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 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--)
1081 if (grp->bb_counters[i] > 0)
1083 /* No need to move between order lists? */
1084 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) ||
1085 i == grp->bb_largest_free_order) {
1086 grp->bb_largest_free_order = i;
1090 if (grp->bb_largest_free_order >= 0) {
1091 write_lock(&sbi->s_mb_largest_free_orders_locks[
1092 grp->bb_largest_free_order]);
1093 list_del_init(&grp->bb_largest_free_order_node);
1094 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1095 grp->bb_largest_free_order]);
1097 grp->bb_largest_free_order = i;
1098 if (grp->bb_largest_free_order >= 0 && grp->bb_free) {
1099 write_lock(&sbi->s_mb_largest_free_orders_locks[
1100 grp->bb_largest_free_order]);
1101 list_add_tail(&grp->bb_largest_free_order_node,
1102 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1103 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1104 grp->bb_largest_free_order]);
1108 static noinline_for_stack
1109 void ext4_mb_generate_buddy(struct super_block *sb,
1110 void *buddy, void *bitmap, ext4_group_t group)
1112 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
1113 struct ext4_sb_info *sbi = EXT4_SB(sb);
1114 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1115 ext4_grpblk_t i = 0;
1116 ext4_grpblk_t first;
1119 unsigned fragments = 0;
1120 unsigned long long period = get_cycles();
1122 /* initialize buddy from bitmap which is aggregation
1123 * of on-disk bitmap and preallocations */
1124 i = mb_find_next_zero_bit(bitmap, max, 0);
1125 grp->bb_first_free = i;
1129 i = mb_find_next_bit(bitmap, max, i);
1133 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1135 grp->bb_counters[0]++;
1137 i = mb_find_next_zero_bit(bitmap, max, i);
1139 grp->bb_fragments = fragments;
1141 if (free != grp->bb_free) {
1142 ext4_grp_locked_error(sb, group, 0, 0,
1143 "block bitmap and bg descriptor "
1144 "inconsistent: %u vs %u free clusters",
1145 free, grp->bb_free);
1147 * If we intend to continue, we consider group descriptor
1148 * corrupt and update bb_free using bitmap value
1150 grp->bb_free = free;
1151 ext4_mark_group_bitmap_corrupted(sb, group,
1152 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1154 mb_set_largest_free_order(sb, grp);
1156 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1158 period = get_cycles() - period;
1159 atomic_inc(&sbi->s_mb_buddies_generated);
1160 atomic64_add(period, &sbi->s_mb_generation_time);
1161 mb_update_avg_fragment_size(sb, grp);
1164 /* The buddy information is attached the buddy cache inode
1165 * for convenience. The information regarding each group
1166 * is loaded via ext4_mb_load_buddy. The information involve
1167 * block bitmap and buddy information. The information are
1168 * stored in the inode as
1171 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1174 * one block each for bitmap and buddy information.
1175 * So for each group we take up 2 blocks. A page can
1176 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
1177 * So it can have information regarding groups_per_page which
1178 * is blocks_per_page/2
1180 * Locking note: This routine takes the block group lock of all groups
1181 * for this page; do not hold this lock when calling this routine!
1184 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
1186 ext4_group_t ngroups;
1188 int blocks_per_page;
1189 int groups_per_page;
1192 ext4_group_t first_group, group;
1194 struct super_block *sb;
1195 struct buffer_head *bhs;
1196 struct buffer_head **bh = NULL;
1197 struct inode *inode;
1200 struct ext4_group_info *grinfo;
1202 inode = page->mapping->host;
1204 ngroups = ext4_get_groups_count(sb);
1205 blocksize = i_blocksize(inode);
1206 blocks_per_page = PAGE_SIZE / blocksize;
1208 mb_debug(sb, "init page %lu\n", page->index);
1210 groups_per_page = blocks_per_page >> 1;
1211 if (groups_per_page == 0)
1212 groups_per_page = 1;
1214 /* allocate buffer_heads to read bitmaps */
1215 if (groups_per_page > 1) {
1216 i = sizeof(struct buffer_head *) * groups_per_page;
1217 bh = kzalloc(i, gfp);
1225 first_group = page->index * blocks_per_page / 2;
1227 /* read all groups the page covers into the cache */
1228 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1229 if (group >= ngroups)
1232 grinfo = ext4_get_group_info(sb, group);
1234 * If page is uptodate then we came here after online resize
1235 * which added some new uninitialized group info structs, so
1236 * we must skip all initialized uptodate buddies on the page,
1237 * which may be currently in use by an allocating task.
1239 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1243 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1244 if (IS_ERR(bh[i])) {
1245 err = PTR_ERR(bh[i]);
1249 mb_debug(sb, "read bitmap for group %u\n", group);
1252 /* wait for I/O completion */
1253 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1258 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1263 first_block = page->index * blocks_per_page;
1264 for (i = 0; i < blocks_per_page; i++) {
1265 group = (first_block + i) >> 1;
1266 if (group >= ngroups)
1269 if (!bh[group - first_group])
1270 /* skip initialized uptodate buddy */
1273 if (!buffer_verified(bh[group - first_group]))
1274 /* Skip faulty bitmaps */
1279 * data carry information regarding this
1280 * particular group in the format specified
1284 data = page_address(page) + (i * blocksize);
1285 bitmap = bh[group - first_group]->b_data;
1288 * We place the buddy block and bitmap block
1291 if ((first_block + i) & 1) {
1292 /* this is block of buddy */
1293 BUG_ON(incore == NULL);
1294 mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
1295 group, page->index, i * blocksize);
1296 trace_ext4_mb_buddy_bitmap_load(sb, group);
1297 grinfo = ext4_get_group_info(sb, group);
1298 grinfo->bb_fragments = 0;
1299 memset(grinfo->bb_counters, 0,
1300 sizeof(*grinfo->bb_counters) *
1301 (MB_NUM_ORDERS(sb)));
1303 * incore got set to the group block bitmap below
1305 ext4_lock_group(sb, group);
1306 /* init the buddy */
1307 memset(data, 0xff, blocksize);
1308 ext4_mb_generate_buddy(sb, data, incore, group);
1309 ext4_unlock_group(sb, group);
1312 /* this is block of bitmap */
1313 BUG_ON(incore != NULL);
1314 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
1315 group, page->index, i * blocksize);
1316 trace_ext4_mb_bitmap_load(sb, group);
1318 /* see comments in ext4_mb_put_pa() */
1319 ext4_lock_group(sb, group);
1320 memcpy(data, bitmap, blocksize);
1322 /* mark all preallocated blks used in in-core bitmap */
1323 ext4_mb_generate_from_pa(sb, data, group);
1324 ext4_mb_generate_from_freelist(sb, data, group);
1325 ext4_unlock_group(sb, group);
1327 /* set incore so that the buddy information can be
1328 * generated using this
1333 SetPageUptodate(page);
1337 for (i = 0; i < groups_per_page; i++)
1346 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1347 * on the same buddy page doesn't happen whild holding the buddy page lock.
1348 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1349 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1351 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1352 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1354 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1355 int block, pnum, poff;
1356 int blocks_per_page;
1359 e4b->bd_buddy_page = NULL;
1360 e4b->bd_bitmap_page = NULL;
1362 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1364 * the buddy cache inode stores the block bitmap
1365 * and buddy information in consecutive blocks.
1366 * So for each group we need two blocks.
1369 pnum = block / blocks_per_page;
1370 poff = block % blocks_per_page;
1371 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1374 BUG_ON(page->mapping != inode->i_mapping);
1375 e4b->bd_bitmap_page = page;
1376 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1378 if (blocks_per_page >= 2) {
1379 /* buddy and bitmap are on the same page */
1384 pnum = block / blocks_per_page;
1385 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1388 BUG_ON(page->mapping != inode->i_mapping);
1389 e4b->bd_buddy_page = page;
1393 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1395 if (e4b->bd_bitmap_page) {
1396 unlock_page(e4b->bd_bitmap_page);
1397 put_page(e4b->bd_bitmap_page);
1399 if (e4b->bd_buddy_page) {
1400 unlock_page(e4b->bd_buddy_page);
1401 put_page(e4b->bd_buddy_page);
1406 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1407 * block group lock of all groups for this page; do not hold the BG lock when
1408 * calling this routine!
1410 static noinline_for_stack
1411 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1414 struct ext4_group_info *this_grp;
1415 struct ext4_buddy e4b;
1420 mb_debug(sb, "init group %u\n", group);
1421 this_grp = ext4_get_group_info(sb, group);
1423 * This ensures that we don't reinit the buddy cache
1424 * page which map to the group from which we are already
1425 * allocating. If we are looking at the buddy cache we would
1426 * have taken a reference using ext4_mb_load_buddy and that
1427 * would have pinned buddy page to page cache.
1428 * The call to ext4_mb_get_buddy_page_lock will mark the
1431 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1432 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1434 * somebody initialized the group
1435 * return without doing anything
1440 page = e4b.bd_bitmap_page;
1441 ret = ext4_mb_init_cache(page, NULL, gfp);
1444 if (!PageUptodate(page)) {
1449 if (e4b.bd_buddy_page == NULL) {
1451 * If both the bitmap and buddy are in
1452 * the same page we don't need to force
1458 /* init buddy cache */
1459 page = e4b.bd_buddy_page;
1460 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1463 if (!PageUptodate(page)) {
1468 ext4_mb_put_buddy_page_lock(&e4b);
1473 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1474 * block group lock of all groups for this page; do not hold the BG lock when
1475 * calling this routine!
1477 static noinline_for_stack int
1478 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1479 struct ext4_buddy *e4b, gfp_t gfp)
1481 int blocks_per_page;
1487 struct ext4_group_info *grp;
1488 struct ext4_sb_info *sbi = EXT4_SB(sb);
1489 struct inode *inode = sbi->s_buddy_cache;
1492 mb_debug(sb, "load group %u\n", group);
1494 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1495 grp = ext4_get_group_info(sb, group);
1497 e4b->bd_blkbits = sb->s_blocksize_bits;
1500 e4b->bd_group = group;
1501 e4b->bd_buddy_page = NULL;
1502 e4b->bd_bitmap_page = NULL;
1504 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1506 * we need full data about the group
1507 * to make a good selection
1509 ret = ext4_mb_init_group(sb, group, gfp);
1515 * the buddy cache inode stores the block bitmap
1516 * and buddy information in consecutive blocks.
1517 * So for each group we need two blocks.
1520 pnum = block / blocks_per_page;
1521 poff = block % blocks_per_page;
1523 /* we could use find_or_create_page(), but it locks page
1524 * what we'd like to avoid in fast path ... */
1525 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1526 if (page == NULL || !PageUptodate(page)) {
1529 * drop the page reference and try
1530 * to get the page with lock. If we
1531 * are not uptodate that implies
1532 * somebody just created the page but
1533 * is yet to initialize the same. So
1534 * wait for it to initialize.
1537 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1539 BUG_ON(page->mapping != inode->i_mapping);
1540 if (!PageUptodate(page)) {
1541 ret = ext4_mb_init_cache(page, NULL, gfp);
1546 mb_cmp_bitmaps(e4b, page_address(page) +
1547 (poff * sb->s_blocksize));
1556 if (!PageUptodate(page)) {
1561 /* Pages marked accessed already */
1562 e4b->bd_bitmap_page = page;
1563 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1566 pnum = block / blocks_per_page;
1567 poff = block % blocks_per_page;
1569 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1570 if (page == NULL || !PageUptodate(page)) {
1573 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1575 BUG_ON(page->mapping != inode->i_mapping);
1576 if (!PageUptodate(page)) {
1577 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1591 if (!PageUptodate(page)) {
1596 /* Pages marked accessed already */
1597 e4b->bd_buddy_page = page;
1598 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1605 if (e4b->bd_bitmap_page)
1606 put_page(e4b->bd_bitmap_page);
1607 if (e4b->bd_buddy_page)
1608 put_page(e4b->bd_buddy_page);
1609 e4b->bd_buddy = NULL;
1610 e4b->bd_bitmap = NULL;
1614 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1615 struct ext4_buddy *e4b)
1617 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1620 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1622 if (e4b->bd_bitmap_page)
1623 put_page(e4b->bd_bitmap_page);
1624 if (e4b->bd_buddy_page)
1625 put_page(e4b->bd_buddy_page);
1629 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1634 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1635 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1637 while (order <= e4b->bd_blkbits + 1) {
1638 bb = mb_find_buddy(e4b, order, &max);
1639 if (!mb_test_bit(block >> order, bb)) {
1640 /* this block is part of buddy of order 'order' */
1648 static void mb_clear_bits(void *bm, int cur, int len)
1654 if ((cur & 31) == 0 && (len - cur) >= 32) {
1655 /* fast path: clear whole word at once */
1656 addr = bm + (cur >> 3);
1661 mb_clear_bit(cur, bm);
1666 /* clear bits in given range
1667 * will return first found zero bit if any, -1 otherwise
1669 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1676 if ((cur & 31) == 0 && (len - cur) >= 32) {
1677 /* fast path: clear whole word at once */
1678 addr = bm + (cur >> 3);
1679 if (*addr != (__u32)(-1) && zero_bit == -1)
1680 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1685 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1693 void mb_set_bits(void *bm, int cur, int len)
1699 if ((cur & 31) == 0 && (len - cur) >= 32) {
1700 /* fast path: set whole word at once */
1701 addr = bm + (cur >> 3);
1706 mb_set_bit(cur, bm);
1711 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1713 if (mb_test_bit(*bit + side, bitmap)) {
1714 mb_clear_bit(*bit, bitmap);
1720 mb_set_bit(*bit, bitmap);
1725 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1729 void *buddy = mb_find_buddy(e4b, order, &max);
1734 /* Bits in range [first; last] are known to be set since
1735 * corresponding blocks were allocated. Bits in range
1736 * (first; last) will stay set because they form buddies on
1737 * upper layer. We just deal with borders if they don't
1738 * align with upper layer and then go up.
1739 * Releasing entire group is all about clearing
1740 * single bit of highest order buddy.
1744 * ---------------------------------
1746 * ---------------------------------
1747 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1748 * ---------------------------------
1750 * \_____________________/
1752 * Neither [1] nor [6] is aligned to above layer.
1753 * Left neighbour [0] is free, so mark it busy,
1754 * decrease bb_counters and extend range to
1756 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1757 * mark [6] free, increase bb_counters and shrink range to
1759 * Then shift range to [0; 2], go up and do the same.
1764 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1766 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1771 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1772 mb_clear_bits(buddy, first, last - first + 1);
1773 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1782 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1783 int first, int count)
1785 int left_is_free = 0;
1786 int right_is_free = 0;
1788 int last = first + count - 1;
1789 struct super_block *sb = e4b->bd_sb;
1791 if (WARN_ON(count == 0))
1793 BUG_ON(last >= (sb->s_blocksize << 3));
1794 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1795 /* Don't bother if the block group is corrupt. */
1796 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1799 mb_check_buddy(e4b);
1800 mb_free_blocks_double(inode, e4b, first, count);
1802 this_cpu_inc(discard_pa_seq);
1803 e4b->bd_info->bb_free += count;
1804 if (first < e4b->bd_info->bb_first_free)
1805 e4b->bd_info->bb_first_free = first;
1807 /* access memory sequentially: check left neighbour,
1808 * clear range and then check right neighbour
1811 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1812 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1813 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1814 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1816 if (unlikely(block != -1)) {
1817 struct ext4_sb_info *sbi = EXT4_SB(sb);
1818 ext4_fsblk_t blocknr;
1820 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1821 blocknr += EXT4_C2B(sbi, block);
1822 if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) {
1823 ext4_grp_locked_error(sb, e4b->bd_group,
1824 inode ? inode->i_ino : 0,
1826 "freeing already freed block (bit %u); block bitmap corrupt.",
1828 ext4_mark_group_bitmap_corrupted(
1830 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1835 /* let's maintain fragments counter */
1836 if (left_is_free && right_is_free)
1837 e4b->bd_info->bb_fragments--;
1838 else if (!left_is_free && !right_is_free)
1839 e4b->bd_info->bb_fragments++;
1841 /* buddy[0] == bd_bitmap is a special case, so handle
1842 * it right away and let mb_buddy_mark_free stay free of
1843 * zero order checks.
1844 * Check if neighbours are to be coaleasced,
1845 * adjust bitmap bb_counters and borders appropriately.
1848 first += !left_is_free;
1849 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1852 last -= !right_is_free;
1853 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1857 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1860 mb_set_largest_free_order(sb, e4b->bd_info);
1861 mb_update_avg_fragment_size(sb, e4b->bd_info);
1862 mb_check_buddy(e4b);
1865 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1866 int needed, struct ext4_free_extent *ex)
1872 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1875 buddy = mb_find_buddy(e4b, 0, &max);
1876 BUG_ON(buddy == NULL);
1877 BUG_ON(block >= max);
1878 if (mb_test_bit(block, buddy)) {
1885 /* find actual order */
1886 order = mb_find_order_for_block(e4b, block);
1887 block = block >> order;
1889 ex->fe_len = 1 << order;
1890 ex->fe_start = block << order;
1891 ex->fe_group = e4b->bd_group;
1893 /* calc difference from given start */
1894 next = next - ex->fe_start;
1896 ex->fe_start += next;
1898 while (needed > ex->fe_len &&
1899 mb_find_buddy(e4b, order, &max)) {
1901 if (block + 1 >= max)
1904 next = (block + 1) * (1 << order);
1905 if (mb_test_bit(next, e4b->bd_bitmap))
1908 order = mb_find_order_for_block(e4b, next);
1910 block = next >> order;
1911 ex->fe_len += 1 << order;
1914 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1915 /* Should never happen! (but apparently sometimes does?!?) */
1917 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
1918 "corruption or bug in mb_find_extent "
1919 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1920 block, order, needed, ex->fe_group, ex->fe_start,
1921 ex->fe_len, ex->fe_logical);
1929 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1935 int start = ex->fe_start;
1936 int len = ex->fe_len;
1941 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1942 BUG_ON(e4b->bd_group != ex->fe_group);
1943 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1944 mb_check_buddy(e4b);
1945 mb_mark_used_double(e4b, start, len);
1947 this_cpu_inc(discard_pa_seq);
1948 e4b->bd_info->bb_free -= len;
1949 if (e4b->bd_info->bb_first_free == start)
1950 e4b->bd_info->bb_first_free += len;
1952 /* let's maintain fragments counter */
1954 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1955 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1956 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1958 e4b->bd_info->bb_fragments++;
1959 else if (!mlen && !max)
1960 e4b->bd_info->bb_fragments--;
1962 /* let's maintain buddy itself */
1964 ord = mb_find_order_for_block(e4b, start);
1966 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1967 /* the whole chunk may be allocated at once! */
1969 buddy = mb_find_buddy(e4b, ord, &max);
1970 BUG_ON((start >> ord) >= max);
1971 mb_set_bit(start >> ord, buddy);
1972 e4b->bd_info->bb_counters[ord]--;
1979 /* store for history */
1981 ret = len | (ord << 16);
1983 /* we have to split large buddy */
1985 buddy = mb_find_buddy(e4b, ord, &max);
1986 mb_set_bit(start >> ord, buddy);
1987 e4b->bd_info->bb_counters[ord]--;
1990 cur = (start >> ord) & ~1U;
1991 buddy = mb_find_buddy(e4b, ord, &max);
1992 mb_clear_bit(cur, buddy);
1993 mb_clear_bit(cur + 1, buddy);
1994 e4b->bd_info->bb_counters[ord]++;
1995 e4b->bd_info->bb_counters[ord]++;
1997 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1999 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
2000 mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
2001 mb_check_buddy(e4b);
2007 * Must be called under group lock!
2009 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
2010 struct ext4_buddy *e4b)
2012 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2015 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
2016 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2018 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
2019 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
2020 ret = mb_mark_used(e4b, &ac->ac_b_ex);
2022 /* preallocation can change ac_b_ex, thus we store actually
2023 * allocated blocks for history */
2024 ac->ac_f_ex = ac->ac_b_ex;
2026 ac->ac_status = AC_STATUS_FOUND;
2027 ac->ac_tail = ret & 0xffff;
2028 ac->ac_buddy = ret >> 16;
2031 * take the page reference. We want the page to be pinned
2032 * so that we don't get a ext4_mb_init_cache_call for this
2033 * group until we update the bitmap. That would mean we
2034 * double allocate blocks. The reference is dropped
2035 * in ext4_mb_release_context
2037 ac->ac_bitmap_page = e4b->bd_bitmap_page;
2038 get_page(ac->ac_bitmap_page);
2039 ac->ac_buddy_page = e4b->bd_buddy_page;
2040 get_page(ac->ac_buddy_page);
2041 /* store last allocated for subsequent stream allocation */
2042 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2043 spin_lock(&sbi->s_md_lock);
2044 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2045 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2046 spin_unlock(&sbi->s_md_lock);
2049 * As we've just preallocated more space than
2050 * user requested originally, we store allocated
2051 * space in a special descriptor.
2053 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2054 ext4_mb_new_preallocation(ac);
2058 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2059 struct ext4_buddy *e4b,
2062 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2063 struct ext4_free_extent *bex = &ac->ac_b_ex;
2064 struct ext4_free_extent *gex = &ac->ac_g_ex;
2065 struct ext4_free_extent ex;
2068 if (ac->ac_status == AC_STATUS_FOUND)
2071 * We don't want to scan for a whole year
2073 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2074 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2075 ac->ac_status = AC_STATUS_BREAK;
2080 * Haven't found good chunk so far, let's continue
2082 if (bex->fe_len < gex->fe_len)
2085 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2086 && bex->fe_group == e4b->bd_group) {
2087 /* recheck chunk's availability - we don't know
2088 * when it was found (within this lock-unlock
2090 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
2091 if (max >= gex->fe_len) {
2092 ext4_mb_use_best_found(ac, e4b);
2099 * The routine checks whether found extent is good enough. If it is,
2100 * then the extent gets marked used and flag is set to the context
2101 * to stop scanning. Otherwise, the extent is compared with the
2102 * previous found extent and if new one is better, then it's stored
2103 * in the context. Later, the best found extent will be used, if
2104 * mballoc can't find good enough extent.
2106 * FIXME: real allocation policy is to be designed yet!
2108 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2109 struct ext4_free_extent *ex,
2110 struct ext4_buddy *e4b)
2112 struct ext4_free_extent *bex = &ac->ac_b_ex;
2113 struct ext4_free_extent *gex = &ac->ac_g_ex;
2115 BUG_ON(ex->fe_len <= 0);
2116 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2117 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2118 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2123 * The special case - take what you catch first
2125 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2127 ext4_mb_use_best_found(ac, e4b);
2132 * Let's check whether the chuck is good enough
2134 if (ex->fe_len == gex->fe_len) {
2136 ext4_mb_use_best_found(ac, e4b);
2141 * If this is first found extent, just store it in the context
2143 if (bex->fe_len == 0) {
2149 * If new found extent is better, store it in the context
2151 if (bex->fe_len < gex->fe_len) {
2152 /* if the request isn't satisfied, any found extent
2153 * larger than previous best one is better */
2154 if (ex->fe_len > bex->fe_len)
2156 } else if (ex->fe_len > gex->fe_len) {
2157 /* if the request is satisfied, then we try to find
2158 * an extent that still satisfy the request, but is
2159 * smaller than previous one */
2160 if (ex->fe_len < bex->fe_len)
2164 ext4_mb_check_limits(ac, e4b, 0);
2167 static noinline_for_stack
2168 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2169 struct ext4_buddy *e4b)
2171 struct ext4_free_extent ex = ac->ac_b_ex;
2172 ext4_group_t group = ex.fe_group;
2176 BUG_ON(ex.fe_len <= 0);
2177 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2181 ext4_lock_group(ac->ac_sb, group);
2182 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2186 ext4_mb_use_best_found(ac, e4b);
2189 ext4_unlock_group(ac->ac_sb, group);
2190 ext4_mb_unload_buddy(e4b);
2195 static noinline_for_stack
2196 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2197 struct ext4_buddy *e4b)
2199 ext4_group_t group = ac->ac_g_ex.fe_group;
2202 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2203 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2204 struct ext4_free_extent ex;
2206 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
2208 if (grp->bb_free == 0)
2211 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2215 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
2216 ext4_mb_unload_buddy(e4b);
2220 ext4_lock_group(ac->ac_sb, group);
2221 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2222 ac->ac_g_ex.fe_len, &ex);
2223 ex.fe_logical = 0xDEADFA11; /* debug value */
2225 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
2228 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
2230 /* use do_div to get remainder (would be 64-bit modulo) */
2231 if (do_div(start, sbi->s_stripe) == 0) {
2234 ext4_mb_use_best_found(ac, e4b);
2236 } else if (max >= ac->ac_g_ex.fe_len) {
2237 BUG_ON(ex.fe_len <= 0);
2238 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2239 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2242 ext4_mb_use_best_found(ac, e4b);
2243 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2244 /* Sometimes, caller may want to merge even small
2245 * number of blocks to an existing extent */
2246 BUG_ON(ex.fe_len <= 0);
2247 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2248 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2251 ext4_mb_use_best_found(ac, e4b);
2253 ext4_unlock_group(ac->ac_sb, group);
2254 ext4_mb_unload_buddy(e4b);
2260 * The routine scans buddy structures (not bitmap!) from given order
2261 * to max order and tries to find big enough chunk to satisfy the req
2263 static noinline_for_stack
2264 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2265 struct ext4_buddy *e4b)
2267 struct super_block *sb = ac->ac_sb;
2268 struct ext4_group_info *grp = e4b->bd_info;
2274 BUG_ON(ac->ac_2order <= 0);
2275 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2276 if (grp->bb_counters[i] == 0)
2279 buddy = mb_find_buddy(e4b, i, &max);
2280 BUG_ON(buddy == NULL);
2282 k = mb_find_next_zero_bit(buddy, max, 0);
2284 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2285 "%d free clusters of order %d. But found 0",
2286 grp->bb_counters[i], i);
2287 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2289 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2294 ac->ac_b_ex.fe_len = 1 << i;
2295 ac->ac_b_ex.fe_start = k << i;
2296 ac->ac_b_ex.fe_group = e4b->bd_group;
2298 ext4_mb_use_best_found(ac, e4b);
2300 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2302 if (EXT4_SB(sb)->s_mb_stats)
2303 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2310 * The routine scans the group and measures all found extents.
2311 * In order to optimize scanning, caller must pass number of
2312 * free blocks in the group, so the routine can know upper limit.
2314 static noinline_for_stack
2315 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2316 struct ext4_buddy *e4b)
2318 struct super_block *sb = ac->ac_sb;
2319 void *bitmap = e4b->bd_bitmap;
2320 struct ext4_free_extent ex;
2324 free = e4b->bd_info->bb_free;
2325 if (WARN_ON(free <= 0))
2328 i = e4b->bd_info->bb_first_free;
2330 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2331 i = mb_find_next_zero_bit(bitmap,
2332 EXT4_CLUSTERS_PER_GROUP(sb), i);
2333 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2335 * IF we have corrupt bitmap, we won't find any
2336 * free blocks even though group info says we
2339 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2340 "%d free clusters as per "
2341 "group info. But bitmap says 0",
2343 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2344 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2348 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2349 if (WARN_ON(ex.fe_len <= 0))
2351 if (free < ex.fe_len) {
2352 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2353 "%d free clusters as per "
2354 "group info. But got %d blocks",
2356 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2357 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2359 * The number of free blocks differs. This mostly
2360 * indicate that the bitmap is corrupt. So exit
2361 * without claiming the space.
2365 ex.fe_logical = 0xDEADC0DE; /* debug value */
2366 ext4_mb_measure_extent(ac, &ex, e4b);
2372 ext4_mb_check_limits(ac, e4b, 1);
2376 * This is a special case for storages like raid5
2377 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2379 static noinline_for_stack
2380 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2381 struct ext4_buddy *e4b)
2383 struct super_block *sb = ac->ac_sb;
2384 struct ext4_sb_info *sbi = EXT4_SB(sb);
2385 void *bitmap = e4b->bd_bitmap;
2386 struct ext4_free_extent ex;
2387 ext4_fsblk_t first_group_block;
2392 BUG_ON(sbi->s_stripe == 0);
2394 /* find first stripe-aligned block in group */
2395 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2397 a = first_group_block + sbi->s_stripe - 1;
2398 do_div(a, sbi->s_stripe);
2399 i = (a * sbi->s_stripe) - first_group_block;
2401 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2402 if (!mb_test_bit(i, bitmap)) {
2403 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2404 if (max >= sbi->s_stripe) {
2406 ex.fe_logical = 0xDEADF00D; /* debug value */
2408 ext4_mb_use_best_found(ac, e4b);
2417 * This is also called BEFORE we load the buddy bitmap.
2418 * Returns either 1 or 0 indicating that the group is either suitable
2419 * for the allocation or not.
2421 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2422 ext4_group_t group, int cr)
2424 ext4_grpblk_t free, fragments;
2425 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2426 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2428 BUG_ON(cr < 0 || cr >= 4);
2430 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2433 free = grp->bb_free;
2437 fragments = grp->bb_fragments;
2443 BUG_ON(ac->ac_2order == 0);
2445 /* Avoid using the first bg of a flexgroup for data files */
2446 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2447 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2448 ((group % flex_size) == 0))
2451 if (free < ac->ac_g_ex.fe_len)
2454 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2457 if (grp->bb_largest_free_order < ac->ac_2order)
2462 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2466 if (free >= ac->ac_g_ex.fe_len)
2479 * This could return negative error code if something goes wrong
2480 * during ext4_mb_init_group(). This should not be called with
2481 * ext4_lock_group() held.
2483 * Note: because we are conditionally operating with the group lock in
2484 * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this
2485 * function using __acquire and __release. This means we need to be
2486 * super careful before messing with the error path handling via "goto
2489 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2490 ext4_group_t group, int cr)
2492 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2493 struct super_block *sb = ac->ac_sb;
2494 struct ext4_sb_info *sbi = EXT4_SB(sb);
2495 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2499 if (sbi->s_mb_stats)
2500 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2502 ext4_lock_group(sb, group);
2503 __release(ext4_group_lock_ptr(sb, group));
2505 free = grp->bb_free;
2508 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2510 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2513 __acquire(ext4_group_lock_ptr(sb, group));
2514 ext4_unlock_group(sb, group);
2517 /* We only do this if the grp has never been initialized */
2518 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2519 struct ext4_group_desc *gdp =
2520 ext4_get_group_desc(sb, group, NULL);
2523 /* cr=0/1 is a very optimistic search to find large
2524 * good chunks almost for free. If buddy data is not
2525 * ready, then this optimization makes no sense. But
2526 * we never skip the first block group in a flex_bg,
2527 * since this gets used for metadata block allocation,
2528 * and we want to make sure we locate metadata blocks
2529 * in the first block group in the flex_bg if possible.
2532 (!sbi->s_log_groups_per_flex ||
2533 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2534 !(ext4_has_group_desc_csum(sb) &&
2535 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2537 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2543 ext4_lock_group(sb, group);
2544 __release(ext4_group_lock_ptr(sb, group));
2546 ret = ext4_mb_good_group(ac, group, cr);
2549 __acquire(ext4_group_lock_ptr(sb, group));
2550 ext4_unlock_group(sb, group);
2556 * Start prefetching @nr block bitmaps starting at @group.
2557 * Return the next group which needs to be prefetched.
2559 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2560 unsigned int nr, int *cnt)
2562 ext4_group_t ngroups = ext4_get_groups_count(sb);
2563 struct buffer_head *bh;
2564 struct blk_plug plug;
2566 blk_start_plug(&plug);
2568 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2570 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2573 * Prefetch block groups with free blocks; but don't
2574 * bother if it is marked uninitialized on disk, since
2575 * it won't require I/O to read. Also only try to
2576 * prefetch once, so we avoid getblk() call, which can
2579 if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2580 EXT4_MB_GRP_NEED_INIT(grp) &&
2581 ext4_free_group_clusters(sb, gdp) > 0 &&
2582 !(ext4_has_group_desc_csum(sb) &&
2583 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2584 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2585 if (bh && !IS_ERR(bh)) {
2586 if (!buffer_uptodate(bh) && cnt)
2591 if (++group >= ngroups)
2594 blk_finish_plug(&plug);
2599 * Prefetching reads the block bitmap into the buffer cache; but we
2600 * need to make sure that the buddy bitmap in the page cache has been
2601 * initialized. Note that ext4_mb_init_group() will block if the I/O
2602 * is not yet completed, or indeed if it was not initiated by
2603 * ext4_mb_prefetch did not start the I/O.
2605 * TODO: We should actually kick off the buddy bitmap setup in a work
2606 * queue when the buffer I/O is completed, so that we don't block
2607 * waiting for the block allocation bitmap read to finish when
2608 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2610 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2614 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2616 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2619 group = ext4_get_groups_count(sb);
2621 grp = ext4_get_group_info(sb, group);
2623 if (EXT4_MB_GRP_NEED_INIT(grp) &&
2624 ext4_free_group_clusters(sb, gdp) > 0 &&
2625 !(ext4_has_group_desc_csum(sb) &&
2626 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2627 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2633 static noinline_for_stack int
2634 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2636 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2637 int cr = -1, new_cr;
2638 int err = 0, first_err = 0;
2639 unsigned int nr = 0, prefetch_ios = 0;
2640 struct ext4_sb_info *sbi;
2641 struct super_block *sb;
2642 struct ext4_buddy e4b;
2647 ngroups = ext4_get_groups_count(sb);
2648 /* non-extent files are limited to low blocks/groups */
2649 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2650 ngroups = sbi->s_blockfile_groups;
2652 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2654 /* first, try the goal */
2655 err = ext4_mb_find_by_goal(ac, &e4b);
2656 if (err || ac->ac_status == AC_STATUS_FOUND)
2659 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2663 * ac->ac_2order is set only if the fe_len is a power of 2
2664 * if ac->ac_2order is set we also set criteria to 0 so that we
2665 * try exact allocation using buddy.
2667 i = fls(ac->ac_g_ex.fe_len);
2670 * We search using buddy data only if the order of the request
2671 * is greater than equal to the sbi_s_mb_order2_reqs
2672 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2673 * We also support searching for power-of-two requests only for
2674 * requests upto maximum buddy size we have constructed.
2676 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2678 * This should tell if fe_len is exactly power of 2
2680 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2681 ac->ac_2order = array_index_nospec(i - 1,
2685 /* if stream allocation is enabled, use global goal */
2686 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2687 /* TBD: may be hot point */
2688 spin_lock(&sbi->s_md_lock);
2689 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2690 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2691 spin_unlock(&sbi->s_md_lock);
2694 /* Let's just scan groups to find more-less suitable blocks */
2695 cr = ac->ac_2order ? 0 : 1;
2697 * cr == 0 try to get exact allocation,
2698 * cr == 3 try to get anything
2701 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2702 ac->ac_criteria = cr;
2704 * searching for the right group start
2705 * from the goal value specified
2707 group = ac->ac_g_ex.fe_group;
2708 ac->ac_last_optimal_group = group;
2709 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2710 prefetch_grp = group;
2712 for (i = 0, new_cr = cr; i < ngroups; i++,
2713 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) {
2723 * Batch reads of the block allocation bitmaps
2724 * to get multiple READs in flight; limit
2725 * prefetching at cr=0/1, otherwise mballoc can
2726 * spend a lot of time loading imperfect groups
2728 if ((prefetch_grp == group) &&
2730 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2731 unsigned int curr_ios = prefetch_ios;
2733 nr = sbi->s_mb_prefetch;
2734 if (ext4_has_feature_flex_bg(sb)) {
2735 nr = 1 << sbi->s_log_groups_per_flex;
2736 nr -= group & (nr - 1);
2737 nr = min(nr, sbi->s_mb_prefetch);
2739 prefetch_grp = ext4_mb_prefetch(sb, group,
2741 if (prefetch_ios == curr_ios)
2745 /* This now checks without needing the buddy page */
2746 ret = ext4_mb_good_group_nolock(ac, group, cr);
2753 err = ext4_mb_load_buddy(sb, group, &e4b);
2757 ext4_lock_group(sb, group);
2760 * We need to check again after locking the
2763 ret = ext4_mb_good_group(ac, group, cr);
2765 ext4_unlock_group(sb, group);
2766 ext4_mb_unload_buddy(&e4b);
2770 ac->ac_groups_scanned++;
2772 ext4_mb_simple_scan_group(ac, &e4b);
2773 else if (cr == 1 && sbi->s_stripe &&
2774 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2775 ext4_mb_scan_aligned(ac, &e4b);
2777 ext4_mb_complex_scan_group(ac, &e4b);
2779 ext4_unlock_group(sb, group);
2780 ext4_mb_unload_buddy(&e4b);
2782 if (ac->ac_status != AC_STATUS_CONTINUE)
2785 /* Processed all groups and haven't found blocks */
2786 if (sbi->s_mb_stats && i == ngroups)
2787 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2790 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2791 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2793 * We've been searching too long. Let's try to allocate
2794 * the best chunk we've found so far
2796 ext4_mb_try_best_found(ac, &e4b);
2797 if (ac->ac_status != AC_STATUS_FOUND) {
2799 * Someone more lucky has already allocated it.
2800 * The only thing we can do is just take first
2803 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2804 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2805 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2806 ac->ac_b_ex.fe_len, lost);
2808 ac->ac_b_ex.fe_group = 0;
2809 ac->ac_b_ex.fe_start = 0;
2810 ac->ac_b_ex.fe_len = 0;
2811 ac->ac_status = AC_STATUS_CONTINUE;
2812 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2818 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2819 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2821 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2824 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2825 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2826 ac->ac_flags, cr, err);
2829 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2834 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2836 struct super_block *sb = pde_data(file_inode(seq->file));
2839 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2842 return (void *) ((unsigned long) group);
2845 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2847 struct super_block *sb = pde_data(file_inode(seq->file));
2851 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2854 return (void *) ((unsigned long) group);
2857 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2859 struct super_block *sb = pde_data(file_inode(seq->file));
2860 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2862 int err, buddy_loaded = 0;
2863 struct ext4_buddy e4b;
2864 struct ext4_group_info *grinfo;
2865 unsigned char blocksize_bits = min_t(unsigned char,
2866 sb->s_blocksize_bits,
2867 EXT4_MAX_BLOCK_LOG_SIZE);
2869 struct ext4_group_info info;
2870 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2875 seq_puts(seq, "#group: free frags first ["
2876 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2877 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2879 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2880 sizeof(struct ext4_group_info);
2882 grinfo = ext4_get_group_info(sb, group);
2883 /* Load the group info in memory only if not already loaded. */
2884 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2885 err = ext4_mb_load_buddy(sb, group, &e4b);
2887 seq_printf(seq, "#%-5u: I/O error\n", group);
2893 memcpy(&sg, ext4_get_group_info(sb, group), i);
2896 ext4_mb_unload_buddy(&e4b);
2898 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2899 sg.info.bb_fragments, sg.info.bb_first_free);
2900 for (i = 0; i <= 13; i++)
2901 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2902 sg.info.bb_counters[i] : 0);
2903 seq_puts(seq, " ]\n");
2908 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2912 const struct seq_operations ext4_mb_seq_groups_ops = {
2913 .start = ext4_mb_seq_groups_start,
2914 .next = ext4_mb_seq_groups_next,
2915 .stop = ext4_mb_seq_groups_stop,
2916 .show = ext4_mb_seq_groups_show,
2919 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
2921 struct super_block *sb = seq->private;
2922 struct ext4_sb_info *sbi = EXT4_SB(sb);
2924 seq_puts(seq, "mballoc:\n");
2925 if (!sbi->s_mb_stats) {
2926 seq_puts(seq, "\tmb stats collection turned off.\n");
2927 seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
2930 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
2931 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
2933 seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned));
2935 seq_puts(seq, "\tcr0_stats:\n");
2936 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0]));
2937 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2938 atomic64_read(&sbi->s_bal_cX_groups_considered[0]));
2939 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2940 atomic64_read(&sbi->s_bal_cX_failed[0]));
2941 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2942 atomic_read(&sbi->s_bal_cr0_bad_suggestions));
2944 seq_puts(seq, "\tcr1_stats:\n");
2945 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1]));
2946 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2947 atomic64_read(&sbi->s_bal_cX_groups_considered[1]));
2948 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2949 atomic64_read(&sbi->s_bal_cX_failed[1]));
2950 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2951 atomic_read(&sbi->s_bal_cr1_bad_suggestions));
2953 seq_puts(seq, "\tcr2_stats:\n");
2954 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2]));
2955 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2956 atomic64_read(&sbi->s_bal_cX_groups_considered[2]));
2957 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2958 atomic64_read(&sbi->s_bal_cX_failed[2]));
2960 seq_puts(seq, "\tcr3_stats:\n");
2961 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3]));
2962 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2963 atomic64_read(&sbi->s_bal_cX_groups_considered[3]));
2964 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2965 atomic64_read(&sbi->s_bal_cX_failed[3]));
2966 seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned));
2967 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
2968 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
2969 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
2970 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
2972 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
2973 atomic_read(&sbi->s_mb_buddies_generated),
2974 ext4_get_groups_count(sb));
2975 seq_printf(seq, "\tbuddies_time_used: %llu\n",
2976 atomic64_read(&sbi->s_mb_generation_time));
2977 seq_printf(seq, "\tpreallocated: %u\n",
2978 atomic_read(&sbi->s_mb_preallocated));
2979 seq_printf(seq, "\tdiscarded: %u\n",
2980 atomic_read(&sbi->s_mb_discarded));
2984 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
2985 __acquires(&EXT4_SB(sb)->s_mb_rb_lock)
2987 struct super_block *sb = pde_data(file_inode(seq->file));
2988 unsigned long position;
2990 read_lock(&EXT4_SB(sb)->s_mb_rb_lock);
2992 if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
2994 position = *pos + 1;
2995 return (void *) ((unsigned long) position);
2998 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
3000 struct super_block *sb = pde_data(file_inode(seq->file));
3001 unsigned long position;
3004 if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
3006 position = *pos + 1;
3007 return (void *) ((unsigned long) position);
3010 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
3012 struct super_block *sb = pde_data(file_inode(seq->file));
3013 struct ext4_sb_info *sbi = EXT4_SB(sb);
3014 unsigned long position = ((unsigned long) v);
3015 struct ext4_group_info *grp;
3017 unsigned int count, min, max;
3020 if (position >= MB_NUM_ORDERS(sb)) {
3021 seq_puts(seq, "fragment_size_tree:\n");
3022 n = rb_first(&sbi->s_mb_avg_fragment_size_root);
3024 seq_puts(seq, "\ttree_min: 0\n\ttree_max: 0\n\ttree_nodes: 0\n");
3027 grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3028 min = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3030 while (rb_next(n)) {
3034 grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3035 max = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3037 seq_printf(seq, "\ttree_min: %u\n\ttree_max: %u\n\ttree_nodes: %u\n",
3042 if (position == 0) {
3043 seq_printf(seq, "optimize_scan: %d\n",
3044 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
3045 seq_puts(seq, "max_free_order_lists:\n");
3048 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
3049 bb_largest_free_order_node)
3051 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3052 (unsigned int)position, count);
3057 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3058 __releases(&EXT4_SB(sb)->s_mb_rb_lock)
3060 struct super_block *sb = pde_data(file_inode(seq->file));
3062 read_unlock(&EXT4_SB(sb)->s_mb_rb_lock);
3065 const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3066 .start = ext4_mb_seq_structs_summary_start,
3067 .next = ext4_mb_seq_structs_summary_next,
3068 .stop = ext4_mb_seq_structs_summary_stop,
3069 .show = ext4_mb_seq_structs_summary_show,
3072 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3074 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3075 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3082 * Allocate the top-level s_group_info array for the specified number
3085 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3087 struct ext4_sb_info *sbi = EXT4_SB(sb);
3089 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3091 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3092 EXT4_DESC_PER_BLOCK_BITS(sb);
3093 if (size <= sbi->s_group_info_size)
3096 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3097 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3098 if (!new_groupinfo) {
3099 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3103 old_groupinfo = rcu_dereference(sbi->s_group_info);
3105 memcpy(new_groupinfo, old_groupinfo,
3106 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3108 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3109 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3111 ext4_kvfree_array_rcu(old_groupinfo);
3112 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3113 sbi->s_group_info_size);
3117 /* Create and initialize ext4_group_info data for the given group. */
3118 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3119 struct ext4_group_desc *desc)
3123 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3124 struct ext4_sb_info *sbi = EXT4_SB(sb);
3125 struct ext4_group_info **meta_group_info;
3126 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3129 * First check if this group is the first of a reserved block.
3130 * If it's true, we have to allocate a new table of pointers
3131 * to ext4_group_info structures
3133 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3134 metalen = sizeof(*meta_group_info) <<
3135 EXT4_DESC_PER_BLOCK_BITS(sb);
3136 meta_group_info = kmalloc(metalen, GFP_NOFS);
3137 if (meta_group_info == NULL) {
3138 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3139 "for a buddy group");
3140 goto exit_meta_group_info;
3143 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3147 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3148 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3150 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3151 if (meta_group_info[i] == NULL) {
3152 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3153 goto exit_group_info;
3155 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3156 &(meta_group_info[i]->bb_state));
3159 * initialize bb_free to be able to skip
3160 * empty groups without initialization
3162 if (ext4_has_group_desc_csum(sb) &&
3163 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3164 meta_group_info[i]->bb_free =
3165 ext4_free_clusters_after_init(sb, group, desc);
3167 meta_group_info[i]->bb_free =
3168 ext4_free_group_clusters(sb, desc);
3171 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3172 init_rwsem(&meta_group_info[i]->alloc_sem);
3173 meta_group_info[i]->bb_free_root = RB_ROOT;
3174 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3175 RB_CLEAR_NODE(&meta_group_info[i]->bb_avg_fragment_size_rb);
3176 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3177 meta_group_info[i]->bb_group = group;
3179 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3183 /* If a meta_group_info table has been allocated, release it now */
3184 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3185 struct ext4_group_info ***group_info;
3188 group_info = rcu_dereference(sbi->s_group_info);
3189 kfree(group_info[idx]);
3190 group_info[idx] = NULL;
3193 exit_meta_group_info:
3195 } /* ext4_mb_add_groupinfo */
3197 static int ext4_mb_init_backend(struct super_block *sb)
3199 ext4_group_t ngroups = ext4_get_groups_count(sb);
3201 struct ext4_sb_info *sbi = EXT4_SB(sb);
3203 struct ext4_group_desc *desc;
3204 struct ext4_group_info ***group_info;
3205 struct kmem_cache *cachep;
3207 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3211 sbi->s_buddy_cache = new_inode(sb);
3212 if (sbi->s_buddy_cache == NULL) {
3213 ext4_msg(sb, KERN_ERR, "can't get new inode");
3216 /* To avoid potentially colliding with an valid on-disk inode number,
3217 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3218 * not in the inode hash, so it should never be found by iget(), but
3219 * this will avoid confusion if it ever shows up during debugging. */
3220 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3221 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3222 for (i = 0; i < ngroups; i++) {
3224 desc = ext4_get_group_desc(sb, i, NULL);
3226 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3229 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3233 if (ext4_has_feature_flex_bg(sb)) {
3234 /* a single flex group is supposed to be read by a single IO.
3235 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3236 * unsigned integer, so the maximum shift is 32.
3238 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3239 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3242 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3243 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3244 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3246 sbi->s_mb_prefetch = 32;
3248 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3249 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3250 /* now many real IOs to prefetch within a single allocation at cr=0
3251 * given cr=0 is an CPU-related optimization we shouldn't try to
3252 * load too many groups, at some point we should start to use what
3253 * we've got in memory.
3254 * with an average random access time 5ms, it'd take a second to get
3255 * 200 groups (* N with flex_bg), so let's make this limit 4
3257 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3258 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3259 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3264 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3266 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
3267 i = sbi->s_group_info_size;
3269 group_info = rcu_dereference(sbi->s_group_info);
3271 kfree(group_info[i]);
3273 iput(sbi->s_buddy_cache);
3276 kvfree(rcu_dereference(sbi->s_group_info));
3281 static void ext4_groupinfo_destroy_slabs(void)
3285 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3286 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3287 ext4_groupinfo_caches[i] = NULL;
3291 static int ext4_groupinfo_create_slab(size_t size)
3293 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3295 int blocksize_bits = order_base_2(size);
3296 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3297 struct kmem_cache *cachep;
3299 if (cache_index >= NR_GRPINFO_CACHES)
3302 if (unlikely(cache_index < 0))
3305 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3306 if (ext4_groupinfo_caches[cache_index]) {
3307 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3308 return 0; /* Already created */
3311 slab_size = offsetof(struct ext4_group_info,
3312 bb_counters[blocksize_bits + 2]);
3314 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3315 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3318 ext4_groupinfo_caches[cache_index] = cachep;
3320 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3323 "EXT4-fs: no memory for groupinfo slab cache\n");
3330 static void ext4_discard_work(struct work_struct *work)
3332 struct ext4_sb_info *sbi = container_of(work,
3333 struct ext4_sb_info, s_discard_work);
3334 struct super_block *sb = sbi->s_sb;
3335 struct ext4_free_data *fd, *nfd;
3336 struct ext4_buddy e4b;
3337 struct list_head discard_list;
3338 ext4_group_t grp, load_grp;
3341 INIT_LIST_HEAD(&discard_list);
3342 spin_lock(&sbi->s_md_lock);
3343 list_splice_init(&sbi->s_discard_list, &discard_list);
3344 spin_unlock(&sbi->s_md_lock);
3346 load_grp = UINT_MAX;
3347 list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
3349 * If filesystem is umounting or no memory or suffering
3350 * from no space, give up the discard
3352 if ((sb->s_flags & SB_ACTIVE) && !err &&
3353 !atomic_read(&sbi->s_retry_alloc_pending)) {
3354 grp = fd->efd_group;
3355 if (grp != load_grp) {
3356 if (load_grp != UINT_MAX)
3357 ext4_mb_unload_buddy(&e4b);
3359 err = ext4_mb_load_buddy(sb, grp, &e4b);
3361 kmem_cache_free(ext4_free_data_cachep, fd);
3362 load_grp = UINT_MAX;
3369 ext4_lock_group(sb, grp);
3370 ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
3371 fd->efd_start_cluster + fd->efd_count - 1, 1);
3372 ext4_unlock_group(sb, grp);
3374 kmem_cache_free(ext4_free_data_cachep, fd);
3377 if (load_grp != UINT_MAX)
3378 ext4_mb_unload_buddy(&e4b);
3381 int ext4_mb_init(struct super_block *sb)
3383 struct ext4_sb_info *sbi = EXT4_SB(sb);
3385 unsigned offset, offset_incr;
3389 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3391 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3392 if (sbi->s_mb_offsets == NULL) {
3397 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3398 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3399 if (sbi->s_mb_maxs == NULL) {
3404 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3408 /* order 0 is regular bitmap */
3409 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3410 sbi->s_mb_offsets[0] = 0;
3414 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3415 max = sb->s_blocksize << 2;
3417 sbi->s_mb_offsets[i] = offset;
3418 sbi->s_mb_maxs[i] = max;
3419 offset += offset_incr;
3420 offset_incr = offset_incr >> 1;
3423 } while (i < MB_NUM_ORDERS(sb));
3425 sbi->s_mb_avg_fragment_size_root = RB_ROOT;
3426 sbi->s_mb_largest_free_orders =
3427 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3429 if (!sbi->s_mb_largest_free_orders) {
3433 sbi->s_mb_largest_free_orders_locks =
3434 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3436 if (!sbi->s_mb_largest_free_orders_locks) {
3440 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3441 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3442 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3444 rwlock_init(&sbi->s_mb_rb_lock);
3446 spin_lock_init(&sbi->s_md_lock);
3447 sbi->s_mb_free_pending = 0;
3448 INIT_LIST_HEAD(&sbi->s_freed_data_list);
3449 INIT_LIST_HEAD(&sbi->s_discard_list);
3450 INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
3451 atomic_set(&sbi->s_retry_alloc_pending, 0);
3453 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3454 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3455 sbi->s_mb_stats = MB_DEFAULT_STATS;
3456 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3457 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3458 sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
3460 * The default group preallocation is 512, which for 4k block
3461 * sizes translates to 2 megabytes. However for bigalloc file
3462 * systems, this is probably too big (i.e, if the cluster size
3463 * is 1 megabyte, then group preallocation size becomes half a
3464 * gigabyte!). As a default, we will keep a two megabyte
3465 * group pralloc size for cluster sizes up to 64k, and after
3466 * that, we will force a minimum group preallocation size of
3467 * 32 clusters. This translates to 8 megs when the cluster
3468 * size is 256k, and 32 megs when the cluster size is 1 meg,
3469 * which seems reasonable as a default.
3471 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3472 sbi->s_cluster_bits, 32);
3474 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3475 * to the lowest multiple of s_stripe which is bigger than
3476 * the s_mb_group_prealloc as determined above. We want
3477 * the preallocation size to be an exact multiple of the
3478 * RAID stripe size so that preallocations don't fragment
3481 if (sbi->s_stripe > 1) {
3482 sbi->s_mb_group_prealloc = roundup(
3483 sbi->s_mb_group_prealloc, sbi->s_stripe);
3486 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3487 if (sbi->s_locality_groups == NULL) {
3491 for_each_possible_cpu(i) {
3492 struct ext4_locality_group *lg;
3493 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3494 mutex_init(&lg->lg_mutex);
3495 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3496 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3497 spin_lock_init(&lg->lg_prealloc_lock);
3500 if (bdev_nonrot(sb->s_bdev))
3501 sbi->s_mb_max_linear_groups = 0;
3503 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3504 /* init file for buddy data */
3505 ret = ext4_mb_init_backend(sb);
3507 goto out_free_locality_groups;
3511 out_free_locality_groups:
3512 free_percpu(sbi->s_locality_groups);
3513 sbi->s_locality_groups = NULL;
3515 kfree(sbi->s_mb_largest_free_orders);
3516 kfree(sbi->s_mb_largest_free_orders_locks);
3517 kfree(sbi->s_mb_offsets);
3518 sbi->s_mb_offsets = NULL;
3519 kfree(sbi->s_mb_maxs);
3520 sbi->s_mb_maxs = NULL;
3524 /* need to called with the ext4 group lock held */
3525 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3527 struct ext4_prealloc_space *pa;
3528 struct list_head *cur, *tmp;
3531 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3532 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3533 list_del(&pa->pa_group_list);
3535 kmem_cache_free(ext4_pspace_cachep, pa);
3540 int ext4_mb_release(struct super_block *sb)
3542 ext4_group_t ngroups = ext4_get_groups_count(sb);
3544 int num_meta_group_infos;
3545 struct ext4_group_info *grinfo, ***group_info;
3546 struct ext4_sb_info *sbi = EXT4_SB(sb);
3547 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3550 if (test_opt(sb, DISCARD)) {
3552 * wait the discard work to drain all of ext4_free_data
3554 flush_work(&sbi->s_discard_work);
3555 WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
3558 if (sbi->s_group_info) {
3559 for (i = 0; i < ngroups; i++) {
3561 grinfo = ext4_get_group_info(sb, i);
3562 mb_group_bb_bitmap_free(grinfo);
3563 ext4_lock_group(sb, i);
3564 count = ext4_mb_cleanup_pa(grinfo);
3566 mb_debug(sb, "mballoc: %d PAs left\n",
3568 ext4_unlock_group(sb, i);
3569 kmem_cache_free(cachep, grinfo);
3571 num_meta_group_infos = (ngroups +
3572 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3573 EXT4_DESC_PER_BLOCK_BITS(sb);
3575 group_info = rcu_dereference(sbi->s_group_info);
3576 for (i = 0; i < num_meta_group_infos; i++)
3577 kfree(group_info[i]);
3581 kfree(sbi->s_mb_largest_free_orders);
3582 kfree(sbi->s_mb_largest_free_orders_locks);
3583 kfree(sbi->s_mb_offsets);
3584 kfree(sbi->s_mb_maxs);
3585 iput(sbi->s_buddy_cache);
3586 if (sbi->s_mb_stats) {
3587 ext4_msg(sb, KERN_INFO,
3588 "mballoc: %u blocks %u reqs (%u success)",
3589 atomic_read(&sbi->s_bal_allocated),
3590 atomic_read(&sbi->s_bal_reqs),
3591 atomic_read(&sbi->s_bal_success));
3592 ext4_msg(sb, KERN_INFO,
3593 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3594 "%u 2^N hits, %u breaks, %u lost",
3595 atomic_read(&sbi->s_bal_ex_scanned),
3596 atomic_read(&sbi->s_bal_groups_scanned),
3597 atomic_read(&sbi->s_bal_goals),
3598 atomic_read(&sbi->s_bal_2orders),
3599 atomic_read(&sbi->s_bal_breaks),
3600 atomic_read(&sbi->s_mb_lost_chunks));
3601 ext4_msg(sb, KERN_INFO,
3602 "mballoc: %u generated and it took %llu",
3603 atomic_read(&sbi->s_mb_buddies_generated),
3604 atomic64_read(&sbi->s_mb_generation_time));
3605 ext4_msg(sb, KERN_INFO,
3606 "mballoc: %u preallocated, %u discarded",
3607 atomic_read(&sbi->s_mb_preallocated),
3608 atomic_read(&sbi->s_mb_discarded));
3611 free_percpu(sbi->s_locality_groups);
3616 static inline int ext4_issue_discard(struct super_block *sb,
3617 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3620 ext4_fsblk_t discard_block;
3622 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3623 ext4_group_first_block_no(sb, block_group));
3624 count = EXT4_C2B(EXT4_SB(sb), count);
3625 trace_ext4_discard_blocks(sb,
3626 (unsigned long long) discard_block, count);
3628 return __blkdev_issue_discard(sb->s_bdev,
3629 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3630 (sector_t)count << (sb->s_blocksize_bits - 9),
3633 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3636 static void ext4_free_data_in_buddy(struct super_block *sb,
3637 struct ext4_free_data *entry)
3639 struct ext4_buddy e4b;
3640 struct ext4_group_info *db;
3641 int err, count = 0, count2 = 0;
3643 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3644 entry->efd_count, entry->efd_group, entry);
3646 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3647 /* we expect to find existing buddy because it's pinned */
3650 spin_lock(&EXT4_SB(sb)->s_md_lock);
3651 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3652 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3655 /* there are blocks to put in buddy to make them really free */
3656 count += entry->efd_count;
3658 ext4_lock_group(sb, entry->efd_group);
3659 /* Take it out of per group rb tree */
3660 rb_erase(&entry->efd_node, &(db->bb_free_root));
3661 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3664 * Clear the trimmed flag for the group so that the next
3665 * ext4_trim_fs can trim it.
3666 * If the volume is mounted with -o discard, online discard
3667 * is supported and the free blocks will be trimmed online.
3669 if (!test_opt(sb, DISCARD))
3670 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3672 if (!db->bb_free_root.rb_node) {
3673 /* No more items in the per group rb tree
3674 * balance refcounts from ext4_mb_free_metadata()
3676 put_page(e4b.bd_buddy_page);
3677 put_page(e4b.bd_bitmap_page);
3679 ext4_unlock_group(sb, entry->efd_group);
3680 ext4_mb_unload_buddy(&e4b);
3682 mb_debug(sb, "freed %d blocks in %d structures\n", count,
3687 * This function is called by the jbd2 layer once the commit has finished,
3688 * so we know we can free the blocks that were released with that commit.
3690 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3692 struct ext4_sb_info *sbi = EXT4_SB(sb);
3693 struct ext4_free_data *entry, *tmp;
3694 struct list_head freed_data_list;
3695 struct list_head *cut_pos = NULL;
3698 INIT_LIST_HEAD(&freed_data_list);
3700 spin_lock(&sbi->s_md_lock);
3701 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3702 if (entry->efd_tid != commit_tid)
3704 cut_pos = &entry->efd_list;
3707 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3709 spin_unlock(&sbi->s_md_lock);
3711 list_for_each_entry(entry, &freed_data_list, efd_list)
3712 ext4_free_data_in_buddy(sb, entry);
3714 if (test_opt(sb, DISCARD)) {
3715 spin_lock(&sbi->s_md_lock);
3716 wake = list_empty(&sbi->s_discard_list);
3717 list_splice_tail(&freed_data_list, &sbi->s_discard_list);
3718 spin_unlock(&sbi->s_md_lock);
3720 queue_work(system_unbound_wq, &sbi->s_discard_work);
3722 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3723 kmem_cache_free(ext4_free_data_cachep, entry);
3727 int __init ext4_init_mballoc(void)
3729 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3730 SLAB_RECLAIM_ACCOUNT);
3731 if (ext4_pspace_cachep == NULL)
3734 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3735 SLAB_RECLAIM_ACCOUNT);
3736 if (ext4_ac_cachep == NULL)
3739 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3740 SLAB_RECLAIM_ACCOUNT);
3741 if (ext4_free_data_cachep == NULL)
3747 kmem_cache_destroy(ext4_ac_cachep);
3749 kmem_cache_destroy(ext4_pspace_cachep);
3754 void ext4_exit_mballoc(void)
3757 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3758 * before destroying the slab cache.
3761 kmem_cache_destroy(ext4_pspace_cachep);
3762 kmem_cache_destroy(ext4_ac_cachep);
3763 kmem_cache_destroy(ext4_free_data_cachep);
3764 ext4_groupinfo_destroy_slabs();
3769 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3770 * Returns 0 if success or error code
3772 static noinline_for_stack int
3773 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3774 handle_t *handle, unsigned int reserv_clstrs)
3776 struct buffer_head *bitmap_bh = NULL;
3777 struct ext4_group_desc *gdp;
3778 struct buffer_head *gdp_bh;
3779 struct ext4_sb_info *sbi;
3780 struct super_block *sb;
3784 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3785 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3790 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3791 if (IS_ERR(bitmap_bh)) {
3792 err = PTR_ERR(bitmap_bh);
3797 BUFFER_TRACE(bitmap_bh, "getting write access");
3798 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
3804 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3808 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3809 ext4_free_group_clusters(sb, gdp));
3811 BUFFER_TRACE(gdp_bh, "get_write_access");
3812 err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE);
3816 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3818 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3819 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3820 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3821 "fs metadata", block, block+len);
3822 /* File system mounted not to panic on error
3823 * Fix the bitmap and return EFSCORRUPTED
3824 * We leak some of the blocks here.
3826 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3827 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3828 ac->ac_b_ex.fe_len);
3829 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3830 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3832 err = -EFSCORRUPTED;
3836 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3837 #ifdef AGGRESSIVE_CHECK
3840 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3841 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3842 bitmap_bh->b_data));
3846 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3847 ac->ac_b_ex.fe_len);
3848 if (ext4_has_group_desc_csum(sb) &&
3849 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3850 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3851 ext4_free_group_clusters_set(sb, gdp,
3852 ext4_free_clusters_after_init(sb,
3853 ac->ac_b_ex.fe_group, gdp));
3855 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3856 ext4_free_group_clusters_set(sb, gdp, len);
3857 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3858 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3860 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3861 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3863 * Now reduce the dirty block count also. Should not go negative
3865 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3866 /* release all the reserved blocks if non delalloc */
3867 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3870 if (sbi->s_log_groups_per_flex) {
3871 ext4_group_t flex_group = ext4_flex_group(sbi,
3872 ac->ac_b_ex.fe_group);
3873 atomic64_sub(ac->ac_b_ex.fe_len,
3874 &sbi_array_rcu_deref(sbi, s_flex_groups,
3875 flex_group)->free_clusters);
3878 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3881 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3889 * Idempotent helper for Ext4 fast commit replay path to set the state of
3890 * blocks in bitmaps and update counters.
3892 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
3895 struct buffer_head *bitmap_bh = NULL;
3896 struct ext4_group_desc *gdp;
3897 struct buffer_head *gdp_bh;
3898 struct ext4_sb_info *sbi = EXT4_SB(sb);
3900 ext4_grpblk_t blkoff;
3903 unsigned int clen, clen_changed, thisgrp_len;
3906 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
3909 * Check to see if we are freeing blocks across a group
3911 * In case of flex_bg, this can happen that (block, len) may
3912 * span across more than one group. In that case we need to
3913 * get the corresponding group metadata to work with.
3914 * For this we have goto again loop.
3916 thisgrp_len = min_t(unsigned int, (unsigned int)len,
3917 EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
3918 clen = EXT4_NUM_B2C(sbi, thisgrp_len);
3920 if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
3921 ext4_error(sb, "Marking blocks in system zone - "
3922 "Block = %llu, len = %u",
3923 block, thisgrp_len);
3928 bitmap_bh = ext4_read_block_bitmap(sb, group);
3929 if (IS_ERR(bitmap_bh)) {
3930 err = PTR_ERR(bitmap_bh);
3936 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
3940 ext4_lock_group(sb, group);
3942 for (i = 0; i < clen; i++)
3943 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
3947 clen_changed = clen - already;
3949 mb_set_bits(bitmap_bh->b_data, blkoff, clen);
3951 mb_clear_bits(bitmap_bh->b_data, blkoff, clen);
3952 if (ext4_has_group_desc_csum(sb) &&
3953 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3954 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3955 ext4_free_group_clusters_set(sb, gdp,
3956 ext4_free_clusters_after_init(sb, group, gdp));
3959 clen = ext4_free_group_clusters(sb, gdp) - clen_changed;
3961 clen = ext4_free_group_clusters(sb, gdp) + clen_changed;
3963 ext4_free_group_clusters_set(sb, gdp, clen);
3964 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
3965 ext4_group_desc_csum_set(sb, group, gdp);
3967 ext4_unlock_group(sb, group);
3969 if (sbi->s_log_groups_per_flex) {
3970 ext4_group_t flex_group = ext4_flex_group(sbi, group);
3971 struct flex_groups *fg = sbi_array_rcu_deref(sbi,
3972 s_flex_groups, flex_group);
3975 atomic64_sub(clen_changed, &fg->free_clusters);
3977 atomic64_add(clen_changed, &fg->free_clusters);
3981 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
3984 sync_dirty_buffer(bitmap_bh);
3985 err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
3986 sync_dirty_buffer(gdp_bh);
3990 block += thisgrp_len;
4001 * here we normalize request for locality group
4002 * Group request are normalized to s_mb_group_prealloc, which goes to
4003 * s_strip if we set the same via mount option.
4004 * s_mb_group_prealloc can be configured via
4005 * /sys/fs/ext4/<partition>/mb_group_prealloc
4007 * XXX: should we try to preallocate more than the group has now?
4009 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
4011 struct super_block *sb = ac->ac_sb;
4012 struct ext4_locality_group *lg = ac->ac_lg;
4015 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
4016 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
4020 * Normalization means making request better in terms of
4021 * size and alignment
4023 static noinline_for_stack void
4024 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
4025 struct ext4_allocation_request *ar)
4027 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4030 loff_t size, start_off;
4031 loff_t orig_size __maybe_unused;
4033 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4034 struct ext4_prealloc_space *pa;
4036 /* do normalize only data requests, metadata requests
4037 do not need preallocation */
4038 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4041 /* sometime caller may want exact blocks */
4042 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4045 /* caller may indicate that preallocation isn't
4046 * required (it's a tail, for example) */
4047 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
4050 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
4051 ext4_mb_normalize_group_request(ac);
4055 bsbits = ac->ac_sb->s_blocksize_bits;
4057 /* first, let's learn actual file size
4058 * given current request is allocated */
4059 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4060 size = size << bsbits;
4061 if (size < i_size_read(ac->ac_inode))
4062 size = i_size_read(ac->ac_inode);
4065 /* max size of free chunks */
4068 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
4069 (req <= (size) || max <= (chunk_size))
4071 /* first, try to predict filesize */
4072 /* XXX: should this table be tunable? */
4074 if (size <= 16 * 1024) {
4076 } else if (size <= 32 * 1024) {
4078 } else if (size <= 64 * 1024) {
4080 } else if (size <= 128 * 1024) {
4082 } else if (size <= 256 * 1024) {
4084 } else if (size <= 512 * 1024) {
4086 } else if (size <= 1024 * 1024) {
4088 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
4089 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4090 (21 - bsbits)) << 21;
4091 size = 2 * 1024 * 1024;
4092 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
4093 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4094 (22 - bsbits)) << 22;
4095 size = 4 * 1024 * 1024;
4096 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
4097 (8<<20)>>bsbits, max, 8 * 1024)) {
4098 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4099 (23 - bsbits)) << 23;
4100 size = 8 * 1024 * 1024;
4102 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
4103 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
4104 ac->ac_o_ex.fe_len) << bsbits;
4106 size = size >> bsbits;
4107 start = start_off >> bsbits;
4110 * For tiny groups (smaller than 8MB) the chosen allocation
4111 * alignment may be larger than group size. Make sure the
4112 * alignment does not move allocation to a different group which
4113 * makes mballoc fail assertions later.
4115 start = max(start, rounddown(ac->ac_o_ex.fe_logical,
4116 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
4118 /* don't cover already allocated blocks in selected range */
4119 if (ar->pleft && start <= ar->lleft) {
4120 size -= ar->lleft + 1 - start;
4121 start = ar->lleft + 1;
4123 if (ar->pright && start + size - 1 >= ar->lright)
4124 size -= start + size - ar->lright;
4127 * Trim allocation request for filesystems with artificially small
4130 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4131 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4135 /* check we don't cross already preallocated blocks */
4137 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4142 spin_lock(&pa->pa_lock);
4143 if (pa->pa_deleted) {
4144 spin_unlock(&pa->pa_lock);
4148 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4151 /* PA must not overlap original request */
4152 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
4153 ac->ac_o_ex.fe_logical < pa->pa_lstart));
4155 /* skip PAs this normalized request doesn't overlap with */
4156 if (pa->pa_lstart >= end || pa_end <= start) {
4157 spin_unlock(&pa->pa_lock);
4160 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
4162 /* adjust start or end to be adjacent to this pa */
4163 if (pa_end <= ac->ac_o_ex.fe_logical) {
4164 BUG_ON(pa_end < start);
4166 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4167 BUG_ON(pa->pa_lstart > end);
4168 end = pa->pa_lstart;
4170 spin_unlock(&pa->pa_lock);
4175 /* XXX: extra loop to check we really don't overlap preallocations */
4177 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4180 spin_lock(&pa->pa_lock);
4181 if (pa->pa_deleted == 0) {
4182 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4184 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
4186 spin_unlock(&pa->pa_lock);
4191 * In this function "start" and "size" are normalized for better
4192 * alignment and length such that we could preallocate more blocks.
4193 * This normalization is done such that original request of
4194 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
4195 * "size" boundaries.
4196 * (Note fe_len can be relaxed since FS block allocation API does not
4197 * provide gurantee on number of contiguous blocks allocation since that
4198 * depends upon free space left, etc).
4199 * In case of inode pa, later we use the allocated blocks
4200 * [pa_start + fe_logical - pa_lstart, fe_len/size] from the preallocated
4201 * range of goal/best blocks [start, size] to put it at the
4202 * ac_o_ex.fe_logical extent of this inode.
4203 * (See ext4_mb_use_inode_pa() for more details)
4205 if (start + size <= ac->ac_o_ex.fe_logical ||
4206 start > ac->ac_o_ex.fe_logical) {
4207 ext4_msg(ac->ac_sb, KERN_ERR,
4208 "start %lu, size %lu, fe_logical %lu",
4209 (unsigned long) start, (unsigned long) size,
4210 (unsigned long) ac->ac_o_ex.fe_logical);
4213 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4215 /* now prepare goal request */
4217 /* XXX: is it better to align blocks WRT to logical
4218 * placement or satisfy big request as is */
4219 ac->ac_g_ex.fe_logical = start;
4220 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4222 /* define goal start in order to merge */
4223 if (ar->pright && (ar->lright == (start + size))) {
4224 /* merge to the right */
4225 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4226 &ac->ac_f_ex.fe_group,
4227 &ac->ac_f_ex.fe_start);
4228 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4230 if (ar->pleft && (ar->lleft + 1 == start)) {
4231 /* merge to the left */
4232 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4233 &ac->ac_f_ex.fe_group,
4234 &ac->ac_f_ex.fe_start);
4235 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4238 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4242 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4244 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4246 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4247 atomic_inc(&sbi->s_bal_reqs);
4248 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4249 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4250 atomic_inc(&sbi->s_bal_success);
4251 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4252 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4253 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4254 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4255 atomic_inc(&sbi->s_bal_goals);
4256 if (ac->ac_found > sbi->s_mb_max_to_scan)
4257 atomic_inc(&sbi->s_bal_breaks);
4260 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4261 trace_ext4_mballoc_alloc(ac);
4263 trace_ext4_mballoc_prealloc(ac);
4267 * Called on failure; free up any blocks from the inode PA for this
4268 * context. We don't need this for MB_GROUP_PA because we only change
4269 * pa_free in ext4_mb_release_context(), but on failure, we've already
4270 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4272 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4274 struct ext4_prealloc_space *pa = ac->ac_pa;
4275 struct ext4_buddy e4b;
4279 if (ac->ac_f_ex.fe_len == 0)
4281 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4284 * This should never happen since we pin the
4285 * pages in the ext4_allocation_context so
4286 * ext4_mb_load_buddy() should never fail.
4288 WARN(1, "mb_load_buddy failed (%d)", err);
4291 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4292 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4293 ac->ac_f_ex.fe_len);
4294 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4295 ext4_mb_unload_buddy(&e4b);
4298 if (pa->pa_type == MB_INODE_PA)
4299 pa->pa_free += ac->ac_b_ex.fe_len;
4303 * use blocks preallocated to inode
4305 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4306 struct ext4_prealloc_space *pa)
4308 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4313 /* found preallocated blocks, use them */
4314 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4315 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4316 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4317 len = EXT4_NUM_B2C(sbi, end - start);
4318 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4319 &ac->ac_b_ex.fe_start);
4320 ac->ac_b_ex.fe_len = len;
4321 ac->ac_status = AC_STATUS_FOUND;
4324 BUG_ON(start < pa->pa_pstart);
4325 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4326 BUG_ON(pa->pa_free < len);
4329 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4333 * use blocks preallocated to locality group
4335 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4336 struct ext4_prealloc_space *pa)
4338 unsigned int len = ac->ac_o_ex.fe_len;
4340 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4341 &ac->ac_b_ex.fe_group,
4342 &ac->ac_b_ex.fe_start);
4343 ac->ac_b_ex.fe_len = len;
4344 ac->ac_status = AC_STATUS_FOUND;
4347 /* we don't correct pa_pstart or pa_plen here to avoid
4348 * possible race when the group is being loaded concurrently
4349 * instead we correct pa later, after blocks are marked
4350 * in on-disk bitmap -- see ext4_mb_release_context()
4351 * Other CPUs are prevented from allocating from this pa by lg_mutex
4353 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4354 pa->pa_lstart-len, len, pa);
4358 * Return the prealloc space that have minimal distance
4359 * from the goal block. @cpa is the prealloc
4360 * space that is having currently known minimal distance
4361 * from the goal block.
4363 static struct ext4_prealloc_space *
4364 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4365 struct ext4_prealloc_space *pa,
4366 struct ext4_prealloc_space *cpa)
4368 ext4_fsblk_t cur_distance, new_distance;
4371 atomic_inc(&pa->pa_count);
4374 cur_distance = abs(goal_block - cpa->pa_pstart);
4375 new_distance = abs(goal_block - pa->pa_pstart);
4377 if (cur_distance <= new_distance)
4380 /* drop the previous reference */
4381 atomic_dec(&cpa->pa_count);
4382 atomic_inc(&pa->pa_count);
4387 * search goal blocks in preallocated space
4389 static noinline_for_stack bool
4390 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4392 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4394 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4395 struct ext4_locality_group *lg;
4396 struct ext4_prealloc_space *pa, *cpa = NULL;
4397 ext4_fsblk_t goal_block;
4399 /* only data can be preallocated */
4400 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4403 /* first, try per-file preallocation */
4405 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4407 /* all fields in this condition don't change,
4408 * so we can skip locking for them */
4409 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
4410 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
4411 EXT4_C2B(sbi, pa->pa_len)))
4414 /* non-extent files can't have physical blocks past 2^32 */
4415 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4416 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
4417 EXT4_MAX_BLOCK_FILE_PHYS))
4420 /* found preallocated blocks, use them */
4421 spin_lock(&pa->pa_lock);
4422 if (pa->pa_deleted == 0 && pa->pa_free) {
4423 atomic_inc(&pa->pa_count);
4424 ext4_mb_use_inode_pa(ac, pa);
4425 spin_unlock(&pa->pa_lock);
4426 ac->ac_criteria = 10;
4430 spin_unlock(&pa->pa_lock);
4434 /* can we use group allocation? */
4435 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4438 /* inode may have no locality group for some reason */
4442 order = fls(ac->ac_o_ex.fe_len) - 1;
4443 if (order > PREALLOC_TB_SIZE - 1)
4444 /* The max size of hash table is PREALLOC_TB_SIZE */
4445 order = PREALLOC_TB_SIZE - 1;
4447 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4449 * search for the prealloc space that is having
4450 * minimal distance from the goal block.
4452 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4454 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
4456 spin_lock(&pa->pa_lock);
4457 if (pa->pa_deleted == 0 &&
4458 pa->pa_free >= ac->ac_o_ex.fe_len) {
4460 cpa = ext4_mb_check_group_pa(goal_block,
4463 spin_unlock(&pa->pa_lock);
4468 ext4_mb_use_group_pa(ac, cpa);
4469 ac->ac_criteria = 20;
4476 * the function goes through all block freed in the group
4477 * but not yet committed and marks them used in in-core bitmap.
4478 * buddy must be generated from this bitmap
4479 * Need to be called with the ext4 group lock held
4481 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
4485 struct ext4_group_info *grp;
4486 struct ext4_free_data *entry;
4488 grp = ext4_get_group_info(sb, group);
4489 n = rb_first(&(grp->bb_free_root));
4492 entry = rb_entry(n, struct ext4_free_data, efd_node);
4493 mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
4500 * the function goes through all preallocation in this group and marks them
4501 * used in in-core bitmap. buddy must be generated from this bitmap
4502 * Need to be called with ext4 group lock held
4504 static noinline_for_stack
4505 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4508 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4509 struct ext4_prealloc_space *pa;
4510 struct list_head *cur;
4511 ext4_group_t groupnr;
4512 ext4_grpblk_t start;
4513 int preallocated = 0;
4516 /* all form of preallocation discards first load group,
4517 * so the only competing code is preallocation use.
4518 * we don't need any locking here
4519 * notice we do NOT ignore preallocations with pa_deleted
4520 * otherwise we could leave used blocks available for
4521 * allocation in buddy when concurrent ext4_mb_put_pa()
4522 * is dropping preallocation
4524 list_for_each(cur, &grp->bb_prealloc_list) {
4525 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
4526 spin_lock(&pa->pa_lock);
4527 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4530 spin_unlock(&pa->pa_lock);
4531 if (unlikely(len == 0))
4533 BUG_ON(groupnr != group);
4534 mb_set_bits(bitmap, start, len);
4535 preallocated += len;
4537 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
4540 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
4541 struct ext4_prealloc_space *pa)
4543 struct ext4_inode_info *ei;
4545 if (pa->pa_deleted) {
4546 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
4547 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
4554 if (pa->pa_type == MB_INODE_PA) {
4555 ei = EXT4_I(pa->pa_inode);
4556 atomic_dec(&ei->i_prealloc_active);
4560 static void ext4_mb_pa_callback(struct rcu_head *head)
4562 struct ext4_prealloc_space *pa;
4563 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
4565 BUG_ON(atomic_read(&pa->pa_count));
4566 BUG_ON(pa->pa_deleted == 0);
4567 kmem_cache_free(ext4_pspace_cachep, pa);
4571 * drops a reference to preallocated space descriptor
4572 * if this was the last reference and the space is consumed
4574 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
4575 struct super_block *sb, struct ext4_prealloc_space *pa)
4578 ext4_fsblk_t grp_blk;
4580 /* in this short window concurrent discard can set pa_deleted */
4581 spin_lock(&pa->pa_lock);
4582 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
4583 spin_unlock(&pa->pa_lock);
4587 if (pa->pa_deleted == 1) {
4588 spin_unlock(&pa->pa_lock);
4592 ext4_mb_mark_pa_deleted(sb, pa);
4593 spin_unlock(&pa->pa_lock);
4595 grp_blk = pa->pa_pstart;
4597 * If doing group-based preallocation, pa_pstart may be in the
4598 * next group when pa is used up
4600 if (pa->pa_type == MB_GROUP_PA)
4603 grp = ext4_get_group_number(sb, grp_blk);
4608 * P1 (buddy init) P2 (regular allocation)
4609 * find block B in PA
4610 * copy on-disk bitmap to buddy
4611 * mark B in on-disk bitmap
4612 * drop PA from group
4613 * mark all PAs in buddy
4615 * thus, P1 initializes buddy with B available. to prevent this
4616 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
4619 ext4_lock_group(sb, grp);
4620 list_del(&pa->pa_group_list);
4621 ext4_unlock_group(sb, grp);
4623 spin_lock(pa->pa_obj_lock);
4624 list_del_rcu(&pa->pa_inode_list);
4625 spin_unlock(pa->pa_obj_lock);
4627 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4631 * creates new preallocated space for given inode
4633 static noinline_for_stack void
4634 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
4636 struct super_block *sb = ac->ac_sb;
4637 struct ext4_sb_info *sbi = EXT4_SB(sb);
4638 struct ext4_prealloc_space *pa;
4639 struct ext4_group_info *grp;
4640 struct ext4_inode_info *ei;
4642 /* preallocate only when found space is larger then requested */
4643 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4644 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4645 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4646 BUG_ON(ac->ac_pa == NULL);
4650 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
4656 /* we can't allocate as much as normalizer wants.
4657 * so, found space must get proper lstart
4658 * to cover original request */
4659 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
4660 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
4662 /* we're limited by original request in that
4663 * logical block must be covered any way
4664 * winl is window we can move our chunk within */
4665 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
4667 /* also, we should cover whole original request */
4668 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
4670 /* the smallest one defines real window */
4671 win = min(winl, wins);
4673 offs = ac->ac_o_ex.fe_logical %
4674 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4675 if (offs && offs < win)
4678 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
4679 EXT4_NUM_B2C(sbi, win);
4680 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
4681 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
4684 /* preallocation can change ac_b_ex, thus we store actually
4685 * allocated blocks for history */
4686 ac->ac_f_ex = ac->ac_b_ex;
4688 pa->pa_lstart = ac->ac_b_ex.fe_logical;
4689 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4690 pa->pa_len = ac->ac_b_ex.fe_len;
4691 pa->pa_free = pa->pa_len;
4692 spin_lock_init(&pa->pa_lock);
4693 INIT_LIST_HEAD(&pa->pa_inode_list);
4694 INIT_LIST_HEAD(&pa->pa_group_list);
4696 pa->pa_type = MB_INODE_PA;
4698 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4699 pa->pa_len, pa->pa_lstart);
4700 trace_ext4_mb_new_inode_pa(ac, pa);
4702 ext4_mb_use_inode_pa(ac, pa);
4703 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
4705 ei = EXT4_I(ac->ac_inode);
4706 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4708 pa->pa_obj_lock = &ei->i_prealloc_lock;
4709 pa->pa_inode = ac->ac_inode;
4711 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4713 spin_lock(pa->pa_obj_lock);
4714 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
4715 spin_unlock(pa->pa_obj_lock);
4716 atomic_inc(&ei->i_prealloc_active);
4720 * creates new preallocated space for locality group inodes belongs to
4722 static noinline_for_stack void
4723 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
4725 struct super_block *sb = ac->ac_sb;
4726 struct ext4_locality_group *lg;
4727 struct ext4_prealloc_space *pa;
4728 struct ext4_group_info *grp;
4730 /* preallocate only when found space is larger then requested */
4731 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4732 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4733 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4734 BUG_ON(ac->ac_pa == NULL);
4738 /* preallocation can change ac_b_ex, thus we store actually
4739 * allocated blocks for history */
4740 ac->ac_f_ex = ac->ac_b_ex;
4742 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4743 pa->pa_lstart = pa->pa_pstart;
4744 pa->pa_len = ac->ac_b_ex.fe_len;
4745 pa->pa_free = pa->pa_len;
4746 spin_lock_init(&pa->pa_lock);
4747 INIT_LIST_HEAD(&pa->pa_inode_list);
4748 INIT_LIST_HEAD(&pa->pa_group_list);
4750 pa->pa_type = MB_GROUP_PA;
4752 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4753 pa->pa_len, pa->pa_lstart);
4754 trace_ext4_mb_new_group_pa(ac, pa);
4756 ext4_mb_use_group_pa(ac, pa);
4757 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4759 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4763 pa->pa_obj_lock = &lg->lg_prealloc_lock;
4764 pa->pa_inode = NULL;
4766 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4769 * We will later add the new pa to the right bucket
4770 * after updating the pa_free in ext4_mb_release_context
4774 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4776 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4777 ext4_mb_new_group_pa(ac);
4779 ext4_mb_new_inode_pa(ac);
4783 * finds all unused blocks in on-disk bitmap, frees them in
4784 * in-core bitmap and buddy.
4785 * @pa must be unlinked from inode and group lists, so that
4786 * nobody else can find/use it.
4787 * the caller MUST hold group/inode locks.
4788 * TODO: optimize the case when there are no in-core structures yet
4790 static noinline_for_stack int
4791 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4792 struct ext4_prealloc_space *pa)
4794 struct super_block *sb = e4b->bd_sb;
4795 struct ext4_sb_info *sbi = EXT4_SB(sb);
4800 unsigned long long grp_blk_start;
4803 BUG_ON(pa->pa_deleted == 0);
4804 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4805 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4806 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4807 end = bit + pa->pa_len;
4810 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4813 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4814 mb_debug(sb, "free preallocated %u/%u in group %u\n",
4815 (unsigned) ext4_group_first_block_no(sb, group) + bit,
4816 (unsigned) next - bit, (unsigned) group);
4819 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4820 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4821 EXT4_C2B(sbi, bit)),
4823 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4826 if (free != pa->pa_free) {
4827 ext4_msg(e4b->bd_sb, KERN_CRIT,
4828 "pa %p: logic %lu, phys. %lu, len %d",
4829 pa, (unsigned long) pa->pa_lstart,
4830 (unsigned long) pa->pa_pstart,
4832 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4835 * pa is already deleted so we use the value obtained
4836 * from the bitmap and continue.
4839 atomic_add(free, &sbi->s_mb_discarded);
4844 static noinline_for_stack int
4845 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4846 struct ext4_prealloc_space *pa)
4848 struct super_block *sb = e4b->bd_sb;
4852 trace_ext4_mb_release_group_pa(sb, pa);
4853 BUG_ON(pa->pa_deleted == 0);
4854 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4855 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4856 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4857 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4858 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4864 * releases all preallocations in given group
4866 * first, we need to decide discard policy:
4867 * - when do we discard
4869 * - how many do we discard
4870 * 1) how many requested
4872 static noinline_for_stack int
4873 ext4_mb_discard_group_preallocations(struct super_block *sb,
4874 ext4_group_t group, int *busy)
4876 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4877 struct buffer_head *bitmap_bh = NULL;
4878 struct ext4_prealloc_space *pa, *tmp;
4879 struct list_head list;
4880 struct ext4_buddy e4b;
4884 mb_debug(sb, "discard preallocation for group %u\n", group);
4885 if (list_empty(&grp->bb_prealloc_list))
4888 bitmap_bh = ext4_read_block_bitmap(sb, group);
4889 if (IS_ERR(bitmap_bh)) {
4890 err = PTR_ERR(bitmap_bh);
4891 ext4_error_err(sb, -err,
4892 "Error %d reading block bitmap for %u",
4897 err = ext4_mb_load_buddy(sb, group, &e4b);
4899 ext4_warning(sb, "Error %d loading buddy information for %u",
4905 INIT_LIST_HEAD(&list);
4906 ext4_lock_group(sb, group);
4907 list_for_each_entry_safe(pa, tmp,
4908 &grp->bb_prealloc_list, pa_group_list) {
4909 spin_lock(&pa->pa_lock);
4910 if (atomic_read(&pa->pa_count)) {
4911 spin_unlock(&pa->pa_lock);
4915 if (pa->pa_deleted) {
4916 spin_unlock(&pa->pa_lock);
4920 /* seems this one can be freed ... */
4921 ext4_mb_mark_pa_deleted(sb, pa);
4924 this_cpu_inc(discard_pa_seq);
4926 /* we can trust pa_free ... */
4927 free += pa->pa_free;
4929 spin_unlock(&pa->pa_lock);
4931 list_del(&pa->pa_group_list);
4932 list_add(&pa->u.pa_tmp_list, &list);
4935 /* now free all selected PAs */
4936 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4938 /* remove from object (inode or locality group) */
4939 spin_lock(pa->pa_obj_lock);
4940 list_del_rcu(&pa->pa_inode_list);
4941 spin_unlock(pa->pa_obj_lock);
4943 if (pa->pa_type == MB_GROUP_PA)
4944 ext4_mb_release_group_pa(&e4b, pa);
4946 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4948 list_del(&pa->u.pa_tmp_list);
4949 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4952 ext4_unlock_group(sb, group);
4953 ext4_mb_unload_buddy(&e4b);
4956 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4957 free, group, grp->bb_free);
4962 * releases all non-used preallocated blocks for given inode
4964 * It's important to discard preallocations under i_data_sem
4965 * We don't want another block to be served from the prealloc
4966 * space when we are discarding the inode prealloc space.
4968 * FIXME!! Make sure it is valid at all the call sites
4970 void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
4972 struct ext4_inode_info *ei = EXT4_I(inode);
4973 struct super_block *sb = inode->i_sb;
4974 struct buffer_head *bitmap_bh = NULL;
4975 struct ext4_prealloc_space *pa, *tmp;
4976 ext4_group_t group = 0;
4977 struct list_head list;
4978 struct ext4_buddy e4b;
4981 if (!S_ISREG(inode->i_mode)) {
4982 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4986 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
4989 mb_debug(sb, "discard preallocation for inode %lu\n",
4991 trace_ext4_discard_preallocations(inode,
4992 atomic_read(&ei->i_prealloc_active), needed);
4994 INIT_LIST_HEAD(&list);
5000 /* first, collect all pa's in the inode */
5001 spin_lock(&ei->i_prealloc_lock);
5002 while (!list_empty(&ei->i_prealloc_list) && needed) {
5003 pa = list_entry(ei->i_prealloc_list.prev,
5004 struct ext4_prealloc_space, pa_inode_list);
5005 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
5006 spin_lock(&pa->pa_lock);
5007 if (atomic_read(&pa->pa_count)) {
5008 /* this shouldn't happen often - nobody should
5009 * use preallocation while we're discarding it */
5010 spin_unlock(&pa->pa_lock);
5011 spin_unlock(&ei->i_prealloc_lock);
5012 ext4_msg(sb, KERN_ERR,
5013 "uh-oh! used pa while discarding");
5015 schedule_timeout_uninterruptible(HZ);
5019 if (pa->pa_deleted == 0) {
5020 ext4_mb_mark_pa_deleted(sb, pa);
5021 spin_unlock(&pa->pa_lock);
5022 list_del_rcu(&pa->pa_inode_list);
5023 list_add(&pa->u.pa_tmp_list, &list);
5028 /* someone is deleting pa right now */
5029 spin_unlock(&pa->pa_lock);
5030 spin_unlock(&ei->i_prealloc_lock);
5032 /* we have to wait here because pa_deleted
5033 * doesn't mean pa is already unlinked from
5034 * the list. as we might be called from
5035 * ->clear_inode() the inode will get freed
5036 * and concurrent thread which is unlinking
5037 * pa from inode's list may access already
5038 * freed memory, bad-bad-bad */
5040 /* XXX: if this happens too often, we can
5041 * add a flag to force wait only in case
5042 * of ->clear_inode(), but not in case of
5043 * regular truncate */
5044 schedule_timeout_uninterruptible(HZ);
5047 spin_unlock(&ei->i_prealloc_lock);
5049 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5050 BUG_ON(pa->pa_type != MB_INODE_PA);
5051 group = ext4_get_group_number(sb, pa->pa_pstart);
5053 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5054 GFP_NOFS|__GFP_NOFAIL);
5056 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5061 bitmap_bh = ext4_read_block_bitmap(sb, group);
5062 if (IS_ERR(bitmap_bh)) {
5063 err = PTR_ERR(bitmap_bh);
5064 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
5066 ext4_mb_unload_buddy(&e4b);
5070 ext4_lock_group(sb, group);
5071 list_del(&pa->pa_group_list);
5072 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5073 ext4_unlock_group(sb, group);
5075 ext4_mb_unload_buddy(&e4b);
5078 list_del(&pa->u.pa_tmp_list);
5079 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5083 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
5085 struct ext4_prealloc_space *pa;
5087 BUG_ON(ext4_pspace_cachep == NULL);
5088 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
5091 atomic_set(&pa->pa_count, 1);
5096 static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
5098 struct ext4_prealloc_space *pa = ac->ac_pa;
5102 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
5103 kmem_cache_free(ext4_pspace_cachep, pa);
5106 #ifdef CONFIG_EXT4_DEBUG
5107 static inline void ext4_mb_show_pa(struct super_block *sb)
5109 ext4_group_t i, ngroups;
5111 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5114 ngroups = ext4_get_groups_count(sb);
5115 mb_debug(sb, "groups: ");
5116 for (i = 0; i < ngroups; i++) {
5117 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5118 struct ext4_prealloc_space *pa;
5119 ext4_grpblk_t start;
5120 struct list_head *cur;
5121 ext4_lock_group(sb, i);
5122 list_for_each(cur, &grp->bb_prealloc_list) {
5123 pa = list_entry(cur, struct ext4_prealloc_space,
5125 spin_lock(&pa->pa_lock);
5126 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5128 spin_unlock(&pa->pa_lock);
5129 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5132 ext4_unlock_group(sb, i);
5133 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5138 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5140 struct super_block *sb = ac->ac_sb;
5142 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5145 mb_debug(sb, "Can't allocate:"
5146 " Allocation context details:");
5147 mb_debug(sb, "status %u flags 0x%x",
5148 ac->ac_status, ac->ac_flags);
5149 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5150 "goal %lu/%lu/%lu@%lu, "
5151 "best %lu/%lu/%lu@%lu cr %d",
5152 (unsigned long)ac->ac_o_ex.fe_group,
5153 (unsigned long)ac->ac_o_ex.fe_start,
5154 (unsigned long)ac->ac_o_ex.fe_len,
5155 (unsigned long)ac->ac_o_ex.fe_logical,
5156 (unsigned long)ac->ac_g_ex.fe_group,
5157 (unsigned long)ac->ac_g_ex.fe_start,
5158 (unsigned long)ac->ac_g_ex.fe_len,
5159 (unsigned long)ac->ac_g_ex.fe_logical,
5160 (unsigned long)ac->ac_b_ex.fe_group,
5161 (unsigned long)ac->ac_b_ex.fe_start,
5162 (unsigned long)ac->ac_b_ex.fe_len,
5163 (unsigned long)ac->ac_b_ex.fe_logical,
5164 (int)ac->ac_criteria);
5165 mb_debug(sb, "%u found", ac->ac_found);
5166 ext4_mb_show_pa(sb);
5169 static inline void ext4_mb_show_pa(struct super_block *sb)
5173 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5175 ext4_mb_show_pa(ac->ac_sb);
5181 * We use locality group preallocation for small size file. The size of the
5182 * file is determined by the current size or the resulting size after
5183 * allocation which ever is larger
5185 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5187 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5189 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5190 int bsbits = ac->ac_sb->s_blocksize_bits;
5192 bool inode_pa_eligible, group_pa_eligible;
5194 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5197 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5200 group_pa_eligible = sbi->s_mb_group_prealloc > 0;
5201 inode_pa_eligible = true;
5202 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
5203 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5206 /* No point in using inode preallocation for closed files */
5207 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5208 !inode_is_open_for_write(ac->ac_inode))
5209 inode_pa_eligible = false;
5211 size = max(size, isize);
5212 /* Don't use group allocation for large files */
5213 if (size > sbi->s_mb_stream_request)
5214 group_pa_eligible = false;
5216 if (!group_pa_eligible) {
5217 if (inode_pa_eligible)
5218 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5220 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5224 BUG_ON(ac->ac_lg != NULL);
5226 * locality group prealloc space are per cpu. The reason for having
5227 * per cpu locality group is to reduce the contention between block
5228 * request from multiple CPUs.
5230 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5232 /* we're going to use group allocation */
5233 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5235 /* serialize all allocations in the group */
5236 mutex_lock(&ac->ac_lg->lg_mutex);
5239 static noinline_for_stack int
5240 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5241 struct ext4_allocation_request *ar)
5243 struct super_block *sb = ar->inode->i_sb;
5244 struct ext4_sb_info *sbi = EXT4_SB(sb);
5245 struct ext4_super_block *es = sbi->s_es;
5249 ext4_grpblk_t block;
5251 /* we can't allocate > group size */
5254 /* just a dirty hack to filter too big requests */
5255 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5256 len = EXT4_CLUSTERS_PER_GROUP(sb);
5258 /* start searching from the goal */
5260 if (goal < le32_to_cpu(es->s_first_data_block) ||
5261 goal >= ext4_blocks_count(es))
5262 goal = le32_to_cpu(es->s_first_data_block);
5263 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5265 /* set up allocation goals */
5266 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5267 ac->ac_status = AC_STATUS_CONTINUE;
5269 ac->ac_inode = ar->inode;
5270 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5271 ac->ac_o_ex.fe_group = group;
5272 ac->ac_o_ex.fe_start = block;
5273 ac->ac_o_ex.fe_len = len;
5274 ac->ac_g_ex = ac->ac_o_ex;
5275 ac->ac_flags = ar->flags;
5277 /* we have to define context: we'll work with a file or
5278 * locality group. this is a policy, actually */
5279 ext4_mb_group_or_file(ac);
5281 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5282 "left: %u/%u, right %u/%u to %swritable\n",
5283 (unsigned) ar->len, (unsigned) ar->logical,
5284 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5285 (unsigned) ar->lleft, (unsigned) ar->pleft,
5286 (unsigned) ar->lright, (unsigned) ar->pright,
5287 inode_is_open_for_write(ar->inode) ? "" : "non-");
5292 static noinline_for_stack void
5293 ext4_mb_discard_lg_preallocations(struct super_block *sb,
5294 struct ext4_locality_group *lg,
5295 int order, int total_entries)
5297 ext4_group_t group = 0;
5298 struct ext4_buddy e4b;
5299 struct list_head discard_list;
5300 struct ext4_prealloc_space *pa, *tmp;
5302 mb_debug(sb, "discard locality group preallocation\n");
5304 INIT_LIST_HEAD(&discard_list);
5306 spin_lock(&lg->lg_prealloc_lock);
5307 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5309 lockdep_is_held(&lg->lg_prealloc_lock)) {
5310 spin_lock(&pa->pa_lock);
5311 if (atomic_read(&pa->pa_count)) {
5313 * This is the pa that we just used
5314 * for block allocation. So don't
5317 spin_unlock(&pa->pa_lock);
5320 if (pa->pa_deleted) {
5321 spin_unlock(&pa->pa_lock);
5324 /* only lg prealloc space */
5325 BUG_ON(pa->pa_type != MB_GROUP_PA);
5327 /* seems this one can be freed ... */
5328 ext4_mb_mark_pa_deleted(sb, pa);
5329 spin_unlock(&pa->pa_lock);
5331 list_del_rcu(&pa->pa_inode_list);
5332 list_add(&pa->u.pa_tmp_list, &discard_list);
5335 if (total_entries <= 5) {
5337 * we want to keep only 5 entries
5338 * allowing it to grow to 8. This
5339 * mak sure we don't call discard
5340 * soon for this list.
5345 spin_unlock(&lg->lg_prealloc_lock);
5347 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5350 group = ext4_get_group_number(sb, pa->pa_pstart);
5351 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5352 GFP_NOFS|__GFP_NOFAIL);
5354 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5358 ext4_lock_group(sb, group);
5359 list_del(&pa->pa_group_list);
5360 ext4_mb_release_group_pa(&e4b, pa);
5361 ext4_unlock_group(sb, group);
5363 ext4_mb_unload_buddy(&e4b);
5364 list_del(&pa->u.pa_tmp_list);
5365 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5370 * We have incremented pa_count. So it cannot be freed at this
5371 * point. Also we hold lg_mutex. So no parallel allocation is
5372 * possible from this lg. That means pa_free cannot be updated.
5374 * A parallel ext4_mb_discard_group_preallocations is possible.
5375 * which can cause the lg_prealloc_list to be updated.
5378 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5380 int order, added = 0, lg_prealloc_count = 1;
5381 struct super_block *sb = ac->ac_sb;
5382 struct ext4_locality_group *lg = ac->ac_lg;
5383 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5385 order = fls(pa->pa_free) - 1;
5386 if (order > PREALLOC_TB_SIZE - 1)
5387 /* The max size of hash table is PREALLOC_TB_SIZE */
5388 order = PREALLOC_TB_SIZE - 1;
5389 /* Add the prealloc space to lg */
5390 spin_lock(&lg->lg_prealloc_lock);
5391 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5393 lockdep_is_held(&lg->lg_prealloc_lock)) {
5394 spin_lock(&tmp_pa->pa_lock);
5395 if (tmp_pa->pa_deleted) {
5396 spin_unlock(&tmp_pa->pa_lock);
5399 if (!added && pa->pa_free < tmp_pa->pa_free) {
5400 /* Add to the tail of the previous entry */
5401 list_add_tail_rcu(&pa->pa_inode_list,
5402 &tmp_pa->pa_inode_list);
5405 * we want to count the total
5406 * number of entries in the list
5409 spin_unlock(&tmp_pa->pa_lock);
5410 lg_prealloc_count++;
5413 list_add_tail_rcu(&pa->pa_inode_list,
5414 &lg->lg_prealloc_list[order]);
5415 spin_unlock(&lg->lg_prealloc_lock);
5417 /* Now trim the list to be not more than 8 elements */
5418 if (lg_prealloc_count > 8) {
5419 ext4_mb_discard_lg_preallocations(sb, lg,
5420 order, lg_prealloc_count);
5427 * if per-inode prealloc list is too long, trim some PA
5429 static void ext4_mb_trim_inode_pa(struct inode *inode)
5431 struct ext4_inode_info *ei = EXT4_I(inode);
5432 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5435 count = atomic_read(&ei->i_prealloc_active);
5436 delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
5437 if (count > sbi->s_mb_max_inode_prealloc + delta) {
5438 count -= sbi->s_mb_max_inode_prealloc;
5439 ext4_discard_preallocations(inode, count);
5444 * release all resource we used in allocation
5446 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
5448 struct inode *inode = ac->ac_inode;
5449 struct ext4_inode_info *ei = EXT4_I(inode);
5450 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5451 struct ext4_prealloc_space *pa = ac->ac_pa;
5453 if (pa->pa_type == MB_GROUP_PA) {
5454 /* see comment in ext4_mb_use_group_pa() */
5455 spin_lock(&pa->pa_lock);
5456 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5457 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5458 pa->pa_free -= ac->ac_b_ex.fe_len;
5459 pa->pa_len -= ac->ac_b_ex.fe_len;
5460 spin_unlock(&pa->pa_lock);
5463 * We want to add the pa to the right bucket.
5464 * Remove it from the list and while adding
5465 * make sure the list to which we are adding
5468 if (likely(pa->pa_free)) {
5469 spin_lock(pa->pa_obj_lock);
5470 list_del_rcu(&pa->pa_inode_list);
5471 spin_unlock(pa->pa_obj_lock);
5472 ext4_mb_add_n_trim(ac);
5476 if (pa->pa_type == MB_INODE_PA) {
5478 * treat per-inode prealloc list as a lru list, then try
5479 * to trim the least recently used PA.
5481 spin_lock(pa->pa_obj_lock);
5482 list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
5483 spin_unlock(pa->pa_obj_lock);
5486 ext4_mb_put_pa(ac, ac->ac_sb, pa);
5488 if (ac->ac_bitmap_page)
5489 put_page(ac->ac_bitmap_page);
5490 if (ac->ac_buddy_page)
5491 put_page(ac->ac_buddy_page);
5492 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5493 mutex_unlock(&ac->ac_lg->lg_mutex);
5494 ext4_mb_collect_stats(ac);
5495 ext4_mb_trim_inode_pa(inode);
5499 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
5501 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
5503 int freed = 0, busy = 0;
5506 trace_ext4_mb_discard_preallocations(sb, needed);
5509 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
5511 for (i = 0; i < ngroups && needed > 0; i++) {
5512 ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
5518 if (needed > 0 && busy && ++retry < 3) {
5526 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
5527 struct ext4_allocation_context *ac, u64 *seq)
5533 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
5538 seq_retry = ext4_get_discard_pa_seq_sum();
5539 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
5540 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
5546 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
5550 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5551 struct ext4_allocation_request *ar, int *errp);
5554 * Main entry point into mballoc to allocate blocks
5555 * it tries to use preallocation first, then falls back
5556 * to usual allocation
5558 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
5559 struct ext4_allocation_request *ar, int *errp)
5561 struct ext4_allocation_context *ac = NULL;
5562 struct ext4_sb_info *sbi;
5563 struct super_block *sb;
5564 ext4_fsblk_t block = 0;
5565 unsigned int inquota = 0;
5566 unsigned int reserv_clstrs = 0;
5571 sb = ar->inode->i_sb;
5574 trace_ext4_request_blocks(ar);
5575 if (sbi->s_mount_state & EXT4_FC_REPLAY)
5576 return ext4_mb_new_blocks_simple(handle, ar, errp);
5578 /* Allow to use superuser reservation for quota file */
5579 if (ext4_is_quota_file(ar->inode))
5580 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
5582 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
5583 /* Without delayed allocation we need to verify
5584 * there is enough free blocks to do block allocation
5585 * and verify allocation doesn't exceed the quota limits.
5588 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
5590 /* let others to free the space */
5592 ar->len = ar->len >> 1;
5595 ext4_mb_show_pa(sb);
5599 reserv_clstrs = ar->len;
5600 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
5601 dquot_alloc_block_nofail(ar->inode,
5602 EXT4_C2B(sbi, ar->len));
5605 dquot_alloc_block(ar->inode,
5606 EXT4_C2B(sbi, ar->len))) {
5608 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
5619 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
5626 *errp = ext4_mb_initialize_context(ac, ar);
5632 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
5633 seq = this_cpu_read(discard_pa_seq);
5634 if (!ext4_mb_use_preallocated(ac)) {
5635 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
5636 ext4_mb_normalize_request(ac, ar);
5638 *errp = ext4_mb_pa_alloc(ac);
5642 /* allocate space in core */
5643 *errp = ext4_mb_regular_allocator(ac);
5645 * pa allocated above is added to grp->bb_prealloc_list only
5646 * when we were able to allocate some block i.e. when
5647 * ac->ac_status == AC_STATUS_FOUND.
5648 * And error from above mean ac->ac_status != AC_STATUS_FOUND
5649 * So we have to free this pa here itself.
5652 ext4_mb_pa_free(ac);
5653 ext4_discard_allocated_blocks(ac);
5656 if (ac->ac_status == AC_STATUS_FOUND &&
5657 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
5658 ext4_mb_pa_free(ac);
5660 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
5661 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
5663 ext4_discard_allocated_blocks(ac);
5666 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5667 ar->len = ac->ac_b_ex.fe_len;
5670 if (++retries < 3 &&
5671 ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
5674 * If block allocation fails then the pa allocated above
5675 * needs to be freed here itself.
5677 ext4_mb_pa_free(ac);
5683 ac->ac_b_ex.fe_len = 0;
5685 ext4_mb_show_ac(ac);
5687 ext4_mb_release_context(ac);
5690 kmem_cache_free(ext4_ac_cachep, ac);
5691 if (inquota && ar->len < inquota)
5692 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
5694 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
5695 /* release all the reserved blocks if non delalloc */
5696 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
5700 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
5706 * We can merge two free data extents only if the physical blocks
5707 * are contiguous, AND the extents were freed by the same transaction,
5708 * AND the blocks are associated with the same group.
5710 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
5711 struct ext4_free_data *entry,
5712 struct ext4_free_data *new_entry,
5713 struct rb_root *entry_rb_root)
5715 if ((entry->efd_tid != new_entry->efd_tid) ||
5716 (entry->efd_group != new_entry->efd_group))
5718 if (entry->efd_start_cluster + entry->efd_count ==
5719 new_entry->efd_start_cluster) {
5720 new_entry->efd_start_cluster = entry->efd_start_cluster;
5721 new_entry->efd_count += entry->efd_count;
5722 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
5723 entry->efd_start_cluster) {
5724 new_entry->efd_count += entry->efd_count;
5727 spin_lock(&sbi->s_md_lock);
5728 list_del(&entry->efd_list);
5729 spin_unlock(&sbi->s_md_lock);
5730 rb_erase(&entry->efd_node, entry_rb_root);
5731 kmem_cache_free(ext4_free_data_cachep, entry);
5734 static noinline_for_stack int
5735 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
5736 struct ext4_free_data *new_entry)
5738 ext4_group_t group = e4b->bd_group;
5739 ext4_grpblk_t cluster;
5740 ext4_grpblk_t clusters = new_entry->efd_count;
5741 struct ext4_free_data *entry;
5742 struct ext4_group_info *db = e4b->bd_info;
5743 struct super_block *sb = e4b->bd_sb;
5744 struct ext4_sb_info *sbi = EXT4_SB(sb);
5745 struct rb_node **n = &db->bb_free_root.rb_node, *node;
5746 struct rb_node *parent = NULL, *new_node;
5748 BUG_ON(!ext4_handle_valid(handle));
5749 BUG_ON(e4b->bd_bitmap_page == NULL);
5750 BUG_ON(e4b->bd_buddy_page == NULL);
5752 new_node = &new_entry->efd_node;
5753 cluster = new_entry->efd_start_cluster;
5756 /* first free block exent. We need to
5757 protect buddy cache from being freed,
5758 * otherwise we'll refresh it from
5759 * on-disk bitmap and lose not-yet-available
5761 get_page(e4b->bd_buddy_page);
5762 get_page(e4b->bd_bitmap_page);
5766 entry = rb_entry(parent, struct ext4_free_data, efd_node);
5767 if (cluster < entry->efd_start_cluster)
5769 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
5770 n = &(*n)->rb_right;
5772 ext4_grp_locked_error(sb, group, 0,
5773 ext4_group_first_block_no(sb, group) +
5774 EXT4_C2B(sbi, cluster),
5775 "Block already on to-be-freed list");
5776 kmem_cache_free(ext4_free_data_cachep, new_entry);
5781 rb_link_node(new_node, parent, n);
5782 rb_insert_color(new_node, &db->bb_free_root);
5784 /* Now try to see the extent can be merged to left and right */
5785 node = rb_prev(new_node);
5787 entry = rb_entry(node, struct ext4_free_data, efd_node);
5788 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5789 &(db->bb_free_root));
5792 node = rb_next(new_node);
5794 entry = rb_entry(node, struct ext4_free_data, efd_node);
5795 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5796 &(db->bb_free_root));
5799 spin_lock(&sbi->s_md_lock);
5800 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5801 sbi->s_mb_free_pending += clusters;
5802 spin_unlock(&sbi->s_md_lock);
5807 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
5808 * linearly starting at the goal block and also excludes the blocks which
5809 * are going to be in use after fast commit replay.
5811 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5812 struct ext4_allocation_request *ar, int *errp)
5814 struct buffer_head *bitmap_bh;
5815 struct super_block *sb = ar->inode->i_sb;
5817 ext4_grpblk_t blkoff;
5818 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
5819 ext4_grpblk_t i = 0;
5820 ext4_fsblk_t goal, block;
5821 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
5824 if (goal < le32_to_cpu(es->s_first_data_block) ||
5825 goal >= ext4_blocks_count(es))
5826 goal = le32_to_cpu(es->s_first_data_block);
5829 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
5830 for (; group < ext4_get_groups_count(sb); group++) {
5831 bitmap_bh = ext4_read_block_bitmap(sb, group);
5832 if (IS_ERR(bitmap_bh)) {
5833 *errp = PTR_ERR(bitmap_bh);
5834 pr_warn("Failed to read block bitmap\n");
5838 ext4_get_group_no_and_offset(sb,
5839 max(ext4_group_first_block_no(sb, group), goal),
5842 i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
5846 if (ext4_fc_replay_check_excluded(sb,
5847 ext4_group_first_block_no(sb, group) + i)) {
5857 if (group >= ext4_get_groups_count(sb) || i >= max) {
5862 block = ext4_group_first_block_no(sb, group) + i;
5863 ext4_mb_mark_bb(sb, block, 1, 1);
5869 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
5870 unsigned long count)
5872 struct buffer_head *bitmap_bh;
5873 struct super_block *sb = inode->i_sb;
5874 struct ext4_group_desc *gdp;
5875 struct buffer_head *gdp_bh;
5877 ext4_grpblk_t blkoff;
5878 int already_freed = 0, err, i;
5880 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
5881 bitmap_bh = ext4_read_block_bitmap(sb, group);
5882 if (IS_ERR(bitmap_bh)) {
5883 err = PTR_ERR(bitmap_bh);
5884 pr_warn("Failed to read block bitmap\n");
5887 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
5891 for (i = 0; i < count; i++) {
5892 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data))
5895 mb_clear_bits(bitmap_bh->b_data, blkoff, count);
5896 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
5899 ext4_free_group_clusters_set(
5900 sb, gdp, ext4_free_group_clusters(sb, gdp) +
5901 count - already_freed);
5902 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
5903 ext4_group_desc_csum_set(sb, group, gdp);
5904 ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
5905 sync_dirty_buffer(bitmap_bh);
5906 sync_dirty_buffer(gdp_bh);
5911 * ext4_mb_clear_bb() -- helper function for freeing blocks.
5912 * Used by ext4_free_blocks()
5913 * @handle: handle for this transaction
5915 * @block: starting physical block to be freed
5916 * @count: number of blocks to be freed
5917 * @flags: flags used by ext4_free_blocks
5919 static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
5920 ext4_fsblk_t block, unsigned long count,
5923 struct buffer_head *bitmap_bh = NULL;
5924 struct super_block *sb = inode->i_sb;
5925 struct ext4_group_desc *gdp;
5926 unsigned int overflow;
5928 struct buffer_head *gd_bh;
5929 ext4_group_t block_group;
5930 struct ext4_sb_info *sbi;
5931 struct ext4_buddy e4b;
5932 unsigned int count_clusters;
5938 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5939 !ext4_inode_block_valid(inode, block, count)) {
5940 ext4_error(sb, "Freeing blocks in system zone - "
5941 "Block = %llu, count = %lu", block, count);
5942 /* err = 0. ext4_std_error should be a no op */
5945 flags |= EXT4_FREE_BLOCKS_VALIDATED;
5949 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5951 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5952 ext4_get_group_info(sb, block_group))))
5956 * Check to see if we are freeing blocks across a group
5959 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5960 overflow = EXT4_C2B(sbi, bit) + count -
5961 EXT4_BLOCKS_PER_GROUP(sb);
5963 /* The range changed so it's no longer validated */
5964 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
5966 count_clusters = EXT4_NUM_B2C(sbi, count);
5967 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5968 if (IS_ERR(bitmap_bh)) {
5969 err = PTR_ERR(bitmap_bh);
5973 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5979 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5980 !ext4_inode_block_valid(inode, block, count)) {
5981 ext4_error(sb, "Freeing blocks in system zone - "
5982 "Block = %llu, count = %lu", block, count);
5983 /* err = 0. ext4_std_error should be a no op */
5987 BUFFER_TRACE(bitmap_bh, "getting write access");
5988 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
5994 * We are about to modify some metadata. Call the journal APIs
5995 * to unshare ->b_data if a currently-committing transaction is
5998 BUFFER_TRACE(gd_bh, "get_write_access");
5999 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
6002 #ifdef AGGRESSIVE_CHECK
6005 for (i = 0; i < count_clusters; i++)
6006 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
6009 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
6011 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
6012 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
6013 GFP_NOFS|__GFP_NOFAIL);
6018 * We need to make sure we don't reuse the freed block until after the
6019 * transaction is committed. We make an exception if the inode is to be
6020 * written in writeback mode since writeback mode has weak data
6021 * consistency guarantees.
6023 if (ext4_handle_valid(handle) &&
6024 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
6025 !ext4_should_writeback_data(inode))) {
6026 struct ext4_free_data *new_entry;
6028 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
6031 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
6032 GFP_NOFS|__GFP_NOFAIL);
6033 new_entry->efd_start_cluster = bit;
6034 new_entry->efd_group = block_group;
6035 new_entry->efd_count = count_clusters;
6036 new_entry->efd_tid = handle->h_transaction->t_tid;
6038 ext4_lock_group(sb, block_group);
6039 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6040 ext4_mb_free_metadata(handle, &e4b, new_entry);
6042 /* need to update group_info->bb_free and bitmap
6043 * with group lock held. generate_buddy look at
6044 * them with group lock_held
6046 if (test_opt(sb, DISCARD)) {
6047 err = ext4_issue_discard(sb, block_group, bit, count,
6049 if (err && err != -EOPNOTSUPP)
6050 ext4_msg(sb, KERN_WARNING, "discard request in"
6051 " group:%u block:%d count:%lu failed"
6052 " with %d", block_group, bit, count,
6055 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
6057 ext4_lock_group(sb, block_group);
6058 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6059 mb_free_blocks(inode, &e4b, bit, count_clusters);
6062 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
6063 ext4_free_group_clusters_set(sb, gdp, ret);
6064 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
6065 ext4_group_desc_csum_set(sb, block_group, gdp);
6066 ext4_unlock_group(sb, block_group);
6068 if (sbi->s_log_groups_per_flex) {
6069 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6070 atomic64_add(count_clusters,
6071 &sbi_array_rcu_deref(sbi, s_flex_groups,
6072 flex_group)->free_clusters);
6076 * on a bigalloc file system, defer the s_freeclusters_counter
6077 * update to the caller (ext4_remove_space and friends) so they
6078 * can determine if a cluster freed here should be rereserved
6080 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6081 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6082 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6083 percpu_counter_add(&sbi->s_freeclusters_counter,
6087 ext4_mb_unload_buddy(&e4b);
6089 /* We dirtied the bitmap block */
6090 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6091 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6093 /* And the group descriptor block */
6094 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6095 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6099 if (overflow && !err) {
6103 /* The range changed so it's no longer validated */
6104 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6109 ext4_std_error(sb, err);
6114 * ext4_free_blocks() -- Free given blocks and update quota
6115 * @handle: handle for this transaction
6117 * @bh: optional buffer of the block to be freed
6118 * @block: starting physical block to be freed
6119 * @count: number of blocks to be freed
6120 * @flags: flags used by ext4_free_blocks
6122 void ext4_free_blocks(handle_t *handle, struct inode *inode,
6123 struct buffer_head *bh, ext4_fsblk_t block,
6124 unsigned long count, int flags)
6126 struct super_block *sb = inode->i_sb;
6127 unsigned int overflow;
6128 struct ext4_sb_info *sbi;
6132 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
6133 ext4_free_blocks_simple(inode, block, count);
6140 BUG_ON(block != bh->b_blocknr);
6142 block = bh->b_blocknr;
6145 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6146 !ext4_inode_block_valid(inode, block, count)) {
6147 ext4_error(sb, "Freeing blocks not in datazone - "
6148 "block = %llu, count = %lu", block, count);
6151 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6153 ext4_debug("freeing block %llu\n", block);
6154 trace_ext4_free_blocks(inode, block, count, flags);
6156 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6159 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
6164 * If the extent to be freed does not begin on a cluster
6165 * boundary, we need to deal with partial clusters at the
6166 * beginning and end of the extent. Normally we will free
6167 * blocks at the beginning or the end unless we are explicitly
6168 * requested to avoid doing so.
6170 overflow = EXT4_PBLK_COFF(sbi, block);
6172 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
6173 overflow = sbi->s_cluster_ratio - overflow;
6175 if (count > overflow)
6183 /* The range changed so it's no longer validated */
6184 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6186 overflow = EXT4_LBLK_COFF(sbi, count);
6188 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
6189 if (count > overflow)
6194 count += sbi->s_cluster_ratio - overflow;
6195 /* The range changed so it's no longer validated */
6196 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6199 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6201 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
6203 for (i = 0; i < count; i++) {
6206 bh = sb_find_get_block(inode->i_sb, block + i);
6207 ext4_forget(handle, is_metadata, inode, bh, block + i);
6211 ext4_mb_clear_bb(handle, inode, block, count, flags);
6216 * ext4_group_add_blocks() -- Add given blocks to an existing group
6217 * @handle: handle to this transaction
6219 * @block: start physical block to add to the block group
6220 * @count: number of blocks to free
6222 * This marks the blocks as free in the bitmap and buddy.
6224 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6225 ext4_fsblk_t block, unsigned long count)
6227 struct buffer_head *bitmap_bh = NULL;
6228 struct buffer_head *gd_bh;
6229 ext4_group_t block_group;
6232 struct ext4_group_desc *desc;
6233 struct ext4_sb_info *sbi = EXT4_SB(sb);
6234 struct ext4_buddy e4b;
6235 int err = 0, ret, free_clusters_count;
6236 ext4_grpblk_t clusters_freed;
6237 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6238 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6239 unsigned long cluster_count = last_cluster - first_cluster + 1;
6241 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6246 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6248 * Check to see if we are freeing blocks across a group
6251 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6252 ext4_warning(sb, "too many blocks added to group %u",
6258 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
6259 if (IS_ERR(bitmap_bh)) {
6260 err = PTR_ERR(bitmap_bh);
6265 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
6271 if (!ext4_sb_block_valid(sb, NULL, block, count)) {
6272 ext4_error(sb, "Adding blocks in system zones - "
6273 "Block = %llu, count = %lu",
6279 BUFFER_TRACE(bitmap_bh, "getting write access");
6280 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
6286 * We are about to modify some metadata. Call the journal APIs
6287 * to unshare ->b_data if a currently-committing transaction is
6290 BUFFER_TRACE(gd_bh, "get_write_access");
6291 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
6295 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
6296 BUFFER_TRACE(bitmap_bh, "clear bit");
6297 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
6298 ext4_error(sb, "bit already cleared for block %llu",
6299 (ext4_fsblk_t)(block + i));
6300 BUFFER_TRACE(bitmap_bh, "bit already cleared");
6306 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6311 * need to update group_info->bb_free and bitmap
6312 * with group lock held. generate_buddy look at
6313 * them with group lock_held
6315 ext4_lock_group(sb, block_group);
6316 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
6317 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6318 free_clusters_count = clusters_freed +
6319 ext4_free_group_clusters(sb, desc);
6320 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
6321 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
6322 ext4_group_desc_csum_set(sb, block_group, desc);
6323 ext4_unlock_group(sb, block_group);
6324 percpu_counter_add(&sbi->s_freeclusters_counter,
6327 if (sbi->s_log_groups_per_flex) {
6328 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6329 atomic64_add(clusters_freed,
6330 &sbi_array_rcu_deref(sbi, s_flex_groups,
6331 flex_group)->free_clusters);
6334 ext4_mb_unload_buddy(&e4b);
6336 /* We dirtied the bitmap block */
6337 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6338 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6340 /* And the group descriptor block */
6341 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6342 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6348 ext4_std_error(sb, err);
6353 * ext4_trim_extent -- function to TRIM one single free extent in the group
6354 * @sb: super block for the file system
6355 * @start: starting block of the free extent in the alloc. group
6356 * @count: number of blocks to TRIM
6357 * @e4b: ext4 buddy for the group
6359 * Trim "count" blocks starting at "start" in the "group". To assure that no
6360 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6361 * be called with under the group lock.
6363 static int ext4_trim_extent(struct super_block *sb,
6364 int start, int count, struct ext4_buddy *e4b)
6368 struct ext4_free_extent ex;
6369 ext4_group_t group = e4b->bd_group;
6372 trace_ext4_trim_extent(sb, group, start, count);
6374 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6376 ex.fe_start = start;
6377 ex.fe_group = group;
6381 * Mark blocks used, so no one can reuse them while
6384 mb_mark_used(e4b, &ex);
6385 ext4_unlock_group(sb, group);
6386 ret = ext4_issue_discard(sb, group, start, count, NULL);
6387 ext4_lock_group(sb, group);
6388 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6392 static int ext4_try_to_trim_range(struct super_block *sb,
6393 struct ext4_buddy *e4b, ext4_grpblk_t start,
6394 ext4_grpblk_t max, ext4_grpblk_t minblocks)
6395 __acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
6396 __releases(ext4_group_lock_ptr(sb, e4b->bd_group))
6398 ext4_grpblk_t next, count, free_count;
6401 bitmap = e4b->bd_bitmap;
6402 start = (e4b->bd_info->bb_first_free > start) ?
6403 e4b->bd_info->bb_first_free : start;
6407 while (start <= max) {
6408 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6411 next = mb_find_next_bit(bitmap, max + 1, start);
6413 if ((next - start) >= minblocks) {
6414 int ret = ext4_trim_extent(sb, start, next - start, e4b);
6416 if (ret && ret != -EOPNOTSUPP)
6418 count += next - start;
6420 free_count += next - start;
6423 if (fatal_signal_pending(current)) {
6424 count = -ERESTARTSYS;
6428 if (need_resched()) {
6429 ext4_unlock_group(sb, e4b->bd_group);
6431 ext4_lock_group(sb, e4b->bd_group);
6434 if ((e4b->bd_info->bb_free - free_count) < minblocks)
6442 * ext4_trim_all_free -- function to trim all free space in alloc. group
6443 * @sb: super block for file system
6444 * @group: group to be trimmed
6445 * @start: first group block to examine
6446 * @max: last group block to examine
6447 * @minblocks: minimum extent block count
6448 * @set_trimmed: set the trimmed flag if at least one block is trimmed
6450 * ext4_trim_all_free walks through group's block bitmap searching for free
6451 * extents. When the free extent is found, mark it as used in group buddy
6452 * bitmap. Then issue a TRIM command on this extent and free the extent in
6453 * the group buddy bitmap.
6455 static ext4_grpblk_t
6456 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6457 ext4_grpblk_t start, ext4_grpblk_t max,
6458 ext4_grpblk_t minblocks, bool set_trimmed)
6460 struct ext4_buddy e4b;
6463 trace_ext4_trim_all_free(sb, group, start, max);
6465 ret = ext4_mb_load_buddy(sb, group, &e4b);
6467 ext4_warning(sb, "Error %d loading buddy information for %u",
6472 ext4_lock_group(sb, group);
6474 if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
6475 minblocks < EXT4_SB(sb)->s_last_trim_minblks) {
6476 ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
6477 if (ret >= 0 && set_trimmed)
6478 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
6483 ext4_unlock_group(sb, group);
6484 ext4_mb_unload_buddy(&e4b);
6486 ext4_debug("trimmed %d blocks in the group %d\n",
6493 * ext4_trim_fs() -- trim ioctl handle function
6494 * @sb: superblock for filesystem
6495 * @range: fstrim_range structure
6497 * start: First Byte to trim
6498 * len: number of Bytes to trim from start
6499 * minlen: minimum extent length in Bytes
6500 * ext4_trim_fs goes through all allocation groups containing Bytes from
6501 * start to start+len. For each such a group ext4_trim_all_free function
6502 * is invoked to trim all free space.
6504 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6506 unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
6507 struct ext4_group_info *grp;
6508 ext4_group_t group, first_group, last_group;
6509 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6510 uint64_t start, end, minlen, trimmed = 0;
6511 ext4_fsblk_t first_data_blk =
6512 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6513 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6514 bool whole_group, eof = false;
6517 start = range->start >> sb->s_blocksize_bits;
6518 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6519 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6520 range->minlen >> sb->s_blocksize_bits);
6522 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6523 start >= max_blks ||
6524 range->len < sb->s_blocksize)
6526 /* No point to try to trim less than discard granularity */
6527 if (range->minlen < discard_granularity) {
6528 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6529 discard_granularity >> sb->s_blocksize_bits);
6530 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
6533 if (end >= max_blks - 1) {
6537 if (end <= first_data_blk)
6539 if (start < first_data_blk)
6540 start = first_data_blk;
6542 /* Determine first and last group to examine based on start and end */
6543 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6544 &first_group, &first_cluster);
6545 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6546 &last_group, &last_cluster);
6548 /* end now represents the last cluster to discard in this group */
6549 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6552 for (group = first_group; group <= last_group; group++) {
6553 grp = ext4_get_group_info(sb, group);
6554 /* We only do this if the grp has never been initialized */
6555 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6556 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6562 * For all the groups except the last one, last cluster will
6563 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6564 * change it for the last group, note that last_cluster is
6565 * already computed earlier by ext4_get_group_no_and_offset()
6567 if (group == last_group) {
6569 whole_group = eof ? true : end == EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6571 if (grp->bb_free >= minlen) {
6572 cnt = ext4_trim_all_free(sb, group, first_cluster,
6573 end, minlen, whole_group);
6582 * For every group except the first one, we are sure
6583 * that the first cluster to discard will be cluster #0.
6589 EXT4_SB(sb)->s_last_trim_minblks = minlen;
6592 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6596 /* Iterate all the free extents in the group. */
6598 ext4_mballoc_query_range(
6599 struct super_block *sb,
6601 ext4_grpblk_t start,
6603 ext4_mballoc_query_range_fn formatter,
6608 struct ext4_buddy e4b;
6611 error = ext4_mb_load_buddy(sb, group, &e4b);
6614 bitmap = e4b.bd_bitmap;
6616 ext4_lock_group(sb, group);
6618 start = (e4b.bd_info->bb_first_free > start) ?
6619 e4b.bd_info->bb_first_free : start;
6620 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
6621 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6623 while (start <= end) {
6624 start = mb_find_next_zero_bit(bitmap, end + 1, start);
6627 next = mb_find_next_bit(bitmap, end + 1, start);
6629 ext4_unlock_group(sb, group);
6630 error = formatter(sb, group, start, next - start, priv);
6633 ext4_lock_group(sb, group);
6638 ext4_unlock_group(sb, group);
6640 ext4_mb_unload_buddy(&e4b);