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 * 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;
622 if (e4b->bd_info->bb_check_counter++ % 10)
626 buddy = mb_find_buddy(e4b, order, &max);
627 MB_CHECK_ASSERT(buddy);
628 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
629 MB_CHECK_ASSERT(buddy2);
630 MB_CHECK_ASSERT(buddy != buddy2);
631 MB_CHECK_ASSERT(max * 2 == max2);
634 for (i = 0; i < max; i++) {
636 if (mb_test_bit(i, buddy)) {
637 /* only single bit in buddy2 may be 1 */
638 if (!mb_test_bit(i << 1, buddy2)) {
640 mb_test_bit((i<<1)+1, buddy2));
641 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
643 mb_test_bit(i << 1, buddy2));
648 /* both bits in buddy2 must be 1 */
649 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
650 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
652 for (j = 0; j < (1 << order); j++) {
653 k = (i * (1 << order)) + j;
655 !mb_test_bit(k, e4b->bd_bitmap));
659 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
664 buddy = mb_find_buddy(e4b, 0, &max);
665 for (i = 0; i < max; i++) {
666 if (!mb_test_bit(i, buddy)) {
667 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
675 /* check used bits only */
676 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
677 buddy2 = mb_find_buddy(e4b, j, &max2);
679 MB_CHECK_ASSERT(k < max2);
680 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
683 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
684 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
686 grp = ext4_get_group_info(sb, e4b->bd_group);
687 list_for_each(cur, &grp->bb_prealloc_list) {
688 ext4_group_t groupnr;
689 struct ext4_prealloc_space *pa;
690 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
691 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
692 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
693 for (i = 0; i < pa->pa_len; i++)
694 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
698 #undef MB_CHECK_ASSERT
699 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
700 __FILE__, __func__, __LINE__)
702 #define mb_check_buddy(e4b)
706 * Divide blocks started from @first with length @len into
707 * smaller chunks with power of 2 blocks.
708 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
709 * then increase bb_counters[] for corresponded chunk size.
711 static void ext4_mb_mark_free_simple(struct super_block *sb,
712 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
713 struct ext4_group_info *grp)
715 struct ext4_sb_info *sbi = EXT4_SB(sb);
721 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
723 border = 2 << sb->s_blocksize_bits;
726 /* find how many blocks can be covered since this position */
727 max = ffs(first | border) - 1;
729 /* find how many blocks of power 2 we need to mark */
736 /* mark multiblock chunks only */
737 grp->bb_counters[min]++;
739 mb_clear_bit(first >> min,
740 buddy + sbi->s_mb_offsets[min]);
748 * Cache the order of the largest free extent we have available in this block
752 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
757 grp->bb_largest_free_order = -1; /* uninit */
759 bits = sb->s_blocksize_bits + 1;
760 for (i = bits; i >= 0; i--) {
761 if (grp->bb_counters[i] > 0) {
762 grp->bb_largest_free_order = i;
768 static noinline_for_stack
769 void ext4_mb_generate_buddy(struct super_block *sb,
770 void *buddy, void *bitmap, ext4_group_t group)
772 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
773 struct ext4_sb_info *sbi = EXT4_SB(sb);
774 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
779 unsigned fragments = 0;
780 unsigned long long period = get_cycles();
782 /* initialize buddy from bitmap which is aggregation
783 * of on-disk bitmap and preallocations */
784 i = mb_find_next_zero_bit(bitmap, max, 0);
785 grp->bb_first_free = i;
789 i = mb_find_next_bit(bitmap, max, i);
793 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
795 grp->bb_counters[0]++;
797 i = mb_find_next_zero_bit(bitmap, max, i);
799 grp->bb_fragments = fragments;
801 if (free != grp->bb_free) {
802 ext4_grp_locked_error(sb, group, 0, 0,
803 "block bitmap and bg descriptor "
804 "inconsistent: %u vs %u free clusters",
807 * If we intend to continue, we consider group descriptor
808 * corrupt and update bb_free using bitmap value
811 ext4_mark_group_bitmap_corrupted(sb, group,
812 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
814 mb_set_largest_free_order(sb, grp);
816 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
818 period = get_cycles() - period;
819 spin_lock(&sbi->s_bal_lock);
820 sbi->s_mb_buddies_generated++;
821 sbi->s_mb_generation_time += period;
822 spin_unlock(&sbi->s_bal_lock);
825 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
831 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
832 ext4_set_bits(buddy, 0, count);
834 e4b->bd_info->bb_fragments = 0;
835 memset(e4b->bd_info->bb_counters, 0,
836 sizeof(*e4b->bd_info->bb_counters) *
837 (e4b->bd_sb->s_blocksize_bits + 2));
839 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
840 e4b->bd_bitmap, e4b->bd_group);
843 /* The buddy information is attached the buddy cache inode
844 * for convenience. The information regarding each group
845 * is loaded via ext4_mb_load_buddy. The information involve
846 * block bitmap and buddy information. The information are
847 * stored in the inode as
850 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
853 * one block each for bitmap and buddy information.
854 * So for each group we take up 2 blocks. A page can
855 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
856 * So it can have information regarding groups_per_page which
857 * is blocks_per_page/2
859 * Locking note: This routine takes the block group lock of all groups
860 * for this page; do not hold this lock when calling this routine!
863 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
865 ext4_group_t ngroups;
871 ext4_group_t first_group, group;
873 struct super_block *sb;
874 struct buffer_head *bhs;
875 struct buffer_head **bh = NULL;
879 struct ext4_group_info *grinfo;
881 inode = page->mapping->host;
883 ngroups = ext4_get_groups_count(sb);
884 blocksize = i_blocksize(inode);
885 blocks_per_page = PAGE_SIZE / blocksize;
887 mb_debug(sb, "init page %lu\n", page->index);
889 groups_per_page = blocks_per_page >> 1;
890 if (groups_per_page == 0)
893 /* allocate buffer_heads to read bitmaps */
894 if (groups_per_page > 1) {
895 i = sizeof(struct buffer_head *) * groups_per_page;
896 bh = kzalloc(i, gfp);
904 first_group = page->index * blocks_per_page / 2;
906 /* read all groups the page covers into the cache */
907 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
908 if (group >= ngroups)
911 grinfo = ext4_get_group_info(sb, group);
913 * If page is uptodate then we came here after online resize
914 * which added some new uninitialized group info structs, so
915 * we must skip all initialized uptodate buddies on the page,
916 * which may be currently in use by an allocating task.
918 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
922 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
924 err = PTR_ERR(bh[i]);
928 mb_debug(sb, "read bitmap for group %u\n", group);
931 /* wait for I/O completion */
932 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
937 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
942 first_block = page->index * blocks_per_page;
943 for (i = 0; i < blocks_per_page; i++) {
944 group = (first_block + i) >> 1;
945 if (group >= ngroups)
948 if (!bh[group - first_group])
949 /* skip initialized uptodate buddy */
952 if (!buffer_verified(bh[group - first_group]))
953 /* Skip faulty bitmaps */
958 * data carry information regarding this
959 * particular group in the format specified
963 data = page_address(page) + (i * blocksize);
964 bitmap = bh[group - first_group]->b_data;
967 * We place the buddy block and bitmap block
970 if ((first_block + i) & 1) {
971 /* this is block of buddy */
972 BUG_ON(incore == NULL);
973 mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
974 group, page->index, i * blocksize);
975 trace_ext4_mb_buddy_bitmap_load(sb, group);
976 grinfo = ext4_get_group_info(sb, group);
977 grinfo->bb_fragments = 0;
978 memset(grinfo->bb_counters, 0,
979 sizeof(*grinfo->bb_counters) *
980 (sb->s_blocksize_bits+2));
982 * incore got set to the group block bitmap below
984 ext4_lock_group(sb, group);
986 memset(data, 0xff, blocksize);
987 ext4_mb_generate_buddy(sb, data, incore, group);
988 ext4_unlock_group(sb, group);
991 /* this is block of bitmap */
992 BUG_ON(incore != NULL);
993 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
994 group, page->index, i * blocksize);
995 trace_ext4_mb_bitmap_load(sb, group);
997 /* see comments in ext4_mb_put_pa() */
998 ext4_lock_group(sb, group);
999 memcpy(data, bitmap, blocksize);
1001 /* mark all preallocated blks used in in-core bitmap */
1002 ext4_mb_generate_from_pa(sb, data, group);
1003 ext4_mb_generate_from_freelist(sb, data, group);
1004 ext4_unlock_group(sb, group);
1006 /* set incore so that the buddy information can be
1007 * generated using this
1012 SetPageUptodate(page);
1016 for (i = 0; i < groups_per_page; i++)
1025 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1026 * on the same buddy page doesn't happen whild holding the buddy page lock.
1027 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1028 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1030 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1031 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1033 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1034 int block, pnum, poff;
1035 int blocks_per_page;
1038 e4b->bd_buddy_page = NULL;
1039 e4b->bd_bitmap_page = NULL;
1041 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1043 * the buddy cache inode stores the block bitmap
1044 * and buddy information in consecutive blocks.
1045 * So for each group we need two blocks.
1048 pnum = block / blocks_per_page;
1049 poff = block % blocks_per_page;
1050 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1053 BUG_ON(page->mapping != inode->i_mapping);
1054 e4b->bd_bitmap_page = page;
1055 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1057 if (blocks_per_page >= 2) {
1058 /* buddy and bitmap are on the same page */
1063 pnum = block / blocks_per_page;
1064 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1067 BUG_ON(page->mapping != inode->i_mapping);
1068 e4b->bd_buddy_page = page;
1072 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1074 if (e4b->bd_bitmap_page) {
1075 unlock_page(e4b->bd_bitmap_page);
1076 put_page(e4b->bd_bitmap_page);
1078 if (e4b->bd_buddy_page) {
1079 unlock_page(e4b->bd_buddy_page);
1080 put_page(e4b->bd_buddy_page);
1085 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1086 * block group lock of all groups for this page; do not hold the BG lock when
1087 * calling this routine!
1089 static noinline_for_stack
1090 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1093 struct ext4_group_info *this_grp;
1094 struct ext4_buddy e4b;
1099 mb_debug(sb, "init group %u\n", group);
1100 this_grp = ext4_get_group_info(sb, group);
1102 * This ensures that we don't reinit the buddy cache
1103 * page which map to the group from which we are already
1104 * allocating. If we are looking at the buddy cache we would
1105 * have taken a reference using ext4_mb_load_buddy and that
1106 * would have pinned buddy page to page cache.
1107 * The call to ext4_mb_get_buddy_page_lock will mark the
1110 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1111 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1113 * somebody initialized the group
1114 * return without doing anything
1119 page = e4b.bd_bitmap_page;
1120 ret = ext4_mb_init_cache(page, NULL, gfp);
1123 if (!PageUptodate(page)) {
1128 if (e4b.bd_buddy_page == NULL) {
1130 * If both the bitmap and buddy are in
1131 * the same page we don't need to force
1137 /* init buddy cache */
1138 page = e4b.bd_buddy_page;
1139 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1142 if (!PageUptodate(page)) {
1147 ext4_mb_put_buddy_page_lock(&e4b);
1152 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1153 * block group lock of all groups for this page; do not hold the BG lock when
1154 * calling this routine!
1156 static noinline_for_stack int
1157 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1158 struct ext4_buddy *e4b, gfp_t gfp)
1160 int blocks_per_page;
1166 struct ext4_group_info *grp;
1167 struct ext4_sb_info *sbi = EXT4_SB(sb);
1168 struct inode *inode = sbi->s_buddy_cache;
1171 mb_debug(sb, "load group %u\n", group);
1173 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1174 grp = ext4_get_group_info(sb, group);
1176 e4b->bd_blkbits = sb->s_blocksize_bits;
1179 e4b->bd_group = group;
1180 e4b->bd_buddy_page = NULL;
1181 e4b->bd_bitmap_page = NULL;
1183 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1185 * we need full data about the group
1186 * to make a good selection
1188 ret = ext4_mb_init_group(sb, group, gfp);
1194 * the buddy cache inode stores the block bitmap
1195 * and buddy information in consecutive blocks.
1196 * So for each group we need two blocks.
1199 pnum = block / blocks_per_page;
1200 poff = block % blocks_per_page;
1202 /* we could use find_or_create_page(), but it locks page
1203 * what we'd like to avoid in fast path ... */
1204 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1205 if (page == NULL || !PageUptodate(page)) {
1208 * drop the page reference and try
1209 * to get the page with lock. If we
1210 * are not uptodate that implies
1211 * somebody just created the page but
1212 * is yet to initialize the same. So
1213 * wait for it to initialize.
1216 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1218 BUG_ON(page->mapping != inode->i_mapping);
1219 if (!PageUptodate(page)) {
1220 ret = ext4_mb_init_cache(page, NULL, gfp);
1225 mb_cmp_bitmaps(e4b, page_address(page) +
1226 (poff * sb->s_blocksize));
1235 if (!PageUptodate(page)) {
1240 /* Pages marked accessed already */
1241 e4b->bd_bitmap_page = page;
1242 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1245 pnum = block / blocks_per_page;
1246 poff = block % blocks_per_page;
1248 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1249 if (page == NULL || !PageUptodate(page)) {
1252 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1254 BUG_ON(page->mapping != inode->i_mapping);
1255 if (!PageUptodate(page)) {
1256 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1270 if (!PageUptodate(page)) {
1275 /* Pages marked accessed already */
1276 e4b->bd_buddy_page = page;
1277 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1284 if (e4b->bd_bitmap_page)
1285 put_page(e4b->bd_bitmap_page);
1286 if (e4b->bd_buddy_page)
1287 put_page(e4b->bd_buddy_page);
1288 e4b->bd_buddy = NULL;
1289 e4b->bd_bitmap = NULL;
1293 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1294 struct ext4_buddy *e4b)
1296 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1299 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1301 if (e4b->bd_bitmap_page)
1302 put_page(e4b->bd_bitmap_page);
1303 if (e4b->bd_buddy_page)
1304 put_page(e4b->bd_buddy_page);
1308 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1311 int bb_incr = 1 << (e4b->bd_blkbits - 1);
1314 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1315 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1318 while (order <= e4b->bd_blkbits + 1) {
1320 if (!mb_test_bit(block, bb)) {
1321 /* this block is part of buddy of order 'order' */
1331 static void mb_clear_bits(void *bm, int cur, int len)
1337 if ((cur & 31) == 0 && (len - cur) >= 32) {
1338 /* fast path: clear whole word at once */
1339 addr = bm + (cur >> 3);
1344 mb_clear_bit(cur, bm);
1349 /* clear bits in given range
1350 * will return first found zero bit if any, -1 otherwise
1352 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1359 if ((cur & 31) == 0 && (len - cur) >= 32) {
1360 /* fast path: clear whole word at once */
1361 addr = bm + (cur >> 3);
1362 if (*addr != (__u32)(-1) && zero_bit == -1)
1363 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1368 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1376 void ext4_set_bits(void *bm, int cur, int len)
1382 if ((cur & 31) == 0 && (len - cur) >= 32) {
1383 /* fast path: set whole word at once */
1384 addr = bm + (cur >> 3);
1389 mb_set_bit(cur, bm);
1394 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1396 if (mb_test_bit(*bit + side, bitmap)) {
1397 mb_clear_bit(*bit, bitmap);
1403 mb_set_bit(*bit, bitmap);
1408 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1412 void *buddy = mb_find_buddy(e4b, order, &max);
1417 /* Bits in range [first; last] are known to be set since
1418 * corresponding blocks were allocated. Bits in range
1419 * (first; last) will stay set because they form buddies on
1420 * upper layer. We just deal with borders if they don't
1421 * align with upper layer and then go up.
1422 * Releasing entire group is all about clearing
1423 * single bit of highest order buddy.
1427 * ---------------------------------
1429 * ---------------------------------
1430 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1431 * ---------------------------------
1433 * \_____________________/
1435 * Neither [1] nor [6] is aligned to above layer.
1436 * Left neighbour [0] is free, so mark it busy,
1437 * decrease bb_counters and extend range to
1439 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1440 * mark [6] free, increase bb_counters and shrink range to
1442 * Then shift range to [0; 2], go up and do the same.
1447 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1449 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1454 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1455 mb_clear_bits(buddy, first, last - first + 1);
1456 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1465 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1466 int first, int count)
1468 int left_is_free = 0;
1469 int right_is_free = 0;
1471 int last = first + count - 1;
1472 struct super_block *sb = e4b->bd_sb;
1474 if (WARN_ON(count == 0))
1476 BUG_ON(last >= (sb->s_blocksize << 3));
1477 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1478 /* Don't bother if the block group is corrupt. */
1479 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1482 mb_check_buddy(e4b);
1483 mb_free_blocks_double(inode, e4b, first, count);
1485 this_cpu_inc(discard_pa_seq);
1486 e4b->bd_info->bb_free += count;
1487 if (first < e4b->bd_info->bb_first_free)
1488 e4b->bd_info->bb_first_free = first;
1490 /* access memory sequentially: check left neighbour,
1491 * clear range and then check right neighbour
1494 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1495 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1496 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1497 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1499 if (unlikely(block != -1)) {
1500 struct ext4_sb_info *sbi = EXT4_SB(sb);
1501 ext4_fsblk_t blocknr;
1503 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1504 blocknr += EXT4_C2B(sbi, block);
1505 if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) {
1506 ext4_grp_locked_error(sb, e4b->bd_group,
1507 inode ? inode->i_ino : 0,
1509 "freeing already freed block (bit %u); block bitmap corrupt.",
1511 ext4_mark_group_bitmap_corrupted(
1513 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1515 mb_regenerate_buddy(e4b);
1519 /* let's maintain fragments counter */
1520 if (left_is_free && right_is_free)
1521 e4b->bd_info->bb_fragments--;
1522 else if (!left_is_free && !right_is_free)
1523 e4b->bd_info->bb_fragments++;
1525 /* buddy[0] == bd_bitmap is a special case, so handle
1526 * it right away and let mb_buddy_mark_free stay free of
1527 * zero order checks.
1528 * Check if neighbours are to be coaleasced,
1529 * adjust bitmap bb_counters and borders appropriately.
1532 first += !left_is_free;
1533 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1536 last -= !right_is_free;
1537 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1541 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1544 mb_set_largest_free_order(sb, e4b->bd_info);
1545 mb_check_buddy(e4b);
1548 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1549 int needed, struct ext4_free_extent *ex)
1555 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1558 buddy = mb_find_buddy(e4b, 0, &max);
1559 BUG_ON(buddy == NULL);
1560 BUG_ON(block >= max);
1561 if (mb_test_bit(block, buddy)) {
1568 /* find actual order */
1569 order = mb_find_order_for_block(e4b, block);
1570 block = block >> order;
1572 ex->fe_len = 1 << order;
1573 ex->fe_start = block << order;
1574 ex->fe_group = e4b->bd_group;
1576 /* calc difference from given start */
1577 next = next - ex->fe_start;
1579 ex->fe_start += next;
1581 while (needed > ex->fe_len &&
1582 mb_find_buddy(e4b, order, &max)) {
1584 if (block + 1 >= max)
1587 next = (block + 1) * (1 << order);
1588 if (mb_test_bit(next, e4b->bd_bitmap))
1591 order = mb_find_order_for_block(e4b, next);
1593 block = next >> order;
1594 ex->fe_len += 1 << order;
1597 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1598 /* Should never happen! (but apparently sometimes does?!?) */
1600 ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent "
1601 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1602 block, order, needed, ex->fe_group, ex->fe_start,
1603 ex->fe_len, ex->fe_logical);
1611 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1617 int start = ex->fe_start;
1618 int len = ex->fe_len;
1623 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1624 BUG_ON(e4b->bd_group != ex->fe_group);
1625 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1626 mb_check_buddy(e4b);
1627 mb_mark_used_double(e4b, start, len);
1629 this_cpu_inc(discard_pa_seq);
1630 e4b->bd_info->bb_free -= len;
1631 if (e4b->bd_info->bb_first_free == start)
1632 e4b->bd_info->bb_first_free += len;
1634 /* let's maintain fragments counter */
1636 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1637 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1638 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1640 e4b->bd_info->bb_fragments++;
1641 else if (!mlen && !max)
1642 e4b->bd_info->bb_fragments--;
1644 /* let's maintain buddy itself */
1646 ord = mb_find_order_for_block(e4b, start);
1648 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1649 /* the whole chunk may be allocated at once! */
1651 buddy = mb_find_buddy(e4b, ord, &max);
1652 BUG_ON((start >> ord) >= max);
1653 mb_set_bit(start >> ord, buddy);
1654 e4b->bd_info->bb_counters[ord]--;
1661 /* store for history */
1663 ret = len | (ord << 16);
1665 /* we have to split large buddy */
1667 buddy = mb_find_buddy(e4b, ord, &max);
1668 mb_set_bit(start >> ord, buddy);
1669 e4b->bd_info->bb_counters[ord]--;
1672 cur = (start >> ord) & ~1U;
1673 buddy = mb_find_buddy(e4b, ord, &max);
1674 mb_clear_bit(cur, buddy);
1675 mb_clear_bit(cur + 1, buddy);
1676 e4b->bd_info->bb_counters[ord]++;
1677 e4b->bd_info->bb_counters[ord]++;
1679 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1681 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1682 mb_check_buddy(e4b);
1688 * Must be called under group lock!
1690 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1691 struct ext4_buddy *e4b)
1693 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1696 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1697 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1699 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1700 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1701 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1703 /* preallocation can change ac_b_ex, thus we store actually
1704 * allocated blocks for history */
1705 ac->ac_f_ex = ac->ac_b_ex;
1707 ac->ac_status = AC_STATUS_FOUND;
1708 ac->ac_tail = ret & 0xffff;
1709 ac->ac_buddy = ret >> 16;
1712 * take the page reference. We want the page to be pinned
1713 * so that we don't get a ext4_mb_init_cache_call for this
1714 * group until we update the bitmap. That would mean we
1715 * double allocate blocks. The reference is dropped
1716 * in ext4_mb_release_context
1718 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1719 get_page(ac->ac_bitmap_page);
1720 ac->ac_buddy_page = e4b->bd_buddy_page;
1721 get_page(ac->ac_buddy_page);
1722 /* store last allocated for subsequent stream allocation */
1723 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1724 spin_lock(&sbi->s_md_lock);
1725 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1726 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1727 spin_unlock(&sbi->s_md_lock);
1730 * As we've just preallocated more space than
1731 * user requested originally, we store allocated
1732 * space in a special descriptor.
1734 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
1735 ext4_mb_new_preallocation(ac);
1739 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1740 struct ext4_buddy *e4b,
1743 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1744 struct ext4_free_extent *bex = &ac->ac_b_ex;
1745 struct ext4_free_extent *gex = &ac->ac_g_ex;
1746 struct ext4_free_extent ex;
1749 if (ac->ac_status == AC_STATUS_FOUND)
1752 * We don't want to scan for a whole year
1754 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1755 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1756 ac->ac_status = AC_STATUS_BREAK;
1761 * Haven't found good chunk so far, let's continue
1763 if (bex->fe_len < gex->fe_len)
1766 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1767 && bex->fe_group == e4b->bd_group) {
1768 /* recheck chunk's availability - we don't know
1769 * when it was found (within this lock-unlock
1771 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1772 if (max >= gex->fe_len) {
1773 ext4_mb_use_best_found(ac, e4b);
1780 * The routine checks whether found extent is good enough. If it is,
1781 * then the extent gets marked used and flag is set to the context
1782 * to stop scanning. Otherwise, the extent is compared with the
1783 * previous found extent and if new one is better, then it's stored
1784 * in the context. Later, the best found extent will be used, if
1785 * mballoc can't find good enough extent.
1787 * FIXME: real allocation policy is to be designed yet!
1789 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1790 struct ext4_free_extent *ex,
1791 struct ext4_buddy *e4b)
1793 struct ext4_free_extent *bex = &ac->ac_b_ex;
1794 struct ext4_free_extent *gex = &ac->ac_g_ex;
1796 BUG_ON(ex->fe_len <= 0);
1797 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1798 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1799 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1804 * The special case - take what you catch first
1806 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1808 ext4_mb_use_best_found(ac, e4b);
1813 * Let's check whether the chuck is good enough
1815 if (ex->fe_len == gex->fe_len) {
1817 ext4_mb_use_best_found(ac, e4b);
1822 * If this is first found extent, just store it in the context
1824 if (bex->fe_len == 0) {
1830 * If new found extent is better, store it in the context
1832 if (bex->fe_len < gex->fe_len) {
1833 /* if the request isn't satisfied, any found extent
1834 * larger than previous best one is better */
1835 if (ex->fe_len > bex->fe_len)
1837 } else if (ex->fe_len > gex->fe_len) {
1838 /* if the request is satisfied, then we try to find
1839 * an extent that still satisfy the request, but is
1840 * smaller than previous one */
1841 if (ex->fe_len < bex->fe_len)
1845 ext4_mb_check_limits(ac, e4b, 0);
1848 static noinline_for_stack
1849 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1850 struct ext4_buddy *e4b)
1852 struct ext4_free_extent ex = ac->ac_b_ex;
1853 ext4_group_t group = ex.fe_group;
1857 BUG_ON(ex.fe_len <= 0);
1858 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1862 ext4_lock_group(ac->ac_sb, group);
1863 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1867 ext4_mb_use_best_found(ac, e4b);
1870 ext4_unlock_group(ac->ac_sb, group);
1871 ext4_mb_unload_buddy(e4b);
1876 static noinline_for_stack
1877 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1878 struct ext4_buddy *e4b)
1880 ext4_group_t group = ac->ac_g_ex.fe_group;
1883 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1884 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1885 struct ext4_free_extent ex;
1887 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1889 if (grp->bb_free == 0)
1892 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1896 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1897 ext4_mb_unload_buddy(e4b);
1901 ext4_lock_group(ac->ac_sb, group);
1902 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1903 ac->ac_g_ex.fe_len, &ex);
1904 ex.fe_logical = 0xDEADFA11; /* debug value */
1906 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1909 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1911 /* use do_div to get remainder (would be 64-bit modulo) */
1912 if (do_div(start, sbi->s_stripe) == 0) {
1915 ext4_mb_use_best_found(ac, e4b);
1917 } else if (max >= ac->ac_g_ex.fe_len) {
1918 BUG_ON(ex.fe_len <= 0);
1919 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1920 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1923 ext4_mb_use_best_found(ac, e4b);
1924 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1925 /* Sometimes, caller may want to merge even small
1926 * number of blocks to an existing extent */
1927 BUG_ON(ex.fe_len <= 0);
1928 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1929 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1932 ext4_mb_use_best_found(ac, e4b);
1934 ext4_unlock_group(ac->ac_sb, group);
1935 ext4_mb_unload_buddy(e4b);
1941 * The routine scans buddy structures (not bitmap!) from given order
1942 * to max order and tries to find big enough chunk to satisfy the req
1944 static noinline_for_stack
1945 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1946 struct ext4_buddy *e4b)
1948 struct super_block *sb = ac->ac_sb;
1949 struct ext4_group_info *grp = e4b->bd_info;
1955 BUG_ON(ac->ac_2order <= 0);
1956 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1957 if (grp->bb_counters[i] == 0)
1960 buddy = mb_find_buddy(e4b, i, &max);
1961 BUG_ON(buddy == NULL);
1963 k = mb_find_next_zero_bit(buddy, max, 0);
1965 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
1966 "%d free clusters of order %d. But found 0",
1967 grp->bb_counters[i], i);
1968 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
1970 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1975 ac->ac_b_ex.fe_len = 1 << i;
1976 ac->ac_b_ex.fe_start = k << i;
1977 ac->ac_b_ex.fe_group = e4b->bd_group;
1979 ext4_mb_use_best_found(ac, e4b);
1981 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
1983 if (EXT4_SB(sb)->s_mb_stats)
1984 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1991 * The routine scans the group and measures all found extents.
1992 * In order to optimize scanning, caller must pass number of
1993 * free blocks in the group, so the routine can know upper limit.
1995 static noinline_for_stack
1996 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1997 struct ext4_buddy *e4b)
1999 struct super_block *sb = ac->ac_sb;
2000 void *bitmap = e4b->bd_bitmap;
2001 struct ext4_free_extent ex;
2005 free = e4b->bd_info->bb_free;
2006 if (WARN_ON(free <= 0))
2009 i = e4b->bd_info->bb_first_free;
2011 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2012 i = mb_find_next_zero_bit(bitmap,
2013 EXT4_CLUSTERS_PER_GROUP(sb), i);
2014 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2016 * IF we have corrupt bitmap, we won't find any
2017 * free blocks even though group info says we
2020 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2021 "%d free clusters as per "
2022 "group info. But bitmap says 0",
2024 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2025 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2029 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2030 if (WARN_ON(ex.fe_len <= 0))
2032 if (free < ex.fe_len) {
2033 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2034 "%d free clusters as per "
2035 "group info. But got %d blocks",
2037 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2038 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2040 * The number of free blocks differs. This mostly
2041 * indicate that the bitmap is corrupt. So exit
2042 * without claiming the space.
2046 ex.fe_logical = 0xDEADC0DE; /* debug value */
2047 ext4_mb_measure_extent(ac, &ex, e4b);
2053 ext4_mb_check_limits(ac, e4b, 1);
2057 * This is a special case for storages like raid5
2058 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2060 static noinline_for_stack
2061 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2062 struct ext4_buddy *e4b)
2064 struct super_block *sb = ac->ac_sb;
2065 struct ext4_sb_info *sbi = EXT4_SB(sb);
2066 void *bitmap = e4b->bd_bitmap;
2067 struct ext4_free_extent ex;
2068 ext4_fsblk_t first_group_block;
2073 BUG_ON(sbi->s_stripe == 0);
2075 /* find first stripe-aligned block in group */
2076 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2078 a = first_group_block + sbi->s_stripe - 1;
2079 do_div(a, sbi->s_stripe);
2080 i = (a * sbi->s_stripe) - first_group_block;
2082 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2083 if (!mb_test_bit(i, bitmap)) {
2084 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2085 if (max >= sbi->s_stripe) {
2087 ex.fe_logical = 0xDEADF00D; /* debug value */
2089 ext4_mb_use_best_found(ac, e4b);
2098 * This is also called BEFORE we load the buddy bitmap.
2099 * Returns either 1 or 0 indicating that the group is either suitable
2100 * for the allocation or not.
2102 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2103 ext4_group_t group, int cr)
2105 ext4_grpblk_t free, fragments;
2106 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2107 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2109 BUG_ON(cr < 0 || cr >= 4);
2111 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2114 free = grp->bb_free;
2118 fragments = grp->bb_fragments;
2124 BUG_ON(ac->ac_2order == 0);
2126 /* Avoid using the first bg of a flexgroup for data files */
2127 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2128 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2129 ((group % flex_size) == 0))
2132 if (free < ac->ac_g_ex.fe_len)
2135 if (ac->ac_2order > ac->ac_sb->s_blocksize_bits+1)
2138 if (grp->bb_largest_free_order < ac->ac_2order)
2143 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2147 if (free >= ac->ac_g_ex.fe_len)
2160 * This could return negative error code if something goes wrong
2161 * during ext4_mb_init_group(). This should not be called with
2162 * ext4_lock_group() held.
2164 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2165 ext4_group_t group, int cr)
2167 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2168 struct super_block *sb = ac->ac_sb;
2169 struct ext4_sb_info *sbi = EXT4_SB(sb);
2170 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2175 ext4_lock_group(sb, group);
2176 free = grp->bb_free;
2179 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2181 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2184 ext4_unlock_group(sb, group);
2186 /* We only do this if the grp has never been initialized */
2187 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2188 struct ext4_group_desc *gdp =
2189 ext4_get_group_desc(sb, group, NULL);
2192 /* cr=0/1 is a very optimistic search to find large
2193 * good chunks almost for free. If buddy data is not
2194 * ready, then this optimization makes no sense. But
2195 * we never skip the first block group in a flex_bg,
2196 * since this gets used for metadata block allocation,
2197 * and we want to make sure we locate metadata blocks
2198 * in the first block group in the flex_bg if possible.
2201 (!sbi->s_log_groups_per_flex ||
2202 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2203 !(ext4_has_group_desc_csum(sb) &&
2204 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2206 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2212 ext4_lock_group(sb, group);
2213 ret = ext4_mb_good_group(ac, group, cr);
2216 ext4_unlock_group(sb, group);
2221 * Start prefetching @nr block bitmaps starting at @group.
2222 * Return the next group which needs to be prefetched.
2224 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2225 unsigned int nr, int *cnt)
2227 ext4_group_t ngroups = ext4_get_groups_count(sb);
2228 struct buffer_head *bh;
2229 struct blk_plug plug;
2231 blk_start_plug(&plug);
2233 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2235 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2238 * Prefetch block groups with free blocks; but don't
2239 * bother if it is marked uninitialized on disk, since
2240 * it won't require I/O to read. Also only try to
2241 * prefetch once, so we avoid getblk() call, which can
2244 if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2245 EXT4_MB_GRP_NEED_INIT(grp) &&
2246 ext4_free_group_clusters(sb, gdp) > 0 &&
2247 !(ext4_has_group_desc_csum(sb) &&
2248 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2249 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2250 if (bh && !IS_ERR(bh)) {
2251 if (!buffer_uptodate(bh) && cnt)
2256 if (++group >= ngroups)
2259 blk_finish_plug(&plug);
2264 * Prefetching reads the block bitmap into the buffer cache; but we
2265 * need to make sure that the buddy bitmap in the page cache has been
2266 * initialized. Note that ext4_mb_init_group() will block if the I/O
2267 * is not yet completed, or indeed if it was not initiated by
2268 * ext4_mb_prefetch did not start the I/O.
2270 * TODO: We should actually kick off the buddy bitmap setup in a work
2271 * queue when the buffer I/O is completed, so that we don't block
2272 * waiting for the block allocation bitmap read to finish when
2273 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2275 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2279 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2281 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2284 group = ext4_get_groups_count(sb);
2286 grp = ext4_get_group_info(sb, group);
2288 if (EXT4_MB_GRP_NEED_INIT(grp) &&
2289 ext4_free_group_clusters(sb, gdp) > 0 &&
2290 !(ext4_has_group_desc_csum(sb) &&
2291 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2292 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2298 static noinline_for_stack int
2299 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2301 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2303 int err = 0, first_err = 0;
2304 unsigned int nr = 0, prefetch_ios = 0;
2305 struct ext4_sb_info *sbi;
2306 struct super_block *sb;
2307 struct ext4_buddy e4b;
2312 ngroups = ext4_get_groups_count(sb);
2313 /* non-extent files are limited to low blocks/groups */
2314 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2315 ngroups = sbi->s_blockfile_groups;
2317 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2319 /* first, try the goal */
2320 err = ext4_mb_find_by_goal(ac, &e4b);
2321 if (err || ac->ac_status == AC_STATUS_FOUND)
2324 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2328 * ac->ac_2order is set only if the fe_len is a power of 2
2329 * if ac->ac_2order is set we also set criteria to 0 so that we
2330 * try exact allocation using buddy.
2332 i = fls(ac->ac_g_ex.fe_len);
2335 * We search using buddy data only if the order of the request
2336 * is greater than equal to the sbi_s_mb_order2_reqs
2337 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2338 * We also support searching for power-of-two requests only for
2339 * requests upto maximum buddy size we have constructed.
2341 if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) {
2343 * This should tell if fe_len is exactly power of 2
2345 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2346 ac->ac_2order = array_index_nospec(i - 1,
2347 sb->s_blocksize_bits + 2);
2350 /* if stream allocation is enabled, use global goal */
2351 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2352 /* TBD: may be hot point */
2353 spin_lock(&sbi->s_md_lock);
2354 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2355 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2356 spin_unlock(&sbi->s_md_lock);
2359 /* Let's just scan groups to find more-less suitable blocks */
2360 cr = ac->ac_2order ? 0 : 1;
2362 * cr == 0 try to get exact allocation,
2363 * cr == 3 try to get anything
2366 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2367 ac->ac_criteria = cr;
2369 * searching for the right group start
2370 * from the goal value specified
2372 group = ac->ac_g_ex.fe_group;
2373 prefetch_grp = group;
2375 for (i = 0; i < ngroups; group++, i++) {
2379 * Artificially restricted ngroups for non-extent
2380 * files makes group > ngroups possible on first loop.
2382 if (group >= ngroups)
2386 * Batch reads of the block allocation bitmaps
2387 * to get multiple READs in flight; limit
2388 * prefetching at cr=0/1, otherwise mballoc can
2389 * spend a lot of time loading imperfect groups
2391 if ((prefetch_grp == group) &&
2393 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2394 unsigned int curr_ios = prefetch_ios;
2396 nr = sbi->s_mb_prefetch;
2397 if (ext4_has_feature_flex_bg(sb)) {
2398 nr = (group / sbi->s_mb_prefetch) *
2400 nr = nr + sbi->s_mb_prefetch - group;
2402 prefetch_grp = ext4_mb_prefetch(sb, group,
2404 if (prefetch_ios == curr_ios)
2408 /* This now checks without needing the buddy page */
2409 ret = ext4_mb_good_group_nolock(ac, group, cr);
2416 err = ext4_mb_load_buddy(sb, group, &e4b);
2420 ext4_lock_group(sb, group);
2423 * We need to check again after locking the
2426 ret = ext4_mb_good_group(ac, group, cr);
2428 ext4_unlock_group(sb, group);
2429 ext4_mb_unload_buddy(&e4b);
2433 ac->ac_groups_scanned++;
2435 ext4_mb_simple_scan_group(ac, &e4b);
2436 else if (cr == 1 && sbi->s_stripe &&
2437 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2438 ext4_mb_scan_aligned(ac, &e4b);
2440 ext4_mb_complex_scan_group(ac, &e4b);
2442 ext4_unlock_group(sb, group);
2443 ext4_mb_unload_buddy(&e4b);
2445 if (ac->ac_status != AC_STATUS_CONTINUE)
2450 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2451 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2453 * We've been searching too long. Let's try to allocate
2454 * the best chunk we've found so far
2456 ext4_mb_try_best_found(ac, &e4b);
2457 if (ac->ac_status != AC_STATUS_FOUND) {
2459 * Someone more lucky has already allocated it.
2460 * The only thing we can do is just take first
2463 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2464 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2465 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2466 ac->ac_b_ex.fe_len, lost);
2468 ac->ac_b_ex.fe_group = 0;
2469 ac->ac_b_ex.fe_start = 0;
2470 ac->ac_b_ex.fe_len = 0;
2471 ac->ac_status = AC_STATUS_CONTINUE;
2472 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2478 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2481 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2482 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2483 ac->ac_flags, cr, err);
2486 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2491 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2493 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2496 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2499 return (void *) ((unsigned long) group);
2502 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2504 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2508 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2511 return (void *) ((unsigned long) group);
2514 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2516 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2517 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2519 int err, buddy_loaded = 0;
2520 struct ext4_buddy e4b;
2521 struct ext4_group_info *grinfo;
2522 unsigned char blocksize_bits = min_t(unsigned char,
2523 sb->s_blocksize_bits,
2524 EXT4_MAX_BLOCK_LOG_SIZE);
2526 struct ext4_group_info info;
2527 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2532 seq_puts(seq, "#group: free frags first ["
2533 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2534 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2536 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2537 sizeof(struct ext4_group_info);
2539 grinfo = ext4_get_group_info(sb, group);
2540 /* Load the group info in memory only if not already loaded. */
2541 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2542 err = ext4_mb_load_buddy(sb, group, &e4b);
2544 seq_printf(seq, "#%-5u: I/O error\n", group);
2550 memcpy(&sg, ext4_get_group_info(sb, group), i);
2553 ext4_mb_unload_buddy(&e4b);
2555 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2556 sg.info.bb_fragments, sg.info.bb_first_free);
2557 for (i = 0; i <= 13; i++)
2558 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2559 sg.info.bb_counters[i] : 0);
2560 seq_puts(seq, " ]\n");
2565 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2569 const struct seq_operations ext4_mb_seq_groups_ops = {
2570 .start = ext4_mb_seq_groups_start,
2571 .next = ext4_mb_seq_groups_next,
2572 .stop = ext4_mb_seq_groups_stop,
2573 .show = ext4_mb_seq_groups_show,
2576 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2578 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2579 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2586 * Allocate the top-level s_group_info array for the specified number
2589 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2591 struct ext4_sb_info *sbi = EXT4_SB(sb);
2593 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
2595 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2596 EXT4_DESC_PER_BLOCK_BITS(sb);
2597 if (size <= sbi->s_group_info_size)
2600 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2601 new_groupinfo = kvzalloc(size, GFP_KERNEL);
2602 if (!new_groupinfo) {
2603 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2607 old_groupinfo = rcu_dereference(sbi->s_group_info);
2609 memcpy(new_groupinfo, old_groupinfo,
2610 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2612 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
2613 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2615 ext4_kvfree_array_rcu(old_groupinfo);
2616 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2617 sbi->s_group_info_size);
2621 /* Create and initialize ext4_group_info data for the given group. */
2622 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2623 struct ext4_group_desc *desc)
2627 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
2628 struct ext4_sb_info *sbi = EXT4_SB(sb);
2629 struct ext4_group_info **meta_group_info;
2630 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2633 * First check if this group is the first of a reserved block.
2634 * If it's true, we have to allocate a new table of pointers
2635 * to ext4_group_info structures
2637 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2638 metalen = sizeof(*meta_group_info) <<
2639 EXT4_DESC_PER_BLOCK_BITS(sb);
2640 meta_group_info = kmalloc(metalen, GFP_NOFS);
2641 if (meta_group_info == NULL) {
2642 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2643 "for a buddy group");
2644 goto exit_meta_group_info;
2647 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
2651 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
2652 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2654 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2655 if (meta_group_info[i] == NULL) {
2656 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2657 goto exit_group_info;
2659 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2660 &(meta_group_info[i]->bb_state));
2663 * initialize bb_free to be able to skip
2664 * empty groups without initialization
2666 if (ext4_has_group_desc_csum(sb) &&
2667 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
2668 meta_group_info[i]->bb_free =
2669 ext4_free_clusters_after_init(sb, group, desc);
2671 meta_group_info[i]->bb_free =
2672 ext4_free_group_clusters(sb, desc);
2675 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2676 init_rwsem(&meta_group_info[i]->alloc_sem);
2677 meta_group_info[i]->bb_free_root = RB_ROOT;
2678 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2680 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
2684 /* If a meta_group_info table has been allocated, release it now */
2685 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2686 struct ext4_group_info ***group_info;
2689 group_info = rcu_dereference(sbi->s_group_info);
2690 kfree(group_info[idx]);
2691 group_info[idx] = NULL;
2694 exit_meta_group_info:
2696 } /* ext4_mb_add_groupinfo */
2698 static int ext4_mb_init_backend(struct super_block *sb)
2700 ext4_group_t ngroups = ext4_get_groups_count(sb);
2702 struct ext4_sb_info *sbi = EXT4_SB(sb);
2704 struct ext4_group_desc *desc;
2705 struct ext4_group_info ***group_info;
2706 struct kmem_cache *cachep;
2708 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2712 sbi->s_buddy_cache = new_inode(sb);
2713 if (sbi->s_buddy_cache == NULL) {
2714 ext4_msg(sb, KERN_ERR, "can't get new inode");
2717 /* To avoid potentially colliding with an valid on-disk inode number,
2718 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2719 * not in the inode hash, so it should never be found by iget(), but
2720 * this will avoid confusion if it ever shows up during debugging. */
2721 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2722 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2723 for (i = 0; i < ngroups; i++) {
2725 desc = ext4_get_group_desc(sb, i, NULL);
2727 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2730 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2734 if (ext4_has_feature_flex_bg(sb)) {
2735 /* a single flex group is supposed to be read by a single IO */
2736 sbi->s_mb_prefetch = 1 << sbi->s_es->s_log_groups_per_flex;
2737 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
2739 sbi->s_mb_prefetch = 32;
2741 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
2742 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
2743 /* now many real IOs to prefetch within a single allocation at cr=0
2744 * given cr=0 is an CPU-related optimization we shouldn't try to
2745 * load too many groups, at some point we should start to use what
2746 * we've got in memory.
2747 * with an average random access time 5ms, it'd take a second to get
2748 * 200 groups (* N with flex_bg), so let's make this limit 4
2750 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
2751 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
2752 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
2757 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2759 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2760 i = sbi->s_group_info_size;
2762 group_info = rcu_dereference(sbi->s_group_info);
2764 kfree(group_info[i]);
2766 iput(sbi->s_buddy_cache);
2769 kvfree(rcu_dereference(sbi->s_group_info));
2774 static void ext4_groupinfo_destroy_slabs(void)
2778 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2779 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2780 ext4_groupinfo_caches[i] = NULL;
2784 static int ext4_groupinfo_create_slab(size_t size)
2786 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2788 int blocksize_bits = order_base_2(size);
2789 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2790 struct kmem_cache *cachep;
2792 if (cache_index >= NR_GRPINFO_CACHES)
2795 if (unlikely(cache_index < 0))
2798 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2799 if (ext4_groupinfo_caches[cache_index]) {
2800 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2801 return 0; /* Already created */
2804 slab_size = offsetof(struct ext4_group_info,
2805 bb_counters[blocksize_bits + 2]);
2807 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2808 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2811 ext4_groupinfo_caches[cache_index] = cachep;
2813 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2816 "EXT4-fs: no memory for groupinfo slab cache\n");
2823 int ext4_mb_init(struct super_block *sb)
2825 struct ext4_sb_info *sbi = EXT4_SB(sb);
2827 unsigned offset, offset_incr;
2831 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2833 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2834 if (sbi->s_mb_offsets == NULL) {
2839 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2840 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2841 if (sbi->s_mb_maxs == NULL) {
2846 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2850 /* order 0 is regular bitmap */
2851 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2852 sbi->s_mb_offsets[0] = 0;
2856 offset_incr = 1 << (sb->s_blocksize_bits - 1);
2857 max = sb->s_blocksize << 2;
2859 sbi->s_mb_offsets[i] = offset;
2860 sbi->s_mb_maxs[i] = max;
2861 offset += offset_incr;
2862 offset_incr = offset_incr >> 1;
2865 } while (i <= sb->s_blocksize_bits + 1);
2867 spin_lock_init(&sbi->s_md_lock);
2868 spin_lock_init(&sbi->s_bal_lock);
2869 sbi->s_mb_free_pending = 0;
2870 INIT_LIST_HEAD(&sbi->s_freed_data_list);
2872 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2873 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2874 sbi->s_mb_stats = MB_DEFAULT_STATS;
2875 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2876 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2877 sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
2879 * The default group preallocation is 512, which for 4k block
2880 * sizes translates to 2 megabytes. However for bigalloc file
2881 * systems, this is probably too big (i.e, if the cluster size
2882 * is 1 megabyte, then group preallocation size becomes half a
2883 * gigabyte!). As a default, we will keep a two megabyte
2884 * group pralloc size for cluster sizes up to 64k, and after
2885 * that, we will force a minimum group preallocation size of
2886 * 32 clusters. This translates to 8 megs when the cluster
2887 * size is 256k, and 32 megs when the cluster size is 1 meg,
2888 * which seems reasonable as a default.
2890 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2891 sbi->s_cluster_bits, 32);
2893 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2894 * to the lowest multiple of s_stripe which is bigger than
2895 * the s_mb_group_prealloc as determined above. We want
2896 * the preallocation size to be an exact multiple of the
2897 * RAID stripe size so that preallocations don't fragment
2900 if (sbi->s_stripe > 1) {
2901 sbi->s_mb_group_prealloc = roundup(
2902 sbi->s_mb_group_prealloc, sbi->s_stripe);
2905 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2906 if (sbi->s_locality_groups == NULL) {
2910 for_each_possible_cpu(i) {
2911 struct ext4_locality_group *lg;
2912 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2913 mutex_init(&lg->lg_mutex);
2914 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2915 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2916 spin_lock_init(&lg->lg_prealloc_lock);
2919 /* init file for buddy data */
2920 ret = ext4_mb_init_backend(sb);
2922 goto out_free_locality_groups;
2926 out_free_locality_groups:
2927 free_percpu(sbi->s_locality_groups);
2928 sbi->s_locality_groups = NULL;
2930 kfree(sbi->s_mb_offsets);
2931 sbi->s_mb_offsets = NULL;
2932 kfree(sbi->s_mb_maxs);
2933 sbi->s_mb_maxs = NULL;
2937 /* need to called with the ext4 group lock held */
2938 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2940 struct ext4_prealloc_space *pa;
2941 struct list_head *cur, *tmp;
2944 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2945 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2946 list_del(&pa->pa_group_list);
2948 kmem_cache_free(ext4_pspace_cachep, pa);
2953 int ext4_mb_release(struct super_block *sb)
2955 ext4_group_t ngroups = ext4_get_groups_count(sb);
2957 int num_meta_group_infos;
2958 struct ext4_group_info *grinfo, ***group_info;
2959 struct ext4_sb_info *sbi = EXT4_SB(sb);
2960 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2963 if (sbi->s_group_info) {
2964 for (i = 0; i < ngroups; i++) {
2966 grinfo = ext4_get_group_info(sb, i);
2967 mb_group_bb_bitmap_free(grinfo);
2968 ext4_lock_group(sb, i);
2969 count = ext4_mb_cleanup_pa(grinfo);
2971 mb_debug(sb, "mballoc: %d PAs left\n",
2973 ext4_unlock_group(sb, i);
2974 kmem_cache_free(cachep, grinfo);
2976 num_meta_group_infos = (ngroups +
2977 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2978 EXT4_DESC_PER_BLOCK_BITS(sb);
2980 group_info = rcu_dereference(sbi->s_group_info);
2981 for (i = 0; i < num_meta_group_infos; i++)
2982 kfree(group_info[i]);
2986 kfree(sbi->s_mb_offsets);
2987 kfree(sbi->s_mb_maxs);
2988 iput(sbi->s_buddy_cache);
2989 if (sbi->s_mb_stats) {
2990 ext4_msg(sb, KERN_INFO,
2991 "mballoc: %u blocks %u reqs (%u success)",
2992 atomic_read(&sbi->s_bal_allocated),
2993 atomic_read(&sbi->s_bal_reqs),
2994 atomic_read(&sbi->s_bal_success));
2995 ext4_msg(sb, KERN_INFO,
2996 "mballoc: %u extents scanned, %u goal hits, "
2997 "%u 2^N hits, %u breaks, %u lost",
2998 atomic_read(&sbi->s_bal_ex_scanned),
2999 atomic_read(&sbi->s_bal_goals),
3000 atomic_read(&sbi->s_bal_2orders),
3001 atomic_read(&sbi->s_bal_breaks),
3002 atomic_read(&sbi->s_mb_lost_chunks));
3003 ext4_msg(sb, KERN_INFO,
3004 "mballoc: %lu generated and it took %Lu",
3005 sbi->s_mb_buddies_generated,
3006 sbi->s_mb_generation_time);
3007 ext4_msg(sb, KERN_INFO,
3008 "mballoc: %u preallocated, %u discarded",
3009 atomic_read(&sbi->s_mb_preallocated),
3010 atomic_read(&sbi->s_mb_discarded));
3013 free_percpu(sbi->s_locality_groups);
3018 static inline int ext4_issue_discard(struct super_block *sb,
3019 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3022 ext4_fsblk_t discard_block;
3024 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3025 ext4_group_first_block_no(sb, block_group));
3026 count = EXT4_C2B(EXT4_SB(sb), count);
3027 trace_ext4_discard_blocks(sb,
3028 (unsigned long long) discard_block, count);
3030 return __blkdev_issue_discard(sb->s_bdev,
3031 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3032 (sector_t)count << (sb->s_blocksize_bits - 9),
3035 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3038 static void ext4_free_data_in_buddy(struct super_block *sb,
3039 struct ext4_free_data *entry)
3041 struct ext4_buddy e4b;
3042 struct ext4_group_info *db;
3043 int err, count = 0, count2 = 0;
3045 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3046 entry->efd_count, entry->efd_group, entry);
3048 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3049 /* we expect to find existing buddy because it's pinned */
3052 spin_lock(&EXT4_SB(sb)->s_md_lock);
3053 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3054 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3057 /* there are blocks to put in buddy to make them really free */
3058 count += entry->efd_count;
3060 ext4_lock_group(sb, entry->efd_group);
3061 /* Take it out of per group rb tree */
3062 rb_erase(&entry->efd_node, &(db->bb_free_root));
3063 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3066 * Clear the trimmed flag for the group so that the next
3067 * ext4_trim_fs can trim it.
3068 * If the volume is mounted with -o discard, online discard
3069 * is supported and the free blocks will be trimmed online.
3071 if (!test_opt(sb, DISCARD))
3072 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3074 if (!db->bb_free_root.rb_node) {
3075 /* No more items in the per group rb tree
3076 * balance refcounts from ext4_mb_free_metadata()
3078 put_page(e4b.bd_buddy_page);
3079 put_page(e4b.bd_bitmap_page);
3081 ext4_unlock_group(sb, entry->efd_group);
3082 kmem_cache_free(ext4_free_data_cachep, entry);
3083 ext4_mb_unload_buddy(&e4b);
3085 mb_debug(sb, "freed %d blocks in %d structures\n", count,
3090 * This function is called by the jbd2 layer once the commit has finished,
3091 * so we know we can free the blocks that were released with that commit.
3093 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3095 struct ext4_sb_info *sbi = EXT4_SB(sb);
3096 struct ext4_free_data *entry, *tmp;
3097 struct bio *discard_bio = NULL;
3098 struct list_head freed_data_list;
3099 struct list_head *cut_pos = NULL;
3102 INIT_LIST_HEAD(&freed_data_list);
3104 spin_lock(&sbi->s_md_lock);
3105 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3106 if (entry->efd_tid != commit_tid)
3108 cut_pos = &entry->efd_list;
3111 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3113 spin_unlock(&sbi->s_md_lock);
3115 if (test_opt(sb, DISCARD)) {
3116 list_for_each_entry(entry, &freed_data_list, efd_list) {
3117 err = ext4_issue_discard(sb, entry->efd_group,
3118 entry->efd_start_cluster,
3121 if (err && err != -EOPNOTSUPP) {
3122 ext4_msg(sb, KERN_WARNING, "discard request in"
3123 " group:%d block:%d count:%d failed"
3124 " with %d", entry->efd_group,
3125 entry->efd_start_cluster,
3126 entry->efd_count, err);
3127 } else if (err == -EOPNOTSUPP)
3132 submit_bio_wait(discard_bio);
3133 bio_put(discard_bio);
3137 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3138 ext4_free_data_in_buddy(sb, entry);
3141 int __init ext4_init_mballoc(void)
3143 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3144 SLAB_RECLAIM_ACCOUNT);
3145 if (ext4_pspace_cachep == NULL)
3148 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3149 SLAB_RECLAIM_ACCOUNT);
3150 if (ext4_ac_cachep == NULL)
3153 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3154 SLAB_RECLAIM_ACCOUNT);
3155 if (ext4_free_data_cachep == NULL)
3161 kmem_cache_destroy(ext4_ac_cachep);
3163 kmem_cache_destroy(ext4_pspace_cachep);
3168 void ext4_exit_mballoc(void)
3171 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3172 * before destroying the slab cache.
3175 kmem_cache_destroy(ext4_pspace_cachep);
3176 kmem_cache_destroy(ext4_ac_cachep);
3177 kmem_cache_destroy(ext4_free_data_cachep);
3178 ext4_groupinfo_destroy_slabs();
3183 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3184 * Returns 0 if success or error code
3186 static noinline_for_stack int
3187 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3188 handle_t *handle, unsigned int reserv_clstrs)
3190 struct buffer_head *bitmap_bh = NULL;
3191 struct ext4_group_desc *gdp;
3192 struct buffer_head *gdp_bh;
3193 struct ext4_sb_info *sbi;
3194 struct super_block *sb;
3198 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3199 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3204 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3205 if (IS_ERR(bitmap_bh)) {
3206 err = PTR_ERR(bitmap_bh);
3211 BUFFER_TRACE(bitmap_bh, "getting write access");
3212 err = ext4_journal_get_write_access(handle, bitmap_bh);
3217 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3221 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3222 ext4_free_group_clusters(sb, gdp));
3224 BUFFER_TRACE(gdp_bh, "get_write_access");
3225 err = ext4_journal_get_write_access(handle, gdp_bh);
3229 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3231 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3232 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3233 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3234 "fs metadata", block, block+len);
3235 /* File system mounted not to panic on error
3236 * Fix the bitmap and return EFSCORRUPTED
3237 * We leak some of the blocks here.
3239 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3240 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3241 ac->ac_b_ex.fe_len);
3242 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3243 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3245 err = -EFSCORRUPTED;
3249 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3250 #ifdef AGGRESSIVE_CHECK
3253 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3254 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3255 bitmap_bh->b_data));
3259 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3260 ac->ac_b_ex.fe_len);
3261 if (ext4_has_group_desc_csum(sb) &&
3262 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3263 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3264 ext4_free_group_clusters_set(sb, gdp,
3265 ext4_free_clusters_after_init(sb,
3266 ac->ac_b_ex.fe_group, gdp));
3268 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3269 ext4_free_group_clusters_set(sb, gdp, len);
3270 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3271 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3273 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3274 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3276 * Now reduce the dirty block count also. Should not go negative
3278 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3279 /* release all the reserved blocks if non delalloc */
3280 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3283 if (sbi->s_log_groups_per_flex) {
3284 ext4_group_t flex_group = ext4_flex_group(sbi,
3285 ac->ac_b_ex.fe_group);
3286 atomic64_sub(ac->ac_b_ex.fe_len,
3287 &sbi_array_rcu_deref(sbi, s_flex_groups,
3288 flex_group)->free_clusters);
3291 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3294 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3302 * Idempotent helper for Ext4 fast commit replay path to set the state of
3303 * blocks in bitmaps and update counters.
3305 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
3308 struct buffer_head *bitmap_bh = NULL;
3309 struct ext4_group_desc *gdp;
3310 struct buffer_head *gdp_bh;
3311 struct ext4_sb_info *sbi = EXT4_SB(sb);
3313 ext4_grpblk_t blkoff;
3317 clen = EXT4_B2C(sbi, len);
3319 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
3320 bitmap_bh = ext4_read_block_bitmap(sb, group);
3321 if (IS_ERR(bitmap_bh)) {
3322 err = PTR_ERR(bitmap_bh);
3328 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
3332 ext4_lock_group(sb, group);
3334 for (i = 0; i < clen; i++)
3335 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) == !state)
3339 ext4_set_bits(bitmap_bh->b_data, blkoff, clen);
3341 mb_test_and_clear_bits(bitmap_bh->b_data, blkoff, clen);
3342 if (ext4_has_group_desc_csum(sb) &&
3343 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3344 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3345 ext4_free_group_clusters_set(sb, gdp,
3346 ext4_free_clusters_after_init(sb,
3350 clen = ext4_free_group_clusters(sb, gdp) - clen + already;
3352 clen = ext4_free_group_clusters(sb, gdp) + clen - already;
3354 ext4_free_group_clusters_set(sb, gdp, clen);
3355 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
3356 ext4_group_desc_csum_set(sb, group, gdp);
3358 ext4_unlock_group(sb, group);
3360 if (sbi->s_log_groups_per_flex) {
3361 ext4_group_t flex_group = ext4_flex_group(sbi, group);
3364 &sbi_array_rcu_deref(sbi, s_flex_groups,
3365 flex_group)->free_clusters);
3368 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
3371 sync_dirty_buffer(bitmap_bh);
3372 err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
3373 sync_dirty_buffer(gdp_bh);
3380 * here we normalize request for locality group
3381 * Group request are normalized to s_mb_group_prealloc, which goes to
3382 * s_strip if we set the same via mount option.
3383 * s_mb_group_prealloc can be configured via
3384 * /sys/fs/ext4/<partition>/mb_group_prealloc
3386 * XXX: should we try to preallocate more than the group has now?
3388 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3390 struct super_block *sb = ac->ac_sb;
3391 struct ext4_locality_group *lg = ac->ac_lg;
3394 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3395 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
3399 * Normalization means making request better in terms of
3400 * size and alignment
3402 static noinline_for_stack void
3403 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3404 struct ext4_allocation_request *ar)
3406 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3409 loff_t size, start_off;
3410 loff_t orig_size __maybe_unused;
3412 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3413 struct ext4_prealloc_space *pa;
3415 /* do normalize only data requests, metadata requests
3416 do not need preallocation */
3417 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3420 /* sometime caller may want exact blocks */
3421 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3424 /* caller may indicate that preallocation isn't
3425 * required (it's a tail, for example) */
3426 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3429 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3430 ext4_mb_normalize_group_request(ac);
3434 bsbits = ac->ac_sb->s_blocksize_bits;
3436 /* first, let's learn actual file size
3437 * given current request is allocated */
3438 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3439 size = size << bsbits;
3440 if (size < i_size_read(ac->ac_inode))
3441 size = i_size_read(ac->ac_inode);
3444 /* max size of free chunks */
3447 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3448 (req <= (size) || max <= (chunk_size))
3450 /* first, try to predict filesize */
3451 /* XXX: should this table be tunable? */
3453 if (size <= 16 * 1024) {
3455 } else if (size <= 32 * 1024) {
3457 } else if (size <= 64 * 1024) {
3459 } else if (size <= 128 * 1024) {
3461 } else if (size <= 256 * 1024) {
3463 } else if (size <= 512 * 1024) {
3465 } else if (size <= 1024 * 1024) {
3467 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3468 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3469 (21 - bsbits)) << 21;
3470 size = 2 * 1024 * 1024;
3471 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3472 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3473 (22 - bsbits)) << 22;
3474 size = 4 * 1024 * 1024;
3475 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3476 (8<<20)>>bsbits, max, 8 * 1024)) {
3477 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3478 (23 - bsbits)) << 23;
3479 size = 8 * 1024 * 1024;
3481 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3482 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3483 ac->ac_o_ex.fe_len) << bsbits;
3485 size = size >> bsbits;
3486 start = start_off >> bsbits;
3488 /* don't cover already allocated blocks in selected range */
3489 if (ar->pleft && start <= ar->lleft) {
3490 size -= ar->lleft + 1 - start;
3491 start = ar->lleft + 1;
3493 if (ar->pright && start + size - 1 >= ar->lright)
3494 size -= start + size - ar->lright;
3497 * Trim allocation request for filesystems with artificially small
3500 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
3501 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
3505 /* check we don't cross already preallocated blocks */
3507 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3512 spin_lock(&pa->pa_lock);
3513 if (pa->pa_deleted) {
3514 spin_unlock(&pa->pa_lock);
3518 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3521 /* PA must not overlap original request */
3522 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3523 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3525 /* skip PAs this normalized request doesn't overlap with */
3526 if (pa->pa_lstart >= end || pa_end <= start) {
3527 spin_unlock(&pa->pa_lock);
3530 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3532 /* adjust start or end to be adjacent to this pa */
3533 if (pa_end <= ac->ac_o_ex.fe_logical) {
3534 BUG_ON(pa_end < start);
3536 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3537 BUG_ON(pa->pa_lstart > end);
3538 end = pa->pa_lstart;
3540 spin_unlock(&pa->pa_lock);
3545 /* XXX: extra loop to check we really don't overlap preallocations */
3547 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3550 spin_lock(&pa->pa_lock);
3551 if (pa->pa_deleted == 0) {
3552 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3554 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3556 spin_unlock(&pa->pa_lock);
3560 if (start + size <= ac->ac_o_ex.fe_logical &&
3561 start > ac->ac_o_ex.fe_logical) {
3562 ext4_msg(ac->ac_sb, KERN_ERR,
3563 "start %lu, size %lu, fe_logical %lu",
3564 (unsigned long) start, (unsigned long) size,
3565 (unsigned long) ac->ac_o_ex.fe_logical);
3568 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3570 /* now prepare goal request */
3572 /* XXX: is it better to align blocks WRT to logical
3573 * placement or satisfy big request as is */
3574 ac->ac_g_ex.fe_logical = start;
3575 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3577 /* define goal start in order to merge */
3578 if (ar->pright && (ar->lright == (start + size))) {
3579 /* merge to the right */
3580 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3581 &ac->ac_f_ex.fe_group,
3582 &ac->ac_f_ex.fe_start);
3583 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3585 if (ar->pleft && (ar->lleft + 1 == start)) {
3586 /* merge to the left */
3587 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3588 &ac->ac_f_ex.fe_group,
3589 &ac->ac_f_ex.fe_start);
3590 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3593 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
3597 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3599 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3601 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3602 atomic_inc(&sbi->s_bal_reqs);
3603 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3604 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3605 atomic_inc(&sbi->s_bal_success);
3606 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3607 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3608 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3609 atomic_inc(&sbi->s_bal_goals);
3610 if (ac->ac_found > sbi->s_mb_max_to_scan)
3611 atomic_inc(&sbi->s_bal_breaks);
3614 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3615 trace_ext4_mballoc_alloc(ac);
3617 trace_ext4_mballoc_prealloc(ac);
3621 * Called on failure; free up any blocks from the inode PA for this
3622 * context. We don't need this for MB_GROUP_PA because we only change
3623 * pa_free in ext4_mb_release_context(), but on failure, we've already
3624 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3626 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3628 struct ext4_prealloc_space *pa = ac->ac_pa;
3629 struct ext4_buddy e4b;
3633 if (ac->ac_f_ex.fe_len == 0)
3635 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3638 * This should never happen since we pin the
3639 * pages in the ext4_allocation_context so
3640 * ext4_mb_load_buddy() should never fail.
3642 WARN(1, "mb_load_buddy failed (%d)", err);
3645 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3646 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3647 ac->ac_f_ex.fe_len);
3648 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3649 ext4_mb_unload_buddy(&e4b);
3652 if (pa->pa_type == MB_INODE_PA)
3653 pa->pa_free += ac->ac_b_ex.fe_len;
3657 * use blocks preallocated to inode
3659 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3660 struct ext4_prealloc_space *pa)
3662 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3667 /* found preallocated blocks, use them */
3668 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3669 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3670 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3671 len = EXT4_NUM_B2C(sbi, end - start);
3672 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3673 &ac->ac_b_ex.fe_start);
3674 ac->ac_b_ex.fe_len = len;
3675 ac->ac_status = AC_STATUS_FOUND;
3678 BUG_ON(start < pa->pa_pstart);
3679 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3680 BUG_ON(pa->pa_free < len);
3683 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
3687 * use blocks preallocated to locality group
3689 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3690 struct ext4_prealloc_space *pa)
3692 unsigned int len = ac->ac_o_ex.fe_len;
3694 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3695 &ac->ac_b_ex.fe_group,
3696 &ac->ac_b_ex.fe_start);
3697 ac->ac_b_ex.fe_len = len;
3698 ac->ac_status = AC_STATUS_FOUND;
3701 /* we don't correct pa_pstart or pa_plen here to avoid
3702 * possible race when the group is being loaded concurrently
3703 * instead we correct pa later, after blocks are marked
3704 * in on-disk bitmap -- see ext4_mb_release_context()
3705 * Other CPUs are prevented from allocating from this pa by lg_mutex
3707 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
3708 pa->pa_lstart-len, len, pa);
3712 * Return the prealloc space that have minimal distance
3713 * from the goal block. @cpa is the prealloc
3714 * space that is having currently known minimal distance
3715 * from the goal block.
3717 static struct ext4_prealloc_space *
3718 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3719 struct ext4_prealloc_space *pa,
3720 struct ext4_prealloc_space *cpa)
3722 ext4_fsblk_t cur_distance, new_distance;
3725 atomic_inc(&pa->pa_count);
3728 cur_distance = abs(goal_block - cpa->pa_pstart);
3729 new_distance = abs(goal_block - pa->pa_pstart);
3731 if (cur_distance <= new_distance)
3734 /* drop the previous reference */
3735 atomic_dec(&cpa->pa_count);
3736 atomic_inc(&pa->pa_count);
3741 * search goal blocks in preallocated space
3743 static noinline_for_stack bool
3744 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3746 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3748 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3749 struct ext4_locality_group *lg;
3750 struct ext4_prealloc_space *pa, *cpa = NULL;
3751 ext4_fsblk_t goal_block;
3753 /* only data can be preallocated */
3754 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3757 /* first, try per-file preallocation */
3759 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3761 /* all fields in this condition don't change,
3762 * so we can skip locking for them */
3763 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3764 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3765 EXT4_C2B(sbi, pa->pa_len)))
3768 /* non-extent files can't have physical blocks past 2^32 */
3769 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3770 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3771 EXT4_MAX_BLOCK_FILE_PHYS))
3774 /* found preallocated blocks, use them */
3775 spin_lock(&pa->pa_lock);
3776 if (pa->pa_deleted == 0 && pa->pa_free) {
3777 atomic_inc(&pa->pa_count);
3778 ext4_mb_use_inode_pa(ac, pa);
3779 spin_unlock(&pa->pa_lock);
3780 ac->ac_criteria = 10;
3784 spin_unlock(&pa->pa_lock);
3788 /* can we use group allocation? */
3789 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3792 /* inode may have no locality group for some reason */
3796 order = fls(ac->ac_o_ex.fe_len) - 1;
3797 if (order > PREALLOC_TB_SIZE - 1)
3798 /* The max size of hash table is PREALLOC_TB_SIZE */
3799 order = PREALLOC_TB_SIZE - 1;
3801 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3803 * search for the prealloc space that is having
3804 * minimal distance from the goal block.
3806 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3808 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3810 spin_lock(&pa->pa_lock);
3811 if (pa->pa_deleted == 0 &&
3812 pa->pa_free >= ac->ac_o_ex.fe_len) {
3814 cpa = ext4_mb_check_group_pa(goal_block,
3817 spin_unlock(&pa->pa_lock);
3822 ext4_mb_use_group_pa(ac, cpa);
3823 ac->ac_criteria = 20;
3830 * the function goes through all block freed in the group
3831 * but not yet committed and marks them used in in-core bitmap.
3832 * buddy must be generated from this bitmap
3833 * Need to be called with the ext4 group lock held
3835 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3839 struct ext4_group_info *grp;
3840 struct ext4_free_data *entry;
3842 grp = ext4_get_group_info(sb, group);
3843 n = rb_first(&(grp->bb_free_root));
3846 entry = rb_entry(n, struct ext4_free_data, efd_node);
3847 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3854 * the function goes through all preallocation in this group and marks them
3855 * used in in-core bitmap. buddy must be generated from this bitmap
3856 * Need to be called with ext4 group lock held
3858 static noinline_for_stack
3859 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3862 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3863 struct ext4_prealloc_space *pa;
3864 struct list_head *cur;
3865 ext4_group_t groupnr;
3866 ext4_grpblk_t start;
3867 int preallocated = 0;
3870 /* all form of preallocation discards first load group,
3871 * so the only competing code is preallocation use.
3872 * we don't need any locking here
3873 * notice we do NOT ignore preallocations with pa_deleted
3874 * otherwise we could leave used blocks available for
3875 * allocation in buddy when concurrent ext4_mb_put_pa()
3876 * is dropping preallocation
3878 list_for_each(cur, &grp->bb_prealloc_list) {
3879 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3880 spin_lock(&pa->pa_lock);
3881 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3884 spin_unlock(&pa->pa_lock);
3885 if (unlikely(len == 0))
3887 BUG_ON(groupnr != group);
3888 ext4_set_bits(bitmap, start, len);
3889 preallocated += len;
3891 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
3894 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
3895 struct ext4_prealloc_space *pa)
3897 struct ext4_inode_info *ei;
3899 if (pa->pa_deleted) {
3900 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
3901 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
3908 if (pa->pa_type == MB_INODE_PA) {
3909 ei = EXT4_I(pa->pa_inode);
3910 atomic_dec(&ei->i_prealloc_active);
3914 static void ext4_mb_pa_callback(struct rcu_head *head)
3916 struct ext4_prealloc_space *pa;
3917 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3919 BUG_ON(atomic_read(&pa->pa_count));
3920 BUG_ON(pa->pa_deleted == 0);
3921 kmem_cache_free(ext4_pspace_cachep, pa);
3925 * drops a reference to preallocated space descriptor
3926 * if this was the last reference and the space is consumed
3928 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3929 struct super_block *sb, struct ext4_prealloc_space *pa)
3932 ext4_fsblk_t grp_blk;
3934 /* in this short window concurrent discard can set pa_deleted */
3935 spin_lock(&pa->pa_lock);
3936 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3937 spin_unlock(&pa->pa_lock);
3941 if (pa->pa_deleted == 1) {
3942 spin_unlock(&pa->pa_lock);
3946 ext4_mb_mark_pa_deleted(sb, pa);
3947 spin_unlock(&pa->pa_lock);
3949 grp_blk = pa->pa_pstart;
3951 * If doing group-based preallocation, pa_pstart may be in the
3952 * next group when pa is used up
3954 if (pa->pa_type == MB_GROUP_PA)
3957 grp = ext4_get_group_number(sb, grp_blk);
3962 * P1 (buddy init) P2 (regular allocation)
3963 * find block B in PA
3964 * copy on-disk bitmap to buddy
3965 * mark B in on-disk bitmap
3966 * drop PA from group
3967 * mark all PAs in buddy
3969 * thus, P1 initializes buddy with B available. to prevent this
3970 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3973 ext4_lock_group(sb, grp);
3974 list_del(&pa->pa_group_list);
3975 ext4_unlock_group(sb, grp);
3977 spin_lock(pa->pa_obj_lock);
3978 list_del_rcu(&pa->pa_inode_list);
3979 spin_unlock(pa->pa_obj_lock);
3981 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3985 * creates new preallocated space for given inode
3987 static noinline_for_stack void
3988 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3990 struct super_block *sb = ac->ac_sb;
3991 struct ext4_sb_info *sbi = EXT4_SB(sb);
3992 struct ext4_prealloc_space *pa;
3993 struct ext4_group_info *grp;
3994 struct ext4_inode_info *ei;
3996 /* preallocate only when found space is larger then requested */
3997 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3998 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3999 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4000 BUG_ON(ac->ac_pa == NULL);
4004 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
4010 /* we can't allocate as much as normalizer wants.
4011 * so, found space must get proper lstart
4012 * to cover original request */
4013 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
4014 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
4016 /* we're limited by original request in that
4017 * logical block must be covered any way
4018 * winl is window we can move our chunk within */
4019 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
4021 /* also, we should cover whole original request */
4022 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
4024 /* the smallest one defines real window */
4025 win = min(winl, wins);
4027 offs = ac->ac_o_ex.fe_logical %
4028 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4029 if (offs && offs < win)
4032 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
4033 EXT4_NUM_B2C(sbi, win);
4034 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
4035 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
4038 /* preallocation can change ac_b_ex, thus we store actually
4039 * allocated blocks for history */
4040 ac->ac_f_ex = ac->ac_b_ex;
4042 pa->pa_lstart = ac->ac_b_ex.fe_logical;
4043 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4044 pa->pa_len = ac->ac_b_ex.fe_len;
4045 pa->pa_free = pa->pa_len;
4046 spin_lock_init(&pa->pa_lock);
4047 INIT_LIST_HEAD(&pa->pa_inode_list);
4048 INIT_LIST_HEAD(&pa->pa_group_list);
4050 pa->pa_type = MB_INODE_PA;
4052 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4053 pa->pa_len, pa->pa_lstart);
4054 trace_ext4_mb_new_inode_pa(ac, pa);
4056 ext4_mb_use_inode_pa(ac, pa);
4057 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
4059 ei = EXT4_I(ac->ac_inode);
4060 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4062 pa->pa_obj_lock = &ei->i_prealloc_lock;
4063 pa->pa_inode = ac->ac_inode;
4065 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4067 spin_lock(pa->pa_obj_lock);
4068 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
4069 spin_unlock(pa->pa_obj_lock);
4070 atomic_inc(&ei->i_prealloc_active);
4074 * creates new preallocated space for locality group inodes belongs to
4076 static noinline_for_stack void
4077 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
4079 struct super_block *sb = ac->ac_sb;
4080 struct ext4_locality_group *lg;
4081 struct ext4_prealloc_space *pa;
4082 struct ext4_group_info *grp;
4084 /* preallocate only when found space is larger then requested */
4085 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4086 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4087 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4088 BUG_ON(ac->ac_pa == NULL);
4092 /* preallocation can change ac_b_ex, thus we store actually
4093 * allocated blocks for history */
4094 ac->ac_f_ex = ac->ac_b_ex;
4096 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4097 pa->pa_lstart = pa->pa_pstart;
4098 pa->pa_len = ac->ac_b_ex.fe_len;
4099 pa->pa_free = pa->pa_len;
4100 spin_lock_init(&pa->pa_lock);
4101 INIT_LIST_HEAD(&pa->pa_inode_list);
4102 INIT_LIST_HEAD(&pa->pa_group_list);
4104 pa->pa_type = MB_GROUP_PA;
4106 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4107 pa->pa_len, pa->pa_lstart);
4108 trace_ext4_mb_new_group_pa(ac, pa);
4110 ext4_mb_use_group_pa(ac, pa);
4111 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4113 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4117 pa->pa_obj_lock = &lg->lg_prealloc_lock;
4118 pa->pa_inode = NULL;
4120 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4123 * We will later add the new pa to the right bucket
4124 * after updating the pa_free in ext4_mb_release_context
4128 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4130 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4131 ext4_mb_new_group_pa(ac);
4133 ext4_mb_new_inode_pa(ac);
4137 * finds all unused blocks in on-disk bitmap, frees them in
4138 * in-core bitmap and buddy.
4139 * @pa must be unlinked from inode and group lists, so that
4140 * nobody else can find/use it.
4141 * the caller MUST hold group/inode locks.
4142 * TODO: optimize the case when there are no in-core structures yet
4144 static noinline_for_stack int
4145 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4146 struct ext4_prealloc_space *pa)
4148 struct super_block *sb = e4b->bd_sb;
4149 struct ext4_sb_info *sbi = EXT4_SB(sb);
4154 unsigned long long grp_blk_start;
4157 BUG_ON(pa->pa_deleted == 0);
4158 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4159 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4160 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4161 end = bit + pa->pa_len;
4164 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4167 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4168 mb_debug(sb, "free preallocated %u/%u in group %u\n",
4169 (unsigned) ext4_group_first_block_no(sb, group) + bit,
4170 (unsigned) next - bit, (unsigned) group);
4173 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4174 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4175 EXT4_C2B(sbi, bit)),
4177 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4180 if (free != pa->pa_free) {
4181 ext4_msg(e4b->bd_sb, KERN_CRIT,
4182 "pa %p: logic %lu, phys. %lu, len %d",
4183 pa, (unsigned long) pa->pa_lstart,
4184 (unsigned long) pa->pa_pstart,
4186 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4189 * pa is already deleted so we use the value obtained
4190 * from the bitmap and continue.
4193 atomic_add(free, &sbi->s_mb_discarded);
4198 static noinline_for_stack int
4199 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4200 struct ext4_prealloc_space *pa)
4202 struct super_block *sb = e4b->bd_sb;
4206 trace_ext4_mb_release_group_pa(sb, pa);
4207 BUG_ON(pa->pa_deleted == 0);
4208 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4209 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4210 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4211 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4212 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4218 * releases all preallocations in given group
4220 * first, we need to decide discard policy:
4221 * - when do we discard
4223 * - how many do we discard
4224 * 1) how many requested
4226 static noinline_for_stack int
4227 ext4_mb_discard_group_preallocations(struct super_block *sb,
4228 ext4_group_t group, int needed)
4230 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4231 struct buffer_head *bitmap_bh = NULL;
4232 struct ext4_prealloc_space *pa, *tmp;
4233 struct list_head list;
4234 struct ext4_buddy e4b;
4237 int free, free_total = 0;
4239 mb_debug(sb, "discard preallocation for group %u\n", group);
4240 if (list_empty(&grp->bb_prealloc_list))
4243 bitmap_bh = ext4_read_block_bitmap(sb, group);
4244 if (IS_ERR(bitmap_bh)) {
4245 err = PTR_ERR(bitmap_bh);
4246 ext4_error_err(sb, -err,
4247 "Error %d reading block bitmap for %u",
4252 err = ext4_mb_load_buddy(sb, group, &e4b);
4254 ext4_warning(sb, "Error %d loading buddy information for %u",
4261 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
4263 INIT_LIST_HEAD(&list);
4266 ext4_lock_group(sb, group);
4267 list_for_each_entry_safe(pa, tmp,
4268 &grp->bb_prealloc_list, pa_group_list) {
4269 spin_lock(&pa->pa_lock);
4270 if (atomic_read(&pa->pa_count)) {
4271 spin_unlock(&pa->pa_lock);
4275 if (pa->pa_deleted) {
4276 spin_unlock(&pa->pa_lock);
4280 /* seems this one can be freed ... */
4281 ext4_mb_mark_pa_deleted(sb, pa);
4284 this_cpu_inc(discard_pa_seq);
4286 /* we can trust pa_free ... */
4287 free += pa->pa_free;
4289 spin_unlock(&pa->pa_lock);
4291 list_del(&pa->pa_group_list);
4292 list_add(&pa->u.pa_tmp_list, &list);
4295 /* now free all selected PAs */
4296 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4298 /* remove from object (inode or locality group) */
4299 spin_lock(pa->pa_obj_lock);
4300 list_del_rcu(&pa->pa_inode_list);
4301 spin_unlock(pa->pa_obj_lock);
4303 if (pa->pa_type == MB_GROUP_PA)
4304 ext4_mb_release_group_pa(&e4b, pa);
4306 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4308 list_del(&pa->u.pa_tmp_list);
4309 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4314 /* if we still need more blocks and some PAs were used, try again */
4315 if (free_total < needed && busy) {
4316 ext4_unlock_group(sb, group);
4321 ext4_unlock_group(sb, group);
4322 ext4_mb_unload_buddy(&e4b);
4325 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4326 free_total, group, grp->bb_free);
4331 * releases all non-used preallocated blocks for given inode
4333 * It's important to discard preallocations under i_data_sem
4334 * We don't want another block to be served from the prealloc
4335 * space when we are discarding the inode prealloc space.
4337 * FIXME!! Make sure it is valid at all the call sites
4339 void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
4341 struct ext4_inode_info *ei = EXT4_I(inode);
4342 struct super_block *sb = inode->i_sb;
4343 struct buffer_head *bitmap_bh = NULL;
4344 struct ext4_prealloc_space *pa, *tmp;
4345 ext4_group_t group = 0;
4346 struct list_head list;
4347 struct ext4_buddy e4b;
4350 if (!S_ISREG(inode->i_mode)) {
4351 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4355 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
4358 mb_debug(sb, "discard preallocation for inode %lu\n",
4360 trace_ext4_discard_preallocations(inode,
4361 atomic_read(&ei->i_prealloc_active), needed);
4363 INIT_LIST_HEAD(&list);
4369 /* first, collect all pa's in the inode */
4370 spin_lock(&ei->i_prealloc_lock);
4371 while (!list_empty(&ei->i_prealloc_list) && needed) {
4372 pa = list_entry(ei->i_prealloc_list.prev,
4373 struct ext4_prealloc_space, pa_inode_list);
4374 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4375 spin_lock(&pa->pa_lock);
4376 if (atomic_read(&pa->pa_count)) {
4377 /* this shouldn't happen often - nobody should
4378 * use preallocation while we're discarding it */
4379 spin_unlock(&pa->pa_lock);
4380 spin_unlock(&ei->i_prealloc_lock);
4381 ext4_msg(sb, KERN_ERR,
4382 "uh-oh! used pa while discarding");
4384 schedule_timeout_uninterruptible(HZ);
4388 if (pa->pa_deleted == 0) {
4389 ext4_mb_mark_pa_deleted(sb, pa);
4390 spin_unlock(&pa->pa_lock);
4391 list_del_rcu(&pa->pa_inode_list);
4392 list_add(&pa->u.pa_tmp_list, &list);
4397 /* someone is deleting pa right now */
4398 spin_unlock(&pa->pa_lock);
4399 spin_unlock(&ei->i_prealloc_lock);
4401 /* we have to wait here because pa_deleted
4402 * doesn't mean pa is already unlinked from
4403 * the list. as we might be called from
4404 * ->clear_inode() the inode will get freed
4405 * and concurrent thread which is unlinking
4406 * pa from inode's list may access already
4407 * freed memory, bad-bad-bad */
4409 /* XXX: if this happens too often, we can
4410 * add a flag to force wait only in case
4411 * of ->clear_inode(), but not in case of
4412 * regular truncate */
4413 schedule_timeout_uninterruptible(HZ);
4416 spin_unlock(&ei->i_prealloc_lock);
4418 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4419 BUG_ON(pa->pa_type != MB_INODE_PA);
4420 group = ext4_get_group_number(sb, pa->pa_pstart);
4422 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4423 GFP_NOFS|__GFP_NOFAIL);
4425 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
4430 bitmap_bh = ext4_read_block_bitmap(sb, group);
4431 if (IS_ERR(bitmap_bh)) {
4432 err = PTR_ERR(bitmap_bh);
4433 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
4435 ext4_mb_unload_buddy(&e4b);
4439 ext4_lock_group(sb, group);
4440 list_del(&pa->pa_group_list);
4441 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4442 ext4_unlock_group(sb, group);
4444 ext4_mb_unload_buddy(&e4b);
4447 list_del(&pa->u.pa_tmp_list);
4448 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4452 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
4454 struct ext4_prealloc_space *pa;
4456 BUG_ON(ext4_pspace_cachep == NULL);
4457 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
4460 atomic_set(&pa->pa_count, 1);
4465 static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
4467 struct ext4_prealloc_space *pa = ac->ac_pa;
4471 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
4472 kmem_cache_free(ext4_pspace_cachep, pa);
4475 #ifdef CONFIG_EXT4_DEBUG
4476 static inline void ext4_mb_show_pa(struct super_block *sb)
4478 ext4_group_t i, ngroups;
4480 if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
4483 ngroups = ext4_get_groups_count(sb);
4484 mb_debug(sb, "groups: ");
4485 for (i = 0; i < ngroups; i++) {
4486 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4487 struct ext4_prealloc_space *pa;
4488 ext4_grpblk_t start;
4489 struct list_head *cur;
4490 ext4_lock_group(sb, i);
4491 list_for_each(cur, &grp->bb_prealloc_list) {
4492 pa = list_entry(cur, struct ext4_prealloc_space,
4494 spin_lock(&pa->pa_lock);
4495 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4497 spin_unlock(&pa->pa_lock);
4498 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
4501 ext4_unlock_group(sb, i);
4502 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
4507 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4509 struct super_block *sb = ac->ac_sb;
4511 if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
4514 mb_debug(sb, "Can't allocate:"
4515 " Allocation context details:");
4516 mb_debug(sb, "status %u flags 0x%x",
4517 ac->ac_status, ac->ac_flags);
4518 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
4519 "goal %lu/%lu/%lu@%lu, "
4520 "best %lu/%lu/%lu@%lu cr %d",
4521 (unsigned long)ac->ac_o_ex.fe_group,
4522 (unsigned long)ac->ac_o_ex.fe_start,
4523 (unsigned long)ac->ac_o_ex.fe_len,
4524 (unsigned long)ac->ac_o_ex.fe_logical,
4525 (unsigned long)ac->ac_g_ex.fe_group,
4526 (unsigned long)ac->ac_g_ex.fe_start,
4527 (unsigned long)ac->ac_g_ex.fe_len,
4528 (unsigned long)ac->ac_g_ex.fe_logical,
4529 (unsigned long)ac->ac_b_ex.fe_group,
4530 (unsigned long)ac->ac_b_ex.fe_start,
4531 (unsigned long)ac->ac_b_ex.fe_len,
4532 (unsigned long)ac->ac_b_ex.fe_logical,
4533 (int)ac->ac_criteria);
4534 mb_debug(sb, "%u found", ac->ac_found);
4535 ext4_mb_show_pa(sb);
4538 static inline void ext4_mb_show_pa(struct super_block *sb)
4542 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4544 ext4_mb_show_pa(ac->ac_sb);
4550 * We use locality group preallocation for small size file. The size of the
4551 * file is determined by the current size or the resulting size after
4552 * allocation which ever is larger
4554 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4556 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4558 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4559 int bsbits = ac->ac_sb->s_blocksize_bits;
4562 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4565 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4568 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4569 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4572 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
4573 !inode_is_open_for_write(ac->ac_inode)) {
4574 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4578 if (sbi->s_mb_group_prealloc <= 0) {
4579 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4583 /* don't use group allocation for large files */
4584 size = max(size, isize);
4585 if (size > sbi->s_mb_stream_request) {
4586 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4590 BUG_ON(ac->ac_lg != NULL);
4592 * locality group prealloc space are per cpu. The reason for having
4593 * per cpu locality group is to reduce the contention between block
4594 * request from multiple CPUs.
4596 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4598 /* we're going to use group allocation */
4599 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4601 /* serialize all allocations in the group */
4602 mutex_lock(&ac->ac_lg->lg_mutex);
4605 static noinline_for_stack int
4606 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4607 struct ext4_allocation_request *ar)
4609 struct super_block *sb = ar->inode->i_sb;
4610 struct ext4_sb_info *sbi = EXT4_SB(sb);
4611 struct ext4_super_block *es = sbi->s_es;
4615 ext4_grpblk_t block;
4617 /* we can't allocate > group size */
4620 /* just a dirty hack to filter too big requests */
4621 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4622 len = EXT4_CLUSTERS_PER_GROUP(sb);
4624 /* start searching from the goal */
4626 if (goal < le32_to_cpu(es->s_first_data_block) ||
4627 goal >= ext4_blocks_count(es))
4628 goal = le32_to_cpu(es->s_first_data_block);
4629 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4631 /* set up allocation goals */
4632 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4633 ac->ac_status = AC_STATUS_CONTINUE;
4635 ac->ac_inode = ar->inode;
4636 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4637 ac->ac_o_ex.fe_group = group;
4638 ac->ac_o_ex.fe_start = block;
4639 ac->ac_o_ex.fe_len = len;
4640 ac->ac_g_ex = ac->ac_o_ex;
4641 ac->ac_flags = ar->flags;
4643 /* we have to define context: we'll work with a file or
4644 * locality group. this is a policy, actually */
4645 ext4_mb_group_or_file(ac);
4647 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
4648 "left: %u/%u, right %u/%u to %swritable\n",
4649 (unsigned) ar->len, (unsigned) ar->logical,
4650 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4651 (unsigned) ar->lleft, (unsigned) ar->pleft,
4652 (unsigned) ar->lright, (unsigned) ar->pright,
4653 inode_is_open_for_write(ar->inode) ? "" : "non-");
4658 static noinline_for_stack void
4659 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4660 struct ext4_locality_group *lg,
4661 int order, int total_entries)
4663 ext4_group_t group = 0;
4664 struct ext4_buddy e4b;
4665 struct list_head discard_list;
4666 struct ext4_prealloc_space *pa, *tmp;
4668 mb_debug(sb, "discard locality group preallocation\n");
4670 INIT_LIST_HEAD(&discard_list);
4672 spin_lock(&lg->lg_prealloc_lock);
4673 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4675 lockdep_is_held(&lg->lg_prealloc_lock)) {
4676 spin_lock(&pa->pa_lock);
4677 if (atomic_read(&pa->pa_count)) {
4679 * This is the pa that we just used
4680 * for block allocation. So don't
4683 spin_unlock(&pa->pa_lock);
4686 if (pa->pa_deleted) {
4687 spin_unlock(&pa->pa_lock);
4690 /* only lg prealloc space */
4691 BUG_ON(pa->pa_type != MB_GROUP_PA);
4693 /* seems this one can be freed ... */
4694 ext4_mb_mark_pa_deleted(sb, pa);
4695 spin_unlock(&pa->pa_lock);
4697 list_del_rcu(&pa->pa_inode_list);
4698 list_add(&pa->u.pa_tmp_list, &discard_list);
4701 if (total_entries <= 5) {
4703 * we want to keep only 5 entries
4704 * allowing it to grow to 8. This
4705 * mak sure we don't call discard
4706 * soon for this list.
4711 spin_unlock(&lg->lg_prealloc_lock);
4713 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4716 group = ext4_get_group_number(sb, pa->pa_pstart);
4717 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4718 GFP_NOFS|__GFP_NOFAIL);
4720 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
4724 ext4_lock_group(sb, group);
4725 list_del(&pa->pa_group_list);
4726 ext4_mb_release_group_pa(&e4b, pa);
4727 ext4_unlock_group(sb, group);
4729 ext4_mb_unload_buddy(&e4b);
4730 list_del(&pa->u.pa_tmp_list);
4731 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4736 * We have incremented pa_count. So it cannot be freed at this
4737 * point. Also we hold lg_mutex. So no parallel allocation is
4738 * possible from this lg. That means pa_free cannot be updated.
4740 * A parallel ext4_mb_discard_group_preallocations is possible.
4741 * which can cause the lg_prealloc_list to be updated.
4744 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4746 int order, added = 0, lg_prealloc_count = 1;
4747 struct super_block *sb = ac->ac_sb;
4748 struct ext4_locality_group *lg = ac->ac_lg;
4749 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4751 order = fls(pa->pa_free) - 1;
4752 if (order > PREALLOC_TB_SIZE - 1)
4753 /* The max size of hash table is PREALLOC_TB_SIZE */
4754 order = PREALLOC_TB_SIZE - 1;
4755 /* Add the prealloc space to lg */
4756 spin_lock(&lg->lg_prealloc_lock);
4757 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4759 lockdep_is_held(&lg->lg_prealloc_lock)) {
4760 spin_lock(&tmp_pa->pa_lock);
4761 if (tmp_pa->pa_deleted) {
4762 spin_unlock(&tmp_pa->pa_lock);
4765 if (!added && pa->pa_free < tmp_pa->pa_free) {
4766 /* Add to the tail of the previous entry */
4767 list_add_tail_rcu(&pa->pa_inode_list,
4768 &tmp_pa->pa_inode_list);
4771 * we want to count the total
4772 * number of entries in the list
4775 spin_unlock(&tmp_pa->pa_lock);
4776 lg_prealloc_count++;
4779 list_add_tail_rcu(&pa->pa_inode_list,
4780 &lg->lg_prealloc_list[order]);
4781 spin_unlock(&lg->lg_prealloc_lock);
4783 /* Now trim the list to be not more than 8 elements */
4784 if (lg_prealloc_count > 8) {
4785 ext4_mb_discard_lg_preallocations(sb, lg,
4786 order, lg_prealloc_count);
4793 * if per-inode prealloc list is too long, trim some PA
4795 static void ext4_mb_trim_inode_pa(struct inode *inode)
4797 struct ext4_inode_info *ei = EXT4_I(inode);
4798 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4801 count = atomic_read(&ei->i_prealloc_active);
4802 delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
4803 if (count > sbi->s_mb_max_inode_prealloc + delta) {
4804 count -= sbi->s_mb_max_inode_prealloc;
4805 ext4_discard_preallocations(inode, count);
4810 * release all resource we used in allocation
4812 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4814 struct inode *inode = ac->ac_inode;
4815 struct ext4_inode_info *ei = EXT4_I(inode);
4816 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4817 struct ext4_prealloc_space *pa = ac->ac_pa;
4819 if (pa->pa_type == MB_GROUP_PA) {
4820 /* see comment in ext4_mb_use_group_pa() */
4821 spin_lock(&pa->pa_lock);
4822 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4823 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4824 pa->pa_free -= ac->ac_b_ex.fe_len;
4825 pa->pa_len -= ac->ac_b_ex.fe_len;
4826 spin_unlock(&pa->pa_lock);
4829 * We want to add the pa to the right bucket.
4830 * Remove it from the list and while adding
4831 * make sure the list to which we are adding
4834 if (likely(pa->pa_free)) {
4835 spin_lock(pa->pa_obj_lock);
4836 list_del_rcu(&pa->pa_inode_list);
4837 spin_unlock(pa->pa_obj_lock);
4838 ext4_mb_add_n_trim(ac);
4842 if (pa->pa_type == MB_INODE_PA) {
4844 * treat per-inode prealloc list as a lru list, then try
4845 * to trim the least recently used PA.
4847 spin_lock(pa->pa_obj_lock);
4848 list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
4849 spin_unlock(pa->pa_obj_lock);
4852 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4854 if (ac->ac_bitmap_page)
4855 put_page(ac->ac_bitmap_page);
4856 if (ac->ac_buddy_page)
4857 put_page(ac->ac_buddy_page);
4858 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4859 mutex_unlock(&ac->ac_lg->lg_mutex);
4860 ext4_mb_collect_stats(ac);
4861 ext4_mb_trim_inode_pa(inode);
4865 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4867 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4871 trace_ext4_mb_discard_preallocations(sb, needed);
4872 for (i = 0; i < ngroups && needed > 0; i++) {
4873 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4881 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
4882 struct ext4_allocation_context *ac, u64 *seq)
4888 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4893 seq_retry = ext4_get_discard_pa_seq_sum();
4894 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
4895 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
4901 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
4905 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
4906 struct ext4_allocation_request *ar, int *errp);
4909 * Main entry point into mballoc to allocate blocks
4910 * it tries to use preallocation first, then falls back
4911 * to usual allocation
4913 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4914 struct ext4_allocation_request *ar, int *errp)
4916 struct ext4_allocation_context *ac = NULL;
4917 struct ext4_sb_info *sbi;
4918 struct super_block *sb;
4919 ext4_fsblk_t block = 0;
4920 unsigned int inquota = 0;
4921 unsigned int reserv_clstrs = 0;
4925 sb = ar->inode->i_sb;
4928 trace_ext4_request_blocks(ar);
4929 if (sbi->s_mount_state & EXT4_FC_REPLAY)
4930 return ext4_mb_new_blocks_simple(handle, ar, errp);
4932 /* Allow to use superuser reservation for quota file */
4933 if (ext4_is_quota_file(ar->inode))
4934 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4936 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4937 /* Without delayed allocation we need to verify
4938 * there is enough free blocks to do block allocation
4939 * and verify allocation doesn't exceed the quota limits.
4942 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4944 /* let others to free the space */
4946 ar->len = ar->len >> 1;
4949 ext4_mb_show_pa(sb);
4953 reserv_clstrs = ar->len;
4954 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4955 dquot_alloc_block_nofail(ar->inode,
4956 EXT4_C2B(sbi, ar->len));
4959 dquot_alloc_block(ar->inode,
4960 EXT4_C2B(sbi, ar->len))) {
4962 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4973 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4980 *errp = ext4_mb_initialize_context(ac, ar);
4986 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4987 seq = this_cpu_read(discard_pa_seq);
4988 if (!ext4_mb_use_preallocated(ac)) {
4989 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4990 ext4_mb_normalize_request(ac, ar);
4992 *errp = ext4_mb_pa_alloc(ac);
4996 /* allocate space in core */
4997 *errp = ext4_mb_regular_allocator(ac);
4999 * pa allocated above is added to grp->bb_prealloc_list only
5000 * when we were able to allocate some block i.e. when
5001 * ac->ac_status == AC_STATUS_FOUND.
5002 * And error from above mean ac->ac_status != AC_STATUS_FOUND
5003 * So we have to free this pa here itself.
5006 ext4_mb_pa_free(ac);
5007 ext4_discard_allocated_blocks(ac);
5010 if (ac->ac_status == AC_STATUS_FOUND &&
5011 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
5012 ext4_mb_pa_free(ac);
5014 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
5015 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
5017 ext4_discard_allocated_blocks(ac);
5020 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5021 ar->len = ac->ac_b_ex.fe_len;
5024 if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
5027 * If block allocation fails then the pa allocated above
5028 * needs to be freed here itself.
5030 ext4_mb_pa_free(ac);
5036 ac->ac_b_ex.fe_len = 0;
5038 ext4_mb_show_ac(ac);
5040 ext4_mb_release_context(ac);
5043 kmem_cache_free(ext4_ac_cachep, ac);
5044 if (inquota && ar->len < inquota)
5045 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
5047 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
5048 /* release all the reserved blocks if non delalloc */
5049 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
5053 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
5059 * We can merge two free data extents only if the physical blocks
5060 * are contiguous, AND the extents were freed by the same transaction,
5061 * AND the blocks are associated with the same group.
5063 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
5064 struct ext4_free_data *entry,
5065 struct ext4_free_data *new_entry,
5066 struct rb_root *entry_rb_root)
5068 if ((entry->efd_tid != new_entry->efd_tid) ||
5069 (entry->efd_group != new_entry->efd_group))
5071 if (entry->efd_start_cluster + entry->efd_count ==
5072 new_entry->efd_start_cluster) {
5073 new_entry->efd_start_cluster = entry->efd_start_cluster;
5074 new_entry->efd_count += entry->efd_count;
5075 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
5076 entry->efd_start_cluster) {
5077 new_entry->efd_count += entry->efd_count;
5080 spin_lock(&sbi->s_md_lock);
5081 list_del(&entry->efd_list);
5082 spin_unlock(&sbi->s_md_lock);
5083 rb_erase(&entry->efd_node, entry_rb_root);
5084 kmem_cache_free(ext4_free_data_cachep, entry);
5087 static noinline_for_stack int
5088 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
5089 struct ext4_free_data *new_entry)
5091 ext4_group_t group = e4b->bd_group;
5092 ext4_grpblk_t cluster;
5093 ext4_grpblk_t clusters = new_entry->efd_count;
5094 struct ext4_free_data *entry;
5095 struct ext4_group_info *db = e4b->bd_info;
5096 struct super_block *sb = e4b->bd_sb;
5097 struct ext4_sb_info *sbi = EXT4_SB(sb);
5098 struct rb_node **n = &db->bb_free_root.rb_node, *node;
5099 struct rb_node *parent = NULL, *new_node;
5101 BUG_ON(!ext4_handle_valid(handle));
5102 BUG_ON(e4b->bd_bitmap_page == NULL);
5103 BUG_ON(e4b->bd_buddy_page == NULL);
5105 new_node = &new_entry->efd_node;
5106 cluster = new_entry->efd_start_cluster;
5109 /* first free block exent. We need to
5110 protect buddy cache from being freed,
5111 * otherwise we'll refresh it from
5112 * on-disk bitmap and lose not-yet-available
5114 get_page(e4b->bd_buddy_page);
5115 get_page(e4b->bd_bitmap_page);
5119 entry = rb_entry(parent, struct ext4_free_data, efd_node);
5120 if (cluster < entry->efd_start_cluster)
5122 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
5123 n = &(*n)->rb_right;
5125 ext4_grp_locked_error(sb, group, 0,
5126 ext4_group_first_block_no(sb, group) +
5127 EXT4_C2B(sbi, cluster),
5128 "Block already on to-be-freed list");
5133 rb_link_node(new_node, parent, n);
5134 rb_insert_color(new_node, &db->bb_free_root);
5136 /* Now try to see the extent can be merged to left and right */
5137 node = rb_prev(new_node);
5139 entry = rb_entry(node, struct ext4_free_data, efd_node);
5140 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5141 &(db->bb_free_root));
5144 node = rb_next(new_node);
5146 entry = rb_entry(node, struct ext4_free_data, efd_node);
5147 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5148 &(db->bb_free_root));
5151 spin_lock(&sbi->s_md_lock);
5152 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5153 sbi->s_mb_free_pending += clusters;
5154 spin_unlock(&sbi->s_md_lock);
5159 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
5160 * linearly starting at the goal block and also excludes the blocks which
5161 * are going to be in use after fast commit replay.
5163 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5164 struct ext4_allocation_request *ar, int *errp)
5166 struct buffer_head *bitmap_bh;
5167 struct super_block *sb = ar->inode->i_sb;
5169 ext4_grpblk_t blkoff;
5171 ext4_fsblk_t goal, block;
5172 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
5175 if (goal < le32_to_cpu(es->s_first_data_block) ||
5176 goal >= ext4_blocks_count(es))
5177 goal = le32_to_cpu(es->s_first_data_block);
5180 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
5181 for (; group < ext4_get_groups_count(sb); group++) {
5182 bitmap_bh = ext4_read_block_bitmap(sb, group);
5183 if (IS_ERR(bitmap_bh)) {
5184 *errp = PTR_ERR(bitmap_bh);
5185 pr_warn("Failed to read block bitmap\n");
5189 ext4_get_group_no_and_offset(sb,
5190 max(ext4_group_first_block_no(sb, group), goal),
5192 i = mb_find_next_zero_bit(bitmap_bh->b_data, sb->s_blocksize,
5195 if (i >= sb->s_blocksize)
5197 if (ext4_fc_replay_check_excluded(sb,
5198 ext4_group_first_block_no(sb, group) + i))
5203 if (group >= ext4_get_groups_count(sb) && i >= sb->s_blocksize)
5206 block = ext4_group_first_block_no(sb, group) + i;
5207 ext4_mb_mark_bb(sb, block, 1, 1);
5213 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
5214 unsigned long count)
5216 struct buffer_head *bitmap_bh;
5217 struct super_block *sb = inode->i_sb;
5218 struct ext4_group_desc *gdp;
5219 struct buffer_head *gdp_bh;
5221 ext4_grpblk_t blkoff;
5222 int already_freed = 0, err, i;
5224 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
5225 bitmap_bh = ext4_read_block_bitmap(sb, group);
5226 if (IS_ERR(bitmap_bh)) {
5227 err = PTR_ERR(bitmap_bh);
5228 pr_warn("Failed to read block bitmap\n");
5231 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
5235 for (i = 0; i < count; i++) {
5236 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data))
5239 mb_clear_bits(bitmap_bh->b_data, blkoff, count);
5240 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
5243 ext4_free_group_clusters_set(
5244 sb, gdp, ext4_free_group_clusters(sb, gdp) +
5245 count - already_freed);
5246 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
5247 ext4_group_desc_csum_set(sb, group, gdp);
5248 ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
5249 sync_dirty_buffer(bitmap_bh);
5250 sync_dirty_buffer(gdp_bh);
5255 * ext4_free_blocks() -- Free given blocks and update quota
5256 * @handle: handle for this transaction
5258 * @bh: optional buffer of the block to be freed
5259 * @block: starting physical block to be freed
5260 * @count: number of blocks to be freed
5261 * @flags: flags used by ext4_free_blocks
5263 void ext4_free_blocks(handle_t *handle, struct inode *inode,
5264 struct buffer_head *bh, ext4_fsblk_t block,
5265 unsigned long count, int flags)
5267 struct buffer_head *bitmap_bh = NULL;
5268 struct super_block *sb = inode->i_sb;
5269 struct ext4_group_desc *gdp;
5270 unsigned int overflow;
5272 struct buffer_head *gd_bh;
5273 ext4_group_t block_group;
5274 struct ext4_sb_info *sbi;
5275 struct ext4_buddy e4b;
5276 unsigned int count_clusters;
5282 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
5283 ext4_free_blocks_simple(inode, block, count);
5290 BUG_ON(block != bh->b_blocknr);
5292 block = bh->b_blocknr;
5295 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5296 !ext4_inode_block_valid(inode, block, count)) {
5297 ext4_error(sb, "Freeing blocks not in datazone - "
5298 "block = %llu, count = %lu", block, count);
5302 ext4_debug("freeing block %llu\n", block);
5303 trace_ext4_free_blocks(inode, block, count, flags);
5305 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
5308 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
5313 * If the extent to be freed does not begin on a cluster
5314 * boundary, we need to deal with partial clusters at the
5315 * beginning and end of the extent. Normally we will free
5316 * blocks at the beginning or the end unless we are explicitly
5317 * requested to avoid doing so.
5319 overflow = EXT4_PBLK_COFF(sbi, block);
5321 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
5322 overflow = sbi->s_cluster_ratio - overflow;
5324 if (count > overflow)
5333 overflow = EXT4_LBLK_COFF(sbi, count);
5335 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
5336 if (count > overflow)
5341 count += sbi->s_cluster_ratio - overflow;
5344 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
5346 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
5348 for (i = 0; i < count; i++) {
5351 bh = sb_find_get_block(inode->i_sb, block + i);
5352 ext4_forget(handle, is_metadata, inode, bh, block + i);
5358 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5360 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5361 ext4_get_group_info(sb, block_group))))
5365 * Check to see if we are freeing blocks across a group
5368 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5369 overflow = EXT4_C2B(sbi, bit) + count -
5370 EXT4_BLOCKS_PER_GROUP(sb);
5373 count_clusters = EXT4_NUM_B2C(sbi, count);
5374 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5375 if (IS_ERR(bitmap_bh)) {
5376 err = PTR_ERR(bitmap_bh);
5380 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5386 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
5387 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
5388 in_range(block, ext4_inode_table(sb, gdp),
5389 sbi->s_itb_per_group) ||
5390 in_range(block + count - 1, ext4_inode_table(sb, gdp),
5391 sbi->s_itb_per_group)) {
5393 ext4_error(sb, "Freeing blocks in system zone - "
5394 "Block = %llu, count = %lu", block, count);
5395 /* err = 0. ext4_std_error should be a no op */
5399 BUFFER_TRACE(bitmap_bh, "getting write access");
5400 err = ext4_journal_get_write_access(handle, bitmap_bh);
5405 * We are about to modify some metadata. Call the journal APIs
5406 * to unshare ->b_data if a currently-committing transaction is
5409 BUFFER_TRACE(gd_bh, "get_write_access");
5410 err = ext4_journal_get_write_access(handle, gd_bh);
5413 #ifdef AGGRESSIVE_CHECK
5416 for (i = 0; i < count_clusters; i++)
5417 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
5420 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
5422 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
5423 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
5424 GFP_NOFS|__GFP_NOFAIL);
5429 * We need to make sure we don't reuse the freed block until after the
5430 * transaction is committed. We make an exception if the inode is to be
5431 * written in writeback mode since writeback mode has weak data
5432 * consistency guarantees.
5434 if (ext4_handle_valid(handle) &&
5435 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
5436 !ext4_should_writeback_data(inode))) {
5437 struct ext4_free_data *new_entry;
5439 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
5442 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
5443 GFP_NOFS|__GFP_NOFAIL);
5444 new_entry->efd_start_cluster = bit;
5445 new_entry->efd_group = block_group;
5446 new_entry->efd_count = count_clusters;
5447 new_entry->efd_tid = handle->h_transaction->t_tid;
5449 ext4_lock_group(sb, block_group);
5450 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
5451 ext4_mb_free_metadata(handle, &e4b, new_entry);
5453 /* need to update group_info->bb_free and bitmap
5454 * with group lock held. generate_buddy look at
5455 * them with group lock_held
5457 if (test_opt(sb, DISCARD)) {
5458 err = ext4_issue_discard(sb, block_group, bit, count,
5460 if (err && err != -EOPNOTSUPP)
5461 ext4_msg(sb, KERN_WARNING, "discard request in"
5462 " group:%d block:%d count:%lu failed"
5463 " with %d", block_group, bit, count,
5466 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
5468 ext4_lock_group(sb, block_group);
5469 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
5470 mb_free_blocks(inode, &e4b, bit, count_clusters);
5473 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
5474 ext4_free_group_clusters_set(sb, gdp, ret);
5475 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
5476 ext4_group_desc_csum_set(sb, block_group, gdp);
5477 ext4_unlock_group(sb, block_group);
5479 if (sbi->s_log_groups_per_flex) {
5480 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5481 atomic64_add(count_clusters,
5482 &sbi_array_rcu_deref(sbi, s_flex_groups,
5483 flex_group)->free_clusters);
5487 * on a bigalloc file system, defer the s_freeclusters_counter
5488 * update to the caller (ext4_remove_space and friends) so they
5489 * can determine if a cluster freed here should be rereserved
5491 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
5492 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
5493 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
5494 percpu_counter_add(&sbi->s_freeclusters_counter,
5498 ext4_mb_unload_buddy(&e4b);
5500 /* We dirtied the bitmap block */
5501 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5502 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5504 /* And the group descriptor block */
5505 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5506 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5510 if (overflow && !err) {
5518 ext4_std_error(sb, err);
5523 * ext4_group_add_blocks() -- Add given blocks to an existing group
5524 * @handle: handle to this transaction
5526 * @block: start physical block to add to the block group
5527 * @count: number of blocks to free
5529 * This marks the blocks as free in the bitmap and buddy.
5531 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
5532 ext4_fsblk_t block, unsigned long count)
5534 struct buffer_head *bitmap_bh = NULL;
5535 struct buffer_head *gd_bh;
5536 ext4_group_t block_group;
5539 struct ext4_group_desc *desc;
5540 struct ext4_sb_info *sbi = EXT4_SB(sb);
5541 struct ext4_buddy e4b;
5542 int err = 0, ret, free_clusters_count;
5543 ext4_grpblk_t clusters_freed;
5544 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
5545 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
5546 unsigned long cluster_count = last_cluster - first_cluster + 1;
5548 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
5553 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5555 * Check to see if we are freeing blocks across a group
5558 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
5559 ext4_warning(sb, "too many blocks added to group %u",
5565 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5566 if (IS_ERR(bitmap_bh)) {
5567 err = PTR_ERR(bitmap_bh);
5572 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
5578 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
5579 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
5580 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
5581 in_range(block + count - 1, ext4_inode_table(sb, desc),
5582 sbi->s_itb_per_group)) {
5583 ext4_error(sb, "Adding blocks in system zones - "
5584 "Block = %llu, count = %lu",
5590 BUFFER_TRACE(bitmap_bh, "getting write access");
5591 err = ext4_journal_get_write_access(handle, bitmap_bh);
5596 * We are about to modify some metadata. Call the journal APIs
5597 * to unshare ->b_data if a currently-committing transaction is
5600 BUFFER_TRACE(gd_bh, "get_write_access");
5601 err = ext4_journal_get_write_access(handle, gd_bh);
5605 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
5606 BUFFER_TRACE(bitmap_bh, "clear bit");
5607 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
5608 ext4_error(sb, "bit already cleared for block %llu",
5609 (ext4_fsblk_t)(block + i));
5610 BUFFER_TRACE(bitmap_bh, "bit already cleared");
5616 err = ext4_mb_load_buddy(sb, block_group, &e4b);
5621 * need to update group_info->bb_free and bitmap
5622 * with group lock held. generate_buddy look at
5623 * them with group lock_held
5625 ext4_lock_group(sb, block_group);
5626 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
5627 mb_free_blocks(NULL, &e4b, bit, cluster_count);
5628 free_clusters_count = clusters_freed +
5629 ext4_free_group_clusters(sb, desc);
5630 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
5631 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5632 ext4_group_desc_csum_set(sb, block_group, desc);
5633 ext4_unlock_group(sb, block_group);
5634 percpu_counter_add(&sbi->s_freeclusters_counter,
5637 if (sbi->s_log_groups_per_flex) {
5638 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5639 atomic64_add(clusters_freed,
5640 &sbi_array_rcu_deref(sbi, s_flex_groups,
5641 flex_group)->free_clusters);
5644 ext4_mb_unload_buddy(&e4b);
5646 /* We dirtied the bitmap block */
5647 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5648 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5650 /* And the group descriptor block */
5651 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5652 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5658 ext4_std_error(sb, err);
5663 * ext4_trim_extent -- function to TRIM one single free extent in the group
5664 * @sb: super block for the file system
5665 * @start: starting block of the free extent in the alloc. group
5666 * @count: number of blocks to TRIM
5667 * @group: alloc. group we are working with
5668 * @e4b: ext4 buddy for the group
5670 * Trim "count" blocks starting at "start" in the "group". To assure that no
5671 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5672 * be called with under the group lock.
5674 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5675 ext4_group_t group, struct ext4_buddy *e4b)
5679 struct ext4_free_extent ex;
5682 trace_ext4_trim_extent(sb, group, start, count);
5684 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5686 ex.fe_start = start;
5687 ex.fe_group = group;
5691 * Mark blocks used, so no one can reuse them while
5694 mb_mark_used(e4b, &ex);
5695 ext4_unlock_group(sb, group);
5696 ret = ext4_issue_discard(sb, group, start, count, NULL);
5697 ext4_lock_group(sb, group);
5698 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5703 * ext4_trim_all_free -- function to trim all free space in alloc. group
5704 * @sb: super block for file system
5705 * @group: group to be trimmed
5706 * @start: first group block to examine
5707 * @max: last group block to examine
5708 * @minblocks: minimum extent block count
5710 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5711 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5715 * ext4_trim_all_free walks through group's block bitmap searching for free
5716 * extents. When the free extent is found, mark it as used in group buddy
5717 * bitmap. Then issue a TRIM command on this extent and free the extent in
5718 * the group buddy bitmap. This is done until whole group is scanned.
5720 static ext4_grpblk_t
5721 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5722 ext4_grpblk_t start, ext4_grpblk_t max,
5723 ext4_grpblk_t minblocks)
5726 ext4_grpblk_t next, count = 0, free_count = 0;
5727 struct ext4_buddy e4b;
5730 trace_ext4_trim_all_free(sb, group, start, max);
5732 ret = ext4_mb_load_buddy(sb, group, &e4b);
5734 ext4_warning(sb, "Error %d loading buddy information for %u",
5738 bitmap = e4b.bd_bitmap;
5740 ext4_lock_group(sb, group);
5741 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5742 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5745 start = (e4b.bd_info->bb_first_free > start) ?
5746 e4b.bd_info->bb_first_free : start;
5748 while (start <= max) {
5749 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5752 next = mb_find_next_bit(bitmap, max + 1, start);
5754 if ((next - start) >= minblocks) {
5755 ret = ext4_trim_extent(sb, start,
5756 next - start, group, &e4b);
5757 if (ret && ret != -EOPNOTSUPP)
5760 count += next - start;
5762 free_count += next - start;
5765 if (fatal_signal_pending(current)) {
5766 count = -ERESTARTSYS;
5770 if (need_resched()) {
5771 ext4_unlock_group(sb, group);
5773 ext4_lock_group(sb, group);
5776 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5782 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5785 ext4_unlock_group(sb, group);
5786 ext4_mb_unload_buddy(&e4b);
5788 ext4_debug("trimmed %d blocks in the group %d\n",
5795 * ext4_trim_fs() -- trim ioctl handle function
5796 * @sb: superblock for filesystem
5797 * @range: fstrim_range structure
5799 * start: First Byte to trim
5800 * len: number of Bytes to trim from start
5801 * minlen: minimum extent length in Bytes
5802 * ext4_trim_fs goes through all allocation groups containing Bytes from
5803 * start to start+len. For each such a group ext4_trim_all_free function
5804 * is invoked to trim all free space.
5806 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5808 struct ext4_group_info *grp;
5809 ext4_group_t group, first_group, last_group;
5810 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5811 uint64_t start, end, minlen, trimmed = 0;
5812 ext4_fsblk_t first_data_blk =
5813 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5814 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5817 start = range->start >> sb->s_blocksize_bits;
5818 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5819 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5820 range->minlen >> sb->s_blocksize_bits);
5822 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5823 start >= max_blks ||
5824 range->len < sb->s_blocksize)
5826 if (end >= max_blks)
5828 if (end <= first_data_blk)
5830 if (start < first_data_blk)
5831 start = first_data_blk;
5833 /* Determine first and last group to examine based on start and end */
5834 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5835 &first_group, &first_cluster);
5836 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5837 &last_group, &last_cluster);
5839 /* end now represents the last cluster to discard in this group */
5840 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5842 for (group = first_group; group <= last_group; group++) {
5843 grp = ext4_get_group_info(sb, group);
5844 /* We only do this if the grp has never been initialized */
5845 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5846 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5852 * For all the groups except the last one, last cluster will
5853 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5854 * change it for the last group, note that last_cluster is
5855 * already computed earlier by ext4_get_group_no_and_offset()
5857 if (group == last_group)
5860 if (grp->bb_free >= minlen) {
5861 cnt = ext4_trim_all_free(sb, group, first_cluster,
5871 * For every group except the first one, we are sure
5872 * that the first cluster to discard will be cluster #0.
5878 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5881 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5885 /* Iterate all the free extents in the group. */
5887 ext4_mballoc_query_range(
5888 struct super_block *sb,
5890 ext4_grpblk_t start,
5892 ext4_mballoc_query_range_fn formatter,
5897 struct ext4_buddy e4b;
5900 error = ext4_mb_load_buddy(sb, group, &e4b);
5903 bitmap = e4b.bd_bitmap;
5905 ext4_lock_group(sb, group);
5907 start = (e4b.bd_info->bb_first_free > start) ?
5908 e4b.bd_info->bb_first_free : start;
5909 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
5910 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5912 while (start <= end) {
5913 start = mb_find_next_zero_bit(bitmap, end + 1, start);
5916 next = mb_find_next_bit(bitmap, end + 1, start);
5918 ext4_unlock_group(sb, group);
5919 error = formatter(sb, group, start, next - start, priv);
5922 ext4_lock_group(sb, group);
5927 ext4_unlock_group(sb, group);
5929 ext4_mb_unload_buddy(&e4b);