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 * If "mb_optimize_scan" mount option is set, we maintain in memory group info
131 * structures in two data structures:
133 * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders)
135 * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks)
137 * This is an array of lists where the index in the array represents the
138 * largest free order in the buddy bitmap of the participating group infos of
139 * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total
140 * number of buddy bitmap orders possible) number of lists. Group-infos are
141 * placed in appropriate lists.
143 * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size)
145 * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks)
147 * This is an array of lists where in the i-th list there are groups with
148 * average fragment size >= 2^i and < 2^(i+1). The average fragment size
149 * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments.
150 * Note that we don't bother with a special list for completely empty groups
151 * so we only have MB_NUM_ORDERS(sb) lists.
153 * When "mb_optimize_scan" mount option is set, mballoc consults the above data
154 * structures to decide the order in which groups are to be traversed for
155 * fulfilling an allocation request.
157 * At CR = 0, we look for groups which have the largest_free_order >= the order
158 * of the request. We directly look at the largest free order list in the data
159 * structure (1) above where largest_free_order = order of the request. If that
160 * list is empty, we look at remaining list in the increasing order of
161 * largest_free_order. This allows us to perform CR = 0 lookup in O(1) time.
163 * At CR = 1, we only consider groups where average fragment size > request
164 * size. So, we lookup a group which has average fragment size just above or
165 * equal to request size using our average fragment size group lists (data
166 * structure 2) in O(1) time.
168 * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in
169 * linear order which requires O(N) search time for each CR 0 and CR 1 phase.
171 * The regular allocator (using the buddy cache) supports a few tunables.
173 * /sys/fs/ext4/<partition>/mb_min_to_scan
174 * /sys/fs/ext4/<partition>/mb_max_to_scan
175 * /sys/fs/ext4/<partition>/mb_order2_req
176 * /sys/fs/ext4/<partition>/mb_linear_limit
178 * The regular allocator uses buddy scan only if the request len is power of
179 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
180 * value of s_mb_order2_reqs can be tuned via
181 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
182 * stripe size (sbi->s_stripe), we try to search for contiguous block in
183 * stripe size. This should result in better allocation on RAID setups. If
184 * not, we search in the specific group using bitmap for best extents. The
185 * tunable min_to_scan and max_to_scan control the behaviour here.
186 * min_to_scan indicate how long the mballoc __must__ look for a best
187 * extent and max_to_scan indicates how long the mballoc __can__ look for a
188 * best extent in the found extents. Searching for the blocks starts with
189 * the group specified as the goal value in allocation context via
190 * ac_g_ex. Each group is first checked based on the criteria whether it
191 * can be used for allocation. ext4_mb_good_group explains how the groups are
194 * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not
195 * get traversed linearly. That may result in subsequent allocations being not
196 * close to each other. And so, the underlying device may get filled up in a
197 * non-linear fashion. While that may not matter on non-rotational devices, for
198 * rotational devices that may result in higher seek times. "mb_linear_limit"
199 * tells mballoc how many groups mballoc should search linearly before
200 * performing consulting above data structures for more efficient lookups. For
201 * non rotational devices, this value defaults to 0 and for rotational devices
202 * this is set to MB_DEFAULT_LINEAR_LIMIT.
204 * Both the prealloc space are getting populated as above. So for the first
205 * request we will hit the buddy cache which will result in this prealloc
206 * space getting filled. The prealloc space is then later used for the
207 * subsequent request.
211 * mballoc operates on the following data:
213 * - in-core buddy (actually includes buddy and bitmap)
214 * - preallocation descriptors (PAs)
216 * there are two types of preallocations:
218 * assiged to specific inode and can be used for this inode only.
219 * it describes part of inode's space preallocated to specific
220 * physical blocks. any block from that preallocated can be used
221 * independent. the descriptor just tracks number of blocks left
222 * unused. so, before taking some block from descriptor, one must
223 * make sure corresponded logical block isn't allocated yet. this
224 * also means that freeing any block within descriptor's range
225 * must discard all preallocated blocks.
227 * assigned to specific locality group which does not translate to
228 * permanent set of inodes: inode can join and leave group. space
229 * from this type of preallocation can be used for any inode. thus
230 * it's consumed from the beginning to the end.
232 * relation between them can be expressed as:
233 * in-core buddy = on-disk bitmap + preallocation descriptors
235 * this mean blocks mballoc considers used are:
236 * - allocated blocks (persistent)
237 * - preallocated blocks (non-persistent)
239 * consistency in mballoc world means that at any time a block is either
240 * free or used in ALL structures. notice: "any time" should not be read
241 * literally -- time is discrete and delimited by locks.
243 * to keep it simple, we don't use block numbers, instead we count number of
244 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
246 * all operations can be expressed as:
247 * - init buddy: buddy = on-disk + PAs
248 * - new PA: buddy += N; PA = N
249 * - use inode PA: on-disk += N; PA -= N
250 * - discard inode PA buddy -= on-disk - PA; PA = 0
251 * - use locality group PA on-disk += N; PA -= N
252 * - discard locality group PA buddy -= PA; PA = 0
253 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
254 * is used in real operation because we can't know actual used
255 * bits from PA, only from on-disk bitmap
257 * if we follow this strict logic, then all operations above should be atomic.
258 * given some of them can block, we'd have to use something like semaphores
259 * killing performance on high-end SMP hardware. let's try to relax it using
260 * the following knowledge:
261 * 1) if buddy is referenced, it's already initialized
262 * 2) while block is used in buddy and the buddy is referenced,
263 * nobody can re-allocate that block
264 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
265 * bit set and PA claims same block, it's OK. IOW, one can set bit in
266 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
269 * so, now we're building a concurrency table:
272 * blocks for PA are allocated in the buddy, buddy must be referenced
273 * until PA is linked to allocation group to avoid concurrent buddy init
275 * we need to make sure that either on-disk bitmap or PA has uptodate data
276 * given (3) we care that PA-=N operation doesn't interfere with init
278 * the simplest way would be to have buddy initialized by the discard
279 * - use locality group PA
280 * again PA-=N must be serialized with init
281 * - discard locality group PA
282 * the simplest way would be to have buddy initialized by the discard
285 * i_data_sem serializes them
287 * discard process must wait until PA isn't used by another process
288 * - use locality group PA
289 * some mutex should serialize them
290 * - discard locality group PA
291 * discard process must wait until PA isn't used by another process
294 * i_data_sem or another mutex should serializes them
296 * discard process must wait until PA isn't used by another process
297 * - use locality group PA
298 * nothing wrong here -- they're different PAs covering different blocks
299 * - discard locality group PA
300 * discard process must wait until PA isn't used by another process
302 * now we're ready to make few consequences:
303 * - PA is referenced and while it is no discard is possible
304 * - PA is referenced until block isn't marked in on-disk bitmap
305 * - PA changes only after on-disk bitmap
306 * - discard must not compete with init. either init is done before
307 * any discard or they're serialized somehow
308 * - buddy init as sum of on-disk bitmap and PAs is done atomically
310 * a special case when we've used PA to emptiness. no need to modify buddy
311 * in this case, but we should care about concurrent init
316 * Logic in few words:
321 * mark bits in on-disk bitmap
324 * - use preallocation:
325 * find proper PA (per-inode or group)
327 * mark bits in on-disk bitmap
333 * mark bits in on-disk bitmap
336 * - discard preallocations in group:
338 * move them onto local list
339 * load on-disk bitmap
341 * remove PA from object (inode or locality group)
342 * mark free blocks in-core
344 * - discard inode's preallocations:
351 * - bitlock on a group (group)
352 * - object (inode/locality) (object)
354 * - cr0 lists lock (cr0)
355 * - cr1 tree lock (cr1)
365 * - release consumed pa:
370 * - generate in-core bitmap:
374 * - discard all for given object (inode, locality group):
379 * - discard all for given group:
385 * - allocation path (ext4_mb_regular_allocator)
389 static struct kmem_cache *ext4_pspace_cachep;
390 static struct kmem_cache *ext4_ac_cachep;
391 static struct kmem_cache *ext4_free_data_cachep;
393 /* We create slab caches for groupinfo data structures based on the
394 * superblock block size. There will be one per mounted filesystem for
395 * each unique s_blocksize_bits */
396 #define NR_GRPINFO_CACHES 8
397 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
399 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
400 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
401 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
402 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
405 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
407 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
409 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac);
411 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
412 ext4_group_t group, int cr);
414 static int ext4_try_to_trim_range(struct super_block *sb,
415 struct ext4_buddy *e4b, ext4_grpblk_t start,
416 ext4_grpblk_t max, ext4_grpblk_t minblocks);
419 * The algorithm using this percpu seq counter goes below:
420 * 1. We sample the percpu discard_pa_seq counter before trying for block
421 * allocation in ext4_mb_new_blocks().
422 * 2. We increment this percpu discard_pa_seq counter when we either allocate
423 * or free these blocks i.e. while marking those blocks as used/free in
424 * mb_mark_used()/mb_free_blocks().
425 * 3. We also increment this percpu seq counter when we successfully identify
426 * that the bb_prealloc_list is not empty and hence proceed for discarding
427 * of those PAs inside ext4_mb_discard_group_preallocations().
429 * Now to make sure that the regular fast path of block allocation is not
430 * affected, as a small optimization we only sample the percpu seq counter
431 * on that cpu. Only when the block allocation fails and when freed blocks
432 * found were 0, that is when we sample percpu seq counter for all cpus using
433 * below function ext4_get_discard_pa_seq_sum(). This happens after making
434 * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
436 static DEFINE_PER_CPU(u64, discard_pa_seq);
437 static inline u64 ext4_get_discard_pa_seq_sum(void)
442 for_each_possible_cpu(__cpu)
443 __seq += per_cpu(discard_pa_seq, __cpu);
447 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
449 #if BITS_PER_LONG == 64
450 *bit += ((unsigned long) addr & 7UL) << 3;
451 addr = (void *) ((unsigned long) addr & ~7UL);
452 #elif BITS_PER_LONG == 32
453 *bit += ((unsigned long) addr & 3UL) << 3;
454 addr = (void *) ((unsigned long) addr & ~3UL);
456 #error "how many bits you are?!"
461 static inline int mb_test_bit(int bit, void *addr)
464 * ext4_test_bit on architecture like powerpc
465 * needs unsigned long aligned address
467 addr = mb_correct_addr_and_bit(&bit, addr);
468 return ext4_test_bit(bit, addr);
471 static inline void mb_set_bit(int bit, void *addr)
473 addr = mb_correct_addr_and_bit(&bit, addr);
474 ext4_set_bit(bit, addr);
477 static inline void mb_clear_bit(int bit, void *addr)
479 addr = mb_correct_addr_and_bit(&bit, addr);
480 ext4_clear_bit(bit, addr);
483 static inline int mb_test_and_clear_bit(int bit, void *addr)
485 addr = mb_correct_addr_and_bit(&bit, addr);
486 return ext4_test_and_clear_bit(bit, addr);
489 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
491 int fix = 0, ret, tmpmax;
492 addr = mb_correct_addr_and_bit(&fix, addr);
496 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
502 static inline int mb_find_next_bit(void *addr, int max, int start)
504 int fix = 0, ret, tmpmax;
505 addr = mb_correct_addr_and_bit(&fix, addr);
509 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
515 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
519 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
522 if (order > e4b->bd_blkbits + 1) {
527 /* at order 0 we see each particular block */
529 *max = 1 << (e4b->bd_blkbits + 3);
530 return e4b->bd_bitmap;
533 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
534 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
540 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
541 int first, int count)
544 struct super_block *sb = e4b->bd_sb;
546 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
548 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
549 for (i = 0; i < count; i++) {
550 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
551 ext4_fsblk_t blocknr;
553 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
554 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
555 ext4_grp_locked_error(sb, e4b->bd_group,
556 inode ? inode->i_ino : 0,
558 "freeing block already freed "
561 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
562 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
564 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
568 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
572 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
574 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
575 for (i = 0; i < count; i++) {
576 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
577 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
581 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
583 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
585 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
586 unsigned char *b1, *b2;
588 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
589 b2 = (unsigned char *) bitmap;
590 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
591 if (b1[i] != b2[i]) {
592 ext4_msg(e4b->bd_sb, KERN_ERR,
593 "corruption in group %u "
594 "at byte %u(%u): %x in copy != %x "
596 e4b->bd_group, i, i * 8, b1[i], b2[i]);
603 static void mb_group_bb_bitmap_alloc(struct super_block *sb,
604 struct ext4_group_info *grp, ext4_group_t group)
606 struct buffer_head *bh;
608 grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
612 bh = ext4_read_block_bitmap(sb, group);
613 if (IS_ERR_OR_NULL(bh)) {
614 kfree(grp->bb_bitmap);
615 grp->bb_bitmap = NULL;
619 memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
623 static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
625 kfree(grp->bb_bitmap);
629 static inline void mb_free_blocks_double(struct inode *inode,
630 struct ext4_buddy *e4b, int first, int count)
634 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
635 int first, int count)
639 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
644 static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
645 struct ext4_group_info *grp, ext4_group_t group)
650 static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
656 #ifdef AGGRESSIVE_CHECK
658 #define MB_CHECK_ASSERT(assert) \
662 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
663 function, file, line, # assert); \
668 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
669 const char *function, int line)
671 struct super_block *sb = e4b->bd_sb;
672 int order = e4b->bd_blkbits + 1;
679 struct ext4_group_info *grp;
682 struct list_head *cur;
686 if (e4b->bd_info->bb_check_counter++ % 10)
690 buddy = mb_find_buddy(e4b, order, &max);
691 MB_CHECK_ASSERT(buddy);
692 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
693 MB_CHECK_ASSERT(buddy2);
694 MB_CHECK_ASSERT(buddy != buddy2);
695 MB_CHECK_ASSERT(max * 2 == max2);
698 for (i = 0; i < max; i++) {
700 if (mb_test_bit(i, buddy)) {
701 /* only single bit in buddy2 may be 0 */
702 if (!mb_test_bit(i << 1, buddy2)) {
704 mb_test_bit((i<<1)+1, buddy2));
709 /* both bits in buddy2 must be 1 */
710 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
711 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
713 for (j = 0; j < (1 << order); j++) {
714 k = (i * (1 << order)) + j;
716 !mb_test_bit(k, e4b->bd_bitmap));
720 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
725 buddy = mb_find_buddy(e4b, 0, &max);
726 for (i = 0; i < max; i++) {
727 if (!mb_test_bit(i, buddy)) {
728 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
736 /* check used bits only */
737 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
738 buddy2 = mb_find_buddy(e4b, j, &max2);
740 MB_CHECK_ASSERT(k < max2);
741 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
744 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
745 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
747 grp = ext4_get_group_info(sb, e4b->bd_group);
748 list_for_each(cur, &grp->bb_prealloc_list) {
749 ext4_group_t groupnr;
750 struct ext4_prealloc_space *pa;
751 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
752 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
753 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
754 for (i = 0; i < pa->pa_len; i++)
755 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
759 #undef MB_CHECK_ASSERT
760 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
761 __FILE__, __func__, __LINE__)
763 #define mb_check_buddy(e4b)
767 * Divide blocks started from @first with length @len into
768 * smaller chunks with power of 2 blocks.
769 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
770 * then increase bb_counters[] for corresponded chunk size.
772 static void ext4_mb_mark_free_simple(struct super_block *sb,
773 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
774 struct ext4_group_info *grp)
776 struct ext4_sb_info *sbi = EXT4_SB(sb);
782 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
784 border = 2 << sb->s_blocksize_bits;
787 /* find how many blocks can be covered since this position */
788 max = ffs(first | border) - 1;
790 /* find how many blocks of power 2 we need to mark */
797 /* mark multiblock chunks only */
798 grp->bb_counters[min]++;
800 mb_clear_bit(first >> min,
801 buddy + sbi->s_mb_offsets[min]);
808 static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len)
813 * We don't bother with a special lists groups with only 1 block free
814 * extents and for completely empty groups.
816 order = fls(len) - 2;
819 if (order == MB_NUM_ORDERS(sb))
824 /* Move group to appropriate avg_fragment_size list */
826 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
828 struct ext4_sb_info *sbi = EXT4_SB(sb);
831 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0)
834 new_order = mb_avg_fragment_size_order(sb,
835 grp->bb_free / grp->bb_fragments);
836 if (new_order == grp->bb_avg_fragment_size_order)
839 if (grp->bb_avg_fragment_size_order != -1) {
840 write_lock(&sbi->s_mb_avg_fragment_size_locks[
841 grp->bb_avg_fragment_size_order]);
842 list_del(&grp->bb_avg_fragment_size_node);
843 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
844 grp->bb_avg_fragment_size_order]);
846 grp->bb_avg_fragment_size_order = new_order;
847 write_lock(&sbi->s_mb_avg_fragment_size_locks[
848 grp->bb_avg_fragment_size_order]);
849 list_add_tail(&grp->bb_avg_fragment_size_node,
850 &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]);
851 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
852 grp->bb_avg_fragment_size_order]);
856 * Choose next group by traversing largest_free_order lists. Updates *new_cr if
857 * cr level needs an update.
859 static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac,
860 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
862 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
863 struct ext4_group_info *iter, *grp;
866 if (ac->ac_status == AC_STATUS_FOUND)
869 if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMIZED))
870 atomic_inc(&sbi->s_bal_cr0_bad_suggestions);
873 for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) {
874 if (list_empty(&sbi->s_mb_largest_free_orders[i]))
876 read_lock(&sbi->s_mb_largest_free_orders_locks[i]);
877 if (list_empty(&sbi->s_mb_largest_free_orders[i])) {
878 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
882 list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i],
883 bb_largest_free_order_node) {
885 atomic64_inc(&sbi->s_bal_cX_groups_considered[0]);
886 if (likely(ext4_mb_good_group(ac, iter->bb_group, 0))) {
891 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
897 /* Increment cr and search again */
900 *group = grp->bb_group;
901 ac->ac_flags |= EXT4_MB_CR0_OPTIMIZED;
906 * Choose next group by traversing average fragment size list of suitable
907 * order. Updates *new_cr if cr level needs an update.
909 static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac,
910 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
912 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
913 struct ext4_group_info *grp, *iter;
916 if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) {
918 atomic_inc(&sbi->s_bal_cr1_bad_suggestions);
921 for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len);
922 i < MB_NUM_ORDERS(ac->ac_sb); i++) {
923 if (list_empty(&sbi->s_mb_avg_fragment_size[i]))
925 read_lock(&sbi->s_mb_avg_fragment_size_locks[i]);
926 if (list_empty(&sbi->s_mb_avg_fragment_size[i])) {
927 read_unlock(&sbi->s_mb_avg_fragment_size_locks[i]);
931 list_for_each_entry(iter, &sbi->s_mb_avg_fragment_size[i],
932 bb_avg_fragment_size_node) {
934 atomic64_inc(&sbi->s_bal_cX_groups_considered[1]);
935 if (likely(ext4_mb_good_group(ac, iter->bb_group, 1))) {
940 read_unlock(&sbi->s_mb_avg_fragment_size_locks[i]);
946 *group = grp->bb_group;
947 ac->ac_flags |= EXT4_MB_CR1_OPTIMIZED;
953 static inline int should_optimize_scan(struct ext4_allocation_context *ac)
955 if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN)))
957 if (ac->ac_criteria >= 2)
959 if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))
965 * Return next linear group for allocation. If linear traversal should not be
966 * performed, this function just returns the same group
969 next_linear_group(struct ext4_allocation_context *ac, int group, int ngroups)
971 if (!should_optimize_scan(ac))
974 if (ac->ac_groups_linear_remaining) {
975 ac->ac_groups_linear_remaining--;
982 * Artificially restricted ngroups for non-extent
983 * files makes group > ngroups possible on first loop.
985 return group + 1 >= ngroups ? 0 : group + 1;
989 * ext4_mb_choose_next_group: choose next group for allocation.
991 * @ac Allocation Context
992 * @new_cr This is an output parameter. If the there is no good group
993 * available at current CR level, this field is updated to indicate
994 * the new cr level that should be used.
995 * @group This is an input / output parameter. As an input it indicates the
996 * next group that the allocator intends to use for allocation. As
997 * output, this field indicates the next group that should be used as
998 * determined by the optimization functions.
999 * @ngroups Total number of groups
1001 static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac,
1002 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
1004 *new_cr = ac->ac_criteria;
1006 if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining) {
1007 *group = next_linear_group(ac, *group, ngroups);
1012 ext4_mb_choose_next_group_cr0(ac, new_cr, group, ngroups);
1013 } else if (*new_cr == 1) {
1014 ext4_mb_choose_next_group_cr1(ac, new_cr, group, ngroups);
1017 * TODO: For CR=2, we can arrange groups in an rb tree sorted by
1018 * bb_free. But until that happens, we should never come here.
1025 * Cache the order of the largest free extent we have available in this block
1029 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
1031 struct ext4_sb_info *sbi = EXT4_SB(sb);
1034 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--)
1035 if (grp->bb_counters[i] > 0)
1037 /* No need to move between order lists? */
1038 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) ||
1039 i == grp->bb_largest_free_order) {
1040 grp->bb_largest_free_order = i;
1044 if (grp->bb_largest_free_order >= 0) {
1045 write_lock(&sbi->s_mb_largest_free_orders_locks[
1046 grp->bb_largest_free_order]);
1047 list_del_init(&grp->bb_largest_free_order_node);
1048 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1049 grp->bb_largest_free_order]);
1051 grp->bb_largest_free_order = i;
1052 if (grp->bb_largest_free_order >= 0 && grp->bb_free) {
1053 write_lock(&sbi->s_mb_largest_free_orders_locks[
1054 grp->bb_largest_free_order]);
1055 list_add_tail(&grp->bb_largest_free_order_node,
1056 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1057 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1058 grp->bb_largest_free_order]);
1062 static noinline_for_stack
1063 void ext4_mb_generate_buddy(struct super_block *sb,
1064 void *buddy, void *bitmap, ext4_group_t group)
1066 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
1067 struct ext4_sb_info *sbi = EXT4_SB(sb);
1068 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1069 ext4_grpblk_t i = 0;
1070 ext4_grpblk_t first;
1073 unsigned fragments = 0;
1074 unsigned long long period = get_cycles();
1076 /* initialize buddy from bitmap which is aggregation
1077 * of on-disk bitmap and preallocations */
1078 i = mb_find_next_zero_bit(bitmap, max, 0);
1079 grp->bb_first_free = i;
1083 i = mb_find_next_bit(bitmap, max, i);
1087 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1089 grp->bb_counters[0]++;
1091 i = mb_find_next_zero_bit(bitmap, max, i);
1093 grp->bb_fragments = fragments;
1095 if (free != grp->bb_free) {
1096 ext4_grp_locked_error(sb, group, 0, 0,
1097 "block bitmap and bg descriptor "
1098 "inconsistent: %u vs %u free clusters",
1099 free, grp->bb_free);
1101 * If we intend to continue, we consider group descriptor
1102 * corrupt and update bb_free using bitmap value
1104 grp->bb_free = free;
1105 ext4_mark_group_bitmap_corrupted(sb, group,
1106 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1108 mb_set_largest_free_order(sb, grp);
1109 mb_update_avg_fragment_size(sb, grp);
1111 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1113 period = get_cycles() - period;
1114 atomic_inc(&sbi->s_mb_buddies_generated);
1115 atomic64_add(period, &sbi->s_mb_generation_time);
1118 /* The buddy information is attached the buddy cache inode
1119 * for convenience. The information regarding each group
1120 * is loaded via ext4_mb_load_buddy. The information involve
1121 * block bitmap and buddy information. The information are
1122 * stored in the inode as
1125 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1128 * one block each for bitmap and buddy information.
1129 * So for each group we take up 2 blocks. A page can
1130 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
1131 * So it can have information regarding groups_per_page which
1132 * is blocks_per_page/2
1134 * Locking note: This routine takes the block group lock of all groups
1135 * for this page; do not hold this lock when calling this routine!
1138 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
1140 ext4_group_t ngroups;
1142 int blocks_per_page;
1143 int groups_per_page;
1146 ext4_group_t first_group, group;
1148 struct super_block *sb;
1149 struct buffer_head *bhs;
1150 struct buffer_head **bh = NULL;
1151 struct inode *inode;
1154 struct ext4_group_info *grinfo;
1156 inode = page->mapping->host;
1158 ngroups = ext4_get_groups_count(sb);
1159 blocksize = i_blocksize(inode);
1160 blocks_per_page = PAGE_SIZE / blocksize;
1162 mb_debug(sb, "init page %lu\n", page->index);
1164 groups_per_page = blocks_per_page >> 1;
1165 if (groups_per_page == 0)
1166 groups_per_page = 1;
1168 /* allocate buffer_heads to read bitmaps */
1169 if (groups_per_page > 1) {
1170 i = sizeof(struct buffer_head *) * groups_per_page;
1171 bh = kzalloc(i, gfp);
1179 first_group = page->index * blocks_per_page / 2;
1181 /* read all groups the page covers into the cache */
1182 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1183 if (group >= ngroups)
1186 grinfo = ext4_get_group_info(sb, group);
1188 * If page is uptodate then we came here after online resize
1189 * which added some new uninitialized group info structs, so
1190 * we must skip all initialized uptodate buddies on the page,
1191 * which may be currently in use by an allocating task.
1193 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1197 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1198 if (IS_ERR(bh[i])) {
1199 err = PTR_ERR(bh[i]);
1203 mb_debug(sb, "read bitmap for group %u\n", group);
1206 /* wait for I/O completion */
1207 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1212 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1217 first_block = page->index * blocks_per_page;
1218 for (i = 0; i < blocks_per_page; i++) {
1219 group = (first_block + i) >> 1;
1220 if (group >= ngroups)
1223 if (!bh[group - first_group])
1224 /* skip initialized uptodate buddy */
1227 if (!buffer_verified(bh[group - first_group]))
1228 /* Skip faulty bitmaps */
1233 * data carry information regarding this
1234 * particular group in the format specified
1238 data = page_address(page) + (i * blocksize);
1239 bitmap = bh[group - first_group]->b_data;
1242 * We place the buddy block and bitmap block
1245 if ((first_block + i) & 1) {
1246 /* this is block of buddy */
1247 BUG_ON(incore == NULL);
1248 mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
1249 group, page->index, i * blocksize);
1250 trace_ext4_mb_buddy_bitmap_load(sb, group);
1251 grinfo = ext4_get_group_info(sb, group);
1252 grinfo->bb_fragments = 0;
1253 memset(grinfo->bb_counters, 0,
1254 sizeof(*grinfo->bb_counters) *
1255 (MB_NUM_ORDERS(sb)));
1257 * incore got set to the group block bitmap below
1259 ext4_lock_group(sb, group);
1260 /* init the buddy */
1261 memset(data, 0xff, blocksize);
1262 ext4_mb_generate_buddy(sb, data, incore, group);
1263 ext4_unlock_group(sb, group);
1266 /* this is block of bitmap */
1267 BUG_ON(incore != NULL);
1268 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
1269 group, page->index, i * blocksize);
1270 trace_ext4_mb_bitmap_load(sb, group);
1272 /* see comments in ext4_mb_put_pa() */
1273 ext4_lock_group(sb, group);
1274 memcpy(data, bitmap, blocksize);
1276 /* mark all preallocated blks used in in-core bitmap */
1277 ext4_mb_generate_from_pa(sb, data, group);
1278 ext4_mb_generate_from_freelist(sb, data, group);
1279 ext4_unlock_group(sb, group);
1281 /* set incore so that the buddy information can be
1282 * generated using this
1287 SetPageUptodate(page);
1291 for (i = 0; i < groups_per_page; i++)
1300 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1301 * on the same buddy page doesn't happen whild holding the buddy page lock.
1302 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1303 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1305 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1306 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1308 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1309 int block, pnum, poff;
1310 int blocks_per_page;
1313 e4b->bd_buddy_page = NULL;
1314 e4b->bd_bitmap_page = NULL;
1316 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1318 * the buddy cache inode stores the block bitmap
1319 * and buddy information in consecutive blocks.
1320 * So for each group we need two blocks.
1323 pnum = block / blocks_per_page;
1324 poff = block % blocks_per_page;
1325 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1328 BUG_ON(page->mapping != inode->i_mapping);
1329 e4b->bd_bitmap_page = page;
1330 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1332 if (blocks_per_page >= 2) {
1333 /* buddy and bitmap are on the same page */
1338 pnum = block / blocks_per_page;
1339 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1342 BUG_ON(page->mapping != inode->i_mapping);
1343 e4b->bd_buddy_page = page;
1347 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1349 if (e4b->bd_bitmap_page) {
1350 unlock_page(e4b->bd_bitmap_page);
1351 put_page(e4b->bd_bitmap_page);
1353 if (e4b->bd_buddy_page) {
1354 unlock_page(e4b->bd_buddy_page);
1355 put_page(e4b->bd_buddy_page);
1360 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1361 * block group lock of all groups for this page; do not hold the BG lock when
1362 * calling this routine!
1364 static noinline_for_stack
1365 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1368 struct ext4_group_info *this_grp;
1369 struct ext4_buddy e4b;
1374 mb_debug(sb, "init group %u\n", group);
1375 this_grp = ext4_get_group_info(sb, group);
1377 * This ensures that we don't reinit the buddy cache
1378 * page which map to the group from which we are already
1379 * allocating. If we are looking at the buddy cache we would
1380 * have taken a reference using ext4_mb_load_buddy and that
1381 * would have pinned buddy page to page cache.
1382 * The call to ext4_mb_get_buddy_page_lock will mark the
1385 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1386 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1388 * somebody initialized the group
1389 * return without doing anything
1394 page = e4b.bd_bitmap_page;
1395 ret = ext4_mb_init_cache(page, NULL, gfp);
1398 if (!PageUptodate(page)) {
1403 if (e4b.bd_buddy_page == NULL) {
1405 * If both the bitmap and buddy are in
1406 * the same page we don't need to force
1412 /* init buddy cache */
1413 page = e4b.bd_buddy_page;
1414 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1417 if (!PageUptodate(page)) {
1422 ext4_mb_put_buddy_page_lock(&e4b);
1427 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1428 * block group lock of all groups for this page; do not hold the BG lock when
1429 * calling this routine!
1431 static noinline_for_stack int
1432 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1433 struct ext4_buddy *e4b, gfp_t gfp)
1435 int blocks_per_page;
1441 struct ext4_group_info *grp;
1442 struct ext4_sb_info *sbi = EXT4_SB(sb);
1443 struct inode *inode = sbi->s_buddy_cache;
1446 mb_debug(sb, "load group %u\n", group);
1448 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1449 grp = ext4_get_group_info(sb, group);
1451 e4b->bd_blkbits = sb->s_blocksize_bits;
1454 e4b->bd_group = group;
1455 e4b->bd_buddy_page = NULL;
1456 e4b->bd_bitmap_page = NULL;
1458 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1460 * we need full data about the group
1461 * to make a good selection
1463 ret = ext4_mb_init_group(sb, group, gfp);
1469 * the buddy cache inode stores the block bitmap
1470 * and buddy information in consecutive blocks.
1471 * So for each group we need two blocks.
1474 pnum = block / blocks_per_page;
1475 poff = block % blocks_per_page;
1477 /* we could use find_or_create_page(), but it locks page
1478 * what we'd like to avoid in fast path ... */
1479 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1480 if (page == NULL || !PageUptodate(page)) {
1483 * drop the page reference and try
1484 * to get the page with lock. If we
1485 * are not uptodate that implies
1486 * somebody just created the page but
1487 * is yet to initialize the same. So
1488 * wait for it to initialize.
1491 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1493 BUG_ON(page->mapping != inode->i_mapping);
1494 if (!PageUptodate(page)) {
1495 ret = ext4_mb_init_cache(page, NULL, gfp);
1500 mb_cmp_bitmaps(e4b, page_address(page) +
1501 (poff * sb->s_blocksize));
1510 if (!PageUptodate(page)) {
1515 /* Pages marked accessed already */
1516 e4b->bd_bitmap_page = page;
1517 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1520 pnum = block / blocks_per_page;
1521 poff = block % blocks_per_page;
1523 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1524 if (page == NULL || !PageUptodate(page)) {
1527 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1529 BUG_ON(page->mapping != inode->i_mapping);
1530 if (!PageUptodate(page)) {
1531 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1545 if (!PageUptodate(page)) {
1550 /* Pages marked accessed already */
1551 e4b->bd_buddy_page = page;
1552 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1559 if (e4b->bd_bitmap_page)
1560 put_page(e4b->bd_bitmap_page);
1561 if (e4b->bd_buddy_page)
1562 put_page(e4b->bd_buddy_page);
1563 e4b->bd_buddy = NULL;
1564 e4b->bd_bitmap = NULL;
1568 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1569 struct ext4_buddy *e4b)
1571 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1574 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1576 if (e4b->bd_bitmap_page)
1577 put_page(e4b->bd_bitmap_page);
1578 if (e4b->bd_buddy_page)
1579 put_page(e4b->bd_buddy_page);
1583 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1588 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1589 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1591 while (order <= e4b->bd_blkbits + 1) {
1592 bb = mb_find_buddy(e4b, order, &max);
1593 if (!mb_test_bit(block >> order, bb)) {
1594 /* this block is part of buddy of order 'order' */
1602 static void mb_clear_bits(void *bm, int cur, int len)
1608 if ((cur & 31) == 0 && (len - cur) >= 32) {
1609 /* fast path: clear whole word at once */
1610 addr = bm + (cur >> 3);
1615 mb_clear_bit(cur, bm);
1620 /* clear bits in given range
1621 * will return first found zero bit if any, -1 otherwise
1623 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1630 if ((cur & 31) == 0 && (len - cur) >= 32) {
1631 /* fast path: clear whole word at once */
1632 addr = bm + (cur >> 3);
1633 if (*addr != (__u32)(-1) && zero_bit == -1)
1634 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1639 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1647 void mb_set_bits(void *bm, int cur, int len)
1653 if ((cur & 31) == 0 && (len - cur) >= 32) {
1654 /* fast path: set whole word at once */
1655 addr = bm + (cur >> 3);
1660 mb_set_bit(cur, bm);
1665 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1667 if (mb_test_bit(*bit + side, bitmap)) {
1668 mb_clear_bit(*bit, bitmap);
1674 mb_set_bit(*bit, bitmap);
1679 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1683 void *buddy = mb_find_buddy(e4b, order, &max);
1688 /* Bits in range [first; last] are known to be set since
1689 * corresponding blocks were allocated. Bits in range
1690 * (first; last) will stay set because they form buddies on
1691 * upper layer. We just deal with borders if they don't
1692 * align with upper layer and then go up.
1693 * Releasing entire group is all about clearing
1694 * single bit of highest order buddy.
1698 * ---------------------------------
1700 * ---------------------------------
1701 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1702 * ---------------------------------
1704 * \_____________________/
1706 * Neither [1] nor [6] is aligned to above layer.
1707 * Left neighbour [0] is free, so mark it busy,
1708 * decrease bb_counters and extend range to
1710 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1711 * mark [6] free, increase bb_counters and shrink range to
1713 * Then shift range to [0; 2], go up and do the same.
1718 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1720 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1725 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1726 mb_clear_bits(buddy, first, last - first + 1);
1727 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1736 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1737 int first, int count)
1739 int left_is_free = 0;
1740 int right_is_free = 0;
1742 int last = first + count - 1;
1743 struct super_block *sb = e4b->bd_sb;
1745 if (WARN_ON(count == 0))
1747 BUG_ON(last >= (sb->s_blocksize << 3));
1748 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1749 /* Don't bother if the block group is corrupt. */
1750 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1753 mb_check_buddy(e4b);
1754 mb_free_blocks_double(inode, e4b, first, count);
1756 this_cpu_inc(discard_pa_seq);
1757 e4b->bd_info->bb_free += count;
1758 if (first < e4b->bd_info->bb_first_free)
1759 e4b->bd_info->bb_first_free = first;
1761 /* access memory sequentially: check left neighbour,
1762 * clear range and then check right neighbour
1765 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1766 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1767 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1768 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1770 if (unlikely(block != -1)) {
1771 struct ext4_sb_info *sbi = EXT4_SB(sb);
1772 ext4_fsblk_t blocknr;
1774 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1775 blocknr += EXT4_C2B(sbi, block);
1776 if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) {
1777 ext4_grp_locked_error(sb, e4b->bd_group,
1778 inode ? inode->i_ino : 0,
1780 "freeing already freed block (bit %u); block bitmap corrupt.",
1782 ext4_mark_group_bitmap_corrupted(
1784 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1789 /* let's maintain fragments counter */
1790 if (left_is_free && right_is_free)
1791 e4b->bd_info->bb_fragments--;
1792 else if (!left_is_free && !right_is_free)
1793 e4b->bd_info->bb_fragments++;
1795 /* buddy[0] == bd_bitmap is a special case, so handle
1796 * it right away and let mb_buddy_mark_free stay free of
1797 * zero order checks.
1798 * Check if neighbours are to be coaleasced,
1799 * adjust bitmap bb_counters and borders appropriately.
1802 first += !left_is_free;
1803 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1806 last -= !right_is_free;
1807 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1811 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1814 mb_set_largest_free_order(sb, e4b->bd_info);
1815 mb_update_avg_fragment_size(sb, e4b->bd_info);
1816 mb_check_buddy(e4b);
1819 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1820 int needed, struct ext4_free_extent *ex)
1826 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1829 buddy = mb_find_buddy(e4b, 0, &max);
1830 BUG_ON(buddy == NULL);
1831 BUG_ON(block >= max);
1832 if (mb_test_bit(block, buddy)) {
1839 /* find actual order */
1840 order = mb_find_order_for_block(e4b, block);
1841 block = block >> order;
1843 ex->fe_len = 1 << order;
1844 ex->fe_start = block << order;
1845 ex->fe_group = e4b->bd_group;
1847 /* calc difference from given start */
1848 next = next - ex->fe_start;
1850 ex->fe_start += next;
1852 while (needed > ex->fe_len &&
1853 mb_find_buddy(e4b, order, &max)) {
1855 if (block + 1 >= max)
1858 next = (block + 1) * (1 << order);
1859 if (mb_test_bit(next, e4b->bd_bitmap))
1862 order = mb_find_order_for_block(e4b, next);
1864 block = next >> order;
1865 ex->fe_len += 1 << order;
1868 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1869 /* Should never happen! (but apparently sometimes does?!?) */
1871 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
1872 "corruption or bug in mb_find_extent "
1873 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1874 block, order, needed, ex->fe_group, ex->fe_start,
1875 ex->fe_len, ex->fe_logical);
1883 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1889 int start = ex->fe_start;
1890 int len = ex->fe_len;
1895 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1896 BUG_ON(e4b->bd_group != ex->fe_group);
1897 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1898 mb_check_buddy(e4b);
1899 mb_mark_used_double(e4b, start, len);
1901 this_cpu_inc(discard_pa_seq);
1902 e4b->bd_info->bb_free -= len;
1903 if (e4b->bd_info->bb_first_free == start)
1904 e4b->bd_info->bb_first_free += len;
1906 /* let's maintain fragments counter */
1908 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1909 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1910 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1912 e4b->bd_info->bb_fragments++;
1913 else if (!mlen && !max)
1914 e4b->bd_info->bb_fragments--;
1916 /* let's maintain buddy itself */
1918 ord = mb_find_order_for_block(e4b, start);
1920 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1921 /* the whole chunk may be allocated at once! */
1923 buddy = mb_find_buddy(e4b, ord, &max);
1924 BUG_ON((start >> ord) >= max);
1925 mb_set_bit(start >> ord, buddy);
1926 e4b->bd_info->bb_counters[ord]--;
1933 /* store for history */
1935 ret = len | (ord << 16);
1937 /* we have to split large buddy */
1939 buddy = mb_find_buddy(e4b, ord, &max);
1940 mb_set_bit(start >> ord, buddy);
1941 e4b->bd_info->bb_counters[ord]--;
1944 cur = (start >> ord) & ~1U;
1945 buddy = mb_find_buddy(e4b, ord, &max);
1946 mb_clear_bit(cur, buddy);
1947 mb_clear_bit(cur + 1, buddy);
1948 e4b->bd_info->bb_counters[ord]++;
1949 e4b->bd_info->bb_counters[ord]++;
1951 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1953 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
1954 mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1955 mb_check_buddy(e4b);
1961 * Must be called under group lock!
1963 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1964 struct ext4_buddy *e4b)
1966 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1969 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1970 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1972 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1973 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1974 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1976 /* preallocation can change ac_b_ex, thus we store actually
1977 * allocated blocks for history */
1978 ac->ac_f_ex = ac->ac_b_ex;
1980 ac->ac_status = AC_STATUS_FOUND;
1981 ac->ac_tail = ret & 0xffff;
1982 ac->ac_buddy = ret >> 16;
1985 * take the page reference. We want the page to be pinned
1986 * so that we don't get a ext4_mb_init_cache_call for this
1987 * group until we update the bitmap. That would mean we
1988 * double allocate blocks. The reference is dropped
1989 * in ext4_mb_release_context
1991 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1992 get_page(ac->ac_bitmap_page);
1993 ac->ac_buddy_page = e4b->bd_buddy_page;
1994 get_page(ac->ac_buddy_page);
1995 /* store last allocated for subsequent stream allocation */
1996 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1997 spin_lock(&sbi->s_md_lock);
1998 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1999 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2000 spin_unlock(&sbi->s_md_lock);
2003 * As we've just preallocated more space than
2004 * user requested originally, we store allocated
2005 * space in a special descriptor.
2007 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2008 ext4_mb_new_preallocation(ac);
2012 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2013 struct ext4_buddy *e4b,
2016 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2017 struct ext4_free_extent *bex = &ac->ac_b_ex;
2018 struct ext4_free_extent *gex = &ac->ac_g_ex;
2019 struct ext4_free_extent ex;
2022 if (ac->ac_status == AC_STATUS_FOUND)
2025 * We don't want to scan for a whole year
2027 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2028 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2029 ac->ac_status = AC_STATUS_BREAK;
2034 * Haven't found good chunk so far, let's continue
2036 if (bex->fe_len < gex->fe_len)
2039 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2040 && bex->fe_group == e4b->bd_group) {
2041 /* recheck chunk's availability - we don't know
2042 * when it was found (within this lock-unlock
2044 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
2045 if (max >= gex->fe_len) {
2046 ext4_mb_use_best_found(ac, e4b);
2053 * The routine checks whether found extent is good enough. If it is,
2054 * then the extent gets marked used and flag is set to the context
2055 * to stop scanning. Otherwise, the extent is compared with the
2056 * previous found extent and if new one is better, then it's stored
2057 * in the context. Later, the best found extent will be used, if
2058 * mballoc can't find good enough extent.
2060 * FIXME: real allocation policy is to be designed yet!
2062 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2063 struct ext4_free_extent *ex,
2064 struct ext4_buddy *e4b)
2066 struct ext4_free_extent *bex = &ac->ac_b_ex;
2067 struct ext4_free_extent *gex = &ac->ac_g_ex;
2069 BUG_ON(ex->fe_len <= 0);
2070 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2071 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2072 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2077 * The special case - take what you catch first
2079 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2081 ext4_mb_use_best_found(ac, e4b);
2086 * Let's check whether the chuck is good enough
2088 if (ex->fe_len == gex->fe_len) {
2090 ext4_mb_use_best_found(ac, e4b);
2095 * If this is first found extent, just store it in the context
2097 if (bex->fe_len == 0) {
2103 * If new found extent is better, store it in the context
2105 if (bex->fe_len < gex->fe_len) {
2106 /* if the request isn't satisfied, any found extent
2107 * larger than previous best one is better */
2108 if (ex->fe_len > bex->fe_len)
2110 } else if (ex->fe_len > gex->fe_len) {
2111 /* if the request is satisfied, then we try to find
2112 * an extent that still satisfy the request, but is
2113 * smaller than previous one */
2114 if (ex->fe_len < bex->fe_len)
2118 ext4_mb_check_limits(ac, e4b, 0);
2121 static noinline_for_stack
2122 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2123 struct ext4_buddy *e4b)
2125 struct ext4_free_extent ex = ac->ac_b_ex;
2126 ext4_group_t group = ex.fe_group;
2130 BUG_ON(ex.fe_len <= 0);
2131 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2135 ext4_lock_group(ac->ac_sb, group);
2136 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2140 ext4_mb_use_best_found(ac, e4b);
2143 ext4_unlock_group(ac->ac_sb, group);
2144 ext4_mb_unload_buddy(e4b);
2149 static noinline_for_stack
2150 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2151 struct ext4_buddy *e4b)
2153 ext4_group_t group = ac->ac_g_ex.fe_group;
2156 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2157 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2158 struct ext4_free_extent ex;
2160 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
2162 if (grp->bb_free == 0)
2165 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2169 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
2170 ext4_mb_unload_buddy(e4b);
2174 ext4_lock_group(ac->ac_sb, group);
2175 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2176 ac->ac_g_ex.fe_len, &ex);
2177 ex.fe_logical = 0xDEADFA11; /* debug value */
2179 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
2182 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
2184 /* use do_div to get remainder (would be 64-bit modulo) */
2185 if (do_div(start, sbi->s_stripe) == 0) {
2188 ext4_mb_use_best_found(ac, e4b);
2190 } else if (max >= ac->ac_g_ex.fe_len) {
2191 BUG_ON(ex.fe_len <= 0);
2192 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2193 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2196 ext4_mb_use_best_found(ac, e4b);
2197 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2198 /* Sometimes, caller may want to merge even small
2199 * number of blocks to an existing extent */
2200 BUG_ON(ex.fe_len <= 0);
2201 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2202 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2205 ext4_mb_use_best_found(ac, e4b);
2207 ext4_unlock_group(ac->ac_sb, group);
2208 ext4_mb_unload_buddy(e4b);
2214 * The routine scans buddy structures (not bitmap!) from given order
2215 * to max order and tries to find big enough chunk to satisfy the req
2217 static noinline_for_stack
2218 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2219 struct ext4_buddy *e4b)
2221 struct super_block *sb = ac->ac_sb;
2222 struct ext4_group_info *grp = e4b->bd_info;
2228 BUG_ON(ac->ac_2order <= 0);
2229 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2230 if (grp->bb_counters[i] == 0)
2233 buddy = mb_find_buddy(e4b, i, &max);
2234 BUG_ON(buddy == NULL);
2236 k = mb_find_next_zero_bit(buddy, max, 0);
2238 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2239 "%d free clusters of order %d. But found 0",
2240 grp->bb_counters[i], i);
2241 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2243 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2248 ac->ac_b_ex.fe_len = 1 << i;
2249 ac->ac_b_ex.fe_start = k << i;
2250 ac->ac_b_ex.fe_group = e4b->bd_group;
2252 ext4_mb_use_best_found(ac, e4b);
2254 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2256 if (EXT4_SB(sb)->s_mb_stats)
2257 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2264 * The routine scans the group and measures all found extents.
2265 * In order to optimize scanning, caller must pass number of
2266 * free blocks in the group, so the routine can know upper limit.
2268 static noinline_for_stack
2269 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2270 struct ext4_buddy *e4b)
2272 struct super_block *sb = ac->ac_sb;
2273 void *bitmap = e4b->bd_bitmap;
2274 struct ext4_free_extent ex;
2278 free = e4b->bd_info->bb_free;
2279 if (WARN_ON(free <= 0))
2282 i = e4b->bd_info->bb_first_free;
2284 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2285 i = mb_find_next_zero_bit(bitmap,
2286 EXT4_CLUSTERS_PER_GROUP(sb), i);
2287 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2289 * IF we have corrupt bitmap, we won't find any
2290 * free blocks even though group info says we
2293 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2294 "%d free clusters as per "
2295 "group info. But bitmap says 0",
2297 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2298 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2302 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2303 if (WARN_ON(ex.fe_len <= 0))
2305 if (free < ex.fe_len) {
2306 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2307 "%d free clusters as per "
2308 "group info. But got %d blocks",
2310 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2311 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2313 * The number of free blocks differs. This mostly
2314 * indicate that the bitmap is corrupt. So exit
2315 * without claiming the space.
2319 ex.fe_logical = 0xDEADC0DE; /* debug value */
2320 ext4_mb_measure_extent(ac, &ex, e4b);
2326 ext4_mb_check_limits(ac, e4b, 1);
2330 * This is a special case for storages like raid5
2331 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2333 static noinline_for_stack
2334 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2335 struct ext4_buddy *e4b)
2337 struct super_block *sb = ac->ac_sb;
2338 struct ext4_sb_info *sbi = EXT4_SB(sb);
2339 void *bitmap = e4b->bd_bitmap;
2340 struct ext4_free_extent ex;
2341 ext4_fsblk_t first_group_block;
2346 BUG_ON(sbi->s_stripe == 0);
2348 /* find first stripe-aligned block in group */
2349 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2351 a = first_group_block + sbi->s_stripe - 1;
2352 do_div(a, sbi->s_stripe);
2353 i = (a * sbi->s_stripe) - first_group_block;
2355 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2356 if (!mb_test_bit(i, bitmap)) {
2357 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2358 if (max >= sbi->s_stripe) {
2360 ex.fe_logical = 0xDEADF00D; /* debug value */
2362 ext4_mb_use_best_found(ac, e4b);
2371 * This is also called BEFORE we load the buddy bitmap.
2372 * Returns either 1 or 0 indicating that the group is either suitable
2373 * for the allocation or not.
2375 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2376 ext4_group_t group, int cr)
2378 ext4_grpblk_t free, fragments;
2379 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2380 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2382 BUG_ON(cr < 0 || cr >= 4);
2384 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2387 free = grp->bb_free;
2391 fragments = grp->bb_fragments;
2397 BUG_ON(ac->ac_2order == 0);
2399 /* Avoid using the first bg of a flexgroup for data files */
2400 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2401 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2402 ((group % flex_size) == 0))
2405 if (free < ac->ac_g_ex.fe_len)
2408 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2411 if (grp->bb_largest_free_order < ac->ac_2order)
2416 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2420 if (free >= ac->ac_g_ex.fe_len)
2433 * This could return negative error code if something goes wrong
2434 * during ext4_mb_init_group(). This should not be called with
2435 * ext4_lock_group() held.
2437 * Note: because we are conditionally operating with the group lock in
2438 * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this
2439 * function using __acquire and __release. This means we need to be
2440 * super careful before messing with the error path handling via "goto
2443 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2444 ext4_group_t group, int cr)
2446 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2447 struct super_block *sb = ac->ac_sb;
2448 struct ext4_sb_info *sbi = EXT4_SB(sb);
2449 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2453 if (sbi->s_mb_stats)
2454 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2456 ext4_lock_group(sb, group);
2457 __release(ext4_group_lock_ptr(sb, group));
2459 free = grp->bb_free;
2462 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2464 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2467 __acquire(ext4_group_lock_ptr(sb, group));
2468 ext4_unlock_group(sb, group);
2471 /* We only do this if the grp has never been initialized */
2472 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2473 struct ext4_group_desc *gdp =
2474 ext4_get_group_desc(sb, group, NULL);
2477 /* cr=0/1 is a very optimistic search to find large
2478 * good chunks almost for free. If buddy data is not
2479 * ready, then this optimization makes no sense. But
2480 * we never skip the first block group in a flex_bg,
2481 * since this gets used for metadata block allocation,
2482 * and we want to make sure we locate metadata blocks
2483 * in the first block group in the flex_bg if possible.
2486 (!sbi->s_log_groups_per_flex ||
2487 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2488 !(ext4_has_group_desc_csum(sb) &&
2489 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2491 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2497 ext4_lock_group(sb, group);
2498 __release(ext4_group_lock_ptr(sb, group));
2500 ret = ext4_mb_good_group(ac, group, cr);
2503 __acquire(ext4_group_lock_ptr(sb, group));
2504 ext4_unlock_group(sb, group);
2510 * Start prefetching @nr block bitmaps starting at @group.
2511 * Return the next group which needs to be prefetched.
2513 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2514 unsigned int nr, int *cnt)
2516 ext4_group_t ngroups = ext4_get_groups_count(sb);
2517 struct buffer_head *bh;
2518 struct blk_plug plug;
2520 blk_start_plug(&plug);
2522 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2524 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2527 * Prefetch block groups with free blocks; but don't
2528 * bother if it is marked uninitialized on disk, since
2529 * it won't require I/O to read. Also only try to
2530 * prefetch once, so we avoid getblk() call, which can
2533 if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2534 EXT4_MB_GRP_NEED_INIT(grp) &&
2535 ext4_free_group_clusters(sb, gdp) > 0 &&
2536 !(ext4_has_group_desc_csum(sb) &&
2537 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2538 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2539 if (bh && !IS_ERR(bh)) {
2540 if (!buffer_uptodate(bh) && cnt)
2545 if (++group >= ngroups)
2548 blk_finish_plug(&plug);
2553 * Prefetching reads the block bitmap into the buffer cache; but we
2554 * need to make sure that the buddy bitmap in the page cache has been
2555 * initialized. Note that ext4_mb_init_group() will block if the I/O
2556 * is not yet completed, or indeed if it was not initiated by
2557 * ext4_mb_prefetch did not start the I/O.
2559 * TODO: We should actually kick off the buddy bitmap setup in a work
2560 * queue when the buffer I/O is completed, so that we don't block
2561 * waiting for the block allocation bitmap read to finish when
2562 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2564 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2568 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2570 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2573 group = ext4_get_groups_count(sb);
2575 grp = ext4_get_group_info(sb, group);
2577 if (EXT4_MB_GRP_NEED_INIT(grp) &&
2578 ext4_free_group_clusters(sb, gdp) > 0 &&
2579 !(ext4_has_group_desc_csum(sb) &&
2580 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2581 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2587 static noinline_for_stack int
2588 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2590 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2591 int cr = -1, new_cr;
2592 int err = 0, first_err = 0;
2593 unsigned int nr = 0, prefetch_ios = 0;
2594 struct ext4_sb_info *sbi;
2595 struct super_block *sb;
2596 struct ext4_buddy e4b;
2601 ngroups = ext4_get_groups_count(sb);
2602 /* non-extent files are limited to low blocks/groups */
2603 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2604 ngroups = sbi->s_blockfile_groups;
2606 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2608 /* first, try the goal */
2609 err = ext4_mb_find_by_goal(ac, &e4b);
2610 if (err || ac->ac_status == AC_STATUS_FOUND)
2613 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2617 * ac->ac_2order is set only if the fe_len is a power of 2
2618 * if ac->ac_2order is set we also set criteria to 0 so that we
2619 * try exact allocation using buddy.
2621 i = fls(ac->ac_g_ex.fe_len);
2624 * We search using buddy data only if the order of the request
2625 * is greater than equal to the sbi_s_mb_order2_reqs
2626 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2627 * We also support searching for power-of-two requests only for
2628 * requests upto maximum buddy size we have constructed.
2630 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2632 * This should tell if fe_len is exactly power of 2
2634 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2635 ac->ac_2order = array_index_nospec(i - 1,
2639 /* if stream allocation is enabled, use global goal */
2640 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2641 /* TBD: may be hot point */
2642 spin_lock(&sbi->s_md_lock);
2643 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2644 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2645 spin_unlock(&sbi->s_md_lock);
2648 /* Let's just scan groups to find more-less suitable blocks */
2649 cr = ac->ac_2order ? 0 : 1;
2651 * cr == 0 try to get exact allocation,
2652 * cr == 3 try to get anything
2655 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2656 ac->ac_criteria = cr;
2658 * searching for the right group start
2659 * from the goal value specified
2661 group = ac->ac_g_ex.fe_group;
2662 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2663 prefetch_grp = group;
2665 for (i = 0, new_cr = cr; i < ngroups; i++,
2666 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) {
2676 * Batch reads of the block allocation bitmaps
2677 * to get multiple READs in flight; limit
2678 * prefetching at cr=0/1, otherwise mballoc can
2679 * spend a lot of time loading imperfect groups
2681 if ((prefetch_grp == group) &&
2683 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2684 unsigned int curr_ios = prefetch_ios;
2686 nr = sbi->s_mb_prefetch;
2687 if (ext4_has_feature_flex_bg(sb)) {
2688 nr = 1 << sbi->s_log_groups_per_flex;
2689 nr -= group & (nr - 1);
2690 nr = min(nr, sbi->s_mb_prefetch);
2692 prefetch_grp = ext4_mb_prefetch(sb, group,
2694 if (prefetch_ios == curr_ios)
2698 /* This now checks without needing the buddy page */
2699 ret = ext4_mb_good_group_nolock(ac, group, cr);
2706 err = ext4_mb_load_buddy(sb, group, &e4b);
2710 ext4_lock_group(sb, group);
2713 * We need to check again after locking the
2716 ret = ext4_mb_good_group(ac, group, cr);
2718 ext4_unlock_group(sb, group);
2719 ext4_mb_unload_buddy(&e4b);
2723 ac->ac_groups_scanned++;
2725 ext4_mb_simple_scan_group(ac, &e4b);
2726 else if (cr == 1 && sbi->s_stripe &&
2727 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2728 ext4_mb_scan_aligned(ac, &e4b);
2730 ext4_mb_complex_scan_group(ac, &e4b);
2732 ext4_unlock_group(sb, group);
2733 ext4_mb_unload_buddy(&e4b);
2735 if (ac->ac_status != AC_STATUS_CONTINUE)
2738 /* Processed all groups and haven't found blocks */
2739 if (sbi->s_mb_stats && i == ngroups)
2740 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2743 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2744 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2746 * We've been searching too long. Let's try to allocate
2747 * the best chunk we've found so far
2749 ext4_mb_try_best_found(ac, &e4b);
2750 if (ac->ac_status != AC_STATUS_FOUND) {
2752 * Someone more lucky has already allocated it.
2753 * The only thing we can do is just take first
2756 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2757 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2758 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2759 ac->ac_b_ex.fe_len, lost);
2761 ac->ac_b_ex.fe_group = 0;
2762 ac->ac_b_ex.fe_start = 0;
2763 ac->ac_b_ex.fe_len = 0;
2764 ac->ac_status = AC_STATUS_CONTINUE;
2765 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2771 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2772 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2774 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2777 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2778 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2779 ac->ac_flags, cr, err);
2782 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2787 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2789 struct super_block *sb = pde_data(file_inode(seq->file));
2792 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2795 return (void *) ((unsigned long) group);
2798 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2800 struct super_block *sb = pde_data(file_inode(seq->file));
2804 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2807 return (void *) ((unsigned long) group);
2810 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2812 struct super_block *sb = pde_data(file_inode(seq->file));
2813 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2815 int err, buddy_loaded = 0;
2816 struct ext4_buddy e4b;
2817 struct ext4_group_info *grinfo;
2818 unsigned char blocksize_bits = min_t(unsigned char,
2819 sb->s_blocksize_bits,
2820 EXT4_MAX_BLOCK_LOG_SIZE);
2822 struct ext4_group_info info;
2823 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2828 seq_puts(seq, "#group: free frags first ["
2829 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2830 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2832 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2833 sizeof(struct ext4_group_info);
2835 grinfo = ext4_get_group_info(sb, group);
2836 /* Load the group info in memory only if not already loaded. */
2837 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2838 err = ext4_mb_load_buddy(sb, group, &e4b);
2840 seq_printf(seq, "#%-5u: I/O error\n", group);
2846 memcpy(&sg, ext4_get_group_info(sb, group), i);
2849 ext4_mb_unload_buddy(&e4b);
2851 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2852 sg.info.bb_fragments, sg.info.bb_first_free);
2853 for (i = 0; i <= 13; i++)
2854 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2855 sg.info.bb_counters[i] : 0);
2856 seq_puts(seq, " ]\n");
2861 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2865 const struct seq_operations ext4_mb_seq_groups_ops = {
2866 .start = ext4_mb_seq_groups_start,
2867 .next = ext4_mb_seq_groups_next,
2868 .stop = ext4_mb_seq_groups_stop,
2869 .show = ext4_mb_seq_groups_show,
2872 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
2874 struct super_block *sb = seq->private;
2875 struct ext4_sb_info *sbi = EXT4_SB(sb);
2877 seq_puts(seq, "mballoc:\n");
2878 if (!sbi->s_mb_stats) {
2879 seq_puts(seq, "\tmb stats collection turned off.\n");
2880 seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
2883 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
2884 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
2886 seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned));
2888 seq_puts(seq, "\tcr0_stats:\n");
2889 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0]));
2890 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2891 atomic64_read(&sbi->s_bal_cX_groups_considered[0]));
2892 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2893 atomic64_read(&sbi->s_bal_cX_failed[0]));
2894 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2895 atomic_read(&sbi->s_bal_cr0_bad_suggestions));
2897 seq_puts(seq, "\tcr1_stats:\n");
2898 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1]));
2899 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2900 atomic64_read(&sbi->s_bal_cX_groups_considered[1]));
2901 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2902 atomic64_read(&sbi->s_bal_cX_failed[1]));
2903 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2904 atomic_read(&sbi->s_bal_cr1_bad_suggestions));
2906 seq_puts(seq, "\tcr2_stats:\n");
2907 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2]));
2908 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2909 atomic64_read(&sbi->s_bal_cX_groups_considered[2]));
2910 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2911 atomic64_read(&sbi->s_bal_cX_failed[2]));
2913 seq_puts(seq, "\tcr3_stats:\n");
2914 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3]));
2915 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2916 atomic64_read(&sbi->s_bal_cX_groups_considered[3]));
2917 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2918 atomic64_read(&sbi->s_bal_cX_failed[3]));
2919 seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned));
2920 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
2921 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
2922 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
2923 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
2925 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
2926 atomic_read(&sbi->s_mb_buddies_generated),
2927 ext4_get_groups_count(sb));
2928 seq_printf(seq, "\tbuddies_time_used: %llu\n",
2929 atomic64_read(&sbi->s_mb_generation_time));
2930 seq_printf(seq, "\tpreallocated: %u\n",
2931 atomic_read(&sbi->s_mb_preallocated));
2932 seq_printf(seq, "\tdiscarded: %u\n",
2933 atomic_read(&sbi->s_mb_discarded));
2937 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
2938 __acquires(&EXT4_SB(sb)->s_mb_rb_lock)
2940 struct super_block *sb = pde_data(file_inode(seq->file));
2941 unsigned long position;
2943 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
2945 position = *pos + 1;
2946 return (void *) ((unsigned long) position);
2949 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
2951 struct super_block *sb = pde_data(file_inode(seq->file));
2952 unsigned long position;
2955 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
2957 position = *pos + 1;
2958 return (void *) ((unsigned long) position);
2961 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
2963 struct super_block *sb = pde_data(file_inode(seq->file));
2964 struct ext4_sb_info *sbi = EXT4_SB(sb);
2965 unsigned long position = ((unsigned long) v);
2966 struct ext4_group_info *grp;
2970 if (position >= MB_NUM_ORDERS(sb)) {
2971 position -= MB_NUM_ORDERS(sb);
2973 seq_puts(seq, "avg_fragment_size_lists:\n");
2976 read_lock(&sbi->s_mb_avg_fragment_size_locks[position]);
2977 list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position],
2978 bb_avg_fragment_size_node)
2980 read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]);
2981 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
2982 (unsigned int)position, count);
2986 if (position == 0) {
2987 seq_printf(seq, "optimize_scan: %d\n",
2988 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
2989 seq_puts(seq, "max_free_order_lists:\n");
2992 read_lock(&sbi->s_mb_largest_free_orders_locks[position]);
2993 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
2994 bb_largest_free_order_node)
2996 read_unlock(&sbi->s_mb_largest_free_orders_locks[position]);
2997 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
2998 (unsigned int)position, count);
3003 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3007 const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3008 .start = ext4_mb_seq_structs_summary_start,
3009 .next = ext4_mb_seq_structs_summary_next,
3010 .stop = ext4_mb_seq_structs_summary_stop,
3011 .show = ext4_mb_seq_structs_summary_show,
3014 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3016 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3017 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3024 * Allocate the top-level s_group_info array for the specified number
3027 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3029 struct ext4_sb_info *sbi = EXT4_SB(sb);
3031 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3033 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3034 EXT4_DESC_PER_BLOCK_BITS(sb);
3035 if (size <= sbi->s_group_info_size)
3038 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3039 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3040 if (!new_groupinfo) {
3041 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3045 old_groupinfo = rcu_dereference(sbi->s_group_info);
3047 memcpy(new_groupinfo, old_groupinfo,
3048 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3050 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3051 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3053 ext4_kvfree_array_rcu(old_groupinfo);
3054 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3055 sbi->s_group_info_size);
3059 /* Create and initialize ext4_group_info data for the given group. */
3060 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3061 struct ext4_group_desc *desc)
3065 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3066 struct ext4_sb_info *sbi = EXT4_SB(sb);
3067 struct ext4_group_info **meta_group_info;
3068 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3071 * First check if this group is the first of a reserved block.
3072 * If it's true, we have to allocate a new table of pointers
3073 * to ext4_group_info structures
3075 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3076 metalen = sizeof(*meta_group_info) <<
3077 EXT4_DESC_PER_BLOCK_BITS(sb);
3078 meta_group_info = kmalloc(metalen, GFP_NOFS);
3079 if (meta_group_info == NULL) {
3080 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3081 "for a buddy group");
3082 goto exit_meta_group_info;
3085 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3089 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3090 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3092 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3093 if (meta_group_info[i] == NULL) {
3094 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3095 goto exit_group_info;
3097 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3098 &(meta_group_info[i]->bb_state));
3101 * initialize bb_free to be able to skip
3102 * empty groups without initialization
3104 if (ext4_has_group_desc_csum(sb) &&
3105 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3106 meta_group_info[i]->bb_free =
3107 ext4_free_clusters_after_init(sb, group, desc);
3109 meta_group_info[i]->bb_free =
3110 ext4_free_group_clusters(sb, desc);
3113 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3114 init_rwsem(&meta_group_info[i]->alloc_sem);
3115 meta_group_info[i]->bb_free_root = RB_ROOT;
3116 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3117 INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node);
3118 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3119 meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */
3120 meta_group_info[i]->bb_group = group;
3122 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3126 /* If a meta_group_info table has been allocated, release it now */
3127 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3128 struct ext4_group_info ***group_info;
3131 group_info = rcu_dereference(sbi->s_group_info);
3132 kfree(group_info[idx]);
3133 group_info[idx] = NULL;
3136 exit_meta_group_info:
3138 } /* ext4_mb_add_groupinfo */
3140 static int ext4_mb_init_backend(struct super_block *sb)
3142 ext4_group_t ngroups = ext4_get_groups_count(sb);
3144 struct ext4_sb_info *sbi = EXT4_SB(sb);
3146 struct ext4_group_desc *desc;
3147 struct ext4_group_info ***group_info;
3148 struct kmem_cache *cachep;
3150 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3154 sbi->s_buddy_cache = new_inode(sb);
3155 if (sbi->s_buddy_cache == NULL) {
3156 ext4_msg(sb, KERN_ERR, "can't get new inode");
3159 /* To avoid potentially colliding with an valid on-disk inode number,
3160 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3161 * not in the inode hash, so it should never be found by iget(), but
3162 * this will avoid confusion if it ever shows up during debugging. */
3163 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3164 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3165 for (i = 0; i < ngroups; i++) {
3167 desc = ext4_get_group_desc(sb, i, NULL);
3169 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3172 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3176 if (ext4_has_feature_flex_bg(sb)) {
3177 /* a single flex group is supposed to be read by a single IO.
3178 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3179 * unsigned integer, so the maximum shift is 32.
3181 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3182 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3185 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3186 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3187 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3189 sbi->s_mb_prefetch = 32;
3191 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3192 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3193 /* now many real IOs to prefetch within a single allocation at cr=0
3194 * given cr=0 is an CPU-related optimization we shouldn't try to
3195 * load too many groups, at some point we should start to use what
3196 * we've got in memory.
3197 * with an average random access time 5ms, it'd take a second to get
3198 * 200 groups (* N with flex_bg), so let's make this limit 4
3200 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3201 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3202 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3207 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3209 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
3210 i = sbi->s_group_info_size;
3212 group_info = rcu_dereference(sbi->s_group_info);
3214 kfree(group_info[i]);
3216 iput(sbi->s_buddy_cache);
3219 kvfree(rcu_dereference(sbi->s_group_info));
3224 static void ext4_groupinfo_destroy_slabs(void)
3228 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3229 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3230 ext4_groupinfo_caches[i] = NULL;
3234 static int ext4_groupinfo_create_slab(size_t size)
3236 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3238 int blocksize_bits = order_base_2(size);
3239 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3240 struct kmem_cache *cachep;
3242 if (cache_index >= NR_GRPINFO_CACHES)
3245 if (unlikely(cache_index < 0))
3248 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3249 if (ext4_groupinfo_caches[cache_index]) {
3250 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3251 return 0; /* Already created */
3254 slab_size = offsetof(struct ext4_group_info,
3255 bb_counters[blocksize_bits + 2]);
3257 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3258 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3261 ext4_groupinfo_caches[cache_index] = cachep;
3263 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3266 "EXT4-fs: no memory for groupinfo slab cache\n");
3273 static void ext4_discard_work(struct work_struct *work)
3275 struct ext4_sb_info *sbi = container_of(work,
3276 struct ext4_sb_info, s_discard_work);
3277 struct super_block *sb = sbi->s_sb;
3278 struct ext4_free_data *fd, *nfd;
3279 struct ext4_buddy e4b;
3280 struct list_head discard_list;
3281 ext4_group_t grp, load_grp;
3284 INIT_LIST_HEAD(&discard_list);
3285 spin_lock(&sbi->s_md_lock);
3286 list_splice_init(&sbi->s_discard_list, &discard_list);
3287 spin_unlock(&sbi->s_md_lock);
3289 load_grp = UINT_MAX;
3290 list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
3292 * If filesystem is umounting or no memory or suffering
3293 * from no space, give up the discard
3295 if ((sb->s_flags & SB_ACTIVE) && !err &&
3296 !atomic_read(&sbi->s_retry_alloc_pending)) {
3297 grp = fd->efd_group;
3298 if (grp != load_grp) {
3299 if (load_grp != UINT_MAX)
3300 ext4_mb_unload_buddy(&e4b);
3302 err = ext4_mb_load_buddy(sb, grp, &e4b);
3304 kmem_cache_free(ext4_free_data_cachep, fd);
3305 load_grp = UINT_MAX;
3312 ext4_lock_group(sb, grp);
3313 ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
3314 fd->efd_start_cluster + fd->efd_count - 1, 1);
3315 ext4_unlock_group(sb, grp);
3317 kmem_cache_free(ext4_free_data_cachep, fd);
3320 if (load_grp != UINT_MAX)
3321 ext4_mb_unload_buddy(&e4b);
3324 int ext4_mb_init(struct super_block *sb)
3326 struct ext4_sb_info *sbi = EXT4_SB(sb);
3328 unsigned offset, offset_incr;
3332 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3334 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3335 if (sbi->s_mb_offsets == NULL) {
3340 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3341 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3342 if (sbi->s_mb_maxs == NULL) {
3347 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3351 /* order 0 is regular bitmap */
3352 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3353 sbi->s_mb_offsets[0] = 0;
3357 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3358 max = sb->s_blocksize << 2;
3360 sbi->s_mb_offsets[i] = offset;
3361 sbi->s_mb_maxs[i] = max;
3362 offset += offset_incr;
3363 offset_incr = offset_incr >> 1;
3366 } while (i < MB_NUM_ORDERS(sb));
3368 sbi->s_mb_avg_fragment_size =
3369 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3371 if (!sbi->s_mb_avg_fragment_size) {
3375 sbi->s_mb_avg_fragment_size_locks =
3376 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3378 if (!sbi->s_mb_avg_fragment_size_locks) {
3382 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3383 INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]);
3384 rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]);
3386 sbi->s_mb_largest_free_orders =
3387 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3389 if (!sbi->s_mb_largest_free_orders) {
3393 sbi->s_mb_largest_free_orders_locks =
3394 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3396 if (!sbi->s_mb_largest_free_orders_locks) {
3400 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3401 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3402 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3405 spin_lock_init(&sbi->s_md_lock);
3406 sbi->s_mb_free_pending = 0;
3407 INIT_LIST_HEAD(&sbi->s_freed_data_list);
3408 INIT_LIST_HEAD(&sbi->s_discard_list);
3409 INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
3410 atomic_set(&sbi->s_retry_alloc_pending, 0);
3412 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3413 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3414 sbi->s_mb_stats = MB_DEFAULT_STATS;
3415 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3416 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3417 sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
3419 * The default group preallocation is 512, which for 4k block
3420 * sizes translates to 2 megabytes. However for bigalloc file
3421 * systems, this is probably too big (i.e, if the cluster size
3422 * is 1 megabyte, then group preallocation size becomes half a
3423 * gigabyte!). As a default, we will keep a two megabyte
3424 * group pralloc size for cluster sizes up to 64k, and after
3425 * that, we will force a minimum group preallocation size of
3426 * 32 clusters. This translates to 8 megs when the cluster
3427 * size is 256k, and 32 megs when the cluster size is 1 meg,
3428 * which seems reasonable as a default.
3430 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3431 sbi->s_cluster_bits, 32);
3433 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3434 * to the lowest multiple of s_stripe which is bigger than
3435 * the s_mb_group_prealloc as determined above. We want
3436 * the preallocation size to be an exact multiple of the
3437 * RAID stripe size so that preallocations don't fragment
3440 if (sbi->s_stripe > 1) {
3441 sbi->s_mb_group_prealloc = roundup(
3442 sbi->s_mb_group_prealloc, sbi->s_stripe);
3445 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3446 if (sbi->s_locality_groups == NULL) {
3450 for_each_possible_cpu(i) {
3451 struct ext4_locality_group *lg;
3452 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3453 mutex_init(&lg->lg_mutex);
3454 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3455 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3456 spin_lock_init(&lg->lg_prealloc_lock);
3459 if (bdev_nonrot(sb->s_bdev))
3460 sbi->s_mb_max_linear_groups = 0;
3462 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3463 /* init file for buddy data */
3464 ret = ext4_mb_init_backend(sb);
3466 goto out_free_locality_groups;
3470 out_free_locality_groups:
3471 free_percpu(sbi->s_locality_groups);
3472 sbi->s_locality_groups = NULL;
3474 kfree(sbi->s_mb_avg_fragment_size);
3475 kfree(sbi->s_mb_avg_fragment_size_locks);
3476 kfree(sbi->s_mb_largest_free_orders);
3477 kfree(sbi->s_mb_largest_free_orders_locks);
3478 kfree(sbi->s_mb_offsets);
3479 sbi->s_mb_offsets = NULL;
3480 kfree(sbi->s_mb_maxs);
3481 sbi->s_mb_maxs = NULL;
3485 /* need to called with the ext4 group lock held */
3486 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3488 struct ext4_prealloc_space *pa;
3489 struct list_head *cur, *tmp;
3492 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3493 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3494 list_del(&pa->pa_group_list);
3496 kmem_cache_free(ext4_pspace_cachep, pa);
3501 int ext4_mb_release(struct super_block *sb)
3503 ext4_group_t ngroups = ext4_get_groups_count(sb);
3505 int num_meta_group_infos;
3506 struct ext4_group_info *grinfo, ***group_info;
3507 struct ext4_sb_info *sbi = EXT4_SB(sb);
3508 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3511 if (test_opt(sb, DISCARD)) {
3513 * wait the discard work to drain all of ext4_free_data
3515 flush_work(&sbi->s_discard_work);
3516 WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
3519 if (sbi->s_group_info) {
3520 for (i = 0; i < ngroups; i++) {
3522 grinfo = ext4_get_group_info(sb, i);
3523 mb_group_bb_bitmap_free(grinfo);
3524 ext4_lock_group(sb, i);
3525 count = ext4_mb_cleanup_pa(grinfo);
3527 mb_debug(sb, "mballoc: %d PAs left\n",
3529 ext4_unlock_group(sb, i);
3530 kmem_cache_free(cachep, grinfo);
3532 num_meta_group_infos = (ngroups +
3533 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3534 EXT4_DESC_PER_BLOCK_BITS(sb);
3536 group_info = rcu_dereference(sbi->s_group_info);
3537 for (i = 0; i < num_meta_group_infos; i++)
3538 kfree(group_info[i]);
3542 kfree(sbi->s_mb_avg_fragment_size);
3543 kfree(sbi->s_mb_avg_fragment_size_locks);
3544 kfree(sbi->s_mb_largest_free_orders);
3545 kfree(sbi->s_mb_largest_free_orders_locks);
3546 kfree(sbi->s_mb_offsets);
3547 kfree(sbi->s_mb_maxs);
3548 iput(sbi->s_buddy_cache);
3549 if (sbi->s_mb_stats) {
3550 ext4_msg(sb, KERN_INFO,
3551 "mballoc: %u blocks %u reqs (%u success)",
3552 atomic_read(&sbi->s_bal_allocated),
3553 atomic_read(&sbi->s_bal_reqs),
3554 atomic_read(&sbi->s_bal_success));
3555 ext4_msg(sb, KERN_INFO,
3556 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3557 "%u 2^N hits, %u breaks, %u lost",
3558 atomic_read(&sbi->s_bal_ex_scanned),
3559 atomic_read(&sbi->s_bal_groups_scanned),
3560 atomic_read(&sbi->s_bal_goals),
3561 atomic_read(&sbi->s_bal_2orders),
3562 atomic_read(&sbi->s_bal_breaks),
3563 atomic_read(&sbi->s_mb_lost_chunks));
3564 ext4_msg(sb, KERN_INFO,
3565 "mballoc: %u generated and it took %llu",
3566 atomic_read(&sbi->s_mb_buddies_generated),
3567 atomic64_read(&sbi->s_mb_generation_time));
3568 ext4_msg(sb, KERN_INFO,
3569 "mballoc: %u preallocated, %u discarded",
3570 atomic_read(&sbi->s_mb_preallocated),
3571 atomic_read(&sbi->s_mb_discarded));
3574 free_percpu(sbi->s_locality_groups);
3579 static inline int ext4_issue_discard(struct super_block *sb,
3580 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3583 ext4_fsblk_t discard_block;
3585 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3586 ext4_group_first_block_no(sb, block_group));
3587 count = EXT4_C2B(EXT4_SB(sb), count);
3588 trace_ext4_discard_blocks(sb,
3589 (unsigned long long) discard_block, count);
3591 return __blkdev_issue_discard(sb->s_bdev,
3592 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3593 (sector_t)count << (sb->s_blocksize_bits - 9),
3596 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3599 static void ext4_free_data_in_buddy(struct super_block *sb,
3600 struct ext4_free_data *entry)
3602 struct ext4_buddy e4b;
3603 struct ext4_group_info *db;
3604 int err, count = 0, count2 = 0;
3606 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3607 entry->efd_count, entry->efd_group, entry);
3609 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3610 /* we expect to find existing buddy because it's pinned */
3613 spin_lock(&EXT4_SB(sb)->s_md_lock);
3614 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3615 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3618 /* there are blocks to put in buddy to make them really free */
3619 count += entry->efd_count;
3621 ext4_lock_group(sb, entry->efd_group);
3622 /* Take it out of per group rb tree */
3623 rb_erase(&entry->efd_node, &(db->bb_free_root));
3624 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3627 * Clear the trimmed flag for the group so that the next
3628 * ext4_trim_fs can trim it.
3629 * If the volume is mounted with -o discard, online discard
3630 * is supported and the free blocks will be trimmed online.
3632 if (!test_opt(sb, DISCARD))
3633 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3635 if (!db->bb_free_root.rb_node) {
3636 /* No more items in the per group rb tree
3637 * balance refcounts from ext4_mb_free_metadata()
3639 put_page(e4b.bd_buddy_page);
3640 put_page(e4b.bd_bitmap_page);
3642 ext4_unlock_group(sb, entry->efd_group);
3643 ext4_mb_unload_buddy(&e4b);
3645 mb_debug(sb, "freed %d blocks in %d structures\n", count,
3650 * This function is called by the jbd2 layer once the commit has finished,
3651 * so we know we can free the blocks that were released with that commit.
3653 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3655 struct ext4_sb_info *sbi = EXT4_SB(sb);
3656 struct ext4_free_data *entry, *tmp;
3657 struct list_head freed_data_list;
3658 struct list_head *cut_pos = NULL;
3661 INIT_LIST_HEAD(&freed_data_list);
3663 spin_lock(&sbi->s_md_lock);
3664 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3665 if (entry->efd_tid != commit_tid)
3667 cut_pos = &entry->efd_list;
3670 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3672 spin_unlock(&sbi->s_md_lock);
3674 list_for_each_entry(entry, &freed_data_list, efd_list)
3675 ext4_free_data_in_buddy(sb, entry);
3677 if (test_opt(sb, DISCARD)) {
3678 spin_lock(&sbi->s_md_lock);
3679 wake = list_empty(&sbi->s_discard_list);
3680 list_splice_tail(&freed_data_list, &sbi->s_discard_list);
3681 spin_unlock(&sbi->s_md_lock);
3683 queue_work(system_unbound_wq, &sbi->s_discard_work);
3685 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3686 kmem_cache_free(ext4_free_data_cachep, entry);
3690 int __init ext4_init_mballoc(void)
3692 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3693 SLAB_RECLAIM_ACCOUNT);
3694 if (ext4_pspace_cachep == NULL)
3697 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3698 SLAB_RECLAIM_ACCOUNT);
3699 if (ext4_ac_cachep == NULL)
3702 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3703 SLAB_RECLAIM_ACCOUNT);
3704 if (ext4_free_data_cachep == NULL)
3710 kmem_cache_destroy(ext4_ac_cachep);
3712 kmem_cache_destroy(ext4_pspace_cachep);
3717 void ext4_exit_mballoc(void)
3720 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3721 * before destroying the slab cache.
3724 kmem_cache_destroy(ext4_pspace_cachep);
3725 kmem_cache_destroy(ext4_ac_cachep);
3726 kmem_cache_destroy(ext4_free_data_cachep);
3727 ext4_groupinfo_destroy_slabs();
3732 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3733 * Returns 0 if success or error code
3735 static noinline_for_stack int
3736 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3737 handle_t *handle, unsigned int reserv_clstrs)
3739 struct buffer_head *bitmap_bh = NULL;
3740 struct ext4_group_desc *gdp;
3741 struct buffer_head *gdp_bh;
3742 struct ext4_sb_info *sbi;
3743 struct super_block *sb;
3747 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3748 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3753 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3754 if (IS_ERR(bitmap_bh)) {
3755 err = PTR_ERR(bitmap_bh);
3760 BUFFER_TRACE(bitmap_bh, "getting write access");
3761 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
3767 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3771 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3772 ext4_free_group_clusters(sb, gdp));
3774 BUFFER_TRACE(gdp_bh, "get_write_access");
3775 err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE);
3779 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3781 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3782 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3783 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3784 "fs metadata", block, block+len);
3785 /* File system mounted not to panic on error
3786 * Fix the bitmap and return EFSCORRUPTED
3787 * We leak some of the blocks here.
3789 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3790 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3791 ac->ac_b_ex.fe_len);
3792 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3793 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3795 err = -EFSCORRUPTED;
3799 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3800 #ifdef AGGRESSIVE_CHECK
3803 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3804 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3805 bitmap_bh->b_data));
3809 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3810 ac->ac_b_ex.fe_len);
3811 if (ext4_has_group_desc_csum(sb) &&
3812 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3813 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3814 ext4_free_group_clusters_set(sb, gdp,
3815 ext4_free_clusters_after_init(sb,
3816 ac->ac_b_ex.fe_group, gdp));
3818 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3819 ext4_free_group_clusters_set(sb, gdp, len);
3820 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3821 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3823 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3824 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3826 * Now reduce the dirty block count also. Should not go negative
3828 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3829 /* release all the reserved blocks if non delalloc */
3830 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3833 if (sbi->s_log_groups_per_flex) {
3834 ext4_group_t flex_group = ext4_flex_group(sbi,
3835 ac->ac_b_ex.fe_group);
3836 atomic64_sub(ac->ac_b_ex.fe_len,
3837 &sbi_array_rcu_deref(sbi, s_flex_groups,
3838 flex_group)->free_clusters);
3841 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3844 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3852 * Idempotent helper for Ext4 fast commit replay path to set the state of
3853 * blocks in bitmaps and update counters.
3855 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
3858 struct buffer_head *bitmap_bh = NULL;
3859 struct ext4_group_desc *gdp;
3860 struct buffer_head *gdp_bh;
3861 struct ext4_sb_info *sbi = EXT4_SB(sb);
3863 ext4_grpblk_t blkoff;
3866 unsigned int clen, clen_changed, thisgrp_len;
3869 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
3872 * Check to see if we are freeing blocks across a group
3874 * In case of flex_bg, this can happen that (block, len) may
3875 * span across more than one group. In that case we need to
3876 * get the corresponding group metadata to work with.
3877 * For this we have goto again loop.
3879 thisgrp_len = min_t(unsigned int, (unsigned int)len,
3880 EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
3881 clen = EXT4_NUM_B2C(sbi, thisgrp_len);
3883 if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
3884 ext4_error(sb, "Marking blocks in system zone - "
3885 "Block = %llu, len = %u",
3886 block, thisgrp_len);
3891 bitmap_bh = ext4_read_block_bitmap(sb, group);
3892 if (IS_ERR(bitmap_bh)) {
3893 err = PTR_ERR(bitmap_bh);
3899 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
3903 ext4_lock_group(sb, group);
3905 for (i = 0; i < clen; i++)
3906 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
3910 clen_changed = clen - already;
3912 mb_set_bits(bitmap_bh->b_data, blkoff, clen);
3914 mb_clear_bits(bitmap_bh->b_data, blkoff, clen);
3915 if (ext4_has_group_desc_csum(sb) &&
3916 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3917 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3918 ext4_free_group_clusters_set(sb, gdp,
3919 ext4_free_clusters_after_init(sb, group, gdp));
3922 clen = ext4_free_group_clusters(sb, gdp) - clen_changed;
3924 clen = ext4_free_group_clusters(sb, gdp) + clen_changed;
3926 ext4_free_group_clusters_set(sb, gdp, clen);
3927 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
3928 ext4_group_desc_csum_set(sb, group, gdp);
3930 ext4_unlock_group(sb, group);
3932 if (sbi->s_log_groups_per_flex) {
3933 ext4_group_t flex_group = ext4_flex_group(sbi, group);
3934 struct flex_groups *fg = sbi_array_rcu_deref(sbi,
3935 s_flex_groups, flex_group);
3938 atomic64_sub(clen_changed, &fg->free_clusters);
3940 atomic64_add(clen_changed, &fg->free_clusters);
3944 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
3947 sync_dirty_buffer(bitmap_bh);
3948 err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
3949 sync_dirty_buffer(gdp_bh);
3953 block += thisgrp_len;
3964 * here we normalize request for locality group
3965 * Group request are normalized to s_mb_group_prealloc, which goes to
3966 * s_strip if we set the same via mount option.
3967 * s_mb_group_prealloc can be configured via
3968 * /sys/fs/ext4/<partition>/mb_group_prealloc
3970 * XXX: should we try to preallocate more than the group has now?
3972 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3974 struct super_block *sb = ac->ac_sb;
3975 struct ext4_locality_group *lg = ac->ac_lg;
3978 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3979 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
3983 * Normalization means making request better in terms of
3984 * size and alignment
3986 static noinline_for_stack void
3987 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3988 struct ext4_allocation_request *ar)
3990 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3993 loff_t size, start_off;
3994 loff_t orig_size __maybe_unused;
3996 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3997 struct ext4_prealloc_space *pa;
3999 /* do normalize only data requests, metadata requests
4000 do not need preallocation */
4001 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4004 /* sometime caller may want exact blocks */
4005 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4008 /* caller may indicate that preallocation isn't
4009 * required (it's a tail, for example) */
4010 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
4013 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
4014 ext4_mb_normalize_group_request(ac);
4018 bsbits = ac->ac_sb->s_blocksize_bits;
4020 /* first, let's learn actual file size
4021 * given current request is allocated */
4022 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4023 size = size << bsbits;
4024 if (size < i_size_read(ac->ac_inode))
4025 size = i_size_read(ac->ac_inode);
4028 /* max size of free chunks */
4031 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
4032 (req <= (size) || max <= (chunk_size))
4034 /* first, try to predict filesize */
4035 /* XXX: should this table be tunable? */
4037 if (size <= 16 * 1024) {
4039 } else if (size <= 32 * 1024) {
4041 } else if (size <= 64 * 1024) {
4043 } else if (size <= 128 * 1024) {
4045 } else if (size <= 256 * 1024) {
4047 } else if (size <= 512 * 1024) {
4049 } else if (size <= 1024 * 1024) {
4051 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
4052 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4053 (21 - bsbits)) << 21;
4054 size = 2 * 1024 * 1024;
4055 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
4056 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4057 (22 - bsbits)) << 22;
4058 size = 4 * 1024 * 1024;
4059 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
4060 (8<<20)>>bsbits, max, 8 * 1024)) {
4061 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4062 (23 - bsbits)) << 23;
4063 size = 8 * 1024 * 1024;
4065 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
4066 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
4067 ac->ac_o_ex.fe_len) << bsbits;
4069 size = size >> bsbits;
4070 start = start_off >> bsbits;
4073 * For tiny groups (smaller than 8MB) the chosen allocation
4074 * alignment may be larger than group size. Make sure the
4075 * alignment does not move allocation to a different group which
4076 * makes mballoc fail assertions later.
4078 start = max(start, rounddown(ac->ac_o_ex.fe_logical,
4079 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
4081 /* don't cover already allocated blocks in selected range */
4082 if (ar->pleft && start <= ar->lleft) {
4083 size -= ar->lleft + 1 - start;
4084 start = ar->lleft + 1;
4086 if (ar->pright && start + size - 1 >= ar->lright)
4087 size -= start + size - ar->lright;
4090 * Trim allocation request for filesystems with artificially small
4093 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4094 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4098 /* check we don't cross already preallocated blocks */
4100 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4105 spin_lock(&pa->pa_lock);
4106 if (pa->pa_deleted) {
4107 spin_unlock(&pa->pa_lock);
4111 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4114 /* PA must not overlap original request */
4115 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
4116 ac->ac_o_ex.fe_logical < pa->pa_lstart));
4118 /* skip PAs this normalized request doesn't overlap with */
4119 if (pa->pa_lstart >= end || pa_end <= start) {
4120 spin_unlock(&pa->pa_lock);
4123 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
4125 /* adjust start or end to be adjacent to this pa */
4126 if (pa_end <= ac->ac_o_ex.fe_logical) {
4127 BUG_ON(pa_end < start);
4129 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4130 BUG_ON(pa->pa_lstart > end);
4131 end = pa->pa_lstart;
4133 spin_unlock(&pa->pa_lock);
4138 /* XXX: extra loop to check we really don't overlap preallocations */
4140 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4143 spin_lock(&pa->pa_lock);
4144 if (pa->pa_deleted == 0) {
4145 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4147 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
4149 spin_unlock(&pa->pa_lock);
4154 * In this function "start" and "size" are normalized for better
4155 * alignment and length such that we could preallocate more blocks.
4156 * This normalization is done such that original request of
4157 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
4158 * "size" boundaries.
4159 * (Note fe_len can be relaxed since FS block allocation API does not
4160 * provide gurantee on number of contiguous blocks allocation since that
4161 * depends upon free space left, etc).
4162 * In case of inode pa, later we use the allocated blocks
4163 * [pa_start + fe_logical - pa_lstart, fe_len/size] from the preallocated
4164 * range of goal/best blocks [start, size] to put it at the
4165 * ac_o_ex.fe_logical extent of this inode.
4166 * (See ext4_mb_use_inode_pa() for more details)
4168 if (start + size <= ac->ac_o_ex.fe_logical ||
4169 start > ac->ac_o_ex.fe_logical) {
4170 ext4_msg(ac->ac_sb, KERN_ERR,
4171 "start %lu, size %lu, fe_logical %lu",
4172 (unsigned long) start, (unsigned long) size,
4173 (unsigned long) ac->ac_o_ex.fe_logical);
4176 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4178 /* now prepare goal request */
4180 /* XXX: is it better to align blocks WRT to logical
4181 * placement or satisfy big request as is */
4182 ac->ac_g_ex.fe_logical = start;
4183 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4185 /* define goal start in order to merge */
4186 if (ar->pright && (ar->lright == (start + size))) {
4187 /* merge to the right */
4188 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4189 &ac->ac_f_ex.fe_group,
4190 &ac->ac_f_ex.fe_start);
4191 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4193 if (ar->pleft && (ar->lleft + 1 == start)) {
4194 /* merge to the left */
4195 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4196 &ac->ac_f_ex.fe_group,
4197 &ac->ac_f_ex.fe_start);
4198 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4201 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4205 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4207 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4209 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4210 atomic_inc(&sbi->s_bal_reqs);
4211 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4212 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4213 atomic_inc(&sbi->s_bal_success);
4214 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4215 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4216 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4217 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4218 atomic_inc(&sbi->s_bal_goals);
4219 if (ac->ac_found > sbi->s_mb_max_to_scan)
4220 atomic_inc(&sbi->s_bal_breaks);
4223 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4224 trace_ext4_mballoc_alloc(ac);
4226 trace_ext4_mballoc_prealloc(ac);
4230 * Called on failure; free up any blocks from the inode PA for this
4231 * context. We don't need this for MB_GROUP_PA because we only change
4232 * pa_free in ext4_mb_release_context(), but on failure, we've already
4233 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4235 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4237 struct ext4_prealloc_space *pa = ac->ac_pa;
4238 struct ext4_buddy e4b;
4242 if (ac->ac_f_ex.fe_len == 0)
4244 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4247 * This should never happen since we pin the
4248 * pages in the ext4_allocation_context so
4249 * ext4_mb_load_buddy() should never fail.
4251 WARN(1, "mb_load_buddy failed (%d)", err);
4254 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4255 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4256 ac->ac_f_ex.fe_len);
4257 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4258 ext4_mb_unload_buddy(&e4b);
4261 if (pa->pa_type == MB_INODE_PA)
4262 pa->pa_free += ac->ac_b_ex.fe_len;
4266 * use blocks preallocated to inode
4268 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4269 struct ext4_prealloc_space *pa)
4271 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4276 /* found preallocated blocks, use them */
4277 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4278 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4279 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4280 len = EXT4_NUM_B2C(sbi, end - start);
4281 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4282 &ac->ac_b_ex.fe_start);
4283 ac->ac_b_ex.fe_len = len;
4284 ac->ac_status = AC_STATUS_FOUND;
4287 BUG_ON(start < pa->pa_pstart);
4288 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4289 BUG_ON(pa->pa_free < len);
4292 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4296 * use blocks preallocated to locality group
4298 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4299 struct ext4_prealloc_space *pa)
4301 unsigned int len = ac->ac_o_ex.fe_len;
4303 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4304 &ac->ac_b_ex.fe_group,
4305 &ac->ac_b_ex.fe_start);
4306 ac->ac_b_ex.fe_len = len;
4307 ac->ac_status = AC_STATUS_FOUND;
4310 /* we don't correct pa_pstart or pa_plen here to avoid
4311 * possible race when the group is being loaded concurrently
4312 * instead we correct pa later, after blocks are marked
4313 * in on-disk bitmap -- see ext4_mb_release_context()
4314 * Other CPUs are prevented from allocating from this pa by lg_mutex
4316 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4317 pa->pa_lstart-len, len, pa);
4321 * Return the prealloc space that have minimal distance
4322 * from the goal block. @cpa is the prealloc
4323 * space that is having currently known minimal distance
4324 * from the goal block.
4326 static struct ext4_prealloc_space *
4327 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4328 struct ext4_prealloc_space *pa,
4329 struct ext4_prealloc_space *cpa)
4331 ext4_fsblk_t cur_distance, new_distance;
4334 atomic_inc(&pa->pa_count);
4337 cur_distance = abs(goal_block - cpa->pa_pstart);
4338 new_distance = abs(goal_block - pa->pa_pstart);
4340 if (cur_distance <= new_distance)
4343 /* drop the previous reference */
4344 atomic_dec(&cpa->pa_count);
4345 atomic_inc(&pa->pa_count);
4350 * search goal blocks in preallocated space
4352 static noinline_for_stack bool
4353 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4355 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4357 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4358 struct ext4_locality_group *lg;
4359 struct ext4_prealloc_space *pa, *cpa = NULL;
4360 ext4_fsblk_t goal_block;
4362 /* only data can be preallocated */
4363 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4366 /* first, try per-file preallocation */
4368 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4370 /* all fields in this condition don't change,
4371 * so we can skip locking for them */
4372 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
4373 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
4374 EXT4_C2B(sbi, pa->pa_len)))
4377 /* non-extent files can't have physical blocks past 2^32 */
4378 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4379 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
4380 EXT4_MAX_BLOCK_FILE_PHYS))
4383 /* found preallocated blocks, use them */
4384 spin_lock(&pa->pa_lock);
4385 if (pa->pa_deleted == 0 && pa->pa_free) {
4386 atomic_inc(&pa->pa_count);
4387 ext4_mb_use_inode_pa(ac, pa);
4388 spin_unlock(&pa->pa_lock);
4389 ac->ac_criteria = 10;
4393 spin_unlock(&pa->pa_lock);
4397 /* can we use group allocation? */
4398 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4401 /* inode may have no locality group for some reason */
4405 order = fls(ac->ac_o_ex.fe_len) - 1;
4406 if (order > PREALLOC_TB_SIZE - 1)
4407 /* The max size of hash table is PREALLOC_TB_SIZE */
4408 order = PREALLOC_TB_SIZE - 1;
4410 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4412 * search for the prealloc space that is having
4413 * minimal distance from the goal block.
4415 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4417 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
4419 spin_lock(&pa->pa_lock);
4420 if (pa->pa_deleted == 0 &&
4421 pa->pa_free >= ac->ac_o_ex.fe_len) {
4423 cpa = ext4_mb_check_group_pa(goal_block,
4426 spin_unlock(&pa->pa_lock);
4431 ext4_mb_use_group_pa(ac, cpa);
4432 ac->ac_criteria = 20;
4439 * the function goes through all block freed in the group
4440 * but not yet committed and marks them used in in-core bitmap.
4441 * buddy must be generated from this bitmap
4442 * Need to be called with the ext4 group lock held
4444 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
4448 struct ext4_group_info *grp;
4449 struct ext4_free_data *entry;
4451 grp = ext4_get_group_info(sb, group);
4452 n = rb_first(&(grp->bb_free_root));
4455 entry = rb_entry(n, struct ext4_free_data, efd_node);
4456 mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
4463 * the function goes through all preallocation in this group and marks them
4464 * used in in-core bitmap. buddy must be generated from this bitmap
4465 * Need to be called with ext4 group lock held
4467 static noinline_for_stack
4468 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4471 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4472 struct ext4_prealloc_space *pa;
4473 struct list_head *cur;
4474 ext4_group_t groupnr;
4475 ext4_grpblk_t start;
4476 int preallocated = 0;
4479 /* all form of preallocation discards first load group,
4480 * so the only competing code is preallocation use.
4481 * we don't need any locking here
4482 * notice we do NOT ignore preallocations with pa_deleted
4483 * otherwise we could leave used blocks available for
4484 * allocation in buddy when concurrent ext4_mb_put_pa()
4485 * is dropping preallocation
4487 list_for_each(cur, &grp->bb_prealloc_list) {
4488 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
4489 spin_lock(&pa->pa_lock);
4490 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4493 spin_unlock(&pa->pa_lock);
4494 if (unlikely(len == 0))
4496 BUG_ON(groupnr != group);
4497 mb_set_bits(bitmap, start, len);
4498 preallocated += len;
4500 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
4503 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
4504 struct ext4_prealloc_space *pa)
4506 struct ext4_inode_info *ei;
4508 if (pa->pa_deleted) {
4509 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
4510 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
4517 if (pa->pa_type == MB_INODE_PA) {
4518 ei = EXT4_I(pa->pa_inode);
4519 atomic_dec(&ei->i_prealloc_active);
4523 static void ext4_mb_pa_callback(struct rcu_head *head)
4525 struct ext4_prealloc_space *pa;
4526 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
4528 BUG_ON(atomic_read(&pa->pa_count));
4529 BUG_ON(pa->pa_deleted == 0);
4530 kmem_cache_free(ext4_pspace_cachep, pa);
4534 * drops a reference to preallocated space descriptor
4535 * if this was the last reference and the space is consumed
4537 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
4538 struct super_block *sb, struct ext4_prealloc_space *pa)
4541 ext4_fsblk_t grp_blk;
4543 /* in this short window concurrent discard can set pa_deleted */
4544 spin_lock(&pa->pa_lock);
4545 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
4546 spin_unlock(&pa->pa_lock);
4550 if (pa->pa_deleted == 1) {
4551 spin_unlock(&pa->pa_lock);
4555 ext4_mb_mark_pa_deleted(sb, pa);
4556 spin_unlock(&pa->pa_lock);
4558 grp_blk = pa->pa_pstart;
4560 * If doing group-based preallocation, pa_pstart may be in the
4561 * next group when pa is used up
4563 if (pa->pa_type == MB_GROUP_PA)
4566 grp = ext4_get_group_number(sb, grp_blk);
4571 * P1 (buddy init) P2 (regular allocation)
4572 * find block B in PA
4573 * copy on-disk bitmap to buddy
4574 * mark B in on-disk bitmap
4575 * drop PA from group
4576 * mark all PAs in buddy
4578 * thus, P1 initializes buddy with B available. to prevent this
4579 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
4582 ext4_lock_group(sb, grp);
4583 list_del(&pa->pa_group_list);
4584 ext4_unlock_group(sb, grp);
4586 spin_lock(pa->pa_obj_lock);
4587 list_del_rcu(&pa->pa_inode_list);
4588 spin_unlock(pa->pa_obj_lock);
4590 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4594 * creates new preallocated space for given inode
4596 static noinline_for_stack void
4597 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
4599 struct super_block *sb = ac->ac_sb;
4600 struct ext4_sb_info *sbi = EXT4_SB(sb);
4601 struct ext4_prealloc_space *pa;
4602 struct ext4_group_info *grp;
4603 struct ext4_inode_info *ei;
4605 /* preallocate only when found space is larger then requested */
4606 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4607 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4608 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4609 BUG_ON(ac->ac_pa == NULL);
4613 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
4619 /* we can't allocate as much as normalizer wants.
4620 * so, found space must get proper lstart
4621 * to cover original request */
4622 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
4623 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
4625 /* we're limited by original request in that
4626 * logical block must be covered any way
4627 * winl is window we can move our chunk within */
4628 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
4630 /* also, we should cover whole original request */
4631 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
4633 /* the smallest one defines real window */
4634 win = min(winl, wins);
4636 offs = ac->ac_o_ex.fe_logical %
4637 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4638 if (offs && offs < win)
4641 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
4642 EXT4_NUM_B2C(sbi, win);
4643 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
4644 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
4647 /* preallocation can change ac_b_ex, thus we store actually
4648 * allocated blocks for history */
4649 ac->ac_f_ex = ac->ac_b_ex;
4651 pa->pa_lstart = ac->ac_b_ex.fe_logical;
4652 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4653 pa->pa_len = ac->ac_b_ex.fe_len;
4654 pa->pa_free = pa->pa_len;
4655 spin_lock_init(&pa->pa_lock);
4656 INIT_LIST_HEAD(&pa->pa_inode_list);
4657 INIT_LIST_HEAD(&pa->pa_group_list);
4659 pa->pa_type = MB_INODE_PA;
4661 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4662 pa->pa_len, pa->pa_lstart);
4663 trace_ext4_mb_new_inode_pa(ac, pa);
4665 ext4_mb_use_inode_pa(ac, pa);
4666 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
4668 ei = EXT4_I(ac->ac_inode);
4669 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4671 pa->pa_obj_lock = &ei->i_prealloc_lock;
4672 pa->pa_inode = ac->ac_inode;
4674 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4676 spin_lock(pa->pa_obj_lock);
4677 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
4678 spin_unlock(pa->pa_obj_lock);
4679 atomic_inc(&ei->i_prealloc_active);
4683 * creates new preallocated space for locality group inodes belongs to
4685 static noinline_for_stack void
4686 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
4688 struct super_block *sb = ac->ac_sb;
4689 struct ext4_locality_group *lg;
4690 struct ext4_prealloc_space *pa;
4691 struct ext4_group_info *grp;
4693 /* preallocate only when found space is larger then requested */
4694 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4695 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4696 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4697 BUG_ON(ac->ac_pa == NULL);
4701 /* preallocation can change ac_b_ex, thus we store actually
4702 * allocated blocks for history */
4703 ac->ac_f_ex = ac->ac_b_ex;
4705 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4706 pa->pa_lstart = pa->pa_pstart;
4707 pa->pa_len = ac->ac_b_ex.fe_len;
4708 pa->pa_free = pa->pa_len;
4709 spin_lock_init(&pa->pa_lock);
4710 INIT_LIST_HEAD(&pa->pa_inode_list);
4711 INIT_LIST_HEAD(&pa->pa_group_list);
4713 pa->pa_type = MB_GROUP_PA;
4715 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4716 pa->pa_len, pa->pa_lstart);
4717 trace_ext4_mb_new_group_pa(ac, pa);
4719 ext4_mb_use_group_pa(ac, pa);
4720 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4722 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4726 pa->pa_obj_lock = &lg->lg_prealloc_lock;
4727 pa->pa_inode = NULL;
4729 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4732 * We will later add the new pa to the right bucket
4733 * after updating the pa_free in ext4_mb_release_context
4737 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4739 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4740 ext4_mb_new_group_pa(ac);
4742 ext4_mb_new_inode_pa(ac);
4746 * finds all unused blocks in on-disk bitmap, frees them in
4747 * in-core bitmap and buddy.
4748 * @pa must be unlinked from inode and group lists, so that
4749 * nobody else can find/use it.
4750 * the caller MUST hold group/inode locks.
4751 * TODO: optimize the case when there are no in-core structures yet
4753 static noinline_for_stack int
4754 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4755 struct ext4_prealloc_space *pa)
4757 struct super_block *sb = e4b->bd_sb;
4758 struct ext4_sb_info *sbi = EXT4_SB(sb);
4763 unsigned long long grp_blk_start;
4766 BUG_ON(pa->pa_deleted == 0);
4767 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4768 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4769 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4770 end = bit + pa->pa_len;
4773 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4776 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4777 mb_debug(sb, "free preallocated %u/%u in group %u\n",
4778 (unsigned) ext4_group_first_block_no(sb, group) + bit,
4779 (unsigned) next - bit, (unsigned) group);
4782 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4783 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4784 EXT4_C2B(sbi, bit)),
4786 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4789 if (free != pa->pa_free) {
4790 ext4_msg(e4b->bd_sb, KERN_CRIT,
4791 "pa %p: logic %lu, phys. %lu, len %d",
4792 pa, (unsigned long) pa->pa_lstart,
4793 (unsigned long) pa->pa_pstart,
4795 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4798 * pa is already deleted so we use the value obtained
4799 * from the bitmap and continue.
4802 atomic_add(free, &sbi->s_mb_discarded);
4807 static noinline_for_stack int
4808 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4809 struct ext4_prealloc_space *pa)
4811 struct super_block *sb = e4b->bd_sb;
4815 trace_ext4_mb_release_group_pa(sb, pa);
4816 BUG_ON(pa->pa_deleted == 0);
4817 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4818 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4819 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4820 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4821 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4827 * releases all preallocations in given group
4829 * first, we need to decide discard policy:
4830 * - when do we discard
4832 * - how many do we discard
4833 * 1) how many requested
4835 static noinline_for_stack int
4836 ext4_mb_discard_group_preallocations(struct super_block *sb,
4837 ext4_group_t group, int *busy)
4839 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4840 struct buffer_head *bitmap_bh = NULL;
4841 struct ext4_prealloc_space *pa, *tmp;
4842 struct list_head list;
4843 struct ext4_buddy e4b;
4847 mb_debug(sb, "discard preallocation for group %u\n", group);
4848 if (list_empty(&grp->bb_prealloc_list))
4851 bitmap_bh = ext4_read_block_bitmap(sb, group);
4852 if (IS_ERR(bitmap_bh)) {
4853 err = PTR_ERR(bitmap_bh);
4854 ext4_error_err(sb, -err,
4855 "Error %d reading block bitmap for %u",
4860 err = ext4_mb_load_buddy(sb, group, &e4b);
4862 ext4_warning(sb, "Error %d loading buddy information for %u",
4868 INIT_LIST_HEAD(&list);
4869 ext4_lock_group(sb, group);
4870 list_for_each_entry_safe(pa, tmp,
4871 &grp->bb_prealloc_list, pa_group_list) {
4872 spin_lock(&pa->pa_lock);
4873 if (atomic_read(&pa->pa_count)) {
4874 spin_unlock(&pa->pa_lock);
4878 if (pa->pa_deleted) {
4879 spin_unlock(&pa->pa_lock);
4883 /* seems this one can be freed ... */
4884 ext4_mb_mark_pa_deleted(sb, pa);
4887 this_cpu_inc(discard_pa_seq);
4889 /* we can trust pa_free ... */
4890 free += pa->pa_free;
4892 spin_unlock(&pa->pa_lock);
4894 list_del(&pa->pa_group_list);
4895 list_add(&pa->u.pa_tmp_list, &list);
4898 /* now free all selected PAs */
4899 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4901 /* remove from object (inode or locality group) */
4902 spin_lock(pa->pa_obj_lock);
4903 list_del_rcu(&pa->pa_inode_list);
4904 spin_unlock(pa->pa_obj_lock);
4906 if (pa->pa_type == MB_GROUP_PA)
4907 ext4_mb_release_group_pa(&e4b, pa);
4909 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4911 list_del(&pa->u.pa_tmp_list);
4912 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4915 ext4_unlock_group(sb, group);
4916 ext4_mb_unload_buddy(&e4b);
4919 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4920 free, group, grp->bb_free);
4925 * releases all non-used preallocated blocks for given inode
4927 * It's important to discard preallocations under i_data_sem
4928 * We don't want another block to be served from the prealloc
4929 * space when we are discarding the inode prealloc space.
4931 * FIXME!! Make sure it is valid at all the call sites
4933 void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
4935 struct ext4_inode_info *ei = EXT4_I(inode);
4936 struct super_block *sb = inode->i_sb;
4937 struct buffer_head *bitmap_bh = NULL;
4938 struct ext4_prealloc_space *pa, *tmp;
4939 ext4_group_t group = 0;
4940 struct list_head list;
4941 struct ext4_buddy e4b;
4944 if (!S_ISREG(inode->i_mode)) {
4945 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4949 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
4952 mb_debug(sb, "discard preallocation for inode %lu\n",
4954 trace_ext4_discard_preallocations(inode,
4955 atomic_read(&ei->i_prealloc_active), needed);
4957 INIT_LIST_HEAD(&list);
4963 /* first, collect all pa's in the inode */
4964 spin_lock(&ei->i_prealloc_lock);
4965 while (!list_empty(&ei->i_prealloc_list) && needed) {
4966 pa = list_entry(ei->i_prealloc_list.prev,
4967 struct ext4_prealloc_space, pa_inode_list);
4968 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4969 spin_lock(&pa->pa_lock);
4970 if (atomic_read(&pa->pa_count)) {
4971 /* this shouldn't happen often - nobody should
4972 * use preallocation while we're discarding it */
4973 spin_unlock(&pa->pa_lock);
4974 spin_unlock(&ei->i_prealloc_lock);
4975 ext4_msg(sb, KERN_ERR,
4976 "uh-oh! used pa while discarding");
4978 schedule_timeout_uninterruptible(HZ);
4982 if (pa->pa_deleted == 0) {
4983 ext4_mb_mark_pa_deleted(sb, pa);
4984 spin_unlock(&pa->pa_lock);
4985 list_del_rcu(&pa->pa_inode_list);
4986 list_add(&pa->u.pa_tmp_list, &list);
4991 /* someone is deleting pa right now */
4992 spin_unlock(&pa->pa_lock);
4993 spin_unlock(&ei->i_prealloc_lock);
4995 /* we have to wait here because pa_deleted
4996 * doesn't mean pa is already unlinked from
4997 * the list. as we might be called from
4998 * ->clear_inode() the inode will get freed
4999 * and concurrent thread which is unlinking
5000 * pa from inode's list may access already
5001 * freed memory, bad-bad-bad */
5003 /* XXX: if this happens too often, we can
5004 * add a flag to force wait only in case
5005 * of ->clear_inode(), but not in case of
5006 * regular truncate */
5007 schedule_timeout_uninterruptible(HZ);
5010 spin_unlock(&ei->i_prealloc_lock);
5012 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5013 BUG_ON(pa->pa_type != MB_INODE_PA);
5014 group = ext4_get_group_number(sb, pa->pa_pstart);
5016 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5017 GFP_NOFS|__GFP_NOFAIL);
5019 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5024 bitmap_bh = ext4_read_block_bitmap(sb, group);
5025 if (IS_ERR(bitmap_bh)) {
5026 err = PTR_ERR(bitmap_bh);
5027 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
5029 ext4_mb_unload_buddy(&e4b);
5033 ext4_lock_group(sb, group);
5034 list_del(&pa->pa_group_list);
5035 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5036 ext4_unlock_group(sb, group);
5038 ext4_mb_unload_buddy(&e4b);
5041 list_del(&pa->u.pa_tmp_list);
5042 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5046 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
5048 struct ext4_prealloc_space *pa;
5050 BUG_ON(ext4_pspace_cachep == NULL);
5051 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
5054 atomic_set(&pa->pa_count, 1);
5059 static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
5061 struct ext4_prealloc_space *pa = ac->ac_pa;
5065 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
5066 kmem_cache_free(ext4_pspace_cachep, pa);
5069 #ifdef CONFIG_EXT4_DEBUG
5070 static inline void ext4_mb_show_pa(struct super_block *sb)
5072 ext4_group_t i, ngroups;
5074 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5077 ngroups = ext4_get_groups_count(sb);
5078 mb_debug(sb, "groups: ");
5079 for (i = 0; i < ngroups; i++) {
5080 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5081 struct ext4_prealloc_space *pa;
5082 ext4_grpblk_t start;
5083 struct list_head *cur;
5084 ext4_lock_group(sb, i);
5085 list_for_each(cur, &grp->bb_prealloc_list) {
5086 pa = list_entry(cur, struct ext4_prealloc_space,
5088 spin_lock(&pa->pa_lock);
5089 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5091 spin_unlock(&pa->pa_lock);
5092 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5095 ext4_unlock_group(sb, i);
5096 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5101 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5103 struct super_block *sb = ac->ac_sb;
5105 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5108 mb_debug(sb, "Can't allocate:"
5109 " Allocation context details:");
5110 mb_debug(sb, "status %u flags 0x%x",
5111 ac->ac_status, ac->ac_flags);
5112 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5113 "goal %lu/%lu/%lu@%lu, "
5114 "best %lu/%lu/%lu@%lu cr %d",
5115 (unsigned long)ac->ac_o_ex.fe_group,
5116 (unsigned long)ac->ac_o_ex.fe_start,
5117 (unsigned long)ac->ac_o_ex.fe_len,
5118 (unsigned long)ac->ac_o_ex.fe_logical,
5119 (unsigned long)ac->ac_g_ex.fe_group,
5120 (unsigned long)ac->ac_g_ex.fe_start,
5121 (unsigned long)ac->ac_g_ex.fe_len,
5122 (unsigned long)ac->ac_g_ex.fe_logical,
5123 (unsigned long)ac->ac_b_ex.fe_group,
5124 (unsigned long)ac->ac_b_ex.fe_start,
5125 (unsigned long)ac->ac_b_ex.fe_len,
5126 (unsigned long)ac->ac_b_ex.fe_logical,
5127 (int)ac->ac_criteria);
5128 mb_debug(sb, "%u found", ac->ac_found);
5129 ext4_mb_show_pa(sb);
5132 static inline void ext4_mb_show_pa(struct super_block *sb)
5136 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5138 ext4_mb_show_pa(ac->ac_sb);
5144 * We use locality group preallocation for small size file. The size of the
5145 * file is determined by the current size or the resulting size after
5146 * allocation which ever is larger
5148 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5150 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5152 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5153 int bsbits = ac->ac_sb->s_blocksize_bits;
5155 bool inode_pa_eligible, group_pa_eligible;
5157 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5160 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5163 group_pa_eligible = sbi->s_mb_group_prealloc > 0;
5164 inode_pa_eligible = true;
5165 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
5166 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5169 /* No point in using inode preallocation for closed files */
5170 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5171 !inode_is_open_for_write(ac->ac_inode))
5172 inode_pa_eligible = false;
5174 size = max(size, isize);
5175 /* Don't use group allocation for large files */
5176 if (size > sbi->s_mb_stream_request)
5177 group_pa_eligible = false;
5179 if (!group_pa_eligible) {
5180 if (inode_pa_eligible)
5181 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5183 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5187 BUG_ON(ac->ac_lg != NULL);
5189 * locality group prealloc space are per cpu. The reason for having
5190 * per cpu locality group is to reduce the contention between block
5191 * request from multiple CPUs.
5193 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5195 /* we're going to use group allocation */
5196 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5198 /* serialize all allocations in the group */
5199 mutex_lock(&ac->ac_lg->lg_mutex);
5202 static noinline_for_stack int
5203 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5204 struct ext4_allocation_request *ar)
5206 struct super_block *sb = ar->inode->i_sb;
5207 struct ext4_sb_info *sbi = EXT4_SB(sb);
5208 struct ext4_super_block *es = sbi->s_es;
5212 ext4_grpblk_t block;
5214 /* we can't allocate > group size */
5217 /* just a dirty hack to filter too big requests */
5218 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5219 len = EXT4_CLUSTERS_PER_GROUP(sb);
5221 /* start searching from the goal */
5223 if (goal < le32_to_cpu(es->s_first_data_block) ||
5224 goal >= ext4_blocks_count(es))
5225 goal = le32_to_cpu(es->s_first_data_block);
5226 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5228 /* set up allocation goals */
5229 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5230 ac->ac_status = AC_STATUS_CONTINUE;
5232 ac->ac_inode = ar->inode;
5233 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5234 ac->ac_o_ex.fe_group = group;
5235 ac->ac_o_ex.fe_start = block;
5236 ac->ac_o_ex.fe_len = len;
5237 ac->ac_g_ex = ac->ac_o_ex;
5238 ac->ac_flags = ar->flags;
5240 /* we have to define context: we'll work with a file or
5241 * locality group. this is a policy, actually */
5242 ext4_mb_group_or_file(ac);
5244 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5245 "left: %u/%u, right %u/%u to %swritable\n",
5246 (unsigned) ar->len, (unsigned) ar->logical,
5247 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5248 (unsigned) ar->lleft, (unsigned) ar->pleft,
5249 (unsigned) ar->lright, (unsigned) ar->pright,
5250 inode_is_open_for_write(ar->inode) ? "" : "non-");
5255 static noinline_for_stack void
5256 ext4_mb_discard_lg_preallocations(struct super_block *sb,
5257 struct ext4_locality_group *lg,
5258 int order, int total_entries)
5260 ext4_group_t group = 0;
5261 struct ext4_buddy e4b;
5262 struct list_head discard_list;
5263 struct ext4_prealloc_space *pa, *tmp;
5265 mb_debug(sb, "discard locality group preallocation\n");
5267 INIT_LIST_HEAD(&discard_list);
5269 spin_lock(&lg->lg_prealloc_lock);
5270 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5272 lockdep_is_held(&lg->lg_prealloc_lock)) {
5273 spin_lock(&pa->pa_lock);
5274 if (atomic_read(&pa->pa_count)) {
5276 * This is the pa that we just used
5277 * for block allocation. So don't
5280 spin_unlock(&pa->pa_lock);
5283 if (pa->pa_deleted) {
5284 spin_unlock(&pa->pa_lock);
5287 /* only lg prealloc space */
5288 BUG_ON(pa->pa_type != MB_GROUP_PA);
5290 /* seems this one can be freed ... */
5291 ext4_mb_mark_pa_deleted(sb, pa);
5292 spin_unlock(&pa->pa_lock);
5294 list_del_rcu(&pa->pa_inode_list);
5295 list_add(&pa->u.pa_tmp_list, &discard_list);
5298 if (total_entries <= 5) {
5300 * we want to keep only 5 entries
5301 * allowing it to grow to 8. This
5302 * mak sure we don't call discard
5303 * soon for this list.
5308 spin_unlock(&lg->lg_prealloc_lock);
5310 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5313 group = ext4_get_group_number(sb, pa->pa_pstart);
5314 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5315 GFP_NOFS|__GFP_NOFAIL);
5317 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5321 ext4_lock_group(sb, group);
5322 list_del(&pa->pa_group_list);
5323 ext4_mb_release_group_pa(&e4b, pa);
5324 ext4_unlock_group(sb, group);
5326 ext4_mb_unload_buddy(&e4b);
5327 list_del(&pa->u.pa_tmp_list);
5328 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5333 * We have incremented pa_count. So it cannot be freed at this
5334 * point. Also we hold lg_mutex. So no parallel allocation is
5335 * possible from this lg. That means pa_free cannot be updated.
5337 * A parallel ext4_mb_discard_group_preallocations is possible.
5338 * which can cause the lg_prealloc_list to be updated.
5341 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5343 int order, added = 0, lg_prealloc_count = 1;
5344 struct super_block *sb = ac->ac_sb;
5345 struct ext4_locality_group *lg = ac->ac_lg;
5346 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5348 order = fls(pa->pa_free) - 1;
5349 if (order > PREALLOC_TB_SIZE - 1)
5350 /* The max size of hash table is PREALLOC_TB_SIZE */
5351 order = PREALLOC_TB_SIZE - 1;
5352 /* Add the prealloc space to lg */
5353 spin_lock(&lg->lg_prealloc_lock);
5354 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5356 lockdep_is_held(&lg->lg_prealloc_lock)) {
5357 spin_lock(&tmp_pa->pa_lock);
5358 if (tmp_pa->pa_deleted) {
5359 spin_unlock(&tmp_pa->pa_lock);
5362 if (!added && pa->pa_free < tmp_pa->pa_free) {
5363 /* Add to the tail of the previous entry */
5364 list_add_tail_rcu(&pa->pa_inode_list,
5365 &tmp_pa->pa_inode_list);
5368 * we want to count the total
5369 * number of entries in the list
5372 spin_unlock(&tmp_pa->pa_lock);
5373 lg_prealloc_count++;
5376 list_add_tail_rcu(&pa->pa_inode_list,
5377 &lg->lg_prealloc_list[order]);
5378 spin_unlock(&lg->lg_prealloc_lock);
5380 /* Now trim the list to be not more than 8 elements */
5381 if (lg_prealloc_count > 8) {
5382 ext4_mb_discard_lg_preallocations(sb, lg,
5383 order, lg_prealloc_count);
5390 * if per-inode prealloc list is too long, trim some PA
5392 static void ext4_mb_trim_inode_pa(struct inode *inode)
5394 struct ext4_inode_info *ei = EXT4_I(inode);
5395 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5398 count = atomic_read(&ei->i_prealloc_active);
5399 delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
5400 if (count > sbi->s_mb_max_inode_prealloc + delta) {
5401 count -= sbi->s_mb_max_inode_prealloc;
5402 ext4_discard_preallocations(inode, count);
5407 * release all resource we used in allocation
5409 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
5411 struct inode *inode = ac->ac_inode;
5412 struct ext4_inode_info *ei = EXT4_I(inode);
5413 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5414 struct ext4_prealloc_space *pa = ac->ac_pa;
5416 if (pa->pa_type == MB_GROUP_PA) {
5417 /* see comment in ext4_mb_use_group_pa() */
5418 spin_lock(&pa->pa_lock);
5419 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5420 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5421 pa->pa_free -= ac->ac_b_ex.fe_len;
5422 pa->pa_len -= ac->ac_b_ex.fe_len;
5423 spin_unlock(&pa->pa_lock);
5426 * We want to add the pa to the right bucket.
5427 * Remove it from the list and while adding
5428 * make sure the list to which we are adding
5431 if (likely(pa->pa_free)) {
5432 spin_lock(pa->pa_obj_lock);
5433 list_del_rcu(&pa->pa_inode_list);
5434 spin_unlock(pa->pa_obj_lock);
5435 ext4_mb_add_n_trim(ac);
5439 if (pa->pa_type == MB_INODE_PA) {
5441 * treat per-inode prealloc list as a lru list, then try
5442 * to trim the least recently used PA.
5444 spin_lock(pa->pa_obj_lock);
5445 list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
5446 spin_unlock(pa->pa_obj_lock);
5449 ext4_mb_put_pa(ac, ac->ac_sb, pa);
5451 if (ac->ac_bitmap_page)
5452 put_page(ac->ac_bitmap_page);
5453 if (ac->ac_buddy_page)
5454 put_page(ac->ac_buddy_page);
5455 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5456 mutex_unlock(&ac->ac_lg->lg_mutex);
5457 ext4_mb_collect_stats(ac);
5458 ext4_mb_trim_inode_pa(inode);
5462 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
5464 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
5466 int freed = 0, busy = 0;
5469 trace_ext4_mb_discard_preallocations(sb, needed);
5472 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
5474 for (i = 0; i < ngroups && needed > 0; i++) {
5475 ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
5481 if (needed > 0 && busy && ++retry < 3) {
5489 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
5490 struct ext4_allocation_context *ac, u64 *seq)
5496 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
5501 seq_retry = ext4_get_discard_pa_seq_sum();
5502 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
5503 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
5509 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
5513 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5514 struct ext4_allocation_request *ar, int *errp);
5517 * Main entry point into mballoc to allocate blocks
5518 * it tries to use preallocation first, then falls back
5519 * to usual allocation
5521 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
5522 struct ext4_allocation_request *ar, int *errp)
5524 struct ext4_allocation_context *ac = NULL;
5525 struct ext4_sb_info *sbi;
5526 struct super_block *sb;
5527 ext4_fsblk_t block = 0;
5528 unsigned int inquota = 0;
5529 unsigned int reserv_clstrs = 0;
5534 sb = ar->inode->i_sb;
5537 trace_ext4_request_blocks(ar);
5538 if (sbi->s_mount_state & EXT4_FC_REPLAY)
5539 return ext4_mb_new_blocks_simple(handle, ar, errp);
5541 /* Allow to use superuser reservation for quota file */
5542 if (ext4_is_quota_file(ar->inode))
5543 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
5545 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
5546 /* Without delayed allocation we need to verify
5547 * there is enough free blocks to do block allocation
5548 * and verify allocation doesn't exceed the quota limits.
5551 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
5553 /* let others to free the space */
5555 ar->len = ar->len >> 1;
5558 ext4_mb_show_pa(sb);
5562 reserv_clstrs = ar->len;
5563 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
5564 dquot_alloc_block_nofail(ar->inode,
5565 EXT4_C2B(sbi, ar->len));
5568 dquot_alloc_block(ar->inode,
5569 EXT4_C2B(sbi, ar->len))) {
5571 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
5582 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
5589 *errp = ext4_mb_initialize_context(ac, ar);
5595 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
5596 seq = this_cpu_read(discard_pa_seq);
5597 if (!ext4_mb_use_preallocated(ac)) {
5598 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
5599 ext4_mb_normalize_request(ac, ar);
5601 *errp = ext4_mb_pa_alloc(ac);
5605 /* allocate space in core */
5606 *errp = ext4_mb_regular_allocator(ac);
5608 * pa allocated above is added to grp->bb_prealloc_list only
5609 * when we were able to allocate some block i.e. when
5610 * ac->ac_status == AC_STATUS_FOUND.
5611 * And error from above mean ac->ac_status != AC_STATUS_FOUND
5612 * So we have to free this pa here itself.
5615 ext4_mb_pa_free(ac);
5616 ext4_discard_allocated_blocks(ac);
5619 if (ac->ac_status == AC_STATUS_FOUND &&
5620 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
5621 ext4_mb_pa_free(ac);
5623 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
5624 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
5626 ext4_discard_allocated_blocks(ac);
5629 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5630 ar->len = ac->ac_b_ex.fe_len;
5633 if (++retries < 3 &&
5634 ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
5637 * If block allocation fails then the pa allocated above
5638 * needs to be freed here itself.
5640 ext4_mb_pa_free(ac);
5646 ac->ac_b_ex.fe_len = 0;
5648 ext4_mb_show_ac(ac);
5650 ext4_mb_release_context(ac);
5653 kmem_cache_free(ext4_ac_cachep, ac);
5654 if (inquota && ar->len < inquota)
5655 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
5657 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
5658 /* release all the reserved blocks if non delalloc */
5659 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
5663 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
5669 * We can merge two free data extents only if the physical blocks
5670 * are contiguous, AND the extents were freed by the same transaction,
5671 * AND the blocks are associated with the same group.
5673 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
5674 struct ext4_free_data *entry,
5675 struct ext4_free_data *new_entry,
5676 struct rb_root *entry_rb_root)
5678 if ((entry->efd_tid != new_entry->efd_tid) ||
5679 (entry->efd_group != new_entry->efd_group))
5681 if (entry->efd_start_cluster + entry->efd_count ==
5682 new_entry->efd_start_cluster) {
5683 new_entry->efd_start_cluster = entry->efd_start_cluster;
5684 new_entry->efd_count += entry->efd_count;
5685 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
5686 entry->efd_start_cluster) {
5687 new_entry->efd_count += entry->efd_count;
5690 spin_lock(&sbi->s_md_lock);
5691 list_del(&entry->efd_list);
5692 spin_unlock(&sbi->s_md_lock);
5693 rb_erase(&entry->efd_node, entry_rb_root);
5694 kmem_cache_free(ext4_free_data_cachep, entry);
5697 static noinline_for_stack int
5698 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
5699 struct ext4_free_data *new_entry)
5701 ext4_group_t group = e4b->bd_group;
5702 ext4_grpblk_t cluster;
5703 ext4_grpblk_t clusters = new_entry->efd_count;
5704 struct ext4_free_data *entry;
5705 struct ext4_group_info *db = e4b->bd_info;
5706 struct super_block *sb = e4b->bd_sb;
5707 struct ext4_sb_info *sbi = EXT4_SB(sb);
5708 struct rb_node **n = &db->bb_free_root.rb_node, *node;
5709 struct rb_node *parent = NULL, *new_node;
5711 BUG_ON(!ext4_handle_valid(handle));
5712 BUG_ON(e4b->bd_bitmap_page == NULL);
5713 BUG_ON(e4b->bd_buddy_page == NULL);
5715 new_node = &new_entry->efd_node;
5716 cluster = new_entry->efd_start_cluster;
5719 /* first free block exent. We need to
5720 protect buddy cache from being freed,
5721 * otherwise we'll refresh it from
5722 * on-disk bitmap and lose not-yet-available
5724 get_page(e4b->bd_buddy_page);
5725 get_page(e4b->bd_bitmap_page);
5729 entry = rb_entry(parent, struct ext4_free_data, efd_node);
5730 if (cluster < entry->efd_start_cluster)
5732 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
5733 n = &(*n)->rb_right;
5735 ext4_grp_locked_error(sb, group, 0,
5736 ext4_group_first_block_no(sb, group) +
5737 EXT4_C2B(sbi, cluster),
5738 "Block already on to-be-freed list");
5739 kmem_cache_free(ext4_free_data_cachep, new_entry);
5744 rb_link_node(new_node, parent, n);
5745 rb_insert_color(new_node, &db->bb_free_root);
5747 /* Now try to see the extent can be merged to left and right */
5748 node = rb_prev(new_node);
5750 entry = rb_entry(node, struct ext4_free_data, efd_node);
5751 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5752 &(db->bb_free_root));
5755 node = rb_next(new_node);
5757 entry = rb_entry(node, struct ext4_free_data, efd_node);
5758 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5759 &(db->bb_free_root));
5762 spin_lock(&sbi->s_md_lock);
5763 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5764 sbi->s_mb_free_pending += clusters;
5765 spin_unlock(&sbi->s_md_lock);
5770 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
5771 * linearly starting at the goal block and also excludes the blocks which
5772 * are going to be in use after fast commit replay.
5774 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5775 struct ext4_allocation_request *ar, int *errp)
5777 struct buffer_head *bitmap_bh;
5778 struct super_block *sb = ar->inode->i_sb;
5780 ext4_grpblk_t blkoff;
5781 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
5782 ext4_grpblk_t i = 0;
5783 ext4_fsblk_t goal, block;
5784 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
5787 if (goal < le32_to_cpu(es->s_first_data_block) ||
5788 goal >= ext4_blocks_count(es))
5789 goal = le32_to_cpu(es->s_first_data_block);
5792 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
5793 for (; group < ext4_get_groups_count(sb); group++) {
5794 bitmap_bh = ext4_read_block_bitmap(sb, group);
5795 if (IS_ERR(bitmap_bh)) {
5796 *errp = PTR_ERR(bitmap_bh);
5797 pr_warn("Failed to read block bitmap\n");
5801 ext4_get_group_no_and_offset(sb,
5802 max(ext4_group_first_block_no(sb, group), goal),
5805 i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
5809 if (ext4_fc_replay_check_excluded(sb,
5810 ext4_group_first_block_no(sb, group) + i)) {
5820 if (group >= ext4_get_groups_count(sb) || i >= max) {
5825 block = ext4_group_first_block_no(sb, group) + i;
5826 ext4_mb_mark_bb(sb, block, 1, 1);
5832 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
5833 unsigned long count)
5835 struct buffer_head *bitmap_bh;
5836 struct super_block *sb = inode->i_sb;
5837 struct ext4_group_desc *gdp;
5838 struct buffer_head *gdp_bh;
5840 ext4_grpblk_t blkoff;
5841 int already_freed = 0, err, i;
5843 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
5844 bitmap_bh = ext4_read_block_bitmap(sb, group);
5845 if (IS_ERR(bitmap_bh)) {
5846 err = PTR_ERR(bitmap_bh);
5847 pr_warn("Failed to read block bitmap\n");
5850 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
5854 for (i = 0; i < count; i++) {
5855 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data))
5858 mb_clear_bits(bitmap_bh->b_data, blkoff, count);
5859 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
5862 ext4_free_group_clusters_set(
5863 sb, gdp, ext4_free_group_clusters(sb, gdp) +
5864 count - already_freed);
5865 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
5866 ext4_group_desc_csum_set(sb, group, gdp);
5867 ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
5868 sync_dirty_buffer(bitmap_bh);
5869 sync_dirty_buffer(gdp_bh);
5874 * ext4_mb_clear_bb() -- helper function for freeing blocks.
5875 * Used by ext4_free_blocks()
5876 * @handle: handle for this transaction
5878 * @block: starting physical block to be freed
5879 * @count: number of blocks to be freed
5880 * @flags: flags used by ext4_free_blocks
5882 static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
5883 ext4_fsblk_t block, unsigned long count,
5886 struct buffer_head *bitmap_bh = NULL;
5887 struct super_block *sb = inode->i_sb;
5888 struct ext4_group_desc *gdp;
5889 unsigned int overflow;
5891 struct buffer_head *gd_bh;
5892 ext4_group_t block_group;
5893 struct ext4_sb_info *sbi;
5894 struct ext4_buddy e4b;
5895 unsigned int count_clusters;
5901 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5902 !ext4_inode_block_valid(inode, block, count)) {
5903 ext4_error(sb, "Freeing blocks in system zone - "
5904 "Block = %llu, count = %lu", block, count);
5905 /* err = 0. ext4_std_error should be a no op */
5908 flags |= EXT4_FREE_BLOCKS_VALIDATED;
5912 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5914 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5915 ext4_get_group_info(sb, block_group))))
5919 * Check to see if we are freeing blocks across a group
5922 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5923 overflow = EXT4_C2B(sbi, bit) + count -
5924 EXT4_BLOCKS_PER_GROUP(sb);
5926 /* The range changed so it's no longer validated */
5927 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
5929 count_clusters = EXT4_NUM_B2C(sbi, count);
5930 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5931 if (IS_ERR(bitmap_bh)) {
5932 err = PTR_ERR(bitmap_bh);
5936 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5942 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5943 !ext4_inode_block_valid(inode, block, count)) {
5944 ext4_error(sb, "Freeing blocks in system zone - "
5945 "Block = %llu, count = %lu", block, count);
5946 /* err = 0. ext4_std_error should be a no op */
5950 BUFFER_TRACE(bitmap_bh, "getting write access");
5951 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
5957 * We are about to modify some metadata. Call the journal APIs
5958 * to unshare ->b_data if a currently-committing transaction is
5961 BUFFER_TRACE(gd_bh, "get_write_access");
5962 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
5965 #ifdef AGGRESSIVE_CHECK
5968 for (i = 0; i < count_clusters; i++)
5969 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
5972 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
5974 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
5975 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
5976 GFP_NOFS|__GFP_NOFAIL);
5981 * We need to make sure we don't reuse the freed block until after the
5982 * transaction is committed. We make an exception if the inode is to be
5983 * written in writeback mode since writeback mode has weak data
5984 * consistency guarantees.
5986 if (ext4_handle_valid(handle) &&
5987 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
5988 !ext4_should_writeback_data(inode))) {
5989 struct ext4_free_data *new_entry;
5991 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
5994 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
5995 GFP_NOFS|__GFP_NOFAIL);
5996 new_entry->efd_start_cluster = bit;
5997 new_entry->efd_group = block_group;
5998 new_entry->efd_count = count_clusters;
5999 new_entry->efd_tid = handle->h_transaction->t_tid;
6001 ext4_lock_group(sb, block_group);
6002 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6003 ext4_mb_free_metadata(handle, &e4b, new_entry);
6005 /* need to update group_info->bb_free and bitmap
6006 * with group lock held. generate_buddy look at
6007 * them with group lock_held
6009 if (test_opt(sb, DISCARD)) {
6010 err = ext4_issue_discard(sb, block_group, bit, count,
6012 if (err && err != -EOPNOTSUPP)
6013 ext4_msg(sb, KERN_WARNING, "discard request in"
6014 " group:%u block:%d count:%lu failed"
6015 " with %d", block_group, bit, count,
6018 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
6020 ext4_lock_group(sb, block_group);
6021 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6022 mb_free_blocks(inode, &e4b, bit, count_clusters);
6025 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
6026 ext4_free_group_clusters_set(sb, gdp, ret);
6027 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
6028 ext4_group_desc_csum_set(sb, block_group, gdp);
6029 ext4_unlock_group(sb, block_group);
6031 if (sbi->s_log_groups_per_flex) {
6032 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6033 atomic64_add(count_clusters,
6034 &sbi_array_rcu_deref(sbi, s_flex_groups,
6035 flex_group)->free_clusters);
6039 * on a bigalloc file system, defer the s_freeclusters_counter
6040 * update to the caller (ext4_remove_space and friends) so they
6041 * can determine if a cluster freed here should be rereserved
6043 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6044 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6045 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6046 percpu_counter_add(&sbi->s_freeclusters_counter,
6050 ext4_mb_unload_buddy(&e4b);
6052 /* We dirtied the bitmap block */
6053 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6054 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6056 /* And the group descriptor block */
6057 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6058 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6062 if (overflow && !err) {
6066 /* The range changed so it's no longer validated */
6067 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6072 ext4_std_error(sb, err);
6077 * ext4_free_blocks() -- Free given blocks and update quota
6078 * @handle: handle for this transaction
6080 * @bh: optional buffer of the block to be freed
6081 * @block: starting physical block to be freed
6082 * @count: number of blocks to be freed
6083 * @flags: flags used by ext4_free_blocks
6085 void ext4_free_blocks(handle_t *handle, struct inode *inode,
6086 struct buffer_head *bh, ext4_fsblk_t block,
6087 unsigned long count, int flags)
6089 struct super_block *sb = inode->i_sb;
6090 unsigned int overflow;
6091 struct ext4_sb_info *sbi;
6095 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
6096 ext4_free_blocks_simple(inode, block, count);
6103 BUG_ON(block != bh->b_blocknr);
6105 block = bh->b_blocknr;
6108 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6109 !ext4_inode_block_valid(inode, block, count)) {
6110 ext4_error(sb, "Freeing blocks not in datazone - "
6111 "block = %llu, count = %lu", block, count);
6114 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6116 ext4_debug("freeing block %llu\n", block);
6117 trace_ext4_free_blocks(inode, block, count, flags);
6119 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6122 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
6127 * If the extent to be freed does not begin on a cluster
6128 * boundary, we need to deal with partial clusters at the
6129 * beginning and end of the extent. Normally we will free
6130 * blocks at the beginning or the end unless we are explicitly
6131 * requested to avoid doing so.
6133 overflow = EXT4_PBLK_COFF(sbi, block);
6135 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
6136 overflow = sbi->s_cluster_ratio - overflow;
6138 if (count > overflow)
6146 /* The range changed so it's no longer validated */
6147 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6149 overflow = EXT4_LBLK_COFF(sbi, count);
6151 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
6152 if (count > overflow)
6157 count += sbi->s_cluster_ratio - overflow;
6158 /* The range changed so it's no longer validated */
6159 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6162 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6164 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
6166 for (i = 0; i < count; i++) {
6169 bh = sb_find_get_block(inode->i_sb, block + i);
6170 ext4_forget(handle, is_metadata, inode, bh, block + i);
6174 ext4_mb_clear_bb(handle, inode, block, count, flags);
6179 * ext4_group_add_blocks() -- Add given blocks to an existing group
6180 * @handle: handle to this transaction
6182 * @block: start physical block to add to the block group
6183 * @count: number of blocks to free
6185 * This marks the blocks as free in the bitmap and buddy.
6187 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6188 ext4_fsblk_t block, unsigned long count)
6190 struct buffer_head *bitmap_bh = NULL;
6191 struct buffer_head *gd_bh;
6192 ext4_group_t block_group;
6195 struct ext4_group_desc *desc;
6196 struct ext4_sb_info *sbi = EXT4_SB(sb);
6197 struct ext4_buddy e4b;
6198 int err = 0, ret, free_clusters_count;
6199 ext4_grpblk_t clusters_freed;
6200 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6201 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6202 unsigned long cluster_count = last_cluster - first_cluster + 1;
6204 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6209 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6211 * Check to see if we are freeing blocks across a group
6214 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6215 ext4_warning(sb, "too many blocks added to group %u",
6221 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
6222 if (IS_ERR(bitmap_bh)) {
6223 err = PTR_ERR(bitmap_bh);
6228 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
6234 if (!ext4_sb_block_valid(sb, NULL, block, count)) {
6235 ext4_error(sb, "Adding blocks in system zones - "
6236 "Block = %llu, count = %lu",
6242 BUFFER_TRACE(bitmap_bh, "getting write access");
6243 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
6249 * We are about to modify some metadata. Call the journal APIs
6250 * to unshare ->b_data if a currently-committing transaction is
6253 BUFFER_TRACE(gd_bh, "get_write_access");
6254 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
6258 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
6259 BUFFER_TRACE(bitmap_bh, "clear bit");
6260 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
6261 ext4_error(sb, "bit already cleared for block %llu",
6262 (ext4_fsblk_t)(block + i));
6263 BUFFER_TRACE(bitmap_bh, "bit already cleared");
6269 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6274 * need to update group_info->bb_free and bitmap
6275 * with group lock held. generate_buddy look at
6276 * them with group lock_held
6278 ext4_lock_group(sb, block_group);
6279 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
6280 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6281 free_clusters_count = clusters_freed +
6282 ext4_free_group_clusters(sb, desc);
6283 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
6284 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
6285 ext4_group_desc_csum_set(sb, block_group, desc);
6286 ext4_unlock_group(sb, block_group);
6287 percpu_counter_add(&sbi->s_freeclusters_counter,
6290 if (sbi->s_log_groups_per_flex) {
6291 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6292 atomic64_add(clusters_freed,
6293 &sbi_array_rcu_deref(sbi, s_flex_groups,
6294 flex_group)->free_clusters);
6297 ext4_mb_unload_buddy(&e4b);
6299 /* We dirtied the bitmap block */
6300 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6301 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6303 /* And the group descriptor block */
6304 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6305 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6311 ext4_std_error(sb, err);
6316 * ext4_trim_extent -- function to TRIM one single free extent in the group
6317 * @sb: super block for the file system
6318 * @start: starting block of the free extent in the alloc. group
6319 * @count: number of blocks to TRIM
6320 * @e4b: ext4 buddy for the group
6322 * Trim "count" blocks starting at "start" in the "group". To assure that no
6323 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6324 * be called with under the group lock.
6326 static int ext4_trim_extent(struct super_block *sb,
6327 int start, int count, struct ext4_buddy *e4b)
6331 struct ext4_free_extent ex;
6332 ext4_group_t group = e4b->bd_group;
6335 trace_ext4_trim_extent(sb, group, start, count);
6337 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6339 ex.fe_start = start;
6340 ex.fe_group = group;
6344 * Mark blocks used, so no one can reuse them while
6347 mb_mark_used(e4b, &ex);
6348 ext4_unlock_group(sb, group);
6349 ret = ext4_issue_discard(sb, group, start, count, NULL);
6350 ext4_lock_group(sb, group);
6351 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6355 static int ext4_try_to_trim_range(struct super_block *sb,
6356 struct ext4_buddy *e4b, ext4_grpblk_t start,
6357 ext4_grpblk_t max, ext4_grpblk_t minblocks)
6358 __acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
6359 __releases(ext4_group_lock_ptr(sb, e4b->bd_group))
6361 ext4_grpblk_t next, count, free_count;
6364 bitmap = e4b->bd_bitmap;
6365 start = (e4b->bd_info->bb_first_free > start) ?
6366 e4b->bd_info->bb_first_free : start;
6370 while (start <= max) {
6371 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6374 next = mb_find_next_bit(bitmap, max + 1, start);
6376 if ((next - start) >= minblocks) {
6377 int ret = ext4_trim_extent(sb, start, next - start, e4b);
6379 if (ret && ret != -EOPNOTSUPP)
6381 count += next - start;
6383 free_count += next - start;
6386 if (fatal_signal_pending(current)) {
6387 count = -ERESTARTSYS;
6391 if (need_resched()) {
6392 ext4_unlock_group(sb, e4b->bd_group);
6394 ext4_lock_group(sb, e4b->bd_group);
6397 if ((e4b->bd_info->bb_free - free_count) < minblocks)
6405 * ext4_trim_all_free -- function to trim all free space in alloc. group
6406 * @sb: super block for file system
6407 * @group: group to be trimmed
6408 * @start: first group block to examine
6409 * @max: last group block to examine
6410 * @minblocks: minimum extent block count
6411 * @set_trimmed: set the trimmed flag if at least one block is trimmed
6413 * ext4_trim_all_free walks through group's block bitmap searching for free
6414 * extents. When the free extent is found, mark it as used in group buddy
6415 * bitmap. Then issue a TRIM command on this extent and free the extent in
6416 * the group buddy bitmap.
6418 static ext4_grpblk_t
6419 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6420 ext4_grpblk_t start, ext4_grpblk_t max,
6421 ext4_grpblk_t minblocks, bool set_trimmed)
6423 struct ext4_buddy e4b;
6426 trace_ext4_trim_all_free(sb, group, start, max);
6428 ret = ext4_mb_load_buddy(sb, group, &e4b);
6430 ext4_warning(sb, "Error %d loading buddy information for %u",
6435 ext4_lock_group(sb, group);
6437 if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
6438 minblocks < EXT4_SB(sb)->s_last_trim_minblks) {
6439 ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
6440 if (ret >= 0 && set_trimmed)
6441 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
6446 ext4_unlock_group(sb, group);
6447 ext4_mb_unload_buddy(&e4b);
6449 ext4_debug("trimmed %d blocks in the group %d\n",
6456 * ext4_trim_fs() -- trim ioctl handle function
6457 * @sb: superblock for filesystem
6458 * @range: fstrim_range structure
6460 * start: First Byte to trim
6461 * len: number of Bytes to trim from start
6462 * minlen: minimum extent length in Bytes
6463 * ext4_trim_fs goes through all allocation groups containing Bytes from
6464 * start to start+len. For each such a group ext4_trim_all_free function
6465 * is invoked to trim all free space.
6467 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6469 unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
6470 struct ext4_group_info *grp;
6471 ext4_group_t group, first_group, last_group;
6472 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6473 uint64_t start, end, minlen, trimmed = 0;
6474 ext4_fsblk_t first_data_blk =
6475 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6476 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6477 bool whole_group, eof = false;
6480 start = range->start >> sb->s_blocksize_bits;
6481 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6482 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6483 range->minlen >> sb->s_blocksize_bits);
6485 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6486 start >= max_blks ||
6487 range->len < sb->s_blocksize)
6489 /* No point to try to trim less than discard granularity */
6490 if (range->minlen < discard_granularity) {
6491 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6492 discard_granularity >> sb->s_blocksize_bits);
6493 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
6496 if (end >= max_blks - 1) {
6500 if (end <= first_data_blk)
6502 if (start < first_data_blk)
6503 start = first_data_blk;
6505 /* Determine first and last group to examine based on start and end */
6506 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6507 &first_group, &first_cluster);
6508 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6509 &last_group, &last_cluster);
6511 /* end now represents the last cluster to discard in this group */
6512 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6515 for (group = first_group; group <= last_group; group++) {
6516 grp = ext4_get_group_info(sb, group);
6517 /* We only do this if the grp has never been initialized */
6518 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6519 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6525 * For all the groups except the last one, last cluster will
6526 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6527 * change it for the last group, note that last_cluster is
6528 * already computed earlier by ext4_get_group_no_and_offset()
6530 if (group == last_group) {
6532 whole_group = eof ? true : end == EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6534 if (grp->bb_free >= minlen) {
6535 cnt = ext4_trim_all_free(sb, group, first_cluster,
6536 end, minlen, whole_group);
6545 * For every group except the first one, we are sure
6546 * that the first cluster to discard will be cluster #0.
6552 EXT4_SB(sb)->s_last_trim_minblks = minlen;
6555 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6559 /* Iterate all the free extents in the group. */
6561 ext4_mballoc_query_range(
6562 struct super_block *sb,
6564 ext4_grpblk_t start,
6566 ext4_mballoc_query_range_fn formatter,
6571 struct ext4_buddy e4b;
6574 error = ext4_mb_load_buddy(sb, group, &e4b);
6577 bitmap = e4b.bd_bitmap;
6579 ext4_lock_group(sb, group);
6581 start = (e4b.bd_info->bb_first_free > start) ?
6582 e4b.bd_info->bb_first_free : start;
6583 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
6584 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6586 while (start <= end) {
6587 start = mb_find_next_zero_bit(bitmap, end + 1, start);
6590 next = mb_find_next_bit(bitmap, end + 1, start);
6592 ext4_unlock_group(sb, group);
6593 error = formatter(sb, group, start, next - start, priv);
6596 ext4_lock_group(sb, group);
6601 ext4_unlock_group(sb, group);
6603 ext4_mb_unload_buddy(&e4b);