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 rb tree (sbi->s_mb_avg_fragment_size_root)
145 * Locking: sbi->s_mb_rb_lock (rwlock)
147 * This is a red black tree consisting of group infos and the tree is sorted
148 * by average fragment sizes (which is calculated as ext4_group_info->bb_free
149 * / ext4_group_info->bb_fragments).
151 * When "mb_optimize_scan" mount option is set, mballoc consults the above data
152 * structures to decide the order in which groups are to be traversed for
153 * fulfilling an allocation request.
155 * At CR = 0, we look for groups which have the largest_free_order >= the order
156 * of the request. We directly look at the largest free order list in the data
157 * structure (1) above where largest_free_order = order of the request. If that
158 * list is empty, we look at remaining list in the increasing order of
159 * largest_free_order. This allows us to perform CR = 0 lookup in O(1) time.
161 * At CR = 1, we only consider groups where average fragment size > request
162 * size. So, we lookup a group which has average fragment size just above or
163 * equal to request size using our rb tree (data structure 2) in O(log N) time.
165 * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in
166 * linear order which requires O(N) search time for each CR 0 and CR 1 phase.
168 * The regular allocator (using the buddy cache) supports a few tunables.
170 * /sys/fs/ext4/<partition>/mb_min_to_scan
171 * /sys/fs/ext4/<partition>/mb_max_to_scan
172 * /sys/fs/ext4/<partition>/mb_order2_req
173 * /sys/fs/ext4/<partition>/mb_linear_limit
175 * The regular allocator uses buddy scan only if the request len is power of
176 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
177 * value of s_mb_order2_reqs can be tuned via
178 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
179 * stripe size (sbi->s_stripe), we try to search for contiguous block in
180 * stripe size. This should result in better allocation on RAID setups. If
181 * not, we search in the specific group using bitmap for best extents. The
182 * tunable min_to_scan and max_to_scan control the behaviour here.
183 * min_to_scan indicate how long the mballoc __must__ look for a best
184 * extent and max_to_scan indicates how long the mballoc __can__ look for a
185 * best extent in the found extents. Searching for the blocks starts with
186 * the group specified as the goal value in allocation context via
187 * ac_g_ex. Each group is first checked based on the criteria whether it
188 * can be used for allocation. ext4_mb_good_group explains how the groups are
191 * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not
192 * get traversed linearly. That may result in subsequent allocations being not
193 * close to each other. And so, the underlying device may get filled up in a
194 * non-linear fashion. While that may not matter on non-rotational devices, for
195 * rotational devices that may result in higher seek times. "mb_linear_limit"
196 * tells mballoc how many groups mballoc should search linearly before
197 * performing consulting above data structures for more efficient lookups. For
198 * non rotational devices, this value defaults to 0 and for rotational devices
199 * this is set to MB_DEFAULT_LINEAR_LIMIT.
201 * Both the prealloc space are getting populated as above. So for the first
202 * request we will hit the buddy cache which will result in this prealloc
203 * space getting filled. The prealloc space is then later used for the
204 * subsequent request.
208 * mballoc operates on the following data:
210 * - in-core buddy (actually includes buddy and bitmap)
211 * - preallocation descriptors (PAs)
213 * there are two types of preallocations:
215 * assiged to specific inode and can be used for this inode only.
216 * it describes part of inode's space preallocated to specific
217 * physical blocks. any block from that preallocated can be used
218 * independent. the descriptor just tracks number of blocks left
219 * unused. so, before taking some block from descriptor, one must
220 * make sure corresponded logical block isn't allocated yet. this
221 * also means that freeing any block within descriptor's range
222 * must discard all preallocated blocks.
224 * assigned to specific locality group which does not translate to
225 * permanent set of inodes: inode can join and leave group. space
226 * from this type of preallocation can be used for any inode. thus
227 * it's consumed from the beginning to the end.
229 * relation between them can be expressed as:
230 * in-core buddy = on-disk bitmap + preallocation descriptors
232 * this mean blocks mballoc considers used are:
233 * - allocated blocks (persistent)
234 * - preallocated blocks (non-persistent)
236 * consistency in mballoc world means that at any time a block is either
237 * free or used in ALL structures. notice: "any time" should not be read
238 * literally -- time is discrete and delimited by locks.
240 * to keep it simple, we don't use block numbers, instead we count number of
241 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
243 * all operations can be expressed as:
244 * - init buddy: buddy = on-disk + PAs
245 * - new PA: buddy += N; PA = N
246 * - use inode PA: on-disk += N; PA -= N
247 * - discard inode PA buddy -= on-disk - PA; PA = 0
248 * - use locality group PA on-disk += N; PA -= N
249 * - discard locality group PA buddy -= PA; PA = 0
250 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
251 * is used in real operation because we can't know actual used
252 * bits from PA, only from on-disk bitmap
254 * if we follow this strict logic, then all operations above should be atomic.
255 * given some of them can block, we'd have to use something like semaphores
256 * killing performance on high-end SMP hardware. let's try to relax it using
257 * the following knowledge:
258 * 1) if buddy is referenced, it's already initialized
259 * 2) while block is used in buddy and the buddy is referenced,
260 * nobody can re-allocate that block
261 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
262 * bit set and PA claims same block, it's OK. IOW, one can set bit in
263 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
266 * so, now we're building a concurrency table:
269 * blocks for PA are allocated in the buddy, buddy must be referenced
270 * until PA is linked to allocation group to avoid concurrent buddy init
272 * we need to make sure that either on-disk bitmap or PA has uptodate data
273 * given (3) we care that PA-=N operation doesn't interfere with init
275 * the simplest way would be to have buddy initialized by the discard
276 * - use locality group PA
277 * again PA-=N must be serialized with init
278 * - discard locality group PA
279 * the simplest way would be to have buddy initialized by the discard
282 * i_data_sem serializes them
284 * discard process must wait until PA isn't used by another process
285 * - use locality group PA
286 * some mutex should serialize them
287 * - discard locality group PA
288 * discard process must wait until PA isn't used by another process
291 * i_data_sem or another mutex should serializes them
293 * discard process must wait until PA isn't used by another process
294 * - use locality group PA
295 * nothing wrong here -- they're different PAs covering different blocks
296 * - discard locality group PA
297 * discard process must wait until PA isn't used by another process
299 * now we're ready to make few consequences:
300 * - PA is referenced and while it is no discard is possible
301 * - PA is referenced until block isn't marked in on-disk bitmap
302 * - PA changes only after on-disk bitmap
303 * - discard must not compete with init. either init is done before
304 * any discard or they're serialized somehow
305 * - buddy init as sum of on-disk bitmap and PAs is done atomically
307 * a special case when we've used PA to emptiness. no need to modify buddy
308 * in this case, but we should care about concurrent init
313 * Logic in few words:
318 * mark bits in on-disk bitmap
321 * - use preallocation:
322 * find proper PA (per-inode or group)
324 * mark bits in on-disk bitmap
330 * mark bits in on-disk bitmap
333 * - discard preallocations in group:
335 * move them onto local list
336 * load on-disk bitmap
338 * remove PA from object (inode or locality group)
339 * mark free blocks in-core
341 * - discard inode's preallocations:
348 * - bitlock on a group (group)
349 * - object (inode/locality) (object)
351 * - cr0 lists lock (cr0)
352 * - cr1 tree lock (cr1)
362 * - release consumed pa:
367 * - generate in-core bitmap:
371 * - discard all for given object (inode, locality group):
376 * - discard all for given group:
382 * - allocation path (ext4_mb_regular_allocator)
386 static struct kmem_cache *ext4_pspace_cachep;
387 static struct kmem_cache *ext4_ac_cachep;
388 static struct kmem_cache *ext4_free_data_cachep;
390 /* We create slab caches for groupinfo data structures based on the
391 * superblock block size. There will be one per mounted filesystem for
392 * each unique s_blocksize_bits */
393 #define NR_GRPINFO_CACHES 8
394 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
396 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
397 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
398 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
399 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
402 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
404 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
406 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac);
408 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
409 ext4_group_t group, int cr);
411 static int ext4_try_to_trim_range(struct super_block *sb,
412 struct ext4_buddy *e4b, ext4_grpblk_t start,
413 ext4_grpblk_t max, ext4_grpblk_t minblocks);
416 * The algorithm using this percpu seq counter goes below:
417 * 1. We sample the percpu discard_pa_seq counter before trying for block
418 * allocation in ext4_mb_new_blocks().
419 * 2. We increment this percpu discard_pa_seq counter when we either allocate
420 * or free these blocks i.e. while marking those blocks as used/free in
421 * mb_mark_used()/mb_free_blocks().
422 * 3. We also increment this percpu seq counter when we successfully identify
423 * that the bb_prealloc_list is not empty and hence proceed for discarding
424 * of those PAs inside ext4_mb_discard_group_preallocations().
426 * Now to make sure that the regular fast path of block allocation is not
427 * affected, as a small optimization we only sample the percpu seq counter
428 * on that cpu. Only when the block allocation fails and when freed blocks
429 * found were 0, that is when we sample percpu seq counter for all cpus using
430 * below function ext4_get_discard_pa_seq_sum(). This happens after making
431 * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
433 static DEFINE_PER_CPU(u64, discard_pa_seq);
434 static inline u64 ext4_get_discard_pa_seq_sum(void)
439 for_each_possible_cpu(__cpu)
440 __seq += per_cpu(discard_pa_seq, __cpu);
444 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
446 #if BITS_PER_LONG == 64
447 *bit += ((unsigned long) addr & 7UL) << 3;
448 addr = (void *) ((unsigned long) addr & ~7UL);
449 #elif BITS_PER_LONG == 32
450 *bit += ((unsigned long) addr & 3UL) << 3;
451 addr = (void *) ((unsigned long) addr & ~3UL);
453 #error "how many bits you are?!"
458 static inline int mb_test_bit(int bit, void *addr)
461 * ext4_test_bit on architecture like powerpc
462 * needs unsigned long aligned address
464 addr = mb_correct_addr_and_bit(&bit, addr);
465 return ext4_test_bit(bit, addr);
468 static inline void mb_set_bit(int bit, void *addr)
470 addr = mb_correct_addr_and_bit(&bit, addr);
471 ext4_set_bit(bit, addr);
474 static inline void mb_clear_bit(int bit, void *addr)
476 addr = mb_correct_addr_and_bit(&bit, addr);
477 ext4_clear_bit(bit, addr);
480 static inline int mb_test_and_clear_bit(int bit, void *addr)
482 addr = mb_correct_addr_and_bit(&bit, addr);
483 return ext4_test_and_clear_bit(bit, addr);
486 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
488 int fix = 0, ret, tmpmax;
489 addr = mb_correct_addr_and_bit(&fix, addr);
493 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
499 static inline int mb_find_next_bit(void *addr, int max, int start)
501 int fix = 0, ret, tmpmax;
502 addr = mb_correct_addr_and_bit(&fix, addr);
506 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
512 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
516 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
519 if (order > e4b->bd_blkbits + 1) {
524 /* at order 0 we see each particular block */
526 *max = 1 << (e4b->bd_blkbits + 3);
527 return e4b->bd_bitmap;
530 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
531 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
537 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
538 int first, int count)
541 struct super_block *sb = e4b->bd_sb;
543 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
545 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
546 for (i = 0; i < count; i++) {
547 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
548 ext4_fsblk_t blocknr;
550 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
551 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
552 ext4_grp_locked_error(sb, e4b->bd_group,
553 inode ? inode->i_ino : 0,
555 "freeing block already freed "
558 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
559 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
561 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
565 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
569 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
571 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
572 for (i = 0; i < count; i++) {
573 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
574 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
578 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
580 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
582 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
583 unsigned char *b1, *b2;
585 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
586 b2 = (unsigned char *) bitmap;
587 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
588 if (b1[i] != b2[i]) {
589 ext4_msg(e4b->bd_sb, KERN_ERR,
590 "corruption in group %u "
591 "at byte %u(%u): %x in copy != %x "
593 e4b->bd_group, i, i * 8, b1[i], b2[i]);
600 static void mb_group_bb_bitmap_alloc(struct super_block *sb,
601 struct ext4_group_info *grp, ext4_group_t group)
603 struct buffer_head *bh;
605 grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
609 bh = ext4_read_block_bitmap(sb, group);
610 if (IS_ERR_OR_NULL(bh)) {
611 kfree(grp->bb_bitmap);
612 grp->bb_bitmap = NULL;
616 memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
620 static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
622 kfree(grp->bb_bitmap);
626 static inline void mb_free_blocks_double(struct inode *inode,
627 struct ext4_buddy *e4b, int first, int count)
631 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
632 int first, int count)
636 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
641 static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
642 struct ext4_group_info *grp, ext4_group_t group)
647 static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
653 #ifdef AGGRESSIVE_CHECK
655 #define MB_CHECK_ASSERT(assert) \
659 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
660 function, file, line, # assert); \
665 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
666 const char *function, int line)
668 struct super_block *sb = e4b->bd_sb;
669 int order = e4b->bd_blkbits + 1;
676 struct ext4_group_info *grp;
679 struct list_head *cur;
683 if (e4b->bd_info->bb_check_counter++ % 10)
687 buddy = mb_find_buddy(e4b, order, &max);
688 MB_CHECK_ASSERT(buddy);
689 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
690 MB_CHECK_ASSERT(buddy2);
691 MB_CHECK_ASSERT(buddy != buddy2);
692 MB_CHECK_ASSERT(max * 2 == max2);
695 for (i = 0; i < max; i++) {
697 if (mb_test_bit(i, buddy)) {
698 /* only single bit in buddy2 may be 0 */
699 if (!mb_test_bit(i << 1, buddy2)) {
701 mb_test_bit((i<<1)+1, buddy2));
706 /* both bits in buddy2 must be 1 */
707 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
708 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
710 for (j = 0; j < (1 << order); j++) {
711 k = (i * (1 << order)) + j;
713 !mb_test_bit(k, e4b->bd_bitmap));
717 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
722 buddy = mb_find_buddy(e4b, 0, &max);
723 for (i = 0; i < max; i++) {
724 if (!mb_test_bit(i, buddy)) {
725 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
733 /* check used bits only */
734 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
735 buddy2 = mb_find_buddy(e4b, j, &max2);
737 MB_CHECK_ASSERT(k < max2);
738 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
741 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
742 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
744 grp = ext4_get_group_info(sb, e4b->bd_group);
745 list_for_each(cur, &grp->bb_prealloc_list) {
746 ext4_group_t groupnr;
747 struct ext4_prealloc_space *pa;
748 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
749 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
750 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
751 for (i = 0; i < pa->pa_len; i++)
752 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
756 #undef MB_CHECK_ASSERT
757 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
758 __FILE__, __func__, __LINE__)
760 #define mb_check_buddy(e4b)
764 * Divide blocks started from @first with length @len into
765 * smaller chunks with power of 2 blocks.
766 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
767 * then increase bb_counters[] for corresponded chunk size.
769 static void ext4_mb_mark_free_simple(struct super_block *sb,
770 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
771 struct ext4_group_info *grp)
773 struct ext4_sb_info *sbi = EXT4_SB(sb);
779 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
781 border = 2 << sb->s_blocksize_bits;
784 /* find how many blocks can be covered since this position */
785 max = ffs(first | border) - 1;
787 /* find how many blocks of power 2 we need to mark */
794 /* mark multiblock chunks only */
795 grp->bb_counters[min]++;
797 mb_clear_bit(first >> min,
798 buddy + sbi->s_mb_offsets[min]);
805 static void ext4_mb_rb_insert(struct rb_root *root, struct rb_node *new,
806 int (*cmp)(struct rb_node *, struct rb_node *))
808 struct rb_node **iter = &root->rb_node, *parent = NULL;
812 if (cmp(new, *iter) > 0)
813 iter = &((*iter)->rb_left);
815 iter = &((*iter)->rb_right);
818 rb_link_node(new, parent, iter);
819 rb_insert_color(new, root);
823 ext4_mb_avg_fragment_size_cmp(struct rb_node *rb1, struct rb_node *rb2)
825 struct ext4_group_info *grp1 = rb_entry(rb1,
826 struct ext4_group_info,
827 bb_avg_fragment_size_rb);
828 struct ext4_group_info *grp2 = rb_entry(rb2,
829 struct ext4_group_info,
830 bb_avg_fragment_size_rb);
831 int num_frags_1, num_frags_2;
833 num_frags_1 = grp1->bb_fragments ?
834 grp1->bb_free / grp1->bb_fragments : 0;
835 num_frags_2 = grp2->bb_fragments ?
836 grp2->bb_free / grp2->bb_fragments : 0;
838 return (num_frags_2 - num_frags_1);
842 * Reinsert grpinfo into the avg_fragment_size tree with new average
846 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
848 struct ext4_sb_info *sbi = EXT4_SB(sb);
850 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0)
853 write_lock(&sbi->s_mb_rb_lock);
854 if (!RB_EMPTY_NODE(&grp->bb_avg_fragment_size_rb)) {
855 rb_erase(&grp->bb_avg_fragment_size_rb,
856 &sbi->s_mb_avg_fragment_size_root);
857 RB_CLEAR_NODE(&grp->bb_avg_fragment_size_rb);
860 ext4_mb_rb_insert(&sbi->s_mb_avg_fragment_size_root,
861 &grp->bb_avg_fragment_size_rb,
862 ext4_mb_avg_fragment_size_cmp);
863 write_unlock(&sbi->s_mb_rb_lock);
867 * Choose next group by traversing largest_free_order lists. Updates *new_cr if
868 * cr level needs an update.
870 static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac,
871 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
873 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
874 struct ext4_group_info *iter, *grp;
877 if (ac->ac_status == AC_STATUS_FOUND)
880 if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMIZED))
881 atomic_inc(&sbi->s_bal_cr0_bad_suggestions);
884 for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) {
885 if (list_empty(&sbi->s_mb_largest_free_orders[i]))
887 read_lock(&sbi->s_mb_largest_free_orders_locks[i]);
888 if (list_empty(&sbi->s_mb_largest_free_orders[i])) {
889 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
893 list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i],
894 bb_largest_free_order_node) {
896 atomic64_inc(&sbi->s_bal_cX_groups_considered[0]);
897 if (likely(ext4_mb_good_group(ac, iter->bb_group, 0))) {
902 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
908 /* Increment cr and search again */
911 *group = grp->bb_group;
912 ac->ac_last_optimal_group = *group;
913 ac->ac_flags |= EXT4_MB_CR0_OPTIMIZED;
918 * Choose next group by traversing average fragment size tree. Updates *new_cr
919 * if cr lvel needs an update. Sets EXT4_MB_SEARCH_NEXT_LINEAR to indicate that
920 * the linear search should continue for one iteration since there's lock
921 * contention on the rb tree lock.
923 static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac,
924 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
926 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
927 int avg_fragment_size, best_so_far;
928 struct rb_node *node, *found;
929 struct ext4_group_info *grp;
932 * If there is contention on the lock, instead of waiting for the lock
933 * to become available, just continue searching lineraly. We'll resume
934 * our rb tree search later starting at ac->ac_last_optimal_group.
936 if (!read_trylock(&sbi->s_mb_rb_lock)) {
937 ac->ac_flags |= EXT4_MB_SEARCH_NEXT_LINEAR;
941 if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) {
943 atomic_inc(&sbi->s_bal_cr1_bad_suggestions);
944 /* We have found something at CR 1 in the past */
945 grp = ext4_get_group_info(ac->ac_sb, ac->ac_last_optimal_group);
946 for (found = rb_next(&grp->bb_avg_fragment_size_rb); found != NULL;
947 found = rb_next(found)) {
948 grp = rb_entry(found, struct ext4_group_info,
949 bb_avg_fragment_size_rb);
951 atomic64_inc(&sbi->s_bal_cX_groups_considered[1]);
952 if (likely(ext4_mb_good_group(ac, grp->bb_group, 1)))
958 node = sbi->s_mb_avg_fragment_size_root.rb_node;
963 grp = rb_entry(node, struct ext4_group_info,
964 bb_avg_fragment_size_rb);
965 avg_fragment_size = 0;
966 if (ext4_mb_good_group(ac, grp->bb_group, 1)) {
967 avg_fragment_size = grp->bb_fragments ?
968 grp->bb_free / grp->bb_fragments : 0;
969 if (!best_so_far || avg_fragment_size < best_so_far) {
970 best_so_far = avg_fragment_size;
974 if (avg_fragment_size > ac->ac_g_ex.fe_len)
975 node = node->rb_right;
977 node = node->rb_left;
982 grp = rb_entry(found, struct ext4_group_info,
983 bb_avg_fragment_size_rb);
984 *group = grp->bb_group;
985 ac->ac_flags |= EXT4_MB_CR1_OPTIMIZED;
990 read_unlock(&sbi->s_mb_rb_lock);
991 ac->ac_last_optimal_group = *group;
994 static inline int should_optimize_scan(struct ext4_allocation_context *ac)
996 if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN)))
998 if (ac->ac_criteria >= 2)
1000 if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))
1006 * Return next linear group for allocation. If linear traversal should not be
1007 * performed, this function just returns the same group
1010 next_linear_group(struct ext4_allocation_context *ac, int group, int ngroups)
1012 if (!should_optimize_scan(ac))
1013 goto inc_and_return;
1015 if (ac->ac_groups_linear_remaining) {
1016 ac->ac_groups_linear_remaining--;
1017 goto inc_and_return;
1020 if (ac->ac_flags & EXT4_MB_SEARCH_NEXT_LINEAR) {
1021 ac->ac_flags &= ~EXT4_MB_SEARCH_NEXT_LINEAR;
1022 goto inc_and_return;
1028 * Artificially restricted ngroups for non-extent
1029 * files makes group > ngroups possible on first loop.
1031 return group + 1 >= ngroups ? 0 : group + 1;
1035 * ext4_mb_choose_next_group: choose next group for allocation.
1037 * @ac Allocation Context
1038 * @new_cr This is an output parameter. If the there is no good group
1039 * available at current CR level, this field is updated to indicate
1040 * the new cr level that should be used.
1041 * @group This is an input / output parameter. As an input it indicates the
1042 * next group that the allocator intends to use for allocation. As
1043 * output, this field indicates the next group that should be used as
1044 * determined by the optimization functions.
1045 * @ngroups Total number of groups
1047 static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac,
1048 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
1050 *new_cr = ac->ac_criteria;
1052 if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining)
1056 ext4_mb_choose_next_group_cr0(ac, new_cr, group, ngroups);
1057 } else if (*new_cr == 1) {
1058 ext4_mb_choose_next_group_cr1(ac, new_cr, group, ngroups);
1061 * TODO: For CR=2, we can arrange groups in an rb tree sorted by
1062 * bb_free. But until that happens, we should never come here.
1069 * Cache the order of the largest free extent we have available in this block
1073 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
1075 struct ext4_sb_info *sbi = EXT4_SB(sb);
1078 if (test_opt2(sb, MB_OPTIMIZE_SCAN) && grp->bb_largest_free_order >= 0) {
1079 write_lock(&sbi->s_mb_largest_free_orders_locks[
1080 grp->bb_largest_free_order]);
1081 list_del_init(&grp->bb_largest_free_order_node);
1082 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1083 grp->bb_largest_free_order]);
1085 grp->bb_largest_free_order = -1; /* uninit */
1087 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) {
1088 if (grp->bb_counters[i] > 0) {
1089 grp->bb_largest_free_order = i;
1093 if (test_opt2(sb, MB_OPTIMIZE_SCAN) &&
1094 grp->bb_largest_free_order >= 0 && grp->bb_free) {
1095 write_lock(&sbi->s_mb_largest_free_orders_locks[
1096 grp->bb_largest_free_order]);
1097 list_add_tail(&grp->bb_largest_free_order_node,
1098 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1099 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1100 grp->bb_largest_free_order]);
1104 static noinline_for_stack
1105 void ext4_mb_generate_buddy(struct super_block *sb,
1106 void *buddy, void *bitmap, ext4_group_t group)
1108 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
1109 struct ext4_sb_info *sbi = EXT4_SB(sb);
1110 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1111 ext4_grpblk_t i = 0;
1112 ext4_grpblk_t first;
1115 unsigned fragments = 0;
1116 unsigned long long period = get_cycles();
1118 /* initialize buddy from bitmap which is aggregation
1119 * of on-disk bitmap and preallocations */
1120 i = mb_find_next_zero_bit(bitmap, max, 0);
1121 grp->bb_first_free = i;
1125 i = mb_find_next_bit(bitmap, max, i);
1129 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1131 grp->bb_counters[0]++;
1133 i = mb_find_next_zero_bit(bitmap, max, i);
1135 grp->bb_fragments = fragments;
1137 if (free != grp->bb_free) {
1138 ext4_grp_locked_error(sb, group, 0, 0,
1139 "block bitmap and bg descriptor "
1140 "inconsistent: %u vs %u free clusters",
1141 free, grp->bb_free);
1143 * If we intend to continue, we consider group descriptor
1144 * corrupt and update bb_free using bitmap value
1146 grp->bb_free = free;
1147 ext4_mark_group_bitmap_corrupted(sb, group,
1148 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1150 mb_set_largest_free_order(sb, grp);
1152 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1154 period = get_cycles() - period;
1155 atomic_inc(&sbi->s_mb_buddies_generated);
1156 atomic64_add(period, &sbi->s_mb_generation_time);
1157 mb_update_avg_fragment_size(sb, grp);
1160 /* The buddy information is attached the buddy cache inode
1161 * for convenience. The information regarding each group
1162 * is loaded via ext4_mb_load_buddy. The information involve
1163 * block bitmap and buddy information. The information are
1164 * stored in the inode as
1167 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1170 * one block each for bitmap and buddy information.
1171 * So for each group we take up 2 blocks. A page can
1172 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
1173 * So it can have information regarding groups_per_page which
1174 * is blocks_per_page/2
1176 * Locking note: This routine takes the block group lock of all groups
1177 * for this page; do not hold this lock when calling this routine!
1180 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
1182 ext4_group_t ngroups;
1184 int blocks_per_page;
1185 int groups_per_page;
1188 ext4_group_t first_group, group;
1190 struct super_block *sb;
1191 struct buffer_head *bhs;
1192 struct buffer_head **bh = NULL;
1193 struct inode *inode;
1196 struct ext4_group_info *grinfo;
1198 inode = page->mapping->host;
1200 ngroups = ext4_get_groups_count(sb);
1201 blocksize = i_blocksize(inode);
1202 blocks_per_page = PAGE_SIZE / blocksize;
1204 mb_debug(sb, "init page %lu\n", page->index);
1206 groups_per_page = blocks_per_page >> 1;
1207 if (groups_per_page == 0)
1208 groups_per_page = 1;
1210 /* allocate buffer_heads to read bitmaps */
1211 if (groups_per_page > 1) {
1212 i = sizeof(struct buffer_head *) * groups_per_page;
1213 bh = kzalloc(i, gfp);
1221 first_group = page->index * blocks_per_page / 2;
1223 /* read all groups the page covers into the cache */
1224 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1225 if (group >= ngroups)
1228 grinfo = ext4_get_group_info(sb, group);
1230 * If page is uptodate then we came here after online resize
1231 * which added some new uninitialized group info structs, so
1232 * we must skip all initialized uptodate buddies on the page,
1233 * which may be currently in use by an allocating task.
1235 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1239 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1240 if (IS_ERR(bh[i])) {
1241 err = PTR_ERR(bh[i]);
1245 mb_debug(sb, "read bitmap for group %u\n", group);
1248 /* wait for I/O completion */
1249 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1254 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1259 first_block = page->index * blocks_per_page;
1260 for (i = 0; i < blocks_per_page; i++) {
1261 group = (first_block + i) >> 1;
1262 if (group >= ngroups)
1265 if (!bh[group - first_group])
1266 /* skip initialized uptodate buddy */
1269 if (!buffer_verified(bh[group - first_group]))
1270 /* Skip faulty bitmaps */
1275 * data carry information regarding this
1276 * particular group in the format specified
1280 data = page_address(page) + (i * blocksize);
1281 bitmap = bh[group - first_group]->b_data;
1284 * We place the buddy block and bitmap block
1287 if ((first_block + i) & 1) {
1288 /* this is block of buddy */
1289 BUG_ON(incore == NULL);
1290 mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
1291 group, page->index, i * blocksize);
1292 trace_ext4_mb_buddy_bitmap_load(sb, group);
1293 grinfo = ext4_get_group_info(sb, group);
1294 grinfo->bb_fragments = 0;
1295 memset(grinfo->bb_counters, 0,
1296 sizeof(*grinfo->bb_counters) *
1297 (MB_NUM_ORDERS(sb)));
1299 * incore got set to the group block bitmap below
1301 ext4_lock_group(sb, group);
1302 /* init the buddy */
1303 memset(data, 0xff, blocksize);
1304 ext4_mb_generate_buddy(sb, data, incore, group);
1305 ext4_unlock_group(sb, group);
1308 /* this is block of bitmap */
1309 BUG_ON(incore != NULL);
1310 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
1311 group, page->index, i * blocksize);
1312 trace_ext4_mb_bitmap_load(sb, group);
1314 /* see comments in ext4_mb_put_pa() */
1315 ext4_lock_group(sb, group);
1316 memcpy(data, bitmap, blocksize);
1318 /* mark all preallocated blks used in in-core bitmap */
1319 ext4_mb_generate_from_pa(sb, data, group);
1320 ext4_mb_generate_from_freelist(sb, data, group);
1321 ext4_unlock_group(sb, group);
1323 /* set incore so that the buddy information can be
1324 * generated using this
1329 SetPageUptodate(page);
1333 for (i = 0; i < groups_per_page; i++)
1342 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1343 * on the same buddy page doesn't happen whild holding the buddy page lock.
1344 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1345 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1347 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1348 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1350 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1351 int block, pnum, poff;
1352 int blocks_per_page;
1355 e4b->bd_buddy_page = NULL;
1356 e4b->bd_bitmap_page = NULL;
1358 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1360 * the buddy cache inode stores the block bitmap
1361 * and buddy information in consecutive blocks.
1362 * So for each group we need two blocks.
1365 pnum = block / blocks_per_page;
1366 poff = block % blocks_per_page;
1367 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1370 BUG_ON(page->mapping != inode->i_mapping);
1371 e4b->bd_bitmap_page = page;
1372 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1374 if (blocks_per_page >= 2) {
1375 /* buddy and bitmap are on the same page */
1380 pnum = block / blocks_per_page;
1381 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1384 BUG_ON(page->mapping != inode->i_mapping);
1385 e4b->bd_buddy_page = page;
1389 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1391 if (e4b->bd_bitmap_page) {
1392 unlock_page(e4b->bd_bitmap_page);
1393 put_page(e4b->bd_bitmap_page);
1395 if (e4b->bd_buddy_page) {
1396 unlock_page(e4b->bd_buddy_page);
1397 put_page(e4b->bd_buddy_page);
1402 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1403 * block group lock of all groups for this page; do not hold the BG lock when
1404 * calling this routine!
1406 static noinline_for_stack
1407 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1410 struct ext4_group_info *this_grp;
1411 struct ext4_buddy e4b;
1416 mb_debug(sb, "init group %u\n", group);
1417 this_grp = ext4_get_group_info(sb, group);
1419 * This ensures that we don't reinit the buddy cache
1420 * page which map to the group from which we are already
1421 * allocating. If we are looking at the buddy cache we would
1422 * have taken a reference using ext4_mb_load_buddy and that
1423 * would have pinned buddy page to page cache.
1424 * The call to ext4_mb_get_buddy_page_lock will mark the
1427 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1428 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1430 * somebody initialized the group
1431 * return without doing anything
1436 page = e4b.bd_bitmap_page;
1437 ret = ext4_mb_init_cache(page, NULL, gfp);
1440 if (!PageUptodate(page)) {
1445 if (e4b.bd_buddy_page == NULL) {
1447 * If both the bitmap and buddy are in
1448 * the same page we don't need to force
1454 /* init buddy cache */
1455 page = e4b.bd_buddy_page;
1456 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1459 if (!PageUptodate(page)) {
1464 ext4_mb_put_buddy_page_lock(&e4b);
1469 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1470 * block group lock of all groups for this page; do not hold the BG lock when
1471 * calling this routine!
1473 static noinline_for_stack int
1474 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1475 struct ext4_buddy *e4b, gfp_t gfp)
1477 int blocks_per_page;
1483 struct ext4_group_info *grp;
1484 struct ext4_sb_info *sbi = EXT4_SB(sb);
1485 struct inode *inode = sbi->s_buddy_cache;
1488 mb_debug(sb, "load group %u\n", group);
1490 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1491 grp = ext4_get_group_info(sb, group);
1493 e4b->bd_blkbits = sb->s_blocksize_bits;
1496 e4b->bd_group = group;
1497 e4b->bd_buddy_page = NULL;
1498 e4b->bd_bitmap_page = NULL;
1500 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1502 * we need full data about the group
1503 * to make a good selection
1505 ret = ext4_mb_init_group(sb, group, gfp);
1511 * the buddy cache inode stores the block bitmap
1512 * and buddy information in consecutive blocks.
1513 * So for each group we need two blocks.
1516 pnum = block / blocks_per_page;
1517 poff = block % blocks_per_page;
1519 /* we could use find_or_create_page(), but it locks page
1520 * what we'd like to avoid in fast path ... */
1521 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1522 if (page == NULL || !PageUptodate(page)) {
1525 * drop the page reference and try
1526 * to get the page with lock. If we
1527 * are not uptodate that implies
1528 * somebody just created the page but
1529 * is yet to initialize the same. So
1530 * wait for it to initialize.
1533 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1535 BUG_ON(page->mapping != inode->i_mapping);
1536 if (!PageUptodate(page)) {
1537 ret = ext4_mb_init_cache(page, NULL, gfp);
1542 mb_cmp_bitmaps(e4b, page_address(page) +
1543 (poff * sb->s_blocksize));
1552 if (!PageUptodate(page)) {
1557 /* Pages marked accessed already */
1558 e4b->bd_bitmap_page = page;
1559 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1562 pnum = block / blocks_per_page;
1563 poff = block % blocks_per_page;
1565 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1566 if (page == NULL || !PageUptodate(page)) {
1569 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1571 BUG_ON(page->mapping != inode->i_mapping);
1572 if (!PageUptodate(page)) {
1573 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1587 if (!PageUptodate(page)) {
1592 /* Pages marked accessed already */
1593 e4b->bd_buddy_page = page;
1594 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1601 if (e4b->bd_bitmap_page)
1602 put_page(e4b->bd_bitmap_page);
1603 if (e4b->bd_buddy_page)
1604 put_page(e4b->bd_buddy_page);
1605 e4b->bd_buddy = NULL;
1606 e4b->bd_bitmap = NULL;
1610 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1611 struct ext4_buddy *e4b)
1613 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1616 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1618 if (e4b->bd_bitmap_page)
1619 put_page(e4b->bd_bitmap_page);
1620 if (e4b->bd_buddy_page)
1621 put_page(e4b->bd_buddy_page);
1625 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1630 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1631 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1633 while (order <= e4b->bd_blkbits + 1) {
1634 bb = mb_find_buddy(e4b, order, &max);
1635 if (!mb_test_bit(block >> order, bb)) {
1636 /* this block is part of buddy of order 'order' */
1644 static void mb_clear_bits(void *bm, int cur, int len)
1650 if ((cur & 31) == 0 && (len - cur) >= 32) {
1651 /* fast path: clear whole word at once */
1652 addr = bm + (cur >> 3);
1657 mb_clear_bit(cur, bm);
1662 /* clear bits in given range
1663 * will return first found zero bit if any, -1 otherwise
1665 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1672 if ((cur & 31) == 0 && (len - cur) >= 32) {
1673 /* fast path: clear whole word at once */
1674 addr = bm + (cur >> 3);
1675 if (*addr != (__u32)(-1) && zero_bit == -1)
1676 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1681 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1689 void mb_set_bits(void *bm, int cur, int len)
1695 if ((cur & 31) == 0 && (len - cur) >= 32) {
1696 /* fast path: set whole word at once */
1697 addr = bm + (cur >> 3);
1702 mb_set_bit(cur, bm);
1707 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1709 if (mb_test_bit(*bit + side, bitmap)) {
1710 mb_clear_bit(*bit, bitmap);
1716 mb_set_bit(*bit, bitmap);
1721 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1725 void *buddy = mb_find_buddy(e4b, order, &max);
1730 /* Bits in range [first; last] are known to be set since
1731 * corresponding blocks were allocated. Bits in range
1732 * (first; last) will stay set because they form buddies on
1733 * upper layer. We just deal with borders if they don't
1734 * align with upper layer and then go up.
1735 * Releasing entire group is all about clearing
1736 * single bit of highest order buddy.
1740 * ---------------------------------
1742 * ---------------------------------
1743 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1744 * ---------------------------------
1746 * \_____________________/
1748 * Neither [1] nor [6] is aligned to above layer.
1749 * Left neighbour [0] is free, so mark it busy,
1750 * decrease bb_counters and extend range to
1752 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1753 * mark [6] free, increase bb_counters and shrink range to
1755 * Then shift range to [0; 2], go up and do the same.
1760 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1762 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1767 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1768 mb_clear_bits(buddy, first, last - first + 1);
1769 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1778 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1779 int first, int count)
1781 int left_is_free = 0;
1782 int right_is_free = 0;
1784 int last = first + count - 1;
1785 struct super_block *sb = e4b->bd_sb;
1787 if (WARN_ON(count == 0))
1789 BUG_ON(last >= (sb->s_blocksize << 3));
1790 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1791 /* Don't bother if the block group is corrupt. */
1792 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1795 mb_check_buddy(e4b);
1796 mb_free_blocks_double(inode, e4b, first, count);
1798 this_cpu_inc(discard_pa_seq);
1799 e4b->bd_info->bb_free += count;
1800 if (first < e4b->bd_info->bb_first_free)
1801 e4b->bd_info->bb_first_free = first;
1803 /* access memory sequentially: check left neighbour,
1804 * clear range and then check right neighbour
1807 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1808 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1809 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1810 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1812 if (unlikely(block != -1)) {
1813 struct ext4_sb_info *sbi = EXT4_SB(sb);
1814 ext4_fsblk_t blocknr;
1816 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1817 blocknr += EXT4_C2B(sbi, block);
1818 if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) {
1819 ext4_grp_locked_error(sb, e4b->bd_group,
1820 inode ? inode->i_ino : 0,
1822 "freeing already freed block (bit %u); block bitmap corrupt.",
1824 ext4_mark_group_bitmap_corrupted(
1826 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1831 /* let's maintain fragments counter */
1832 if (left_is_free && right_is_free)
1833 e4b->bd_info->bb_fragments--;
1834 else if (!left_is_free && !right_is_free)
1835 e4b->bd_info->bb_fragments++;
1837 /* buddy[0] == bd_bitmap is a special case, so handle
1838 * it right away and let mb_buddy_mark_free stay free of
1839 * zero order checks.
1840 * Check if neighbours are to be coaleasced,
1841 * adjust bitmap bb_counters and borders appropriately.
1844 first += !left_is_free;
1845 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1848 last -= !right_is_free;
1849 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1853 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1856 mb_set_largest_free_order(sb, e4b->bd_info);
1857 mb_update_avg_fragment_size(sb, e4b->bd_info);
1858 mb_check_buddy(e4b);
1861 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1862 int needed, struct ext4_free_extent *ex)
1868 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1871 buddy = mb_find_buddy(e4b, 0, &max);
1872 BUG_ON(buddy == NULL);
1873 BUG_ON(block >= max);
1874 if (mb_test_bit(block, buddy)) {
1881 /* find actual order */
1882 order = mb_find_order_for_block(e4b, block);
1883 block = block >> order;
1885 ex->fe_len = 1 << order;
1886 ex->fe_start = block << order;
1887 ex->fe_group = e4b->bd_group;
1889 /* calc difference from given start */
1890 next = next - ex->fe_start;
1892 ex->fe_start += next;
1894 while (needed > ex->fe_len &&
1895 mb_find_buddy(e4b, order, &max)) {
1897 if (block + 1 >= max)
1900 next = (block + 1) * (1 << order);
1901 if (mb_test_bit(next, e4b->bd_bitmap))
1904 order = mb_find_order_for_block(e4b, next);
1906 block = next >> order;
1907 ex->fe_len += 1 << order;
1910 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1911 /* Should never happen! (but apparently sometimes does?!?) */
1913 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
1914 "corruption or bug in mb_find_extent "
1915 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1916 block, order, needed, ex->fe_group, ex->fe_start,
1917 ex->fe_len, ex->fe_logical);
1925 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1931 int start = ex->fe_start;
1932 int len = ex->fe_len;
1938 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1939 BUG_ON(e4b->bd_group != ex->fe_group);
1940 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1941 mb_check_buddy(e4b);
1942 mb_mark_used_double(e4b, start, len);
1944 this_cpu_inc(discard_pa_seq);
1945 e4b->bd_info->bb_free -= len;
1946 if (e4b->bd_info->bb_first_free == start)
1947 e4b->bd_info->bb_first_free += len;
1949 /* let's maintain fragments counter */
1951 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1952 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1953 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1955 e4b->bd_info->bb_fragments++;
1956 else if (!mlen && !max)
1957 e4b->bd_info->bb_fragments--;
1959 /* let's maintain buddy itself */
1962 ord = mb_find_order_for_block(e4b, start);
1964 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1965 /* the whole chunk may be allocated at once! */
1968 buddy = mb_find_buddy(e4b, ord, &max);
1971 BUG_ON((start >> ord) >= max);
1972 mb_set_bit(start >> ord, buddy);
1973 e4b->bd_info->bb_counters[ord]--;
1980 /* store for history */
1982 ret = len | (ord << 16);
1984 /* we have to split large buddy */
1986 buddy = mb_find_buddy(e4b, ord, &max);
1987 mb_set_bit(start >> ord, buddy);
1988 e4b->bd_info->bb_counters[ord]--;
1991 cur = (start >> ord) & ~1U;
1992 buddy = mb_find_buddy(e4b, ord, &max);
1993 mb_clear_bit(cur, buddy);
1994 mb_clear_bit(cur + 1, buddy);
1995 e4b->bd_info->bb_counters[ord]++;
1996 e4b->bd_info->bb_counters[ord]++;
1999 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
2001 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
2002 mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
2003 mb_check_buddy(e4b);
2009 * Must be called under group lock!
2011 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
2012 struct ext4_buddy *e4b)
2014 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2017 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
2018 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2020 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
2021 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
2022 ret = mb_mark_used(e4b, &ac->ac_b_ex);
2024 /* preallocation can change ac_b_ex, thus we store actually
2025 * allocated blocks for history */
2026 ac->ac_f_ex = ac->ac_b_ex;
2028 ac->ac_status = AC_STATUS_FOUND;
2029 ac->ac_tail = ret & 0xffff;
2030 ac->ac_buddy = ret >> 16;
2033 * take the page reference. We want the page to be pinned
2034 * so that we don't get a ext4_mb_init_cache_call for this
2035 * group until we update the bitmap. That would mean we
2036 * double allocate blocks. The reference is dropped
2037 * in ext4_mb_release_context
2039 ac->ac_bitmap_page = e4b->bd_bitmap_page;
2040 get_page(ac->ac_bitmap_page);
2041 ac->ac_buddy_page = e4b->bd_buddy_page;
2042 get_page(ac->ac_buddy_page);
2043 /* store last allocated for subsequent stream allocation */
2044 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2045 spin_lock(&sbi->s_md_lock);
2046 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2047 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2048 spin_unlock(&sbi->s_md_lock);
2051 * As we've just preallocated more space than
2052 * user requested originally, we store allocated
2053 * space in a special descriptor.
2055 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2056 ext4_mb_new_preallocation(ac);
2060 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2061 struct ext4_buddy *e4b,
2064 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2065 struct ext4_free_extent *bex = &ac->ac_b_ex;
2066 struct ext4_free_extent *gex = &ac->ac_g_ex;
2067 struct ext4_free_extent ex;
2070 if (ac->ac_status == AC_STATUS_FOUND)
2073 * We don't want to scan for a whole year
2075 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2076 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2077 ac->ac_status = AC_STATUS_BREAK;
2082 * Haven't found good chunk so far, let's continue
2084 if (bex->fe_len < gex->fe_len)
2087 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2088 && bex->fe_group == e4b->bd_group) {
2089 /* recheck chunk's availability - we don't know
2090 * when it was found (within this lock-unlock
2092 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
2093 if (max >= gex->fe_len) {
2094 ext4_mb_use_best_found(ac, e4b);
2101 * The routine checks whether found extent is good enough. If it is,
2102 * then the extent gets marked used and flag is set to the context
2103 * to stop scanning. Otherwise, the extent is compared with the
2104 * previous found extent and if new one is better, then it's stored
2105 * in the context. Later, the best found extent will be used, if
2106 * mballoc can't find good enough extent.
2108 * FIXME: real allocation policy is to be designed yet!
2110 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2111 struct ext4_free_extent *ex,
2112 struct ext4_buddy *e4b)
2114 struct ext4_free_extent *bex = &ac->ac_b_ex;
2115 struct ext4_free_extent *gex = &ac->ac_g_ex;
2117 BUG_ON(ex->fe_len <= 0);
2118 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2119 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2120 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2125 * The special case - take what you catch first
2127 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2129 ext4_mb_use_best_found(ac, e4b);
2134 * Let's check whether the chuck is good enough
2136 if (ex->fe_len == gex->fe_len) {
2138 ext4_mb_use_best_found(ac, e4b);
2143 * If this is first found extent, just store it in the context
2145 if (bex->fe_len == 0) {
2151 * If new found extent is better, store it in the context
2153 if (bex->fe_len < gex->fe_len) {
2154 /* if the request isn't satisfied, any found extent
2155 * larger than previous best one is better */
2156 if (ex->fe_len > bex->fe_len)
2158 } else if (ex->fe_len > gex->fe_len) {
2159 /* if the request is satisfied, then we try to find
2160 * an extent that still satisfy the request, but is
2161 * smaller than previous one */
2162 if (ex->fe_len < bex->fe_len)
2166 ext4_mb_check_limits(ac, e4b, 0);
2169 static noinline_for_stack
2170 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2171 struct ext4_buddy *e4b)
2173 struct ext4_free_extent ex = ac->ac_b_ex;
2174 ext4_group_t group = ex.fe_group;
2178 BUG_ON(ex.fe_len <= 0);
2179 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2183 ext4_lock_group(ac->ac_sb, group);
2184 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2188 ext4_mb_use_best_found(ac, e4b);
2191 ext4_unlock_group(ac->ac_sb, group);
2192 ext4_mb_unload_buddy(e4b);
2197 static noinline_for_stack
2198 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2199 struct ext4_buddy *e4b)
2201 ext4_group_t group = ac->ac_g_ex.fe_group;
2204 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2205 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2206 struct ext4_free_extent ex;
2208 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
2210 if (grp->bb_free == 0)
2213 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2217 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
2218 ext4_mb_unload_buddy(e4b);
2222 ext4_lock_group(ac->ac_sb, group);
2223 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2224 ac->ac_g_ex.fe_len, &ex);
2225 ex.fe_logical = 0xDEADFA11; /* debug value */
2227 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
2230 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
2232 /* use do_div to get remainder (would be 64-bit modulo) */
2233 if (do_div(start, sbi->s_stripe) == 0) {
2236 ext4_mb_use_best_found(ac, e4b);
2238 } else if (max >= ac->ac_g_ex.fe_len) {
2239 BUG_ON(ex.fe_len <= 0);
2240 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2241 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2244 ext4_mb_use_best_found(ac, e4b);
2245 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2246 /* Sometimes, caller may want to merge even small
2247 * number of blocks to an existing extent */
2248 BUG_ON(ex.fe_len <= 0);
2249 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2250 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2253 ext4_mb_use_best_found(ac, e4b);
2255 ext4_unlock_group(ac->ac_sb, group);
2256 ext4_mb_unload_buddy(e4b);
2262 * The routine scans buddy structures (not bitmap!) from given order
2263 * to max order and tries to find big enough chunk to satisfy the req
2265 static noinline_for_stack
2266 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2267 struct ext4_buddy *e4b)
2269 struct super_block *sb = ac->ac_sb;
2270 struct ext4_group_info *grp = e4b->bd_info;
2276 BUG_ON(ac->ac_2order <= 0);
2277 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2278 if (grp->bb_counters[i] == 0)
2281 buddy = mb_find_buddy(e4b, i, &max);
2282 BUG_ON(buddy == NULL);
2284 k = mb_find_next_zero_bit(buddy, max, 0);
2286 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2287 "%d free clusters of order %d. But found 0",
2288 grp->bb_counters[i], i);
2289 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2291 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2296 ac->ac_b_ex.fe_len = 1 << i;
2297 ac->ac_b_ex.fe_start = k << i;
2298 ac->ac_b_ex.fe_group = e4b->bd_group;
2300 ext4_mb_use_best_found(ac, e4b);
2302 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2304 if (EXT4_SB(sb)->s_mb_stats)
2305 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2312 * The routine scans the group and measures all found extents.
2313 * In order to optimize scanning, caller must pass number of
2314 * free blocks in the group, so the routine can know upper limit.
2316 static noinline_for_stack
2317 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2318 struct ext4_buddy *e4b)
2320 struct super_block *sb = ac->ac_sb;
2321 void *bitmap = e4b->bd_bitmap;
2322 struct ext4_free_extent ex;
2326 free = e4b->bd_info->bb_free;
2327 if (WARN_ON(free <= 0))
2330 i = e4b->bd_info->bb_first_free;
2332 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2333 i = mb_find_next_zero_bit(bitmap,
2334 EXT4_CLUSTERS_PER_GROUP(sb), i);
2335 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2337 * IF we have corrupt bitmap, we won't find any
2338 * free blocks even though group info says we
2341 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2342 "%d free clusters as per "
2343 "group info. But bitmap says 0",
2345 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2346 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2350 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2351 if (WARN_ON(ex.fe_len <= 0))
2353 if (free < ex.fe_len) {
2354 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2355 "%d free clusters as per "
2356 "group info. But got %d blocks",
2358 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2359 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2361 * The number of free blocks differs. This mostly
2362 * indicate that the bitmap is corrupt. So exit
2363 * without claiming the space.
2367 ex.fe_logical = 0xDEADC0DE; /* debug value */
2368 ext4_mb_measure_extent(ac, &ex, e4b);
2374 ext4_mb_check_limits(ac, e4b, 1);
2378 * This is a special case for storages like raid5
2379 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2381 static noinline_for_stack
2382 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2383 struct ext4_buddy *e4b)
2385 struct super_block *sb = ac->ac_sb;
2386 struct ext4_sb_info *sbi = EXT4_SB(sb);
2387 void *bitmap = e4b->bd_bitmap;
2388 struct ext4_free_extent ex;
2389 ext4_fsblk_t first_group_block;
2394 BUG_ON(sbi->s_stripe == 0);
2396 /* find first stripe-aligned block in group */
2397 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2399 a = first_group_block + sbi->s_stripe - 1;
2400 do_div(a, sbi->s_stripe);
2401 i = (a * sbi->s_stripe) - first_group_block;
2403 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2404 if (!mb_test_bit(i, bitmap)) {
2405 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2406 if (max >= sbi->s_stripe) {
2408 ex.fe_logical = 0xDEADF00D; /* debug value */
2410 ext4_mb_use_best_found(ac, e4b);
2419 * This is also called BEFORE we load the buddy bitmap.
2420 * Returns either 1 or 0 indicating that the group is either suitable
2421 * for the allocation or not.
2423 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2424 ext4_group_t group, int cr)
2426 ext4_grpblk_t free, fragments;
2427 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2428 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2430 BUG_ON(cr < 0 || cr >= 4);
2432 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2435 free = grp->bb_free;
2439 fragments = grp->bb_fragments;
2445 BUG_ON(ac->ac_2order == 0);
2447 /* Avoid using the first bg of a flexgroup for data files */
2448 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2449 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2450 ((group % flex_size) == 0))
2453 if (free < ac->ac_g_ex.fe_len)
2456 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2459 if (grp->bb_largest_free_order < ac->ac_2order)
2464 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2468 if (free >= ac->ac_g_ex.fe_len)
2481 * This could return negative error code if something goes wrong
2482 * during ext4_mb_init_group(). This should not be called with
2483 * ext4_lock_group() held.
2485 * Note: because we are conditionally operating with the group lock in
2486 * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this
2487 * function using __acquire and __release. This means we need to be
2488 * super careful before messing with the error path handling via "goto
2491 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2492 ext4_group_t group, int cr)
2494 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2495 struct super_block *sb = ac->ac_sb;
2496 struct ext4_sb_info *sbi = EXT4_SB(sb);
2497 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2501 if (sbi->s_mb_stats)
2502 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2504 ext4_lock_group(sb, group);
2505 __release(ext4_group_lock_ptr(sb, group));
2507 free = grp->bb_free;
2510 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2512 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2515 __acquire(ext4_group_lock_ptr(sb, group));
2516 ext4_unlock_group(sb, group);
2519 /* We only do this if the grp has never been initialized */
2520 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2521 struct ext4_group_desc *gdp =
2522 ext4_get_group_desc(sb, group, NULL);
2525 /* cr=0/1 is a very optimistic search to find large
2526 * good chunks almost for free. If buddy data is not
2527 * ready, then this optimization makes no sense. But
2528 * we never skip the first block group in a flex_bg,
2529 * since this gets used for metadata block allocation,
2530 * and we want to make sure we locate metadata blocks
2531 * in the first block group in the flex_bg if possible.
2534 (!sbi->s_log_groups_per_flex ||
2535 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2536 !(ext4_has_group_desc_csum(sb) &&
2537 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2539 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2545 ext4_lock_group(sb, group);
2546 __release(ext4_group_lock_ptr(sb, group));
2548 ret = ext4_mb_good_group(ac, group, cr);
2551 __acquire(ext4_group_lock_ptr(sb, group));
2552 ext4_unlock_group(sb, group);
2558 * Start prefetching @nr block bitmaps starting at @group.
2559 * Return the next group which needs to be prefetched.
2561 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2562 unsigned int nr, int *cnt)
2564 ext4_group_t ngroups = ext4_get_groups_count(sb);
2565 struct buffer_head *bh;
2566 struct blk_plug plug;
2568 blk_start_plug(&plug);
2570 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2572 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2575 * Prefetch block groups with free blocks; but don't
2576 * bother if it is marked uninitialized on disk, since
2577 * it won't require I/O to read. Also only try to
2578 * prefetch once, so we avoid getblk() call, which can
2581 if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2582 EXT4_MB_GRP_NEED_INIT(grp) &&
2583 ext4_free_group_clusters(sb, gdp) > 0 &&
2584 !(ext4_has_group_desc_csum(sb) &&
2585 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2586 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2587 if (bh && !IS_ERR(bh)) {
2588 if (!buffer_uptodate(bh) && cnt)
2593 if (++group >= ngroups)
2596 blk_finish_plug(&plug);
2601 * Prefetching reads the block bitmap into the buffer cache; but we
2602 * need to make sure that the buddy bitmap in the page cache has been
2603 * initialized. Note that ext4_mb_init_group() will block if the I/O
2604 * is not yet completed, or indeed if it was not initiated by
2605 * ext4_mb_prefetch did not start the I/O.
2607 * TODO: We should actually kick off the buddy bitmap setup in a work
2608 * queue when the buffer I/O is completed, so that we don't block
2609 * waiting for the block allocation bitmap read to finish when
2610 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2612 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2616 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2618 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2621 group = ext4_get_groups_count(sb);
2623 grp = ext4_get_group_info(sb, group);
2625 if (EXT4_MB_GRP_NEED_INIT(grp) &&
2626 ext4_free_group_clusters(sb, gdp) > 0 &&
2627 !(ext4_has_group_desc_csum(sb) &&
2628 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2629 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2635 static noinline_for_stack int
2636 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2638 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2640 int err = 0, first_err = 0;
2641 unsigned int nr = 0, prefetch_ios = 0;
2642 struct ext4_sb_info *sbi;
2643 struct super_block *sb;
2644 struct ext4_buddy e4b;
2649 ngroups = ext4_get_groups_count(sb);
2650 /* non-extent files are limited to low blocks/groups */
2651 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2652 ngroups = sbi->s_blockfile_groups;
2654 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2656 /* first, try the goal */
2657 err = ext4_mb_find_by_goal(ac, &e4b);
2658 if (err || ac->ac_status == AC_STATUS_FOUND)
2661 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2665 * ac->ac_2order is set only if the fe_len is a power of 2
2666 * if ac->ac_2order is set we also set criteria to 0 so that we
2667 * try exact allocation using buddy.
2669 i = fls(ac->ac_g_ex.fe_len);
2672 * We search using buddy data only if the order of the request
2673 * is greater than equal to the sbi_s_mb_order2_reqs
2674 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2675 * We also support searching for power-of-two requests only for
2676 * requests upto maximum buddy size we have constructed.
2678 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2680 * This should tell if fe_len is exactly power of 2
2682 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2683 ac->ac_2order = array_index_nospec(i - 1,
2687 /* if stream allocation is enabled, use global goal */
2688 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2689 /* TBD: may be hot point */
2690 spin_lock(&sbi->s_md_lock);
2691 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2692 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2693 spin_unlock(&sbi->s_md_lock);
2696 /* Let's just scan groups to find more-less suitable blocks */
2697 cr = ac->ac_2order ? 0 : 1;
2699 * cr == 0 try to get exact allocation,
2700 * cr == 3 try to get anything
2703 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2704 ac->ac_criteria = cr;
2706 * searching for the right group start
2707 * from the goal value specified
2709 group = ac->ac_g_ex.fe_group;
2710 ac->ac_last_optimal_group = group;
2711 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2712 prefetch_grp = group;
2714 for (i = 0; i < ngroups; group = next_linear_group(ac, group, ngroups),
2716 int ret = 0, new_cr;
2720 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups);
2727 * Batch reads of the block allocation bitmaps
2728 * to get multiple READs in flight; limit
2729 * prefetching at cr=0/1, otherwise mballoc can
2730 * spend a lot of time loading imperfect groups
2732 if ((prefetch_grp == group) &&
2734 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2735 unsigned int curr_ios = prefetch_ios;
2737 nr = sbi->s_mb_prefetch;
2738 if (ext4_has_feature_flex_bg(sb)) {
2739 nr = 1 << sbi->s_log_groups_per_flex;
2740 nr -= group & (nr - 1);
2741 nr = min(nr, sbi->s_mb_prefetch);
2743 prefetch_grp = ext4_mb_prefetch(sb, group,
2745 if (prefetch_ios == curr_ios)
2749 /* This now checks without needing the buddy page */
2750 ret = ext4_mb_good_group_nolock(ac, group, cr);
2757 err = ext4_mb_load_buddy(sb, group, &e4b);
2761 ext4_lock_group(sb, group);
2764 * We need to check again after locking the
2767 ret = ext4_mb_good_group(ac, group, cr);
2769 ext4_unlock_group(sb, group);
2770 ext4_mb_unload_buddy(&e4b);
2774 ac->ac_groups_scanned++;
2776 ext4_mb_simple_scan_group(ac, &e4b);
2777 else if (cr == 1 && sbi->s_stripe &&
2778 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2779 ext4_mb_scan_aligned(ac, &e4b);
2781 ext4_mb_complex_scan_group(ac, &e4b);
2783 ext4_unlock_group(sb, group);
2784 ext4_mb_unload_buddy(&e4b);
2786 if (ac->ac_status != AC_STATUS_CONTINUE)
2789 /* Processed all groups and haven't found blocks */
2790 if (sbi->s_mb_stats && i == ngroups)
2791 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2794 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2795 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2797 * We've been searching too long. Let's try to allocate
2798 * the best chunk we've found so far
2800 ext4_mb_try_best_found(ac, &e4b);
2801 if (ac->ac_status != AC_STATUS_FOUND) {
2803 * Someone more lucky has already allocated it.
2804 * The only thing we can do is just take first
2807 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2808 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2809 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2810 ac->ac_b_ex.fe_len, lost);
2812 ac->ac_b_ex.fe_group = 0;
2813 ac->ac_b_ex.fe_start = 0;
2814 ac->ac_b_ex.fe_len = 0;
2815 ac->ac_status = AC_STATUS_CONTINUE;
2816 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2822 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2823 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2825 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2828 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2829 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2830 ac->ac_flags, cr, err);
2833 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2838 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2840 struct super_block *sb = pde_data(file_inode(seq->file));
2843 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2846 return (void *) ((unsigned long) group);
2849 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2851 struct super_block *sb = pde_data(file_inode(seq->file));
2855 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2858 return (void *) ((unsigned long) group);
2861 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2863 struct super_block *sb = pde_data(file_inode(seq->file));
2864 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2866 int err, buddy_loaded = 0;
2867 struct ext4_buddy e4b;
2868 struct ext4_group_info *grinfo;
2869 unsigned char blocksize_bits = min_t(unsigned char,
2870 sb->s_blocksize_bits,
2871 EXT4_MAX_BLOCK_LOG_SIZE);
2873 struct ext4_group_info info;
2874 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2879 seq_puts(seq, "#group: free frags first ["
2880 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2881 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2883 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2884 sizeof(struct ext4_group_info);
2886 grinfo = ext4_get_group_info(sb, group);
2887 /* Load the group info in memory only if not already loaded. */
2888 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2889 err = ext4_mb_load_buddy(sb, group, &e4b);
2891 seq_printf(seq, "#%-5u: I/O error\n", group);
2897 memcpy(&sg, ext4_get_group_info(sb, group), i);
2900 ext4_mb_unload_buddy(&e4b);
2902 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2903 sg.info.bb_fragments, sg.info.bb_first_free);
2904 for (i = 0; i <= 13; i++)
2905 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2906 sg.info.bb_counters[i] : 0);
2907 seq_puts(seq, " ]\n");
2912 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2916 const struct seq_operations ext4_mb_seq_groups_ops = {
2917 .start = ext4_mb_seq_groups_start,
2918 .next = ext4_mb_seq_groups_next,
2919 .stop = ext4_mb_seq_groups_stop,
2920 .show = ext4_mb_seq_groups_show,
2923 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
2925 struct super_block *sb = seq->private;
2926 struct ext4_sb_info *sbi = EXT4_SB(sb);
2928 seq_puts(seq, "mballoc:\n");
2929 if (!sbi->s_mb_stats) {
2930 seq_puts(seq, "\tmb stats collection turned off.\n");
2931 seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
2934 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
2935 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
2937 seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned));
2939 seq_puts(seq, "\tcr0_stats:\n");
2940 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0]));
2941 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2942 atomic64_read(&sbi->s_bal_cX_groups_considered[0]));
2943 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2944 atomic64_read(&sbi->s_bal_cX_failed[0]));
2945 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2946 atomic_read(&sbi->s_bal_cr0_bad_suggestions));
2948 seq_puts(seq, "\tcr1_stats:\n");
2949 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1]));
2950 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2951 atomic64_read(&sbi->s_bal_cX_groups_considered[1]));
2952 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2953 atomic64_read(&sbi->s_bal_cX_failed[1]));
2954 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2955 atomic_read(&sbi->s_bal_cr1_bad_suggestions));
2957 seq_puts(seq, "\tcr2_stats:\n");
2958 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2]));
2959 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2960 atomic64_read(&sbi->s_bal_cX_groups_considered[2]));
2961 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2962 atomic64_read(&sbi->s_bal_cX_failed[2]));
2964 seq_puts(seq, "\tcr3_stats:\n");
2965 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3]));
2966 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2967 atomic64_read(&sbi->s_bal_cX_groups_considered[3]));
2968 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2969 atomic64_read(&sbi->s_bal_cX_failed[3]));
2970 seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned));
2971 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
2972 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
2973 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
2974 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
2976 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
2977 atomic_read(&sbi->s_mb_buddies_generated),
2978 ext4_get_groups_count(sb));
2979 seq_printf(seq, "\tbuddies_time_used: %llu\n",
2980 atomic64_read(&sbi->s_mb_generation_time));
2981 seq_printf(seq, "\tpreallocated: %u\n",
2982 atomic_read(&sbi->s_mb_preallocated));
2983 seq_printf(seq, "\tdiscarded: %u\n",
2984 atomic_read(&sbi->s_mb_discarded));
2988 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
2989 __acquires(&EXT4_SB(sb)->s_mb_rb_lock)
2991 struct super_block *sb = pde_data(file_inode(seq->file));
2992 unsigned long position;
2994 read_lock(&EXT4_SB(sb)->s_mb_rb_lock);
2996 if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
2998 position = *pos + 1;
2999 return (void *) ((unsigned long) position);
3002 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
3004 struct super_block *sb = pde_data(file_inode(seq->file));
3005 unsigned long position;
3008 if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
3010 position = *pos + 1;
3011 return (void *) ((unsigned long) position);
3014 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
3016 struct super_block *sb = pde_data(file_inode(seq->file));
3017 struct ext4_sb_info *sbi = EXT4_SB(sb);
3018 unsigned long position = ((unsigned long) v);
3019 struct ext4_group_info *grp;
3021 unsigned int count, min, max;
3024 if (position >= MB_NUM_ORDERS(sb)) {
3025 seq_puts(seq, "fragment_size_tree:\n");
3026 n = rb_first(&sbi->s_mb_avg_fragment_size_root);
3028 seq_puts(seq, "\ttree_min: 0\n\ttree_max: 0\n\ttree_nodes: 0\n");
3031 grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3032 min = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3034 while (rb_next(n)) {
3038 grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3039 max = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3041 seq_printf(seq, "\ttree_min: %u\n\ttree_max: %u\n\ttree_nodes: %u\n",
3046 if (position == 0) {
3047 seq_printf(seq, "optimize_scan: %d\n",
3048 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
3049 seq_puts(seq, "max_free_order_lists:\n");
3052 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
3053 bb_largest_free_order_node)
3055 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3056 (unsigned int)position, count);
3061 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3062 __releases(&EXT4_SB(sb)->s_mb_rb_lock)
3064 struct super_block *sb = pde_data(file_inode(seq->file));
3066 read_unlock(&EXT4_SB(sb)->s_mb_rb_lock);
3069 const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3070 .start = ext4_mb_seq_structs_summary_start,
3071 .next = ext4_mb_seq_structs_summary_next,
3072 .stop = ext4_mb_seq_structs_summary_stop,
3073 .show = ext4_mb_seq_structs_summary_show,
3076 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3078 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3079 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3086 * Allocate the top-level s_group_info array for the specified number
3089 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3091 struct ext4_sb_info *sbi = EXT4_SB(sb);
3093 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3095 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3096 EXT4_DESC_PER_BLOCK_BITS(sb);
3097 if (size <= sbi->s_group_info_size)
3100 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3101 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3102 if (!new_groupinfo) {
3103 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3107 old_groupinfo = rcu_dereference(sbi->s_group_info);
3109 memcpy(new_groupinfo, old_groupinfo,
3110 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3112 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3113 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3115 ext4_kvfree_array_rcu(old_groupinfo);
3116 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3117 sbi->s_group_info_size);
3121 /* Create and initialize ext4_group_info data for the given group. */
3122 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3123 struct ext4_group_desc *desc)
3127 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3128 struct ext4_sb_info *sbi = EXT4_SB(sb);
3129 struct ext4_group_info **meta_group_info;
3130 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3133 * First check if this group is the first of a reserved block.
3134 * If it's true, we have to allocate a new table of pointers
3135 * to ext4_group_info structures
3137 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3138 metalen = sizeof(*meta_group_info) <<
3139 EXT4_DESC_PER_BLOCK_BITS(sb);
3140 meta_group_info = kmalloc(metalen, GFP_NOFS);
3141 if (meta_group_info == NULL) {
3142 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3143 "for a buddy group");
3144 goto exit_meta_group_info;
3147 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3151 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3152 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3154 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3155 if (meta_group_info[i] == NULL) {
3156 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3157 goto exit_group_info;
3159 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3160 &(meta_group_info[i]->bb_state));
3163 * initialize bb_free to be able to skip
3164 * empty groups without initialization
3166 if (ext4_has_group_desc_csum(sb) &&
3167 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3168 meta_group_info[i]->bb_free =
3169 ext4_free_clusters_after_init(sb, group, desc);
3171 meta_group_info[i]->bb_free =
3172 ext4_free_group_clusters(sb, desc);
3175 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3176 init_rwsem(&meta_group_info[i]->alloc_sem);
3177 meta_group_info[i]->bb_free_root = RB_ROOT;
3178 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3179 RB_CLEAR_NODE(&meta_group_info[i]->bb_avg_fragment_size_rb);
3180 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3181 meta_group_info[i]->bb_group = group;
3183 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3187 /* If a meta_group_info table has been allocated, release it now */
3188 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3189 struct ext4_group_info ***group_info;
3192 group_info = rcu_dereference(sbi->s_group_info);
3193 kfree(group_info[idx]);
3194 group_info[idx] = NULL;
3197 exit_meta_group_info:
3199 } /* ext4_mb_add_groupinfo */
3201 static int ext4_mb_init_backend(struct super_block *sb)
3203 ext4_group_t ngroups = ext4_get_groups_count(sb);
3205 struct ext4_sb_info *sbi = EXT4_SB(sb);
3207 struct ext4_group_desc *desc;
3208 struct ext4_group_info ***group_info;
3209 struct kmem_cache *cachep;
3211 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3215 sbi->s_buddy_cache = new_inode(sb);
3216 if (sbi->s_buddy_cache == NULL) {
3217 ext4_msg(sb, KERN_ERR, "can't get new inode");
3220 /* To avoid potentially colliding with an valid on-disk inode number,
3221 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3222 * not in the inode hash, so it should never be found by iget(), but
3223 * this will avoid confusion if it ever shows up during debugging. */
3224 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3225 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3226 for (i = 0; i < ngroups; i++) {
3228 desc = ext4_get_group_desc(sb, i, NULL);
3230 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3233 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3237 if (ext4_has_feature_flex_bg(sb)) {
3238 /* a single flex group is supposed to be read by a single IO.
3239 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3240 * unsigned integer, so the maximum shift is 32.
3242 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3243 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3246 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3247 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3248 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3250 sbi->s_mb_prefetch = 32;
3252 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3253 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3254 /* now many real IOs to prefetch within a single allocation at cr=0
3255 * given cr=0 is an CPU-related optimization we shouldn't try to
3256 * load too many groups, at some point we should start to use what
3257 * we've got in memory.
3258 * with an average random access time 5ms, it'd take a second to get
3259 * 200 groups (* N with flex_bg), so let's make this limit 4
3261 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3262 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3263 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3268 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3270 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
3271 i = sbi->s_group_info_size;
3273 group_info = rcu_dereference(sbi->s_group_info);
3275 kfree(group_info[i]);
3277 iput(sbi->s_buddy_cache);
3280 kvfree(rcu_dereference(sbi->s_group_info));
3285 static void ext4_groupinfo_destroy_slabs(void)
3289 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3290 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3291 ext4_groupinfo_caches[i] = NULL;
3295 static int ext4_groupinfo_create_slab(size_t size)
3297 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3299 int blocksize_bits = order_base_2(size);
3300 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3301 struct kmem_cache *cachep;
3303 if (cache_index >= NR_GRPINFO_CACHES)
3306 if (unlikely(cache_index < 0))
3309 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3310 if (ext4_groupinfo_caches[cache_index]) {
3311 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3312 return 0; /* Already created */
3315 slab_size = offsetof(struct ext4_group_info,
3316 bb_counters[blocksize_bits + 2]);
3318 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3319 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3322 ext4_groupinfo_caches[cache_index] = cachep;
3324 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3327 "EXT4-fs: no memory for groupinfo slab cache\n");
3334 static void ext4_discard_work(struct work_struct *work)
3336 struct ext4_sb_info *sbi = container_of(work,
3337 struct ext4_sb_info, s_discard_work);
3338 struct super_block *sb = sbi->s_sb;
3339 struct ext4_free_data *fd, *nfd;
3340 struct ext4_buddy e4b;
3341 struct list_head discard_list;
3342 ext4_group_t grp, load_grp;
3345 INIT_LIST_HEAD(&discard_list);
3346 spin_lock(&sbi->s_md_lock);
3347 list_splice_init(&sbi->s_discard_list, &discard_list);
3348 spin_unlock(&sbi->s_md_lock);
3350 load_grp = UINT_MAX;
3351 list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
3353 * If filesystem is umounting or no memory or suffering
3354 * from no space, give up the discard
3356 if ((sb->s_flags & SB_ACTIVE) && !err &&
3357 !atomic_read(&sbi->s_retry_alloc_pending)) {
3358 grp = fd->efd_group;
3359 if (grp != load_grp) {
3360 if (load_grp != UINT_MAX)
3361 ext4_mb_unload_buddy(&e4b);
3363 err = ext4_mb_load_buddy(sb, grp, &e4b);
3365 kmem_cache_free(ext4_free_data_cachep, fd);
3366 load_grp = UINT_MAX;
3373 ext4_lock_group(sb, grp);
3374 ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
3375 fd->efd_start_cluster + fd->efd_count - 1, 1);
3376 ext4_unlock_group(sb, grp);
3378 kmem_cache_free(ext4_free_data_cachep, fd);
3381 if (load_grp != UINT_MAX)
3382 ext4_mb_unload_buddy(&e4b);
3385 int ext4_mb_init(struct super_block *sb)
3387 struct ext4_sb_info *sbi = EXT4_SB(sb);
3389 unsigned offset, offset_incr;
3393 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3395 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3396 if (sbi->s_mb_offsets == NULL) {
3401 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3402 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3403 if (sbi->s_mb_maxs == NULL) {
3408 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3412 /* order 0 is regular bitmap */
3413 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3414 sbi->s_mb_offsets[0] = 0;
3418 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3419 max = sb->s_blocksize << 2;
3421 sbi->s_mb_offsets[i] = offset;
3422 sbi->s_mb_maxs[i] = max;
3423 offset += offset_incr;
3424 offset_incr = offset_incr >> 1;
3427 } while (i < MB_NUM_ORDERS(sb));
3429 sbi->s_mb_avg_fragment_size_root = RB_ROOT;
3430 sbi->s_mb_largest_free_orders =
3431 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3433 if (!sbi->s_mb_largest_free_orders) {
3437 sbi->s_mb_largest_free_orders_locks =
3438 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3440 if (!sbi->s_mb_largest_free_orders_locks) {
3444 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3445 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3446 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3448 rwlock_init(&sbi->s_mb_rb_lock);
3450 spin_lock_init(&sbi->s_md_lock);
3451 sbi->s_mb_free_pending = 0;
3452 INIT_LIST_HEAD(&sbi->s_freed_data_list);
3453 INIT_LIST_HEAD(&sbi->s_discard_list);
3454 INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
3455 atomic_set(&sbi->s_retry_alloc_pending, 0);
3457 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3458 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3459 sbi->s_mb_stats = MB_DEFAULT_STATS;
3460 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3461 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3462 sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
3464 * The default group preallocation is 512, which for 4k block
3465 * sizes translates to 2 megabytes. However for bigalloc file
3466 * systems, this is probably too big (i.e, if the cluster size
3467 * is 1 megabyte, then group preallocation size becomes half a
3468 * gigabyte!). As a default, we will keep a two megabyte
3469 * group pralloc size for cluster sizes up to 64k, and after
3470 * that, we will force a minimum group preallocation size of
3471 * 32 clusters. This translates to 8 megs when the cluster
3472 * size is 256k, and 32 megs when the cluster size is 1 meg,
3473 * which seems reasonable as a default.
3475 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3476 sbi->s_cluster_bits, 32);
3478 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3479 * to the lowest multiple of s_stripe which is bigger than
3480 * the s_mb_group_prealloc as determined above. We want
3481 * the preallocation size to be an exact multiple of the
3482 * RAID stripe size so that preallocations don't fragment
3485 if (sbi->s_stripe > 1) {
3486 sbi->s_mb_group_prealloc = roundup(
3487 sbi->s_mb_group_prealloc, sbi->s_stripe);
3490 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3491 if (sbi->s_locality_groups == NULL) {
3495 for_each_possible_cpu(i) {
3496 struct ext4_locality_group *lg;
3497 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3498 mutex_init(&lg->lg_mutex);
3499 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3500 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3501 spin_lock_init(&lg->lg_prealloc_lock);
3504 if (bdev_nonrot(sb->s_bdev))
3505 sbi->s_mb_max_linear_groups = 0;
3507 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3508 /* init file for buddy data */
3509 ret = ext4_mb_init_backend(sb);
3511 goto out_free_locality_groups;
3515 out_free_locality_groups:
3516 free_percpu(sbi->s_locality_groups);
3517 sbi->s_locality_groups = NULL;
3519 kfree(sbi->s_mb_largest_free_orders);
3520 kfree(sbi->s_mb_largest_free_orders_locks);
3521 kfree(sbi->s_mb_offsets);
3522 sbi->s_mb_offsets = NULL;
3523 kfree(sbi->s_mb_maxs);
3524 sbi->s_mb_maxs = NULL;
3528 /* need to called with the ext4 group lock held */
3529 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3531 struct ext4_prealloc_space *pa;
3532 struct list_head *cur, *tmp;
3535 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3536 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3537 list_del(&pa->pa_group_list);
3539 kmem_cache_free(ext4_pspace_cachep, pa);
3544 int ext4_mb_release(struct super_block *sb)
3546 ext4_group_t ngroups = ext4_get_groups_count(sb);
3548 int num_meta_group_infos;
3549 struct ext4_group_info *grinfo, ***group_info;
3550 struct ext4_sb_info *sbi = EXT4_SB(sb);
3551 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3554 if (test_opt(sb, DISCARD)) {
3556 * wait the discard work to drain all of ext4_free_data
3558 flush_work(&sbi->s_discard_work);
3559 WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
3562 if (sbi->s_group_info) {
3563 for (i = 0; i < ngroups; i++) {
3565 grinfo = ext4_get_group_info(sb, i);
3566 mb_group_bb_bitmap_free(grinfo);
3567 ext4_lock_group(sb, i);
3568 count = ext4_mb_cleanup_pa(grinfo);
3570 mb_debug(sb, "mballoc: %d PAs left\n",
3572 ext4_unlock_group(sb, i);
3573 kmem_cache_free(cachep, grinfo);
3575 num_meta_group_infos = (ngroups +
3576 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3577 EXT4_DESC_PER_BLOCK_BITS(sb);
3579 group_info = rcu_dereference(sbi->s_group_info);
3580 for (i = 0; i < num_meta_group_infos; i++)
3581 kfree(group_info[i]);
3585 kfree(sbi->s_mb_largest_free_orders);
3586 kfree(sbi->s_mb_largest_free_orders_locks);
3587 kfree(sbi->s_mb_offsets);
3588 kfree(sbi->s_mb_maxs);
3589 iput(sbi->s_buddy_cache);
3590 if (sbi->s_mb_stats) {
3591 ext4_msg(sb, KERN_INFO,
3592 "mballoc: %u blocks %u reqs (%u success)",
3593 atomic_read(&sbi->s_bal_allocated),
3594 atomic_read(&sbi->s_bal_reqs),
3595 atomic_read(&sbi->s_bal_success));
3596 ext4_msg(sb, KERN_INFO,
3597 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3598 "%u 2^N hits, %u breaks, %u lost",
3599 atomic_read(&sbi->s_bal_ex_scanned),
3600 atomic_read(&sbi->s_bal_groups_scanned),
3601 atomic_read(&sbi->s_bal_goals),
3602 atomic_read(&sbi->s_bal_2orders),
3603 atomic_read(&sbi->s_bal_breaks),
3604 atomic_read(&sbi->s_mb_lost_chunks));
3605 ext4_msg(sb, KERN_INFO,
3606 "mballoc: %u generated and it took %llu",
3607 atomic_read(&sbi->s_mb_buddies_generated),
3608 atomic64_read(&sbi->s_mb_generation_time));
3609 ext4_msg(sb, KERN_INFO,
3610 "mballoc: %u preallocated, %u discarded",
3611 atomic_read(&sbi->s_mb_preallocated),
3612 atomic_read(&sbi->s_mb_discarded));
3615 free_percpu(sbi->s_locality_groups);
3620 static inline int ext4_issue_discard(struct super_block *sb,
3621 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3624 ext4_fsblk_t discard_block;
3626 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3627 ext4_group_first_block_no(sb, block_group));
3628 count = EXT4_C2B(EXT4_SB(sb), count);
3629 trace_ext4_discard_blocks(sb,
3630 (unsigned long long) discard_block, count);
3632 return __blkdev_issue_discard(sb->s_bdev,
3633 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3634 (sector_t)count << (sb->s_blocksize_bits - 9),
3637 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3640 static void ext4_free_data_in_buddy(struct super_block *sb,
3641 struct ext4_free_data *entry)
3643 struct ext4_buddy e4b;
3644 struct ext4_group_info *db;
3645 int err, count = 0, count2 = 0;
3647 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3648 entry->efd_count, entry->efd_group, entry);
3650 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3651 /* we expect to find existing buddy because it's pinned */
3654 spin_lock(&EXT4_SB(sb)->s_md_lock);
3655 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3656 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3659 /* there are blocks to put in buddy to make them really free */
3660 count += entry->efd_count;
3662 ext4_lock_group(sb, entry->efd_group);
3663 /* Take it out of per group rb tree */
3664 rb_erase(&entry->efd_node, &(db->bb_free_root));
3665 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3668 * Clear the trimmed flag for the group so that the next
3669 * ext4_trim_fs can trim it.
3670 * If the volume is mounted with -o discard, online discard
3671 * is supported and the free blocks will be trimmed online.
3673 if (!test_opt(sb, DISCARD))
3674 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3676 if (!db->bb_free_root.rb_node) {
3677 /* No more items in the per group rb tree
3678 * balance refcounts from ext4_mb_free_metadata()
3680 put_page(e4b.bd_buddy_page);
3681 put_page(e4b.bd_bitmap_page);
3683 ext4_unlock_group(sb, entry->efd_group);
3684 ext4_mb_unload_buddy(&e4b);
3686 mb_debug(sb, "freed %d blocks in %d structures\n", count,
3691 * This function is called by the jbd2 layer once the commit has finished,
3692 * so we know we can free the blocks that were released with that commit.
3694 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3696 struct ext4_sb_info *sbi = EXT4_SB(sb);
3697 struct ext4_free_data *entry, *tmp;
3698 struct list_head freed_data_list;
3699 struct list_head *cut_pos = NULL;
3702 INIT_LIST_HEAD(&freed_data_list);
3704 spin_lock(&sbi->s_md_lock);
3705 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3706 if (entry->efd_tid != commit_tid)
3708 cut_pos = &entry->efd_list;
3711 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3713 spin_unlock(&sbi->s_md_lock);
3715 list_for_each_entry(entry, &freed_data_list, efd_list)
3716 ext4_free_data_in_buddy(sb, entry);
3718 if (test_opt(sb, DISCARD)) {
3719 spin_lock(&sbi->s_md_lock);
3720 wake = list_empty(&sbi->s_discard_list);
3721 list_splice_tail(&freed_data_list, &sbi->s_discard_list);
3722 spin_unlock(&sbi->s_md_lock);
3724 queue_work(system_unbound_wq, &sbi->s_discard_work);
3726 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3727 kmem_cache_free(ext4_free_data_cachep, entry);
3731 int __init ext4_init_mballoc(void)
3733 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3734 SLAB_RECLAIM_ACCOUNT);
3735 if (ext4_pspace_cachep == NULL)
3738 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3739 SLAB_RECLAIM_ACCOUNT);
3740 if (ext4_ac_cachep == NULL)
3743 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3744 SLAB_RECLAIM_ACCOUNT);
3745 if (ext4_free_data_cachep == NULL)
3751 kmem_cache_destroy(ext4_ac_cachep);
3753 kmem_cache_destroy(ext4_pspace_cachep);
3758 void ext4_exit_mballoc(void)
3761 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3762 * before destroying the slab cache.
3765 kmem_cache_destroy(ext4_pspace_cachep);
3766 kmem_cache_destroy(ext4_ac_cachep);
3767 kmem_cache_destroy(ext4_free_data_cachep);
3768 ext4_groupinfo_destroy_slabs();
3773 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3774 * Returns 0 if success or error code
3776 static noinline_for_stack int
3777 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3778 handle_t *handle, unsigned int reserv_clstrs)
3780 struct buffer_head *bitmap_bh = NULL;
3781 struct ext4_group_desc *gdp;
3782 struct buffer_head *gdp_bh;
3783 struct ext4_sb_info *sbi;
3784 struct super_block *sb;
3788 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3789 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3794 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3795 if (IS_ERR(bitmap_bh)) {
3796 err = PTR_ERR(bitmap_bh);
3801 BUFFER_TRACE(bitmap_bh, "getting write access");
3802 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
3808 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3812 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3813 ext4_free_group_clusters(sb, gdp));
3815 BUFFER_TRACE(gdp_bh, "get_write_access");
3816 err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE);
3820 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3822 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3823 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3824 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3825 "fs metadata", block, block+len);
3826 /* File system mounted not to panic on error
3827 * Fix the bitmap and return EFSCORRUPTED
3828 * We leak some of the blocks here.
3830 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3831 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3832 ac->ac_b_ex.fe_len);
3833 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3834 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3836 err = -EFSCORRUPTED;
3840 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3841 #ifdef AGGRESSIVE_CHECK
3844 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3845 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3846 bitmap_bh->b_data));
3850 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3851 ac->ac_b_ex.fe_len);
3852 if (ext4_has_group_desc_csum(sb) &&
3853 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3854 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3855 ext4_free_group_clusters_set(sb, gdp,
3856 ext4_free_clusters_after_init(sb,
3857 ac->ac_b_ex.fe_group, gdp));
3859 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3860 ext4_free_group_clusters_set(sb, gdp, len);
3861 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3862 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3864 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3865 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3867 * Now reduce the dirty block count also. Should not go negative
3869 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3870 /* release all the reserved blocks if non delalloc */
3871 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3874 if (sbi->s_log_groups_per_flex) {
3875 ext4_group_t flex_group = ext4_flex_group(sbi,
3876 ac->ac_b_ex.fe_group);
3877 atomic64_sub(ac->ac_b_ex.fe_len,
3878 &sbi_array_rcu_deref(sbi, s_flex_groups,
3879 flex_group)->free_clusters);
3882 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3885 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3893 * Idempotent helper for Ext4 fast commit replay path to set the state of
3894 * blocks in bitmaps and update counters.
3896 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
3899 struct buffer_head *bitmap_bh = NULL;
3900 struct ext4_group_desc *gdp;
3901 struct buffer_head *gdp_bh;
3902 struct ext4_sb_info *sbi = EXT4_SB(sb);
3904 ext4_grpblk_t blkoff;
3907 unsigned int clen, clen_changed, thisgrp_len;
3910 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
3913 * Check to see if we are freeing blocks across a group
3915 * In case of flex_bg, this can happen that (block, len) may
3916 * span across more than one group. In that case we need to
3917 * get the corresponding group metadata to work with.
3918 * For this we have goto again loop.
3920 thisgrp_len = min_t(unsigned int, (unsigned int)len,
3921 EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
3922 clen = EXT4_NUM_B2C(sbi, thisgrp_len);
3924 if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
3925 ext4_error(sb, "Marking blocks in system zone - "
3926 "Block = %llu, len = %u",
3927 block, thisgrp_len);
3932 bitmap_bh = ext4_read_block_bitmap(sb, group);
3933 if (IS_ERR(bitmap_bh)) {
3934 err = PTR_ERR(bitmap_bh);
3940 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
3944 ext4_lock_group(sb, group);
3946 for (i = 0; i < clen; i++)
3947 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
3951 clen_changed = clen - already;
3953 mb_set_bits(bitmap_bh->b_data, blkoff, clen);
3955 mb_clear_bits(bitmap_bh->b_data, blkoff, clen);
3956 if (ext4_has_group_desc_csum(sb) &&
3957 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3958 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3959 ext4_free_group_clusters_set(sb, gdp,
3960 ext4_free_clusters_after_init(sb, group, gdp));
3963 clen = ext4_free_group_clusters(sb, gdp) - clen_changed;
3965 clen = ext4_free_group_clusters(sb, gdp) + clen_changed;
3967 ext4_free_group_clusters_set(sb, gdp, clen);
3968 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
3969 ext4_group_desc_csum_set(sb, group, gdp);
3971 ext4_unlock_group(sb, group);
3973 if (sbi->s_log_groups_per_flex) {
3974 ext4_group_t flex_group = ext4_flex_group(sbi, group);
3975 struct flex_groups *fg = sbi_array_rcu_deref(sbi,
3976 s_flex_groups, flex_group);
3979 atomic64_sub(clen_changed, &fg->free_clusters);
3981 atomic64_add(clen_changed, &fg->free_clusters);
3985 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
3988 sync_dirty_buffer(bitmap_bh);
3989 err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
3990 sync_dirty_buffer(gdp_bh);
3994 block += thisgrp_len;
4005 * here we normalize request for locality group
4006 * Group request are normalized to s_mb_group_prealloc, which goes to
4007 * s_strip if we set the same via mount option.
4008 * s_mb_group_prealloc can be configured via
4009 * /sys/fs/ext4/<partition>/mb_group_prealloc
4011 * XXX: should we try to preallocate more than the group has now?
4013 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
4015 struct super_block *sb = ac->ac_sb;
4016 struct ext4_locality_group *lg = ac->ac_lg;
4019 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
4020 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
4024 * Normalization means making request better in terms of
4025 * size and alignment
4027 static noinline_for_stack void
4028 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
4029 struct ext4_allocation_request *ar)
4031 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4034 loff_t size, start_off;
4035 loff_t orig_size __maybe_unused;
4037 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4038 struct ext4_prealloc_space *pa;
4040 /* do normalize only data requests, metadata requests
4041 do not need preallocation */
4042 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4045 /* sometime caller may want exact blocks */
4046 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4049 /* caller may indicate that preallocation isn't
4050 * required (it's a tail, for example) */
4051 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
4054 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
4055 ext4_mb_normalize_group_request(ac);
4059 bsbits = ac->ac_sb->s_blocksize_bits;
4061 /* first, let's learn actual file size
4062 * given current request is allocated */
4063 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4064 size = size << bsbits;
4065 if (size < i_size_read(ac->ac_inode))
4066 size = i_size_read(ac->ac_inode);
4069 /* max size of free chunks */
4072 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
4073 (req <= (size) || max <= (chunk_size))
4075 /* first, try to predict filesize */
4076 /* XXX: should this table be tunable? */
4078 if (size <= 16 * 1024) {
4080 } else if (size <= 32 * 1024) {
4082 } else if (size <= 64 * 1024) {
4084 } else if (size <= 128 * 1024) {
4086 } else if (size <= 256 * 1024) {
4088 } else if (size <= 512 * 1024) {
4090 } else if (size <= 1024 * 1024) {
4092 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
4093 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4094 (21 - bsbits)) << 21;
4095 size = 2 * 1024 * 1024;
4096 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
4097 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4098 (22 - bsbits)) << 22;
4099 size = 4 * 1024 * 1024;
4100 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
4101 (8<<20)>>bsbits, max, 8 * 1024)) {
4102 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4103 (23 - bsbits)) << 23;
4104 size = 8 * 1024 * 1024;
4106 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
4107 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
4108 ac->ac_o_ex.fe_len) << bsbits;
4110 size = size >> bsbits;
4111 start = start_off >> bsbits;
4114 * For tiny groups (smaller than 8MB) the chosen allocation
4115 * alignment may be larger than group size. Make sure the
4116 * alignment does not move allocation to a different group which
4117 * makes mballoc fail assertions later.
4119 start = max(start, rounddown(ac->ac_o_ex.fe_logical,
4120 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
4122 /* don't cover already allocated blocks in selected range */
4123 if (ar->pleft && start <= ar->lleft) {
4124 size -= ar->lleft + 1 - start;
4125 start = ar->lleft + 1;
4127 if (ar->pright && start + size - 1 >= ar->lright)
4128 size -= start + size - ar->lright;
4131 * Trim allocation request for filesystems with artificially small
4134 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4135 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4139 /* check we don't cross already preallocated blocks */
4141 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4146 spin_lock(&pa->pa_lock);
4147 if (pa->pa_deleted) {
4148 spin_unlock(&pa->pa_lock);
4152 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4155 /* PA must not overlap original request */
4156 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
4157 ac->ac_o_ex.fe_logical < pa->pa_lstart));
4159 /* skip PAs this normalized request doesn't overlap with */
4160 if (pa->pa_lstart >= end || pa_end <= start) {
4161 spin_unlock(&pa->pa_lock);
4164 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
4166 /* adjust start or end to be adjacent to this pa */
4167 if (pa_end <= ac->ac_o_ex.fe_logical) {
4168 BUG_ON(pa_end < start);
4170 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4171 BUG_ON(pa->pa_lstart > end);
4172 end = pa->pa_lstart;
4174 spin_unlock(&pa->pa_lock);
4179 /* XXX: extra loop to check we really don't overlap preallocations */
4181 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4184 spin_lock(&pa->pa_lock);
4185 if (pa->pa_deleted == 0) {
4186 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4188 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
4190 spin_unlock(&pa->pa_lock);
4195 * In this function "start" and "size" are normalized for better
4196 * alignment and length such that we could preallocate more blocks.
4197 * This normalization is done such that original request of
4198 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
4199 * "size" boundaries.
4200 * (Note fe_len can be relaxed since FS block allocation API does not
4201 * provide gurantee on number of contiguous blocks allocation since that
4202 * depends upon free space left, etc).
4203 * In case of inode pa, later we use the allocated blocks
4204 * [pa_start + fe_logical - pa_lstart, fe_len/size] from the preallocated
4205 * range of goal/best blocks [start, size] to put it at the
4206 * ac_o_ex.fe_logical extent of this inode.
4207 * (See ext4_mb_use_inode_pa() for more details)
4209 if (start + size <= ac->ac_o_ex.fe_logical ||
4210 start > ac->ac_o_ex.fe_logical) {
4211 ext4_msg(ac->ac_sb, KERN_ERR,
4212 "start %lu, size %lu, fe_logical %lu",
4213 (unsigned long) start, (unsigned long) size,
4214 (unsigned long) ac->ac_o_ex.fe_logical);
4217 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4219 /* now prepare goal request */
4221 /* XXX: is it better to align blocks WRT to logical
4222 * placement or satisfy big request as is */
4223 ac->ac_g_ex.fe_logical = start;
4224 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4226 /* define goal start in order to merge */
4227 if (ar->pright && (ar->lright == (start + size))) {
4228 /* merge to the right */
4229 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4230 &ac->ac_f_ex.fe_group,
4231 &ac->ac_f_ex.fe_start);
4232 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4234 if (ar->pleft && (ar->lleft + 1 == start)) {
4235 /* merge to the left */
4236 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4237 &ac->ac_f_ex.fe_group,
4238 &ac->ac_f_ex.fe_start);
4239 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4242 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4246 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4248 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4250 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4251 atomic_inc(&sbi->s_bal_reqs);
4252 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4253 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4254 atomic_inc(&sbi->s_bal_success);
4255 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4256 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4257 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4258 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4259 atomic_inc(&sbi->s_bal_goals);
4260 if (ac->ac_found > sbi->s_mb_max_to_scan)
4261 atomic_inc(&sbi->s_bal_breaks);
4264 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4265 trace_ext4_mballoc_alloc(ac);
4267 trace_ext4_mballoc_prealloc(ac);
4271 * Called on failure; free up any blocks from the inode PA for this
4272 * context. We don't need this for MB_GROUP_PA because we only change
4273 * pa_free in ext4_mb_release_context(), but on failure, we've already
4274 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4276 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4278 struct ext4_prealloc_space *pa = ac->ac_pa;
4279 struct ext4_buddy e4b;
4283 if (ac->ac_f_ex.fe_len == 0)
4285 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4288 * This should never happen since we pin the
4289 * pages in the ext4_allocation_context so
4290 * ext4_mb_load_buddy() should never fail.
4292 WARN(1, "mb_load_buddy failed (%d)", err);
4295 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4296 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4297 ac->ac_f_ex.fe_len);
4298 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4299 ext4_mb_unload_buddy(&e4b);
4302 if (pa->pa_type == MB_INODE_PA)
4303 pa->pa_free += ac->ac_b_ex.fe_len;
4307 * use blocks preallocated to inode
4309 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4310 struct ext4_prealloc_space *pa)
4312 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4317 /* found preallocated blocks, use them */
4318 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4319 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4320 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4321 len = EXT4_NUM_B2C(sbi, end - start);
4322 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4323 &ac->ac_b_ex.fe_start);
4324 ac->ac_b_ex.fe_len = len;
4325 ac->ac_status = AC_STATUS_FOUND;
4328 BUG_ON(start < pa->pa_pstart);
4329 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4330 BUG_ON(pa->pa_free < len);
4333 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4337 * use blocks preallocated to locality group
4339 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4340 struct ext4_prealloc_space *pa)
4342 unsigned int len = ac->ac_o_ex.fe_len;
4344 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4345 &ac->ac_b_ex.fe_group,
4346 &ac->ac_b_ex.fe_start);
4347 ac->ac_b_ex.fe_len = len;
4348 ac->ac_status = AC_STATUS_FOUND;
4351 /* we don't correct pa_pstart or pa_plen here to avoid
4352 * possible race when the group is being loaded concurrently
4353 * instead we correct pa later, after blocks are marked
4354 * in on-disk bitmap -- see ext4_mb_release_context()
4355 * Other CPUs are prevented from allocating from this pa by lg_mutex
4357 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4358 pa->pa_lstart-len, len, pa);
4362 * Return the prealloc space that have minimal distance
4363 * from the goal block. @cpa is the prealloc
4364 * space that is having currently known minimal distance
4365 * from the goal block.
4367 static struct ext4_prealloc_space *
4368 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4369 struct ext4_prealloc_space *pa,
4370 struct ext4_prealloc_space *cpa)
4372 ext4_fsblk_t cur_distance, new_distance;
4375 atomic_inc(&pa->pa_count);
4378 cur_distance = abs(goal_block - cpa->pa_pstart);
4379 new_distance = abs(goal_block - pa->pa_pstart);
4381 if (cur_distance <= new_distance)
4384 /* drop the previous reference */
4385 atomic_dec(&cpa->pa_count);
4386 atomic_inc(&pa->pa_count);
4391 * search goal blocks in preallocated space
4393 static noinline_for_stack bool
4394 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4396 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4398 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4399 struct ext4_locality_group *lg;
4400 struct ext4_prealloc_space *pa, *cpa = NULL;
4401 ext4_fsblk_t goal_block;
4403 /* only data can be preallocated */
4404 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4407 /* first, try per-file preallocation */
4409 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4411 /* all fields in this condition don't change,
4412 * so we can skip locking for them */
4413 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
4414 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
4415 EXT4_C2B(sbi, pa->pa_len)))
4418 /* non-extent files can't have physical blocks past 2^32 */
4419 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4420 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
4421 EXT4_MAX_BLOCK_FILE_PHYS))
4424 /* found preallocated blocks, use them */
4425 spin_lock(&pa->pa_lock);
4426 if (pa->pa_deleted == 0 && pa->pa_free) {
4427 atomic_inc(&pa->pa_count);
4428 ext4_mb_use_inode_pa(ac, pa);
4429 spin_unlock(&pa->pa_lock);
4430 ac->ac_criteria = 10;
4434 spin_unlock(&pa->pa_lock);
4438 /* can we use group allocation? */
4439 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4442 /* inode may have no locality group for some reason */
4446 order = fls(ac->ac_o_ex.fe_len) - 1;
4447 if (order > PREALLOC_TB_SIZE - 1)
4448 /* The max size of hash table is PREALLOC_TB_SIZE */
4449 order = PREALLOC_TB_SIZE - 1;
4451 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4453 * search for the prealloc space that is having
4454 * minimal distance from the goal block.
4456 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4458 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
4460 spin_lock(&pa->pa_lock);
4461 if (pa->pa_deleted == 0 &&
4462 pa->pa_free >= ac->ac_o_ex.fe_len) {
4464 cpa = ext4_mb_check_group_pa(goal_block,
4467 spin_unlock(&pa->pa_lock);
4472 ext4_mb_use_group_pa(ac, cpa);
4473 ac->ac_criteria = 20;
4480 * the function goes through all block freed in the group
4481 * but not yet committed and marks them used in in-core bitmap.
4482 * buddy must be generated from this bitmap
4483 * Need to be called with the ext4 group lock held
4485 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
4489 struct ext4_group_info *grp;
4490 struct ext4_free_data *entry;
4492 grp = ext4_get_group_info(sb, group);
4493 n = rb_first(&(grp->bb_free_root));
4496 entry = rb_entry(n, struct ext4_free_data, efd_node);
4497 mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
4504 * the function goes through all preallocation in this group and marks them
4505 * used in in-core bitmap. buddy must be generated from this bitmap
4506 * Need to be called with ext4 group lock held
4508 static noinline_for_stack
4509 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4512 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4513 struct ext4_prealloc_space *pa;
4514 struct list_head *cur;
4515 ext4_group_t groupnr;
4516 ext4_grpblk_t start;
4517 int preallocated = 0;
4520 /* all form of preallocation discards first load group,
4521 * so the only competing code is preallocation use.
4522 * we don't need any locking here
4523 * notice we do NOT ignore preallocations with pa_deleted
4524 * otherwise we could leave used blocks available for
4525 * allocation in buddy when concurrent ext4_mb_put_pa()
4526 * is dropping preallocation
4528 list_for_each(cur, &grp->bb_prealloc_list) {
4529 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
4530 spin_lock(&pa->pa_lock);
4531 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4534 spin_unlock(&pa->pa_lock);
4535 if (unlikely(len == 0))
4537 BUG_ON(groupnr != group);
4538 mb_set_bits(bitmap, start, len);
4539 preallocated += len;
4541 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
4544 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
4545 struct ext4_prealloc_space *pa)
4547 struct ext4_inode_info *ei;
4549 if (pa->pa_deleted) {
4550 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
4551 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
4558 if (pa->pa_type == MB_INODE_PA) {
4559 ei = EXT4_I(pa->pa_inode);
4560 atomic_dec(&ei->i_prealloc_active);
4564 static void ext4_mb_pa_callback(struct rcu_head *head)
4566 struct ext4_prealloc_space *pa;
4567 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
4569 BUG_ON(atomic_read(&pa->pa_count));
4570 BUG_ON(pa->pa_deleted == 0);
4571 kmem_cache_free(ext4_pspace_cachep, pa);
4575 * drops a reference to preallocated space descriptor
4576 * if this was the last reference and the space is consumed
4578 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
4579 struct super_block *sb, struct ext4_prealloc_space *pa)
4582 ext4_fsblk_t grp_blk;
4584 /* in this short window concurrent discard can set pa_deleted */
4585 spin_lock(&pa->pa_lock);
4586 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
4587 spin_unlock(&pa->pa_lock);
4591 if (pa->pa_deleted == 1) {
4592 spin_unlock(&pa->pa_lock);
4596 ext4_mb_mark_pa_deleted(sb, pa);
4597 spin_unlock(&pa->pa_lock);
4599 grp_blk = pa->pa_pstart;
4601 * If doing group-based preallocation, pa_pstart may be in the
4602 * next group when pa is used up
4604 if (pa->pa_type == MB_GROUP_PA)
4607 grp = ext4_get_group_number(sb, grp_blk);
4612 * P1 (buddy init) P2 (regular allocation)
4613 * find block B in PA
4614 * copy on-disk bitmap to buddy
4615 * mark B in on-disk bitmap
4616 * drop PA from group
4617 * mark all PAs in buddy
4619 * thus, P1 initializes buddy with B available. to prevent this
4620 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
4623 ext4_lock_group(sb, grp);
4624 list_del(&pa->pa_group_list);
4625 ext4_unlock_group(sb, grp);
4627 spin_lock(pa->pa_obj_lock);
4628 list_del_rcu(&pa->pa_inode_list);
4629 spin_unlock(pa->pa_obj_lock);
4631 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4635 * creates new preallocated space for given inode
4637 static noinline_for_stack void
4638 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
4640 struct super_block *sb = ac->ac_sb;
4641 struct ext4_sb_info *sbi = EXT4_SB(sb);
4642 struct ext4_prealloc_space *pa;
4643 struct ext4_group_info *grp;
4644 struct ext4_inode_info *ei;
4646 /* preallocate only when found space is larger then requested */
4647 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4648 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4649 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4650 BUG_ON(ac->ac_pa == NULL);
4654 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
4660 /* we can't allocate as much as normalizer wants.
4661 * so, found space must get proper lstart
4662 * to cover original request */
4663 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
4664 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
4666 /* we're limited by original request in that
4667 * logical block must be covered any way
4668 * winl is window we can move our chunk within */
4669 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
4671 /* also, we should cover whole original request */
4672 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
4674 /* the smallest one defines real window */
4675 win = min(winl, wins);
4677 offs = ac->ac_o_ex.fe_logical %
4678 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4679 if (offs && offs < win)
4682 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
4683 EXT4_NUM_B2C(sbi, win);
4684 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
4685 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
4688 /* preallocation can change ac_b_ex, thus we store actually
4689 * allocated blocks for history */
4690 ac->ac_f_ex = ac->ac_b_ex;
4692 pa->pa_lstart = ac->ac_b_ex.fe_logical;
4693 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4694 pa->pa_len = ac->ac_b_ex.fe_len;
4695 pa->pa_free = pa->pa_len;
4696 spin_lock_init(&pa->pa_lock);
4697 INIT_LIST_HEAD(&pa->pa_inode_list);
4698 INIT_LIST_HEAD(&pa->pa_group_list);
4700 pa->pa_type = MB_INODE_PA;
4702 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4703 pa->pa_len, pa->pa_lstart);
4704 trace_ext4_mb_new_inode_pa(ac, pa);
4706 ext4_mb_use_inode_pa(ac, pa);
4707 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
4709 ei = EXT4_I(ac->ac_inode);
4710 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4712 pa->pa_obj_lock = &ei->i_prealloc_lock;
4713 pa->pa_inode = ac->ac_inode;
4715 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4717 spin_lock(pa->pa_obj_lock);
4718 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
4719 spin_unlock(pa->pa_obj_lock);
4720 atomic_inc(&ei->i_prealloc_active);
4724 * creates new preallocated space for locality group inodes belongs to
4726 static noinline_for_stack void
4727 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
4729 struct super_block *sb = ac->ac_sb;
4730 struct ext4_locality_group *lg;
4731 struct ext4_prealloc_space *pa;
4732 struct ext4_group_info *grp;
4734 /* preallocate only when found space is larger then requested */
4735 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4736 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4737 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4738 BUG_ON(ac->ac_pa == NULL);
4742 /* preallocation can change ac_b_ex, thus we store actually
4743 * allocated blocks for history */
4744 ac->ac_f_ex = ac->ac_b_ex;
4746 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4747 pa->pa_lstart = pa->pa_pstart;
4748 pa->pa_len = ac->ac_b_ex.fe_len;
4749 pa->pa_free = pa->pa_len;
4750 spin_lock_init(&pa->pa_lock);
4751 INIT_LIST_HEAD(&pa->pa_inode_list);
4752 INIT_LIST_HEAD(&pa->pa_group_list);
4754 pa->pa_type = MB_GROUP_PA;
4756 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4757 pa->pa_len, pa->pa_lstart);
4758 trace_ext4_mb_new_group_pa(ac, pa);
4760 ext4_mb_use_group_pa(ac, pa);
4761 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4763 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4767 pa->pa_obj_lock = &lg->lg_prealloc_lock;
4768 pa->pa_inode = NULL;
4770 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4773 * We will later add the new pa to the right bucket
4774 * after updating the pa_free in ext4_mb_release_context
4778 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4780 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4781 ext4_mb_new_group_pa(ac);
4783 ext4_mb_new_inode_pa(ac);
4787 * finds all unused blocks in on-disk bitmap, frees them in
4788 * in-core bitmap and buddy.
4789 * @pa must be unlinked from inode and group lists, so that
4790 * nobody else can find/use it.
4791 * the caller MUST hold group/inode locks.
4792 * TODO: optimize the case when there are no in-core structures yet
4794 static noinline_for_stack int
4795 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4796 struct ext4_prealloc_space *pa)
4798 struct super_block *sb = e4b->bd_sb;
4799 struct ext4_sb_info *sbi = EXT4_SB(sb);
4804 unsigned long long grp_blk_start;
4807 BUG_ON(pa->pa_deleted == 0);
4808 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4809 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4810 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4811 end = bit + pa->pa_len;
4814 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4817 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4818 mb_debug(sb, "free preallocated %u/%u in group %u\n",
4819 (unsigned) ext4_group_first_block_no(sb, group) + bit,
4820 (unsigned) next - bit, (unsigned) group);
4823 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4824 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4825 EXT4_C2B(sbi, bit)),
4827 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4830 if (free != pa->pa_free) {
4831 ext4_msg(e4b->bd_sb, KERN_CRIT,
4832 "pa %p: logic %lu, phys. %lu, len %d",
4833 pa, (unsigned long) pa->pa_lstart,
4834 (unsigned long) pa->pa_pstart,
4836 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4839 * pa is already deleted so we use the value obtained
4840 * from the bitmap and continue.
4843 atomic_add(free, &sbi->s_mb_discarded);
4848 static noinline_for_stack int
4849 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4850 struct ext4_prealloc_space *pa)
4852 struct super_block *sb = e4b->bd_sb;
4856 trace_ext4_mb_release_group_pa(sb, pa);
4857 BUG_ON(pa->pa_deleted == 0);
4858 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4859 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4860 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4861 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4862 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4868 * releases all preallocations in given group
4870 * first, we need to decide discard policy:
4871 * - when do we discard
4873 * - how many do we discard
4874 * 1) how many requested
4876 static noinline_for_stack int
4877 ext4_mb_discard_group_preallocations(struct super_block *sb,
4878 ext4_group_t group, int *busy)
4880 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4881 struct buffer_head *bitmap_bh = NULL;
4882 struct ext4_prealloc_space *pa, *tmp;
4883 struct list_head list;
4884 struct ext4_buddy e4b;
4888 mb_debug(sb, "discard preallocation for group %u\n", group);
4889 if (list_empty(&grp->bb_prealloc_list))
4892 bitmap_bh = ext4_read_block_bitmap(sb, group);
4893 if (IS_ERR(bitmap_bh)) {
4894 err = PTR_ERR(bitmap_bh);
4895 ext4_error_err(sb, -err,
4896 "Error %d reading block bitmap for %u",
4901 err = ext4_mb_load_buddy(sb, group, &e4b);
4903 ext4_warning(sb, "Error %d loading buddy information for %u",
4909 INIT_LIST_HEAD(&list);
4910 ext4_lock_group(sb, group);
4911 list_for_each_entry_safe(pa, tmp,
4912 &grp->bb_prealloc_list, pa_group_list) {
4913 spin_lock(&pa->pa_lock);
4914 if (atomic_read(&pa->pa_count)) {
4915 spin_unlock(&pa->pa_lock);
4919 if (pa->pa_deleted) {
4920 spin_unlock(&pa->pa_lock);
4924 /* seems this one can be freed ... */
4925 ext4_mb_mark_pa_deleted(sb, pa);
4928 this_cpu_inc(discard_pa_seq);
4930 /* we can trust pa_free ... */
4931 free += pa->pa_free;
4933 spin_unlock(&pa->pa_lock);
4935 list_del(&pa->pa_group_list);
4936 list_add(&pa->u.pa_tmp_list, &list);
4939 /* now free all selected PAs */
4940 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4942 /* remove from object (inode or locality group) */
4943 spin_lock(pa->pa_obj_lock);
4944 list_del_rcu(&pa->pa_inode_list);
4945 spin_unlock(pa->pa_obj_lock);
4947 if (pa->pa_type == MB_GROUP_PA)
4948 ext4_mb_release_group_pa(&e4b, pa);
4950 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4952 list_del(&pa->u.pa_tmp_list);
4953 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4956 ext4_unlock_group(sb, group);
4957 ext4_mb_unload_buddy(&e4b);
4960 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4961 free, group, grp->bb_free);
4966 * releases all non-used preallocated blocks for given inode
4968 * It's important to discard preallocations under i_data_sem
4969 * We don't want another block to be served from the prealloc
4970 * space when we are discarding the inode prealloc space.
4972 * FIXME!! Make sure it is valid at all the call sites
4974 void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
4976 struct ext4_inode_info *ei = EXT4_I(inode);
4977 struct super_block *sb = inode->i_sb;
4978 struct buffer_head *bitmap_bh = NULL;
4979 struct ext4_prealloc_space *pa, *tmp;
4980 ext4_group_t group = 0;
4981 struct list_head list;
4982 struct ext4_buddy e4b;
4985 if (!S_ISREG(inode->i_mode)) {
4986 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4990 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
4993 mb_debug(sb, "discard preallocation for inode %lu\n",
4995 trace_ext4_discard_preallocations(inode,
4996 atomic_read(&ei->i_prealloc_active), needed);
4998 INIT_LIST_HEAD(&list);
5004 /* first, collect all pa's in the inode */
5005 spin_lock(&ei->i_prealloc_lock);
5006 while (!list_empty(&ei->i_prealloc_list) && needed) {
5007 pa = list_entry(ei->i_prealloc_list.prev,
5008 struct ext4_prealloc_space, pa_inode_list);
5009 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
5010 spin_lock(&pa->pa_lock);
5011 if (atomic_read(&pa->pa_count)) {
5012 /* this shouldn't happen often - nobody should
5013 * use preallocation while we're discarding it */
5014 spin_unlock(&pa->pa_lock);
5015 spin_unlock(&ei->i_prealloc_lock);
5016 ext4_msg(sb, KERN_ERR,
5017 "uh-oh! used pa while discarding");
5019 schedule_timeout_uninterruptible(HZ);
5023 if (pa->pa_deleted == 0) {
5024 ext4_mb_mark_pa_deleted(sb, pa);
5025 spin_unlock(&pa->pa_lock);
5026 list_del_rcu(&pa->pa_inode_list);
5027 list_add(&pa->u.pa_tmp_list, &list);
5032 /* someone is deleting pa right now */
5033 spin_unlock(&pa->pa_lock);
5034 spin_unlock(&ei->i_prealloc_lock);
5036 /* we have to wait here because pa_deleted
5037 * doesn't mean pa is already unlinked from
5038 * the list. as we might be called from
5039 * ->clear_inode() the inode will get freed
5040 * and concurrent thread which is unlinking
5041 * pa from inode's list may access already
5042 * freed memory, bad-bad-bad */
5044 /* XXX: if this happens too often, we can
5045 * add a flag to force wait only in case
5046 * of ->clear_inode(), but not in case of
5047 * regular truncate */
5048 schedule_timeout_uninterruptible(HZ);
5051 spin_unlock(&ei->i_prealloc_lock);
5053 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5054 BUG_ON(pa->pa_type != MB_INODE_PA);
5055 group = ext4_get_group_number(sb, pa->pa_pstart);
5057 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5058 GFP_NOFS|__GFP_NOFAIL);
5060 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5065 bitmap_bh = ext4_read_block_bitmap(sb, group);
5066 if (IS_ERR(bitmap_bh)) {
5067 err = PTR_ERR(bitmap_bh);
5068 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
5070 ext4_mb_unload_buddy(&e4b);
5074 ext4_lock_group(sb, group);
5075 list_del(&pa->pa_group_list);
5076 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5077 ext4_unlock_group(sb, group);
5079 ext4_mb_unload_buddy(&e4b);
5082 list_del(&pa->u.pa_tmp_list);
5083 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5087 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
5089 struct ext4_prealloc_space *pa;
5091 BUG_ON(ext4_pspace_cachep == NULL);
5092 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
5095 atomic_set(&pa->pa_count, 1);
5100 static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
5102 struct ext4_prealloc_space *pa = ac->ac_pa;
5106 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
5107 kmem_cache_free(ext4_pspace_cachep, pa);
5110 #ifdef CONFIG_EXT4_DEBUG
5111 static inline void ext4_mb_show_pa(struct super_block *sb)
5113 ext4_group_t i, ngroups;
5115 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5118 ngroups = ext4_get_groups_count(sb);
5119 mb_debug(sb, "groups: ");
5120 for (i = 0; i < ngroups; i++) {
5121 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5122 struct ext4_prealloc_space *pa;
5123 ext4_grpblk_t start;
5124 struct list_head *cur;
5125 ext4_lock_group(sb, i);
5126 list_for_each(cur, &grp->bb_prealloc_list) {
5127 pa = list_entry(cur, struct ext4_prealloc_space,
5129 spin_lock(&pa->pa_lock);
5130 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5132 spin_unlock(&pa->pa_lock);
5133 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5136 ext4_unlock_group(sb, i);
5137 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5142 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5144 struct super_block *sb = ac->ac_sb;
5146 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5149 mb_debug(sb, "Can't allocate:"
5150 " Allocation context details:");
5151 mb_debug(sb, "status %u flags 0x%x",
5152 ac->ac_status, ac->ac_flags);
5153 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5154 "goal %lu/%lu/%lu@%lu, "
5155 "best %lu/%lu/%lu@%lu cr %d",
5156 (unsigned long)ac->ac_o_ex.fe_group,
5157 (unsigned long)ac->ac_o_ex.fe_start,
5158 (unsigned long)ac->ac_o_ex.fe_len,
5159 (unsigned long)ac->ac_o_ex.fe_logical,
5160 (unsigned long)ac->ac_g_ex.fe_group,
5161 (unsigned long)ac->ac_g_ex.fe_start,
5162 (unsigned long)ac->ac_g_ex.fe_len,
5163 (unsigned long)ac->ac_g_ex.fe_logical,
5164 (unsigned long)ac->ac_b_ex.fe_group,
5165 (unsigned long)ac->ac_b_ex.fe_start,
5166 (unsigned long)ac->ac_b_ex.fe_len,
5167 (unsigned long)ac->ac_b_ex.fe_logical,
5168 (int)ac->ac_criteria);
5169 mb_debug(sb, "%u found", ac->ac_found);
5170 ext4_mb_show_pa(sb);
5173 static inline void ext4_mb_show_pa(struct super_block *sb)
5177 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5179 ext4_mb_show_pa(ac->ac_sb);
5185 * We use locality group preallocation for small size file. The size of the
5186 * file is determined by the current size or the resulting size after
5187 * allocation which ever is larger
5189 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5191 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5193 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5194 int bsbits = ac->ac_sb->s_blocksize_bits;
5197 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5200 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5203 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
5204 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5207 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5208 !inode_is_open_for_write(ac->ac_inode)) {
5209 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5213 if (sbi->s_mb_group_prealloc <= 0) {
5214 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5218 /* don't use group allocation for large files */
5219 size = max(size, isize);
5220 if (size > sbi->s_mb_stream_request) {
5221 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5225 BUG_ON(ac->ac_lg != NULL);
5227 * locality group prealloc space are per cpu. The reason for having
5228 * per cpu locality group is to reduce the contention between block
5229 * request from multiple CPUs.
5231 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5233 /* we're going to use group allocation */
5234 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5236 /* serialize all allocations in the group */
5237 mutex_lock(&ac->ac_lg->lg_mutex);
5240 static noinline_for_stack int
5241 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5242 struct ext4_allocation_request *ar)
5244 struct super_block *sb = ar->inode->i_sb;
5245 struct ext4_sb_info *sbi = EXT4_SB(sb);
5246 struct ext4_super_block *es = sbi->s_es;
5250 ext4_grpblk_t block;
5252 /* we can't allocate > group size */
5255 /* just a dirty hack to filter too big requests */
5256 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5257 len = EXT4_CLUSTERS_PER_GROUP(sb);
5259 /* start searching from the goal */
5261 if (goal < le32_to_cpu(es->s_first_data_block) ||
5262 goal >= ext4_blocks_count(es))
5263 goal = le32_to_cpu(es->s_first_data_block);
5264 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5266 /* set up allocation goals */
5267 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5268 ac->ac_status = AC_STATUS_CONTINUE;
5270 ac->ac_inode = ar->inode;
5271 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5272 ac->ac_o_ex.fe_group = group;
5273 ac->ac_o_ex.fe_start = block;
5274 ac->ac_o_ex.fe_len = len;
5275 ac->ac_g_ex = ac->ac_o_ex;
5276 ac->ac_flags = ar->flags;
5278 /* we have to define context: we'll work with a file or
5279 * locality group. this is a policy, actually */
5280 ext4_mb_group_or_file(ac);
5282 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5283 "left: %u/%u, right %u/%u to %swritable\n",
5284 (unsigned) ar->len, (unsigned) ar->logical,
5285 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5286 (unsigned) ar->lleft, (unsigned) ar->pleft,
5287 (unsigned) ar->lright, (unsigned) ar->pright,
5288 inode_is_open_for_write(ar->inode) ? "" : "non-");
5293 static noinline_for_stack void
5294 ext4_mb_discard_lg_preallocations(struct super_block *sb,
5295 struct ext4_locality_group *lg,
5296 int order, int total_entries)
5298 ext4_group_t group = 0;
5299 struct ext4_buddy e4b;
5300 struct list_head discard_list;
5301 struct ext4_prealloc_space *pa, *tmp;
5303 mb_debug(sb, "discard locality group preallocation\n");
5305 INIT_LIST_HEAD(&discard_list);
5307 spin_lock(&lg->lg_prealloc_lock);
5308 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5310 lockdep_is_held(&lg->lg_prealloc_lock)) {
5311 spin_lock(&pa->pa_lock);
5312 if (atomic_read(&pa->pa_count)) {
5314 * This is the pa that we just used
5315 * for block allocation. So don't
5318 spin_unlock(&pa->pa_lock);
5321 if (pa->pa_deleted) {
5322 spin_unlock(&pa->pa_lock);
5325 /* only lg prealloc space */
5326 BUG_ON(pa->pa_type != MB_GROUP_PA);
5328 /* seems this one can be freed ... */
5329 ext4_mb_mark_pa_deleted(sb, pa);
5330 spin_unlock(&pa->pa_lock);
5332 list_del_rcu(&pa->pa_inode_list);
5333 list_add(&pa->u.pa_tmp_list, &discard_list);
5336 if (total_entries <= 5) {
5338 * we want to keep only 5 entries
5339 * allowing it to grow to 8. This
5340 * mak sure we don't call discard
5341 * soon for this list.
5346 spin_unlock(&lg->lg_prealloc_lock);
5348 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5351 group = ext4_get_group_number(sb, pa->pa_pstart);
5352 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5353 GFP_NOFS|__GFP_NOFAIL);
5355 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5359 ext4_lock_group(sb, group);
5360 list_del(&pa->pa_group_list);
5361 ext4_mb_release_group_pa(&e4b, pa);
5362 ext4_unlock_group(sb, group);
5364 ext4_mb_unload_buddy(&e4b);
5365 list_del(&pa->u.pa_tmp_list);
5366 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5371 * We have incremented pa_count. So it cannot be freed at this
5372 * point. Also we hold lg_mutex. So no parallel allocation is
5373 * possible from this lg. That means pa_free cannot be updated.
5375 * A parallel ext4_mb_discard_group_preallocations is possible.
5376 * which can cause the lg_prealloc_list to be updated.
5379 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5381 int order, added = 0, lg_prealloc_count = 1;
5382 struct super_block *sb = ac->ac_sb;
5383 struct ext4_locality_group *lg = ac->ac_lg;
5384 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5386 order = fls(pa->pa_free) - 1;
5387 if (order > PREALLOC_TB_SIZE - 1)
5388 /* The max size of hash table is PREALLOC_TB_SIZE */
5389 order = PREALLOC_TB_SIZE - 1;
5390 /* Add the prealloc space to lg */
5391 spin_lock(&lg->lg_prealloc_lock);
5392 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5394 lockdep_is_held(&lg->lg_prealloc_lock)) {
5395 spin_lock(&tmp_pa->pa_lock);
5396 if (tmp_pa->pa_deleted) {
5397 spin_unlock(&tmp_pa->pa_lock);
5400 if (!added && pa->pa_free < tmp_pa->pa_free) {
5401 /* Add to the tail of the previous entry */
5402 list_add_tail_rcu(&pa->pa_inode_list,
5403 &tmp_pa->pa_inode_list);
5406 * we want to count the total
5407 * number of entries in the list
5410 spin_unlock(&tmp_pa->pa_lock);
5411 lg_prealloc_count++;
5414 list_add_tail_rcu(&pa->pa_inode_list,
5415 &lg->lg_prealloc_list[order]);
5416 spin_unlock(&lg->lg_prealloc_lock);
5418 /* Now trim the list to be not more than 8 elements */
5419 if (lg_prealloc_count > 8) {
5420 ext4_mb_discard_lg_preallocations(sb, lg,
5421 order, lg_prealloc_count);
5428 * if per-inode prealloc list is too long, trim some PA
5430 static void ext4_mb_trim_inode_pa(struct inode *inode)
5432 struct ext4_inode_info *ei = EXT4_I(inode);
5433 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5436 count = atomic_read(&ei->i_prealloc_active);
5437 delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
5438 if (count > sbi->s_mb_max_inode_prealloc + delta) {
5439 count -= sbi->s_mb_max_inode_prealloc;
5440 ext4_discard_preallocations(inode, count);
5445 * release all resource we used in allocation
5447 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
5449 struct inode *inode = ac->ac_inode;
5450 struct ext4_inode_info *ei = EXT4_I(inode);
5451 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5452 struct ext4_prealloc_space *pa = ac->ac_pa;
5454 if (pa->pa_type == MB_GROUP_PA) {
5455 /* see comment in ext4_mb_use_group_pa() */
5456 spin_lock(&pa->pa_lock);
5457 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5458 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5459 pa->pa_free -= ac->ac_b_ex.fe_len;
5460 pa->pa_len -= ac->ac_b_ex.fe_len;
5461 spin_unlock(&pa->pa_lock);
5464 * We want to add the pa to the right bucket.
5465 * Remove it from the list and while adding
5466 * make sure the list to which we are adding
5469 if (likely(pa->pa_free)) {
5470 spin_lock(pa->pa_obj_lock);
5471 list_del_rcu(&pa->pa_inode_list);
5472 spin_unlock(pa->pa_obj_lock);
5473 ext4_mb_add_n_trim(ac);
5477 if (pa->pa_type == MB_INODE_PA) {
5479 * treat per-inode prealloc list as a lru list, then try
5480 * to trim the least recently used PA.
5482 spin_lock(pa->pa_obj_lock);
5483 list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
5484 spin_unlock(pa->pa_obj_lock);
5487 ext4_mb_put_pa(ac, ac->ac_sb, pa);
5489 if (ac->ac_bitmap_page)
5490 put_page(ac->ac_bitmap_page);
5491 if (ac->ac_buddy_page)
5492 put_page(ac->ac_buddy_page);
5493 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5494 mutex_unlock(&ac->ac_lg->lg_mutex);
5495 ext4_mb_collect_stats(ac);
5496 ext4_mb_trim_inode_pa(inode);
5500 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
5502 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
5504 int freed = 0, busy = 0;
5507 trace_ext4_mb_discard_preallocations(sb, needed);
5510 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
5512 for (i = 0; i < ngroups && needed > 0; i++) {
5513 ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
5519 if (needed > 0 && busy && ++retry < 3) {
5527 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
5528 struct ext4_allocation_context *ac, u64 *seq)
5534 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
5539 seq_retry = ext4_get_discard_pa_seq_sum();
5540 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
5541 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
5547 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
5551 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5552 struct ext4_allocation_request *ar, int *errp);
5555 * Main entry point into mballoc to allocate blocks
5556 * it tries to use preallocation first, then falls back
5557 * to usual allocation
5559 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
5560 struct ext4_allocation_request *ar, int *errp)
5562 struct ext4_allocation_context *ac = NULL;
5563 struct ext4_sb_info *sbi;
5564 struct super_block *sb;
5565 ext4_fsblk_t block = 0;
5566 unsigned int inquota = 0;
5567 unsigned int reserv_clstrs = 0;
5571 sb = ar->inode->i_sb;
5574 trace_ext4_request_blocks(ar);
5575 if (sbi->s_mount_state & EXT4_FC_REPLAY)
5576 return ext4_mb_new_blocks_simple(handle, ar, errp);
5578 /* Allow to use superuser reservation for quota file */
5579 if (ext4_is_quota_file(ar->inode))
5580 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
5582 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
5583 /* Without delayed allocation we need to verify
5584 * there is enough free blocks to do block allocation
5585 * and verify allocation doesn't exceed the quota limits.
5588 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
5590 /* let others to free the space */
5592 ar->len = ar->len >> 1;
5595 ext4_mb_show_pa(sb);
5599 reserv_clstrs = ar->len;
5600 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
5601 dquot_alloc_block_nofail(ar->inode,
5602 EXT4_C2B(sbi, ar->len));
5605 dquot_alloc_block(ar->inode,
5606 EXT4_C2B(sbi, ar->len))) {
5608 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
5619 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
5626 *errp = ext4_mb_initialize_context(ac, ar);
5632 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
5633 seq = this_cpu_read(discard_pa_seq);
5634 if (!ext4_mb_use_preallocated(ac)) {
5635 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
5636 ext4_mb_normalize_request(ac, ar);
5638 *errp = ext4_mb_pa_alloc(ac);
5642 /* allocate space in core */
5643 *errp = ext4_mb_regular_allocator(ac);
5645 * pa allocated above is added to grp->bb_prealloc_list only
5646 * when we were able to allocate some block i.e. when
5647 * ac->ac_status == AC_STATUS_FOUND.
5648 * And error from above mean ac->ac_status != AC_STATUS_FOUND
5649 * So we have to free this pa here itself.
5652 ext4_mb_pa_free(ac);
5653 ext4_discard_allocated_blocks(ac);
5656 if (ac->ac_status == AC_STATUS_FOUND &&
5657 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
5658 ext4_mb_pa_free(ac);
5660 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
5661 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
5663 ext4_discard_allocated_blocks(ac);
5666 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5667 ar->len = ac->ac_b_ex.fe_len;
5670 if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
5673 * If block allocation fails then the pa allocated above
5674 * needs to be freed here itself.
5676 ext4_mb_pa_free(ac);
5682 ac->ac_b_ex.fe_len = 0;
5684 ext4_mb_show_ac(ac);
5686 ext4_mb_release_context(ac);
5689 kmem_cache_free(ext4_ac_cachep, ac);
5690 if (inquota && ar->len < inquota)
5691 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
5693 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
5694 /* release all the reserved blocks if non delalloc */
5695 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
5699 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
5705 * We can merge two free data extents only if the physical blocks
5706 * are contiguous, AND the extents were freed by the same transaction,
5707 * AND the blocks are associated with the same group.
5709 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
5710 struct ext4_free_data *entry,
5711 struct ext4_free_data *new_entry,
5712 struct rb_root *entry_rb_root)
5714 if ((entry->efd_tid != new_entry->efd_tid) ||
5715 (entry->efd_group != new_entry->efd_group))
5717 if (entry->efd_start_cluster + entry->efd_count ==
5718 new_entry->efd_start_cluster) {
5719 new_entry->efd_start_cluster = entry->efd_start_cluster;
5720 new_entry->efd_count += entry->efd_count;
5721 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
5722 entry->efd_start_cluster) {
5723 new_entry->efd_count += entry->efd_count;
5726 spin_lock(&sbi->s_md_lock);
5727 list_del(&entry->efd_list);
5728 spin_unlock(&sbi->s_md_lock);
5729 rb_erase(&entry->efd_node, entry_rb_root);
5730 kmem_cache_free(ext4_free_data_cachep, entry);
5733 static noinline_for_stack int
5734 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
5735 struct ext4_free_data *new_entry)
5737 ext4_group_t group = e4b->bd_group;
5738 ext4_grpblk_t cluster;
5739 ext4_grpblk_t clusters = new_entry->efd_count;
5740 struct ext4_free_data *entry;
5741 struct ext4_group_info *db = e4b->bd_info;
5742 struct super_block *sb = e4b->bd_sb;
5743 struct ext4_sb_info *sbi = EXT4_SB(sb);
5744 struct rb_node **n = &db->bb_free_root.rb_node, *node;
5745 struct rb_node *parent = NULL, *new_node;
5747 BUG_ON(!ext4_handle_valid(handle));
5748 BUG_ON(e4b->bd_bitmap_page == NULL);
5749 BUG_ON(e4b->bd_buddy_page == NULL);
5751 new_node = &new_entry->efd_node;
5752 cluster = new_entry->efd_start_cluster;
5755 /* first free block exent. We need to
5756 protect buddy cache from being freed,
5757 * otherwise we'll refresh it from
5758 * on-disk bitmap and lose not-yet-available
5760 get_page(e4b->bd_buddy_page);
5761 get_page(e4b->bd_bitmap_page);
5765 entry = rb_entry(parent, struct ext4_free_data, efd_node);
5766 if (cluster < entry->efd_start_cluster)
5768 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
5769 n = &(*n)->rb_right;
5771 ext4_grp_locked_error(sb, group, 0,
5772 ext4_group_first_block_no(sb, group) +
5773 EXT4_C2B(sbi, cluster),
5774 "Block already on to-be-freed list");
5775 kmem_cache_free(ext4_free_data_cachep, new_entry);
5780 rb_link_node(new_node, parent, n);
5781 rb_insert_color(new_node, &db->bb_free_root);
5783 /* Now try to see the extent can be merged to left and right */
5784 node = rb_prev(new_node);
5786 entry = rb_entry(node, struct ext4_free_data, efd_node);
5787 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5788 &(db->bb_free_root));
5791 node = rb_next(new_node);
5793 entry = rb_entry(node, struct ext4_free_data, efd_node);
5794 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5795 &(db->bb_free_root));
5798 spin_lock(&sbi->s_md_lock);
5799 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5800 sbi->s_mb_free_pending += clusters;
5801 spin_unlock(&sbi->s_md_lock);
5806 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
5807 * linearly starting at the goal block and also excludes the blocks which
5808 * are going to be in use after fast commit replay.
5810 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5811 struct ext4_allocation_request *ar, int *errp)
5813 struct buffer_head *bitmap_bh;
5814 struct super_block *sb = ar->inode->i_sb;
5816 ext4_grpblk_t blkoff;
5817 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
5818 ext4_grpblk_t i = 0;
5819 ext4_fsblk_t goal, block;
5820 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
5823 if (goal < le32_to_cpu(es->s_first_data_block) ||
5824 goal >= ext4_blocks_count(es))
5825 goal = le32_to_cpu(es->s_first_data_block);
5828 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
5829 for (; group < ext4_get_groups_count(sb); group++) {
5830 bitmap_bh = ext4_read_block_bitmap(sb, group);
5831 if (IS_ERR(bitmap_bh)) {
5832 *errp = PTR_ERR(bitmap_bh);
5833 pr_warn("Failed to read block bitmap\n");
5837 ext4_get_group_no_and_offset(sb,
5838 max(ext4_group_first_block_no(sb, group), goal),
5841 i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
5845 if (ext4_fc_replay_check_excluded(sb,
5846 ext4_group_first_block_no(sb, group) + i)) {
5856 if (group >= ext4_get_groups_count(sb) || i >= max) {
5861 block = ext4_group_first_block_no(sb, group) + i;
5862 ext4_mb_mark_bb(sb, block, 1, 1);
5868 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
5869 unsigned long count)
5871 struct buffer_head *bitmap_bh;
5872 struct super_block *sb = inode->i_sb;
5873 struct ext4_group_desc *gdp;
5874 struct buffer_head *gdp_bh;
5876 ext4_grpblk_t blkoff;
5877 int already_freed = 0, err, i;
5879 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
5880 bitmap_bh = ext4_read_block_bitmap(sb, group);
5881 if (IS_ERR(bitmap_bh)) {
5882 err = PTR_ERR(bitmap_bh);
5883 pr_warn("Failed to read block bitmap\n");
5886 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
5890 for (i = 0; i < count; i++) {
5891 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data))
5894 mb_clear_bits(bitmap_bh->b_data, blkoff, count);
5895 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
5898 ext4_free_group_clusters_set(
5899 sb, gdp, ext4_free_group_clusters(sb, gdp) +
5900 count - already_freed);
5901 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
5902 ext4_group_desc_csum_set(sb, group, gdp);
5903 ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
5904 sync_dirty_buffer(bitmap_bh);
5905 sync_dirty_buffer(gdp_bh);
5910 * ext4_mb_clear_bb() -- helper function for freeing blocks.
5911 * Used by ext4_free_blocks()
5912 * @handle: handle for this transaction
5914 * @block: starting physical block to be freed
5915 * @count: number of blocks to be freed
5916 * @flags: flags used by ext4_free_blocks
5918 static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
5919 ext4_fsblk_t block, unsigned long count,
5922 struct buffer_head *bitmap_bh = NULL;
5923 struct super_block *sb = inode->i_sb;
5924 struct ext4_group_desc *gdp;
5925 unsigned int overflow;
5927 struct buffer_head *gd_bh;
5928 ext4_group_t block_group;
5929 struct ext4_sb_info *sbi;
5930 struct ext4_buddy e4b;
5931 unsigned int count_clusters;
5937 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5938 !ext4_inode_block_valid(inode, block, count)) {
5939 ext4_error(sb, "Freeing blocks in system zone - "
5940 "Block = %llu, count = %lu", block, count);
5941 /* err = 0. ext4_std_error should be a no op */
5944 flags |= EXT4_FREE_BLOCKS_VALIDATED;
5948 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5950 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5951 ext4_get_group_info(sb, block_group))))
5955 * Check to see if we are freeing blocks across a group
5958 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5959 overflow = EXT4_C2B(sbi, bit) + count -
5960 EXT4_BLOCKS_PER_GROUP(sb);
5962 /* The range changed so it's no longer validated */
5963 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
5965 count_clusters = EXT4_NUM_B2C(sbi, count);
5966 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5967 if (IS_ERR(bitmap_bh)) {
5968 err = PTR_ERR(bitmap_bh);
5972 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5978 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5979 !ext4_inode_block_valid(inode, block, count)) {
5980 ext4_error(sb, "Freeing blocks in system zone - "
5981 "Block = %llu, count = %lu", block, count);
5982 /* err = 0. ext4_std_error should be a no op */
5986 BUFFER_TRACE(bitmap_bh, "getting write access");
5987 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
5993 * We are about to modify some metadata. Call the journal APIs
5994 * to unshare ->b_data if a currently-committing transaction is
5997 BUFFER_TRACE(gd_bh, "get_write_access");
5998 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
6001 #ifdef AGGRESSIVE_CHECK
6004 for (i = 0; i < count_clusters; i++)
6005 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
6008 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
6010 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
6011 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
6012 GFP_NOFS|__GFP_NOFAIL);
6017 * We need to make sure we don't reuse the freed block until after the
6018 * transaction is committed. We make an exception if the inode is to be
6019 * written in writeback mode since writeback mode has weak data
6020 * consistency guarantees.
6022 if (ext4_handle_valid(handle) &&
6023 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
6024 !ext4_should_writeback_data(inode))) {
6025 struct ext4_free_data *new_entry;
6027 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
6030 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
6031 GFP_NOFS|__GFP_NOFAIL);
6032 new_entry->efd_start_cluster = bit;
6033 new_entry->efd_group = block_group;
6034 new_entry->efd_count = count_clusters;
6035 new_entry->efd_tid = handle->h_transaction->t_tid;
6037 ext4_lock_group(sb, block_group);
6038 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6039 ext4_mb_free_metadata(handle, &e4b, new_entry);
6041 /* need to update group_info->bb_free and bitmap
6042 * with group lock held. generate_buddy look at
6043 * them with group lock_held
6045 if (test_opt(sb, DISCARD)) {
6046 err = ext4_issue_discard(sb, block_group, bit, count,
6048 if (err && err != -EOPNOTSUPP)
6049 ext4_msg(sb, KERN_WARNING, "discard request in"
6050 " group:%u block:%d count:%lu failed"
6051 " with %d", block_group, bit, count,
6054 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
6056 ext4_lock_group(sb, block_group);
6057 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6058 mb_free_blocks(inode, &e4b, bit, count_clusters);
6061 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
6062 ext4_free_group_clusters_set(sb, gdp, ret);
6063 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
6064 ext4_group_desc_csum_set(sb, block_group, gdp);
6065 ext4_unlock_group(sb, block_group);
6067 if (sbi->s_log_groups_per_flex) {
6068 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6069 atomic64_add(count_clusters,
6070 &sbi_array_rcu_deref(sbi, s_flex_groups,
6071 flex_group)->free_clusters);
6075 * on a bigalloc file system, defer the s_freeclusters_counter
6076 * update to the caller (ext4_remove_space and friends) so they
6077 * can determine if a cluster freed here should be rereserved
6079 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6080 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6081 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6082 percpu_counter_add(&sbi->s_freeclusters_counter,
6086 ext4_mb_unload_buddy(&e4b);
6088 /* We dirtied the bitmap block */
6089 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6090 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6092 /* And the group descriptor block */
6093 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6094 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6098 if (overflow && !err) {
6102 /* The range changed so it's no longer validated */
6103 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6108 ext4_std_error(sb, err);
6113 * ext4_free_blocks() -- Free given blocks and update quota
6114 * @handle: handle for this transaction
6116 * @bh: optional buffer of the block to be freed
6117 * @block: starting physical block to be freed
6118 * @count: number of blocks to be freed
6119 * @flags: flags used by ext4_free_blocks
6121 void ext4_free_blocks(handle_t *handle, struct inode *inode,
6122 struct buffer_head *bh, ext4_fsblk_t block,
6123 unsigned long count, int flags)
6125 struct super_block *sb = inode->i_sb;
6126 unsigned int overflow;
6127 struct ext4_sb_info *sbi;
6131 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
6132 ext4_free_blocks_simple(inode, block, count);
6139 BUG_ON(block != bh->b_blocknr);
6141 block = bh->b_blocknr;
6144 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6145 !ext4_inode_block_valid(inode, block, count)) {
6146 ext4_error(sb, "Freeing blocks not in datazone - "
6147 "block = %llu, count = %lu", block, count);
6150 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6152 ext4_debug("freeing block %llu\n", block);
6153 trace_ext4_free_blocks(inode, block, count, flags);
6155 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6158 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
6163 * If the extent to be freed does not begin on a cluster
6164 * boundary, we need to deal with partial clusters at the
6165 * beginning and end of the extent. Normally we will free
6166 * blocks at the beginning or the end unless we are explicitly
6167 * requested to avoid doing so.
6169 overflow = EXT4_PBLK_COFF(sbi, block);
6171 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
6172 overflow = sbi->s_cluster_ratio - overflow;
6174 if (count > overflow)
6182 /* The range changed so it's no longer validated */
6183 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6185 overflow = EXT4_LBLK_COFF(sbi, count);
6187 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
6188 if (count > overflow)
6193 count += sbi->s_cluster_ratio - overflow;
6194 /* The range changed so it's no longer validated */
6195 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6198 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6200 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
6202 for (i = 0; i < count; i++) {
6205 bh = sb_find_get_block(inode->i_sb, block + i);
6206 ext4_forget(handle, is_metadata, inode, bh, block + i);
6210 ext4_mb_clear_bb(handle, inode, block, count, flags);
6215 * ext4_group_add_blocks() -- Add given blocks to an existing group
6216 * @handle: handle to this transaction
6218 * @block: start physical block to add to the block group
6219 * @count: number of blocks to free
6221 * This marks the blocks as free in the bitmap and buddy.
6223 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6224 ext4_fsblk_t block, unsigned long count)
6226 struct buffer_head *bitmap_bh = NULL;
6227 struct buffer_head *gd_bh;
6228 ext4_group_t block_group;
6231 struct ext4_group_desc *desc;
6232 struct ext4_sb_info *sbi = EXT4_SB(sb);
6233 struct ext4_buddy e4b;
6234 int err = 0, ret, free_clusters_count;
6235 ext4_grpblk_t clusters_freed;
6236 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6237 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6238 unsigned long cluster_count = last_cluster - first_cluster + 1;
6240 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6245 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6247 * Check to see if we are freeing blocks across a group
6250 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6251 ext4_warning(sb, "too many blocks added to group %u",
6257 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
6258 if (IS_ERR(bitmap_bh)) {
6259 err = PTR_ERR(bitmap_bh);
6264 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
6270 if (!ext4_sb_block_valid(sb, NULL, block, count)) {
6271 ext4_error(sb, "Adding blocks in system zones - "
6272 "Block = %llu, count = %lu",
6278 BUFFER_TRACE(bitmap_bh, "getting write access");
6279 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
6285 * We are about to modify some metadata. Call the journal APIs
6286 * to unshare ->b_data if a currently-committing transaction is
6289 BUFFER_TRACE(gd_bh, "get_write_access");
6290 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
6294 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
6295 BUFFER_TRACE(bitmap_bh, "clear bit");
6296 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
6297 ext4_error(sb, "bit already cleared for block %llu",
6298 (ext4_fsblk_t)(block + i));
6299 BUFFER_TRACE(bitmap_bh, "bit already cleared");
6305 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6310 * need to update group_info->bb_free and bitmap
6311 * with group lock held. generate_buddy look at
6312 * them with group lock_held
6314 ext4_lock_group(sb, block_group);
6315 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
6316 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6317 free_clusters_count = clusters_freed +
6318 ext4_free_group_clusters(sb, desc);
6319 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
6320 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
6321 ext4_group_desc_csum_set(sb, block_group, desc);
6322 ext4_unlock_group(sb, block_group);
6323 percpu_counter_add(&sbi->s_freeclusters_counter,
6326 if (sbi->s_log_groups_per_flex) {
6327 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6328 atomic64_add(clusters_freed,
6329 &sbi_array_rcu_deref(sbi, s_flex_groups,
6330 flex_group)->free_clusters);
6333 ext4_mb_unload_buddy(&e4b);
6335 /* We dirtied the bitmap block */
6336 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6337 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6339 /* And the group descriptor block */
6340 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6341 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6347 ext4_std_error(sb, err);
6352 * ext4_trim_extent -- function to TRIM one single free extent in the group
6353 * @sb: super block for the file system
6354 * @start: starting block of the free extent in the alloc. group
6355 * @count: number of blocks to TRIM
6356 * @e4b: ext4 buddy for the group
6358 * Trim "count" blocks starting at "start" in the "group". To assure that no
6359 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6360 * be called with under the group lock.
6362 static int ext4_trim_extent(struct super_block *sb,
6363 int start, int count, struct ext4_buddy *e4b)
6367 struct ext4_free_extent ex;
6368 ext4_group_t group = e4b->bd_group;
6371 trace_ext4_trim_extent(sb, group, start, count);
6373 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6375 ex.fe_start = start;
6376 ex.fe_group = group;
6380 * Mark blocks used, so no one can reuse them while
6383 mb_mark_used(e4b, &ex);
6384 ext4_unlock_group(sb, group);
6385 ret = ext4_issue_discard(sb, group, start, count, NULL);
6386 ext4_lock_group(sb, group);
6387 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6391 static int ext4_try_to_trim_range(struct super_block *sb,
6392 struct ext4_buddy *e4b, ext4_grpblk_t start,
6393 ext4_grpblk_t max, ext4_grpblk_t minblocks)
6394 __acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
6395 __releases(ext4_group_lock_ptr(sb, e4b->bd_group))
6397 ext4_grpblk_t next, count, free_count;
6400 bitmap = e4b->bd_bitmap;
6401 start = (e4b->bd_info->bb_first_free > start) ?
6402 e4b->bd_info->bb_first_free : start;
6406 while (start <= max) {
6407 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6410 next = mb_find_next_bit(bitmap, max + 1, start);
6412 if ((next - start) >= minblocks) {
6413 int ret = ext4_trim_extent(sb, start, next - start, e4b);
6415 if (ret && ret != -EOPNOTSUPP)
6417 count += next - start;
6419 free_count += next - start;
6422 if (fatal_signal_pending(current)) {
6423 count = -ERESTARTSYS;
6427 if (need_resched()) {
6428 ext4_unlock_group(sb, e4b->bd_group);
6430 ext4_lock_group(sb, e4b->bd_group);
6433 if ((e4b->bd_info->bb_free - free_count) < minblocks)
6441 * ext4_trim_all_free -- function to trim all free space in alloc. group
6442 * @sb: super block for file system
6443 * @group: group to be trimmed
6444 * @start: first group block to examine
6445 * @max: last group block to examine
6446 * @minblocks: minimum extent block count
6447 * @set_trimmed: set the trimmed flag if at least one block is trimmed
6449 * ext4_trim_all_free walks through group's block bitmap searching for free
6450 * extents. When the free extent is found, mark it as used in group buddy
6451 * bitmap. Then issue a TRIM command on this extent and free the extent in
6452 * the group buddy bitmap.
6454 static ext4_grpblk_t
6455 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6456 ext4_grpblk_t start, ext4_grpblk_t max,
6457 ext4_grpblk_t minblocks, bool set_trimmed)
6459 struct ext4_buddy e4b;
6462 trace_ext4_trim_all_free(sb, group, start, max);
6464 ret = ext4_mb_load_buddy(sb, group, &e4b);
6466 ext4_warning(sb, "Error %d loading buddy information for %u",
6471 ext4_lock_group(sb, group);
6473 if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
6474 minblocks < EXT4_SB(sb)->s_last_trim_minblks) {
6475 ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
6476 if (ret >= 0 && set_trimmed)
6477 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
6482 ext4_unlock_group(sb, group);
6483 ext4_mb_unload_buddy(&e4b);
6485 ext4_debug("trimmed %d blocks in the group %d\n",
6492 * ext4_trim_fs() -- trim ioctl handle function
6493 * @sb: superblock for filesystem
6494 * @range: fstrim_range structure
6496 * start: First Byte to trim
6497 * len: number of Bytes to trim from start
6498 * minlen: minimum extent length in Bytes
6499 * ext4_trim_fs goes through all allocation groups containing Bytes from
6500 * start to start+len. For each such a group ext4_trim_all_free function
6501 * is invoked to trim all free space.
6503 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6505 unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
6506 struct ext4_group_info *grp;
6507 ext4_group_t group, first_group, last_group;
6508 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6509 uint64_t start, end, minlen, trimmed = 0;
6510 ext4_fsblk_t first_data_blk =
6511 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6512 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6513 bool whole_group, eof = false;
6516 start = range->start >> sb->s_blocksize_bits;
6517 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6518 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6519 range->minlen >> sb->s_blocksize_bits);
6521 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6522 start >= max_blks ||
6523 range->len < sb->s_blocksize)
6525 /* No point to try to trim less than discard granularity */
6526 if (range->minlen < discard_granularity) {
6527 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6528 discard_granularity >> sb->s_blocksize_bits);
6529 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
6532 if (end >= max_blks - 1) {
6536 if (end <= first_data_blk)
6538 if (start < first_data_blk)
6539 start = first_data_blk;
6541 /* Determine first and last group to examine based on start and end */
6542 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6543 &first_group, &first_cluster);
6544 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6545 &last_group, &last_cluster);
6547 /* end now represents the last cluster to discard in this group */
6548 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6551 for (group = first_group; group <= last_group; group++) {
6552 grp = ext4_get_group_info(sb, group);
6553 /* We only do this if the grp has never been initialized */
6554 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6555 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6561 * For all the groups except the last one, last cluster will
6562 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6563 * change it for the last group, note that last_cluster is
6564 * already computed earlier by ext4_get_group_no_and_offset()
6566 if (group == last_group) {
6568 whole_group = eof ? true : end == EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6570 if (grp->bb_free >= minlen) {
6571 cnt = ext4_trim_all_free(sb, group, first_cluster,
6572 end, minlen, whole_group);
6581 * For every group except the first one, we are sure
6582 * that the first cluster to discard will be cluster #0.
6588 EXT4_SB(sb)->s_last_trim_minblks = minlen;
6591 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6595 /* Iterate all the free extents in the group. */
6597 ext4_mballoc_query_range(
6598 struct super_block *sb,
6600 ext4_grpblk_t start,
6602 ext4_mballoc_query_range_fn formatter,
6607 struct ext4_buddy e4b;
6610 error = ext4_mb_load_buddy(sb, group, &e4b);
6613 bitmap = e4b.bd_bitmap;
6615 ext4_lock_group(sb, group);
6617 start = (e4b.bd_info->bb_first_free > start) ?
6618 e4b.bd_info->bb_first_free : start;
6619 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
6620 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6622 while (start <= end) {
6623 start = mb_find_next_zero_bit(bitmap, end + 1, start);
6626 next = mb_find_next_bit(bitmap, end + 1, start);
6628 ext4_unlock_group(sb, group);
6629 error = formatter(sb, group, start, next - start, priv);
6632 ext4_lock_group(sb, group);
6637 ext4_unlock_group(sb, group);
6639 ext4_mb_unload_buddy(&e4b);