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_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
1913 "corruption or bug in mb_find_extent "
1914 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1915 block, order, needed, ex->fe_group, ex->fe_start,
1916 ex->fe_len, ex->fe_logical);
1924 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1930 int start = ex->fe_start;
1931 int len = ex->fe_len;
1936 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1937 BUG_ON(e4b->bd_group != ex->fe_group);
1938 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1939 mb_check_buddy(e4b);
1940 mb_mark_used_double(e4b, start, len);
1942 this_cpu_inc(discard_pa_seq);
1943 e4b->bd_info->bb_free -= len;
1944 if (e4b->bd_info->bb_first_free == start)
1945 e4b->bd_info->bb_first_free += len;
1947 /* let's maintain fragments counter */
1949 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1950 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1951 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1953 e4b->bd_info->bb_fragments++;
1954 else if (!mlen && !max)
1955 e4b->bd_info->bb_fragments--;
1957 /* let's maintain buddy itself */
1959 ord = mb_find_order_for_block(e4b, start);
1961 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1962 /* the whole chunk may be allocated at once! */
1964 buddy = mb_find_buddy(e4b, ord, &max);
1965 BUG_ON((start >> ord) >= max);
1966 mb_set_bit(start >> ord, buddy);
1967 e4b->bd_info->bb_counters[ord]--;
1974 /* store for history */
1976 ret = len | (ord << 16);
1978 /* we have to split large buddy */
1980 buddy = mb_find_buddy(e4b, ord, &max);
1981 mb_set_bit(start >> ord, buddy);
1982 e4b->bd_info->bb_counters[ord]--;
1985 cur = (start >> ord) & ~1U;
1986 buddy = mb_find_buddy(e4b, ord, &max);
1987 mb_clear_bit(cur, buddy);
1988 mb_clear_bit(cur + 1, buddy);
1989 e4b->bd_info->bb_counters[ord]++;
1990 e4b->bd_info->bb_counters[ord]++;
1992 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1994 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
1995 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1996 mb_check_buddy(e4b);
2002 * Must be called under group lock!
2004 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
2005 struct ext4_buddy *e4b)
2007 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2010 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
2011 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2013 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
2014 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
2015 ret = mb_mark_used(e4b, &ac->ac_b_ex);
2017 /* preallocation can change ac_b_ex, thus we store actually
2018 * allocated blocks for history */
2019 ac->ac_f_ex = ac->ac_b_ex;
2021 ac->ac_status = AC_STATUS_FOUND;
2022 ac->ac_tail = ret & 0xffff;
2023 ac->ac_buddy = ret >> 16;
2026 * take the page reference. We want the page to be pinned
2027 * so that we don't get a ext4_mb_init_cache_call for this
2028 * group until we update the bitmap. That would mean we
2029 * double allocate blocks. The reference is dropped
2030 * in ext4_mb_release_context
2032 ac->ac_bitmap_page = e4b->bd_bitmap_page;
2033 get_page(ac->ac_bitmap_page);
2034 ac->ac_buddy_page = e4b->bd_buddy_page;
2035 get_page(ac->ac_buddy_page);
2036 /* store last allocated for subsequent stream allocation */
2037 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2038 spin_lock(&sbi->s_md_lock);
2039 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2040 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2041 spin_unlock(&sbi->s_md_lock);
2044 * As we've just preallocated more space than
2045 * user requested originally, we store allocated
2046 * space in a special descriptor.
2048 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2049 ext4_mb_new_preallocation(ac);
2053 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2054 struct ext4_buddy *e4b,
2057 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2058 struct ext4_free_extent *bex = &ac->ac_b_ex;
2059 struct ext4_free_extent *gex = &ac->ac_g_ex;
2060 struct ext4_free_extent ex;
2063 if (ac->ac_status == AC_STATUS_FOUND)
2066 * We don't want to scan for a whole year
2068 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2069 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2070 ac->ac_status = AC_STATUS_BREAK;
2075 * Haven't found good chunk so far, let's continue
2077 if (bex->fe_len < gex->fe_len)
2080 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2081 && bex->fe_group == e4b->bd_group) {
2082 /* recheck chunk's availability - we don't know
2083 * when it was found (within this lock-unlock
2085 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
2086 if (max >= gex->fe_len) {
2087 ext4_mb_use_best_found(ac, e4b);
2094 * The routine checks whether found extent is good enough. If it is,
2095 * then the extent gets marked used and flag is set to the context
2096 * to stop scanning. Otherwise, the extent is compared with the
2097 * previous found extent and if new one is better, then it's stored
2098 * in the context. Later, the best found extent will be used, if
2099 * mballoc can't find good enough extent.
2101 * FIXME: real allocation policy is to be designed yet!
2103 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2104 struct ext4_free_extent *ex,
2105 struct ext4_buddy *e4b)
2107 struct ext4_free_extent *bex = &ac->ac_b_ex;
2108 struct ext4_free_extent *gex = &ac->ac_g_ex;
2110 BUG_ON(ex->fe_len <= 0);
2111 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2112 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2113 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2118 * The special case - take what you catch first
2120 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2122 ext4_mb_use_best_found(ac, e4b);
2127 * Let's check whether the chuck is good enough
2129 if (ex->fe_len == gex->fe_len) {
2131 ext4_mb_use_best_found(ac, e4b);
2136 * If this is first found extent, just store it in the context
2138 if (bex->fe_len == 0) {
2144 * If new found extent is better, store it in the context
2146 if (bex->fe_len < gex->fe_len) {
2147 /* if the request isn't satisfied, any found extent
2148 * larger than previous best one is better */
2149 if (ex->fe_len > bex->fe_len)
2151 } else if (ex->fe_len > gex->fe_len) {
2152 /* if the request is satisfied, then we try to find
2153 * an extent that still satisfy the request, but is
2154 * smaller than previous one */
2155 if (ex->fe_len < bex->fe_len)
2159 ext4_mb_check_limits(ac, e4b, 0);
2162 static noinline_for_stack
2163 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2164 struct ext4_buddy *e4b)
2166 struct ext4_free_extent ex = ac->ac_b_ex;
2167 ext4_group_t group = ex.fe_group;
2171 BUG_ON(ex.fe_len <= 0);
2172 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2176 ext4_lock_group(ac->ac_sb, group);
2177 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2181 ext4_mb_use_best_found(ac, e4b);
2184 ext4_unlock_group(ac->ac_sb, group);
2185 ext4_mb_unload_buddy(e4b);
2190 static noinline_for_stack
2191 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2192 struct ext4_buddy *e4b)
2194 ext4_group_t group = ac->ac_g_ex.fe_group;
2197 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2198 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2199 struct ext4_free_extent ex;
2201 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
2203 if (grp->bb_free == 0)
2206 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2210 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
2211 ext4_mb_unload_buddy(e4b);
2215 ext4_lock_group(ac->ac_sb, group);
2216 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2217 ac->ac_g_ex.fe_len, &ex);
2218 ex.fe_logical = 0xDEADFA11; /* debug value */
2220 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
2223 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
2225 /* use do_div to get remainder (would be 64-bit modulo) */
2226 if (do_div(start, sbi->s_stripe) == 0) {
2229 ext4_mb_use_best_found(ac, e4b);
2231 } else if (max >= ac->ac_g_ex.fe_len) {
2232 BUG_ON(ex.fe_len <= 0);
2233 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2234 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2237 ext4_mb_use_best_found(ac, e4b);
2238 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2239 /* Sometimes, caller may want to merge even small
2240 * number of blocks to an existing extent */
2241 BUG_ON(ex.fe_len <= 0);
2242 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2243 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2246 ext4_mb_use_best_found(ac, e4b);
2248 ext4_unlock_group(ac->ac_sb, group);
2249 ext4_mb_unload_buddy(e4b);
2255 * The routine scans buddy structures (not bitmap!) from given order
2256 * to max order and tries to find big enough chunk to satisfy the req
2258 static noinline_for_stack
2259 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2260 struct ext4_buddy *e4b)
2262 struct super_block *sb = ac->ac_sb;
2263 struct ext4_group_info *grp = e4b->bd_info;
2269 BUG_ON(ac->ac_2order <= 0);
2270 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2271 if (grp->bb_counters[i] == 0)
2274 buddy = mb_find_buddy(e4b, i, &max);
2275 BUG_ON(buddy == NULL);
2277 k = mb_find_next_zero_bit(buddy, max, 0);
2279 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2280 "%d free clusters of order %d. But found 0",
2281 grp->bb_counters[i], i);
2282 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2284 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2289 ac->ac_b_ex.fe_len = 1 << i;
2290 ac->ac_b_ex.fe_start = k << i;
2291 ac->ac_b_ex.fe_group = e4b->bd_group;
2293 ext4_mb_use_best_found(ac, e4b);
2295 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2297 if (EXT4_SB(sb)->s_mb_stats)
2298 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2305 * The routine scans the group and measures all found extents.
2306 * In order to optimize scanning, caller must pass number of
2307 * free blocks in the group, so the routine can know upper limit.
2309 static noinline_for_stack
2310 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2311 struct ext4_buddy *e4b)
2313 struct super_block *sb = ac->ac_sb;
2314 void *bitmap = e4b->bd_bitmap;
2315 struct ext4_free_extent ex;
2319 free = e4b->bd_info->bb_free;
2320 if (WARN_ON(free <= 0))
2323 i = e4b->bd_info->bb_first_free;
2325 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2326 i = mb_find_next_zero_bit(bitmap,
2327 EXT4_CLUSTERS_PER_GROUP(sb), i);
2328 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2330 * IF we have corrupt bitmap, we won't find any
2331 * free blocks even though group info says we
2334 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2335 "%d free clusters as per "
2336 "group info. But bitmap says 0",
2338 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2339 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2343 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2344 if (WARN_ON(ex.fe_len <= 0))
2346 if (free < ex.fe_len) {
2347 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2348 "%d free clusters as per "
2349 "group info. But got %d blocks",
2351 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2352 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2354 * The number of free blocks differs. This mostly
2355 * indicate that the bitmap is corrupt. So exit
2356 * without claiming the space.
2360 ex.fe_logical = 0xDEADC0DE; /* debug value */
2361 ext4_mb_measure_extent(ac, &ex, e4b);
2367 ext4_mb_check_limits(ac, e4b, 1);
2371 * This is a special case for storages like raid5
2372 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2374 static noinline_for_stack
2375 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2376 struct ext4_buddy *e4b)
2378 struct super_block *sb = ac->ac_sb;
2379 struct ext4_sb_info *sbi = EXT4_SB(sb);
2380 void *bitmap = e4b->bd_bitmap;
2381 struct ext4_free_extent ex;
2382 ext4_fsblk_t first_group_block;
2387 BUG_ON(sbi->s_stripe == 0);
2389 /* find first stripe-aligned block in group */
2390 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2392 a = first_group_block + sbi->s_stripe - 1;
2393 do_div(a, sbi->s_stripe);
2394 i = (a * sbi->s_stripe) - first_group_block;
2396 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2397 if (!mb_test_bit(i, bitmap)) {
2398 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2399 if (max >= sbi->s_stripe) {
2401 ex.fe_logical = 0xDEADF00D; /* debug value */
2403 ext4_mb_use_best_found(ac, e4b);
2412 * This is also called BEFORE we load the buddy bitmap.
2413 * Returns either 1 or 0 indicating that the group is either suitable
2414 * for the allocation or not.
2416 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2417 ext4_group_t group, int cr)
2419 ext4_grpblk_t free, fragments;
2420 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2421 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2423 BUG_ON(cr < 0 || cr >= 4);
2425 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2428 free = grp->bb_free;
2432 fragments = grp->bb_fragments;
2438 BUG_ON(ac->ac_2order == 0);
2440 /* Avoid using the first bg of a flexgroup for data files */
2441 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2442 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2443 ((group % flex_size) == 0))
2446 if (free < ac->ac_g_ex.fe_len)
2449 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2452 if (grp->bb_largest_free_order < ac->ac_2order)
2457 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2461 if (free >= ac->ac_g_ex.fe_len)
2474 * This could return negative error code if something goes wrong
2475 * during ext4_mb_init_group(). This should not be called with
2476 * ext4_lock_group() held.
2478 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2479 ext4_group_t group, int cr)
2481 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2482 struct super_block *sb = ac->ac_sb;
2483 struct ext4_sb_info *sbi = EXT4_SB(sb);
2484 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2488 if (sbi->s_mb_stats)
2489 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2491 ext4_lock_group(sb, group);
2492 free = grp->bb_free;
2495 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2497 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2500 ext4_unlock_group(sb, group);
2502 /* We only do this if the grp has never been initialized */
2503 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2504 struct ext4_group_desc *gdp =
2505 ext4_get_group_desc(sb, group, NULL);
2508 /* cr=0/1 is a very optimistic search to find large
2509 * good chunks almost for free. If buddy data is not
2510 * ready, then this optimization makes no sense. But
2511 * we never skip the first block group in a flex_bg,
2512 * since this gets used for metadata block allocation,
2513 * and we want to make sure we locate metadata blocks
2514 * in the first block group in the flex_bg if possible.
2517 (!sbi->s_log_groups_per_flex ||
2518 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2519 !(ext4_has_group_desc_csum(sb) &&
2520 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2522 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2528 ext4_lock_group(sb, group);
2529 ret = ext4_mb_good_group(ac, group, cr);
2532 ext4_unlock_group(sb, group);
2537 * Start prefetching @nr block bitmaps starting at @group.
2538 * Return the next group which needs to be prefetched.
2540 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2541 unsigned int nr, int *cnt)
2543 ext4_group_t ngroups = ext4_get_groups_count(sb);
2544 struct buffer_head *bh;
2545 struct blk_plug plug;
2547 blk_start_plug(&plug);
2549 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2551 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2554 * Prefetch block groups with free blocks; but don't
2555 * bother if it is marked uninitialized on disk, since
2556 * it won't require I/O to read. Also only try to
2557 * prefetch once, so we avoid getblk() call, which can
2560 if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2561 EXT4_MB_GRP_NEED_INIT(grp) &&
2562 ext4_free_group_clusters(sb, gdp) > 0 &&
2563 !(ext4_has_group_desc_csum(sb) &&
2564 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2565 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2566 if (bh && !IS_ERR(bh)) {
2567 if (!buffer_uptodate(bh) && cnt)
2572 if (++group >= ngroups)
2575 blk_finish_plug(&plug);
2580 * Prefetching reads the block bitmap into the buffer cache; but we
2581 * need to make sure that the buddy bitmap in the page cache has been
2582 * initialized. Note that ext4_mb_init_group() will block if the I/O
2583 * is not yet completed, or indeed if it was not initiated by
2584 * ext4_mb_prefetch did not start the I/O.
2586 * TODO: We should actually kick off the buddy bitmap setup in a work
2587 * queue when the buffer I/O is completed, so that we don't block
2588 * waiting for the block allocation bitmap read to finish when
2589 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2591 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2595 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2597 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2600 group = ext4_get_groups_count(sb);
2602 grp = ext4_get_group_info(sb, group);
2604 if (EXT4_MB_GRP_NEED_INIT(grp) &&
2605 ext4_free_group_clusters(sb, gdp) > 0 &&
2606 !(ext4_has_group_desc_csum(sb) &&
2607 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2608 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2614 static noinline_for_stack int
2615 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2617 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2619 int err = 0, first_err = 0;
2620 unsigned int nr = 0, prefetch_ios = 0;
2621 struct ext4_sb_info *sbi;
2622 struct super_block *sb;
2623 struct ext4_buddy e4b;
2628 ngroups = ext4_get_groups_count(sb);
2629 /* non-extent files are limited to low blocks/groups */
2630 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2631 ngroups = sbi->s_blockfile_groups;
2633 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2635 /* first, try the goal */
2636 err = ext4_mb_find_by_goal(ac, &e4b);
2637 if (err || ac->ac_status == AC_STATUS_FOUND)
2640 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2644 * ac->ac_2order is set only if the fe_len is a power of 2
2645 * if ac->ac_2order is set we also set criteria to 0 so that we
2646 * try exact allocation using buddy.
2648 i = fls(ac->ac_g_ex.fe_len);
2651 * We search using buddy data only if the order of the request
2652 * is greater than equal to the sbi_s_mb_order2_reqs
2653 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2654 * We also support searching for power-of-two requests only for
2655 * requests upto maximum buddy size we have constructed.
2657 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2659 * This should tell if fe_len is exactly power of 2
2661 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2662 ac->ac_2order = array_index_nospec(i - 1,
2666 /* if stream allocation is enabled, use global goal */
2667 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2668 /* TBD: may be hot point */
2669 spin_lock(&sbi->s_md_lock);
2670 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2671 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2672 spin_unlock(&sbi->s_md_lock);
2675 /* Let's just scan groups to find more-less suitable blocks */
2676 cr = ac->ac_2order ? 0 : 1;
2678 * cr == 0 try to get exact allocation,
2679 * cr == 3 try to get anything
2682 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2683 ac->ac_criteria = cr;
2685 * searching for the right group start
2686 * from the goal value specified
2688 group = ac->ac_g_ex.fe_group;
2689 ac->ac_last_optimal_group = group;
2690 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2691 prefetch_grp = group;
2693 for (i = 0; i < ngroups; group = next_linear_group(ac, group, ngroups),
2695 int ret = 0, new_cr;
2699 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups);
2706 * Batch reads of the block allocation bitmaps
2707 * to get multiple READs in flight; limit
2708 * prefetching at cr=0/1, otherwise mballoc can
2709 * spend a lot of time loading imperfect groups
2711 if ((prefetch_grp == group) &&
2713 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2714 unsigned int curr_ios = prefetch_ios;
2716 nr = sbi->s_mb_prefetch;
2717 if (ext4_has_feature_flex_bg(sb)) {
2718 nr = 1 << sbi->s_log_groups_per_flex;
2719 nr -= group & (nr - 1);
2720 nr = min(nr, sbi->s_mb_prefetch);
2722 prefetch_grp = ext4_mb_prefetch(sb, group,
2724 if (prefetch_ios == curr_ios)
2728 /* This now checks without needing the buddy page */
2729 ret = ext4_mb_good_group_nolock(ac, group, cr);
2736 err = ext4_mb_load_buddy(sb, group, &e4b);
2740 ext4_lock_group(sb, group);
2743 * We need to check again after locking the
2746 ret = ext4_mb_good_group(ac, group, cr);
2748 ext4_unlock_group(sb, group);
2749 ext4_mb_unload_buddy(&e4b);
2753 ac->ac_groups_scanned++;
2755 ext4_mb_simple_scan_group(ac, &e4b);
2756 else if (cr == 1 && sbi->s_stripe &&
2757 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2758 ext4_mb_scan_aligned(ac, &e4b);
2760 ext4_mb_complex_scan_group(ac, &e4b);
2762 ext4_unlock_group(sb, group);
2763 ext4_mb_unload_buddy(&e4b);
2765 if (ac->ac_status != AC_STATUS_CONTINUE)
2768 /* Processed all groups and haven't found blocks */
2769 if (sbi->s_mb_stats && i == ngroups)
2770 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2773 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2774 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2776 * We've been searching too long. Let's try to allocate
2777 * the best chunk we've found so far
2779 ext4_mb_try_best_found(ac, &e4b);
2780 if (ac->ac_status != AC_STATUS_FOUND) {
2782 * Someone more lucky has already allocated it.
2783 * The only thing we can do is just take first
2786 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2787 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2788 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2789 ac->ac_b_ex.fe_len, lost);
2791 ac->ac_b_ex.fe_group = 0;
2792 ac->ac_b_ex.fe_start = 0;
2793 ac->ac_b_ex.fe_len = 0;
2794 ac->ac_status = AC_STATUS_CONTINUE;
2795 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2801 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2802 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2804 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2807 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2808 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2809 ac->ac_flags, cr, err);
2812 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2817 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2819 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2822 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2825 return (void *) ((unsigned long) group);
2828 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2830 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2834 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2837 return (void *) ((unsigned long) group);
2840 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2842 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2843 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2845 int err, buddy_loaded = 0;
2846 struct ext4_buddy e4b;
2847 struct ext4_group_info *grinfo;
2848 unsigned char blocksize_bits = min_t(unsigned char,
2849 sb->s_blocksize_bits,
2850 EXT4_MAX_BLOCK_LOG_SIZE);
2852 struct ext4_group_info info;
2853 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2858 seq_puts(seq, "#group: free frags first ["
2859 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2860 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2862 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2863 sizeof(struct ext4_group_info);
2865 grinfo = ext4_get_group_info(sb, group);
2866 /* Load the group info in memory only if not already loaded. */
2867 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2868 err = ext4_mb_load_buddy(sb, group, &e4b);
2870 seq_printf(seq, "#%-5u: I/O error\n", group);
2876 memcpy(&sg, ext4_get_group_info(sb, group), i);
2879 ext4_mb_unload_buddy(&e4b);
2881 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2882 sg.info.bb_fragments, sg.info.bb_first_free);
2883 for (i = 0; i <= 13; i++)
2884 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2885 sg.info.bb_counters[i] : 0);
2886 seq_puts(seq, " ]\n");
2891 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2895 const struct seq_operations ext4_mb_seq_groups_ops = {
2896 .start = ext4_mb_seq_groups_start,
2897 .next = ext4_mb_seq_groups_next,
2898 .stop = ext4_mb_seq_groups_stop,
2899 .show = ext4_mb_seq_groups_show,
2902 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
2904 struct super_block *sb = (struct super_block *)seq->private;
2905 struct ext4_sb_info *sbi = EXT4_SB(sb);
2907 seq_puts(seq, "mballoc:\n");
2908 if (!sbi->s_mb_stats) {
2909 seq_puts(seq, "\tmb stats collection turned off.\n");
2910 seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
2913 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
2914 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
2916 seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned));
2918 seq_puts(seq, "\tcr0_stats:\n");
2919 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0]));
2920 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2921 atomic64_read(&sbi->s_bal_cX_groups_considered[0]));
2922 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2923 atomic64_read(&sbi->s_bal_cX_failed[0]));
2924 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2925 atomic_read(&sbi->s_bal_cr0_bad_suggestions));
2927 seq_puts(seq, "\tcr1_stats:\n");
2928 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1]));
2929 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2930 atomic64_read(&sbi->s_bal_cX_groups_considered[1]));
2931 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2932 atomic64_read(&sbi->s_bal_cX_failed[1]));
2933 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2934 atomic_read(&sbi->s_bal_cr1_bad_suggestions));
2936 seq_puts(seq, "\tcr2_stats:\n");
2937 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2]));
2938 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2939 atomic64_read(&sbi->s_bal_cX_groups_considered[2]));
2940 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2941 atomic64_read(&sbi->s_bal_cX_failed[2]));
2943 seq_puts(seq, "\tcr3_stats:\n");
2944 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3]));
2945 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2946 atomic64_read(&sbi->s_bal_cX_groups_considered[3]));
2947 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2948 atomic64_read(&sbi->s_bal_cX_failed[3]));
2949 seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned));
2950 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
2951 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
2952 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
2953 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
2955 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
2956 atomic_read(&sbi->s_mb_buddies_generated),
2957 ext4_get_groups_count(sb));
2958 seq_printf(seq, "\tbuddies_time_used: %llu\n",
2959 atomic64_read(&sbi->s_mb_generation_time));
2960 seq_printf(seq, "\tpreallocated: %u\n",
2961 atomic_read(&sbi->s_mb_preallocated));
2962 seq_printf(seq, "\tdiscarded: %u\n",
2963 atomic_read(&sbi->s_mb_discarded));
2967 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
2969 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2970 unsigned long position;
2972 read_lock(&EXT4_SB(sb)->s_mb_rb_lock);
2974 if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
2976 position = *pos + 1;
2977 return (void *) ((unsigned long) position);
2980 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
2982 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2983 unsigned long position;
2986 if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
2988 position = *pos + 1;
2989 return (void *) ((unsigned long) position);
2992 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
2994 struct super_block *sb = PDE_DATA(file_inode(seq->file));
2995 struct ext4_sb_info *sbi = EXT4_SB(sb);
2996 unsigned long position = ((unsigned long) v);
2997 struct ext4_group_info *grp;
2999 unsigned int count, min, max;
3002 if (position >= MB_NUM_ORDERS(sb)) {
3003 seq_puts(seq, "fragment_size_tree:\n");
3004 n = rb_first(&sbi->s_mb_avg_fragment_size_root);
3006 seq_puts(seq, "\ttree_min: 0\n\ttree_max: 0\n\ttree_nodes: 0\n");
3009 grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3010 min = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3012 while (rb_next(n)) {
3016 grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3017 max = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3019 seq_printf(seq, "\ttree_min: %u\n\ttree_max: %u\n\ttree_nodes: %u\n",
3024 if (position == 0) {
3025 seq_printf(seq, "optimize_scan: %d\n",
3026 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
3027 seq_puts(seq, "max_free_order_lists:\n");
3030 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
3031 bb_largest_free_order_node)
3033 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3034 (unsigned int)position, count);
3039 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3041 struct super_block *sb = PDE_DATA(file_inode(seq->file));
3043 read_unlock(&EXT4_SB(sb)->s_mb_rb_lock);
3046 const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3047 .start = ext4_mb_seq_structs_summary_start,
3048 .next = ext4_mb_seq_structs_summary_next,
3049 .stop = ext4_mb_seq_structs_summary_stop,
3050 .show = ext4_mb_seq_structs_summary_show,
3053 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3055 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3056 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3063 * Allocate the top-level s_group_info array for the specified number
3066 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3068 struct ext4_sb_info *sbi = EXT4_SB(sb);
3070 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3072 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3073 EXT4_DESC_PER_BLOCK_BITS(sb);
3074 if (size <= sbi->s_group_info_size)
3077 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3078 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3079 if (!new_groupinfo) {
3080 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3084 old_groupinfo = rcu_dereference(sbi->s_group_info);
3086 memcpy(new_groupinfo, old_groupinfo,
3087 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3089 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3090 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3092 ext4_kvfree_array_rcu(old_groupinfo);
3093 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3094 sbi->s_group_info_size);
3098 /* Create and initialize ext4_group_info data for the given group. */
3099 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3100 struct ext4_group_desc *desc)
3104 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3105 struct ext4_sb_info *sbi = EXT4_SB(sb);
3106 struct ext4_group_info **meta_group_info;
3107 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3110 * First check if this group is the first of a reserved block.
3111 * If it's true, we have to allocate a new table of pointers
3112 * to ext4_group_info structures
3114 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3115 metalen = sizeof(*meta_group_info) <<
3116 EXT4_DESC_PER_BLOCK_BITS(sb);
3117 meta_group_info = kmalloc(metalen, GFP_NOFS);
3118 if (meta_group_info == NULL) {
3119 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3120 "for a buddy group");
3121 goto exit_meta_group_info;
3124 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3128 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3129 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3131 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3132 if (meta_group_info[i] == NULL) {
3133 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3134 goto exit_group_info;
3136 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3137 &(meta_group_info[i]->bb_state));
3140 * initialize bb_free to be able to skip
3141 * empty groups without initialization
3143 if (ext4_has_group_desc_csum(sb) &&
3144 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3145 meta_group_info[i]->bb_free =
3146 ext4_free_clusters_after_init(sb, group, desc);
3148 meta_group_info[i]->bb_free =
3149 ext4_free_group_clusters(sb, desc);
3152 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3153 init_rwsem(&meta_group_info[i]->alloc_sem);
3154 meta_group_info[i]->bb_free_root = RB_ROOT;
3155 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3156 RB_CLEAR_NODE(&meta_group_info[i]->bb_avg_fragment_size_rb);
3157 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3158 meta_group_info[i]->bb_group = group;
3160 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3164 /* If a meta_group_info table has been allocated, release it now */
3165 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3166 struct ext4_group_info ***group_info;
3169 group_info = rcu_dereference(sbi->s_group_info);
3170 kfree(group_info[idx]);
3171 group_info[idx] = NULL;
3174 exit_meta_group_info:
3176 } /* ext4_mb_add_groupinfo */
3178 static int ext4_mb_init_backend(struct super_block *sb)
3180 ext4_group_t ngroups = ext4_get_groups_count(sb);
3182 struct ext4_sb_info *sbi = EXT4_SB(sb);
3184 struct ext4_group_desc *desc;
3185 struct ext4_group_info ***group_info;
3186 struct kmem_cache *cachep;
3188 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3192 sbi->s_buddy_cache = new_inode(sb);
3193 if (sbi->s_buddy_cache == NULL) {
3194 ext4_msg(sb, KERN_ERR, "can't get new inode");
3197 /* To avoid potentially colliding with an valid on-disk inode number,
3198 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3199 * not in the inode hash, so it should never be found by iget(), but
3200 * this will avoid confusion if it ever shows up during debugging. */
3201 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3202 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3203 for (i = 0; i < ngroups; i++) {
3205 desc = ext4_get_group_desc(sb, i, NULL);
3207 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3210 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3214 if (ext4_has_feature_flex_bg(sb)) {
3215 /* a single flex group is supposed to be read by a single IO.
3216 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3217 * unsigned integer, so the maximum shift is 32.
3219 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3220 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3223 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3224 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3225 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3227 sbi->s_mb_prefetch = 32;
3229 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3230 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3231 /* now many real IOs to prefetch within a single allocation at cr=0
3232 * given cr=0 is an CPU-related optimization we shouldn't try to
3233 * load too many groups, at some point we should start to use what
3234 * we've got in memory.
3235 * with an average random access time 5ms, it'd take a second to get
3236 * 200 groups (* N with flex_bg), so let's make this limit 4
3238 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3239 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3240 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3245 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3247 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
3248 i = sbi->s_group_info_size;
3250 group_info = rcu_dereference(sbi->s_group_info);
3252 kfree(group_info[i]);
3254 iput(sbi->s_buddy_cache);
3257 kvfree(rcu_dereference(sbi->s_group_info));
3262 static void ext4_groupinfo_destroy_slabs(void)
3266 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3267 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3268 ext4_groupinfo_caches[i] = NULL;
3272 static int ext4_groupinfo_create_slab(size_t size)
3274 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3276 int blocksize_bits = order_base_2(size);
3277 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3278 struct kmem_cache *cachep;
3280 if (cache_index >= NR_GRPINFO_CACHES)
3283 if (unlikely(cache_index < 0))
3286 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3287 if (ext4_groupinfo_caches[cache_index]) {
3288 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3289 return 0; /* Already created */
3292 slab_size = offsetof(struct ext4_group_info,
3293 bb_counters[blocksize_bits + 2]);
3295 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3296 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3299 ext4_groupinfo_caches[cache_index] = cachep;
3301 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3304 "EXT4-fs: no memory for groupinfo slab cache\n");
3311 int ext4_mb_init(struct super_block *sb)
3313 struct ext4_sb_info *sbi = EXT4_SB(sb);
3315 unsigned offset, offset_incr;
3319 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3321 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3322 if (sbi->s_mb_offsets == NULL) {
3327 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3328 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3329 if (sbi->s_mb_maxs == NULL) {
3334 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3338 /* order 0 is regular bitmap */
3339 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3340 sbi->s_mb_offsets[0] = 0;
3344 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3345 max = sb->s_blocksize << 2;
3347 sbi->s_mb_offsets[i] = offset;
3348 sbi->s_mb_maxs[i] = max;
3349 offset += offset_incr;
3350 offset_incr = offset_incr >> 1;
3353 } while (i < MB_NUM_ORDERS(sb));
3355 sbi->s_mb_avg_fragment_size_root = RB_ROOT;
3356 sbi->s_mb_largest_free_orders =
3357 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3359 if (!sbi->s_mb_largest_free_orders) {
3363 sbi->s_mb_largest_free_orders_locks =
3364 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3366 if (!sbi->s_mb_largest_free_orders_locks) {
3370 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3371 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3372 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3374 rwlock_init(&sbi->s_mb_rb_lock);
3376 spin_lock_init(&sbi->s_md_lock);
3377 sbi->s_mb_free_pending = 0;
3378 INIT_LIST_HEAD(&sbi->s_freed_data_list);
3380 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3381 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3382 sbi->s_mb_stats = MB_DEFAULT_STATS;
3383 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3384 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3385 sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
3387 * The default group preallocation is 512, which for 4k block
3388 * sizes translates to 2 megabytes. However for bigalloc file
3389 * systems, this is probably too big (i.e, if the cluster size
3390 * is 1 megabyte, then group preallocation size becomes half a
3391 * gigabyte!). As a default, we will keep a two megabyte
3392 * group pralloc size for cluster sizes up to 64k, and after
3393 * that, we will force a minimum group preallocation size of
3394 * 32 clusters. This translates to 8 megs when the cluster
3395 * size is 256k, and 32 megs when the cluster size is 1 meg,
3396 * which seems reasonable as a default.
3398 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3399 sbi->s_cluster_bits, 32);
3401 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3402 * to the lowest multiple of s_stripe which is bigger than
3403 * the s_mb_group_prealloc as determined above. We want
3404 * the preallocation size to be an exact multiple of the
3405 * RAID stripe size so that preallocations don't fragment
3408 if (sbi->s_stripe > 1) {
3409 sbi->s_mb_group_prealloc = roundup(
3410 sbi->s_mb_group_prealloc, sbi->s_stripe);
3413 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3414 if (sbi->s_locality_groups == NULL) {
3418 for_each_possible_cpu(i) {
3419 struct ext4_locality_group *lg;
3420 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3421 mutex_init(&lg->lg_mutex);
3422 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3423 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3424 spin_lock_init(&lg->lg_prealloc_lock);
3427 if (blk_queue_nonrot(bdev_get_queue(sb->s_bdev)))
3428 sbi->s_mb_max_linear_groups = 0;
3430 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3431 /* init file for buddy data */
3432 ret = ext4_mb_init_backend(sb);
3434 goto out_free_locality_groups;
3438 out_free_locality_groups:
3439 free_percpu(sbi->s_locality_groups);
3440 sbi->s_locality_groups = NULL;
3442 kfree(sbi->s_mb_largest_free_orders);
3443 kfree(sbi->s_mb_largest_free_orders_locks);
3444 kfree(sbi->s_mb_offsets);
3445 sbi->s_mb_offsets = NULL;
3446 kfree(sbi->s_mb_maxs);
3447 sbi->s_mb_maxs = NULL;
3451 /* need to called with the ext4 group lock held */
3452 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3454 struct ext4_prealloc_space *pa;
3455 struct list_head *cur, *tmp;
3458 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3459 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3460 list_del(&pa->pa_group_list);
3462 kmem_cache_free(ext4_pspace_cachep, pa);
3467 int ext4_mb_release(struct super_block *sb)
3469 ext4_group_t ngroups = ext4_get_groups_count(sb);
3471 int num_meta_group_infos;
3472 struct ext4_group_info *grinfo, ***group_info;
3473 struct ext4_sb_info *sbi = EXT4_SB(sb);
3474 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3477 if (sbi->s_group_info) {
3478 for (i = 0; i < ngroups; i++) {
3480 grinfo = ext4_get_group_info(sb, i);
3481 mb_group_bb_bitmap_free(grinfo);
3482 ext4_lock_group(sb, i);
3483 count = ext4_mb_cleanup_pa(grinfo);
3485 mb_debug(sb, "mballoc: %d PAs left\n",
3487 ext4_unlock_group(sb, i);
3488 kmem_cache_free(cachep, grinfo);
3490 num_meta_group_infos = (ngroups +
3491 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3492 EXT4_DESC_PER_BLOCK_BITS(sb);
3494 group_info = rcu_dereference(sbi->s_group_info);
3495 for (i = 0; i < num_meta_group_infos; i++)
3496 kfree(group_info[i]);
3500 kfree(sbi->s_mb_largest_free_orders);
3501 kfree(sbi->s_mb_largest_free_orders_locks);
3502 kfree(sbi->s_mb_offsets);
3503 kfree(sbi->s_mb_maxs);
3504 iput(sbi->s_buddy_cache);
3505 if (sbi->s_mb_stats) {
3506 ext4_msg(sb, KERN_INFO,
3507 "mballoc: %u blocks %u reqs (%u success)",
3508 atomic_read(&sbi->s_bal_allocated),
3509 atomic_read(&sbi->s_bal_reqs),
3510 atomic_read(&sbi->s_bal_success));
3511 ext4_msg(sb, KERN_INFO,
3512 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3513 "%u 2^N hits, %u breaks, %u lost",
3514 atomic_read(&sbi->s_bal_ex_scanned),
3515 atomic_read(&sbi->s_bal_groups_scanned),
3516 atomic_read(&sbi->s_bal_goals),
3517 atomic_read(&sbi->s_bal_2orders),
3518 atomic_read(&sbi->s_bal_breaks),
3519 atomic_read(&sbi->s_mb_lost_chunks));
3520 ext4_msg(sb, KERN_INFO,
3521 "mballoc: %u generated and it took %llu",
3522 atomic_read(&sbi->s_mb_buddies_generated),
3523 atomic64_read(&sbi->s_mb_generation_time));
3524 ext4_msg(sb, KERN_INFO,
3525 "mballoc: %u preallocated, %u discarded",
3526 atomic_read(&sbi->s_mb_preallocated),
3527 atomic_read(&sbi->s_mb_discarded));
3530 free_percpu(sbi->s_locality_groups);
3535 static inline int ext4_issue_discard(struct super_block *sb,
3536 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3539 ext4_fsblk_t discard_block;
3541 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3542 ext4_group_first_block_no(sb, block_group));
3543 count = EXT4_C2B(EXT4_SB(sb), count);
3544 trace_ext4_discard_blocks(sb,
3545 (unsigned long long) discard_block, count);
3547 return __blkdev_issue_discard(sb->s_bdev,
3548 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3549 (sector_t)count << (sb->s_blocksize_bits - 9),
3552 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3555 static void ext4_free_data_in_buddy(struct super_block *sb,
3556 struct ext4_free_data *entry)
3558 struct ext4_buddy e4b;
3559 struct ext4_group_info *db;
3560 int err, count = 0, count2 = 0;
3562 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3563 entry->efd_count, entry->efd_group, entry);
3565 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3566 /* we expect to find existing buddy because it's pinned */
3569 spin_lock(&EXT4_SB(sb)->s_md_lock);
3570 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3571 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3574 /* there are blocks to put in buddy to make them really free */
3575 count += entry->efd_count;
3577 ext4_lock_group(sb, entry->efd_group);
3578 /* Take it out of per group rb tree */
3579 rb_erase(&entry->efd_node, &(db->bb_free_root));
3580 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3583 * Clear the trimmed flag for the group so that the next
3584 * ext4_trim_fs can trim it.
3585 * If the volume is mounted with -o discard, online discard
3586 * is supported and the free blocks will be trimmed online.
3588 if (!test_opt(sb, DISCARD))
3589 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3591 if (!db->bb_free_root.rb_node) {
3592 /* No more items in the per group rb tree
3593 * balance refcounts from ext4_mb_free_metadata()
3595 put_page(e4b.bd_buddy_page);
3596 put_page(e4b.bd_bitmap_page);
3598 ext4_unlock_group(sb, entry->efd_group);
3599 kmem_cache_free(ext4_free_data_cachep, entry);
3600 ext4_mb_unload_buddy(&e4b);
3602 mb_debug(sb, "freed %d blocks in %d structures\n", count,
3607 * This function is called by the jbd2 layer once the commit has finished,
3608 * so we know we can free the blocks that were released with that commit.
3610 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3612 struct ext4_sb_info *sbi = EXT4_SB(sb);
3613 struct ext4_free_data *entry, *tmp;
3614 struct bio *discard_bio = NULL;
3615 struct list_head freed_data_list;
3616 struct list_head *cut_pos = NULL;
3619 INIT_LIST_HEAD(&freed_data_list);
3621 spin_lock(&sbi->s_md_lock);
3622 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3623 if (entry->efd_tid != commit_tid)
3625 cut_pos = &entry->efd_list;
3628 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3630 spin_unlock(&sbi->s_md_lock);
3632 if (test_opt(sb, DISCARD)) {
3633 list_for_each_entry(entry, &freed_data_list, efd_list) {
3634 err = ext4_issue_discard(sb, entry->efd_group,
3635 entry->efd_start_cluster,
3638 if (err && err != -EOPNOTSUPP) {
3639 ext4_msg(sb, KERN_WARNING, "discard request in"
3640 " group:%d block:%d count:%d failed"
3641 " with %d", entry->efd_group,
3642 entry->efd_start_cluster,
3643 entry->efd_count, err);
3644 } else if (err == -EOPNOTSUPP)
3649 submit_bio_wait(discard_bio);
3650 bio_put(discard_bio);
3654 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3655 ext4_free_data_in_buddy(sb, entry);
3658 int __init ext4_init_mballoc(void)
3660 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3661 SLAB_RECLAIM_ACCOUNT);
3662 if (ext4_pspace_cachep == NULL)
3665 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3666 SLAB_RECLAIM_ACCOUNT);
3667 if (ext4_ac_cachep == NULL)
3670 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3671 SLAB_RECLAIM_ACCOUNT);
3672 if (ext4_free_data_cachep == NULL)
3678 kmem_cache_destroy(ext4_ac_cachep);
3680 kmem_cache_destroy(ext4_pspace_cachep);
3685 void ext4_exit_mballoc(void)
3688 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3689 * before destroying the slab cache.
3692 kmem_cache_destroy(ext4_pspace_cachep);
3693 kmem_cache_destroy(ext4_ac_cachep);
3694 kmem_cache_destroy(ext4_free_data_cachep);
3695 ext4_groupinfo_destroy_slabs();
3700 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3701 * Returns 0 if success or error code
3703 static noinline_for_stack int
3704 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3705 handle_t *handle, unsigned int reserv_clstrs)
3707 struct buffer_head *bitmap_bh = NULL;
3708 struct ext4_group_desc *gdp;
3709 struct buffer_head *gdp_bh;
3710 struct ext4_sb_info *sbi;
3711 struct super_block *sb;
3715 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3716 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3721 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3722 if (IS_ERR(bitmap_bh)) {
3723 err = PTR_ERR(bitmap_bh);
3728 BUFFER_TRACE(bitmap_bh, "getting write access");
3729 err = ext4_journal_get_write_access(handle, bitmap_bh);
3734 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3738 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3739 ext4_free_group_clusters(sb, gdp));
3741 BUFFER_TRACE(gdp_bh, "get_write_access");
3742 err = ext4_journal_get_write_access(handle, gdp_bh);
3746 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3748 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3749 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3750 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3751 "fs metadata", block, block+len);
3752 /* File system mounted not to panic on error
3753 * Fix the bitmap and return EFSCORRUPTED
3754 * We leak some of the blocks here.
3756 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3757 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3758 ac->ac_b_ex.fe_len);
3759 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3760 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3762 err = -EFSCORRUPTED;
3766 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3767 #ifdef AGGRESSIVE_CHECK
3770 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3771 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3772 bitmap_bh->b_data));
3776 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3777 ac->ac_b_ex.fe_len);
3778 if (ext4_has_group_desc_csum(sb) &&
3779 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3780 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3781 ext4_free_group_clusters_set(sb, gdp,
3782 ext4_free_clusters_after_init(sb,
3783 ac->ac_b_ex.fe_group, gdp));
3785 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3786 ext4_free_group_clusters_set(sb, gdp, len);
3787 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3788 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3790 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3791 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3793 * Now reduce the dirty block count also. Should not go negative
3795 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3796 /* release all the reserved blocks if non delalloc */
3797 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3800 if (sbi->s_log_groups_per_flex) {
3801 ext4_group_t flex_group = ext4_flex_group(sbi,
3802 ac->ac_b_ex.fe_group);
3803 atomic64_sub(ac->ac_b_ex.fe_len,
3804 &sbi_array_rcu_deref(sbi, s_flex_groups,
3805 flex_group)->free_clusters);
3808 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3811 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3819 * Idempotent helper for Ext4 fast commit replay path to set the state of
3820 * blocks in bitmaps and update counters.
3822 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
3825 struct buffer_head *bitmap_bh = NULL;
3826 struct ext4_group_desc *gdp;
3827 struct buffer_head *gdp_bh;
3828 struct ext4_sb_info *sbi = EXT4_SB(sb);
3830 ext4_grpblk_t blkoff;
3834 clen = EXT4_B2C(sbi, len);
3836 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
3837 bitmap_bh = ext4_read_block_bitmap(sb, group);
3838 if (IS_ERR(bitmap_bh)) {
3839 err = PTR_ERR(bitmap_bh);
3845 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
3849 ext4_lock_group(sb, group);
3851 for (i = 0; i < clen; i++)
3852 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) == !state)
3856 ext4_set_bits(bitmap_bh->b_data, blkoff, clen);
3858 mb_test_and_clear_bits(bitmap_bh->b_data, blkoff, clen);
3859 if (ext4_has_group_desc_csum(sb) &&
3860 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3861 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3862 ext4_free_group_clusters_set(sb, gdp,
3863 ext4_free_clusters_after_init(sb,
3867 clen = ext4_free_group_clusters(sb, gdp) - clen + already;
3869 clen = ext4_free_group_clusters(sb, gdp) + clen - already;
3871 ext4_free_group_clusters_set(sb, gdp, clen);
3872 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
3873 ext4_group_desc_csum_set(sb, group, gdp);
3875 ext4_unlock_group(sb, group);
3877 if (sbi->s_log_groups_per_flex) {
3878 ext4_group_t flex_group = ext4_flex_group(sbi, group);
3881 &sbi_array_rcu_deref(sbi, s_flex_groups,
3882 flex_group)->free_clusters);
3885 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
3888 sync_dirty_buffer(bitmap_bh);
3889 err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
3890 sync_dirty_buffer(gdp_bh);
3897 * here we normalize request for locality group
3898 * Group request are normalized to s_mb_group_prealloc, which goes to
3899 * s_strip if we set the same via mount option.
3900 * s_mb_group_prealloc can be configured via
3901 * /sys/fs/ext4/<partition>/mb_group_prealloc
3903 * XXX: should we try to preallocate more than the group has now?
3905 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3907 struct super_block *sb = ac->ac_sb;
3908 struct ext4_locality_group *lg = ac->ac_lg;
3911 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3912 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
3916 * Normalization means making request better in terms of
3917 * size and alignment
3919 static noinline_for_stack void
3920 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3921 struct ext4_allocation_request *ar)
3923 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3926 loff_t size, start_off;
3927 loff_t orig_size __maybe_unused;
3929 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3930 struct ext4_prealloc_space *pa;
3932 /* do normalize only data requests, metadata requests
3933 do not need preallocation */
3934 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3937 /* sometime caller may want exact blocks */
3938 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3941 /* caller may indicate that preallocation isn't
3942 * required (it's a tail, for example) */
3943 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3946 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3947 ext4_mb_normalize_group_request(ac);
3951 bsbits = ac->ac_sb->s_blocksize_bits;
3953 /* first, let's learn actual file size
3954 * given current request is allocated */
3955 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3956 size = size << bsbits;
3957 if (size < i_size_read(ac->ac_inode))
3958 size = i_size_read(ac->ac_inode);
3961 /* max size of free chunks */
3964 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3965 (req <= (size) || max <= (chunk_size))
3967 /* first, try to predict filesize */
3968 /* XXX: should this table be tunable? */
3970 if (size <= 16 * 1024) {
3972 } else if (size <= 32 * 1024) {
3974 } else if (size <= 64 * 1024) {
3976 } else if (size <= 128 * 1024) {
3978 } else if (size <= 256 * 1024) {
3980 } else if (size <= 512 * 1024) {
3982 } else if (size <= 1024 * 1024) {
3984 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3985 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3986 (21 - bsbits)) << 21;
3987 size = 2 * 1024 * 1024;
3988 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3989 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3990 (22 - bsbits)) << 22;
3991 size = 4 * 1024 * 1024;
3992 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3993 (8<<20)>>bsbits, max, 8 * 1024)) {
3994 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3995 (23 - bsbits)) << 23;
3996 size = 8 * 1024 * 1024;
3998 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3999 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
4000 ac->ac_o_ex.fe_len) << bsbits;
4002 size = size >> bsbits;
4003 start = start_off >> bsbits;
4005 /* don't cover already allocated blocks in selected range */
4006 if (ar->pleft && start <= ar->lleft) {
4007 size -= ar->lleft + 1 - start;
4008 start = ar->lleft + 1;
4010 if (ar->pright && start + size - 1 >= ar->lright)
4011 size -= start + size - ar->lright;
4014 * Trim allocation request for filesystems with artificially small
4017 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4018 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4022 /* check we don't cross already preallocated blocks */
4024 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4029 spin_lock(&pa->pa_lock);
4030 if (pa->pa_deleted) {
4031 spin_unlock(&pa->pa_lock);
4035 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4038 /* PA must not overlap original request */
4039 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
4040 ac->ac_o_ex.fe_logical < pa->pa_lstart));
4042 /* skip PAs this normalized request doesn't overlap with */
4043 if (pa->pa_lstart >= end || pa_end <= start) {
4044 spin_unlock(&pa->pa_lock);
4047 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
4049 /* adjust start or end to be adjacent to this pa */
4050 if (pa_end <= ac->ac_o_ex.fe_logical) {
4051 BUG_ON(pa_end < start);
4053 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4054 BUG_ON(pa->pa_lstart > end);
4055 end = pa->pa_lstart;
4057 spin_unlock(&pa->pa_lock);
4062 /* XXX: extra loop to check we really don't overlap preallocations */
4064 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4067 spin_lock(&pa->pa_lock);
4068 if (pa->pa_deleted == 0) {
4069 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4071 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
4073 spin_unlock(&pa->pa_lock);
4077 if (start + size <= ac->ac_o_ex.fe_logical &&
4078 start > ac->ac_o_ex.fe_logical) {
4079 ext4_msg(ac->ac_sb, KERN_ERR,
4080 "start %lu, size %lu, fe_logical %lu",
4081 (unsigned long) start, (unsigned long) size,
4082 (unsigned long) ac->ac_o_ex.fe_logical);
4085 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4087 /* now prepare goal request */
4089 /* XXX: is it better to align blocks WRT to logical
4090 * placement or satisfy big request as is */
4091 ac->ac_g_ex.fe_logical = start;
4092 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4094 /* define goal start in order to merge */
4095 if (ar->pright && (ar->lright == (start + size))) {
4096 /* merge to the right */
4097 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4098 &ac->ac_f_ex.fe_group,
4099 &ac->ac_f_ex.fe_start);
4100 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4102 if (ar->pleft && (ar->lleft + 1 == start)) {
4103 /* merge to the left */
4104 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4105 &ac->ac_f_ex.fe_group,
4106 &ac->ac_f_ex.fe_start);
4107 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4110 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4114 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4116 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4118 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4119 atomic_inc(&sbi->s_bal_reqs);
4120 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4121 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4122 atomic_inc(&sbi->s_bal_success);
4123 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4124 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4125 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4126 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4127 atomic_inc(&sbi->s_bal_goals);
4128 if (ac->ac_found > sbi->s_mb_max_to_scan)
4129 atomic_inc(&sbi->s_bal_breaks);
4132 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4133 trace_ext4_mballoc_alloc(ac);
4135 trace_ext4_mballoc_prealloc(ac);
4139 * Called on failure; free up any blocks from the inode PA for this
4140 * context. We don't need this for MB_GROUP_PA because we only change
4141 * pa_free in ext4_mb_release_context(), but on failure, we've already
4142 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4144 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4146 struct ext4_prealloc_space *pa = ac->ac_pa;
4147 struct ext4_buddy e4b;
4151 if (ac->ac_f_ex.fe_len == 0)
4153 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4156 * This should never happen since we pin the
4157 * pages in the ext4_allocation_context so
4158 * ext4_mb_load_buddy() should never fail.
4160 WARN(1, "mb_load_buddy failed (%d)", err);
4163 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4164 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4165 ac->ac_f_ex.fe_len);
4166 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4167 ext4_mb_unload_buddy(&e4b);
4170 if (pa->pa_type == MB_INODE_PA)
4171 pa->pa_free += ac->ac_b_ex.fe_len;
4175 * use blocks preallocated to inode
4177 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4178 struct ext4_prealloc_space *pa)
4180 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4185 /* found preallocated blocks, use them */
4186 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4187 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4188 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4189 len = EXT4_NUM_B2C(sbi, end - start);
4190 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4191 &ac->ac_b_ex.fe_start);
4192 ac->ac_b_ex.fe_len = len;
4193 ac->ac_status = AC_STATUS_FOUND;
4196 BUG_ON(start < pa->pa_pstart);
4197 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4198 BUG_ON(pa->pa_free < len);
4201 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4205 * use blocks preallocated to locality group
4207 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4208 struct ext4_prealloc_space *pa)
4210 unsigned int len = ac->ac_o_ex.fe_len;
4212 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4213 &ac->ac_b_ex.fe_group,
4214 &ac->ac_b_ex.fe_start);
4215 ac->ac_b_ex.fe_len = len;
4216 ac->ac_status = AC_STATUS_FOUND;
4219 /* we don't correct pa_pstart or pa_plen here to avoid
4220 * possible race when the group is being loaded concurrently
4221 * instead we correct pa later, after blocks are marked
4222 * in on-disk bitmap -- see ext4_mb_release_context()
4223 * Other CPUs are prevented from allocating from this pa by lg_mutex
4225 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4226 pa->pa_lstart-len, len, pa);
4230 * Return the prealloc space that have minimal distance
4231 * from the goal block. @cpa is the prealloc
4232 * space that is having currently known minimal distance
4233 * from the goal block.
4235 static struct ext4_prealloc_space *
4236 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4237 struct ext4_prealloc_space *pa,
4238 struct ext4_prealloc_space *cpa)
4240 ext4_fsblk_t cur_distance, new_distance;
4243 atomic_inc(&pa->pa_count);
4246 cur_distance = abs(goal_block - cpa->pa_pstart);
4247 new_distance = abs(goal_block - pa->pa_pstart);
4249 if (cur_distance <= new_distance)
4252 /* drop the previous reference */
4253 atomic_dec(&cpa->pa_count);
4254 atomic_inc(&pa->pa_count);
4259 * search goal blocks in preallocated space
4261 static noinline_for_stack bool
4262 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4264 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4266 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4267 struct ext4_locality_group *lg;
4268 struct ext4_prealloc_space *pa, *cpa = NULL;
4269 ext4_fsblk_t goal_block;
4271 /* only data can be preallocated */
4272 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4275 /* first, try per-file preallocation */
4277 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4279 /* all fields in this condition don't change,
4280 * so we can skip locking for them */
4281 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
4282 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
4283 EXT4_C2B(sbi, pa->pa_len)))
4286 /* non-extent files can't have physical blocks past 2^32 */
4287 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4288 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
4289 EXT4_MAX_BLOCK_FILE_PHYS))
4292 /* found preallocated blocks, use them */
4293 spin_lock(&pa->pa_lock);
4294 if (pa->pa_deleted == 0 && pa->pa_free) {
4295 atomic_inc(&pa->pa_count);
4296 ext4_mb_use_inode_pa(ac, pa);
4297 spin_unlock(&pa->pa_lock);
4298 ac->ac_criteria = 10;
4302 spin_unlock(&pa->pa_lock);
4306 /* can we use group allocation? */
4307 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4310 /* inode may have no locality group for some reason */
4314 order = fls(ac->ac_o_ex.fe_len) - 1;
4315 if (order > PREALLOC_TB_SIZE - 1)
4316 /* The max size of hash table is PREALLOC_TB_SIZE */
4317 order = PREALLOC_TB_SIZE - 1;
4319 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4321 * search for the prealloc space that is having
4322 * minimal distance from the goal block.
4324 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4326 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
4328 spin_lock(&pa->pa_lock);
4329 if (pa->pa_deleted == 0 &&
4330 pa->pa_free >= ac->ac_o_ex.fe_len) {
4332 cpa = ext4_mb_check_group_pa(goal_block,
4335 spin_unlock(&pa->pa_lock);
4340 ext4_mb_use_group_pa(ac, cpa);
4341 ac->ac_criteria = 20;
4348 * the function goes through all block freed in the group
4349 * but not yet committed and marks them used in in-core bitmap.
4350 * buddy must be generated from this bitmap
4351 * Need to be called with the ext4 group lock held
4353 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
4357 struct ext4_group_info *grp;
4358 struct ext4_free_data *entry;
4360 grp = ext4_get_group_info(sb, group);
4361 n = rb_first(&(grp->bb_free_root));
4364 entry = rb_entry(n, struct ext4_free_data, efd_node);
4365 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
4372 * the function goes through all preallocation in this group and marks them
4373 * used in in-core bitmap. buddy must be generated from this bitmap
4374 * Need to be called with ext4 group lock held
4376 static noinline_for_stack
4377 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4380 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4381 struct ext4_prealloc_space *pa;
4382 struct list_head *cur;
4383 ext4_group_t groupnr;
4384 ext4_grpblk_t start;
4385 int preallocated = 0;
4388 /* all form of preallocation discards first load group,
4389 * so the only competing code is preallocation use.
4390 * we don't need any locking here
4391 * notice we do NOT ignore preallocations with pa_deleted
4392 * otherwise we could leave used blocks available for
4393 * allocation in buddy when concurrent ext4_mb_put_pa()
4394 * is dropping preallocation
4396 list_for_each(cur, &grp->bb_prealloc_list) {
4397 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
4398 spin_lock(&pa->pa_lock);
4399 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4402 spin_unlock(&pa->pa_lock);
4403 if (unlikely(len == 0))
4405 BUG_ON(groupnr != group);
4406 ext4_set_bits(bitmap, start, len);
4407 preallocated += len;
4409 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
4412 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
4413 struct ext4_prealloc_space *pa)
4415 struct ext4_inode_info *ei;
4417 if (pa->pa_deleted) {
4418 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
4419 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
4426 if (pa->pa_type == MB_INODE_PA) {
4427 ei = EXT4_I(pa->pa_inode);
4428 atomic_dec(&ei->i_prealloc_active);
4432 static void ext4_mb_pa_callback(struct rcu_head *head)
4434 struct ext4_prealloc_space *pa;
4435 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
4437 BUG_ON(atomic_read(&pa->pa_count));
4438 BUG_ON(pa->pa_deleted == 0);
4439 kmem_cache_free(ext4_pspace_cachep, pa);
4443 * drops a reference to preallocated space descriptor
4444 * if this was the last reference and the space is consumed
4446 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
4447 struct super_block *sb, struct ext4_prealloc_space *pa)
4450 ext4_fsblk_t grp_blk;
4452 /* in this short window concurrent discard can set pa_deleted */
4453 spin_lock(&pa->pa_lock);
4454 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
4455 spin_unlock(&pa->pa_lock);
4459 if (pa->pa_deleted == 1) {
4460 spin_unlock(&pa->pa_lock);
4464 ext4_mb_mark_pa_deleted(sb, pa);
4465 spin_unlock(&pa->pa_lock);
4467 grp_blk = pa->pa_pstart;
4469 * If doing group-based preallocation, pa_pstart may be in the
4470 * next group when pa is used up
4472 if (pa->pa_type == MB_GROUP_PA)
4475 grp = ext4_get_group_number(sb, grp_blk);
4480 * P1 (buddy init) P2 (regular allocation)
4481 * find block B in PA
4482 * copy on-disk bitmap to buddy
4483 * mark B in on-disk bitmap
4484 * drop PA from group
4485 * mark all PAs in buddy
4487 * thus, P1 initializes buddy with B available. to prevent this
4488 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
4491 ext4_lock_group(sb, grp);
4492 list_del(&pa->pa_group_list);
4493 ext4_unlock_group(sb, grp);
4495 spin_lock(pa->pa_obj_lock);
4496 list_del_rcu(&pa->pa_inode_list);
4497 spin_unlock(pa->pa_obj_lock);
4499 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4503 * creates new preallocated space for given inode
4505 static noinline_for_stack void
4506 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
4508 struct super_block *sb = ac->ac_sb;
4509 struct ext4_sb_info *sbi = EXT4_SB(sb);
4510 struct ext4_prealloc_space *pa;
4511 struct ext4_group_info *grp;
4512 struct ext4_inode_info *ei;
4514 /* preallocate only when found space is larger then requested */
4515 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4516 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4517 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4518 BUG_ON(ac->ac_pa == NULL);
4522 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
4528 /* we can't allocate as much as normalizer wants.
4529 * so, found space must get proper lstart
4530 * to cover original request */
4531 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
4532 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
4534 /* we're limited by original request in that
4535 * logical block must be covered any way
4536 * winl is window we can move our chunk within */
4537 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
4539 /* also, we should cover whole original request */
4540 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
4542 /* the smallest one defines real window */
4543 win = min(winl, wins);
4545 offs = ac->ac_o_ex.fe_logical %
4546 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4547 if (offs && offs < win)
4550 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
4551 EXT4_NUM_B2C(sbi, win);
4552 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
4553 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
4556 /* preallocation can change ac_b_ex, thus we store actually
4557 * allocated blocks for history */
4558 ac->ac_f_ex = ac->ac_b_ex;
4560 pa->pa_lstart = ac->ac_b_ex.fe_logical;
4561 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4562 pa->pa_len = ac->ac_b_ex.fe_len;
4563 pa->pa_free = pa->pa_len;
4564 spin_lock_init(&pa->pa_lock);
4565 INIT_LIST_HEAD(&pa->pa_inode_list);
4566 INIT_LIST_HEAD(&pa->pa_group_list);
4568 pa->pa_type = MB_INODE_PA;
4570 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4571 pa->pa_len, pa->pa_lstart);
4572 trace_ext4_mb_new_inode_pa(ac, pa);
4574 ext4_mb_use_inode_pa(ac, pa);
4575 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
4577 ei = EXT4_I(ac->ac_inode);
4578 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4580 pa->pa_obj_lock = &ei->i_prealloc_lock;
4581 pa->pa_inode = ac->ac_inode;
4583 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4585 spin_lock(pa->pa_obj_lock);
4586 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
4587 spin_unlock(pa->pa_obj_lock);
4588 atomic_inc(&ei->i_prealloc_active);
4592 * creates new preallocated space for locality group inodes belongs to
4594 static noinline_for_stack void
4595 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
4597 struct super_block *sb = ac->ac_sb;
4598 struct ext4_locality_group *lg;
4599 struct ext4_prealloc_space *pa;
4600 struct ext4_group_info *grp;
4602 /* preallocate only when found space is larger then requested */
4603 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4604 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4605 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4606 BUG_ON(ac->ac_pa == NULL);
4610 /* preallocation can change ac_b_ex, thus we store actually
4611 * allocated blocks for history */
4612 ac->ac_f_ex = ac->ac_b_ex;
4614 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4615 pa->pa_lstart = pa->pa_pstart;
4616 pa->pa_len = ac->ac_b_ex.fe_len;
4617 pa->pa_free = pa->pa_len;
4618 spin_lock_init(&pa->pa_lock);
4619 INIT_LIST_HEAD(&pa->pa_inode_list);
4620 INIT_LIST_HEAD(&pa->pa_group_list);
4622 pa->pa_type = MB_GROUP_PA;
4624 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4625 pa->pa_len, pa->pa_lstart);
4626 trace_ext4_mb_new_group_pa(ac, pa);
4628 ext4_mb_use_group_pa(ac, pa);
4629 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4631 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4635 pa->pa_obj_lock = &lg->lg_prealloc_lock;
4636 pa->pa_inode = NULL;
4638 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4641 * We will later add the new pa to the right bucket
4642 * after updating the pa_free in ext4_mb_release_context
4646 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4648 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4649 ext4_mb_new_group_pa(ac);
4651 ext4_mb_new_inode_pa(ac);
4655 * finds all unused blocks in on-disk bitmap, frees them in
4656 * in-core bitmap and buddy.
4657 * @pa must be unlinked from inode and group lists, so that
4658 * nobody else can find/use it.
4659 * the caller MUST hold group/inode locks.
4660 * TODO: optimize the case when there are no in-core structures yet
4662 static noinline_for_stack int
4663 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4664 struct ext4_prealloc_space *pa)
4666 struct super_block *sb = e4b->bd_sb;
4667 struct ext4_sb_info *sbi = EXT4_SB(sb);
4672 unsigned long long grp_blk_start;
4675 BUG_ON(pa->pa_deleted == 0);
4676 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4677 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4678 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4679 end = bit + pa->pa_len;
4682 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4685 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4686 mb_debug(sb, "free preallocated %u/%u in group %u\n",
4687 (unsigned) ext4_group_first_block_no(sb, group) + bit,
4688 (unsigned) next - bit, (unsigned) group);
4691 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4692 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4693 EXT4_C2B(sbi, bit)),
4695 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4698 if (free != pa->pa_free) {
4699 ext4_msg(e4b->bd_sb, KERN_CRIT,
4700 "pa %p: logic %lu, phys. %lu, len %d",
4701 pa, (unsigned long) pa->pa_lstart,
4702 (unsigned long) pa->pa_pstart,
4704 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4707 * pa is already deleted so we use the value obtained
4708 * from the bitmap and continue.
4711 atomic_add(free, &sbi->s_mb_discarded);
4716 static noinline_for_stack int
4717 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4718 struct ext4_prealloc_space *pa)
4720 struct super_block *sb = e4b->bd_sb;
4724 trace_ext4_mb_release_group_pa(sb, pa);
4725 BUG_ON(pa->pa_deleted == 0);
4726 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4727 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4728 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4729 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4730 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4736 * releases all preallocations in given group
4738 * first, we need to decide discard policy:
4739 * - when do we discard
4741 * - how many do we discard
4742 * 1) how many requested
4744 static noinline_for_stack int
4745 ext4_mb_discard_group_preallocations(struct super_block *sb,
4746 ext4_group_t group, int needed)
4748 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4749 struct buffer_head *bitmap_bh = NULL;
4750 struct ext4_prealloc_space *pa, *tmp;
4751 struct list_head list;
4752 struct ext4_buddy e4b;
4755 int free, free_total = 0;
4757 mb_debug(sb, "discard preallocation for group %u\n", group);
4758 if (list_empty(&grp->bb_prealloc_list))
4761 bitmap_bh = ext4_read_block_bitmap(sb, group);
4762 if (IS_ERR(bitmap_bh)) {
4763 err = PTR_ERR(bitmap_bh);
4764 ext4_error_err(sb, -err,
4765 "Error %d reading block bitmap for %u",
4770 err = ext4_mb_load_buddy(sb, group, &e4b);
4772 ext4_warning(sb, "Error %d loading buddy information for %u",
4779 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
4781 INIT_LIST_HEAD(&list);
4784 ext4_lock_group(sb, group);
4785 list_for_each_entry_safe(pa, tmp,
4786 &grp->bb_prealloc_list, pa_group_list) {
4787 spin_lock(&pa->pa_lock);
4788 if (atomic_read(&pa->pa_count)) {
4789 spin_unlock(&pa->pa_lock);
4793 if (pa->pa_deleted) {
4794 spin_unlock(&pa->pa_lock);
4798 /* seems this one can be freed ... */
4799 ext4_mb_mark_pa_deleted(sb, pa);
4802 this_cpu_inc(discard_pa_seq);
4804 /* we can trust pa_free ... */
4805 free += pa->pa_free;
4807 spin_unlock(&pa->pa_lock);
4809 list_del(&pa->pa_group_list);
4810 list_add(&pa->u.pa_tmp_list, &list);
4813 /* now free all selected PAs */
4814 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4816 /* remove from object (inode or locality group) */
4817 spin_lock(pa->pa_obj_lock);
4818 list_del_rcu(&pa->pa_inode_list);
4819 spin_unlock(pa->pa_obj_lock);
4821 if (pa->pa_type == MB_GROUP_PA)
4822 ext4_mb_release_group_pa(&e4b, pa);
4824 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4826 list_del(&pa->u.pa_tmp_list);
4827 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4832 /* if we still need more blocks and some PAs were used, try again */
4833 if (free_total < needed && busy) {
4834 ext4_unlock_group(sb, group);
4839 ext4_unlock_group(sb, group);
4840 ext4_mb_unload_buddy(&e4b);
4843 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4844 free_total, group, grp->bb_free);
4849 * releases all non-used preallocated blocks for given inode
4851 * It's important to discard preallocations under i_data_sem
4852 * We don't want another block to be served from the prealloc
4853 * space when we are discarding the inode prealloc space.
4855 * FIXME!! Make sure it is valid at all the call sites
4857 void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
4859 struct ext4_inode_info *ei = EXT4_I(inode);
4860 struct super_block *sb = inode->i_sb;
4861 struct buffer_head *bitmap_bh = NULL;
4862 struct ext4_prealloc_space *pa, *tmp;
4863 ext4_group_t group = 0;
4864 struct list_head list;
4865 struct ext4_buddy e4b;
4868 if (!S_ISREG(inode->i_mode)) {
4869 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4873 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
4876 mb_debug(sb, "discard preallocation for inode %lu\n",
4878 trace_ext4_discard_preallocations(inode,
4879 atomic_read(&ei->i_prealloc_active), needed);
4881 INIT_LIST_HEAD(&list);
4887 /* first, collect all pa's in the inode */
4888 spin_lock(&ei->i_prealloc_lock);
4889 while (!list_empty(&ei->i_prealloc_list) && needed) {
4890 pa = list_entry(ei->i_prealloc_list.prev,
4891 struct ext4_prealloc_space, pa_inode_list);
4892 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4893 spin_lock(&pa->pa_lock);
4894 if (atomic_read(&pa->pa_count)) {
4895 /* this shouldn't happen often - nobody should
4896 * use preallocation while we're discarding it */
4897 spin_unlock(&pa->pa_lock);
4898 spin_unlock(&ei->i_prealloc_lock);
4899 ext4_msg(sb, KERN_ERR,
4900 "uh-oh! used pa while discarding");
4902 schedule_timeout_uninterruptible(HZ);
4906 if (pa->pa_deleted == 0) {
4907 ext4_mb_mark_pa_deleted(sb, pa);
4908 spin_unlock(&pa->pa_lock);
4909 list_del_rcu(&pa->pa_inode_list);
4910 list_add(&pa->u.pa_tmp_list, &list);
4915 /* someone is deleting pa right now */
4916 spin_unlock(&pa->pa_lock);
4917 spin_unlock(&ei->i_prealloc_lock);
4919 /* we have to wait here because pa_deleted
4920 * doesn't mean pa is already unlinked from
4921 * the list. as we might be called from
4922 * ->clear_inode() the inode will get freed
4923 * and concurrent thread which is unlinking
4924 * pa from inode's list may access already
4925 * freed memory, bad-bad-bad */
4927 /* XXX: if this happens too often, we can
4928 * add a flag to force wait only in case
4929 * of ->clear_inode(), but not in case of
4930 * regular truncate */
4931 schedule_timeout_uninterruptible(HZ);
4934 spin_unlock(&ei->i_prealloc_lock);
4936 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4937 BUG_ON(pa->pa_type != MB_INODE_PA);
4938 group = ext4_get_group_number(sb, pa->pa_pstart);
4940 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4941 GFP_NOFS|__GFP_NOFAIL);
4943 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
4948 bitmap_bh = ext4_read_block_bitmap(sb, group);
4949 if (IS_ERR(bitmap_bh)) {
4950 err = PTR_ERR(bitmap_bh);
4951 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
4953 ext4_mb_unload_buddy(&e4b);
4957 ext4_lock_group(sb, group);
4958 list_del(&pa->pa_group_list);
4959 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4960 ext4_unlock_group(sb, group);
4962 ext4_mb_unload_buddy(&e4b);
4965 list_del(&pa->u.pa_tmp_list);
4966 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4970 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
4972 struct ext4_prealloc_space *pa;
4974 BUG_ON(ext4_pspace_cachep == NULL);
4975 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
4978 atomic_set(&pa->pa_count, 1);
4983 static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
4985 struct ext4_prealloc_space *pa = ac->ac_pa;
4989 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
4990 kmem_cache_free(ext4_pspace_cachep, pa);
4993 #ifdef CONFIG_EXT4_DEBUG
4994 static inline void ext4_mb_show_pa(struct super_block *sb)
4996 ext4_group_t i, ngroups;
4998 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5001 ngroups = ext4_get_groups_count(sb);
5002 mb_debug(sb, "groups: ");
5003 for (i = 0; i < ngroups; i++) {
5004 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5005 struct ext4_prealloc_space *pa;
5006 ext4_grpblk_t start;
5007 struct list_head *cur;
5008 ext4_lock_group(sb, i);
5009 list_for_each(cur, &grp->bb_prealloc_list) {
5010 pa = list_entry(cur, struct ext4_prealloc_space,
5012 spin_lock(&pa->pa_lock);
5013 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5015 spin_unlock(&pa->pa_lock);
5016 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5019 ext4_unlock_group(sb, i);
5020 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5025 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5027 struct super_block *sb = ac->ac_sb;
5029 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5032 mb_debug(sb, "Can't allocate:"
5033 " Allocation context details:");
5034 mb_debug(sb, "status %u flags 0x%x",
5035 ac->ac_status, ac->ac_flags);
5036 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5037 "goal %lu/%lu/%lu@%lu, "
5038 "best %lu/%lu/%lu@%lu cr %d",
5039 (unsigned long)ac->ac_o_ex.fe_group,
5040 (unsigned long)ac->ac_o_ex.fe_start,
5041 (unsigned long)ac->ac_o_ex.fe_len,
5042 (unsigned long)ac->ac_o_ex.fe_logical,
5043 (unsigned long)ac->ac_g_ex.fe_group,
5044 (unsigned long)ac->ac_g_ex.fe_start,
5045 (unsigned long)ac->ac_g_ex.fe_len,
5046 (unsigned long)ac->ac_g_ex.fe_logical,
5047 (unsigned long)ac->ac_b_ex.fe_group,
5048 (unsigned long)ac->ac_b_ex.fe_start,
5049 (unsigned long)ac->ac_b_ex.fe_len,
5050 (unsigned long)ac->ac_b_ex.fe_logical,
5051 (int)ac->ac_criteria);
5052 mb_debug(sb, "%u found", ac->ac_found);
5053 ext4_mb_show_pa(sb);
5056 static inline void ext4_mb_show_pa(struct super_block *sb)
5060 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5062 ext4_mb_show_pa(ac->ac_sb);
5068 * We use locality group preallocation for small size file. The size of the
5069 * file is determined by the current size or the resulting size after
5070 * allocation which ever is larger
5072 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5074 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5076 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5077 int bsbits = ac->ac_sb->s_blocksize_bits;
5080 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5083 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5086 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
5087 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5090 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5091 !inode_is_open_for_write(ac->ac_inode)) {
5092 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5096 if (sbi->s_mb_group_prealloc <= 0) {
5097 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5101 /* don't use group allocation for large files */
5102 size = max(size, isize);
5103 if (size > sbi->s_mb_stream_request) {
5104 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5108 BUG_ON(ac->ac_lg != NULL);
5110 * locality group prealloc space are per cpu. The reason for having
5111 * per cpu locality group is to reduce the contention between block
5112 * request from multiple CPUs.
5114 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5116 /* we're going to use group allocation */
5117 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5119 /* serialize all allocations in the group */
5120 mutex_lock(&ac->ac_lg->lg_mutex);
5123 static noinline_for_stack int
5124 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5125 struct ext4_allocation_request *ar)
5127 struct super_block *sb = ar->inode->i_sb;
5128 struct ext4_sb_info *sbi = EXT4_SB(sb);
5129 struct ext4_super_block *es = sbi->s_es;
5133 ext4_grpblk_t block;
5135 /* we can't allocate > group size */
5138 /* just a dirty hack to filter too big requests */
5139 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5140 len = EXT4_CLUSTERS_PER_GROUP(sb);
5142 /* start searching from the goal */
5144 if (goal < le32_to_cpu(es->s_first_data_block) ||
5145 goal >= ext4_blocks_count(es))
5146 goal = le32_to_cpu(es->s_first_data_block);
5147 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5149 /* set up allocation goals */
5150 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5151 ac->ac_status = AC_STATUS_CONTINUE;
5153 ac->ac_inode = ar->inode;
5154 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5155 ac->ac_o_ex.fe_group = group;
5156 ac->ac_o_ex.fe_start = block;
5157 ac->ac_o_ex.fe_len = len;
5158 ac->ac_g_ex = ac->ac_o_ex;
5159 ac->ac_flags = ar->flags;
5161 /* we have to define context: we'll work with a file or
5162 * locality group. this is a policy, actually */
5163 ext4_mb_group_or_file(ac);
5165 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5166 "left: %u/%u, right %u/%u to %swritable\n",
5167 (unsigned) ar->len, (unsigned) ar->logical,
5168 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5169 (unsigned) ar->lleft, (unsigned) ar->pleft,
5170 (unsigned) ar->lright, (unsigned) ar->pright,
5171 inode_is_open_for_write(ar->inode) ? "" : "non-");
5176 static noinline_for_stack void
5177 ext4_mb_discard_lg_preallocations(struct super_block *sb,
5178 struct ext4_locality_group *lg,
5179 int order, int total_entries)
5181 ext4_group_t group = 0;
5182 struct ext4_buddy e4b;
5183 struct list_head discard_list;
5184 struct ext4_prealloc_space *pa, *tmp;
5186 mb_debug(sb, "discard locality group preallocation\n");
5188 INIT_LIST_HEAD(&discard_list);
5190 spin_lock(&lg->lg_prealloc_lock);
5191 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5193 lockdep_is_held(&lg->lg_prealloc_lock)) {
5194 spin_lock(&pa->pa_lock);
5195 if (atomic_read(&pa->pa_count)) {
5197 * This is the pa that we just used
5198 * for block allocation. So don't
5201 spin_unlock(&pa->pa_lock);
5204 if (pa->pa_deleted) {
5205 spin_unlock(&pa->pa_lock);
5208 /* only lg prealloc space */
5209 BUG_ON(pa->pa_type != MB_GROUP_PA);
5211 /* seems this one can be freed ... */
5212 ext4_mb_mark_pa_deleted(sb, pa);
5213 spin_unlock(&pa->pa_lock);
5215 list_del_rcu(&pa->pa_inode_list);
5216 list_add(&pa->u.pa_tmp_list, &discard_list);
5219 if (total_entries <= 5) {
5221 * we want to keep only 5 entries
5222 * allowing it to grow to 8. This
5223 * mak sure we don't call discard
5224 * soon for this list.
5229 spin_unlock(&lg->lg_prealloc_lock);
5231 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5234 group = ext4_get_group_number(sb, pa->pa_pstart);
5235 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5236 GFP_NOFS|__GFP_NOFAIL);
5238 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5242 ext4_lock_group(sb, group);
5243 list_del(&pa->pa_group_list);
5244 ext4_mb_release_group_pa(&e4b, pa);
5245 ext4_unlock_group(sb, group);
5247 ext4_mb_unload_buddy(&e4b);
5248 list_del(&pa->u.pa_tmp_list);
5249 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5254 * We have incremented pa_count. So it cannot be freed at this
5255 * point. Also we hold lg_mutex. So no parallel allocation is
5256 * possible from this lg. That means pa_free cannot be updated.
5258 * A parallel ext4_mb_discard_group_preallocations is possible.
5259 * which can cause the lg_prealloc_list to be updated.
5262 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5264 int order, added = 0, lg_prealloc_count = 1;
5265 struct super_block *sb = ac->ac_sb;
5266 struct ext4_locality_group *lg = ac->ac_lg;
5267 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5269 order = fls(pa->pa_free) - 1;
5270 if (order > PREALLOC_TB_SIZE - 1)
5271 /* The max size of hash table is PREALLOC_TB_SIZE */
5272 order = PREALLOC_TB_SIZE - 1;
5273 /* Add the prealloc space to lg */
5274 spin_lock(&lg->lg_prealloc_lock);
5275 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5277 lockdep_is_held(&lg->lg_prealloc_lock)) {
5278 spin_lock(&tmp_pa->pa_lock);
5279 if (tmp_pa->pa_deleted) {
5280 spin_unlock(&tmp_pa->pa_lock);
5283 if (!added && pa->pa_free < tmp_pa->pa_free) {
5284 /* Add to the tail of the previous entry */
5285 list_add_tail_rcu(&pa->pa_inode_list,
5286 &tmp_pa->pa_inode_list);
5289 * we want to count the total
5290 * number of entries in the list
5293 spin_unlock(&tmp_pa->pa_lock);
5294 lg_prealloc_count++;
5297 list_add_tail_rcu(&pa->pa_inode_list,
5298 &lg->lg_prealloc_list[order]);
5299 spin_unlock(&lg->lg_prealloc_lock);
5301 /* Now trim the list to be not more than 8 elements */
5302 if (lg_prealloc_count > 8) {
5303 ext4_mb_discard_lg_preallocations(sb, lg,
5304 order, lg_prealloc_count);
5311 * if per-inode prealloc list is too long, trim some PA
5313 static void ext4_mb_trim_inode_pa(struct inode *inode)
5315 struct ext4_inode_info *ei = EXT4_I(inode);
5316 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5319 count = atomic_read(&ei->i_prealloc_active);
5320 delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
5321 if (count > sbi->s_mb_max_inode_prealloc + delta) {
5322 count -= sbi->s_mb_max_inode_prealloc;
5323 ext4_discard_preallocations(inode, count);
5328 * release all resource we used in allocation
5330 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
5332 struct inode *inode = ac->ac_inode;
5333 struct ext4_inode_info *ei = EXT4_I(inode);
5334 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5335 struct ext4_prealloc_space *pa = ac->ac_pa;
5337 if (pa->pa_type == MB_GROUP_PA) {
5338 /* see comment in ext4_mb_use_group_pa() */
5339 spin_lock(&pa->pa_lock);
5340 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5341 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5342 pa->pa_free -= ac->ac_b_ex.fe_len;
5343 pa->pa_len -= ac->ac_b_ex.fe_len;
5344 spin_unlock(&pa->pa_lock);
5347 * We want to add the pa to the right bucket.
5348 * Remove it from the list and while adding
5349 * make sure the list to which we are adding
5352 if (likely(pa->pa_free)) {
5353 spin_lock(pa->pa_obj_lock);
5354 list_del_rcu(&pa->pa_inode_list);
5355 spin_unlock(pa->pa_obj_lock);
5356 ext4_mb_add_n_trim(ac);
5360 if (pa->pa_type == MB_INODE_PA) {
5362 * treat per-inode prealloc list as a lru list, then try
5363 * to trim the least recently used PA.
5365 spin_lock(pa->pa_obj_lock);
5366 list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
5367 spin_unlock(pa->pa_obj_lock);
5370 ext4_mb_put_pa(ac, ac->ac_sb, pa);
5372 if (ac->ac_bitmap_page)
5373 put_page(ac->ac_bitmap_page);
5374 if (ac->ac_buddy_page)
5375 put_page(ac->ac_buddy_page);
5376 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5377 mutex_unlock(&ac->ac_lg->lg_mutex);
5378 ext4_mb_collect_stats(ac);
5379 ext4_mb_trim_inode_pa(inode);
5383 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
5385 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
5389 trace_ext4_mb_discard_preallocations(sb, needed);
5390 for (i = 0; i < ngroups && needed > 0; i++) {
5391 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
5399 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
5400 struct ext4_allocation_context *ac, u64 *seq)
5406 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
5411 seq_retry = ext4_get_discard_pa_seq_sum();
5412 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
5413 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
5419 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
5423 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5424 struct ext4_allocation_request *ar, int *errp);
5427 * Main entry point into mballoc to allocate blocks
5428 * it tries to use preallocation first, then falls back
5429 * to usual allocation
5431 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
5432 struct ext4_allocation_request *ar, int *errp)
5434 struct ext4_allocation_context *ac = NULL;
5435 struct ext4_sb_info *sbi;
5436 struct super_block *sb;
5437 ext4_fsblk_t block = 0;
5438 unsigned int inquota = 0;
5439 unsigned int reserv_clstrs = 0;
5443 sb = ar->inode->i_sb;
5446 trace_ext4_request_blocks(ar);
5447 if (sbi->s_mount_state & EXT4_FC_REPLAY)
5448 return ext4_mb_new_blocks_simple(handle, ar, errp);
5450 /* Allow to use superuser reservation for quota file */
5451 if (ext4_is_quota_file(ar->inode))
5452 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
5454 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
5455 /* Without delayed allocation we need to verify
5456 * there is enough free blocks to do block allocation
5457 * and verify allocation doesn't exceed the quota limits.
5460 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
5462 /* let others to free the space */
5464 ar->len = ar->len >> 1;
5467 ext4_mb_show_pa(sb);
5471 reserv_clstrs = ar->len;
5472 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
5473 dquot_alloc_block_nofail(ar->inode,
5474 EXT4_C2B(sbi, ar->len));
5477 dquot_alloc_block(ar->inode,
5478 EXT4_C2B(sbi, ar->len))) {
5480 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
5491 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
5498 *errp = ext4_mb_initialize_context(ac, ar);
5504 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
5505 seq = this_cpu_read(discard_pa_seq);
5506 if (!ext4_mb_use_preallocated(ac)) {
5507 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
5508 ext4_mb_normalize_request(ac, ar);
5510 *errp = ext4_mb_pa_alloc(ac);
5514 /* allocate space in core */
5515 *errp = ext4_mb_regular_allocator(ac);
5517 * pa allocated above is added to grp->bb_prealloc_list only
5518 * when we were able to allocate some block i.e. when
5519 * ac->ac_status == AC_STATUS_FOUND.
5520 * And error from above mean ac->ac_status != AC_STATUS_FOUND
5521 * So we have to free this pa here itself.
5524 ext4_mb_pa_free(ac);
5525 ext4_discard_allocated_blocks(ac);
5528 if (ac->ac_status == AC_STATUS_FOUND &&
5529 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
5530 ext4_mb_pa_free(ac);
5532 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
5533 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
5535 ext4_discard_allocated_blocks(ac);
5538 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5539 ar->len = ac->ac_b_ex.fe_len;
5542 if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
5545 * If block allocation fails then the pa allocated above
5546 * needs to be freed here itself.
5548 ext4_mb_pa_free(ac);
5554 ac->ac_b_ex.fe_len = 0;
5556 ext4_mb_show_ac(ac);
5558 ext4_mb_release_context(ac);
5561 kmem_cache_free(ext4_ac_cachep, ac);
5562 if (inquota && ar->len < inquota)
5563 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
5565 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
5566 /* release all the reserved blocks if non delalloc */
5567 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
5571 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
5577 * We can merge two free data extents only if the physical blocks
5578 * are contiguous, AND the extents were freed by the same transaction,
5579 * AND the blocks are associated with the same group.
5581 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
5582 struct ext4_free_data *entry,
5583 struct ext4_free_data *new_entry,
5584 struct rb_root *entry_rb_root)
5586 if ((entry->efd_tid != new_entry->efd_tid) ||
5587 (entry->efd_group != new_entry->efd_group))
5589 if (entry->efd_start_cluster + entry->efd_count ==
5590 new_entry->efd_start_cluster) {
5591 new_entry->efd_start_cluster = entry->efd_start_cluster;
5592 new_entry->efd_count += entry->efd_count;
5593 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
5594 entry->efd_start_cluster) {
5595 new_entry->efd_count += entry->efd_count;
5598 spin_lock(&sbi->s_md_lock);
5599 list_del(&entry->efd_list);
5600 spin_unlock(&sbi->s_md_lock);
5601 rb_erase(&entry->efd_node, entry_rb_root);
5602 kmem_cache_free(ext4_free_data_cachep, entry);
5605 static noinline_for_stack int
5606 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
5607 struct ext4_free_data *new_entry)
5609 ext4_group_t group = e4b->bd_group;
5610 ext4_grpblk_t cluster;
5611 ext4_grpblk_t clusters = new_entry->efd_count;
5612 struct ext4_free_data *entry;
5613 struct ext4_group_info *db = e4b->bd_info;
5614 struct super_block *sb = e4b->bd_sb;
5615 struct ext4_sb_info *sbi = EXT4_SB(sb);
5616 struct rb_node **n = &db->bb_free_root.rb_node, *node;
5617 struct rb_node *parent = NULL, *new_node;
5619 BUG_ON(!ext4_handle_valid(handle));
5620 BUG_ON(e4b->bd_bitmap_page == NULL);
5621 BUG_ON(e4b->bd_buddy_page == NULL);
5623 new_node = &new_entry->efd_node;
5624 cluster = new_entry->efd_start_cluster;
5627 /* first free block exent. We need to
5628 protect buddy cache from being freed,
5629 * otherwise we'll refresh it from
5630 * on-disk bitmap and lose not-yet-available
5632 get_page(e4b->bd_buddy_page);
5633 get_page(e4b->bd_bitmap_page);
5637 entry = rb_entry(parent, struct ext4_free_data, efd_node);
5638 if (cluster < entry->efd_start_cluster)
5640 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
5641 n = &(*n)->rb_right;
5643 ext4_grp_locked_error(sb, group, 0,
5644 ext4_group_first_block_no(sb, group) +
5645 EXT4_C2B(sbi, cluster),
5646 "Block already on to-be-freed list");
5647 kmem_cache_free(ext4_free_data_cachep, new_entry);
5652 rb_link_node(new_node, parent, n);
5653 rb_insert_color(new_node, &db->bb_free_root);
5655 /* Now try to see the extent can be merged to left and right */
5656 node = rb_prev(new_node);
5658 entry = rb_entry(node, struct ext4_free_data, efd_node);
5659 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5660 &(db->bb_free_root));
5663 node = rb_next(new_node);
5665 entry = rb_entry(node, struct ext4_free_data, efd_node);
5666 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5667 &(db->bb_free_root));
5670 spin_lock(&sbi->s_md_lock);
5671 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5672 sbi->s_mb_free_pending += clusters;
5673 spin_unlock(&sbi->s_md_lock);
5678 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
5679 * linearly starting at the goal block and also excludes the blocks which
5680 * are going to be in use after fast commit replay.
5682 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5683 struct ext4_allocation_request *ar, int *errp)
5685 struct buffer_head *bitmap_bh;
5686 struct super_block *sb = ar->inode->i_sb;
5688 ext4_grpblk_t blkoff;
5689 int i = sb->s_blocksize;
5690 ext4_fsblk_t goal, block;
5691 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
5694 if (goal < le32_to_cpu(es->s_first_data_block) ||
5695 goal >= ext4_blocks_count(es))
5696 goal = le32_to_cpu(es->s_first_data_block);
5699 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
5700 for (; group < ext4_get_groups_count(sb); group++) {
5701 bitmap_bh = ext4_read_block_bitmap(sb, group);
5702 if (IS_ERR(bitmap_bh)) {
5703 *errp = PTR_ERR(bitmap_bh);
5704 pr_warn("Failed to read block bitmap\n");
5708 ext4_get_group_no_and_offset(sb,
5709 max(ext4_group_first_block_no(sb, group), goal),
5711 i = mb_find_next_zero_bit(bitmap_bh->b_data, sb->s_blocksize,
5714 if (i >= sb->s_blocksize)
5716 if (ext4_fc_replay_check_excluded(sb,
5717 ext4_group_first_block_no(sb, group) + i))
5722 if (group >= ext4_get_groups_count(sb) && i >= sb->s_blocksize)
5725 block = ext4_group_first_block_no(sb, group) + i;
5726 ext4_mb_mark_bb(sb, block, 1, 1);
5732 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
5733 unsigned long count)
5735 struct buffer_head *bitmap_bh;
5736 struct super_block *sb = inode->i_sb;
5737 struct ext4_group_desc *gdp;
5738 struct buffer_head *gdp_bh;
5740 ext4_grpblk_t blkoff;
5741 int already_freed = 0, err, i;
5743 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
5744 bitmap_bh = ext4_read_block_bitmap(sb, group);
5745 if (IS_ERR(bitmap_bh)) {
5746 err = PTR_ERR(bitmap_bh);
5747 pr_warn("Failed to read block bitmap\n");
5750 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
5754 for (i = 0; i < count; i++) {
5755 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data))
5758 mb_clear_bits(bitmap_bh->b_data, blkoff, count);
5759 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
5762 ext4_free_group_clusters_set(
5763 sb, gdp, ext4_free_group_clusters(sb, gdp) +
5764 count - already_freed);
5765 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
5766 ext4_group_desc_csum_set(sb, group, gdp);
5767 ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
5768 sync_dirty_buffer(bitmap_bh);
5769 sync_dirty_buffer(gdp_bh);
5774 * ext4_free_blocks() -- Free given blocks and update quota
5775 * @handle: handle for this transaction
5777 * @bh: optional buffer of the block to be freed
5778 * @block: starting physical block to be freed
5779 * @count: number of blocks to be freed
5780 * @flags: flags used by ext4_free_blocks
5782 void ext4_free_blocks(handle_t *handle, struct inode *inode,
5783 struct buffer_head *bh, ext4_fsblk_t block,
5784 unsigned long count, int flags)
5786 struct buffer_head *bitmap_bh = NULL;
5787 struct super_block *sb = inode->i_sb;
5788 struct ext4_group_desc *gdp;
5789 unsigned int overflow;
5791 struct buffer_head *gd_bh;
5792 ext4_group_t block_group;
5793 struct ext4_sb_info *sbi;
5794 struct ext4_buddy e4b;
5795 unsigned int count_clusters;
5801 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
5802 ext4_free_blocks_simple(inode, block, count);
5809 BUG_ON(block != bh->b_blocknr);
5811 block = bh->b_blocknr;
5814 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5815 !ext4_inode_block_valid(inode, block, count)) {
5816 ext4_error(sb, "Freeing blocks not in datazone - "
5817 "block = %llu, count = %lu", block, count);
5821 ext4_debug("freeing block %llu\n", block);
5822 trace_ext4_free_blocks(inode, block, count, flags);
5824 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
5827 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
5832 * If the extent to be freed does not begin on a cluster
5833 * boundary, we need to deal with partial clusters at the
5834 * beginning and end of the extent. Normally we will free
5835 * blocks at the beginning or the end unless we are explicitly
5836 * requested to avoid doing so.
5838 overflow = EXT4_PBLK_COFF(sbi, block);
5840 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
5841 overflow = sbi->s_cluster_ratio - overflow;
5843 if (count > overflow)
5852 overflow = EXT4_LBLK_COFF(sbi, count);
5854 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
5855 if (count > overflow)
5860 count += sbi->s_cluster_ratio - overflow;
5863 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
5865 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
5867 for (i = 0; i < count; i++) {
5870 bh = sb_find_get_block(inode->i_sb, block + i);
5871 ext4_forget(handle, is_metadata, inode, bh, block + i);
5877 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5879 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5880 ext4_get_group_info(sb, block_group))))
5884 * Check to see if we are freeing blocks across a group
5887 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5888 overflow = EXT4_C2B(sbi, bit) + count -
5889 EXT4_BLOCKS_PER_GROUP(sb);
5892 count_clusters = EXT4_NUM_B2C(sbi, count);
5893 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5894 if (IS_ERR(bitmap_bh)) {
5895 err = PTR_ERR(bitmap_bh);
5899 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5905 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
5906 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
5907 in_range(block, ext4_inode_table(sb, gdp),
5908 sbi->s_itb_per_group) ||
5909 in_range(block + count - 1, ext4_inode_table(sb, gdp),
5910 sbi->s_itb_per_group)) {
5912 ext4_error(sb, "Freeing blocks in system zone - "
5913 "Block = %llu, count = %lu", block, count);
5914 /* err = 0. ext4_std_error should be a no op */
5918 BUFFER_TRACE(bitmap_bh, "getting write access");
5919 err = ext4_journal_get_write_access(handle, bitmap_bh);
5924 * We are about to modify some metadata. Call the journal APIs
5925 * to unshare ->b_data if a currently-committing transaction is
5928 BUFFER_TRACE(gd_bh, "get_write_access");
5929 err = ext4_journal_get_write_access(handle, gd_bh);
5932 #ifdef AGGRESSIVE_CHECK
5935 for (i = 0; i < count_clusters; i++)
5936 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
5939 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
5941 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
5942 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
5943 GFP_NOFS|__GFP_NOFAIL);
5948 * We need to make sure we don't reuse the freed block until after the
5949 * transaction is committed. We make an exception if the inode is to be
5950 * written in writeback mode since writeback mode has weak data
5951 * consistency guarantees.
5953 if (ext4_handle_valid(handle) &&
5954 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
5955 !ext4_should_writeback_data(inode))) {
5956 struct ext4_free_data *new_entry;
5958 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
5961 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
5962 GFP_NOFS|__GFP_NOFAIL);
5963 new_entry->efd_start_cluster = bit;
5964 new_entry->efd_group = block_group;
5965 new_entry->efd_count = count_clusters;
5966 new_entry->efd_tid = handle->h_transaction->t_tid;
5968 ext4_lock_group(sb, block_group);
5969 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
5970 ext4_mb_free_metadata(handle, &e4b, new_entry);
5972 /* need to update group_info->bb_free and bitmap
5973 * with group lock held. generate_buddy look at
5974 * them with group lock_held
5976 if (test_opt(sb, DISCARD)) {
5977 err = ext4_issue_discard(sb, block_group, bit, count,
5979 if (err && err != -EOPNOTSUPP)
5980 ext4_msg(sb, KERN_WARNING, "discard request in"
5981 " group:%d block:%d count:%lu failed"
5982 " with %d", block_group, bit, count,
5985 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
5987 ext4_lock_group(sb, block_group);
5988 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
5989 mb_free_blocks(inode, &e4b, bit, count_clusters);
5992 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
5993 ext4_free_group_clusters_set(sb, gdp, ret);
5994 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
5995 ext4_group_desc_csum_set(sb, block_group, gdp);
5996 ext4_unlock_group(sb, block_group);
5998 if (sbi->s_log_groups_per_flex) {
5999 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6000 atomic64_add(count_clusters,
6001 &sbi_array_rcu_deref(sbi, s_flex_groups,
6002 flex_group)->free_clusters);
6006 * on a bigalloc file system, defer the s_freeclusters_counter
6007 * update to the caller (ext4_remove_space and friends) so they
6008 * can determine if a cluster freed here should be rereserved
6010 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6011 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6012 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6013 percpu_counter_add(&sbi->s_freeclusters_counter,
6017 ext4_mb_unload_buddy(&e4b);
6019 /* We dirtied the bitmap block */
6020 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6021 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6023 /* And the group descriptor block */
6024 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6025 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6029 if (overflow && !err) {
6037 ext4_std_error(sb, err);
6042 * ext4_group_add_blocks() -- Add given blocks to an existing group
6043 * @handle: handle to this transaction
6045 * @block: start physical block to add to the block group
6046 * @count: number of blocks to free
6048 * This marks the blocks as free in the bitmap and buddy.
6050 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6051 ext4_fsblk_t block, unsigned long count)
6053 struct buffer_head *bitmap_bh = NULL;
6054 struct buffer_head *gd_bh;
6055 ext4_group_t block_group;
6058 struct ext4_group_desc *desc;
6059 struct ext4_sb_info *sbi = EXT4_SB(sb);
6060 struct ext4_buddy e4b;
6061 int err = 0, ret, free_clusters_count;
6062 ext4_grpblk_t clusters_freed;
6063 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6064 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6065 unsigned long cluster_count = last_cluster - first_cluster + 1;
6067 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6072 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6074 * Check to see if we are freeing blocks across a group
6077 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6078 ext4_warning(sb, "too many blocks added to group %u",
6084 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
6085 if (IS_ERR(bitmap_bh)) {
6086 err = PTR_ERR(bitmap_bh);
6091 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
6097 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
6098 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
6099 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
6100 in_range(block + count - 1, ext4_inode_table(sb, desc),
6101 sbi->s_itb_per_group)) {
6102 ext4_error(sb, "Adding blocks in system zones - "
6103 "Block = %llu, count = %lu",
6109 BUFFER_TRACE(bitmap_bh, "getting write access");
6110 err = ext4_journal_get_write_access(handle, bitmap_bh);
6115 * We are about to modify some metadata. Call the journal APIs
6116 * to unshare ->b_data if a currently-committing transaction is
6119 BUFFER_TRACE(gd_bh, "get_write_access");
6120 err = ext4_journal_get_write_access(handle, gd_bh);
6124 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
6125 BUFFER_TRACE(bitmap_bh, "clear bit");
6126 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
6127 ext4_error(sb, "bit already cleared for block %llu",
6128 (ext4_fsblk_t)(block + i));
6129 BUFFER_TRACE(bitmap_bh, "bit already cleared");
6135 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6140 * need to update group_info->bb_free and bitmap
6141 * with group lock held. generate_buddy look at
6142 * them with group lock_held
6144 ext4_lock_group(sb, block_group);
6145 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
6146 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6147 free_clusters_count = clusters_freed +
6148 ext4_free_group_clusters(sb, desc);
6149 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
6150 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
6151 ext4_group_desc_csum_set(sb, block_group, desc);
6152 ext4_unlock_group(sb, block_group);
6153 percpu_counter_add(&sbi->s_freeclusters_counter,
6156 if (sbi->s_log_groups_per_flex) {
6157 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6158 atomic64_add(clusters_freed,
6159 &sbi_array_rcu_deref(sbi, s_flex_groups,
6160 flex_group)->free_clusters);
6163 ext4_mb_unload_buddy(&e4b);
6165 /* We dirtied the bitmap block */
6166 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6167 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6169 /* And the group descriptor block */
6170 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6171 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6177 ext4_std_error(sb, err);
6182 * ext4_trim_extent -- function to TRIM one single free extent in the group
6183 * @sb: super block for the file system
6184 * @start: starting block of the free extent in the alloc. group
6185 * @count: number of blocks to TRIM
6186 * @group: alloc. group we are working with
6187 * @e4b: ext4 buddy for the group
6189 * Trim "count" blocks starting at "start" in the "group". To assure that no
6190 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6191 * be called with under the group lock.
6193 static int ext4_trim_extent(struct super_block *sb, int start, int count,
6194 ext4_group_t group, struct ext4_buddy *e4b)
6198 struct ext4_free_extent ex;
6201 trace_ext4_trim_extent(sb, group, start, count);
6203 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6205 ex.fe_start = start;
6206 ex.fe_group = group;
6210 * Mark blocks used, so no one can reuse them while
6213 mb_mark_used(e4b, &ex);
6214 ext4_unlock_group(sb, group);
6215 ret = ext4_issue_discard(sb, group, start, count, NULL);
6216 ext4_lock_group(sb, group);
6217 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6222 * ext4_trim_all_free -- function to trim all free space in alloc. group
6223 * @sb: super block for file system
6224 * @group: group to be trimmed
6225 * @start: first group block to examine
6226 * @max: last group block to examine
6227 * @minblocks: minimum extent block count
6229 * ext4_trim_all_free walks through group's buddy bitmap searching for free
6230 * extents. When the free block is found, ext4_trim_extent is called to TRIM
6234 * ext4_trim_all_free walks through group's block bitmap searching for free
6235 * extents. When the free extent is found, mark it as used in group buddy
6236 * bitmap. Then issue a TRIM command on this extent and free the extent in
6237 * the group buddy bitmap. This is done until whole group is scanned.
6239 static ext4_grpblk_t
6240 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6241 ext4_grpblk_t start, ext4_grpblk_t max,
6242 ext4_grpblk_t minblocks)
6245 ext4_grpblk_t next, count = 0, free_count = 0;
6246 struct ext4_buddy e4b;
6249 trace_ext4_trim_all_free(sb, group, start, max);
6251 ret = ext4_mb_load_buddy(sb, group, &e4b);
6253 ext4_warning(sb, "Error %d loading buddy information for %u",
6257 bitmap = e4b.bd_bitmap;
6259 ext4_lock_group(sb, group);
6260 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
6261 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
6264 start = (e4b.bd_info->bb_first_free > start) ?
6265 e4b.bd_info->bb_first_free : start;
6267 while (start <= max) {
6268 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6271 next = mb_find_next_bit(bitmap, max + 1, start);
6273 if ((next - start) >= minblocks) {
6274 ret = ext4_trim_extent(sb, start,
6275 next - start, group, &e4b);
6276 if (ret && ret != -EOPNOTSUPP)
6279 count += next - start;
6281 free_count += next - start;
6284 if (fatal_signal_pending(current)) {
6285 count = -ERESTARTSYS;
6289 if (need_resched()) {
6290 ext4_unlock_group(sb, group);
6292 ext4_lock_group(sb, group);
6295 if ((e4b.bd_info->bb_free - free_count) < minblocks)
6301 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
6304 ext4_unlock_group(sb, group);
6305 ext4_mb_unload_buddy(&e4b);
6307 ext4_debug("trimmed %d blocks in the group %d\n",
6314 * ext4_trim_fs() -- trim ioctl handle function
6315 * @sb: superblock for filesystem
6316 * @range: fstrim_range structure
6318 * start: First Byte to trim
6319 * len: number of Bytes to trim from start
6320 * minlen: minimum extent length in Bytes
6321 * ext4_trim_fs goes through all allocation groups containing Bytes from
6322 * start to start+len. For each such a group ext4_trim_all_free function
6323 * is invoked to trim all free space.
6325 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6327 struct ext4_group_info *grp;
6328 ext4_group_t group, first_group, last_group;
6329 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6330 uint64_t start, end, minlen, trimmed = 0;
6331 ext4_fsblk_t first_data_blk =
6332 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6333 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6336 start = range->start >> sb->s_blocksize_bits;
6337 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6338 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6339 range->minlen >> sb->s_blocksize_bits);
6341 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6342 start >= max_blks ||
6343 range->len < sb->s_blocksize)
6345 if (end >= max_blks)
6347 if (end <= first_data_blk)
6349 if (start < first_data_blk)
6350 start = first_data_blk;
6352 /* Determine first and last group to examine based on start and end */
6353 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6354 &first_group, &first_cluster);
6355 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6356 &last_group, &last_cluster);
6358 /* end now represents the last cluster to discard in this group */
6359 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6361 for (group = first_group; group <= last_group; group++) {
6362 grp = ext4_get_group_info(sb, group);
6363 /* We only do this if the grp has never been initialized */
6364 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6365 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6371 * For all the groups except the last one, last cluster will
6372 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6373 * change it for the last group, note that last_cluster is
6374 * already computed earlier by ext4_get_group_no_and_offset()
6376 if (group == last_group)
6379 if (grp->bb_free >= minlen) {
6380 cnt = ext4_trim_all_free(sb, group, first_cluster,
6390 * For every group except the first one, we are sure
6391 * that the first cluster to discard will be cluster #0.
6397 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
6400 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6404 /* Iterate all the free extents in the group. */
6406 ext4_mballoc_query_range(
6407 struct super_block *sb,
6409 ext4_grpblk_t start,
6411 ext4_mballoc_query_range_fn formatter,
6416 struct ext4_buddy e4b;
6419 error = ext4_mb_load_buddy(sb, group, &e4b);
6422 bitmap = e4b.bd_bitmap;
6424 ext4_lock_group(sb, group);
6426 start = (e4b.bd_info->bb_first_free > start) ?
6427 e4b.bd_info->bb_first_free : start;
6428 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
6429 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6431 while (start <= end) {
6432 start = mb_find_next_zero_bit(bitmap, end + 1, start);
6435 next = mb_find_next_bit(bitmap, end + 1, start);
6437 ext4_unlock_group(sb, group);
6438 error = formatter(sb, group, start, next - start, priv);
6441 ext4_lock_group(sb, group);
6446 ext4_unlock_group(sb, group);
6448 ext4_mb_unload_buddy(&e4b);