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>
23 * - test ext4_ext_search_left() and ext4_ext_search_right()
24 * - search for metadata in few groups
27 * - normalization should take into account whether file is still open
28 * - discard preallocations if no free space left (policy?)
29 * - don't normalize tails
31 * - reservation for superuser
34 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
35 * - track min/max extents in each group for better group selection
36 * - mb_mark_used() may allocate chunk right after splitting buddy
37 * - tree of groups sorted by number of free blocks
42 * The allocation request involve request for multiple number of blocks
43 * near to the goal(block) value specified.
45 * During initialization phase of the allocator we decide to use the
46 * group preallocation or inode preallocation depending on the size of
47 * the file. The size of the file could be the resulting file size we
48 * would have after allocation, or the current file size, which ever
49 * is larger. If the size is less than sbi->s_mb_stream_request we
50 * select to use the group preallocation. The default value of
51 * s_mb_stream_request is 16 blocks. This can also be tuned via
52 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
53 * terms of number of blocks.
55 * The main motivation for having small file use group preallocation is to
56 * ensure that we have small files closer together on the disk.
58 * First stage the allocator looks at the inode prealloc list,
59 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
60 * spaces for this particular inode. The inode prealloc space is
63 * pa_lstart -> the logical start block for this prealloc space
64 * pa_pstart -> the physical start block for this prealloc space
65 * pa_len -> length for this prealloc space (in clusters)
66 * pa_free -> free space available in this prealloc space (in clusters)
68 * The inode preallocation space is used looking at the _logical_ start
69 * block. If only the logical file block falls within the range of prealloc
70 * space we will consume the particular prealloc space. This makes sure that
71 * we have contiguous physical blocks representing the file blocks
73 * The important thing to be noted in case of inode prealloc space is that
74 * we don't modify the values associated to inode prealloc space except
77 * If we are not able to find blocks in the inode prealloc space and if we
78 * have the group allocation flag set then we look at the locality group
79 * prealloc space. These are per CPU prealloc list represented as
81 * ext4_sb_info.s_locality_groups[smp_processor_id()]
83 * The reason for having a per cpu locality group is to reduce the contention
84 * between CPUs. It is possible to get scheduled at this point.
86 * The locality group prealloc space is used looking at whether we have
87 * enough free space (pa_free) within the prealloc space.
89 * If we can't allocate blocks via inode prealloc or/and locality group
90 * prealloc then we look at the buddy cache. The buddy cache is represented
91 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
92 * mapped to the buddy and bitmap information regarding different
93 * groups. The buddy information is attached to buddy cache inode so that
94 * we can access them through the page cache. The information regarding
95 * each group is loaded via ext4_mb_load_buddy. The information involve
96 * block bitmap and buddy information. The information are stored in the
100 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
103 * one block each for bitmap and buddy information. So for each group we
104 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
105 * blocksize) blocks. So it can have information regarding groups_per_page
106 * which is blocks_per_page/2
108 * The buddy cache inode is not stored on disk. The inode is thrown
109 * away when the filesystem is unmounted.
111 * We look for count number of blocks in the buddy cache. If we were able
112 * to locate that many free blocks we return with additional information
113 * regarding rest of the contiguous physical block available
115 * Before allocating blocks via buddy cache we normalize the request
116 * blocks. This ensure we ask for more blocks that we needed. The extra
117 * blocks that we get after allocation is added to the respective prealloc
118 * list. In case of inode preallocation we follow a list of heuristics
119 * based on file size. This can be found in ext4_mb_normalize_request. If
120 * we are doing a group prealloc we try to normalize the request to
121 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
122 * dependent on the cluster size; for non-bigalloc file systems, it is
123 * 512 blocks. This can be tuned via
124 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
125 * terms of number of blocks. If we have mounted the file system with -O
126 * stripe=<value> option the group prealloc request is normalized to the
127 * the smallest multiple of the stripe value (sbi->s_stripe) which is
128 * greater than the default mb_group_prealloc.
130 * The regular allocator (using the buddy cache) supports a few tunables.
132 * /sys/fs/ext4/<partition>/mb_min_to_scan
133 * /sys/fs/ext4/<partition>/mb_max_to_scan
134 * /sys/fs/ext4/<partition>/mb_order2_req
136 * The regular allocator uses buddy scan only if the request len is power of
137 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
138 * value of s_mb_order2_reqs can be tuned via
139 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
140 * stripe size (sbi->s_stripe), we try to search for contiguous block in
141 * stripe size. This should result in better allocation on RAID setups. If
142 * not, we search in the specific group using bitmap for best extents. The
143 * tunable min_to_scan and max_to_scan control the behaviour here.
144 * min_to_scan indicate how long the mballoc __must__ look for a best
145 * extent and max_to_scan indicates how long the mballoc __can__ look for a
146 * best extent in the found extents. Searching for the blocks starts with
147 * the group specified as the goal value in allocation context via
148 * ac_g_ex. Each group is first checked based on the criteria whether it
149 * can be used for allocation. ext4_mb_good_group explains how the groups are
152 * Both the prealloc space are getting populated as above. So for the first
153 * request we will hit the buddy cache which will result in this prealloc
154 * space getting filled. The prealloc space is then later used for the
155 * subsequent request.
159 * mballoc operates on the following data:
161 * - in-core buddy (actually includes buddy and bitmap)
162 * - preallocation descriptors (PAs)
164 * there are two types of preallocations:
166 * assiged to specific inode and can be used for this inode only.
167 * it describes part of inode's space preallocated to specific
168 * physical blocks. any block from that preallocated can be used
169 * independent. the descriptor just tracks number of blocks left
170 * unused. so, before taking some block from descriptor, one must
171 * make sure corresponded logical block isn't allocated yet. this
172 * also means that freeing any block within descriptor's range
173 * must discard all preallocated blocks.
175 * assigned to specific locality group which does not translate to
176 * permanent set of inodes: inode can join and leave group. space
177 * from this type of preallocation can be used for any inode. thus
178 * it's consumed from the beginning to the end.
180 * relation between them can be expressed as:
181 * in-core buddy = on-disk bitmap + preallocation descriptors
183 * this mean blocks mballoc considers used are:
184 * - allocated blocks (persistent)
185 * - preallocated blocks (non-persistent)
187 * consistency in mballoc world means that at any time a block is either
188 * free or used in ALL structures. notice: "any time" should not be read
189 * literally -- time is discrete and delimited by locks.
191 * to keep it simple, we don't use block numbers, instead we count number of
192 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
194 * all operations can be expressed as:
195 * - init buddy: buddy = on-disk + PAs
196 * - new PA: buddy += N; PA = N
197 * - use inode PA: on-disk += N; PA -= N
198 * - discard inode PA buddy -= on-disk - PA; PA = 0
199 * - use locality group PA on-disk += N; PA -= N
200 * - discard locality group PA buddy -= PA; PA = 0
201 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
202 * is used in real operation because we can't know actual used
203 * bits from PA, only from on-disk bitmap
205 * if we follow this strict logic, then all operations above should be atomic.
206 * given some of them can block, we'd have to use something like semaphores
207 * killing performance on high-end SMP hardware. let's try to relax it using
208 * the following knowledge:
209 * 1) if buddy is referenced, it's already initialized
210 * 2) while block is used in buddy and the buddy is referenced,
211 * nobody can re-allocate that block
212 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
213 * bit set and PA claims same block, it's OK. IOW, one can set bit in
214 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
217 * so, now we're building a concurrency table:
220 * blocks for PA are allocated in the buddy, buddy must be referenced
221 * until PA is linked to allocation group to avoid concurrent buddy init
223 * we need to make sure that either on-disk bitmap or PA has uptodate data
224 * given (3) we care that PA-=N operation doesn't interfere with init
226 * the simplest way would be to have buddy initialized by the discard
227 * - use locality group PA
228 * again PA-=N must be serialized with init
229 * - discard locality group PA
230 * the simplest way would be to have buddy initialized by the discard
233 * i_data_sem serializes them
235 * discard process must wait until PA isn't used by another process
236 * - use locality group PA
237 * some mutex should serialize them
238 * - discard locality group PA
239 * discard process must wait until PA isn't used by another process
242 * i_data_sem or another mutex should serializes them
244 * discard process must wait until PA isn't used by another process
245 * - use locality group PA
246 * nothing wrong here -- they're different PAs covering different blocks
247 * - discard locality group PA
248 * discard process must wait until PA isn't used by another process
250 * now we're ready to make few consequences:
251 * - PA is referenced and while it is no discard is possible
252 * - PA is referenced until block isn't marked in on-disk bitmap
253 * - PA changes only after on-disk bitmap
254 * - discard must not compete with init. either init is done before
255 * any discard or they're serialized somehow
256 * - buddy init as sum of on-disk bitmap and PAs is done atomically
258 * a special case when we've used PA to emptiness. no need to modify buddy
259 * in this case, but we should care about concurrent init
264 * Logic in few words:
269 * mark bits in on-disk bitmap
272 * - use preallocation:
273 * find proper PA (per-inode or group)
275 * mark bits in on-disk bitmap
281 * mark bits in on-disk bitmap
284 * - discard preallocations in group:
286 * move them onto local list
287 * load on-disk bitmap
289 * remove PA from object (inode or locality group)
290 * mark free blocks in-core
292 * - discard inode's preallocations:
299 * - bitlock on a group (group)
300 * - object (inode/locality) (object)
311 * - release consumed pa:
316 * - generate in-core bitmap:
320 * - discard all for given object (inode, locality group):
325 * - discard all for given group:
332 static struct kmem_cache *ext4_pspace_cachep;
333 static struct kmem_cache *ext4_ac_cachep;
334 static struct kmem_cache *ext4_free_data_cachep;
336 /* We create slab caches for groupinfo data structures based on the
337 * superblock block size. There will be one per mounted filesystem for
338 * each unique s_blocksize_bits */
339 #define NR_GRPINFO_CACHES 8
340 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
342 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
343 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
344 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
345 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
348 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
350 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
352 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac);
355 * The algorithm using this percpu seq counter goes below:
356 * 1. We sample the percpu discard_pa_seq counter before trying for block
357 * allocation in ext4_mb_new_blocks().
358 * 2. We increment this percpu discard_pa_seq counter when we either allocate
359 * or free these blocks i.e. while marking those blocks as used/free in
360 * mb_mark_used()/mb_free_blocks().
361 * 3. We also increment this percpu seq counter when we successfully identify
362 * that the bb_prealloc_list is not empty and hence proceed for discarding
363 * of those PAs inside ext4_mb_discard_group_preallocations().
365 * Now to make sure that the regular fast path of block allocation is not
366 * affected, as a small optimization we only sample the percpu seq counter
367 * on that cpu. Only when the block allocation fails and when freed blocks
368 * found were 0, that is when we sample percpu seq counter for all cpus using
369 * below function ext4_get_discard_pa_seq_sum(). This happens after making
370 * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
372 static DEFINE_PER_CPU(u64, discard_pa_seq);
373 static inline u64 ext4_get_discard_pa_seq_sum(void)
378 for_each_possible_cpu(__cpu)
379 __seq += per_cpu(discard_pa_seq, __cpu);
383 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
385 #if BITS_PER_LONG == 64
386 *bit += ((unsigned long) addr & 7UL) << 3;
387 addr = (void *) ((unsigned long) addr & ~7UL);
388 #elif BITS_PER_LONG == 32
389 *bit += ((unsigned long) addr & 3UL) << 3;
390 addr = (void *) ((unsigned long) addr & ~3UL);
392 #error "how many bits you are?!"
397 static inline int mb_test_bit(int bit, void *addr)
400 * ext4_test_bit on architecture like powerpc
401 * needs unsigned long aligned address
403 addr = mb_correct_addr_and_bit(&bit, addr);
404 return ext4_test_bit(bit, addr);
407 static inline void mb_set_bit(int bit, void *addr)
409 addr = mb_correct_addr_and_bit(&bit, addr);
410 ext4_set_bit(bit, addr);
413 static inline void mb_clear_bit(int bit, void *addr)
415 addr = mb_correct_addr_and_bit(&bit, addr);
416 ext4_clear_bit(bit, addr);
419 static inline int mb_test_and_clear_bit(int bit, void *addr)
421 addr = mb_correct_addr_and_bit(&bit, addr);
422 return ext4_test_and_clear_bit(bit, addr);
425 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
427 int fix = 0, ret, tmpmax;
428 addr = mb_correct_addr_and_bit(&fix, addr);
432 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
438 static inline int mb_find_next_bit(void *addr, int max, int start)
440 int fix = 0, ret, tmpmax;
441 addr = mb_correct_addr_and_bit(&fix, addr);
445 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
451 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
455 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
458 if (order > e4b->bd_blkbits + 1) {
463 /* at order 0 we see each particular block */
465 *max = 1 << (e4b->bd_blkbits + 3);
466 return e4b->bd_bitmap;
469 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
470 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
476 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
477 int first, int count)
480 struct super_block *sb = e4b->bd_sb;
482 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
484 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
485 for (i = 0; i < count; i++) {
486 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
487 ext4_fsblk_t blocknr;
489 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
490 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
491 ext4_grp_locked_error(sb, e4b->bd_group,
492 inode ? inode->i_ino : 0,
494 "freeing block already freed "
497 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
498 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
500 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
504 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
508 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
510 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
511 for (i = 0; i < count; i++) {
512 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
513 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
517 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
519 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
521 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
522 unsigned char *b1, *b2;
524 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
525 b2 = (unsigned char *) bitmap;
526 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
527 if (b1[i] != b2[i]) {
528 ext4_msg(e4b->bd_sb, KERN_ERR,
529 "corruption in group %u "
530 "at byte %u(%u): %x in copy != %x "
532 e4b->bd_group, i, i * 8, b1[i], b2[i]);
539 static void mb_group_bb_bitmap_alloc(struct super_block *sb,
540 struct ext4_group_info *grp, ext4_group_t group)
542 struct buffer_head *bh;
544 grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
548 bh = ext4_read_block_bitmap(sb, group);
549 if (IS_ERR_OR_NULL(bh)) {
550 kfree(grp->bb_bitmap);
551 grp->bb_bitmap = NULL;
555 memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
559 static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
561 kfree(grp->bb_bitmap);
565 static inline void mb_free_blocks_double(struct inode *inode,
566 struct ext4_buddy *e4b, int first, int count)
570 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
571 int first, int count)
575 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
580 static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
581 struct ext4_group_info *grp, ext4_group_t group)
586 static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
592 #ifdef AGGRESSIVE_CHECK
594 #define MB_CHECK_ASSERT(assert) \
598 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
599 function, file, line, # assert); \
604 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
605 const char *function, int line)
607 struct super_block *sb = e4b->bd_sb;
608 int order = e4b->bd_blkbits + 1;
615 struct ext4_group_info *grp;
618 struct list_head *cur;
623 static int mb_check_counter;
624 if (mb_check_counter++ % 100 != 0)
629 buddy = mb_find_buddy(e4b, order, &max);
630 MB_CHECK_ASSERT(buddy);
631 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
632 MB_CHECK_ASSERT(buddy2);
633 MB_CHECK_ASSERT(buddy != buddy2);
634 MB_CHECK_ASSERT(max * 2 == max2);
637 for (i = 0; i < max; i++) {
639 if (mb_test_bit(i, buddy)) {
640 /* only single bit in buddy2 may be 1 */
641 if (!mb_test_bit(i << 1, buddy2)) {
643 mb_test_bit((i<<1)+1, buddy2));
644 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
646 mb_test_bit(i << 1, buddy2));
651 /* both bits in buddy2 must be 1 */
652 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
653 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
655 for (j = 0; j < (1 << order); j++) {
656 k = (i * (1 << order)) + j;
658 !mb_test_bit(k, e4b->bd_bitmap));
662 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
667 buddy = mb_find_buddy(e4b, 0, &max);
668 for (i = 0; i < max; i++) {
669 if (!mb_test_bit(i, buddy)) {
670 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
678 /* check used bits only */
679 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
680 buddy2 = mb_find_buddy(e4b, j, &max2);
682 MB_CHECK_ASSERT(k < max2);
683 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
686 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
687 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
689 grp = ext4_get_group_info(sb, e4b->bd_group);
690 list_for_each(cur, &grp->bb_prealloc_list) {
691 ext4_group_t groupnr;
692 struct ext4_prealloc_space *pa;
693 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
694 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
695 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
696 for (i = 0; i < pa->pa_len; i++)
697 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
701 #undef MB_CHECK_ASSERT
702 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
703 __FILE__, __func__, __LINE__)
705 #define mb_check_buddy(e4b)
709 * Divide blocks started from @first with length @len into
710 * smaller chunks with power of 2 blocks.
711 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
712 * then increase bb_counters[] for corresponded chunk size.
714 static void ext4_mb_mark_free_simple(struct super_block *sb,
715 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
716 struct ext4_group_info *grp)
718 struct ext4_sb_info *sbi = EXT4_SB(sb);
724 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
726 border = 2 << sb->s_blocksize_bits;
729 /* find how many blocks can be covered since this position */
730 max = ffs(first | border) - 1;
732 /* find how many blocks of power 2 we need to mark */
739 /* mark multiblock chunks only */
740 grp->bb_counters[min]++;
742 mb_clear_bit(first >> min,
743 buddy + sbi->s_mb_offsets[min]);
751 * Cache the order of the largest free extent we have available in this block
755 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
760 grp->bb_largest_free_order = -1; /* uninit */
762 bits = sb->s_blocksize_bits + 1;
763 for (i = bits; i >= 0; i--) {
764 if (grp->bb_counters[i] > 0) {
765 grp->bb_largest_free_order = i;
771 static noinline_for_stack
772 void ext4_mb_generate_buddy(struct super_block *sb,
773 void *buddy, void *bitmap, ext4_group_t group)
775 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
776 struct ext4_sb_info *sbi = EXT4_SB(sb);
777 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
782 unsigned fragments = 0;
783 unsigned long long period = get_cycles();
785 /* initialize buddy from bitmap which is aggregation
786 * of on-disk bitmap and preallocations */
787 i = mb_find_next_zero_bit(bitmap, max, 0);
788 grp->bb_first_free = i;
792 i = mb_find_next_bit(bitmap, max, i);
796 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
798 grp->bb_counters[0]++;
800 i = mb_find_next_zero_bit(bitmap, max, i);
802 grp->bb_fragments = fragments;
804 if (free != grp->bb_free) {
805 ext4_grp_locked_error(sb, group, 0, 0,
806 "block bitmap and bg descriptor "
807 "inconsistent: %u vs %u free clusters",
810 * If we intend to continue, we consider group descriptor
811 * corrupt and update bb_free using bitmap value
814 ext4_mark_group_bitmap_corrupted(sb, group,
815 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
817 mb_set_largest_free_order(sb, grp);
819 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
821 period = get_cycles() - period;
822 spin_lock(&sbi->s_bal_lock);
823 sbi->s_mb_buddies_generated++;
824 sbi->s_mb_generation_time += period;
825 spin_unlock(&sbi->s_bal_lock);
828 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
834 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
835 ext4_set_bits(buddy, 0, count);
837 e4b->bd_info->bb_fragments = 0;
838 memset(e4b->bd_info->bb_counters, 0,
839 sizeof(*e4b->bd_info->bb_counters) *
840 (e4b->bd_sb->s_blocksize_bits + 2));
842 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
843 e4b->bd_bitmap, e4b->bd_group);
846 /* The buddy information is attached the buddy cache inode
847 * for convenience. The information regarding each group
848 * is loaded via ext4_mb_load_buddy. The information involve
849 * block bitmap and buddy information. The information are
850 * stored in the inode as
853 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
856 * one block each for bitmap and buddy information.
857 * So for each group we take up 2 blocks. A page can
858 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
859 * So it can have information regarding groups_per_page which
860 * is blocks_per_page/2
862 * Locking note: This routine takes the block group lock of all groups
863 * for this page; do not hold this lock when calling this routine!
866 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
868 ext4_group_t ngroups;
874 ext4_group_t first_group, group;
876 struct super_block *sb;
877 struct buffer_head *bhs;
878 struct buffer_head **bh = NULL;
882 struct ext4_group_info *grinfo;
884 inode = page->mapping->host;
886 ngroups = ext4_get_groups_count(sb);
887 blocksize = i_blocksize(inode);
888 blocks_per_page = PAGE_SIZE / blocksize;
890 mb_debug(sb, "init page %lu\n", page->index);
892 groups_per_page = blocks_per_page >> 1;
893 if (groups_per_page == 0)
896 /* allocate buffer_heads to read bitmaps */
897 if (groups_per_page > 1) {
898 i = sizeof(struct buffer_head *) * groups_per_page;
899 bh = kzalloc(i, gfp);
907 first_group = page->index * blocks_per_page / 2;
909 /* read all groups the page covers into the cache */
910 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
911 if (group >= ngroups)
914 grinfo = ext4_get_group_info(sb, group);
916 * If page is uptodate then we came here after online resize
917 * which added some new uninitialized group info structs, so
918 * we must skip all initialized uptodate buddies on the page,
919 * which may be currently in use by an allocating task.
921 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
925 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
927 err = PTR_ERR(bh[i]);
931 mb_debug(sb, "read bitmap for group %u\n", group);
934 /* wait for I/O completion */
935 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
940 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
945 first_block = page->index * blocks_per_page;
946 for (i = 0; i < blocks_per_page; i++) {
947 group = (first_block + i) >> 1;
948 if (group >= ngroups)
951 if (!bh[group - first_group])
952 /* skip initialized uptodate buddy */
955 if (!buffer_verified(bh[group - first_group]))
956 /* Skip faulty bitmaps */
961 * data carry information regarding this
962 * particular group in the format specified
966 data = page_address(page) + (i * blocksize);
967 bitmap = bh[group - first_group]->b_data;
970 * We place the buddy block and bitmap block
973 if ((first_block + i) & 1) {
974 /* this is block of buddy */
975 BUG_ON(incore == NULL);
976 mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
977 group, page->index, i * blocksize);
978 trace_ext4_mb_buddy_bitmap_load(sb, group);
979 grinfo = ext4_get_group_info(sb, group);
980 grinfo->bb_fragments = 0;
981 memset(grinfo->bb_counters, 0,
982 sizeof(*grinfo->bb_counters) *
983 (sb->s_blocksize_bits+2));
985 * incore got set to the group block bitmap below
987 ext4_lock_group(sb, group);
989 memset(data, 0xff, blocksize);
990 ext4_mb_generate_buddy(sb, data, incore, group);
991 ext4_unlock_group(sb, group);
994 /* this is block of bitmap */
995 BUG_ON(incore != NULL);
996 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
997 group, page->index, i * blocksize);
998 trace_ext4_mb_bitmap_load(sb, group);
1000 /* see comments in ext4_mb_put_pa() */
1001 ext4_lock_group(sb, group);
1002 memcpy(data, bitmap, blocksize);
1004 /* mark all preallocated blks used in in-core bitmap */
1005 ext4_mb_generate_from_pa(sb, data, group);
1006 ext4_mb_generate_from_freelist(sb, data, group);
1007 ext4_unlock_group(sb, group);
1009 /* set incore so that the buddy information can be
1010 * generated using this
1015 SetPageUptodate(page);
1019 for (i = 0; i < groups_per_page; i++)
1028 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1029 * on the same buddy page doesn't happen whild holding the buddy page lock.
1030 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1031 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1033 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1034 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1036 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1037 int block, pnum, poff;
1038 int blocks_per_page;
1041 e4b->bd_buddy_page = NULL;
1042 e4b->bd_bitmap_page = NULL;
1044 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1046 * the buddy cache inode stores the block bitmap
1047 * and buddy information in consecutive blocks.
1048 * So for each group we need two blocks.
1051 pnum = block / blocks_per_page;
1052 poff = block % blocks_per_page;
1053 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1056 BUG_ON(page->mapping != inode->i_mapping);
1057 e4b->bd_bitmap_page = page;
1058 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1060 if (blocks_per_page >= 2) {
1061 /* buddy and bitmap are on the same page */
1066 pnum = block / blocks_per_page;
1067 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1070 BUG_ON(page->mapping != inode->i_mapping);
1071 e4b->bd_buddy_page = page;
1075 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1077 if (e4b->bd_bitmap_page) {
1078 unlock_page(e4b->bd_bitmap_page);
1079 put_page(e4b->bd_bitmap_page);
1081 if (e4b->bd_buddy_page) {
1082 unlock_page(e4b->bd_buddy_page);
1083 put_page(e4b->bd_buddy_page);
1088 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1089 * block group lock of all groups for this page; do not hold the BG lock when
1090 * calling this routine!
1092 static noinline_for_stack
1093 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1096 struct ext4_group_info *this_grp;
1097 struct ext4_buddy e4b;
1102 mb_debug(sb, "init group %u\n", group);
1103 this_grp = ext4_get_group_info(sb, group);
1105 * This ensures that we don't reinit the buddy cache
1106 * page which map to the group from which we are already
1107 * allocating. If we are looking at the buddy cache we would
1108 * have taken a reference using ext4_mb_load_buddy and that
1109 * would have pinned buddy page to page cache.
1110 * The call to ext4_mb_get_buddy_page_lock will mark the
1113 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1114 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1116 * somebody initialized the group
1117 * return without doing anything
1122 page = e4b.bd_bitmap_page;
1123 ret = ext4_mb_init_cache(page, NULL, gfp);
1126 if (!PageUptodate(page)) {
1131 if (e4b.bd_buddy_page == NULL) {
1133 * If both the bitmap and buddy are in
1134 * the same page we don't need to force
1140 /* init buddy cache */
1141 page = e4b.bd_buddy_page;
1142 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1145 if (!PageUptodate(page)) {
1150 ext4_mb_put_buddy_page_lock(&e4b);
1155 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1156 * block group lock of all groups for this page; do not hold the BG lock when
1157 * calling this routine!
1159 static noinline_for_stack int
1160 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1161 struct ext4_buddy *e4b, gfp_t gfp)
1163 int blocks_per_page;
1169 struct ext4_group_info *grp;
1170 struct ext4_sb_info *sbi = EXT4_SB(sb);
1171 struct inode *inode = sbi->s_buddy_cache;
1174 mb_debug(sb, "load group %u\n", group);
1176 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1177 grp = ext4_get_group_info(sb, group);
1179 e4b->bd_blkbits = sb->s_blocksize_bits;
1182 e4b->bd_group = group;
1183 e4b->bd_buddy_page = NULL;
1184 e4b->bd_bitmap_page = NULL;
1186 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1188 * we need full data about the group
1189 * to make a good selection
1191 ret = ext4_mb_init_group(sb, group, gfp);
1197 * the buddy cache inode stores the block bitmap
1198 * and buddy information in consecutive blocks.
1199 * So for each group we need two blocks.
1202 pnum = block / blocks_per_page;
1203 poff = block % blocks_per_page;
1205 /* we could use find_or_create_page(), but it locks page
1206 * what we'd like to avoid in fast path ... */
1207 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1208 if (page == NULL || !PageUptodate(page)) {
1211 * drop the page reference and try
1212 * to get the page with lock. If we
1213 * are not uptodate that implies
1214 * somebody just created the page but
1215 * is yet to initialize the same. So
1216 * wait for it to initialize.
1219 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1221 BUG_ON(page->mapping != inode->i_mapping);
1222 if (!PageUptodate(page)) {
1223 ret = ext4_mb_init_cache(page, NULL, gfp);
1228 mb_cmp_bitmaps(e4b, page_address(page) +
1229 (poff * sb->s_blocksize));
1238 if (!PageUptodate(page)) {
1243 /* Pages marked accessed already */
1244 e4b->bd_bitmap_page = page;
1245 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1248 pnum = block / blocks_per_page;
1249 poff = block % blocks_per_page;
1251 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1252 if (page == NULL || !PageUptodate(page)) {
1255 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1257 BUG_ON(page->mapping != inode->i_mapping);
1258 if (!PageUptodate(page)) {
1259 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1273 if (!PageUptodate(page)) {
1278 /* Pages marked accessed already */
1279 e4b->bd_buddy_page = page;
1280 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1287 if (e4b->bd_bitmap_page)
1288 put_page(e4b->bd_bitmap_page);
1289 if (e4b->bd_buddy_page)
1290 put_page(e4b->bd_buddy_page);
1291 e4b->bd_buddy = NULL;
1292 e4b->bd_bitmap = NULL;
1296 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1297 struct ext4_buddy *e4b)
1299 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1302 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1304 if (e4b->bd_bitmap_page)
1305 put_page(e4b->bd_bitmap_page);
1306 if (e4b->bd_buddy_page)
1307 put_page(e4b->bd_buddy_page);
1311 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1314 int bb_incr = 1 << (e4b->bd_blkbits - 1);
1317 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1318 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1321 while (order <= e4b->bd_blkbits + 1) {
1323 if (!mb_test_bit(block, bb)) {
1324 /* this block is part of buddy of order 'order' */
1334 static void mb_clear_bits(void *bm, int cur, int len)
1340 if ((cur & 31) == 0 && (len - cur) >= 32) {
1341 /* fast path: clear whole word at once */
1342 addr = bm + (cur >> 3);
1347 mb_clear_bit(cur, bm);
1352 /* clear bits in given range
1353 * will return first found zero bit if any, -1 otherwise
1355 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1362 if ((cur & 31) == 0 && (len - cur) >= 32) {
1363 /* fast path: clear whole word at once */
1364 addr = bm + (cur >> 3);
1365 if (*addr != (__u32)(-1) && zero_bit == -1)
1366 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1371 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1379 void ext4_set_bits(void *bm, int cur, int len)
1385 if ((cur & 31) == 0 && (len - cur) >= 32) {
1386 /* fast path: set whole word at once */
1387 addr = bm + (cur >> 3);
1392 mb_set_bit(cur, bm);
1398 * _________________________________________________________________ */
1400 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1402 if (mb_test_bit(*bit + side, bitmap)) {
1403 mb_clear_bit(*bit, bitmap);
1409 mb_set_bit(*bit, bitmap);
1414 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1418 void *buddy = mb_find_buddy(e4b, order, &max);
1423 /* Bits in range [first; last] are known to be set since
1424 * corresponding blocks were allocated. Bits in range
1425 * (first; last) will stay set because they form buddies on
1426 * upper layer. We just deal with borders if they don't
1427 * align with upper layer and then go up.
1428 * Releasing entire group is all about clearing
1429 * single bit of highest order buddy.
1433 * ---------------------------------
1435 * ---------------------------------
1436 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1437 * ---------------------------------
1439 * \_____________________/
1441 * Neither [1] nor [6] is aligned to above layer.
1442 * Left neighbour [0] is free, so mark it busy,
1443 * decrease bb_counters and extend range to
1445 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1446 * mark [6] free, increase bb_counters and shrink range to
1448 * Then shift range to [0; 2], go up and do the same.
1453 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1455 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1460 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1461 mb_clear_bits(buddy, first, last - first + 1);
1462 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1471 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1472 int first, int count)
1474 int left_is_free = 0;
1475 int right_is_free = 0;
1477 int last = first + count - 1;
1478 struct super_block *sb = e4b->bd_sb;
1480 if (WARN_ON(count == 0))
1482 BUG_ON(last >= (sb->s_blocksize << 3));
1483 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1484 /* Don't bother if the block group is corrupt. */
1485 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1488 mb_check_buddy(e4b);
1489 mb_free_blocks_double(inode, e4b, first, count);
1491 this_cpu_inc(discard_pa_seq);
1492 e4b->bd_info->bb_free += count;
1493 if (first < e4b->bd_info->bb_first_free)
1494 e4b->bd_info->bb_first_free = first;
1496 /* access memory sequentially: check left neighbour,
1497 * clear range and then check right neighbour
1500 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1501 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1502 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1503 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1505 if (unlikely(block != -1)) {
1506 struct ext4_sb_info *sbi = EXT4_SB(sb);
1507 ext4_fsblk_t blocknr;
1509 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1510 blocknr += EXT4_C2B(sbi, block);
1511 ext4_grp_locked_error(sb, e4b->bd_group,
1512 inode ? inode->i_ino : 0,
1514 "freeing already freed block "
1515 "(bit %u); block bitmap corrupt.",
1517 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1518 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1519 mb_regenerate_buddy(e4b);
1523 /* let's maintain fragments counter */
1524 if (left_is_free && right_is_free)
1525 e4b->bd_info->bb_fragments--;
1526 else if (!left_is_free && !right_is_free)
1527 e4b->bd_info->bb_fragments++;
1529 /* buddy[0] == bd_bitmap is a special case, so handle
1530 * it right away and let mb_buddy_mark_free stay free of
1531 * zero order checks.
1532 * Check if neighbours are to be coaleasced,
1533 * adjust bitmap bb_counters and borders appropriately.
1536 first += !left_is_free;
1537 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1540 last -= !right_is_free;
1541 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1545 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1548 mb_set_largest_free_order(sb, e4b->bd_info);
1549 mb_check_buddy(e4b);
1552 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1553 int needed, struct ext4_free_extent *ex)
1559 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1562 buddy = mb_find_buddy(e4b, 0, &max);
1563 BUG_ON(buddy == NULL);
1564 BUG_ON(block >= max);
1565 if (mb_test_bit(block, buddy)) {
1572 /* find actual order */
1573 order = mb_find_order_for_block(e4b, block);
1574 block = block >> order;
1576 ex->fe_len = 1 << order;
1577 ex->fe_start = block << order;
1578 ex->fe_group = e4b->bd_group;
1580 /* calc difference from given start */
1581 next = next - ex->fe_start;
1583 ex->fe_start += next;
1585 while (needed > ex->fe_len &&
1586 mb_find_buddy(e4b, order, &max)) {
1588 if (block + 1 >= max)
1591 next = (block + 1) * (1 << order);
1592 if (mb_test_bit(next, e4b->bd_bitmap))
1595 order = mb_find_order_for_block(e4b, next);
1597 block = next >> order;
1598 ex->fe_len += 1 << order;
1601 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1602 /* Should never happen! (but apparently sometimes does?!?) */
1604 ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent "
1605 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1606 block, order, needed, ex->fe_group, ex->fe_start,
1607 ex->fe_len, ex->fe_logical);
1615 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1621 int start = ex->fe_start;
1622 int len = ex->fe_len;
1627 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1628 BUG_ON(e4b->bd_group != ex->fe_group);
1629 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1630 mb_check_buddy(e4b);
1631 mb_mark_used_double(e4b, start, len);
1633 this_cpu_inc(discard_pa_seq);
1634 e4b->bd_info->bb_free -= len;
1635 if (e4b->bd_info->bb_first_free == start)
1636 e4b->bd_info->bb_first_free += len;
1638 /* let's maintain fragments counter */
1640 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1641 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1642 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1644 e4b->bd_info->bb_fragments++;
1645 else if (!mlen && !max)
1646 e4b->bd_info->bb_fragments--;
1648 /* let's maintain buddy itself */
1650 ord = mb_find_order_for_block(e4b, start);
1652 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1653 /* the whole chunk may be allocated at once! */
1655 buddy = mb_find_buddy(e4b, ord, &max);
1656 BUG_ON((start >> ord) >= max);
1657 mb_set_bit(start >> ord, buddy);
1658 e4b->bd_info->bb_counters[ord]--;
1665 /* store for history */
1667 ret = len | (ord << 16);
1669 /* we have to split large buddy */
1671 buddy = mb_find_buddy(e4b, ord, &max);
1672 mb_set_bit(start >> ord, buddy);
1673 e4b->bd_info->bb_counters[ord]--;
1676 cur = (start >> ord) & ~1U;
1677 buddy = mb_find_buddy(e4b, ord, &max);
1678 mb_clear_bit(cur, buddy);
1679 mb_clear_bit(cur + 1, buddy);
1680 e4b->bd_info->bb_counters[ord]++;
1681 e4b->bd_info->bb_counters[ord]++;
1683 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1685 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1686 mb_check_buddy(e4b);
1692 * Must be called under group lock!
1694 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1695 struct ext4_buddy *e4b)
1697 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1700 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1701 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1703 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1704 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1705 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1707 /* preallocation can change ac_b_ex, thus we store actually
1708 * allocated blocks for history */
1709 ac->ac_f_ex = ac->ac_b_ex;
1711 ac->ac_status = AC_STATUS_FOUND;
1712 ac->ac_tail = ret & 0xffff;
1713 ac->ac_buddy = ret >> 16;
1716 * take the page reference. We want the page to be pinned
1717 * so that we don't get a ext4_mb_init_cache_call for this
1718 * group until we update the bitmap. That would mean we
1719 * double allocate blocks. The reference is dropped
1720 * in ext4_mb_release_context
1722 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1723 get_page(ac->ac_bitmap_page);
1724 ac->ac_buddy_page = e4b->bd_buddy_page;
1725 get_page(ac->ac_buddy_page);
1726 /* store last allocated for subsequent stream allocation */
1727 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1728 spin_lock(&sbi->s_md_lock);
1729 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1730 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1731 spin_unlock(&sbi->s_md_lock);
1734 * As we've just preallocated more space than
1735 * user requested originally, we store allocated
1736 * space in a special descriptor.
1738 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
1739 ext4_mb_new_preallocation(ac);
1744 * regular allocator, for general purposes allocation
1747 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1748 struct ext4_buddy *e4b,
1751 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1752 struct ext4_free_extent *bex = &ac->ac_b_ex;
1753 struct ext4_free_extent *gex = &ac->ac_g_ex;
1754 struct ext4_free_extent ex;
1757 if (ac->ac_status == AC_STATUS_FOUND)
1760 * We don't want to scan for a whole year
1762 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1763 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1764 ac->ac_status = AC_STATUS_BREAK;
1769 * Haven't found good chunk so far, let's continue
1771 if (bex->fe_len < gex->fe_len)
1774 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1775 && bex->fe_group == e4b->bd_group) {
1776 /* recheck chunk's availability - we don't know
1777 * when it was found (within this lock-unlock
1779 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1780 if (max >= gex->fe_len) {
1781 ext4_mb_use_best_found(ac, e4b);
1788 * The routine checks whether found extent is good enough. If it is,
1789 * then the extent gets marked used and flag is set to the context
1790 * to stop scanning. Otherwise, the extent is compared with the
1791 * previous found extent and if new one is better, then it's stored
1792 * in the context. Later, the best found extent will be used, if
1793 * mballoc can't find good enough extent.
1795 * FIXME: real allocation policy is to be designed yet!
1797 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1798 struct ext4_free_extent *ex,
1799 struct ext4_buddy *e4b)
1801 struct ext4_free_extent *bex = &ac->ac_b_ex;
1802 struct ext4_free_extent *gex = &ac->ac_g_ex;
1804 BUG_ON(ex->fe_len <= 0);
1805 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1806 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1807 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1812 * The special case - take what you catch first
1814 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1816 ext4_mb_use_best_found(ac, e4b);
1821 * Let's check whether the chuck is good enough
1823 if (ex->fe_len == gex->fe_len) {
1825 ext4_mb_use_best_found(ac, e4b);
1830 * If this is first found extent, just store it in the context
1832 if (bex->fe_len == 0) {
1838 * If new found extent is better, store it in the context
1840 if (bex->fe_len < gex->fe_len) {
1841 /* if the request isn't satisfied, any found extent
1842 * larger than previous best one is better */
1843 if (ex->fe_len > bex->fe_len)
1845 } else if (ex->fe_len > gex->fe_len) {
1846 /* if the request is satisfied, then we try to find
1847 * an extent that still satisfy the request, but is
1848 * smaller than previous one */
1849 if (ex->fe_len < bex->fe_len)
1853 ext4_mb_check_limits(ac, e4b, 0);
1856 static noinline_for_stack
1857 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1858 struct ext4_buddy *e4b)
1860 struct ext4_free_extent ex = ac->ac_b_ex;
1861 ext4_group_t group = ex.fe_group;
1865 BUG_ON(ex.fe_len <= 0);
1866 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1870 ext4_lock_group(ac->ac_sb, group);
1871 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1875 ext4_mb_use_best_found(ac, e4b);
1878 ext4_unlock_group(ac->ac_sb, group);
1879 ext4_mb_unload_buddy(e4b);
1884 static noinline_for_stack
1885 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1886 struct ext4_buddy *e4b)
1888 ext4_group_t group = ac->ac_g_ex.fe_group;
1891 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1892 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1893 struct ext4_free_extent ex;
1895 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1897 if (grp->bb_free == 0)
1900 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1904 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1905 ext4_mb_unload_buddy(e4b);
1909 ext4_lock_group(ac->ac_sb, group);
1910 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1911 ac->ac_g_ex.fe_len, &ex);
1912 ex.fe_logical = 0xDEADFA11; /* debug value */
1914 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1917 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1919 /* use do_div to get remainder (would be 64-bit modulo) */
1920 if (do_div(start, sbi->s_stripe) == 0) {
1923 ext4_mb_use_best_found(ac, e4b);
1925 } else if (max >= ac->ac_g_ex.fe_len) {
1926 BUG_ON(ex.fe_len <= 0);
1927 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1928 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1931 ext4_mb_use_best_found(ac, e4b);
1932 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1933 /* Sometimes, caller may want to merge even small
1934 * number of blocks to an existing extent */
1935 BUG_ON(ex.fe_len <= 0);
1936 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1937 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1940 ext4_mb_use_best_found(ac, e4b);
1942 ext4_unlock_group(ac->ac_sb, group);
1943 ext4_mb_unload_buddy(e4b);
1949 * The routine scans buddy structures (not bitmap!) from given order
1950 * to max order and tries to find big enough chunk to satisfy the req
1952 static noinline_for_stack
1953 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1954 struct ext4_buddy *e4b)
1956 struct super_block *sb = ac->ac_sb;
1957 struct ext4_group_info *grp = e4b->bd_info;
1963 BUG_ON(ac->ac_2order <= 0);
1964 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1965 if (grp->bb_counters[i] == 0)
1968 buddy = mb_find_buddy(e4b, i, &max);
1969 BUG_ON(buddy == NULL);
1971 k = mb_find_next_zero_bit(buddy, max, 0);
1973 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
1974 "%d free clusters of order %d. But found 0",
1975 grp->bb_counters[i], i);
1976 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
1978 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1983 ac->ac_b_ex.fe_len = 1 << i;
1984 ac->ac_b_ex.fe_start = k << i;
1985 ac->ac_b_ex.fe_group = e4b->bd_group;
1987 ext4_mb_use_best_found(ac, e4b);
1989 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
1991 if (EXT4_SB(sb)->s_mb_stats)
1992 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1999 * The routine scans the group and measures all found extents.
2000 * In order to optimize scanning, caller must pass number of
2001 * free blocks in the group, so the routine can know upper limit.
2003 static noinline_for_stack
2004 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2005 struct ext4_buddy *e4b)
2007 struct super_block *sb = ac->ac_sb;
2008 void *bitmap = e4b->bd_bitmap;
2009 struct ext4_free_extent ex;
2013 free = e4b->bd_info->bb_free;
2014 if (WARN_ON(free <= 0))
2017 i = e4b->bd_info->bb_first_free;
2019 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2020 i = mb_find_next_zero_bit(bitmap,
2021 EXT4_CLUSTERS_PER_GROUP(sb), i);
2022 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2024 * IF we have corrupt bitmap, we won't find any
2025 * free blocks even though group info says we
2026 * we have free blocks
2028 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2029 "%d free clusters as per "
2030 "group info. But bitmap says 0",
2032 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2033 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2037 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2038 if (WARN_ON(ex.fe_len <= 0))
2040 if (free < ex.fe_len) {
2041 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2042 "%d free clusters as per "
2043 "group info. But got %d blocks",
2045 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2046 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2048 * The number of free blocks differs. This mostly
2049 * indicate that the bitmap is corrupt. So exit
2050 * without claiming the space.
2054 ex.fe_logical = 0xDEADC0DE; /* debug value */
2055 ext4_mb_measure_extent(ac, &ex, e4b);
2061 ext4_mb_check_limits(ac, e4b, 1);
2065 * This is a special case for storages like raid5
2066 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2068 static noinline_for_stack
2069 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2070 struct ext4_buddy *e4b)
2072 struct super_block *sb = ac->ac_sb;
2073 struct ext4_sb_info *sbi = EXT4_SB(sb);
2074 void *bitmap = e4b->bd_bitmap;
2075 struct ext4_free_extent ex;
2076 ext4_fsblk_t first_group_block;
2081 BUG_ON(sbi->s_stripe == 0);
2083 /* find first stripe-aligned block in group */
2084 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2086 a = first_group_block + sbi->s_stripe - 1;
2087 do_div(a, sbi->s_stripe);
2088 i = (a * sbi->s_stripe) - first_group_block;
2090 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2091 if (!mb_test_bit(i, bitmap)) {
2092 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2093 if (max >= sbi->s_stripe) {
2095 ex.fe_logical = 0xDEADF00D; /* debug value */
2097 ext4_mb_use_best_found(ac, e4b);
2106 * This is also called BEFORE we load the buddy bitmap.
2107 * Returns either 1 or 0 indicating that the group is either suitable
2108 * for the allocation or not.
2110 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2111 ext4_group_t group, int cr)
2113 ext4_grpblk_t free, fragments;
2114 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2115 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2117 BUG_ON(cr < 0 || cr >= 4);
2119 free = grp->bb_free;
2122 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2125 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2128 fragments = grp->bb_fragments;
2134 BUG_ON(ac->ac_2order == 0);
2136 /* Avoid using the first bg of a flexgroup for data files */
2137 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2138 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2139 ((group % flex_size) == 0))
2142 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2143 (free / fragments) >= ac->ac_g_ex.fe_len)
2146 if (grp->bb_largest_free_order < ac->ac_2order)
2151 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2155 if (free >= ac->ac_g_ex.fe_len)
2168 * This could return negative error code if something goes wrong
2169 * during ext4_mb_init_group(). This should not be called with
2170 * ext4_lock_group() held.
2172 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2173 ext4_group_t group, int cr)
2175 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2176 struct super_block *sb = ac->ac_sb;
2177 struct ext4_sb_info *sbi = EXT4_SB(sb);
2178 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2183 ext4_lock_group(sb, group);
2184 free = grp->bb_free;
2187 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2189 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2192 ext4_unlock_group(sb, group);
2194 /* We only do this if the grp has never been initialized */
2195 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2196 struct ext4_group_desc *gdp =
2197 ext4_get_group_desc(sb, group, NULL);
2200 /* cr=0/1 is a very optimistic search to find large
2201 * good chunks almost for free. If buddy data is not
2202 * ready, then this optimization makes no sense. But
2203 * we never skip the first block group in a flex_bg,
2204 * since this gets used for metadata block allocation,
2205 * and we want to make sure we locate metadata blocks
2206 * in the first block group in the flex_bg if possible.
2209 (!sbi->s_log_groups_per_flex ||
2210 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2211 !(ext4_has_group_desc_csum(sb) &&
2212 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2214 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2220 ext4_lock_group(sb, group);
2221 ret = ext4_mb_good_group(ac, group, cr);
2224 ext4_unlock_group(sb, group);
2229 * Start prefetching @nr block bitmaps starting at @group.
2230 * Return the next group which needs to be prefetched.
2232 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2233 unsigned int nr, int *cnt)
2235 ext4_group_t ngroups = ext4_get_groups_count(sb);
2236 struct buffer_head *bh;
2237 struct blk_plug plug;
2239 blk_start_plug(&plug);
2241 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2243 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2246 * Prefetch block groups with free blocks; but don't
2247 * bother if it is marked uninitialized on disk, since
2248 * it won't require I/O to read. Also only try to
2249 * prefetch once, so we avoid getblk() call, which can
2252 if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2253 EXT4_MB_GRP_NEED_INIT(grp) &&
2254 ext4_free_group_clusters(sb, gdp) > 0 &&
2255 !(ext4_has_group_desc_csum(sb) &&
2256 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2257 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2258 if (bh && !IS_ERR(bh)) {
2259 if (!buffer_uptodate(bh) && cnt)
2264 if (++group >= ngroups)
2267 blk_finish_plug(&plug);
2272 * Prefetching reads the block bitmap into the buffer cache; but we
2273 * need to make sure that the buddy bitmap in the page cache has been
2274 * initialized. Note that ext4_mb_init_group() will block if the I/O
2275 * is not yet completed, or indeed if it was not initiated by
2276 * ext4_mb_prefetch did not start the I/O.
2278 * TODO: We should actually kick off the buddy bitmap setup in a work
2279 * queue when the buffer I/O is completed, so that we don't block
2280 * waiting for the block allocation bitmap read to finish when
2281 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2283 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2287 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2289 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2292 group = ext4_get_groups_count(sb);
2294 grp = ext4_get_group_info(sb, group);
2296 if (EXT4_MB_GRP_NEED_INIT(grp) &&
2297 ext4_free_group_clusters(sb, gdp) > 0 &&
2298 !(ext4_has_group_desc_csum(sb) &&
2299 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2300 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2306 static noinline_for_stack int
2307 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2309 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2311 int err = 0, first_err = 0;
2312 unsigned int nr = 0, prefetch_ios = 0;
2313 struct ext4_sb_info *sbi;
2314 struct super_block *sb;
2315 struct ext4_buddy e4b;
2319 ngroups = ext4_get_groups_count(sb);
2320 /* non-extent files are limited to low blocks/groups */
2321 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2322 ngroups = sbi->s_blockfile_groups;
2324 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2326 /* first, try the goal */
2327 err = ext4_mb_find_by_goal(ac, &e4b);
2328 if (err || ac->ac_status == AC_STATUS_FOUND)
2331 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2335 * ac->ac2_order is set only if the fe_len is a power of 2
2336 * if ac2_order is set we also set criteria to 0 so that we
2337 * try exact allocation using buddy.
2339 i = fls(ac->ac_g_ex.fe_len);
2342 * We search using buddy data only if the order of the request
2343 * is greater than equal to the sbi_s_mb_order2_reqs
2344 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2345 * We also support searching for power-of-two requests only for
2346 * requests upto maximum buddy size we have constructed.
2348 if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) {
2350 * This should tell if fe_len is exactly power of 2
2352 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2353 ac->ac_2order = array_index_nospec(i - 1,
2354 sb->s_blocksize_bits + 2);
2357 /* if stream allocation is enabled, use global goal */
2358 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2359 /* TBD: may be hot point */
2360 spin_lock(&sbi->s_md_lock);
2361 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2362 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2363 spin_unlock(&sbi->s_md_lock);
2366 /* Let's just scan groups to find more-less suitable blocks */
2367 cr = ac->ac_2order ? 0 : 1;
2369 * cr == 0 try to get exact allocation,
2370 * cr == 3 try to get anything
2373 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2374 ac->ac_criteria = cr;
2376 * searching for the right group start
2377 * from the goal value specified
2379 group = ac->ac_g_ex.fe_group;
2380 prefetch_grp = group;
2382 for (i = 0; i < ngroups; group++, i++) {
2386 * Artificially restricted ngroups for non-extent
2387 * files makes group > ngroups possible on first loop.
2389 if (group >= ngroups)
2393 * Batch reads of the block allocation bitmaps
2394 * to get multiple READs in flight; limit
2395 * prefetching at cr=0/1, otherwise mballoc can
2396 * spend a lot of time loading imperfect groups
2398 if ((prefetch_grp == group) &&
2400 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2401 unsigned int curr_ios = prefetch_ios;
2403 nr = sbi->s_mb_prefetch;
2404 if (ext4_has_feature_flex_bg(sb)) {
2405 nr = (group / sbi->s_mb_prefetch) *
2407 nr = nr + sbi->s_mb_prefetch - group;
2409 prefetch_grp = ext4_mb_prefetch(sb, group,
2411 if (prefetch_ios == curr_ios)
2415 /* This now checks without needing the buddy page */
2416 ret = ext4_mb_good_group_nolock(ac, group, cr);
2423 err = ext4_mb_load_buddy(sb, group, &e4b);
2427 ext4_lock_group(sb, group);
2430 * We need to check again after locking the
2433 ret = ext4_mb_good_group(ac, group, cr);
2435 ext4_unlock_group(sb, group);
2436 ext4_mb_unload_buddy(&e4b);
2440 ac->ac_groups_scanned++;
2442 ext4_mb_simple_scan_group(ac, &e4b);
2443 else if (cr == 1 && sbi->s_stripe &&
2444 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2445 ext4_mb_scan_aligned(ac, &e4b);
2447 ext4_mb_complex_scan_group(ac, &e4b);
2449 ext4_unlock_group(sb, group);
2450 ext4_mb_unload_buddy(&e4b);
2452 if (ac->ac_status != AC_STATUS_CONTINUE)
2457 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2458 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2460 * We've been searching too long. Let's try to allocate
2461 * the best chunk we've found so far
2464 ext4_mb_try_best_found(ac, &e4b);
2465 if (ac->ac_status != AC_STATUS_FOUND) {
2467 * Someone more lucky has already allocated it.
2468 * The only thing we can do is just take first
2470 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2472 ac->ac_b_ex.fe_group = 0;
2473 ac->ac_b_ex.fe_start = 0;
2474 ac->ac_b_ex.fe_len = 0;
2475 ac->ac_status = AC_STATUS_CONTINUE;
2476 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2478 atomic_inc(&sbi->s_mb_lost_chunks);
2483 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2486 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2487 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2488 ac->ac_flags, cr, err);
2491 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2496 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2498 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2501 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2504 return (void *) ((unsigned long) group);
2507 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2509 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2513 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2516 return (void *) ((unsigned long) group);
2519 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2521 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2522 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2524 int err, buddy_loaded = 0;
2525 struct ext4_buddy e4b;
2526 struct ext4_group_info *grinfo;
2527 unsigned char blocksize_bits = min_t(unsigned char,
2528 sb->s_blocksize_bits,
2529 EXT4_MAX_BLOCK_LOG_SIZE);
2531 struct ext4_group_info info;
2532 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2537 seq_puts(seq, "#group: free frags first ["
2538 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2539 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2541 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2542 sizeof(struct ext4_group_info);
2544 grinfo = ext4_get_group_info(sb, group);
2545 /* Load the group info in memory only if not already loaded. */
2546 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2547 err = ext4_mb_load_buddy(sb, group, &e4b);
2549 seq_printf(seq, "#%-5u: I/O error\n", group);
2555 memcpy(&sg, ext4_get_group_info(sb, group), i);
2558 ext4_mb_unload_buddy(&e4b);
2560 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2561 sg.info.bb_fragments, sg.info.bb_first_free);
2562 for (i = 0; i <= 13; i++)
2563 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2564 sg.info.bb_counters[i] : 0);
2565 seq_printf(seq, " ]\n");
2570 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2574 const struct seq_operations ext4_mb_seq_groups_ops = {
2575 .start = ext4_mb_seq_groups_start,
2576 .next = ext4_mb_seq_groups_next,
2577 .stop = ext4_mb_seq_groups_stop,
2578 .show = ext4_mb_seq_groups_show,
2581 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2583 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2584 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2591 * Allocate the top-level s_group_info array for the specified number
2594 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2596 struct ext4_sb_info *sbi = EXT4_SB(sb);
2598 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
2600 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2601 EXT4_DESC_PER_BLOCK_BITS(sb);
2602 if (size <= sbi->s_group_info_size)
2605 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2606 new_groupinfo = kvzalloc(size, GFP_KERNEL);
2607 if (!new_groupinfo) {
2608 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2612 old_groupinfo = rcu_dereference(sbi->s_group_info);
2614 memcpy(new_groupinfo, old_groupinfo,
2615 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2617 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
2618 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2620 ext4_kvfree_array_rcu(old_groupinfo);
2621 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2622 sbi->s_group_info_size);
2626 /* Create and initialize ext4_group_info data for the given group. */
2627 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2628 struct ext4_group_desc *desc)
2632 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
2633 struct ext4_sb_info *sbi = EXT4_SB(sb);
2634 struct ext4_group_info **meta_group_info;
2635 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2638 * First check if this group is the first of a reserved block.
2639 * If it's true, we have to allocate a new table of pointers
2640 * to ext4_group_info structures
2642 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2643 metalen = sizeof(*meta_group_info) <<
2644 EXT4_DESC_PER_BLOCK_BITS(sb);
2645 meta_group_info = kmalloc(metalen, GFP_NOFS);
2646 if (meta_group_info == NULL) {
2647 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2648 "for a buddy group");
2649 goto exit_meta_group_info;
2652 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
2656 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
2657 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2659 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2660 if (meta_group_info[i] == NULL) {
2661 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2662 goto exit_group_info;
2664 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2665 &(meta_group_info[i]->bb_state));
2668 * initialize bb_free to be able to skip
2669 * empty groups without initialization
2671 if (ext4_has_group_desc_csum(sb) &&
2672 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
2673 meta_group_info[i]->bb_free =
2674 ext4_free_clusters_after_init(sb, group, desc);
2676 meta_group_info[i]->bb_free =
2677 ext4_free_group_clusters(sb, desc);
2680 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2681 init_rwsem(&meta_group_info[i]->alloc_sem);
2682 meta_group_info[i]->bb_free_root = RB_ROOT;
2683 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2685 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
2689 /* If a meta_group_info table has been allocated, release it now */
2690 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2691 struct ext4_group_info ***group_info;
2694 group_info = rcu_dereference(sbi->s_group_info);
2695 kfree(group_info[idx]);
2696 group_info[idx] = NULL;
2699 exit_meta_group_info:
2701 } /* ext4_mb_add_groupinfo */
2703 static int ext4_mb_init_backend(struct super_block *sb)
2705 ext4_group_t ngroups = ext4_get_groups_count(sb);
2707 struct ext4_sb_info *sbi = EXT4_SB(sb);
2709 struct ext4_group_desc *desc;
2710 struct ext4_group_info ***group_info;
2711 struct kmem_cache *cachep;
2713 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2717 sbi->s_buddy_cache = new_inode(sb);
2718 if (sbi->s_buddy_cache == NULL) {
2719 ext4_msg(sb, KERN_ERR, "can't get new inode");
2722 /* To avoid potentially colliding with an valid on-disk inode number,
2723 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2724 * not in the inode hash, so it should never be found by iget(), but
2725 * this will avoid confusion if it ever shows up during debugging. */
2726 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2727 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2728 for (i = 0; i < ngroups; i++) {
2730 desc = ext4_get_group_desc(sb, i, NULL);
2732 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2735 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2739 if (ext4_has_feature_flex_bg(sb)) {
2740 /* a single flex group is supposed to be read by a single IO */
2741 sbi->s_mb_prefetch = 1 << sbi->s_es->s_log_groups_per_flex;
2742 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
2744 sbi->s_mb_prefetch = 32;
2746 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
2747 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
2748 /* now many real IOs to prefetch within a single allocation at cr=0
2749 * given cr=0 is an CPU-related optimization we shouldn't try to
2750 * load too many groups, at some point we should start to use what
2751 * we've got in memory.
2752 * with an average random access time 5ms, it'd take a second to get
2753 * 200 groups (* N with flex_bg), so let's make this limit 4
2755 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
2756 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
2757 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
2762 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2764 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2765 i = sbi->s_group_info_size;
2767 group_info = rcu_dereference(sbi->s_group_info);
2769 kfree(group_info[i]);
2771 iput(sbi->s_buddy_cache);
2774 kvfree(rcu_dereference(sbi->s_group_info));
2779 static void ext4_groupinfo_destroy_slabs(void)
2783 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2784 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2785 ext4_groupinfo_caches[i] = NULL;
2789 static int ext4_groupinfo_create_slab(size_t size)
2791 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2793 int blocksize_bits = order_base_2(size);
2794 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2795 struct kmem_cache *cachep;
2797 if (cache_index >= NR_GRPINFO_CACHES)
2800 if (unlikely(cache_index < 0))
2803 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2804 if (ext4_groupinfo_caches[cache_index]) {
2805 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2806 return 0; /* Already created */
2809 slab_size = offsetof(struct ext4_group_info,
2810 bb_counters[blocksize_bits + 2]);
2812 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2813 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2816 ext4_groupinfo_caches[cache_index] = cachep;
2818 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2821 "EXT4-fs: no memory for groupinfo slab cache\n");
2828 int ext4_mb_init(struct super_block *sb)
2830 struct ext4_sb_info *sbi = EXT4_SB(sb);
2832 unsigned offset, offset_incr;
2836 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2838 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2839 if (sbi->s_mb_offsets == NULL) {
2844 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2845 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2846 if (sbi->s_mb_maxs == NULL) {
2851 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2855 /* order 0 is regular bitmap */
2856 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2857 sbi->s_mb_offsets[0] = 0;
2861 offset_incr = 1 << (sb->s_blocksize_bits - 1);
2862 max = sb->s_blocksize << 2;
2864 sbi->s_mb_offsets[i] = offset;
2865 sbi->s_mb_maxs[i] = max;
2866 offset += offset_incr;
2867 offset_incr = offset_incr >> 1;
2870 } while (i <= sb->s_blocksize_bits + 1);
2872 spin_lock_init(&sbi->s_md_lock);
2873 spin_lock_init(&sbi->s_bal_lock);
2874 sbi->s_mb_free_pending = 0;
2875 INIT_LIST_HEAD(&sbi->s_freed_data_list);
2877 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2878 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2879 sbi->s_mb_stats = MB_DEFAULT_STATS;
2880 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2881 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2883 * The default group preallocation is 512, which for 4k block
2884 * sizes translates to 2 megabytes. However for bigalloc file
2885 * systems, this is probably too big (i.e, if the cluster size
2886 * is 1 megabyte, then group preallocation size becomes half a
2887 * gigabyte!). As a default, we will keep a two megabyte
2888 * group pralloc size for cluster sizes up to 64k, and after
2889 * that, we will force a minimum group preallocation size of
2890 * 32 clusters. This translates to 8 megs when the cluster
2891 * size is 256k, and 32 megs when the cluster size is 1 meg,
2892 * which seems reasonable as a default.
2894 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2895 sbi->s_cluster_bits, 32);
2897 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2898 * to the lowest multiple of s_stripe which is bigger than
2899 * the s_mb_group_prealloc as determined above. We want
2900 * the preallocation size to be an exact multiple of the
2901 * RAID stripe size so that preallocations don't fragment
2904 if (sbi->s_stripe > 1) {
2905 sbi->s_mb_group_prealloc = roundup(
2906 sbi->s_mb_group_prealloc, sbi->s_stripe);
2909 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2910 if (sbi->s_locality_groups == NULL) {
2914 for_each_possible_cpu(i) {
2915 struct ext4_locality_group *lg;
2916 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2917 mutex_init(&lg->lg_mutex);
2918 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2919 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2920 spin_lock_init(&lg->lg_prealloc_lock);
2923 /* init file for buddy data */
2924 ret = ext4_mb_init_backend(sb);
2926 goto out_free_locality_groups;
2930 out_free_locality_groups:
2931 free_percpu(sbi->s_locality_groups);
2932 sbi->s_locality_groups = NULL;
2934 kfree(sbi->s_mb_offsets);
2935 sbi->s_mb_offsets = NULL;
2936 kfree(sbi->s_mb_maxs);
2937 sbi->s_mb_maxs = NULL;
2941 /* need to called with the ext4 group lock held */
2942 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2944 struct ext4_prealloc_space *pa;
2945 struct list_head *cur, *tmp;
2948 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2949 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2950 list_del(&pa->pa_group_list);
2952 kmem_cache_free(ext4_pspace_cachep, pa);
2957 int ext4_mb_release(struct super_block *sb)
2959 ext4_group_t ngroups = ext4_get_groups_count(sb);
2961 int num_meta_group_infos;
2962 struct ext4_group_info *grinfo, ***group_info;
2963 struct ext4_sb_info *sbi = EXT4_SB(sb);
2964 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2967 if (sbi->s_group_info) {
2968 for (i = 0; i < ngroups; i++) {
2970 grinfo = ext4_get_group_info(sb, i);
2971 mb_group_bb_bitmap_free(grinfo);
2972 ext4_lock_group(sb, i);
2973 count = ext4_mb_cleanup_pa(grinfo);
2975 mb_debug(sb, "mballoc: %d PAs left\n",
2977 ext4_unlock_group(sb, i);
2978 kmem_cache_free(cachep, grinfo);
2980 num_meta_group_infos = (ngroups +
2981 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2982 EXT4_DESC_PER_BLOCK_BITS(sb);
2984 group_info = rcu_dereference(sbi->s_group_info);
2985 for (i = 0; i < num_meta_group_infos; i++)
2986 kfree(group_info[i]);
2990 kfree(sbi->s_mb_offsets);
2991 kfree(sbi->s_mb_maxs);
2992 iput(sbi->s_buddy_cache);
2993 if (sbi->s_mb_stats) {
2994 ext4_msg(sb, KERN_INFO,
2995 "mballoc: %u blocks %u reqs (%u success)",
2996 atomic_read(&sbi->s_bal_allocated),
2997 atomic_read(&sbi->s_bal_reqs),
2998 atomic_read(&sbi->s_bal_success));
2999 ext4_msg(sb, KERN_INFO,
3000 "mballoc: %u extents scanned, %u goal hits, "
3001 "%u 2^N hits, %u breaks, %u lost",
3002 atomic_read(&sbi->s_bal_ex_scanned),
3003 atomic_read(&sbi->s_bal_goals),
3004 atomic_read(&sbi->s_bal_2orders),
3005 atomic_read(&sbi->s_bal_breaks),
3006 atomic_read(&sbi->s_mb_lost_chunks));
3007 ext4_msg(sb, KERN_INFO,
3008 "mballoc: %lu generated and it took %Lu",
3009 sbi->s_mb_buddies_generated,
3010 sbi->s_mb_generation_time);
3011 ext4_msg(sb, KERN_INFO,
3012 "mballoc: %u preallocated, %u discarded",
3013 atomic_read(&sbi->s_mb_preallocated),
3014 atomic_read(&sbi->s_mb_discarded));
3017 free_percpu(sbi->s_locality_groups);
3022 static inline int ext4_issue_discard(struct super_block *sb,
3023 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3026 ext4_fsblk_t discard_block;
3028 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3029 ext4_group_first_block_no(sb, block_group));
3030 count = EXT4_C2B(EXT4_SB(sb), count);
3031 trace_ext4_discard_blocks(sb,
3032 (unsigned long long) discard_block, count);
3034 return __blkdev_issue_discard(sb->s_bdev,
3035 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3036 (sector_t)count << (sb->s_blocksize_bits - 9),
3039 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3042 static void ext4_free_data_in_buddy(struct super_block *sb,
3043 struct ext4_free_data *entry)
3045 struct ext4_buddy e4b;
3046 struct ext4_group_info *db;
3047 int err, count = 0, count2 = 0;
3049 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3050 entry->efd_count, entry->efd_group, entry);
3052 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3053 /* we expect to find existing buddy because it's pinned */
3056 spin_lock(&EXT4_SB(sb)->s_md_lock);
3057 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3058 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3061 /* there are blocks to put in buddy to make them really free */
3062 count += entry->efd_count;
3064 ext4_lock_group(sb, entry->efd_group);
3065 /* Take it out of per group rb tree */
3066 rb_erase(&entry->efd_node, &(db->bb_free_root));
3067 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3070 * Clear the trimmed flag for the group so that the next
3071 * ext4_trim_fs can trim it.
3072 * If the volume is mounted with -o discard, online discard
3073 * is supported and the free blocks will be trimmed online.
3075 if (!test_opt(sb, DISCARD))
3076 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3078 if (!db->bb_free_root.rb_node) {
3079 /* No more items in the per group rb tree
3080 * balance refcounts from ext4_mb_free_metadata()
3082 put_page(e4b.bd_buddy_page);
3083 put_page(e4b.bd_bitmap_page);
3085 ext4_unlock_group(sb, entry->efd_group);
3086 kmem_cache_free(ext4_free_data_cachep, entry);
3087 ext4_mb_unload_buddy(&e4b);
3089 mb_debug(sb, "freed %d blocks in %d structures\n", count,
3094 * This function is called by the jbd2 layer once the commit has finished,
3095 * so we know we can free the blocks that were released with that commit.
3097 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3099 struct ext4_sb_info *sbi = EXT4_SB(sb);
3100 struct ext4_free_data *entry, *tmp;
3101 struct bio *discard_bio = NULL;
3102 struct list_head freed_data_list;
3103 struct list_head *cut_pos = NULL;
3106 INIT_LIST_HEAD(&freed_data_list);
3108 spin_lock(&sbi->s_md_lock);
3109 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3110 if (entry->efd_tid != commit_tid)
3112 cut_pos = &entry->efd_list;
3115 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3117 spin_unlock(&sbi->s_md_lock);
3119 if (test_opt(sb, DISCARD)) {
3120 list_for_each_entry(entry, &freed_data_list, efd_list) {
3121 err = ext4_issue_discard(sb, entry->efd_group,
3122 entry->efd_start_cluster,
3125 if (err && err != -EOPNOTSUPP) {
3126 ext4_msg(sb, KERN_WARNING, "discard request in"
3127 " group:%d block:%d count:%d failed"
3128 " with %d", entry->efd_group,
3129 entry->efd_start_cluster,
3130 entry->efd_count, err);
3131 } else if (err == -EOPNOTSUPP)
3136 submit_bio_wait(discard_bio);
3137 bio_put(discard_bio);
3141 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3142 ext4_free_data_in_buddy(sb, entry);
3145 int __init ext4_init_mballoc(void)
3147 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3148 SLAB_RECLAIM_ACCOUNT);
3149 if (ext4_pspace_cachep == NULL)
3152 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3153 SLAB_RECLAIM_ACCOUNT);
3154 if (ext4_ac_cachep == NULL)
3157 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3158 SLAB_RECLAIM_ACCOUNT);
3159 if (ext4_free_data_cachep == NULL)
3165 kmem_cache_destroy(ext4_ac_cachep);
3167 kmem_cache_destroy(ext4_pspace_cachep);
3172 void ext4_exit_mballoc(void)
3175 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3176 * before destroying the slab cache.
3179 kmem_cache_destroy(ext4_pspace_cachep);
3180 kmem_cache_destroy(ext4_ac_cachep);
3181 kmem_cache_destroy(ext4_free_data_cachep);
3182 ext4_groupinfo_destroy_slabs();
3187 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3188 * Returns 0 if success or error code
3190 static noinline_for_stack int
3191 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3192 handle_t *handle, unsigned int reserv_clstrs)
3194 struct buffer_head *bitmap_bh = NULL;
3195 struct ext4_group_desc *gdp;
3196 struct buffer_head *gdp_bh;
3197 struct ext4_sb_info *sbi;
3198 struct super_block *sb;
3202 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3203 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3208 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3209 if (IS_ERR(bitmap_bh)) {
3210 err = PTR_ERR(bitmap_bh);
3215 BUFFER_TRACE(bitmap_bh, "getting write access");
3216 err = ext4_journal_get_write_access(handle, bitmap_bh);
3221 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3225 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3226 ext4_free_group_clusters(sb, gdp));
3228 BUFFER_TRACE(gdp_bh, "get_write_access");
3229 err = ext4_journal_get_write_access(handle, gdp_bh);
3233 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3235 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3236 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3237 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3238 "fs metadata", block, block+len);
3239 /* File system mounted not to panic on error
3240 * Fix the bitmap and return EFSCORRUPTED
3241 * We leak some of the blocks here.
3243 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3244 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3245 ac->ac_b_ex.fe_len);
3246 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3247 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3249 err = -EFSCORRUPTED;
3253 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3254 #ifdef AGGRESSIVE_CHECK
3257 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3258 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3259 bitmap_bh->b_data));
3263 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3264 ac->ac_b_ex.fe_len);
3265 if (ext4_has_group_desc_csum(sb) &&
3266 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3267 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3268 ext4_free_group_clusters_set(sb, gdp,
3269 ext4_free_clusters_after_init(sb,
3270 ac->ac_b_ex.fe_group, gdp));
3272 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3273 ext4_free_group_clusters_set(sb, gdp, len);
3274 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3275 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3277 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3278 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3280 * Now reduce the dirty block count also. Should not go negative
3282 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3283 /* release all the reserved blocks if non delalloc */
3284 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3287 if (sbi->s_log_groups_per_flex) {
3288 ext4_group_t flex_group = ext4_flex_group(sbi,
3289 ac->ac_b_ex.fe_group);
3290 atomic64_sub(ac->ac_b_ex.fe_len,
3291 &sbi_array_rcu_deref(sbi, s_flex_groups,
3292 flex_group)->free_clusters);
3295 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3298 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3306 * here we normalize request for locality group
3307 * Group request are normalized to s_mb_group_prealloc, which goes to
3308 * s_strip if we set the same via mount option.
3309 * s_mb_group_prealloc can be configured via
3310 * /sys/fs/ext4/<partition>/mb_group_prealloc
3312 * XXX: should we try to preallocate more than the group has now?
3314 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3316 struct super_block *sb = ac->ac_sb;
3317 struct ext4_locality_group *lg = ac->ac_lg;
3320 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3321 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
3325 * Normalization means making request better in terms of
3326 * size and alignment
3328 static noinline_for_stack void
3329 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3330 struct ext4_allocation_request *ar)
3332 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3335 loff_t size, start_off;
3336 loff_t orig_size __maybe_unused;
3338 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3339 struct ext4_prealloc_space *pa;
3341 /* do normalize only data requests, metadata requests
3342 do not need preallocation */
3343 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3346 /* sometime caller may want exact blocks */
3347 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3350 /* caller may indicate that preallocation isn't
3351 * required (it's a tail, for example) */
3352 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3355 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3356 ext4_mb_normalize_group_request(ac);
3360 bsbits = ac->ac_sb->s_blocksize_bits;
3362 /* first, let's learn actual file size
3363 * given current request is allocated */
3364 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3365 size = size << bsbits;
3366 if (size < i_size_read(ac->ac_inode))
3367 size = i_size_read(ac->ac_inode);
3370 /* max size of free chunks */
3373 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3374 (req <= (size) || max <= (chunk_size))
3376 /* first, try to predict filesize */
3377 /* XXX: should this table be tunable? */
3379 if (size <= 16 * 1024) {
3381 } else if (size <= 32 * 1024) {
3383 } else if (size <= 64 * 1024) {
3385 } else if (size <= 128 * 1024) {
3387 } else if (size <= 256 * 1024) {
3389 } else if (size <= 512 * 1024) {
3391 } else if (size <= 1024 * 1024) {
3393 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3394 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3395 (21 - bsbits)) << 21;
3396 size = 2 * 1024 * 1024;
3397 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3398 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3399 (22 - bsbits)) << 22;
3400 size = 4 * 1024 * 1024;
3401 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3402 (8<<20)>>bsbits, max, 8 * 1024)) {
3403 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3404 (23 - bsbits)) << 23;
3405 size = 8 * 1024 * 1024;
3407 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3408 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3409 ac->ac_o_ex.fe_len) << bsbits;
3411 size = size >> bsbits;
3412 start = start_off >> bsbits;
3414 /* don't cover already allocated blocks in selected range */
3415 if (ar->pleft && start <= ar->lleft) {
3416 size -= ar->lleft + 1 - start;
3417 start = ar->lleft + 1;
3419 if (ar->pright && start + size - 1 >= ar->lright)
3420 size -= start + size - ar->lright;
3423 * Trim allocation request for filesystems with artificially small
3426 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
3427 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
3431 /* check we don't cross already preallocated blocks */
3433 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3438 spin_lock(&pa->pa_lock);
3439 if (pa->pa_deleted) {
3440 spin_unlock(&pa->pa_lock);
3444 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3447 /* PA must not overlap original request */
3448 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3449 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3451 /* skip PAs this normalized request doesn't overlap with */
3452 if (pa->pa_lstart >= end || pa_end <= start) {
3453 spin_unlock(&pa->pa_lock);
3456 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3458 /* adjust start or end to be adjacent to this pa */
3459 if (pa_end <= ac->ac_o_ex.fe_logical) {
3460 BUG_ON(pa_end < start);
3462 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3463 BUG_ON(pa->pa_lstart > end);
3464 end = pa->pa_lstart;
3466 spin_unlock(&pa->pa_lock);
3471 /* XXX: extra loop to check we really don't overlap preallocations */
3473 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3476 spin_lock(&pa->pa_lock);
3477 if (pa->pa_deleted == 0) {
3478 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3480 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3482 spin_unlock(&pa->pa_lock);
3486 if (start + size <= ac->ac_o_ex.fe_logical &&
3487 start > ac->ac_o_ex.fe_logical) {
3488 ext4_msg(ac->ac_sb, KERN_ERR,
3489 "start %lu, size %lu, fe_logical %lu",
3490 (unsigned long) start, (unsigned long) size,
3491 (unsigned long) ac->ac_o_ex.fe_logical);
3494 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3496 /* now prepare goal request */
3498 /* XXX: is it better to align blocks WRT to logical
3499 * placement or satisfy big request as is */
3500 ac->ac_g_ex.fe_logical = start;
3501 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3503 /* define goal start in order to merge */
3504 if (ar->pright && (ar->lright == (start + size))) {
3505 /* merge to the right */
3506 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3507 &ac->ac_f_ex.fe_group,
3508 &ac->ac_f_ex.fe_start);
3509 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3511 if (ar->pleft && (ar->lleft + 1 == start)) {
3512 /* merge to the left */
3513 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3514 &ac->ac_f_ex.fe_group,
3515 &ac->ac_f_ex.fe_start);
3516 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3519 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
3523 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3525 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3527 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3528 atomic_inc(&sbi->s_bal_reqs);
3529 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3530 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3531 atomic_inc(&sbi->s_bal_success);
3532 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3533 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3534 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3535 atomic_inc(&sbi->s_bal_goals);
3536 if (ac->ac_found > sbi->s_mb_max_to_scan)
3537 atomic_inc(&sbi->s_bal_breaks);
3540 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3541 trace_ext4_mballoc_alloc(ac);
3543 trace_ext4_mballoc_prealloc(ac);
3547 * Called on failure; free up any blocks from the inode PA for this
3548 * context. We don't need this for MB_GROUP_PA because we only change
3549 * pa_free in ext4_mb_release_context(), but on failure, we've already
3550 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3552 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3554 struct ext4_prealloc_space *pa = ac->ac_pa;
3555 struct ext4_buddy e4b;
3559 if (ac->ac_f_ex.fe_len == 0)
3561 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3564 * This should never happen since we pin the
3565 * pages in the ext4_allocation_context so
3566 * ext4_mb_load_buddy() should never fail.
3568 WARN(1, "mb_load_buddy failed (%d)", err);
3571 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3572 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3573 ac->ac_f_ex.fe_len);
3574 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3575 ext4_mb_unload_buddy(&e4b);
3578 if (pa->pa_type == MB_INODE_PA)
3579 pa->pa_free += ac->ac_b_ex.fe_len;
3583 * use blocks preallocated to inode
3585 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3586 struct ext4_prealloc_space *pa)
3588 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3593 /* found preallocated blocks, use them */
3594 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3595 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3596 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3597 len = EXT4_NUM_B2C(sbi, end - start);
3598 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3599 &ac->ac_b_ex.fe_start);
3600 ac->ac_b_ex.fe_len = len;
3601 ac->ac_status = AC_STATUS_FOUND;
3604 BUG_ON(start < pa->pa_pstart);
3605 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3606 BUG_ON(pa->pa_free < len);
3609 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
3613 * use blocks preallocated to locality group
3615 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3616 struct ext4_prealloc_space *pa)
3618 unsigned int len = ac->ac_o_ex.fe_len;
3620 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3621 &ac->ac_b_ex.fe_group,
3622 &ac->ac_b_ex.fe_start);
3623 ac->ac_b_ex.fe_len = len;
3624 ac->ac_status = AC_STATUS_FOUND;
3627 /* we don't correct pa_pstart or pa_plen here to avoid
3628 * possible race when the group is being loaded concurrently
3629 * instead we correct pa later, after blocks are marked
3630 * in on-disk bitmap -- see ext4_mb_release_context()
3631 * Other CPUs are prevented from allocating from this pa by lg_mutex
3633 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
3634 pa->pa_lstart-len, len, pa);
3638 * Return the prealloc space that have minimal distance
3639 * from the goal block. @cpa is the prealloc
3640 * space that is having currently known minimal distance
3641 * from the goal block.
3643 static struct ext4_prealloc_space *
3644 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3645 struct ext4_prealloc_space *pa,
3646 struct ext4_prealloc_space *cpa)
3648 ext4_fsblk_t cur_distance, new_distance;
3651 atomic_inc(&pa->pa_count);
3654 cur_distance = abs(goal_block - cpa->pa_pstart);
3655 new_distance = abs(goal_block - pa->pa_pstart);
3657 if (cur_distance <= new_distance)
3660 /* drop the previous reference */
3661 atomic_dec(&cpa->pa_count);
3662 atomic_inc(&pa->pa_count);
3667 * search goal blocks in preallocated space
3669 static noinline_for_stack bool
3670 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3672 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3674 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3675 struct ext4_locality_group *lg;
3676 struct ext4_prealloc_space *pa, *cpa = NULL;
3677 ext4_fsblk_t goal_block;
3679 /* only data can be preallocated */
3680 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3683 /* first, try per-file preallocation */
3685 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3687 /* all fields in this condition don't change,
3688 * so we can skip locking for them */
3689 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3690 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3691 EXT4_C2B(sbi, pa->pa_len)))
3694 /* non-extent files can't have physical blocks past 2^32 */
3695 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3696 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3697 EXT4_MAX_BLOCK_FILE_PHYS))
3700 /* found preallocated blocks, use them */
3701 spin_lock(&pa->pa_lock);
3702 if (pa->pa_deleted == 0 && pa->pa_free) {
3703 atomic_inc(&pa->pa_count);
3704 ext4_mb_use_inode_pa(ac, pa);
3705 spin_unlock(&pa->pa_lock);
3706 ac->ac_criteria = 10;
3710 spin_unlock(&pa->pa_lock);
3714 /* can we use group allocation? */
3715 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3718 /* inode may have no locality group for some reason */
3722 order = fls(ac->ac_o_ex.fe_len) - 1;
3723 if (order > PREALLOC_TB_SIZE - 1)
3724 /* The max size of hash table is PREALLOC_TB_SIZE */
3725 order = PREALLOC_TB_SIZE - 1;
3727 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3729 * search for the prealloc space that is having
3730 * minimal distance from the goal block.
3732 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3734 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3736 spin_lock(&pa->pa_lock);
3737 if (pa->pa_deleted == 0 &&
3738 pa->pa_free >= ac->ac_o_ex.fe_len) {
3740 cpa = ext4_mb_check_group_pa(goal_block,
3743 spin_unlock(&pa->pa_lock);
3748 ext4_mb_use_group_pa(ac, cpa);
3749 ac->ac_criteria = 20;
3756 * the function goes through all block freed in the group
3757 * but not yet committed and marks them used in in-core bitmap.
3758 * buddy must be generated from this bitmap
3759 * Need to be called with the ext4 group lock held
3761 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3765 struct ext4_group_info *grp;
3766 struct ext4_free_data *entry;
3768 grp = ext4_get_group_info(sb, group);
3769 n = rb_first(&(grp->bb_free_root));
3772 entry = rb_entry(n, struct ext4_free_data, efd_node);
3773 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3780 * the function goes through all preallocation in this group and marks them
3781 * used in in-core bitmap. buddy must be generated from this bitmap
3782 * Need to be called with ext4 group lock held
3784 static noinline_for_stack
3785 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3788 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3789 struct ext4_prealloc_space *pa;
3790 struct list_head *cur;
3791 ext4_group_t groupnr;
3792 ext4_grpblk_t start;
3793 int preallocated = 0;
3796 /* all form of preallocation discards first load group,
3797 * so the only competing code is preallocation use.
3798 * we don't need any locking here
3799 * notice we do NOT ignore preallocations with pa_deleted
3800 * otherwise we could leave used blocks available for
3801 * allocation in buddy when concurrent ext4_mb_put_pa()
3802 * is dropping preallocation
3804 list_for_each(cur, &grp->bb_prealloc_list) {
3805 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3806 spin_lock(&pa->pa_lock);
3807 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3810 spin_unlock(&pa->pa_lock);
3811 if (unlikely(len == 0))
3813 BUG_ON(groupnr != group);
3814 ext4_set_bits(bitmap, start, len);
3815 preallocated += len;
3817 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
3820 static void ext4_mb_pa_callback(struct rcu_head *head)
3822 struct ext4_prealloc_space *pa;
3823 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3825 BUG_ON(atomic_read(&pa->pa_count));
3826 BUG_ON(pa->pa_deleted == 0);
3827 kmem_cache_free(ext4_pspace_cachep, pa);
3831 * drops a reference to preallocated space descriptor
3832 * if this was the last reference and the space is consumed
3834 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3835 struct super_block *sb, struct ext4_prealloc_space *pa)
3838 ext4_fsblk_t grp_blk;
3840 /* in this short window concurrent discard can set pa_deleted */
3841 spin_lock(&pa->pa_lock);
3842 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3843 spin_unlock(&pa->pa_lock);
3847 if (pa->pa_deleted == 1) {
3848 spin_unlock(&pa->pa_lock);
3853 spin_unlock(&pa->pa_lock);
3855 grp_blk = pa->pa_pstart;
3857 * If doing group-based preallocation, pa_pstart may be in the
3858 * next group when pa is used up
3860 if (pa->pa_type == MB_GROUP_PA)
3863 grp = ext4_get_group_number(sb, grp_blk);
3868 * P1 (buddy init) P2 (regular allocation)
3869 * find block B in PA
3870 * copy on-disk bitmap to buddy
3871 * mark B in on-disk bitmap
3872 * drop PA from group
3873 * mark all PAs in buddy
3875 * thus, P1 initializes buddy with B available. to prevent this
3876 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3879 ext4_lock_group(sb, grp);
3880 list_del(&pa->pa_group_list);
3881 ext4_unlock_group(sb, grp);
3883 spin_lock(pa->pa_obj_lock);
3884 list_del_rcu(&pa->pa_inode_list);
3885 spin_unlock(pa->pa_obj_lock);
3887 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3891 * creates new preallocated space for given inode
3893 static noinline_for_stack void
3894 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3896 struct super_block *sb = ac->ac_sb;
3897 struct ext4_sb_info *sbi = EXT4_SB(sb);
3898 struct ext4_prealloc_space *pa;
3899 struct ext4_group_info *grp;
3900 struct ext4_inode_info *ei;
3902 /* preallocate only when found space is larger then requested */
3903 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3904 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3905 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3906 BUG_ON(ac->ac_pa == NULL);
3910 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3916 /* we can't allocate as much as normalizer wants.
3917 * so, found space must get proper lstart
3918 * to cover original request */
3919 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3920 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3922 /* we're limited by original request in that
3923 * logical block must be covered any way
3924 * winl is window we can move our chunk within */
3925 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3927 /* also, we should cover whole original request */
3928 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3930 /* the smallest one defines real window */
3931 win = min(winl, wins);
3933 offs = ac->ac_o_ex.fe_logical %
3934 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3935 if (offs && offs < win)
3938 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3939 EXT4_NUM_B2C(sbi, win);
3940 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3941 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3944 /* preallocation can change ac_b_ex, thus we store actually
3945 * allocated blocks for history */
3946 ac->ac_f_ex = ac->ac_b_ex;
3948 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3949 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3950 pa->pa_len = ac->ac_b_ex.fe_len;
3951 pa->pa_free = pa->pa_len;
3952 spin_lock_init(&pa->pa_lock);
3953 INIT_LIST_HEAD(&pa->pa_inode_list);
3954 INIT_LIST_HEAD(&pa->pa_group_list);
3956 pa->pa_type = MB_INODE_PA;
3958 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
3959 pa->pa_len, pa->pa_lstart);
3960 trace_ext4_mb_new_inode_pa(ac, pa);
3962 ext4_mb_use_inode_pa(ac, pa);
3963 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3965 ei = EXT4_I(ac->ac_inode);
3966 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3968 pa->pa_obj_lock = &ei->i_prealloc_lock;
3969 pa->pa_inode = ac->ac_inode;
3971 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3973 spin_lock(pa->pa_obj_lock);
3974 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3975 spin_unlock(pa->pa_obj_lock);
3979 * creates new preallocated space for locality group inodes belongs to
3981 static noinline_for_stack void
3982 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3984 struct super_block *sb = ac->ac_sb;
3985 struct ext4_locality_group *lg;
3986 struct ext4_prealloc_space *pa;
3987 struct ext4_group_info *grp;
3989 /* preallocate only when found space is larger then requested */
3990 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3991 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3992 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3993 BUG_ON(ac->ac_pa == NULL);
3997 /* preallocation can change ac_b_ex, thus we store actually
3998 * allocated blocks for history */
3999 ac->ac_f_ex = ac->ac_b_ex;
4001 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4002 pa->pa_lstart = pa->pa_pstart;
4003 pa->pa_len = ac->ac_b_ex.fe_len;
4004 pa->pa_free = pa->pa_len;
4005 spin_lock_init(&pa->pa_lock);
4006 INIT_LIST_HEAD(&pa->pa_inode_list);
4007 INIT_LIST_HEAD(&pa->pa_group_list);
4009 pa->pa_type = MB_GROUP_PA;
4011 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4012 pa->pa_len, pa->pa_lstart);
4013 trace_ext4_mb_new_group_pa(ac, pa);
4015 ext4_mb_use_group_pa(ac, pa);
4016 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4018 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4022 pa->pa_obj_lock = &lg->lg_prealloc_lock;
4023 pa->pa_inode = NULL;
4025 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4028 * We will later add the new pa to the right bucket
4029 * after updating the pa_free in ext4_mb_release_context
4033 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4035 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4036 ext4_mb_new_group_pa(ac);
4038 ext4_mb_new_inode_pa(ac);
4042 * finds all unused blocks in on-disk bitmap, frees them in
4043 * in-core bitmap and buddy.
4044 * @pa must be unlinked from inode and group lists, so that
4045 * nobody else can find/use it.
4046 * the caller MUST hold group/inode locks.
4047 * TODO: optimize the case when there are no in-core structures yet
4049 static noinline_for_stack int
4050 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4051 struct ext4_prealloc_space *pa)
4053 struct super_block *sb = e4b->bd_sb;
4054 struct ext4_sb_info *sbi = EXT4_SB(sb);
4059 unsigned long long grp_blk_start;
4062 BUG_ON(pa->pa_deleted == 0);
4063 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4064 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4065 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4066 end = bit + pa->pa_len;
4069 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4072 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4073 mb_debug(sb, "free preallocated %u/%u in group %u\n",
4074 (unsigned) ext4_group_first_block_no(sb, group) + bit,
4075 (unsigned) next - bit, (unsigned) group);
4078 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4079 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4080 EXT4_C2B(sbi, bit)),
4082 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4085 if (free != pa->pa_free) {
4086 ext4_msg(e4b->bd_sb, KERN_CRIT,
4087 "pa %p: logic %lu, phys. %lu, len %d",
4088 pa, (unsigned long) pa->pa_lstart,
4089 (unsigned long) pa->pa_pstart,
4091 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4094 * pa is already deleted so we use the value obtained
4095 * from the bitmap and continue.
4098 atomic_add(free, &sbi->s_mb_discarded);
4103 static noinline_for_stack int
4104 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4105 struct ext4_prealloc_space *pa)
4107 struct super_block *sb = e4b->bd_sb;
4111 trace_ext4_mb_release_group_pa(sb, pa);
4112 BUG_ON(pa->pa_deleted == 0);
4113 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4114 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4115 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4116 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4117 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4123 * releases all preallocations in given group
4125 * first, we need to decide discard policy:
4126 * - when do we discard
4128 * - how many do we discard
4129 * 1) how many requested
4131 static noinline_for_stack int
4132 ext4_mb_discard_group_preallocations(struct super_block *sb,
4133 ext4_group_t group, int needed)
4135 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4136 struct buffer_head *bitmap_bh = NULL;
4137 struct ext4_prealloc_space *pa, *tmp;
4138 struct list_head list;
4139 struct ext4_buddy e4b;
4144 mb_debug(sb, "discard preallocation for group %u\n", group);
4145 if (list_empty(&grp->bb_prealloc_list))
4148 bitmap_bh = ext4_read_block_bitmap(sb, group);
4149 if (IS_ERR(bitmap_bh)) {
4150 err = PTR_ERR(bitmap_bh);
4151 ext4_error_err(sb, -err,
4152 "Error %d reading block bitmap for %u",
4157 err = ext4_mb_load_buddy(sb, group, &e4b);
4159 ext4_warning(sb, "Error %d loading buddy information for %u",
4166 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
4168 INIT_LIST_HEAD(&list);
4170 ext4_lock_group(sb, group);
4171 this_cpu_inc(discard_pa_seq);
4172 list_for_each_entry_safe(pa, tmp,
4173 &grp->bb_prealloc_list, pa_group_list) {
4174 spin_lock(&pa->pa_lock);
4175 if (atomic_read(&pa->pa_count)) {
4176 spin_unlock(&pa->pa_lock);
4180 if (pa->pa_deleted) {
4181 spin_unlock(&pa->pa_lock);
4185 /* seems this one can be freed ... */
4188 /* we can trust pa_free ... */
4189 free += pa->pa_free;
4191 spin_unlock(&pa->pa_lock);
4193 list_del(&pa->pa_group_list);
4194 list_add(&pa->u.pa_tmp_list, &list);
4197 /* if we still need more blocks and some PAs were used, try again */
4198 if (free < needed && busy) {
4200 ext4_unlock_group(sb, group);
4205 /* found anything to free? */
4206 if (list_empty(&list)) {
4208 mb_debug(sb, "Someone else may have freed PA for this group %u\n",
4213 /* now free all selected PAs */
4214 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4216 /* remove from object (inode or locality group) */
4217 spin_lock(pa->pa_obj_lock);
4218 list_del_rcu(&pa->pa_inode_list);
4219 spin_unlock(pa->pa_obj_lock);
4221 if (pa->pa_type == MB_GROUP_PA)
4222 ext4_mb_release_group_pa(&e4b, pa);
4224 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4226 list_del(&pa->u.pa_tmp_list);
4227 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4231 ext4_unlock_group(sb, group);
4232 ext4_mb_unload_buddy(&e4b);
4235 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4236 free, group, grp->bb_free);
4241 * releases all non-used preallocated blocks for given inode
4243 * It's important to discard preallocations under i_data_sem
4244 * We don't want another block to be served from the prealloc
4245 * space when we are discarding the inode prealloc space.
4247 * FIXME!! Make sure it is valid at all the call sites
4249 void ext4_discard_preallocations(struct inode *inode)
4251 struct ext4_inode_info *ei = EXT4_I(inode);
4252 struct super_block *sb = inode->i_sb;
4253 struct buffer_head *bitmap_bh = NULL;
4254 struct ext4_prealloc_space *pa, *tmp;
4255 ext4_group_t group = 0;
4256 struct list_head list;
4257 struct ext4_buddy e4b;
4260 if (!S_ISREG(inode->i_mode)) {
4261 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4265 mb_debug(sb, "discard preallocation for inode %lu\n",
4267 trace_ext4_discard_preallocations(inode);
4269 INIT_LIST_HEAD(&list);
4272 /* first, collect all pa's in the inode */
4273 spin_lock(&ei->i_prealloc_lock);
4274 while (!list_empty(&ei->i_prealloc_list)) {
4275 pa = list_entry(ei->i_prealloc_list.next,
4276 struct ext4_prealloc_space, pa_inode_list);
4277 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4278 spin_lock(&pa->pa_lock);
4279 if (atomic_read(&pa->pa_count)) {
4280 /* this shouldn't happen often - nobody should
4281 * use preallocation while we're discarding it */
4282 spin_unlock(&pa->pa_lock);
4283 spin_unlock(&ei->i_prealloc_lock);
4284 ext4_msg(sb, KERN_ERR,
4285 "uh-oh! used pa while discarding");
4287 schedule_timeout_uninterruptible(HZ);
4291 if (pa->pa_deleted == 0) {
4293 spin_unlock(&pa->pa_lock);
4294 list_del_rcu(&pa->pa_inode_list);
4295 list_add(&pa->u.pa_tmp_list, &list);
4299 /* someone is deleting pa right now */
4300 spin_unlock(&pa->pa_lock);
4301 spin_unlock(&ei->i_prealloc_lock);
4303 /* we have to wait here because pa_deleted
4304 * doesn't mean pa is already unlinked from
4305 * the list. as we might be called from
4306 * ->clear_inode() the inode will get freed
4307 * and concurrent thread which is unlinking
4308 * pa from inode's list may access already
4309 * freed memory, bad-bad-bad */
4311 /* XXX: if this happens too often, we can
4312 * add a flag to force wait only in case
4313 * of ->clear_inode(), but not in case of
4314 * regular truncate */
4315 schedule_timeout_uninterruptible(HZ);
4318 spin_unlock(&ei->i_prealloc_lock);
4320 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4321 BUG_ON(pa->pa_type != MB_INODE_PA);
4322 group = ext4_get_group_number(sb, pa->pa_pstart);
4324 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4325 GFP_NOFS|__GFP_NOFAIL);
4327 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
4332 bitmap_bh = ext4_read_block_bitmap(sb, group);
4333 if (IS_ERR(bitmap_bh)) {
4334 err = PTR_ERR(bitmap_bh);
4335 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
4337 ext4_mb_unload_buddy(&e4b);
4341 ext4_lock_group(sb, group);
4342 list_del(&pa->pa_group_list);
4343 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4344 ext4_unlock_group(sb, group);
4346 ext4_mb_unload_buddy(&e4b);
4349 list_del(&pa->u.pa_tmp_list);
4350 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4354 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
4356 struct ext4_prealloc_space *pa;
4358 BUG_ON(ext4_pspace_cachep == NULL);
4359 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
4362 atomic_set(&pa->pa_count, 1);
4367 static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
4369 struct ext4_prealloc_space *pa = ac->ac_pa;
4373 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
4374 kmem_cache_free(ext4_pspace_cachep, pa);
4377 #ifdef CONFIG_EXT4_DEBUG
4378 static inline void ext4_mb_show_pa(struct super_block *sb)
4380 ext4_group_t i, ngroups;
4382 if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
4385 ngroups = ext4_get_groups_count(sb);
4386 mb_debug(sb, "groups: ");
4387 for (i = 0; i < ngroups; i++) {
4388 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4389 struct ext4_prealloc_space *pa;
4390 ext4_grpblk_t start;
4391 struct list_head *cur;
4392 ext4_lock_group(sb, i);
4393 list_for_each(cur, &grp->bb_prealloc_list) {
4394 pa = list_entry(cur, struct ext4_prealloc_space,
4396 spin_lock(&pa->pa_lock);
4397 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4399 spin_unlock(&pa->pa_lock);
4400 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
4403 ext4_unlock_group(sb, i);
4404 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
4409 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4411 struct super_block *sb = ac->ac_sb;
4413 if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
4416 mb_debug(sb, "Can't allocate:"
4417 " Allocation context details:");
4418 mb_debug(sb, "status %u flags 0x%x",
4419 ac->ac_status, ac->ac_flags);
4420 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
4421 "goal %lu/%lu/%lu@%lu, "
4422 "best %lu/%lu/%lu@%lu cr %d",
4423 (unsigned long)ac->ac_o_ex.fe_group,
4424 (unsigned long)ac->ac_o_ex.fe_start,
4425 (unsigned long)ac->ac_o_ex.fe_len,
4426 (unsigned long)ac->ac_o_ex.fe_logical,
4427 (unsigned long)ac->ac_g_ex.fe_group,
4428 (unsigned long)ac->ac_g_ex.fe_start,
4429 (unsigned long)ac->ac_g_ex.fe_len,
4430 (unsigned long)ac->ac_g_ex.fe_logical,
4431 (unsigned long)ac->ac_b_ex.fe_group,
4432 (unsigned long)ac->ac_b_ex.fe_start,
4433 (unsigned long)ac->ac_b_ex.fe_len,
4434 (unsigned long)ac->ac_b_ex.fe_logical,
4435 (int)ac->ac_criteria);
4436 mb_debug(sb, "%u found", ac->ac_found);
4437 ext4_mb_show_pa(sb);
4440 static inline void ext4_mb_show_pa(struct super_block *sb)
4444 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4446 ext4_mb_show_pa(ac->ac_sb);
4452 * We use locality group preallocation for small size file. The size of the
4453 * file is determined by the current size or the resulting size after
4454 * allocation which ever is larger
4456 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4458 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4460 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4461 int bsbits = ac->ac_sb->s_blocksize_bits;
4464 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4467 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4470 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4471 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4474 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
4475 !inode_is_open_for_write(ac->ac_inode)) {
4476 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4480 if (sbi->s_mb_group_prealloc <= 0) {
4481 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4485 /* don't use group allocation for large files */
4486 size = max(size, isize);
4487 if (size > sbi->s_mb_stream_request) {
4488 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4492 BUG_ON(ac->ac_lg != NULL);
4494 * locality group prealloc space are per cpu. The reason for having
4495 * per cpu locality group is to reduce the contention between block
4496 * request from multiple CPUs.
4498 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4500 /* we're going to use group allocation */
4501 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4503 /* serialize all allocations in the group */
4504 mutex_lock(&ac->ac_lg->lg_mutex);
4507 static noinline_for_stack int
4508 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4509 struct ext4_allocation_request *ar)
4511 struct super_block *sb = ar->inode->i_sb;
4512 struct ext4_sb_info *sbi = EXT4_SB(sb);
4513 struct ext4_super_block *es = sbi->s_es;
4517 ext4_grpblk_t block;
4519 /* we can't allocate > group size */
4522 /* just a dirty hack to filter too big requests */
4523 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4524 len = EXT4_CLUSTERS_PER_GROUP(sb);
4526 /* start searching from the goal */
4528 if (goal < le32_to_cpu(es->s_first_data_block) ||
4529 goal >= ext4_blocks_count(es))
4530 goal = le32_to_cpu(es->s_first_data_block);
4531 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4533 /* set up allocation goals */
4534 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4535 ac->ac_status = AC_STATUS_CONTINUE;
4537 ac->ac_inode = ar->inode;
4538 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4539 ac->ac_o_ex.fe_group = group;
4540 ac->ac_o_ex.fe_start = block;
4541 ac->ac_o_ex.fe_len = len;
4542 ac->ac_g_ex = ac->ac_o_ex;
4543 ac->ac_flags = ar->flags;
4545 /* we have to define context: we'll work with a file or
4546 * locality group. this is a policy, actually */
4547 ext4_mb_group_or_file(ac);
4549 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
4550 "left: %u/%u, right %u/%u to %swritable\n",
4551 (unsigned) ar->len, (unsigned) ar->logical,
4552 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4553 (unsigned) ar->lleft, (unsigned) ar->pleft,
4554 (unsigned) ar->lright, (unsigned) ar->pright,
4555 inode_is_open_for_write(ar->inode) ? "" : "non-");
4560 static noinline_for_stack void
4561 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4562 struct ext4_locality_group *lg,
4563 int order, int total_entries)
4565 ext4_group_t group = 0;
4566 struct ext4_buddy e4b;
4567 struct list_head discard_list;
4568 struct ext4_prealloc_space *pa, *tmp;
4570 mb_debug(sb, "discard locality group preallocation\n");
4572 INIT_LIST_HEAD(&discard_list);
4574 spin_lock(&lg->lg_prealloc_lock);
4575 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4577 lockdep_is_held(&lg->lg_prealloc_lock)) {
4578 spin_lock(&pa->pa_lock);
4579 if (atomic_read(&pa->pa_count)) {
4581 * This is the pa that we just used
4582 * for block allocation. So don't
4585 spin_unlock(&pa->pa_lock);
4588 if (pa->pa_deleted) {
4589 spin_unlock(&pa->pa_lock);
4592 /* only lg prealloc space */
4593 BUG_ON(pa->pa_type != MB_GROUP_PA);
4595 /* seems this one can be freed ... */
4597 spin_unlock(&pa->pa_lock);
4599 list_del_rcu(&pa->pa_inode_list);
4600 list_add(&pa->u.pa_tmp_list, &discard_list);
4603 if (total_entries <= 5) {
4605 * we want to keep only 5 entries
4606 * allowing it to grow to 8. This
4607 * mak sure we don't call discard
4608 * soon for this list.
4613 spin_unlock(&lg->lg_prealloc_lock);
4615 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4618 group = ext4_get_group_number(sb, pa->pa_pstart);
4619 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4620 GFP_NOFS|__GFP_NOFAIL);
4622 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
4626 ext4_lock_group(sb, group);
4627 list_del(&pa->pa_group_list);
4628 ext4_mb_release_group_pa(&e4b, pa);
4629 ext4_unlock_group(sb, group);
4631 ext4_mb_unload_buddy(&e4b);
4632 list_del(&pa->u.pa_tmp_list);
4633 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4638 * We have incremented pa_count. So it cannot be freed at this
4639 * point. Also we hold lg_mutex. So no parallel allocation is
4640 * possible from this lg. That means pa_free cannot be updated.
4642 * A parallel ext4_mb_discard_group_preallocations is possible.
4643 * which can cause the lg_prealloc_list to be updated.
4646 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4648 int order, added = 0, lg_prealloc_count = 1;
4649 struct super_block *sb = ac->ac_sb;
4650 struct ext4_locality_group *lg = ac->ac_lg;
4651 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4653 order = fls(pa->pa_free) - 1;
4654 if (order > PREALLOC_TB_SIZE - 1)
4655 /* The max size of hash table is PREALLOC_TB_SIZE */
4656 order = PREALLOC_TB_SIZE - 1;
4657 /* Add the prealloc space to lg */
4658 spin_lock(&lg->lg_prealloc_lock);
4659 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4661 lockdep_is_held(&lg->lg_prealloc_lock)) {
4662 spin_lock(&tmp_pa->pa_lock);
4663 if (tmp_pa->pa_deleted) {
4664 spin_unlock(&tmp_pa->pa_lock);
4667 if (!added && pa->pa_free < tmp_pa->pa_free) {
4668 /* Add to the tail of the previous entry */
4669 list_add_tail_rcu(&pa->pa_inode_list,
4670 &tmp_pa->pa_inode_list);
4673 * we want to count the total
4674 * number of entries in the list
4677 spin_unlock(&tmp_pa->pa_lock);
4678 lg_prealloc_count++;
4681 list_add_tail_rcu(&pa->pa_inode_list,
4682 &lg->lg_prealloc_list[order]);
4683 spin_unlock(&lg->lg_prealloc_lock);
4685 /* Now trim the list to be not more than 8 elements */
4686 if (lg_prealloc_count > 8) {
4687 ext4_mb_discard_lg_preallocations(sb, lg,
4688 order, lg_prealloc_count);
4695 * release all resource we used in allocation
4697 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4699 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4700 struct ext4_prealloc_space *pa = ac->ac_pa;
4702 if (pa->pa_type == MB_GROUP_PA) {
4703 /* see comment in ext4_mb_use_group_pa() */
4704 spin_lock(&pa->pa_lock);
4705 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4706 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4707 pa->pa_free -= ac->ac_b_ex.fe_len;
4708 pa->pa_len -= ac->ac_b_ex.fe_len;
4709 spin_unlock(&pa->pa_lock);
4714 * We want to add the pa to the right bucket.
4715 * Remove it from the list and while adding
4716 * make sure the list to which we are adding
4719 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4720 spin_lock(pa->pa_obj_lock);
4721 list_del_rcu(&pa->pa_inode_list);
4722 spin_unlock(pa->pa_obj_lock);
4723 ext4_mb_add_n_trim(ac);
4725 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4727 if (ac->ac_bitmap_page)
4728 put_page(ac->ac_bitmap_page);
4729 if (ac->ac_buddy_page)
4730 put_page(ac->ac_buddy_page);
4731 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4732 mutex_unlock(&ac->ac_lg->lg_mutex);
4733 ext4_mb_collect_stats(ac);
4737 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4739 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4743 trace_ext4_mb_discard_preallocations(sb, needed);
4744 for (i = 0; i < ngroups && needed > 0; i++) {
4745 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4753 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
4754 struct ext4_allocation_context *ac, u64 *seq)
4760 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4765 seq_retry = ext4_get_discard_pa_seq_sum();
4766 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
4767 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
4773 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
4778 * Main entry point into mballoc to allocate blocks
4779 * it tries to use preallocation first, then falls back
4780 * to usual allocation
4782 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4783 struct ext4_allocation_request *ar, int *errp)
4785 struct ext4_allocation_context *ac = NULL;
4786 struct ext4_sb_info *sbi;
4787 struct super_block *sb;
4788 ext4_fsblk_t block = 0;
4789 unsigned int inquota = 0;
4790 unsigned int reserv_clstrs = 0;
4794 sb = ar->inode->i_sb;
4797 trace_ext4_request_blocks(ar);
4799 /* Allow to use superuser reservation for quota file */
4800 if (ext4_is_quota_file(ar->inode))
4801 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4803 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4804 /* Without delayed allocation we need to verify
4805 * there is enough free blocks to do block allocation
4806 * and verify allocation doesn't exceed the quota limits.
4809 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4811 /* let others to free the space */
4813 ar->len = ar->len >> 1;
4816 ext4_mb_show_pa(sb);
4820 reserv_clstrs = ar->len;
4821 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4822 dquot_alloc_block_nofail(ar->inode,
4823 EXT4_C2B(sbi, ar->len));
4826 dquot_alloc_block(ar->inode,
4827 EXT4_C2B(sbi, ar->len))) {
4829 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4840 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4847 *errp = ext4_mb_initialize_context(ac, ar);
4853 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4854 seq = this_cpu_read(discard_pa_seq);
4855 if (!ext4_mb_use_preallocated(ac)) {
4856 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4857 ext4_mb_normalize_request(ac, ar);
4859 *errp = ext4_mb_pa_alloc(ac);
4863 /* allocate space in core */
4864 *errp = ext4_mb_regular_allocator(ac);
4866 * pa allocated above is added to grp->bb_prealloc_list only
4867 * when we were able to allocate some block i.e. when
4868 * ac->ac_status == AC_STATUS_FOUND.
4869 * And error from above mean ac->ac_status != AC_STATUS_FOUND
4870 * So we have to free this pa here itself.
4873 ext4_mb_pa_free(ac);
4874 ext4_discard_allocated_blocks(ac);
4877 if (ac->ac_status == AC_STATUS_FOUND &&
4878 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
4879 ext4_mb_pa_free(ac);
4881 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4882 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4884 ext4_discard_allocated_blocks(ac);
4887 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4888 ar->len = ac->ac_b_ex.fe_len;
4891 if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
4894 * If block allocation fails then the pa allocated above
4895 * needs to be freed here itself.
4897 ext4_mb_pa_free(ac);
4903 ac->ac_b_ex.fe_len = 0;
4905 ext4_mb_show_ac(ac);
4907 ext4_mb_release_context(ac);
4910 kmem_cache_free(ext4_ac_cachep, ac);
4911 if (inquota && ar->len < inquota)
4912 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4914 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4915 /* release all the reserved blocks if non delalloc */
4916 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4920 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4926 * We can merge two free data extents only if the physical blocks
4927 * are contiguous, AND the extents were freed by the same transaction,
4928 * AND the blocks are associated with the same group.
4930 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
4931 struct ext4_free_data *entry,
4932 struct ext4_free_data *new_entry,
4933 struct rb_root *entry_rb_root)
4935 if ((entry->efd_tid != new_entry->efd_tid) ||
4936 (entry->efd_group != new_entry->efd_group))
4938 if (entry->efd_start_cluster + entry->efd_count ==
4939 new_entry->efd_start_cluster) {
4940 new_entry->efd_start_cluster = entry->efd_start_cluster;
4941 new_entry->efd_count += entry->efd_count;
4942 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
4943 entry->efd_start_cluster) {
4944 new_entry->efd_count += entry->efd_count;
4947 spin_lock(&sbi->s_md_lock);
4948 list_del(&entry->efd_list);
4949 spin_unlock(&sbi->s_md_lock);
4950 rb_erase(&entry->efd_node, entry_rb_root);
4951 kmem_cache_free(ext4_free_data_cachep, entry);
4954 static noinline_for_stack int
4955 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4956 struct ext4_free_data *new_entry)
4958 ext4_group_t group = e4b->bd_group;
4959 ext4_grpblk_t cluster;
4960 ext4_grpblk_t clusters = new_entry->efd_count;
4961 struct ext4_free_data *entry;
4962 struct ext4_group_info *db = e4b->bd_info;
4963 struct super_block *sb = e4b->bd_sb;
4964 struct ext4_sb_info *sbi = EXT4_SB(sb);
4965 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4966 struct rb_node *parent = NULL, *new_node;
4968 BUG_ON(!ext4_handle_valid(handle));
4969 BUG_ON(e4b->bd_bitmap_page == NULL);
4970 BUG_ON(e4b->bd_buddy_page == NULL);
4972 new_node = &new_entry->efd_node;
4973 cluster = new_entry->efd_start_cluster;
4976 /* first free block exent. We need to
4977 protect buddy cache from being freed,
4978 * otherwise we'll refresh it from
4979 * on-disk bitmap and lose not-yet-available
4981 get_page(e4b->bd_buddy_page);
4982 get_page(e4b->bd_bitmap_page);
4986 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4987 if (cluster < entry->efd_start_cluster)
4989 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4990 n = &(*n)->rb_right;
4992 ext4_grp_locked_error(sb, group, 0,
4993 ext4_group_first_block_no(sb, group) +
4994 EXT4_C2B(sbi, cluster),
4995 "Block already on to-be-freed list");
5000 rb_link_node(new_node, parent, n);
5001 rb_insert_color(new_node, &db->bb_free_root);
5003 /* Now try to see the extent can be merged to left and right */
5004 node = rb_prev(new_node);
5006 entry = rb_entry(node, struct ext4_free_data, efd_node);
5007 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5008 &(db->bb_free_root));
5011 node = rb_next(new_node);
5013 entry = rb_entry(node, struct ext4_free_data, efd_node);
5014 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5015 &(db->bb_free_root));
5018 spin_lock(&sbi->s_md_lock);
5019 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5020 sbi->s_mb_free_pending += clusters;
5021 spin_unlock(&sbi->s_md_lock);
5026 * ext4_free_blocks() -- Free given blocks and update quota
5027 * @handle: handle for this transaction
5029 * @bh: optional buffer of the block to be freed
5030 * @block: starting physical block to be freed
5031 * @count: number of blocks to be freed
5032 * @flags: flags used by ext4_free_blocks
5034 void ext4_free_blocks(handle_t *handle, struct inode *inode,
5035 struct buffer_head *bh, ext4_fsblk_t block,
5036 unsigned long count, int flags)
5038 struct buffer_head *bitmap_bh = NULL;
5039 struct super_block *sb = inode->i_sb;
5040 struct ext4_group_desc *gdp;
5041 unsigned int overflow;
5043 struct buffer_head *gd_bh;
5044 ext4_group_t block_group;
5045 struct ext4_sb_info *sbi;
5046 struct ext4_buddy e4b;
5047 unsigned int count_clusters;
5054 BUG_ON(block != bh->b_blocknr);
5056 block = bh->b_blocknr;
5060 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5061 !ext4_inode_block_valid(inode, block, count)) {
5062 ext4_error(sb, "Freeing blocks not in datazone - "
5063 "block = %llu, count = %lu", block, count);
5067 ext4_debug("freeing block %llu\n", block);
5068 trace_ext4_free_blocks(inode, block, count, flags);
5070 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
5073 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
5078 * If the extent to be freed does not begin on a cluster
5079 * boundary, we need to deal with partial clusters at the
5080 * beginning and end of the extent. Normally we will free
5081 * blocks at the beginning or the end unless we are explicitly
5082 * requested to avoid doing so.
5084 overflow = EXT4_PBLK_COFF(sbi, block);
5086 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
5087 overflow = sbi->s_cluster_ratio - overflow;
5089 if (count > overflow)
5098 overflow = EXT4_LBLK_COFF(sbi, count);
5100 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
5101 if (count > overflow)
5106 count += sbi->s_cluster_ratio - overflow;
5109 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
5111 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
5113 for (i = 0; i < count; i++) {
5116 bh = sb_find_get_block(inode->i_sb, block + i);
5117 ext4_forget(handle, is_metadata, inode, bh, block + i);
5123 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5125 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5126 ext4_get_group_info(sb, block_group))))
5130 * Check to see if we are freeing blocks across a group
5133 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5134 overflow = EXT4_C2B(sbi, bit) + count -
5135 EXT4_BLOCKS_PER_GROUP(sb);
5138 count_clusters = EXT4_NUM_B2C(sbi, count);
5139 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5140 if (IS_ERR(bitmap_bh)) {
5141 err = PTR_ERR(bitmap_bh);
5145 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5151 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
5152 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
5153 in_range(block, ext4_inode_table(sb, gdp),
5154 sbi->s_itb_per_group) ||
5155 in_range(block + count - 1, ext4_inode_table(sb, gdp),
5156 sbi->s_itb_per_group)) {
5158 ext4_error(sb, "Freeing blocks in system zone - "
5159 "Block = %llu, count = %lu", block, count);
5160 /* err = 0. ext4_std_error should be a no op */
5164 BUFFER_TRACE(bitmap_bh, "getting write access");
5165 err = ext4_journal_get_write_access(handle, bitmap_bh);
5170 * We are about to modify some metadata. Call the journal APIs
5171 * to unshare ->b_data if a currently-committing transaction is
5174 BUFFER_TRACE(gd_bh, "get_write_access");
5175 err = ext4_journal_get_write_access(handle, gd_bh);
5178 #ifdef AGGRESSIVE_CHECK
5181 for (i = 0; i < count_clusters; i++)
5182 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
5185 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
5187 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
5188 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
5189 GFP_NOFS|__GFP_NOFAIL);
5194 * We need to make sure we don't reuse the freed block until after the
5195 * transaction is committed. We make an exception if the inode is to be
5196 * written in writeback mode since writeback mode has weak data
5197 * consistency guarantees.
5199 if (ext4_handle_valid(handle) &&
5200 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
5201 !ext4_should_writeback_data(inode))) {
5202 struct ext4_free_data *new_entry;
5204 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
5207 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
5208 GFP_NOFS|__GFP_NOFAIL);
5209 new_entry->efd_start_cluster = bit;
5210 new_entry->efd_group = block_group;
5211 new_entry->efd_count = count_clusters;
5212 new_entry->efd_tid = handle->h_transaction->t_tid;
5214 ext4_lock_group(sb, block_group);
5215 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
5216 ext4_mb_free_metadata(handle, &e4b, new_entry);
5218 /* need to update group_info->bb_free and bitmap
5219 * with group lock held. generate_buddy look at
5220 * them with group lock_held
5222 if (test_opt(sb, DISCARD)) {
5223 err = ext4_issue_discard(sb, block_group, bit, count,
5225 if (err && err != -EOPNOTSUPP)
5226 ext4_msg(sb, KERN_WARNING, "discard request in"
5227 " group:%d block:%d count:%lu failed"
5228 " with %d", block_group, bit, count,
5231 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
5233 ext4_lock_group(sb, block_group);
5234 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
5235 mb_free_blocks(inode, &e4b, bit, count_clusters);
5238 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
5239 ext4_free_group_clusters_set(sb, gdp, ret);
5240 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
5241 ext4_group_desc_csum_set(sb, block_group, gdp);
5242 ext4_unlock_group(sb, block_group);
5244 if (sbi->s_log_groups_per_flex) {
5245 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5246 atomic64_add(count_clusters,
5247 &sbi_array_rcu_deref(sbi, s_flex_groups,
5248 flex_group)->free_clusters);
5252 * on a bigalloc file system, defer the s_freeclusters_counter
5253 * update to the caller (ext4_remove_space and friends) so they
5254 * can determine if a cluster freed here should be rereserved
5256 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
5257 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
5258 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
5259 percpu_counter_add(&sbi->s_freeclusters_counter,
5263 ext4_mb_unload_buddy(&e4b);
5265 /* We dirtied the bitmap block */
5266 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5267 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5269 /* And the group descriptor block */
5270 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5271 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5275 if (overflow && !err) {
5283 ext4_std_error(sb, err);
5288 * ext4_group_add_blocks() -- Add given blocks to an existing group
5289 * @handle: handle to this transaction
5291 * @block: start physical block to add to the block group
5292 * @count: number of blocks to free
5294 * This marks the blocks as free in the bitmap and buddy.
5296 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
5297 ext4_fsblk_t block, unsigned long count)
5299 struct buffer_head *bitmap_bh = NULL;
5300 struct buffer_head *gd_bh;
5301 ext4_group_t block_group;
5304 struct ext4_group_desc *desc;
5305 struct ext4_sb_info *sbi = EXT4_SB(sb);
5306 struct ext4_buddy e4b;
5307 int err = 0, ret, free_clusters_count;
5308 ext4_grpblk_t clusters_freed;
5309 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
5310 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
5311 unsigned long cluster_count = last_cluster - first_cluster + 1;
5313 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
5318 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5320 * Check to see if we are freeing blocks across a group
5323 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
5324 ext4_warning(sb, "too many blocks added to group %u",
5330 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5331 if (IS_ERR(bitmap_bh)) {
5332 err = PTR_ERR(bitmap_bh);
5337 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
5343 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
5344 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
5345 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
5346 in_range(block + count - 1, ext4_inode_table(sb, desc),
5347 sbi->s_itb_per_group)) {
5348 ext4_error(sb, "Adding blocks in system zones - "
5349 "Block = %llu, count = %lu",
5355 BUFFER_TRACE(bitmap_bh, "getting write access");
5356 err = ext4_journal_get_write_access(handle, bitmap_bh);
5361 * We are about to modify some metadata. Call the journal APIs
5362 * to unshare ->b_data if a currently-committing transaction is
5365 BUFFER_TRACE(gd_bh, "get_write_access");
5366 err = ext4_journal_get_write_access(handle, gd_bh);
5370 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
5371 BUFFER_TRACE(bitmap_bh, "clear bit");
5372 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
5373 ext4_error(sb, "bit already cleared for block %llu",
5374 (ext4_fsblk_t)(block + i));
5375 BUFFER_TRACE(bitmap_bh, "bit already cleared");
5381 err = ext4_mb_load_buddy(sb, block_group, &e4b);
5386 * need to update group_info->bb_free and bitmap
5387 * with group lock held. generate_buddy look at
5388 * them with group lock_held
5390 ext4_lock_group(sb, block_group);
5391 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
5392 mb_free_blocks(NULL, &e4b, bit, cluster_count);
5393 free_clusters_count = clusters_freed +
5394 ext4_free_group_clusters(sb, desc);
5395 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
5396 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5397 ext4_group_desc_csum_set(sb, block_group, desc);
5398 ext4_unlock_group(sb, block_group);
5399 percpu_counter_add(&sbi->s_freeclusters_counter,
5402 if (sbi->s_log_groups_per_flex) {
5403 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5404 atomic64_add(clusters_freed,
5405 &sbi_array_rcu_deref(sbi, s_flex_groups,
5406 flex_group)->free_clusters);
5409 ext4_mb_unload_buddy(&e4b);
5411 /* We dirtied the bitmap block */
5412 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5413 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5415 /* And the group descriptor block */
5416 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5417 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5423 ext4_std_error(sb, err);
5428 * ext4_trim_extent -- function to TRIM one single free extent in the group
5429 * @sb: super block for the file system
5430 * @start: starting block of the free extent in the alloc. group
5431 * @count: number of blocks to TRIM
5432 * @group: alloc. group we are working with
5433 * @e4b: ext4 buddy for the group
5435 * Trim "count" blocks starting at "start" in the "group". To assure that no
5436 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5437 * be called with under the group lock.
5439 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5440 ext4_group_t group, struct ext4_buddy *e4b)
5444 struct ext4_free_extent ex;
5447 trace_ext4_trim_extent(sb, group, start, count);
5449 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5451 ex.fe_start = start;
5452 ex.fe_group = group;
5456 * Mark blocks used, so no one can reuse them while
5459 mb_mark_used(e4b, &ex);
5460 ext4_unlock_group(sb, group);
5461 ret = ext4_issue_discard(sb, group, start, count, NULL);
5462 ext4_lock_group(sb, group);
5463 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5468 * ext4_trim_all_free -- function to trim all free space in alloc. group
5469 * @sb: super block for file system
5470 * @group: group to be trimmed
5471 * @start: first group block to examine
5472 * @max: last group block to examine
5473 * @minblocks: minimum extent block count
5475 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5476 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5480 * ext4_trim_all_free walks through group's block bitmap searching for free
5481 * extents. When the free extent is found, mark it as used in group buddy
5482 * bitmap. Then issue a TRIM command on this extent and free the extent in
5483 * the group buddy bitmap. This is done until whole group is scanned.
5485 static ext4_grpblk_t
5486 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5487 ext4_grpblk_t start, ext4_grpblk_t max,
5488 ext4_grpblk_t minblocks)
5491 ext4_grpblk_t next, count = 0, free_count = 0;
5492 struct ext4_buddy e4b;
5495 trace_ext4_trim_all_free(sb, group, start, max);
5497 ret = ext4_mb_load_buddy(sb, group, &e4b);
5499 ext4_warning(sb, "Error %d loading buddy information for %u",
5503 bitmap = e4b.bd_bitmap;
5505 ext4_lock_group(sb, group);
5506 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5507 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5510 start = (e4b.bd_info->bb_first_free > start) ?
5511 e4b.bd_info->bb_first_free : start;
5513 while (start <= max) {
5514 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5517 next = mb_find_next_bit(bitmap, max + 1, start);
5519 if ((next - start) >= minblocks) {
5520 ret = ext4_trim_extent(sb, start,
5521 next - start, group, &e4b);
5522 if (ret && ret != -EOPNOTSUPP)
5525 count += next - start;
5527 free_count += next - start;
5530 if (fatal_signal_pending(current)) {
5531 count = -ERESTARTSYS;
5535 if (need_resched()) {
5536 ext4_unlock_group(sb, group);
5538 ext4_lock_group(sb, group);
5541 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5547 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5550 ext4_unlock_group(sb, group);
5551 ext4_mb_unload_buddy(&e4b);
5553 ext4_debug("trimmed %d blocks in the group %d\n",
5560 * ext4_trim_fs() -- trim ioctl handle function
5561 * @sb: superblock for filesystem
5562 * @range: fstrim_range structure
5564 * start: First Byte to trim
5565 * len: number of Bytes to trim from start
5566 * minlen: minimum extent length in Bytes
5567 * ext4_trim_fs goes through all allocation groups containing Bytes from
5568 * start to start+len. For each such a group ext4_trim_all_free function
5569 * is invoked to trim all free space.
5571 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5573 struct ext4_group_info *grp;
5574 ext4_group_t group, first_group, last_group;
5575 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5576 uint64_t start, end, minlen, trimmed = 0;
5577 ext4_fsblk_t first_data_blk =
5578 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5579 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5582 start = range->start >> sb->s_blocksize_bits;
5583 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5584 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5585 range->minlen >> sb->s_blocksize_bits);
5587 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5588 start >= max_blks ||
5589 range->len < sb->s_blocksize)
5591 if (end >= max_blks)
5593 if (end <= first_data_blk)
5595 if (start < first_data_blk)
5596 start = first_data_blk;
5598 /* Determine first and last group to examine based on start and end */
5599 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5600 &first_group, &first_cluster);
5601 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5602 &last_group, &last_cluster);
5604 /* end now represents the last cluster to discard in this group */
5605 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5607 for (group = first_group; group <= last_group; group++) {
5608 grp = ext4_get_group_info(sb, group);
5609 /* We only do this if the grp has never been initialized */
5610 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5611 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5617 * For all the groups except the last one, last cluster will
5618 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5619 * change it for the last group, note that last_cluster is
5620 * already computed earlier by ext4_get_group_no_and_offset()
5622 if (group == last_group)
5625 if (grp->bb_free >= minlen) {
5626 cnt = ext4_trim_all_free(sb, group, first_cluster,
5636 * For every group except the first one, we are sure
5637 * that the first cluster to discard will be cluster #0.
5643 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5646 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5650 /* Iterate all the free extents in the group. */
5652 ext4_mballoc_query_range(
5653 struct super_block *sb,
5655 ext4_grpblk_t start,
5657 ext4_mballoc_query_range_fn formatter,
5662 struct ext4_buddy e4b;
5665 error = ext4_mb_load_buddy(sb, group, &e4b);
5668 bitmap = e4b.bd_bitmap;
5670 ext4_lock_group(sb, group);
5672 start = (e4b.bd_info->bb_first_free > start) ?
5673 e4b.bd_info->bb_first_free : start;
5674 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
5675 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5677 while (start <= end) {
5678 start = mb_find_next_zero_bit(bitmap, end + 1, start);
5681 next = mb_find_next_bit(bitmap, end + 1, start);
5683 ext4_unlock_group(sb, group);
5684 error = formatter(sb, group, start, next - start, priv);
5687 ext4_lock_group(sb, group);
5692 ext4_unlock_group(sb, group);
5694 ext4_mb_unload_buddy(&e4b);