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);
412 * The algorithm using this percpu seq counter goes below:
413 * 1. We sample the percpu discard_pa_seq counter before trying for block
414 * allocation in ext4_mb_new_blocks().
415 * 2. We increment this percpu discard_pa_seq counter when we either allocate
416 * or free these blocks i.e. while marking those blocks as used/free in
417 * mb_mark_used()/mb_free_blocks().
418 * 3. We also increment this percpu seq counter when we successfully identify
419 * that the bb_prealloc_list is not empty and hence proceed for discarding
420 * of those PAs inside ext4_mb_discard_group_preallocations().
422 * Now to make sure that the regular fast path of block allocation is not
423 * affected, as a small optimization we only sample the percpu seq counter
424 * on that cpu. Only when the block allocation fails and when freed blocks
425 * found were 0, that is when we sample percpu seq counter for all cpus using
426 * below function ext4_get_discard_pa_seq_sum(). This happens after making
427 * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
429 static DEFINE_PER_CPU(u64, discard_pa_seq);
430 static inline u64 ext4_get_discard_pa_seq_sum(void)
435 for_each_possible_cpu(__cpu)
436 __seq += per_cpu(discard_pa_seq, __cpu);
440 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
442 #if BITS_PER_LONG == 64
443 *bit += ((unsigned long) addr & 7UL) << 3;
444 addr = (void *) ((unsigned long) addr & ~7UL);
445 #elif BITS_PER_LONG == 32
446 *bit += ((unsigned long) addr & 3UL) << 3;
447 addr = (void *) ((unsigned long) addr & ~3UL);
449 #error "how many bits you are?!"
454 static inline int mb_test_bit(int bit, void *addr)
457 * ext4_test_bit on architecture like powerpc
458 * needs unsigned long aligned address
460 addr = mb_correct_addr_and_bit(&bit, addr);
461 return ext4_test_bit(bit, addr);
464 static inline void mb_set_bit(int bit, void *addr)
466 addr = mb_correct_addr_and_bit(&bit, addr);
467 ext4_set_bit(bit, addr);
470 static inline void mb_clear_bit(int bit, void *addr)
472 addr = mb_correct_addr_and_bit(&bit, addr);
473 ext4_clear_bit(bit, addr);
476 static inline int mb_test_and_clear_bit(int bit, void *addr)
478 addr = mb_correct_addr_and_bit(&bit, addr);
479 return ext4_test_and_clear_bit(bit, addr);
482 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
484 int fix = 0, ret, tmpmax;
485 addr = mb_correct_addr_and_bit(&fix, addr);
489 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
495 static inline int mb_find_next_bit(void *addr, int max, int start)
497 int fix = 0, ret, tmpmax;
498 addr = mb_correct_addr_and_bit(&fix, addr);
502 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
508 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
512 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
515 if (order > e4b->bd_blkbits + 1) {
520 /* at order 0 we see each particular block */
522 *max = 1 << (e4b->bd_blkbits + 3);
523 return e4b->bd_bitmap;
526 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
527 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
533 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
534 int first, int count)
537 struct super_block *sb = e4b->bd_sb;
539 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
541 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
542 for (i = 0; i < count; i++) {
543 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
544 ext4_fsblk_t blocknr;
546 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
547 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
548 ext4_grp_locked_error(sb, e4b->bd_group,
549 inode ? inode->i_ino : 0,
551 "freeing block already freed "
554 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
555 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
557 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
561 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
565 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
567 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
568 for (i = 0; i < count; i++) {
569 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
570 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
574 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
576 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
578 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
579 unsigned char *b1, *b2;
581 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
582 b2 = (unsigned char *) bitmap;
583 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
584 if (b1[i] != b2[i]) {
585 ext4_msg(e4b->bd_sb, KERN_ERR,
586 "corruption in group %u "
587 "at byte %u(%u): %x in copy != %x "
589 e4b->bd_group, i, i * 8, b1[i], b2[i]);
596 static void mb_group_bb_bitmap_alloc(struct super_block *sb,
597 struct ext4_group_info *grp, ext4_group_t group)
599 struct buffer_head *bh;
601 grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
605 bh = ext4_read_block_bitmap(sb, group);
606 if (IS_ERR_OR_NULL(bh)) {
607 kfree(grp->bb_bitmap);
608 grp->bb_bitmap = NULL;
612 memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
616 static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
618 kfree(grp->bb_bitmap);
622 static inline void mb_free_blocks_double(struct inode *inode,
623 struct ext4_buddy *e4b, int first, int count)
627 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
628 int first, int count)
632 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
637 static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
638 struct ext4_group_info *grp, ext4_group_t group)
643 static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
649 #ifdef AGGRESSIVE_CHECK
651 #define MB_CHECK_ASSERT(assert) \
655 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
656 function, file, line, # assert); \
661 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
662 const char *function, int line)
664 struct super_block *sb = e4b->bd_sb;
665 int order = e4b->bd_blkbits + 1;
672 struct ext4_group_info *grp;
675 struct list_head *cur;
679 if (e4b->bd_info->bb_check_counter++ % 10)
683 buddy = mb_find_buddy(e4b, order, &max);
684 MB_CHECK_ASSERT(buddy);
685 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
686 MB_CHECK_ASSERT(buddy2);
687 MB_CHECK_ASSERT(buddy != buddy2);
688 MB_CHECK_ASSERT(max * 2 == max2);
691 for (i = 0; i < max; i++) {
693 if (mb_test_bit(i, buddy)) {
694 /* only single bit in buddy2 may be 1 */
695 if (!mb_test_bit(i << 1, buddy2)) {
697 mb_test_bit((i<<1)+1, buddy2));
698 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
700 mb_test_bit(i << 1, buddy2));
705 /* both bits in buddy2 must be 1 */
706 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
707 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
709 for (j = 0; j < (1 << order); j++) {
710 k = (i * (1 << order)) + j;
712 !mb_test_bit(k, e4b->bd_bitmap));
716 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
721 buddy = mb_find_buddy(e4b, 0, &max);
722 for (i = 0; i < max; i++) {
723 if (!mb_test_bit(i, buddy)) {
724 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
732 /* check used bits only */
733 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
734 buddy2 = mb_find_buddy(e4b, j, &max2);
736 MB_CHECK_ASSERT(k < max2);
737 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
740 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
741 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
743 grp = ext4_get_group_info(sb, e4b->bd_group);
744 list_for_each(cur, &grp->bb_prealloc_list) {
745 ext4_group_t groupnr;
746 struct ext4_prealloc_space *pa;
747 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
748 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
749 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
750 for (i = 0; i < pa->pa_len; i++)
751 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
755 #undef MB_CHECK_ASSERT
756 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
757 __FILE__, __func__, __LINE__)
759 #define mb_check_buddy(e4b)
763 * Divide blocks started from @first with length @len into
764 * smaller chunks with power of 2 blocks.
765 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
766 * then increase bb_counters[] for corresponded chunk size.
768 static void ext4_mb_mark_free_simple(struct super_block *sb,
769 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
770 struct ext4_group_info *grp)
772 struct ext4_sb_info *sbi = EXT4_SB(sb);
778 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
780 border = 2 << sb->s_blocksize_bits;
783 /* find how many blocks can be covered since this position */
784 max = ffs(first | border) - 1;
786 /* find how many blocks of power 2 we need to mark */
793 /* mark multiblock chunks only */
794 grp->bb_counters[min]++;
796 mb_clear_bit(first >> min,
797 buddy + sbi->s_mb_offsets[min]);
804 static void ext4_mb_rb_insert(struct rb_root *root, struct rb_node *new,
805 int (*cmp)(struct rb_node *, struct rb_node *))
807 struct rb_node **iter = &root->rb_node, *parent = NULL;
811 if (cmp(new, *iter) > 0)
812 iter = &((*iter)->rb_left);
814 iter = &((*iter)->rb_right);
817 rb_link_node(new, parent, iter);
818 rb_insert_color(new, root);
822 ext4_mb_avg_fragment_size_cmp(struct rb_node *rb1, struct rb_node *rb2)
824 struct ext4_group_info *grp1 = rb_entry(rb1,
825 struct ext4_group_info,
826 bb_avg_fragment_size_rb);
827 struct ext4_group_info *grp2 = rb_entry(rb2,
828 struct ext4_group_info,
829 bb_avg_fragment_size_rb);
830 int num_frags_1, num_frags_2;
832 num_frags_1 = grp1->bb_fragments ?
833 grp1->bb_free / grp1->bb_fragments : 0;
834 num_frags_2 = grp2->bb_fragments ?
835 grp2->bb_free / grp2->bb_fragments : 0;
837 return (num_frags_2 - num_frags_1);
841 * Reinsert grpinfo into the avg_fragment_size tree with new average
845 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
847 struct ext4_sb_info *sbi = EXT4_SB(sb);
849 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0)
852 write_lock(&sbi->s_mb_rb_lock);
853 if (!RB_EMPTY_NODE(&grp->bb_avg_fragment_size_rb)) {
854 rb_erase(&grp->bb_avg_fragment_size_rb,
855 &sbi->s_mb_avg_fragment_size_root);
856 RB_CLEAR_NODE(&grp->bb_avg_fragment_size_rb);
859 ext4_mb_rb_insert(&sbi->s_mb_avg_fragment_size_root,
860 &grp->bb_avg_fragment_size_rb,
861 ext4_mb_avg_fragment_size_cmp);
862 write_unlock(&sbi->s_mb_rb_lock);
866 * Choose next group by traversing largest_free_order lists. Updates *new_cr if
867 * cr level needs an update.
869 static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac,
870 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
872 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
873 struct ext4_group_info *iter, *grp;
876 if (ac->ac_status == AC_STATUS_FOUND)
879 if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMIZED))
880 atomic_inc(&sbi->s_bal_cr0_bad_suggestions);
883 for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) {
884 if (list_empty(&sbi->s_mb_largest_free_orders[i]))
886 read_lock(&sbi->s_mb_largest_free_orders_locks[i]);
887 if (list_empty(&sbi->s_mb_largest_free_orders[i])) {
888 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
892 list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i],
893 bb_largest_free_order_node) {
895 atomic64_inc(&sbi->s_bal_cX_groups_considered[0]);
896 if (likely(ext4_mb_good_group(ac, iter->bb_group, 0))) {
901 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
907 /* Increment cr and search again */
910 *group = grp->bb_group;
911 ac->ac_last_optimal_group = *group;
912 ac->ac_flags |= EXT4_MB_CR0_OPTIMIZED;
917 * Choose next group by traversing average fragment size tree. Updates *new_cr
918 * if cr lvel needs an update. Sets EXT4_MB_SEARCH_NEXT_LINEAR to indicate that
919 * the linear search should continue for one iteration since there's lock
920 * contention on the rb tree lock.
922 static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac,
923 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
925 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
926 int avg_fragment_size, best_so_far;
927 struct rb_node *node, *found;
928 struct ext4_group_info *grp;
931 * If there is contention on the lock, instead of waiting for the lock
932 * to become available, just continue searching lineraly. We'll resume
933 * our rb tree search later starting at ac->ac_last_optimal_group.
935 if (!read_trylock(&sbi->s_mb_rb_lock)) {
936 ac->ac_flags |= EXT4_MB_SEARCH_NEXT_LINEAR;
940 if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) {
942 atomic_inc(&sbi->s_bal_cr1_bad_suggestions);
943 /* We have found something at CR 1 in the past */
944 grp = ext4_get_group_info(ac->ac_sb, ac->ac_last_optimal_group);
945 for (found = rb_next(&grp->bb_avg_fragment_size_rb); found != NULL;
946 found = rb_next(found)) {
947 grp = rb_entry(found, struct ext4_group_info,
948 bb_avg_fragment_size_rb);
950 atomic64_inc(&sbi->s_bal_cX_groups_considered[1]);
951 if (likely(ext4_mb_good_group(ac, grp->bb_group, 1)))
957 node = sbi->s_mb_avg_fragment_size_root.rb_node;
962 grp = rb_entry(node, struct ext4_group_info,
963 bb_avg_fragment_size_rb);
964 avg_fragment_size = 0;
965 if (ext4_mb_good_group(ac, grp->bb_group, 1)) {
966 avg_fragment_size = grp->bb_fragments ?
967 grp->bb_free / grp->bb_fragments : 0;
968 if (!best_so_far || avg_fragment_size < best_so_far) {
969 best_so_far = avg_fragment_size;
973 if (avg_fragment_size > ac->ac_g_ex.fe_len)
974 node = node->rb_right;
976 node = node->rb_left;
981 grp = rb_entry(found, struct ext4_group_info,
982 bb_avg_fragment_size_rb);
983 *group = grp->bb_group;
984 ac->ac_flags |= EXT4_MB_CR1_OPTIMIZED;
989 read_unlock(&sbi->s_mb_rb_lock);
990 ac->ac_last_optimal_group = *group;
993 static inline int should_optimize_scan(struct ext4_allocation_context *ac)
995 if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN)))
997 if (ac->ac_criteria >= 2)
999 if (ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))
1005 * Return next linear group for allocation. If linear traversal should not be
1006 * performed, this function just returns the same group
1009 next_linear_group(struct ext4_allocation_context *ac, int group, int ngroups)
1011 if (!should_optimize_scan(ac))
1012 goto inc_and_return;
1014 if (ac->ac_groups_linear_remaining) {
1015 ac->ac_groups_linear_remaining--;
1016 goto inc_and_return;
1019 if (ac->ac_flags & EXT4_MB_SEARCH_NEXT_LINEAR) {
1020 ac->ac_flags &= ~EXT4_MB_SEARCH_NEXT_LINEAR;
1021 goto inc_and_return;
1027 * Artificially restricted ngroups for non-extent
1028 * files makes group > ngroups possible on first loop.
1030 return group + 1 >= ngroups ? 0 : group + 1;
1034 * ext4_mb_choose_next_group: choose next group for allocation.
1036 * @ac Allocation Context
1037 * @new_cr This is an output parameter. If the there is no good group
1038 * available at current CR level, this field is updated to indicate
1039 * the new cr level that should be used.
1040 * @group This is an input / output parameter. As an input it indicates the
1041 * next group that the allocator intends to use for allocation. As
1042 * output, this field indicates the next group that should be used as
1043 * determined by the optimization functions.
1044 * @ngroups Total number of groups
1046 static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac,
1047 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
1049 *new_cr = ac->ac_criteria;
1051 if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining)
1055 ext4_mb_choose_next_group_cr0(ac, new_cr, group, ngroups);
1056 } else if (*new_cr == 1) {
1057 ext4_mb_choose_next_group_cr1(ac, new_cr, group, ngroups);
1060 * TODO: For CR=2, we can arrange groups in an rb tree sorted by
1061 * bb_free. But until that happens, we should never come here.
1068 * Cache the order of the largest free extent we have available in this block
1072 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
1074 struct ext4_sb_info *sbi = EXT4_SB(sb);
1077 if (test_opt2(sb, MB_OPTIMIZE_SCAN) && grp->bb_largest_free_order >= 0) {
1078 write_lock(&sbi->s_mb_largest_free_orders_locks[
1079 grp->bb_largest_free_order]);
1080 list_del_init(&grp->bb_largest_free_order_node);
1081 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1082 grp->bb_largest_free_order]);
1084 grp->bb_largest_free_order = -1; /* uninit */
1086 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) {
1087 if (grp->bb_counters[i] > 0) {
1088 grp->bb_largest_free_order = i;
1092 if (test_opt2(sb, MB_OPTIMIZE_SCAN) &&
1093 grp->bb_largest_free_order >= 0 && grp->bb_free) {
1094 write_lock(&sbi->s_mb_largest_free_orders_locks[
1095 grp->bb_largest_free_order]);
1096 list_add_tail(&grp->bb_largest_free_order_node,
1097 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1098 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1099 grp->bb_largest_free_order]);
1103 static noinline_for_stack
1104 void ext4_mb_generate_buddy(struct super_block *sb,
1105 void *buddy, void *bitmap, ext4_group_t group)
1107 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
1108 struct ext4_sb_info *sbi = EXT4_SB(sb);
1109 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1110 ext4_grpblk_t i = 0;
1111 ext4_grpblk_t first;
1114 unsigned fragments = 0;
1115 unsigned long long period = get_cycles();
1117 /* initialize buddy from bitmap which is aggregation
1118 * of on-disk bitmap and preallocations */
1119 i = mb_find_next_zero_bit(bitmap, max, 0);
1120 grp->bb_first_free = i;
1124 i = mb_find_next_bit(bitmap, max, i);
1128 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1130 grp->bb_counters[0]++;
1132 i = mb_find_next_zero_bit(bitmap, max, i);
1134 grp->bb_fragments = fragments;
1136 if (free != grp->bb_free) {
1137 ext4_grp_locked_error(sb, group, 0, 0,
1138 "block bitmap and bg descriptor "
1139 "inconsistent: %u vs %u free clusters",
1140 free, grp->bb_free);
1142 * If we intend to continue, we consider group descriptor
1143 * corrupt and update bb_free using bitmap value
1145 grp->bb_free = free;
1146 ext4_mark_group_bitmap_corrupted(sb, group,
1147 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1149 mb_set_largest_free_order(sb, grp);
1151 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1153 period = get_cycles() - period;
1154 atomic_inc(&sbi->s_mb_buddies_generated);
1155 atomic64_add(period, &sbi->s_mb_generation_time);
1156 mb_update_avg_fragment_size(sb, grp);
1159 /* The buddy information is attached the buddy cache inode
1160 * for convenience. The information regarding each group
1161 * is loaded via ext4_mb_load_buddy. The information involve
1162 * block bitmap and buddy information. The information are
1163 * stored in the inode as
1166 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1169 * one block each for bitmap and buddy information.
1170 * So for each group we take up 2 blocks. A page can
1171 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
1172 * So it can have information regarding groups_per_page which
1173 * is blocks_per_page/2
1175 * Locking note: This routine takes the block group lock of all groups
1176 * for this page; do not hold this lock when calling this routine!
1179 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
1181 ext4_group_t ngroups;
1183 int blocks_per_page;
1184 int groups_per_page;
1187 ext4_group_t first_group, group;
1189 struct super_block *sb;
1190 struct buffer_head *bhs;
1191 struct buffer_head **bh = NULL;
1192 struct inode *inode;
1195 struct ext4_group_info *grinfo;
1197 inode = page->mapping->host;
1199 ngroups = ext4_get_groups_count(sb);
1200 blocksize = i_blocksize(inode);
1201 blocks_per_page = PAGE_SIZE / blocksize;
1203 mb_debug(sb, "init page %lu\n", page->index);
1205 groups_per_page = blocks_per_page >> 1;
1206 if (groups_per_page == 0)
1207 groups_per_page = 1;
1209 /* allocate buffer_heads to read bitmaps */
1210 if (groups_per_page > 1) {
1211 i = sizeof(struct buffer_head *) * groups_per_page;
1212 bh = kzalloc(i, gfp);
1220 first_group = page->index * blocks_per_page / 2;
1222 /* read all groups the page covers into the cache */
1223 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1224 if (group >= ngroups)
1227 grinfo = ext4_get_group_info(sb, group);
1229 * If page is uptodate then we came here after online resize
1230 * which added some new uninitialized group info structs, so
1231 * we must skip all initialized uptodate buddies on the page,
1232 * which may be currently in use by an allocating task.
1234 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1238 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1239 if (IS_ERR(bh[i])) {
1240 err = PTR_ERR(bh[i]);
1244 mb_debug(sb, "read bitmap for group %u\n", group);
1247 /* wait for I/O completion */
1248 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1253 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1258 first_block = page->index * blocks_per_page;
1259 for (i = 0; i < blocks_per_page; i++) {
1260 group = (first_block + i) >> 1;
1261 if (group >= ngroups)
1264 if (!bh[group - first_group])
1265 /* skip initialized uptodate buddy */
1268 if (!buffer_verified(bh[group - first_group]))
1269 /* Skip faulty bitmaps */
1274 * data carry information regarding this
1275 * particular group in the format specified
1279 data = page_address(page) + (i * blocksize);
1280 bitmap = bh[group - first_group]->b_data;
1283 * We place the buddy block and bitmap block
1286 if ((first_block + i) & 1) {
1287 /* this is block of buddy */
1288 BUG_ON(incore == NULL);
1289 mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
1290 group, page->index, i * blocksize);
1291 trace_ext4_mb_buddy_bitmap_load(sb, group);
1292 grinfo = ext4_get_group_info(sb, group);
1293 grinfo->bb_fragments = 0;
1294 memset(grinfo->bb_counters, 0,
1295 sizeof(*grinfo->bb_counters) *
1296 (MB_NUM_ORDERS(sb)));
1298 * incore got set to the group block bitmap below
1300 ext4_lock_group(sb, group);
1301 /* init the buddy */
1302 memset(data, 0xff, blocksize);
1303 ext4_mb_generate_buddy(sb, data, incore, group);
1304 ext4_unlock_group(sb, group);
1307 /* this is block of bitmap */
1308 BUG_ON(incore != NULL);
1309 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
1310 group, page->index, i * blocksize);
1311 trace_ext4_mb_bitmap_load(sb, group);
1313 /* see comments in ext4_mb_put_pa() */
1314 ext4_lock_group(sb, group);
1315 memcpy(data, bitmap, blocksize);
1317 /* mark all preallocated blks used in in-core bitmap */
1318 ext4_mb_generate_from_pa(sb, data, group);
1319 ext4_mb_generate_from_freelist(sb, data, group);
1320 ext4_unlock_group(sb, group);
1322 /* set incore so that the buddy information can be
1323 * generated using this
1328 SetPageUptodate(page);
1332 for (i = 0; i < groups_per_page; i++)
1341 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1342 * on the same buddy page doesn't happen whild holding the buddy page lock.
1343 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1344 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1346 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1347 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1349 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1350 int block, pnum, poff;
1351 int blocks_per_page;
1354 e4b->bd_buddy_page = NULL;
1355 e4b->bd_bitmap_page = NULL;
1357 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1359 * the buddy cache inode stores the block bitmap
1360 * and buddy information in consecutive blocks.
1361 * So for each group we need two blocks.
1364 pnum = block / blocks_per_page;
1365 poff = block % blocks_per_page;
1366 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1369 BUG_ON(page->mapping != inode->i_mapping);
1370 e4b->bd_bitmap_page = page;
1371 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1373 if (blocks_per_page >= 2) {
1374 /* buddy and bitmap are on the same page */
1379 pnum = block / blocks_per_page;
1380 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1383 BUG_ON(page->mapping != inode->i_mapping);
1384 e4b->bd_buddy_page = page;
1388 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1390 if (e4b->bd_bitmap_page) {
1391 unlock_page(e4b->bd_bitmap_page);
1392 put_page(e4b->bd_bitmap_page);
1394 if (e4b->bd_buddy_page) {
1395 unlock_page(e4b->bd_buddy_page);
1396 put_page(e4b->bd_buddy_page);
1401 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1402 * block group lock of all groups for this page; do not hold the BG lock when
1403 * calling this routine!
1405 static noinline_for_stack
1406 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1409 struct ext4_group_info *this_grp;
1410 struct ext4_buddy e4b;
1415 mb_debug(sb, "init group %u\n", group);
1416 this_grp = ext4_get_group_info(sb, group);
1418 * This ensures that we don't reinit the buddy cache
1419 * page which map to the group from which we are already
1420 * allocating. If we are looking at the buddy cache we would
1421 * have taken a reference using ext4_mb_load_buddy and that
1422 * would have pinned buddy page to page cache.
1423 * The call to ext4_mb_get_buddy_page_lock will mark the
1426 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1427 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1429 * somebody initialized the group
1430 * return without doing anything
1435 page = e4b.bd_bitmap_page;
1436 ret = ext4_mb_init_cache(page, NULL, gfp);
1439 if (!PageUptodate(page)) {
1444 if (e4b.bd_buddy_page == NULL) {
1446 * If both the bitmap and buddy are in
1447 * the same page we don't need to force
1453 /* init buddy cache */
1454 page = e4b.bd_buddy_page;
1455 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1458 if (!PageUptodate(page)) {
1463 ext4_mb_put_buddy_page_lock(&e4b);
1468 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1469 * block group lock of all groups for this page; do not hold the BG lock when
1470 * calling this routine!
1472 static noinline_for_stack int
1473 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1474 struct ext4_buddy *e4b, gfp_t gfp)
1476 int blocks_per_page;
1482 struct ext4_group_info *grp;
1483 struct ext4_sb_info *sbi = EXT4_SB(sb);
1484 struct inode *inode = sbi->s_buddy_cache;
1487 mb_debug(sb, "load group %u\n", group);
1489 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1490 grp = ext4_get_group_info(sb, group);
1492 e4b->bd_blkbits = sb->s_blocksize_bits;
1495 e4b->bd_group = group;
1496 e4b->bd_buddy_page = NULL;
1497 e4b->bd_bitmap_page = NULL;
1499 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1501 * we need full data about the group
1502 * to make a good selection
1504 ret = ext4_mb_init_group(sb, group, gfp);
1510 * the buddy cache inode stores the block bitmap
1511 * and buddy information in consecutive blocks.
1512 * So for each group we need two blocks.
1515 pnum = block / blocks_per_page;
1516 poff = block % blocks_per_page;
1518 /* we could use find_or_create_page(), but it locks page
1519 * what we'd like to avoid in fast path ... */
1520 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1521 if (page == NULL || !PageUptodate(page)) {
1524 * drop the page reference and try
1525 * to get the page with lock. If we
1526 * are not uptodate that implies
1527 * somebody just created the page but
1528 * is yet to initialize the same. So
1529 * wait for it to initialize.
1532 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1534 BUG_ON(page->mapping != inode->i_mapping);
1535 if (!PageUptodate(page)) {
1536 ret = ext4_mb_init_cache(page, NULL, gfp);
1541 mb_cmp_bitmaps(e4b, page_address(page) +
1542 (poff * sb->s_blocksize));
1551 if (!PageUptodate(page)) {
1556 /* Pages marked accessed already */
1557 e4b->bd_bitmap_page = page;
1558 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1561 pnum = block / blocks_per_page;
1562 poff = block % blocks_per_page;
1564 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1565 if (page == NULL || !PageUptodate(page)) {
1568 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1570 BUG_ON(page->mapping != inode->i_mapping);
1571 if (!PageUptodate(page)) {
1572 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1586 if (!PageUptodate(page)) {
1591 /* Pages marked accessed already */
1592 e4b->bd_buddy_page = page;
1593 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1600 if (e4b->bd_bitmap_page)
1601 put_page(e4b->bd_bitmap_page);
1602 if (e4b->bd_buddy_page)
1603 put_page(e4b->bd_buddy_page);
1604 e4b->bd_buddy = NULL;
1605 e4b->bd_bitmap = NULL;
1609 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1610 struct ext4_buddy *e4b)
1612 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1615 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1617 if (e4b->bd_bitmap_page)
1618 put_page(e4b->bd_bitmap_page);
1619 if (e4b->bd_buddy_page)
1620 put_page(e4b->bd_buddy_page);
1624 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1629 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1630 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1632 while (order <= e4b->bd_blkbits + 1) {
1633 bb = mb_find_buddy(e4b, order, &max);
1634 if (!mb_test_bit(block >> order, bb)) {
1635 /* this block is part of buddy of order 'order' */
1643 static void mb_clear_bits(void *bm, int cur, int len)
1649 if ((cur & 31) == 0 && (len - cur) >= 32) {
1650 /* fast path: clear whole word at once */
1651 addr = bm + (cur >> 3);
1656 mb_clear_bit(cur, bm);
1661 /* clear bits in given range
1662 * will return first found zero bit if any, -1 otherwise
1664 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1671 if ((cur & 31) == 0 && (len - cur) >= 32) {
1672 /* fast path: clear whole word at once */
1673 addr = bm + (cur >> 3);
1674 if (*addr != (__u32)(-1) && zero_bit == -1)
1675 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1680 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1688 void ext4_set_bits(void *bm, int cur, int len)
1694 if ((cur & 31) == 0 && (len - cur) >= 32) {
1695 /* fast path: set whole word at once */
1696 addr = bm + (cur >> 3);
1701 mb_set_bit(cur, bm);
1706 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1708 if (mb_test_bit(*bit + side, bitmap)) {
1709 mb_clear_bit(*bit, bitmap);
1715 mb_set_bit(*bit, bitmap);
1720 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1724 void *buddy = mb_find_buddy(e4b, order, &max);
1729 /* Bits in range [first; last] are known to be set since
1730 * corresponding blocks were allocated. Bits in range
1731 * (first; last) will stay set because they form buddies on
1732 * upper layer. We just deal with borders if they don't
1733 * align with upper layer and then go up.
1734 * Releasing entire group is all about clearing
1735 * single bit of highest order buddy.
1739 * ---------------------------------
1741 * ---------------------------------
1742 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1743 * ---------------------------------
1745 * \_____________________/
1747 * Neither [1] nor [6] is aligned to above layer.
1748 * Left neighbour [0] is free, so mark it busy,
1749 * decrease bb_counters and extend range to
1751 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1752 * mark [6] free, increase bb_counters and shrink range to
1754 * Then shift range to [0; 2], go up and do the same.
1759 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1761 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1766 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1767 mb_clear_bits(buddy, first, last - first + 1);
1768 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1777 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1778 int first, int count)
1780 int left_is_free = 0;
1781 int right_is_free = 0;
1783 int last = first + count - 1;
1784 struct super_block *sb = e4b->bd_sb;
1786 if (WARN_ON(count == 0))
1788 BUG_ON(last >= (sb->s_blocksize << 3));
1789 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1790 /* Don't bother if the block group is corrupt. */
1791 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1794 mb_check_buddy(e4b);
1795 mb_free_blocks_double(inode, e4b, first, count);
1797 this_cpu_inc(discard_pa_seq);
1798 e4b->bd_info->bb_free += count;
1799 if (first < e4b->bd_info->bb_first_free)
1800 e4b->bd_info->bb_first_free = first;
1802 /* access memory sequentially: check left neighbour,
1803 * clear range and then check right neighbour
1806 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1807 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1808 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1809 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1811 if (unlikely(block != -1)) {
1812 struct ext4_sb_info *sbi = EXT4_SB(sb);
1813 ext4_fsblk_t blocknr;
1815 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1816 blocknr += EXT4_C2B(sbi, block);
1817 if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) {
1818 ext4_grp_locked_error(sb, e4b->bd_group,
1819 inode ? inode->i_ino : 0,
1821 "freeing already freed block (bit %u); block bitmap corrupt.",
1823 ext4_mark_group_bitmap_corrupted(
1825 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1830 /* let's maintain fragments counter */
1831 if (left_is_free && right_is_free)
1832 e4b->bd_info->bb_fragments--;
1833 else if (!left_is_free && !right_is_free)
1834 e4b->bd_info->bb_fragments++;
1836 /* buddy[0] == bd_bitmap is a special case, so handle
1837 * it right away and let mb_buddy_mark_free stay free of
1838 * zero order checks.
1839 * Check if neighbours are to be coaleasced,
1840 * adjust bitmap bb_counters and borders appropriately.
1843 first += !left_is_free;
1844 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1847 last -= !right_is_free;
1848 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1852 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1855 mb_set_largest_free_order(sb, e4b->bd_info);
1856 mb_update_avg_fragment_size(sb, e4b->bd_info);
1857 mb_check_buddy(e4b);
1860 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1861 int needed, struct ext4_free_extent *ex)
1867 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1870 buddy = mb_find_buddy(e4b, 0, &max);
1871 BUG_ON(buddy == NULL);
1872 BUG_ON(block >= max);
1873 if (mb_test_bit(block, buddy)) {
1880 /* find actual order */
1881 order = mb_find_order_for_block(e4b, block);
1882 block = block >> order;
1884 ex->fe_len = 1 << order;
1885 ex->fe_start = block << order;
1886 ex->fe_group = e4b->bd_group;
1888 /* calc difference from given start */
1889 next = next - ex->fe_start;
1891 ex->fe_start += next;
1893 while (needed > ex->fe_len &&
1894 mb_find_buddy(e4b, order, &max)) {
1896 if (block + 1 >= max)
1899 next = (block + 1) * (1 << order);
1900 if (mb_test_bit(next, e4b->bd_bitmap))
1903 order = mb_find_order_for_block(e4b, next);
1905 block = next >> order;
1906 ex->fe_len += 1 << order;
1909 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1910 /* Should never happen! (but apparently sometimes does?!?) */
1912 ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent "
1913 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1914 block, order, needed, ex->fe_group, ex->fe_start,
1915 ex->fe_len, ex->fe_logical);
1923 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1929 int start = ex->fe_start;
1930 int len = ex->fe_len;
1935 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1936 BUG_ON(e4b->bd_group != ex->fe_group);
1937 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1938 mb_check_buddy(e4b);
1939 mb_mark_used_double(e4b, start, len);
1941 this_cpu_inc(discard_pa_seq);
1942 e4b->bd_info->bb_free -= len;
1943 if (e4b->bd_info->bb_first_free == start)
1944 e4b->bd_info->bb_first_free += len;
1946 /* let's maintain fragments counter */
1948 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1949 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1950 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1952 e4b->bd_info->bb_fragments++;
1953 else if (!mlen && !max)
1954 e4b->bd_info->bb_fragments--;
1956 /* let's maintain buddy itself */
1958 ord = mb_find_order_for_block(e4b, start);
1960 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1961 /* the whole chunk may be allocated at once! */
1963 buddy = mb_find_buddy(e4b, ord, &max);
1964 BUG_ON((start >> ord) >= max);
1965 mb_set_bit(start >> ord, buddy);
1966 e4b->bd_info->bb_counters[ord]--;
1973 /* store for history */
1975 ret = len | (ord << 16);
1977 /* we have to split large buddy */
1979 buddy = mb_find_buddy(e4b, ord, &max);
1980 mb_set_bit(start >> ord, buddy);
1981 e4b->bd_info->bb_counters[ord]--;
1984 cur = (start >> ord) & ~1U;
1985 buddy = mb_find_buddy(e4b, ord, &max);
1986 mb_clear_bit(cur, buddy);
1987 mb_clear_bit(cur + 1, buddy);
1988 e4b->bd_info->bb_counters[ord]++;
1989 e4b->bd_info->bb_counters[ord]++;
1991 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1993 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
1994 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1995 mb_check_buddy(e4b);
2001 * Must be called under group lock!
2003 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
2004 struct ext4_buddy *e4b)
2006 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2009 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
2010 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2012 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
2013 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
2014 ret = mb_mark_used(e4b, &ac->ac_b_ex);
2016 /* preallocation can change ac_b_ex, thus we store actually
2017 * allocated blocks for history */
2018 ac->ac_f_ex = ac->ac_b_ex;
2020 ac->ac_status = AC_STATUS_FOUND;
2021 ac->ac_tail = ret & 0xffff;
2022 ac->ac_buddy = ret >> 16;
2025 * take the page reference. We want the page to be pinned
2026 * so that we don't get a ext4_mb_init_cache_call for this
2027 * group until we update the bitmap. That would mean we
2028 * double allocate blocks. The reference is dropped
2029 * in ext4_mb_release_context
2031 ac->ac_bitmap_page = e4b->bd_bitmap_page;
2032 get_page(ac->ac_bitmap_page);
2033 ac->ac_buddy_page = e4b->bd_buddy_page;
2034 get_page(ac->ac_buddy_page);
2035 /* store last allocated for subsequent stream allocation */
2036 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2037 spin_lock(&sbi->s_md_lock);
2038 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2039 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2040 spin_unlock(&sbi->s_md_lock);
2043 * As we've just preallocated more space than
2044 * user requested originally, we store allocated
2045 * space in a special descriptor.
2047 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2048 ext4_mb_new_preallocation(ac);
2052 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2053 struct ext4_buddy *e4b,
2056 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2057 struct ext4_free_extent *bex = &ac->ac_b_ex;
2058 struct ext4_free_extent *gex = &ac->ac_g_ex;
2059 struct ext4_free_extent ex;
2062 if (ac->ac_status == AC_STATUS_FOUND)
2065 * We don't want to scan for a whole year
2067 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2068 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2069 ac->ac_status = AC_STATUS_BREAK;
2074 * Haven't found good chunk so far, let's continue
2076 if (bex->fe_len < gex->fe_len)
2079 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2080 && bex->fe_group == e4b->bd_group) {
2081 /* recheck chunk's availability - we don't know
2082 * when it was found (within this lock-unlock
2084 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
2085 if (max >= gex->fe_len) {
2086 ext4_mb_use_best_found(ac, e4b);
2093 * The routine checks whether found extent is good enough. If it is,
2094 * then the extent gets marked used and flag is set to the context
2095 * to stop scanning. Otherwise, the extent is compared with the
2096 * previous found extent and if new one is better, then it's stored
2097 * in the context. Later, the best found extent will be used, if
2098 * mballoc can't find good enough extent.
2100 * FIXME: real allocation policy is to be designed yet!
2102 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2103 struct ext4_free_extent *ex,
2104 struct ext4_buddy *e4b)
2106 struct ext4_free_extent *bex = &ac->ac_b_ex;
2107 struct ext4_free_extent *gex = &ac->ac_g_ex;
2109 BUG_ON(ex->fe_len <= 0);
2110 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2111 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2112 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2117 * The special case - take what you catch first
2119 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2121 ext4_mb_use_best_found(ac, e4b);
2126 * Let's check whether the chuck is good enough
2128 if (ex->fe_len == gex->fe_len) {
2130 ext4_mb_use_best_found(ac, e4b);
2135 * If this is first found extent, just store it in the context
2137 if (bex->fe_len == 0) {
2143 * If new found extent is better, store it in the context
2145 if (bex->fe_len < gex->fe_len) {
2146 /* if the request isn't satisfied, any found extent
2147 * larger than previous best one is better */
2148 if (ex->fe_len > bex->fe_len)
2150 } else if (ex->fe_len > gex->fe_len) {
2151 /* if the request is satisfied, then we try to find
2152 * an extent that still satisfy the request, but is
2153 * smaller than previous one */
2154 if (ex->fe_len < bex->fe_len)
2158 ext4_mb_check_limits(ac, e4b, 0);
2161 static noinline_for_stack
2162 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2163 struct ext4_buddy *e4b)
2165 struct ext4_free_extent ex = ac->ac_b_ex;
2166 ext4_group_t group = ex.fe_group;
2170 BUG_ON(ex.fe_len <= 0);
2171 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2175 ext4_lock_group(ac->ac_sb, group);
2176 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2180 ext4_mb_use_best_found(ac, e4b);
2183 ext4_unlock_group(ac->ac_sb, group);
2184 ext4_mb_unload_buddy(e4b);
2189 static noinline_for_stack
2190 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2191 struct ext4_buddy *e4b)
2193 ext4_group_t group = ac->ac_g_ex.fe_group;
2196 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2197 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2198 struct ext4_free_extent ex;
2200 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
2202 if (grp->bb_free == 0)
2205 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2209 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
2210 ext4_mb_unload_buddy(e4b);
2214 ext4_lock_group(ac->ac_sb, group);
2215 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2216 ac->ac_g_ex.fe_len, &ex);
2217 ex.fe_logical = 0xDEADFA11; /* debug value */
2219 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
2222 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
2224 /* use do_div to get remainder (would be 64-bit modulo) */
2225 if (do_div(start, sbi->s_stripe) == 0) {
2228 ext4_mb_use_best_found(ac, e4b);
2230 } else if (max >= ac->ac_g_ex.fe_len) {
2231 BUG_ON(ex.fe_len <= 0);
2232 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2233 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2236 ext4_mb_use_best_found(ac, e4b);
2237 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2238 /* Sometimes, caller may want to merge even small
2239 * number of blocks to an existing extent */
2240 BUG_ON(ex.fe_len <= 0);
2241 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2242 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2245 ext4_mb_use_best_found(ac, e4b);
2247 ext4_unlock_group(ac->ac_sb, group);
2248 ext4_mb_unload_buddy(e4b);
2254 * The routine scans buddy structures (not bitmap!) from given order
2255 * to max order and tries to find big enough chunk to satisfy the req
2257 static noinline_for_stack
2258 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2259 struct ext4_buddy *e4b)
2261 struct super_block *sb = ac->ac_sb;
2262 struct ext4_group_info *grp = e4b->bd_info;
2268 BUG_ON(ac->ac_2order <= 0);
2269 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2270 if (grp->bb_counters[i] == 0)
2273 buddy = mb_find_buddy(e4b, i, &max);
2274 BUG_ON(buddy == NULL);
2276 k = mb_find_next_zero_bit(buddy, max, 0);
2278 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2279 "%d free clusters of order %d. But found 0",
2280 grp->bb_counters[i], i);
2281 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2283 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2288 ac->ac_b_ex.fe_len = 1 << i;
2289 ac->ac_b_ex.fe_start = k << i;
2290 ac->ac_b_ex.fe_group = e4b->bd_group;
2292 ext4_mb_use_best_found(ac, e4b);
2294 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2296 if (EXT4_SB(sb)->s_mb_stats)
2297 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2304 * The routine scans the group and measures all found extents.
2305 * In order to optimize scanning, caller must pass number of
2306 * free blocks in the group, so the routine can know upper limit.
2308 static noinline_for_stack
2309 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2310 struct ext4_buddy *e4b)
2312 struct super_block *sb = ac->ac_sb;
2313 void *bitmap = e4b->bd_bitmap;
2314 struct ext4_free_extent ex;
2318 free = e4b->bd_info->bb_free;
2319 if (WARN_ON(free <= 0))
2322 i = e4b->bd_info->bb_first_free;
2324 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2325 i = mb_find_next_zero_bit(bitmap,
2326 EXT4_CLUSTERS_PER_GROUP(sb), i);
2327 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2329 * IF we have corrupt bitmap, we won't find any
2330 * free blocks even though group info says we
2333 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2334 "%d free clusters as per "
2335 "group info. But bitmap says 0",
2337 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2338 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2342 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2343 if (WARN_ON(ex.fe_len <= 0))
2345 if (free < ex.fe_len) {
2346 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2347 "%d free clusters as per "
2348 "group info. But got %d blocks",
2350 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2351 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2353 * The number of free blocks differs. This mostly
2354 * indicate that the bitmap is corrupt. So exit
2355 * without claiming the space.
2359 ex.fe_logical = 0xDEADC0DE; /* debug value */
2360 ext4_mb_measure_extent(ac, &ex, e4b);
2366 ext4_mb_check_limits(ac, e4b, 1);
2370 * This is a special case for storages like raid5
2371 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2373 static noinline_for_stack
2374 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2375 struct ext4_buddy *e4b)
2377 struct super_block *sb = ac->ac_sb;
2378 struct ext4_sb_info *sbi = EXT4_SB(sb);
2379 void *bitmap = e4b->bd_bitmap;
2380 struct ext4_free_extent ex;
2381 ext4_fsblk_t first_group_block;
2386 BUG_ON(sbi->s_stripe == 0);
2388 /* find first stripe-aligned block in group */
2389 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2391 a = first_group_block + sbi->s_stripe - 1;
2392 do_div(a, sbi->s_stripe);
2393 i = (a * sbi->s_stripe) - first_group_block;
2395 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2396 if (!mb_test_bit(i, bitmap)) {
2397 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2398 if (max >= sbi->s_stripe) {
2400 ex.fe_logical = 0xDEADF00D; /* debug value */
2402 ext4_mb_use_best_found(ac, e4b);
2411 * This is also called BEFORE we load the buddy bitmap.
2412 * Returns either 1 or 0 indicating that the group is either suitable
2413 * for the allocation or not.
2415 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2416 ext4_group_t group, int cr)
2418 ext4_grpblk_t free, fragments;
2419 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2420 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2422 BUG_ON(cr < 0 || cr >= 4);
2424 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2427 free = grp->bb_free;
2431 fragments = grp->bb_fragments;
2437 BUG_ON(ac->ac_2order == 0);
2439 /* Avoid using the first bg of a flexgroup for data files */
2440 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2441 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2442 ((group % flex_size) == 0))
2445 if (free < ac->ac_g_ex.fe_len)
2448 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2451 if (grp->bb_largest_free_order < ac->ac_2order)
2456 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2460 if (free >= ac->ac_g_ex.fe_len)
2473 * This could return negative error code if something goes wrong
2474 * during ext4_mb_init_group(). This should not be called with
2475 * ext4_lock_group() held.
2477 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2478 ext4_group_t group, int cr)
2480 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2481 struct super_block *sb = ac->ac_sb;
2482 struct ext4_sb_info *sbi = EXT4_SB(sb);
2483 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2487 if (sbi->s_mb_stats)
2488 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2490 ext4_lock_group(sb, group);
2491 free = grp->bb_free;
2494 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2496 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2499 ext4_unlock_group(sb, group);
2501 /* We only do this if the grp has never been initialized */
2502 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2503 struct ext4_group_desc *gdp =
2504 ext4_get_group_desc(sb, group, NULL);
2507 /* cr=0/1 is a very optimistic search to find large
2508 * good chunks almost for free. If buddy data is not
2509 * ready, then this optimization makes no sense. But
2510 * we never skip the first block group in a flex_bg,
2511 * since this gets used for metadata block allocation,
2512 * and we want to make sure we locate metadata blocks
2513 * in the first block group in the flex_bg if possible.
2516 (!sbi->s_log_groups_per_flex ||
2517 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2518 !(ext4_has_group_desc_csum(sb) &&
2519 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2521 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2527 ext4_lock_group(sb, group);
2528 ret = ext4_mb_good_group(ac, group, cr);
2531 ext4_unlock_group(sb, group);
2536 * Start prefetching @nr block bitmaps starting at @group.
2537 * Return the next group which needs to be prefetched.
2539 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2540 unsigned int nr, int *cnt)
2542 ext4_group_t ngroups = ext4_get_groups_count(sb);
2543 struct buffer_head *bh;
2544 struct blk_plug plug;
2546 blk_start_plug(&plug);
2548 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2550 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2553 * Prefetch block groups with free blocks; but don't
2554 * bother if it is marked uninitialized on disk, since
2555 * it won't require I/O to read. Also only try to
2556 * prefetch once, so we avoid getblk() call, which can
2559 if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2560 EXT4_MB_GRP_NEED_INIT(grp) &&
2561 ext4_free_group_clusters(sb, gdp) > 0 &&
2562 !(ext4_has_group_desc_csum(sb) &&
2563 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2564 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2565 if (bh && !IS_ERR(bh)) {
2566 if (!buffer_uptodate(bh) && cnt)
2571 if (++group >= ngroups)
2574 blk_finish_plug(&plug);
2579 * Prefetching reads the block bitmap into the buffer cache; but we
2580 * need to make sure that the buddy bitmap in the page cache has been
2581 * initialized. Note that ext4_mb_init_group() will block if the I/O
2582 * is not yet completed, or indeed if it was not initiated by
2583 * ext4_mb_prefetch did not start the I/O.
2585 * TODO: We should actually kick off the buddy bitmap setup in a work
2586 * queue when the buffer I/O is completed, so that we don't block
2587 * waiting for the block allocation bitmap read to finish when
2588 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2590 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2594 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2596 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2599 group = ext4_get_groups_count(sb);
2601 grp = ext4_get_group_info(sb, group);
2603 if (EXT4_MB_GRP_NEED_INIT(grp) &&
2604 ext4_free_group_clusters(sb, gdp) > 0 &&
2605 !(ext4_has_group_desc_csum(sb) &&
2606 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2607 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2613 static noinline_for_stack int
2614 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2616 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2618 int err = 0, first_err = 0;
2619 unsigned int nr = 0, prefetch_ios = 0;
2620 struct ext4_sb_info *sbi;
2621 struct super_block *sb;
2622 struct ext4_buddy e4b;
2627 ngroups = ext4_get_groups_count(sb);
2628 /* non-extent files are limited to low blocks/groups */
2629 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2630 ngroups = sbi->s_blockfile_groups;
2632 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2634 /* first, try the goal */
2635 err = ext4_mb_find_by_goal(ac, &e4b);
2636 if (err || ac->ac_status == AC_STATUS_FOUND)
2639 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2643 * ac->ac_2order is set only if the fe_len is a power of 2
2644 * if ac->ac_2order is set we also set criteria to 0 so that we
2645 * try exact allocation using buddy.
2647 i = fls(ac->ac_g_ex.fe_len);
2650 * We search using buddy data only if the order of the request
2651 * is greater than equal to the sbi_s_mb_order2_reqs
2652 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2653 * We also support searching for power-of-two requests only for
2654 * requests upto maximum buddy size we have constructed.
2656 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2658 * This should tell if fe_len is exactly power of 2
2660 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2661 ac->ac_2order = array_index_nospec(i - 1,
2665 /* if stream allocation is enabled, use global goal */
2666 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2667 /* TBD: may be hot point */
2668 spin_lock(&sbi->s_md_lock);
2669 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2670 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2671 spin_unlock(&sbi->s_md_lock);
2674 /* Let's just scan groups to find more-less suitable blocks */
2675 cr = ac->ac_2order ? 0 : 1;
2677 * cr == 0 try to get exact allocation,
2678 * cr == 3 try to get anything
2681 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2682 ac->ac_criteria = cr;
2684 * searching for the right group start
2685 * from the goal value specified
2687 group = ac->ac_g_ex.fe_group;
2688 ac->ac_last_optimal_group = group;
2689 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2690 prefetch_grp = group;
2692 for (i = 0; i < ngroups; group = next_linear_group(ac, group, ngroups),
2694 int ret = 0, new_cr;
2698 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups);
2705 * Batch reads of the block allocation bitmaps
2706 * to get multiple READs in flight; limit
2707 * prefetching at cr=0/1, otherwise mballoc can
2708 * spend a lot of time loading imperfect groups
2710 if ((prefetch_grp == group) &&
2712 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2713 unsigned int curr_ios = prefetch_ios;
2715 nr = sbi->s_mb_prefetch;
2716 if (ext4_has_feature_flex_bg(sb)) {
2717 nr = 1 << sbi->s_log_groups_per_flex;
2718 nr -= group & (nr - 1);
2719 nr = min(nr, sbi->s_mb_prefetch);
2721 prefetch_grp = ext4_mb_prefetch(sb, group,
2723 if (prefetch_ios == curr_ios)
2727 /* This now checks without needing the buddy page */
2728 ret = ext4_mb_good_group_nolock(ac, group, cr);
2735 err = ext4_mb_load_buddy(sb, group, &e4b);
2739 ext4_lock_group(sb, group);
2742 * We need to check again after locking the
2745 ret = ext4_mb_good_group(ac, group, cr);
2747 ext4_unlock_group(sb, group);
2748 ext4_mb_unload_buddy(&e4b);
2752 ac->ac_groups_scanned++;
2754 ext4_mb_simple_scan_group(ac, &e4b);
2755 else if (cr == 1 && sbi->s_stripe &&
2756 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2757 ext4_mb_scan_aligned(ac, &e4b);
2759 ext4_mb_complex_scan_group(ac, &e4b);
2761 ext4_unlock_group(sb, group);
2762 ext4_mb_unload_buddy(&e4b);
2764 if (ac->ac_status != AC_STATUS_CONTINUE)
2767 /* Processed all groups and haven't found blocks */
2768 if (sbi->s_mb_stats && i == ngroups)
2769 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2772 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2773 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2775 * We've been searching too long. Let's try to allocate
2776 * the best chunk we've found so far
2778 ext4_mb_try_best_found(ac, &e4b);
2779 if (ac->ac_status != AC_STATUS_FOUND) {
2781 * Someone more lucky has already allocated it.
2782 * The only thing we can do is just take first
2785 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2786 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2787 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2788 ac->ac_b_ex.fe_len, lost);
2790 ac->ac_b_ex.fe_group = 0;
2791 ac->ac_b_ex.fe_start = 0;
2792 ac->ac_b_ex.fe_len = 0;
2793 ac->ac_status = AC_STATUS_CONTINUE;
2794 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2800 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2801 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2803 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2806 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2807 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2808 ac->ac_flags, cr, err);
2811 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2816 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2818 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2821 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2824 return (void *) ((unsigned long) group);
2827 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2829 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2833 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2836 return (void *) ((unsigned long) group);
2839 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2841 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2842 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2844 int err, buddy_loaded = 0;
2845 struct ext4_buddy e4b;
2846 struct ext4_group_info *grinfo;
2847 unsigned char blocksize_bits = min_t(unsigned char,
2848 sb->s_blocksize_bits,
2849 EXT4_MAX_BLOCK_LOG_SIZE);
2851 struct ext4_group_info info;
2852 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2857 seq_puts(seq, "#group: free frags first ["
2858 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2859 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2861 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2862 sizeof(struct ext4_group_info);
2864 grinfo = ext4_get_group_info(sb, group);
2865 /* Load the group info in memory only if not already loaded. */
2866 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2867 err = ext4_mb_load_buddy(sb, group, &e4b);
2869 seq_printf(seq, "#%-5u: I/O error\n", group);
2875 memcpy(&sg, ext4_get_group_info(sb, group), i);
2878 ext4_mb_unload_buddy(&e4b);
2880 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2881 sg.info.bb_fragments, sg.info.bb_first_free);
2882 for (i = 0; i <= 13; i++)
2883 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2884 sg.info.bb_counters[i] : 0);
2885 seq_puts(seq, " ]\n");
2890 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2894 const struct seq_operations ext4_mb_seq_groups_ops = {
2895 .start = ext4_mb_seq_groups_start,
2896 .next = ext4_mb_seq_groups_next,
2897 .stop = ext4_mb_seq_groups_stop,
2898 .show = ext4_mb_seq_groups_show,
2901 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
2903 struct super_block *sb = (struct super_block *)seq->private;
2904 struct ext4_sb_info *sbi = EXT4_SB(sb);
2906 seq_puts(seq, "mballoc:\n");
2907 if (!sbi->s_mb_stats) {
2908 seq_puts(seq, "\tmb stats collection turned off.\n");
2909 seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
2912 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
2913 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
2915 seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned));
2917 seq_puts(seq, "\tcr0_stats:\n");
2918 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0]));
2919 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2920 atomic64_read(&sbi->s_bal_cX_groups_considered[0]));
2921 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2922 atomic64_read(&sbi->s_bal_cX_failed[0]));
2923 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2924 atomic_read(&sbi->s_bal_cr0_bad_suggestions));
2926 seq_puts(seq, "\tcr1_stats:\n");
2927 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1]));
2928 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2929 atomic64_read(&sbi->s_bal_cX_groups_considered[1]));
2930 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2931 atomic64_read(&sbi->s_bal_cX_failed[1]));
2932 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2933 atomic_read(&sbi->s_bal_cr1_bad_suggestions));
2935 seq_puts(seq, "\tcr2_stats:\n");
2936 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2]));
2937 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2938 atomic64_read(&sbi->s_bal_cX_groups_considered[2]));
2939 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2940 atomic64_read(&sbi->s_bal_cX_failed[2]));
2942 seq_puts(seq, "\tcr3_stats:\n");
2943 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3]));
2944 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2945 atomic64_read(&sbi->s_bal_cX_groups_considered[3]));
2946 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2947 atomic64_read(&sbi->s_bal_cX_failed[3]));
2948 seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned));
2949 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
2950 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
2951 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
2952 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
2954 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
2955 atomic_read(&sbi->s_mb_buddies_generated),
2956 ext4_get_groups_count(sb));
2957 seq_printf(seq, "\tbuddies_time_used: %llu\n",
2958 atomic64_read(&sbi->s_mb_generation_time));
2959 seq_printf(seq, "\tpreallocated: %u\n",
2960 atomic_read(&sbi->s_mb_preallocated));
2961 seq_printf(seq, "\tdiscarded: %u\n",
2962 atomic_read(&sbi->s_mb_discarded));
2966 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
2968 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2969 unsigned long position;
2971 read_lock(&EXT4_SB(sb)->s_mb_rb_lock);
2973 if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
2975 position = *pos + 1;
2976 return (void *) ((unsigned long) position);
2979 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
2981 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2982 unsigned long position;
2985 if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
2987 position = *pos + 1;
2988 return (void *) ((unsigned long) position);
2991 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
2993 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2994 struct ext4_sb_info *sbi = EXT4_SB(sb);
2995 unsigned long position = ((unsigned long) v);
2996 struct ext4_group_info *grp;
2998 unsigned int count, min, max;
3001 if (position >= MB_NUM_ORDERS(sb)) {
3002 seq_puts(seq, "fragment_size_tree:\n");
3003 n = rb_first(&sbi->s_mb_avg_fragment_size_root);
3005 seq_puts(seq, "\ttree_min: 0\n\ttree_max: 0\n\ttree_nodes: 0\n");
3008 grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3009 min = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3011 while (rb_next(n)) {
3015 grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3016 max = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3018 seq_printf(seq, "\ttree_min: %u\n\ttree_max: %u\n\ttree_nodes: %u\n",
3023 if (position == 0) {
3024 seq_printf(seq, "optimize_scan: %d\n",
3025 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
3026 seq_puts(seq, "max_free_order_lists:\n");
3029 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
3030 bb_largest_free_order_node)
3032 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3033 (unsigned int)position, count);
3038 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3040 struct super_block *sb = PDE_DATA(file_inode(seq->file));
3042 read_unlock(&EXT4_SB(sb)->s_mb_rb_lock);
3045 const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3046 .start = ext4_mb_seq_structs_summary_start,
3047 .next = ext4_mb_seq_structs_summary_next,
3048 .stop = ext4_mb_seq_structs_summary_stop,
3049 .show = ext4_mb_seq_structs_summary_show,
3052 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3054 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3055 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3062 * Allocate the top-level s_group_info array for the specified number
3065 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3067 struct ext4_sb_info *sbi = EXT4_SB(sb);
3069 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3071 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3072 EXT4_DESC_PER_BLOCK_BITS(sb);
3073 if (size <= sbi->s_group_info_size)
3076 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3077 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3078 if (!new_groupinfo) {
3079 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3083 old_groupinfo = rcu_dereference(sbi->s_group_info);
3085 memcpy(new_groupinfo, old_groupinfo,
3086 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3088 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3089 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3091 ext4_kvfree_array_rcu(old_groupinfo);
3092 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3093 sbi->s_group_info_size);
3097 /* Create and initialize ext4_group_info data for the given group. */
3098 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3099 struct ext4_group_desc *desc)
3103 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3104 struct ext4_sb_info *sbi = EXT4_SB(sb);
3105 struct ext4_group_info **meta_group_info;
3106 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3109 * First check if this group is the first of a reserved block.
3110 * If it's true, we have to allocate a new table of pointers
3111 * to ext4_group_info structures
3113 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3114 metalen = sizeof(*meta_group_info) <<
3115 EXT4_DESC_PER_BLOCK_BITS(sb);
3116 meta_group_info = kmalloc(metalen, GFP_NOFS);
3117 if (meta_group_info == NULL) {
3118 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3119 "for a buddy group");
3120 goto exit_meta_group_info;
3123 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3127 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3128 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3130 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3131 if (meta_group_info[i] == NULL) {
3132 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3133 goto exit_group_info;
3135 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3136 &(meta_group_info[i]->bb_state));
3139 * initialize bb_free to be able to skip
3140 * empty groups without initialization
3142 if (ext4_has_group_desc_csum(sb) &&
3143 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3144 meta_group_info[i]->bb_free =
3145 ext4_free_clusters_after_init(sb, group, desc);
3147 meta_group_info[i]->bb_free =
3148 ext4_free_group_clusters(sb, desc);
3151 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3152 init_rwsem(&meta_group_info[i]->alloc_sem);
3153 meta_group_info[i]->bb_free_root = RB_ROOT;
3154 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3155 RB_CLEAR_NODE(&meta_group_info[i]->bb_avg_fragment_size_rb);
3156 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3157 meta_group_info[i]->bb_group = group;
3159 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3163 /* If a meta_group_info table has been allocated, release it now */
3164 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3165 struct ext4_group_info ***group_info;
3168 group_info = rcu_dereference(sbi->s_group_info);
3169 kfree(group_info[idx]);
3170 group_info[idx] = NULL;
3173 exit_meta_group_info:
3175 } /* ext4_mb_add_groupinfo */
3177 static int ext4_mb_init_backend(struct super_block *sb)
3179 ext4_group_t ngroups = ext4_get_groups_count(sb);
3181 struct ext4_sb_info *sbi = EXT4_SB(sb);
3183 struct ext4_group_desc *desc;
3184 struct ext4_group_info ***group_info;
3185 struct kmem_cache *cachep;
3187 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3191 sbi->s_buddy_cache = new_inode(sb);
3192 if (sbi->s_buddy_cache == NULL) {
3193 ext4_msg(sb, KERN_ERR, "can't get new inode");
3196 /* To avoid potentially colliding with an valid on-disk inode number,
3197 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3198 * not in the inode hash, so it should never be found by iget(), but
3199 * this will avoid confusion if it ever shows up during debugging. */
3200 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3201 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3202 for (i = 0; i < ngroups; i++) {
3204 desc = ext4_get_group_desc(sb, i, NULL);
3206 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3209 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3213 if (ext4_has_feature_flex_bg(sb)) {
3214 /* a single flex group is supposed to be read by a single IO.
3215 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3216 * unsigned integer, so the maximum shift is 32.
3218 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3219 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3222 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3223 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3224 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3226 sbi->s_mb_prefetch = 32;
3228 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3229 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3230 /* now many real IOs to prefetch within a single allocation at cr=0
3231 * given cr=0 is an CPU-related optimization we shouldn't try to
3232 * load too many groups, at some point we should start to use what
3233 * we've got in memory.
3234 * with an average random access time 5ms, it'd take a second to get
3235 * 200 groups (* N with flex_bg), so let's make this limit 4
3237 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3238 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3239 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3244 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3246 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
3247 i = sbi->s_group_info_size;
3249 group_info = rcu_dereference(sbi->s_group_info);
3251 kfree(group_info[i]);
3253 iput(sbi->s_buddy_cache);
3256 kvfree(rcu_dereference(sbi->s_group_info));
3261 static void ext4_groupinfo_destroy_slabs(void)
3265 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3266 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3267 ext4_groupinfo_caches[i] = NULL;
3271 static int ext4_groupinfo_create_slab(size_t size)
3273 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3275 int blocksize_bits = order_base_2(size);
3276 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3277 struct kmem_cache *cachep;
3279 if (cache_index >= NR_GRPINFO_CACHES)
3282 if (unlikely(cache_index < 0))
3285 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3286 if (ext4_groupinfo_caches[cache_index]) {
3287 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3288 return 0; /* Already created */
3291 slab_size = offsetof(struct ext4_group_info,
3292 bb_counters[blocksize_bits + 2]);
3294 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3295 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3298 ext4_groupinfo_caches[cache_index] = cachep;
3300 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3303 "EXT4-fs: no memory for groupinfo slab cache\n");
3310 int ext4_mb_init(struct super_block *sb)
3312 struct ext4_sb_info *sbi = EXT4_SB(sb);
3314 unsigned offset, offset_incr;
3318 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3320 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3321 if (sbi->s_mb_offsets == NULL) {
3326 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3327 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3328 if (sbi->s_mb_maxs == NULL) {
3333 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3337 /* order 0 is regular bitmap */
3338 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3339 sbi->s_mb_offsets[0] = 0;
3343 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3344 max = sb->s_blocksize << 2;
3346 sbi->s_mb_offsets[i] = offset;
3347 sbi->s_mb_maxs[i] = max;
3348 offset += offset_incr;
3349 offset_incr = offset_incr >> 1;
3352 } while (i < MB_NUM_ORDERS(sb));
3354 sbi->s_mb_avg_fragment_size_root = RB_ROOT;
3355 sbi->s_mb_largest_free_orders =
3356 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3358 if (!sbi->s_mb_largest_free_orders) {
3362 sbi->s_mb_largest_free_orders_locks =
3363 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3365 if (!sbi->s_mb_largest_free_orders_locks) {
3369 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3370 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3371 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3373 rwlock_init(&sbi->s_mb_rb_lock);
3375 spin_lock_init(&sbi->s_md_lock);
3376 sbi->s_mb_free_pending = 0;
3377 INIT_LIST_HEAD(&sbi->s_freed_data_list);
3379 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3380 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3381 sbi->s_mb_stats = MB_DEFAULT_STATS;
3382 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3383 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3384 sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
3386 * The default group preallocation is 512, which for 4k block
3387 * sizes translates to 2 megabytes. However for bigalloc file
3388 * systems, this is probably too big (i.e, if the cluster size
3389 * is 1 megabyte, then group preallocation size becomes half a
3390 * gigabyte!). As a default, we will keep a two megabyte
3391 * group pralloc size for cluster sizes up to 64k, and after
3392 * that, we will force a minimum group preallocation size of
3393 * 32 clusters. This translates to 8 megs when the cluster
3394 * size is 256k, and 32 megs when the cluster size is 1 meg,
3395 * which seems reasonable as a default.
3397 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3398 sbi->s_cluster_bits, 32);
3400 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3401 * to the lowest multiple of s_stripe which is bigger than
3402 * the s_mb_group_prealloc as determined above. We want
3403 * the preallocation size to be an exact multiple of the
3404 * RAID stripe size so that preallocations don't fragment
3407 if (sbi->s_stripe > 1) {
3408 sbi->s_mb_group_prealloc = roundup(
3409 sbi->s_mb_group_prealloc, sbi->s_stripe);
3412 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3413 if (sbi->s_locality_groups == NULL) {
3417 for_each_possible_cpu(i) {
3418 struct ext4_locality_group *lg;
3419 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3420 mutex_init(&lg->lg_mutex);
3421 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3422 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3423 spin_lock_init(&lg->lg_prealloc_lock);
3426 if (blk_queue_nonrot(bdev_get_queue(sb->s_bdev)))
3427 sbi->s_mb_max_linear_groups = 0;
3429 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3430 /* init file for buddy data */
3431 ret = ext4_mb_init_backend(sb);
3433 goto out_free_locality_groups;
3437 out_free_locality_groups:
3438 free_percpu(sbi->s_locality_groups);
3439 sbi->s_locality_groups = NULL;
3441 kfree(sbi->s_mb_largest_free_orders);
3442 kfree(sbi->s_mb_largest_free_orders_locks);
3443 kfree(sbi->s_mb_offsets);
3444 sbi->s_mb_offsets = NULL;
3445 kfree(sbi->s_mb_maxs);
3446 sbi->s_mb_maxs = NULL;
3450 /* need to called with the ext4 group lock held */
3451 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3453 struct ext4_prealloc_space *pa;
3454 struct list_head *cur, *tmp;
3457 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3458 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3459 list_del(&pa->pa_group_list);
3461 kmem_cache_free(ext4_pspace_cachep, pa);
3466 int ext4_mb_release(struct super_block *sb)
3468 ext4_group_t ngroups = ext4_get_groups_count(sb);
3470 int num_meta_group_infos;
3471 struct ext4_group_info *grinfo, ***group_info;
3472 struct ext4_sb_info *sbi = EXT4_SB(sb);
3473 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3476 if (sbi->s_group_info) {
3477 for (i = 0; i < ngroups; i++) {
3479 grinfo = ext4_get_group_info(sb, i);
3480 mb_group_bb_bitmap_free(grinfo);
3481 ext4_lock_group(sb, i);
3482 count = ext4_mb_cleanup_pa(grinfo);
3484 mb_debug(sb, "mballoc: %d PAs left\n",
3486 ext4_unlock_group(sb, i);
3487 kmem_cache_free(cachep, grinfo);
3489 num_meta_group_infos = (ngroups +
3490 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3491 EXT4_DESC_PER_BLOCK_BITS(sb);
3493 group_info = rcu_dereference(sbi->s_group_info);
3494 for (i = 0; i < num_meta_group_infos; i++)
3495 kfree(group_info[i]);
3499 kfree(sbi->s_mb_largest_free_orders);
3500 kfree(sbi->s_mb_largest_free_orders_locks);
3501 kfree(sbi->s_mb_offsets);
3502 kfree(sbi->s_mb_maxs);
3503 iput(sbi->s_buddy_cache);
3504 if (sbi->s_mb_stats) {
3505 ext4_msg(sb, KERN_INFO,
3506 "mballoc: %u blocks %u reqs (%u success)",
3507 atomic_read(&sbi->s_bal_allocated),
3508 atomic_read(&sbi->s_bal_reqs),
3509 atomic_read(&sbi->s_bal_success));
3510 ext4_msg(sb, KERN_INFO,
3511 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3512 "%u 2^N hits, %u breaks, %u lost",
3513 atomic_read(&sbi->s_bal_ex_scanned),
3514 atomic_read(&sbi->s_bal_groups_scanned),
3515 atomic_read(&sbi->s_bal_goals),
3516 atomic_read(&sbi->s_bal_2orders),
3517 atomic_read(&sbi->s_bal_breaks),
3518 atomic_read(&sbi->s_mb_lost_chunks));
3519 ext4_msg(sb, KERN_INFO,
3520 "mballoc: %u generated and it took %llu",
3521 atomic_read(&sbi->s_mb_buddies_generated),
3522 atomic64_read(&sbi->s_mb_generation_time));
3523 ext4_msg(sb, KERN_INFO,
3524 "mballoc: %u preallocated, %u discarded",
3525 atomic_read(&sbi->s_mb_preallocated),
3526 atomic_read(&sbi->s_mb_discarded));
3529 free_percpu(sbi->s_locality_groups);
3534 static inline int ext4_issue_discard(struct super_block *sb,
3535 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3538 ext4_fsblk_t discard_block;
3540 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3541 ext4_group_first_block_no(sb, block_group));
3542 count = EXT4_C2B(EXT4_SB(sb), count);
3543 trace_ext4_discard_blocks(sb,
3544 (unsigned long long) discard_block, count);
3546 return __blkdev_issue_discard(sb->s_bdev,
3547 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3548 (sector_t)count << (sb->s_blocksize_bits - 9),
3551 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3554 static void ext4_free_data_in_buddy(struct super_block *sb,
3555 struct ext4_free_data *entry)
3557 struct ext4_buddy e4b;
3558 struct ext4_group_info *db;
3559 int err, count = 0, count2 = 0;
3561 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3562 entry->efd_count, entry->efd_group, entry);
3564 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3565 /* we expect to find existing buddy because it's pinned */
3568 spin_lock(&EXT4_SB(sb)->s_md_lock);
3569 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3570 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3573 /* there are blocks to put in buddy to make them really free */
3574 count += entry->efd_count;
3576 ext4_lock_group(sb, entry->efd_group);
3577 /* Take it out of per group rb tree */
3578 rb_erase(&entry->efd_node, &(db->bb_free_root));
3579 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3582 * Clear the trimmed flag for the group so that the next
3583 * ext4_trim_fs can trim it.
3584 * If the volume is mounted with -o discard, online discard
3585 * is supported and the free blocks will be trimmed online.
3587 if (!test_opt(sb, DISCARD))
3588 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3590 if (!db->bb_free_root.rb_node) {
3591 /* No more items in the per group rb tree
3592 * balance refcounts from ext4_mb_free_metadata()
3594 put_page(e4b.bd_buddy_page);
3595 put_page(e4b.bd_bitmap_page);
3597 ext4_unlock_group(sb, entry->efd_group);
3598 kmem_cache_free(ext4_free_data_cachep, entry);
3599 ext4_mb_unload_buddy(&e4b);
3601 mb_debug(sb, "freed %d blocks in %d structures\n", count,
3606 * This function is called by the jbd2 layer once the commit has finished,
3607 * so we know we can free the blocks that were released with that commit.
3609 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3611 struct ext4_sb_info *sbi = EXT4_SB(sb);
3612 struct ext4_free_data *entry, *tmp;
3613 struct bio *discard_bio = NULL;
3614 struct list_head freed_data_list;
3615 struct list_head *cut_pos = NULL;
3618 INIT_LIST_HEAD(&freed_data_list);
3620 spin_lock(&sbi->s_md_lock);
3621 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3622 if (entry->efd_tid != commit_tid)
3624 cut_pos = &entry->efd_list;
3627 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3629 spin_unlock(&sbi->s_md_lock);
3631 if (test_opt(sb, DISCARD)) {
3632 list_for_each_entry(entry, &freed_data_list, efd_list) {
3633 err = ext4_issue_discard(sb, entry->efd_group,
3634 entry->efd_start_cluster,
3637 if (err && err != -EOPNOTSUPP) {
3638 ext4_msg(sb, KERN_WARNING, "discard request in"
3639 " group:%d block:%d count:%d failed"
3640 " with %d", entry->efd_group,
3641 entry->efd_start_cluster,
3642 entry->efd_count, err);
3643 } else if (err == -EOPNOTSUPP)
3648 submit_bio_wait(discard_bio);
3649 bio_put(discard_bio);
3653 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3654 ext4_free_data_in_buddy(sb, entry);
3657 int __init ext4_init_mballoc(void)
3659 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3660 SLAB_RECLAIM_ACCOUNT);
3661 if (ext4_pspace_cachep == NULL)
3664 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3665 SLAB_RECLAIM_ACCOUNT);
3666 if (ext4_ac_cachep == NULL)
3669 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3670 SLAB_RECLAIM_ACCOUNT);
3671 if (ext4_free_data_cachep == NULL)
3677 kmem_cache_destroy(ext4_ac_cachep);
3679 kmem_cache_destroy(ext4_pspace_cachep);
3684 void ext4_exit_mballoc(void)
3687 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3688 * before destroying the slab cache.
3691 kmem_cache_destroy(ext4_pspace_cachep);
3692 kmem_cache_destroy(ext4_ac_cachep);
3693 kmem_cache_destroy(ext4_free_data_cachep);
3694 ext4_groupinfo_destroy_slabs();
3699 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3700 * Returns 0 if success or error code
3702 static noinline_for_stack int
3703 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3704 handle_t *handle, unsigned int reserv_clstrs)
3706 struct buffer_head *bitmap_bh = NULL;
3707 struct ext4_group_desc *gdp;
3708 struct buffer_head *gdp_bh;
3709 struct ext4_sb_info *sbi;
3710 struct super_block *sb;
3714 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3715 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3720 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3721 if (IS_ERR(bitmap_bh)) {
3722 err = PTR_ERR(bitmap_bh);
3727 BUFFER_TRACE(bitmap_bh, "getting write access");
3728 err = ext4_journal_get_write_access(handle, bitmap_bh);
3733 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3737 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3738 ext4_free_group_clusters(sb, gdp));
3740 BUFFER_TRACE(gdp_bh, "get_write_access");
3741 err = ext4_journal_get_write_access(handle, gdp_bh);
3745 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3747 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3748 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3749 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3750 "fs metadata", block, block+len);
3751 /* File system mounted not to panic on error
3752 * Fix the bitmap and return EFSCORRUPTED
3753 * We leak some of the blocks here.
3755 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3756 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3757 ac->ac_b_ex.fe_len);
3758 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3759 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3761 err = -EFSCORRUPTED;
3765 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3766 #ifdef AGGRESSIVE_CHECK
3769 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3770 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3771 bitmap_bh->b_data));
3775 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3776 ac->ac_b_ex.fe_len);
3777 if (ext4_has_group_desc_csum(sb) &&
3778 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3779 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3780 ext4_free_group_clusters_set(sb, gdp,
3781 ext4_free_clusters_after_init(sb,
3782 ac->ac_b_ex.fe_group, gdp));
3784 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3785 ext4_free_group_clusters_set(sb, gdp, len);
3786 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3787 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3789 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3790 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3792 * Now reduce the dirty block count also. Should not go negative
3794 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3795 /* release all the reserved blocks if non delalloc */
3796 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3799 if (sbi->s_log_groups_per_flex) {
3800 ext4_group_t flex_group = ext4_flex_group(sbi,
3801 ac->ac_b_ex.fe_group);
3802 atomic64_sub(ac->ac_b_ex.fe_len,
3803 &sbi_array_rcu_deref(sbi, s_flex_groups,
3804 flex_group)->free_clusters);
3807 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3810 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3818 * Idempotent helper for Ext4 fast commit replay path to set the state of
3819 * blocks in bitmaps and update counters.
3821 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
3824 struct buffer_head *bitmap_bh = NULL;
3825 struct ext4_group_desc *gdp;
3826 struct buffer_head *gdp_bh;
3827 struct ext4_sb_info *sbi = EXT4_SB(sb);
3829 ext4_grpblk_t blkoff;
3833 clen = EXT4_B2C(sbi, len);
3835 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
3836 bitmap_bh = ext4_read_block_bitmap(sb, group);
3837 if (IS_ERR(bitmap_bh)) {
3838 err = PTR_ERR(bitmap_bh);
3844 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
3848 ext4_lock_group(sb, group);
3850 for (i = 0; i < clen; i++)
3851 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) == !state)
3855 ext4_set_bits(bitmap_bh->b_data, blkoff, clen);
3857 mb_test_and_clear_bits(bitmap_bh->b_data, blkoff, clen);
3858 if (ext4_has_group_desc_csum(sb) &&
3859 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3860 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3861 ext4_free_group_clusters_set(sb, gdp,
3862 ext4_free_clusters_after_init(sb,
3866 clen = ext4_free_group_clusters(sb, gdp) - clen + already;
3868 clen = ext4_free_group_clusters(sb, gdp) + clen - already;
3870 ext4_free_group_clusters_set(sb, gdp, clen);
3871 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
3872 ext4_group_desc_csum_set(sb, group, gdp);
3874 ext4_unlock_group(sb, group);
3876 if (sbi->s_log_groups_per_flex) {
3877 ext4_group_t flex_group = ext4_flex_group(sbi, group);
3880 &sbi_array_rcu_deref(sbi, s_flex_groups,
3881 flex_group)->free_clusters);
3884 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
3887 sync_dirty_buffer(bitmap_bh);
3888 err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
3889 sync_dirty_buffer(gdp_bh);
3896 * here we normalize request for locality group
3897 * Group request are normalized to s_mb_group_prealloc, which goes to
3898 * s_strip if we set the same via mount option.
3899 * s_mb_group_prealloc can be configured via
3900 * /sys/fs/ext4/<partition>/mb_group_prealloc
3902 * XXX: should we try to preallocate more than the group has now?
3904 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3906 struct super_block *sb = ac->ac_sb;
3907 struct ext4_locality_group *lg = ac->ac_lg;
3910 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3911 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
3915 * Normalization means making request better in terms of
3916 * size and alignment
3918 static noinline_for_stack void
3919 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3920 struct ext4_allocation_request *ar)
3922 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3925 loff_t size, start_off;
3926 loff_t orig_size __maybe_unused;
3928 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3929 struct ext4_prealloc_space *pa;
3931 /* do normalize only data requests, metadata requests
3932 do not need preallocation */
3933 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3936 /* sometime caller may want exact blocks */
3937 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3940 /* caller may indicate that preallocation isn't
3941 * required (it's a tail, for example) */
3942 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3945 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3946 ext4_mb_normalize_group_request(ac);
3950 bsbits = ac->ac_sb->s_blocksize_bits;
3952 /* first, let's learn actual file size
3953 * given current request is allocated */
3954 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3955 size = size << bsbits;
3956 if (size < i_size_read(ac->ac_inode))
3957 size = i_size_read(ac->ac_inode);
3960 /* max size of free chunks */
3963 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3964 (req <= (size) || max <= (chunk_size))
3966 /* first, try to predict filesize */
3967 /* XXX: should this table be tunable? */
3969 if (size <= 16 * 1024) {
3971 } else if (size <= 32 * 1024) {
3973 } else if (size <= 64 * 1024) {
3975 } else if (size <= 128 * 1024) {
3977 } else if (size <= 256 * 1024) {
3979 } else if (size <= 512 * 1024) {
3981 } else if (size <= 1024 * 1024) {
3983 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3984 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3985 (21 - bsbits)) << 21;
3986 size = 2 * 1024 * 1024;
3987 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3988 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3989 (22 - bsbits)) << 22;
3990 size = 4 * 1024 * 1024;
3991 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3992 (8<<20)>>bsbits, max, 8 * 1024)) {
3993 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3994 (23 - bsbits)) << 23;
3995 size = 8 * 1024 * 1024;
3997 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3998 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3999 ac->ac_o_ex.fe_len) << bsbits;
4001 size = size >> bsbits;
4002 start = start_off >> bsbits;
4004 /* don't cover already allocated blocks in selected range */
4005 if (ar->pleft && start <= ar->lleft) {
4006 size -= ar->lleft + 1 - start;
4007 start = ar->lleft + 1;
4009 if (ar->pright && start + size - 1 >= ar->lright)
4010 size -= start + size - ar->lright;
4013 * Trim allocation request for filesystems with artificially small
4016 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4017 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4021 /* check we don't cross already preallocated blocks */
4023 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4028 spin_lock(&pa->pa_lock);
4029 if (pa->pa_deleted) {
4030 spin_unlock(&pa->pa_lock);
4034 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4037 /* PA must not overlap original request */
4038 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
4039 ac->ac_o_ex.fe_logical < pa->pa_lstart));
4041 /* skip PAs this normalized request doesn't overlap with */
4042 if (pa->pa_lstart >= end || pa_end <= start) {
4043 spin_unlock(&pa->pa_lock);
4046 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
4048 /* adjust start or end to be adjacent to this pa */
4049 if (pa_end <= ac->ac_o_ex.fe_logical) {
4050 BUG_ON(pa_end < start);
4052 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4053 BUG_ON(pa->pa_lstart > end);
4054 end = pa->pa_lstart;
4056 spin_unlock(&pa->pa_lock);
4061 /* XXX: extra loop to check we really don't overlap preallocations */
4063 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4066 spin_lock(&pa->pa_lock);
4067 if (pa->pa_deleted == 0) {
4068 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4070 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
4072 spin_unlock(&pa->pa_lock);
4076 if (start + size <= ac->ac_o_ex.fe_logical &&
4077 start > ac->ac_o_ex.fe_logical) {
4078 ext4_msg(ac->ac_sb, KERN_ERR,
4079 "start %lu, size %lu, fe_logical %lu",
4080 (unsigned long) start, (unsigned long) size,
4081 (unsigned long) ac->ac_o_ex.fe_logical);
4084 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4086 /* now prepare goal request */
4088 /* XXX: is it better to align blocks WRT to logical
4089 * placement or satisfy big request as is */
4090 ac->ac_g_ex.fe_logical = start;
4091 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4093 /* define goal start in order to merge */
4094 if (ar->pright && (ar->lright == (start + size))) {
4095 /* merge to the right */
4096 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4097 &ac->ac_f_ex.fe_group,
4098 &ac->ac_f_ex.fe_start);
4099 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4101 if (ar->pleft && (ar->lleft + 1 == start)) {
4102 /* merge to the left */
4103 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4104 &ac->ac_f_ex.fe_group,
4105 &ac->ac_f_ex.fe_start);
4106 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4109 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4113 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4115 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4117 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4118 atomic_inc(&sbi->s_bal_reqs);
4119 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4120 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4121 atomic_inc(&sbi->s_bal_success);
4122 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4123 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4124 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4125 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4126 atomic_inc(&sbi->s_bal_goals);
4127 if (ac->ac_found > sbi->s_mb_max_to_scan)
4128 atomic_inc(&sbi->s_bal_breaks);
4131 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4132 trace_ext4_mballoc_alloc(ac);
4134 trace_ext4_mballoc_prealloc(ac);
4138 * Called on failure; free up any blocks from the inode PA for this
4139 * context. We don't need this for MB_GROUP_PA because we only change
4140 * pa_free in ext4_mb_release_context(), but on failure, we've already
4141 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4143 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4145 struct ext4_prealloc_space *pa = ac->ac_pa;
4146 struct ext4_buddy e4b;
4150 if (ac->ac_f_ex.fe_len == 0)
4152 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4155 * This should never happen since we pin the
4156 * pages in the ext4_allocation_context so
4157 * ext4_mb_load_buddy() should never fail.
4159 WARN(1, "mb_load_buddy failed (%d)", err);
4162 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4163 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4164 ac->ac_f_ex.fe_len);
4165 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4166 ext4_mb_unload_buddy(&e4b);
4169 if (pa->pa_type == MB_INODE_PA)
4170 pa->pa_free += ac->ac_b_ex.fe_len;
4174 * use blocks preallocated to inode
4176 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4177 struct ext4_prealloc_space *pa)
4179 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4184 /* found preallocated blocks, use them */
4185 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4186 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4187 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4188 len = EXT4_NUM_B2C(sbi, end - start);
4189 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4190 &ac->ac_b_ex.fe_start);
4191 ac->ac_b_ex.fe_len = len;
4192 ac->ac_status = AC_STATUS_FOUND;
4195 BUG_ON(start < pa->pa_pstart);
4196 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4197 BUG_ON(pa->pa_free < len);
4200 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4204 * use blocks preallocated to locality group
4206 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4207 struct ext4_prealloc_space *pa)
4209 unsigned int len = ac->ac_o_ex.fe_len;
4211 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4212 &ac->ac_b_ex.fe_group,
4213 &ac->ac_b_ex.fe_start);
4214 ac->ac_b_ex.fe_len = len;
4215 ac->ac_status = AC_STATUS_FOUND;
4218 /* we don't correct pa_pstart or pa_plen here to avoid
4219 * possible race when the group is being loaded concurrently
4220 * instead we correct pa later, after blocks are marked
4221 * in on-disk bitmap -- see ext4_mb_release_context()
4222 * Other CPUs are prevented from allocating from this pa by lg_mutex
4224 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4225 pa->pa_lstart-len, len, pa);
4229 * Return the prealloc space that have minimal distance
4230 * from the goal block. @cpa is the prealloc
4231 * space that is having currently known minimal distance
4232 * from the goal block.
4234 static struct ext4_prealloc_space *
4235 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4236 struct ext4_prealloc_space *pa,
4237 struct ext4_prealloc_space *cpa)
4239 ext4_fsblk_t cur_distance, new_distance;
4242 atomic_inc(&pa->pa_count);
4245 cur_distance = abs(goal_block - cpa->pa_pstart);
4246 new_distance = abs(goal_block - pa->pa_pstart);
4248 if (cur_distance <= new_distance)
4251 /* drop the previous reference */
4252 atomic_dec(&cpa->pa_count);
4253 atomic_inc(&pa->pa_count);
4258 * search goal blocks in preallocated space
4260 static noinline_for_stack bool
4261 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4263 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4265 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4266 struct ext4_locality_group *lg;
4267 struct ext4_prealloc_space *pa, *cpa = NULL;
4268 ext4_fsblk_t goal_block;
4270 /* only data can be preallocated */
4271 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4274 /* first, try per-file preallocation */
4276 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4278 /* all fields in this condition don't change,
4279 * so we can skip locking for them */
4280 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
4281 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
4282 EXT4_C2B(sbi, pa->pa_len)))
4285 /* non-extent files can't have physical blocks past 2^32 */
4286 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4287 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
4288 EXT4_MAX_BLOCK_FILE_PHYS))
4291 /* found preallocated blocks, use them */
4292 spin_lock(&pa->pa_lock);
4293 if (pa->pa_deleted == 0 && pa->pa_free) {
4294 atomic_inc(&pa->pa_count);
4295 ext4_mb_use_inode_pa(ac, pa);
4296 spin_unlock(&pa->pa_lock);
4297 ac->ac_criteria = 10;
4301 spin_unlock(&pa->pa_lock);
4305 /* can we use group allocation? */
4306 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4309 /* inode may have no locality group for some reason */
4313 order = fls(ac->ac_o_ex.fe_len) - 1;
4314 if (order > PREALLOC_TB_SIZE - 1)
4315 /* The max size of hash table is PREALLOC_TB_SIZE */
4316 order = PREALLOC_TB_SIZE - 1;
4318 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4320 * search for the prealloc space that is having
4321 * minimal distance from the goal block.
4323 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4325 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
4327 spin_lock(&pa->pa_lock);
4328 if (pa->pa_deleted == 0 &&
4329 pa->pa_free >= ac->ac_o_ex.fe_len) {
4331 cpa = ext4_mb_check_group_pa(goal_block,
4334 spin_unlock(&pa->pa_lock);
4339 ext4_mb_use_group_pa(ac, cpa);
4340 ac->ac_criteria = 20;
4347 * the function goes through all block freed in the group
4348 * but not yet committed and marks them used in in-core bitmap.
4349 * buddy must be generated from this bitmap
4350 * Need to be called with the ext4 group lock held
4352 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
4356 struct ext4_group_info *grp;
4357 struct ext4_free_data *entry;
4359 grp = ext4_get_group_info(sb, group);
4360 n = rb_first(&(grp->bb_free_root));
4363 entry = rb_entry(n, struct ext4_free_data, efd_node);
4364 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
4371 * the function goes through all preallocation in this group and marks them
4372 * used in in-core bitmap. buddy must be generated from this bitmap
4373 * Need to be called with ext4 group lock held
4375 static noinline_for_stack
4376 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4379 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4380 struct ext4_prealloc_space *pa;
4381 struct list_head *cur;
4382 ext4_group_t groupnr;
4383 ext4_grpblk_t start;
4384 int preallocated = 0;
4387 /* all form of preallocation discards first load group,
4388 * so the only competing code is preallocation use.
4389 * we don't need any locking here
4390 * notice we do NOT ignore preallocations with pa_deleted
4391 * otherwise we could leave used blocks available for
4392 * allocation in buddy when concurrent ext4_mb_put_pa()
4393 * is dropping preallocation
4395 list_for_each(cur, &grp->bb_prealloc_list) {
4396 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
4397 spin_lock(&pa->pa_lock);
4398 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4401 spin_unlock(&pa->pa_lock);
4402 if (unlikely(len == 0))
4404 BUG_ON(groupnr != group);
4405 ext4_set_bits(bitmap, start, len);
4406 preallocated += len;
4408 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
4411 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
4412 struct ext4_prealloc_space *pa)
4414 struct ext4_inode_info *ei;
4416 if (pa->pa_deleted) {
4417 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
4418 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
4425 if (pa->pa_type == MB_INODE_PA) {
4426 ei = EXT4_I(pa->pa_inode);
4427 atomic_dec(&ei->i_prealloc_active);
4431 static void ext4_mb_pa_callback(struct rcu_head *head)
4433 struct ext4_prealloc_space *pa;
4434 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
4436 BUG_ON(atomic_read(&pa->pa_count));
4437 BUG_ON(pa->pa_deleted == 0);
4438 kmem_cache_free(ext4_pspace_cachep, pa);
4442 * drops a reference to preallocated space descriptor
4443 * if this was the last reference and the space is consumed
4445 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
4446 struct super_block *sb, struct ext4_prealloc_space *pa)
4449 ext4_fsblk_t grp_blk;
4451 /* in this short window concurrent discard can set pa_deleted */
4452 spin_lock(&pa->pa_lock);
4453 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
4454 spin_unlock(&pa->pa_lock);
4458 if (pa->pa_deleted == 1) {
4459 spin_unlock(&pa->pa_lock);
4463 ext4_mb_mark_pa_deleted(sb, pa);
4464 spin_unlock(&pa->pa_lock);
4466 grp_blk = pa->pa_pstart;
4468 * If doing group-based preallocation, pa_pstart may be in the
4469 * next group when pa is used up
4471 if (pa->pa_type == MB_GROUP_PA)
4474 grp = ext4_get_group_number(sb, grp_blk);
4479 * P1 (buddy init) P2 (regular allocation)
4480 * find block B in PA
4481 * copy on-disk bitmap to buddy
4482 * mark B in on-disk bitmap
4483 * drop PA from group
4484 * mark all PAs in buddy
4486 * thus, P1 initializes buddy with B available. to prevent this
4487 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
4490 ext4_lock_group(sb, grp);
4491 list_del(&pa->pa_group_list);
4492 ext4_unlock_group(sb, grp);
4494 spin_lock(pa->pa_obj_lock);
4495 list_del_rcu(&pa->pa_inode_list);
4496 spin_unlock(pa->pa_obj_lock);
4498 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4502 * creates new preallocated space for given inode
4504 static noinline_for_stack void
4505 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
4507 struct super_block *sb = ac->ac_sb;
4508 struct ext4_sb_info *sbi = EXT4_SB(sb);
4509 struct ext4_prealloc_space *pa;
4510 struct ext4_group_info *grp;
4511 struct ext4_inode_info *ei;
4513 /* preallocate only when found space is larger then requested */
4514 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4515 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4516 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4517 BUG_ON(ac->ac_pa == NULL);
4521 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
4527 /* we can't allocate as much as normalizer wants.
4528 * so, found space must get proper lstart
4529 * to cover original request */
4530 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
4531 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
4533 /* we're limited by original request in that
4534 * logical block must be covered any way
4535 * winl is window we can move our chunk within */
4536 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
4538 /* also, we should cover whole original request */
4539 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
4541 /* the smallest one defines real window */
4542 win = min(winl, wins);
4544 offs = ac->ac_o_ex.fe_logical %
4545 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4546 if (offs && offs < win)
4549 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
4550 EXT4_NUM_B2C(sbi, win);
4551 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
4552 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
4555 /* preallocation can change ac_b_ex, thus we store actually
4556 * allocated blocks for history */
4557 ac->ac_f_ex = ac->ac_b_ex;
4559 pa->pa_lstart = ac->ac_b_ex.fe_logical;
4560 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4561 pa->pa_len = ac->ac_b_ex.fe_len;
4562 pa->pa_free = pa->pa_len;
4563 spin_lock_init(&pa->pa_lock);
4564 INIT_LIST_HEAD(&pa->pa_inode_list);
4565 INIT_LIST_HEAD(&pa->pa_group_list);
4567 pa->pa_type = MB_INODE_PA;
4569 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4570 pa->pa_len, pa->pa_lstart);
4571 trace_ext4_mb_new_inode_pa(ac, pa);
4573 ext4_mb_use_inode_pa(ac, pa);
4574 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
4576 ei = EXT4_I(ac->ac_inode);
4577 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4579 pa->pa_obj_lock = &ei->i_prealloc_lock;
4580 pa->pa_inode = ac->ac_inode;
4582 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4584 spin_lock(pa->pa_obj_lock);
4585 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
4586 spin_unlock(pa->pa_obj_lock);
4587 atomic_inc(&ei->i_prealloc_active);
4591 * creates new preallocated space for locality group inodes belongs to
4593 static noinline_for_stack void
4594 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
4596 struct super_block *sb = ac->ac_sb;
4597 struct ext4_locality_group *lg;
4598 struct ext4_prealloc_space *pa;
4599 struct ext4_group_info *grp;
4601 /* preallocate only when found space is larger then requested */
4602 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4603 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4604 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4605 BUG_ON(ac->ac_pa == NULL);
4609 /* preallocation can change ac_b_ex, thus we store actually
4610 * allocated blocks for history */
4611 ac->ac_f_ex = ac->ac_b_ex;
4613 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4614 pa->pa_lstart = pa->pa_pstart;
4615 pa->pa_len = ac->ac_b_ex.fe_len;
4616 pa->pa_free = pa->pa_len;
4617 spin_lock_init(&pa->pa_lock);
4618 INIT_LIST_HEAD(&pa->pa_inode_list);
4619 INIT_LIST_HEAD(&pa->pa_group_list);
4621 pa->pa_type = MB_GROUP_PA;
4623 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4624 pa->pa_len, pa->pa_lstart);
4625 trace_ext4_mb_new_group_pa(ac, pa);
4627 ext4_mb_use_group_pa(ac, pa);
4628 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4630 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4634 pa->pa_obj_lock = &lg->lg_prealloc_lock;
4635 pa->pa_inode = NULL;
4637 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4640 * We will later add the new pa to the right bucket
4641 * after updating the pa_free in ext4_mb_release_context
4645 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4647 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4648 ext4_mb_new_group_pa(ac);
4650 ext4_mb_new_inode_pa(ac);
4654 * finds all unused blocks in on-disk bitmap, frees them in
4655 * in-core bitmap and buddy.
4656 * @pa must be unlinked from inode and group lists, so that
4657 * nobody else can find/use it.
4658 * the caller MUST hold group/inode locks.
4659 * TODO: optimize the case when there are no in-core structures yet
4661 static noinline_for_stack int
4662 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4663 struct ext4_prealloc_space *pa)
4665 struct super_block *sb = e4b->bd_sb;
4666 struct ext4_sb_info *sbi = EXT4_SB(sb);
4671 unsigned long long grp_blk_start;
4674 BUG_ON(pa->pa_deleted == 0);
4675 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4676 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4677 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4678 end = bit + pa->pa_len;
4681 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4684 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4685 mb_debug(sb, "free preallocated %u/%u in group %u\n",
4686 (unsigned) ext4_group_first_block_no(sb, group) + bit,
4687 (unsigned) next - bit, (unsigned) group);
4690 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4691 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4692 EXT4_C2B(sbi, bit)),
4694 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4697 if (free != pa->pa_free) {
4698 ext4_msg(e4b->bd_sb, KERN_CRIT,
4699 "pa %p: logic %lu, phys. %lu, len %d",
4700 pa, (unsigned long) pa->pa_lstart,
4701 (unsigned long) pa->pa_pstart,
4703 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4706 * pa is already deleted so we use the value obtained
4707 * from the bitmap and continue.
4710 atomic_add(free, &sbi->s_mb_discarded);
4715 static noinline_for_stack int
4716 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4717 struct ext4_prealloc_space *pa)
4719 struct super_block *sb = e4b->bd_sb;
4723 trace_ext4_mb_release_group_pa(sb, pa);
4724 BUG_ON(pa->pa_deleted == 0);
4725 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4726 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4727 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4728 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4729 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4735 * releases all preallocations in given group
4737 * first, we need to decide discard policy:
4738 * - when do we discard
4740 * - how many do we discard
4741 * 1) how many requested
4743 static noinline_for_stack int
4744 ext4_mb_discard_group_preallocations(struct super_block *sb,
4745 ext4_group_t group, int needed)
4747 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4748 struct buffer_head *bitmap_bh = NULL;
4749 struct ext4_prealloc_space *pa, *tmp;
4750 struct list_head list;
4751 struct ext4_buddy e4b;
4754 int free, free_total = 0;
4756 mb_debug(sb, "discard preallocation for group %u\n", group);
4757 if (list_empty(&grp->bb_prealloc_list))
4760 bitmap_bh = ext4_read_block_bitmap(sb, group);
4761 if (IS_ERR(bitmap_bh)) {
4762 err = PTR_ERR(bitmap_bh);
4763 ext4_error_err(sb, -err,
4764 "Error %d reading block bitmap for %u",
4769 err = ext4_mb_load_buddy(sb, group, &e4b);
4771 ext4_warning(sb, "Error %d loading buddy information for %u",
4778 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
4780 INIT_LIST_HEAD(&list);
4783 ext4_lock_group(sb, group);
4784 list_for_each_entry_safe(pa, tmp,
4785 &grp->bb_prealloc_list, pa_group_list) {
4786 spin_lock(&pa->pa_lock);
4787 if (atomic_read(&pa->pa_count)) {
4788 spin_unlock(&pa->pa_lock);
4792 if (pa->pa_deleted) {
4793 spin_unlock(&pa->pa_lock);
4797 /* seems this one can be freed ... */
4798 ext4_mb_mark_pa_deleted(sb, pa);
4801 this_cpu_inc(discard_pa_seq);
4803 /* we can trust pa_free ... */
4804 free += pa->pa_free;
4806 spin_unlock(&pa->pa_lock);
4808 list_del(&pa->pa_group_list);
4809 list_add(&pa->u.pa_tmp_list, &list);
4812 /* now free all selected PAs */
4813 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4815 /* remove from object (inode or locality group) */
4816 spin_lock(pa->pa_obj_lock);
4817 list_del_rcu(&pa->pa_inode_list);
4818 spin_unlock(pa->pa_obj_lock);
4820 if (pa->pa_type == MB_GROUP_PA)
4821 ext4_mb_release_group_pa(&e4b, pa);
4823 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4825 list_del(&pa->u.pa_tmp_list);
4826 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4831 /* if we still need more blocks and some PAs were used, try again */
4832 if (free_total < needed && busy) {
4833 ext4_unlock_group(sb, group);
4838 ext4_unlock_group(sb, group);
4839 ext4_mb_unload_buddy(&e4b);
4842 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4843 free_total, group, grp->bb_free);
4848 * releases all non-used preallocated blocks for given inode
4850 * It's important to discard preallocations under i_data_sem
4851 * We don't want another block to be served from the prealloc
4852 * space when we are discarding the inode prealloc space.
4854 * FIXME!! Make sure it is valid at all the call sites
4856 void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
4858 struct ext4_inode_info *ei = EXT4_I(inode);
4859 struct super_block *sb = inode->i_sb;
4860 struct buffer_head *bitmap_bh = NULL;
4861 struct ext4_prealloc_space *pa, *tmp;
4862 ext4_group_t group = 0;
4863 struct list_head list;
4864 struct ext4_buddy e4b;
4867 if (!S_ISREG(inode->i_mode)) {
4868 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4872 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
4875 mb_debug(sb, "discard preallocation for inode %lu\n",
4877 trace_ext4_discard_preallocations(inode,
4878 atomic_read(&ei->i_prealloc_active), needed);
4880 INIT_LIST_HEAD(&list);
4886 /* first, collect all pa's in the inode */
4887 spin_lock(&ei->i_prealloc_lock);
4888 while (!list_empty(&ei->i_prealloc_list) && needed) {
4889 pa = list_entry(ei->i_prealloc_list.prev,
4890 struct ext4_prealloc_space, pa_inode_list);
4891 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4892 spin_lock(&pa->pa_lock);
4893 if (atomic_read(&pa->pa_count)) {
4894 /* this shouldn't happen often - nobody should
4895 * use preallocation while we're discarding it */
4896 spin_unlock(&pa->pa_lock);
4897 spin_unlock(&ei->i_prealloc_lock);
4898 ext4_msg(sb, KERN_ERR,
4899 "uh-oh! used pa while discarding");
4901 schedule_timeout_uninterruptible(HZ);
4905 if (pa->pa_deleted == 0) {
4906 ext4_mb_mark_pa_deleted(sb, pa);
4907 spin_unlock(&pa->pa_lock);
4908 list_del_rcu(&pa->pa_inode_list);
4909 list_add(&pa->u.pa_tmp_list, &list);
4914 /* someone is deleting pa right now */
4915 spin_unlock(&pa->pa_lock);
4916 spin_unlock(&ei->i_prealloc_lock);
4918 /* we have to wait here because pa_deleted
4919 * doesn't mean pa is already unlinked from
4920 * the list. as we might be called from
4921 * ->clear_inode() the inode will get freed
4922 * and concurrent thread which is unlinking
4923 * pa from inode's list may access already
4924 * freed memory, bad-bad-bad */
4926 /* XXX: if this happens too often, we can
4927 * add a flag to force wait only in case
4928 * of ->clear_inode(), but not in case of
4929 * regular truncate */
4930 schedule_timeout_uninterruptible(HZ);
4933 spin_unlock(&ei->i_prealloc_lock);
4935 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4936 BUG_ON(pa->pa_type != MB_INODE_PA);
4937 group = ext4_get_group_number(sb, pa->pa_pstart);
4939 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4940 GFP_NOFS|__GFP_NOFAIL);
4942 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
4947 bitmap_bh = ext4_read_block_bitmap(sb, group);
4948 if (IS_ERR(bitmap_bh)) {
4949 err = PTR_ERR(bitmap_bh);
4950 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
4952 ext4_mb_unload_buddy(&e4b);
4956 ext4_lock_group(sb, group);
4957 list_del(&pa->pa_group_list);
4958 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4959 ext4_unlock_group(sb, group);
4961 ext4_mb_unload_buddy(&e4b);
4964 list_del(&pa->u.pa_tmp_list);
4965 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4969 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
4971 struct ext4_prealloc_space *pa;
4973 BUG_ON(ext4_pspace_cachep == NULL);
4974 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
4977 atomic_set(&pa->pa_count, 1);
4982 static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
4984 struct ext4_prealloc_space *pa = ac->ac_pa;
4988 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
4989 kmem_cache_free(ext4_pspace_cachep, pa);
4992 #ifdef CONFIG_EXT4_DEBUG
4993 static inline void ext4_mb_show_pa(struct super_block *sb)
4995 ext4_group_t i, ngroups;
4997 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5000 ngroups = ext4_get_groups_count(sb);
5001 mb_debug(sb, "groups: ");
5002 for (i = 0; i < ngroups; i++) {
5003 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5004 struct ext4_prealloc_space *pa;
5005 ext4_grpblk_t start;
5006 struct list_head *cur;
5007 ext4_lock_group(sb, i);
5008 list_for_each(cur, &grp->bb_prealloc_list) {
5009 pa = list_entry(cur, struct ext4_prealloc_space,
5011 spin_lock(&pa->pa_lock);
5012 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5014 spin_unlock(&pa->pa_lock);
5015 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5018 ext4_unlock_group(sb, i);
5019 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5024 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5026 struct super_block *sb = ac->ac_sb;
5028 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5031 mb_debug(sb, "Can't allocate:"
5032 " Allocation context details:");
5033 mb_debug(sb, "status %u flags 0x%x",
5034 ac->ac_status, ac->ac_flags);
5035 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5036 "goal %lu/%lu/%lu@%lu, "
5037 "best %lu/%lu/%lu@%lu cr %d",
5038 (unsigned long)ac->ac_o_ex.fe_group,
5039 (unsigned long)ac->ac_o_ex.fe_start,
5040 (unsigned long)ac->ac_o_ex.fe_len,
5041 (unsigned long)ac->ac_o_ex.fe_logical,
5042 (unsigned long)ac->ac_g_ex.fe_group,
5043 (unsigned long)ac->ac_g_ex.fe_start,
5044 (unsigned long)ac->ac_g_ex.fe_len,
5045 (unsigned long)ac->ac_g_ex.fe_logical,
5046 (unsigned long)ac->ac_b_ex.fe_group,
5047 (unsigned long)ac->ac_b_ex.fe_start,
5048 (unsigned long)ac->ac_b_ex.fe_len,
5049 (unsigned long)ac->ac_b_ex.fe_logical,
5050 (int)ac->ac_criteria);
5051 mb_debug(sb, "%u found", ac->ac_found);
5052 ext4_mb_show_pa(sb);
5055 static inline void ext4_mb_show_pa(struct super_block *sb)
5059 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5061 ext4_mb_show_pa(ac->ac_sb);
5067 * We use locality group preallocation for small size file. The size of the
5068 * file is determined by the current size or the resulting size after
5069 * allocation which ever is larger
5071 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5073 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5075 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5076 int bsbits = ac->ac_sb->s_blocksize_bits;
5079 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5082 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5085 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
5086 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5089 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5090 !inode_is_open_for_write(ac->ac_inode)) {
5091 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5095 if (sbi->s_mb_group_prealloc <= 0) {
5096 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5100 /* don't use group allocation for large files */
5101 size = max(size, isize);
5102 if (size > sbi->s_mb_stream_request) {
5103 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5107 BUG_ON(ac->ac_lg != NULL);
5109 * locality group prealloc space are per cpu. The reason for having
5110 * per cpu locality group is to reduce the contention between block
5111 * request from multiple CPUs.
5113 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5115 /* we're going to use group allocation */
5116 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5118 /* serialize all allocations in the group */
5119 mutex_lock(&ac->ac_lg->lg_mutex);
5122 static noinline_for_stack int
5123 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5124 struct ext4_allocation_request *ar)
5126 struct super_block *sb = ar->inode->i_sb;
5127 struct ext4_sb_info *sbi = EXT4_SB(sb);
5128 struct ext4_super_block *es = sbi->s_es;
5132 ext4_grpblk_t block;
5134 /* we can't allocate > group size */
5137 /* just a dirty hack to filter too big requests */
5138 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5139 len = EXT4_CLUSTERS_PER_GROUP(sb);
5141 /* start searching from the goal */
5143 if (goal < le32_to_cpu(es->s_first_data_block) ||
5144 goal >= ext4_blocks_count(es))
5145 goal = le32_to_cpu(es->s_first_data_block);
5146 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5148 /* set up allocation goals */
5149 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5150 ac->ac_status = AC_STATUS_CONTINUE;
5152 ac->ac_inode = ar->inode;
5153 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5154 ac->ac_o_ex.fe_group = group;
5155 ac->ac_o_ex.fe_start = block;
5156 ac->ac_o_ex.fe_len = len;
5157 ac->ac_g_ex = ac->ac_o_ex;
5158 ac->ac_flags = ar->flags;
5160 /* we have to define context: we'll work with a file or
5161 * locality group. this is a policy, actually */
5162 ext4_mb_group_or_file(ac);
5164 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5165 "left: %u/%u, right %u/%u to %swritable\n",
5166 (unsigned) ar->len, (unsigned) ar->logical,
5167 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5168 (unsigned) ar->lleft, (unsigned) ar->pleft,
5169 (unsigned) ar->lright, (unsigned) ar->pright,
5170 inode_is_open_for_write(ar->inode) ? "" : "non-");
5175 static noinline_for_stack void
5176 ext4_mb_discard_lg_preallocations(struct super_block *sb,
5177 struct ext4_locality_group *lg,
5178 int order, int total_entries)
5180 ext4_group_t group = 0;
5181 struct ext4_buddy e4b;
5182 struct list_head discard_list;
5183 struct ext4_prealloc_space *pa, *tmp;
5185 mb_debug(sb, "discard locality group preallocation\n");
5187 INIT_LIST_HEAD(&discard_list);
5189 spin_lock(&lg->lg_prealloc_lock);
5190 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5192 lockdep_is_held(&lg->lg_prealloc_lock)) {
5193 spin_lock(&pa->pa_lock);
5194 if (atomic_read(&pa->pa_count)) {
5196 * This is the pa that we just used
5197 * for block allocation. So don't
5200 spin_unlock(&pa->pa_lock);
5203 if (pa->pa_deleted) {
5204 spin_unlock(&pa->pa_lock);
5207 /* only lg prealloc space */
5208 BUG_ON(pa->pa_type != MB_GROUP_PA);
5210 /* seems this one can be freed ... */
5211 ext4_mb_mark_pa_deleted(sb, pa);
5212 spin_unlock(&pa->pa_lock);
5214 list_del_rcu(&pa->pa_inode_list);
5215 list_add(&pa->u.pa_tmp_list, &discard_list);
5218 if (total_entries <= 5) {
5220 * we want to keep only 5 entries
5221 * allowing it to grow to 8. This
5222 * mak sure we don't call discard
5223 * soon for this list.
5228 spin_unlock(&lg->lg_prealloc_lock);
5230 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5233 group = ext4_get_group_number(sb, pa->pa_pstart);
5234 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5235 GFP_NOFS|__GFP_NOFAIL);
5237 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5241 ext4_lock_group(sb, group);
5242 list_del(&pa->pa_group_list);
5243 ext4_mb_release_group_pa(&e4b, pa);
5244 ext4_unlock_group(sb, group);
5246 ext4_mb_unload_buddy(&e4b);
5247 list_del(&pa->u.pa_tmp_list);
5248 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5253 * We have incremented pa_count. So it cannot be freed at this
5254 * point. Also we hold lg_mutex. So no parallel allocation is
5255 * possible from this lg. That means pa_free cannot be updated.
5257 * A parallel ext4_mb_discard_group_preallocations is possible.
5258 * which can cause the lg_prealloc_list to be updated.
5261 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5263 int order, added = 0, lg_prealloc_count = 1;
5264 struct super_block *sb = ac->ac_sb;
5265 struct ext4_locality_group *lg = ac->ac_lg;
5266 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5268 order = fls(pa->pa_free) - 1;
5269 if (order > PREALLOC_TB_SIZE - 1)
5270 /* The max size of hash table is PREALLOC_TB_SIZE */
5271 order = PREALLOC_TB_SIZE - 1;
5272 /* Add the prealloc space to lg */
5273 spin_lock(&lg->lg_prealloc_lock);
5274 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5276 lockdep_is_held(&lg->lg_prealloc_lock)) {
5277 spin_lock(&tmp_pa->pa_lock);
5278 if (tmp_pa->pa_deleted) {
5279 spin_unlock(&tmp_pa->pa_lock);
5282 if (!added && pa->pa_free < tmp_pa->pa_free) {
5283 /* Add to the tail of the previous entry */
5284 list_add_tail_rcu(&pa->pa_inode_list,
5285 &tmp_pa->pa_inode_list);
5288 * we want to count the total
5289 * number of entries in the list
5292 spin_unlock(&tmp_pa->pa_lock);
5293 lg_prealloc_count++;
5296 list_add_tail_rcu(&pa->pa_inode_list,
5297 &lg->lg_prealloc_list[order]);
5298 spin_unlock(&lg->lg_prealloc_lock);
5300 /* Now trim the list to be not more than 8 elements */
5301 if (lg_prealloc_count > 8) {
5302 ext4_mb_discard_lg_preallocations(sb, lg,
5303 order, lg_prealloc_count);
5310 * if per-inode prealloc list is too long, trim some PA
5312 static void ext4_mb_trim_inode_pa(struct inode *inode)
5314 struct ext4_inode_info *ei = EXT4_I(inode);
5315 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5318 count = atomic_read(&ei->i_prealloc_active);
5319 delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
5320 if (count > sbi->s_mb_max_inode_prealloc + delta) {
5321 count -= sbi->s_mb_max_inode_prealloc;
5322 ext4_discard_preallocations(inode, count);
5327 * release all resource we used in allocation
5329 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
5331 struct inode *inode = ac->ac_inode;
5332 struct ext4_inode_info *ei = EXT4_I(inode);
5333 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5334 struct ext4_prealloc_space *pa = ac->ac_pa;
5336 if (pa->pa_type == MB_GROUP_PA) {
5337 /* see comment in ext4_mb_use_group_pa() */
5338 spin_lock(&pa->pa_lock);
5339 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5340 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5341 pa->pa_free -= ac->ac_b_ex.fe_len;
5342 pa->pa_len -= ac->ac_b_ex.fe_len;
5343 spin_unlock(&pa->pa_lock);
5346 * We want to add the pa to the right bucket.
5347 * Remove it from the list and while adding
5348 * make sure the list to which we are adding
5351 if (likely(pa->pa_free)) {
5352 spin_lock(pa->pa_obj_lock);
5353 list_del_rcu(&pa->pa_inode_list);
5354 spin_unlock(pa->pa_obj_lock);
5355 ext4_mb_add_n_trim(ac);
5359 if (pa->pa_type == MB_INODE_PA) {
5361 * treat per-inode prealloc list as a lru list, then try
5362 * to trim the least recently used PA.
5364 spin_lock(pa->pa_obj_lock);
5365 list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
5366 spin_unlock(pa->pa_obj_lock);
5369 ext4_mb_put_pa(ac, ac->ac_sb, pa);
5371 if (ac->ac_bitmap_page)
5372 put_page(ac->ac_bitmap_page);
5373 if (ac->ac_buddy_page)
5374 put_page(ac->ac_buddy_page);
5375 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5376 mutex_unlock(&ac->ac_lg->lg_mutex);
5377 ext4_mb_collect_stats(ac);
5378 ext4_mb_trim_inode_pa(inode);
5382 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
5384 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
5388 trace_ext4_mb_discard_preallocations(sb, needed);
5389 for (i = 0; i < ngroups && needed > 0; i++) {
5390 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
5398 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
5399 struct ext4_allocation_context *ac, u64 *seq)
5405 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
5410 seq_retry = ext4_get_discard_pa_seq_sum();
5411 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
5412 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
5418 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
5422 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5423 struct ext4_allocation_request *ar, int *errp);
5426 * Main entry point into mballoc to allocate blocks
5427 * it tries to use preallocation first, then falls back
5428 * to usual allocation
5430 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
5431 struct ext4_allocation_request *ar, int *errp)
5433 struct ext4_allocation_context *ac = NULL;
5434 struct ext4_sb_info *sbi;
5435 struct super_block *sb;
5436 ext4_fsblk_t block = 0;
5437 unsigned int inquota = 0;
5438 unsigned int reserv_clstrs = 0;
5442 sb = ar->inode->i_sb;
5445 trace_ext4_request_blocks(ar);
5446 if (sbi->s_mount_state & EXT4_FC_REPLAY)
5447 return ext4_mb_new_blocks_simple(handle, ar, errp);
5449 /* Allow to use superuser reservation for quota file */
5450 if (ext4_is_quota_file(ar->inode))
5451 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
5453 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
5454 /* Without delayed allocation we need to verify
5455 * there is enough free blocks to do block allocation
5456 * and verify allocation doesn't exceed the quota limits.
5459 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
5461 /* let others to free the space */
5463 ar->len = ar->len >> 1;
5466 ext4_mb_show_pa(sb);
5470 reserv_clstrs = ar->len;
5471 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
5472 dquot_alloc_block_nofail(ar->inode,
5473 EXT4_C2B(sbi, ar->len));
5476 dquot_alloc_block(ar->inode,
5477 EXT4_C2B(sbi, ar->len))) {
5479 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
5490 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
5497 *errp = ext4_mb_initialize_context(ac, ar);
5503 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
5504 seq = this_cpu_read(discard_pa_seq);
5505 if (!ext4_mb_use_preallocated(ac)) {
5506 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
5507 ext4_mb_normalize_request(ac, ar);
5509 *errp = ext4_mb_pa_alloc(ac);
5513 /* allocate space in core */
5514 *errp = ext4_mb_regular_allocator(ac);
5516 * pa allocated above is added to grp->bb_prealloc_list only
5517 * when we were able to allocate some block i.e. when
5518 * ac->ac_status == AC_STATUS_FOUND.
5519 * And error from above mean ac->ac_status != AC_STATUS_FOUND
5520 * So we have to free this pa here itself.
5523 ext4_mb_pa_free(ac);
5524 ext4_discard_allocated_blocks(ac);
5527 if (ac->ac_status == AC_STATUS_FOUND &&
5528 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
5529 ext4_mb_pa_free(ac);
5531 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
5532 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
5534 ext4_discard_allocated_blocks(ac);
5537 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5538 ar->len = ac->ac_b_ex.fe_len;
5541 if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
5544 * If block allocation fails then the pa allocated above
5545 * needs to be freed here itself.
5547 ext4_mb_pa_free(ac);
5553 ac->ac_b_ex.fe_len = 0;
5555 ext4_mb_show_ac(ac);
5557 ext4_mb_release_context(ac);
5560 kmem_cache_free(ext4_ac_cachep, ac);
5561 if (inquota && ar->len < inquota)
5562 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
5564 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
5565 /* release all the reserved blocks if non delalloc */
5566 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
5570 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
5576 * We can merge two free data extents only if the physical blocks
5577 * are contiguous, AND the extents were freed by the same transaction,
5578 * AND the blocks are associated with the same group.
5580 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
5581 struct ext4_free_data *entry,
5582 struct ext4_free_data *new_entry,
5583 struct rb_root *entry_rb_root)
5585 if ((entry->efd_tid != new_entry->efd_tid) ||
5586 (entry->efd_group != new_entry->efd_group))
5588 if (entry->efd_start_cluster + entry->efd_count ==
5589 new_entry->efd_start_cluster) {
5590 new_entry->efd_start_cluster = entry->efd_start_cluster;
5591 new_entry->efd_count += entry->efd_count;
5592 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
5593 entry->efd_start_cluster) {
5594 new_entry->efd_count += entry->efd_count;
5597 spin_lock(&sbi->s_md_lock);
5598 list_del(&entry->efd_list);
5599 spin_unlock(&sbi->s_md_lock);
5600 rb_erase(&entry->efd_node, entry_rb_root);
5601 kmem_cache_free(ext4_free_data_cachep, entry);
5604 static noinline_for_stack int
5605 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
5606 struct ext4_free_data *new_entry)
5608 ext4_group_t group = e4b->bd_group;
5609 ext4_grpblk_t cluster;
5610 ext4_grpblk_t clusters = new_entry->efd_count;
5611 struct ext4_free_data *entry;
5612 struct ext4_group_info *db = e4b->bd_info;
5613 struct super_block *sb = e4b->bd_sb;
5614 struct ext4_sb_info *sbi = EXT4_SB(sb);
5615 struct rb_node **n = &db->bb_free_root.rb_node, *node;
5616 struct rb_node *parent = NULL, *new_node;
5618 BUG_ON(!ext4_handle_valid(handle));
5619 BUG_ON(e4b->bd_bitmap_page == NULL);
5620 BUG_ON(e4b->bd_buddy_page == NULL);
5622 new_node = &new_entry->efd_node;
5623 cluster = new_entry->efd_start_cluster;
5626 /* first free block exent. We need to
5627 protect buddy cache from being freed,
5628 * otherwise we'll refresh it from
5629 * on-disk bitmap and lose not-yet-available
5631 get_page(e4b->bd_buddy_page);
5632 get_page(e4b->bd_bitmap_page);
5636 entry = rb_entry(parent, struct ext4_free_data, efd_node);
5637 if (cluster < entry->efd_start_cluster)
5639 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
5640 n = &(*n)->rb_right;
5642 ext4_grp_locked_error(sb, group, 0,
5643 ext4_group_first_block_no(sb, group) +
5644 EXT4_C2B(sbi, cluster),
5645 "Block already on to-be-freed list");
5646 kmem_cache_free(ext4_free_data_cachep, new_entry);
5651 rb_link_node(new_node, parent, n);
5652 rb_insert_color(new_node, &db->bb_free_root);
5654 /* Now try to see the extent can be merged to left and right */
5655 node = rb_prev(new_node);
5657 entry = rb_entry(node, struct ext4_free_data, efd_node);
5658 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5659 &(db->bb_free_root));
5662 node = rb_next(new_node);
5664 entry = rb_entry(node, struct ext4_free_data, efd_node);
5665 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5666 &(db->bb_free_root));
5669 spin_lock(&sbi->s_md_lock);
5670 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5671 sbi->s_mb_free_pending += clusters;
5672 spin_unlock(&sbi->s_md_lock);
5677 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
5678 * linearly starting at the goal block and also excludes the blocks which
5679 * are going to be in use after fast commit replay.
5681 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5682 struct ext4_allocation_request *ar, int *errp)
5684 struct buffer_head *bitmap_bh;
5685 struct super_block *sb = ar->inode->i_sb;
5687 ext4_grpblk_t blkoff;
5688 int i = sb->s_blocksize;
5689 ext4_fsblk_t goal, block;
5690 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
5693 if (goal < le32_to_cpu(es->s_first_data_block) ||
5694 goal >= ext4_blocks_count(es))
5695 goal = le32_to_cpu(es->s_first_data_block);
5698 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
5699 for (; group < ext4_get_groups_count(sb); group++) {
5700 bitmap_bh = ext4_read_block_bitmap(sb, group);
5701 if (IS_ERR(bitmap_bh)) {
5702 *errp = PTR_ERR(bitmap_bh);
5703 pr_warn("Failed to read block bitmap\n");
5707 ext4_get_group_no_and_offset(sb,
5708 max(ext4_group_first_block_no(sb, group), goal),
5710 i = mb_find_next_zero_bit(bitmap_bh->b_data, sb->s_blocksize,
5713 if (i >= sb->s_blocksize)
5715 if (ext4_fc_replay_check_excluded(sb,
5716 ext4_group_first_block_no(sb, group) + i))
5721 if (group >= ext4_get_groups_count(sb) && i >= sb->s_blocksize)
5724 block = ext4_group_first_block_no(sb, group) + i;
5725 ext4_mb_mark_bb(sb, block, 1, 1);
5731 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
5732 unsigned long count)
5734 struct buffer_head *bitmap_bh;
5735 struct super_block *sb = inode->i_sb;
5736 struct ext4_group_desc *gdp;
5737 struct buffer_head *gdp_bh;
5739 ext4_grpblk_t blkoff;
5740 int already_freed = 0, err, i;
5742 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
5743 bitmap_bh = ext4_read_block_bitmap(sb, group);
5744 if (IS_ERR(bitmap_bh)) {
5745 err = PTR_ERR(bitmap_bh);
5746 pr_warn("Failed to read block bitmap\n");
5749 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
5753 for (i = 0; i < count; i++) {
5754 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data))
5757 mb_clear_bits(bitmap_bh->b_data, blkoff, count);
5758 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
5761 ext4_free_group_clusters_set(
5762 sb, gdp, ext4_free_group_clusters(sb, gdp) +
5763 count - already_freed);
5764 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
5765 ext4_group_desc_csum_set(sb, group, gdp);
5766 ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
5767 sync_dirty_buffer(bitmap_bh);
5768 sync_dirty_buffer(gdp_bh);
5773 * ext4_free_blocks() -- Free given blocks and update quota
5774 * @handle: handle for this transaction
5776 * @bh: optional buffer of the block to be freed
5777 * @block: starting physical block to be freed
5778 * @count: number of blocks to be freed
5779 * @flags: flags used by ext4_free_blocks
5781 void ext4_free_blocks(handle_t *handle, struct inode *inode,
5782 struct buffer_head *bh, ext4_fsblk_t block,
5783 unsigned long count, int flags)
5785 struct buffer_head *bitmap_bh = NULL;
5786 struct super_block *sb = inode->i_sb;
5787 struct ext4_group_desc *gdp;
5788 unsigned int overflow;
5790 struct buffer_head *gd_bh;
5791 ext4_group_t block_group;
5792 struct ext4_sb_info *sbi;
5793 struct ext4_buddy e4b;
5794 unsigned int count_clusters;
5800 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
5801 ext4_free_blocks_simple(inode, block, count);
5808 BUG_ON(block != bh->b_blocknr);
5810 block = bh->b_blocknr;
5813 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5814 !ext4_inode_block_valid(inode, block, count)) {
5815 ext4_error(sb, "Freeing blocks not in datazone - "
5816 "block = %llu, count = %lu", block, count);
5820 ext4_debug("freeing block %llu\n", block);
5821 trace_ext4_free_blocks(inode, block, count, flags);
5823 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
5826 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
5831 * If the extent to be freed does not begin on a cluster
5832 * boundary, we need to deal with partial clusters at the
5833 * beginning and end of the extent. Normally we will free
5834 * blocks at the beginning or the end unless we are explicitly
5835 * requested to avoid doing so.
5837 overflow = EXT4_PBLK_COFF(sbi, block);
5839 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
5840 overflow = sbi->s_cluster_ratio - overflow;
5842 if (count > overflow)
5851 overflow = EXT4_LBLK_COFF(sbi, count);
5853 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
5854 if (count > overflow)
5859 count += sbi->s_cluster_ratio - overflow;
5862 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
5864 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
5866 for (i = 0; i < count; i++) {
5869 bh = sb_find_get_block(inode->i_sb, block + i);
5870 ext4_forget(handle, is_metadata, inode, bh, block + i);
5876 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5878 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5879 ext4_get_group_info(sb, block_group))))
5883 * Check to see if we are freeing blocks across a group
5886 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5887 overflow = EXT4_C2B(sbi, bit) + count -
5888 EXT4_BLOCKS_PER_GROUP(sb);
5891 count_clusters = EXT4_NUM_B2C(sbi, count);
5892 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5893 if (IS_ERR(bitmap_bh)) {
5894 err = PTR_ERR(bitmap_bh);
5898 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5904 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
5905 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
5906 in_range(block, ext4_inode_table(sb, gdp),
5907 sbi->s_itb_per_group) ||
5908 in_range(block + count - 1, ext4_inode_table(sb, gdp),
5909 sbi->s_itb_per_group)) {
5911 ext4_error(sb, "Freeing blocks in system zone - "
5912 "Block = %llu, count = %lu", block, count);
5913 /* err = 0. ext4_std_error should be a no op */
5917 BUFFER_TRACE(bitmap_bh, "getting write access");
5918 err = ext4_journal_get_write_access(handle, bitmap_bh);
5923 * We are about to modify some metadata. Call the journal APIs
5924 * to unshare ->b_data if a currently-committing transaction is
5927 BUFFER_TRACE(gd_bh, "get_write_access");
5928 err = ext4_journal_get_write_access(handle, gd_bh);
5931 #ifdef AGGRESSIVE_CHECK
5934 for (i = 0; i < count_clusters; i++)
5935 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
5938 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
5940 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
5941 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
5942 GFP_NOFS|__GFP_NOFAIL);
5947 * We need to make sure we don't reuse the freed block until after the
5948 * transaction is committed. We make an exception if the inode is to be
5949 * written in writeback mode since writeback mode has weak data
5950 * consistency guarantees.
5952 if (ext4_handle_valid(handle) &&
5953 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
5954 !ext4_should_writeback_data(inode))) {
5955 struct ext4_free_data *new_entry;
5957 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
5960 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
5961 GFP_NOFS|__GFP_NOFAIL);
5962 new_entry->efd_start_cluster = bit;
5963 new_entry->efd_group = block_group;
5964 new_entry->efd_count = count_clusters;
5965 new_entry->efd_tid = handle->h_transaction->t_tid;
5967 ext4_lock_group(sb, block_group);
5968 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
5969 ext4_mb_free_metadata(handle, &e4b, new_entry);
5971 /* need to update group_info->bb_free and bitmap
5972 * with group lock held. generate_buddy look at
5973 * them with group lock_held
5975 if (test_opt(sb, DISCARD)) {
5976 err = ext4_issue_discard(sb, block_group, bit, count,
5978 if (err && err != -EOPNOTSUPP)
5979 ext4_msg(sb, KERN_WARNING, "discard request in"
5980 " group:%d block:%d count:%lu failed"
5981 " with %d", block_group, bit, count,
5984 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
5986 ext4_lock_group(sb, block_group);
5987 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
5988 mb_free_blocks(inode, &e4b, bit, count_clusters);
5991 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
5992 ext4_free_group_clusters_set(sb, gdp, ret);
5993 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
5994 ext4_group_desc_csum_set(sb, block_group, gdp);
5995 ext4_unlock_group(sb, block_group);
5997 if (sbi->s_log_groups_per_flex) {
5998 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5999 atomic64_add(count_clusters,
6000 &sbi_array_rcu_deref(sbi, s_flex_groups,
6001 flex_group)->free_clusters);
6005 * on a bigalloc file system, defer the s_freeclusters_counter
6006 * update to the caller (ext4_remove_space and friends) so they
6007 * can determine if a cluster freed here should be rereserved
6009 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6010 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6011 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6012 percpu_counter_add(&sbi->s_freeclusters_counter,
6016 ext4_mb_unload_buddy(&e4b);
6018 /* We dirtied the bitmap block */
6019 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6020 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6022 /* And the group descriptor block */
6023 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6024 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6028 if (overflow && !err) {
6036 ext4_std_error(sb, err);
6041 * ext4_group_add_blocks() -- Add given blocks to an existing group
6042 * @handle: handle to this transaction
6044 * @block: start physical block to add to the block group
6045 * @count: number of blocks to free
6047 * This marks the blocks as free in the bitmap and buddy.
6049 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6050 ext4_fsblk_t block, unsigned long count)
6052 struct buffer_head *bitmap_bh = NULL;
6053 struct buffer_head *gd_bh;
6054 ext4_group_t block_group;
6057 struct ext4_group_desc *desc;
6058 struct ext4_sb_info *sbi = EXT4_SB(sb);
6059 struct ext4_buddy e4b;
6060 int err = 0, ret, free_clusters_count;
6061 ext4_grpblk_t clusters_freed;
6062 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6063 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6064 unsigned long cluster_count = last_cluster - first_cluster + 1;
6066 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6071 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6073 * Check to see if we are freeing blocks across a group
6076 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6077 ext4_warning(sb, "too many blocks added to group %u",
6083 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
6084 if (IS_ERR(bitmap_bh)) {
6085 err = PTR_ERR(bitmap_bh);
6090 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
6096 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
6097 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
6098 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
6099 in_range(block + count - 1, ext4_inode_table(sb, desc),
6100 sbi->s_itb_per_group)) {
6101 ext4_error(sb, "Adding blocks in system zones - "
6102 "Block = %llu, count = %lu",
6108 BUFFER_TRACE(bitmap_bh, "getting write access");
6109 err = ext4_journal_get_write_access(handle, bitmap_bh);
6114 * We are about to modify some metadata. Call the journal APIs
6115 * to unshare ->b_data if a currently-committing transaction is
6118 BUFFER_TRACE(gd_bh, "get_write_access");
6119 err = ext4_journal_get_write_access(handle, gd_bh);
6123 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
6124 BUFFER_TRACE(bitmap_bh, "clear bit");
6125 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
6126 ext4_error(sb, "bit already cleared for block %llu",
6127 (ext4_fsblk_t)(block + i));
6128 BUFFER_TRACE(bitmap_bh, "bit already cleared");
6134 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6139 * need to update group_info->bb_free and bitmap
6140 * with group lock held. generate_buddy look at
6141 * them with group lock_held
6143 ext4_lock_group(sb, block_group);
6144 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
6145 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6146 free_clusters_count = clusters_freed +
6147 ext4_free_group_clusters(sb, desc);
6148 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
6149 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
6150 ext4_group_desc_csum_set(sb, block_group, desc);
6151 ext4_unlock_group(sb, block_group);
6152 percpu_counter_add(&sbi->s_freeclusters_counter,
6155 if (sbi->s_log_groups_per_flex) {
6156 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6157 atomic64_add(clusters_freed,
6158 &sbi_array_rcu_deref(sbi, s_flex_groups,
6159 flex_group)->free_clusters);
6162 ext4_mb_unload_buddy(&e4b);
6164 /* We dirtied the bitmap block */
6165 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6166 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6168 /* And the group descriptor block */
6169 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6170 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6176 ext4_std_error(sb, err);
6181 * ext4_trim_extent -- function to TRIM one single free extent in the group
6182 * @sb: super block for the file system
6183 * @start: starting block of the free extent in the alloc. group
6184 * @count: number of blocks to TRIM
6185 * @group: alloc. group we are working with
6186 * @e4b: ext4 buddy for the group
6188 * Trim "count" blocks starting at "start" in the "group". To assure that no
6189 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6190 * be called with under the group lock.
6192 static int ext4_trim_extent(struct super_block *sb, int start, int count,
6193 ext4_group_t group, struct ext4_buddy *e4b)
6197 struct ext4_free_extent ex;
6200 trace_ext4_trim_extent(sb, group, start, count);
6202 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6204 ex.fe_start = start;
6205 ex.fe_group = group;
6209 * Mark blocks used, so no one can reuse them while
6212 mb_mark_used(e4b, &ex);
6213 ext4_unlock_group(sb, group);
6214 ret = ext4_issue_discard(sb, group, start, count, NULL);
6215 ext4_lock_group(sb, group);
6216 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6221 * ext4_trim_all_free -- function to trim all free space in alloc. group
6222 * @sb: super block for file system
6223 * @group: group to be trimmed
6224 * @start: first group block to examine
6225 * @max: last group block to examine
6226 * @minblocks: minimum extent block count
6228 * ext4_trim_all_free walks through group's buddy bitmap searching for free
6229 * extents. When the free block is found, ext4_trim_extent is called to TRIM
6233 * ext4_trim_all_free walks through group's block bitmap searching for free
6234 * extents. When the free extent is found, mark it as used in group buddy
6235 * bitmap. Then issue a TRIM command on this extent and free the extent in
6236 * the group buddy bitmap. This is done until whole group is scanned.
6238 static ext4_grpblk_t
6239 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6240 ext4_grpblk_t start, ext4_grpblk_t max,
6241 ext4_grpblk_t minblocks)
6244 ext4_grpblk_t next, count = 0, free_count = 0;
6245 struct ext4_buddy e4b;
6248 trace_ext4_trim_all_free(sb, group, start, max);
6250 ret = ext4_mb_load_buddy(sb, group, &e4b);
6252 ext4_warning(sb, "Error %d loading buddy information for %u",
6256 bitmap = e4b.bd_bitmap;
6258 ext4_lock_group(sb, group);
6259 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
6260 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
6263 start = (e4b.bd_info->bb_first_free > start) ?
6264 e4b.bd_info->bb_first_free : start;
6266 while (start <= max) {
6267 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6270 next = mb_find_next_bit(bitmap, max + 1, start);
6272 if ((next - start) >= minblocks) {
6273 ret = ext4_trim_extent(sb, start,
6274 next - start, group, &e4b);
6275 if (ret && ret != -EOPNOTSUPP)
6278 count += next - start;
6280 free_count += next - start;
6283 if (fatal_signal_pending(current)) {
6284 count = -ERESTARTSYS;
6288 if (need_resched()) {
6289 ext4_unlock_group(sb, group);
6291 ext4_lock_group(sb, group);
6294 if ((e4b.bd_info->bb_free - free_count) < minblocks)
6300 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
6303 ext4_unlock_group(sb, group);
6304 ext4_mb_unload_buddy(&e4b);
6306 ext4_debug("trimmed %d blocks in the group %d\n",
6313 * ext4_trim_fs() -- trim ioctl handle function
6314 * @sb: superblock for filesystem
6315 * @range: fstrim_range structure
6317 * start: First Byte to trim
6318 * len: number of Bytes to trim from start
6319 * minlen: minimum extent length in Bytes
6320 * ext4_trim_fs goes through all allocation groups containing Bytes from
6321 * start to start+len. For each such a group ext4_trim_all_free function
6322 * is invoked to trim all free space.
6324 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6326 struct ext4_group_info *grp;
6327 ext4_group_t group, first_group, last_group;
6328 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6329 uint64_t start, end, minlen, trimmed = 0;
6330 ext4_fsblk_t first_data_blk =
6331 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6332 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6335 start = range->start >> sb->s_blocksize_bits;
6336 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6337 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6338 range->minlen >> sb->s_blocksize_bits);
6340 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6341 start >= max_blks ||
6342 range->len < sb->s_blocksize)
6344 if (end >= max_blks)
6346 if (end <= first_data_blk)
6348 if (start < first_data_blk)
6349 start = first_data_blk;
6351 /* Determine first and last group to examine based on start and end */
6352 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6353 &first_group, &first_cluster);
6354 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6355 &last_group, &last_cluster);
6357 /* end now represents the last cluster to discard in this group */
6358 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6360 for (group = first_group; group <= last_group; group++) {
6361 grp = ext4_get_group_info(sb, group);
6362 /* We only do this if the grp has never been initialized */
6363 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6364 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6370 * For all the groups except the last one, last cluster will
6371 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6372 * change it for the last group, note that last_cluster is
6373 * already computed earlier by ext4_get_group_no_and_offset()
6375 if (group == last_group)
6378 if (grp->bb_free >= minlen) {
6379 cnt = ext4_trim_all_free(sb, group, first_cluster,
6389 * For every group except the first one, we are sure
6390 * that the first cluster to discard will be cluster #0.
6396 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
6399 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6403 /* Iterate all the free extents in the group. */
6405 ext4_mballoc_query_range(
6406 struct super_block *sb,
6408 ext4_grpblk_t start,
6410 ext4_mballoc_query_range_fn formatter,
6415 struct ext4_buddy e4b;
6418 error = ext4_mb_load_buddy(sb, group, &e4b);
6421 bitmap = e4b.bd_bitmap;
6423 ext4_lock_group(sb, group);
6425 start = (e4b.bd_info->bb_first_free > start) ?
6426 e4b.bd_info->bb_first_free : start;
6427 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
6428 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6430 while (start <= end) {
6431 start = mb_find_next_zero_bit(bitmap, end + 1, start);
6434 next = mb_find_next_bit(bitmap, end + 1, start);
6436 ext4_unlock_group(sb, group);
6437 error = formatter(sb, group, start, next - start, priv);
6440 ext4_lock_group(sb, group);
6445 ext4_unlock_group(sb, group);
6447 ext4_mb_unload_buddy(&e4b);