex->fe_start = block << order;
1518 ex->fe_group = e4b->bd_group;
1520 /* calc difference from given start */
1521 next = next - ex->fe_start;
1523 ex->fe_start += next;
1525 while (needed > ex->fe_len &&
1526 mb_find_buddy(e4b, order, &max)) {
1528 if (block + 1 >= max)
1531 next = (block + 1) * (1 << order);
1532 if (mb_test_bit(next, e4b->bd_bitmap))
1535 order = mb_find_order_for_block(e4b, next);
1537 block = next >> order;
1538 ex->fe_len += 1 << order;
1541 if (ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))) {
1542 /* Should never happen! (but apparently sometimes does?!?) */
1544 ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent "
1545 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1546 block, order, needed, ex->fe_group, ex->fe_start,
1547 ex->fe_len, ex->fe_logical);
1555 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1561 int start = ex->fe_start;
1562 int len = ex->fe_len;
1567 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1568 BUG_ON(e4b->bd_group != ex->fe_group);
1569 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1570 mb_check_buddy(e4b);
1571 mb_mark_used_double(e4b, start, len);
1573 e4b->bd_info->bb_free -= len;
1574 if (e4b->bd_info->bb_first_free == start)
1575 e4b->bd_info->bb_first_free += len;
1577 /* let's maintain fragments counter */
1579 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1580 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1581 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1583 e4b->bd_info->bb_fragments++;
1584 else if (!mlen && !max)
1585 e4b->bd_info->bb_fragments--;
1587 /* let's maintain buddy itself */
1589 ord = mb_find_order_for_block(e4b, start);
1591 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1592 /* the whole chunk may be allocated at once! */
1594 buddy = mb_find_buddy(e4b, ord, &max);
1595 BUG_ON((start >> ord) >= max);
1596 mb_set_bit(start >> ord, buddy);
1597 e4b->bd_info->bb_counters[ord]--;
1604 /* store for history */
1606 ret = len | (ord << 16);
1608 /* we have to split large buddy */
1610 buddy = mb_find_buddy(e4b, ord, &max);
1611 mb_set_bit(start >> ord, buddy);
1612 e4b->bd_info->bb_counters[ord]--;
1615 cur = (start >> ord) & ~1U;
1616 buddy = mb_find_buddy(e4b, ord, &max);
1617 mb_clear_bit(cur, buddy);
1618 mb_clear_bit(cur + 1, buddy);
1619 e4b->bd_info->bb_counters[ord]++;
1620 e4b->bd_info->bb_counters[ord]++;
1622 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1624 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1625 mb_check_buddy(e4b);
1631 * Must be called under group lock!
1633 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1634 struct ext4_buddy *e4b)
1636 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1639 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1640 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1642 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1643 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1644 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1646 /* preallocation can change ac_b_ex, thus we store actually
1647 * allocated blocks for history */
1648 ac->ac_f_ex = ac->ac_b_ex;
1650 ac->ac_status = AC_STATUS_FOUND;
1651 ac->ac_tail = ret & 0xffff;
1652 ac->ac_buddy = ret >> 16;
1655 * take the page reference. We want the page to be pinned
1656 * so that we don't get a ext4_mb_init_cache_call for this
1657 * group until we update the bitmap. That would mean we
1658 * double allocate blocks. The reference is dropped
1659 * in ext4_mb_release_context
1661 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1662 get_page(ac->ac_bitmap_page);
1663 ac->ac_buddy_page = e4b->bd_buddy_page;
1664 get_page(ac->ac_buddy_page);
1665 /* store last allocated for subsequent stream allocation */
1666 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1667 spin_lock(&sbi->s_md_lock);
1668 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1669 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1670 spin_unlock(&sbi->s_md_lock);
1675 * regular allocator, for general purposes allocation
1678 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1679 struct ext4_buddy *e4b,
1682 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1683 struct ext4_free_extent *bex = &ac->ac_b_ex;
1684 struct ext4_free_extent *gex = &ac->ac_g_ex;
1685 struct ext4_free_extent ex;
1688 if (ac->ac_status == AC_STATUS_FOUND)
1691 * We don't want to scan for a whole year
1693 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1694 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1695 ac->ac_status = AC_STATUS_BREAK;
1700 * Haven't found good chunk so far, let's continue
1702 if (bex->fe_len < gex->fe_len)
1705 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1706 && bex->fe_group == e4b->bd_group) {
1707 /* recheck chunk's availability - we don't know
1708 * when it was found (within this lock-unlock
1710 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1711 if (max >= gex->fe_len) {
1712 ext4_mb_use_best_found(ac, e4b);
1719 * The routine checks whether found extent is good enough. If it is,
1720 * then the extent gets marked used and flag is set to the context
1721 * to stop scanning. Otherwise, the extent is compared with the
1722 * previous found extent and if new one is better, then it's stored
1723 * in the context. Later, the best found extent will be used, if
1724 * mballoc can't find good enough extent.
1726 * FIXME: real allocation policy is to be designed yet!
1728 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1729 struct ext4_free_extent *ex,
1730 struct ext4_buddy *e4b)
1732 struct ext4_free_extent *bex = &ac->ac_b_ex;
1733 struct ext4_free_extent *gex = &ac->ac_g_ex;
1735 BUG_ON(ex->fe_len <= 0);
1736 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1737 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1738 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1743 * The special case - take what you catch first
1745 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1747 ext4_mb_use_best_found(ac, e4b);
1752 * Let's check whether the chuck is good enough
1754 if (ex->fe_len == gex->fe_len) {
1756 ext4_mb_use_best_found(ac, e4b);
1761 * If this is first found extent, just store it in the context
1763 if (bex->fe_len == 0) {
1769 * If new found extent is better, store it in the context
1771 if (bex->fe_len < gex->fe_len) {
1772 /* if the request isn't satisfied, any found extent
1773 * larger than previous best one is better */
1774 if (ex->fe_len > bex->fe_len)
1776 } else if (ex->fe_len > gex->fe_len) {
1777 /* if the request is satisfied, then we try to find
1778 * an extent that still satisfy the request, but is
1779 * smaller than previous one */
1780 if (ex->fe_len < bex->fe_len)
1784 ext4_mb_check_limits(ac, e4b, 0);
1787 static noinline_for_stack
1788 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1789 struct ext4_buddy *e4b)
1791 struct ext4_free_extent ex = ac->ac_b_ex;
1792 ext4_group_t group = ex.fe_group;
1796 BUG_ON(ex.fe_len <= 0);
1797 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1801 ext4_lock_group(ac->ac_sb, group);
1802 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1806 ext4_mb_use_best_found(ac, e4b);
1809 ext4_unlock_group(ac->ac_sb, group);
1810 ext4_mb_unload_buddy(e4b);
1815 static noinline_for_stack
1816 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1817 struct ext4_buddy *e4b)
1819 ext4_group_t group = ac->ac_g_ex.fe_group;
1822 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1823 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1824 struct ext4_free_extent ex;
1826 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1828 if (grp->bb_free == 0)
1831 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1835 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1836 ext4_mb_unload_buddy(e4b);
1840 ext4_lock_group(ac->ac_sb, group);
1841 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1842 ac->ac_g_ex.fe_len, &ex);
1843 ex.fe_logical = 0xDEADFA11; /* debug value */
1845 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1848 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1850 /* use do_div to get remainder (would be 64-bit modulo) */
1851 if (do_div(start, sbi->s_stripe) == 0) {
1854 ext4_mb_use_best_found(ac, e4b);
1856 } else if (max >= ac->ac_g_ex.fe_len) {
1857 BUG_ON(ex.fe_len <= 0);
1858 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1859 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1862 ext4_mb_use_best_found(ac, e4b);
1863 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1864 /* Sometimes, caller may want to merge even small
1865 * number of blocks to an existing extent */
1866 BUG_ON(ex.fe_len <= 0);
1867 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1868 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1871 ext4_mb_use_best_found(ac, e4b);
1873 ext4_unlock_group(ac->ac_sb, group);
1874 ext4_mb_unload_buddy(e4b);
1880 * The routine scans buddy structures (not bitmap!) from given order
1881 * to max order and tries to find big enough chunk to satisfy the req
1883 static noinline_for_stack
1884 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1885 struct ext4_buddy *e4b)
1887 struct super_block *sb = ac->ac_sb;
1888 struct ext4_group_info *grp = e4b->bd_info;
1894 BUG_ON(ac->ac_2order <= 0);
1895 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1896 if (grp->bb_counters[i] == 0)
1899 buddy = mb_find_buddy(e4b, i, &max);
1900 BUG_ON(buddy == NULL);
1902 k = mb_find_next_zero_bit(buddy, max, 0);
1907 ac->ac_b_ex.fe_len = 1 << i;
1908 ac->ac_b_ex.fe_start = k << i;
1909 ac->ac_b_ex.fe_group = e4b->bd_group;
1911 ext4_mb_use_best_found(ac, e4b);
1913 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1915 if (EXT4_SB(sb)->s_mb_stats)
1916 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1923 * The routine scans the group and measures all found extents.
1924 * In order to optimize scanning, caller must pass number of
1925 * free blocks in the group, so the routine can know upper limit.
1927 static noinline_for_stack
1928 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1929 struct ext4_buddy *e4b)
1931 struct super_block *sb = ac->ac_sb;
1932 void *bitmap = e4b->bd_bitmap;
1933 struct ext4_free_extent ex;
1937 free = e4b->bd_info->bb_free;
1940 i = e4b->bd_info->bb_first_free;
1942 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1943 i = mb_find_next_zero_bit(bitmap,
1944 EXT4_CLUSTERS_PER_GROUP(sb), i);
1945 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1947 * IF we have corrupt bitmap, we won't find any
1948 * free blocks even though group info says we
1949 * we have free blocks
1951 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1952 "%d free clusters as per "
1953 "group info. But bitmap says 0",
1955 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1956 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1960 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1961 BUG_ON(ex.fe_len <= 0);
1962 if (free < ex.fe_len) {
1963 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1964 "%d free clusters as per "
1965 "group info. But got %d blocks",
1967 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1968 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1970 * The number of free blocks differs. This mostly
1971 * indicate that the bitmap is corrupt. So exit
1972 * without claiming the space.
1976 ex.fe_logical = 0xDEADC0DE; /* debug value */
1977 ext4_mb_measure_extent(ac, &ex, e4b);
1983 ext4_mb_check_limits(ac, e4b, 1);
1987 * This is a special case for storages like raid5
1988 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1990 static noinline_for_stack
1991 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1992 struct ext4_buddy *e4b)
1994 struct super_block *sb = ac->ac_sb;
1995 struct ext4_sb_info *sbi = EXT4_SB(sb);
1996 void *bitmap = e4b->bd_bitmap;
1997 struct ext4_free_extent ex;
1998 ext4_fsblk_t first_group_block;
2003 BUG_ON(sbi->s_stripe == 0);
2005 /* find first stripe-aligned block in group */
2006 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2008 a = first_group_block + sbi->s_stripe - 1;
2009 do_div(a, sbi->s_stripe);
2010 i = (a * sbi->s_stripe) - first_group_block;
2012 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2013 if (!mb_test_bit(i, bitmap)) {
2014 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2015 if (max >= sbi->s_stripe) {
2017 ex.fe_logical = 0xDEADF00D; /* debug value */
2019 ext4_mb_use_best_found(ac, e4b);
2028 * This is now called BEFORE we load the buddy bitmap.
2029 * Returns either 1 or 0 indicating that the group is either suitable
2030 * for the allocation or not. In addition it can also return negative
2031 * error code when something goes wrong.
2033 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2034 ext4_group_t group, int cr)
2036 unsigned free, fragments;
2037 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2038 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2040 BUG_ON(cr < 0 || cr >= 4);
2042 free = grp->bb_free;
2045 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2048 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2051 /* We only do this if the grp has never been initialized */
2052 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2053 int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS);
2058 fragments = grp->bb_fragments;
2064 BUG_ON(ac->ac_2order == 0);
2066 /* Avoid using the first bg of a flexgroup for data files */
2067 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2068 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2069 ((group % flex_size) == 0))
2072 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2073 (free / fragments) >= ac->ac_g_ex.fe_len)
2076 if (grp->bb_largest_free_order < ac->ac_2order)
2081 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2085 if (free >= ac->ac_g_ex.fe_len)
2097 static noinline_for_stack int
2098 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2100 ext4_group_t ngroups, group, i;
2102 int err = 0, first_err = 0;
2103 struct ext4_sb_info *sbi;
2104 struct super_block *sb;
2105 struct ext4_buddy e4b;
2109 ngroups = ext4_get_groups_count(sb);
2110 /* non-extent files are limited to low blocks/groups */
2111 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2112 ngroups = sbi->s_blockfile_groups;
2114 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2116 /* first, try the goal */
2117 err = ext4_mb_find_by_goal(ac, &e4b);
2118 if (err || ac->ac_status == AC_STATUS_FOUND)
2121 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2125 * ac->ac2_order is set only if the fe_len is a power of 2
2126 * if ac2_order is set we also set criteria to 0 so that we
2127 * try exact allocation using buddy.
2129 i = fls(ac->ac_g_ex.fe_len);
2132 * We search using buddy data only if the order of the request
2133 * is greater than equal to the sbi_s_mb_order2_reqs
2134 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2135 * We also support searching for power-of-two requests only for
2136 * requests upto maximum buddy size we have constructed.
2138 if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) {
2140 * This should tell if fe_len is exactly power of 2
2142 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2143 ac->ac_2order = i - 1;
2146 /* if stream allocation is enabled, use global goal */
2147 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2148 /* TBD: may be hot point */
2149 spin_lock(&sbi->s_md_lock);
2150 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2151 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2152 spin_unlock(&sbi->s_md_lock);
2155 /* Let's just scan groups to find more-less suitable blocks */
2156 cr = ac->ac_2order ? 0 : 1;
2158 * cr == 0 try to get exact allocation,
2159 * cr == 3 try to get anything
2162 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2163 ac->ac_criteria = cr;
2165 * searching for the right group start
2166 * from the goal value specified
2168 group = ac->ac_g_ex.fe_group;
2170 for (i = 0; i < ngroups; group++, i++) {
2174 * Artificially restricted ngroups for non-extent
2175 * files makes group > ngroups possible on first loop.
2177 if (group >= ngroups)
2180 /* This now checks without needing the buddy page */
2181 ret = ext4_mb_good_group(ac, group, cr);
2188 err = ext4_mb_load_buddy(sb, group, &e4b);
2192 ext4_lock_group(sb, group);
2195 * We need to check again after locking the
2198 ret = ext4_mb_good_group(ac, group, cr);
2200 ext4_unlock_group(sb, group);
2201 ext4_mb_unload_buddy(&e4b);
2207 ac->ac_groups_scanned++;
2209 ext4_mb_simple_scan_group(ac, &e4b);
2210 else if (cr == 1 && sbi->s_stripe &&
2211 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2212 ext4_mb_scan_aligned(ac, &e4b);
2214 ext4_mb_complex_scan_group(ac, &e4b);
2216 ext4_unlock_group(sb, group);
2217 ext4_mb_unload_buddy(&e4b);
2219 if (ac->ac_status != AC_STATUS_CONTINUE)
2224 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2225 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2227 * We've been searching too long. Let's try to allocate
2228 * the best chunk we've found so far
2231 ext4_mb_try_best_found(ac, &e4b);
2232 if (ac->ac_status != AC_STATUS_FOUND) {
2234 * Someone more lucky has already allocated it.
2235 * The only thing we can do is just take first
2237 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2239 ac->ac_b_ex.fe_group = 0;
2240 ac->ac_b_ex.fe_start = 0;
2241 ac->ac_b_ex.fe_len = 0;
2242 ac->ac_status = AC_STATUS_CONTINUE;
2243 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2245 atomic_inc(&sbi->s_mb_lost_chunks);
2250 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2255 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2257 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2260 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2263 return (void *) ((unsigned long) group);
2266 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2268 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2272 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2275 return (void *) ((unsigned long) group);
2278 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2280 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2281 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2283 int err, buddy_loaded = 0;
2284 struct ext4_buddy e4b;
2285 struct ext4_group_info *grinfo;
2286 unsigned char blocksize_bits = min_t(unsigned char,
2287 sb->s_blocksize_bits,
2288 EXT4_MAX_BLOCK_LOG_SIZE);
2290 struct ext4_group_info info;
2291 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2296 seq_puts(seq, "#group: free frags first ["
2297 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2298 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2300 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2301 sizeof(struct ext4_group_info);
2303 grinfo = ext4_get_group_info(sb, group);
2304 /* Load the group info in memory only if not already loaded. */
2305 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2306 err = ext4_mb_load_buddy(sb, group, &e4b);
2308 seq_printf(seq, "#%-5u: I/O error\n", group);
2314 memcpy(&sg, ext4_get_group_info(sb, group), i);
2317 ext4_mb_unload_buddy(&e4b);
2319 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2320 sg.info.bb_fragments, sg.info.bb_first_free);
2321 for (i = 0; i <= 13; i++)
2322 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2323 sg.info.bb_counters[i] : 0);
2324 seq_printf(seq, " ]\n");
2329 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2333 const struct seq_operations ext4_mb_seq_groups_ops = {
2334 .start = ext4_mb_seq_groups_start,
2335 .next = ext4_mb_seq_groups_next,
2336 .stop = ext4_mb_seq_groups_stop,
2337 .show = ext4_mb_seq_groups_show,
2340 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2342 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2343 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2350 * Allocate the top-level s_group_info array for the specified number
2353 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2355 struct ext4_sb_info *sbi = EXT4_SB(sb);
2357 struct ext4_group_info ***new_groupinfo;
2359 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2360 EXT4_DESC_PER_BLOCK_BITS(sb);
2361 if (size <= sbi->s_group_info_size)
2364 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2365 new_groupinfo = kvzalloc(size, GFP_KERNEL);
2366 if (!new_groupinfo) {
2367 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2370 if (sbi->s_group_info) {
2371 memcpy(new_groupinfo, sbi->s_group_info,
2372 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2373 kvfree(sbi->s_group_info);
2375 sbi->s_group_info = new_groupinfo;
2376 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2377 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2378 sbi->s_group_info_size);
2382 /* Create and initialize ext4_group_info data for the given group. */
2383 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2384 struct ext4_group_desc *desc)
2388 struct ext4_sb_info *sbi = EXT4_SB(sb);
2389 struct ext4_group_info **meta_group_info;
2390 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2393 * First check if this group is the first of a reserved block.
2394 * If it's true, we have to allocate a new table of pointers
2395 * to ext4_group_info structures
2397 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2398 metalen = sizeof(*meta_group_info) <<
2399 EXT4_DESC_PER_BLOCK_BITS(sb);
2400 meta_group_info = kmalloc(metalen, GFP_NOFS);
2401 if (meta_group_info == NULL) {
2402 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2403 "for a buddy group");
2404 goto exit_meta_group_info;
2406 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2411 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2412 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2414 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2415 if (meta_group_info[i] == NULL) {
2416 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2417 goto exit_group_info;
2419 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2420 &(meta_group_info[i]->bb_state));
2423 * initialize bb_free to be able to skip
2424 * empty groups without initialization
2426 if (ext4_has_group_desc_csum(sb) &&
2427 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
2428 meta_group_info[i]->bb_free =
2429 ext4_free_clusters_after_init(sb, group, desc);
2431 meta_group_info[i]->bb_free =
2432 ext4_free_group_clusters(sb, desc);
2435 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2436 init_rwsem(&meta_group_info[i]->alloc_sem);
2437 meta_group_info[i]->bb_free_root = RB_ROOT;
2438 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2442 struct buffer_head *bh;
2443 meta_group_info[i]->bb_bitmap =
2444 kmalloc(sb->s_blocksize, GFP_NOFS);
2445 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2446 bh = ext4_read_block_bitmap(sb, group);
2447 BUG_ON(IS_ERR_OR_NULL(bh));
2448 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2457 /* If a meta_group_info table has been allocated, release it now */
2458 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2459 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2460 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2462 exit_meta_group_info:
2464 } /* ext4_mb_add_groupinfo */
2466 static int ext4_mb_init_backend(struct super_block *sb)
2468 ext4_group_t ngroups = ext4_get_groups_count(sb);
2470 struct ext4_sb_info *sbi = EXT4_SB(sb);
2472 struct ext4_group_desc *desc;
2473 struct kmem_cache *cachep;
2475 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2479 sbi->s_buddy_cache = new_inode(sb);
2480 if (sbi->s_buddy_cache == NULL) {
2481 ext4_msg(sb, KERN_ERR, "can't get new inode");
2484 /* To avoid potentially colliding with an valid on-disk inode number,
2485 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2486 * not in the inode hash, so it should never be found by iget(), but
2487 * this will avoid confusion if it ever shows up during debugging. */
2488 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2489 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2490 for (i = 0; i < ngroups; i++) {
2491 desc = ext4_get_group_desc(sb, i, NULL);
2493 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2496 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2503 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2505 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2506 i = sbi->s_group_info_size;
2508 kfree(sbi->s_group_info[i]);
2509 iput(sbi->s_buddy_cache);
2511 kvfree(sbi->s_group_info);
2515 static void ext4_groupinfo_destroy_slabs(void)
2519 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2520 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2521 ext4_groupinfo_caches[i] = NULL;
2525 static int ext4_groupinfo_create_slab(size_t size)
2527 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2529 int blocksize_bits = order_base_2(size);
2530 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2531 struct kmem_cache *cachep;
2533 if (cache_index >= NR_GRPINFO_CACHES)
2536 if (unlikely(cache_index < 0))
2539 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2540 if (ext4_groupinfo_caches[cache_index]) {
2541 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2542 return 0; /* Already created */
2545 slab_size = offsetof(struct ext4_group_info,
2546 bb_counters[blocksize_bits + 2]);
2548 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2549 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2552 ext4_groupinfo_caches[cache_index] = cachep;
2554 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2557 "EXT4-fs: no memory for groupinfo slab cache\n");
2564 int ext4_mb_init(struct super_block *sb)
2566 struct ext4_sb_info *sbi = EXT4_SB(sb);
2568 unsigned offset, offset_incr;
2572 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2574 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2575 if (sbi->s_mb_offsets == NULL) {
2580 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2581 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2582 if (sbi->s_mb_maxs == NULL) {
2587 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2591 /* order 0 is regular bitmap */
2592 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2593 sbi->s_mb_offsets[0] = 0;
2597 offset_incr = 1 << (sb->s_blocksize_bits - 1);
2598 max = sb->s_blocksize << 2;
2600 sbi->s_mb_offsets[i] = offset;
2601 sbi->s_mb_maxs[i] = max;
2602 offset += offset_incr;
2603 offset_incr = offset_incr >> 1;
2606 } while (i <= sb->s_blocksize_bits + 1);
2608 spin_lock_init(&sbi->s_md_lock);
2609 spin_lock_init(&sbi->s_bal_lock);
2610 sbi->s_mb_free_pending = 0;
2611 INIT_LIST_HEAD(&sbi->s_freed_data_list);
2613 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2614 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2615 sbi->s_mb_stats = MB_DEFAULT_STATS;
2616 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2617 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2619 * The default group preallocation is 512, which for 4k block
2620 * sizes translates to 2 megabytes. However for bigalloc file
2621 * systems, this is probably too big (i.e, if the cluster size
2622 * is 1 megabyte, then group preallocation size becomes half a
2623 * gigabyte!). As a default, we will keep a two megabyte
2624 * group pralloc size for cluster sizes up to 64k, and after
2625 * that, we will force a minimum group preallocation size of
2626 * 32 clusters. This translates to 8 megs when the cluster
2627 * size is 256k, and 32 megs when the cluster size is 1 meg,
2628 * which seems reasonable as a default.
2630 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2631 sbi->s_cluster_bits, 32);
2633 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2634 * to the lowest multiple of s_stripe which is bigger than
2635 * the s_mb_group_prealloc as determined above. We want
2636 * the preallocation size to be an exact multiple of the
2637 * RAID stripe size so that preallocations don't fragment
2640 if (sbi->s_stripe > 1) {
2641 sbi->s_mb_group_prealloc = roundup(
2642 sbi->s_mb_group_prealloc, sbi->s_stripe);
2645 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2646 if (sbi->s_locality_groups == NULL) {
2650 for_each_possible_cpu(i) {
2651 struct ext4_locality_group *lg;
2652 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2653 mutex_init(&lg->lg_mutex);
2654 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2655 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2656 spin_lock_init(&lg->lg_prealloc_lock);
2659 /* init file for buddy data */
2660 ret = ext4_mb_init_backend(sb);
2662 goto out_free_locality_groups;
2666 out_free_locality_groups:
2667 free_percpu(sbi->s_locality_groups);
2668 sbi->s_locality_groups = NULL;
2670 kfree(sbi->s_mb_offsets);
2671 sbi->s_mb_offsets = NULL;
2672 kfree(sbi->s_mb_maxs);
2673 sbi->s_mb_maxs = NULL;
2677 /* need to called with the ext4 group lock held */
2678 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2680 struct ext4_prealloc_space *pa;
2681 struct list_head *cur, *tmp;
2684 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2685 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2686 list_del(&pa->pa_group_list);
2688 kmem_cache_free(ext4_pspace_cachep, pa);
2691 mb_debug(1, "mballoc: %u PAs left\n", count);
2695 int ext4_mb_release(struct super_block *sb)
2697 ext4_group_t ngroups = ext4_get_groups_count(sb);
2699 int num_meta_group_infos;
2700 struct ext4_group_info *grinfo;
2701 struct ext4_sb_info *sbi = EXT4_SB(sb);
2702 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2704 if (sbi->s_group_info) {
2705 for (i = 0; i < ngroups; i++) {
2706 grinfo = ext4_get_group_info(sb, i);
2708 kfree(grinfo->bb_bitmap);
2710 ext4_lock_group(sb, i);
2711 ext4_mb_cleanup_pa(grinfo);
2712 ext4_unlock_group(sb, i);
2713 kmem_cache_free(cachep, grinfo);
2715 num_meta_group_infos = (ngroups +
2716 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2717 EXT4_DESC_PER_BLOCK_BITS(sb);
2718 for (i = 0; i < num_meta_group_infos; i++)
2719 kfree(sbi->s_group_info[i]);
2720 kvfree(sbi->s_group_info);
2722 kfree(sbi->s_mb_offsets);
2723 kfree(sbi->s_mb_maxs);
2724 iput(sbi->s_buddy_cache);
2725 if (sbi->s_mb_stats) {
2726 ext4_msg(sb, KERN_INFO,
2727 "mballoc: %u blocks %u reqs (%u success)",
2728 atomic_read(&sbi->s_bal_allocated),
2729 atomic_read(&sbi->s_bal_reqs),
2730 atomic_read(&sbi->s_bal_success));
2731 ext4_msg(sb, KERN_INFO,
2732 "mballoc: %u extents scanned, %u goal hits, "
2733 "%u 2^N hits, %u breaks, %u lost",
2734 atomic_read(&sbi->s_bal_ex_scanned),
2735 atomic_read(&sbi->s_bal_goals),
2736 atomic_read(&sbi->s_bal_2orders),
2737 atomic_read(&sbi->s_bal_breaks),
2738 atomic_read(&sbi->s_mb_lost_chunks));
2739 ext4_msg(sb, KERN_INFO,
2740 "mballoc: %lu generated and it took %Lu",
2741 sbi->s_mb_buddies_generated,
2742 sbi->s_mb_generation_time);
2743 ext4_msg(sb, KERN_INFO,
2744 "mballoc: %u preallocated, %u discarded",
2745 atomic_read(&sbi->s_mb_preallocated),
2746 atomic_read(&sbi->s_mb_discarded));
2749 free_percpu(sbi->s_locality_groups);
2754 static inline int ext4_issue_discard(struct super_block *sb,
2755 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
2758 ext4_fsblk_t discard_block;
2760 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2761 ext4_group_first_block_no(sb, block_group));
2762 count = EXT4_C2B(EXT4_SB(sb), count);
2763 trace_ext4_discard_blocks(sb,
2764 (unsigned long long) discard_block, count);
2766 return __blkdev_issue_discard(sb->s_bdev,
2767 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
2768 (sector_t)count << (sb->s_blocksize_bits - 9),
2771 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2774 static void ext4_free_data_in_buddy(struct super_block *sb,
2775 struct ext4_free_data *entry)
2777 struct ext4_buddy e4b;
2778 struct ext4_group_info *db;
2779 int err, count = 0, count2 = 0;
2781 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2782 entry->efd_count, entry->efd_group, entry);
2784 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2785 /* we expect to find existing buddy because it's pinned */
2788 spin_lock(&EXT4_SB(sb)->s_md_lock);
2789 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
2790 spin_unlock(&EXT4_SB(sb)->s_md_lock);
2793 /* there are blocks to put in buddy to make them really free */
2794 count += entry->efd_count;
2796 ext4_lock_group(sb, entry->efd_group);
2797 /* Take it out of per group rb tree */
2798 rb_erase(&entry->efd_node, &(db->bb_free_root));
2799 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2802 * Clear the trimmed flag for the group so that the next
2803 * ext4_trim_fs can trim it.
2804 * If the volume is mounted with -o discard, online discard
2805 * is supported and the free blocks will be trimmed online.
2807 if (!test_opt(sb, DISCARD))
2808 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2810 if (!db->bb_free_root.rb_node) {
2811 /* No more items in the per group rb tree
2812 * balance refcounts from ext4_mb_free_metadata()
2814 put_page(e4b.bd_buddy_page);
2815 put_page(e4b.bd_bitmap_page);
2817 ext4_unlock_group(sb, entry->efd_group);
2818 kmem_cache_free(ext4_free_data_cachep, entry);
2819 ext4_mb_unload_buddy(&e4b);
2821 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2825 * This function is called by the jbd2 layer once the commit has finished,
2826 * so we know we can free the blocks that were released with that commit.
2828 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
2830 struct ext4_sb_info *sbi = EXT4_SB(sb);
2831 struct ext4_free_data *entry, *tmp;
2832 struct bio *discard_bio = NULL;
2833 struct list_head freed_data_list;
2834 struct list_head *cut_pos = NULL;
2837 INIT_LIST_HEAD(&freed_data_list);
2839 spin_lock(&sbi->s_md_lock);
2840 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
2841 if (entry->efd_tid != commit_tid)
2843 cut_pos = &entry->efd_list;
2846 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
2848 spin_unlock(&sbi->s_md_lock);
2850 if (test_opt(sb, DISCARD)) {
2851 list_for_each_entry(entry, &freed_data_list, efd_list) {
2852 err = ext4_issue_discard(sb, entry->efd_group,
2853 entry->efd_start_cluster,
2856 if (err && err != -EOPNOTSUPP) {
2857 ext4_msg(sb, KERN_WARNING, "discard request in"
2858 " group:%d block:%d count:%d failed"
2859 " with %d", entry->efd_group,
2860 entry->efd_start_cluster,
2861 entry->efd_count, err);
2862 } else if (err == -EOPNOTSUPP)
2867 submit_bio_wait(discard_bio);
2868 bio_put(discard_bio);
2872 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
2873 ext4_free_data_in_buddy(sb, entry);
2876 int __init ext4_init_mballoc(void)
2878 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2879 SLAB_RECLAIM_ACCOUNT);
2880 if (ext4_pspace_cachep == NULL)
2883 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2884 SLAB_RECLAIM_ACCOUNT);
2885 if (ext4_ac_cachep == NULL) {
2886 kmem_cache_destroy(ext4_pspace_cachep);
2890 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2891 SLAB_RECLAIM_ACCOUNT);
2892 if (ext4_free_data_cachep == NULL) {
2893 kmem_cache_destroy(ext4_pspace_cachep);
2894 kmem_cache_destroy(ext4_ac_cachep);
2900 void ext4_exit_mballoc(void)
2903 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2904 * before destroying the slab cache.
2907 kmem_cache_destroy(ext4_pspace_cachep);
2908 kmem_cache_destroy(ext4_ac_cachep);
2909 kmem_cache_destroy(ext4_free_data_cachep);
2910 ext4_groupinfo_destroy_slabs();
2915 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2916 * Returns 0 if success or error code
2918 static noinline_for_stack int
2919 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2920 handle_t *handle, unsigned int reserv_clstrs)
2922 struct buffer_head *bitmap_bh = NULL;
2923 struct ext4_group_desc *gdp;
2924 struct buffer_head *gdp_bh;
2925 struct ext4_sb_info *sbi;
2926 struct super_block *sb;
2930 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2931 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2936 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2937 if (IS_ERR(bitmap_bh)) {
2938 err = PTR_ERR(bitmap_bh);
2943 BUFFER_TRACE(bitmap_bh, "getting write access");
2944 err = ext4_journal_get_write_access(handle, bitmap_bh);
2949 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2953 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2954 ext4_free_group_clusters(sb, gdp));
2956 BUFFER_TRACE(gdp_bh, "get_write_access");
2957 err = ext4_journal_get_write_access(handle, gdp_bh);
2961 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2963 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2964 if (!ext4_data_block_valid(sbi, block, len)) {
2965 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2966 "fs metadata", block, block+len);
2967 /* File system mounted not to panic on error
2968 * Fix the bitmap and return EFSCORRUPTED
2969 * We leak some of the blocks here.
2971 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2972 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2973 ac->ac_b_ex.fe_len);
2974 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2975 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2977 err = -EFSCORRUPTED;
2981 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2982 #ifdef AGGRESSIVE_CHECK
2985 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2986 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2987 bitmap_bh->b_data));
2991 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2992 ac->ac_b_ex.fe_len);
2993 if (ext4_has_group_desc_csum(sb) &&
2994 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
2995 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2996 ext4_free_group_clusters_set(sb, gdp,
2997 ext4_free_clusters_after_init(sb,
2998 ac->ac_b_ex.fe_group, gdp));
3000 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3001 ext4_free_group_clusters_set(sb, gdp, len);
3002 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3003 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3005 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3006 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3008 * Now reduce the dirty block count also. Should not go negative
3010 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3011 /* release all the reserved blocks if non delalloc */
3012 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3015 if (sbi->s_log_groups_per_flex) {
3016 ext4_group_t flex_group = ext4_flex_group(sbi,
3017 ac->ac_b_ex.fe_group);
3018 atomic64_sub(ac->ac_b_ex.fe_len,
3019 &sbi->s_flex_groups[flex_group].free_clusters);
3022 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3025 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3033 * here we normalize request for locality group
3034 * Group request are normalized to s_mb_group_prealloc, which goes to
3035 * s_strip if we set the same via mount option.
3036 * s_mb_group_prealloc can be configured via
3037 * /sys/fs/ext4/<partition>/mb_group_prealloc
3039 * XXX: should we try to preallocate more than the group has now?
3041 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3043 struct super_block *sb = ac->ac_sb;
3044 struct ext4_locality_group *lg = ac->ac_lg;
3047 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3048 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3049 current->pid, ac->ac_g_ex.fe_len);
3053 * Normalization means making request better in terms of
3054 * size and alignment
3056 static noinline_for_stack void
3057 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3058 struct ext4_allocation_request *ar)
3060 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3063 loff_t size, start_off;
3064 loff_t orig_size __maybe_unused;
3066 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3067 struct ext4_prealloc_space *pa;
3069 /* do normalize only data requests, metadata requests
3070 do not need preallocation */
3071 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3074 /* sometime caller may want exact blocks */
3075 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3078 /* caller may indicate that preallocation isn't
3079 * required (it's a tail, for example) */
3080 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3083 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3084 ext4_mb_normalize_group_request(ac);
3088 bsbits = ac->ac_sb->s_blocksize_bits;
3090 /* first, let's learn actual file size
3091 * given current request is allocated */
3092 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3093 size = size << bsbits;
3094 if (size < i_size_read(ac->ac_inode))
3095 size = i_size_read(ac->ac_inode);
3098 /* max size of free chunks */
3101 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3102 (req <= (size) || max <= (chunk_size))
3104 /* first, try to predict filesize */
3105 /* XXX: should this table be tunable? */
3107 if (size <= 16 * 1024) {
3109 } else if (size <= 32 * 1024) {
3111 } else if (size <= 64 * 1024) {
3113 } else if (size <= 128 * 1024) {
3115 } else if (size <= 256 * 1024) {
3117 } else if (size <= 512 * 1024) {
3119 } else if (size <= 1024 * 1024) {
3121 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3122 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3123 (21 - bsbits)) << 21;
3124 size = 2 * 1024 * 1024;
3125 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3126 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3127 (22 - bsbits)) << 22;
3128 size = 4 * 1024 * 1024;
3129 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3130 (8<<20)>>bsbits, max, 8 * 1024)) {
3131 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3132 (23 - bsbits)) << 23;
3133 size = 8 * 1024 * 1024;
3135 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3136 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3137 ac->ac_o_ex.fe_len) << bsbits;
3139 size = size >> bsbits;
3140 start = start_off >> bsbits;
3142 /* don't cover already allocated blocks in selected range */
3143 if (ar->pleft && start <= ar->lleft) {
3144 size -= ar->lleft + 1 - start;
3145 start = ar->lleft + 1;
3147 if (ar->pright && start + size - 1 >= ar->lright)
3148 size -= start + size - ar->lright;
3151 * Trim allocation request for filesystems with artificially small
3154 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
3155 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
3159 /* check we don't cross already preallocated blocks */
3161 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3166 spin_lock(&pa->pa_lock);
3167 if (pa->pa_deleted) {
3168 spin_unlock(&pa->pa_lock);
3172 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3175 /* PA must not overlap original request */
3176 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3177 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3179 /* skip PAs this normalized request doesn't overlap with */
3180 if (pa->pa_lstart >= end || pa_end <= start) {
3181 spin_unlock(&pa->pa_lock);
3184 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3186 /* adjust start or end to be adjacent to this pa */
3187 if (pa_end <= ac->ac_o_ex.fe_logical) {
3188 BUG_ON(pa_end < start);
3190 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3191 BUG_ON(pa->pa_lstart > end);
3192 end = pa->pa_lstart;
3194 spin_unlock(&pa->pa_lock);
3199 /* XXX: extra loop to check we really don't overlap preallocations */
3201 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3204 spin_lock(&pa->pa_lock);
3205 if (pa->pa_deleted == 0) {
3206 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3208 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3210 spin_unlock(&pa->pa_lock);
3214 if (start + size <= ac->ac_o_ex.fe_logical &&
3215 start > ac->ac_o_ex.fe_logical) {
3216 ext4_msg(ac->ac_sb, KERN_ERR,
3217 "start %lu, size %lu, fe_logical %lu",
3218 (unsigned long) start, (unsigned long) size,
3219 (unsigned long) ac->ac_o_ex.fe_logical);
3222 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3224 /* now prepare goal request */
3226 /* XXX: is it better to align blocks WRT to logical
3227 * placement or satisfy big request as is */
3228 ac->ac_g_ex.fe_logical = start;
3229 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3231 /* define goal start in order to merge */
3232 if (ar->pright && (ar->lright == (start + size))) {
3233 /* merge to the right */
3234 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3235 &ac->ac_f_ex.fe_group,
3236 &ac->ac_f_ex.fe_start);
3237 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3239 if (ar->pleft && (ar->lleft + 1 == start)) {
3240 /* merge to the left */
3241 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3242 &ac->ac_f_ex.fe_group,
3243 &ac->ac_f_ex.fe_start);
3244 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3247 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3248 (unsigned) orig_size, (unsigned) start);
3251 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3253 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3255 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3256 atomic_inc(&sbi->s_bal_reqs);
3257 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3258 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3259 atomic_inc(&sbi->s_bal_success);
3260 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3261 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3262 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3263 atomic_inc(&sbi->s_bal_goals);
3264 if (ac->ac_found > sbi->s_mb_max_to_scan)
3265 atomic_inc(&sbi->s_bal_breaks);
3268 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3269 trace_ext4_mballoc_alloc(ac);
3271 trace_ext4_mballoc_prealloc(ac);
3275 * Called on failure; free up any blocks from the inode PA for this
3276 * context. We don't need this for MB_GROUP_PA because we only change
3277 * pa_free in ext4_mb_release_context(), but on failure, we've already
3278 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3280 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3282 struct ext4_prealloc_space *pa = ac->ac_pa;
3283 struct ext4_buddy e4b;
3287 if (ac->ac_f_ex.fe_len == 0)
3289 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3292 * This should never happen since we pin the
3293 * pages in the ext4_allocation_context so
3294 * ext4_mb_load_buddy() should never fail.
3296 WARN(1, "mb_load_buddy failed (%d)", err);
3299 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3300 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3301 ac->ac_f_ex.fe_len);
3302 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3303 ext4_mb_unload_buddy(&e4b);
3306 if (pa->pa_type == MB_INODE_PA)
3307 pa->pa_free += ac->ac_b_ex.fe_len;
3311 * use blocks preallocated to inode
3313 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3314 struct ext4_prealloc_space *pa)
3316 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3321 /* found preallocated blocks, use them */
3322 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3323 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3324 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3325 len = EXT4_NUM_B2C(sbi, end - start);
3326 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3327 &ac->ac_b_ex.fe_start);
3328 ac->ac_b_ex.fe_len = len;
3329 ac->ac_status = AC_STATUS_FOUND;
3332 BUG_ON(start < pa->pa_pstart);
3333 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3334 BUG_ON(pa->pa_free < len);
3337 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3341 * use blocks preallocated to locality group
3343 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3344 struct ext4_prealloc_space *pa)
3346 unsigned int len = ac->ac_o_ex.fe_len;
3348 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3349 &ac->ac_b_ex.fe_group,
3350 &ac->ac_b_ex.fe_start);
3351 ac->ac_b_ex.fe_len = len;
3352 ac->ac_status = AC_STATUS_FOUND;
3355 /* we don't correct pa_pstart or pa_plen here to avoid
3356 * possible race when the group is being loaded concurrently
3357 * instead we correct pa later, after blocks are marked
3358 * in on-disk bitmap -- see ext4_mb_release_context()
3359 * Other CPUs are prevented from allocating from this pa by lg_mutex
3361 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3365 * Return the prealloc space that have minimal distance
3366 * from the goal block. @cpa is the prealloc
3367 * space that is having currently known minimal distance
3368 * from the goal block.
3370 static struct ext4_prealloc_space *
3371 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3372 struct ext4_prealloc_space *pa,
3373 struct ext4_prealloc_space *cpa)
3375 ext4_fsblk_t cur_distance, new_distance;
3378 atomic_inc(&pa->pa_count);
3381 cur_distance = abs(goal_block - cpa->pa_pstart);
3382 new_distance = abs(goal_block - pa->pa_pstart);
3384 if (cur_distance <= new_distance)
3387 /* drop the previous reference */
3388 atomic_dec(&cpa->pa_count);
3389 atomic_inc(&pa->pa_count);
3394 * search goal blocks in preallocated space
3396 static noinline_for_stack int
3397 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3399 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3401 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3402 struct ext4_locality_group *lg;
3403 struct ext4_prealloc_space *pa, *cpa = NULL;
3404 ext4_fsblk_t goal_block;
3406 /* only data can be preallocated */
3407 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3410 /* first, try per-file preallocation */
3412 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3414 /* all fields in this condition don't change,
3415 * so we can skip locking for them */
3416 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3417 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3418 EXT4_C2B(sbi, pa->pa_len)))
3421 /* non-extent files can't have physical blocks past 2^32 */
3422 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3423 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3424 EXT4_MAX_BLOCK_FILE_PHYS))
3427 /* found preallocated blocks, use them */
3428 spin_lock(&pa->pa_lock);
3429 if (pa->pa_deleted == 0 && pa->pa_free) {
3430 atomic_inc(&pa->pa_count);
3431 ext4_mb_use_inode_pa(ac, pa);
3432 spin_unlock(&pa->pa_lock);
3433 ac->ac_criteria = 10;
3437 spin_unlock(&pa->pa_lock);
3441 /* can we use group allocation? */
3442 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3445 /* inode may have no locality group for some reason */
3449 order = fls(ac->ac_o_ex.fe_len) - 1;
3450 if (order > PREALLOC_TB_SIZE - 1)
3451 /* The max size of hash table is PREALLOC_TB_SIZE */
3452 order = PREALLOC_TB_SIZE - 1;
3454 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3456 * search for the prealloc space that is having
3457 * minimal distance from the goal block.
3459 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3461 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3463 spin_lock(&pa->pa_lock);
3464 if (pa->pa_deleted == 0 &&
3465 pa->pa_free >= ac->ac_o_ex.fe_len) {
3467 cpa = ext4_mb_check_group_pa(goal_block,
3470 spin_unlock(&pa->pa_lock);
3475 ext4_mb_use_group_pa(ac, cpa);
3476 ac->ac_criteria = 20;
3483 * the function goes through all block freed in the group
3484 * but not yet committed and marks them used in in-core bitmap.
3485 * buddy must be generated from this bitmap
3486 * Need to be called with the ext4 group lock held
3488 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3492 struct ext4_group_info *grp;
3493 struct ext4_free_data *entry;
3495 grp = ext4_get_group_info(sb, group);
3496 n = rb_first(&(grp->bb_free_root));
3499 entry = rb_entry(n, struct ext4_free_data, efd_node);
3500 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3507 * the function goes through all preallocation in this group and marks them
3508 * used in in-core bitmap. buddy must be generated from this bitmap
3509 * Need to be called with ext4 group lock held
3511 static noinline_for_stack
3512 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3515 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3516 struct ext4_prealloc_space *pa;
3517 struct list_head *cur;
3518 ext4_group_t groupnr;
3519 ext4_grpblk_t start;
3520 int preallocated = 0;
3523 /* all form of preallocation discards first load group,
3524 * so the only competing code is preallocation use.
3525 * we don't need any locking here
3526 * notice we do NOT ignore preallocations with pa_deleted
3527 * otherwise we could leave used blocks available for
3528 * allocation in buddy when concurrent ext4_mb_put_pa()
3529 * is dropping preallocation
3531 list_for_each(cur, &grp->bb_prealloc_list) {
3532 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3533 spin_lock(&pa->pa_lock);
3534 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3537 spin_unlock(&pa->pa_lock);
3538 if (unlikely(len == 0))
3540 BUG_ON(groupnr != group);
3541 ext4_set_bits(bitmap, start, len);
3542 preallocated += len;
3544 mb_debug(1, "preallocated %u for group %u\n", preallocated, group);
3547 static void ext4_mb_pa_callback(struct rcu_head *head)
3549 struct ext4_prealloc_space *pa;
3550 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3552 BUG_ON(atomic_read(&pa->pa_count));
3553 BUG_ON(pa->pa_deleted == 0);
3554 kmem_cache_free(ext4_pspace_cachep, pa);
3558 * drops a reference to preallocated space descriptor
3559 * if this was the last reference and the space is consumed
3561 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3562 struct super_block *sb, struct ext4_prealloc_space *pa)
3565 ext4_fsblk_t grp_blk;
3567 /* in this short window concurrent discard can set pa_deleted */
3568 spin_lock(&pa->pa_lock);
3569 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3570 spin_unlock(&pa->pa_lock);
3574 if (pa->pa_deleted == 1) {
3575 spin_unlock(&pa->pa_lock);
3580 spin_unlock(&pa->pa_lock);
3582 grp_blk = pa->pa_pstart;
3584 * If doing group-based preallocation, pa_pstart may be in the
3585 * next group when pa is used up
3587 if (pa->pa_type == MB_GROUP_PA)
3590 grp = ext4_get_group_number(sb, grp_blk);
3595 * P1 (buddy init) P2 (regular allocation)
3596 * find block B in PA
3597 * copy on-disk bitmap to buddy
3598 * mark B in on-disk bitmap
3599 * drop PA from group
3600 * mark all PAs in buddy
3602 * thus, P1 initializes buddy with B available. to prevent this
3603 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3606 ext4_lock_group(sb, grp);
3607 list_del(&pa->pa_group_list);
3608 ext4_unlock_group(sb, grp);
3610 spin_lock(pa->pa_obj_lock);
3611 list_del_rcu(&pa->pa_inode_list);
3612 spin_unlock(pa->pa_obj_lock);
3614 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3618 * creates new preallocated space for given inode
3620 static noinline_for_stack int
3621 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3623 struct super_block *sb = ac->ac_sb;
3624 struct ext4_sb_info *sbi = EXT4_SB(sb);
3625 struct ext4_prealloc_space *pa;
3626 struct ext4_group_info *grp;
3627 struct ext4_inode_info *ei;
3629 /* preallocate only when found space is larger then requested */
3630 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3631 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3632 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3634 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3638 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3644 /* we can't allocate as much as normalizer wants.
3645 * so, found space must get proper lstart
3646 * to cover original request */
3647 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3648 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3650 /* we're limited by original request in that
3651 * logical block must be covered any way
3652 * winl is window we can move our chunk within */
3653 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3655 /* also, we should cover whole original request */
3656 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3658 /* the smallest one defines real window */
3659 win = min(winl, wins);
3661 offs = ac->ac_o_ex.fe_logical %
3662 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3663 if (offs && offs < win)
3666 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3667 EXT4_NUM_B2C(sbi, win);
3668 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3669 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3672 /* preallocation can change ac_b_ex, thus we store actually
3673 * allocated blocks for history */
3674 ac->ac_f_ex = ac->ac_b_ex;
3676 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3677 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3678 pa->pa_len = ac->ac_b_ex.fe_len;
3679 pa->pa_free = pa->pa_len;
3680 atomic_set(&pa->pa_count, 1);
3681 spin_lock_init(&pa->pa_lock);
3682 INIT_LIST_HEAD(&pa->pa_inode_list);
3683 INIT_LIST_HEAD(&pa->pa_group_list);
3685 pa->pa_type = MB_INODE_PA;
3687 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3688 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3689 trace_ext4_mb_new_inode_pa(ac, pa);
3691 ext4_mb_use_inode_pa(ac, pa);
3692 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3694 ei = EXT4_I(ac->ac_inode);
3695 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3697 pa->pa_obj_lock = &ei->i_prealloc_lock;
3698 pa->pa_inode = ac->ac_inode;
3700 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3701 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3702 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3704 spin_lock(pa->pa_obj_lock);
3705 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3706 spin_unlock(pa->pa_obj_lock);
3712 * creates new preallocated space for locality group inodes belongs to
3714 static noinline_for_stack int
3715 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3717 struct super_block *sb = ac->ac_sb;
3718 struct ext4_locality_group *lg;
3719 struct ext4_prealloc_space *pa;
3720 struct ext4_group_info *grp;
3722 /* preallocate only when found space is larger then requested */
3723 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3724 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3725 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3727 BUG_ON(ext4_pspace_cachep == NULL);
3728 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3732 /* preallocation can change ac_b_ex, thus we store actually
3733 * allocated blocks for history */
3734 ac->ac_f_ex = ac->ac_b_ex;
3736 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3737 pa->pa_lstart = pa->pa_pstart;
3738 pa->pa_len = ac->ac_b_ex.fe_len;
3739 pa->pa_free = pa->pa_len;
3740 atomic_set(&pa->pa_count, 1);
3741 spin_lock_init(&pa->pa_lock);
3742 INIT_LIST_HEAD(&pa->pa_inode_list);
3743 INIT_LIST_HEAD(&pa->pa_group_list);
3745 pa->pa_type = MB_GROUP_PA;
3747 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3748 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3749 trace_ext4_mb_new_group_pa(ac, pa);
3751 ext4_mb_use_group_pa(ac, pa);
3752 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3754 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3758 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3759 pa->pa_inode = NULL;
3761 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3762 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3763 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3766 * We will later add the new pa to the right bucket
3767 * after updating the pa_free in ext4_mb_release_context
3772 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3776 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3777 err = ext4_mb_new_group_pa(ac);
3779 err = ext4_mb_new_inode_pa(ac);
3784 * finds all unused blocks in on-disk bitmap, frees them in
3785 * in-core bitmap and buddy.
3786 * @pa must be unlinked from inode and group lists, so that
3787 * nobody else can find/use it.
3788 * the caller MUST hold group/inode locks.
3789 * TODO: optimize the case when there are no in-core structures yet
3791 static noinline_for_stack int
3792 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3793 struct ext4_prealloc_space *pa)
3795 struct super_block *sb = e4b->bd_sb;
3796 struct ext4_sb_info *sbi = EXT4_SB(sb);
3801 unsigned long long grp_blk_start;
3805 BUG_ON(pa->pa_deleted == 0);
3806 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3807 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3808 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3809 end = bit + pa->pa_len;
3812 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3815 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3816 mb_debug(1, " free preallocated %u/%u in group %u\n",
3817 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3818 (unsigned) next - bit, (unsigned) group);
3821 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3822 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3823 EXT4_C2B(sbi, bit)),
3825 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3828 if (free != pa->pa_free) {
3829 ext4_msg(e4b->bd_sb, KERN_CRIT,
3830 "pa %p: logic %lu, phys. %lu, len %lu",
3831 pa, (unsigned long) pa->pa_lstart,
3832 (unsigned long) pa->pa_pstart,
3833 (unsigned long) pa->pa_len);
3834 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3837 * pa is already deleted so we use the value obtained
3838 * from the bitmap and continue.
3841 atomic_add(free, &sbi->s_mb_discarded);
3846 static noinline_for_stack int
3847 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3848 struct ext4_prealloc_space *pa)
3850 struct super_block *sb = e4b->bd_sb;
3854 trace_ext4_mb_release_group_pa(sb, pa);
3855 BUG_ON(pa->pa_deleted == 0);
3856 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3857 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3858 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3859 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3860 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3866 * releases all preallocations in given group
3868 * first, we need to decide discard policy:
3869 * - when do we discard
3871 * - how many do we discard
3872 * 1) how many requested
3874 static noinline_for_stack int
3875 ext4_mb_discard_group_preallocations(struct super_block *sb,
3876 ext4_group_t group, int needed)
3878 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3879 struct buffer_head *bitmap_bh = NULL;
3880 struct ext4_prealloc_space *pa, *tmp;
3881 struct list_head list;
3882 struct ext4_buddy e4b;
3887 mb_debug(1, "discard preallocation for group %u\n", group);
3889 if (list_empty(&grp->bb_prealloc_list))
3892 bitmap_bh = ext4_read_block_bitmap(sb, group);
3893 if (IS_ERR(bitmap_bh)) {
3894 err = PTR_ERR(bitmap_bh);
3895 ext4_error(sb, "Error %d reading block bitmap for %u",
3900 err = ext4_mb_load_buddy(sb, group, &e4b);
3902 ext4_warning(sb, "Error %d loading buddy information for %u",
3909 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3911 INIT_LIST_HEAD(&list);
3913 ext4_lock_group(sb, group);
3914 list_for_each_entry_safe(pa, tmp,
3915 &grp->bb_prealloc_list, pa_group_list) {
3916 spin_lock(&pa->pa_lock);
3917 if (atomic_read(&pa->pa_count)) {
3918 spin_unlock(&pa->pa_lock);
3922 if (pa->pa_deleted) {
3923 spin_unlock(&pa->pa_lock);
3927 /* seems this one can be freed ... */
3930 /* we can trust pa_free ... */
3931 free += pa->pa_free;
3933 spin_unlock(&pa->pa_lock);
3935 list_del(&pa->pa_group_list);
3936 list_add(&pa->u.pa_tmp_list, &list);
3939 /* if we still need more blocks and some PAs were used, try again */
3940 if (free < needed && busy) {
3942 ext4_unlock_group(sb, group);
3947 /* found anything to free? */
3948 if (list_empty(&list)) {
3953 /* now free all selected PAs */
3954 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3956 /* remove from object (inode or locality group) */
3957 spin_lock(pa->pa_obj_lock);
3958 list_del_rcu(&pa->pa_inode_list);
3959 spin_unlock(pa->pa_obj_lock);
3961 if (pa->pa_type == MB_GROUP_PA)
3962 ext4_mb_release_group_pa(&e4b, pa);
3964 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3966 list_del(&pa->u.pa_tmp_list);
3967 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3971 ext4_unlock_group(sb, group);
3972 ext4_mb_unload_buddy(&e4b);
3978 * releases all non-used preallocated blocks for given inode
3980 * It's important to discard preallocations under i_data_sem
3981 * We don't want another block to be served from the prealloc
3982 * space when we are discarding the inode prealloc space.
3984 * FIXME!! Make sure it is valid at all the call sites
3986 void ext4_discard_preallocations(struct inode *inode)
3988 struct ext4_inode_info *ei = EXT4_I(inode);
3989 struct super_block *sb = inode->i_sb;
3990 struct buffer_head *bitmap_bh = NULL;
3991 struct ext4_prealloc_space *pa, *tmp;
3992 ext4_group_t group = 0;
3993 struct list_head list;
3994 struct ext4_buddy e4b;
3997 if (!S_ISREG(inode->i_mode)) {
3998 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4002 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
4003 trace_ext4_discard_preallocations(inode);
4005 INIT_LIST_HEAD(&list);
4008 /* first, collect all pa's in the inode */
4009 spin_lock(&ei->i_prealloc_lock);
4010 while (!list_empty(&ei->i_prealloc_list)) {
4011 pa = list_entry(ei->i_prealloc_list.next,
4012 struct ext4_prealloc_space, pa_inode_list);
4013 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4014 spin_lock(&pa->pa_lock);
4015 if (atomic_read(&pa->pa_count)) {
4016 /* this shouldn't happen often - nobody should
4017 * use preallocation while we're discarding it */
4018 spin_unlock(&pa->pa_lock);
4019 spin_unlock(&ei->i_prealloc_lock);
4020 ext4_msg(sb, KERN_ERR,
4021 "uh-oh! used pa while discarding");
4023 schedule_timeout_uninterruptible(HZ);
4027 if (pa->pa_deleted == 0) {
4029 spin_unlock(&pa->pa_lock);
4030 list_del_rcu(&pa->pa_inode_list);
4031 list_add(&pa->u.pa_tmp_list, &list);
4035 /* someone is deleting pa right now */
4036 spin_unlock(&pa->pa_lock);
4037 spin_unlock(&ei->i_prealloc_lock);
4039 /* we have to wait here because pa_deleted
4040 * doesn't mean pa is already unlinked from
4041 * the list. as we might be called from
4042 * ->clear_inode() the inode will get freed
4043 * and concurrent thread which is unlinking
4044 * pa from inode's list may access already
4045 * freed memory, bad-bad-bad */
4047 /* XXX: if this happens too often, we can
4048 * add a flag to force wait only in case
4049 * of ->clear_inode(), but not in case of
4050 * regular truncate */
4051 schedule_timeout_uninterruptible(HZ);
4054 spin_unlock(&ei->i_prealloc_lock);
4056 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4057 BUG_ON(pa->pa_type != MB_INODE_PA);
4058 group = ext4_get_group_number(sb, pa->pa_pstart);
4060 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4061 GFP_NOFS|__GFP_NOFAIL);
4063 ext4_error(sb, "Error %d loading buddy information for %u",
4068 bitmap_bh = ext4_read_block_bitmap(sb, group);
4069 if (IS_ERR(bitmap_bh)) {
4070 err = PTR_ERR(bitmap_bh);
4071 ext4_error(sb, "Error %d reading block bitmap for %u",
4073 ext4_mb_unload_buddy(&e4b);
4077 ext4_lock_group(sb, group);
4078 list_del(&pa->pa_group_list);
4079 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4080 ext4_unlock_group(sb, group);
4082 ext4_mb_unload_buddy(&e4b);
4085 list_del(&pa->u.pa_tmp_list);
4086 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4090 #ifdef CONFIG_EXT4_DEBUG
4091 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4093 struct super_block *sb = ac->ac_sb;
4094 ext4_group_t ngroups, i;
4096 if (!ext4_mballoc_debug ||
4097 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4100 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4101 " Allocation context details:");
4102 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4103 ac->ac_status, ac->ac_flags);
4104 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4105 "goal %lu/%lu/%lu@%lu, "
4106 "best %lu/%lu/%lu@%lu cr %d",
4107 (unsigned long)ac->ac_o_ex.fe_group,
4108 (unsigned long)ac->ac_o_ex.fe_start,
4109 (unsigned long)ac->ac_o_ex.fe_len,
4110 (unsigned long)ac->ac_o_ex.fe_logical,
4111 (unsigned long)ac->ac_g_ex.fe_group,
4112 (unsigned long)ac->ac_g_ex.fe_start,
4113 (unsigned long)ac->ac_g_ex.fe_len,
4114 (unsigned long)ac->ac_g_ex.fe_logical,
4115 (unsigned long)ac->ac_b_ex.fe_group,
4116 (unsigned long)ac->ac_b_ex.fe_start,
4117 (unsigned long)ac->ac_b_ex.fe_len,
4118 (unsigned long)ac->ac_b_ex.fe_logical,
4119 (int)ac->ac_criteria);
4120 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4121 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4122 ngroups = ext4_get_groups_count(sb);
4123 for (i = 0; i < ngroups; i++) {
4124 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4125 struct ext4_prealloc_space *pa;
4126 ext4_grpblk_t start;
4127 struct list_head *cur;
4128 ext4_lock_group(sb, i);
4129 list_for_each(cur, &grp->bb_prealloc_list) {
4130 pa = list_entry(cur, struct ext4_prealloc_space,
4132 spin_lock(&pa->pa_lock);
4133 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4135 spin_unlock(&pa->pa_lock);
4136 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4139 ext4_unlock_group(sb, i);
4141 if (grp->bb_free == 0)
4143 printk(KERN_ERR "%u: %d/%d \n",
4144 i, grp->bb_free, grp->bb_fragments);
4146 printk(KERN_ERR "\n");
4149 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4156 * We use locality group preallocation for small size file. The size of the
4157 * file is determined by the current size or the resulting size after
4158 * allocation which ever is larger
4160 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4162 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4164 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4165 int bsbits = ac->ac_sb->s_blocksize_bits;
4168 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4171 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4174 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4175 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4178 if ((size == isize) &&
4179 !ext4_fs_is_busy(sbi) &&
4180 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4181 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4185 if (sbi->s_mb_group_prealloc <= 0) {
4186 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4190 /* don't use group allocation for large files */
4191 size = max(size, isize);
4192 if (size > sbi->s_mb_stream_request) {
4193 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4197 BUG_ON(ac->ac_lg != NULL);
4199 * locality group prealloc space are per cpu. The reason for having
4200 * per cpu locality group is to reduce the contention between block
4201 * request from multiple CPUs.
4203 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4205 /* we're going to use group allocation */
4206 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4208 /* serialize all allocations in the group */
4209 mutex_lock(&ac->ac_lg->lg_mutex);
4212 static noinline_for_stack int
4213 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4214 struct ext4_allocation_request *ar)
4216 struct super_block *sb = ar->inode->i_sb;
4217 struct ext4_sb_info *sbi = EXT4_SB(sb);
4218 struct ext4_super_block *es = sbi->s_es;
4222 ext4_grpblk_t block;
4224 /* we can't allocate > group size */
4227 /* just a dirty hack to filter too big requests */
4228 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4229 len = EXT4_CLUSTERS_PER_GROUP(sb);
4231 /* start searching from the goal */
4233 if (goal < le32_to_cpu(es->s_first_data_block) ||
4234 goal >= ext4_blocks_count(es))
4235 goal = le32_to_cpu(es->s_first_data_block);
4236 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4238 /* set up allocation goals */
4239 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4240 ac->ac_status = AC_STATUS_CONTINUE;
4242 ac->ac_inode = ar->inode;
4243 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4244 ac->ac_o_ex.fe_group = group;
4245 ac->ac_o_ex.fe_start = block;
4246 ac->ac_o_ex.fe_len = len;
4247 ac->ac_g_ex = ac->ac_o_ex;
4248 ac->ac_flags = ar->flags;
4250 /* we have to define context: we'll we work with a file or
4251 * locality group. this is a policy, actually */
4252 ext4_mb_group_or_file(ac);
4254 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4255 "left: %u/%u, right %u/%u to %swritable\n",
4256 (unsigned) ar->len, (unsigned) ar->logical,
4257 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4258 (unsigned) ar->lleft, (unsigned) ar->pleft,
4259 (unsigned) ar->lright, (unsigned) ar->pright,
4260 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4265 static noinline_for_stack void
4266 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4267 struct ext4_locality_group *lg,
4268 int order, int total_entries)
4270 ext4_group_t group = 0;
4271 struct ext4_buddy e4b;
4272 struct list_head discard_list;
4273 struct ext4_prealloc_space *pa, *tmp;
4275 mb_debug(1, "discard locality group preallocation\n");
4277 INIT_LIST_HEAD(&discard_list);
4279 spin_lock(&lg->lg_prealloc_lock);
4280 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4282 spin_lock(&pa->pa_lock);
4283 if (atomic_read(&pa->pa_count)) {
4285 * This is the pa that we just used
4286 * for block allocation. So don't
4289 spin_unlock(&pa->pa_lock);
4292 if (pa->pa_deleted) {
4293 spin_unlock(&pa->pa_lock);
4296 /* only lg prealloc space */
4297 BUG_ON(pa->pa_type != MB_GROUP_PA);
4299 /* seems this one can be freed ... */
4301 spin_unlock(&pa->pa_lock);
4303 list_del_rcu(&pa->pa_inode_list);
4304 list_add(&pa->u.pa_tmp_list, &discard_list);
4307 if (total_entries <= 5) {
4309 * we want to keep only 5 entries
4310 * allowing it to grow to 8. This
4311 * mak sure we don't call discard
4312 * soon for this list.
4317 spin_unlock(&lg->lg_prealloc_lock);
4319 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4322 group = ext4_get_group_number(sb, pa->pa_pstart);
4323 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4324 GFP_NOFS|__GFP_NOFAIL);
4326 ext4_error(sb, "Error %d loading buddy information for %u",
4330 ext4_lock_group(sb, group);
4331 list_del(&pa->pa_group_list);
4332 ext4_mb_release_group_pa(&e4b, pa);
4333 ext4_unlock_group(sb, group);
4335 ext4_mb_unload_buddy(&e4b);
4336 list_del(&pa->u.pa_tmp_list);
4337 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4342 * We have incremented pa_count. So it cannot be freed at this
4343 * point. Also we hold lg_mutex. So no parallel allocation is
4344 * possible from this lg. That means pa_free cannot be updated.
4346 * A parallel ext4_mb_discard_group_preallocations is possible.
4347 * which can cause the lg_prealloc_list to be updated.
4350 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4352 int order, added = 0, lg_prealloc_count = 1;
4353 struct super_block *sb = ac->ac_sb;
4354 struct ext4_locality_group *lg = ac->ac_lg;
4355 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4357 order = fls(pa->pa_free) - 1;
4358 if (order > PREALLOC_TB_SIZE - 1)
4359 /* The max size of hash table is PREALLOC_TB_SIZE */
4360 order = PREALLOC_TB_SIZE - 1;
4361 /* Add the prealloc space to lg */
4362 spin_lock(&lg->lg_prealloc_lock);
4363 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4365 spin_lock(&tmp_pa->pa_lock);
4366 if (tmp_pa->pa_deleted) {
4367 spin_unlock(&tmp_pa->pa_lock);
4370 if (!added && pa->pa_free < tmp_pa->pa_free) {
4371 /* Add to the tail of the previous entry */
4372 list_add_tail_rcu(&pa->pa_inode_list,
4373 &tmp_pa->pa_inode_list);
4376 * we want to count the total
4377 * number of entries in the list
4380 spin_unlock(&tmp_pa->pa_lock);
4381 lg_prealloc_count++;
4384 list_add_tail_rcu(&pa->pa_inode_list,
4385 &lg->lg_prealloc_list[order]);
4386 spin_unlock(&lg->lg_prealloc_lock);
4388 /* Now trim the list to be not more than 8 elements */
4389 if (lg_prealloc_count > 8) {
4390 ext4_mb_discard_lg_preallocations(sb, lg,
4391 order, lg_prealloc_count);
4398 * release all resource we used in allocation
4400 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4402 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4403 struct ext4_prealloc_space *pa = ac->ac_pa;
4405 if (pa->pa_type == MB_GROUP_PA) {
4406 /* see comment in ext4_mb_use_group_pa() */
4407 spin_lock(&pa->pa_lock);
4408 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4409 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4410 pa->pa_free -= ac->ac_b_ex.fe_len;
4411 pa->pa_len -= ac->ac_b_ex.fe_len;
4412 spin_unlock(&pa->pa_lock);
4417 * We want to add the pa to the right bucket.
4418 * Remove it from the list and while adding
4419 * make sure the list to which we are adding
4422 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4423 spin_lock(pa->pa_obj_lock);
4424 list_del_rcu(&pa->pa_inode_list);
4425 spin_unlock(pa->pa_obj_lock);
4426 ext4_mb_add_n_trim(ac);
4428 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4430 if (ac->ac_bitmap_page)
4431 put_page(ac->ac_bitmap_page);
4432 if (ac->ac_buddy_page)
4433 put_page(ac->ac_buddy_page);
4434 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4435 mutex_unlock(&ac->ac_lg->lg_mutex);
4436 ext4_mb_collect_stats(ac);
4440 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4442 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4446 trace_ext4_mb_discard_preallocations(sb, needed);
4447 for (i = 0; i < ngroups && needed > 0; i++) {
4448 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4457 * Main entry point into mballoc to allocate blocks
4458 * it tries to use preallocation first, then falls back
4459 * to usual allocation
4461 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4462 struct ext4_allocation_request *ar, int *errp)
4465 struct ext4_allocation_context *ac = NULL;
4466 struct ext4_sb_info *sbi;
4467 struct super_block *sb;
4468 ext4_fsblk_t block = 0;
4469 unsigned int inquota = 0;
4470 unsigned int reserv_clstrs = 0;
4473 sb = ar->inode->i_sb;
4476 trace_ext4_request_blocks(ar);
4478 /* Allow to use superuser reservation for quota file */
4479 if (ext4_is_quota_file(ar->inode))
4480 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4482 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4483 /* Without delayed allocation we need to verify
4484 * there is enough free blocks to do block allocation
4485 * and verify allocation doesn't exceed the quota limits.
4488 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4490 /* let others to free the space */
4492 ar->len = ar->len >> 1;
4498 reserv_clstrs = ar->len;
4499 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4500 dquot_alloc_block_nofail(ar->inode,
4501 EXT4_C2B(sbi, ar->len));
4504 dquot_alloc_block(ar->inode,
4505 EXT4_C2B(sbi, ar->len))) {
4507 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4518 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4525 *errp = ext4_mb_initialize_context(ac, ar);
4531 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4532 if (!ext4_mb_use_preallocated(ac)) {
4533 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4534 ext4_mb_normalize_request(ac, ar);
4536 /* allocate space in core */
4537 *errp = ext4_mb_regular_allocator(ac);
4539 goto discard_and_exit;
4541 /* as we've just preallocated more space than
4542 * user requested originally, we store allocated
4543 * space in a special descriptor */
4544 if (ac->ac_status == AC_STATUS_FOUND &&
4545 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4546 *errp = ext4_mb_new_preallocation(ac);
4549 ext4_discard_allocated_blocks(ac);
4553 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4554 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4556 ext4_discard_allocated_blocks(ac);
4559 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4560 ar->len = ac->ac_b_ex.fe_len;
4563 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4571 ac->ac_b_ex.fe_len = 0;
4573 ext4_mb_show_ac(ac);
4575 ext4_mb_release_context(ac);
4578 kmem_cache_free(ext4_ac_cachep, ac);
4579 if (inquota && ar->len < inquota)
4580 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4582 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4583 /* release all the reserved blocks if non delalloc */
4584 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4588 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4594 * We can merge two free data extents only if the physical blocks
4595 * are contiguous, AND the extents were freed by the same transaction,
4596 * AND the blocks are associated with the same group.
4598 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
4599 struct ext4_free_data *entry,
4600 struct ext4_free_data *new_entry,
4601 struct rb_root *entry_rb_root)
4603 if ((entry->efd_tid != new_entry->efd_tid) ||
4604 (entry->efd_group != new_entry->efd_group))
4606 if (entry->efd_start_cluster + entry->efd_count ==
4607 new_entry->efd_start_cluster) {
4608 new_entry->efd_start_cluster = entry->efd_start_cluster;
4609 new_entry->efd_count += entry->efd_count;
4610 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
4611 entry->efd_start_cluster) {
4612 new_entry->efd_count += entry->efd_count;
4615 spin_lock(&sbi->s_md_lock);
4616 list_del(&entry->efd_list);
4617 spin_unlock(&sbi->s_md_lock);
4618 rb_erase(&entry->efd_node, entry_rb_root);
4619 kmem_cache_free(ext4_free_data_cachep, entry);
4622 static noinline_for_stack int
4623 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4624 struct ext4_free_data *new_entry)
4626 ext4_group_t group = e4b->bd_group;
4627 ext4_grpblk_t cluster;
4628 ext4_grpblk_t clusters = new_entry->efd_count;
4629 struct ext4_free_data *entry;
4630 struct ext4_group_info *db = e4b->bd_info;
4631 struct super_block *sb = e4b->bd_sb;
4632 struct ext4_sb_info *sbi = EXT4_SB(sb);
4633 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4634 struct rb_node *parent = NULL, *new_node;
4636 BUG_ON(!ext4_handle_valid(handle));
4637 BUG_ON(e4b->bd_bitmap_page == NULL);
4638 BUG_ON(e4b->bd_buddy_page == NULL);
4640 new_node = &new_entry->efd_node;
4641 cluster = new_entry->efd_start_cluster;
4644 /* first free block exent. We need to
4645 protect buddy cache from being freed,
4646 * otherwise we'll refresh it from
4647 * on-disk bitmap and lose not-yet-available
4649 get_page(e4b->bd_buddy_page);
4650 get_page(e4b->bd_bitmap_page);
4654 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4655 if (cluster < entry->efd_start_cluster)
4657 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4658 n = &(*n)->rb_right;
4660 ext4_grp_locked_error(sb, group, 0,
4661 ext4_group_first_block_no(sb, group) +
4662 EXT4_C2B(sbi, cluster),
4663 "Block already on to-be-freed list");
4668 rb_link_node(new_node, parent, n);
4669 rb_insert_color(new_node, &db->bb_free_root);
4671 /* Now try to see the extent can be merged to left and right */
4672 node = rb_prev(new_node);
4674 entry = rb_entry(node, struct ext4_free_data, efd_node);
4675 ext4_try_merge_freed_extent(sbi, entry, new_entry,
4676 &(db->bb_free_root));
4679 node = rb_next(new_node);
4681 entry = rb_entry(node, struct ext4_free_data, efd_node);
4682 ext4_try_merge_freed_extent(sbi, entry, new_entry,
4683 &(db->bb_free_root));
4686 spin_lock(&sbi->s_md_lock);
4687 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
4688 sbi->s_mb_free_pending += clusters;
4689 spin_unlock(&sbi->s_md_lock);
4694 * ext4_free_blocks() -- Free given blocks and update quota
4695 * @handle: handle for this transaction
4697 * @block: start physical block to free
4698 * @count: number of blocks to count
4699 * @flags: flags used by ext4_free_blocks
4701 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4702 struct buffer_head *bh, ext4_fsblk_t block,
4703 unsigned long count, int flags)
4705 struct buffer_head *bitmap_bh = NULL;
4706 struct super_block *sb = inode->i_sb;
4707 struct ext4_group_desc *gdp;
4708 unsigned int overflow;
4710 struct buffer_head *gd_bh;
4711 ext4_group_t block_group;
4712 struct ext4_sb_info *sbi;
4713 struct ext4_buddy e4b;
4714 unsigned int count_clusters;
4721 BUG_ON(block != bh->b_blocknr);
4723 block = bh->b_blocknr;
4727 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4728 !ext4_data_block_valid(sbi, block, count)) {
4729 ext4_error(sb, "Freeing blocks not in datazone - "
4730 "block = %llu, count = %lu", block, count);
4734 ext4_debug("freeing block %llu\n", block);
4735 trace_ext4_free_blocks(inode, block, count, flags);
4737 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4740 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4745 * If the extent to be freed does not begin on a cluster
4746 * boundary, we need to deal with partial clusters at the
4747 * beginning and end of the extent. Normally we will free
4748 * blocks at the beginning or the end unless we are explicitly
4749 * requested to avoid doing so.
4751 overflow = EXT4_PBLK_COFF(sbi, block);
4753 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4754 overflow = sbi->s_cluster_ratio - overflow;
4756 if (count > overflow)
4765 overflow = EXT4_LBLK_COFF(sbi, count);
4767 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4768 if (count > overflow)
4773 count += sbi->s_cluster_ratio - overflow;
4776 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4778 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
4780 for (i = 0; i < count; i++) {
4783 bh = sb_find_get_block(inode->i_sb, block + i);
4784 ext4_forget(handle, is_metadata, inode, bh, block + i);
4790 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4792 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4793 ext4_get_group_info(sb, block_group))))
4797 * Check to see if we are freeing blocks across a group
4800 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4801 overflow = EXT4_C2B(sbi, bit) + count -
4802 EXT4_BLOCKS_PER_GROUP(sb);
4805 count_clusters = EXT4_NUM_B2C(sbi, count);
4806 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4807 if (IS_ERR(bitmap_bh)) {
4808 err = PTR_ERR(bitmap_bh);
4812 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4818 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4819 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4820 in_range(block, ext4_inode_table(sb, gdp),
4821 sbi->s_itb_per_group) ||
4822 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4823 sbi->s_itb_per_group)) {
4825 ext4_error(sb, "Freeing blocks in system zone - "
4826 "Block = %llu, count = %lu", block, count);
4827 /* err = 0. ext4_std_error should be a no op */
4831 BUFFER_TRACE(bitmap_bh, "getting write access");
4832 err = ext4_journal_get_write_access(handle, bitmap_bh);
4837 * We are about to modify some metadata. Call the journal APIs
4838 * to unshare ->b_data if a currently-committing transaction is
4841 BUFFER_TRACE(gd_bh, "get_write_access");
4842 err = ext4_journal_get_write_access(handle, gd_bh);
4845 #ifdef AGGRESSIVE_CHECK
4848 for (i = 0; i < count_clusters; i++)
4849 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4852 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4854 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4855 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
4856 GFP_NOFS|__GFP_NOFAIL);
4861 * We need to make sure we don't reuse the freed block until after the
4862 * transaction is committed. We make an exception if the inode is to be
4863 * written in writeback mode since writeback mode has weak data
4864 * consistency guarantees.
4866 if (ext4_handle_valid(handle) &&
4867 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
4868 !ext4_should_writeback_data(inode))) {
4869 struct ext4_free_data *new_entry;
4871 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4874 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4875 GFP_NOFS|__GFP_NOFAIL);
4876 new_entry->efd_start_cluster = bit;
4877 new_entry->efd_group = block_group;
4878 new_entry->efd_count = count_clusters;
4879 new_entry->efd_tid = handle->h_transaction->t_tid;
4881 ext4_lock_group(sb, block_group);
4882 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4883 ext4_mb_free_metadata(handle, &e4b, new_entry);
4885 /* need to update group_info->bb_free and bitmap
4886 * with group lock held. generate_buddy look at
4887 * them with group lock_held
4889 if (test_opt(sb, DISCARD)) {
4890 err = ext4_issue_discard(sb, block_group, bit, count,
4892 if (err && err != -EOPNOTSUPP)
4893 ext4_msg(sb, KERN_WARNING, "discard request in"
4894 " group:%d block:%d count:%lu failed"
4895 " with %d", block_group, bit, count,
4898 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4900 ext4_lock_group(sb, block_group);
4901 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4902 mb_free_blocks(inode, &e4b, bit, count_clusters);
4905 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4906 ext4_free_group_clusters_set(sb, gdp, ret);
4907 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4908 ext4_group_desc_csum_set(sb, block_group, gdp);
4909 ext4_unlock_group(sb, block_group);
4911 if (sbi->s_log_groups_per_flex) {
4912 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4913 atomic64_add(count_clusters,
4914 &sbi->s_flex_groups[flex_group].free_clusters);
4917 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4918 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4919 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4921 ext4_mb_unload_buddy(&e4b);
4923 /* We dirtied the bitmap block */
4924 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4925 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4927 /* And the group descriptor block */
4928 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4929 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4933 if (overflow && !err) {
4941 ext4_std_error(sb, err);
4946 * ext4_group_add_blocks() -- Add given blocks to an existing group
4947 * @handle: handle to this transaction
4949 * @block: start physical block to add to the block group
4950 * @count: number of blocks to free
4952 * This marks the blocks as free in the bitmap and buddy.
4954 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4955 ext4_fsblk_t block, unsigned long count)
4957 struct buffer_head *bitmap_bh = NULL;
4958 struct buffer_head *gd_bh;
4959 ext4_group_t block_group;
4962 struct ext4_group_desc *desc;
4963 struct ext4_sb_info *sbi = EXT4_SB(sb);
4964 struct ext4_buddy e4b;
4965 int err = 0, ret, free_clusters_count;
4966 ext4_grpblk_t clusters_freed;
4967 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
4968 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
4969 unsigned long cluster_count = last_cluster - first_cluster + 1;
4971 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4976 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4978 * Check to see if we are freeing blocks across a group
4981 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
4982 ext4_warning(sb, "too many blocks added to group %u",
4988 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4989 if (IS_ERR(bitmap_bh)) {
4990 err = PTR_ERR(bitmap_bh);
4995 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
5001 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
5002 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
5003 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
5004 in_range(block + count - 1, ext4_inode_table(sb, desc),
5005 sbi->s_itb_per_group)) {
5006 ext4_error(sb, "Adding blocks in system zones - "
5007 "Block = %llu, count = %lu",
5013 BUFFER_TRACE(bitmap_bh, "getting write access");
5014 err = ext4_journal_get_write_access(handle, bitmap_bh);
5019 * We are about to modify some metadata. Call the journal APIs
5020 * to unshare ->b_data if a currently-committing transaction is
5023 BUFFER_TRACE(gd_bh, "get_write_access");
5024 err = ext4_journal_get_write_access(handle, gd_bh);
5028 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
5029 BUFFER_TRACE(bitmap_bh, "clear bit");
5030 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
5031 ext4_error(sb, "bit already cleared for block %llu",
5032 (ext4_fsblk_t)(block + i));
5033 BUFFER_TRACE(bitmap_bh, "bit already cleared");
5039 err = ext4_mb_load_buddy(sb, block_group, &e4b);
5044 * need to update group_info->bb_free and bitmap
5045 * with group lock held. generate_buddy look at
5046 * them with group lock_held
5048 ext4_lock_group(sb, block_group);
5049 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
5050 mb_free_blocks(NULL, &e4b, bit, cluster_count);
5051 free_clusters_count = clusters_freed +
5052 ext4_free_group_clusters(sb, desc);
5053 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
5054 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5055 ext4_group_desc_csum_set(sb, block_group, desc);
5056 ext4_unlock_group(sb, block_group);
5057 percpu_counter_add(&sbi->s_freeclusters_counter,
5060 if (sbi->s_log_groups_per_flex) {
5061 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5062 atomic64_add(clusters_freed,
5063 &sbi->s_flex_groups[flex_group].free_clusters);
5066 ext4_mb_unload_buddy(&e4b);
5068 /* We dirtied the bitmap block */
5069 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5070 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5072 /* And the group descriptor block */
5073 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5074 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5080 ext4_std_error(sb, err);
5085 * ext4_trim_extent -- function to TRIM one single free extent in the group
5086 * @sb: super block for the file system
5087 * @start: starting block of the free extent in the alloc. group
5088 * @count: number of blocks to TRIM
5089 * @group: alloc. group we are working with
5090 * @e4b: ext4 buddy for the group
5092 * Trim "count" blocks starting at "start" in the "group". To assure that no
5093 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5094 * be called with under the group lock.
5096 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5097 ext4_group_t group, struct ext4_buddy *e4b)
5101 struct ext4_free_extent ex;
5104 trace_ext4_trim_extent(sb, group, start, count);
5106 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5108 ex.fe_start = start;
5109 ex.fe_group = group;
5113 * Mark blocks used, so no one can reuse them while
5116 mb_mark_used(e4b, &ex);
5117 ext4_unlock_group(sb, group);
5118 ret = ext4_issue_discard(sb, group, start, count, NULL);
5119 ext4_lock_group(sb, group);
5120 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5125 * ext4_trim_all_free -- function to trim all free space in alloc. group
5126 * @sb: super block for file system
5127 * @group: group to be trimmed
5128 * @start: first group block to examine
5129 * @max: last group block to examine
5130 * @minblocks: minimum extent block count
5132 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5133 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5137 * ext4_trim_all_free walks through group's block bitmap searching for free
5138 * extents. When the free extent is found, mark it as used in group buddy
5139 * bitmap. Then issue a TRIM command on this extent and free the extent in
5140 * the group buddy bitmap. This is done until whole group is scanned.
5142 static ext4_grpblk_t
5143 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5144 ext4_grpblk_t start, ext4_grpblk_t max,
5145 ext4_grpblk_t minblocks)
5148 ext4_grpblk_t next, count = 0, free_count = 0;
5149 struct ext4_buddy e4b;
5152 trace_ext4_trim_all_free(sb, group, start, max);
5154 ret = ext4_mb_load_buddy(sb, group, &e4b);
5156 ext4_warning(sb, "Error %d loading buddy information for %u",
5160 bitmap = e4b.bd_bitmap;
5162 ext4_lock_group(sb, group);
5163 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5164 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5167 start = (e4b.bd_info->bb_first_free > start) ?
5168 e4b.bd_info->bb_first_free : start;
5170 while (start <= max) {
5171 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5174 next = mb_find_next_bit(bitmap, max + 1, start);
5176 if ((next - start) >= minblocks) {
5177 ret = ext4_trim_extent(sb, start,
5178 next - start, group, &e4b);
5179 if (ret && ret != -EOPNOTSUPP)
5182 count += next - start;
5184 free_count += next - start;
5187 if (fatal_signal_pending(current)) {
5188 count = -ERESTARTSYS;
5192 if (need_resched()) {
5193 ext4_unlock_group(sb, group);
5195 ext4_lock_group(sb, group);
5198 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5204 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5207 ext4_unlock_group(sb, group);
5208 ext4_mb_unload_buddy(&e4b);
5210 ext4_debug("trimmed %d blocks in the group %d\n",
5217 * ext4_trim_fs() -- trim ioctl handle function
5218 * @sb: superblock for filesystem
5219 * @range: fstrim_range structure
5221 * start: First Byte to trim
5222 * len: number of Bytes to trim from start
5223 * minlen: minimum extent length in Bytes
5224 * ext4_trim_fs goes through all allocation groups containing Bytes from
5225 * start to start+len. For each such a group ext4_trim_all_free function
5226 * is invoked to trim all free space.
5228 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5230 struct ext4_group_info *grp;
5231 ext4_group_t group, first_group, last_group;
5232 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5233 uint64_t start, end, minlen, trimmed = 0;
5234 ext4_fsblk_t first_data_blk =
5235 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5236 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5239 start = range->start >> sb->s_blocksize_bits;
5240 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5241 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5242 range->minlen >> sb->s_blocksize_bits);
5244 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5245 start >= max_blks ||
5246 range->len < sb->s_blocksize)
5248 if (end >= max_blks)
5250 if (end <= first_data_blk)
5252 if (start < first_data_blk)
5253 start = first_data_blk;
5255 /* Determine first and last group to examine based on start and end */
5256 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5257 &first_group, &first_cluster);
5258 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5259 &last_group, &last_cluster);
5261 /* end now represents the last cluster to discard in this group */
5262 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5264 for (group = first_group; group <= last_group; group++) {
5265 grp = ext4_get_group_info(sb, group);
5266 /* We only do this if the grp has never been initialized */
5267 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5268 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5274 * For all the groups except the last one, last cluster will
5275 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5276 * change it for the last group, note that last_cluster is
5277 * already computed earlier by ext4_get_group_no_and_offset()
5279 if (group == last_group)
5282 if (grp->bb_free >= minlen) {
5283 cnt = ext4_trim_all_free(sb, group, first_cluster,
5293 * For every group except the first one, we are sure
5294 * that the first cluster to discard will be cluster #0.
5300 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5303 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5307 /* Iterate all the free extents in the group. */
5309 ext4_mballoc_query_range(
5310 struct super_block *sb,
5312 ext4_grpblk_t start,
5314 ext4_mballoc_query_range_fn formatter,
5319 struct ext4_buddy e4b;
5322 error = ext4_mb_load_buddy(sb, group, &e4b);
5325 bitmap = e4b.bd_bitmap;
5327 ext4_lock_group(sb, group);
5329 start = (e4b.bd_info->bb_first_free > start) ?
5330 e4b.bd_info->bb_first_free : start;
5331 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
5332 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5334 while (start <= end) {
5335 start = mb_find_next_zero_bit(bitmap, end + 1, start);
5338 next = mb_find_next_bit(bitmap, end + 1, start);
5340 ext4_unlock_group(sb, group);
5341 error = formatter(sb, group, start, next - start, priv);
5344 ext4_lock_group(sb, group);
5349 ext4_unlock_group(sb, group);
5351 ext4_mb_unload_buddy(&e4b);