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) {
1053 *group = next_linear_group(ac, *group, ngroups);
1058 ext4_mb_choose_next_group_cr0(ac, new_cr, group, ngroups);
1059 } else if (*new_cr == 1) {
1060 ext4_mb_choose_next_group_cr1(ac, new_cr, group, ngroups);
1063 * TODO: For CR=2, we can arrange groups in an rb tree sorted by
1064 * bb_free. But until that happens, we should never come here.
1071 * Cache the order of the largest free extent we have available in this block
1075 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
1077 struct ext4_sb_info *sbi = EXT4_SB(sb);
1080 if (test_opt2(sb, MB_OPTIMIZE_SCAN) && grp->bb_largest_free_order >= 0) {
1081 write_lock(&sbi->s_mb_largest_free_orders_locks[
1082 grp->bb_largest_free_order]);
1083 list_del_init(&grp->bb_largest_free_order_node);
1084 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1085 grp->bb_largest_free_order]);
1087 grp->bb_largest_free_order = -1; /* uninit */
1089 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) {
1090 if (grp->bb_counters[i] > 0) {
1091 grp->bb_largest_free_order = i;
1095 if (test_opt2(sb, MB_OPTIMIZE_SCAN) &&
1096 grp->bb_largest_free_order >= 0 && grp->bb_free) {
1097 write_lock(&sbi->s_mb_largest_free_orders_locks[
1098 grp->bb_largest_free_order]);
1099 list_add_tail(&grp->bb_largest_free_order_node,
1100 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1101 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1102 grp->bb_largest_free_order]);
1106 static noinline_for_stack
1107 void ext4_mb_generate_buddy(struct super_block *sb,
1108 void *buddy, void *bitmap, ext4_group_t group)
1110 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
1111 struct ext4_sb_info *sbi = EXT4_SB(sb);
1112 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1113 ext4_grpblk_t i = 0;
1114 ext4_grpblk_t first;
1117 unsigned fragments = 0;
1118 unsigned long long period = get_cycles();
1120 /* initialize buddy from bitmap which is aggregation
1121 * of on-disk bitmap and preallocations */
1122 i = mb_find_next_zero_bit(bitmap, max, 0);
1123 grp->bb_first_free = i;
1127 i = mb_find_next_bit(bitmap, max, i);
1131 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1133 grp->bb_counters[0]++;
1135 i = mb_find_next_zero_bit(bitmap, max, i);
1137 grp->bb_fragments = fragments;
1139 if (free != grp->bb_free) {
1140 ext4_grp_locked_error(sb, group, 0, 0,
1141 "block bitmap and bg descriptor "
1142 "inconsistent: %u vs %u free clusters",
1143 free, grp->bb_free);
1145 * If we intend to continue, we consider group descriptor
1146 * corrupt and update bb_free using bitmap value
1148 grp->bb_free = free;
1149 ext4_mark_group_bitmap_corrupted(sb, group,
1150 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1152 mb_set_largest_free_order(sb, grp);
1154 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1156 period = get_cycles() - period;
1157 atomic_inc(&sbi->s_mb_buddies_generated);
1158 atomic64_add(period, &sbi->s_mb_generation_time);
1159 mb_update_avg_fragment_size(sb, grp);
1162 /* The buddy information is attached the buddy cache inode
1163 * for convenience. The information regarding each group
1164 * is loaded via ext4_mb_load_buddy. The information involve
1165 * block bitmap and buddy information. The information are
1166 * stored in the inode as
1169 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1172 * one block each for bitmap and buddy information.
1173 * So for each group we take up 2 blocks. A page can
1174 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
1175 * So it can have information regarding groups_per_page which
1176 * is blocks_per_page/2
1178 * Locking note: This routine takes the block group lock of all groups
1179 * for this page; do not hold this lock when calling this routine!
1182 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
1184 ext4_group_t ngroups;
1186 int blocks_per_page;
1187 int groups_per_page;
1190 ext4_group_t first_group, group;
1192 struct super_block *sb;
1193 struct buffer_head *bhs;
1194 struct buffer_head **bh = NULL;
1195 struct inode *inode;
1198 struct ext4_group_info *grinfo;
1200 inode = page->mapping->host;
1202 ngroups = ext4_get_groups_count(sb);
1203 blocksize = i_blocksize(inode);
1204 blocks_per_page = PAGE_SIZE / blocksize;
1206 mb_debug(sb, "init page %lu\n", page->index);
1208 groups_per_page = blocks_per_page >> 1;
1209 if (groups_per_page == 0)
1210 groups_per_page = 1;
1212 /* allocate buffer_heads to read bitmaps */
1213 if (groups_per_page > 1) {
1214 i = sizeof(struct buffer_head *) * groups_per_page;
1215 bh = kzalloc(i, gfp);
1223 first_group = page->index * blocks_per_page / 2;
1225 /* read all groups the page covers into the cache */
1226 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1227 if (group >= ngroups)
1230 grinfo = ext4_get_group_info(sb, group);
1232 * If page is uptodate then we came here after online resize
1233 * which added some new uninitialized group info structs, so
1234 * we must skip all initialized uptodate buddies on the page,
1235 * which may be currently in use by an allocating task.
1237 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1241 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1242 if (IS_ERR(bh[i])) {
1243 err = PTR_ERR(bh[i]);
1247 mb_debug(sb, "read bitmap for group %u\n", group);
1250 /* wait for I/O completion */
1251 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1256 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1261 first_block = page->index * blocks_per_page;
1262 for (i = 0; i < blocks_per_page; i++) {
1263 group = (first_block + i) >> 1;
1264 if (group >= ngroups)
1267 if (!bh[group - first_group])
1268 /* skip initialized uptodate buddy */
1271 if (!buffer_verified(bh[group - first_group]))
1272 /* Skip faulty bitmaps */
1277 * data carry information regarding this
1278 * particular group in the format specified
1282 data = page_address(page) + (i * blocksize);
1283 bitmap = bh[group - first_group]->b_data;
1286 * We place the buddy block and bitmap block
1289 if ((first_block + i) & 1) {
1290 /* this is block of buddy */
1291 BUG_ON(incore == NULL);
1292 mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
1293 group, page->index, i * blocksize);
1294 trace_ext4_mb_buddy_bitmap_load(sb, group);
1295 grinfo = ext4_get_group_info(sb, group);
1296 grinfo->bb_fragments = 0;
1297 memset(grinfo->bb_counters, 0,
1298 sizeof(*grinfo->bb_counters) *
1299 (MB_NUM_ORDERS(sb)));
1301 * incore got set to the group block bitmap below
1303 ext4_lock_group(sb, group);
1304 /* init the buddy */
1305 memset(data, 0xff, blocksize);
1306 ext4_mb_generate_buddy(sb, data, incore, group);
1307 ext4_unlock_group(sb, group);
1310 /* this is block of bitmap */
1311 BUG_ON(incore != NULL);
1312 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
1313 group, page->index, i * blocksize);
1314 trace_ext4_mb_bitmap_load(sb, group);
1316 /* see comments in ext4_mb_put_pa() */
1317 ext4_lock_group(sb, group);
1318 memcpy(data, bitmap, blocksize);
1320 /* mark all preallocated blks used in in-core bitmap */
1321 ext4_mb_generate_from_pa(sb, data, group);
1322 ext4_mb_generate_from_freelist(sb, data, group);
1323 ext4_unlock_group(sb, group);
1325 /* set incore so that the buddy information can be
1326 * generated using this
1331 SetPageUptodate(page);
1335 for (i = 0; i < groups_per_page; i++)
1344 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1345 * on the same buddy page doesn't happen whild holding the buddy page lock.
1346 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1347 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1349 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1350 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1352 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1353 int block, pnum, poff;
1354 int blocks_per_page;
1357 e4b->bd_buddy_page = NULL;
1358 e4b->bd_bitmap_page = NULL;
1360 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1362 * the buddy cache inode stores the block bitmap
1363 * and buddy information in consecutive blocks.
1364 * So for each group we need two blocks.
1367 pnum = block / blocks_per_page;
1368 poff = block % blocks_per_page;
1369 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1372 BUG_ON(page->mapping != inode->i_mapping);
1373 e4b->bd_bitmap_page = page;
1374 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1376 if (blocks_per_page >= 2) {
1377 /* buddy and bitmap are on the same page */
1382 pnum = block / blocks_per_page;
1383 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1386 BUG_ON(page->mapping != inode->i_mapping);
1387 e4b->bd_buddy_page = page;
1391 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1393 if (e4b->bd_bitmap_page) {
1394 unlock_page(e4b->bd_bitmap_page);
1395 put_page(e4b->bd_bitmap_page);
1397 if (e4b->bd_buddy_page) {
1398 unlock_page(e4b->bd_buddy_page);
1399 put_page(e4b->bd_buddy_page);
1404 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1405 * block group lock of all groups for this page; do not hold the BG lock when
1406 * calling this routine!
1408 static noinline_for_stack
1409 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1412 struct ext4_group_info *this_grp;
1413 struct ext4_buddy e4b;
1418 mb_debug(sb, "init group %u\n", group);
1419 this_grp = ext4_get_group_info(sb, group);
1421 * This ensures that we don't reinit the buddy cache
1422 * page which map to the group from which we are already
1423 * allocating. If we are looking at the buddy cache we would
1424 * have taken a reference using ext4_mb_load_buddy and that
1425 * would have pinned buddy page to page cache.
1426 * The call to ext4_mb_get_buddy_page_lock will mark the
1429 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1430 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1432 * somebody initialized the group
1433 * return without doing anything
1438 page = e4b.bd_bitmap_page;
1439 ret = ext4_mb_init_cache(page, NULL, gfp);
1442 if (!PageUptodate(page)) {
1447 if (e4b.bd_buddy_page == NULL) {
1449 * If both the bitmap and buddy are in
1450 * the same page we don't need to force
1456 /* init buddy cache */
1457 page = e4b.bd_buddy_page;
1458 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1461 if (!PageUptodate(page)) {
1466 ext4_mb_put_buddy_page_lock(&e4b);
1471 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1472 * block group lock of all groups for this page; do not hold the BG lock when
1473 * calling this routine!
1475 static noinline_for_stack int
1476 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1477 struct ext4_buddy *e4b, gfp_t gfp)
1479 int blocks_per_page;
1485 struct ext4_group_info *grp;
1486 struct ext4_sb_info *sbi = EXT4_SB(sb);
1487 struct inode *inode = sbi->s_buddy_cache;
1490 mb_debug(sb, "load group %u\n", group);
1492 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1493 grp = ext4_get_group_info(sb, group);
1495 e4b->bd_blkbits = sb->s_blocksize_bits;
1498 e4b->bd_group = group;
1499 e4b->bd_buddy_page = NULL;
1500 e4b->bd_bitmap_page = NULL;
1502 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1504 * we need full data about the group
1505 * to make a good selection
1507 ret = ext4_mb_init_group(sb, group, gfp);
1513 * the buddy cache inode stores the block bitmap
1514 * and buddy information in consecutive blocks.
1515 * So for each group we need two blocks.
1518 pnum = block / blocks_per_page;
1519 poff = block % blocks_per_page;
1521 /* we could use find_or_create_page(), but it locks page
1522 * what we'd like to avoid in fast path ... */
1523 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1524 if (page == NULL || !PageUptodate(page)) {
1527 * drop the page reference and try
1528 * to get the page with lock. If we
1529 * are not uptodate that implies
1530 * somebody just created the page but
1531 * is yet to initialize the same. So
1532 * wait for it to initialize.
1535 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1537 BUG_ON(page->mapping != inode->i_mapping);
1538 if (!PageUptodate(page)) {
1539 ret = ext4_mb_init_cache(page, NULL, gfp);
1544 mb_cmp_bitmaps(e4b, page_address(page) +
1545 (poff * sb->s_blocksize));
1554 if (!PageUptodate(page)) {
1559 /* Pages marked accessed already */
1560 e4b->bd_bitmap_page = page;
1561 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1564 pnum = block / blocks_per_page;
1565 poff = block % blocks_per_page;
1567 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1568 if (page == NULL || !PageUptodate(page)) {
1571 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1573 BUG_ON(page->mapping != inode->i_mapping);
1574 if (!PageUptodate(page)) {
1575 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1589 if (!PageUptodate(page)) {
1594 /* Pages marked accessed already */
1595 e4b->bd_buddy_page = page;
1596 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1603 if (e4b->bd_bitmap_page)
1604 put_page(e4b->bd_bitmap_page);
1605 if (e4b->bd_buddy_page)
1606 put_page(e4b->bd_buddy_page);
1607 e4b->bd_buddy = NULL;
1608 e4b->bd_bitmap = NULL;
1612 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1613 struct ext4_buddy *e4b)
1615 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1618 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1620 if (e4b->bd_bitmap_page)
1621 put_page(e4b->bd_bitmap_page);
1622 if (e4b->bd_buddy_page)
1623 put_page(e4b->bd_buddy_page);
1627 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1632 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1633 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1635 while (order <= e4b->bd_blkbits + 1) {
1636 bb = mb_find_buddy(e4b, order, &max);
1637 if (!mb_test_bit(block >> order, bb)) {
1638 /* this block is part of buddy of order 'order' */
1646 static void mb_clear_bits(void *bm, int cur, int len)
1652 if ((cur & 31) == 0 && (len - cur) >= 32) {
1653 /* fast path: clear whole word at once */
1654 addr = bm + (cur >> 3);
1659 mb_clear_bit(cur, bm);
1664 /* clear bits in given range
1665 * will return first found zero bit if any, -1 otherwise
1667 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1674 if ((cur & 31) == 0 && (len - cur) >= 32) {
1675 /* fast path: clear whole word at once */
1676 addr = bm + (cur >> 3);
1677 if (*addr != (__u32)(-1) && zero_bit == -1)
1678 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1683 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1691 void mb_set_bits(void *bm, int cur, int len)
1697 if ((cur & 31) == 0 && (len - cur) >= 32) {
1698 /* fast path: set whole word at once */
1699 addr = bm + (cur >> 3);
1704 mb_set_bit(cur, bm);
1709 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1711 if (mb_test_bit(*bit + side, bitmap)) {
1712 mb_clear_bit(*bit, bitmap);
1718 mb_set_bit(*bit, bitmap);
1723 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1727 void *buddy = mb_find_buddy(e4b, order, &max);
1732 /* Bits in range [first; last] are known to be set since
1733 * corresponding blocks were allocated. Bits in range
1734 * (first; last) will stay set because they form buddies on
1735 * upper layer. We just deal with borders if they don't
1736 * align with upper layer and then go up.
1737 * Releasing entire group is all about clearing
1738 * single bit of highest order buddy.
1742 * ---------------------------------
1744 * ---------------------------------
1745 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1746 * ---------------------------------
1748 * \_____________________/
1750 * Neither [1] nor [6] is aligned to above layer.
1751 * Left neighbour [0] is free, so mark it busy,
1752 * decrease bb_counters and extend range to
1754 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1755 * mark [6] free, increase bb_counters and shrink range to
1757 * Then shift range to [0; 2], go up and do the same.
1762 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1764 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1769 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1770 mb_clear_bits(buddy, first, last - first + 1);
1771 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1780 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1781 int first, int count)
1783 int left_is_free = 0;
1784 int right_is_free = 0;
1786 int last = first + count - 1;
1787 struct super_block *sb = e4b->bd_sb;
1789 if (WARN_ON(count == 0))
1791 BUG_ON(last >= (sb->s_blocksize << 3));
1792 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1793 /* Don't bother if the block group is corrupt. */
1794 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1797 mb_check_buddy(e4b);
1798 mb_free_blocks_double(inode, e4b, first, count);
1800 this_cpu_inc(discard_pa_seq);
1801 e4b->bd_info->bb_free += count;
1802 if (first < e4b->bd_info->bb_first_free)
1803 e4b->bd_info->bb_first_free = first;
1805 /* access memory sequentially: check left neighbour,
1806 * clear range and then check right neighbour
1809 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1810 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1811 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1812 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1814 if (unlikely(block != -1)) {
1815 struct ext4_sb_info *sbi = EXT4_SB(sb);
1816 ext4_fsblk_t blocknr;
1818 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1819 blocknr += EXT4_C2B(sbi, block);
1820 if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) {
1821 ext4_grp_locked_error(sb, e4b->bd_group,
1822 inode ? inode->i_ino : 0,
1824 "freeing already freed block (bit %u); block bitmap corrupt.",
1826 ext4_mark_group_bitmap_corrupted(
1828 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1833 /* let's maintain fragments counter */
1834 if (left_is_free && right_is_free)
1835 e4b->bd_info->bb_fragments--;
1836 else if (!left_is_free && !right_is_free)
1837 e4b->bd_info->bb_fragments++;
1839 /* buddy[0] == bd_bitmap is a special case, so handle
1840 * it right away and let mb_buddy_mark_free stay free of
1841 * zero order checks.
1842 * Check if neighbours are to be coaleasced,
1843 * adjust bitmap bb_counters and borders appropriately.
1846 first += !left_is_free;
1847 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1850 last -= !right_is_free;
1851 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1855 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1858 mb_set_largest_free_order(sb, e4b->bd_info);
1859 mb_update_avg_fragment_size(sb, e4b->bd_info);
1860 mb_check_buddy(e4b);
1863 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1864 int needed, struct ext4_free_extent *ex)
1870 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1873 buddy = mb_find_buddy(e4b, 0, &max);
1874 BUG_ON(buddy == NULL);
1875 BUG_ON(block >= max);
1876 if (mb_test_bit(block, buddy)) {
1883 /* find actual order */
1884 order = mb_find_order_for_block(e4b, block);
1885 block = block >> order;
1887 ex->fe_len = 1 << order;
1888 ex->fe_start = block << order;
1889 ex->fe_group = e4b->bd_group;
1891 /* calc difference from given start */
1892 next = next - ex->fe_start;
1894 ex->fe_start += next;
1896 while (needed > ex->fe_len &&
1897 mb_find_buddy(e4b, order, &max)) {
1899 if (block + 1 >= max)
1902 next = (block + 1) * (1 << order);
1903 if (mb_test_bit(next, e4b->bd_bitmap))
1906 order = mb_find_order_for_block(e4b, next);
1908 block = next >> order;
1909 ex->fe_len += 1 << order;
1912 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1913 /* Should never happen! (but apparently sometimes does?!?) */
1915 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
1916 "corruption or bug in mb_find_extent "
1917 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1918 block, order, needed, ex->fe_group, ex->fe_start,
1919 ex->fe_len, ex->fe_logical);
1927 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1933 int start = ex->fe_start;
1934 int len = ex->fe_len;
1939 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1940 BUG_ON(e4b->bd_group != ex->fe_group);
1941 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1942 mb_check_buddy(e4b);
1943 mb_mark_used_double(e4b, start, len);
1945 this_cpu_inc(discard_pa_seq);
1946 e4b->bd_info->bb_free -= len;
1947 if (e4b->bd_info->bb_first_free == start)
1948 e4b->bd_info->bb_first_free += len;
1950 /* let's maintain fragments counter */
1952 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1953 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1954 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1956 e4b->bd_info->bb_fragments++;
1957 else if (!mlen && !max)
1958 e4b->bd_info->bb_fragments--;
1960 /* 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! */
1967 buddy = mb_find_buddy(e4b, ord, &max);
1968 BUG_ON((start >> ord) >= max);
1969 mb_set_bit(start >> ord, buddy);
1970 e4b->bd_info->bb_counters[ord]--;
1977 /* store for history */
1979 ret = len | (ord << 16);
1981 /* we have to split large buddy */
1983 buddy = mb_find_buddy(e4b, ord, &max);
1984 mb_set_bit(start >> ord, buddy);
1985 e4b->bd_info->bb_counters[ord]--;
1988 cur = (start >> ord) & ~1U;
1989 buddy = mb_find_buddy(e4b, ord, &max);
1990 mb_clear_bit(cur, buddy);
1991 mb_clear_bit(cur + 1, buddy);
1992 e4b->bd_info->bb_counters[ord]++;
1993 e4b->bd_info->bb_counters[ord]++;
1995 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1997 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
1998 mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1999 mb_check_buddy(e4b);
2005 * Must be called under group lock!
2007 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
2008 struct ext4_buddy *e4b)
2010 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2013 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
2014 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2016 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
2017 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
2018 ret = mb_mark_used(e4b, &ac->ac_b_ex);
2020 /* preallocation can change ac_b_ex, thus we store actually
2021 * allocated blocks for history */
2022 ac->ac_f_ex = ac->ac_b_ex;
2024 ac->ac_status = AC_STATUS_FOUND;
2025 ac->ac_tail = ret & 0xffff;
2026 ac->ac_buddy = ret >> 16;
2029 * take the page reference. We want the page to be pinned
2030 * so that we don't get a ext4_mb_init_cache_call for this
2031 * group until we update the bitmap. That would mean we
2032 * double allocate blocks. The reference is dropped
2033 * in ext4_mb_release_context
2035 ac->ac_bitmap_page = e4b->bd_bitmap_page;
2036 get_page(ac->ac_bitmap_page);
2037 ac->ac_buddy_page = e4b->bd_buddy_page;
2038 get_page(ac->ac_buddy_page);
2039 /* store last allocated for subsequent stream allocation */
2040 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2041 spin_lock(&sbi->s_md_lock);
2042 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2043 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2044 spin_unlock(&sbi->s_md_lock);
2047 * As we've just preallocated more space than
2048 * user requested originally, we store allocated
2049 * space in a special descriptor.
2051 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2052 ext4_mb_new_preallocation(ac);
2056 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2057 struct ext4_buddy *e4b,
2060 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2061 struct ext4_free_extent *bex = &ac->ac_b_ex;
2062 struct ext4_free_extent *gex = &ac->ac_g_ex;
2063 struct ext4_free_extent ex;
2066 if (ac->ac_status == AC_STATUS_FOUND)
2069 * We don't want to scan for a whole year
2071 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2072 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2073 ac->ac_status = AC_STATUS_BREAK;
2078 * Haven't found good chunk so far, let's continue
2080 if (bex->fe_len < gex->fe_len)
2083 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2084 && bex->fe_group == e4b->bd_group) {
2085 /* recheck chunk's availability - we don't know
2086 * when it was found (within this lock-unlock
2088 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
2089 if (max >= gex->fe_len) {
2090 ext4_mb_use_best_found(ac, e4b);
2097 * The routine checks whether found extent is good enough. If it is,
2098 * then the extent gets marked used and flag is set to the context
2099 * to stop scanning. Otherwise, the extent is compared with the
2100 * previous found extent and if new one is better, then it's stored
2101 * in the context. Later, the best found extent will be used, if
2102 * mballoc can't find good enough extent.
2104 * FIXME: real allocation policy is to be designed yet!
2106 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2107 struct ext4_free_extent *ex,
2108 struct ext4_buddy *e4b)
2110 struct ext4_free_extent *bex = &ac->ac_b_ex;
2111 struct ext4_free_extent *gex = &ac->ac_g_ex;
2113 BUG_ON(ex->fe_len <= 0);
2114 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2115 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2116 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2121 * The special case - take what you catch first
2123 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2125 ext4_mb_use_best_found(ac, e4b);
2130 * Let's check whether the chuck is good enough
2132 if (ex->fe_len == gex->fe_len) {
2134 ext4_mb_use_best_found(ac, e4b);
2139 * If this is first found extent, just store it in the context
2141 if (bex->fe_len == 0) {
2147 * If new found extent is better, store it in the context
2149 if (bex->fe_len < gex->fe_len) {
2150 /* if the request isn't satisfied, any found extent
2151 * larger than previous best one is better */
2152 if (ex->fe_len > bex->fe_len)
2154 } else if (ex->fe_len > gex->fe_len) {
2155 /* if the request is satisfied, then we try to find
2156 * an extent that still satisfy the request, but is
2157 * smaller than previous one */
2158 if (ex->fe_len < bex->fe_len)
2162 ext4_mb_check_limits(ac, e4b, 0);
2165 static noinline_for_stack
2166 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2167 struct ext4_buddy *e4b)
2169 struct ext4_free_extent ex = ac->ac_b_ex;
2170 ext4_group_t group = ex.fe_group;
2174 BUG_ON(ex.fe_len <= 0);
2175 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2179 ext4_lock_group(ac->ac_sb, group);
2180 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2184 ext4_mb_use_best_found(ac, e4b);
2187 ext4_unlock_group(ac->ac_sb, group);
2188 ext4_mb_unload_buddy(e4b);
2193 static noinline_for_stack
2194 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2195 struct ext4_buddy *e4b)
2197 ext4_group_t group = ac->ac_g_ex.fe_group;
2200 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2201 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2202 struct ext4_free_extent ex;
2204 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
2206 if (grp->bb_free == 0)
2209 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2213 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
2214 ext4_mb_unload_buddy(e4b);
2218 ext4_lock_group(ac->ac_sb, group);
2219 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2220 ac->ac_g_ex.fe_len, &ex);
2221 ex.fe_logical = 0xDEADFA11; /* debug value */
2223 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
2226 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
2228 /* use do_div to get remainder (would be 64-bit modulo) */
2229 if (do_div(start, sbi->s_stripe) == 0) {
2232 ext4_mb_use_best_found(ac, e4b);
2234 } else if (max >= ac->ac_g_ex.fe_len) {
2235 BUG_ON(ex.fe_len <= 0);
2236 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2237 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2240 ext4_mb_use_best_found(ac, e4b);
2241 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2242 /* Sometimes, caller may want to merge even small
2243 * number of blocks to an existing extent */
2244 BUG_ON(ex.fe_len <= 0);
2245 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2246 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2249 ext4_mb_use_best_found(ac, e4b);
2251 ext4_unlock_group(ac->ac_sb, group);
2252 ext4_mb_unload_buddy(e4b);
2258 * The routine scans buddy structures (not bitmap!) from given order
2259 * to max order and tries to find big enough chunk to satisfy the req
2261 static noinline_for_stack
2262 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2263 struct ext4_buddy *e4b)
2265 struct super_block *sb = ac->ac_sb;
2266 struct ext4_group_info *grp = e4b->bd_info;
2272 BUG_ON(ac->ac_2order <= 0);
2273 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2274 if (grp->bb_counters[i] == 0)
2277 buddy = mb_find_buddy(e4b, i, &max);
2278 BUG_ON(buddy == NULL);
2280 k = mb_find_next_zero_bit(buddy, max, 0);
2282 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2283 "%d free clusters of order %d. But found 0",
2284 grp->bb_counters[i], i);
2285 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2287 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2292 ac->ac_b_ex.fe_len = 1 << i;
2293 ac->ac_b_ex.fe_start = k << i;
2294 ac->ac_b_ex.fe_group = e4b->bd_group;
2296 ext4_mb_use_best_found(ac, e4b);
2298 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2300 if (EXT4_SB(sb)->s_mb_stats)
2301 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2308 * The routine scans the group and measures all found extents.
2309 * In order to optimize scanning, caller must pass number of
2310 * free blocks in the group, so the routine can know upper limit.
2312 static noinline_for_stack
2313 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2314 struct ext4_buddy *e4b)
2316 struct super_block *sb = ac->ac_sb;
2317 void *bitmap = e4b->bd_bitmap;
2318 struct ext4_free_extent ex;
2322 free = e4b->bd_info->bb_free;
2323 if (WARN_ON(free <= 0))
2326 i = e4b->bd_info->bb_first_free;
2328 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2329 i = mb_find_next_zero_bit(bitmap,
2330 EXT4_CLUSTERS_PER_GROUP(sb), i);
2331 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2333 * IF we have corrupt bitmap, we won't find any
2334 * free blocks even though group info says we
2337 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2338 "%d free clusters as per "
2339 "group info. But bitmap says 0",
2341 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2342 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2346 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2347 if (WARN_ON(ex.fe_len <= 0))
2349 if (free < ex.fe_len) {
2350 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2351 "%d free clusters as per "
2352 "group info. But got %d blocks",
2354 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2355 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2357 * The number of free blocks differs. This mostly
2358 * indicate that the bitmap is corrupt. So exit
2359 * without claiming the space.
2363 ex.fe_logical = 0xDEADC0DE; /* debug value */
2364 ext4_mb_measure_extent(ac, &ex, e4b);
2370 ext4_mb_check_limits(ac, e4b, 1);
2374 * This is a special case for storages like raid5
2375 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2377 static noinline_for_stack
2378 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2379 struct ext4_buddy *e4b)
2381 struct super_block *sb = ac->ac_sb;
2382 struct ext4_sb_info *sbi = EXT4_SB(sb);
2383 void *bitmap = e4b->bd_bitmap;
2384 struct ext4_free_extent ex;
2385 ext4_fsblk_t first_group_block;
2390 BUG_ON(sbi->s_stripe == 0);
2392 /* find first stripe-aligned block in group */
2393 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2395 a = first_group_block + sbi->s_stripe - 1;
2396 do_div(a, sbi->s_stripe);
2397 i = (a * sbi->s_stripe) - first_group_block;
2399 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2400 if (!mb_test_bit(i, bitmap)) {
2401 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2402 if (max >= sbi->s_stripe) {
2404 ex.fe_logical = 0xDEADF00D; /* debug value */
2406 ext4_mb_use_best_found(ac, e4b);
2415 * This is also called BEFORE we load the buddy bitmap.
2416 * Returns either 1 or 0 indicating that the group is either suitable
2417 * for the allocation or not.
2419 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2420 ext4_group_t group, int cr)
2422 ext4_grpblk_t free, fragments;
2423 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2424 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2426 BUG_ON(cr < 0 || cr >= 4);
2428 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2431 free = grp->bb_free;
2435 fragments = grp->bb_fragments;
2441 BUG_ON(ac->ac_2order == 0);
2443 /* Avoid using the first bg of a flexgroup for data files */
2444 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2445 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2446 ((group % flex_size) == 0))
2449 if (free < ac->ac_g_ex.fe_len)
2452 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2455 if (grp->bb_largest_free_order < ac->ac_2order)
2460 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2464 if (free >= ac->ac_g_ex.fe_len)
2477 * This could return negative error code if something goes wrong
2478 * during ext4_mb_init_group(). This should not be called with
2479 * ext4_lock_group() held.
2481 * Note: because we are conditionally operating with the group lock in
2482 * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this
2483 * function using __acquire and __release. This means we need to be
2484 * super careful before messing with the error path handling via "goto
2487 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2488 ext4_group_t group, int cr)
2490 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2491 struct super_block *sb = ac->ac_sb;
2492 struct ext4_sb_info *sbi = EXT4_SB(sb);
2493 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2497 if (sbi->s_mb_stats)
2498 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2500 ext4_lock_group(sb, group);
2501 __release(ext4_group_lock_ptr(sb, group));
2503 free = grp->bb_free;
2506 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2508 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2511 __acquire(ext4_group_lock_ptr(sb, group));
2512 ext4_unlock_group(sb, group);
2515 /* We only do this if the grp has never been initialized */
2516 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2517 struct ext4_group_desc *gdp =
2518 ext4_get_group_desc(sb, group, NULL);
2521 /* cr=0/1 is a very optimistic search to find large
2522 * good chunks almost for free. If buddy data is not
2523 * ready, then this optimization makes no sense. But
2524 * we never skip the first block group in a flex_bg,
2525 * since this gets used for metadata block allocation,
2526 * and we want to make sure we locate metadata blocks
2527 * in the first block group in the flex_bg if possible.
2530 (!sbi->s_log_groups_per_flex ||
2531 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2532 !(ext4_has_group_desc_csum(sb) &&
2533 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2535 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2541 ext4_lock_group(sb, group);
2542 __release(ext4_group_lock_ptr(sb, group));
2544 ret = ext4_mb_good_group(ac, group, cr);
2547 __acquire(ext4_group_lock_ptr(sb, group));
2548 ext4_unlock_group(sb, group);
2554 * Start prefetching @nr block bitmaps starting at @group.
2555 * Return the next group which needs to be prefetched.
2557 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2558 unsigned int nr, int *cnt)
2560 ext4_group_t ngroups = ext4_get_groups_count(sb);
2561 struct buffer_head *bh;
2562 struct blk_plug plug;
2564 blk_start_plug(&plug);
2566 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2568 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2571 * Prefetch block groups with free blocks; but don't
2572 * bother if it is marked uninitialized on disk, since
2573 * it won't require I/O to read. Also only try to
2574 * prefetch once, so we avoid getblk() call, which can
2577 if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2578 EXT4_MB_GRP_NEED_INIT(grp) &&
2579 ext4_free_group_clusters(sb, gdp) > 0 &&
2580 !(ext4_has_group_desc_csum(sb) &&
2581 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2582 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2583 if (bh && !IS_ERR(bh)) {
2584 if (!buffer_uptodate(bh) && cnt)
2589 if (++group >= ngroups)
2592 blk_finish_plug(&plug);
2597 * Prefetching reads the block bitmap into the buffer cache; but we
2598 * need to make sure that the buddy bitmap in the page cache has been
2599 * initialized. Note that ext4_mb_init_group() will block if the I/O
2600 * is not yet completed, or indeed if it was not initiated by
2601 * ext4_mb_prefetch did not start the I/O.
2603 * TODO: We should actually kick off the buddy bitmap setup in a work
2604 * queue when the buffer I/O is completed, so that we don't block
2605 * waiting for the block allocation bitmap read to finish when
2606 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2608 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2612 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2614 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2617 group = ext4_get_groups_count(sb);
2619 grp = ext4_get_group_info(sb, group);
2621 if (EXT4_MB_GRP_NEED_INIT(grp) &&
2622 ext4_free_group_clusters(sb, gdp) > 0 &&
2623 !(ext4_has_group_desc_csum(sb) &&
2624 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2625 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2631 static noinline_for_stack int
2632 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2634 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2635 int cr = -1, new_cr;
2636 int err = 0, first_err = 0;
2637 unsigned int nr = 0, prefetch_ios = 0;
2638 struct ext4_sb_info *sbi;
2639 struct super_block *sb;
2640 struct ext4_buddy e4b;
2645 ngroups = ext4_get_groups_count(sb);
2646 /* non-extent files are limited to low blocks/groups */
2647 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2648 ngroups = sbi->s_blockfile_groups;
2650 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2652 /* first, try the goal */
2653 err = ext4_mb_find_by_goal(ac, &e4b);
2654 if (err || ac->ac_status == AC_STATUS_FOUND)
2657 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2661 * ac->ac_2order is set only if the fe_len is a power of 2
2662 * if ac->ac_2order is set we also set criteria to 0 so that we
2663 * try exact allocation using buddy.
2665 i = fls(ac->ac_g_ex.fe_len);
2668 * We search using buddy data only if the order of the request
2669 * is greater than equal to the sbi_s_mb_order2_reqs
2670 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2671 * We also support searching for power-of-two requests only for
2672 * requests upto maximum buddy size we have constructed.
2674 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2676 * This should tell if fe_len is exactly power of 2
2678 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2679 ac->ac_2order = array_index_nospec(i - 1,
2683 /* if stream allocation is enabled, use global goal */
2684 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2685 /* TBD: may be hot point */
2686 spin_lock(&sbi->s_md_lock);
2687 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2688 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2689 spin_unlock(&sbi->s_md_lock);
2692 /* Let's just scan groups to find more-less suitable blocks */
2693 cr = ac->ac_2order ? 0 : 1;
2695 * cr == 0 try to get exact allocation,
2696 * cr == 3 try to get anything
2699 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2700 ac->ac_criteria = cr;
2702 * searching for the right group start
2703 * from the goal value specified
2705 group = ac->ac_g_ex.fe_group;
2706 ac->ac_last_optimal_group = group;
2707 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2708 prefetch_grp = group;
2710 for (i = 0, new_cr = cr; i < ngroups; i++,
2711 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) {
2721 * Batch reads of the block allocation bitmaps
2722 * to get multiple READs in flight; limit
2723 * prefetching at cr=0/1, otherwise mballoc can
2724 * spend a lot of time loading imperfect groups
2726 if ((prefetch_grp == group) &&
2728 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2729 unsigned int curr_ios = prefetch_ios;
2731 nr = sbi->s_mb_prefetch;
2732 if (ext4_has_feature_flex_bg(sb)) {
2733 nr = 1 << sbi->s_log_groups_per_flex;
2734 nr -= group & (nr - 1);
2735 nr = min(nr, sbi->s_mb_prefetch);
2737 prefetch_grp = ext4_mb_prefetch(sb, group,
2739 if (prefetch_ios == curr_ios)
2743 /* This now checks without needing the buddy page */
2744 ret = ext4_mb_good_group_nolock(ac, group, cr);
2751 err = ext4_mb_load_buddy(sb, group, &e4b);
2755 ext4_lock_group(sb, group);
2758 * We need to check again after locking the
2761 ret = ext4_mb_good_group(ac, group, cr);
2763 ext4_unlock_group(sb, group);
2764 ext4_mb_unload_buddy(&e4b);
2768 ac->ac_groups_scanned++;
2770 ext4_mb_simple_scan_group(ac, &e4b);
2771 else if (cr == 1 && sbi->s_stripe &&
2772 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2773 ext4_mb_scan_aligned(ac, &e4b);
2775 ext4_mb_complex_scan_group(ac, &e4b);
2777 ext4_unlock_group(sb, group);
2778 ext4_mb_unload_buddy(&e4b);
2780 if (ac->ac_status != AC_STATUS_CONTINUE)
2783 /* Processed all groups and haven't found blocks */
2784 if (sbi->s_mb_stats && i == ngroups)
2785 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2788 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2789 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2791 * We've been searching too long. Let's try to allocate
2792 * the best chunk we've found so far
2794 ext4_mb_try_best_found(ac, &e4b);
2795 if (ac->ac_status != AC_STATUS_FOUND) {
2797 * Someone more lucky has already allocated it.
2798 * The only thing we can do is just take first
2801 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2802 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2803 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2804 ac->ac_b_ex.fe_len, lost);
2806 ac->ac_b_ex.fe_group = 0;
2807 ac->ac_b_ex.fe_start = 0;
2808 ac->ac_b_ex.fe_len = 0;
2809 ac->ac_status = AC_STATUS_CONTINUE;
2810 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2816 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2817 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2819 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2822 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2823 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2824 ac->ac_flags, cr, err);
2827 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2832 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2834 struct super_block *sb = pde_data(file_inode(seq->file));
2837 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2840 return (void *) ((unsigned long) group);
2843 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2845 struct super_block *sb = pde_data(file_inode(seq->file));
2849 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2852 return (void *) ((unsigned long) group);
2855 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2857 struct super_block *sb = pde_data(file_inode(seq->file));
2858 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2860 int err, buddy_loaded = 0;
2861 struct ext4_buddy e4b;
2862 struct ext4_group_info *grinfo;
2863 unsigned char blocksize_bits = min_t(unsigned char,
2864 sb->s_blocksize_bits,
2865 EXT4_MAX_BLOCK_LOG_SIZE);
2867 struct ext4_group_info info;
2868 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2873 seq_puts(seq, "#group: free frags first ["
2874 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2875 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2877 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2878 sizeof(struct ext4_group_info);
2880 grinfo = ext4_get_group_info(sb, group);
2881 /* Load the group info in memory only if not already loaded. */
2882 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2883 err = ext4_mb_load_buddy(sb, group, &e4b);
2885 seq_printf(seq, "#%-5u: I/O error\n", group);
2891 memcpy(&sg, ext4_get_group_info(sb, group), i);
2894 ext4_mb_unload_buddy(&e4b);
2896 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2897 sg.info.bb_fragments, sg.info.bb_first_free);
2898 for (i = 0; i <= 13; i++)
2899 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2900 sg.info.bb_counters[i] : 0);
2901 seq_puts(seq, " ]\n");
2906 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2910 const struct seq_operations ext4_mb_seq_groups_ops = {
2911 .start = ext4_mb_seq_groups_start,
2912 .next = ext4_mb_seq_groups_next,
2913 .stop = ext4_mb_seq_groups_stop,
2914 .show = ext4_mb_seq_groups_show,
2917 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
2919 struct super_block *sb = seq->private;
2920 struct ext4_sb_info *sbi = EXT4_SB(sb);
2922 seq_puts(seq, "mballoc:\n");
2923 if (!sbi->s_mb_stats) {
2924 seq_puts(seq, "\tmb stats collection turned off.\n");
2925 seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
2928 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
2929 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
2931 seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned));
2933 seq_puts(seq, "\tcr0_stats:\n");
2934 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0]));
2935 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2936 atomic64_read(&sbi->s_bal_cX_groups_considered[0]));
2937 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2938 atomic64_read(&sbi->s_bal_cX_failed[0]));
2939 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2940 atomic_read(&sbi->s_bal_cr0_bad_suggestions));
2942 seq_puts(seq, "\tcr1_stats:\n");
2943 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1]));
2944 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2945 atomic64_read(&sbi->s_bal_cX_groups_considered[1]));
2946 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2947 atomic64_read(&sbi->s_bal_cX_failed[1]));
2948 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2949 atomic_read(&sbi->s_bal_cr1_bad_suggestions));
2951 seq_puts(seq, "\tcr2_stats:\n");
2952 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2]));
2953 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2954 atomic64_read(&sbi->s_bal_cX_groups_considered[2]));
2955 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2956 atomic64_read(&sbi->s_bal_cX_failed[2]));
2958 seq_puts(seq, "\tcr3_stats:\n");
2959 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3]));
2960 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2961 atomic64_read(&sbi->s_bal_cX_groups_considered[3]));
2962 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2963 atomic64_read(&sbi->s_bal_cX_failed[3]));
2964 seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned));
2965 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
2966 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
2967 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
2968 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
2970 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
2971 atomic_read(&sbi->s_mb_buddies_generated),
2972 ext4_get_groups_count(sb));
2973 seq_printf(seq, "\tbuddies_time_used: %llu\n",
2974 atomic64_read(&sbi->s_mb_generation_time));
2975 seq_printf(seq, "\tpreallocated: %u\n",
2976 atomic_read(&sbi->s_mb_preallocated));
2977 seq_printf(seq, "\tdiscarded: %u\n",
2978 atomic_read(&sbi->s_mb_discarded));
2982 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
2983 __acquires(&EXT4_SB(sb)->s_mb_rb_lock)
2985 struct super_block *sb = pde_data(file_inode(seq->file));
2986 unsigned long position;
2988 read_lock(&EXT4_SB(sb)->s_mb_rb_lock);
2990 if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
2992 position = *pos + 1;
2993 return (void *) ((unsigned long) position);
2996 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
2998 struct super_block *sb = pde_data(file_inode(seq->file));
2999 unsigned long position;
3002 if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
3004 position = *pos + 1;
3005 return (void *) ((unsigned long) position);
3008 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
3010 struct super_block *sb = pde_data(file_inode(seq->file));
3011 struct ext4_sb_info *sbi = EXT4_SB(sb);
3012 unsigned long position = ((unsigned long) v);
3013 struct ext4_group_info *grp;
3015 unsigned int count, min, max;
3018 if (position >= MB_NUM_ORDERS(sb)) {
3019 seq_puts(seq, "fragment_size_tree:\n");
3020 n = rb_first(&sbi->s_mb_avg_fragment_size_root);
3022 seq_puts(seq, "\ttree_min: 0\n\ttree_max: 0\n\ttree_nodes: 0\n");
3025 grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3026 min = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3028 while (rb_next(n)) {
3032 grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3033 max = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3035 seq_printf(seq, "\ttree_min: %u\n\ttree_max: %u\n\ttree_nodes: %u\n",
3040 if (position == 0) {
3041 seq_printf(seq, "optimize_scan: %d\n",
3042 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
3043 seq_puts(seq, "max_free_order_lists:\n");
3046 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
3047 bb_largest_free_order_node)
3049 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3050 (unsigned int)position, count);
3055 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3056 __releases(&EXT4_SB(sb)->s_mb_rb_lock)
3058 struct super_block *sb = pde_data(file_inode(seq->file));
3060 read_unlock(&EXT4_SB(sb)->s_mb_rb_lock);
3063 const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3064 .start = ext4_mb_seq_structs_summary_start,
3065 .next = ext4_mb_seq_structs_summary_next,
3066 .stop = ext4_mb_seq_structs_summary_stop,
3067 .show = ext4_mb_seq_structs_summary_show,
3070 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3072 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3073 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3080 * Allocate the top-level s_group_info array for the specified number
3083 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3085 struct ext4_sb_info *sbi = EXT4_SB(sb);
3087 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3089 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3090 EXT4_DESC_PER_BLOCK_BITS(sb);
3091 if (size <= sbi->s_group_info_size)
3094 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3095 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3096 if (!new_groupinfo) {
3097 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3101 old_groupinfo = rcu_dereference(sbi->s_group_info);
3103 memcpy(new_groupinfo, old_groupinfo,
3104 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3106 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3107 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3109 ext4_kvfree_array_rcu(old_groupinfo);
3110 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3111 sbi->s_group_info_size);
3115 /* Create and initialize ext4_group_info data for the given group. */
3116 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3117 struct ext4_group_desc *desc)
3121 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3122 struct ext4_sb_info *sbi = EXT4_SB(sb);
3123 struct ext4_group_info **meta_group_info;
3124 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3127 * First check if this group is the first of a reserved block.
3128 * If it's true, we have to allocate a new table of pointers
3129 * to ext4_group_info structures
3131 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3132 metalen = sizeof(*meta_group_info) <<
3133 EXT4_DESC_PER_BLOCK_BITS(sb);
3134 meta_group_info = kmalloc(metalen, GFP_NOFS);
3135 if (meta_group_info == NULL) {
3136 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3137 "for a buddy group");
3138 goto exit_meta_group_info;
3141 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3145 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3146 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3148 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3149 if (meta_group_info[i] == NULL) {
3150 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3151 goto exit_group_info;
3153 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3154 &(meta_group_info[i]->bb_state));
3157 * initialize bb_free to be able to skip
3158 * empty groups without initialization
3160 if (ext4_has_group_desc_csum(sb) &&
3161 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3162 meta_group_info[i]->bb_free =
3163 ext4_free_clusters_after_init(sb, group, desc);
3165 meta_group_info[i]->bb_free =
3166 ext4_free_group_clusters(sb, desc);
3169 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3170 init_rwsem(&meta_group_info[i]->alloc_sem);
3171 meta_group_info[i]->bb_free_root = RB_ROOT;
3172 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3173 RB_CLEAR_NODE(&meta_group_info[i]->bb_avg_fragment_size_rb);
3174 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3175 meta_group_info[i]->bb_group = group;
3177 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3181 /* If a meta_group_info table has been allocated, release it now */
3182 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3183 struct ext4_group_info ***group_info;
3186 group_info = rcu_dereference(sbi->s_group_info);
3187 kfree(group_info[idx]);
3188 group_info[idx] = NULL;
3191 exit_meta_group_info:
3193 } /* ext4_mb_add_groupinfo */
3195 static int ext4_mb_init_backend(struct super_block *sb)
3197 ext4_group_t ngroups = ext4_get_groups_count(sb);
3199 struct ext4_sb_info *sbi = EXT4_SB(sb);
3201 struct ext4_group_desc *desc;
3202 struct ext4_group_info ***group_info;
3203 struct kmem_cache *cachep;
3205 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3209 sbi->s_buddy_cache = new_inode(sb);
3210 if (sbi->s_buddy_cache == NULL) {
3211 ext4_msg(sb, KERN_ERR, "can't get new inode");
3214 /* To avoid potentially colliding with an valid on-disk inode number,
3215 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3216 * not in the inode hash, so it should never be found by iget(), but
3217 * this will avoid confusion if it ever shows up during debugging. */
3218 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3219 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3220 for (i = 0; i < ngroups; i++) {
3222 desc = ext4_get_group_desc(sb, i, NULL);
3224 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3227 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3231 if (ext4_has_feature_flex_bg(sb)) {
3232 /* a single flex group is supposed to be read by a single IO.
3233 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3234 * unsigned integer, so the maximum shift is 32.
3236 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3237 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3240 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3241 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3242 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3244 sbi->s_mb_prefetch = 32;
3246 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3247 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3248 /* now many real IOs to prefetch within a single allocation at cr=0
3249 * given cr=0 is an CPU-related optimization we shouldn't try to
3250 * load too many groups, at some point we should start to use what
3251 * we've got in memory.
3252 * with an average random access time 5ms, it'd take a second to get
3253 * 200 groups (* N with flex_bg), so let's make this limit 4
3255 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3256 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3257 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3262 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3264 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
3265 i = sbi->s_group_info_size;
3267 group_info = rcu_dereference(sbi->s_group_info);
3269 kfree(group_info[i]);
3271 iput(sbi->s_buddy_cache);
3274 kvfree(rcu_dereference(sbi->s_group_info));
3279 static void ext4_groupinfo_destroy_slabs(void)
3283 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3284 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3285 ext4_groupinfo_caches[i] = NULL;
3289 static int ext4_groupinfo_create_slab(size_t size)
3291 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3293 int blocksize_bits = order_base_2(size);
3294 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3295 struct kmem_cache *cachep;
3297 if (cache_index >= NR_GRPINFO_CACHES)
3300 if (unlikely(cache_index < 0))
3303 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3304 if (ext4_groupinfo_caches[cache_index]) {
3305 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3306 return 0; /* Already created */
3309 slab_size = offsetof(struct ext4_group_info,
3310 bb_counters[blocksize_bits + 2]);
3312 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3313 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3316 ext4_groupinfo_caches[cache_index] = cachep;
3318 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3321 "EXT4-fs: no memory for groupinfo slab cache\n");
3328 static void ext4_discard_work(struct work_struct *work)
3330 struct ext4_sb_info *sbi = container_of(work,
3331 struct ext4_sb_info, s_discard_work);
3332 struct super_block *sb = sbi->s_sb;
3333 struct ext4_free_data *fd, *nfd;
3334 struct ext4_buddy e4b;
3335 struct list_head discard_list;
3336 ext4_group_t grp, load_grp;
3339 INIT_LIST_HEAD(&discard_list);
3340 spin_lock(&sbi->s_md_lock);
3341 list_splice_init(&sbi->s_discard_list, &discard_list);
3342 spin_unlock(&sbi->s_md_lock);
3344 load_grp = UINT_MAX;
3345 list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
3347 * If filesystem is umounting or no memory or suffering
3348 * from no space, give up the discard
3350 if ((sb->s_flags & SB_ACTIVE) && !err &&
3351 !atomic_read(&sbi->s_retry_alloc_pending)) {
3352 grp = fd->efd_group;
3353 if (grp != load_grp) {
3354 if (load_grp != UINT_MAX)
3355 ext4_mb_unload_buddy(&e4b);
3357 err = ext4_mb_load_buddy(sb, grp, &e4b);
3359 kmem_cache_free(ext4_free_data_cachep, fd);
3360 load_grp = UINT_MAX;
3367 ext4_lock_group(sb, grp);
3368 ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
3369 fd->efd_start_cluster + fd->efd_count - 1, 1);
3370 ext4_unlock_group(sb, grp);
3372 kmem_cache_free(ext4_free_data_cachep, fd);
3375 if (load_grp != UINT_MAX)
3376 ext4_mb_unload_buddy(&e4b);
3379 int ext4_mb_init(struct super_block *sb)
3381 struct ext4_sb_info *sbi = EXT4_SB(sb);
3383 unsigned offset, offset_incr;
3387 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3389 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3390 if (sbi->s_mb_offsets == NULL) {
3395 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3396 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3397 if (sbi->s_mb_maxs == NULL) {
3402 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3406 /* order 0 is regular bitmap */
3407 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3408 sbi->s_mb_offsets[0] = 0;
3412 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3413 max = sb->s_blocksize << 2;
3415 sbi->s_mb_offsets[i] = offset;
3416 sbi->s_mb_maxs[i] = max;
3417 offset += offset_incr;
3418 offset_incr = offset_incr >> 1;
3421 } while (i < MB_NUM_ORDERS(sb));
3423 sbi->s_mb_avg_fragment_size_root = RB_ROOT;
3424 sbi->s_mb_largest_free_orders =
3425 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3427 if (!sbi->s_mb_largest_free_orders) {
3431 sbi->s_mb_largest_free_orders_locks =
3432 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3434 if (!sbi->s_mb_largest_free_orders_locks) {
3438 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3439 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3440 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3442 rwlock_init(&sbi->s_mb_rb_lock);
3444 spin_lock_init(&sbi->s_md_lock);
3445 sbi->s_mb_free_pending = 0;
3446 INIT_LIST_HEAD(&sbi->s_freed_data_list);
3447 INIT_LIST_HEAD(&sbi->s_discard_list);
3448 INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
3449 atomic_set(&sbi->s_retry_alloc_pending, 0);
3451 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3452 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3453 sbi->s_mb_stats = MB_DEFAULT_STATS;
3454 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3455 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3456 sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
3458 * The default group preallocation is 512, which for 4k block
3459 * sizes translates to 2 megabytes. However for bigalloc file
3460 * systems, this is probably too big (i.e, if the cluster size
3461 * is 1 megabyte, then group preallocation size becomes half a
3462 * gigabyte!). As a default, we will keep a two megabyte
3463 * group pralloc size for cluster sizes up to 64k, and after
3464 * that, we will force a minimum group preallocation size of
3465 * 32 clusters. This translates to 8 megs when the cluster
3466 * size is 256k, and 32 megs when the cluster size is 1 meg,
3467 * which seems reasonable as a default.
3469 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3470 sbi->s_cluster_bits, 32);
3472 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3473 * to the lowest multiple of s_stripe which is bigger than
3474 * the s_mb_group_prealloc as determined above. We want
3475 * the preallocation size to be an exact multiple of the
3476 * RAID stripe size so that preallocations don't fragment
3479 if (sbi->s_stripe > 1) {
3480 sbi->s_mb_group_prealloc = roundup(
3481 sbi->s_mb_group_prealloc, sbi->s_stripe);
3484 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3485 if (sbi->s_locality_groups == NULL) {
3489 for_each_possible_cpu(i) {
3490 struct ext4_locality_group *lg;
3491 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3492 mutex_init(&lg->lg_mutex);
3493 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3494 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3495 spin_lock_init(&lg->lg_prealloc_lock);
3498 if (bdev_nonrot(sb->s_bdev))
3499 sbi->s_mb_max_linear_groups = 0;
3501 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3502 /* init file for buddy data */
3503 ret = ext4_mb_init_backend(sb);
3505 goto out_free_locality_groups;
3509 out_free_locality_groups:
3510 free_percpu(sbi->s_locality_groups);
3511 sbi->s_locality_groups = NULL;
3513 kfree(sbi->s_mb_largest_free_orders);
3514 kfree(sbi->s_mb_largest_free_orders_locks);
3515 kfree(sbi->s_mb_offsets);
3516 sbi->s_mb_offsets = NULL;
3517 kfree(sbi->s_mb_maxs);
3518 sbi->s_mb_maxs = NULL;
3522 /* need to called with the ext4 group lock held */
3523 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3525 struct ext4_prealloc_space *pa;
3526 struct list_head *cur, *tmp;
3529 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3530 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3531 list_del(&pa->pa_group_list);
3533 kmem_cache_free(ext4_pspace_cachep, pa);
3538 int ext4_mb_release(struct super_block *sb)
3540 ext4_group_t ngroups = ext4_get_groups_count(sb);
3542 int num_meta_group_infos;
3543 struct ext4_group_info *grinfo, ***group_info;
3544 struct ext4_sb_info *sbi = EXT4_SB(sb);
3545 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3548 if (test_opt(sb, DISCARD)) {
3550 * wait the discard work to drain all of ext4_free_data
3552 flush_work(&sbi->s_discard_work);
3553 WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
3556 if (sbi->s_group_info) {
3557 for (i = 0; i < ngroups; i++) {
3559 grinfo = ext4_get_group_info(sb, i);
3560 mb_group_bb_bitmap_free(grinfo);
3561 ext4_lock_group(sb, i);
3562 count = ext4_mb_cleanup_pa(grinfo);
3564 mb_debug(sb, "mballoc: %d PAs left\n",
3566 ext4_unlock_group(sb, i);
3567 kmem_cache_free(cachep, grinfo);
3569 num_meta_group_infos = (ngroups +
3570 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3571 EXT4_DESC_PER_BLOCK_BITS(sb);
3573 group_info = rcu_dereference(sbi->s_group_info);
3574 for (i = 0; i < num_meta_group_infos; i++)
3575 kfree(group_info[i]);
3579 kfree(sbi->s_mb_largest_free_orders);
3580 kfree(sbi->s_mb_largest_free_orders_locks);
3581 kfree(sbi->s_mb_offsets);
3582 kfree(sbi->s_mb_maxs);
3583 iput(sbi->s_buddy_cache);
3584 if (sbi->s_mb_stats) {
3585 ext4_msg(sb, KERN_INFO,
3586 "mballoc: %u blocks %u reqs (%u success)",
3587 atomic_read(&sbi->s_bal_allocated),
3588 atomic_read(&sbi->s_bal_reqs),
3589 atomic_read(&sbi->s_bal_success));
3590 ext4_msg(sb, KERN_INFO,
3591 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3592 "%u 2^N hits, %u breaks, %u lost",
3593 atomic_read(&sbi->s_bal_ex_scanned),
3594 atomic_read(&sbi->s_bal_groups_scanned),
3595 atomic_read(&sbi->s_bal_goals),
3596 atomic_read(&sbi->s_bal_2orders),
3597 atomic_read(&sbi->s_bal_breaks),
3598 atomic_read(&sbi->s_mb_lost_chunks));
3599 ext4_msg(sb, KERN_INFO,
3600 "mballoc: %u generated and it took %llu",
3601 atomic_read(&sbi->s_mb_buddies_generated),
3602 atomic64_read(&sbi->s_mb_generation_time));
3603 ext4_msg(sb, KERN_INFO,
3604 "mballoc: %u preallocated, %u discarded",
3605 atomic_read(&sbi->s_mb_preallocated),
3606 atomic_read(&sbi->s_mb_discarded));
3609 free_percpu(sbi->s_locality_groups);
3614 static inline int ext4_issue_discard(struct super_block *sb,
3615 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3618 ext4_fsblk_t discard_block;
3620 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3621 ext4_group_first_block_no(sb, block_group));
3622 count = EXT4_C2B(EXT4_SB(sb), count);
3623 trace_ext4_discard_blocks(sb,
3624 (unsigned long long) discard_block, count);
3626 return __blkdev_issue_discard(sb->s_bdev,
3627 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3628 (sector_t)count << (sb->s_blocksize_bits - 9),
3631 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3634 static void ext4_free_data_in_buddy(struct super_block *sb,
3635 struct ext4_free_data *entry)
3637 struct ext4_buddy e4b;
3638 struct ext4_group_info *db;
3639 int err, count = 0, count2 = 0;
3641 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3642 entry->efd_count, entry->efd_group, entry);
3644 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3645 /* we expect to find existing buddy because it's pinned */
3648 spin_lock(&EXT4_SB(sb)->s_md_lock);
3649 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3650 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3653 /* there are blocks to put in buddy to make them really free */
3654 count += entry->efd_count;
3656 ext4_lock_group(sb, entry->efd_group);
3657 /* Take it out of per group rb tree */
3658 rb_erase(&entry->efd_node, &(db->bb_free_root));
3659 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3662 * Clear the trimmed flag for the group so that the next
3663 * ext4_trim_fs can trim it.
3664 * If the volume is mounted with -o discard, online discard
3665 * is supported and the free blocks will be trimmed online.
3667 if (!test_opt(sb, DISCARD))
3668 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3670 if (!db->bb_free_root.rb_node) {
3671 /* No more items in the per group rb tree
3672 * balance refcounts from ext4_mb_free_metadata()
3674 put_page(e4b.bd_buddy_page);
3675 put_page(e4b.bd_bitmap_page);
3677 ext4_unlock_group(sb, entry->efd_group);
3678 ext4_mb_unload_buddy(&e4b);
3680 mb_debug(sb, "freed %d blocks in %d structures\n", count,
3685 * This function is called by the jbd2 layer once the commit has finished,
3686 * so we know we can free the blocks that were released with that commit.
3688 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3690 struct ext4_sb_info *sbi = EXT4_SB(sb);
3691 struct ext4_free_data *entry, *tmp;
3692 struct list_head freed_data_list;
3693 struct list_head *cut_pos = NULL;
3696 INIT_LIST_HEAD(&freed_data_list);
3698 spin_lock(&sbi->s_md_lock);
3699 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3700 if (entry->efd_tid != commit_tid)
3702 cut_pos = &entry->efd_list;
3705 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3707 spin_unlock(&sbi->s_md_lock);
3709 list_for_each_entry(entry, &freed_data_list, efd_list)
3710 ext4_free_data_in_buddy(sb, entry);
3712 if (test_opt(sb, DISCARD)) {
3713 spin_lock(&sbi->s_md_lock);
3714 wake = list_empty(&sbi->s_discard_list);
3715 list_splice_tail(&freed_data_list, &sbi->s_discard_list);
3716 spin_unlock(&sbi->s_md_lock);
3718 queue_work(system_unbound_wq, &sbi->s_discard_work);
3720 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3721 kmem_cache_free(ext4_free_data_cachep, entry);
3725 int __init ext4_init_mballoc(void)
3727 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3728 SLAB_RECLAIM_ACCOUNT);
3729 if (ext4_pspace_cachep == NULL)
3732 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3733 SLAB_RECLAIM_ACCOUNT);
3734 if (ext4_ac_cachep == NULL)
3737 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3738 SLAB_RECLAIM_ACCOUNT);
3739 if (ext4_free_data_cachep == NULL)
3745 kmem_cache_destroy(ext4_ac_cachep);
3747 kmem_cache_destroy(ext4_pspace_cachep);
3752 void ext4_exit_mballoc(void)
3755 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3756 * before destroying the slab cache.
3759 kmem_cache_destroy(ext4_pspace_cachep);
3760 kmem_cache_destroy(ext4_ac_cachep);
3761 kmem_cache_destroy(ext4_free_data_cachep);
3762 ext4_groupinfo_destroy_slabs();
3767 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3768 * Returns 0 if success or error code
3770 static noinline_for_stack int
3771 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3772 handle_t *handle, unsigned int reserv_clstrs)
3774 struct buffer_head *bitmap_bh = NULL;
3775 struct ext4_group_desc *gdp;
3776 struct buffer_head *gdp_bh;
3777 struct ext4_sb_info *sbi;
3778 struct super_block *sb;
3782 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3783 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3788 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3789 if (IS_ERR(bitmap_bh)) {
3790 err = PTR_ERR(bitmap_bh);
3795 BUFFER_TRACE(bitmap_bh, "getting write access");
3796 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
3802 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3806 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3807 ext4_free_group_clusters(sb, gdp));
3809 BUFFER_TRACE(gdp_bh, "get_write_access");
3810 err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE);
3814 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3816 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3817 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3818 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3819 "fs metadata", block, block+len);
3820 /* File system mounted not to panic on error
3821 * Fix the bitmap and return EFSCORRUPTED
3822 * We leak some of the blocks here.
3824 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3825 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3826 ac->ac_b_ex.fe_len);
3827 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3828 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3830 err = -EFSCORRUPTED;
3834 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3835 #ifdef AGGRESSIVE_CHECK
3838 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3839 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3840 bitmap_bh->b_data));
3844 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3845 ac->ac_b_ex.fe_len);
3846 if (ext4_has_group_desc_csum(sb) &&
3847 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3848 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3849 ext4_free_group_clusters_set(sb, gdp,
3850 ext4_free_clusters_after_init(sb,
3851 ac->ac_b_ex.fe_group, gdp));
3853 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3854 ext4_free_group_clusters_set(sb, gdp, len);
3855 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3856 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3858 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3859 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3861 * Now reduce the dirty block count also. Should not go negative
3863 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3864 /* release all the reserved blocks if non delalloc */
3865 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3868 if (sbi->s_log_groups_per_flex) {
3869 ext4_group_t flex_group = ext4_flex_group(sbi,
3870 ac->ac_b_ex.fe_group);
3871 atomic64_sub(ac->ac_b_ex.fe_len,
3872 &sbi_array_rcu_deref(sbi, s_flex_groups,
3873 flex_group)->free_clusters);
3876 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3879 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3887 * Idempotent helper for Ext4 fast commit replay path to set the state of
3888 * blocks in bitmaps and update counters.
3890 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
3893 struct buffer_head *bitmap_bh = NULL;
3894 struct ext4_group_desc *gdp;
3895 struct buffer_head *gdp_bh;
3896 struct ext4_sb_info *sbi = EXT4_SB(sb);
3898 ext4_grpblk_t blkoff;
3901 unsigned int clen, clen_changed, thisgrp_len;
3904 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
3907 * Check to see if we are freeing blocks across a group
3909 * In case of flex_bg, this can happen that (block, len) may
3910 * span across more than one group. In that case we need to
3911 * get the corresponding group metadata to work with.
3912 * For this we have goto again loop.
3914 thisgrp_len = min_t(unsigned int, (unsigned int)len,
3915 EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
3916 clen = EXT4_NUM_B2C(sbi, thisgrp_len);
3918 if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
3919 ext4_error(sb, "Marking blocks in system zone - "
3920 "Block = %llu, len = %u",
3921 block, thisgrp_len);
3926 bitmap_bh = ext4_read_block_bitmap(sb, group);
3927 if (IS_ERR(bitmap_bh)) {
3928 err = PTR_ERR(bitmap_bh);
3934 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
3938 ext4_lock_group(sb, group);
3940 for (i = 0; i < clen; i++)
3941 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
3945 clen_changed = clen - already;
3947 mb_set_bits(bitmap_bh->b_data, blkoff, clen);
3949 mb_clear_bits(bitmap_bh->b_data, blkoff, clen);
3950 if (ext4_has_group_desc_csum(sb) &&
3951 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3952 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3953 ext4_free_group_clusters_set(sb, gdp,
3954 ext4_free_clusters_after_init(sb, group, gdp));
3957 clen = ext4_free_group_clusters(sb, gdp) - clen_changed;
3959 clen = ext4_free_group_clusters(sb, gdp) + clen_changed;
3961 ext4_free_group_clusters_set(sb, gdp, clen);
3962 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
3963 ext4_group_desc_csum_set(sb, group, gdp);
3965 ext4_unlock_group(sb, group);
3967 if (sbi->s_log_groups_per_flex) {
3968 ext4_group_t flex_group = ext4_flex_group(sbi, group);
3969 struct flex_groups *fg = sbi_array_rcu_deref(sbi,
3970 s_flex_groups, flex_group);
3973 atomic64_sub(clen_changed, &fg->free_clusters);
3975 atomic64_add(clen_changed, &fg->free_clusters);
3979 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
3982 sync_dirty_buffer(bitmap_bh);
3983 err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
3984 sync_dirty_buffer(gdp_bh);
3988 block += thisgrp_len;
3999 * here we normalize request for locality group
4000 * Group request are normalized to s_mb_group_prealloc, which goes to
4001 * s_strip if we set the same via mount option.
4002 * s_mb_group_prealloc can be configured via
4003 * /sys/fs/ext4/<partition>/mb_group_prealloc
4005 * XXX: should we try to preallocate more than the group has now?
4007 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
4009 struct super_block *sb = ac->ac_sb;
4010 struct ext4_locality_group *lg = ac->ac_lg;
4013 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
4014 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
4018 * Normalization means making request better in terms of
4019 * size and alignment
4021 static noinline_for_stack void
4022 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
4023 struct ext4_allocation_request *ar)
4025 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4028 loff_t size, start_off;
4029 loff_t orig_size __maybe_unused;
4031 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4032 struct ext4_prealloc_space *pa;
4034 /* do normalize only data requests, metadata requests
4035 do not need preallocation */
4036 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4039 /* sometime caller may want exact blocks */
4040 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4043 /* caller may indicate that preallocation isn't
4044 * required (it's a tail, for example) */
4045 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
4048 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
4049 ext4_mb_normalize_group_request(ac);
4053 bsbits = ac->ac_sb->s_blocksize_bits;
4055 /* first, let's learn actual file size
4056 * given current request is allocated */
4057 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4058 size = size << bsbits;
4059 if (size < i_size_read(ac->ac_inode))
4060 size = i_size_read(ac->ac_inode);
4063 /* max size of free chunks */
4066 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
4067 (req <= (size) || max <= (chunk_size))
4069 /* first, try to predict filesize */
4070 /* XXX: should this table be tunable? */
4072 if (size <= 16 * 1024) {
4074 } else if (size <= 32 * 1024) {
4076 } else if (size <= 64 * 1024) {
4078 } else if (size <= 128 * 1024) {
4080 } else if (size <= 256 * 1024) {
4082 } else if (size <= 512 * 1024) {
4084 } else if (size <= 1024 * 1024) {
4086 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
4087 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4088 (21 - bsbits)) << 21;
4089 size = 2 * 1024 * 1024;
4090 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
4091 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4092 (22 - bsbits)) << 22;
4093 size = 4 * 1024 * 1024;
4094 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
4095 (8<<20)>>bsbits, max, 8 * 1024)) {
4096 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4097 (23 - bsbits)) << 23;
4098 size = 8 * 1024 * 1024;
4100 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
4101 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
4102 ac->ac_o_ex.fe_len) << bsbits;
4104 size = size >> bsbits;
4105 start = start_off >> bsbits;
4108 * For tiny groups (smaller than 8MB) the chosen allocation
4109 * alignment may be larger than group size. Make sure the
4110 * alignment does not move allocation to a different group which
4111 * makes mballoc fail assertions later.
4113 start = max(start, rounddown(ac->ac_o_ex.fe_logical,
4114 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
4116 /* don't cover already allocated blocks in selected range */
4117 if (ar->pleft && start <= ar->lleft) {
4118 size -= ar->lleft + 1 - start;
4119 start = ar->lleft + 1;
4121 if (ar->pright && start + size - 1 >= ar->lright)
4122 size -= start + size - ar->lright;
4125 * Trim allocation request for filesystems with artificially small
4128 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4129 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4133 /* check we don't cross already preallocated blocks */
4135 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4140 spin_lock(&pa->pa_lock);
4141 if (pa->pa_deleted) {
4142 spin_unlock(&pa->pa_lock);
4146 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4149 /* PA must not overlap original request */
4150 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
4151 ac->ac_o_ex.fe_logical < pa->pa_lstart));
4153 /* skip PAs this normalized request doesn't overlap with */
4154 if (pa->pa_lstart >= end || pa_end <= start) {
4155 spin_unlock(&pa->pa_lock);
4158 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
4160 /* adjust start or end to be adjacent to this pa */
4161 if (pa_end <= ac->ac_o_ex.fe_logical) {
4162 BUG_ON(pa_end < start);
4164 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4165 BUG_ON(pa->pa_lstart > end);
4166 end = pa->pa_lstart;
4168 spin_unlock(&pa->pa_lock);
4173 /* XXX: extra loop to check we really don't overlap preallocations */
4175 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4178 spin_lock(&pa->pa_lock);
4179 if (pa->pa_deleted == 0) {
4180 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4182 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
4184 spin_unlock(&pa->pa_lock);
4189 * In this function "start" and "size" are normalized for better
4190 * alignment and length such that we could preallocate more blocks.
4191 * This normalization is done such that original request of
4192 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
4193 * "size" boundaries.
4194 * (Note fe_len can be relaxed since FS block allocation API does not
4195 * provide gurantee on number of contiguous blocks allocation since that
4196 * depends upon free space left, etc).
4197 * In case of inode pa, later we use the allocated blocks
4198 * [pa_start + fe_logical - pa_lstart, fe_len/size] from the preallocated
4199 * range of goal/best blocks [start, size] to put it at the
4200 * ac_o_ex.fe_logical extent of this inode.
4201 * (See ext4_mb_use_inode_pa() for more details)
4203 if (start + size <= ac->ac_o_ex.fe_logical ||
4204 start > ac->ac_o_ex.fe_logical) {
4205 ext4_msg(ac->ac_sb, KERN_ERR,
4206 "start %lu, size %lu, fe_logical %lu",
4207 (unsigned long) start, (unsigned long) size,
4208 (unsigned long) ac->ac_o_ex.fe_logical);
4211 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4213 /* now prepare goal request */
4215 /* XXX: is it better to align blocks WRT to logical
4216 * placement or satisfy big request as is */
4217 ac->ac_g_ex.fe_logical = start;
4218 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4220 /* define goal start in order to merge */
4221 if (ar->pright && (ar->lright == (start + size))) {
4222 /* merge to the right */
4223 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4224 &ac->ac_f_ex.fe_group,
4225 &ac->ac_f_ex.fe_start);
4226 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4228 if (ar->pleft && (ar->lleft + 1 == start)) {
4229 /* merge to the left */
4230 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4231 &ac->ac_f_ex.fe_group,
4232 &ac->ac_f_ex.fe_start);
4233 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4236 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4240 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4242 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4244 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4245 atomic_inc(&sbi->s_bal_reqs);
4246 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4247 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4248 atomic_inc(&sbi->s_bal_success);
4249 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4250 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4251 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4252 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4253 atomic_inc(&sbi->s_bal_goals);
4254 if (ac->ac_found > sbi->s_mb_max_to_scan)
4255 atomic_inc(&sbi->s_bal_breaks);
4258 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4259 trace_ext4_mballoc_alloc(ac);
4261 trace_ext4_mballoc_prealloc(ac);
4265 * Called on failure; free up any blocks from the inode PA for this
4266 * context. We don't need this for MB_GROUP_PA because we only change
4267 * pa_free in ext4_mb_release_context(), but on failure, we've already
4268 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4270 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4272 struct ext4_prealloc_space *pa = ac->ac_pa;
4273 struct ext4_buddy e4b;
4277 if (ac->ac_f_ex.fe_len == 0)
4279 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4282 * This should never happen since we pin the
4283 * pages in the ext4_allocation_context so
4284 * ext4_mb_load_buddy() should never fail.
4286 WARN(1, "mb_load_buddy failed (%d)", err);
4289 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4290 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4291 ac->ac_f_ex.fe_len);
4292 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4293 ext4_mb_unload_buddy(&e4b);
4296 if (pa->pa_type == MB_INODE_PA)
4297 pa->pa_free += ac->ac_b_ex.fe_len;
4301 * use blocks preallocated to inode
4303 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4304 struct ext4_prealloc_space *pa)
4306 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4311 /* found preallocated blocks, use them */
4312 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4313 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4314 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4315 len = EXT4_NUM_B2C(sbi, end - start);
4316 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4317 &ac->ac_b_ex.fe_start);
4318 ac->ac_b_ex.fe_len = len;
4319 ac->ac_status = AC_STATUS_FOUND;
4322 BUG_ON(start < pa->pa_pstart);
4323 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4324 BUG_ON(pa->pa_free < len);
4327 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4331 * use blocks preallocated to locality group
4333 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4334 struct ext4_prealloc_space *pa)
4336 unsigned int len = ac->ac_o_ex.fe_len;
4338 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4339 &ac->ac_b_ex.fe_group,
4340 &ac->ac_b_ex.fe_start);
4341 ac->ac_b_ex.fe_len = len;
4342 ac->ac_status = AC_STATUS_FOUND;
4345 /* we don't correct pa_pstart or pa_plen here to avoid
4346 * possible race when the group is being loaded concurrently
4347 * instead we correct pa later, after blocks are marked
4348 * in on-disk bitmap -- see ext4_mb_release_context()
4349 * Other CPUs are prevented from allocating from this pa by lg_mutex
4351 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4352 pa->pa_lstart-len, len, pa);
4356 * Return the prealloc space that have minimal distance
4357 * from the goal block. @cpa is the prealloc
4358 * space that is having currently known minimal distance
4359 * from the goal block.
4361 static struct ext4_prealloc_space *
4362 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4363 struct ext4_prealloc_space *pa,
4364 struct ext4_prealloc_space *cpa)
4366 ext4_fsblk_t cur_distance, new_distance;
4369 atomic_inc(&pa->pa_count);
4372 cur_distance = abs(goal_block - cpa->pa_pstart);
4373 new_distance = abs(goal_block - pa->pa_pstart);
4375 if (cur_distance <= new_distance)
4378 /* drop the previous reference */
4379 atomic_dec(&cpa->pa_count);
4380 atomic_inc(&pa->pa_count);
4385 * search goal blocks in preallocated space
4387 static noinline_for_stack bool
4388 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4390 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4392 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4393 struct ext4_locality_group *lg;
4394 struct ext4_prealloc_space *pa, *cpa = NULL;
4395 ext4_fsblk_t goal_block;
4397 /* only data can be preallocated */
4398 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4401 /* first, try per-file preallocation */
4403 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4405 /* all fields in this condition don't change,
4406 * so we can skip locking for them */
4407 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
4408 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
4409 EXT4_C2B(sbi, pa->pa_len)))
4412 /* non-extent files can't have physical blocks past 2^32 */
4413 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4414 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
4415 EXT4_MAX_BLOCK_FILE_PHYS))
4418 /* found preallocated blocks, use them */
4419 spin_lock(&pa->pa_lock);
4420 if (pa->pa_deleted == 0 && pa->pa_free) {
4421 atomic_inc(&pa->pa_count);
4422 ext4_mb_use_inode_pa(ac, pa);
4423 spin_unlock(&pa->pa_lock);
4424 ac->ac_criteria = 10;
4428 spin_unlock(&pa->pa_lock);
4432 /* can we use group allocation? */
4433 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4436 /* inode may have no locality group for some reason */
4440 order = fls(ac->ac_o_ex.fe_len) - 1;
4441 if (order > PREALLOC_TB_SIZE - 1)
4442 /* The max size of hash table is PREALLOC_TB_SIZE */
4443 order = PREALLOC_TB_SIZE - 1;
4445 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4447 * search for the prealloc space that is having
4448 * minimal distance from the goal block.
4450 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4452 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
4454 spin_lock(&pa->pa_lock);
4455 if (pa->pa_deleted == 0 &&
4456 pa->pa_free >= ac->ac_o_ex.fe_len) {
4458 cpa = ext4_mb_check_group_pa(goal_block,
4461 spin_unlock(&pa->pa_lock);
4466 ext4_mb_use_group_pa(ac, cpa);
4467 ac->ac_criteria = 20;
4474 * the function goes through all block freed in the group
4475 * but not yet committed and marks them used in in-core bitmap.
4476 * buddy must be generated from this bitmap
4477 * Need to be called with the ext4 group lock held
4479 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
4483 struct ext4_group_info *grp;
4484 struct ext4_free_data *entry;
4486 grp = ext4_get_group_info(sb, group);
4487 n = rb_first(&(grp->bb_free_root));
4490 entry = rb_entry(n, struct ext4_free_data, efd_node);
4491 mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
4498 * the function goes through all preallocation in this group and marks them
4499 * used in in-core bitmap. buddy must be generated from this bitmap
4500 * Need to be called with ext4 group lock held
4502 static noinline_for_stack
4503 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4506 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4507 struct ext4_prealloc_space *pa;
4508 struct list_head *cur;
4509 ext4_group_t groupnr;
4510 ext4_grpblk_t start;
4511 int preallocated = 0;
4514 /* all form of preallocation discards first load group,
4515 * so the only competing code is preallocation use.
4516 * we don't need any locking here
4517 * notice we do NOT ignore preallocations with pa_deleted
4518 * otherwise we could leave used blocks available for
4519 * allocation in buddy when concurrent ext4_mb_put_pa()
4520 * is dropping preallocation
4522 list_for_each(cur, &grp->bb_prealloc_list) {
4523 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
4524 spin_lock(&pa->pa_lock);
4525 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4528 spin_unlock(&pa->pa_lock);
4529 if (unlikely(len == 0))
4531 BUG_ON(groupnr != group);
4532 mb_set_bits(bitmap, start, len);
4533 preallocated += len;
4535 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
4538 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
4539 struct ext4_prealloc_space *pa)
4541 struct ext4_inode_info *ei;
4543 if (pa->pa_deleted) {
4544 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
4545 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
4552 if (pa->pa_type == MB_INODE_PA) {
4553 ei = EXT4_I(pa->pa_inode);
4554 atomic_dec(&ei->i_prealloc_active);
4558 static void ext4_mb_pa_callback(struct rcu_head *head)
4560 struct ext4_prealloc_space *pa;
4561 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
4563 BUG_ON(atomic_read(&pa->pa_count));
4564 BUG_ON(pa->pa_deleted == 0);
4565 kmem_cache_free(ext4_pspace_cachep, pa);
4569 * drops a reference to preallocated space descriptor
4570 * if this was the last reference and the space is consumed
4572 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
4573 struct super_block *sb, struct ext4_prealloc_space *pa)
4576 ext4_fsblk_t grp_blk;
4578 /* in this short window concurrent discard can set pa_deleted */
4579 spin_lock(&pa->pa_lock);
4580 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
4581 spin_unlock(&pa->pa_lock);
4585 if (pa->pa_deleted == 1) {
4586 spin_unlock(&pa->pa_lock);
4590 ext4_mb_mark_pa_deleted(sb, pa);
4591 spin_unlock(&pa->pa_lock);
4593 grp_blk = pa->pa_pstart;
4595 * If doing group-based preallocation, pa_pstart may be in the
4596 * next group when pa is used up
4598 if (pa->pa_type == MB_GROUP_PA)
4601 grp = ext4_get_group_number(sb, grp_blk);
4606 * P1 (buddy init) P2 (regular allocation)
4607 * find block B in PA
4608 * copy on-disk bitmap to buddy
4609 * mark B in on-disk bitmap
4610 * drop PA from group
4611 * mark all PAs in buddy
4613 * thus, P1 initializes buddy with B available. to prevent this
4614 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
4617 ext4_lock_group(sb, grp);
4618 list_del(&pa->pa_group_list);
4619 ext4_unlock_group(sb, grp);
4621 spin_lock(pa->pa_obj_lock);
4622 list_del_rcu(&pa->pa_inode_list);
4623 spin_unlock(pa->pa_obj_lock);
4625 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4629 * creates new preallocated space for given inode
4631 static noinline_for_stack void
4632 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
4634 struct super_block *sb = ac->ac_sb;
4635 struct ext4_sb_info *sbi = EXT4_SB(sb);
4636 struct ext4_prealloc_space *pa;
4637 struct ext4_group_info *grp;
4638 struct ext4_inode_info *ei;
4640 /* preallocate only when found space is larger then requested */
4641 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4642 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4643 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4644 BUG_ON(ac->ac_pa == NULL);
4648 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
4654 /* we can't allocate as much as normalizer wants.
4655 * so, found space must get proper lstart
4656 * to cover original request */
4657 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
4658 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
4660 /* we're limited by original request in that
4661 * logical block must be covered any way
4662 * winl is window we can move our chunk within */
4663 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
4665 /* also, we should cover whole original request */
4666 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
4668 /* the smallest one defines real window */
4669 win = min(winl, wins);
4671 offs = ac->ac_o_ex.fe_logical %
4672 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4673 if (offs && offs < win)
4676 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
4677 EXT4_NUM_B2C(sbi, win);
4678 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
4679 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
4682 /* preallocation can change ac_b_ex, thus we store actually
4683 * allocated blocks for history */
4684 ac->ac_f_ex = ac->ac_b_ex;
4686 pa->pa_lstart = ac->ac_b_ex.fe_logical;
4687 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4688 pa->pa_len = ac->ac_b_ex.fe_len;
4689 pa->pa_free = pa->pa_len;
4690 spin_lock_init(&pa->pa_lock);
4691 INIT_LIST_HEAD(&pa->pa_inode_list);
4692 INIT_LIST_HEAD(&pa->pa_group_list);
4694 pa->pa_type = MB_INODE_PA;
4696 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4697 pa->pa_len, pa->pa_lstart);
4698 trace_ext4_mb_new_inode_pa(ac, pa);
4700 ext4_mb_use_inode_pa(ac, pa);
4701 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
4703 ei = EXT4_I(ac->ac_inode);
4704 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4706 pa->pa_obj_lock = &ei->i_prealloc_lock;
4707 pa->pa_inode = ac->ac_inode;
4709 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4711 spin_lock(pa->pa_obj_lock);
4712 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
4713 spin_unlock(pa->pa_obj_lock);
4714 atomic_inc(&ei->i_prealloc_active);
4718 * creates new preallocated space for locality group inodes belongs to
4720 static noinline_for_stack void
4721 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
4723 struct super_block *sb = ac->ac_sb;
4724 struct ext4_locality_group *lg;
4725 struct ext4_prealloc_space *pa;
4726 struct ext4_group_info *grp;
4728 /* preallocate only when found space is larger then requested */
4729 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4730 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4731 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4732 BUG_ON(ac->ac_pa == NULL);
4736 /* preallocation can change ac_b_ex, thus we store actually
4737 * allocated blocks for history */
4738 ac->ac_f_ex = ac->ac_b_ex;
4740 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4741 pa->pa_lstart = pa->pa_pstart;
4742 pa->pa_len = ac->ac_b_ex.fe_len;
4743 pa->pa_free = pa->pa_len;
4744 spin_lock_init(&pa->pa_lock);
4745 INIT_LIST_HEAD(&pa->pa_inode_list);
4746 INIT_LIST_HEAD(&pa->pa_group_list);
4748 pa->pa_type = MB_GROUP_PA;
4750 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4751 pa->pa_len, pa->pa_lstart);
4752 trace_ext4_mb_new_group_pa(ac, pa);
4754 ext4_mb_use_group_pa(ac, pa);
4755 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4757 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4761 pa->pa_obj_lock = &lg->lg_prealloc_lock;
4762 pa->pa_inode = NULL;
4764 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4767 * We will later add the new pa to the right bucket
4768 * after updating the pa_free in ext4_mb_release_context
4772 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4774 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4775 ext4_mb_new_group_pa(ac);
4777 ext4_mb_new_inode_pa(ac);
4781 * finds all unused blocks in on-disk bitmap, frees them in
4782 * in-core bitmap and buddy.
4783 * @pa must be unlinked from inode and group lists, so that
4784 * nobody else can find/use it.
4785 * the caller MUST hold group/inode locks.
4786 * TODO: optimize the case when there are no in-core structures yet
4788 static noinline_for_stack int
4789 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4790 struct ext4_prealloc_space *pa)
4792 struct super_block *sb = e4b->bd_sb;
4793 struct ext4_sb_info *sbi = EXT4_SB(sb);
4798 unsigned long long grp_blk_start;
4801 BUG_ON(pa->pa_deleted == 0);
4802 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4803 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4804 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4805 end = bit + pa->pa_len;
4808 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4811 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4812 mb_debug(sb, "free preallocated %u/%u in group %u\n",
4813 (unsigned) ext4_group_first_block_no(sb, group) + bit,
4814 (unsigned) next - bit, (unsigned) group);
4817 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4818 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4819 EXT4_C2B(sbi, bit)),
4821 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4824 if (free != pa->pa_free) {
4825 ext4_msg(e4b->bd_sb, KERN_CRIT,
4826 "pa %p: logic %lu, phys. %lu, len %d",
4827 pa, (unsigned long) pa->pa_lstart,
4828 (unsigned long) pa->pa_pstart,
4830 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4833 * pa is already deleted so we use the value obtained
4834 * from the bitmap and continue.
4837 atomic_add(free, &sbi->s_mb_discarded);
4842 static noinline_for_stack int
4843 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4844 struct ext4_prealloc_space *pa)
4846 struct super_block *sb = e4b->bd_sb;
4850 trace_ext4_mb_release_group_pa(sb, pa);
4851 BUG_ON(pa->pa_deleted == 0);
4852 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4853 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4854 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4855 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4856 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4862 * releases all preallocations in given group
4864 * first, we need to decide discard policy:
4865 * - when do we discard
4867 * - how many do we discard
4868 * 1) how many requested
4870 static noinline_for_stack int
4871 ext4_mb_discard_group_preallocations(struct super_block *sb,
4872 ext4_group_t group, int *busy)
4874 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4875 struct buffer_head *bitmap_bh = NULL;
4876 struct ext4_prealloc_space *pa, *tmp;
4877 struct list_head list;
4878 struct ext4_buddy e4b;
4882 mb_debug(sb, "discard preallocation for group %u\n", group);
4883 if (list_empty(&grp->bb_prealloc_list))
4886 bitmap_bh = ext4_read_block_bitmap(sb, group);
4887 if (IS_ERR(bitmap_bh)) {
4888 err = PTR_ERR(bitmap_bh);
4889 ext4_error_err(sb, -err,
4890 "Error %d reading block bitmap for %u",
4895 err = ext4_mb_load_buddy(sb, group, &e4b);
4897 ext4_warning(sb, "Error %d loading buddy information for %u",
4903 INIT_LIST_HEAD(&list);
4904 ext4_lock_group(sb, group);
4905 list_for_each_entry_safe(pa, tmp,
4906 &grp->bb_prealloc_list, pa_group_list) {
4907 spin_lock(&pa->pa_lock);
4908 if (atomic_read(&pa->pa_count)) {
4909 spin_unlock(&pa->pa_lock);
4913 if (pa->pa_deleted) {
4914 spin_unlock(&pa->pa_lock);
4918 /* seems this one can be freed ... */
4919 ext4_mb_mark_pa_deleted(sb, pa);
4922 this_cpu_inc(discard_pa_seq);
4924 /* we can trust pa_free ... */
4925 free += pa->pa_free;
4927 spin_unlock(&pa->pa_lock);
4929 list_del(&pa->pa_group_list);
4930 list_add(&pa->u.pa_tmp_list, &list);
4933 /* now free all selected PAs */
4934 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4936 /* remove from object (inode or locality group) */
4937 spin_lock(pa->pa_obj_lock);
4938 list_del_rcu(&pa->pa_inode_list);
4939 spin_unlock(pa->pa_obj_lock);
4941 if (pa->pa_type == MB_GROUP_PA)
4942 ext4_mb_release_group_pa(&e4b, pa);
4944 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4946 list_del(&pa->u.pa_tmp_list);
4947 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4950 ext4_unlock_group(sb, group);
4951 ext4_mb_unload_buddy(&e4b);
4954 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4955 free, group, grp->bb_free);
4960 * releases all non-used preallocated blocks for given inode
4962 * It's important to discard preallocations under i_data_sem
4963 * We don't want another block to be served from the prealloc
4964 * space when we are discarding the inode prealloc space.
4966 * FIXME!! Make sure it is valid at all the call sites
4968 void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
4970 struct ext4_inode_info *ei = EXT4_I(inode);
4971 struct super_block *sb = inode->i_sb;
4972 struct buffer_head *bitmap_bh = NULL;
4973 struct ext4_prealloc_space *pa, *tmp;
4974 ext4_group_t group = 0;
4975 struct list_head list;
4976 struct ext4_buddy e4b;
4979 if (!S_ISREG(inode->i_mode)) {
4980 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4984 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
4987 mb_debug(sb, "discard preallocation for inode %lu\n",
4989 trace_ext4_discard_preallocations(inode,
4990 atomic_read(&ei->i_prealloc_active), needed);
4992 INIT_LIST_HEAD(&list);
4998 /* first, collect all pa's in the inode */
4999 spin_lock(&ei->i_prealloc_lock);
5000 while (!list_empty(&ei->i_prealloc_list) && needed) {
5001 pa = list_entry(ei->i_prealloc_list.prev,
5002 struct ext4_prealloc_space, pa_inode_list);
5003 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
5004 spin_lock(&pa->pa_lock);
5005 if (atomic_read(&pa->pa_count)) {
5006 /* this shouldn't happen often - nobody should
5007 * use preallocation while we're discarding it */
5008 spin_unlock(&pa->pa_lock);
5009 spin_unlock(&ei->i_prealloc_lock);
5010 ext4_msg(sb, KERN_ERR,
5011 "uh-oh! used pa while discarding");
5013 schedule_timeout_uninterruptible(HZ);
5017 if (pa->pa_deleted == 0) {
5018 ext4_mb_mark_pa_deleted(sb, pa);
5019 spin_unlock(&pa->pa_lock);
5020 list_del_rcu(&pa->pa_inode_list);
5021 list_add(&pa->u.pa_tmp_list, &list);
5026 /* someone is deleting pa right now */
5027 spin_unlock(&pa->pa_lock);
5028 spin_unlock(&ei->i_prealloc_lock);
5030 /* we have to wait here because pa_deleted
5031 * doesn't mean pa is already unlinked from
5032 * the list. as we might be called from
5033 * ->clear_inode() the inode will get freed
5034 * and concurrent thread which is unlinking
5035 * pa from inode's list may access already
5036 * freed memory, bad-bad-bad */
5038 /* XXX: if this happens too often, we can
5039 * add a flag to force wait only in case
5040 * of ->clear_inode(), but not in case of
5041 * regular truncate */
5042 schedule_timeout_uninterruptible(HZ);
5045 spin_unlock(&ei->i_prealloc_lock);
5047 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5048 BUG_ON(pa->pa_type != MB_INODE_PA);
5049 group = ext4_get_group_number(sb, pa->pa_pstart);
5051 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5052 GFP_NOFS|__GFP_NOFAIL);
5054 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5059 bitmap_bh = ext4_read_block_bitmap(sb, group);
5060 if (IS_ERR(bitmap_bh)) {
5061 err = PTR_ERR(bitmap_bh);
5062 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
5064 ext4_mb_unload_buddy(&e4b);
5068 ext4_lock_group(sb, group);
5069 list_del(&pa->pa_group_list);
5070 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5071 ext4_unlock_group(sb, group);
5073 ext4_mb_unload_buddy(&e4b);
5076 list_del(&pa->u.pa_tmp_list);
5077 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5081 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
5083 struct ext4_prealloc_space *pa;
5085 BUG_ON(ext4_pspace_cachep == NULL);
5086 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
5089 atomic_set(&pa->pa_count, 1);
5094 static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
5096 struct ext4_prealloc_space *pa = ac->ac_pa;
5100 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
5101 kmem_cache_free(ext4_pspace_cachep, pa);
5104 #ifdef CONFIG_EXT4_DEBUG
5105 static inline void ext4_mb_show_pa(struct super_block *sb)
5107 ext4_group_t i, ngroups;
5109 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5112 ngroups = ext4_get_groups_count(sb);
5113 mb_debug(sb, "groups: ");
5114 for (i = 0; i < ngroups; i++) {
5115 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5116 struct ext4_prealloc_space *pa;
5117 ext4_grpblk_t start;
5118 struct list_head *cur;
5119 ext4_lock_group(sb, i);
5120 list_for_each(cur, &grp->bb_prealloc_list) {
5121 pa = list_entry(cur, struct ext4_prealloc_space,
5123 spin_lock(&pa->pa_lock);
5124 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5126 spin_unlock(&pa->pa_lock);
5127 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5130 ext4_unlock_group(sb, i);
5131 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5136 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5138 struct super_block *sb = ac->ac_sb;
5140 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5143 mb_debug(sb, "Can't allocate:"
5144 " Allocation context details:");
5145 mb_debug(sb, "status %u flags 0x%x",
5146 ac->ac_status, ac->ac_flags);
5147 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5148 "goal %lu/%lu/%lu@%lu, "
5149 "best %lu/%lu/%lu@%lu cr %d",
5150 (unsigned long)ac->ac_o_ex.fe_group,
5151 (unsigned long)ac->ac_o_ex.fe_start,
5152 (unsigned long)ac->ac_o_ex.fe_len,
5153 (unsigned long)ac->ac_o_ex.fe_logical,
5154 (unsigned long)ac->ac_g_ex.fe_group,
5155 (unsigned long)ac->ac_g_ex.fe_start,
5156 (unsigned long)ac->ac_g_ex.fe_len,
5157 (unsigned long)ac->ac_g_ex.fe_logical,
5158 (unsigned long)ac->ac_b_ex.fe_group,
5159 (unsigned long)ac->ac_b_ex.fe_start,
5160 (unsigned long)ac->ac_b_ex.fe_len,
5161 (unsigned long)ac->ac_b_ex.fe_logical,
5162 (int)ac->ac_criteria);
5163 mb_debug(sb, "%u found", ac->ac_found);
5164 ext4_mb_show_pa(sb);
5167 static inline void ext4_mb_show_pa(struct super_block *sb)
5171 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5173 ext4_mb_show_pa(ac->ac_sb);
5179 * We use locality group preallocation for small size file. The size of the
5180 * file is determined by the current size or the resulting size after
5181 * allocation which ever is larger
5183 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5185 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5187 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5188 int bsbits = ac->ac_sb->s_blocksize_bits;
5191 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5194 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5197 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
5198 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5201 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5202 !inode_is_open_for_write(ac->ac_inode)) {
5203 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5207 if (sbi->s_mb_group_prealloc <= 0) {
5208 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5212 /* don't use group allocation for large files */
5213 size = max(size, isize);
5214 if (size > sbi->s_mb_stream_request) {
5215 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5219 BUG_ON(ac->ac_lg != NULL);
5221 * locality group prealloc space are per cpu. The reason for having
5222 * per cpu locality group is to reduce the contention between block
5223 * request from multiple CPUs.
5225 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5227 /* we're going to use group allocation */
5228 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5230 /* serialize all allocations in the group */
5231 mutex_lock(&ac->ac_lg->lg_mutex);
5234 static noinline_for_stack int
5235 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5236 struct ext4_allocation_request *ar)
5238 struct super_block *sb = ar->inode->i_sb;
5239 struct ext4_sb_info *sbi = EXT4_SB(sb);
5240 struct ext4_super_block *es = sbi->s_es;
5244 ext4_grpblk_t block;
5246 /* we can't allocate > group size */
5249 /* just a dirty hack to filter too big requests */
5250 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5251 len = EXT4_CLUSTERS_PER_GROUP(sb);
5253 /* start searching from the goal */
5255 if (goal < le32_to_cpu(es->s_first_data_block) ||
5256 goal >= ext4_blocks_count(es))
5257 goal = le32_to_cpu(es->s_first_data_block);
5258 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5260 /* set up allocation goals */
5261 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5262 ac->ac_status = AC_STATUS_CONTINUE;
5264 ac->ac_inode = ar->inode;
5265 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5266 ac->ac_o_ex.fe_group = group;
5267 ac->ac_o_ex.fe_start = block;
5268 ac->ac_o_ex.fe_len = len;
5269 ac->ac_g_ex = ac->ac_o_ex;
5270 ac->ac_flags = ar->flags;
5272 /* we have to define context: we'll work with a file or
5273 * locality group. this is a policy, actually */
5274 ext4_mb_group_or_file(ac);
5276 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5277 "left: %u/%u, right %u/%u to %swritable\n",
5278 (unsigned) ar->len, (unsigned) ar->logical,
5279 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5280 (unsigned) ar->lleft, (unsigned) ar->pleft,
5281 (unsigned) ar->lright, (unsigned) ar->pright,
5282 inode_is_open_for_write(ar->inode) ? "" : "non-");
5287 static noinline_for_stack void
5288 ext4_mb_discard_lg_preallocations(struct super_block *sb,
5289 struct ext4_locality_group *lg,
5290 int order, int total_entries)
5292 ext4_group_t group = 0;
5293 struct ext4_buddy e4b;
5294 struct list_head discard_list;
5295 struct ext4_prealloc_space *pa, *tmp;
5297 mb_debug(sb, "discard locality group preallocation\n");
5299 INIT_LIST_HEAD(&discard_list);
5301 spin_lock(&lg->lg_prealloc_lock);
5302 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5304 lockdep_is_held(&lg->lg_prealloc_lock)) {
5305 spin_lock(&pa->pa_lock);
5306 if (atomic_read(&pa->pa_count)) {
5308 * This is the pa that we just used
5309 * for block allocation. So don't
5312 spin_unlock(&pa->pa_lock);
5315 if (pa->pa_deleted) {
5316 spin_unlock(&pa->pa_lock);
5319 /* only lg prealloc space */
5320 BUG_ON(pa->pa_type != MB_GROUP_PA);
5322 /* seems this one can be freed ... */
5323 ext4_mb_mark_pa_deleted(sb, pa);
5324 spin_unlock(&pa->pa_lock);
5326 list_del_rcu(&pa->pa_inode_list);
5327 list_add(&pa->u.pa_tmp_list, &discard_list);
5330 if (total_entries <= 5) {
5332 * we want to keep only 5 entries
5333 * allowing it to grow to 8. This
5334 * mak sure we don't call discard
5335 * soon for this list.
5340 spin_unlock(&lg->lg_prealloc_lock);
5342 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5345 group = ext4_get_group_number(sb, pa->pa_pstart);
5346 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5347 GFP_NOFS|__GFP_NOFAIL);
5349 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5353 ext4_lock_group(sb, group);
5354 list_del(&pa->pa_group_list);
5355 ext4_mb_release_group_pa(&e4b, pa);
5356 ext4_unlock_group(sb, group);
5358 ext4_mb_unload_buddy(&e4b);
5359 list_del(&pa->u.pa_tmp_list);
5360 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5365 * We have incremented pa_count. So it cannot be freed at this
5366 * point. Also we hold lg_mutex. So no parallel allocation is
5367 * possible from this lg. That means pa_free cannot be updated.
5369 * A parallel ext4_mb_discard_group_preallocations is possible.
5370 * which can cause the lg_prealloc_list to be updated.
5373 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5375 int order, added = 0, lg_prealloc_count = 1;
5376 struct super_block *sb = ac->ac_sb;
5377 struct ext4_locality_group *lg = ac->ac_lg;
5378 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5380 order = fls(pa->pa_free) - 1;
5381 if (order > PREALLOC_TB_SIZE - 1)
5382 /* The max size of hash table is PREALLOC_TB_SIZE */
5383 order = PREALLOC_TB_SIZE - 1;
5384 /* Add the prealloc space to lg */
5385 spin_lock(&lg->lg_prealloc_lock);
5386 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5388 lockdep_is_held(&lg->lg_prealloc_lock)) {
5389 spin_lock(&tmp_pa->pa_lock);
5390 if (tmp_pa->pa_deleted) {
5391 spin_unlock(&tmp_pa->pa_lock);
5394 if (!added && pa->pa_free < tmp_pa->pa_free) {
5395 /* Add to the tail of the previous entry */
5396 list_add_tail_rcu(&pa->pa_inode_list,
5397 &tmp_pa->pa_inode_list);
5400 * we want to count the total
5401 * number of entries in the list
5404 spin_unlock(&tmp_pa->pa_lock);
5405 lg_prealloc_count++;
5408 list_add_tail_rcu(&pa->pa_inode_list,
5409 &lg->lg_prealloc_list[order]);
5410 spin_unlock(&lg->lg_prealloc_lock);
5412 /* Now trim the list to be not more than 8 elements */
5413 if (lg_prealloc_count > 8) {
5414 ext4_mb_discard_lg_preallocations(sb, lg,
5415 order, lg_prealloc_count);
5422 * if per-inode prealloc list is too long, trim some PA
5424 static void ext4_mb_trim_inode_pa(struct inode *inode)
5426 struct ext4_inode_info *ei = EXT4_I(inode);
5427 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5430 count = atomic_read(&ei->i_prealloc_active);
5431 delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
5432 if (count > sbi->s_mb_max_inode_prealloc + delta) {
5433 count -= sbi->s_mb_max_inode_prealloc;
5434 ext4_discard_preallocations(inode, count);
5439 * release all resource we used in allocation
5441 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
5443 struct inode *inode = ac->ac_inode;
5444 struct ext4_inode_info *ei = EXT4_I(inode);
5445 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5446 struct ext4_prealloc_space *pa = ac->ac_pa;
5448 if (pa->pa_type == MB_GROUP_PA) {
5449 /* see comment in ext4_mb_use_group_pa() */
5450 spin_lock(&pa->pa_lock);
5451 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5452 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5453 pa->pa_free -= ac->ac_b_ex.fe_len;
5454 pa->pa_len -= ac->ac_b_ex.fe_len;
5455 spin_unlock(&pa->pa_lock);
5458 * We want to add the pa to the right bucket.
5459 * Remove it from the list and while adding
5460 * make sure the list to which we are adding
5463 if (likely(pa->pa_free)) {
5464 spin_lock(pa->pa_obj_lock);
5465 list_del_rcu(&pa->pa_inode_list);
5466 spin_unlock(pa->pa_obj_lock);
5467 ext4_mb_add_n_trim(ac);
5471 if (pa->pa_type == MB_INODE_PA) {
5473 * treat per-inode prealloc list as a lru list, then try
5474 * to trim the least recently used PA.
5476 spin_lock(pa->pa_obj_lock);
5477 list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
5478 spin_unlock(pa->pa_obj_lock);
5481 ext4_mb_put_pa(ac, ac->ac_sb, pa);
5483 if (ac->ac_bitmap_page)
5484 put_page(ac->ac_bitmap_page);
5485 if (ac->ac_buddy_page)
5486 put_page(ac->ac_buddy_page);
5487 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5488 mutex_unlock(&ac->ac_lg->lg_mutex);
5489 ext4_mb_collect_stats(ac);
5490 ext4_mb_trim_inode_pa(inode);
5494 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
5496 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
5498 int freed = 0, busy = 0;
5501 trace_ext4_mb_discard_preallocations(sb, needed);
5504 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
5506 for (i = 0; i < ngroups && needed > 0; i++) {
5507 ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
5513 if (needed > 0 && busy && ++retry < 3) {
5521 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
5522 struct ext4_allocation_context *ac, u64 *seq)
5528 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
5533 seq_retry = ext4_get_discard_pa_seq_sum();
5534 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
5535 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
5541 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
5545 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5546 struct ext4_allocation_request *ar, int *errp);
5549 * Main entry point into mballoc to allocate blocks
5550 * it tries to use preallocation first, then falls back
5551 * to usual allocation
5553 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
5554 struct ext4_allocation_request *ar, int *errp)
5556 struct ext4_allocation_context *ac = NULL;
5557 struct ext4_sb_info *sbi;
5558 struct super_block *sb;
5559 ext4_fsblk_t block = 0;
5560 unsigned int inquota = 0;
5561 unsigned int reserv_clstrs = 0;
5566 sb = ar->inode->i_sb;
5569 trace_ext4_request_blocks(ar);
5570 if (sbi->s_mount_state & EXT4_FC_REPLAY)
5571 return ext4_mb_new_blocks_simple(handle, ar, errp);
5573 /* Allow to use superuser reservation for quota file */
5574 if (ext4_is_quota_file(ar->inode))
5575 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
5577 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
5578 /* Without delayed allocation we need to verify
5579 * there is enough free blocks to do block allocation
5580 * and verify allocation doesn't exceed the quota limits.
5583 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
5585 /* let others to free the space */
5587 ar->len = ar->len >> 1;
5590 ext4_mb_show_pa(sb);
5594 reserv_clstrs = ar->len;
5595 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
5596 dquot_alloc_block_nofail(ar->inode,
5597 EXT4_C2B(sbi, ar->len));
5600 dquot_alloc_block(ar->inode,
5601 EXT4_C2B(sbi, ar->len))) {
5603 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
5614 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
5621 *errp = ext4_mb_initialize_context(ac, ar);
5627 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
5628 seq = this_cpu_read(discard_pa_seq);
5629 if (!ext4_mb_use_preallocated(ac)) {
5630 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
5631 ext4_mb_normalize_request(ac, ar);
5633 *errp = ext4_mb_pa_alloc(ac);
5637 /* allocate space in core */
5638 *errp = ext4_mb_regular_allocator(ac);
5640 * pa allocated above is added to grp->bb_prealloc_list only
5641 * when we were able to allocate some block i.e. when
5642 * ac->ac_status == AC_STATUS_FOUND.
5643 * And error from above mean ac->ac_status != AC_STATUS_FOUND
5644 * So we have to free this pa here itself.
5647 ext4_mb_pa_free(ac);
5648 ext4_discard_allocated_blocks(ac);
5651 if (ac->ac_status == AC_STATUS_FOUND &&
5652 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
5653 ext4_mb_pa_free(ac);
5655 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
5656 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
5658 ext4_discard_allocated_blocks(ac);
5661 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5662 ar->len = ac->ac_b_ex.fe_len;
5665 if (++retries < 3 &&
5666 ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
5669 * If block allocation fails then the pa allocated above
5670 * needs to be freed here itself.
5672 ext4_mb_pa_free(ac);
5678 ac->ac_b_ex.fe_len = 0;
5680 ext4_mb_show_ac(ac);
5682 ext4_mb_release_context(ac);
5685 kmem_cache_free(ext4_ac_cachep, ac);
5686 if (inquota && ar->len < inquota)
5687 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
5689 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
5690 /* release all the reserved blocks if non delalloc */
5691 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
5695 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
5701 * We can merge two free data extents only if the physical blocks
5702 * are contiguous, AND the extents were freed by the same transaction,
5703 * AND the blocks are associated with the same group.
5705 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
5706 struct ext4_free_data *entry,
5707 struct ext4_free_data *new_entry,
5708 struct rb_root *entry_rb_root)
5710 if ((entry->efd_tid != new_entry->efd_tid) ||
5711 (entry->efd_group != new_entry->efd_group))
5713 if (entry->efd_start_cluster + entry->efd_count ==
5714 new_entry->efd_start_cluster) {
5715 new_entry->efd_start_cluster = entry->efd_start_cluster;
5716 new_entry->efd_count += entry->efd_count;
5717 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
5718 entry->efd_start_cluster) {
5719 new_entry->efd_count += entry->efd_count;
5722 spin_lock(&sbi->s_md_lock);
5723 list_del(&entry->efd_list);
5724 spin_unlock(&sbi->s_md_lock);
5725 rb_erase(&entry->efd_node, entry_rb_root);
5726 kmem_cache_free(ext4_free_data_cachep, entry);
5729 static noinline_for_stack int
5730 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
5731 struct ext4_free_data *new_entry)
5733 ext4_group_t group = e4b->bd_group;
5734 ext4_grpblk_t cluster;
5735 ext4_grpblk_t clusters = new_entry->efd_count;
5736 struct ext4_free_data *entry;
5737 struct ext4_group_info *db = e4b->bd_info;
5738 struct super_block *sb = e4b->bd_sb;
5739 struct ext4_sb_info *sbi = EXT4_SB(sb);
5740 struct rb_node **n = &db->bb_free_root.rb_node, *node;
5741 struct rb_node *parent = NULL, *new_node;
5743 BUG_ON(!ext4_handle_valid(handle));
5744 BUG_ON(e4b->bd_bitmap_page == NULL);
5745 BUG_ON(e4b->bd_buddy_page == NULL);
5747 new_node = &new_entry->efd_node;
5748 cluster = new_entry->efd_start_cluster;
5751 /* first free block exent. We need to
5752 protect buddy cache from being freed,
5753 * otherwise we'll refresh it from
5754 * on-disk bitmap and lose not-yet-available
5756 get_page(e4b->bd_buddy_page);
5757 get_page(e4b->bd_bitmap_page);
5761 entry = rb_entry(parent, struct ext4_free_data, efd_node);
5762 if (cluster < entry->efd_start_cluster)
5764 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
5765 n = &(*n)->rb_right;
5767 ext4_grp_locked_error(sb, group, 0,
5768 ext4_group_first_block_no(sb, group) +
5769 EXT4_C2B(sbi, cluster),
5770 "Block already on to-be-freed list");
5771 kmem_cache_free(ext4_free_data_cachep, new_entry);
5776 rb_link_node(new_node, parent, n);
5777 rb_insert_color(new_node, &db->bb_free_root);
5779 /* Now try to see the extent can be merged to left and right */
5780 node = rb_prev(new_node);
5782 entry = rb_entry(node, struct ext4_free_data, efd_node);
5783 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5784 &(db->bb_free_root));
5787 node = rb_next(new_node);
5789 entry = rb_entry(node, struct ext4_free_data, efd_node);
5790 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5791 &(db->bb_free_root));
5794 spin_lock(&sbi->s_md_lock);
5795 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5796 sbi->s_mb_free_pending += clusters;
5797 spin_unlock(&sbi->s_md_lock);
5802 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
5803 * linearly starting at the goal block and also excludes the blocks which
5804 * are going to be in use after fast commit replay.
5806 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5807 struct ext4_allocation_request *ar, int *errp)
5809 struct buffer_head *bitmap_bh;
5810 struct super_block *sb = ar->inode->i_sb;
5812 ext4_grpblk_t blkoff;
5813 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
5814 ext4_grpblk_t i = 0;
5815 ext4_fsblk_t goal, block;
5816 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
5819 if (goal < le32_to_cpu(es->s_first_data_block) ||
5820 goal >= ext4_blocks_count(es))
5821 goal = le32_to_cpu(es->s_first_data_block);
5824 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
5825 for (; group < ext4_get_groups_count(sb); group++) {
5826 bitmap_bh = ext4_read_block_bitmap(sb, group);
5827 if (IS_ERR(bitmap_bh)) {
5828 *errp = PTR_ERR(bitmap_bh);
5829 pr_warn("Failed to read block bitmap\n");
5833 ext4_get_group_no_and_offset(sb,
5834 max(ext4_group_first_block_no(sb, group), goal),
5837 i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
5841 if (ext4_fc_replay_check_excluded(sb,
5842 ext4_group_first_block_no(sb, group) + i)) {
5852 if (group >= ext4_get_groups_count(sb) || i >= max) {
5857 block = ext4_group_first_block_no(sb, group) + i;
5858 ext4_mb_mark_bb(sb, block, 1, 1);
5864 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
5865 unsigned long count)
5867 struct buffer_head *bitmap_bh;
5868 struct super_block *sb = inode->i_sb;
5869 struct ext4_group_desc *gdp;
5870 struct buffer_head *gdp_bh;
5872 ext4_grpblk_t blkoff;
5873 int already_freed = 0, err, i;
5875 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
5876 bitmap_bh = ext4_read_block_bitmap(sb, group);
5877 if (IS_ERR(bitmap_bh)) {
5878 err = PTR_ERR(bitmap_bh);
5879 pr_warn("Failed to read block bitmap\n");
5882 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
5886 for (i = 0; i < count; i++) {
5887 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data))
5890 mb_clear_bits(bitmap_bh->b_data, blkoff, count);
5891 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
5894 ext4_free_group_clusters_set(
5895 sb, gdp, ext4_free_group_clusters(sb, gdp) +
5896 count - already_freed);
5897 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
5898 ext4_group_desc_csum_set(sb, group, gdp);
5899 ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
5900 sync_dirty_buffer(bitmap_bh);
5901 sync_dirty_buffer(gdp_bh);
5906 * ext4_mb_clear_bb() -- helper function for freeing blocks.
5907 * Used by ext4_free_blocks()
5908 * @handle: handle for this transaction
5910 * @block: starting physical block to be freed
5911 * @count: number of blocks to be freed
5912 * @flags: flags used by ext4_free_blocks
5914 static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
5915 ext4_fsblk_t block, unsigned long count,
5918 struct buffer_head *bitmap_bh = NULL;
5919 struct super_block *sb = inode->i_sb;
5920 struct ext4_group_desc *gdp;
5921 unsigned int overflow;
5923 struct buffer_head *gd_bh;
5924 ext4_group_t block_group;
5925 struct ext4_sb_info *sbi;
5926 struct ext4_buddy e4b;
5927 unsigned int count_clusters;
5933 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5934 !ext4_inode_block_valid(inode, block, count)) {
5935 ext4_error(sb, "Freeing blocks in system zone - "
5936 "Block = %llu, count = %lu", block, count);
5937 /* err = 0. ext4_std_error should be a no op */
5940 flags |= EXT4_FREE_BLOCKS_VALIDATED;
5944 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5946 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5947 ext4_get_group_info(sb, block_group))))
5951 * Check to see if we are freeing blocks across a group
5954 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5955 overflow = EXT4_C2B(sbi, bit) + count -
5956 EXT4_BLOCKS_PER_GROUP(sb);
5958 /* The range changed so it's no longer validated */
5959 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
5961 count_clusters = EXT4_NUM_B2C(sbi, count);
5962 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5963 if (IS_ERR(bitmap_bh)) {
5964 err = PTR_ERR(bitmap_bh);
5968 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5974 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5975 !ext4_inode_block_valid(inode, block, count)) {
5976 ext4_error(sb, "Freeing blocks in system zone - "
5977 "Block = %llu, count = %lu", block, count);
5978 /* err = 0. ext4_std_error should be a no op */
5982 BUFFER_TRACE(bitmap_bh, "getting write access");
5983 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
5989 * We are about to modify some metadata. Call the journal APIs
5990 * to unshare ->b_data if a currently-committing transaction is
5993 BUFFER_TRACE(gd_bh, "get_write_access");
5994 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
5997 #ifdef AGGRESSIVE_CHECK
6000 for (i = 0; i < count_clusters; i++)
6001 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
6004 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
6006 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
6007 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
6008 GFP_NOFS|__GFP_NOFAIL);
6013 * We need to make sure we don't reuse the freed block until after the
6014 * transaction is committed. We make an exception if the inode is to be
6015 * written in writeback mode since writeback mode has weak data
6016 * consistency guarantees.
6018 if (ext4_handle_valid(handle) &&
6019 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
6020 !ext4_should_writeback_data(inode))) {
6021 struct ext4_free_data *new_entry;
6023 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
6026 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
6027 GFP_NOFS|__GFP_NOFAIL);
6028 new_entry->efd_start_cluster = bit;
6029 new_entry->efd_group = block_group;
6030 new_entry->efd_count = count_clusters;
6031 new_entry->efd_tid = handle->h_transaction->t_tid;
6033 ext4_lock_group(sb, block_group);
6034 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6035 ext4_mb_free_metadata(handle, &e4b, new_entry);
6037 /* need to update group_info->bb_free and bitmap
6038 * with group lock held. generate_buddy look at
6039 * them with group lock_held
6041 if (test_opt(sb, DISCARD)) {
6042 err = ext4_issue_discard(sb, block_group, bit, count,
6044 if (err && err != -EOPNOTSUPP)
6045 ext4_msg(sb, KERN_WARNING, "discard request in"
6046 " group:%u block:%d count:%lu failed"
6047 " with %d", block_group, bit, count,
6050 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
6052 ext4_lock_group(sb, block_group);
6053 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6054 mb_free_blocks(inode, &e4b, bit, count_clusters);
6057 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
6058 ext4_free_group_clusters_set(sb, gdp, ret);
6059 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
6060 ext4_group_desc_csum_set(sb, block_group, gdp);
6061 ext4_unlock_group(sb, block_group);
6063 if (sbi->s_log_groups_per_flex) {
6064 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6065 atomic64_add(count_clusters,
6066 &sbi_array_rcu_deref(sbi, s_flex_groups,
6067 flex_group)->free_clusters);
6071 * on a bigalloc file system, defer the s_freeclusters_counter
6072 * update to the caller (ext4_remove_space and friends) so they
6073 * can determine if a cluster freed here should be rereserved
6075 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6076 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6077 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6078 percpu_counter_add(&sbi->s_freeclusters_counter,
6082 ext4_mb_unload_buddy(&e4b);
6084 /* We dirtied the bitmap block */
6085 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6086 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6088 /* And the group descriptor block */
6089 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6090 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6094 if (overflow && !err) {
6098 /* The range changed so it's no longer validated */
6099 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6104 ext4_std_error(sb, err);
6109 * ext4_free_blocks() -- Free given blocks and update quota
6110 * @handle: handle for this transaction
6112 * @bh: optional buffer of the block to be freed
6113 * @block: starting physical block to be freed
6114 * @count: number of blocks to be freed
6115 * @flags: flags used by ext4_free_blocks
6117 void ext4_free_blocks(handle_t *handle, struct inode *inode,
6118 struct buffer_head *bh, ext4_fsblk_t block,
6119 unsigned long count, int flags)
6121 struct super_block *sb = inode->i_sb;
6122 unsigned int overflow;
6123 struct ext4_sb_info *sbi;
6127 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
6128 ext4_free_blocks_simple(inode, block, count);
6135 BUG_ON(block != bh->b_blocknr);
6137 block = bh->b_blocknr;
6140 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6141 !ext4_inode_block_valid(inode, block, count)) {
6142 ext4_error(sb, "Freeing blocks not in datazone - "
6143 "block = %llu, count = %lu", block, count);
6146 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6148 ext4_debug("freeing block %llu\n", block);
6149 trace_ext4_free_blocks(inode, block, count, flags);
6151 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6154 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
6159 * If the extent to be freed does not begin on a cluster
6160 * boundary, we need to deal with partial clusters at the
6161 * beginning and end of the extent. Normally we will free
6162 * blocks at the beginning or the end unless we are explicitly
6163 * requested to avoid doing so.
6165 overflow = EXT4_PBLK_COFF(sbi, block);
6167 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
6168 overflow = sbi->s_cluster_ratio - overflow;
6170 if (count > overflow)
6178 /* The range changed so it's no longer validated */
6179 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6181 overflow = EXT4_LBLK_COFF(sbi, count);
6183 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
6184 if (count > overflow)
6189 count += sbi->s_cluster_ratio - overflow;
6190 /* The range changed so it's no longer validated */
6191 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6194 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6196 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
6198 for (i = 0; i < count; i++) {
6201 bh = sb_find_get_block(inode->i_sb, block + i);
6202 ext4_forget(handle, is_metadata, inode, bh, block + i);
6206 ext4_mb_clear_bb(handle, inode, block, count, flags);
6211 * ext4_group_add_blocks() -- Add given blocks to an existing group
6212 * @handle: handle to this transaction
6214 * @block: start physical block to add to the block group
6215 * @count: number of blocks to free
6217 * This marks the blocks as free in the bitmap and buddy.
6219 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6220 ext4_fsblk_t block, unsigned long count)
6222 struct buffer_head *bitmap_bh = NULL;
6223 struct buffer_head *gd_bh;
6224 ext4_group_t block_group;
6227 struct ext4_group_desc *desc;
6228 struct ext4_sb_info *sbi = EXT4_SB(sb);
6229 struct ext4_buddy e4b;
6230 int err = 0, ret, free_clusters_count;
6231 ext4_grpblk_t clusters_freed;
6232 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6233 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6234 unsigned long cluster_count = last_cluster - first_cluster + 1;
6236 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6241 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6243 * Check to see if we are freeing blocks across a group
6246 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6247 ext4_warning(sb, "too many blocks added to group %u",
6253 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
6254 if (IS_ERR(bitmap_bh)) {
6255 err = PTR_ERR(bitmap_bh);
6260 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
6266 if (!ext4_sb_block_valid(sb, NULL, block, count)) {
6267 ext4_error(sb, "Adding blocks in system zones - "
6268 "Block = %llu, count = %lu",
6274 BUFFER_TRACE(bitmap_bh, "getting write access");
6275 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
6281 * We are about to modify some metadata. Call the journal APIs
6282 * to unshare ->b_data if a currently-committing transaction is
6285 BUFFER_TRACE(gd_bh, "get_write_access");
6286 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
6290 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
6291 BUFFER_TRACE(bitmap_bh, "clear bit");
6292 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
6293 ext4_error(sb, "bit already cleared for block %llu",
6294 (ext4_fsblk_t)(block + i));
6295 BUFFER_TRACE(bitmap_bh, "bit already cleared");
6301 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6306 * need to update group_info->bb_free and bitmap
6307 * with group lock held. generate_buddy look at
6308 * them with group lock_held
6310 ext4_lock_group(sb, block_group);
6311 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
6312 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6313 free_clusters_count = clusters_freed +
6314 ext4_free_group_clusters(sb, desc);
6315 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
6316 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
6317 ext4_group_desc_csum_set(sb, block_group, desc);
6318 ext4_unlock_group(sb, block_group);
6319 percpu_counter_add(&sbi->s_freeclusters_counter,
6322 if (sbi->s_log_groups_per_flex) {
6323 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6324 atomic64_add(clusters_freed,
6325 &sbi_array_rcu_deref(sbi, s_flex_groups,
6326 flex_group)->free_clusters);
6329 ext4_mb_unload_buddy(&e4b);
6331 /* We dirtied the bitmap block */
6332 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6333 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6335 /* And the group descriptor block */
6336 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6337 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6343 ext4_std_error(sb, err);
6348 * ext4_trim_extent -- function to TRIM one single free extent in the group
6349 * @sb: super block for the file system
6350 * @start: starting block of the free extent in the alloc. group
6351 * @count: number of blocks to TRIM
6352 * @e4b: ext4 buddy for the group
6354 * Trim "count" blocks starting at "start" in the "group". To assure that no
6355 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6356 * be called with under the group lock.
6358 static int ext4_trim_extent(struct super_block *sb,
6359 int start, int count, struct ext4_buddy *e4b)
6363 struct ext4_free_extent ex;
6364 ext4_group_t group = e4b->bd_group;
6367 trace_ext4_trim_extent(sb, group, start, count);
6369 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6371 ex.fe_start = start;
6372 ex.fe_group = group;
6376 * Mark blocks used, so no one can reuse them while
6379 mb_mark_used(e4b, &ex);
6380 ext4_unlock_group(sb, group);
6381 ret = ext4_issue_discard(sb, group, start, count, NULL);
6382 ext4_lock_group(sb, group);
6383 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6387 static int ext4_try_to_trim_range(struct super_block *sb,
6388 struct ext4_buddy *e4b, ext4_grpblk_t start,
6389 ext4_grpblk_t max, ext4_grpblk_t minblocks)
6390 __acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
6391 __releases(ext4_group_lock_ptr(sb, e4b->bd_group))
6393 ext4_grpblk_t next, count, free_count;
6396 bitmap = e4b->bd_bitmap;
6397 start = (e4b->bd_info->bb_first_free > start) ?
6398 e4b->bd_info->bb_first_free : start;
6402 while (start <= max) {
6403 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6406 next = mb_find_next_bit(bitmap, max + 1, start);
6408 if ((next - start) >= minblocks) {
6409 int ret = ext4_trim_extent(sb, start, next - start, e4b);
6411 if (ret && ret != -EOPNOTSUPP)
6413 count += next - start;
6415 free_count += next - start;
6418 if (fatal_signal_pending(current)) {
6419 count = -ERESTARTSYS;
6423 if (need_resched()) {
6424 ext4_unlock_group(sb, e4b->bd_group);
6426 ext4_lock_group(sb, e4b->bd_group);
6429 if ((e4b->bd_info->bb_free - free_count) < minblocks)
6437 * ext4_trim_all_free -- function to trim all free space in alloc. group
6438 * @sb: super block for file system
6439 * @group: group to be trimmed
6440 * @start: first group block to examine
6441 * @max: last group block to examine
6442 * @minblocks: minimum extent block count
6443 * @set_trimmed: set the trimmed flag if at least one block is trimmed
6445 * ext4_trim_all_free walks through group's block bitmap searching for free
6446 * extents. When the free extent is found, mark it as used in group buddy
6447 * bitmap. Then issue a TRIM command on this extent and free the extent in
6448 * the group buddy bitmap.
6450 static ext4_grpblk_t
6451 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6452 ext4_grpblk_t start, ext4_grpblk_t max,
6453 ext4_grpblk_t minblocks, bool set_trimmed)
6455 struct ext4_buddy e4b;
6458 trace_ext4_trim_all_free(sb, group, start, max);
6460 ret = ext4_mb_load_buddy(sb, group, &e4b);
6462 ext4_warning(sb, "Error %d loading buddy information for %u",
6467 ext4_lock_group(sb, group);
6469 if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
6470 minblocks < EXT4_SB(sb)->s_last_trim_minblks) {
6471 ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
6472 if (ret >= 0 && set_trimmed)
6473 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
6478 ext4_unlock_group(sb, group);
6479 ext4_mb_unload_buddy(&e4b);
6481 ext4_debug("trimmed %d blocks in the group %d\n",
6488 * ext4_trim_fs() -- trim ioctl handle function
6489 * @sb: superblock for filesystem
6490 * @range: fstrim_range structure
6492 * start: First Byte to trim
6493 * len: number of Bytes to trim from start
6494 * minlen: minimum extent length in Bytes
6495 * ext4_trim_fs goes through all allocation groups containing Bytes from
6496 * start to start+len. For each such a group ext4_trim_all_free function
6497 * is invoked to trim all free space.
6499 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6501 unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
6502 struct ext4_group_info *grp;
6503 ext4_group_t group, first_group, last_group;
6504 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6505 uint64_t start, end, minlen, trimmed = 0;
6506 ext4_fsblk_t first_data_blk =
6507 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6508 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6509 bool whole_group, eof = false;
6512 start = range->start >> sb->s_blocksize_bits;
6513 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6514 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6515 range->minlen >> sb->s_blocksize_bits);
6517 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6518 start >= max_blks ||
6519 range->len < sb->s_blocksize)
6521 /* No point to try to trim less than discard granularity */
6522 if (range->minlen < discard_granularity) {
6523 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6524 discard_granularity >> sb->s_blocksize_bits);
6525 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
6528 if (end >= max_blks - 1) {
6532 if (end <= first_data_blk)
6534 if (start < first_data_blk)
6535 start = first_data_blk;
6537 /* Determine first and last group to examine based on start and end */
6538 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6539 &first_group, &first_cluster);
6540 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6541 &last_group, &last_cluster);
6543 /* end now represents the last cluster to discard in this group */
6544 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6547 for (group = first_group; group <= last_group; group++) {
6548 grp = ext4_get_group_info(sb, group);
6549 /* We only do this if the grp has never been initialized */
6550 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6551 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6557 * For all the groups except the last one, last cluster will
6558 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6559 * change it for the last group, note that last_cluster is
6560 * already computed earlier by ext4_get_group_no_and_offset()
6562 if (group == last_group) {
6564 whole_group = eof ? true : end == EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6566 if (grp->bb_free >= minlen) {
6567 cnt = ext4_trim_all_free(sb, group, first_cluster,
6568 end, minlen, whole_group);
6577 * For every group except the first one, we are sure
6578 * that the first cluster to discard will be cluster #0.
6584 EXT4_SB(sb)->s_last_trim_minblks = minlen;
6587 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6591 /* Iterate all the free extents in the group. */
6593 ext4_mballoc_query_range(
6594 struct super_block *sb,
6596 ext4_grpblk_t start,
6598 ext4_mballoc_query_range_fn formatter,
6603 struct ext4_buddy e4b;
6606 error = ext4_mb_load_buddy(sb, group, &e4b);
6609 bitmap = e4b.bd_bitmap;
6611 ext4_lock_group(sb, group);
6613 start = (e4b.bd_info->bb_first_free > start) ?
6614 e4b.bd_info->bb_first_free : start;
6615 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
6616 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6618 while (start <= end) {
6619 start = mb_find_next_zero_bit(bitmap, end + 1, start);
6622 next = mb_find_next_bit(bitmap, end + 1, start);
6624 ext4_unlock_group(sb, group);
6625 error = formatter(sb, group, start, next - start, priv);
6628 ext4_lock_group(sb, group);
6633 ext4_unlock_group(sb, group);
6635 ext4_mb_unload_buddy(&e4b);