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 <linux/freezer.h>
20 #include <trace/events/ext4.h>
21 #include <kunit/static_stub.h>
25 * - test ext4_ext_search_left() and ext4_ext_search_right()
26 * - search for metadata in few groups
29 * - normalization should take into account whether file is still open
30 * - discard preallocations if no free space left (policy?)
31 * - don't normalize tails
33 * - reservation for superuser
36 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
37 * - track min/max extents in each group for better group selection
38 * - mb_mark_used() may allocate chunk right after splitting buddy
39 * - tree of groups sorted by number of free blocks
44 * The allocation request involve request for multiple number of blocks
45 * near to the goal(block) value specified.
47 * During initialization phase of the allocator we decide to use the
48 * group preallocation or inode preallocation depending on the size of
49 * the file. The size of the file could be the resulting file size we
50 * would have after allocation, or the current file size, which ever
51 * is larger. If the size is less than sbi->s_mb_stream_request we
52 * select to use the group preallocation. The default value of
53 * s_mb_stream_request is 16 blocks. This can also be tuned via
54 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
55 * terms of number of blocks.
57 * The main motivation for having small file use group preallocation is to
58 * ensure that we have small files closer together on the disk.
60 * First stage the allocator looks at the inode prealloc list,
61 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
62 * spaces for this particular inode. The inode prealloc space is
65 * pa_lstart -> the logical start block for this prealloc space
66 * pa_pstart -> the physical start block for this prealloc space
67 * pa_len -> length for this prealloc space (in clusters)
68 * pa_free -> free space available in this prealloc space (in clusters)
70 * The inode preallocation space is used looking at the _logical_ start
71 * block. If only the logical file block falls within the range of prealloc
72 * space we will consume the particular prealloc space. This makes sure that
73 * we have contiguous physical blocks representing the file blocks
75 * The important thing to be noted in case of inode prealloc space is that
76 * we don't modify the values associated to inode prealloc space except
79 * If we are not able to find blocks in the inode prealloc space and if we
80 * have the group allocation flag set then we look at the locality group
81 * prealloc space. These are per CPU prealloc list represented as
83 * ext4_sb_info.s_locality_groups[smp_processor_id()]
85 * The reason for having a per cpu locality group is to reduce the contention
86 * between CPUs. It is possible to get scheduled at this point.
88 * The locality group prealloc space is used looking at whether we have
89 * enough free space (pa_free) within the prealloc space.
91 * If we can't allocate blocks via inode prealloc or/and locality group
92 * prealloc then we look at the buddy cache. The buddy cache is represented
93 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
94 * mapped to the buddy and bitmap information regarding different
95 * groups. The buddy information is attached to buddy cache inode so that
96 * we can access them through the page cache. The information regarding
97 * each group is loaded via ext4_mb_load_buddy. The information involve
98 * block bitmap and buddy information. The information are stored in the
102 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
105 * one block each for bitmap and buddy information. So for each group we
106 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
107 * blocksize) blocks. So it can have information regarding groups_per_page
108 * which is blocks_per_page/2
110 * The buddy cache inode is not stored on disk. The inode is thrown
111 * away when the filesystem is unmounted.
113 * We look for count number of blocks in the buddy cache. If we were able
114 * to locate that many free blocks we return with additional information
115 * regarding rest of the contiguous physical block available
117 * Before allocating blocks via buddy cache we normalize the request
118 * blocks. This ensure we ask for more blocks that we needed. The extra
119 * blocks that we get after allocation is added to the respective prealloc
120 * list. In case of inode preallocation we follow a list of heuristics
121 * based on file size. This can be found in ext4_mb_normalize_request. If
122 * we are doing a group prealloc we try to normalize the request to
123 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
124 * dependent on the cluster size; for non-bigalloc file systems, it is
125 * 512 blocks. This can be tuned via
126 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
127 * terms of number of blocks. If we have mounted the file system with -O
128 * stripe=<value> option the group prealloc request is normalized to the
129 * smallest multiple of the stripe value (sbi->s_stripe) which is
130 * greater than the default mb_group_prealloc.
132 * If "mb_optimize_scan" mount option is set, we maintain in memory group info
133 * structures in two data structures:
135 * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders)
137 * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks)
139 * This is an array of lists where the index in the array represents the
140 * largest free order in the buddy bitmap of the participating group infos of
141 * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total
142 * number of buddy bitmap orders possible) number of lists. Group-infos are
143 * placed in appropriate lists.
145 * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size)
147 * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks)
149 * This is an array of lists where in the i-th list there are groups with
150 * average fragment size >= 2^i and < 2^(i+1). The average fragment size
151 * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments.
152 * Note that we don't bother with a special list for completely empty groups
153 * so we only have MB_NUM_ORDERS(sb) lists.
155 * When "mb_optimize_scan" mount option is set, mballoc consults the above data
156 * structures to decide the order in which groups are to be traversed for
157 * fulfilling an allocation request.
159 * At CR_POWER2_ALIGNED , we look for groups which have the largest_free_order
160 * >= the order of the request. We directly look at the largest free order list
161 * in the data structure (1) above where largest_free_order = order of the
162 * request. If that list is empty, we look at remaining list in the increasing
163 * order of largest_free_order. This allows us to perform CR_POWER2_ALIGNED
164 * lookup in O(1) time.
166 * At CR_GOAL_LEN_FAST, we only consider groups where
167 * average fragment size > request size. So, we lookup a group which has average
168 * fragment size just above or equal to request size using our average fragment
169 * size group lists (data structure 2) in O(1) time.
171 * At CR_BEST_AVAIL_LEN, we aim to optimize allocations which can't be satisfied
172 * in CR_GOAL_LEN_FAST. The fact that we couldn't find a group in
173 * CR_GOAL_LEN_FAST suggests that there is no BG that has avg
174 * fragment size > goal length. So before falling to the slower
175 * CR_GOAL_LEN_SLOW, in CR_BEST_AVAIL_LEN we proactively trim goal length and
176 * then use the same fragment lists as CR_GOAL_LEN_FAST to find a BG with a big
177 * enough average fragment size. This increases the chances of finding a
178 * suitable block group in O(1) time and results in faster allocation at the
179 * cost of reduced size of allocation.
181 * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in
182 * linear order which requires O(N) search time for each CR_POWER2_ALIGNED and
183 * CR_GOAL_LEN_FAST phase.
185 * The regular allocator (using the buddy cache) supports a few tunables.
187 * /sys/fs/ext4/<partition>/mb_min_to_scan
188 * /sys/fs/ext4/<partition>/mb_max_to_scan
189 * /sys/fs/ext4/<partition>/mb_order2_req
190 * /sys/fs/ext4/<partition>/mb_linear_limit
192 * The regular allocator uses buddy scan only if the request len is power of
193 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
194 * value of s_mb_order2_reqs can be tuned via
195 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
196 * stripe size (sbi->s_stripe), we try to search for contiguous block in
197 * stripe size. This should result in better allocation on RAID setups. If
198 * not, we search in the specific group using bitmap for best extents. The
199 * tunable min_to_scan and max_to_scan control the behaviour here.
200 * min_to_scan indicate how long the mballoc __must__ look for a best
201 * extent and max_to_scan indicates how long the mballoc __can__ look for a
202 * best extent in the found extents. Searching for the blocks starts with
203 * the group specified as the goal value in allocation context via
204 * ac_g_ex. Each group is first checked based on the criteria whether it
205 * can be used for allocation. ext4_mb_good_group explains how the groups are
208 * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not
209 * get traversed linearly. That may result in subsequent allocations being not
210 * close to each other. And so, the underlying device may get filled up in a
211 * non-linear fashion. While that may not matter on non-rotational devices, for
212 * rotational devices that may result in higher seek times. "mb_linear_limit"
213 * tells mballoc how many groups mballoc should search linearly before
214 * performing consulting above data structures for more efficient lookups. For
215 * non rotational devices, this value defaults to 0 and for rotational devices
216 * this is set to MB_DEFAULT_LINEAR_LIMIT.
218 * Both the prealloc space are getting populated as above. So for the first
219 * request we will hit the buddy cache which will result in this prealloc
220 * space getting filled. The prealloc space is then later used for the
221 * subsequent request.
225 * mballoc operates on the following data:
227 * - in-core buddy (actually includes buddy and bitmap)
228 * - preallocation descriptors (PAs)
230 * there are two types of preallocations:
232 * assiged to specific inode and can be used for this inode only.
233 * it describes part of inode's space preallocated to specific
234 * physical blocks. any block from that preallocated can be used
235 * independent. the descriptor just tracks number of blocks left
236 * unused. so, before taking some block from descriptor, one must
237 * make sure corresponded logical block isn't allocated yet. this
238 * also means that freeing any block within descriptor's range
239 * must discard all preallocated blocks.
241 * assigned to specific locality group which does not translate to
242 * permanent set of inodes: inode can join and leave group. space
243 * from this type of preallocation can be used for any inode. thus
244 * it's consumed from the beginning to the end.
246 * relation between them can be expressed as:
247 * in-core buddy = on-disk bitmap + preallocation descriptors
249 * this mean blocks mballoc considers used are:
250 * - allocated blocks (persistent)
251 * - preallocated blocks (non-persistent)
253 * consistency in mballoc world means that at any time a block is either
254 * free or used in ALL structures. notice: "any time" should not be read
255 * literally -- time is discrete and delimited by locks.
257 * to keep it simple, we don't use block numbers, instead we count number of
258 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
260 * all operations can be expressed as:
261 * - init buddy: buddy = on-disk + PAs
262 * - new PA: buddy += N; PA = N
263 * - use inode PA: on-disk += N; PA -= N
264 * - discard inode PA buddy -= on-disk - PA; PA = 0
265 * - use locality group PA on-disk += N; PA -= N
266 * - discard locality group PA buddy -= PA; PA = 0
267 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
268 * is used in real operation because we can't know actual used
269 * bits from PA, only from on-disk bitmap
271 * if we follow this strict logic, then all operations above should be atomic.
272 * given some of them can block, we'd have to use something like semaphores
273 * killing performance on high-end SMP hardware. let's try to relax it using
274 * the following knowledge:
275 * 1) if buddy is referenced, it's already initialized
276 * 2) while block is used in buddy and the buddy is referenced,
277 * nobody can re-allocate that block
278 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
279 * bit set and PA claims same block, it's OK. IOW, one can set bit in
280 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
283 * so, now we're building a concurrency table:
286 * blocks for PA are allocated in the buddy, buddy must be referenced
287 * until PA is linked to allocation group to avoid concurrent buddy init
289 * we need to make sure that either on-disk bitmap or PA has uptodate data
290 * given (3) we care that PA-=N operation doesn't interfere with init
292 * the simplest way would be to have buddy initialized by the discard
293 * - use locality group PA
294 * again PA-=N must be serialized with init
295 * - discard locality group PA
296 * the simplest way would be to have buddy initialized by the discard
299 * i_data_sem serializes them
301 * discard process must wait until PA isn't used by another process
302 * - use locality group PA
303 * some mutex should serialize them
304 * - discard locality group PA
305 * discard process must wait until PA isn't used by another process
308 * i_data_sem or another mutex should serializes them
310 * discard process must wait until PA isn't used by another process
311 * - use locality group PA
312 * nothing wrong here -- they're different PAs covering different blocks
313 * - discard locality group PA
314 * discard process must wait until PA isn't used by another process
316 * now we're ready to make few consequences:
317 * - PA is referenced and while it is no discard is possible
318 * - PA is referenced until block isn't marked in on-disk bitmap
319 * - PA changes only after on-disk bitmap
320 * - discard must not compete with init. either init is done before
321 * any discard or they're serialized somehow
322 * - buddy init as sum of on-disk bitmap and PAs is done atomically
324 * a special case when we've used PA to emptiness. no need to modify buddy
325 * in this case, but we should care about concurrent init
330 * Logic in few words:
335 * mark bits in on-disk bitmap
338 * - use preallocation:
339 * find proper PA (per-inode or group)
341 * mark bits in on-disk bitmap
347 * mark bits in on-disk bitmap
350 * - discard preallocations in group:
352 * move them onto local list
353 * load on-disk bitmap
355 * remove PA from object (inode or locality group)
356 * mark free blocks in-core
358 * - discard inode's preallocations:
365 * - bitlock on a group (group)
366 * - object (inode/locality) (object)
368 * - cr_power2_aligned lists lock (cr_power2_aligned)
369 * - cr_goal_len_fast lists lock (cr_goal_len_fast)
379 * - release consumed pa:
384 * - generate in-core bitmap:
388 * - discard all for given object (inode, locality group):
393 * - discard all for given group:
399 * - allocation path (ext4_mb_regular_allocator)
401 * cr_power2_aligned/cr_goal_len_fast
403 static struct kmem_cache *ext4_pspace_cachep;
404 static struct kmem_cache *ext4_ac_cachep;
405 static struct kmem_cache *ext4_free_data_cachep;
407 /* We create slab caches for groupinfo data structures based on the
408 * superblock block size. There will be one per mounted filesystem for
409 * each unique s_blocksize_bits */
410 #define NR_GRPINFO_CACHES 8
411 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
413 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
414 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
415 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
416 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
419 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
421 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac);
423 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
424 ext4_group_t group, enum criteria cr);
426 static int ext4_try_to_trim_range(struct super_block *sb,
427 struct ext4_buddy *e4b, ext4_grpblk_t start,
428 ext4_grpblk_t max, ext4_grpblk_t minblocks);
431 * The algorithm using this percpu seq counter goes below:
432 * 1. We sample the percpu discard_pa_seq counter before trying for block
433 * allocation in ext4_mb_new_blocks().
434 * 2. We increment this percpu discard_pa_seq counter when we either allocate
435 * or free these blocks i.e. while marking those blocks as used/free in
436 * mb_mark_used()/mb_free_blocks().
437 * 3. We also increment this percpu seq counter when we successfully identify
438 * that the bb_prealloc_list is not empty and hence proceed for discarding
439 * of those PAs inside ext4_mb_discard_group_preallocations().
441 * Now to make sure that the regular fast path of block allocation is not
442 * affected, as a small optimization we only sample the percpu seq counter
443 * on that cpu. Only when the block allocation fails and when freed blocks
444 * found were 0, that is when we sample percpu seq counter for all cpus using
445 * below function ext4_get_discard_pa_seq_sum(). This happens after making
446 * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
448 static DEFINE_PER_CPU(u64, discard_pa_seq);
449 static inline u64 ext4_get_discard_pa_seq_sum(void)
454 for_each_possible_cpu(__cpu)
455 __seq += per_cpu(discard_pa_seq, __cpu);
459 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
461 #if BITS_PER_LONG == 64
462 *bit += ((unsigned long) addr & 7UL) << 3;
463 addr = (void *) ((unsigned long) addr & ~7UL);
464 #elif BITS_PER_LONG == 32
465 *bit += ((unsigned long) addr & 3UL) << 3;
466 addr = (void *) ((unsigned long) addr & ~3UL);
468 #error "how many bits you are?!"
473 static inline int mb_test_bit(int bit, void *addr)
476 * ext4_test_bit on architecture like powerpc
477 * needs unsigned long aligned address
479 addr = mb_correct_addr_and_bit(&bit, addr);
480 return ext4_test_bit(bit, addr);
483 static inline void mb_set_bit(int bit, void *addr)
485 addr = mb_correct_addr_and_bit(&bit, addr);
486 ext4_set_bit(bit, addr);
489 static inline void mb_clear_bit(int bit, void *addr)
491 addr = mb_correct_addr_and_bit(&bit, addr);
492 ext4_clear_bit(bit, addr);
495 static inline int mb_test_and_clear_bit(int bit, void *addr)
497 addr = mb_correct_addr_and_bit(&bit, addr);
498 return ext4_test_and_clear_bit(bit, addr);
501 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
503 int fix = 0, ret, tmpmax;
504 addr = mb_correct_addr_and_bit(&fix, addr);
508 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
514 static inline int mb_find_next_bit(void *addr, int max, int start)
516 int fix = 0, ret, tmpmax;
517 addr = mb_correct_addr_and_bit(&fix, addr);
521 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
527 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
531 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
534 if (order > e4b->bd_blkbits + 1) {
539 /* at order 0 we see each particular block */
541 *max = 1 << (e4b->bd_blkbits + 3);
542 return e4b->bd_bitmap;
545 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
546 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
552 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
553 int first, int count)
556 struct super_block *sb = e4b->bd_sb;
558 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
560 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
561 for (i = 0; i < count; i++) {
562 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
563 ext4_fsblk_t blocknr;
565 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
566 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
567 ext4_grp_locked_error(sb, e4b->bd_group,
568 inode ? inode->i_ino : 0,
570 "freeing block already freed "
573 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
574 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
576 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
580 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
584 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
586 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
587 for (i = 0; i < count; i++) {
588 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
589 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
593 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
595 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
597 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
598 unsigned char *b1, *b2;
600 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
601 b2 = (unsigned char *) bitmap;
602 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
603 if (b1[i] != b2[i]) {
604 ext4_msg(e4b->bd_sb, KERN_ERR,
605 "corruption in group %u "
606 "at byte %u(%u): %x in copy != %x "
608 e4b->bd_group, i, i * 8, b1[i], b2[i]);
615 static void mb_group_bb_bitmap_alloc(struct super_block *sb,
616 struct ext4_group_info *grp, ext4_group_t group)
618 struct buffer_head *bh;
620 grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
624 bh = ext4_read_block_bitmap(sb, group);
625 if (IS_ERR_OR_NULL(bh)) {
626 kfree(grp->bb_bitmap);
627 grp->bb_bitmap = NULL;
631 memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
635 static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
637 kfree(grp->bb_bitmap);
641 static inline void mb_free_blocks_double(struct inode *inode,
642 struct ext4_buddy *e4b, int first, int count)
646 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
647 int first, int count)
651 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
656 static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
657 struct ext4_group_info *grp, ext4_group_t group)
662 static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
668 #ifdef AGGRESSIVE_CHECK
670 #define MB_CHECK_ASSERT(assert) \
674 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
675 function, file, line, # assert); \
680 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
681 const char *function, int line)
683 struct super_block *sb = e4b->bd_sb;
684 int order = e4b->bd_blkbits + 1;
691 struct ext4_group_info *grp;
694 struct list_head *cur;
698 if (e4b->bd_info->bb_check_counter++ % 10)
702 buddy = mb_find_buddy(e4b, order, &max);
703 MB_CHECK_ASSERT(buddy);
704 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
705 MB_CHECK_ASSERT(buddy2);
706 MB_CHECK_ASSERT(buddy != buddy2);
707 MB_CHECK_ASSERT(max * 2 == max2);
710 for (i = 0; i < max; i++) {
712 if (mb_test_bit(i, buddy)) {
713 /* only single bit in buddy2 may be 0 */
714 if (!mb_test_bit(i << 1, buddy2)) {
716 mb_test_bit((i<<1)+1, buddy2));
721 /* both bits in buddy2 must be 1 */
722 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
723 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
725 for (j = 0; j < (1 << order); j++) {
726 k = (i * (1 << order)) + j;
728 !mb_test_bit(k, e4b->bd_bitmap));
732 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
737 buddy = mb_find_buddy(e4b, 0, &max);
738 for (i = 0; i < max; i++) {
739 if (!mb_test_bit(i, buddy)) {
740 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
748 /* check used bits only */
749 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
750 buddy2 = mb_find_buddy(e4b, j, &max2);
752 MB_CHECK_ASSERT(k < max2);
753 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
756 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
757 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
759 grp = ext4_get_group_info(sb, e4b->bd_group);
762 list_for_each(cur, &grp->bb_prealloc_list) {
763 ext4_group_t groupnr;
764 struct ext4_prealloc_space *pa;
765 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
766 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
767 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
768 for (i = 0; i < pa->pa_len; i++)
769 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
773 #undef MB_CHECK_ASSERT
774 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
775 __FILE__, __func__, __LINE__)
777 #define mb_check_buddy(e4b)
781 * Divide blocks started from @first with length @len into
782 * smaller chunks with power of 2 blocks.
783 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
784 * then increase bb_counters[] for corresponded chunk size.
786 static void ext4_mb_mark_free_simple(struct super_block *sb,
787 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
788 struct ext4_group_info *grp)
790 struct ext4_sb_info *sbi = EXT4_SB(sb);
796 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
798 border = 2 << sb->s_blocksize_bits;
801 /* find how many blocks can be covered since this position */
802 max = ffs(first | border) - 1;
804 /* find how many blocks of power 2 we need to mark */
811 /* mark multiblock chunks only */
812 grp->bb_counters[min]++;
814 mb_clear_bit(first >> min,
815 buddy + sbi->s_mb_offsets[min]);
822 static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len)
827 * We don't bother with a special lists groups with only 1 block free
828 * extents and for completely empty groups.
830 order = fls(len) - 2;
833 if (order == MB_NUM_ORDERS(sb))
838 /* Move group to appropriate avg_fragment_size list */
840 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
842 struct ext4_sb_info *sbi = EXT4_SB(sb);
845 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_fragments == 0)
848 new_order = mb_avg_fragment_size_order(sb,
849 grp->bb_free / grp->bb_fragments);
850 if (new_order == grp->bb_avg_fragment_size_order)
853 if (grp->bb_avg_fragment_size_order != -1) {
854 write_lock(&sbi->s_mb_avg_fragment_size_locks[
855 grp->bb_avg_fragment_size_order]);
856 list_del(&grp->bb_avg_fragment_size_node);
857 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
858 grp->bb_avg_fragment_size_order]);
860 grp->bb_avg_fragment_size_order = new_order;
861 write_lock(&sbi->s_mb_avg_fragment_size_locks[
862 grp->bb_avg_fragment_size_order]);
863 list_add_tail(&grp->bb_avg_fragment_size_node,
864 &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]);
865 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
866 grp->bb_avg_fragment_size_order]);
870 * Choose next group by traversing largest_free_order lists. Updates *new_cr if
871 * cr level needs an update.
873 static void ext4_mb_choose_next_group_p2_aligned(struct ext4_allocation_context *ac,
874 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups)
876 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
877 struct ext4_group_info *iter;
880 if (ac->ac_status == AC_STATUS_FOUND)
883 if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED))
884 atomic_inc(&sbi->s_bal_p2_aligned_bad_suggestions);
886 for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) {
887 if (list_empty(&sbi->s_mb_largest_free_orders[i]))
889 read_lock(&sbi->s_mb_largest_free_orders_locks[i]);
890 if (list_empty(&sbi->s_mb_largest_free_orders[i])) {
891 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
894 list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i],
895 bb_largest_free_order_node) {
897 atomic64_inc(&sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]);
898 if (likely(ext4_mb_good_group(ac, iter->bb_group, CR_POWER2_ALIGNED))) {
899 *group = iter->bb_group;
900 ac->ac_flags |= EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED;
901 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
905 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
908 /* Increment cr and search again if no group is found */
909 *new_cr = CR_GOAL_LEN_FAST;
913 * Find a suitable group of given order from the average fragments list.
915 static struct ext4_group_info *
916 ext4_mb_find_good_group_avg_frag_lists(struct ext4_allocation_context *ac, int order)
918 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
919 struct list_head *frag_list = &sbi->s_mb_avg_fragment_size[order];
920 rwlock_t *frag_list_lock = &sbi->s_mb_avg_fragment_size_locks[order];
921 struct ext4_group_info *grp = NULL, *iter;
922 enum criteria cr = ac->ac_criteria;
924 if (list_empty(frag_list))
926 read_lock(frag_list_lock);
927 if (list_empty(frag_list)) {
928 read_unlock(frag_list_lock);
931 list_for_each_entry(iter, frag_list, bb_avg_fragment_size_node) {
933 atomic64_inc(&sbi->s_bal_cX_groups_considered[cr]);
934 if (likely(ext4_mb_good_group(ac, iter->bb_group, cr))) {
939 read_unlock(frag_list_lock);
944 * Choose next group by traversing average fragment size list of suitable
945 * order. Updates *new_cr if cr level needs an update.
947 static void ext4_mb_choose_next_group_goal_fast(struct ext4_allocation_context *ac,
948 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups)
950 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
951 struct ext4_group_info *grp = NULL;
954 if (unlikely(ac->ac_flags & EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED)) {
956 atomic_inc(&sbi->s_bal_goal_fast_bad_suggestions);
959 for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len);
960 i < MB_NUM_ORDERS(ac->ac_sb); i++) {
961 grp = ext4_mb_find_good_group_avg_frag_lists(ac, i);
963 *group = grp->bb_group;
964 ac->ac_flags |= EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED;
970 * CR_BEST_AVAIL_LEN works based on the concept that we have
971 * a larger normalized goal len request which can be trimmed to
972 * a smaller goal len such that it can still satisfy original
973 * request len. However, allocation request for non-regular
974 * files never gets normalized.
975 * See function ext4_mb_normalize_request() (EXT4_MB_HINT_DATA).
977 if (ac->ac_flags & EXT4_MB_HINT_DATA)
978 *new_cr = CR_BEST_AVAIL_LEN;
980 *new_cr = CR_GOAL_LEN_SLOW;
984 * We couldn't find a group in CR_GOAL_LEN_FAST so try to find the highest free fragment
985 * order we have and proactively trim the goal request length to that order to
986 * find a suitable group faster.
988 * This optimizes allocation speed at the cost of slightly reduced
989 * preallocations. However, we make sure that we don't trim the request too
990 * much and fall to CR_GOAL_LEN_SLOW in that case.
992 static void ext4_mb_choose_next_group_best_avail(struct ext4_allocation_context *ac,
993 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups)
995 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
996 struct ext4_group_info *grp = NULL;
997 int i, order, min_order;
998 unsigned long num_stripe_clusters = 0;
1000 if (unlikely(ac->ac_flags & EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED)) {
1001 if (sbi->s_mb_stats)
1002 atomic_inc(&sbi->s_bal_best_avail_bad_suggestions);
1006 * mb_avg_fragment_size_order() returns order in a way that makes
1007 * retrieving back the length using (1 << order) inaccurate. Hence, use
1008 * fls() instead since we need to know the actual length while modifying
1011 order = fls(ac->ac_g_ex.fe_len) - 1;
1012 min_order = order - sbi->s_mb_best_avail_max_trim_order;
1016 if (sbi->s_stripe > 0) {
1018 * We are assuming that stripe size is always a multiple of
1019 * cluster ratio otherwise __ext4_fill_super exists early.
1021 num_stripe_clusters = EXT4_NUM_B2C(sbi, sbi->s_stripe);
1022 if (1 << min_order < num_stripe_clusters)
1024 * We consider 1 order less because later we round
1025 * up the goal len to num_stripe_clusters
1027 min_order = fls(num_stripe_clusters) - 1;
1030 if (1 << min_order < ac->ac_o_ex.fe_len)
1031 min_order = fls(ac->ac_o_ex.fe_len);
1033 for (i = order; i >= min_order; i--) {
1036 * Scale down goal len to make sure we find something
1037 * in the free fragments list. Basically, reduce
1040 ac->ac_g_ex.fe_len = 1 << i;
1042 if (num_stripe_clusters > 0) {
1044 * Try to round up the adjusted goal length to
1045 * stripe size (in cluster units) multiple for
1048 ac->ac_g_ex.fe_len = roundup(ac->ac_g_ex.fe_len,
1049 num_stripe_clusters);
1052 frag_order = mb_avg_fragment_size_order(ac->ac_sb,
1053 ac->ac_g_ex.fe_len);
1055 grp = ext4_mb_find_good_group_avg_frag_lists(ac, frag_order);
1057 *group = grp->bb_group;
1058 ac->ac_flags |= EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED;
1063 /* Reset goal length to original goal length before falling into CR_GOAL_LEN_SLOW */
1064 ac->ac_g_ex.fe_len = ac->ac_orig_goal_len;
1065 *new_cr = CR_GOAL_LEN_SLOW;
1068 static inline int should_optimize_scan(struct ext4_allocation_context *ac)
1070 if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN)))
1072 if (ac->ac_criteria >= CR_GOAL_LEN_SLOW)
1074 if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))
1080 * Return next linear group for allocation. If linear traversal should not be
1081 * performed, this function just returns the same group
1084 next_linear_group(struct ext4_allocation_context *ac, ext4_group_t group,
1085 ext4_group_t ngroups)
1087 if (!should_optimize_scan(ac))
1088 goto inc_and_return;
1090 if (ac->ac_groups_linear_remaining) {
1091 ac->ac_groups_linear_remaining--;
1092 goto inc_and_return;
1098 * Artificially restricted ngroups for non-extent
1099 * files makes group > ngroups possible on first loop.
1101 return group + 1 >= ngroups ? 0 : group + 1;
1105 * ext4_mb_choose_next_group: choose next group for allocation.
1107 * @ac Allocation Context
1108 * @new_cr This is an output parameter. If the there is no good group
1109 * available at current CR level, this field is updated to indicate
1110 * the new cr level that should be used.
1111 * @group This is an input / output parameter. As an input it indicates the
1112 * next group that the allocator intends to use for allocation. As
1113 * output, this field indicates the next group that should be used as
1114 * determined by the optimization functions.
1115 * @ngroups Total number of groups
1117 static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac,
1118 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups)
1120 *new_cr = ac->ac_criteria;
1122 if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining) {
1123 *group = next_linear_group(ac, *group, ngroups);
1127 if (*new_cr == CR_POWER2_ALIGNED) {
1128 ext4_mb_choose_next_group_p2_aligned(ac, new_cr, group, ngroups);
1129 } else if (*new_cr == CR_GOAL_LEN_FAST) {
1130 ext4_mb_choose_next_group_goal_fast(ac, new_cr, group, ngroups);
1131 } else if (*new_cr == CR_BEST_AVAIL_LEN) {
1132 ext4_mb_choose_next_group_best_avail(ac, new_cr, group, ngroups);
1135 * TODO: For CR=2, we can arrange groups in an rb tree sorted by
1136 * bb_free. But until that happens, we should never come here.
1143 * Cache the order of the largest free extent we have available in this block
1147 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
1149 struct ext4_sb_info *sbi = EXT4_SB(sb);
1152 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--)
1153 if (grp->bb_counters[i] > 0)
1155 /* No need to move between order lists? */
1156 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) ||
1157 i == grp->bb_largest_free_order) {
1158 grp->bb_largest_free_order = i;
1162 if (grp->bb_largest_free_order >= 0) {
1163 write_lock(&sbi->s_mb_largest_free_orders_locks[
1164 grp->bb_largest_free_order]);
1165 list_del_init(&grp->bb_largest_free_order_node);
1166 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1167 grp->bb_largest_free_order]);
1169 grp->bb_largest_free_order = i;
1170 if (grp->bb_largest_free_order >= 0 && grp->bb_free) {
1171 write_lock(&sbi->s_mb_largest_free_orders_locks[
1172 grp->bb_largest_free_order]);
1173 list_add_tail(&grp->bb_largest_free_order_node,
1174 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1175 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1176 grp->bb_largest_free_order]);
1180 static noinline_for_stack
1181 void ext4_mb_generate_buddy(struct super_block *sb,
1182 void *buddy, void *bitmap, ext4_group_t group,
1183 struct ext4_group_info *grp)
1185 struct ext4_sb_info *sbi = EXT4_SB(sb);
1186 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1187 ext4_grpblk_t i = 0;
1188 ext4_grpblk_t first;
1191 unsigned fragments = 0;
1192 unsigned long long period = get_cycles();
1194 /* initialize buddy from bitmap which is aggregation
1195 * of on-disk bitmap and preallocations */
1196 i = mb_find_next_zero_bit(bitmap, max, 0);
1197 grp->bb_first_free = i;
1201 i = mb_find_next_bit(bitmap, max, i);
1205 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1207 grp->bb_counters[0]++;
1209 i = mb_find_next_zero_bit(bitmap, max, i);
1211 grp->bb_fragments = fragments;
1213 if (free != grp->bb_free) {
1214 ext4_grp_locked_error(sb, group, 0, 0,
1215 "block bitmap and bg descriptor "
1216 "inconsistent: %u vs %u free clusters",
1217 free, grp->bb_free);
1219 * If we intend to continue, we consider group descriptor
1220 * corrupt and update bb_free using bitmap value
1222 grp->bb_free = free;
1223 ext4_mark_group_bitmap_corrupted(sb, group,
1224 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1226 mb_set_largest_free_order(sb, grp);
1227 mb_update_avg_fragment_size(sb, grp);
1229 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1231 period = get_cycles() - period;
1232 atomic_inc(&sbi->s_mb_buddies_generated);
1233 atomic64_add(period, &sbi->s_mb_generation_time);
1236 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
1242 while ((buddy = mb_find_buddy(e4b, order++, &count)))
1243 mb_set_bits(buddy, 0, count);
1245 e4b->bd_info->bb_fragments = 0;
1246 memset(e4b->bd_info->bb_counters, 0,
1247 sizeof(*e4b->bd_info->bb_counters) *
1248 (e4b->bd_sb->s_blocksize_bits + 2));
1250 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
1251 e4b->bd_bitmap, e4b->bd_group, e4b->bd_info);
1254 /* The buddy information is attached the buddy cache inode
1255 * for convenience. The information regarding each group
1256 * is loaded via ext4_mb_load_buddy. The information involve
1257 * block bitmap and buddy information. The information are
1258 * stored in the inode as
1261 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1264 * one block each for bitmap and buddy information.
1265 * So for each group we take up 2 blocks. A page can
1266 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
1267 * So it can have information regarding groups_per_page which
1268 * is blocks_per_page/2
1270 * Locking note: This routine takes the block group lock of all groups
1271 * for this page; do not hold this lock when calling this routine!
1274 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
1276 ext4_group_t ngroups;
1277 unsigned int blocksize;
1278 int blocks_per_page;
1279 int groups_per_page;
1282 ext4_group_t first_group, group;
1284 struct super_block *sb;
1285 struct buffer_head *bhs;
1286 struct buffer_head **bh = NULL;
1287 struct inode *inode;
1290 struct ext4_group_info *grinfo;
1292 inode = page->mapping->host;
1294 ngroups = ext4_get_groups_count(sb);
1295 blocksize = i_blocksize(inode);
1296 blocks_per_page = PAGE_SIZE / blocksize;
1298 mb_debug(sb, "init page %lu\n", page->index);
1300 groups_per_page = blocks_per_page >> 1;
1301 if (groups_per_page == 0)
1302 groups_per_page = 1;
1304 /* allocate buffer_heads to read bitmaps */
1305 if (groups_per_page > 1) {
1306 i = sizeof(struct buffer_head *) * groups_per_page;
1307 bh = kzalloc(i, gfp);
1313 first_group = page->index * blocks_per_page / 2;
1315 /* read all groups the page covers into the cache */
1316 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1317 if (group >= ngroups)
1320 grinfo = ext4_get_group_info(sb, group);
1324 * If page is uptodate then we came here after online resize
1325 * which added some new uninitialized group info structs, so
1326 * we must skip all initialized uptodate buddies on the page,
1327 * which may be currently in use by an allocating task.
1329 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1333 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1334 if (IS_ERR(bh[i])) {
1335 err = PTR_ERR(bh[i]);
1339 mb_debug(sb, "read bitmap for group %u\n", group);
1342 /* wait for I/O completion */
1343 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1348 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1353 first_block = page->index * blocks_per_page;
1354 for (i = 0; i < blocks_per_page; i++) {
1355 group = (first_block + i) >> 1;
1356 if (group >= ngroups)
1359 if (!bh[group - first_group])
1360 /* skip initialized uptodate buddy */
1363 if (!buffer_verified(bh[group - first_group]))
1364 /* Skip faulty bitmaps */
1369 * data carry information regarding this
1370 * particular group in the format specified
1374 data = page_address(page) + (i * blocksize);
1375 bitmap = bh[group - first_group]->b_data;
1378 * We place the buddy block and bitmap block
1381 grinfo = ext4_get_group_info(sb, group);
1383 err = -EFSCORRUPTED;
1386 if ((first_block + i) & 1) {
1387 /* this is block of buddy */
1388 BUG_ON(incore == NULL);
1389 mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
1390 group, page->index, i * blocksize);
1391 trace_ext4_mb_buddy_bitmap_load(sb, group);
1392 grinfo->bb_fragments = 0;
1393 memset(grinfo->bb_counters, 0,
1394 sizeof(*grinfo->bb_counters) *
1395 (MB_NUM_ORDERS(sb)));
1397 * incore got set to the group block bitmap below
1399 ext4_lock_group(sb, group);
1400 /* init the buddy */
1401 memset(data, 0xff, blocksize);
1402 ext4_mb_generate_buddy(sb, data, incore, group, grinfo);
1403 ext4_unlock_group(sb, group);
1406 /* this is block of bitmap */
1407 BUG_ON(incore != NULL);
1408 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
1409 group, page->index, i * blocksize);
1410 trace_ext4_mb_bitmap_load(sb, group);
1412 /* see comments in ext4_mb_put_pa() */
1413 ext4_lock_group(sb, group);
1414 memcpy(data, bitmap, blocksize);
1416 /* mark all preallocated blks used in in-core bitmap */
1417 ext4_mb_generate_from_pa(sb, data, group);
1418 WARN_ON_ONCE(!RB_EMPTY_ROOT(&grinfo->bb_free_root));
1419 ext4_unlock_group(sb, group);
1421 /* set incore so that the buddy information can be
1422 * generated using this
1427 SetPageUptodate(page);
1431 for (i = 0; i < groups_per_page; i++)
1440 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1441 * on the same buddy page doesn't happen whild holding the buddy page lock.
1442 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1443 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1445 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1446 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1448 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1449 int block, pnum, poff;
1450 int blocks_per_page;
1453 e4b->bd_buddy_page = NULL;
1454 e4b->bd_bitmap_page = NULL;
1456 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1458 * the buddy cache inode stores the block bitmap
1459 * and buddy information in consecutive blocks.
1460 * So for each group we need two blocks.
1463 pnum = block / blocks_per_page;
1464 poff = block % blocks_per_page;
1465 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1468 BUG_ON(page->mapping != inode->i_mapping);
1469 e4b->bd_bitmap_page = page;
1470 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1472 if (blocks_per_page >= 2) {
1473 /* buddy and bitmap are on the same page */
1478 pnum = block / blocks_per_page;
1479 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1482 BUG_ON(page->mapping != inode->i_mapping);
1483 e4b->bd_buddy_page = page;
1487 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1489 if (e4b->bd_bitmap_page) {
1490 unlock_page(e4b->bd_bitmap_page);
1491 put_page(e4b->bd_bitmap_page);
1493 if (e4b->bd_buddy_page) {
1494 unlock_page(e4b->bd_buddy_page);
1495 put_page(e4b->bd_buddy_page);
1500 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1501 * block group lock of all groups for this page; do not hold the BG lock when
1502 * calling this routine!
1504 static noinline_for_stack
1505 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1508 struct ext4_group_info *this_grp;
1509 struct ext4_buddy e4b;
1514 mb_debug(sb, "init group %u\n", group);
1515 this_grp = ext4_get_group_info(sb, group);
1517 return -EFSCORRUPTED;
1520 * This ensures that we don't reinit the buddy cache
1521 * page which map to the group from which we are already
1522 * allocating. If we are looking at the buddy cache we would
1523 * have taken a reference using ext4_mb_load_buddy and that
1524 * would have pinned buddy page to page cache.
1525 * The call to ext4_mb_get_buddy_page_lock will mark the
1528 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1529 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1531 * somebody initialized the group
1532 * return without doing anything
1537 page = e4b.bd_bitmap_page;
1538 ret = ext4_mb_init_cache(page, NULL, gfp);
1541 if (!PageUptodate(page)) {
1546 if (e4b.bd_buddy_page == NULL) {
1548 * If both the bitmap and buddy are in
1549 * the same page we don't need to force
1555 /* init buddy cache */
1556 page = e4b.bd_buddy_page;
1557 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1560 if (!PageUptodate(page)) {
1565 ext4_mb_put_buddy_page_lock(&e4b);
1570 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1571 * block group lock of all groups for this page; do not hold the BG lock when
1572 * calling this routine!
1574 static noinline_for_stack int
1575 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1576 struct ext4_buddy *e4b, gfp_t gfp)
1578 int blocks_per_page;
1584 struct ext4_group_info *grp;
1585 struct ext4_sb_info *sbi = EXT4_SB(sb);
1586 struct inode *inode = sbi->s_buddy_cache;
1589 mb_debug(sb, "load group %u\n", group);
1591 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1592 grp = ext4_get_group_info(sb, group);
1594 return -EFSCORRUPTED;
1596 e4b->bd_blkbits = sb->s_blocksize_bits;
1599 e4b->bd_group = group;
1600 e4b->bd_buddy_page = NULL;
1601 e4b->bd_bitmap_page = NULL;
1603 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1605 * we need full data about the group
1606 * to make a good selection
1608 ret = ext4_mb_init_group(sb, group, gfp);
1614 * the buddy cache inode stores the block bitmap
1615 * and buddy information in consecutive blocks.
1616 * So for each group we need two blocks.
1619 pnum = block / blocks_per_page;
1620 poff = block % blocks_per_page;
1622 /* we could use find_or_create_page(), but it locks page
1623 * what we'd like to avoid in fast path ... */
1624 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1625 if (page == NULL || !PageUptodate(page)) {
1628 * drop the page reference and try
1629 * to get the page with lock. If we
1630 * are not uptodate that implies
1631 * somebody just created the page but
1632 * is yet to initialize the same. So
1633 * wait for it to initialize.
1636 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1638 if (WARN_RATELIMIT(page->mapping != inode->i_mapping,
1639 "ext4: bitmap's paging->mapping != inode->i_mapping\n")) {
1640 /* should never happen */
1645 if (!PageUptodate(page)) {
1646 ret = ext4_mb_init_cache(page, NULL, gfp);
1651 mb_cmp_bitmaps(e4b, page_address(page) +
1652 (poff * sb->s_blocksize));
1661 if (!PageUptodate(page)) {
1666 /* Pages marked accessed already */
1667 e4b->bd_bitmap_page = page;
1668 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1671 pnum = block / blocks_per_page;
1672 poff = block % blocks_per_page;
1674 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1675 if (page == NULL || !PageUptodate(page)) {
1678 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1680 if (WARN_RATELIMIT(page->mapping != inode->i_mapping,
1681 "ext4: buddy bitmap's page->mapping != inode->i_mapping\n")) {
1682 /* should never happen */
1687 if (!PageUptodate(page)) {
1688 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1702 if (!PageUptodate(page)) {
1707 /* Pages marked accessed already */
1708 e4b->bd_buddy_page = page;
1709 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1716 if (e4b->bd_bitmap_page)
1717 put_page(e4b->bd_bitmap_page);
1719 e4b->bd_buddy = NULL;
1720 e4b->bd_bitmap = NULL;
1724 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1725 struct ext4_buddy *e4b)
1727 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1730 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1732 if (e4b->bd_bitmap_page)
1733 put_page(e4b->bd_bitmap_page);
1734 if (e4b->bd_buddy_page)
1735 put_page(e4b->bd_buddy_page);
1739 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1744 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1745 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1747 while (order <= e4b->bd_blkbits + 1) {
1748 bb = mb_find_buddy(e4b, order, &max);
1749 if (!mb_test_bit(block >> order, bb)) {
1750 /* this block is part of buddy of order 'order' */
1758 static void mb_clear_bits(void *bm, int cur, int len)
1764 if ((cur & 31) == 0 && (len - cur) >= 32) {
1765 /* fast path: clear whole word at once */
1766 addr = bm + (cur >> 3);
1771 mb_clear_bit(cur, bm);
1776 /* clear bits in given range
1777 * will return first found zero bit if any, -1 otherwise
1779 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1786 if ((cur & 31) == 0 && (len - cur) >= 32) {
1787 /* fast path: clear whole word at once */
1788 addr = bm + (cur >> 3);
1789 if (*addr != (__u32)(-1) && zero_bit == -1)
1790 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1795 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1803 void mb_set_bits(void *bm, int cur, int len)
1809 if ((cur & 31) == 0 && (len - cur) >= 32) {
1810 /* fast path: set whole word at once */
1811 addr = bm + (cur >> 3);
1816 mb_set_bit(cur, bm);
1821 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1823 if (mb_test_bit(*bit + side, bitmap)) {
1824 mb_clear_bit(*bit, bitmap);
1830 mb_set_bit(*bit, bitmap);
1835 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1839 void *buddy = mb_find_buddy(e4b, order, &max);
1844 /* Bits in range [first; last] are known to be set since
1845 * corresponding blocks were allocated. Bits in range
1846 * (first; last) will stay set because they form buddies on
1847 * upper layer. We just deal with borders if they don't
1848 * align with upper layer and then go up.
1849 * Releasing entire group is all about clearing
1850 * single bit of highest order buddy.
1854 * ---------------------------------
1856 * ---------------------------------
1857 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1858 * ---------------------------------
1860 * \_____________________/
1862 * Neither [1] nor [6] is aligned to above layer.
1863 * Left neighbour [0] is free, so mark it busy,
1864 * decrease bb_counters and extend range to
1866 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1867 * mark [6] free, increase bb_counters and shrink range to
1869 * Then shift range to [0; 2], go up and do the same.
1874 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1876 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1881 buddy2 = mb_find_buddy(e4b, order, &max);
1883 mb_clear_bits(buddy, first, last - first + 1);
1884 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1893 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1894 int first, int count)
1896 int left_is_free = 0;
1897 int right_is_free = 0;
1899 int last = first + count - 1;
1900 struct super_block *sb = e4b->bd_sb;
1902 if (WARN_ON(count == 0))
1904 BUG_ON(last >= (sb->s_blocksize << 3));
1905 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1906 /* Don't bother if the block group is corrupt. */
1907 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1910 mb_check_buddy(e4b);
1911 mb_free_blocks_double(inode, e4b, first, count);
1913 /* access memory sequentially: check left neighbour,
1914 * clear range and then check right neighbour
1917 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1918 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1919 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1920 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1922 if (unlikely(block != -1)) {
1923 struct ext4_sb_info *sbi = EXT4_SB(sb);
1924 ext4_fsblk_t blocknr;
1927 * Fastcommit replay can free already freed blocks which
1928 * corrupts allocation info. Regenerate it.
1930 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
1931 mb_regenerate_buddy(e4b);
1935 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1936 blocknr += EXT4_C2B(sbi, block);
1937 ext4_grp_locked_error(sb, e4b->bd_group,
1938 inode ? inode->i_ino : 0, blocknr,
1939 "freeing already freed block (bit %u); block bitmap corrupt.",
1941 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1942 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1946 this_cpu_inc(discard_pa_seq);
1947 e4b->bd_info->bb_free += count;
1948 if (first < e4b->bd_info->bb_first_free)
1949 e4b->bd_info->bb_first_free = first;
1951 /* let's maintain fragments counter */
1952 if (left_is_free && right_is_free)
1953 e4b->bd_info->bb_fragments--;
1954 else if (!left_is_free && !right_is_free)
1955 e4b->bd_info->bb_fragments++;
1957 /* buddy[0] == bd_bitmap is a special case, so handle
1958 * it right away and let mb_buddy_mark_free stay free of
1959 * zero order checks.
1960 * Check if neighbours are to be coaleasced,
1961 * adjust bitmap bb_counters and borders appropriately.
1964 first += !left_is_free;
1965 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1968 last -= !right_is_free;
1969 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1973 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1975 mb_set_largest_free_order(sb, e4b->bd_info);
1976 mb_update_avg_fragment_size(sb, e4b->bd_info);
1978 mb_check_buddy(e4b);
1981 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1982 int needed, struct ext4_free_extent *ex)
1988 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1991 buddy = mb_find_buddy(e4b, 0, &max);
1992 BUG_ON(buddy == NULL);
1993 BUG_ON(block >= max);
1994 if (mb_test_bit(block, buddy)) {
2001 /* find actual order */
2002 order = mb_find_order_for_block(e4b, block);
2003 block = block >> order;
2005 ex->fe_len = 1 << order;
2006 ex->fe_start = block << order;
2007 ex->fe_group = e4b->bd_group;
2009 /* calc difference from given start */
2010 next = next - ex->fe_start;
2012 ex->fe_start += next;
2014 while (needed > ex->fe_len &&
2015 mb_find_buddy(e4b, order, &max)) {
2017 if (block + 1 >= max)
2020 next = (block + 1) * (1 << order);
2021 if (mb_test_bit(next, e4b->bd_bitmap))
2024 order = mb_find_order_for_block(e4b, next);
2026 block = next >> order;
2027 ex->fe_len += 1 << order;
2030 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
2031 /* Should never happen! (but apparently sometimes does?!?) */
2033 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
2034 "corruption or bug in mb_find_extent "
2035 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
2036 block, order, needed, ex->fe_group, ex->fe_start,
2037 ex->fe_len, ex->fe_logical);
2045 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
2051 int start = ex->fe_start;
2052 int len = ex->fe_len;
2058 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
2059 BUG_ON(e4b->bd_group != ex->fe_group);
2060 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
2061 mb_check_buddy(e4b);
2062 mb_mark_used_double(e4b, start, len);
2064 this_cpu_inc(discard_pa_seq);
2065 e4b->bd_info->bb_free -= len;
2066 if (e4b->bd_info->bb_first_free == start)
2067 e4b->bd_info->bb_first_free += len;
2069 /* let's maintain fragments counter */
2071 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
2072 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
2073 max = !mb_test_bit(start + len, e4b->bd_bitmap);
2075 e4b->bd_info->bb_fragments++;
2076 else if (!mlen && !max)
2077 e4b->bd_info->bb_fragments--;
2079 /* let's maintain buddy itself */
2082 ord = mb_find_order_for_block(e4b, start);
2084 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
2085 /* the whole chunk may be allocated at once! */
2088 buddy = mb_find_buddy(e4b, ord, &max);
2091 BUG_ON((start >> ord) >= max);
2092 mb_set_bit(start >> ord, buddy);
2093 e4b->bd_info->bb_counters[ord]--;
2100 /* store for history */
2102 ret = len | (ord << 16);
2104 /* we have to split large buddy */
2106 buddy = mb_find_buddy(e4b, ord, &max);
2107 mb_set_bit(start >> ord, buddy);
2108 e4b->bd_info->bb_counters[ord]--;
2111 cur = (start >> ord) & ~1U;
2112 buddy = mb_find_buddy(e4b, ord, &max);
2113 mb_clear_bit(cur, buddy);
2114 mb_clear_bit(cur + 1, buddy);
2115 e4b->bd_info->bb_counters[ord]++;
2116 e4b->bd_info->bb_counters[ord]++;
2119 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
2121 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
2122 mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
2123 mb_check_buddy(e4b);
2129 * Must be called under group lock!
2131 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
2132 struct ext4_buddy *e4b)
2134 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2137 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
2138 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2140 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
2141 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
2142 ret = mb_mark_used(e4b, &ac->ac_b_ex);
2144 /* preallocation can change ac_b_ex, thus we store actually
2145 * allocated blocks for history */
2146 ac->ac_f_ex = ac->ac_b_ex;
2148 ac->ac_status = AC_STATUS_FOUND;
2149 ac->ac_tail = ret & 0xffff;
2150 ac->ac_buddy = ret >> 16;
2153 * take the page reference. We want the page to be pinned
2154 * so that we don't get a ext4_mb_init_cache_call for this
2155 * group until we update the bitmap. That would mean we
2156 * double allocate blocks. The reference is dropped
2157 * in ext4_mb_release_context
2159 ac->ac_bitmap_page = e4b->bd_bitmap_page;
2160 get_page(ac->ac_bitmap_page);
2161 ac->ac_buddy_page = e4b->bd_buddy_page;
2162 get_page(ac->ac_buddy_page);
2163 /* store last allocated for subsequent stream allocation */
2164 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2165 spin_lock(&sbi->s_md_lock);
2166 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2167 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2168 spin_unlock(&sbi->s_md_lock);
2171 * As we've just preallocated more space than
2172 * user requested originally, we store allocated
2173 * space in a special descriptor.
2175 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2176 ext4_mb_new_preallocation(ac);
2180 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2181 struct ext4_buddy *e4b,
2184 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2185 struct ext4_free_extent *bex = &ac->ac_b_ex;
2186 struct ext4_free_extent *gex = &ac->ac_g_ex;
2188 if (ac->ac_status == AC_STATUS_FOUND)
2191 * We don't want to scan for a whole year
2193 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2194 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2195 ac->ac_status = AC_STATUS_BREAK;
2200 * Haven't found good chunk so far, let's continue
2202 if (bex->fe_len < gex->fe_len)
2205 if (finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2206 ext4_mb_use_best_found(ac, e4b);
2210 * The routine checks whether found extent is good enough. If it is,
2211 * then the extent gets marked used and flag is set to the context
2212 * to stop scanning. Otherwise, the extent is compared with the
2213 * previous found extent and if new one is better, then it's stored
2214 * in the context. Later, the best found extent will be used, if
2215 * mballoc can't find good enough extent.
2217 * The algorithm used is roughly as follows:
2219 * * If free extent found is exactly as big as goal, then
2220 * stop the scan and use it immediately
2222 * * If free extent found is smaller than goal, then keep retrying
2223 * upto a max of sbi->s_mb_max_to_scan times (default 200). After
2224 * that stop scanning and use whatever we have.
2226 * * If free extent found is bigger than goal, then keep retrying
2227 * upto a max of sbi->s_mb_min_to_scan times (default 10) before
2228 * stopping the scan and using the extent.
2231 * FIXME: real allocation policy is to be designed yet!
2233 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2234 struct ext4_free_extent *ex,
2235 struct ext4_buddy *e4b)
2237 struct ext4_free_extent *bex = &ac->ac_b_ex;
2238 struct ext4_free_extent *gex = &ac->ac_g_ex;
2240 BUG_ON(ex->fe_len <= 0);
2241 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2242 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2243 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2246 ac->ac_cX_found[ac->ac_criteria]++;
2249 * The special case - take what you catch first
2251 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2253 ext4_mb_use_best_found(ac, e4b);
2258 * Let's check whether the chuck is good enough
2260 if (ex->fe_len == gex->fe_len) {
2262 ext4_mb_use_best_found(ac, e4b);
2267 * If this is first found extent, just store it in the context
2269 if (bex->fe_len == 0) {
2275 * If new found extent is better, store it in the context
2277 if (bex->fe_len < gex->fe_len) {
2278 /* if the request isn't satisfied, any found extent
2279 * larger than previous best one is better */
2280 if (ex->fe_len > bex->fe_len)
2282 } else if (ex->fe_len > gex->fe_len) {
2283 /* if the request is satisfied, then we try to find
2284 * an extent that still satisfy the request, but is
2285 * smaller than previous one */
2286 if (ex->fe_len < bex->fe_len)
2290 ext4_mb_check_limits(ac, e4b, 0);
2293 static noinline_for_stack
2294 void ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2295 struct ext4_buddy *e4b)
2297 struct ext4_free_extent ex = ac->ac_b_ex;
2298 ext4_group_t group = ex.fe_group;
2302 BUG_ON(ex.fe_len <= 0);
2303 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2307 ext4_lock_group(ac->ac_sb, group);
2308 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
2311 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2315 ext4_mb_use_best_found(ac, e4b);
2319 ext4_unlock_group(ac->ac_sb, group);
2320 ext4_mb_unload_buddy(e4b);
2323 static noinline_for_stack
2324 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2325 struct ext4_buddy *e4b)
2327 ext4_group_t group = ac->ac_g_ex.fe_group;
2330 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2331 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2332 struct ext4_free_extent ex;
2335 return -EFSCORRUPTED;
2336 if (!(ac->ac_flags & (EXT4_MB_HINT_TRY_GOAL | EXT4_MB_HINT_GOAL_ONLY)))
2338 if (grp->bb_free == 0)
2341 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2345 ext4_lock_group(ac->ac_sb, group);
2346 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
2349 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2350 ac->ac_g_ex.fe_len, &ex);
2351 ex.fe_logical = 0xDEADFA11; /* debug value */
2353 if (max >= ac->ac_g_ex.fe_len &&
2354 ac->ac_g_ex.fe_len == EXT4_B2C(sbi, sbi->s_stripe)) {
2357 start = ext4_grp_offs_to_block(ac->ac_sb, &ex);
2358 /* use do_div to get remainder (would be 64-bit modulo) */
2359 if (do_div(start, sbi->s_stripe) == 0) {
2362 ext4_mb_use_best_found(ac, e4b);
2364 } else if (max >= ac->ac_g_ex.fe_len) {
2365 BUG_ON(ex.fe_len <= 0);
2366 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2367 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2370 ext4_mb_use_best_found(ac, e4b);
2371 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2372 /* Sometimes, caller may want to merge even small
2373 * number of blocks to an existing extent */
2374 BUG_ON(ex.fe_len <= 0);
2375 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2376 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2379 ext4_mb_use_best_found(ac, e4b);
2382 ext4_unlock_group(ac->ac_sb, group);
2383 ext4_mb_unload_buddy(e4b);
2389 * The routine scans buddy structures (not bitmap!) from given order
2390 * to max order and tries to find big enough chunk to satisfy the req
2392 static noinline_for_stack
2393 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2394 struct ext4_buddy *e4b)
2396 struct super_block *sb = ac->ac_sb;
2397 struct ext4_group_info *grp = e4b->bd_info;
2403 BUG_ON(ac->ac_2order <= 0);
2404 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2405 if (grp->bb_counters[i] == 0)
2408 buddy = mb_find_buddy(e4b, i, &max);
2409 if (WARN_RATELIMIT(buddy == NULL,
2410 "ext4: mb_simple_scan_group: mb_find_buddy failed, (%d)\n", i))
2413 k = mb_find_next_zero_bit(buddy, max, 0);
2415 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2416 "%d free clusters of order %d. But found 0",
2417 grp->bb_counters[i], i);
2418 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2420 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2424 ac->ac_cX_found[ac->ac_criteria]++;
2426 ac->ac_b_ex.fe_len = 1 << i;
2427 ac->ac_b_ex.fe_start = k << i;
2428 ac->ac_b_ex.fe_group = e4b->bd_group;
2430 ext4_mb_use_best_found(ac, e4b);
2432 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2434 if (EXT4_SB(sb)->s_mb_stats)
2435 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2442 * The routine scans the group and measures all found extents.
2443 * In order to optimize scanning, caller must pass number of
2444 * free blocks in the group, so the routine can know upper limit.
2446 static noinline_for_stack
2447 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2448 struct ext4_buddy *e4b)
2450 struct super_block *sb = ac->ac_sb;
2451 void *bitmap = e4b->bd_bitmap;
2452 struct ext4_free_extent ex;
2456 free = e4b->bd_info->bb_free;
2457 if (WARN_ON(free <= 0))
2460 i = e4b->bd_info->bb_first_free;
2462 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2463 i = mb_find_next_zero_bit(bitmap,
2464 EXT4_CLUSTERS_PER_GROUP(sb), i);
2465 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2467 * IF we have corrupt bitmap, we won't find any
2468 * free blocks even though group info says we
2471 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2472 "%d free clusters as per "
2473 "group info. But bitmap says 0",
2475 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2476 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2480 if (!ext4_mb_cr_expensive(ac->ac_criteria)) {
2482 * In CR_GOAL_LEN_FAST and CR_BEST_AVAIL_LEN, we are
2483 * sure that this group will have a large enough
2484 * continuous free extent, so skip over the smaller free
2487 j = mb_find_next_bit(bitmap,
2488 EXT4_CLUSTERS_PER_GROUP(sb), i);
2491 if (freelen < ac->ac_g_ex.fe_len) {
2498 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2499 if (WARN_ON(ex.fe_len <= 0))
2501 if (free < ex.fe_len) {
2502 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2503 "%d free clusters as per "
2504 "group info. But got %d blocks",
2506 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2507 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2509 * The number of free blocks differs. This mostly
2510 * indicate that the bitmap is corrupt. So exit
2511 * without claiming the space.
2515 ex.fe_logical = 0xDEADC0DE; /* debug value */
2516 ext4_mb_measure_extent(ac, &ex, e4b);
2522 ext4_mb_check_limits(ac, e4b, 1);
2526 * This is a special case for storages like raid5
2527 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2529 static noinline_for_stack
2530 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2531 struct ext4_buddy *e4b)
2533 struct super_block *sb = ac->ac_sb;
2534 struct ext4_sb_info *sbi = EXT4_SB(sb);
2535 void *bitmap = e4b->bd_bitmap;
2536 struct ext4_free_extent ex;
2537 ext4_fsblk_t first_group_block;
2539 ext4_grpblk_t i, stripe;
2542 BUG_ON(sbi->s_stripe == 0);
2544 /* find first stripe-aligned block in group */
2545 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2547 a = first_group_block + sbi->s_stripe - 1;
2548 do_div(a, sbi->s_stripe);
2549 i = (a * sbi->s_stripe) - first_group_block;
2551 stripe = EXT4_B2C(sbi, sbi->s_stripe);
2552 i = EXT4_B2C(sbi, i);
2553 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2554 if (!mb_test_bit(i, bitmap)) {
2555 max = mb_find_extent(e4b, i, stripe, &ex);
2556 if (max >= stripe) {
2558 ac->ac_cX_found[ac->ac_criteria]++;
2559 ex.fe_logical = 0xDEADF00D; /* debug value */
2561 ext4_mb_use_best_found(ac, e4b);
2570 * This is also called BEFORE we load the buddy bitmap.
2571 * Returns either 1 or 0 indicating that the group is either suitable
2572 * for the allocation or not.
2574 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2575 ext4_group_t group, enum criteria cr)
2577 ext4_grpblk_t free, fragments;
2578 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2579 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2581 BUG_ON(cr < CR_POWER2_ALIGNED || cr >= EXT4_MB_NUM_CRS);
2583 if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2586 free = grp->bb_free;
2590 fragments = grp->bb_fragments;
2595 case CR_POWER2_ALIGNED:
2596 BUG_ON(ac->ac_2order == 0);
2598 /* Avoid using the first bg of a flexgroup for data files */
2599 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2600 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2601 ((group % flex_size) == 0))
2604 if (free < ac->ac_g_ex.fe_len)
2607 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2610 if (grp->bb_largest_free_order < ac->ac_2order)
2614 case CR_GOAL_LEN_FAST:
2615 case CR_BEST_AVAIL_LEN:
2616 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2619 case CR_GOAL_LEN_SLOW:
2620 if (free >= ac->ac_g_ex.fe_len)
2633 * This could return negative error code if something goes wrong
2634 * during ext4_mb_init_group(). This should not be called with
2635 * ext4_lock_group() held.
2637 * Note: because we are conditionally operating with the group lock in
2638 * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this
2639 * function using __acquire and __release. This means we need to be
2640 * super careful before messing with the error path handling via "goto
2643 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2644 ext4_group_t group, enum criteria cr)
2646 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2647 struct super_block *sb = ac->ac_sb;
2648 struct ext4_sb_info *sbi = EXT4_SB(sb);
2649 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2654 return -EFSCORRUPTED;
2655 if (sbi->s_mb_stats)
2656 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2658 ext4_lock_group(sb, group);
2659 __release(ext4_group_lock_ptr(sb, group));
2661 free = grp->bb_free;
2665 * In all criterias except CR_ANY_FREE we try to avoid groups that
2666 * can't possibly satisfy the full goal request due to insufficient
2669 if (cr < CR_ANY_FREE && free < ac->ac_g_ex.fe_len)
2671 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2674 __acquire(ext4_group_lock_ptr(sb, group));
2675 ext4_unlock_group(sb, group);
2678 /* We only do this if the grp has never been initialized */
2679 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2680 struct ext4_group_desc *gdp =
2681 ext4_get_group_desc(sb, group, NULL);
2685 * cr=CR_POWER2_ALIGNED/CR_GOAL_LEN_FAST is a very optimistic
2686 * search to find large good chunks almost for free. If buddy
2687 * data is not ready, then this optimization makes no sense. But
2688 * we never skip the first block group in a flex_bg, since this
2689 * gets used for metadata block allocation, and we want to make
2690 * sure we locate metadata blocks in the first block group in
2691 * the flex_bg if possible.
2693 if (!ext4_mb_cr_expensive(cr) &&
2694 (!sbi->s_log_groups_per_flex ||
2695 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2696 !(ext4_has_group_desc_csum(sb) &&
2697 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2699 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2705 ext4_lock_group(sb, group);
2706 __release(ext4_group_lock_ptr(sb, group));
2708 ret = ext4_mb_good_group(ac, group, cr);
2711 __acquire(ext4_group_lock_ptr(sb, group));
2712 ext4_unlock_group(sb, group);
2718 * Start prefetching @nr block bitmaps starting at @group.
2719 * Return the next group which needs to be prefetched.
2721 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2722 unsigned int nr, int *cnt)
2724 ext4_group_t ngroups = ext4_get_groups_count(sb);
2725 struct buffer_head *bh;
2726 struct blk_plug plug;
2728 blk_start_plug(&plug);
2730 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2732 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2735 * Prefetch block groups with free blocks; but don't
2736 * bother if it is marked uninitialized on disk, since
2737 * it won't require I/O to read. Also only try to
2738 * prefetch once, so we avoid getblk() call, which can
2741 if (gdp && grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2742 EXT4_MB_GRP_NEED_INIT(grp) &&
2743 ext4_free_group_clusters(sb, gdp) > 0 ) {
2744 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2745 if (bh && !IS_ERR(bh)) {
2746 if (!buffer_uptodate(bh) && cnt)
2751 if (++group >= ngroups)
2754 blk_finish_plug(&plug);
2759 * Prefetching reads the block bitmap into the buffer cache; but we
2760 * need to make sure that the buddy bitmap in the page cache has been
2761 * initialized. Note that ext4_mb_init_group() will block if the I/O
2762 * is not yet completed, or indeed if it was not initiated by
2763 * ext4_mb_prefetch did not start the I/O.
2765 * TODO: We should actually kick off the buddy bitmap setup in a work
2766 * queue when the buffer I/O is completed, so that we don't block
2767 * waiting for the block allocation bitmap read to finish when
2768 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2770 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2773 struct ext4_group_desc *gdp;
2774 struct ext4_group_info *grp;
2778 group = ext4_get_groups_count(sb);
2780 gdp = ext4_get_group_desc(sb, group, NULL);
2781 grp = ext4_get_group_info(sb, group);
2783 if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) &&
2784 ext4_free_group_clusters(sb, gdp) > 0) {
2785 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2791 static noinline_for_stack int
2792 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2794 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2795 enum criteria new_cr, cr = CR_GOAL_LEN_FAST;
2796 int err = 0, first_err = 0;
2797 unsigned int nr = 0, prefetch_ios = 0;
2798 struct ext4_sb_info *sbi;
2799 struct super_block *sb;
2800 struct ext4_buddy e4b;
2805 ngroups = ext4_get_groups_count(sb);
2806 /* non-extent files are limited to low blocks/groups */
2807 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2808 ngroups = sbi->s_blockfile_groups;
2810 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2812 /* first, try the goal */
2813 err = ext4_mb_find_by_goal(ac, &e4b);
2814 if (err || ac->ac_status == AC_STATUS_FOUND)
2817 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2821 * ac->ac_2order is set only if the fe_len is a power of 2
2822 * if ac->ac_2order is set we also set criteria to CR_POWER2_ALIGNED
2823 * so that we try exact allocation using buddy.
2825 i = fls(ac->ac_g_ex.fe_len);
2828 * We search using buddy data only if the order of the request
2829 * is greater than equal to the sbi_s_mb_order2_reqs
2830 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2831 * We also support searching for power-of-two requests only for
2832 * requests upto maximum buddy size we have constructed.
2834 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2835 if (is_power_of_2(ac->ac_g_ex.fe_len))
2836 ac->ac_2order = array_index_nospec(i - 1,
2840 /* if stream allocation is enabled, use global goal */
2841 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2842 /* TBD: may be hot point */
2843 spin_lock(&sbi->s_md_lock);
2844 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2845 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2846 spin_unlock(&sbi->s_md_lock);
2850 * Let's just scan groups to find more-less suitable blocks We
2851 * start with CR_GOAL_LEN_FAST, unless it is power of 2
2852 * aligned, in which case let's do that faster approach first.
2855 cr = CR_POWER2_ALIGNED;
2857 for (; cr < EXT4_MB_NUM_CRS && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2858 ac->ac_criteria = cr;
2860 * searching for the right group start
2861 * from the goal value specified
2863 group = ac->ac_g_ex.fe_group;
2864 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2865 prefetch_grp = group;
2867 for (i = 0, new_cr = cr; i < ngroups; i++,
2868 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) {
2878 * Batch reads of the block allocation bitmaps
2879 * to get multiple READs in flight; limit
2880 * prefetching at inexpensive CR, otherwise mballoc
2881 * can spend a lot of time loading imperfect groups
2883 if ((prefetch_grp == group) &&
2884 (ext4_mb_cr_expensive(cr) ||
2885 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2886 nr = sbi->s_mb_prefetch;
2887 if (ext4_has_feature_flex_bg(sb)) {
2888 nr = 1 << sbi->s_log_groups_per_flex;
2889 nr -= group & (nr - 1);
2890 nr = min(nr, sbi->s_mb_prefetch);
2892 prefetch_grp = ext4_mb_prefetch(sb, group,
2896 /* This now checks without needing the buddy page */
2897 ret = ext4_mb_good_group_nolock(ac, group, cr);
2904 err = ext4_mb_load_buddy(sb, group, &e4b);
2908 ext4_lock_group(sb, group);
2911 * We need to check again after locking the
2914 ret = ext4_mb_good_group(ac, group, cr);
2916 ext4_unlock_group(sb, group);
2917 ext4_mb_unload_buddy(&e4b);
2921 ac->ac_groups_scanned++;
2922 if (cr == CR_POWER2_ALIGNED)
2923 ext4_mb_simple_scan_group(ac, &e4b);
2924 else if ((cr == CR_GOAL_LEN_FAST ||
2925 cr == CR_BEST_AVAIL_LEN) &&
2927 !(ac->ac_g_ex.fe_len %
2928 EXT4_B2C(sbi, sbi->s_stripe)))
2929 ext4_mb_scan_aligned(ac, &e4b);
2931 ext4_mb_complex_scan_group(ac, &e4b);
2933 ext4_unlock_group(sb, group);
2934 ext4_mb_unload_buddy(&e4b);
2936 if (ac->ac_status != AC_STATUS_CONTINUE)
2939 /* Processed all groups and haven't found blocks */
2940 if (sbi->s_mb_stats && i == ngroups)
2941 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2943 if (i == ngroups && ac->ac_criteria == CR_BEST_AVAIL_LEN)
2944 /* Reset goal length to original goal length before
2945 * falling into CR_GOAL_LEN_SLOW */
2946 ac->ac_g_ex.fe_len = ac->ac_orig_goal_len;
2949 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2950 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2952 * We've been searching too long. Let's try to allocate
2953 * the best chunk we've found so far
2955 ext4_mb_try_best_found(ac, &e4b);
2956 if (ac->ac_status != AC_STATUS_FOUND) {
2958 * Someone more lucky has already allocated it.
2959 * The only thing we can do is just take first
2962 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2963 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2964 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2965 ac->ac_b_ex.fe_len, lost);
2967 ac->ac_b_ex.fe_group = 0;
2968 ac->ac_b_ex.fe_start = 0;
2969 ac->ac_b_ex.fe_len = 0;
2970 ac->ac_status = AC_STATUS_CONTINUE;
2971 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2977 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2978 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2980 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2983 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2984 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2985 ac->ac_flags, cr, err);
2988 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2993 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2995 struct super_block *sb = pde_data(file_inode(seq->file));
2998 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
3001 return (void *) ((unsigned long) group);
3004 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
3006 struct super_block *sb = pde_data(file_inode(seq->file));
3010 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
3013 return (void *) ((unsigned long) group);
3016 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
3018 struct super_block *sb = pde_data(file_inode(seq->file));
3019 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
3021 int err, buddy_loaded = 0;
3022 struct ext4_buddy e4b;
3023 struct ext4_group_info *grinfo;
3024 unsigned char blocksize_bits = min_t(unsigned char,
3025 sb->s_blocksize_bits,
3026 EXT4_MAX_BLOCK_LOG_SIZE);
3028 struct ext4_group_info info;
3029 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
3034 seq_puts(seq, "#group: free frags first ["
3035 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
3036 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
3038 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
3039 sizeof(struct ext4_group_info);
3041 grinfo = ext4_get_group_info(sb, group);
3044 /* Load the group info in memory only if not already loaded. */
3045 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
3046 err = ext4_mb_load_buddy(sb, group, &e4b);
3048 seq_printf(seq, "#%-5u: I/O error\n", group);
3054 memcpy(&sg, grinfo, i);
3057 ext4_mb_unload_buddy(&e4b);
3059 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
3060 sg.info.bb_fragments, sg.info.bb_first_free);
3061 for (i = 0; i <= 13; i++)
3062 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
3063 sg.info.bb_counters[i] : 0);
3064 seq_puts(seq, " ]\n");
3069 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
3073 const struct seq_operations ext4_mb_seq_groups_ops = {
3074 .start = ext4_mb_seq_groups_start,
3075 .next = ext4_mb_seq_groups_next,
3076 .stop = ext4_mb_seq_groups_stop,
3077 .show = ext4_mb_seq_groups_show,
3080 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
3082 struct super_block *sb = seq->private;
3083 struct ext4_sb_info *sbi = EXT4_SB(sb);
3085 seq_puts(seq, "mballoc:\n");
3086 if (!sbi->s_mb_stats) {
3087 seq_puts(seq, "\tmb stats collection turned off.\n");
3090 "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
3093 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
3094 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
3096 seq_printf(seq, "\tgroups_scanned: %u\n",
3097 atomic_read(&sbi->s_bal_groups_scanned));
3099 /* CR_POWER2_ALIGNED stats */
3100 seq_puts(seq, "\tcr_p2_aligned_stats:\n");
3101 seq_printf(seq, "\t\thits: %llu\n",
3102 atomic64_read(&sbi->s_bal_cX_hits[CR_POWER2_ALIGNED]));
3104 seq, "\t\tgroups_considered: %llu\n",
3106 &sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]));
3107 seq_printf(seq, "\t\textents_scanned: %u\n",
3108 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_POWER2_ALIGNED]));
3109 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3110 atomic64_read(&sbi->s_bal_cX_failed[CR_POWER2_ALIGNED]));
3111 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3112 atomic_read(&sbi->s_bal_p2_aligned_bad_suggestions));
3114 /* CR_GOAL_LEN_FAST stats */
3115 seq_puts(seq, "\tcr_goal_fast_stats:\n");
3116 seq_printf(seq, "\t\thits: %llu\n",
3117 atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_FAST]));
3118 seq_printf(seq, "\t\tgroups_considered: %llu\n",
3120 &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_FAST]));
3121 seq_printf(seq, "\t\textents_scanned: %u\n",
3122 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_FAST]));
3123 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3124 atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_FAST]));
3125 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3126 atomic_read(&sbi->s_bal_goal_fast_bad_suggestions));
3128 /* CR_BEST_AVAIL_LEN stats */
3129 seq_puts(seq, "\tcr_best_avail_stats:\n");
3130 seq_printf(seq, "\t\thits: %llu\n",
3131 atomic64_read(&sbi->s_bal_cX_hits[CR_BEST_AVAIL_LEN]));
3133 seq, "\t\tgroups_considered: %llu\n",
3135 &sbi->s_bal_cX_groups_considered[CR_BEST_AVAIL_LEN]));
3136 seq_printf(seq, "\t\textents_scanned: %u\n",
3137 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_BEST_AVAIL_LEN]));
3138 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3139 atomic64_read(&sbi->s_bal_cX_failed[CR_BEST_AVAIL_LEN]));
3140 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3141 atomic_read(&sbi->s_bal_best_avail_bad_suggestions));
3143 /* CR_GOAL_LEN_SLOW stats */
3144 seq_puts(seq, "\tcr_goal_slow_stats:\n");
3145 seq_printf(seq, "\t\thits: %llu\n",
3146 atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_SLOW]));
3147 seq_printf(seq, "\t\tgroups_considered: %llu\n",
3149 &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_SLOW]));
3150 seq_printf(seq, "\t\textents_scanned: %u\n",
3151 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_SLOW]));
3152 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3153 atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_SLOW]));
3155 /* CR_ANY_FREE stats */
3156 seq_puts(seq, "\tcr_any_free_stats:\n");
3157 seq_printf(seq, "\t\thits: %llu\n",
3158 atomic64_read(&sbi->s_bal_cX_hits[CR_ANY_FREE]));
3160 seq, "\t\tgroups_considered: %llu\n",
3161 atomic64_read(&sbi->s_bal_cX_groups_considered[CR_ANY_FREE]));
3162 seq_printf(seq, "\t\textents_scanned: %u\n",
3163 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_ANY_FREE]));
3164 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3165 atomic64_read(&sbi->s_bal_cX_failed[CR_ANY_FREE]));
3168 seq_printf(seq, "\textents_scanned: %u\n",
3169 atomic_read(&sbi->s_bal_ex_scanned));
3170 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
3171 seq_printf(seq, "\t\tlen_goal_hits: %u\n",
3172 atomic_read(&sbi->s_bal_len_goals));
3173 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
3174 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
3175 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
3176 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
3177 atomic_read(&sbi->s_mb_buddies_generated),
3178 ext4_get_groups_count(sb));
3179 seq_printf(seq, "\tbuddies_time_used: %llu\n",
3180 atomic64_read(&sbi->s_mb_generation_time));
3181 seq_printf(seq, "\tpreallocated: %u\n",
3182 atomic_read(&sbi->s_mb_preallocated));
3183 seq_printf(seq, "\tdiscarded: %u\n", atomic_read(&sbi->s_mb_discarded));
3187 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
3188 __acquires(&EXT4_SB(sb)->s_mb_rb_lock)
3190 struct super_block *sb = pde_data(file_inode(seq->file));
3191 unsigned long position;
3193 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
3195 position = *pos + 1;
3196 return (void *) ((unsigned long) position);
3199 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
3201 struct super_block *sb = pde_data(file_inode(seq->file));
3202 unsigned long position;
3205 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
3207 position = *pos + 1;
3208 return (void *) ((unsigned long) position);
3211 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
3213 struct super_block *sb = pde_data(file_inode(seq->file));
3214 struct ext4_sb_info *sbi = EXT4_SB(sb);
3215 unsigned long position = ((unsigned long) v);
3216 struct ext4_group_info *grp;
3220 if (position >= MB_NUM_ORDERS(sb)) {
3221 position -= MB_NUM_ORDERS(sb);
3223 seq_puts(seq, "avg_fragment_size_lists:\n");
3226 read_lock(&sbi->s_mb_avg_fragment_size_locks[position]);
3227 list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position],
3228 bb_avg_fragment_size_node)
3230 read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]);
3231 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3232 (unsigned int)position, count);
3236 if (position == 0) {
3237 seq_printf(seq, "optimize_scan: %d\n",
3238 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
3239 seq_puts(seq, "max_free_order_lists:\n");
3242 read_lock(&sbi->s_mb_largest_free_orders_locks[position]);
3243 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
3244 bb_largest_free_order_node)
3246 read_unlock(&sbi->s_mb_largest_free_orders_locks[position]);
3247 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3248 (unsigned int)position, count);
3253 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3257 const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3258 .start = ext4_mb_seq_structs_summary_start,
3259 .next = ext4_mb_seq_structs_summary_next,
3260 .stop = ext4_mb_seq_structs_summary_stop,
3261 .show = ext4_mb_seq_structs_summary_show,
3264 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3266 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3267 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3274 * Allocate the top-level s_group_info array for the specified number
3277 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3279 struct ext4_sb_info *sbi = EXT4_SB(sb);
3281 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3283 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3284 EXT4_DESC_PER_BLOCK_BITS(sb);
3285 if (size <= sbi->s_group_info_size)
3288 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3289 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3290 if (!new_groupinfo) {
3291 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3295 old_groupinfo = rcu_dereference(sbi->s_group_info);
3297 memcpy(new_groupinfo, old_groupinfo,
3298 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3300 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3301 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3303 ext4_kvfree_array_rcu(old_groupinfo);
3304 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3305 sbi->s_group_info_size);
3309 /* Create and initialize ext4_group_info data for the given group. */
3310 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3311 struct ext4_group_desc *desc)
3315 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3316 struct ext4_sb_info *sbi = EXT4_SB(sb);
3317 struct ext4_group_info **meta_group_info;
3318 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3321 * First check if this group is the first of a reserved block.
3322 * If it's true, we have to allocate a new table of pointers
3323 * to ext4_group_info structures
3325 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3326 metalen = sizeof(*meta_group_info) <<
3327 EXT4_DESC_PER_BLOCK_BITS(sb);
3328 meta_group_info = kmalloc(metalen, GFP_NOFS);
3329 if (meta_group_info == NULL) {
3330 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3331 "for a buddy group");
3335 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3339 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3340 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3342 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3343 if (meta_group_info[i] == NULL) {
3344 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3345 goto exit_group_info;
3347 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3348 &(meta_group_info[i]->bb_state));
3351 * initialize bb_free to be able to skip
3352 * empty groups without initialization
3354 if (ext4_has_group_desc_csum(sb) &&
3355 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3356 meta_group_info[i]->bb_free =
3357 ext4_free_clusters_after_init(sb, group, desc);
3359 meta_group_info[i]->bb_free =
3360 ext4_free_group_clusters(sb, desc);
3363 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3364 init_rwsem(&meta_group_info[i]->alloc_sem);
3365 meta_group_info[i]->bb_free_root = RB_ROOT;
3366 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3367 INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node);
3368 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3369 meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */
3370 meta_group_info[i]->bb_group = group;
3372 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3376 /* If a meta_group_info table has been allocated, release it now */
3377 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3378 struct ext4_group_info ***group_info;
3381 group_info = rcu_dereference(sbi->s_group_info);
3382 kfree(group_info[idx]);
3383 group_info[idx] = NULL;
3387 } /* ext4_mb_add_groupinfo */
3389 static int ext4_mb_init_backend(struct super_block *sb)
3391 ext4_group_t ngroups = ext4_get_groups_count(sb);
3393 struct ext4_sb_info *sbi = EXT4_SB(sb);
3395 struct ext4_group_desc *desc;
3396 struct ext4_group_info ***group_info;
3397 struct kmem_cache *cachep;
3399 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3403 sbi->s_buddy_cache = new_inode(sb);
3404 if (sbi->s_buddy_cache == NULL) {
3405 ext4_msg(sb, KERN_ERR, "can't get new inode");
3408 /* To avoid potentially colliding with an valid on-disk inode number,
3409 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3410 * not in the inode hash, so it should never be found by iget(), but
3411 * this will avoid confusion if it ever shows up during debugging. */
3412 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3413 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3414 for (i = 0; i < ngroups; i++) {
3416 desc = ext4_get_group_desc(sb, i, NULL);
3418 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3421 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3425 if (ext4_has_feature_flex_bg(sb)) {
3426 /* a single flex group is supposed to be read by a single IO.
3427 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3428 * unsigned integer, so the maximum shift is 32.
3430 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3431 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3434 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3435 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3436 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3438 sbi->s_mb_prefetch = 32;
3440 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3441 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3442 /* now many real IOs to prefetch within a single allocation at cr=0
3443 * given cr=0 is an CPU-related optimization we shouldn't try to
3444 * load too many groups, at some point we should start to use what
3445 * we've got in memory.
3446 * with an average random access time 5ms, it'd take a second to get
3447 * 200 groups (* N with flex_bg), so let's make this limit 4
3449 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3450 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3451 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3456 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3458 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3461 kmem_cache_free(cachep, grp);
3463 i = sbi->s_group_info_size;
3465 group_info = rcu_dereference(sbi->s_group_info);
3467 kfree(group_info[i]);
3469 iput(sbi->s_buddy_cache);
3472 kvfree(rcu_dereference(sbi->s_group_info));
3477 static void ext4_groupinfo_destroy_slabs(void)
3481 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3482 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3483 ext4_groupinfo_caches[i] = NULL;
3487 static int ext4_groupinfo_create_slab(size_t size)
3489 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3491 int blocksize_bits = order_base_2(size);
3492 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3493 struct kmem_cache *cachep;
3495 if (cache_index >= NR_GRPINFO_CACHES)
3498 if (unlikely(cache_index < 0))
3501 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3502 if (ext4_groupinfo_caches[cache_index]) {
3503 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3504 return 0; /* Already created */
3507 slab_size = offsetof(struct ext4_group_info,
3508 bb_counters[blocksize_bits + 2]);
3510 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3511 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3514 ext4_groupinfo_caches[cache_index] = cachep;
3516 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3519 "EXT4-fs: no memory for groupinfo slab cache\n");
3526 static void ext4_discard_work(struct work_struct *work)
3528 struct ext4_sb_info *sbi = container_of(work,
3529 struct ext4_sb_info, s_discard_work);
3530 struct super_block *sb = sbi->s_sb;
3531 struct ext4_free_data *fd, *nfd;
3532 struct ext4_buddy e4b;
3533 LIST_HEAD(discard_list);
3534 ext4_group_t grp, load_grp;
3537 spin_lock(&sbi->s_md_lock);
3538 list_splice_init(&sbi->s_discard_list, &discard_list);
3539 spin_unlock(&sbi->s_md_lock);
3541 load_grp = UINT_MAX;
3542 list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
3544 * If filesystem is umounting or no memory or suffering
3545 * from no space, give up the discard
3547 if ((sb->s_flags & SB_ACTIVE) && !err &&
3548 !atomic_read(&sbi->s_retry_alloc_pending)) {
3549 grp = fd->efd_group;
3550 if (grp != load_grp) {
3551 if (load_grp != UINT_MAX)
3552 ext4_mb_unload_buddy(&e4b);
3554 err = ext4_mb_load_buddy(sb, grp, &e4b);
3556 kmem_cache_free(ext4_free_data_cachep, fd);
3557 load_grp = UINT_MAX;
3564 ext4_lock_group(sb, grp);
3565 ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
3566 fd->efd_start_cluster + fd->efd_count - 1, 1);
3567 ext4_unlock_group(sb, grp);
3569 kmem_cache_free(ext4_free_data_cachep, fd);
3572 if (load_grp != UINT_MAX)
3573 ext4_mb_unload_buddy(&e4b);
3576 int ext4_mb_init(struct super_block *sb)
3578 struct ext4_sb_info *sbi = EXT4_SB(sb);
3580 unsigned offset, offset_incr;
3584 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3586 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3587 if (sbi->s_mb_offsets == NULL) {
3592 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3593 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3594 if (sbi->s_mb_maxs == NULL) {
3599 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3603 /* order 0 is regular bitmap */
3604 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3605 sbi->s_mb_offsets[0] = 0;
3609 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3610 max = sb->s_blocksize << 2;
3612 sbi->s_mb_offsets[i] = offset;
3613 sbi->s_mb_maxs[i] = max;
3614 offset += offset_incr;
3615 offset_incr = offset_incr >> 1;
3618 } while (i < MB_NUM_ORDERS(sb));
3620 sbi->s_mb_avg_fragment_size =
3621 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3623 if (!sbi->s_mb_avg_fragment_size) {
3627 sbi->s_mb_avg_fragment_size_locks =
3628 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3630 if (!sbi->s_mb_avg_fragment_size_locks) {
3634 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3635 INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]);
3636 rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]);
3638 sbi->s_mb_largest_free_orders =
3639 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3641 if (!sbi->s_mb_largest_free_orders) {
3645 sbi->s_mb_largest_free_orders_locks =
3646 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3648 if (!sbi->s_mb_largest_free_orders_locks) {
3652 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3653 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3654 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3657 spin_lock_init(&sbi->s_md_lock);
3658 sbi->s_mb_free_pending = 0;
3659 INIT_LIST_HEAD(&sbi->s_freed_data_list[0]);
3660 INIT_LIST_HEAD(&sbi->s_freed_data_list[1]);
3661 INIT_LIST_HEAD(&sbi->s_discard_list);
3662 INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
3663 atomic_set(&sbi->s_retry_alloc_pending, 0);
3665 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3666 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3667 sbi->s_mb_stats = MB_DEFAULT_STATS;
3668 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3669 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3670 sbi->s_mb_best_avail_max_trim_order = MB_DEFAULT_BEST_AVAIL_TRIM_ORDER;
3673 * The default group preallocation is 512, which for 4k block
3674 * sizes translates to 2 megabytes. However for bigalloc file
3675 * systems, this is probably too big (i.e, if the cluster size
3676 * is 1 megabyte, then group preallocation size becomes half a
3677 * gigabyte!). As a default, we will keep a two megabyte
3678 * group pralloc size for cluster sizes up to 64k, and after
3679 * that, we will force a minimum group preallocation size of
3680 * 32 clusters. This translates to 8 megs when the cluster
3681 * size is 256k, and 32 megs when the cluster size is 1 meg,
3682 * which seems reasonable as a default.
3684 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3685 sbi->s_cluster_bits, 32);
3687 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3688 * to the lowest multiple of s_stripe which is bigger than
3689 * the s_mb_group_prealloc as determined above. We want
3690 * the preallocation size to be an exact multiple of the
3691 * RAID stripe size so that preallocations don't fragment
3694 if (sbi->s_stripe > 1) {
3695 sbi->s_mb_group_prealloc = roundup(
3696 sbi->s_mb_group_prealloc, EXT4_B2C(sbi, sbi->s_stripe));
3699 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3700 if (sbi->s_locality_groups == NULL) {
3704 for_each_possible_cpu(i) {
3705 struct ext4_locality_group *lg;
3706 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3707 mutex_init(&lg->lg_mutex);
3708 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3709 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3710 spin_lock_init(&lg->lg_prealloc_lock);
3713 if (bdev_nonrot(sb->s_bdev))
3714 sbi->s_mb_max_linear_groups = 0;
3716 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3717 /* init file for buddy data */
3718 ret = ext4_mb_init_backend(sb);
3720 goto out_free_locality_groups;
3724 out_free_locality_groups:
3725 free_percpu(sbi->s_locality_groups);
3726 sbi->s_locality_groups = NULL;
3728 kfree(sbi->s_mb_avg_fragment_size);
3729 kfree(sbi->s_mb_avg_fragment_size_locks);
3730 kfree(sbi->s_mb_largest_free_orders);
3731 kfree(sbi->s_mb_largest_free_orders_locks);
3732 kfree(sbi->s_mb_offsets);
3733 sbi->s_mb_offsets = NULL;
3734 kfree(sbi->s_mb_maxs);
3735 sbi->s_mb_maxs = NULL;
3739 /* need to called with the ext4 group lock held */
3740 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3742 struct ext4_prealloc_space *pa;
3743 struct list_head *cur, *tmp;
3746 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3747 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3748 list_del(&pa->pa_group_list);
3750 kmem_cache_free(ext4_pspace_cachep, pa);
3755 int ext4_mb_release(struct super_block *sb)
3757 ext4_group_t ngroups = ext4_get_groups_count(sb);
3759 int num_meta_group_infos;
3760 struct ext4_group_info *grinfo, ***group_info;
3761 struct ext4_sb_info *sbi = EXT4_SB(sb);
3762 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3765 if (test_opt(sb, DISCARD)) {
3767 * wait the discard work to drain all of ext4_free_data
3769 flush_work(&sbi->s_discard_work);
3770 WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
3773 if (sbi->s_group_info) {
3774 for (i = 0; i < ngroups; i++) {
3776 grinfo = ext4_get_group_info(sb, i);
3779 mb_group_bb_bitmap_free(grinfo);
3780 ext4_lock_group(sb, i);
3781 count = ext4_mb_cleanup_pa(grinfo);
3783 mb_debug(sb, "mballoc: %d PAs left\n",
3785 ext4_unlock_group(sb, i);
3786 kmem_cache_free(cachep, grinfo);
3788 num_meta_group_infos = (ngroups +
3789 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3790 EXT4_DESC_PER_BLOCK_BITS(sb);
3792 group_info = rcu_dereference(sbi->s_group_info);
3793 for (i = 0; i < num_meta_group_infos; i++)
3794 kfree(group_info[i]);
3798 kfree(sbi->s_mb_avg_fragment_size);
3799 kfree(sbi->s_mb_avg_fragment_size_locks);
3800 kfree(sbi->s_mb_largest_free_orders);
3801 kfree(sbi->s_mb_largest_free_orders_locks);
3802 kfree(sbi->s_mb_offsets);
3803 kfree(sbi->s_mb_maxs);
3804 iput(sbi->s_buddy_cache);
3805 if (sbi->s_mb_stats) {
3806 ext4_msg(sb, KERN_INFO,
3807 "mballoc: %u blocks %u reqs (%u success)",
3808 atomic_read(&sbi->s_bal_allocated),
3809 atomic_read(&sbi->s_bal_reqs),
3810 atomic_read(&sbi->s_bal_success));
3811 ext4_msg(sb, KERN_INFO,
3812 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3813 "%u 2^N hits, %u breaks, %u lost",
3814 atomic_read(&sbi->s_bal_ex_scanned),
3815 atomic_read(&sbi->s_bal_groups_scanned),
3816 atomic_read(&sbi->s_bal_goals),
3817 atomic_read(&sbi->s_bal_2orders),
3818 atomic_read(&sbi->s_bal_breaks),
3819 atomic_read(&sbi->s_mb_lost_chunks));
3820 ext4_msg(sb, KERN_INFO,
3821 "mballoc: %u generated and it took %llu",
3822 atomic_read(&sbi->s_mb_buddies_generated),
3823 atomic64_read(&sbi->s_mb_generation_time));
3824 ext4_msg(sb, KERN_INFO,
3825 "mballoc: %u preallocated, %u discarded",
3826 atomic_read(&sbi->s_mb_preallocated),
3827 atomic_read(&sbi->s_mb_discarded));
3830 free_percpu(sbi->s_locality_groups);
3835 static inline int ext4_issue_discard(struct super_block *sb,
3836 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3839 ext4_fsblk_t discard_block;
3841 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3842 ext4_group_first_block_no(sb, block_group));
3843 count = EXT4_C2B(EXT4_SB(sb), count);
3844 trace_ext4_discard_blocks(sb,
3845 (unsigned long long) discard_block, count);
3847 return __blkdev_issue_discard(sb->s_bdev,
3848 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3849 (sector_t)count << (sb->s_blocksize_bits - 9),
3852 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3855 static void ext4_free_data_in_buddy(struct super_block *sb,
3856 struct ext4_free_data *entry)
3858 struct ext4_buddy e4b;
3859 struct ext4_group_info *db;
3862 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3863 entry->efd_count, entry->efd_group, entry);
3865 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3866 /* we expect to find existing buddy because it's pinned */
3869 spin_lock(&EXT4_SB(sb)->s_md_lock);
3870 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3871 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3874 /* there are blocks to put in buddy to make them really free */
3875 count += entry->efd_count;
3876 ext4_lock_group(sb, entry->efd_group);
3877 /* Take it out of per group rb tree */
3878 rb_erase(&entry->efd_node, &(db->bb_free_root));
3879 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3882 * Clear the trimmed flag for the group so that the next
3883 * ext4_trim_fs can trim it.
3884 * If the volume is mounted with -o discard, online discard
3885 * is supported and the free blocks will be trimmed online.
3887 if (!test_opt(sb, DISCARD))
3888 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3890 if (!db->bb_free_root.rb_node) {
3891 /* No more items in the per group rb tree
3892 * balance refcounts from ext4_mb_free_metadata()
3894 put_page(e4b.bd_buddy_page);
3895 put_page(e4b.bd_bitmap_page);
3897 ext4_unlock_group(sb, entry->efd_group);
3898 ext4_mb_unload_buddy(&e4b);
3900 mb_debug(sb, "freed %d blocks in 1 structures\n", count);
3904 * This function is called by the jbd2 layer once the commit has finished,
3905 * so we know we can free the blocks that were released with that commit.
3907 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3909 struct ext4_sb_info *sbi = EXT4_SB(sb);
3910 struct ext4_free_data *entry, *tmp;
3911 LIST_HEAD(freed_data_list);
3912 struct list_head *s_freed_head = &sbi->s_freed_data_list[commit_tid & 1];
3915 list_replace_init(s_freed_head, &freed_data_list);
3917 list_for_each_entry(entry, &freed_data_list, efd_list)
3918 ext4_free_data_in_buddy(sb, entry);
3920 if (test_opt(sb, DISCARD)) {
3921 spin_lock(&sbi->s_md_lock);
3922 wake = list_empty(&sbi->s_discard_list);
3923 list_splice_tail(&freed_data_list, &sbi->s_discard_list);
3924 spin_unlock(&sbi->s_md_lock);
3926 queue_work(system_unbound_wq, &sbi->s_discard_work);
3928 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3929 kmem_cache_free(ext4_free_data_cachep, entry);
3933 int __init ext4_init_mballoc(void)
3935 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3936 SLAB_RECLAIM_ACCOUNT);
3937 if (ext4_pspace_cachep == NULL)
3940 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3941 SLAB_RECLAIM_ACCOUNT);
3942 if (ext4_ac_cachep == NULL)
3945 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3946 SLAB_RECLAIM_ACCOUNT);
3947 if (ext4_free_data_cachep == NULL)
3953 kmem_cache_destroy(ext4_ac_cachep);
3955 kmem_cache_destroy(ext4_pspace_cachep);
3960 void ext4_exit_mballoc(void)
3963 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3964 * before destroying the slab cache.
3967 kmem_cache_destroy(ext4_pspace_cachep);
3968 kmem_cache_destroy(ext4_ac_cachep);
3969 kmem_cache_destroy(ext4_free_data_cachep);
3970 ext4_groupinfo_destroy_slabs();
3973 #define EXT4_MB_BITMAP_MARKED_CHECK 0x0001
3974 #define EXT4_MB_SYNC_UPDATE 0x0002
3976 ext4_mb_mark_context(handle_t *handle, struct super_block *sb, bool state,
3977 ext4_group_t group, ext4_grpblk_t blkoff,
3978 ext4_grpblk_t len, int flags, ext4_grpblk_t *ret_changed)
3980 struct ext4_sb_info *sbi = EXT4_SB(sb);
3981 struct buffer_head *bitmap_bh = NULL;
3982 struct ext4_group_desc *gdp;
3983 struct buffer_head *gdp_bh;
3985 unsigned int i, already, changed = len;
3987 KUNIT_STATIC_STUB_REDIRECT(ext4_mb_mark_context,
3988 handle, sb, state, group, blkoff, len,
3989 flags, ret_changed);
3993 bitmap_bh = ext4_read_block_bitmap(sb, group);
3994 if (IS_ERR(bitmap_bh))
3995 return PTR_ERR(bitmap_bh);
3998 BUFFER_TRACE(bitmap_bh, "getting write access");
3999 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
4006 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
4011 BUFFER_TRACE(gdp_bh, "get_write_access");
4012 err = ext4_journal_get_write_access(handle, sb, gdp_bh,
4018 ext4_lock_group(sb, group);
4019 if (ext4_has_group_desc_csum(sb) &&
4020 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
4021 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
4022 ext4_free_group_clusters_set(sb, gdp,
4023 ext4_free_clusters_after_init(sb, group, gdp));
4026 if (flags & EXT4_MB_BITMAP_MARKED_CHECK) {
4028 for (i = 0; i < len; i++)
4029 if (mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
4032 changed = len - already;
4036 mb_set_bits(bitmap_bh->b_data, blkoff, len);
4037 ext4_free_group_clusters_set(sb, gdp,
4038 ext4_free_group_clusters(sb, gdp) - changed);
4040 mb_clear_bits(bitmap_bh->b_data, blkoff, len);
4041 ext4_free_group_clusters_set(sb, gdp,
4042 ext4_free_group_clusters(sb, gdp) + changed);
4045 ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh);
4046 ext4_group_desc_csum_set(sb, group, gdp);
4047 ext4_unlock_group(sb, group);
4049 *ret_changed = changed;
4051 if (sbi->s_log_groups_per_flex) {
4052 ext4_group_t flex_group = ext4_flex_group(sbi, group);
4053 struct flex_groups *fg = sbi_array_rcu_deref(sbi,
4054 s_flex_groups, flex_group);
4057 atomic64_sub(changed, &fg->free_clusters);
4059 atomic64_add(changed, &fg->free_clusters);
4062 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4065 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
4069 if (flags & EXT4_MB_SYNC_UPDATE) {
4070 sync_dirty_buffer(bitmap_bh);
4071 sync_dirty_buffer(gdp_bh);
4080 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
4081 * Returns 0 if success or error code
4083 static noinline_for_stack int
4084 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
4085 handle_t *handle, unsigned int reserv_clstrs)
4087 struct ext4_group_desc *gdp;
4088 struct ext4_sb_info *sbi;
4089 struct super_block *sb;
4093 ext4_grpblk_t changed;
4095 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4096 BUG_ON(ac->ac_b_ex.fe_len <= 0);
4101 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, NULL);
4104 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
4105 ext4_free_group_clusters(sb, gdp));
4107 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4108 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4109 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
4110 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
4111 "fs metadata", block, block+len);
4112 /* File system mounted not to panic on error
4113 * Fix the bitmap and return EFSCORRUPTED
4114 * We leak some of the blocks here.
4116 err = ext4_mb_mark_context(handle, sb, true,
4117 ac->ac_b_ex.fe_group,
4118 ac->ac_b_ex.fe_start,
4122 err = -EFSCORRUPTED;
4126 #ifdef AGGRESSIVE_CHECK
4127 flags |= EXT4_MB_BITMAP_MARKED_CHECK;
4129 err = ext4_mb_mark_context(handle, sb, true, ac->ac_b_ex.fe_group,
4130 ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len,
4133 if (err && changed == 0)
4136 #ifdef AGGRESSIVE_CHECK
4137 BUG_ON(changed != ac->ac_b_ex.fe_len);
4139 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
4141 * Now reduce the dirty block count also. Should not go negative
4143 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
4144 /* release all the reserved blocks if non delalloc */
4145 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4152 * Idempotent helper for Ext4 fast commit replay path to set the state of
4153 * blocks in bitmaps and update counters.
4155 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
4156 int len, bool state)
4158 struct ext4_sb_info *sbi = EXT4_SB(sb);
4160 ext4_grpblk_t blkoff;
4162 unsigned int clen, thisgrp_len;
4165 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
4168 * Check to see if we are freeing blocks across a group
4170 * In case of flex_bg, this can happen that (block, len) may
4171 * span across more than one group. In that case we need to
4172 * get the corresponding group metadata to work with.
4173 * For this we have goto again loop.
4175 thisgrp_len = min_t(unsigned int, (unsigned int)len,
4176 EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
4177 clen = EXT4_NUM_B2C(sbi, thisgrp_len);
4179 if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
4180 ext4_error(sb, "Marking blocks in system zone - "
4181 "Block = %llu, len = %u",
4182 block, thisgrp_len);
4186 err = ext4_mb_mark_context(NULL, sb, state,
4187 group, blkoff, clen,
4188 EXT4_MB_BITMAP_MARKED_CHECK |
4189 EXT4_MB_SYNC_UPDATE,
4194 block += thisgrp_len;
4201 * here we normalize request for locality group
4202 * Group request are normalized to s_mb_group_prealloc, which goes to
4203 * s_strip if we set the same via mount option.
4204 * s_mb_group_prealloc can be configured via
4205 * /sys/fs/ext4/<partition>/mb_group_prealloc
4207 * XXX: should we try to preallocate more than the group has now?
4209 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
4211 struct super_block *sb = ac->ac_sb;
4212 struct ext4_locality_group *lg = ac->ac_lg;
4215 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
4216 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
4220 * This function returns the next element to look at during inode
4221 * PA rbtree walk. We assume that we have held the inode PA rbtree lock
4222 * (ei->i_prealloc_lock)
4224 * new_start The start of the range we want to compare
4225 * cur_start The existing start that we are comparing against
4226 * node The node of the rb_tree
4228 static inline struct rb_node*
4229 ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node)
4231 if (new_start < cur_start)
4232 return node->rb_left;
4234 return node->rb_right;
4238 ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac,
4239 ext4_lblk_t start, loff_t end)
4241 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4242 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4243 struct ext4_prealloc_space *tmp_pa;
4244 ext4_lblk_t tmp_pa_start;
4246 struct rb_node *iter;
4248 read_lock(&ei->i_prealloc_lock);
4249 for (iter = ei->i_prealloc_node.rb_node; iter;
4250 iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) {
4251 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4252 pa_node.inode_node);
4253 tmp_pa_start = tmp_pa->pa_lstart;
4254 tmp_pa_end = pa_logical_end(sbi, tmp_pa);
4256 spin_lock(&tmp_pa->pa_lock);
4257 if (tmp_pa->pa_deleted == 0)
4258 BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start));
4259 spin_unlock(&tmp_pa->pa_lock);
4261 read_unlock(&ei->i_prealloc_lock);
4265 * Given an allocation context "ac" and a range "start", "end", check
4266 * and adjust boundaries if the range overlaps with any of the existing
4267 * preallocatoins stored in the corresponding inode of the allocation context.
4270 * ac allocation context
4271 * start start of the new range
4272 * end end of the new range
4275 ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac,
4276 ext4_lblk_t *start, loff_t *end)
4278 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4279 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4280 struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL;
4281 struct rb_node *iter;
4282 ext4_lblk_t new_start, tmp_pa_start, right_pa_start = -1;
4283 loff_t new_end, tmp_pa_end, left_pa_end = -1;
4289 * Adjust the normalized range so that it doesn't overlap with any
4290 * existing preallocated blocks(PAs). Make sure to hold the rbtree lock
4291 * so it doesn't change underneath us.
4293 read_lock(&ei->i_prealloc_lock);
4295 /* Step 1: find any one immediate neighboring PA of the normalized range */
4296 for (iter = ei->i_prealloc_node.rb_node; iter;
4297 iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical,
4298 tmp_pa_start, iter)) {
4299 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4300 pa_node.inode_node);
4301 tmp_pa_start = tmp_pa->pa_lstart;
4302 tmp_pa_end = pa_logical_end(sbi, tmp_pa);
4304 /* PA must not overlap original request */
4305 spin_lock(&tmp_pa->pa_lock);
4306 if (tmp_pa->pa_deleted == 0)
4307 BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end ||
4308 ac->ac_o_ex.fe_logical < tmp_pa_start));
4309 spin_unlock(&tmp_pa->pa_lock);
4313 * Step 2: check if the found PA is left or right neighbor and
4314 * get the other neighbor
4317 if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) {
4318 struct rb_node *tmp;
4321 tmp = rb_next(&left_pa->pa_node.inode_node);
4323 right_pa = rb_entry(tmp,
4324 struct ext4_prealloc_space,
4325 pa_node.inode_node);
4328 struct rb_node *tmp;
4331 tmp = rb_prev(&right_pa->pa_node.inode_node);
4333 left_pa = rb_entry(tmp,
4334 struct ext4_prealloc_space,
4335 pa_node.inode_node);
4340 /* Step 3: get the non deleted neighbors */
4342 for (iter = &left_pa->pa_node.inode_node;;
4343 iter = rb_prev(iter)) {
4349 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4350 pa_node.inode_node);
4352 spin_lock(&tmp_pa->pa_lock);
4353 if (tmp_pa->pa_deleted == 0) {
4354 spin_unlock(&tmp_pa->pa_lock);
4357 spin_unlock(&tmp_pa->pa_lock);
4362 for (iter = &right_pa->pa_node.inode_node;;
4363 iter = rb_next(iter)) {
4369 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4370 pa_node.inode_node);
4372 spin_lock(&tmp_pa->pa_lock);
4373 if (tmp_pa->pa_deleted == 0) {
4374 spin_unlock(&tmp_pa->pa_lock);
4377 spin_unlock(&tmp_pa->pa_lock);
4382 left_pa_end = pa_logical_end(sbi, left_pa);
4383 BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical);
4387 right_pa_start = right_pa->pa_lstart;
4388 BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical);
4391 /* Step 4: trim our normalized range to not overlap with the neighbors */
4393 if (left_pa_end > new_start)
4394 new_start = left_pa_end;
4398 if (right_pa_start < new_end)
4399 new_end = right_pa_start;
4401 read_unlock(&ei->i_prealloc_lock);
4403 /* XXX: extra loop to check we really don't overlap preallocations */
4404 ext4_mb_pa_assert_overlap(ac, new_start, new_end);
4411 * Normalization means making request better in terms of
4412 * size and alignment
4414 static noinline_for_stack void
4415 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
4416 struct ext4_allocation_request *ar)
4418 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4419 struct ext4_super_block *es = sbi->s_es;
4421 loff_t size, start_off, end;
4422 loff_t orig_size __maybe_unused;
4425 /* do normalize only data requests, metadata requests
4426 do not need preallocation */
4427 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4430 /* sometime caller may want exact blocks */
4431 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4434 /* caller may indicate that preallocation isn't
4435 * required (it's a tail, for example) */
4436 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
4439 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
4440 ext4_mb_normalize_group_request(ac);
4444 bsbits = ac->ac_sb->s_blocksize_bits;
4446 /* first, let's learn actual file size
4447 * given current request is allocated */
4448 size = extent_logical_end(sbi, &ac->ac_o_ex);
4449 size = size << bsbits;
4450 if (size < i_size_read(ac->ac_inode))
4451 size = i_size_read(ac->ac_inode);
4454 /* max size of free chunks */
4457 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
4458 (req <= (size) || max <= (chunk_size))
4460 /* first, try to predict filesize */
4461 /* XXX: should this table be tunable? */
4463 if (size <= 16 * 1024) {
4465 } else if (size <= 32 * 1024) {
4467 } else if (size <= 64 * 1024) {
4469 } else if (size <= 128 * 1024) {
4471 } else if (size <= 256 * 1024) {
4473 } else if (size <= 512 * 1024) {
4475 } else if (size <= 1024 * 1024) {
4477 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
4478 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4479 (21 - bsbits)) << 21;
4480 size = 2 * 1024 * 1024;
4481 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
4482 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4483 (22 - bsbits)) << 22;
4484 size = 4 * 1024 * 1024;
4485 } else if (NRL_CHECK_SIZE(EXT4_C2B(sbi, ac->ac_o_ex.fe_len),
4486 (8<<20)>>bsbits, max, 8 * 1024)) {
4487 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4488 (23 - bsbits)) << 23;
4489 size = 8 * 1024 * 1024;
4491 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
4492 size = (loff_t) EXT4_C2B(sbi,
4493 ac->ac_o_ex.fe_len) << bsbits;
4495 size = size >> bsbits;
4496 start = start_off >> bsbits;
4499 * For tiny groups (smaller than 8MB) the chosen allocation
4500 * alignment may be larger than group size. Make sure the
4501 * alignment does not move allocation to a different group which
4502 * makes mballoc fail assertions later.
4504 start = max(start, rounddown(ac->ac_o_ex.fe_logical,
4505 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
4507 /* avoid unnecessary preallocation that may trigger assertions */
4508 if (start + size > EXT_MAX_BLOCKS)
4509 size = EXT_MAX_BLOCKS - start;
4511 /* don't cover already allocated blocks in selected range */
4512 if (ar->pleft && start <= ar->lleft) {
4513 size -= ar->lleft + 1 - start;
4514 start = ar->lleft + 1;
4516 if (ar->pright && start + size - 1 >= ar->lright)
4517 size -= start + size - ar->lright;
4520 * Trim allocation request for filesystems with artificially small
4523 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4524 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4528 ext4_mb_pa_adjust_overlap(ac, &start, &end);
4533 * In this function "start" and "size" are normalized for better
4534 * alignment and length such that we could preallocate more blocks.
4535 * This normalization is done such that original request of
4536 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
4537 * "size" boundaries.
4538 * (Note fe_len can be relaxed since FS block allocation API does not
4539 * provide gurantee on number of contiguous blocks allocation since that
4540 * depends upon free space left, etc).
4541 * In case of inode pa, later we use the allocated blocks
4542 * [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated
4543 * range of goal/best blocks [start, size] to put it at the
4544 * ac_o_ex.fe_logical extent of this inode.
4545 * (See ext4_mb_use_inode_pa() for more details)
4547 if (start + size <= ac->ac_o_ex.fe_logical ||
4548 start > ac->ac_o_ex.fe_logical) {
4549 ext4_msg(ac->ac_sb, KERN_ERR,
4550 "start %lu, size %lu, fe_logical %lu",
4551 (unsigned long) start, (unsigned long) size,
4552 (unsigned long) ac->ac_o_ex.fe_logical);
4555 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4557 /* now prepare goal request */
4559 /* XXX: is it better to align blocks WRT to logical
4560 * placement or satisfy big request as is */
4561 ac->ac_g_ex.fe_logical = start;
4562 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4563 ac->ac_orig_goal_len = ac->ac_g_ex.fe_len;
4565 /* define goal start in order to merge */
4566 if (ar->pright && (ar->lright == (start + size)) &&
4567 ar->pright >= size &&
4568 ar->pright - size >= le32_to_cpu(es->s_first_data_block)) {
4569 /* merge to the right */
4570 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4571 &ac->ac_g_ex.fe_group,
4572 &ac->ac_g_ex.fe_start);
4573 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4575 if (ar->pleft && (ar->lleft + 1 == start) &&
4576 ar->pleft + 1 < ext4_blocks_count(es)) {
4577 /* merge to the left */
4578 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4579 &ac->ac_g_ex.fe_group,
4580 &ac->ac_g_ex.fe_start);
4581 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4584 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4588 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4590 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4592 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4593 atomic_inc(&sbi->s_bal_reqs);
4594 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4595 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4596 atomic_inc(&sbi->s_bal_success);
4598 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4599 for (int i=0; i<EXT4_MB_NUM_CRS; i++) {
4600 atomic_add(ac->ac_cX_found[i], &sbi->s_bal_cX_ex_scanned[i]);
4603 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4604 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4605 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4606 atomic_inc(&sbi->s_bal_goals);
4607 /* did we allocate as much as normalizer originally wanted? */
4608 if (ac->ac_f_ex.fe_len == ac->ac_orig_goal_len)
4609 atomic_inc(&sbi->s_bal_len_goals);
4611 if (ac->ac_found > sbi->s_mb_max_to_scan)
4612 atomic_inc(&sbi->s_bal_breaks);
4615 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4616 trace_ext4_mballoc_alloc(ac);
4618 trace_ext4_mballoc_prealloc(ac);
4622 * Called on failure; free up any blocks from the inode PA for this
4623 * context. We don't need this for MB_GROUP_PA because we only change
4624 * pa_free in ext4_mb_release_context(), but on failure, we've already
4625 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4627 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4629 struct ext4_prealloc_space *pa = ac->ac_pa;
4630 struct ext4_buddy e4b;
4634 if (ac->ac_f_ex.fe_len == 0)
4636 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4637 if (WARN_RATELIMIT(err,
4638 "ext4: mb_load_buddy failed (%d)", err))
4640 * This should never happen since we pin the
4641 * pages in the ext4_allocation_context so
4642 * ext4_mb_load_buddy() should never fail.
4645 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4646 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4647 ac->ac_f_ex.fe_len);
4648 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4649 ext4_mb_unload_buddy(&e4b);
4652 if (pa->pa_type == MB_INODE_PA) {
4653 spin_lock(&pa->pa_lock);
4654 pa->pa_free += ac->ac_b_ex.fe_len;
4655 spin_unlock(&pa->pa_lock);
4660 * use blocks preallocated to inode
4662 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4663 struct ext4_prealloc_space *pa)
4665 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4670 /* found preallocated blocks, use them */
4671 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4672 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4673 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4674 len = EXT4_NUM_B2C(sbi, end - start);
4675 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4676 &ac->ac_b_ex.fe_start);
4677 ac->ac_b_ex.fe_len = len;
4678 ac->ac_status = AC_STATUS_FOUND;
4681 BUG_ON(start < pa->pa_pstart);
4682 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4683 BUG_ON(pa->pa_free < len);
4684 BUG_ON(ac->ac_b_ex.fe_len <= 0);
4687 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4691 * use blocks preallocated to locality group
4693 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4694 struct ext4_prealloc_space *pa)
4696 unsigned int len = ac->ac_o_ex.fe_len;
4698 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4699 &ac->ac_b_ex.fe_group,
4700 &ac->ac_b_ex.fe_start);
4701 ac->ac_b_ex.fe_len = len;
4702 ac->ac_status = AC_STATUS_FOUND;
4705 /* we don't correct pa_pstart or pa_len here to avoid
4706 * possible race when the group is being loaded concurrently
4707 * instead we correct pa later, after blocks are marked
4708 * in on-disk bitmap -- see ext4_mb_release_context()
4709 * Other CPUs are prevented from allocating from this pa by lg_mutex
4711 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4712 pa->pa_lstart, len, pa);
4716 * Return the prealloc space that have minimal distance
4717 * from the goal block. @cpa is the prealloc
4718 * space that is having currently known minimal distance
4719 * from the goal block.
4721 static struct ext4_prealloc_space *
4722 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4723 struct ext4_prealloc_space *pa,
4724 struct ext4_prealloc_space *cpa)
4726 ext4_fsblk_t cur_distance, new_distance;
4729 atomic_inc(&pa->pa_count);
4732 cur_distance = abs(goal_block - cpa->pa_pstart);
4733 new_distance = abs(goal_block - pa->pa_pstart);
4735 if (cur_distance <= new_distance)
4738 /* drop the previous reference */
4739 atomic_dec(&cpa->pa_count);
4740 atomic_inc(&pa->pa_count);
4745 * check if found pa meets EXT4_MB_HINT_GOAL_ONLY
4748 ext4_mb_pa_goal_check(struct ext4_allocation_context *ac,
4749 struct ext4_prealloc_space *pa)
4751 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4754 if (likely(!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)))
4758 * If EXT4_MB_HINT_GOAL_ONLY is set, ac_g_ex will not be adjusted
4759 * in ext4_mb_normalize_request and will keep same with ac_o_ex
4760 * from ext4_mb_initialize_context. Choose ac_g_ex here to keep
4761 * consistent with ext4_mb_find_by_goal.
4763 start = pa->pa_pstart +
4764 (ac->ac_g_ex.fe_logical - pa->pa_lstart);
4765 if (ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex) != start)
4768 if (ac->ac_g_ex.fe_len > pa->pa_len -
4769 EXT4_B2C(sbi, ac->ac_g_ex.fe_logical - pa->pa_lstart))
4776 * search goal blocks in preallocated space
4778 static noinline_for_stack bool
4779 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4781 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4783 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4784 struct ext4_locality_group *lg;
4785 struct ext4_prealloc_space *tmp_pa = NULL, *cpa = NULL;
4786 struct rb_node *iter;
4787 ext4_fsblk_t goal_block;
4789 /* only data can be preallocated */
4790 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4794 * first, try per-file preallocation by searching the inode pa rbtree.
4796 * Here, we can't do a direct traversal of the tree because
4797 * ext4_mb_discard_group_preallocation() can paralelly mark the pa
4798 * deleted and that can cause direct traversal to skip some entries.
4800 read_lock(&ei->i_prealloc_lock);
4802 if (RB_EMPTY_ROOT(&ei->i_prealloc_node)) {
4807 * Step 1: Find a pa with logical start immediately adjacent to the
4808 * original logical start. This could be on the left or right.
4810 * (tmp_pa->pa_lstart never changes so we can skip locking for it).
4812 for (iter = ei->i_prealloc_node.rb_node; iter;
4813 iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical,
4814 tmp_pa->pa_lstart, iter)) {
4815 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4816 pa_node.inode_node);
4820 * Step 2: The adjacent pa might be to the right of logical start, find
4821 * the left adjacent pa. After this step we'd have a valid tmp_pa whose
4822 * logical start is towards the left of original request's logical start
4824 if (tmp_pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4825 struct rb_node *tmp;
4826 tmp = rb_prev(&tmp_pa->pa_node.inode_node);
4829 tmp_pa = rb_entry(tmp, struct ext4_prealloc_space,
4830 pa_node.inode_node);
4833 * If there is no adjacent pa to the left then finding
4834 * an overlapping pa is not possible hence stop searching
4841 BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical));
4844 * Step 3: If the left adjacent pa is deleted, keep moving left to find
4845 * the first non deleted adjacent pa. After this step we should have a
4846 * valid tmp_pa which is guaranteed to be non deleted.
4848 for (iter = &tmp_pa->pa_node.inode_node;; iter = rb_prev(iter)) {
4851 * no non deleted left adjacent pa, so stop searching
4856 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4857 pa_node.inode_node);
4858 spin_lock(&tmp_pa->pa_lock);
4859 if (tmp_pa->pa_deleted == 0) {
4861 * We will keep holding the pa_lock from
4862 * this point on because we don't want group discard
4863 * to delete this pa underneath us. Since group
4864 * discard is anyways an ENOSPC operation it
4865 * should be okay for it to wait a few more cycles.
4869 spin_unlock(&tmp_pa->pa_lock);
4873 BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical));
4874 BUG_ON(tmp_pa->pa_deleted == 1);
4877 * Step 4: We now have the non deleted left adjacent pa. Only this
4878 * pa can possibly satisfy the request hence check if it overlaps
4879 * original logical start and stop searching if it doesn't.
4881 if (ac->ac_o_ex.fe_logical >= pa_logical_end(sbi, tmp_pa)) {
4882 spin_unlock(&tmp_pa->pa_lock);
4886 /* non-extent files can't have physical blocks past 2^32 */
4887 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4888 (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) >
4889 EXT4_MAX_BLOCK_FILE_PHYS)) {
4891 * Since PAs don't overlap, we won't find any other PA to
4894 spin_unlock(&tmp_pa->pa_lock);
4898 if (tmp_pa->pa_free && likely(ext4_mb_pa_goal_check(ac, tmp_pa))) {
4899 atomic_inc(&tmp_pa->pa_count);
4900 ext4_mb_use_inode_pa(ac, tmp_pa);
4901 spin_unlock(&tmp_pa->pa_lock);
4902 read_unlock(&ei->i_prealloc_lock);
4906 * We found a valid overlapping pa but couldn't use it because
4907 * it had no free blocks. This should ideally never happen
4910 * 1. When a new inode pa is added to rbtree it must have
4911 * pa_free > 0 since otherwise we won't actually need
4914 * 2. An inode pa that is in the rbtree can only have it's
4915 * pa_free become zero when another thread calls:
4916 * ext4_mb_new_blocks
4917 * ext4_mb_use_preallocated
4918 * ext4_mb_use_inode_pa
4920 * 3. Further, after the above calls make pa_free == 0, we will
4921 * immediately remove it from the rbtree in:
4922 * ext4_mb_new_blocks
4923 * ext4_mb_release_context
4926 * 4. Since the pa_free becoming 0 and pa_free getting removed
4927 * from tree both happen in ext4_mb_new_blocks, which is always
4928 * called with i_data_sem held for data allocations, we can be
4929 * sure that another process will never see a pa in rbtree with
4932 WARN_ON_ONCE(tmp_pa->pa_free == 0);
4934 spin_unlock(&tmp_pa->pa_lock);
4936 read_unlock(&ei->i_prealloc_lock);
4938 /* can we use group allocation? */
4939 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4942 /* inode may have no locality group for some reason */
4946 order = fls(ac->ac_o_ex.fe_len) - 1;
4947 if (order > PREALLOC_TB_SIZE - 1)
4948 /* The max size of hash table is PREALLOC_TB_SIZE */
4949 order = PREALLOC_TB_SIZE - 1;
4951 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4953 * search for the prealloc space that is having
4954 * minimal distance from the goal block.
4956 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4958 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i],
4960 spin_lock(&tmp_pa->pa_lock);
4961 if (tmp_pa->pa_deleted == 0 &&
4962 tmp_pa->pa_free >= ac->ac_o_ex.fe_len) {
4964 cpa = ext4_mb_check_group_pa(goal_block,
4967 spin_unlock(&tmp_pa->pa_lock);
4972 ext4_mb_use_group_pa(ac, cpa);
4979 * the function goes through all preallocation in this group and marks them
4980 * used in in-core bitmap. buddy must be generated from this bitmap
4981 * Need to be called with ext4 group lock held
4983 static noinline_for_stack
4984 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4987 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4988 struct ext4_prealloc_space *pa;
4989 struct list_head *cur;
4990 ext4_group_t groupnr;
4991 ext4_grpblk_t start;
4992 int preallocated = 0;
4998 /* all form of preallocation discards first load group,
4999 * so the only competing code is preallocation use.
5000 * we don't need any locking here
5001 * notice we do NOT ignore preallocations with pa_deleted
5002 * otherwise we could leave used blocks available for
5003 * allocation in buddy when concurrent ext4_mb_put_pa()
5004 * is dropping preallocation
5006 list_for_each(cur, &grp->bb_prealloc_list) {
5007 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
5008 spin_lock(&pa->pa_lock);
5009 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5012 spin_unlock(&pa->pa_lock);
5013 if (unlikely(len == 0))
5015 BUG_ON(groupnr != group);
5016 mb_set_bits(bitmap, start, len);
5017 preallocated += len;
5019 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
5022 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
5023 struct ext4_prealloc_space *pa)
5025 struct ext4_inode_info *ei;
5027 if (pa->pa_deleted) {
5028 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
5029 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
5036 if (pa->pa_type == MB_INODE_PA) {
5037 ei = EXT4_I(pa->pa_inode);
5038 atomic_dec(&ei->i_prealloc_active);
5042 static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa)
5045 BUG_ON(atomic_read(&pa->pa_count));
5046 BUG_ON(pa->pa_deleted == 0);
5047 kmem_cache_free(ext4_pspace_cachep, pa);
5050 static void ext4_mb_pa_callback(struct rcu_head *head)
5052 struct ext4_prealloc_space *pa;
5054 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
5055 ext4_mb_pa_free(pa);
5059 * drops a reference to preallocated space descriptor
5060 * if this was the last reference and the space is consumed
5062 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
5063 struct super_block *sb, struct ext4_prealloc_space *pa)
5066 ext4_fsblk_t grp_blk;
5067 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
5069 /* in this short window concurrent discard can set pa_deleted */
5070 spin_lock(&pa->pa_lock);
5071 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
5072 spin_unlock(&pa->pa_lock);
5076 if (pa->pa_deleted == 1) {
5077 spin_unlock(&pa->pa_lock);
5081 ext4_mb_mark_pa_deleted(sb, pa);
5082 spin_unlock(&pa->pa_lock);
5084 grp_blk = pa->pa_pstart;
5086 * If doing group-based preallocation, pa_pstart may be in the
5087 * next group when pa is used up
5089 if (pa->pa_type == MB_GROUP_PA)
5092 grp = ext4_get_group_number(sb, grp_blk);
5097 * P1 (buddy init) P2 (regular allocation)
5098 * find block B in PA
5099 * copy on-disk bitmap to buddy
5100 * mark B in on-disk bitmap
5101 * drop PA from group
5102 * mark all PAs in buddy
5104 * thus, P1 initializes buddy with B available. to prevent this
5105 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
5108 ext4_lock_group(sb, grp);
5109 list_del(&pa->pa_group_list);
5110 ext4_unlock_group(sb, grp);
5112 if (pa->pa_type == MB_INODE_PA) {
5113 write_lock(pa->pa_node_lock.inode_lock);
5114 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5115 write_unlock(pa->pa_node_lock.inode_lock);
5116 ext4_mb_pa_free(pa);
5118 spin_lock(pa->pa_node_lock.lg_lock);
5119 list_del_rcu(&pa->pa_node.lg_list);
5120 spin_unlock(pa->pa_node_lock.lg_lock);
5121 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5125 static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new)
5127 struct rb_node **iter = &root->rb_node, *parent = NULL;
5128 struct ext4_prealloc_space *iter_pa, *new_pa;
5129 ext4_lblk_t iter_start, new_start;
5132 iter_pa = rb_entry(*iter, struct ext4_prealloc_space,
5133 pa_node.inode_node);
5134 new_pa = rb_entry(new, struct ext4_prealloc_space,
5135 pa_node.inode_node);
5136 iter_start = iter_pa->pa_lstart;
5137 new_start = new_pa->pa_lstart;
5140 if (new_start < iter_start)
5141 iter = &((*iter)->rb_left);
5143 iter = &((*iter)->rb_right);
5146 rb_link_node(new, parent, iter);
5147 rb_insert_color(new, root);
5151 * creates new preallocated space for given inode
5153 static noinline_for_stack void
5154 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
5156 struct super_block *sb = ac->ac_sb;
5157 struct ext4_sb_info *sbi = EXT4_SB(sb);
5158 struct ext4_prealloc_space *pa;
5159 struct ext4_group_info *grp;
5160 struct ext4_inode_info *ei;
5162 /* preallocate only when found space is larger then requested */
5163 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
5164 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
5165 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
5166 BUG_ON(ac->ac_pa == NULL);
5170 if (ac->ac_b_ex.fe_len < ac->ac_orig_goal_len) {
5171 struct ext4_free_extent ex = {
5172 .fe_logical = ac->ac_g_ex.fe_logical,
5173 .fe_len = ac->ac_orig_goal_len,
5175 loff_t orig_goal_end = extent_logical_end(sbi, &ex);
5177 /* we can't allocate as much as normalizer wants.
5178 * so, found space must get proper lstart
5179 * to cover original request */
5180 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
5181 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
5184 * Use the below logic for adjusting best extent as it keeps
5185 * fragmentation in check while ensuring logical range of best
5186 * extent doesn't overflow out of goal extent:
5188 * 1. Check if best ex can be kept at end of goal (before
5189 * cr_best_avail trimmed it) and still cover original start
5190 * 2. Else, check if best ex can be kept at start of goal and
5191 * still cover original start
5192 * 3. Else, keep the best ex at start of original request.
5194 ex.fe_len = ac->ac_b_ex.fe_len;
5196 ex.fe_logical = orig_goal_end - EXT4_C2B(sbi, ex.fe_len);
5197 if (ac->ac_o_ex.fe_logical >= ex.fe_logical)
5200 ex.fe_logical = ac->ac_g_ex.fe_logical;
5201 if (ac->ac_o_ex.fe_logical < extent_logical_end(sbi, &ex))
5204 ex.fe_logical = ac->ac_o_ex.fe_logical;
5206 ac->ac_b_ex.fe_logical = ex.fe_logical;
5208 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
5209 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
5210 BUG_ON(extent_logical_end(sbi, &ex) > orig_goal_end);
5213 pa->pa_lstart = ac->ac_b_ex.fe_logical;
5214 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5215 pa->pa_len = ac->ac_b_ex.fe_len;
5216 pa->pa_free = pa->pa_len;
5217 spin_lock_init(&pa->pa_lock);
5218 INIT_LIST_HEAD(&pa->pa_group_list);
5220 pa->pa_type = MB_INODE_PA;
5222 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
5223 pa->pa_len, pa->pa_lstart);
5224 trace_ext4_mb_new_inode_pa(ac, pa);
5226 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
5227 ext4_mb_use_inode_pa(ac, pa);
5229 ei = EXT4_I(ac->ac_inode);
5230 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
5234 pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock;
5235 pa->pa_inode = ac->ac_inode;
5237 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
5239 write_lock(pa->pa_node_lock.inode_lock);
5240 ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node);
5241 write_unlock(pa->pa_node_lock.inode_lock);
5242 atomic_inc(&ei->i_prealloc_active);
5246 * creates new preallocated space for locality group inodes belongs to
5248 static noinline_for_stack void
5249 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
5251 struct super_block *sb = ac->ac_sb;
5252 struct ext4_locality_group *lg;
5253 struct ext4_prealloc_space *pa;
5254 struct ext4_group_info *grp;
5256 /* preallocate only when found space is larger then requested */
5257 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
5258 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
5259 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
5260 BUG_ON(ac->ac_pa == NULL);
5264 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5265 pa->pa_lstart = pa->pa_pstart;
5266 pa->pa_len = ac->ac_b_ex.fe_len;
5267 pa->pa_free = pa->pa_len;
5268 spin_lock_init(&pa->pa_lock);
5269 INIT_LIST_HEAD(&pa->pa_node.lg_list);
5270 INIT_LIST_HEAD(&pa->pa_group_list);
5272 pa->pa_type = MB_GROUP_PA;
5274 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
5275 pa->pa_len, pa->pa_lstart);
5276 trace_ext4_mb_new_group_pa(ac, pa);
5278 ext4_mb_use_group_pa(ac, pa);
5279 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
5281 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
5287 pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock;
5288 pa->pa_inode = NULL;
5290 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
5293 * We will later add the new pa to the right bucket
5294 * after updating the pa_free in ext4_mb_release_context
5298 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
5300 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5301 ext4_mb_new_group_pa(ac);
5303 ext4_mb_new_inode_pa(ac);
5307 * finds all unused blocks in on-disk bitmap, frees them in
5308 * in-core bitmap and buddy.
5309 * @pa must be unlinked from inode and group lists, so that
5310 * nobody else can find/use it.
5311 * the caller MUST hold group/inode locks.
5312 * TODO: optimize the case when there are no in-core structures yet
5314 static noinline_for_stack int
5315 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
5316 struct ext4_prealloc_space *pa)
5318 struct super_block *sb = e4b->bd_sb;
5319 struct ext4_sb_info *sbi = EXT4_SB(sb);
5324 unsigned long long grp_blk_start;
5327 BUG_ON(pa->pa_deleted == 0);
5328 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
5329 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
5330 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
5331 end = bit + pa->pa_len;
5334 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
5337 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
5338 mb_debug(sb, "free preallocated %u/%u in group %u\n",
5339 (unsigned) ext4_group_first_block_no(sb, group) + bit,
5340 (unsigned) next - bit, (unsigned) group);
5343 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
5344 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
5345 EXT4_C2B(sbi, bit)),
5347 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
5350 if (free != pa->pa_free) {
5351 ext4_msg(e4b->bd_sb, KERN_CRIT,
5352 "pa %p: logic %lu, phys. %lu, len %d",
5353 pa, (unsigned long) pa->pa_lstart,
5354 (unsigned long) pa->pa_pstart,
5356 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
5359 * pa is already deleted so we use the value obtained
5360 * from the bitmap and continue.
5363 atomic_add(free, &sbi->s_mb_discarded);
5368 static noinline_for_stack int
5369 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
5370 struct ext4_prealloc_space *pa)
5372 struct super_block *sb = e4b->bd_sb;
5376 trace_ext4_mb_release_group_pa(sb, pa);
5377 BUG_ON(pa->pa_deleted == 0);
5378 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
5379 if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) {
5380 ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu",
5381 e4b->bd_group, group, pa->pa_pstart);
5384 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
5385 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
5386 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
5392 * releases all preallocations in given group
5394 * first, we need to decide discard policy:
5395 * - when do we discard
5397 * - how many do we discard
5398 * 1) how many requested
5400 static noinline_for_stack int
5401 ext4_mb_discard_group_preallocations(struct super_block *sb,
5402 ext4_group_t group, int *busy)
5404 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
5405 struct buffer_head *bitmap_bh = NULL;
5406 struct ext4_prealloc_space *pa, *tmp;
5408 struct ext4_buddy e4b;
5409 struct ext4_inode_info *ei;
5415 mb_debug(sb, "discard preallocation for group %u\n", group);
5416 if (list_empty(&grp->bb_prealloc_list))
5419 bitmap_bh = ext4_read_block_bitmap(sb, group);
5420 if (IS_ERR(bitmap_bh)) {
5421 err = PTR_ERR(bitmap_bh);
5422 ext4_error_err(sb, -err,
5423 "Error %d reading block bitmap for %u",
5428 err = ext4_mb_load_buddy(sb, group, &e4b);
5430 ext4_warning(sb, "Error %d loading buddy information for %u",
5436 ext4_lock_group(sb, group);
5437 list_for_each_entry_safe(pa, tmp,
5438 &grp->bb_prealloc_list, pa_group_list) {
5439 spin_lock(&pa->pa_lock);
5440 if (atomic_read(&pa->pa_count)) {
5441 spin_unlock(&pa->pa_lock);
5445 if (pa->pa_deleted) {
5446 spin_unlock(&pa->pa_lock);
5450 /* seems this one can be freed ... */
5451 ext4_mb_mark_pa_deleted(sb, pa);
5454 this_cpu_inc(discard_pa_seq);
5456 /* we can trust pa_free ... */
5457 free += pa->pa_free;
5459 spin_unlock(&pa->pa_lock);
5461 list_del(&pa->pa_group_list);
5462 list_add(&pa->u.pa_tmp_list, &list);
5465 /* now free all selected PAs */
5466 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5468 /* remove from object (inode or locality group) */
5469 if (pa->pa_type == MB_GROUP_PA) {
5470 spin_lock(pa->pa_node_lock.lg_lock);
5471 list_del_rcu(&pa->pa_node.lg_list);
5472 spin_unlock(pa->pa_node_lock.lg_lock);
5474 write_lock(pa->pa_node_lock.inode_lock);
5475 ei = EXT4_I(pa->pa_inode);
5476 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5477 write_unlock(pa->pa_node_lock.inode_lock);
5480 list_del(&pa->u.pa_tmp_list);
5482 if (pa->pa_type == MB_GROUP_PA) {
5483 ext4_mb_release_group_pa(&e4b, pa);
5484 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5486 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5487 ext4_mb_pa_free(pa);
5491 ext4_unlock_group(sb, group);
5492 ext4_mb_unload_buddy(&e4b);
5495 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
5496 free, group, grp->bb_free);
5501 * releases all non-used preallocated blocks for given inode
5503 * It's important to discard preallocations under i_data_sem
5504 * We don't want another block to be served from the prealloc
5505 * space when we are discarding the inode prealloc space.
5507 * FIXME!! Make sure it is valid at all the call sites
5509 void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
5511 struct ext4_inode_info *ei = EXT4_I(inode);
5512 struct super_block *sb = inode->i_sb;
5513 struct buffer_head *bitmap_bh = NULL;
5514 struct ext4_prealloc_space *pa, *tmp;
5515 ext4_group_t group = 0;
5517 struct ext4_buddy e4b;
5518 struct rb_node *iter;
5521 if (!S_ISREG(inode->i_mode)) {
5525 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
5528 mb_debug(sb, "discard preallocation for inode %lu\n",
5530 trace_ext4_discard_preallocations(inode,
5531 atomic_read(&ei->i_prealloc_active), needed);
5537 /* first, collect all pa's in the inode */
5538 write_lock(&ei->i_prealloc_lock);
5539 for (iter = rb_first(&ei->i_prealloc_node); iter && needed;
5540 iter = rb_next(iter)) {
5541 pa = rb_entry(iter, struct ext4_prealloc_space,
5542 pa_node.inode_node);
5543 BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock);
5545 spin_lock(&pa->pa_lock);
5546 if (atomic_read(&pa->pa_count)) {
5547 /* this shouldn't happen often - nobody should
5548 * use preallocation while we're discarding it */
5549 spin_unlock(&pa->pa_lock);
5550 write_unlock(&ei->i_prealloc_lock);
5551 ext4_msg(sb, KERN_ERR,
5552 "uh-oh! used pa while discarding");
5554 schedule_timeout_uninterruptible(HZ);
5558 if (pa->pa_deleted == 0) {
5559 ext4_mb_mark_pa_deleted(sb, pa);
5560 spin_unlock(&pa->pa_lock);
5561 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5562 list_add(&pa->u.pa_tmp_list, &list);
5567 /* someone is deleting pa right now */
5568 spin_unlock(&pa->pa_lock);
5569 write_unlock(&ei->i_prealloc_lock);
5571 /* we have to wait here because pa_deleted
5572 * doesn't mean pa is already unlinked from
5573 * the list. as we might be called from
5574 * ->clear_inode() the inode will get freed
5575 * and concurrent thread which is unlinking
5576 * pa from inode's list may access already
5577 * freed memory, bad-bad-bad */
5579 /* XXX: if this happens too often, we can
5580 * add a flag to force wait only in case
5581 * of ->clear_inode(), but not in case of
5582 * regular truncate */
5583 schedule_timeout_uninterruptible(HZ);
5586 write_unlock(&ei->i_prealloc_lock);
5588 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5589 BUG_ON(pa->pa_type != MB_INODE_PA);
5590 group = ext4_get_group_number(sb, pa->pa_pstart);
5592 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5593 GFP_NOFS|__GFP_NOFAIL);
5595 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5600 bitmap_bh = ext4_read_block_bitmap(sb, group);
5601 if (IS_ERR(bitmap_bh)) {
5602 err = PTR_ERR(bitmap_bh);
5603 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
5605 ext4_mb_unload_buddy(&e4b);
5609 ext4_lock_group(sb, group);
5610 list_del(&pa->pa_group_list);
5611 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5612 ext4_unlock_group(sb, group);
5614 ext4_mb_unload_buddy(&e4b);
5617 list_del(&pa->u.pa_tmp_list);
5618 ext4_mb_pa_free(pa);
5622 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
5624 struct ext4_prealloc_space *pa;
5626 BUG_ON(ext4_pspace_cachep == NULL);
5627 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
5630 atomic_set(&pa->pa_count, 1);
5635 static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac)
5637 struct ext4_prealloc_space *pa = ac->ac_pa;
5641 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
5643 * current function is only called due to an error or due to
5644 * len of found blocks < len of requested blocks hence the PA has not
5645 * been added to grp->bb_prealloc_list. So we don't need to lock it
5648 ext4_mb_pa_free(pa);
5651 #ifdef CONFIG_EXT4_DEBUG
5652 static inline void ext4_mb_show_pa(struct super_block *sb)
5654 ext4_group_t i, ngroups;
5656 if (ext4_forced_shutdown(sb))
5659 ngroups = ext4_get_groups_count(sb);
5660 mb_debug(sb, "groups: ");
5661 for (i = 0; i < ngroups; i++) {
5662 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5663 struct ext4_prealloc_space *pa;
5664 ext4_grpblk_t start;
5665 struct list_head *cur;
5669 ext4_lock_group(sb, i);
5670 list_for_each(cur, &grp->bb_prealloc_list) {
5671 pa = list_entry(cur, struct ext4_prealloc_space,
5673 spin_lock(&pa->pa_lock);
5674 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5676 spin_unlock(&pa->pa_lock);
5677 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5680 ext4_unlock_group(sb, i);
5681 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5686 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5688 struct super_block *sb = ac->ac_sb;
5690 if (ext4_forced_shutdown(sb))
5693 mb_debug(sb, "Can't allocate:"
5694 " Allocation context details:");
5695 mb_debug(sb, "status %u flags 0x%x",
5696 ac->ac_status, ac->ac_flags);
5697 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5698 "goal %lu/%lu/%lu@%lu, "
5699 "best %lu/%lu/%lu@%lu cr %d",
5700 (unsigned long)ac->ac_o_ex.fe_group,
5701 (unsigned long)ac->ac_o_ex.fe_start,
5702 (unsigned long)ac->ac_o_ex.fe_len,
5703 (unsigned long)ac->ac_o_ex.fe_logical,
5704 (unsigned long)ac->ac_g_ex.fe_group,
5705 (unsigned long)ac->ac_g_ex.fe_start,
5706 (unsigned long)ac->ac_g_ex.fe_len,
5707 (unsigned long)ac->ac_g_ex.fe_logical,
5708 (unsigned long)ac->ac_b_ex.fe_group,
5709 (unsigned long)ac->ac_b_ex.fe_start,
5710 (unsigned long)ac->ac_b_ex.fe_len,
5711 (unsigned long)ac->ac_b_ex.fe_logical,
5712 (int)ac->ac_criteria);
5713 mb_debug(sb, "%u found", ac->ac_found);
5714 mb_debug(sb, "used pa: %s, ", ac->ac_pa ? "yes" : "no");
5716 mb_debug(sb, "pa_type %s\n", ac->ac_pa->pa_type == MB_GROUP_PA ?
5717 "group pa" : "inode pa");
5718 ext4_mb_show_pa(sb);
5721 static inline void ext4_mb_show_pa(struct super_block *sb)
5724 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5726 ext4_mb_show_pa(ac->ac_sb);
5731 * We use locality group preallocation for small size file. The size of the
5732 * file is determined by the current size or the resulting size after
5733 * allocation which ever is larger
5735 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5737 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5739 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5740 int bsbits = ac->ac_sb->s_blocksize_bits;
5742 bool inode_pa_eligible, group_pa_eligible;
5744 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5747 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5750 group_pa_eligible = sbi->s_mb_group_prealloc > 0;
5751 inode_pa_eligible = true;
5752 size = extent_logical_end(sbi, &ac->ac_o_ex);
5753 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5756 /* No point in using inode preallocation for closed files */
5757 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5758 !inode_is_open_for_write(ac->ac_inode))
5759 inode_pa_eligible = false;
5761 size = max(size, isize);
5762 /* Don't use group allocation for large files */
5763 if (size > sbi->s_mb_stream_request)
5764 group_pa_eligible = false;
5766 if (!group_pa_eligible) {
5767 if (inode_pa_eligible)
5768 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5770 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5774 BUG_ON(ac->ac_lg != NULL);
5776 * locality group prealloc space are per cpu. The reason for having
5777 * per cpu locality group is to reduce the contention between block
5778 * request from multiple CPUs.
5780 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5782 /* we're going to use group allocation */
5783 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5785 /* serialize all allocations in the group */
5786 mutex_lock(&ac->ac_lg->lg_mutex);
5789 static noinline_for_stack void
5790 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5791 struct ext4_allocation_request *ar)
5793 struct super_block *sb = ar->inode->i_sb;
5794 struct ext4_sb_info *sbi = EXT4_SB(sb);
5795 struct ext4_super_block *es = sbi->s_es;
5799 ext4_grpblk_t block;
5801 /* we can't allocate > group size */
5804 /* just a dirty hack to filter too big requests */
5805 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5806 len = EXT4_CLUSTERS_PER_GROUP(sb);
5808 /* start searching from the goal */
5810 if (goal < le32_to_cpu(es->s_first_data_block) ||
5811 goal >= ext4_blocks_count(es))
5812 goal = le32_to_cpu(es->s_first_data_block);
5813 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5815 /* set up allocation goals */
5816 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5817 ac->ac_status = AC_STATUS_CONTINUE;
5819 ac->ac_inode = ar->inode;
5820 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5821 ac->ac_o_ex.fe_group = group;
5822 ac->ac_o_ex.fe_start = block;
5823 ac->ac_o_ex.fe_len = len;
5824 ac->ac_g_ex = ac->ac_o_ex;
5825 ac->ac_orig_goal_len = ac->ac_g_ex.fe_len;
5826 ac->ac_flags = ar->flags;
5828 /* we have to define context: we'll work with a file or
5829 * locality group. this is a policy, actually */
5830 ext4_mb_group_or_file(ac);
5832 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5833 "left: %u/%u, right %u/%u to %swritable\n",
5834 (unsigned) ar->len, (unsigned) ar->logical,
5835 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5836 (unsigned) ar->lleft, (unsigned) ar->pleft,
5837 (unsigned) ar->lright, (unsigned) ar->pright,
5838 inode_is_open_for_write(ar->inode) ? "" : "non-");
5841 static noinline_for_stack void
5842 ext4_mb_discard_lg_preallocations(struct super_block *sb,
5843 struct ext4_locality_group *lg,
5844 int order, int total_entries)
5846 ext4_group_t group = 0;
5847 struct ext4_buddy e4b;
5848 LIST_HEAD(discard_list);
5849 struct ext4_prealloc_space *pa, *tmp;
5851 mb_debug(sb, "discard locality group preallocation\n");
5853 spin_lock(&lg->lg_prealloc_lock);
5854 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5856 lockdep_is_held(&lg->lg_prealloc_lock)) {
5857 spin_lock(&pa->pa_lock);
5858 if (atomic_read(&pa->pa_count)) {
5860 * This is the pa that we just used
5861 * for block allocation. So don't
5864 spin_unlock(&pa->pa_lock);
5867 if (pa->pa_deleted) {
5868 spin_unlock(&pa->pa_lock);
5871 /* only lg prealloc space */
5872 BUG_ON(pa->pa_type != MB_GROUP_PA);
5874 /* seems this one can be freed ... */
5875 ext4_mb_mark_pa_deleted(sb, pa);
5876 spin_unlock(&pa->pa_lock);
5878 list_del_rcu(&pa->pa_node.lg_list);
5879 list_add(&pa->u.pa_tmp_list, &discard_list);
5882 if (total_entries <= 5) {
5884 * we want to keep only 5 entries
5885 * allowing it to grow to 8. This
5886 * mak sure we don't call discard
5887 * soon for this list.
5892 spin_unlock(&lg->lg_prealloc_lock);
5894 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5897 group = ext4_get_group_number(sb, pa->pa_pstart);
5898 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5899 GFP_NOFS|__GFP_NOFAIL);
5901 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5905 ext4_lock_group(sb, group);
5906 list_del(&pa->pa_group_list);
5907 ext4_mb_release_group_pa(&e4b, pa);
5908 ext4_unlock_group(sb, group);
5910 ext4_mb_unload_buddy(&e4b);
5911 list_del(&pa->u.pa_tmp_list);
5912 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5917 * We have incremented pa_count. So it cannot be freed at this
5918 * point. Also we hold lg_mutex. So no parallel allocation is
5919 * possible from this lg. That means pa_free cannot be updated.
5921 * A parallel ext4_mb_discard_group_preallocations is possible.
5922 * which can cause the lg_prealloc_list to be updated.
5925 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5927 int order, added = 0, lg_prealloc_count = 1;
5928 struct super_block *sb = ac->ac_sb;
5929 struct ext4_locality_group *lg = ac->ac_lg;
5930 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5932 order = fls(pa->pa_free) - 1;
5933 if (order > PREALLOC_TB_SIZE - 1)
5934 /* The max size of hash table is PREALLOC_TB_SIZE */
5935 order = PREALLOC_TB_SIZE - 1;
5936 /* Add the prealloc space to lg */
5937 spin_lock(&lg->lg_prealloc_lock);
5938 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5940 lockdep_is_held(&lg->lg_prealloc_lock)) {
5941 spin_lock(&tmp_pa->pa_lock);
5942 if (tmp_pa->pa_deleted) {
5943 spin_unlock(&tmp_pa->pa_lock);
5946 if (!added && pa->pa_free < tmp_pa->pa_free) {
5947 /* Add to the tail of the previous entry */
5948 list_add_tail_rcu(&pa->pa_node.lg_list,
5949 &tmp_pa->pa_node.lg_list);
5952 * we want to count the total
5953 * number of entries in the list
5956 spin_unlock(&tmp_pa->pa_lock);
5957 lg_prealloc_count++;
5960 list_add_tail_rcu(&pa->pa_node.lg_list,
5961 &lg->lg_prealloc_list[order]);
5962 spin_unlock(&lg->lg_prealloc_lock);
5964 /* Now trim the list to be not more than 8 elements */
5965 if (lg_prealloc_count > 8)
5966 ext4_mb_discard_lg_preallocations(sb, lg,
5967 order, lg_prealloc_count);
5971 * release all resource we used in allocation
5973 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
5975 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5976 struct ext4_prealloc_space *pa = ac->ac_pa;
5978 if (pa->pa_type == MB_GROUP_PA) {
5979 /* see comment in ext4_mb_use_group_pa() */
5980 spin_lock(&pa->pa_lock);
5981 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5982 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5983 pa->pa_free -= ac->ac_b_ex.fe_len;
5984 pa->pa_len -= ac->ac_b_ex.fe_len;
5985 spin_unlock(&pa->pa_lock);
5988 * We want to add the pa to the right bucket.
5989 * Remove it from the list and while adding
5990 * make sure the list to which we are adding
5993 if (likely(pa->pa_free)) {
5994 spin_lock(pa->pa_node_lock.lg_lock);
5995 list_del_rcu(&pa->pa_node.lg_list);
5996 spin_unlock(pa->pa_node_lock.lg_lock);
5997 ext4_mb_add_n_trim(ac);
6001 ext4_mb_put_pa(ac, ac->ac_sb, pa);
6003 if (ac->ac_bitmap_page)
6004 put_page(ac->ac_bitmap_page);
6005 if (ac->ac_buddy_page)
6006 put_page(ac->ac_buddy_page);
6007 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
6008 mutex_unlock(&ac->ac_lg->lg_mutex);
6009 ext4_mb_collect_stats(ac);
6013 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
6015 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
6017 int freed = 0, busy = 0;
6020 trace_ext4_mb_discard_preallocations(sb, needed);
6023 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
6025 for (i = 0; i < ngroups && needed > 0; i++) {
6026 ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
6032 if (needed > 0 && busy && ++retry < 3) {
6040 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
6041 struct ext4_allocation_context *ac, u64 *seq)
6047 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
6052 seq_retry = ext4_get_discard_pa_seq_sum();
6053 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
6054 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
6060 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
6065 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
6066 * linearly starting at the goal block and also excludes the blocks which
6067 * are going to be in use after fast commit replay.
6070 ext4_mb_new_blocks_simple(struct ext4_allocation_request *ar, int *errp)
6072 struct buffer_head *bitmap_bh;
6073 struct super_block *sb = ar->inode->i_sb;
6074 struct ext4_sb_info *sbi = EXT4_SB(sb);
6075 ext4_group_t group, nr;
6076 ext4_grpblk_t blkoff;
6077 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
6078 ext4_grpblk_t i = 0;
6079 ext4_fsblk_t goal, block;
6080 struct ext4_super_block *es = sbi->s_es;
6083 if (goal < le32_to_cpu(es->s_first_data_block) ||
6084 goal >= ext4_blocks_count(es))
6085 goal = le32_to_cpu(es->s_first_data_block);
6088 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
6089 for (nr = ext4_get_groups_count(sb); nr > 0; nr--) {
6090 bitmap_bh = ext4_read_block_bitmap(sb, group);
6091 if (IS_ERR(bitmap_bh)) {
6092 *errp = PTR_ERR(bitmap_bh);
6093 pr_warn("Failed to read block bitmap\n");
6098 i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
6102 if (ext4_fc_replay_check_excluded(sb,
6103 ext4_group_first_block_no(sb, group) +
6104 EXT4_C2B(sbi, i))) {
6113 if (++group >= ext4_get_groups_count(sb))
6124 block = ext4_group_first_block_no(sb, group) + EXT4_C2B(sbi, i);
6125 ext4_mb_mark_bb(sb, block, 1, true);
6132 * Main entry point into mballoc to allocate blocks
6133 * it tries to use preallocation first, then falls back
6134 * to usual allocation
6136 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
6137 struct ext4_allocation_request *ar, int *errp)
6139 struct ext4_allocation_context *ac = NULL;
6140 struct ext4_sb_info *sbi;
6141 struct super_block *sb;
6142 ext4_fsblk_t block = 0;
6143 unsigned int inquota = 0;
6144 unsigned int reserv_clstrs = 0;
6149 sb = ar->inode->i_sb;
6152 trace_ext4_request_blocks(ar);
6153 if (sbi->s_mount_state & EXT4_FC_REPLAY)
6154 return ext4_mb_new_blocks_simple(ar, errp);
6156 /* Allow to use superuser reservation for quota file */
6157 if (ext4_is_quota_file(ar->inode))
6158 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
6160 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
6161 /* Without delayed allocation we need to verify
6162 * there is enough free blocks to do block allocation
6163 * and verify allocation doesn't exceed the quota limits.
6166 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
6168 /* let others to free the space */
6170 ar->len = ar->len >> 1;
6173 ext4_mb_show_pa(sb);
6177 reserv_clstrs = ar->len;
6178 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
6179 dquot_alloc_block_nofail(ar->inode,
6180 EXT4_C2B(sbi, ar->len));
6183 dquot_alloc_block(ar->inode,
6184 EXT4_C2B(sbi, ar->len))) {
6186 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
6197 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
6204 ext4_mb_initialize_context(ac, ar);
6206 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
6207 seq = this_cpu_read(discard_pa_seq);
6208 if (!ext4_mb_use_preallocated(ac)) {
6209 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
6210 ext4_mb_normalize_request(ac, ar);
6212 *errp = ext4_mb_pa_alloc(ac);
6216 /* allocate space in core */
6217 *errp = ext4_mb_regular_allocator(ac);
6219 * pa allocated above is added to grp->bb_prealloc_list only
6220 * when we were able to allocate some block i.e. when
6221 * ac->ac_status == AC_STATUS_FOUND.
6222 * And error from above mean ac->ac_status != AC_STATUS_FOUND
6223 * So we have to free this pa here itself.
6226 ext4_mb_pa_put_free(ac);
6227 ext4_discard_allocated_blocks(ac);
6230 if (ac->ac_status == AC_STATUS_FOUND &&
6231 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
6232 ext4_mb_pa_put_free(ac);
6234 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
6235 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
6237 ext4_discard_allocated_blocks(ac);
6240 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
6241 ar->len = ac->ac_b_ex.fe_len;
6244 if (++retries < 3 &&
6245 ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
6248 * If block allocation fails then the pa allocated above
6249 * needs to be freed here itself.
6251 ext4_mb_pa_put_free(ac);
6257 ac->ac_b_ex.fe_len = 0;
6259 ext4_mb_show_ac(ac);
6261 ext4_mb_release_context(ac);
6262 kmem_cache_free(ext4_ac_cachep, ac);
6264 if (inquota && ar->len < inquota)
6265 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
6267 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
6268 /* release all the reserved blocks if non delalloc */
6269 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
6273 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
6279 * We can merge two free data extents only if the physical blocks
6280 * are contiguous, AND the extents were freed by the same transaction,
6281 * AND the blocks are associated with the same group.
6283 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
6284 struct ext4_free_data *entry,
6285 struct ext4_free_data *new_entry,
6286 struct rb_root *entry_rb_root)
6288 if ((entry->efd_tid != new_entry->efd_tid) ||
6289 (entry->efd_group != new_entry->efd_group))
6291 if (entry->efd_start_cluster + entry->efd_count ==
6292 new_entry->efd_start_cluster) {
6293 new_entry->efd_start_cluster = entry->efd_start_cluster;
6294 new_entry->efd_count += entry->efd_count;
6295 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
6296 entry->efd_start_cluster) {
6297 new_entry->efd_count += entry->efd_count;
6300 spin_lock(&sbi->s_md_lock);
6301 list_del(&entry->efd_list);
6302 spin_unlock(&sbi->s_md_lock);
6303 rb_erase(&entry->efd_node, entry_rb_root);
6304 kmem_cache_free(ext4_free_data_cachep, entry);
6307 static noinline_for_stack void
6308 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
6309 struct ext4_free_data *new_entry)
6311 ext4_group_t group = e4b->bd_group;
6312 ext4_grpblk_t cluster;
6313 ext4_grpblk_t clusters = new_entry->efd_count;
6314 struct ext4_free_data *entry;
6315 struct ext4_group_info *db = e4b->bd_info;
6316 struct super_block *sb = e4b->bd_sb;
6317 struct ext4_sb_info *sbi = EXT4_SB(sb);
6318 struct rb_node **n = &db->bb_free_root.rb_node, *node;
6319 struct rb_node *parent = NULL, *new_node;
6321 BUG_ON(!ext4_handle_valid(handle));
6322 BUG_ON(e4b->bd_bitmap_page == NULL);
6323 BUG_ON(e4b->bd_buddy_page == NULL);
6325 new_node = &new_entry->efd_node;
6326 cluster = new_entry->efd_start_cluster;
6329 /* first free block exent. We need to
6330 protect buddy cache from being freed,
6331 * otherwise we'll refresh it from
6332 * on-disk bitmap and lose not-yet-available
6334 get_page(e4b->bd_buddy_page);
6335 get_page(e4b->bd_bitmap_page);
6339 entry = rb_entry(parent, struct ext4_free_data, efd_node);
6340 if (cluster < entry->efd_start_cluster)
6342 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
6343 n = &(*n)->rb_right;
6345 ext4_grp_locked_error(sb, group, 0,
6346 ext4_group_first_block_no(sb, group) +
6347 EXT4_C2B(sbi, cluster),
6348 "Block already on to-be-freed list");
6349 kmem_cache_free(ext4_free_data_cachep, new_entry);
6354 rb_link_node(new_node, parent, n);
6355 rb_insert_color(new_node, &db->bb_free_root);
6357 /* Now try to see the extent can be merged to left and right */
6358 node = rb_prev(new_node);
6360 entry = rb_entry(node, struct ext4_free_data, efd_node);
6361 ext4_try_merge_freed_extent(sbi, entry, new_entry,
6362 &(db->bb_free_root));
6365 node = rb_next(new_node);
6367 entry = rb_entry(node, struct ext4_free_data, efd_node);
6368 ext4_try_merge_freed_extent(sbi, entry, new_entry,
6369 &(db->bb_free_root));
6372 spin_lock(&sbi->s_md_lock);
6373 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list[new_entry->efd_tid & 1]);
6374 sbi->s_mb_free_pending += clusters;
6375 spin_unlock(&sbi->s_md_lock);
6378 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
6379 unsigned long count)
6381 struct super_block *sb = inode->i_sb;
6383 ext4_grpblk_t blkoff;
6385 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
6386 ext4_mb_mark_context(NULL, sb, false, group, blkoff, count,
6387 EXT4_MB_BITMAP_MARKED_CHECK |
6388 EXT4_MB_SYNC_UPDATE,
6393 * ext4_mb_clear_bb() -- helper function for freeing blocks.
6394 * Used by ext4_free_blocks()
6395 * @handle: handle for this transaction
6397 * @block: starting physical block to be freed
6398 * @count: number of blocks to be freed
6399 * @flags: flags used by ext4_free_blocks
6401 static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
6402 ext4_fsblk_t block, unsigned long count,
6405 struct super_block *sb = inode->i_sb;
6406 struct ext4_group_info *grp;
6407 unsigned int overflow;
6409 ext4_group_t block_group;
6410 struct ext4_sb_info *sbi;
6411 struct ext4_buddy e4b;
6412 unsigned int count_clusters;
6415 ext4_grpblk_t changed;
6419 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6420 !ext4_inode_block_valid(inode, block, count)) {
6421 ext4_error(sb, "Freeing blocks in system zone - "
6422 "Block = %llu, count = %lu", block, count);
6423 /* err = 0. ext4_std_error should be a no op */
6426 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6430 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6432 grp = ext4_get_group_info(sb, block_group);
6433 if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
6437 * Check to see if we are freeing blocks across a group
6440 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
6441 overflow = EXT4_C2B(sbi, bit) + count -
6442 EXT4_BLOCKS_PER_GROUP(sb);
6444 /* The range changed so it's no longer validated */
6445 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6447 count_clusters = EXT4_NUM_B2C(sbi, count);
6448 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
6450 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
6451 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
6452 GFP_NOFS|__GFP_NOFAIL);
6456 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6457 !ext4_inode_block_valid(inode, block, count)) {
6458 ext4_error(sb, "Freeing blocks in system zone - "
6459 "Block = %llu, count = %lu", block, count);
6460 /* err = 0. ext4_std_error should be a no op */
6464 #ifdef AGGRESSIVE_CHECK
6465 mark_flags |= EXT4_MB_BITMAP_MARKED_CHECK;
6467 err = ext4_mb_mark_context(handle, sb, false, block_group, bit,
6468 count_clusters, mark_flags, &changed);
6471 if (err && changed == 0)
6474 #ifdef AGGRESSIVE_CHECK
6475 BUG_ON(changed != count_clusters);
6479 * We need to make sure we don't reuse the freed block until after the
6480 * transaction is committed. We make an exception if the inode is to be
6481 * written in writeback mode since writeback mode has weak data
6482 * consistency guarantees.
6484 if (ext4_handle_valid(handle) &&
6485 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
6486 !ext4_should_writeback_data(inode))) {
6487 struct ext4_free_data *new_entry;
6489 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
6492 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
6493 GFP_NOFS|__GFP_NOFAIL);
6494 new_entry->efd_start_cluster = bit;
6495 new_entry->efd_group = block_group;
6496 new_entry->efd_count = count_clusters;
6497 new_entry->efd_tid = handle->h_transaction->t_tid;
6499 ext4_lock_group(sb, block_group);
6500 ext4_mb_free_metadata(handle, &e4b, new_entry);
6502 if (test_opt(sb, DISCARD)) {
6503 err = ext4_issue_discard(sb, block_group, bit,
6504 count_clusters, NULL);
6505 if (err && err != -EOPNOTSUPP)
6506 ext4_msg(sb, KERN_WARNING, "discard request in"
6507 " group:%u block:%d count:%lu failed"
6508 " with %d", block_group, bit, count,
6511 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
6513 ext4_lock_group(sb, block_group);
6514 mb_free_blocks(inode, &e4b, bit, count_clusters);
6517 ext4_unlock_group(sb, block_group);
6520 * on a bigalloc file system, defer the s_freeclusters_counter
6521 * update to the caller (ext4_remove_space and friends) so they
6522 * can determine if a cluster freed here should be rereserved
6524 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6525 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6526 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6527 percpu_counter_add(&sbi->s_freeclusters_counter,
6531 if (overflow && !err) {
6534 ext4_mb_unload_buddy(&e4b);
6535 /* The range changed so it's no longer validated */
6536 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6541 ext4_mb_unload_buddy(&e4b);
6543 ext4_std_error(sb, err);
6547 * ext4_free_blocks() -- Free given blocks and update quota
6548 * @handle: handle for this transaction
6550 * @bh: optional buffer of the block to be freed
6551 * @block: starting physical block to be freed
6552 * @count: number of blocks to be freed
6553 * @flags: flags used by ext4_free_blocks
6555 void ext4_free_blocks(handle_t *handle, struct inode *inode,
6556 struct buffer_head *bh, ext4_fsblk_t block,
6557 unsigned long count, int flags)
6559 struct super_block *sb = inode->i_sb;
6560 unsigned int overflow;
6561 struct ext4_sb_info *sbi;
6567 BUG_ON(block != bh->b_blocknr);
6569 block = bh->b_blocknr;
6572 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
6573 ext4_free_blocks_simple(inode, block, EXT4_NUM_B2C(sbi, count));
6579 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6580 !ext4_inode_block_valid(inode, block, count)) {
6581 ext4_error(sb, "Freeing blocks not in datazone - "
6582 "block = %llu, count = %lu", block, count);
6585 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6587 ext4_debug("freeing block %llu\n", block);
6588 trace_ext4_free_blocks(inode, block, count, flags);
6590 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6593 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
6598 * If the extent to be freed does not begin on a cluster
6599 * boundary, we need to deal with partial clusters at the
6600 * beginning and end of the extent. Normally we will free
6601 * blocks at the beginning or the end unless we are explicitly
6602 * requested to avoid doing so.
6604 overflow = EXT4_PBLK_COFF(sbi, block);
6606 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
6607 overflow = sbi->s_cluster_ratio - overflow;
6609 if (count > overflow)
6617 /* The range changed so it's no longer validated */
6618 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6620 overflow = EXT4_LBLK_COFF(sbi, count);
6622 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
6623 if (count > overflow)
6628 count += sbi->s_cluster_ratio - overflow;
6629 /* The range changed so it's no longer validated */
6630 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6633 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6635 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
6637 for (i = 0; i < count; i++) {
6640 bh = sb_find_get_block(inode->i_sb, block + i);
6641 ext4_forget(handle, is_metadata, inode, bh, block + i);
6645 ext4_mb_clear_bb(handle, inode, block, count, flags);
6649 * ext4_group_add_blocks() -- Add given blocks to an existing group
6650 * @handle: handle to this transaction
6652 * @block: start physical block to add to the block group
6653 * @count: number of blocks to free
6655 * This marks the blocks as free in the bitmap and buddy.
6657 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6658 ext4_fsblk_t block, unsigned long count)
6660 ext4_group_t block_group;
6662 struct ext4_sb_info *sbi = EXT4_SB(sb);
6663 struct ext4_buddy e4b;
6665 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6666 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6667 unsigned long cluster_count = last_cluster - first_cluster + 1;
6668 ext4_grpblk_t changed;
6670 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6672 if (cluster_count == 0)
6675 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6677 * Check to see if we are freeing blocks across a group
6680 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6681 ext4_warning(sb, "too many blocks added to group %u",
6687 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6691 if (!ext4_sb_block_valid(sb, NULL, block, count)) {
6692 ext4_error(sb, "Adding blocks in system zones - "
6693 "Block = %llu, count = %lu",
6699 err = ext4_mb_mark_context(handle, sb, false, block_group, bit,
6700 cluster_count, EXT4_MB_BITMAP_MARKED_CHECK,
6702 if (err && changed == 0)
6705 if (changed != cluster_count)
6706 ext4_error(sb, "bit already cleared in group %u", block_group);
6708 ext4_lock_group(sb, block_group);
6709 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6710 ext4_unlock_group(sb, block_group);
6711 percpu_counter_add(&sbi->s_freeclusters_counter,
6715 ext4_mb_unload_buddy(&e4b);
6717 ext4_std_error(sb, err);
6722 * ext4_trim_extent -- function to TRIM one single free extent in the group
6723 * @sb: super block for the file system
6724 * @start: starting block of the free extent in the alloc. group
6725 * @count: number of blocks to TRIM
6726 * @e4b: ext4 buddy for the group
6728 * Trim "count" blocks starting at "start" in the "group". To assure that no
6729 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6730 * be called with under the group lock.
6732 static int ext4_trim_extent(struct super_block *sb,
6733 int start, int count, struct ext4_buddy *e4b)
6737 struct ext4_free_extent ex;
6738 ext4_group_t group = e4b->bd_group;
6741 trace_ext4_trim_extent(sb, group, start, count);
6743 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6745 ex.fe_start = start;
6746 ex.fe_group = group;
6750 * Mark blocks used, so no one can reuse them while
6753 mb_mark_used(e4b, &ex);
6754 ext4_unlock_group(sb, group);
6755 ret = ext4_issue_discard(sb, group, start, count, NULL);
6756 ext4_lock_group(sb, group);
6757 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6761 static ext4_grpblk_t ext4_last_grp_cluster(struct super_block *sb,
6764 unsigned long nr_clusters_in_group;
6766 if (grp < (ext4_get_groups_count(sb) - 1))
6767 nr_clusters_in_group = EXT4_CLUSTERS_PER_GROUP(sb);
6769 nr_clusters_in_group = (ext4_blocks_count(EXT4_SB(sb)->s_es) -
6770 ext4_group_first_block_no(sb, grp))
6771 >> EXT4_CLUSTER_BITS(sb);
6773 return nr_clusters_in_group - 1;
6776 static bool ext4_trim_interrupted(void)
6778 return fatal_signal_pending(current) || freezing(current);
6781 static int ext4_try_to_trim_range(struct super_block *sb,
6782 struct ext4_buddy *e4b, ext4_grpblk_t start,
6783 ext4_grpblk_t max, ext4_grpblk_t minblocks)
6784 __acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
6785 __releases(ext4_group_lock_ptr(sb, e4b->bd_group))
6787 ext4_grpblk_t next, count, free_count, last, origin_start;
6788 bool set_trimmed = false;
6791 last = ext4_last_grp_cluster(sb, e4b->bd_group);
6792 bitmap = e4b->bd_bitmap;
6793 if (start == 0 && max >= last)
6795 origin_start = start;
6796 start = max(e4b->bd_info->bb_first_free, start);
6800 while (start <= max) {
6801 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6805 next = mb_find_next_bit(bitmap, last + 1, start);
6806 if (origin_start == 0 && next >= last)
6809 if ((next - start) >= minblocks) {
6810 int ret = ext4_trim_extent(sb, start, next - start, e4b);
6812 if (ret && ret != -EOPNOTSUPP)
6814 count += next - start;
6816 free_count += next - start;
6819 if (ext4_trim_interrupted())
6822 if (need_resched()) {
6823 ext4_unlock_group(sb, e4b->bd_group);
6825 ext4_lock_group(sb, e4b->bd_group);
6828 if ((e4b->bd_info->bb_free - free_count) < minblocks)
6833 EXT4_MB_GRP_SET_TRIMMED(e4b->bd_info);
6839 * ext4_trim_all_free -- function to trim all free space in alloc. group
6840 * @sb: super block for file system
6841 * @group: group to be trimmed
6842 * @start: first group block to examine
6843 * @max: last group block to examine
6844 * @minblocks: minimum extent block count
6846 * ext4_trim_all_free walks through group's block bitmap searching for free
6847 * extents. When the free extent is found, mark it as used in group buddy
6848 * bitmap. Then issue a TRIM command on this extent and free the extent in
6849 * the group buddy bitmap.
6851 static ext4_grpblk_t
6852 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6853 ext4_grpblk_t start, ext4_grpblk_t max,
6854 ext4_grpblk_t minblocks)
6856 struct ext4_buddy e4b;
6859 trace_ext4_trim_all_free(sb, group, start, max);
6861 ret = ext4_mb_load_buddy(sb, group, &e4b);
6863 ext4_warning(sb, "Error %d loading buddy information for %u",
6868 ext4_lock_group(sb, group);
6870 if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
6871 minblocks < EXT4_SB(sb)->s_last_trim_minblks)
6872 ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
6876 ext4_unlock_group(sb, group);
6877 ext4_mb_unload_buddy(&e4b);
6879 ext4_debug("trimmed %d blocks in the group %d\n",
6886 * ext4_trim_fs() -- trim ioctl handle function
6887 * @sb: superblock for filesystem
6888 * @range: fstrim_range structure
6890 * start: First Byte to trim
6891 * len: number of Bytes to trim from start
6892 * minlen: minimum extent length in Bytes
6893 * ext4_trim_fs goes through all allocation groups containing Bytes from
6894 * start to start+len. For each such a group ext4_trim_all_free function
6895 * is invoked to trim all free space.
6897 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6899 unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
6900 struct ext4_group_info *grp;
6901 ext4_group_t group, first_group, last_group;
6902 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6903 uint64_t start, end, minlen, trimmed = 0;
6904 ext4_fsblk_t first_data_blk =
6905 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6906 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6909 start = range->start >> sb->s_blocksize_bits;
6910 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6911 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6912 range->minlen >> sb->s_blocksize_bits);
6914 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6915 start >= max_blks ||
6916 range->len < sb->s_blocksize)
6918 /* No point to try to trim less than discard granularity */
6919 if (range->minlen < discard_granularity) {
6920 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6921 discard_granularity >> sb->s_blocksize_bits);
6922 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
6925 if (end >= max_blks - 1)
6927 if (end <= first_data_blk)
6929 if (start < first_data_blk)
6930 start = first_data_blk;
6932 /* Determine first and last group to examine based on start and end */
6933 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6934 &first_group, &first_cluster);
6935 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6936 &last_group, &last_cluster);
6938 /* end now represents the last cluster to discard in this group */
6939 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6941 for (group = first_group; group <= last_group; group++) {
6942 if (ext4_trim_interrupted())
6944 grp = ext4_get_group_info(sb, group);
6947 /* We only do this if the grp has never been initialized */
6948 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6949 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6955 * For all the groups except the last one, last cluster will
6956 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6957 * change it for the last group, note that last_cluster is
6958 * already computed earlier by ext4_get_group_no_and_offset()
6960 if (group == last_group)
6962 if (grp->bb_free >= minlen) {
6963 cnt = ext4_trim_all_free(sb, group, first_cluster,
6973 * For every group except the first one, we are sure
6974 * that the first cluster to discard will be cluster #0.
6980 EXT4_SB(sb)->s_last_trim_minblks = minlen;
6983 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6987 /* Iterate all the free extents in the group. */
6989 ext4_mballoc_query_range(
6990 struct super_block *sb,
6992 ext4_grpblk_t start,
6994 ext4_mballoc_query_range_fn formatter,
6999 struct ext4_buddy e4b;
7002 error = ext4_mb_load_buddy(sb, group, &e4b);
7005 bitmap = e4b.bd_bitmap;
7007 ext4_lock_group(sb, group);
7009 start = max(e4b.bd_info->bb_first_free, start);
7010 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
7011 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
7013 while (start <= end) {
7014 start = mb_find_next_zero_bit(bitmap, end + 1, start);
7017 next = mb_find_next_bit(bitmap, end + 1, start);
7019 ext4_unlock_group(sb, group);
7020 error = formatter(sb, group, start, next - start, priv);
7023 ext4_lock_group(sb, group);
7028 ext4_unlock_group(sb, group);
7030 ext4_mb_unload_buddy(&e4b);
7035 #ifdef CONFIG_EXT4_KUNIT_TESTS
7036 #include "mballoc-test.c"