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/nospec.h>
18 #include <linux/backing-dev.h>
19 #include <trace/events/ext4.h>
21 #ifdef CONFIG_EXT4_DEBUG
22 ushort ext4_mballoc_debug __read_mostly;
24 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
25 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
30 * - test ext4_ext_search_left() and ext4_ext_search_right()
31 * - search for metadata in few groups
34 * - normalization should take into account whether file is still open
35 * - discard preallocations if no free space left (policy?)
36 * - don't normalize tails
38 * - reservation for superuser
41 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
42 * - track min/max extents in each group for better group selection
43 * - mb_mark_used() may allocate chunk right after splitting buddy
44 * - tree of groups sorted by number of free blocks
49 * The allocation request involve request for multiple number of blocks
50 * near to the goal(block) value specified.
52 * During initialization phase of the allocator we decide to use the
53 * group preallocation or inode preallocation depending on the size of
54 * the file. The size of the file could be the resulting file size we
55 * would have after allocation, or the current file size, which ever
56 * is larger. If the size is less than sbi->s_mb_stream_request we
57 * select to use the group preallocation. The default value of
58 * s_mb_stream_request is 16 blocks. This can also be tuned via
59 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
60 * terms of number of blocks.
62 * The main motivation for having small file use group preallocation is to
63 * ensure that we have small files closer together on the disk.
65 * First stage the allocator looks at the inode prealloc list,
66 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
67 * spaces for this particular inode. The inode prealloc space is
70 * pa_lstart -> the logical start block for this prealloc space
71 * pa_pstart -> the physical start block for this prealloc space
72 * pa_len -> length for this prealloc space (in clusters)
73 * pa_free -> free space available in this prealloc space (in clusters)
75 * The inode preallocation space is used looking at the _logical_ start
76 * block. If only the logical file block falls within the range of prealloc
77 * space we will consume the particular prealloc space. This makes sure that
78 * we have contiguous physical blocks representing the file blocks
80 * The important thing to be noted in case of inode prealloc space is that
81 * we don't modify the values associated to inode prealloc space except
84 * If we are not able to find blocks in the inode prealloc space and if we
85 * have the group allocation flag set then we look at the locality group
86 * prealloc space. These are per CPU prealloc list represented as
88 * ext4_sb_info.s_locality_groups[smp_processor_id()]
90 * The reason for having a per cpu locality group is to reduce the contention
91 * between CPUs. It is possible to get scheduled at this point.
93 * The locality group prealloc space is used looking at whether we have
94 * enough free space (pa_free) within the prealloc space.
96 * If we can't allocate blocks via inode prealloc or/and locality group
97 * prealloc then we look at the buddy cache. The buddy cache is represented
98 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
99 * mapped to the buddy and bitmap information regarding different
100 * groups. The buddy information is attached to buddy cache inode so that
101 * we can access them through the page cache. The information regarding
102 * each group is loaded via ext4_mb_load_buddy. The information involve
103 * block bitmap and buddy information. The information are stored in the
107 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
110 * one block each for bitmap and buddy information. So for each group we
111 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
112 * blocksize) blocks. So it can have information regarding groups_per_page
113 * which is blocks_per_page/2
115 * The buddy cache inode is not stored on disk. The inode is thrown
116 * away when the filesystem is unmounted.
118 * We look for count number of blocks in the buddy cache. If we were able
119 * to locate that many free blocks we return with additional information
120 * regarding rest of the contiguous physical block available
122 * Before allocating blocks via buddy cache we normalize the request
123 * blocks. This ensure we ask for more blocks that we needed. The extra
124 * blocks that we get after allocation is added to the respective prealloc
125 * list. In case of inode preallocation we follow a list of heuristics
126 * based on file size. This can be found in ext4_mb_normalize_request. If
127 * we are doing a group prealloc we try to normalize the request to
128 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
129 * dependent on the cluster size; for non-bigalloc file systems, it is
130 * 512 blocks. This can be tuned via
131 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
132 * terms of number of blocks. If we have mounted the file system with -O
133 * stripe=<value> option the group prealloc request is normalized to the
134 * the smallest multiple of the stripe value (sbi->s_stripe) which is
135 * greater than the default mb_group_prealloc.
137 * The regular allocator (using the buddy cache) supports a few tunables.
139 * /sys/fs/ext4/<partition>/mb_min_to_scan
140 * /sys/fs/ext4/<partition>/mb_max_to_scan
141 * /sys/fs/ext4/<partition>/mb_order2_req
143 * The regular allocator uses buddy scan only if the request len is power of
144 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
145 * value of s_mb_order2_reqs can be tuned via
146 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
147 * stripe size (sbi->s_stripe), we try to search for contiguous block in
148 * stripe size. This should result in better allocation on RAID setups. If
149 * not, we search in the specific group using bitmap for best extents. The
150 * tunable min_to_scan and max_to_scan control the behaviour here.
151 * min_to_scan indicate how long the mballoc __must__ look for a best
152 * extent and max_to_scan indicates how long the mballoc __can__ look for a
153 * best extent in the found extents. Searching for the blocks starts with
154 * the group specified as the goal value in allocation context via
155 * ac_g_ex. Each group is first checked based on the criteria whether it
156 * can be used for allocation. ext4_mb_good_group explains how the groups are
159 * Both the prealloc space are getting populated as above. So for the first
160 * request we will hit the buddy cache which will result in this prealloc
161 * space getting filled. The prealloc space is then later used for the
162 * subsequent request.
166 * mballoc operates on the following data:
168 * - in-core buddy (actually includes buddy and bitmap)
169 * - preallocation descriptors (PAs)
171 * there are two types of preallocations:
173 * assiged to specific inode and can be used for this inode only.
174 * it describes part of inode's space preallocated to specific
175 * physical blocks. any block from that preallocated can be used
176 * independent. the descriptor just tracks number of blocks left
177 * unused. so, before taking some block from descriptor, one must
178 * make sure corresponded logical block isn't allocated yet. this
179 * also means that freeing any block within descriptor's range
180 * must discard all preallocated blocks.
182 * assigned to specific locality group which does not translate to
183 * permanent set of inodes: inode can join and leave group. space
184 * from this type of preallocation can be used for any inode. thus
185 * it's consumed from the beginning to the end.
187 * relation between them can be expressed as:
188 * in-core buddy = on-disk bitmap + preallocation descriptors
190 * this mean blocks mballoc considers used are:
191 * - allocated blocks (persistent)
192 * - preallocated blocks (non-persistent)
194 * consistency in mballoc world means that at any time a block is either
195 * free or used in ALL structures. notice: "any time" should not be read
196 * literally -- time is discrete and delimited by locks.
198 * to keep it simple, we don't use block numbers, instead we count number of
199 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
201 * all operations can be expressed as:
202 * - init buddy: buddy = on-disk + PAs
203 * - new PA: buddy += N; PA = N
204 * - use inode PA: on-disk += N; PA -= N
205 * - discard inode PA buddy -= on-disk - PA; PA = 0
206 * - use locality group PA on-disk += N; PA -= N
207 * - discard locality group PA buddy -= PA; PA = 0
208 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
209 * is used in real operation because we can't know actual used
210 * bits from PA, only from on-disk bitmap
212 * if we follow this strict logic, then all operations above should be atomic.
213 * given some of them can block, we'd have to use something like semaphores
214 * killing performance on high-end SMP hardware. let's try to relax it using
215 * the following knowledge:
216 * 1) if buddy is referenced, it's already initialized
217 * 2) while block is used in buddy and the buddy is referenced,
218 * nobody can re-allocate that block
219 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
220 * bit set and PA claims same block, it's OK. IOW, one can set bit in
221 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
224 * so, now we're building a concurrency table:
227 * blocks for PA are allocated in the buddy, buddy must be referenced
228 * until PA is linked to allocation group to avoid concurrent buddy init
230 * we need to make sure that either on-disk bitmap or PA has uptodate data
231 * given (3) we care that PA-=N operation doesn't interfere with init
233 * the simplest way would be to have buddy initialized by the discard
234 * - use locality group PA
235 * again PA-=N must be serialized with init
236 * - discard locality group PA
237 * the simplest way would be to have buddy initialized by the discard
240 * i_data_sem serializes them
242 * discard process must wait until PA isn't used by another process
243 * - use locality group PA
244 * some mutex should serialize them
245 * - discard locality group PA
246 * discard process must wait until PA isn't used by another process
249 * i_data_sem or another mutex should serializes them
251 * discard process must wait until PA isn't used by another process
252 * - use locality group PA
253 * nothing wrong here -- they're different PAs covering different blocks
254 * - discard locality group PA
255 * discard process must wait until PA isn't used by another process
257 * now we're ready to make few consequences:
258 * - PA is referenced and while it is no discard is possible
259 * - PA is referenced until block isn't marked in on-disk bitmap
260 * - PA changes only after on-disk bitmap
261 * - discard must not compete with init. either init is done before
262 * any discard or they're serialized somehow
263 * - buddy init as sum of on-disk bitmap and PAs is done atomically
265 * a special case when we've used PA to emptiness. no need to modify buddy
266 * in this case, but we should care about concurrent init
271 * Logic in few words:
276 * mark bits in on-disk bitmap
279 * - use preallocation:
280 * find proper PA (per-inode or group)
282 * mark bits in on-disk bitmap
288 * mark bits in on-disk bitmap
291 * - discard preallocations in group:
293 * move them onto local list
294 * load on-disk bitmap
296 * remove PA from object (inode or locality group)
297 * mark free blocks in-core
299 * - discard inode's preallocations:
306 * - bitlock on a group (group)
307 * - object (inode/locality) (object)
318 * - release consumed pa:
323 * - generate in-core bitmap:
327 * - discard all for given object (inode, locality group):
332 * - discard all for given group:
339 static struct kmem_cache *ext4_pspace_cachep;
340 static struct kmem_cache *ext4_ac_cachep;
341 static struct kmem_cache *ext4_free_data_cachep;
343 /* We create slab caches for groupinfo data structures based on the
344 * superblock block size. There will be one per mounted filesystem for
345 * each unique s_blocksize_bits */
346 #define NR_GRPINFO_CACHES 8
347 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
349 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
350 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
351 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
352 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
355 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
357 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
360 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
362 #if BITS_PER_LONG == 64
363 *bit += ((unsigned long) addr & 7UL) << 3;
364 addr = (void *) ((unsigned long) addr & ~7UL);
365 #elif BITS_PER_LONG == 32
366 *bit += ((unsigned long) addr & 3UL) << 3;
367 addr = (void *) ((unsigned long) addr & ~3UL);
369 #error "how many bits you are?!"
374 static inline int mb_test_bit(int bit, void *addr)
377 * ext4_test_bit on architecture like powerpc
378 * needs unsigned long aligned address
380 addr = mb_correct_addr_and_bit(&bit, addr);
381 return ext4_test_bit(bit, addr);
384 static inline void mb_set_bit(int bit, void *addr)
386 addr = mb_correct_addr_and_bit(&bit, addr);
387 ext4_set_bit(bit, addr);
390 static inline void mb_clear_bit(int bit, void *addr)
392 addr = mb_correct_addr_and_bit(&bit, addr);
393 ext4_clear_bit(bit, addr);
396 static inline int mb_test_and_clear_bit(int bit, void *addr)
398 addr = mb_correct_addr_and_bit(&bit, addr);
399 return ext4_test_and_clear_bit(bit, addr);
402 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
404 int fix = 0, ret, tmpmax;
405 addr = mb_correct_addr_and_bit(&fix, addr);
409 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
415 static inline int mb_find_next_bit(void *addr, int max, int start)
417 int fix = 0, ret, tmpmax;
418 addr = mb_correct_addr_and_bit(&fix, addr);
422 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
428 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
432 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
435 if (order > e4b->bd_blkbits + 1) {
440 /* at order 0 we see each particular block */
442 *max = 1 << (e4b->bd_blkbits + 3);
443 return e4b->bd_bitmap;
446 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
447 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
453 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
454 int first, int count)
457 struct super_block *sb = e4b->bd_sb;
459 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
461 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
462 for (i = 0; i < count; i++) {
463 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
464 ext4_fsblk_t blocknr;
466 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
467 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
468 ext4_grp_locked_error(sb, e4b->bd_group,
469 inode ? inode->i_ino : 0,
471 "freeing block already freed "
474 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
475 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
477 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
481 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
485 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
487 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
488 for (i = 0; i < count; i++) {
489 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
490 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
494 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
496 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
497 unsigned char *b1, *b2;
499 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
500 b2 = (unsigned char *) bitmap;
501 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
502 if (b1[i] != b2[i]) {
503 ext4_msg(e4b->bd_sb, KERN_ERR,
504 "corruption in group %u "
505 "at byte %u(%u): %x in copy != %x "
507 e4b->bd_group, i, i * 8, b1[i], b2[i]);
515 static inline void mb_free_blocks_double(struct inode *inode,
516 struct ext4_buddy *e4b, int first, int count)
520 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
521 int first, int count)
525 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
531 #ifdef AGGRESSIVE_CHECK
533 #define MB_CHECK_ASSERT(assert) \
537 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
538 function, file, line, # assert); \
543 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
544 const char *function, int line)
546 struct super_block *sb = e4b->bd_sb;
547 int order = e4b->bd_blkbits + 1;
554 struct ext4_group_info *grp;
557 struct list_head *cur;
562 static int mb_check_counter;
563 if (mb_check_counter++ % 100 != 0)
568 buddy = mb_find_buddy(e4b, order, &max);
569 MB_CHECK_ASSERT(buddy);
570 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
571 MB_CHECK_ASSERT(buddy2);
572 MB_CHECK_ASSERT(buddy != buddy2);
573 MB_CHECK_ASSERT(max * 2 == max2);
576 for (i = 0; i < max; i++) {
578 if (mb_test_bit(i, buddy)) {
579 /* only single bit in buddy2 may be 1 */
580 if (!mb_test_bit(i << 1, buddy2)) {
582 mb_test_bit((i<<1)+1, buddy2));
583 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
585 mb_test_bit(i << 1, buddy2));
590 /* both bits in buddy2 must be 1 */
591 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
592 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
594 for (j = 0; j < (1 << order); j++) {
595 k = (i * (1 << order)) + j;
597 !mb_test_bit(k, e4b->bd_bitmap));
601 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
606 buddy = mb_find_buddy(e4b, 0, &max);
607 for (i = 0; i < max; i++) {
608 if (!mb_test_bit(i, buddy)) {
609 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
617 /* check used bits only */
618 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
619 buddy2 = mb_find_buddy(e4b, j, &max2);
621 MB_CHECK_ASSERT(k < max2);
622 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
625 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
626 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
628 grp = ext4_get_group_info(sb, e4b->bd_group);
629 list_for_each(cur, &grp->bb_prealloc_list) {
630 ext4_group_t groupnr;
631 struct ext4_prealloc_space *pa;
632 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
633 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
634 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
635 for (i = 0; i < pa->pa_len; i++)
636 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
640 #undef MB_CHECK_ASSERT
641 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
642 __FILE__, __func__, __LINE__)
644 #define mb_check_buddy(e4b)
648 * Divide blocks started from @first with length @len into
649 * smaller chunks with power of 2 blocks.
650 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
651 * then increase bb_counters[] for corresponded chunk size.
653 static void ext4_mb_mark_free_simple(struct super_block *sb,
654 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
655 struct ext4_group_info *grp)
657 struct ext4_sb_info *sbi = EXT4_SB(sb);
663 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
665 border = 2 << sb->s_blocksize_bits;
668 /* find how many blocks can be covered since this position */
669 max = ffs(first | border) - 1;
671 /* find how many blocks of power 2 we need to mark */
678 /* mark multiblock chunks only */
679 grp->bb_counters[min]++;
681 mb_clear_bit(first >> min,
682 buddy + sbi->s_mb_offsets[min]);
690 * Cache the order of the largest free extent we have available in this block
694 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
699 grp->bb_largest_free_order = -1; /* uninit */
701 bits = sb->s_blocksize_bits + 1;
702 for (i = bits; i >= 0; i--) {
703 if (grp->bb_counters[i] > 0) {
704 grp->bb_largest_free_order = i;
710 static noinline_for_stack
711 void ext4_mb_generate_buddy(struct super_block *sb,
712 void *buddy, void *bitmap, ext4_group_t group)
714 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
715 struct ext4_sb_info *sbi = EXT4_SB(sb);
716 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
721 unsigned fragments = 0;
722 unsigned long long period = get_cycles();
724 /* initialize buddy from bitmap which is aggregation
725 * of on-disk bitmap and preallocations */
726 i = mb_find_next_zero_bit(bitmap, max, 0);
727 grp->bb_first_free = i;
731 i = mb_find_next_bit(bitmap, max, i);
735 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
737 grp->bb_counters[0]++;
739 i = mb_find_next_zero_bit(bitmap, max, i);
741 grp->bb_fragments = fragments;
743 if (free != grp->bb_free) {
744 ext4_grp_locked_error(sb, group, 0, 0,
745 "block bitmap and bg descriptor "
746 "inconsistent: %u vs %u free clusters",
749 * If we intend to continue, we consider group descriptor
750 * corrupt and update bb_free using bitmap value
753 ext4_mark_group_bitmap_corrupted(sb, group,
754 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
756 mb_set_largest_free_order(sb, grp);
758 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
760 period = get_cycles() - period;
761 spin_lock(&sbi->s_bal_lock);
762 sbi->s_mb_buddies_generated++;
763 sbi->s_mb_generation_time += period;
764 spin_unlock(&sbi->s_bal_lock);
767 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
773 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
774 ext4_set_bits(buddy, 0, count);
776 e4b->bd_info->bb_fragments = 0;
777 memset(e4b->bd_info->bb_counters, 0,
778 sizeof(*e4b->bd_info->bb_counters) *
779 (e4b->bd_sb->s_blocksize_bits + 2));
781 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
782 e4b->bd_bitmap, e4b->bd_group);
785 /* The buddy information is attached the buddy cache inode
786 * for convenience. The information regarding each group
787 * is loaded via ext4_mb_load_buddy. The information involve
788 * block bitmap and buddy information. The information are
789 * stored in the inode as
792 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
795 * one block each for bitmap and buddy information.
796 * So for each group we take up 2 blocks. A page can
797 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
798 * So it can have information regarding groups_per_page which
799 * is blocks_per_page/2
801 * Locking note: This routine takes the block group lock of all groups
802 * for this page; do not hold this lock when calling this routine!
805 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
807 ext4_group_t ngroups;
813 ext4_group_t first_group, group;
815 struct super_block *sb;
816 struct buffer_head *bhs;
817 struct buffer_head **bh = NULL;
821 struct ext4_group_info *grinfo;
823 mb_debug(1, "init page %lu\n", page->index);
825 inode = page->mapping->host;
827 ngroups = ext4_get_groups_count(sb);
828 blocksize = i_blocksize(inode);
829 blocks_per_page = PAGE_SIZE / blocksize;
831 groups_per_page = blocks_per_page >> 1;
832 if (groups_per_page == 0)
835 /* allocate buffer_heads to read bitmaps */
836 if (groups_per_page > 1) {
837 i = sizeof(struct buffer_head *) * groups_per_page;
838 bh = kzalloc(i, gfp);
846 first_group = page->index * blocks_per_page / 2;
848 /* read all groups the page covers into the cache */
849 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
850 if (group >= ngroups)
853 grinfo = ext4_get_group_info(sb, group);
855 * If page is uptodate then we came here after online resize
856 * which added some new uninitialized group info structs, so
857 * we must skip all initialized uptodate buddies on the page,
858 * which may be currently in use by an allocating task.
860 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
864 bh[i] = ext4_read_block_bitmap_nowait(sb, group);
866 err = PTR_ERR(bh[i]);
870 mb_debug(1, "read bitmap for group %u\n", group);
873 /* wait for I/O completion */
874 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
879 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
884 first_block = page->index * blocks_per_page;
885 for (i = 0; i < blocks_per_page; i++) {
886 group = (first_block + i) >> 1;
887 if (group >= ngroups)
890 if (!bh[group - first_group])
891 /* skip initialized uptodate buddy */
894 if (!buffer_verified(bh[group - first_group]))
895 /* Skip faulty bitmaps */
900 * data carry information regarding this
901 * particular group in the format specified
905 data = page_address(page) + (i * blocksize);
906 bitmap = bh[group - first_group]->b_data;
909 * We place the buddy block and bitmap block
912 if ((first_block + i) & 1) {
913 /* this is block of buddy */
914 BUG_ON(incore == NULL);
915 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
916 group, page->index, i * blocksize);
917 trace_ext4_mb_buddy_bitmap_load(sb, group);
918 grinfo = ext4_get_group_info(sb, group);
919 grinfo->bb_fragments = 0;
920 memset(grinfo->bb_counters, 0,
921 sizeof(*grinfo->bb_counters) *
922 (sb->s_blocksize_bits+2));
924 * incore got set to the group block bitmap below
926 ext4_lock_group(sb, group);
928 memset(data, 0xff, blocksize);
929 ext4_mb_generate_buddy(sb, data, incore, group);
930 ext4_unlock_group(sb, group);
933 /* this is block of bitmap */
934 BUG_ON(incore != NULL);
935 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
936 group, page->index, i * blocksize);
937 trace_ext4_mb_bitmap_load(sb, group);
939 /* see comments in ext4_mb_put_pa() */
940 ext4_lock_group(sb, group);
941 memcpy(data, bitmap, blocksize);
943 /* mark all preallocated blks used in in-core bitmap */
944 ext4_mb_generate_from_pa(sb, data, group);
945 ext4_mb_generate_from_freelist(sb, data, group);
946 ext4_unlock_group(sb, group);
948 /* set incore so that the buddy information can be
949 * generated using this
954 SetPageUptodate(page);
958 for (i = 0; i < groups_per_page; i++)
967 * Lock the buddy and bitmap pages. This make sure other parallel init_group
968 * on the same buddy page doesn't happen whild holding the buddy page lock.
969 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
970 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
972 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
973 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
975 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
976 int block, pnum, poff;
980 e4b->bd_buddy_page = NULL;
981 e4b->bd_bitmap_page = NULL;
983 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
985 * the buddy cache inode stores the block bitmap
986 * and buddy information in consecutive blocks.
987 * So for each group we need two blocks.
990 pnum = block / blocks_per_page;
991 poff = block % blocks_per_page;
992 page = find_or_create_page(inode->i_mapping, pnum, gfp);
995 BUG_ON(page->mapping != inode->i_mapping);
996 e4b->bd_bitmap_page = page;
997 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
999 if (blocks_per_page >= 2) {
1000 /* buddy and bitmap are on the same page */
1005 pnum = block / blocks_per_page;
1006 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1009 BUG_ON(page->mapping != inode->i_mapping);
1010 e4b->bd_buddy_page = page;
1014 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1016 if (e4b->bd_bitmap_page) {
1017 unlock_page(e4b->bd_bitmap_page);
1018 put_page(e4b->bd_bitmap_page);
1020 if (e4b->bd_buddy_page) {
1021 unlock_page(e4b->bd_buddy_page);
1022 put_page(e4b->bd_buddy_page);
1027 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1028 * block group lock of all groups for this page; do not hold the BG lock when
1029 * calling this routine!
1031 static noinline_for_stack
1032 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1035 struct ext4_group_info *this_grp;
1036 struct ext4_buddy e4b;
1041 mb_debug(1, "init group %u\n", group);
1042 this_grp = ext4_get_group_info(sb, group);
1044 * This ensures that we don't reinit the buddy cache
1045 * page which map to the group from which we are already
1046 * allocating. If we are looking at the buddy cache we would
1047 * have taken a reference using ext4_mb_load_buddy and that
1048 * would have pinned buddy page to page cache.
1049 * The call to ext4_mb_get_buddy_page_lock will mark the
1052 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1053 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1055 * somebody initialized the group
1056 * return without doing anything
1061 page = e4b.bd_bitmap_page;
1062 ret = ext4_mb_init_cache(page, NULL, gfp);
1065 if (!PageUptodate(page)) {
1070 if (e4b.bd_buddy_page == NULL) {
1072 * If both the bitmap and buddy are in
1073 * the same page we don't need to force
1079 /* init buddy cache */
1080 page = e4b.bd_buddy_page;
1081 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1084 if (!PageUptodate(page)) {
1089 ext4_mb_put_buddy_page_lock(&e4b);
1094 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1095 * block group lock of all groups for this page; do not hold the BG lock when
1096 * calling this routine!
1098 static noinline_for_stack int
1099 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1100 struct ext4_buddy *e4b, gfp_t gfp)
1102 int blocks_per_page;
1108 struct ext4_group_info *grp;
1109 struct ext4_sb_info *sbi = EXT4_SB(sb);
1110 struct inode *inode = sbi->s_buddy_cache;
1113 mb_debug(1, "load group %u\n", group);
1115 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1116 grp = ext4_get_group_info(sb, group);
1118 e4b->bd_blkbits = sb->s_blocksize_bits;
1121 e4b->bd_group = group;
1122 e4b->bd_buddy_page = NULL;
1123 e4b->bd_bitmap_page = NULL;
1125 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1127 * we need full data about the group
1128 * to make a good selection
1130 ret = ext4_mb_init_group(sb, group, gfp);
1136 * the buddy cache inode stores the block bitmap
1137 * and buddy information in consecutive blocks.
1138 * So for each group we need two blocks.
1141 pnum = block / blocks_per_page;
1142 poff = block % blocks_per_page;
1144 /* we could use find_or_create_page(), but it locks page
1145 * what we'd like to avoid in fast path ... */
1146 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1147 if (page == NULL || !PageUptodate(page)) {
1150 * drop the page reference and try
1151 * to get the page with lock. If we
1152 * are not uptodate that implies
1153 * somebody just created the page but
1154 * is yet to initialize the same. So
1155 * wait for it to initialize.
1158 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1160 BUG_ON(page->mapping != inode->i_mapping);
1161 if (!PageUptodate(page)) {
1162 ret = ext4_mb_init_cache(page, NULL, gfp);
1167 mb_cmp_bitmaps(e4b, page_address(page) +
1168 (poff * sb->s_blocksize));
1177 if (!PageUptodate(page)) {
1182 /* Pages marked accessed already */
1183 e4b->bd_bitmap_page = page;
1184 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1187 pnum = block / blocks_per_page;
1188 poff = block % blocks_per_page;
1190 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1191 if (page == NULL || !PageUptodate(page)) {
1194 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1196 BUG_ON(page->mapping != inode->i_mapping);
1197 if (!PageUptodate(page)) {
1198 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1212 if (!PageUptodate(page)) {
1217 /* Pages marked accessed already */
1218 e4b->bd_buddy_page = page;
1219 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1221 BUG_ON(e4b->bd_bitmap_page == NULL);
1222 BUG_ON(e4b->bd_buddy_page == NULL);
1229 if (e4b->bd_bitmap_page)
1230 put_page(e4b->bd_bitmap_page);
1231 if (e4b->bd_buddy_page)
1232 put_page(e4b->bd_buddy_page);
1233 e4b->bd_buddy = NULL;
1234 e4b->bd_bitmap = NULL;
1238 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1239 struct ext4_buddy *e4b)
1241 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1244 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1246 if (e4b->bd_bitmap_page)
1247 put_page(e4b->bd_bitmap_page);
1248 if (e4b->bd_buddy_page)
1249 put_page(e4b->bd_buddy_page);
1253 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1256 int bb_incr = 1 << (e4b->bd_blkbits - 1);
1259 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1260 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1263 while (order <= e4b->bd_blkbits + 1) {
1265 if (!mb_test_bit(block, bb)) {
1266 /* this block is part of buddy of order 'order' */
1276 static void mb_clear_bits(void *bm, int cur, int len)
1282 if ((cur & 31) == 0 && (len - cur) >= 32) {
1283 /* fast path: clear whole word at once */
1284 addr = bm + (cur >> 3);
1289 mb_clear_bit(cur, bm);
1294 /* clear bits in given range
1295 * will return first found zero bit if any, -1 otherwise
1297 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1304 if ((cur & 31) == 0 && (len - cur) >= 32) {
1305 /* fast path: clear whole word at once */
1306 addr = bm + (cur >> 3);
1307 if (*addr != (__u32)(-1) && zero_bit == -1)
1308 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1313 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1321 void ext4_set_bits(void *bm, int cur, int len)
1327 if ((cur & 31) == 0 && (len - cur) >= 32) {
1328 /* fast path: set whole word at once */
1329 addr = bm + (cur >> 3);
1334 mb_set_bit(cur, bm);
1340 * _________________________________________________________________ */
1342 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1344 if (mb_test_bit(*bit + side, bitmap)) {
1345 mb_clear_bit(*bit, bitmap);
1351 mb_set_bit(*bit, bitmap);
1356 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1360 void *buddy = mb_find_buddy(e4b, order, &max);
1365 /* Bits in range [first; last] are known to be set since
1366 * corresponding blocks were allocated. Bits in range
1367 * (first; last) will stay set because they form buddies on
1368 * upper layer. We just deal with borders if they don't
1369 * align with upper layer and then go up.
1370 * Releasing entire group is all about clearing
1371 * single bit of highest order buddy.
1375 * ---------------------------------
1377 * ---------------------------------
1378 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1379 * ---------------------------------
1381 * \_____________________/
1383 * Neither [1] nor [6] is aligned to above layer.
1384 * Left neighbour [0] is free, so mark it busy,
1385 * decrease bb_counters and extend range to
1387 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1388 * mark [6] free, increase bb_counters and shrink range to
1390 * Then shift range to [0; 2], go up and do the same.
1395 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1397 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1402 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1403 mb_clear_bits(buddy, first, last - first + 1);
1404 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1413 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1414 int first, int count)
1416 int left_is_free = 0;
1417 int right_is_free = 0;
1419 int last = first + count - 1;
1420 struct super_block *sb = e4b->bd_sb;
1422 if (WARN_ON(count == 0))
1424 BUG_ON(last >= (sb->s_blocksize << 3));
1425 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1426 /* Don't bother if the block group is corrupt. */
1427 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1430 mb_check_buddy(e4b);
1431 mb_free_blocks_double(inode, e4b, first, count);
1433 e4b->bd_info->bb_free += count;
1434 if (first < e4b->bd_info->bb_first_free)
1435 e4b->bd_info->bb_first_free = first;
1437 /* access memory sequentially: check left neighbour,
1438 * clear range and then check right neighbour
1441 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1442 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1443 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1444 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1446 if (unlikely(block != -1)) {
1447 struct ext4_sb_info *sbi = EXT4_SB(sb);
1448 ext4_fsblk_t blocknr;
1450 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1451 blocknr += EXT4_C2B(sbi, block);
1452 ext4_grp_locked_error(sb, e4b->bd_group,
1453 inode ? inode->i_ino : 0,
1455 "freeing already freed block "
1456 "(bit %u); block bitmap corrupt.",
1458 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1459 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1460 mb_regenerate_buddy(e4b);
1464 /* let's maintain fragments counter */
1465 if (left_is_free && right_is_free)
1466 e4b->bd_info->bb_fragments--;
1467 else if (!left_is_free && !right_is_free)
1468 e4b->bd_info->bb_fragments++;
1470 /* buddy[0] == bd_bitmap is a special case, so handle
1471 * it right away and let mb_buddy_mark_free stay free of
1472 * zero order checks.
1473 * Check if neighbours are to be coaleasced,
1474 * adjust bitmap bb_counters and borders appropriately.
1477 first += !left_is_free;
1478 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1481 last -= !right_is_free;
1482 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1486 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1489 mb_set_largest_free_order(sb, e4b->bd_info);
1490 mb_check_buddy(e4b);
1493 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1494 int needed, struct ext4_free_extent *ex)
1500 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1503 buddy = mb_find_buddy(e4b, 0, &max);
1504 BUG_ON(buddy == NULL);
1505 BUG_ON(block >= max);
1506 if (mb_test_bit(block, buddy)) {
1513 /* find actual order */
1514 order = mb_find_order_for_block(e4b, block);
1515 block = block >> order;
1517 ex->fe_len = 1 << order;
1518 ex->fe_start = block << order;
1519 ex->fe_group = e4b->bd_group;
1521 /* calc difference from given start */
1522 next = next - ex->fe_start;
1524 ex->fe_start += next;
1526 while (needed > ex->fe_len &&
1527 mb_find_buddy(e4b, order, &max)) {
1529 if (block + 1 >= max)
1532 next = (block + 1) * (1 << order);
1533 if (mb_test_bit(next, e4b->bd_bitmap))
1536 order = mb_find_order_for_block(e4b, next);
1538 block = next >> order;
1539 ex->fe_len += 1 << order;
1542 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1543 /* Should never happen! (but apparently sometimes does?!?) */
1545 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
1546 "corruption or bug in mb_find_extent "
1547 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1548 block, order, needed, ex->fe_group, ex->fe_start,
1549 ex->fe_len, ex->fe_logical);
1557 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1563 int start = ex->fe_start;
1564 int len = ex->fe_len;
1569 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1570 BUG_ON(e4b->bd_group != ex->fe_group);
1571 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1572 mb_check_buddy(e4b);
1573 mb_mark_used_double(e4b, start, len);
1575 e4b->bd_info->bb_free -= len;
1576 if (e4b->bd_info->bb_first_free == start)
1577 e4b->bd_info->bb_first_free += len;
1579 /* let's maintain fragments counter */
1581 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1582 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1583 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1585 e4b->bd_info->bb_fragments++;
1586 else if (!mlen && !max)
1587 e4b->bd_info->bb_fragments--;
1589 /* let's maintain buddy itself */
1591 ord = mb_find_order_for_block(e4b, start);
1593 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1594 /* the whole chunk may be allocated at once! */
1596 buddy = mb_find_buddy(e4b, ord, &max);
1597 BUG_ON((start >> ord) >= max);
1598 mb_set_bit(start >> ord, buddy);
1599 e4b->bd_info->bb_counters[ord]--;
1606 /* store for history */
1608 ret = len | (ord << 16);
1610 /* we have to split large buddy */
1612 buddy = mb_find_buddy(e4b, ord, &max);
1613 mb_set_bit(start >> ord, buddy);
1614 e4b->bd_info->bb_counters[ord]--;
1617 cur = (start >> ord) & ~1U;
1618 buddy = mb_find_buddy(e4b, ord, &max);
1619 mb_clear_bit(cur, buddy);
1620 mb_clear_bit(cur + 1, buddy);
1621 e4b->bd_info->bb_counters[ord]++;
1622 e4b->bd_info->bb_counters[ord]++;
1624 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1626 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1627 mb_check_buddy(e4b);
1633 * Must be called under group lock!
1635 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1636 struct ext4_buddy *e4b)
1638 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1641 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1642 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1644 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1645 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1646 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1648 /* preallocation can change ac_b_ex, thus we store actually
1649 * allocated blocks for history */
1650 ac->ac_f_ex = ac->ac_b_ex;
1652 ac->ac_status = AC_STATUS_FOUND;
1653 ac->ac_tail = ret & 0xffff;
1654 ac->ac_buddy = ret >> 16;
1657 * take the page reference. We want the page to be pinned
1658 * so that we don't get a ext4_mb_init_cache_call for this
1659 * group until we update the bitmap. That would mean we
1660 * double allocate blocks. The reference is dropped
1661 * in ext4_mb_release_context
1663 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1664 get_page(ac->ac_bitmap_page);
1665 ac->ac_buddy_page = e4b->bd_buddy_page;
1666 get_page(ac->ac_buddy_page);
1667 /* store last allocated for subsequent stream allocation */
1668 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1669 spin_lock(&sbi->s_md_lock);
1670 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1671 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1672 spin_unlock(&sbi->s_md_lock);
1677 * regular allocator, for general purposes allocation
1680 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1681 struct ext4_buddy *e4b,
1684 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1685 struct ext4_free_extent *bex = &ac->ac_b_ex;
1686 struct ext4_free_extent *gex = &ac->ac_g_ex;
1687 struct ext4_free_extent ex;
1690 if (ac->ac_status == AC_STATUS_FOUND)
1693 * We don't want to scan for a whole year
1695 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1696 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1697 ac->ac_status = AC_STATUS_BREAK;
1702 * Haven't found good chunk so far, let's continue
1704 if (bex->fe_len < gex->fe_len)
1707 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1708 && bex->fe_group == e4b->bd_group) {
1709 /* recheck chunk's availability - we don't know
1710 * when it was found (within this lock-unlock
1712 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1713 if (max >= gex->fe_len) {
1714 ext4_mb_use_best_found(ac, e4b);
1721 * The routine checks whether found extent is good enough. If it is,
1722 * then the extent gets marked used and flag is set to the context
1723 * to stop scanning. Otherwise, the extent is compared with the
1724 * previous found extent and if new one is better, then it's stored
1725 * in the context. Later, the best found extent will be used, if
1726 * mballoc can't find good enough extent.
1728 * FIXME: real allocation policy is to be designed yet!
1730 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1731 struct ext4_free_extent *ex,
1732 struct ext4_buddy *e4b)
1734 struct ext4_free_extent *bex = &ac->ac_b_ex;
1735 struct ext4_free_extent *gex = &ac->ac_g_ex;
1737 BUG_ON(ex->fe_len <= 0);
1738 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1739 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1740 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1745 * The special case - take what you catch first
1747 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1749 ext4_mb_use_best_found(ac, e4b);
1754 * Let's check whether the chuck is good enough
1756 if (ex->fe_len == gex->fe_len) {
1758 ext4_mb_use_best_found(ac, e4b);
1763 * If this is first found extent, just store it in the context
1765 if (bex->fe_len == 0) {
1771 * If new found extent is better, store it in the context
1773 if (bex->fe_len < gex->fe_len) {
1774 /* if the request isn't satisfied, any found extent
1775 * larger than previous best one is better */
1776 if (ex->fe_len > bex->fe_len)
1778 } else if (ex->fe_len > gex->fe_len) {
1779 /* if the request is satisfied, then we try to find
1780 * an extent that still satisfy the request, but is
1781 * smaller than previous one */
1782 if (ex->fe_len < bex->fe_len)
1786 ext4_mb_check_limits(ac, e4b, 0);
1789 static noinline_for_stack
1790 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1791 struct ext4_buddy *e4b)
1793 struct ext4_free_extent ex = ac->ac_b_ex;
1794 ext4_group_t group = ex.fe_group;
1798 BUG_ON(ex.fe_len <= 0);
1799 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1803 ext4_lock_group(ac->ac_sb, group);
1804 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1808 ext4_mb_use_best_found(ac, e4b);
1811 ext4_unlock_group(ac->ac_sb, group);
1812 ext4_mb_unload_buddy(e4b);
1817 static noinline_for_stack
1818 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1819 struct ext4_buddy *e4b)
1821 ext4_group_t group = ac->ac_g_ex.fe_group;
1824 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1825 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1826 struct ext4_free_extent ex;
1828 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1830 if (grp->bb_free == 0)
1833 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1837 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1838 ext4_mb_unload_buddy(e4b);
1842 ext4_lock_group(ac->ac_sb, group);
1843 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1844 ac->ac_g_ex.fe_len, &ex);
1845 ex.fe_logical = 0xDEADFA11; /* debug value */
1847 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1850 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1852 /* use do_div to get remainder (would be 64-bit modulo) */
1853 if (do_div(start, sbi->s_stripe) == 0) {
1856 ext4_mb_use_best_found(ac, e4b);
1858 } else if (max >= ac->ac_g_ex.fe_len) {
1859 BUG_ON(ex.fe_len <= 0);
1860 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1861 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1864 ext4_mb_use_best_found(ac, e4b);
1865 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1866 /* Sometimes, caller may want to merge even small
1867 * number of blocks to an existing extent */
1868 BUG_ON(ex.fe_len <= 0);
1869 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1870 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1873 ext4_mb_use_best_found(ac, e4b);
1875 ext4_unlock_group(ac->ac_sb, group);
1876 ext4_mb_unload_buddy(e4b);
1882 * The routine scans buddy structures (not bitmap!) from given order
1883 * to max order and tries to find big enough chunk to satisfy the req
1885 static noinline_for_stack
1886 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1887 struct ext4_buddy *e4b)
1889 struct super_block *sb = ac->ac_sb;
1890 struct ext4_group_info *grp = e4b->bd_info;
1896 BUG_ON(ac->ac_2order <= 0);
1897 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1898 if (grp->bb_counters[i] == 0)
1901 buddy = mb_find_buddy(e4b, i, &max);
1902 BUG_ON(buddy == NULL);
1904 k = mb_find_next_zero_bit(buddy, max, 0);
1906 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
1907 "%d free clusters of order %d. But found 0",
1908 grp->bb_counters[i], i);
1909 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
1911 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1916 ac->ac_b_ex.fe_len = 1 << i;
1917 ac->ac_b_ex.fe_start = k << i;
1918 ac->ac_b_ex.fe_group = e4b->bd_group;
1920 ext4_mb_use_best_found(ac, e4b);
1922 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1924 if (EXT4_SB(sb)->s_mb_stats)
1925 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1932 * The routine scans the group and measures all found extents.
1933 * In order to optimize scanning, caller must pass number of
1934 * free blocks in the group, so the routine can know upper limit.
1936 static noinline_for_stack
1937 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1938 struct ext4_buddy *e4b)
1940 struct super_block *sb = ac->ac_sb;
1941 void *bitmap = e4b->bd_bitmap;
1942 struct ext4_free_extent ex;
1946 free = e4b->bd_info->bb_free;
1947 if (WARN_ON(free <= 0))
1950 i = e4b->bd_info->bb_first_free;
1952 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1953 i = mb_find_next_zero_bit(bitmap,
1954 EXT4_CLUSTERS_PER_GROUP(sb), i);
1955 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1957 * IF we have corrupt bitmap, we won't find any
1958 * free blocks even though group info says we
1959 * we have free blocks
1961 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1962 "%d free clusters as per "
1963 "group info. But bitmap says 0",
1965 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1966 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1970 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1971 if (WARN_ON(ex.fe_len <= 0))
1973 if (free < ex.fe_len) {
1974 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1975 "%d free clusters as per "
1976 "group info. But got %d blocks",
1978 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1979 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1981 * The number of free blocks differs. This mostly
1982 * indicate that the bitmap is corrupt. So exit
1983 * without claiming the space.
1987 ex.fe_logical = 0xDEADC0DE; /* debug value */
1988 ext4_mb_measure_extent(ac, &ex, e4b);
1994 ext4_mb_check_limits(ac, e4b, 1);
1998 * This is a special case for storages like raid5
1999 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2001 static noinline_for_stack
2002 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2003 struct ext4_buddy *e4b)
2005 struct super_block *sb = ac->ac_sb;
2006 struct ext4_sb_info *sbi = EXT4_SB(sb);
2007 void *bitmap = e4b->bd_bitmap;
2008 struct ext4_free_extent ex;
2009 ext4_fsblk_t first_group_block;
2014 BUG_ON(sbi->s_stripe == 0);
2016 /* find first stripe-aligned block in group */
2017 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2019 a = first_group_block + sbi->s_stripe - 1;
2020 do_div(a, sbi->s_stripe);
2021 i = (a * sbi->s_stripe) - first_group_block;
2023 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2024 if (!mb_test_bit(i, bitmap)) {
2025 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2026 if (max >= sbi->s_stripe) {
2028 ex.fe_logical = 0xDEADF00D; /* debug value */
2030 ext4_mb_use_best_found(ac, e4b);
2039 * This is now called BEFORE we load the buddy bitmap.
2040 * Returns either 1 or 0 indicating that the group is either suitable
2041 * for the allocation or not. In addition it can also return negative
2042 * error code when something goes wrong.
2044 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2045 ext4_group_t group, int cr)
2047 unsigned free, fragments;
2048 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2049 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2051 BUG_ON(cr < 0 || cr >= 4);
2053 free = grp->bb_free;
2056 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2059 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2062 /* We only do this if the grp has never been initialized */
2063 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2064 int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS);
2069 fragments = grp->bb_fragments;
2075 BUG_ON(ac->ac_2order == 0);
2077 /* Avoid using the first bg of a flexgroup for data files */
2078 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2079 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2080 ((group % flex_size) == 0))
2083 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2084 (free / fragments) >= ac->ac_g_ex.fe_len)
2087 if (grp->bb_largest_free_order < ac->ac_2order)
2092 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2096 if (free >= ac->ac_g_ex.fe_len)
2108 static noinline_for_stack int
2109 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2111 ext4_group_t ngroups, group, i;
2113 int err = 0, first_err = 0;
2114 struct ext4_sb_info *sbi;
2115 struct super_block *sb;
2116 struct ext4_buddy e4b;
2120 ngroups = ext4_get_groups_count(sb);
2121 /* non-extent files are limited to low blocks/groups */
2122 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2123 ngroups = sbi->s_blockfile_groups;
2125 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2127 /* first, try the goal */
2128 err = ext4_mb_find_by_goal(ac, &e4b);
2129 if (err || ac->ac_status == AC_STATUS_FOUND)
2132 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2136 * ac->ac2_order is set only if the fe_len is a power of 2
2137 * if ac2_order is set we also set criteria to 0 so that we
2138 * try exact allocation using buddy.
2140 i = fls(ac->ac_g_ex.fe_len);
2143 * We search using buddy data only if the order of the request
2144 * is greater than equal to the sbi_s_mb_order2_reqs
2145 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2146 * We also support searching for power-of-two requests only for
2147 * requests upto maximum buddy size we have constructed.
2149 if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) {
2151 * This should tell if fe_len is exactly power of 2
2153 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2154 ac->ac_2order = array_index_nospec(i - 1,
2155 sb->s_blocksize_bits + 2);
2158 /* if stream allocation is enabled, use global goal */
2159 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2160 /* TBD: may be hot point */
2161 spin_lock(&sbi->s_md_lock);
2162 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2163 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2164 spin_unlock(&sbi->s_md_lock);
2167 /* Let's just scan groups to find more-less suitable blocks */
2168 cr = ac->ac_2order ? 0 : 1;
2170 * cr == 0 try to get exact allocation,
2171 * cr == 3 try to get anything
2174 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2175 ac->ac_criteria = cr;
2177 * searching for the right group start
2178 * from the goal value specified
2180 group = ac->ac_g_ex.fe_group;
2182 for (i = 0; i < ngroups; group++, i++) {
2186 * Artificially restricted ngroups for non-extent
2187 * files makes group > ngroups possible on first loop.
2189 if (group >= ngroups)
2192 /* This now checks without needing the buddy page */
2193 ret = ext4_mb_good_group(ac, group, cr);
2200 err = ext4_mb_load_buddy(sb, group, &e4b);
2204 ext4_lock_group(sb, group);
2207 * We need to check again after locking the
2210 ret = ext4_mb_good_group(ac, group, cr);
2212 ext4_unlock_group(sb, group);
2213 ext4_mb_unload_buddy(&e4b);
2219 ac->ac_groups_scanned++;
2221 ext4_mb_simple_scan_group(ac, &e4b);
2222 else if (cr == 1 && sbi->s_stripe &&
2223 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2224 ext4_mb_scan_aligned(ac, &e4b);
2226 ext4_mb_complex_scan_group(ac, &e4b);
2228 ext4_unlock_group(sb, group);
2229 ext4_mb_unload_buddy(&e4b);
2231 if (ac->ac_status != AC_STATUS_CONTINUE)
2236 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2237 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2239 * We've been searching too long. Let's try to allocate
2240 * the best chunk we've found so far
2243 ext4_mb_try_best_found(ac, &e4b);
2244 if (ac->ac_status != AC_STATUS_FOUND) {
2246 * Someone more lucky has already allocated it.
2247 * The only thing we can do is just take first
2249 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2251 ac->ac_b_ex.fe_group = 0;
2252 ac->ac_b_ex.fe_start = 0;
2253 ac->ac_b_ex.fe_len = 0;
2254 ac->ac_status = AC_STATUS_CONTINUE;
2255 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2257 atomic_inc(&sbi->s_mb_lost_chunks);
2262 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2267 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2269 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 void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2280 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2284 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2287 return (void *) ((unsigned long) group);
2290 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2292 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2293 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2295 int err, buddy_loaded = 0;
2296 struct ext4_buddy e4b;
2297 struct ext4_group_info *grinfo;
2298 unsigned char blocksize_bits = min_t(unsigned char,
2299 sb->s_blocksize_bits,
2300 EXT4_MAX_BLOCK_LOG_SIZE);
2302 struct ext4_group_info info;
2303 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2308 seq_puts(seq, "#group: free frags first ["
2309 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2310 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2312 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2313 sizeof(struct ext4_group_info);
2315 grinfo = ext4_get_group_info(sb, group);
2316 /* Load the group info in memory only if not already loaded. */
2317 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2318 err = ext4_mb_load_buddy(sb, group, &e4b);
2320 seq_printf(seq, "#%-5u: I/O error\n", group);
2326 memcpy(&sg, ext4_get_group_info(sb, group), i);
2329 ext4_mb_unload_buddy(&e4b);
2331 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2332 sg.info.bb_fragments, sg.info.bb_first_free);
2333 for (i = 0; i <= 13; i++)
2334 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2335 sg.info.bb_counters[i] : 0);
2336 seq_printf(seq, " ]\n");
2341 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2345 const struct seq_operations ext4_mb_seq_groups_ops = {
2346 .start = ext4_mb_seq_groups_start,
2347 .next = ext4_mb_seq_groups_next,
2348 .stop = ext4_mb_seq_groups_stop,
2349 .show = ext4_mb_seq_groups_show,
2352 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2354 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2355 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2362 * Allocate the top-level s_group_info array for the specified number
2365 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2367 struct ext4_sb_info *sbi = EXT4_SB(sb);
2369 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
2371 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2372 EXT4_DESC_PER_BLOCK_BITS(sb);
2373 if (size <= sbi->s_group_info_size)
2376 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2377 new_groupinfo = kvzalloc(size, GFP_KERNEL);
2378 if (!new_groupinfo) {
2379 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2383 old_groupinfo = rcu_dereference(sbi->s_group_info);
2385 memcpy(new_groupinfo, old_groupinfo,
2386 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2388 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
2389 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2391 ext4_kvfree_array_rcu(old_groupinfo);
2392 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2393 sbi->s_group_info_size);
2397 /* Create and initialize ext4_group_info data for the given group. */
2398 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2399 struct ext4_group_desc *desc)
2403 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
2404 struct ext4_sb_info *sbi = EXT4_SB(sb);
2405 struct ext4_group_info **meta_group_info;
2406 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2409 * First check if this group is the first of a reserved block.
2410 * If it's true, we have to allocate a new table of pointers
2411 * to ext4_group_info structures
2413 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2414 metalen = sizeof(*meta_group_info) <<
2415 EXT4_DESC_PER_BLOCK_BITS(sb);
2416 meta_group_info = kmalloc(metalen, GFP_NOFS);
2417 if (meta_group_info == NULL) {
2418 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2419 "for a buddy group");
2420 goto exit_meta_group_info;
2423 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
2427 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
2428 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2430 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2431 if (meta_group_info[i] == NULL) {
2432 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2433 goto exit_group_info;
2435 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2436 &(meta_group_info[i]->bb_state));
2439 * initialize bb_free to be able to skip
2440 * empty groups without initialization
2442 if (ext4_has_group_desc_csum(sb) &&
2443 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
2444 meta_group_info[i]->bb_free =
2445 ext4_free_clusters_after_init(sb, group, desc);
2447 meta_group_info[i]->bb_free =
2448 ext4_free_group_clusters(sb, desc);
2451 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2452 init_rwsem(&meta_group_info[i]->alloc_sem);
2453 meta_group_info[i]->bb_free_root = RB_ROOT;
2454 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2458 struct buffer_head *bh;
2459 meta_group_info[i]->bb_bitmap =
2460 kmalloc(sb->s_blocksize, GFP_NOFS);
2461 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2462 bh = ext4_read_block_bitmap(sb, group);
2463 BUG_ON(IS_ERR_OR_NULL(bh));
2464 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2473 /* If a meta_group_info table has been allocated, release it now */
2474 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2475 struct ext4_group_info ***group_info;
2478 group_info = rcu_dereference(sbi->s_group_info);
2479 kfree(group_info[idx]);
2480 group_info[idx] = NULL;
2483 exit_meta_group_info:
2485 } /* ext4_mb_add_groupinfo */
2487 static int ext4_mb_init_backend(struct super_block *sb)
2489 ext4_group_t ngroups = ext4_get_groups_count(sb);
2491 struct ext4_sb_info *sbi = EXT4_SB(sb);
2493 struct ext4_group_desc *desc;
2494 struct ext4_group_info ***group_info;
2495 struct kmem_cache *cachep;
2497 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2501 sbi->s_buddy_cache = new_inode(sb);
2502 if (sbi->s_buddy_cache == NULL) {
2503 ext4_msg(sb, KERN_ERR, "can't get new inode");
2506 /* To avoid potentially colliding with an valid on-disk inode number,
2507 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2508 * not in the inode hash, so it should never be found by iget(), but
2509 * this will avoid confusion if it ever shows up during debugging. */
2510 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2511 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2512 for (i = 0; i < ngroups; i++) {
2514 desc = ext4_get_group_desc(sb, i, NULL);
2516 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2519 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2526 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2528 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2529 i = sbi->s_group_info_size;
2531 group_info = rcu_dereference(sbi->s_group_info);
2533 kfree(group_info[i]);
2535 iput(sbi->s_buddy_cache);
2538 kvfree(rcu_dereference(sbi->s_group_info));
2543 static void ext4_groupinfo_destroy_slabs(void)
2547 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2548 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2549 ext4_groupinfo_caches[i] = NULL;
2553 static int ext4_groupinfo_create_slab(size_t size)
2555 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2557 int blocksize_bits = order_base_2(size);
2558 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2559 struct kmem_cache *cachep;
2561 if (cache_index >= NR_GRPINFO_CACHES)
2564 if (unlikely(cache_index < 0))
2567 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2568 if (ext4_groupinfo_caches[cache_index]) {
2569 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2570 return 0; /* Already created */
2573 slab_size = offsetof(struct ext4_group_info,
2574 bb_counters[blocksize_bits + 2]);
2576 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2577 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2580 ext4_groupinfo_caches[cache_index] = cachep;
2582 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2585 "EXT4-fs: no memory for groupinfo slab cache\n");
2592 int ext4_mb_init(struct super_block *sb)
2594 struct ext4_sb_info *sbi = EXT4_SB(sb);
2596 unsigned offset, offset_incr;
2600 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2602 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2603 if (sbi->s_mb_offsets == NULL) {
2608 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2609 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2610 if (sbi->s_mb_maxs == NULL) {
2615 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2619 /* order 0 is regular bitmap */
2620 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2621 sbi->s_mb_offsets[0] = 0;
2625 offset_incr = 1 << (sb->s_blocksize_bits - 1);
2626 max = sb->s_blocksize << 2;
2628 sbi->s_mb_offsets[i] = offset;
2629 sbi->s_mb_maxs[i] = max;
2630 offset += offset_incr;
2631 offset_incr = offset_incr >> 1;
2634 } while (i <= sb->s_blocksize_bits + 1);
2636 spin_lock_init(&sbi->s_md_lock);
2637 spin_lock_init(&sbi->s_bal_lock);
2638 sbi->s_mb_free_pending = 0;
2639 INIT_LIST_HEAD(&sbi->s_freed_data_list);
2641 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2642 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2643 sbi->s_mb_stats = MB_DEFAULT_STATS;
2644 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2645 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2647 * The default group preallocation is 512, which for 4k block
2648 * sizes translates to 2 megabytes. However for bigalloc file
2649 * systems, this is probably too big (i.e, if the cluster size
2650 * is 1 megabyte, then group preallocation size becomes half a
2651 * gigabyte!). As a default, we will keep a two megabyte
2652 * group pralloc size for cluster sizes up to 64k, and after
2653 * that, we will force a minimum group preallocation size of
2654 * 32 clusters. This translates to 8 megs when the cluster
2655 * size is 256k, and 32 megs when the cluster size is 1 meg,
2656 * which seems reasonable as a default.
2658 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2659 sbi->s_cluster_bits, 32);
2661 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2662 * to the lowest multiple of s_stripe which is bigger than
2663 * the s_mb_group_prealloc as determined above. We want
2664 * the preallocation size to be an exact multiple of the
2665 * RAID stripe size so that preallocations don't fragment
2668 if (sbi->s_stripe > 1) {
2669 sbi->s_mb_group_prealloc = roundup(
2670 sbi->s_mb_group_prealloc, sbi->s_stripe);
2673 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2674 if (sbi->s_locality_groups == NULL) {
2678 for_each_possible_cpu(i) {
2679 struct ext4_locality_group *lg;
2680 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2681 mutex_init(&lg->lg_mutex);
2682 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2683 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2684 spin_lock_init(&lg->lg_prealloc_lock);
2687 /* init file for buddy data */
2688 ret = ext4_mb_init_backend(sb);
2690 goto out_free_locality_groups;
2694 out_free_locality_groups:
2695 free_percpu(sbi->s_locality_groups);
2696 sbi->s_locality_groups = NULL;
2698 kfree(sbi->s_mb_offsets);
2699 sbi->s_mb_offsets = NULL;
2700 kfree(sbi->s_mb_maxs);
2701 sbi->s_mb_maxs = NULL;
2705 /* need to called with the ext4 group lock held */
2706 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2708 struct ext4_prealloc_space *pa;
2709 struct list_head *cur, *tmp;
2712 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2713 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2714 list_del(&pa->pa_group_list);
2716 kmem_cache_free(ext4_pspace_cachep, pa);
2719 mb_debug(1, "mballoc: %u PAs left\n", count);
2723 int ext4_mb_release(struct super_block *sb)
2725 ext4_group_t ngroups = ext4_get_groups_count(sb);
2727 int num_meta_group_infos;
2728 struct ext4_group_info *grinfo, ***group_info;
2729 struct ext4_sb_info *sbi = EXT4_SB(sb);
2730 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2732 if (sbi->s_group_info) {
2733 for (i = 0; i < ngroups; i++) {
2735 grinfo = ext4_get_group_info(sb, i);
2737 kfree(grinfo->bb_bitmap);
2739 ext4_lock_group(sb, i);
2740 ext4_mb_cleanup_pa(grinfo);
2741 ext4_unlock_group(sb, i);
2742 kmem_cache_free(cachep, grinfo);
2744 num_meta_group_infos = (ngroups +
2745 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2746 EXT4_DESC_PER_BLOCK_BITS(sb);
2748 group_info = rcu_dereference(sbi->s_group_info);
2749 for (i = 0; i < num_meta_group_infos; i++)
2750 kfree(group_info[i]);
2754 kfree(sbi->s_mb_offsets);
2755 kfree(sbi->s_mb_maxs);
2756 iput(sbi->s_buddy_cache);
2757 if (sbi->s_mb_stats) {
2758 ext4_msg(sb, KERN_INFO,
2759 "mballoc: %u blocks %u reqs (%u success)",
2760 atomic_read(&sbi->s_bal_allocated),
2761 atomic_read(&sbi->s_bal_reqs),
2762 atomic_read(&sbi->s_bal_success));
2763 ext4_msg(sb, KERN_INFO,
2764 "mballoc: %u extents scanned, %u goal hits, "
2765 "%u 2^N hits, %u breaks, %u lost",
2766 atomic_read(&sbi->s_bal_ex_scanned),
2767 atomic_read(&sbi->s_bal_goals),
2768 atomic_read(&sbi->s_bal_2orders),
2769 atomic_read(&sbi->s_bal_breaks),
2770 atomic_read(&sbi->s_mb_lost_chunks));
2771 ext4_msg(sb, KERN_INFO,
2772 "mballoc: %lu generated and it took %Lu",
2773 sbi->s_mb_buddies_generated,
2774 sbi->s_mb_generation_time);
2775 ext4_msg(sb, KERN_INFO,
2776 "mballoc: %u preallocated, %u discarded",
2777 atomic_read(&sbi->s_mb_preallocated),
2778 atomic_read(&sbi->s_mb_discarded));
2781 free_percpu(sbi->s_locality_groups);
2786 static inline int ext4_issue_discard(struct super_block *sb,
2787 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
2790 ext4_fsblk_t discard_block;
2792 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2793 ext4_group_first_block_no(sb, block_group));
2794 count = EXT4_C2B(EXT4_SB(sb), count);
2795 trace_ext4_discard_blocks(sb,
2796 (unsigned long long) discard_block, count);
2798 return __blkdev_issue_discard(sb->s_bdev,
2799 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
2800 (sector_t)count << (sb->s_blocksize_bits - 9),
2803 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2806 static void ext4_free_data_in_buddy(struct super_block *sb,
2807 struct ext4_free_data *entry)
2809 struct ext4_buddy e4b;
2810 struct ext4_group_info *db;
2811 int err, count = 0, count2 = 0;
2813 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2814 entry->efd_count, entry->efd_group, entry);
2816 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2817 /* we expect to find existing buddy because it's pinned */
2820 spin_lock(&EXT4_SB(sb)->s_md_lock);
2821 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
2822 spin_unlock(&EXT4_SB(sb)->s_md_lock);
2825 /* there are blocks to put in buddy to make them really free */
2826 count += entry->efd_count;
2828 ext4_lock_group(sb, entry->efd_group);
2829 /* Take it out of per group rb tree */
2830 rb_erase(&entry->efd_node, &(db->bb_free_root));
2831 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2834 * Clear the trimmed flag for the group so that the next
2835 * ext4_trim_fs can trim it.
2836 * If the volume is mounted with -o discard, online discard
2837 * is supported and the free blocks will be trimmed online.
2839 if (!test_opt(sb, DISCARD))
2840 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2842 if (!db->bb_free_root.rb_node) {
2843 /* No more items in the per group rb tree
2844 * balance refcounts from ext4_mb_free_metadata()
2846 put_page(e4b.bd_buddy_page);
2847 put_page(e4b.bd_bitmap_page);
2849 ext4_unlock_group(sb, entry->efd_group);
2850 kmem_cache_free(ext4_free_data_cachep, entry);
2851 ext4_mb_unload_buddy(&e4b);
2853 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2857 * This function is called by the jbd2 layer once the commit has finished,
2858 * so we know we can free the blocks that were released with that commit.
2860 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
2862 struct ext4_sb_info *sbi = EXT4_SB(sb);
2863 struct ext4_free_data *entry, *tmp;
2864 struct bio *discard_bio = NULL;
2865 struct list_head freed_data_list;
2866 struct list_head *cut_pos = NULL;
2869 INIT_LIST_HEAD(&freed_data_list);
2871 spin_lock(&sbi->s_md_lock);
2872 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
2873 if (entry->efd_tid != commit_tid)
2875 cut_pos = &entry->efd_list;
2878 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
2880 spin_unlock(&sbi->s_md_lock);
2882 if (test_opt(sb, DISCARD)) {
2883 list_for_each_entry(entry, &freed_data_list, efd_list) {
2884 err = ext4_issue_discard(sb, entry->efd_group,
2885 entry->efd_start_cluster,
2888 if (err && err != -EOPNOTSUPP) {
2889 ext4_msg(sb, KERN_WARNING, "discard request in"
2890 " group:%d block:%d count:%d failed"
2891 " with %d", entry->efd_group,
2892 entry->efd_start_cluster,
2893 entry->efd_count, err);
2894 } else if (err == -EOPNOTSUPP)
2899 submit_bio_wait(discard_bio);
2900 bio_put(discard_bio);
2904 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
2905 ext4_free_data_in_buddy(sb, entry);
2908 int __init ext4_init_mballoc(void)
2910 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2911 SLAB_RECLAIM_ACCOUNT);
2912 if (ext4_pspace_cachep == NULL)
2915 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2916 SLAB_RECLAIM_ACCOUNT);
2917 if (ext4_ac_cachep == NULL) {
2918 kmem_cache_destroy(ext4_pspace_cachep);
2922 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2923 SLAB_RECLAIM_ACCOUNT);
2924 if (ext4_free_data_cachep == NULL) {
2925 kmem_cache_destroy(ext4_pspace_cachep);
2926 kmem_cache_destroy(ext4_ac_cachep);
2932 void ext4_exit_mballoc(void)
2935 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2936 * before destroying the slab cache.
2939 kmem_cache_destroy(ext4_pspace_cachep);
2940 kmem_cache_destroy(ext4_ac_cachep);
2941 kmem_cache_destroy(ext4_free_data_cachep);
2942 ext4_groupinfo_destroy_slabs();
2947 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2948 * Returns 0 if success or error code
2950 static noinline_for_stack int
2951 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2952 handle_t *handle, unsigned int reserv_clstrs)
2954 struct buffer_head *bitmap_bh = NULL;
2955 struct ext4_group_desc *gdp;
2956 struct buffer_head *gdp_bh;
2957 struct ext4_sb_info *sbi;
2958 struct super_block *sb;
2962 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2963 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2968 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2969 if (IS_ERR(bitmap_bh)) {
2970 err = PTR_ERR(bitmap_bh);
2975 BUFFER_TRACE(bitmap_bh, "getting write access");
2976 err = ext4_journal_get_write_access(handle, bitmap_bh);
2981 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2985 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2986 ext4_free_group_clusters(sb, gdp));
2988 BUFFER_TRACE(gdp_bh, "get_write_access");
2989 err = ext4_journal_get_write_access(handle, gdp_bh);
2993 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2995 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2996 if (!ext4_data_block_valid(sbi, block, len)) {
2997 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2998 "fs metadata", block, block+len);
2999 /* File system mounted not to panic on error
3000 * Fix the bitmap and return EFSCORRUPTED
3001 * We leak some of the blocks here.
3003 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3004 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3005 ac->ac_b_ex.fe_len);
3006 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3007 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3009 err = -EFSCORRUPTED;
3013 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3014 #ifdef AGGRESSIVE_CHECK
3017 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3018 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3019 bitmap_bh->b_data));
3023 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3024 ac->ac_b_ex.fe_len);
3025 if (ext4_has_group_desc_csum(sb) &&
3026 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3027 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3028 ext4_free_group_clusters_set(sb, gdp,
3029 ext4_free_clusters_after_init(sb,
3030 ac->ac_b_ex.fe_group, gdp));
3032 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3033 ext4_free_group_clusters_set(sb, gdp, len);
3034 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3035 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3037 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3038 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3040 * Now reduce the dirty block count also. Should not go negative
3042 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3043 /* release all the reserved blocks if non delalloc */
3044 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3047 if (sbi->s_log_groups_per_flex) {
3048 ext4_group_t flex_group = ext4_flex_group(sbi,
3049 ac->ac_b_ex.fe_group);
3050 atomic64_sub(ac->ac_b_ex.fe_len,
3051 &sbi_array_rcu_deref(sbi, s_flex_groups,
3052 flex_group)->free_clusters);
3055 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3058 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3066 * here we normalize request for locality group
3067 * Group request are normalized to s_mb_group_prealloc, which goes to
3068 * s_strip if we set the same via mount option.
3069 * s_mb_group_prealloc can be configured via
3070 * /sys/fs/ext4/<partition>/mb_group_prealloc
3072 * XXX: should we try to preallocate more than the group has now?
3074 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3076 struct super_block *sb = ac->ac_sb;
3077 struct ext4_locality_group *lg = ac->ac_lg;
3080 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3081 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3082 current->pid, ac->ac_g_ex.fe_len);
3086 * Normalization means making request better in terms of
3087 * size and alignment
3089 static noinline_for_stack void
3090 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3091 struct ext4_allocation_request *ar)
3093 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3094 struct ext4_super_block *es = sbi->s_es;
3097 loff_t size, start_off;
3098 loff_t orig_size __maybe_unused;
3100 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3101 struct ext4_prealloc_space *pa;
3103 /* do normalize only data requests, metadata requests
3104 do not need preallocation */
3105 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3108 /* sometime caller may want exact blocks */
3109 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3112 /* caller may indicate that preallocation isn't
3113 * required (it's a tail, for example) */
3114 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3117 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3118 ext4_mb_normalize_group_request(ac);
3122 bsbits = ac->ac_sb->s_blocksize_bits;
3124 /* first, let's learn actual file size
3125 * given current request is allocated */
3126 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3127 size = size << bsbits;
3128 if (size < i_size_read(ac->ac_inode))
3129 size = i_size_read(ac->ac_inode);
3132 /* max size of free chunks */
3135 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3136 (req <= (size) || max <= (chunk_size))
3138 /* first, try to predict filesize */
3139 /* XXX: should this table be tunable? */
3141 if (size <= 16 * 1024) {
3143 } else if (size <= 32 * 1024) {
3145 } else if (size <= 64 * 1024) {
3147 } else if (size <= 128 * 1024) {
3149 } else if (size <= 256 * 1024) {
3151 } else if (size <= 512 * 1024) {
3153 } else if (size <= 1024 * 1024) {
3155 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3156 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3157 (21 - bsbits)) << 21;
3158 size = 2 * 1024 * 1024;
3159 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3160 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3161 (22 - bsbits)) << 22;
3162 size = 4 * 1024 * 1024;
3163 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3164 (8<<20)>>bsbits, max, 8 * 1024)) {
3165 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3166 (23 - bsbits)) << 23;
3167 size = 8 * 1024 * 1024;
3169 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3170 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3171 ac->ac_o_ex.fe_len) << bsbits;
3173 size = size >> bsbits;
3174 start = start_off >> bsbits;
3177 * For tiny groups (smaller than 8MB) the chosen allocation
3178 * alignment may be larger than group size. Make sure the
3179 * alignment does not move allocation to a different group which
3180 * makes mballoc fail assertions later.
3182 start = max(start, rounddown(ac->ac_o_ex.fe_logical,
3183 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
3185 /* don't cover already allocated blocks in selected range */
3186 if (ar->pleft && start <= ar->lleft) {
3187 size -= ar->lleft + 1 - start;
3188 start = ar->lleft + 1;
3190 if (ar->pright && start + size - 1 >= ar->lright)
3191 size -= start + size - ar->lright;
3194 * Trim allocation request for filesystems with artificially small
3197 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
3198 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
3202 /* check we don't cross already preallocated blocks */
3204 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3209 spin_lock(&pa->pa_lock);
3210 if (pa->pa_deleted) {
3211 spin_unlock(&pa->pa_lock);
3215 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3218 /* PA must not overlap original request */
3219 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3220 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3222 /* skip PAs this normalized request doesn't overlap with */
3223 if (pa->pa_lstart >= end || pa_end <= start) {
3224 spin_unlock(&pa->pa_lock);
3227 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3229 /* adjust start or end to be adjacent to this pa */
3230 if (pa_end <= ac->ac_o_ex.fe_logical) {
3231 BUG_ON(pa_end < start);
3233 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3234 BUG_ON(pa->pa_lstart > end);
3235 end = pa->pa_lstart;
3237 spin_unlock(&pa->pa_lock);
3242 /* XXX: extra loop to check we really don't overlap preallocations */
3244 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3247 spin_lock(&pa->pa_lock);
3248 if (pa->pa_deleted == 0) {
3249 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3251 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3253 spin_unlock(&pa->pa_lock);
3257 if (start + size <= ac->ac_o_ex.fe_logical &&
3258 start > ac->ac_o_ex.fe_logical) {
3259 ext4_msg(ac->ac_sb, KERN_ERR,
3260 "start %lu, size %lu, fe_logical %lu",
3261 (unsigned long) start, (unsigned long) size,
3262 (unsigned long) ac->ac_o_ex.fe_logical);
3265 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3267 /* now prepare goal request */
3269 /* XXX: is it better to align blocks WRT to logical
3270 * placement or satisfy big request as is */
3271 ac->ac_g_ex.fe_logical = start;
3272 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3274 /* define goal start in order to merge */
3275 if (ar->pright && (ar->lright == (start + size)) &&
3276 ar->pright >= size &&
3277 ar->pright - size >= le32_to_cpu(es->s_first_data_block)) {
3278 /* merge to the right */
3279 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3280 &ac->ac_g_ex.fe_group,
3281 &ac->ac_g_ex.fe_start);
3282 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3284 if (ar->pleft && (ar->lleft + 1 == start) &&
3285 ar->pleft + 1 < ext4_blocks_count(es)) {
3286 /* merge to the left */
3287 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3288 &ac->ac_g_ex.fe_group,
3289 &ac->ac_g_ex.fe_start);
3290 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3293 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3294 (unsigned) orig_size, (unsigned) start);
3297 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3299 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3301 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3302 atomic_inc(&sbi->s_bal_reqs);
3303 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3304 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3305 atomic_inc(&sbi->s_bal_success);
3306 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3307 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3308 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3309 atomic_inc(&sbi->s_bal_goals);
3310 if (ac->ac_found > sbi->s_mb_max_to_scan)
3311 atomic_inc(&sbi->s_bal_breaks);
3314 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3315 trace_ext4_mballoc_alloc(ac);
3317 trace_ext4_mballoc_prealloc(ac);
3321 * Called on failure; free up any blocks from the inode PA for this
3322 * context. We don't need this for MB_GROUP_PA because we only change
3323 * pa_free in ext4_mb_release_context(), but on failure, we've already
3324 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3326 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3328 struct ext4_prealloc_space *pa = ac->ac_pa;
3329 struct ext4_buddy e4b;
3333 if (ac->ac_f_ex.fe_len == 0)
3335 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3338 * This should never happen since we pin the
3339 * pages in the ext4_allocation_context so
3340 * ext4_mb_load_buddy() should never fail.
3342 WARN(1, "mb_load_buddy failed (%d)", err);
3345 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3346 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3347 ac->ac_f_ex.fe_len);
3348 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3349 ext4_mb_unload_buddy(&e4b);
3352 if (pa->pa_type == MB_INODE_PA)
3353 pa->pa_free += ac->ac_b_ex.fe_len;
3357 * use blocks preallocated to inode
3359 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3360 struct ext4_prealloc_space *pa)
3362 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3367 /* found preallocated blocks, use them */
3368 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3369 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3370 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3371 len = EXT4_NUM_B2C(sbi, end - start);
3372 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3373 &ac->ac_b_ex.fe_start);
3374 ac->ac_b_ex.fe_len = len;
3375 ac->ac_status = AC_STATUS_FOUND;
3378 BUG_ON(start < pa->pa_pstart);
3379 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3380 BUG_ON(pa->pa_free < len);
3381 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3384 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3388 * use blocks preallocated to locality group
3390 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3391 struct ext4_prealloc_space *pa)
3393 unsigned int len = ac->ac_o_ex.fe_len;
3395 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3396 &ac->ac_b_ex.fe_group,
3397 &ac->ac_b_ex.fe_start);
3398 ac->ac_b_ex.fe_len = len;
3399 ac->ac_status = AC_STATUS_FOUND;
3402 /* we don't correct pa_pstart or pa_plen here to avoid
3403 * possible race when the group is being loaded concurrently
3404 * instead we correct pa later, after blocks are marked
3405 * in on-disk bitmap -- see ext4_mb_release_context()
3406 * Other CPUs are prevented from allocating from this pa by lg_mutex
3408 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3412 * Return the prealloc space that have minimal distance
3413 * from the goal block. @cpa is the prealloc
3414 * space that is having currently known minimal distance
3415 * from the goal block.
3417 static struct ext4_prealloc_space *
3418 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3419 struct ext4_prealloc_space *pa,
3420 struct ext4_prealloc_space *cpa)
3422 ext4_fsblk_t cur_distance, new_distance;
3425 atomic_inc(&pa->pa_count);
3428 cur_distance = abs(goal_block - cpa->pa_pstart);
3429 new_distance = abs(goal_block - pa->pa_pstart);
3431 if (cur_distance <= new_distance)
3434 /* drop the previous reference */
3435 atomic_dec(&cpa->pa_count);
3436 atomic_inc(&pa->pa_count);
3441 * search goal blocks in preallocated space
3443 static noinline_for_stack int
3444 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3446 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3448 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3449 struct ext4_locality_group *lg;
3450 struct ext4_prealloc_space *pa, *cpa = NULL;
3451 ext4_fsblk_t goal_block;
3453 /* only data can be preallocated */
3454 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3457 /* first, try per-file preallocation */
3459 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3461 /* all fields in this condition don't change,
3462 * so we can skip locking for them */
3463 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3464 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3465 EXT4_C2B(sbi, pa->pa_len)))
3468 /* non-extent files can't have physical blocks past 2^32 */
3469 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3470 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3471 EXT4_MAX_BLOCK_FILE_PHYS))
3474 /* found preallocated blocks, use them */
3475 spin_lock(&pa->pa_lock);
3476 if (pa->pa_deleted == 0 && pa->pa_free) {
3477 atomic_inc(&pa->pa_count);
3478 ext4_mb_use_inode_pa(ac, pa);
3479 spin_unlock(&pa->pa_lock);
3480 ac->ac_criteria = 10;
3484 spin_unlock(&pa->pa_lock);
3488 /* can we use group allocation? */
3489 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3492 /* inode may have no locality group for some reason */
3496 order = fls(ac->ac_o_ex.fe_len) - 1;
3497 if (order > PREALLOC_TB_SIZE - 1)
3498 /* The max size of hash table is PREALLOC_TB_SIZE */
3499 order = PREALLOC_TB_SIZE - 1;
3501 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3503 * search for the prealloc space that is having
3504 * minimal distance from the goal block.
3506 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3508 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3510 spin_lock(&pa->pa_lock);
3511 if (pa->pa_deleted == 0 &&
3512 pa->pa_free >= ac->ac_o_ex.fe_len) {
3514 cpa = ext4_mb_check_group_pa(goal_block,
3517 spin_unlock(&pa->pa_lock);
3522 ext4_mb_use_group_pa(ac, cpa);
3523 ac->ac_criteria = 20;
3530 * the function goes through all block freed in the group
3531 * but not yet committed and marks them used in in-core bitmap.
3532 * buddy must be generated from this bitmap
3533 * Need to be called with the ext4 group lock held
3535 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3539 struct ext4_group_info *grp;
3540 struct ext4_free_data *entry;
3542 grp = ext4_get_group_info(sb, group);
3543 n = rb_first(&(grp->bb_free_root));
3546 entry = rb_entry(n, struct ext4_free_data, efd_node);
3547 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3554 * the function goes through all preallocation in this group and marks them
3555 * used in in-core bitmap. buddy must be generated from this bitmap
3556 * Need to be called with ext4 group lock held
3558 static noinline_for_stack
3559 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3562 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3563 struct ext4_prealloc_space *pa;
3564 struct list_head *cur;
3565 ext4_group_t groupnr;
3566 ext4_grpblk_t start;
3567 int preallocated = 0;
3570 /* all form of preallocation discards first load group,
3571 * so the only competing code is preallocation use.
3572 * we don't need any locking here
3573 * notice we do NOT ignore preallocations with pa_deleted
3574 * otherwise we could leave used blocks available for
3575 * allocation in buddy when concurrent ext4_mb_put_pa()
3576 * is dropping preallocation
3578 list_for_each(cur, &grp->bb_prealloc_list) {
3579 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3580 spin_lock(&pa->pa_lock);
3581 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3584 spin_unlock(&pa->pa_lock);
3585 if (unlikely(len == 0))
3587 BUG_ON(groupnr != group);
3588 ext4_set_bits(bitmap, start, len);
3589 preallocated += len;
3591 mb_debug(1, "preallocated %u for group %u\n", preallocated, group);
3594 static void ext4_mb_pa_callback(struct rcu_head *head)
3596 struct ext4_prealloc_space *pa;
3597 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3599 BUG_ON(atomic_read(&pa->pa_count));
3600 BUG_ON(pa->pa_deleted == 0);
3601 kmem_cache_free(ext4_pspace_cachep, pa);
3605 * drops a reference to preallocated space descriptor
3606 * if this was the last reference and the space is consumed
3608 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3609 struct super_block *sb, struct ext4_prealloc_space *pa)
3612 ext4_fsblk_t grp_blk;
3614 /* in this short window concurrent discard can set pa_deleted */
3615 spin_lock(&pa->pa_lock);
3616 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3617 spin_unlock(&pa->pa_lock);
3621 if (pa->pa_deleted == 1) {
3622 spin_unlock(&pa->pa_lock);
3627 spin_unlock(&pa->pa_lock);
3629 grp_blk = pa->pa_pstart;
3631 * If doing group-based preallocation, pa_pstart may be in the
3632 * next group when pa is used up
3634 if (pa->pa_type == MB_GROUP_PA)
3637 grp = ext4_get_group_number(sb, grp_blk);
3642 * P1 (buddy init) P2 (regular allocation)
3643 * find block B in PA
3644 * copy on-disk bitmap to buddy
3645 * mark B in on-disk bitmap
3646 * drop PA from group
3647 * mark all PAs in buddy
3649 * thus, P1 initializes buddy with B available. to prevent this
3650 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3653 ext4_lock_group(sb, grp);
3654 list_del(&pa->pa_group_list);
3655 ext4_unlock_group(sb, grp);
3657 spin_lock(pa->pa_obj_lock);
3658 list_del_rcu(&pa->pa_inode_list);
3659 spin_unlock(pa->pa_obj_lock);
3661 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3665 * creates new preallocated space for given inode
3667 static noinline_for_stack int
3668 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3670 struct super_block *sb = ac->ac_sb;
3671 struct ext4_sb_info *sbi = EXT4_SB(sb);
3672 struct ext4_prealloc_space *pa;
3673 struct ext4_group_info *grp;
3674 struct ext4_inode_info *ei;
3676 /* preallocate only when found space is larger then requested */
3677 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3678 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3679 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3681 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3685 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3689 /* we can't allocate as much as normalizer wants.
3690 * so, found space must get proper lstart
3691 * to cover original request */
3692 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3693 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3696 * Use the below logic for adjusting best extent as it keeps
3697 * fragmentation in check while ensuring logical range of best
3698 * extent doesn't overflow out of goal extent:
3700 * 1. Check if best ex can be kept at end of goal and still
3701 * cover original start
3702 * 2. Else, check if best ex can be kept at start of goal and
3703 * still cover original start
3704 * 3. Else, keep the best ex at start of original request.
3706 new_bex_end = ac->ac_g_ex.fe_logical +
3707 EXT4_C2B(sbi, ac->ac_g_ex.fe_len);
3708 new_bex_start = new_bex_end - EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3709 if (ac->ac_o_ex.fe_logical >= new_bex_start)
3712 new_bex_start = ac->ac_g_ex.fe_logical;
3714 new_bex_start + EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3715 if (ac->ac_o_ex.fe_logical < new_bex_end)
3718 new_bex_start = ac->ac_o_ex.fe_logical;
3720 new_bex_start + EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3723 ac->ac_b_ex.fe_logical = new_bex_start;
3725 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3726 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3727 BUG_ON(new_bex_end > (ac->ac_g_ex.fe_logical +
3728 EXT4_C2B(sbi, ac->ac_g_ex.fe_len)));
3731 /* preallocation can change ac_b_ex, thus we store actually
3732 * allocated blocks for history */
3733 ac->ac_f_ex = ac->ac_b_ex;
3735 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3736 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3737 pa->pa_len = ac->ac_b_ex.fe_len;
3738 pa->pa_free = pa->pa_len;
3739 atomic_set(&pa->pa_count, 1);
3740 spin_lock_init(&pa->pa_lock);
3741 INIT_LIST_HEAD(&pa->pa_inode_list);
3742 INIT_LIST_HEAD(&pa->pa_group_list);
3744 pa->pa_type = MB_INODE_PA;
3746 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3747 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3748 trace_ext4_mb_new_inode_pa(ac, pa);
3750 ext4_mb_use_inode_pa(ac, pa);
3751 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3753 ei = EXT4_I(ac->ac_inode);
3754 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3756 pa->pa_obj_lock = &ei->i_prealloc_lock;
3757 pa->pa_inode = ac->ac_inode;
3759 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3760 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3761 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3763 spin_lock(pa->pa_obj_lock);
3764 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3765 spin_unlock(pa->pa_obj_lock);
3771 * creates new preallocated space for locality group inodes belongs to
3773 static noinline_for_stack int
3774 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3776 struct super_block *sb = ac->ac_sb;
3777 struct ext4_locality_group *lg;
3778 struct ext4_prealloc_space *pa;
3779 struct ext4_group_info *grp;
3781 /* preallocate only when found space is larger then requested */
3782 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3783 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3784 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3786 BUG_ON(ext4_pspace_cachep == NULL);
3787 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3791 /* preallocation can change ac_b_ex, thus we store actually
3792 * allocated blocks for history */
3793 ac->ac_f_ex = ac->ac_b_ex;
3795 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3796 pa->pa_lstart = pa->pa_pstart;
3797 pa->pa_len = ac->ac_b_ex.fe_len;
3798 pa->pa_free = pa->pa_len;
3799 atomic_set(&pa->pa_count, 1);
3800 spin_lock_init(&pa->pa_lock);
3801 INIT_LIST_HEAD(&pa->pa_inode_list);
3802 INIT_LIST_HEAD(&pa->pa_group_list);
3804 pa->pa_type = MB_GROUP_PA;
3806 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3807 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3808 trace_ext4_mb_new_group_pa(ac, pa);
3810 ext4_mb_use_group_pa(ac, pa);
3811 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3813 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3817 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3818 pa->pa_inode = NULL;
3820 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3821 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3822 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3825 * We will later add the new pa to the right bucket
3826 * after updating the pa_free in ext4_mb_release_context
3831 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3835 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3836 err = ext4_mb_new_group_pa(ac);
3838 err = ext4_mb_new_inode_pa(ac);
3843 * finds all unused blocks in on-disk bitmap, frees them in
3844 * in-core bitmap and buddy.
3845 * @pa must be unlinked from inode and group lists, so that
3846 * nobody else can find/use it.
3847 * the caller MUST hold group/inode locks.
3848 * TODO: optimize the case when there are no in-core structures yet
3850 static noinline_for_stack int
3851 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3852 struct ext4_prealloc_space *pa)
3854 struct super_block *sb = e4b->bd_sb;
3855 struct ext4_sb_info *sbi = EXT4_SB(sb);
3860 unsigned long long grp_blk_start;
3863 BUG_ON(pa->pa_deleted == 0);
3864 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3865 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3866 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3867 end = bit + pa->pa_len;
3870 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3873 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3874 mb_debug(1, " free preallocated %u/%u in group %u\n",
3875 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3876 (unsigned) next - bit, (unsigned) group);
3879 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3880 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3881 EXT4_C2B(sbi, bit)),
3883 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3886 if (free != pa->pa_free) {
3887 ext4_msg(e4b->bd_sb, KERN_CRIT,
3888 "pa %p: logic %lu, phys. %lu, len %lu",
3889 pa, (unsigned long) pa->pa_lstart,
3890 (unsigned long) pa->pa_pstart,
3891 (unsigned long) pa->pa_len);
3892 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3895 * pa is already deleted so we use the value obtained
3896 * from the bitmap and continue.
3899 atomic_add(free, &sbi->s_mb_discarded);
3904 static noinline_for_stack int
3905 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3906 struct ext4_prealloc_space *pa)
3908 struct super_block *sb = e4b->bd_sb;
3912 trace_ext4_mb_release_group_pa(sb, pa);
3913 BUG_ON(pa->pa_deleted == 0);
3914 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3915 if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) {
3916 ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu",
3917 e4b->bd_group, group, pa->pa_pstart);
3920 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3921 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3922 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3928 * releases all preallocations in given group
3930 * first, we need to decide discard policy:
3931 * - when do we discard
3933 * - how many do we discard
3934 * 1) how many requested
3936 static noinline_for_stack int
3937 ext4_mb_discard_group_preallocations(struct super_block *sb,
3938 ext4_group_t group, int needed)
3940 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3941 struct buffer_head *bitmap_bh = NULL;
3942 struct ext4_prealloc_space *pa, *tmp;
3943 struct list_head list;
3944 struct ext4_buddy e4b;
3949 mb_debug(1, "discard preallocation for group %u\n", group);
3951 if (list_empty(&grp->bb_prealloc_list))
3954 bitmap_bh = ext4_read_block_bitmap(sb, group);
3955 if (IS_ERR(bitmap_bh)) {
3956 err = PTR_ERR(bitmap_bh);
3957 ext4_error(sb, "Error %d reading block bitmap for %u",
3962 err = ext4_mb_load_buddy(sb, group, &e4b);
3964 ext4_warning(sb, "Error %d loading buddy information for %u",
3971 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3973 INIT_LIST_HEAD(&list);
3975 ext4_lock_group(sb, group);
3976 list_for_each_entry_safe(pa, tmp,
3977 &grp->bb_prealloc_list, pa_group_list) {
3978 spin_lock(&pa->pa_lock);
3979 if (atomic_read(&pa->pa_count)) {
3980 spin_unlock(&pa->pa_lock);
3984 if (pa->pa_deleted) {
3985 spin_unlock(&pa->pa_lock);
3989 /* seems this one can be freed ... */
3992 /* we can trust pa_free ... */
3993 free += pa->pa_free;
3995 spin_unlock(&pa->pa_lock);
3997 list_del(&pa->pa_group_list);
3998 list_add(&pa->u.pa_tmp_list, &list);
4001 /* if we still need more blocks and some PAs were used, try again */
4002 if (free < needed && busy) {
4004 ext4_unlock_group(sb, group);
4009 /* found anything to free? */
4010 if (list_empty(&list)) {
4015 /* now free all selected PAs */
4016 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4018 /* remove from object (inode or locality group) */
4019 spin_lock(pa->pa_obj_lock);
4020 list_del_rcu(&pa->pa_inode_list);
4021 spin_unlock(pa->pa_obj_lock);
4023 if (pa->pa_type == MB_GROUP_PA)
4024 ext4_mb_release_group_pa(&e4b, pa);
4026 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4028 list_del(&pa->u.pa_tmp_list);
4029 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4033 ext4_unlock_group(sb, group);
4034 ext4_mb_unload_buddy(&e4b);
4040 * releases all non-used preallocated blocks for given inode
4042 * It's important to discard preallocations under i_data_sem
4043 * We don't want another block to be served from the prealloc
4044 * space when we are discarding the inode prealloc space.
4046 * FIXME!! Make sure it is valid at all the call sites
4048 void ext4_discard_preallocations(struct inode *inode)
4050 struct ext4_inode_info *ei = EXT4_I(inode);
4051 struct super_block *sb = inode->i_sb;
4052 struct buffer_head *bitmap_bh = NULL;
4053 struct ext4_prealloc_space *pa, *tmp;
4054 ext4_group_t group = 0;
4055 struct list_head list;
4056 struct ext4_buddy e4b;
4059 if (!S_ISREG(inode->i_mode)) {
4060 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4064 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
4065 trace_ext4_discard_preallocations(inode);
4067 INIT_LIST_HEAD(&list);
4070 /* first, collect all pa's in the inode */
4071 spin_lock(&ei->i_prealloc_lock);
4072 while (!list_empty(&ei->i_prealloc_list)) {
4073 pa = list_entry(ei->i_prealloc_list.next,
4074 struct ext4_prealloc_space, pa_inode_list);
4075 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4076 spin_lock(&pa->pa_lock);
4077 if (atomic_read(&pa->pa_count)) {
4078 /* this shouldn't happen often - nobody should
4079 * use preallocation while we're discarding it */
4080 spin_unlock(&pa->pa_lock);
4081 spin_unlock(&ei->i_prealloc_lock);
4082 ext4_msg(sb, KERN_ERR,
4083 "uh-oh! used pa while discarding");
4085 schedule_timeout_uninterruptible(HZ);
4089 if (pa->pa_deleted == 0) {
4091 spin_unlock(&pa->pa_lock);
4092 list_del_rcu(&pa->pa_inode_list);
4093 list_add(&pa->u.pa_tmp_list, &list);
4097 /* someone is deleting pa right now */
4098 spin_unlock(&pa->pa_lock);
4099 spin_unlock(&ei->i_prealloc_lock);
4101 /* we have to wait here because pa_deleted
4102 * doesn't mean pa is already unlinked from
4103 * the list. as we might be called from
4104 * ->clear_inode() the inode will get freed
4105 * and concurrent thread which is unlinking
4106 * pa from inode's list may access already
4107 * freed memory, bad-bad-bad */
4109 /* XXX: if this happens too often, we can
4110 * add a flag to force wait only in case
4111 * of ->clear_inode(), but not in case of
4112 * regular truncate */
4113 schedule_timeout_uninterruptible(HZ);
4116 spin_unlock(&ei->i_prealloc_lock);
4118 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4119 BUG_ON(pa->pa_type != MB_INODE_PA);
4120 group = ext4_get_group_number(sb, pa->pa_pstart);
4122 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4123 GFP_NOFS|__GFP_NOFAIL);
4125 ext4_error(sb, "Error %d loading buddy information for %u",
4130 bitmap_bh = ext4_read_block_bitmap(sb, group);
4131 if (IS_ERR(bitmap_bh)) {
4132 err = PTR_ERR(bitmap_bh);
4133 ext4_error(sb, "Error %d reading block bitmap for %u",
4135 ext4_mb_unload_buddy(&e4b);
4139 ext4_lock_group(sb, group);
4140 list_del(&pa->pa_group_list);
4141 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4142 ext4_unlock_group(sb, group);
4144 ext4_mb_unload_buddy(&e4b);
4147 list_del(&pa->u.pa_tmp_list);
4148 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4152 #ifdef CONFIG_EXT4_DEBUG
4153 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4155 struct super_block *sb = ac->ac_sb;
4156 ext4_group_t ngroups, i;
4158 if (!ext4_mballoc_debug ||
4159 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4162 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4163 " Allocation context details:");
4164 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4165 ac->ac_status, ac->ac_flags);
4166 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4167 "goal %lu/%lu/%lu@%lu, "
4168 "best %lu/%lu/%lu@%lu cr %d",
4169 (unsigned long)ac->ac_o_ex.fe_group,
4170 (unsigned long)ac->ac_o_ex.fe_start,
4171 (unsigned long)ac->ac_o_ex.fe_len,
4172 (unsigned long)ac->ac_o_ex.fe_logical,
4173 (unsigned long)ac->ac_g_ex.fe_group,
4174 (unsigned long)ac->ac_g_ex.fe_start,
4175 (unsigned long)ac->ac_g_ex.fe_len,
4176 (unsigned long)ac->ac_g_ex.fe_logical,
4177 (unsigned long)ac->ac_b_ex.fe_group,
4178 (unsigned long)ac->ac_b_ex.fe_start,
4179 (unsigned long)ac->ac_b_ex.fe_len,
4180 (unsigned long)ac->ac_b_ex.fe_logical,
4181 (int)ac->ac_criteria);
4182 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4183 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4184 ngroups = ext4_get_groups_count(sb);
4185 for (i = 0; i < ngroups; i++) {
4186 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4187 struct ext4_prealloc_space *pa;
4188 ext4_grpblk_t start;
4189 struct list_head *cur;
4190 ext4_lock_group(sb, i);
4191 list_for_each(cur, &grp->bb_prealloc_list) {
4192 pa = list_entry(cur, struct ext4_prealloc_space,
4194 spin_lock(&pa->pa_lock);
4195 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4197 spin_unlock(&pa->pa_lock);
4198 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4201 ext4_unlock_group(sb, i);
4203 if (grp->bb_free == 0)
4205 printk(KERN_ERR "%u: %d/%d \n",
4206 i, grp->bb_free, grp->bb_fragments);
4208 printk(KERN_ERR "\n");
4211 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4218 * We use locality group preallocation for small size file. The size of the
4219 * file is determined by the current size or the resulting size after
4220 * allocation which ever is larger
4222 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4224 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4226 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4227 int bsbits = ac->ac_sb->s_blocksize_bits;
4230 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4233 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4236 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4237 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4240 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
4241 !inode_is_open_for_write(ac->ac_inode)) {
4242 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4246 if (sbi->s_mb_group_prealloc <= 0) {
4247 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4251 /* don't use group allocation for large files */
4252 size = max(size, isize);
4253 if (size > sbi->s_mb_stream_request) {
4254 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4258 BUG_ON(ac->ac_lg != NULL);
4260 * locality group prealloc space are per cpu. The reason for having
4261 * per cpu locality group is to reduce the contention between block
4262 * request from multiple CPUs.
4264 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4266 /* we're going to use group allocation */
4267 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4269 /* serialize all allocations in the group */
4270 mutex_lock(&ac->ac_lg->lg_mutex);
4273 static noinline_for_stack int
4274 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4275 struct ext4_allocation_request *ar)
4277 struct super_block *sb = ar->inode->i_sb;
4278 struct ext4_sb_info *sbi = EXT4_SB(sb);
4279 struct ext4_super_block *es = sbi->s_es;
4283 ext4_grpblk_t block;
4285 /* we can't allocate > group size */
4288 /* just a dirty hack to filter too big requests */
4289 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4290 len = EXT4_CLUSTERS_PER_GROUP(sb);
4292 /* start searching from the goal */
4294 if (goal < le32_to_cpu(es->s_first_data_block) ||
4295 goal >= ext4_blocks_count(es))
4296 goal = le32_to_cpu(es->s_first_data_block);
4297 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4299 /* set up allocation goals */
4300 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4301 ac->ac_status = AC_STATUS_CONTINUE;
4303 ac->ac_inode = ar->inode;
4304 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4305 ac->ac_o_ex.fe_group = group;
4306 ac->ac_o_ex.fe_start = block;
4307 ac->ac_o_ex.fe_len = len;
4308 ac->ac_g_ex = ac->ac_o_ex;
4309 ac->ac_flags = ar->flags;
4311 /* we have to define context: we'll we work with a file or
4312 * locality group. this is a policy, actually */
4313 ext4_mb_group_or_file(ac);
4315 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4316 "left: %u/%u, right %u/%u to %swritable\n",
4317 (unsigned) ar->len, (unsigned) ar->logical,
4318 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4319 (unsigned) ar->lleft, (unsigned) ar->pleft,
4320 (unsigned) ar->lright, (unsigned) ar->pright,
4321 inode_is_open_for_write(ar->inode) ? "" : "non-");
4326 static noinline_for_stack void
4327 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4328 struct ext4_locality_group *lg,
4329 int order, int total_entries)
4331 ext4_group_t group = 0;
4332 struct ext4_buddy e4b;
4333 struct list_head discard_list;
4334 struct ext4_prealloc_space *pa, *tmp;
4336 mb_debug(1, "discard locality group preallocation\n");
4338 INIT_LIST_HEAD(&discard_list);
4340 spin_lock(&lg->lg_prealloc_lock);
4341 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4343 spin_lock(&pa->pa_lock);
4344 if (atomic_read(&pa->pa_count)) {
4346 * This is the pa that we just used
4347 * for block allocation. So don't
4350 spin_unlock(&pa->pa_lock);
4353 if (pa->pa_deleted) {
4354 spin_unlock(&pa->pa_lock);
4357 /* only lg prealloc space */
4358 BUG_ON(pa->pa_type != MB_GROUP_PA);
4360 /* seems this one can be freed ... */
4362 spin_unlock(&pa->pa_lock);
4364 list_del_rcu(&pa->pa_inode_list);
4365 list_add(&pa->u.pa_tmp_list, &discard_list);
4368 if (total_entries <= 5) {
4370 * we want to keep only 5 entries
4371 * allowing it to grow to 8. This
4372 * mak sure we don't call discard
4373 * soon for this list.
4378 spin_unlock(&lg->lg_prealloc_lock);
4380 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4383 group = ext4_get_group_number(sb, pa->pa_pstart);
4384 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4385 GFP_NOFS|__GFP_NOFAIL);
4387 ext4_error(sb, "Error %d loading buddy information for %u",
4391 ext4_lock_group(sb, group);
4392 list_del(&pa->pa_group_list);
4393 ext4_mb_release_group_pa(&e4b, pa);
4394 ext4_unlock_group(sb, group);
4396 ext4_mb_unload_buddy(&e4b);
4397 list_del(&pa->u.pa_tmp_list);
4398 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4403 * We have incremented pa_count. So it cannot be freed at this
4404 * point. Also we hold lg_mutex. So no parallel allocation is
4405 * possible from this lg. That means pa_free cannot be updated.
4407 * A parallel ext4_mb_discard_group_preallocations is possible.
4408 * which can cause the lg_prealloc_list to be updated.
4411 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4413 int order, added = 0, lg_prealloc_count = 1;
4414 struct super_block *sb = ac->ac_sb;
4415 struct ext4_locality_group *lg = ac->ac_lg;
4416 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4418 order = fls(pa->pa_free) - 1;
4419 if (order > PREALLOC_TB_SIZE - 1)
4420 /* The max size of hash table is PREALLOC_TB_SIZE */
4421 order = PREALLOC_TB_SIZE - 1;
4422 /* Add the prealloc space to lg */
4423 spin_lock(&lg->lg_prealloc_lock);
4424 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4426 spin_lock(&tmp_pa->pa_lock);
4427 if (tmp_pa->pa_deleted) {
4428 spin_unlock(&tmp_pa->pa_lock);
4431 if (!added && pa->pa_free < tmp_pa->pa_free) {
4432 /* Add to the tail of the previous entry */
4433 list_add_tail_rcu(&pa->pa_inode_list,
4434 &tmp_pa->pa_inode_list);
4437 * we want to count the total
4438 * number of entries in the list
4441 spin_unlock(&tmp_pa->pa_lock);
4442 lg_prealloc_count++;
4445 list_add_tail_rcu(&pa->pa_inode_list,
4446 &lg->lg_prealloc_list[order]);
4447 spin_unlock(&lg->lg_prealloc_lock);
4449 /* Now trim the list to be not more than 8 elements */
4450 if (lg_prealloc_count > 8) {
4451 ext4_mb_discard_lg_preallocations(sb, lg,
4452 order, lg_prealloc_count);
4459 * release all resource we used in allocation
4461 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4463 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4464 struct ext4_prealloc_space *pa = ac->ac_pa;
4466 if (pa->pa_type == MB_GROUP_PA) {
4467 /* see comment in ext4_mb_use_group_pa() */
4468 spin_lock(&pa->pa_lock);
4469 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4470 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4471 pa->pa_free -= ac->ac_b_ex.fe_len;
4472 pa->pa_len -= ac->ac_b_ex.fe_len;
4473 spin_unlock(&pa->pa_lock);
4478 * We want to add the pa to the right bucket.
4479 * Remove it from the list and while adding
4480 * make sure the list to which we are adding
4483 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4484 spin_lock(pa->pa_obj_lock);
4485 list_del_rcu(&pa->pa_inode_list);
4486 spin_unlock(pa->pa_obj_lock);
4487 ext4_mb_add_n_trim(ac);
4489 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4491 if (ac->ac_bitmap_page)
4492 put_page(ac->ac_bitmap_page);
4493 if (ac->ac_buddy_page)
4494 put_page(ac->ac_buddy_page);
4495 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4496 mutex_unlock(&ac->ac_lg->lg_mutex);
4497 ext4_mb_collect_stats(ac);
4501 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4503 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4507 trace_ext4_mb_discard_preallocations(sb, needed);
4508 for (i = 0; i < ngroups && needed > 0; i++) {
4509 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4518 * Main entry point into mballoc to allocate blocks
4519 * it tries to use preallocation first, then falls back
4520 * to usual allocation
4522 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4523 struct ext4_allocation_request *ar, int *errp)
4526 struct ext4_allocation_context *ac = NULL;
4527 struct ext4_sb_info *sbi;
4528 struct super_block *sb;
4529 ext4_fsblk_t block = 0;
4530 unsigned int inquota = 0;
4531 unsigned int reserv_clstrs = 0;
4534 sb = ar->inode->i_sb;
4537 trace_ext4_request_blocks(ar);
4539 /* Allow to use superuser reservation for quota file */
4540 if (ext4_is_quota_file(ar->inode))
4541 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4543 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4544 /* Without delayed allocation we need to verify
4545 * there is enough free blocks to do block allocation
4546 * and verify allocation doesn't exceed the quota limits.
4549 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4551 /* let others to free the space */
4553 ar->len = ar->len >> 1;
4559 reserv_clstrs = ar->len;
4560 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4561 dquot_alloc_block_nofail(ar->inode,
4562 EXT4_C2B(sbi, ar->len));
4565 dquot_alloc_block(ar->inode,
4566 EXT4_C2B(sbi, ar->len))) {
4568 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4579 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4586 *errp = ext4_mb_initialize_context(ac, ar);
4592 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4593 if (!ext4_mb_use_preallocated(ac)) {
4594 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4595 ext4_mb_normalize_request(ac, ar);
4597 /* allocate space in core */
4598 *errp = ext4_mb_regular_allocator(ac);
4600 goto discard_and_exit;
4602 /* as we've just preallocated more space than
4603 * user requested originally, we store allocated
4604 * space in a special descriptor */
4605 if (ac->ac_status == AC_STATUS_FOUND &&
4606 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4607 *errp = ext4_mb_new_preallocation(ac);
4610 ext4_discard_allocated_blocks(ac);
4614 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4615 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4617 ext4_discard_allocated_blocks(ac);
4620 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4621 ar->len = ac->ac_b_ex.fe_len;
4624 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4632 ac->ac_b_ex.fe_len = 0;
4634 ext4_mb_show_ac(ac);
4636 ext4_mb_release_context(ac);
4639 kmem_cache_free(ext4_ac_cachep, ac);
4640 if (inquota && ar->len < inquota)
4641 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4643 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4644 /* release all the reserved blocks if non delalloc */
4645 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4649 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4655 * We can merge two free data extents only if the physical blocks
4656 * are contiguous, AND the extents were freed by the same transaction,
4657 * AND the blocks are associated with the same group.
4659 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
4660 struct ext4_free_data *entry,
4661 struct ext4_free_data *new_entry,
4662 struct rb_root *entry_rb_root)
4664 if ((entry->efd_tid != new_entry->efd_tid) ||
4665 (entry->efd_group != new_entry->efd_group))
4667 if (entry->efd_start_cluster + entry->efd_count ==
4668 new_entry->efd_start_cluster) {
4669 new_entry->efd_start_cluster = entry->efd_start_cluster;
4670 new_entry->efd_count += entry->efd_count;
4671 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
4672 entry->efd_start_cluster) {
4673 new_entry->efd_count += entry->efd_count;
4676 spin_lock(&sbi->s_md_lock);
4677 list_del(&entry->efd_list);
4678 spin_unlock(&sbi->s_md_lock);
4679 rb_erase(&entry->efd_node, entry_rb_root);
4680 kmem_cache_free(ext4_free_data_cachep, entry);
4683 static noinline_for_stack int
4684 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4685 struct ext4_free_data *new_entry)
4687 ext4_group_t group = e4b->bd_group;
4688 ext4_grpblk_t cluster;
4689 ext4_grpblk_t clusters = new_entry->efd_count;
4690 struct ext4_free_data *entry;
4691 struct ext4_group_info *db = e4b->bd_info;
4692 struct super_block *sb = e4b->bd_sb;
4693 struct ext4_sb_info *sbi = EXT4_SB(sb);
4694 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4695 struct rb_node *parent = NULL, *new_node;
4697 BUG_ON(!ext4_handle_valid(handle));
4698 BUG_ON(e4b->bd_bitmap_page == NULL);
4699 BUG_ON(e4b->bd_buddy_page == NULL);
4701 new_node = &new_entry->efd_node;
4702 cluster = new_entry->efd_start_cluster;
4705 /* first free block exent. We need to
4706 protect buddy cache from being freed,
4707 * otherwise we'll refresh it from
4708 * on-disk bitmap and lose not-yet-available
4710 get_page(e4b->bd_buddy_page);
4711 get_page(e4b->bd_bitmap_page);
4715 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4716 if (cluster < entry->efd_start_cluster)
4718 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4719 n = &(*n)->rb_right;
4721 ext4_grp_locked_error(sb, group, 0,
4722 ext4_group_first_block_no(sb, group) +
4723 EXT4_C2B(sbi, cluster),
4724 "Block already on to-be-freed list");
4725 kmem_cache_free(ext4_free_data_cachep, new_entry);
4730 rb_link_node(new_node, parent, n);
4731 rb_insert_color(new_node, &db->bb_free_root);
4733 /* Now try to see the extent can be merged to left and right */
4734 node = rb_prev(new_node);
4736 entry = rb_entry(node, struct ext4_free_data, efd_node);
4737 ext4_try_merge_freed_extent(sbi, entry, new_entry,
4738 &(db->bb_free_root));
4741 node = rb_next(new_node);
4743 entry = rb_entry(node, struct ext4_free_data, efd_node);
4744 ext4_try_merge_freed_extent(sbi, entry, new_entry,
4745 &(db->bb_free_root));
4748 spin_lock(&sbi->s_md_lock);
4749 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
4750 sbi->s_mb_free_pending += clusters;
4751 spin_unlock(&sbi->s_md_lock);
4756 * ext4_free_blocks() -- Free given blocks and update quota
4757 * @handle: handle for this transaction
4759 * @bh: optional buffer of the block to be freed
4760 * @block: starting physical block to be freed
4761 * @count: number of blocks to be freed
4762 * @flags: flags used by ext4_free_blocks
4764 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4765 struct buffer_head *bh, ext4_fsblk_t block,
4766 unsigned long count, int flags)
4768 struct buffer_head *bitmap_bh = NULL;
4769 struct super_block *sb = inode->i_sb;
4770 struct ext4_group_desc *gdp;
4771 unsigned int overflow;
4773 struct buffer_head *gd_bh;
4774 ext4_group_t block_group;
4775 struct ext4_sb_info *sbi;
4776 struct ext4_buddy e4b;
4777 unsigned int count_clusters;
4784 BUG_ON(block != bh->b_blocknr);
4786 block = bh->b_blocknr;
4790 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4791 !ext4_data_block_valid(sbi, block, count)) {
4792 ext4_error(sb, "Freeing blocks not in datazone - "
4793 "block = %llu, count = %lu", block, count);
4797 ext4_debug("freeing block %llu\n", block);
4798 trace_ext4_free_blocks(inode, block, count, flags);
4800 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4803 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4808 * If the extent to be freed does not begin on a cluster
4809 * boundary, we need to deal with partial clusters at the
4810 * beginning and end of the extent. Normally we will free
4811 * blocks at the beginning or the end unless we are explicitly
4812 * requested to avoid doing so.
4814 overflow = EXT4_PBLK_COFF(sbi, block);
4816 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4817 overflow = sbi->s_cluster_ratio - overflow;
4819 if (count > overflow)
4828 overflow = EXT4_LBLK_COFF(sbi, count);
4830 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4831 if (count > overflow)
4836 count += sbi->s_cluster_ratio - overflow;
4839 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4841 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
4843 for (i = 0; i < count; i++) {
4846 bh = sb_find_get_block(inode->i_sb, block + i);
4847 ext4_forget(handle, is_metadata, inode, bh, block + i);
4853 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4855 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4856 ext4_get_group_info(sb, block_group))))
4860 * Check to see if we are freeing blocks across a group
4863 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4864 overflow = EXT4_C2B(sbi, bit) + count -
4865 EXT4_BLOCKS_PER_GROUP(sb);
4868 count_clusters = EXT4_NUM_B2C(sbi, count);
4869 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4870 if (IS_ERR(bitmap_bh)) {
4871 err = PTR_ERR(bitmap_bh);
4875 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4881 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4882 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4883 in_range(block, ext4_inode_table(sb, gdp),
4884 sbi->s_itb_per_group) ||
4885 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4886 sbi->s_itb_per_group)) {
4888 ext4_error(sb, "Freeing blocks in system zone - "
4889 "Block = %llu, count = %lu", block, count);
4890 /* err = 0. ext4_std_error should be a no op */
4894 BUFFER_TRACE(bitmap_bh, "getting write access");
4895 err = ext4_journal_get_write_access(handle, bitmap_bh);
4900 * We are about to modify some metadata. Call the journal APIs
4901 * to unshare ->b_data if a currently-committing transaction is
4904 BUFFER_TRACE(gd_bh, "get_write_access");
4905 err = ext4_journal_get_write_access(handle, gd_bh);
4908 #ifdef AGGRESSIVE_CHECK
4911 for (i = 0; i < count_clusters; i++)
4912 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4915 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4917 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4918 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
4919 GFP_NOFS|__GFP_NOFAIL);
4924 * We need to make sure we don't reuse the freed block until after the
4925 * transaction is committed. We make an exception if the inode is to be
4926 * written in writeback mode since writeback mode has weak data
4927 * consistency guarantees.
4929 if (ext4_handle_valid(handle) &&
4930 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
4931 !ext4_should_writeback_data(inode))) {
4932 struct ext4_free_data *new_entry;
4934 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4937 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4938 GFP_NOFS|__GFP_NOFAIL);
4939 new_entry->efd_start_cluster = bit;
4940 new_entry->efd_group = block_group;
4941 new_entry->efd_count = count_clusters;
4942 new_entry->efd_tid = handle->h_transaction->t_tid;
4944 ext4_lock_group(sb, block_group);
4945 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4946 ext4_mb_free_metadata(handle, &e4b, new_entry);
4948 /* need to update group_info->bb_free and bitmap
4949 * with group lock held. generate_buddy look at
4950 * them with group lock_held
4952 if (test_opt(sb, DISCARD)) {
4953 err = ext4_issue_discard(sb, block_group, bit,
4954 count_clusters, NULL);
4955 if (err && err != -EOPNOTSUPP)
4956 ext4_msg(sb, KERN_WARNING, "discard request in"
4957 " group:%d block:%d count:%lu failed"
4958 " with %d", block_group, bit, count,
4961 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4963 ext4_lock_group(sb, block_group);
4964 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4965 mb_free_blocks(inode, &e4b, bit, count_clusters);
4968 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4969 ext4_free_group_clusters_set(sb, gdp, ret);
4970 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4971 ext4_group_desc_csum_set(sb, block_group, gdp);
4972 ext4_unlock_group(sb, block_group);
4974 if (sbi->s_log_groups_per_flex) {
4975 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4976 atomic64_add(count_clusters,
4977 &sbi_array_rcu_deref(sbi, s_flex_groups,
4978 flex_group)->free_clusters);
4982 * on a bigalloc file system, defer the s_freeclusters_counter
4983 * update to the caller (ext4_remove_space and friends) so they
4984 * can determine if a cluster freed here should be rereserved
4986 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
4987 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4988 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4989 percpu_counter_add(&sbi->s_freeclusters_counter,
4993 ext4_mb_unload_buddy(&e4b);
4995 /* We dirtied the bitmap block */
4996 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4997 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4999 /* And the group descriptor block */
5000 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5001 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5005 if (overflow && !err) {
5013 ext4_std_error(sb, err);
5018 * ext4_group_add_blocks() -- Add given blocks to an existing group
5019 * @handle: handle to this transaction
5021 * @block: start physical block to add to the block group
5022 * @count: number of blocks to free
5024 * This marks the blocks as free in the bitmap and buddy.
5026 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
5027 ext4_fsblk_t block, unsigned long count)
5029 struct buffer_head *bitmap_bh = NULL;
5030 struct buffer_head *gd_bh;
5031 ext4_group_t block_group;
5034 struct ext4_group_desc *desc;
5035 struct ext4_sb_info *sbi = EXT4_SB(sb);
5036 struct ext4_buddy e4b;
5037 int err = 0, ret, free_clusters_count;
5038 ext4_grpblk_t clusters_freed;
5039 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
5040 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
5041 unsigned long cluster_count = last_cluster - first_cluster + 1;
5043 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
5048 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5050 * Check to see if we are freeing blocks across a group
5053 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
5054 ext4_warning(sb, "too many blocks added to group %u",
5060 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5061 if (IS_ERR(bitmap_bh)) {
5062 err = PTR_ERR(bitmap_bh);
5067 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
5073 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
5074 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
5075 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
5076 in_range(block + count - 1, ext4_inode_table(sb, desc),
5077 sbi->s_itb_per_group)) {
5078 ext4_error(sb, "Adding blocks in system zones - "
5079 "Block = %llu, count = %lu",
5085 BUFFER_TRACE(bitmap_bh, "getting write access");
5086 err = ext4_journal_get_write_access(handle, bitmap_bh);
5091 * We are about to modify some metadata. Call the journal APIs
5092 * to unshare ->b_data if a currently-committing transaction is
5095 BUFFER_TRACE(gd_bh, "get_write_access");
5096 err = ext4_journal_get_write_access(handle, gd_bh);
5100 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
5101 BUFFER_TRACE(bitmap_bh, "clear bit");
5102 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
5103 ext4_error(sb, "bit already cleared for block %llu",
5104 (ext4_fsblk_t)(block + i));
5105 BUFFER_TRACE(bitmap_bh, "bit already cleared");
5111 err = ext4_mb_load_buddy(sb, block_group, &e4b);
5116 * need to update group_info->bb_free and bitmap
5117 * with group lock held. generate_buddy look at
5118 * them with group lock_held
5120 ext4_lock_group(sb, block_group);
5121 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
5122 mb_free_blocks(NULL, &e4b, bit, cluster_count);
5123 free_clusters_count = clusters_freed +
5124 ext4_free_group_clusters(sb, desc);
5125 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
5126 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5127 ext4_group_desc_csum_set(sb, block_group, desc);
5128 ext4_unlock_group(sb, block_group);
5129 percpu_counter_add(&sbi->s_freeclusters_counter,
5132 if (sbi->s_log_groups_per_flex) {
5133 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5134 atomic64_add(clusters_freed,
5135 &sbi_array_rcu_deref(sbi, s_flex_groups,
5136 flex_group)->free_clusters);
5139 ext4_mb_unload_buddy(&e4b);
5141 /* We dirtied the bitmap block */
5142 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5143 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5145 /* And the group descriptor block */
5146 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5147 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5153 ext4_std_error(sb, err);
5158 * ext4_trim_extent -- function to TRIM one single free extent in the group
5159 * @sb: super block for the file system
5160 * @start: starting block of the free extent in the alloc. group
5161 * @count: number of blocks to TRIM
5162 * @group: alloc. group we are working with
5163 * @e4b: ext4 buddy for the group
5165 * Trim "count" blocks starting at "start" in the "group". To assure that no
5166 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5167 * be called with under the group lock.
5169 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5170 ext4_group_t group, struct ext4_buddy *e4b)
5174 struct ext4_free_extent ex;
5177 trace_ext4_trim_extent(sb, group, start, count);
5179 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5181 ex.fe_start = start;
5182 ex.fe_group = group;
5186 * Mark blocks used, so no one can reuse them while
5189 mb_mark_used(e4b, &ex);
5190 ext4_unlock_group(sb, group);
5191 ret = ext4_issue_discard(sb, group, start, count, NULL);
5192 ext4_lock_group(sb, group);
5193 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5198 * ext4_trim_all_free -- function to trim all free space in alloc. group
5199 * @sb: super block for file system
5200 * @group: group to be trimmed
5201 * @start: first group block to examine
5202 * @max: last group block to examine
5203 * @minblocks: minimum extent block count
5205 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5206 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5210 * ext4_trim_all_free walks through group's block bitmap searching for free
5211 * extents. When the free extent is found, mark it as used in group buddy
5212 * bitmap. Then issue a TRIM command on this extent and free the extent in
5213 * the group buddy bitmap. This is done until whole group is scanned.
5215 static ext4_grpblk_t
5216 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5217 ext4_grpblk_t start, ext4_grpblk_t max,
5218 ext4_grpblk_t minblocks)
5221 ext4_grpblk_t next, count = 0, free_count = 0;
5222 struct ext4_buddy e4b;
5225 trace_ext4_trim_all_free(sb, group, start, max);
5227 ret = ext4_mb_load_buddy(sb, group, &e4b);
5229 ext4_warning(sb, "Error %d loading buddy information for %u",
5233 bitmap = e4b.bd_bitmap;
5235 ext4_lock_group(sb, group);
5236 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5237 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5240 start = (e4b.bd_info->bb_first_free > start) ?
5241 e4b.bd_info->bb_first_free : start;
5243 while (start <= max) {
5244 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5247 next = mb_find_next_bit(bitmap, max + 1, start);
5249 if ((next - start) >= minblocks) {
5250 ret = ext4_trim_extent(sb, start,
5251 next - start, group, &e4b);
5252 if (ret && ret != -EOPNOTSUPP)
5255 count += next - start;
5257 free_count += next - start;
5260 if (fatal_signal_pending(current)) {
5261 count = -ERESTARTSYS;
5265 if (need_resched()) {
5266 ext4_unlock_group(sb, group);
5268 ext4_lock_group(sb, group);
5271 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5277 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5280 ext4_unlock_group(sb, group);
5281 ext4_mb_unload_buddy(&e4b);
5283 ext4_debug("trimmed %d blocks in the group %d\n",
5290 * ext4_trim_fs() -- trim ioctl handle function
5291 * @sb: superblock for filesystem
5292 * @range: fstrim_range structure
5294 * start: First Byte to trim
5295 * len: number of Bytes to trim from start
5296 * minlen: minimum extent length in Bytes
5297 * ext4_trim_fs goes through all allocation groups containing Bytes from
5298 * start to start+len. For each such a group ext4_trim_all_free function
5299 * is invoked to trim all free space.
5301 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5303 struct request_queue *q = bdev_get_queue(sb->s_bdev);
5304 struct ext4_group_info *grp;
5305 ext4_group_t group, first_group, last_group;
5306 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5307 uint64_t start, end, minlen, trimmed = 0;
5308 ext4_fsblk_t first_data_blk =
5309 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5310 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5313 start = range->start >> sb->s_blocksize_bits;
5314 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5315 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5316 range->minlen >> sb->s_blocksize_bits);
5318 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5319 start >= max_blks ||
5320 range->len < sb->s_blocksize)
5322 /* No point to try to trim less than discard granularity */
5323 if (range->minlen < q->limits.discard_granularity) {
5324 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5325 q->limits.discard_granularity >> sb->s_blocksize_bits);
5326 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
5329 if (end >= max_blks)
5331 if (end <= first_data_blk)
5333 if (start < first_data_blk)
5334 start = first_data_blk;
5336 /* Determine first and last group to examine based on start and end */
5337 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5338 &first_group, &first_cluster);
5339 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5340 &last_group, &last_cluster);
5342 /* end now represents the last cluster to discard in this group */
5343 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5345 for (group = first_group; group <= last_group; group++) {
5346 grp = ext4_get_group_info(sb, group);
5347 /* We only do this if the grp has never been initialized */
5348 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5349 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5355 * For all the groups except the last one, last cluster will
5356 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5357 * change it for the last group, note that last_cluster is
5358 * already computed earlier by ext4_get_group_no_and_offset()
5360 if (group == last_group)
5363 if (grp->bb_free >= minlen) {
5364 cnt = ext4_trim_all_free(sb, group, first_cluster,
5374 * For every group except the first one, we are sure
5375 * that the first cluster to discard will be cluster #0.
5381 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5384 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5388 /* Iterate all the free extents in the group. */
5390 ext4_mballoc_query_range(
5391 struct super_block *sb,
5393 ext4_grpblk_t start,
5395 ext4_mballoc_query_range_fn formatter,
5400 struct ext4_buddy e4b;
5403 error = ext4_mb_load_buddy(sb, group, &e4b);
5406 bitmap = e4b.bd_bitmap;
5408 ext4_lock_group(sb, group);
5410 start = (e4b.bd_info->bb_first_free > start) ?
5411 e4b.bd_info->bb_first_free : start;
5412 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
5413 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5415 while (start <= end) {
5416 start = mb_find_next_zero_bit(bitmap, end + 1, start);
5419 next = mb_find_next_bit(bitmap, end + 1, start);
5421 ext4_unlock_group(sb, group);
5422 error = formatter(sb, group, start, next - start, priv);
5425 ext4_lock_group(sb, group);
5430 ext4_unlock_group(sb, group);
5432 ext4_mb_unload_buddy(&e4b);