IPVS: netns, svc counters moved in ip_vs_ctl,c
[linux/fpc-iii.git] / fs / ext4 / mballoc.c
blob5b4d4e3a4d58e3506c15e891e7b8df394bc0838f
1 /*
2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
24 #include "mballoc.h"
25 #include <linux/debugfs.h>
26 #include <linux/slab.h>
27 #include <trace/events/ext4.h>
30 * MUSTDO:
31 * - test ext4_ext_search_left() and ext4_ext_search_right()
32 * - search for metadata in few groups
34 * TODO v4:
35 * - normalization should take into account whether file is still open
36 * - discard preallocations if no free space left (policy?)
37 * - don't normalize tails
38 * - quota
39 * - reservation for superuser
41 * TODO v3:
42 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
43 * - track min/max extents in each group for better group selection
44 * - mb_mark_used() may allocate chunk right after splitting buddy
45 * - tree of groups sorted by number of free blocks
46 * - error handling
50 * The allocation request involve request for multiple number of blocks
51 * near to the goal(block) value specified.
53 * During initialization phase of the allocator we decide to use the
54 * group preallocation or inode preallocation depending on the size of
55 * the file. The size of the file could be the resulting file size we
56 * would have after allocation, or the current file size, which ever
57 * is larger. If the size is less than sbi->s_mb_stream_request we
58 * select to use the group preallocation. The default value of
59 * s_mb_stream_request is 16 blocks. This can also be tuned via
60 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
61 * terms of number of blocks.
63 * The main motivation for having small file use group preallocation is to
64 * ensure that we have small files closer together on the disk.
66 * First stage the allocator looks at the inode prealloc list,
67 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
68 * spaces for this particular inode. The inode prealloc space is
69 * represented as:
71 * pa_lstart -> the logical start block for this prealloc space
72 * pa_pstart -> the physical start block for this prealloc space
73 * pa_len -> length for this prealloc space
74 * pa_free -> free space available in this prealloc space
76 * The inode preallocation space is used looking at the _logical_ start
77 * block. If only the logical file block falls within the range of prealloc
78 * space we will consume the particular prealloc space. This make sure that
79 * that the we have contiguous physical blocks representing the file blocks
81 * The important thing to be noted in case of inode prealloc space is that
82 * we don't modify the values associated to inode prealloc space except
83 * pa_free.
85 * If we are not able to find blocks in the inode prealloc space and if we
86 * have the group allocation flag set then we look at the locality group
87 * prealloc space. These are per CPU prealloc list repreasented as
89 * ext4_sb_info.s_locality_groups[smp_processor_id()]
91 * The reason for having a per cpu locality group is to reduce the contention
92 * between CPUs. It is possible to get scheduled at this point.
94 * The locality group prealloc space is used looking at whether we have
95 * enough free space (pa_free) withing the prealloc space.
97 * If we can't allocate blocks via inode prealloc or/and locality group
98 * prealloc then we look at the buddy cache. The buddy cache is represented
99 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
100 * mapped to the buddy and bitmap information regarding different
101 * groups. The buddy information is attached to buddy cache inode so that
102 * we can access them through the page cache. The information regarding
103 * each group is loaded via ext4_mb_load_buddy. The information involve
104 * block bitmap and buddy information. The information are stored in the
105 * inode as:
107 * { page }
108 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
111 * one block each for bitmap and buddy information. So for each group we
112 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
113 * blocksize) blocks. So it can have information regarding groups_per_page
114 * which is blocks_per_page/2
116 * The buddy cache inode is not stored on disk. The inode is thrown
117 * away when the filesystem is unmounted.
119 * We look for count number of blocks in the buddy cache. If we were able
120 * to locate that many free blocks we return with additional information
121 * regarding rest of the contiguous physical block available
123 * Before allocating blocks via buddy cache we normalize the request
124 * blocks. This ensure we ask for more blocks that we needed. The extra
125 * blocks that we get after allocation is added to the respective prealloc
126 * list. In case of inode preallocation we follow a list of heuristics
127 * based on file size. This can be found in ext4_mb_normalize_request. If
128 * we are doing a group prealloc we try to normalize the request to
129 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
130 * 512 blocks. This can be tuned via
131 * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
132 * terms of number of blocks. If we have mounted the file system with -O
133 * stripe=<value> option the group prealloc request is normalized to the
134 * stripe value (sbi->s_stripe)
136 * The regular allocator(using the buddy cache) supports few tunables.
138 * /sys/fs/ext4/<partition>/mb_min_to_scan
139 * /sys/fs/ext4/<partition>/mb_max_to_scan
140 * /sys/fs/ext4/<partition>/mb_order2_req
142 * The regular allocator uses buddy scan only if the request len is power of
143 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
144 * value of s_mb_order2_reqs can be tuned via
145 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
146 * stripe size (sbi->s_stripe), we try to search for contiguous block in
147 * stripe size. This should result in better allocation on RAID setups. If
148 * not, we search in the specific group using bitmap for best extents. The
149 * tunable min_to_scan and max_to_scan control the behaviour here.
150 * min_to_scan indicate how long the mballoc __must__ look for a best
151 * extent and max_to_scan indicates how long the mballoc __can__ look for a
152 * best extent in the found extents. Searching for the blocks starts with
153 * the group specified as the goal value in allocation context via
154 * ac_g_ex. Each group is first checked based on the criteria whether it
155 * can used for allocation. ext4_mb_good_group explains how the groups are
156 * checked.
158 * Both the prealloc space are getting populated as above. So for the first
159 * request we will hit the buddy cache which will result in this prealloc
160 * space getting filled. The prealloc space is then later used for the
161 * subsequent request.
165 * mballoc operates on the following data:
166 * - on-disk bitmap
167 * - in-core buddy (actually includes buddy and bitmap)
168 * - preallocation descriptors (PAs)
170 * there are two types of preallocations:
171 * - inode
172 * assiged to specific inode and can be used for this inode only.
173 * it describes part of inode's space preallocated to specific
174 * physical blocks. any block from that preallocated can be used
175 * independent. the descriptor just tracks number of blocks left
176 * unused. so, before taking some block from descriptor, one must
177 * make sure corresponded logical block isn't allocated yet. this
178 * also means that freeing any block within descriptor's range
179 * must discard all preallocated blocks.
180 * - locality group
181 * assigned to specific locality group which does not translate to
182 * permanent set of inodes: inode can join and leave group. space
183 * from this type of preallocation can be used for any inode. thus
184 * it's consumed from the beginning to the end.
186 * relation between them can be expressed as:
187 * in-core buddy = on-disk bitmap + preallocation descriptors
189 * this mean blocks mballoc considers used are:
190 * - allocated blocks (persistent)
191 * - preallocated blocks (non-persistent)
193 * consistency in mballoc world means that at any time a block is either
194 * free or used in ALL structures. notice: "any time" should not be read
195 * literally -- time is discrete and delimited by locks.
197 * to keep it simple, we don't use block numbers, instead we count number of
198 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
200 * all operations can be expressed as:
201 * - init buddy: buddy = on-disk + PAs
202 * - new PA: buddy += N; PA = N
203 * - use inode PA: on-disk += N; PA -= N
204 * - discard inode PA buddy -= on-disk - PA; PA = 0
205 * - use locality group PA on-disk += N; PA -= N
206 * - discard locality group PA buddy -= PA; PA = 0
207 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
208 * is used in real operation because we can't know actual used
209 * bits from PA, only from on-disk bitmap
211 * if we follow this strict logic, then all operations above should be atomic.
212 * given some of them can block, we'd have to use something like semaphores
213 * killing performance on high-end SMP hardware. let's try to relax it using
214 * the following knowledge:
215 * 1) if buddy is referenced, it's already initialized
216 * 2) while block is used in buddy and the buddy is referenced,
217 * nobody can re-allocate that block
218 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
219 * bit set and PA claims same block, it's OK. IOW, one can set bit in
220 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
221 * block
223 * so, now we're building a concurrency table:
224 * - init buddy vs.
225 * - new PA
226 * blocks for PA are allocated in the buddy, buddy must be referenced
227 * until PA is linked to allocation group to avoid concurrent buddy init
228 * - use inode PA
229 * we need to make sure that either on-disk bitmap or PA has uptodate data
230 * given (3) we care that PA-=N operation doesn't interfere with init
231 * - discard inode PA
232 * the simplest way would be to have buddy initialized by the discard
233 * - use locality group PA
234 * again PA-=N must be serialized with init
235 * - discard locality group PA
236 * the simplest way would be to have buddy initialized by the discard
237 * - new PA vs.
238 * - use inode PA
239 * i_data_sem serializes them
240 * - discard inode PA
241 * discard process must wait until PA isn't used by another process
242 * - use locality group PA
243 * some mutex should serialize them
244 * - discard locality group PA
245 * discard process must wait until PA isn't used by another process
246 * - use inode PA
247 * - use inode PA
248 * i_data_sem or another mutex should serializes them
249 * - discard inode PA
250 * discard process must wait until PA isn't used by another process
251 * - use locality group PA
252 * nothing wrong here -- they're different PAs covering different blocks
253 * - discard locality group PA
254 * discard process must wait until PA isn't used by another process
256 * now we're ready to make few consequences:
257 * - PA is referenced and while it is no discard is possible
258 * - PA is referenced until block isn't marked in on-disk bitmap
259 * - PA changes only after on-disk bitmap
260 * - discard must not compete with init. either init is done before
261 * any discard or they're serialized somehow
262 * - buddy init as sum of on-disk bitmap and PAs is done atomically
264 * a special case when we've used PA to emptiness. no need to modify buddy
265 * in this case, but we should care about concurrent init
270 * Logic in few words:
272 * - allocation:
273 * load group
274 * find blocks
275 * mark bits in on-disk bitmap
276 * release group
278 * - use preallocation:
279 * find proper PA (per-inode or group)
280 * load group
281 * mark bits in on-disk bitmap
282 * release group
283 * release PA
285 * - free:
286 * load group
287 * mark bits in on-disk bitmap
288 * release group
290 * - discard preallocations in group:
291 * mark PAs deleted
292 * move them onto local list
293 * load on-disk bitmap
294 * load group
295 * remove PA from object (inode or locality group)
296 * mark free blocks in-core
298 * - discard inode's preallocations:
302 * Locking rules
304 * Locks:
305 * - bitlock on a group (group)
306 * - object (inode/locality) (object)
307 * - per-pa lock (pa)
309 * Paths:
310 * - new pa
311 * object
312 * group
314 * - find and use pa:
315 * pa
317 * - release consumed pa:
318 * pa
319 * group
320 * object
322 * - generate in-core bitmap:
323 * group
324 * pa
326 * - discard all for given object (inode, locality group):
327 * object
328 * pa
329 * group
331 * - discard all for given group:
332 * group
333 * pa
334 * group
335 * object
338 static struct kmem_cache *ext4_pspace_cachep;
339 static struct kmem_cache *ext4_ac_cachep;
340 static struct kmem_cache *ext4_free_ext_cachep;
342 /* We create slab caches for groupinfo data structures based on the
343 * superblock block size. There will be one per mounted filesystem for
344 * each unique s_blocksize_bits */
345 #define NR_GRPINFO_CACHES \
346 (EXT4_MAX_BLOCK_LOG_SIZE - EXT4_MIN_BLOCK_LOG_SIZE + 1)
347 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
349 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
350 ext4_group_t group);
351 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
352 ext4_group_t group);
353 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
355 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
357 #if BITS_PER_LONG == 64
358 *bit += ((unsigned long) addr & 7UL) << 3;
359 addr = (void *) ((unsigned long) addr & ~7UL);
360 #elif BITS_PER_LONG == 32
361 *bit += ((unsigned long) addr & 3UL) << 3;
362 addr = (void *) ((unsigned long) addr & ~3UL);
363 #else
364 #error "how many bits you are?!"
365 #endif
366 return addr;
369 static inline int mb_test_bit(int bit, void *addr)
372 * ext4_test_bit on architecture like powerpc
373 * needs unsigned long aligned address
375 addr = mb_correct_addr_and_bit(&bit, addr);
376 return ext4_test_bit(bit, addr);
379 static inline void mb_set_bit(int bit, void *addr)
381 addr = mb_correct_addr_and_bit(&bit, addr);
382 ext4_set_bit(bit, addr);
385 static inline void mb_clear_bit(int bit, void *addr)
387 addr = mb_correct_addr_and_bit(&bit, addr);
388 ext4_clear_bit(bit, addr);
391 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
393 int fix = 0, ret, tmpmax;
394 addr = mb_correct_addr_and_bit(&fix, addr);
395 tmpmax = max + fix;
396 start += fix;
398 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
399 if (ret > max)
400 return max;
401 return ret;
404 static inline int mb_find_next_bit(void *addr, int max, int start)
406 int fix = 0, ret, tmpmax;
407 addr = mb_correct_addr_and_bit(&fix, addr);
408 tmpmax = max + fix;
409 start += fix;
411 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
412 if (ret > max)
413 return max;
414 return ret;
417 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
419 char *bb;
421 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
422 BUG_ON(max == NULL);
424 if (order > e4b->bd_blkbits + 1) {
425 *max = 0;
426 return NULL;
429 /* at order 0 we see each particular block */
430 *max = 1 << (e4b->bd_blkbits + 3);
431 if (order == 0)
432 return EXT4_MB_BITMAP(e4b);
434 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
435 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
437 return bb;
440 #ifdef DOUBLE_CHECK
441 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
442 int first, int count)
444 int i;
445 struct super_block *sb = e4b->bd_sb;
447 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
448 return;
449 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
450 for (i = 0; i < count; i++) {
451 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
452 ext4_fsblk_t blocknr;
454 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
455 blocknr += first + i;
456 ext4_grp_locked_error(sb, e4b->bd_group,
457 inode ? inode->i_ino : 0,
458 blocknr,
459 "freeing block already freed "
460 "(bit %u)",
461 first + i);
463 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
467 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
469 int i;
471 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
472 return;
473 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
474 for (i = 0; i < count; i++) {
475 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
476 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
480 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
482 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
483 unsigned char *b1, *b2;
484 int i;
485 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
486 b2 = (unsigned char *) bitmap;
487 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
488 if (b1[i] != b2[i]) {
489 printk(KERN_ERR "corruption in group %u "
490 "at byte %u(%u): %x in copy != %x "
491 "on disk/prealloc\n",
492 e4b->bd_group, i, i * 8, b1[i], b2[i]);
493 BUG();
499 #else
500 static inline void mb_free_blocks_double(struct inode *inode,
501 struct ext4_buddy *e4b, int first, int count)
503 return;
505 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
506 int first, int count)
508 return;
510 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
512 return;
514 #endif
516 #ifdef AGGRESSIVE_CHECK
518 #define MB_CHECK_ASSERT(assert) \
519 do { \
520 if (!(assert)) { \
521 printk(KERN_EMERG \
522 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
523 function, file, line, # assert); \
524 BUG(); \
526 } while (0)
528 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
529 const char *function, int line)
531 struct super_block *sb = e4b->bd_sb;
532 int order = e4b->bd_blkbits + 1;
533 int max;
534 int max2;
535 int i;
536 int j;
537 int k;
538 int count;
539 struct ext4_group_info *grp;
540 int fragments = 0;
541 int fstart;
542 struct list_head *cur;
543 void *buddy;
544 void *buddy2;
547 static int mb_check_counter;
548 if (mb_check_counter++ % 100 != 0)
549 return 0;
552 while (order > 1) {
553 buddy = mb_find_buddy(e4b, order, &max);
554 MB_CHECK_ASSERT(buddy);
555 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
556 MB_CHECK_ASSERT(buddy2);
557 MB_CHECK_ASSERT(buddy != buddy2);
558 MB_CHECK_ASSERT(max * 2 == max2);
560 count = 0;
561 for (i = 0; i < max; i++) {
563 if (mb_test_bit(i, buddy)) {
564 /* only single bit in buddy2 may be 1 */
565 if (!mb_test_bit(i << 1, buddy2)) {
566 MB_CHECK_ASSERT(
567 mb_test_bit((i<<1)+1, buddy2));
568 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
569 MB_CHECK_ASSERT(
570 mb_test_bit(i << 1, buddy2));
572 continue;
575 /* both bits in buddy2 must be 0 */
576 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
577 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
579 for (j = 0; j < (1 << order); j++) {
580 k = (i * (1 << order)) + j;
581 MB_CHECK_ASSERT(
582 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
584 count++;
586 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
587 order--;
590 fstart = -1;
591 buddy = mb_find_buddy(e4b, 0, &max);
592 for (i = 0; i < max; i++) {
593 if (!mb_test_bit(i, buddy)) {
594 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
595 if (fstart == -1) {
596 fragments++;
597 fstart = i;
599 continue;
601 fstart = -1;
602 /* check used bits only */
603 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
604 buddy2 = mb_find_buddy(e4b, j, &max2);
605 k = i >> j;
606 MB_CHECK_ASSERT(k < max2);
607 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
610 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
611 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
613 grp = ext4_get_group_info(sb, e4b->bd_group);
614 buddy = mb_find_buddy(e4b, 0, &max);
615 list_for_each(cur, &grp->bb_prealloc_list) {
616 ext4_group_t groupnr;
617 struct ext4_prealloc_space *pa;
618 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
619 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
620 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
621 for (i = 0; i < pa->pa_len; i++)
622 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
624 return 0;
626 #undef MB_CHECK_ASSERT
627 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
628 __FILE__, __func__, __LINE__)
629 #else
630 #define mb_check_buddy(e4b)
631 #endif
633 /* FIXME!! need more doc */
634 static void ext4_mb_mark_free_simple(struct super_block *sb,
635 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
636 struct ext4_group_info *grp)
638 struct ext4_sb_info *sbi = EXT4_SB(sb);
639 ext4_grpblk_t min;
640 ext4_grpblk_t max;
641 ext4_grpblk_t chunk;
642 unsigned short border;
644 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
646 border = 2 << sb->s_blocksize_bits;
648 while (len > 0) {
649 /* find how many blocks can be covered since this position */
650 max = ffs(first | border) - 1;
652 /* find how many blocks of power 2 we need to mark */
653 min = fls(len) - 1;
655 if (max < min)
656 min = max;
657 chunk = 1 << min;
659 /* mark multiblock chunks only */
660 grp->bb_counters[min]++;
661 if (min > 0)
662 mb_clear_bit(first >> min,
663 buddy + sbi->s_mb_offsets[min]);
665 len -= chunk;
666 first += chunk;
671 * Cache the order of the largest free extent we have available in this block
672 * group.
674 static void
675 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
677 int i;
678 int bits;
680 grp->bb_largest_free_order = -1; /* uninit */
682 bits = sb->s_blocksize_bits + 1;
683 for (i = bits; i >= 0; i--) {
684 if (grp->bb_counters[i] > 0) {
685 grp->bb_largest_free_order = i;
686 break;
691 static noinline_for_stack
692 void ext4_mb_generate_buddy(struct super_block *sb,
693 void *buddy, void *bitmap, ext4_group_t group)
695 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
696 ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb);
697 ext4_grpblk_t i = 0;
698 ext4_grpblk_t first;
699 ext4_grpblk_t len;
700 unsigned free = 0;
701 unsigned fragments = 0;
702 unsigned long long period = get_cycles();
704 /* initialize buddy from bitmap which is aggregation
705 * of on-disk bitmap and preallocations */
706 i = mb_find_next_zero_bit(bitmap, max, 0);
707 grp->bb_first_free = i;
708 while (i < max) {
709 fragments++;
710 first = i;
711 i = mb_find_next_bit(bitmap, max, i);
712 len = i - first;
713 free += len;
714 if (len > 1)
715 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
716 else
717 grp->bb_counters[0]++;
718 if (i < max)
719 i = mb_find_next_zero_bit(bitmap, max, i);
721 grp->bb_fragments = fragments;
723 if (free != grp->bb_free) {
724 ext4_grp_locked_error(sb, group, 0, 0,
725 "%u blocks in bitmap, %u in gd",
726 free, grp->bb_free);
728 * If we intent to continue, we consider group descritor
729 * corrupt and update bb_free using bitmap value
731 grp->bb_free = free;
733 mb_set_largest_free_order(sb, grp);
735 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
737 period = get_cycles() - period;
738 spin_lock(&EXT4_SB(sb)->s_bal_lock);
739 EXT4_SB(sb)->s_mb_buddies_generated++;
740 EXT4_SB(sb)->s_mb_generation_time += period;
741 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
744 /* The buddy information is attached the buddy cache inode
745 * for convenience. The information regarding each group
746 * is loaded via ext4_mb_load_buddy. The information involve
747 * block bitmap and buddy information. The information are
748 * stored in the inode as
750 * { page }
751 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
754 * one block each for bitmap and buddy information.
755 * So for each group we take up 2 blocks. A page can
756 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
757 * So it can have information regarding groups_per_page which
758 * is blocks_per_page/2
760 * Locking note: This routine takes the block group lock of all groups
761 * for this page; do not hold this lock when calling this routine!
764 static int ext4_mb_init_cache(struct page *page, char *incore)
766 ext4_group_t ngroups;
767 int blocksize;
768 int blocks_per_page;
769 int groups_per_page;
770 int err = 0;
771 int i;
772 ext4_group_t first_group;
773 int first_block;
774 struct super_block *sb;
775 struct buffer_head *bhs;
776 struct buffer_head **bh;
777 struct inode *inode;
778 char *data;
779 char *bitmap;
781 mb_debug(1, "init page %lu\n", page->index);
783 inode = page->mapping->host;
784 sb = inode->i_sb;
785 ngroups = ext4_get_groups_count(sb);
786 blocksize = 1 << inode->i_blkbits;
787 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
789 groups_per_page = blocks_per_page >> 1;
790 if (groups_per_page == 0)
791 groups_per_page = 1;
793 /* allocate buffer_heads to read bitmaps */
794 if (groups_per_page > 1) {
795 err = -ENOMEM;
796 i = sizeof(struct buffer_head *) * groups_per_page;
797 bh = kzalloc(i, GFP_NOFS);
798 if (bh == NULL)
799 goto out;
800 } else
801 bh = &bhs;
803 first_group = page->index * blocks_per_page / 2;
805 /* read all groups the page covers into the cache */
806 for (i = 0; i < groups_per_page; i++) {
807 struct ext4_group_desc *desc;
809 if (first_group + i >= ngroups)
810 break;
812 err = -EIO;
813 desc = ext4_get_group_desc(sb, first_group + i, NULL);
814 if (desc == NULL)
815 goto out;
817 err = -ENOMEM;
818 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
819 if (bh[i] == NULL)
820 goto out;
822 if (bitmap_uptodate(bh[i]))
823 continue;
825 lock_buffer(bh[i]);
826 if (bitmap_uptodate(bh[i])) {
827 unlock_buffer(bh[i]);
828 continue;
830 ext4_lock_group(sb, first_group + i);
831 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
832 ext4_init_block_bitmap(sb, bh[i],
833 first_group + i, desc);
834 set_bitmap_uptodate(bh[i]);
835 set_buffer_uptodate(bh[i]);
836 ext4_unlock_group(sb, first_group + i);
837 unlock_buffer(bh[i]);
838 continue;
840 ext4_unlock_group(sb, first_group + i);
841 if (buffer_uptodate(bh[i])) {
843 * if not uninit if bh is uptodate,
844 * bitmap is also uptodate
846 set_bitmap_uptodate(bh[i]);
847 unlock_buffer(bh[i]);
848 continue;
850 get_bh(bh[i]);
852 * submit the buffer_head for read. We can
853 * safely mark the bitmap as uptodate now.
854 * We do it here so the bitmap uptodate bit
855 * get set with buffer lock held.
857 set_bitmap_uptodate(bh[i]);
858 bh[i]->b_end_io = end_buffer_read_sync;
859 submit_bh(READ, bh[i]);
860 mb_debug(1, "read bitmap for group %u\n", first_group + i);
863 /* wait for I/O completion */
864 for (i = 0; i < groups_per_page && bh[i]; i++)
865 wait_on_buffer(bh[i]);
867 err = -EIO;
868 for (i = 0; i < groups_per_page && bh[i]; i++)
869 if (!buffer_uptodate(bh[i]))
870 goto out;
872 err = 0;
873 first_block = page->index * blocks_per_page;
874 /* init the page */
875 memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
876 for (i = 0; i < blocks_per_page; i++) {
877 int group;
878 struct ext4_group_info *grinfo;
880 group = (first_block + i) >> 1;
881 if (group >= ngroups)
882 break;
885 * data carry information regarding this
886 * particular group in the format specified
887 * above
890 data = page_address(page) + (i * blocksize);
891 bitmap = bh[group - first_group]->b_data;
894 * We place the buddy block and bitmap block
895 * close together
897 if ((first_block + i) & 1) {
898 /* this is block of buddy */
899 BUG_ON(incore == NULL);
900 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
901 group, page->index, i * blocksize);
902 trace_ext4_mb_buddy_bitmap_load(sb, group);
903 grinfo = ext4_get_group_info(sb, group);
904 grinfo->bb_fragments = 0;
905 memset(grinfo->bb_counters, 0,
906 sizeof(*grinfo->bb_counters) *
907 (sb->s_blocksize_bits+2));
909 * incore got set to the group block bitmap below
911 ext4_lock_group(sb, group);
912 ext4_mb_generate_buddy(sb, data, incore, group);
913 ext4_unlock_group(sb, group);
914 incore = NULL;
915 } else {
916 /* this is block of bitmap */
917 BUG_ON(incore != NULL);
918 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
919 group, page->index, i * blocksize);
920 trace_ext4_mb_bitmap_load(sb, group);
922 /* see comments in ext4_mb_put_pa() */
923 ext4_lock_group(sb, group);
924 memcpy(data, bitmap, blocksize);
926 /* mark all preallocated blks used in in-core bitmap */
927 ext4_mb_generate_from_pa(sb, data, group);
928 ext4_mb_generate_from_freelist(sb, data, group);
929 ext4_unlock_group(sb, group);
931 /* set incore so that the buddy information can be
932 * generated using this
934 incore = data;
937 SetPageUptodate(page);
939 out:
940 if (bh) {
941 for (i = 0; i < groups_per_page && bh[i]; i++)
942 brelse(bh[i]);
943 if (bh != &bhs)
944 kfree(bh);
946 return err;
950 * lock the group_info alloc_sem of all the groups
951 * belonging to the same buddy cache page. This
952 * make sure other parallel operation on the buddy
953 * cache doesn't happen whild holding the buddy cache
954 * lock
956 static int ext4_mb_get_buddy_cache_lock(struct super_block *sb,
957 ext4_group_t group)
959 int i;
960 int block, pnum;
961 int blocks_per_page;
962 int groups_per_page;
963 ext4_group_t ngroups = ext4_get_groups_count(sb);
964 ext4_group_t first_group;
965 struct ext4_group_info *grp;
967 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
969 * the buddy cache inode stores the block bitmap
970 * and buddy information in consecutive blocks.
971 * So for each group we need two blocks.
973 block = group * 2;
974 pnum = block / blocks_per_page;
975 first_group = pnum * blocks_per_page / 2;
977 groups_per_page = blocks_per_page >> 1;
978 if (groups_per_page == 0)
979 groups_per_page = 1;
980 /* read all groups the page covers into the cache */
981 for (i = 0; i < groups_per_page; i++) {
983 if ((first_group + i) >= ngroups)
984 break;
985 grp = ext4_get_group_info(sb, first_group + i);
986 /* take all groups write allocation
987 * semaphore. This make sure there is
988 * no block allocation going on in any
989 * of that groups
991 down_write_nested(&grp->alloc_sem, i);
993 return i;
996 static void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
997 ext4_group_t group, int locked_group)
999 int i;
1000 int block, pnum;
1001 int blocks_per_page;
1002 ext4_group_t first_group;
1003 struct ext4_group_info *grp;
1005 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1007 * the buddy cache inode stores the block bitmap
1008 * and buddy information in consecutive blocks.
1009 * So for each group we need two blocks.
1011 block = group * 2;
1012 pnum = block / blocks_per_page;
1013 first_group = pnum * blocks_per_page / 2;
1014 /* release locks on all the groups */
1015 for (i = 0; i < locked_group; i++) {
1017 grp = ext4_get_group_info(sb, first_group + i);
1018 /* take all groups write allocation
1019 * semaphore. This make sure there is
1020 * no block allocation going on in any
1021 * of that groups
1023 up_write(&grp->alloc_sem);
1029 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1030 * block group lock of all groups for this page; do not hold the BG lock when
1031 * calling this routine!
1033 static noinline_for_stack
1034 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1037 int ret = 0;
1038 void *bitmap;
1039 int blocks_per_page;
1040 int block, pnum, poff;
1041 int num_grp_locked = 0;
1042 struct ext4_group_info *this_grp;
1043 struct ext4_sb_info *sbi = EXT4_SB(sb);
1044 struct inode *inode = sbi->s_buddy_cache;
1045 struct page *page = NULL, *bitmap_page = NULL;
1047 mb_debug(1, "init group %u\n", group);
1048 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1049 this_grp = ext4_get_group_info(sb, group);
1051 * This ensures that we don't reinit the buddy cache
1052 * page which map to the group from which we are already
1053 * allocating. If we are looking at the buddy cache we would
1054 * have taken a reference using ext4_mb_load_buddy and that
1055 * would have taken the alloc_sem lock.
1057 num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group);
1058 if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
1060 * somebody initialized the group
1061 * return without doing anything
1063 ret = 0;
1064 goto err;
1067 * the buddy cache inode stores the block bitmap
1068 * and buddy information in consecutive blocks.
1069 * So for each group we need two blocks.
1071 block = group * 2;
1072 pnum = block / blocks_per_page;
1073 poff = block % blocks_per_page;
1074 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1075 if (page) {
1076 BUG_ON(page->mapping != inode->i_mapping);
1077 ret = ext4_mb_init_cache(page, NULL);
1078 if (ret) {
1079 unlock_page(page);
1080 goto err;
1082 unlock_page(page);
1084 if (page == NULL || !PageUptodate(page)) {
1085 ret = -EIO;
1086 goto err;
1088 mark_page_accessed(page);
1089 bitmap_page = page;
1090 bitmap = page_address(page) + (poff * sb->s_blocksize);
1092 /* init buddy cache */
1093 block++;
1094 pnum = block / blocks_per_page;
1095 poff = block % blocks_per_page;
1096 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1097 if (page == bitmap_page) {
1099 * If both the bitmap and buddy are in
1100 * the same page we don't need to force
1101 * init the buddy
1103 unlock_page(page);
1104 } else if (page) {
1105 BUG_ON(page->mapping != inode->i_mapping);
1106 ret = ext4_mb_init_cache(page, bitmap);
1107 if (ret) {
1108 unlock_page(page);
1109 goto err;
1111 unlock_page(page);
1113 if (page == NULL || !PageUptodate(page)) {
1114 ret = -EIO;
1115 goto err;
1117 mark_page_accessed(page);
1118 err:
1119 ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
1120 if (bitmap_page)
1121 page_cache_release(bitmap_page);
1122 if (page)
1123 page_cache_release(page);
1124 return ret;
1128 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1129 * block group lock of all groups for this page; do not hold the BG lock when
1130 * calling this routine!
1132 static noinline_for_stack int
1133 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1134 struct ext4_buddy *e4b)
1136 int blocks_per_page;
1137 int block;
1138 int pnum;
1139 int poff;
1140 struct page *page;
1141 int ret;
1142 struct ext4_group_info *grp;
1143 struct ext4_sb_info *sbi = EXT4_SB(sb);
1144 struct inode *inode = sbi->s_buddy_cache;
1146 mb_debug(1, "load group %u\n", group);
1148 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1149 grp = ext4_get_group_info(sb, group);
1151 e4b->bd_blkbits = sb->s_blocksize_bits;
1152 e4b->bd_info = ext4_get_group_info(sb, group);
1153 e4b->bd_sb = sb;
1154 e4b->bd_group = group;
1155 e4b->bd_buddy_page = NULL;
1156 e4b->bd_bitmap_page = NULL;
1157 e4b->alloc_semp = &grp->alloc_sem;
1159 /* Take the read lock on the group alloc
1160 * sem. This would make sure a parallel
1161 * ext4_mb_init_group happening on other
1162 * groups mapped by the page is blocked
1163 * till we are done with allocation
1165 repeat_load_buddy:
1166 down_read(e4b->alloc_semp);
1168 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1169 /* we need to check for group need init flag
1170 * with alloc_semp held so that we can be sure
1171 * that new blocks didn't get added to the group
1172 * when we are loading the buddy cache
1174 up_read(e4b->alloc_semp);
1176 * we need full data about the group
1177 * to make a good selection
1179 ret = ext4_mb_init_group(sb, group);
1180 if (ret)
1181 return ret;
1182 goto repeat_load_buddy;
1186 * the buddy cache inode stores the block bitmap
1187 * and buddy information in consecutive blocks.
1188 * So for each group we need two blocks.
1190 block = group * 2;
1191 pnum = block / blocks_per_page;
1192 poff = block % blocks_per_page;
1194 /* we could use find_or_create_page(), but it locks page
1195 * what we'd like to avoid in fast path ... */
1196 page = find_get_page(inode->i_mapping, pnum);
1197 if (page == NULL || !PageUptodate(page)) {
1198 if (page)
1200 * drop the page reference and try
1201 * to get the page with lock. If we
1202 * are not uptodate that implies
1203 * somebody just created the page but
1204 * is yet to initialize the same. So
1205 * wait for it to initialize.
1207 page_cache_release(page);
1208 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1209 if (page) {
1210 BUG_ON(page->mapping != inode->i_mapping);
1211 if (!PageUptodate(page)) {
1212 ret = ext4_mb_init_cache(page, NULL);
1213 if (ret) {
1214 unlock_page(page);
1215 goto err;
1217 mb_cmp_bitmaps(e4b, page_address(page) +
1218 (poff * sb->s_blocksize));
1220 unlock_page(page);
1223 if (page == NULL || !PageUptodate(page)) {
1224 ret = -EIO;
1225 goto err;
1227 e4b->bd_bitmap_page = page;
1228 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1229 mark_page_accessed(page);
1231 block++;
1232 pnum = block / blocks_per_page;
1233 poff = block % blocks_per_page;
1235 page = find_get_page(inode->i_mapping, pnum);
1236 if (page == NULL || !PageUptodate(page)) {
1237 if (page)
1238 page_cache_release(page);
1239 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1240 if (page) {
1241 BUG_ON(page->mapping != inode->i_mapping);
1242 if (!PageUptodate(page)) {
1243 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1244 if (ret) {
1245 unlock_page(page);
1246 goto err;
1249 unlock_page(page);
1252 if (page == NULL || !PageUptodate(page)) {
1253 ret = -EIO;
1254 goto err;
1256 e4b->bd_buddy_page = page;
1257 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1258 mark_page_accessed(page);
1260 BUG_ON(e4b->bd_bitmap_page == NULL);
1261 BUG_ON(e4b->bd_buddy_page == NULL);
1263 return 0;
1265 err:
1266 if (e4b->bd_bitmap_page)
1267 page_cache_release(e4b->bd_bitmap_page);
1268 if (e4b->bd_buddy_page)
1269 page_cache_release(e4b->bd_buddy_page);
1270 e4b->bd_buddy = NULL;
1271 e4b->bd_bitmap = NULL;
1273 /* Done with the buddy cache */
1274 up_read(e4b->alloc_semp);
1275 return ret;
1278 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1280 if (e4b->bd_bitmap_page)
1281 page_cache_release(e4b->bd_bitmap_page);
1282 if (e4b->bd_buddy_page)
1283 page_cache_release(e4b->bd_buddy_page);
1284 /* Done with the buddy cache */
1285 if (e4b->alloc_semp)
1286 up_read(e4b->alloc_semp);
1290 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1292 int order = 1;
1293 void *bb;
1295 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1296 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1298 bb = EXT4_MB_BUDDY(e4b);
1299 while (order <= e4b->bd_blkbits + 1) {
1300 block = block >> 1;
1301 if (!mb_test_bit(block, bb)) {
1302 /* this block is part of buddy of order 'order' */
1303 return order;
1305 bb += 1 << (e4b->bd_blkbits - order);
1306 order++;
1308 return 0;
1311 static void mb_clear_bits(void *bm, int cur, int len)
1313 __u32 *addr;
1315 len = cur + len;
1316 while (cur < len) {
1317 if ((cur & 31) == 0 && (len - cur) >= 32) {
1318 /* fast path: clear whole word at once */
1319 addr = bm + (cur >> 3);
1320 *addr = 0;
1321 cur += 32;
1322 continue;
1324 mb_clear_bit(cur, bm);
1325 cur++;
1329 static void mb_set_bits(void *bm, int cur, int len)
1331 __u32 *addr;
1333 len = cur + len;
1334 while (cur < len) {
1335 if ((cur & 31) == 0 && (len - cur) >= 32) {
1336 /* fast path: set whole word at once */
1337 addr = bm + (cur >> 3);
1338 *addr = 0xffffffff;
1339 cur += 32;
1340 continue;
1342 mb_set_bit(cur, bm);
1343 cur++;
1347 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1348 int first, int count)
1350 int block = 0;
1351 int max = 0;
1352 int order;
1353 void *buddy;
1354 void *buddy2;
1355 struct super_block *sb = e4b->bd_sb;
1357 BUG_ON(first + count > (sb->s_blocksize << 3));
1358 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1359 mb_check_buddy(e4b);
1360 mb_free_blocks_double(inode, e4b, first, count);
1362 e4b->bd_info->bb_free += count;
1363 if (first < e4b->bd_info->bb_first_free)
1364 e4b->bd_info->bb_first_free = first;
1366 /* let's maintain fragments counter */
1367 if (first != 0)
1368 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1369 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1370 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1371 if (block && max)
1372 e4b->bd_info->bb_fragments--;
1373 else if (!block && !max)
1374 e4b->bd_info->bb_fragments++;
1376 /* let's maintain buddy itself */
1377 while (count-- > 0) {
1378 block = first++;
1379 order = 0;
1381 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1382 ext4_fsblk_t blocknr;
1384 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1385 blocknr += block;
1386 ext4_grp_locked_error(sb, e4b->bd_group,
1387 inode ? inode->i_ino : 0,
1388 blocknr,
1389 "freeing already freed block "
1390 "(bit %u)", block);
1392 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1393 e4b->bd_info->bb_counters[order]++;
1395 /* start of the buddy */
1396 buddy = mb_find_buddy(e4b, order, &max);
1398 do {
1399 block &= ~1UL;
1400 if (mb_test_bit(block, buddy) ||
1401 mb_test_bit(block + 1, buddy))
1402 break;
1404 /* both the buddies are free, try to coalesce them */
1405 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1407 if (!buddy2)
1408 break;
1410 if (order > 0) {
1411 /* for special purposes, we don't set
1412 * free bits in bitmap */
1413 mb_set_bit(block, buddy);
1414 mb_set_bit(block + 1, buddy);
1416 e4b->bd_info->bb_counters[order]--;
1417 e4b->bd_info->bb_counters[order]--;
1419 block = block >> 1;
1420 order++;
1421 e4b->bd_info->bb_counters[order]++;
1423 mb_clear_bit(block, buddy2);
1424 buddy = buddy2;
1425 } while (1);
1427 mb_set_largest_free_order(sb, e4b->bd_info);
1428 mb_check_buddy(e4b);
1431 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1432 int needed, struct ext4_free_extent *ex)
1434 int next = block;
1435 int max;
1436 int ord;
1437 void *buddy;
1439 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1440 BUG_ON(ex == NULL);
1442 buddy = mb_find_buddy(e4b, order, &max);
1443 BUG_ON(buddy == NULL);
1444 BUG_ON(block >= max);
1445 if (mb_test_bit(block, buddy)) {
1446 ex->fe_len = 0;
1447 ex->fe_start = 0;
1448 ex->fe_group = 0;
1449 return 0;
1452 /* FIXME dorp order completely ? */
1453 if (likely(order == 0)) {
1454 /* find actual order */
1455 order = mb_find_order_for_block(e4b, block);
1456 block = block >> order;
1459 ex->fe_len = 1 << order;
1460 ex->fe_start = block << order;
1461 ex->fe_group = e4b->bd_group;
1463 /* calc difference from given start */
1464 next = next - ex->fe_start;
1465 ex->fe_len -= next;
1466 ex->fe_start += next;
1468 while (needed > ex->fe_len &&
1469 (buddy = mb_find_buddy(e4b, order, &max))) {
1471 if (block + 1 >= max)
1472 break;
1474 next = (block + 1) * (1 << order);
1475 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1476 break;
1478 ord = mb_find_order_for_block(e4b, next);
1480 order = ord;
1481 block = next >> order;
1482 ex->fe_len += 1 << order;
1485 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1486 return ex->fe_len;
1489 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1491 int ord;
1492 int mlen = 0;
1493 int max = 0;
1494 int cur;
1495 int start = ex->fe_start;
1496 int len = ex->fe_len;
1497 unsigned ret = 0;
1498 int len0 = len;
1499 void *buddy;
1501 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1502 BUG_ON(e4b->bd_group != ex->fe_group);
1503 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1504 mb_check_buddy(e4b);
1505 mb_mark_used_double(e4b, start, len);
1507 e4b->bd_info->bb_free -= len;
1508 if (e4b->bd_info->bb_first_free == start)
1509 e4b->bd_info->bb_first_free += len;
1511 /* let's maintain fragments counter */
1512 if (start != 0)
1513 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1514 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1515 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1516 if (mlen && max)
1517 e4b->bd_info->bb_fragments++;
1518 else if (!mlen && !max)
1519 e4b->bd_info->bb_fragments--;
1521 /* let's maintain buddy itself */
1522 while (len) {
1523 ord = mb_find_order_for_block(e4b, start);
1525 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1526 /* the whole chunk may be allocated at once! */
1527 mlen = 1 << ord;
1528 buddy = mb_find_buddy(e4b, ord, &max);
1529 BUG_ON((start >> ord) >= max);
1530 mb_set_bit(start >> ord, buddy);
1531 e4b->bd_info->bb_counters[ord]--;
1532 start += mlen;
1533 len -= mlen;
1534 BUG_ON(len < 0);
1535 continue;
1538 /* store for history */
1539 if (ret == 0)
1540 ret = len | (ord << 16);
1542 /* we have to split large buddy */
1543 BUG_ON(ord <= 0);
1544 buddy = mb_find_buddy(e4b, ord, &max);
1545 mb_set_bit(start >> ord, buddy);
1546 e4b->bd_info->bb_counters[ord]--;
1548 ord--;
1549 cur = (start >> ord) & ~1U;
1550 buddy = mb_find_buddy(e4b, ord, &max);
1551 mb_clear_bit(cur, buddy);
1552 mb_clear_bit(cur + 1, buddy);
1553 e4b->bd_info->bb_counters[ord]++;
1554 e4b->bd_info->bb_counters[ord]++;
1556 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1558 mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1559 mb_check_buddy(e4b);
1561 return ret;
1565 * Must be called under group lock!
1567 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1568 struct ext4_buddy *e4b)
1570 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1571 int ret;
1573 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1574 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1576 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1577 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1578 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1580 /* preallocation can change ac_b_ex, thus we store actually
1581 * allocated blocks for history */
1582 ac->ac_f_ex = ac->ac_b_ex;
1584 ac->ac_status = AC_STATUS_FOUND;
1585 ac->ac_tail = ret & 0xffff;
1586 ac->ac_buddy = ret >> 16;
1589 * take the page reference. We want the page to be pinned
1590 * so that we don't get a ext4_mb_init_cache_call for this
1591 * group until we update the bitmap. That would mean we
1592 * double allocate blocks. The reference is dropped
1593 * in ext4_mb_release_context
1595 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1596 get_page(ac->ac_bitmap_page);
1597 ac->ac_buddy_page = e4b->bd_buddy_page;
1598 get_page(ac->ac_buddy_page);
1599 /* on allocation we use ac to track the held semaphore */
1600 ac->alloc_semp = e4b->alloc_semp;
1601 e4b->alloc_semp = NULL;
1602 /* store last allocated for subsequent stream allocation */
1603 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1604 spin_lock(&sbi->s_md_lock);
1605 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1606 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1607 spin_unlock(&sbi->s_md_lock);
1612 * regular allocator, for general purposes allocation
1615 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1616 struct ext4_buddy *e4b,
1617 int finish_group)
1619 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1620 struct ext4_free_extent *bex = &ac->ac_b_ex;
1621 struct ext4_free_extent *gex = &ac->ac_g_ex;
1622 struct ext4_free_extent ex;
1623 int max;
1625 if (ac->ac_status == AC_STATUS_FOUND)
1626 return;
1628 * We don't want to scan for a whole year
1630 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1631 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1632 ac->ac_status = AC_STATUS_BREAK;
1633 return;
1637 * Haven't found good chunk so far, let's continue
1639 if (bex->fe_len < gex->fe_len)
1640 return;
1642 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1643 && bex->fe_group == e4b->bd_group) {
1644 /* recheck chunk's availability - we don't know
1645 * when it was found (within this lock-unlock
1646 * period or not) */
1647 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1648 if (max >= gex->fe_len) {
1649 ext4_mb_use_best_found(ac, e4b);
1650 return;
1656 * The routine checks whether found extent is good enough. If it is,
1657 * then the extent gets marked used and flag is set to the context
1658 * to stop scanning. Otherwise, the extent is compared with the
1659 * previous found extent and if new one is better, then it's stored
1660 * in the context. Later, the best found extent will be used, if
1661 * mballoc can't find good enough extent.
1663 * FIXME: real allocation policy is to be designed yet!
1665 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1666 struct ext4_free_extent *ex,
1667 struct ext4_buddy *e4b)
1669 struct ext4_free_extent *bex = &ac->ac_b_ex;
1670 struct ext4_free_extent *gex = &ac->ac_g_ex;
1672 BUG_ON(ex->fe_len <= 0);
1673 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1674 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1675 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1677 ac->ac_found++;
1680 * The special case - take what you catch first
1682 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1683 *bex = *ex;
1684 ext4_mb_use_best_found(ac, e4b);
1685 return;
1689 * Let's check whether the chuck is good enough
1691 if (ex->fe_len == gex->fe_len) {
1692 *bex = *ex;
1693 ext4_mb_use_best_found(ac, e4b);
1694 return;
1698 * If this is first found extent, just store it in the context
1700 if (bex->fe_len == 0) {
1701 *bex = *ex;
1702 return;
1706 * If new found extent is better, store it in the context
1708 if (bex->fe_len < gex->fe_len) {
1709 /* if the request isn't satisfied, any found extent
1710 * larger than previous best one is better */
1711 if (ex->fe_len > bex->fe_len)
1712 *bex = *ex;
1713 } else if (ex->fe_len > gex->fe_len) {
1714 /* if the request is satisfied, then we try to find
1715 * an extent that still satisfy the request, but is
1716 * smaller than previous one */
1717 if (ex->fe_len < bex->fe_len)
1718 *bex = *ex;
1721 ext4_mb_check_limits(ac, e4b, 0);
1724 static noinline_for_stack
1725 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1726 struct ext4_buddy *e4b)
1728 struct ext4_free_extent ex = ac->ac_b_ex;
1729 ext4_group_t group = ex.fe_group;
1730 int max;
1731 int err;
1733 BUG_ON(ex.fe_len <= 0);
1734 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1735 if (err)
1736 return err;
1738 ext4_lock_group(ac->ac_sb, group);
1739 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1741 if (max > 0) {
1742 ac->ac_b_ex = ex;
1743 ext4_mb_use_best_found(ac, e4b);
1746 ext4_unlock_group(ac->ac_sb, group);
1747 ext4_mb_unload_buddy(e4b);
1749 return 0;
1752 static noinline_for_stack
1753 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1754 struct ext4_buddy *e4b)
1756 ext4_group_t group = ac->ac_g_ex.fe_group;
1757 int max;
1758 int err;
1759 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1760 struct ext4_free_extent ex;
1762 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1763 return 0;
1765 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1766 if (err)
1767 return err;
1769 ext4_lock_group(ac->ac_sb, group);
1770 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1771 ac->ac_g_ex.fe_len, &ex);
1773 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1774 ext4_fsblk_t start;
1776 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1777 ex.fe_start;
1778 /* use do_div to get remainder (would be 64-bit modulo) */
1779 if (do_div(start, sbi->s_stripe) == 0) {
1780 ac->ac_found++;
1781 ac->ac_b_ex = ex;
1782 ext4_mb_use_best_found(ac, e4b);
1784 } else if (max >= ac->ac_g_ex.fe_len) {
1785 BUG_ON(ex.fe_len <= 0);
1786 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1787 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1788 ac->ac_found++;
1789 ac->ac_b_ex = ex;
1790 ext4_mb_use_best_found(ac, e4b);
1791 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1792 /* Sometimes, caller may want to merge even small
1793 * number of blocks to an existing extent */
1794 BUG_ON(ex.fe_len <= 0);
1795 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1796 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1797 ac->ac_found++;
1798 ac->ac_b_ex = ex;
1799 ext4_mb_use_best_found(ac, e4b);
1801 ext4_unlock_group(ac->ac_sb, group);
1802 ext4_mb_unload_buddy(e4b);
1804 return 0;
1808 * The routine scans buddy structures (not bitmap!) from given order
1809 * to max order and tries to find big enough chunk to satisfy the req
1811 static noinline_for_stack
1812 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1813 struct ext4_buddy *e4b)
1815 struct super_block *sb = ac->ac_sb;
1816 struct ext4_group_info *grp = e4b->bd_info;
1817 void *buddy;
1818 int i;
1819 int k;
1820 int max;
1822 BUG_ON(ac->ac_2order <= 0);
1823 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1824 if (grp->bb_counters[i] == 0)
1825 continue;
1827 buddy = mb_find_buddy(e4b, i, &max);
1828 BUG_ON(buddy == NULL);
1830 k = mb_find_next_zero_bit(buddy, max, 0);
1831 BUG_ON(k >= max);
1833 ac->ac_found++;
1835 ac->ac_b_ex.fe_len = 1 << i;
1836 ac->ac_b_ex.fe_start = k << i;
1837 ac->ac_b_ex.fe_group = e4b->bd_group;
1839 ext4_mb_use_best_found(ac, e4b);
1841 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1843 if (EXT4_SB(sb)->s_mb_stats)
1844 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1846 break;
1851 * The routine scans the group and measures all found extents.
1852 * In order to optimize scanning, caller must pass number of
1853 * free blocks in the group, so the routine can know upper limit.
1855 static noinline_for_stack
1856 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1857 struct ext4_buddy *e4b)
1859 struct super_block *sb = ac->ac_sb;
1860 void *bitmap = EXT4_MB_BITMAP(e4b);
1861 struct ext4_free_extent ex;
1862 int i;
1863 int free;
1865 free = e4b->bd_info->bb_free;
1866 BUG_ON(free <= 0);
1868 i = e4b->bd_info->bb_first_free;
1870 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1871 i = mb_find_next_zero_bit(bitmap,
1872 EXT4_BLOCKS_PER_GROUP(sb), i);
1873 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1875 * IF we have corrupt bitmap, we won't find any
1876 * free blocks even though group info says we
1877 * we have free blocks
1879 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1880 "%d free blocks as per "
1881 "group info. But bitmap says 0",
1882 free);
1883 break;
1886 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1887 BUG_ON(ex.fe_len <= 0);
1888 if (free < ex.fe_len) {
1889 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1890 "%d free blocks as per "
1891 "group info. But got %d blocks",
1892 free, ex.fe_len);
1894 * The number of free blocks differs. This mostly
1895 * indicate that the bitmap is corrupt. So exit
1896 * without claiming the space.
1898 break;
1901 ext4_mb_measure_extent(ac, &ex, e4b);
1903 i += ex.fe_len;
1904 free -= ex.fe_len;
1907 ext4_mb_check_limits(ac, e4b, 1);
1911 * This is a special case for storages like raid5
1912 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1914 static noinline_for_stack
1915 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1916 struct ext4_buddy *e4b)
1918 struct super_block *sb = ac->ac_sb;
1919 struct ext4_sb_info *sbi = EXT4_SB(sb);
1920 void *bitmap = EXT4_MB_BITMAP(e4b);
1921 struct ext4_free_extent ex;
1922 ext4_fsblk_t first_group_block;
1923 ext4_fsblk_t a;
1924 ext4_grpblk_t i;
1925 int max;
1927 BUG_ON(sbi->s_stripe == 0);
1929 /* find first stripe-aligned block in group */
1930 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1932 a = first_group_block + sbi->s_stripe - 1;
1933 do_div(a, sbi->s_stripe);
1934 i = (a * sbi->s_stripe) - first_group_block;
1936 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1937 if (!mb_test_bit(i, bitmap)) {
1938 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1939 if (max >= sbi->s_stripe) {
1940 ac->ac_found++;
1941 ac->ac_b_ex = ex;
1942 ext4_mb_use_best_found(ac, e4b);
1943 break;
1946 i += sbi->s_stripe;
1950 /* This is now called BEFORE we load the buddy bitmap. */
1951 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1952 ext4_group_t group, int cr)
1954 unsigned free, fragments;
1955 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1956 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1958 BUG_ON(cr < 0 || cr >= 4);
1960 /* We only do this if the grp has never been initialized */
1961 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1962 int ret = ext4_mb_init_group(ac->ac_sb, group);
1963 if (ret)
1964 return 0;
1967 free = grp->bb_free;
1968 fragments = grp->bb_fragments;
1969 if (free == 0)
1970 return 0;
1971 if (fragments == 0)
1972 return 0;
1974 switch (cr) {
1975 case 0:
1976 BUG_ON(ac->ac_2order == 0);
1978 if (grp->bb_largest_free_order < ac->ac_2order)
1979 return 0;
1981 /* Avoid using the first bg of a flexgroup for data files */
1982 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1983 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1984 ((group % flex_size) == 0))
1985 return 0;
1987 return 1;
1988 case 1:
1989 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1990 return 1;
1991 break;
1992 case 2:
1993 if (free >= ac->ac_g_ex.fe_len)
1994 return 1;
1995 break;
1996 case 3:
1997 return 1;
1998 default:
1999 BUG();
2002 return 0;
2005 static noinline_for_stack int
2006 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2008 ext4_group_t ngroups, group, i;
2009 int cr;
2010 int err = 0;
2011 struct ext4_sb_info *sbi;
2012 struct super_block *sb;
2013 struct ext4_buddy e4b;
2015 sb = ac->ac_sb;
2016 sbi = EXT4_SB(sb);
2017 ngroups = ext4_get_groups_count(sb);
2018 /* non-extent files are limited to low blocks/groups */
2019 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2020 ngroups = sbi->s_blockfile_groups;
2022 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2024 /* first, try the goal */
2025 err = ext4_mb_find_by_goal(ac, &e4b);
2026 if (err || ac->ac_status == AC_STATUS_FOUND)
2027 goto out;
2029 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2030 goto out;
2033 * ac->ac2_order is set only if the fe_len is a power of 2
2034 * if ac2_order is set we also set criteria to 0 so that we
2035 * try exact allocation using buddy.
2037 i = fls(ac->ac_g_ex.fe_len);
2038 ac->ac_2order = 0;
2040 * We search using buddy data only if the order of the request
2041 * is greater than equal to the sbi_s_mb_order2_reqs
2042 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2044 if (i >= sbi->s_mb_order2_reqs) {
2046 * This should tell if fe_len is exactly power of 2
2048 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2049 ac->ac_2order = i - 1;
2052 /* if stream allocation is enabled, use global goal */
2053 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2054 /* TBD: may be hot point */
2055 spin_lock(&sbi->s_md_lock);
2056 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2057 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2058 spin_unlock(&sbi->s_md_lock);
2061 /* Let's just scan groups to find more-less suitable blocks */
2062 cr = ac->ac_2order ? 0 : 1;
2064 * cr == 0 try to get exact allocation,
2065 * cr == 3 try to get anything
2067 repeat:
2068 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2069 ac->ac_criteria = cr;
2071 * searching for the right group start
2072 * from the goal value specified
2074 group = ac->ac_g_ex.fe_group;
2076 for (i = 0; i < ngroups; group++, i++) {
2077 if (group == ngroups)
2078 group = 0;
2080 /* This now checks without needing the buddy page */
2081 if (!ext4_mb_good_group(ac, group, cr))
2082 continue;
2084 err = ext4_mb_load_buddy(sb, group, &e4b);
2085 if (err)
2086 goto out;
2088 ext4_lock_group(sb, group);
2091 * We need to check again after locking the
2092 * block group
2094 if (!ext4_mb_good_group(ac, group, cr)) {
2095 ext4_unlock_group(sb, group);
2096 ext4_mb_unload_buddy(&e4b);
2097 continue;
2100 ac->ac_groups_scanned++;
2101 if (cr == 0)
2102 ext4_mb_simple_scan_group(ac, &e4b);
2103 else if (cr == 1 && sbi->s_stripe &&
2104 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2105 ext4_mb_scan_aligned(ac, &e4b);
2106 else
2107 ext4_mb_complex_scan_group(ac, &e4b);
2109 ext4_unlock_group(sb, group);
2110 ext4_mb_unload_buddy(&e4b);
2112 if (ac->ac_status != AC_STATUS_CONTINUE)
2113 break;
2117 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2118 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2120 * We've been searching too long. Let's try to allocate
2121 * the best chunk we've found so far
2124 ext4_mb_try_best_found(ac, &e4b);
2125 if (ac->ac_status != AC_STATUS_FOUND) {
2127 * Someone more lucky has already allocated it.
2128 * The only thing we can do is just take first
2129 * found block(s)
2130 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2132 ac->ac_b_ex.fe_group = 0;
2133 ac->ac_b_ex.fe_start = 0;
2134 ac->ac_b_ex.fe_len = 0;
2135 ac->ac_status = AC_STATUS_CONTINUE;
2136 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2137 cr = 3;
2138 atomic_inc(&sbi->s_mb_lost_chunks);
2139 goto repeat;
2142 out:
2143 return err;
2146 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2148 struct super_block *sb = seq->private;
2149 ext4_group_t group;
2151 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2152 return NULL;
2153 group = *pos + 1;
2154 return (void *) ((unsigned long) group);
2157 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2159 struct super_block *sb = seq->private;
2160 ext4_group_t group;
2162 ++*pos;
2163 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2164 return NULL;
2165 group = *pos + 1;
2166 return (void *) ((unsigned long) group);
2169 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2171 struct super_block *sb = seq->private;
2172 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2173 int i;
2174 int err;
2175 struct ext4_buddy e4b;
2176 struct sg {
2177 struct ext4_group_info info;
2178 ext4_grpblk_t counters[16];
2179 } sg;
2181 group--;
2182 if (group == 0)
2183 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2184 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2185 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2186 "group", "free", "frags", "first",
2187 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2188 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2190 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2191 sizeof(struct ext4_group_info);
2192 err = ext4_mb_load_buddy(sb, group, &e4b);
2193 if (err) {
2194 seq_printf(seq, "#%-5u: I/O error\n", group);
2195 return 0;
2197 ext4_lock_group(sb, group);
2198 memcpy(&sg, ext4_get_group_info(sb, group), i);
2199 ext4_unlock_group(sb, group);
2200 ext4_mb_unload_buddy(&e4b);
2202 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2203 sg.info.bb_fragments, sg.info.bb_first_free);
2204 for (i = 0; i <= 13; i++)
2205 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2206 sg.info.bb_counters[i] : 0);
2207 seq_printf(seq, " ]\n");
2209 return 0;
2212 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2216 static const struct seq_operations ext4_mb_seq_groups_ops = {
2217 .start = ext4_mb_seq_groups_start,
2218 .next = ext4_mb_seq_groups_next,
2219 .stop = ext4_mb_seq_groups_stop,
2220 .show = ext4_mb_seq_groups_show,
2223 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2225 struct super_block *sb = PDE(inode)->data;
2226 int rc;
2228 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2229 if (rc == 0) {
2230 struct seq_file *m = file->private_data;
2231 m->private = sb;
2233 return rc;
2237 static const struct file_operations ext4_mb_seq_groups_fops = {
2238 .owner = THIS_MODULE,
2239 .open = ext4_mb_seq_groups_open,
2240 .read = seq_read,
2241 .llseek = seq_lseek,
2242 .release = seq_release,
2245 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2247 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2248 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2250 BUG_ON(!cachep);
2251 return cachep;
2254 /* Create and initialize ext4_group_info data for the given group. */
2255 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2256 struct ext4_group_desc *desc)
2258 int i;
2259 int metalen = 0;
2260 struct ext4_sb_info *sbi = EXT4_SB(sb);
2261 struct ext4_group_info **meta_group_info;
2262 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2265 * First check if this group is the first of a reserved block.
2266 * If it's true, we have to allocate a new table of pointers
2267 * to ext4_group_info structures
2269 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2270 metalen = sizeof(*meta_group_info) <<
2271 EXT4_DESC_PER_BLOCK_BITS(sb);
2272 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2273 if (meta_group_info == NULL) {
2274 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2275 "buddy group\n");
2276 goto exit_meta_group_info;
2278 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2279 meta_group_info;
2282 meta_group_info =
2283 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2284 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2286 meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL);
2287 if (meta_group_info[i] == NULL) {
2288 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2289 goto exit_group_info;
2291 memset(meta_group_info[i], 0, kmem_cache_size(cachep));
2292 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2293 &(meta_group_info[i]->bb_state));
2296 * initialize bb_free to be able to skip
2297 * empty groups without initialization
2299 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2300 meta_group_info[i]->bb_free =
2301 ext4_free_blocks_after_init(sb, group, desc);
2302 } else {
2303 meta_group_info[i]->bb_free =
2304 ext4_free_blks_count(sb, desc);
2307 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2308 init_rwsem(&meta_group_info[i]->alloc_sem);
2309 meta_group_info[i]->bb_free_root = RB_ROOT;
2310 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2312 #ifdef DOUBLE_CHECK
2314 struct buffer_head *bh;
2315 meta_group_info[i]->bb_bitmap =
2316 kmalloc(sb->s_blocksize, GFP_KERNEL);
2317 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2318 bh = ext4_read_block_bitmap(sb, group);
2319 BUG_ON(bh == NULL);
2320 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2321 sb->s_blocksize);
2322 put_bh(bh);
2324 #endif
2326 return 0;
2328 exit_group_info:
2329 /* If a meta_group_info table has been allocated, release it now */
2330 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2331 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2332 exit_meta_group_info:
2333 return -ENOMEM;
2334 } /* ext4_mb_add_groupinfo */
2336 static int ext4_mb_init_backend(struct super_block *sb)
2338 ext4_group_t ngroups = ext4_get_groups_count(sb);
2339 ext4_group_t i;
2340 struct ext4_sb_info *sbi = EXT4_SB(sb);
2341 struct ext4_super_block *es = sbi->s_es;
2342 int num_meta_group_infos;
2343 int num_meta_group_infos_max;
2344 int array_size;
2345 struct ext4_group_desc *desc;
2346 struct kmem_cache *cachep;
2348 /* This is the number of blocks used by GDT */
2349 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2350 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2353 * This is the total number of blocks used by GDT including
2354 * the number of reserved blocks for GDT.
2355 * The s_group_info array is allocated with this value
2356 * to allow a clean online resize without a complex
2357 * manipulation of pointer.
2358 * The drawback is the unused memory when no resize
2359 * occurs but it's very low in terms of pages
2360 * (see comments below)
2361 * Need to handle this properly when META_BG resizing is allowed
2363 num_meta_group_infos_max = num_meta_group_infos +
2364 le16_to_cpu(es->s_reserved_gdt_blocks);
2367 * array_size is the size of s_group_info array. We round it
2368 * to the next power of two because this approximation is done
2369 * internally by kmalloc so we can have some more memory
2370 * for free here (e.g. may be used for META_BG resize).
2372 array_size = 1;
2373 while (array_size < sizeof(*sbi->s_group_info) *
2374 num_meta_group_infos_max)
2375 array_size = array_size << 1;
2376 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2377 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2378 * So a two level scheme suffices for now. */
2379 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2380 if (sbi->s_group_info == NULL) {
2381 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2382 return -ENOMEM;
2384 sbi->s_buddy_cache = new_inode(sb);
2385 if (sbi->s_buddy_cache == NULL) {
2386 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2387 goto err_freesgi;
2389 sbi->s_buddy_cache->i_ino = get_next_ino();
2390 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2391 for (i = 0; i < ngroups; i++) {
2392 desc = ext4_get_group_desc(sb, i, NULL);
2393 if (desc == NULL) {
2394 printk(KERN_ERR
2395 "EXT4-fs: can't read descriptor %u\n", i);
2396 goto err_freebuddy;
2398 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2399 goto err_freebuddy;
2402 return 0;
2404 err_freebuddy:
2405 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2406 while (i-- > 0)
2407 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2408 i = num_meta_group_infos;
2409 while (i-- > 0)
2410 kfree(sbi->s_group_info[i]);
2411 iput(sbi->s_buddy_cache);
2412 err_freesgi:
2413 kfree(sbi->s_group_info);
2414 return -ENOMEM;
2417 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2419 struct ext4_sb_info *sbi = EXT4_SB(sb);
2420 unsigned i, j;
2421 unsigned offset;
2422 unsigned max;
2423 int ret;
2424 int cache_index;
2425 struct kmem_cache *cachep;
2426 char *namep = NULL;
2428 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2430 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2431 if (sbi->s_mb_offsets == NULL) {
2432 ret = -ENOMEM;
2433 goto out;
2436 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2437 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2438 if (sbi->s_mb_maxs == NULL) {
2439 ret = -ENOMEM;
2440 goto out;
2443 cache_index = sb->s_blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2444 cachep = ext4_groupinfo_caches[cache_index];
2445 if (!cachep) {
2446 char name[32];
2447 int len = offsetof(struct ext4_group_info,
2448 bb_counters[sb->s_blocksize_bits + 2]);
2450 sprintf(name, "ext4_groupinfo_%d", sb->s_blocksize_bits);
2451 namep = kstrdup(name, GFP_KERNEL);
2452 if (!namep) {
2453 ret = -ENOMEM;
2454 goto out;
2457 /* Need to free the kmem_cache_name() when we
2458 * destroy the slab */
2459 cachep = kmem_cache_create(namep, len, 0,
2460 SLAB_RECLAIM_ACCOUNT, NULL);
2461 if (!cachep) {
2462 ret = -ENOMEM;
2463 goto out;
2465 ext4_groupinfo_caches[cache_index] = cachep;
2468 /* order 0 is regular bitmap */
2469 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2470 sbi->s_mb_offsets[0] = 0;
2472 i = 1;
2473 offset = 0;
2474 max = sb->s_blocksize << 2;
2475 do {
2476 sbi->s_mb_offsets[i] = offset;
2477 sbi->s_mb_maxs[i] = max;
2478 offset += 1 << (sb->s_blocksize_bits - i);
2479 max = max >> 1;
2480 i++;
2481 } while (i <= sb->s_blocksize_bits + 1);
2483 /* init file for buddy data */
2484 ret = ext4_mb_init_backend(sb);
2485 if (ret != 0) {
2486 goto out;
2489 spin_lock_init(&sbi->s_md_lock);
2490 spin_lock_init(&sbi->s_bal_lock);
2492 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2493 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2494 sbi->s_mb_stats = MB_DEFAULT_STATS;
2495 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2496 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2497 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2499 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2500 if (sbi->s_locality_groups == NULL) {
2501 ret = -ENOMEM;
2502 goto out;
2504 for_each_possible_cpu(i) {
2505 struct ext4_locality_group *lg;
2506 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2507 mutex_init(&lg->lg_mutex);
2508 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2509 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2510 spin_lock_init(&lg->lg_prealloc_lock);
2513 if (sbi->s_proc)
2514 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2515 &ext4_mb_seq_groups_fops, sb);
2517 if (sbi->s_journal)
2518 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2519 out:
2520 if (ret) {
2521 kfree(sbi->s_mb_offsets);
2522 kfree(sbi->s_mb_maxs);
2523 kfree(namep);
2525 return ret;
2528 /* need to called with the ext4 group lock held */
2529 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2531 struct ext4_prealloc_space *pa;
2532 struct list_head *cur, *tmp;
2533 int count = 0;
2535 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2536 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2537 list_del(&pa->pa_group_list);
2538 count++;
2539 kmem_cache_free(ext4_pspace_cachep, pa);
2541 if (count)
2542 mb_debug(1, "mballoc: %u PAs left\n", count);
2546 int ext4_mb_release(struct super_block *sb)
2548 ext4_group_t ngroups = ext4_get_groups_count(sb);
2549 ext4_group_t i;
2550 int num_meta_group_infos;
2551 struct ext4_group_info *grinfo;
2552 struct ext4_sb_info *sbi = EXT4_SB(sb);
2553 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2555 if (sbi->s_group_info) {
2556 for (i = 0; i < ngroups; i++) {
2557 grinfo = ext4_get_group_info(sb, i);
2558 #ifdef DOUBLE_CHECK
2559 kfree(grinfo->bb_bitmap);
2560 #endif
2561 ext4_lock_group(sb, i);
2562 ext4_mb_cleanup_pa(grinfo);
2563 ext4_unlock_group(sb, i);
2564 kmem_cache_free(cachep, grinfo);
2566 num_meta_group_infos = (ngroups +
2567 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2568 EXT4_DESC_PER_BLOCK_BITS(sb);
2569 for (i = 0; i < num_meta_group_infos; i++)
2570 kfree(sbi->s_group_info[i]);
2571 kfree(sbi->s_group_info);
2573 kfree(sbi->s_mb_offsets);
2574 kfree(sbi->s_mb_maxs);
2575 if (sbi->s_buddy_cache)
2576 iput(sbi->s_buddy_cache);
2577 if (sbi->s_mb_stats) {
2578 printk(KERN_INFO
2579 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2580 atomic_read(&sbi->s_bal_allocated),
2581 atomic_read(&sbi->s_bal_reqs),
2582 atomic_read(&sbi->s_bal_success));
2583 printk(KERN_INFO
2584 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2585 "%u 2^N hits, %u breaks, %u lost\n",
2586 atomic_read(&sbi->s_bal_ex_scanned),
2587 atomic_read(&sbi->s_bal_goals),
2588 atomic_read(&sbi->s_bal_2orders),
2589 atomic_read(&sbi->s_bal_breaks),
2590 atomic_read(&sbi->s_mb_lost_chunks));
2591 printk(KERN_INFO
2592 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2593 sbi->s_mb_buddies_generated++,
2594 sbi->s_mb_generation_time);
2595 printk(KERN_INFO
2596 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2597 atomic_read(&sbi->s_mb_preallocated),
2598 atomic_read(&sbi->s_mb_discarded));
2601 free_percpu(sbi->s_locality_groups);
2602 if (sbi->s_proc)
2603 remove_proc_entry("mb_groups", sbi->s_proc);
2605 return 0;
2608 static inline int ext4_issue_discard(struct super_block *sb,
2609 ext4_group_t block_group, ext4_grpblk_t block, int count)
2611 int ret;
2612 ext4_fsblk_t discard_block;
2614 discard_block = block + ext4_group_first_block_no(sb, block_group);
2615 trace_ext4_discard_blocks(sb,
2616 (unsigned long long) discard_block, count);
2617 ret = sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2618 if (ret == -EOPNOTSUPP) {
2619 ext4_warning(sb, "discard not supported, disabling");
2620 clear_opt(EXT4_SB(sb)->s_mount_opt, DISCARD);
2622 return ret;
2626 * This function is called by the jbd2 layer once the commit has finished,
2627 * so we know we can free the blocks that were released with that commit.
2629 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2631 struct super_block *sb = journal->j_private;
2632 struct ext4_buddy e4b;
2633 struct ext4_group_info *db;
2634 int err, count = 0, count2 = 0;
2635 struct ext4_free_data *entry;
2636 struct list_head *l, *ltmp;
2638 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2639 entry = list_entry(l, struct ext4_free_data, list);
2641 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2642 entry->count, entry->group, entry);
2644 if (test_opt(sb, DISCARD))
2645 ext4_issue_discard(sb, entry->group,
2646 entry->start_blk, entry->count);
2648 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2649 /* we expect to find existing buddy because it's pinned */
2650 BUG_ON(err != 0);
2652 db = e4b.bd_info;
2653 /* there are blocks to put in buddy to make them really free */
2654 count += entry->count;
2655 count2++;
2656 ext4_lock_group(sb, entry->group);
2657 /* Take it out of per group rb tree */
2658 rb_erase(&entry->node, &(db->bb_free_root));
2659 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2661 if (!db->bb_free_root.rb_node) {
2662 /* No more items in the per group rb tree
2663 * balance refcounts from ext4_mb_free_metadata()
2665 page_cache_release(e4b.bd_buddy_page);
2666 page_cache_release(e4b.bd_bitmap_page);
2668 ext4_unlock_group(sb, entry->group);
2669 kmem_cache_free(ext4_free_ext_cachep, entry);
2670 ext4_mb_unload_buddy(&e4b);
2673 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2676 #ifdef CONFIG_EXT4_DEBUG
2677 u8 mb_enable_debug __read_mostly;
2679 static struct dentry *debugfs_dir;
2680 static struct dentry *debugfs_debug;
2682 static void __init ext4_create_debugfs_entry(void)
2684 debugfs_dir = debugfs_create_dir("ext4", NULL);
2685 if (debugfs_dir)
2686 debugfs_debug = debugfs_create_u8("mballoc-debug",
2687 S_IRUGO | S_IWUSR,
2688 debugfs_dir,
2689 &mb_enable_debug);
2692 static void ext4_remove_debugfs_entry(void)
2694 debugfs_remove(debugfs_debug);
2695 debugfs_remove(debugfs_dir);
2698 #else
2700 static void __init ext4_create_debugfs_entry(void)
2704 static void ext4_remove_debugfs_entry(void)
2708 #endif
2710 int __init ext4_init_mballoc(void)
2712 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2713 SLAB_RECLAIM_ACCOUNT);
2714 if (ext4_pspace_cachep == NULL)
2715 return -ENOMEM;
2717 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2718 SLAB_RECLAIM_ACCOUNT);
2719 if (ext4_ac_cachep == NULL) {
2720 kmem_cache_destroy(ext4_pspace_cachep);
2721 return -ENOMEM;
2724 ext4_free_ext_cachep = KMEM_CACHE(ext4_free_data,
2725 SLAB_RECLAIM_ACCOUNT);
2726 if (ext4_free_ext_cachep == NULL) {
2727 kmem_cache_destroy(ext4_pspace_cachep);
2728 kmem_cache_destroy(ext4_ac_cachep);
2729 return -ENOMEM;
2731 ext4_create_debugfs_entry();
2732 return 0;
2735 void ext4_exit_mballoc(void)
2737 int i;
2739 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2740 * before destroying the slab cache.
2742 rcu_barrier();
2743 kmem_cache_destroy(ext4_pspace_cachep);
2744 kmem_cache_destroy(ext4_ac_cachep);
2745 kmem_cache_destroy(ext4_free_ext_cachep);
2747 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2748 struct kmem_cache *cachep = ext4_groupinfo_caches[i];
2749 if (cachep) {
2750 char *name = (char *)kmem_cache_name(cachep);
2751 kmem_cache_destroy(cachep);
2752 kfree(name);
2755 ext4_remove_debugfs_entry();
2760 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2761 * Returns 0 if success or error code
2763 static noinline_for_stack int
2764 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2765 handle_t *handle, unsigned int reserv_blks)
2767 struct buffer_head *bitmap_bh = NULL;
2768 struct ext4_group_desc *gdp;
2769 struct buffer_head *gdp_bh;
2770 struct ext4_sb_info *sbi;
2771 struct super_block *sb;
2772 ext4_fsblk_t block;
2773 int err, len;
2775 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2776 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2778 sb = ac->ac_sb;
2779 sbi = EXT4_SB(sb);
2781 err = -EIO;
2782 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2783 if (!bitmap_bh)
2784 goto out_err;
2786 err = ext4_journal_get_write_access(handle, bitmap_bh);
2787 if (err)
2788 goto out_err;
2790 err = -EIO;
2791 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2792 if (!gdp)
2793 goto out_err;
2795 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2796 ext4_free_blks_count(sb, gdp));
2798 err = ext4_journal_get_write_access(handle, gdp_bh);
2799 if (err)
2800 goto out_err;
2802 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2804 len = ac->ac_b_ex.fe_len;
2805 if (!ext4_data_block_valid(sbi, block, len)) {
2806 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2807 "fs metadata\n", block, block+len);
2808 /* File system mounted not to panic on error
2809 * Fix the bitmap and repeat the block allocation
2810 * We leak some of the blocks here.
2812 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2813 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2814 ac->ac_b_ex.fe_len);
2815 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2816 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2817 if (!err)
2818 err = -EAGAIN;
2819 goto out_err;
2822 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2823 #ifdef AGGRESSIVE_CHECK
2825 int i;
2826 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2827 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2828 bitmap_bh->b_data));
2831 #endif
2832 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
2833 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2834 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2835 ext4_free_blks_set(sb, gdp,
2836 ext4_free_blocks_after_init(sb,
2837 ac->ac_b_ex.fe_group, gdp));
2839 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
2840 ext4_free_blks_set(sb, gdp, len);
2841 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2843 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2844 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
2846 * Now reduce the dirty block count also. Should not go negative
2848 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2849 /* release all the reserved blocks if non delalloc */
2850 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
2852 if (sbi->s_log_groups_per_flex) {
2853 ext4_group_t flex_group = ext4_flex_group(sbi,
2854 ac->ac_b_ex.fe_group);
2855 atomic_sub(ac->ac_b_ex.fe_len,
2856 &sbi->s_flex_groups[flex_group].free_blocks);
2859 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2860 if (err)
2861 goto out_err;
2862 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2864 out_err:
2865 ext4_mark_super_dirty(sb);
2866 brelse(bitmap_bh);
2867 return err;
2871 * here we normalize request for locality group
2872 * Group request are normalized to s_strip size if we set the same via mount
2873 * option. If not we set it to s_mb_group_prealloc which can be configured via
2874 * /sys/fs/ext4/<partition>/mb_group_prealloc
2876 * XXX: should we try to preallocate more than the group has now?
2878 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2880 struct super_block *sb = ac->ac_sb;
2881 struct ext4_locality_group *lg = ac->ac_lg;
2883 BUG_ON(lg == NULL);
2884 if (EXT4_SB(sb)->s_stripe)
2885 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
2886 else
2887 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2888 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2889 current->pid, ac->ac_g_ex.fe_len);
2893 * Normalization means making request better in terms of
2894 * size and alignment
2896 static noinline_for_stack void
2897 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2898 struct ext4_allocation_request *ar)
2900 int bsbits, max;
2901 ext4_lblk_t end;
2902 loff_t size, orig_size, start_off;
2903 ext4_lblk_t start;
2904 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2905 struct ext4_prealloc_space *pa;
2907 /* do normalize only data requests, metadata requests
2908 do not need preallocation */
2909 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2910 return;
2912 /* sometime caller may want exact blocks */
2913 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2914 return;
2916 /* caller may indicate that preallocation isn't
2917 * required (it's a tail, for example) */
2918 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2919 return;
2921 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2922 ext4_mb_normalize_group_request(ac);
2923 return ;
2926 bsbits = ac->ac_sb->s_blocksize_bits;
2928 /* first, let's learn actual file size
2929 * given current request is allocated */
2930 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2931 size = size << bsbits;
2932 if (size < i_size_read(ac->ac_inode))
2933 size = i_size_read(ac->ac_inode);
2934 orig_size = size;
2936 /* max size of free chunks */
2937 max = 2 << bsbits;
2939 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2940 (req <= (size) || max <= (chunk_size))
2942 /* first, try to predict filesize */
2943 /* XXX: should this table be tunable? */
2944 start_off = 0;
2945 if (size <= 16 * 1024) {
2946 size = 16 * 1024;
2947 } else if (size <= 32 * 1024) {
2948 size = 32 * 1024;
2949 } else if (size <= 64 * 1024) {
2950 size = 64 * 1024;
2951 } else if (size <= 128 * 1024) {
2952 size = 128 * 1024;
2953 } else if (size <= 256 * 1024) {
2954 size = 256 * 1024;
2955 } else if (size <= 512 * 1024) {
2956 size = 512 * 1024;
2957 } else if (size <= 1024 * 1024) {
2958 size = 1024 * 1024;
2959 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2960 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2961 (21 - bsbits)) << 21;
2962 size = 2 * 1024 * 1024;
2963 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2964 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2965 (22 - bsbits)) << 22;
2966 size = 4 * 1024 * 1024;
2967 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2968 (8<<20)>>bsbits, max, 8 * 1024)) {
2969 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2970 (23 - bsbits)) << 23;
2971 size = 8 * 1024 * 1024;
2972 } else {
2973 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2974 size = ac->ac_o_ex.fe_len << bsbits;
2976 size = size >> bsbits;
2977 start = start_off >> bsbits;
2979 /* don't cover already allocated blocks in selected range */
2980 if (ar->pleft && start <= ar->lleft) {
2981 size -= ar->lleft + 1 - start;
2982 start = ar->lleft + 1;
2984 if (ar->pright && start + size - 1 >= ar->lright)
2985 size -= start + size - ar->lright;
2987 end = start + size;
2989 /* check we don't cross already preallocated blocks */
2990 rcu_read_lock();
2991 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2992 ext4_lblk_t pa_end;
2994 if (pa->pa_deleted)
2995 continue;
2996 spin_lock(&pa->pa_lock);
2997 if (pa->pa_deleted) {
2998 spin_unlock(&pa->pa_lock);
2999 continue;
3002 pa_end = pa->pa_lstart + pa->pa_len;
3004 /* PA must not overlap original request */
3005 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3006 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3008 /* skip PAs this normalized request doesn't overlap with */
3009 if (pa->pa_lstart >= end || pa_end <= start) {
3010 spin_unlock(&pa->pa_lock);
3011 continue;
3013 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3015 /* adjust start or end to be adjacent to this pa */
3016 if (pa_end <= ac->ac_o_ex.fe_logical) {
3017 BUG_ON(pa_end < start);
3018 start = pa_end;
3019 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3020 BUG_ON(pa->pa_lstart > end);
3021 end = pa->pa_lstart;
3023 spin_unlock(&pa->pa_lock);
3025 rcu_read_unlock();
3026 size = end - start;
3028 /* XXX: extra loop to check we really don't overlap preallocations */
3029 rcu_read_lock();
3030 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3031 ext4_lblk_t pa_end;
3032 spin_lock(&pa->pa_lock);
3033 if (pa->pa_deleted == 0) {
3034 pa_end = pa->pa_lstart + pa->pa_len;
3035 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3037 spin_unlock(&pa->pa_lock);
3039 rcu_read_unlock();
3041 if (start + size <= ac->ac_o_ex.fe_logical &&
3042 start > ac->ac_o_ex.fe_logical) {
3043 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3044 (unsigned long) start, (unsigned long) size,
3045 (unsigned long) ac->ac_o_ex.fe_logical);
3047 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3048 start > ac->ac_o_ex.fe_logical);
3049 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3051 /* now prepare goal request */
3053 /* XXX: is it better to align blocks WRT to logical
3054 * placement or satisfy big request as is */
3055 ac->ac_g_ex.fe_logical = start;
3056 ac->ac_g_ex.fe_len = size;
3058 /* define goal start in order to merge */
3059 if (ar->pright && (ar->lright == (start + size))) {
3060 /* merge to the right */
3061 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3062 &ac->ac_f_ex.fe_group,
3063 &ac->ac_f_ex.fe_start);
3064 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3066 if (ar->pleft && (ar->lleft + 1 == start)) {
3067 /* merge to the left */
3068 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3069 &ac->ac_f_ex.fe_group,
3070 &ac->ac_f_ex.fe_start);
3071 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3074 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3075 (unsigned) orig_size, (unsigned) start);
3078 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3080 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3082 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3083 atomic_inc(&sbi->s_bal_reqs);
3084 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3085 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3086 atomic_inc(&sbi->s_bal_success);
3087 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3088 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3089 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3090 atomic_inc(&sbi->s_bal_goals);
3091 if (ac->ac_found > sbi->s_mb_max_to_scan)
3092 atomic_inc(&sbi->s_bal_breaks);
3095 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3096 trace_ext4_mballoc_alloc(ac);
3097 else
3098 trace_ext4_mballoc_prealloc(ac);
3102 * Called on failure; free up any blocks from the inode PA for this
3103 * context. We don't need this for MB_GROUP_PA because we only change
3104 * pa_free in ext4_mb_release_context(), but on failure, we've already
3105 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3107 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3109 struct ext4_prealloc_space *pa = ac->ac_pa;
3110 int len;
3112 if (pa && pa->pa_type == MB_INODE_PA) {
3113 len = ac->ac_b_ex.fe_len;
3114 pa->pa_free += len;
3120 * use blocks preallocated to inode
3122 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3123 struct ext4_prealloc_space *pa)
3125 ext4_fsblk_t start;
3126 ext4_fsblk_t end;
3127 int len;
3129 /* found preallocated blocks, use them */
3130 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3131 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3132 len = end - start;
3133 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3134 &ac->ac_b_ex.fe_start);
3135 ac->ac_b_ex.fe_len = len;
3136 ac->ac_status = AC_STATUS_FOUND;
3137 ac->ac_pa = pa;
3139 BUG_ON(start < pa->pa_pstart);
3140 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3141 BUG_ON(pa->pa_free < len);
3142 pa->pa_free -= len;
3144 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3148 * use blocks preallocated to locality group
3150 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3151 struct ext4_prealloc_space *pa)
3153 unsigned int len = ac->ac_o_ex.fe_len;
3155 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3156 &ac->ac_b_ex.fe_group,
3157 &ac->ac_b_ex.fe_start);
3158 ac->ac_b_ex.fe_len = len;
3159 ac->ac_status = AC_STATUS_FOUND;
3160 ac->ac_pa = pa;
3162 /* we don't correct pa_pstart or pa_plen here to avoid
3163 * possible race when the group is being loaded concurrently
3164 * instead we correct pa later, after blocks are marked
3165 * in on-disk bitmap -- see ext4_mb_release_context()
3166 * Other CPUs are prevented from allocating from this pa by lg_mutex
3168 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3172 * Return the prealloc space that have minimal distance
3173 * from the goal block. @cpa is the prealloc
3174 * space that is having currently known minimal distance
3175 * from the goal block.
3177 static struct ext4_prealloc_space *
3178 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3179 struct ext4_prealloc_space *pa,
3180 struct ext4_prealloc_space *cpa)
3182 ext4_fsblk_t cur_distance, new_distance;
3184 if (cpa == NULL) {
3185 atomic_inc(&pa->pa_count);
3186 return pa;
3188 cur_distance = abs(goal_block - cpa->pa_pstart);
3189 new_distance = abs(goal_block - pa->pa_pstart);
3191 if (cur_distance < new_distance)
3192 return cpa;
3194 /* drop the previous reference */
3195 atomic_dec(&cpa->pa_count);
3196 atomic_inc(&pa->pa_count);
3197 return pa;
3201 * search goal blocks in preallocated space
3203 static noinline_for_stack int
3204 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3206 int order, i;
3207 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3208 struct ext4_locality_group *lg;
3209 struct ext4_prealloc_space *pa, *cpa = NULL;
3210 ext4_fsblk_t goal_block;
3212 /* only data can be preallocated */
3213 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3214 return 0;
3216 /* first, try per-file preallocation */
3217 rcu_read_lock();
3218 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3220 /* all fields in this condition don't change,
3221 * so we can skip locking for them */
3222 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3223 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3224 continue;
3226 /* non-extent files can't have physical blocks past 2^32 */
3227 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3228 pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3229 continue;
3231 /* found preallocated blocks, use them */
3232 spin_lock(&pa->pa_lock);
3233 if (pa->pa_deleted == 0 && pa->pa_free) {
3234 atomic_inc(&pa->pa_count);
3235 ext4_mb_use_inode_pa(ac, pa);
3236 spin_unlock(&pa->pa_lock);
3237 ac->ac_criteria = 10;
3238 rcu_read_unlock();
3239 return 1;
3241 spin_unlock(&pa->pa_lock);
3243 rcu_read_unlock();
3245 /* can we use group allocation? */
3246 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3247 return 0;
3249 /* inode may have no locality group for some reason */
3250 lg = ac->ac_lg;
3251 if (lg == NULL)
3252 return 0;
3253 order = fls(ac->ac_o_ex.fe_len) - 1;
3254 if (order > PREALLOC_TB_SIZE - 1)
3255 /* The max size of hash table is PREALLOC_TB_SIZE */
3256 order = PREALLOC_TB_SIZE - 1;
3258 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3260 * search for the prealloc space that is having
3261 * minimal distance from the goal block.
3263 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3264 rcu_read_lock();
3265 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3266 pa_inode_list) {
3267 spin_lock(&pa->pa_lock);
3268 if (pa->pa_deleted == 0 &&
3269 pa->pa_free >= ac->ac_o_ex.fe_len) {
3271 cpa = ext4_mb_check_group_pa(goal_block,
3272 pa, cpa);
3274 spin_unlock(&pa->pa_lock);
3276 rcu_read_unlock();
3278 if (cpa) {
3279 ext4_mb_use_group_pa(ac, cpa);
3280 ac->ac_criteria = 20;
3281 return 1;
3283 return 0;
3287 * the function goes through all block freed in the group
3288 * but not yet committed and marks them used in in-core bitmap.
3289 * buddy must be generated from this bitmap
3290 * Need to be called with the ext4 group lock held
3292 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3293 ext4_group_t group)
3295 struct rb_node *n;
3296 struct ext4_group_info *grp;
3297 struct ext4_free_data *entry;
3299 grp = ext4_get_group_info(sb, group);
3300 n = rb_first(&(grp->bb_free_root));
3302 while (n) {
3303 entry = rb_entry(n, struct ext4_free_data, node);
3304 mb_set_bits(bitmap, entry->start_blk, entry->count);
3305 n = rb_next(n);
3307 return;
3311 * the function goes through all preallocation in this group and marks them
3312 * used in in-core bitmap. buddy must be generated from this bitmap
3313 * Need to be called with ext4 group lock held
3315 static noinline_for_stack
3316 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3317 ext4_group_t group)
3319 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3320 struct ext4_prealloc_space *pa;
3321 struct list_head *cur;
3322 ext4_group_t groupnr;
3323 ext4_grpblk_t start;
3324 int preallocated = 0;
3325 int count = 0;
3326 int len;
3328 /* all form of preallocation discards first load group,
3329 * so the only competing code is preallocation use.
3330 * we don't need any locking here
3331 * notice we do NOT ignore preallocations with pa_deleted
3332 * otherwise we could leave used blocks available for
3333 * allocation in buddy when concurrent ext4_mb_put_pa()
3334 * is dropping preallocation
3336 list_for_each(cur, &grp->bb_prealloc_list) {
3337 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3338 spin_lock(&pa->pa_lock);
3339 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3340 &groupnr, &start);
3341 len = pa->pa_len;
3342 spin_unlock(&pa->pa_lock);
3343 if (unlikely(len == 0))
3344 continue;
3345 BUG_ON(groupnr != group);
3346 mb_set_bits(bitmap, start, len);
3347 preallocated += len;
3348 count++;
3350 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3353 static void ext4_mb_pa_callback(struct rcu_head *head)
3355 struct ext4_prealloc_space *pa;
3356 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3357 kmem_cache_free(ext4_pspace_cachep, pa);
3361 * drops a reference to preallocated space descriptor
3362 * if this was the last reference and the space is consumed
3364 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3365 struct super_block *sb, struct ext4_prealloc_space *pa)
3367 ext4_group_t grp;
3368 ext4_fsblk_t grp_blk;
3370 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3371 return;
3373 /* in this short window concurrent discard can set pa_deleted */
3374 spin_lock(&pa->pa_lock);
3375 if (pa->pa_deleted == 1) {
3376 spin_unlock(&pa->pa_lock);
3377 return;
3380 pa->pa_deleted = 1;
3381 spin_unlock(&pa->pa_lock);
3383 grp_blk = pa->pa_pstart;
3385 * If doing group-based preallocation, pa_pstart may be in the
3386 * next group when pa is used up
3388 if (pa->pa_type == MB_GROUP_PA)
3389 grp_blk--;
3391 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3394 * possible race:
3396 * P1 (buddy init) P2 (regular allocation)
3397 * find block B in PA
3398 * copy on-disk bitmap to buddy
3399 * mark B in on-disk bitmap
3400 * drop PA from group
3401 * mark all PAs in buddy
3403 * thus, P1 initializes buddy with B available. to prevent this
3404 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3405 * against that pair
3407 ext4_lock_group(sb, grp);
3408 list_del(&pa->pa_group_list);
3409 ext4_unlock_group(sb, grp);
3411 spin_lock(pa->pa_obj_lock);
3412 list_del_rcu(&pa->pa_inode_list);
3413 spin_unlock(pa->pa_obj_lock);
3415 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3419 * creates new preallocated space for given inode
3421 static noinline_for_stack int
3422 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3424 struct super_block *sb = ac->ac_sb;
3425 struct ext4_prealloc_space *pa;
3426 struct ext4_group_info *grp;
3427 struct ext4_inode_info *ei;
3429 /* preallocate only when found space is larger then requested */
3430 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3431 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3432 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3434 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3435 if (pa == NULL)
3436 return -ENOMEM;
3438 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3439 int winl;
3440 int wins;
3441 int win;
3442 int offs;
3444 /* we can't allocate as much as normalizer wants.
3445 * so, found space must get proper lstart
3446 * to cover original request */
3447 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3448 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3450 /* we're limited by original request in that
3451 * logical block must be covered any way
3452 * winl is window we can move our chunk within */
3453 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3455 /* also, we should cover whole original request */
3456 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3458 /* the smallest one defines real window */
3459 win = min(winl, wins);
3461 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3462 if (offs && offs < win)
3463 win = offs;
3465 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3466 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3467 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3470 /* preallocation can change ac_b_ex, thus we store actually
3471 * allocated blocks for history */
3472 ac->ac_f_ex = ac->ac_b_ex;
3474 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3475 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3476 pa->pa_len = ac->ac_b_ex.fe_len;
3477 pa->pa_free = pa->pa_len;
3478 atomic_set(&pa->pa_count, 1);
3479 spin_lock_init(&pa->pa_lock);
3480 INIT_LIST_HEAD(&pa->pa_inode_list);
3481 INIT_LIST_HEAD(&pa->pa_group_list);
3482 pa->pa_deleted = 0;
3483 pa->pa_type = MB_INODE_PA;
3485 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3486 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3487 trace_ext4_mb_new_inode_pa(ac, pa);
3489 ext4_mb_use_inode_pa(ac, pa);
3490 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3492 ei = EXT4_I(ac->ac_inode);
3493 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3495 pa->pa_obj_lock = &ei->i_prealloc_lock;
3496 pa->pa_inode = ac->ac_inode;
3498 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3499 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3500 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3502 spin_lock(pa->pa_obj_lock);
3503 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3504 spin_unlock(pa->pa_obj_lock);
3506 return 0;
3510 * creates new preallocated space for locality group inodes belongs to
3512 static noinline_for_stack int
3513 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3515 struct super_block *sb = ac->ac_sb;
3516 struct ext4_locality_group *lg;
3517 struct ext4_prealloc_space *pa;
3518 struct ext4_group_info *grp;
3520 /* preallocate only when found space is larger then requested */
3521 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3522 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3523 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3525 BUG_ON(ext4_pspace_cachep == NULL);
3526 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3527 if (pa == NULL)
3528 return -ENOMEM;
3530 /* preallocation can change ac_b_ex, thus we store actually
3531 * allocated blocks for history */
3532 ac->ac_f_ex = ac->ac_b_ex;
3534 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3535 pa->pa_lstart = pa->pa_pstart;
3536 pa->pa_len = ac->ac_b_ex.fe_len;
3537 pa->pa_free = pa->pa_len;
3538 atomic_set(&pa->pa_count, 1);
3539 spin_lock_init(&pa->pa_lock);
3540 INIT_LIST_HEAD(&pa->pa_inode_list);
3541 INIT_LIST_HEAD(&pa->pa_group_list);
3542 pa->pa_deleted = 0;
3543 pa->pa_type = MB_GROUP_PA;
3545 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3546 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3547 trace_ext4_mb_new_group_pa(ac, pa);
3549 ext4_mb_use_group_pa(ac, pa);
3550 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3552 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3553 lg = ac->ac_lg;
3554 BUG_ON(lg == NULL);
3556 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3557 pa->pa_inode = NULL;
3559 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3560 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3561 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3564 * We will later add the new pa to the right bucket
3565 * after updating the pa_free in ext4_mb_release_context
3567 return 0;
3570 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3572 int err;
3574 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3575 err = ext4_mb_new_group_pa(ac);
3576 else
3577 err = ext4_mb_new_inode_pa(ac);
3578 return err;
3582 * finds all unused blocks in on-disk bitmap, frees them in
3583 * in-core bitmap and buddy.
3584 * @pa must be unlinked from inode and group lists, so that
3585 * nobody else can find/use it.
3586 * the caller MUST hold group/inode locks.
3587 * TODO: optimize the case when there are no in-core structures yet
3589 static noinline_for_stack int
3590 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3591 struct ext4_prealloc_space *pa)
3593 struct super_block *sb = e4b->bd_sb;
3594 struct ext4_sb_info *sbi = EXT4_SB(sb);
3595 unsigned int end;
3596 unsigned int next;
3597 ext4_group_t group;
3598 ext4_grpblk_t bit;
3599 unsigned long long grp_blk_start;
3600 int err = 0;
3601 int free = 0;
3603 BUG_ON(pa->pa_deleted == 0);
3604 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3605 grp_blk_start = pa->pa_pstart - bit;
3606 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3607 end = bit + pa->pa_len;
3609 while (bit < end) {
3610 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3611 if (bit >= end)
3612 break;
3613 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3614 mb_debug(1, " free preallocated %u/%u in group %u\n",
3615 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3616 (unsigned) next - bit, (unsigned) group);
3617 free += next - bit;
3619 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3620 trace_ext4_mb_release_inode_pa(sb, pa->pa_inode, pa,
3621 grp_blk_start + bit, next - bit);
3622 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3623 bit = next + 1;
3625 if (free != pa->pa_free) {
3626 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3627 pa, (unsigned long) pa->pa_lstart,
3628 (unsigned long) pa->pa_pstart,
3629 (unsigned long) pa->pa_len);
3630 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3631 free, pa->pa_free);
3633 * pa is already deleted so we use the value obtained
3634 * from the bitmap and continue.
3637 atomic_add(free, &sbi->s_mb_discarded);
3639 return err;
3642 static noinline_for_stack int
3643 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3644 struct ext4_prealloc_space *pa)
3646 struct super_block *sb = e4b->bd_sb;
3647 ext4_group_t group;
3648 ext4_grpblk_t bit;
3650 trace_ext4_mb_release_group_pa(sb, pa);
3651 BUG_ON(pa->pa_deleted == 0);
3652 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3653 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3654 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3655 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3656 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3658 return 0;
3662 * releases all preallocations in given group
3664 * first, we need to decide discard policy:
3665 * - when do we discard
3666 * 1) ENOSPC
3667 * - how many do we discard
3668 * 1) how many requested
3670 static noinline_for_stack int
3671 ext4_mb_discard_group_preallocations(struct super_block *sb,
3672 ext4_group_t group, int needed)
3674 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3675 struct buffer_head *bitmap_bh = NULL;
3676 struct ext4_prealloc_space *pa, *tmp;
3677 struct list_head list;
3678 struct ext4_buddy e4b;
3679 int err;
3680 int busy = 0;
3681 int free = 0;
3683 mb_debug(1, "discard preallocation for group %u\n", group);
3685 if (list_empty(&grp->bb_prealloc_list))
3686 return 0;
3688 bitmap_bh = ext4_read_block_bitmap(sb, group);
3689 if (bitmap_bh == NULL) {
3690 ext4_error(sb, "Error reading block bitmap for %u", group);
3691 return 0;
3694 err = ext4_mb_load_buddy(sb, group, &e4b);
3695 if (err) {
3696 ext4_error(sb, "Error loading buddy information for %u", group);
3697 put_bh(bitmap_bh);
3698 return 0;
3701 if (needed == 0)
3702 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3704 INIT_LIST_HEAD(&list);
3705 repeat:
3706 ext4_lock_group(sb, group);
3707 list_for_each_entry_safe(pa, tmp,
3708 &grp->bb_prealloc_list, pa_group_list) {
3709 spin_lock(&pa->pa_lock);
3710 if (atomic_read(&pa->pa_count)) {
3711 spin_unlock(&pa->pa_lock);
3712 busy = 1;
3713 continue;
3715 if (pa->pa_deleted) {
3716 spin_unlock(&pa->pa_lock);
3717 continue;
3720 /* seems this one can be freed ... */
3721 pa->pa_deleted = 1;
3723 /* we can trust pa_free ... */
3724 free += pa->pa_free;
3726 spin_unlock(&pa->pa_lock);
3728 list_del(&pa->pa_group_list);
3729 list_add(&pa->u.pa_tmp_list, &list);
3732 /* if we still need more blocks and some PAs were used, try again */
3733 if (free < needed && busy) {
3734 busy = 0;
3735 ext4_unlock_group(sb, group);
3737 * Yield the CPU here so that we don't get soft lockup
3738 * in non preempt case.
3740 yield();
3741 goto repeat;
3744 /* found anything to free? */
3745 if (list_empty(&list)) {
3746 BUG_ON(free != 0);
3747 goto out;
3750 /* now free all selected PAs */
3751 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3753 /* remove from object (inode or locality group) */
3754 spin_lock(pa->pa_obj_lock);
3755 list_del_rcu(&pa->pa_inode_list);
3756 spin_unlock(pa->pa_obj_lock);
3758 if (pa->pa_type == MB_GROUP_PA)
3759 ext4_mb_release_group_pa(&e4b, pa);
3760 else
3761 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3763 list_del(&pa->u.pa_tmp_list);
3764 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3767 out:
3768 ext4_unlock_group(sb, group);
3769 ext4_mb_unload_buddy(&e4b);
3770 put_bh(bitmap_bh);
3771 return free;
3775 * releases all non-used preallocated blocks for given inode
3777 * It's important to discard preallocations under i_data_sem
3778 * We don't want another block to be served from the prealloc
3779 * space when we are discarding the inode prealloc space.
3781 * FIXME!! Make sure it is valid at all the call sites
3783 void ext4_discard_preallocations(struct inode *inode)
3785 struct ext4_inode_info *ei = EXT4_I(inode);
3786 struct super_block *sb = inode->i_sb;
3787 struct buffer_head *bitmap_bh = NULL;
3788 struct ext4_prealloc_space *pa, *tmp;
3789 ext4_group_t group = 0;
3790 struct list_head list;
3791 struct ext4_buddy e4b;
3792 int err;
3794 if (!S_ISREG(inode->i_mode)) {
3795 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3796 return;
3799 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3800 trace_ext4_discard_preallocations(inode);
3802 INIT_LIST_HEAD(&list);
3804 repeat:
3805 /* first, collect all pa's in the inode */
3806 spin_lock(&ei->i_prealloc_lock);
3807 while (!list_empty(&ei->i_prealloc_list)) {
3808 pa = list_entry(ei->i_prealloc_list.next,
3809 struct ext4_prealloc_space, pa_inode_list);
3810 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3811 spin_lock(&pa->pa_lock);
3812 if (atomic_read(&pa->pa_count)) {
3813 /* this shouldn't happen often - nobody should
3814 * use preallocation while we're discarding it */
3815 spin_unlock(&pa->pa_lock);
3816 spin_unlock(&ei->i_prealloc_lock);
3817 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3818 WARN_ON(1);
3819 schedule_timeout_uninterruptible(HZ);
3820 goto repeat;
3823 if (pa->pa_deleted == 0) {
3824 pa->pa_deleted = 1;
3825 spin_unlock(&pa->pa_lock);
3826 list_del_rcu(&pa->pa_inode_list);
3827 list_add(&pa->u.pa_tmp_list, &list);
3828 continue;
3831 /* someone is deleting pa right now */
3832 spin_unlock(&pa->pa_lock);
3833 spin_unlock(&ei->i_prealloc_lock);
3835 /* we have to wait here because pa_deleted
3836 * doesn't mean pa is already unlinked from
3837 * the list. as we might be called from
3838 * ->clear_inode() the inode will get freed
3839 * and concurrent thread which is unlinking
3840 * pa from inode's list may access already
3841 * freed memory, bad-bad-bad */
3843 /* XXX: if this happens too often, we can
3844 * add a flag to force wait only in case
3845 * of ->clear_inode(), but not in case of
3846 * regular truncate */
3847 schedule_timeout_uninterruptible(HZ);
3848 goto repeat;
3850 spin_unlock(&ei->i_prealloc_lock);
3852 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3853 BUG_ON(pa->pa_type != MB_INODE_PA);
3854 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3856 err = ext4_mb_load_buddy(sb, group, &e4b);
3857 if (err) {
3858 ext4_error(sb, "Error loading buddy information for %u",
3859 group);
3860 continue;
3863 bitmap_bh = ext4_read_block_bitmap(sb, group);
3864 if (bitmap_bh == NULL) {
3865 ext4_error(sb, "Error reading block bitmap for %u",
3866 group);
3867 ext4_mb_unload_buddy(&e4b);
3868 continue;
3871 ext4_lock_group(sb, group);
3872 list_del(&pa->pa_group_list);
3873 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3874 ext4_unlock_group(sb, group);
3876 ext4_mb_unload_buddy(&e4b);
3877 put_bh(bitmap_bh);
3879 list_del(&pa->u.pa_tmp_list);
3880 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3885 * finds all preallocated spaces and return blocks being freed to them
3886 * if preallocated space becomes full (no block is used from the space)
3887 * then the function frees space in buddy
3888 * XXX: at the moment, truncate (which is the only way to free blocks)
3889 * discards all preallocations
3891 static void ext4_mb_return_to_preallocation(struct inode *inode,
3892 struct ext4_buddy *e4b,
3893 sector_t block, int count)
3895 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
3897 #ifdef CONFIG_EXT4_DEBUG
3898 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3900 struct super_block *sb = ac->ac_sb;
3901 ext4_group_t ngroups, i;
3903 if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
3904 return;
3906 printk(KERN_ERR "EXT4-fs: Can't allocate:"
3907 " Allocation context details:\n");
3908 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
3909 ac->ac_status, ac->ac_flags);
3910 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3911 "best %lu/%lu/%lu@%lu cr %d\n",
3912 (unsigned long)ac->ac_o_ex.fe_group,
3913 (unsigned long)ac->ac_o_ex.fe_start,
3914 (unsigned long)ac->ac_o_ex.fe_len,
3915 (unsigned long)ac->ac_o_ex.fe_logical,
3916 (unsigned long)ac->ac_g_ex.fe_group,
3917 (unsigned long)ac->ac_g_ex.fe_start,
3918 (unsigned long)ac->ac_g_ex.fe_len,
3919 (unsigned long)ac->ac_g_ex.fe_logical,
3920 (unsigned long)ac->ac_b_ex.fe_group,
3921 (unsigned long)ac->ac_b_ex.fe_start,
3922 (unsigned long)ac->ac_b_ex.fe_len,
3923 (unsigned long)ac->ac_b_ex.fe_logical,
3924 (int)ac->ac_criteria);
3925 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
3926 ac->ac_found);
3927 printk(KERN_ERR "EXT4-fs: groups: \n");
3928 ngroups = ext4_get_groups_count(sb);
3929 for (i = 0; i < ngroups; i++) {
3930 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3931 struct ext4_prealloc_space *pa;
3932 ext4_grpblk_t start;
3933 struct list_head *cur;
3934 ext4_lock_group(sb, i);
3935 list_for_each(cur, &grp->bb_prealloc_list) {
3936 pa = list_entry(cur, struct ext4_prealloc_space,
3937 pa_group_list);
3938 spin_lock(&pa->pa_lock);
3939 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3940 NULL, &start);
3941 spin_unlock(&pa->pa_lock);
3942 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3943 start, pa->pa_len);
3945 ext4_unlock_group(sb, i);
3947 if (grp->bb_free == 0)
3948 continue;
3949 printk(KERN_ERR "%u: %d/%d \n",
3950 i, grp->bb_free, grp->bb_fragments);
3952 printk(KERN_ERR "\n");
3954 #else
3955 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3957 return;
3959 #endif
3962 * We use locality group preallocation for small size file. The size of the
3963 * file is determined by the current size or the resulting size after
3964 * allocation which ever is larger
3966 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3968 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3970 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3971 int bsbits = ac->ac_sb->s_blocksize_bits;
3972 loff_t size, isize;
3974 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3975 return;
3977 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3978 return;
3980 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3981 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3982 >> bsbits;
3984 if ((size == isize) &&
3985 !ext4_fs_is_busy(sbi) &&
3986 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3987 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3988 return;
3991 /* don't use group allocation for large files */
3992 size = max(size, isize);
3993 if (size > sbi->s_mb_stream_request) {
3994 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3995 return;
3998 BUG_ON(ac->ac_lg != NULL);
4000 * locality group prealloc space are per cpu. The reason for having
4001 * per cpu locality group is to reduce the contention between block
4002 * request from multiple CPUs.
4004 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
4006 /* we're going to use group allocation */
4007 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4009 /* serialize all allocations in the group */
4010 mutex_lock(&ac->ac_lg->lg_mutex);
4013 static noinline_for_stack int
4014 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4015 struct ext4_allocation_request *ar)
4017 struct super_block *sb = ar->inode->i_sb;
4018 struct ext4_sb_info *sbi = EXT4_SB(sb);
4019 struct ext4_super_block *es = sbi->s_es;
4020 ext4_group_t group;
4021 unsigned int len;
4022 ext4_fsblk_t goal;
4023 ext4_grpblk_t block;
4025 /* we can't allocate > group size */
4026 len = ar->len;
4028 /* just a dirty hack to filter too big requests */
4029 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4030 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4032 /* start searching from the goal */
4033 goal = ar->goal;
4034 if (goal < le32_to_cpu(es->s_first_data_block) ||
4035 goal >= ext4_blocks_count(es))
4036 goal = le32_to_cpu(es->s_first_data_block);
4037 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4039 /* set up allocation goals */
4040 memset(ac, 0, sizeof(struct ext4_allocation_context));
4041 ac->ac_b_ex.fe_logical = ar->logical;
4042 ac->ac_status = AC_STATUS_CONTINUE;
4043 ac->ac_sb = sb;
4044 ac->ac_inode = ar->inode;
4045 ac->ac_o_ex.fe_logical = ar->logical;
4046 ac->ac_o_ex.fe_group = group;
4047 ac->ac_o_ex.fe_start = block;
4048 ac->ac_o_ex.fe_len = len;
4049 ac->ac_g_ex.fe_logical = ar->logical;
4050 ac->ac_g_ex.fe_group = group;
4051 ac->ac_g_ex.fe_start = block;
4052 ac->ac_g_ex.fe_len = len;
4053 ac->ac_flags = ar->flags;
4055 /* we have to define context: we'll we work with a file or
4056 * locality group. this is a policy, actually */
4057 ext4_mb_group_or_file(ac);
4059 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4060 "left: %u/%u, right %u/%u to %swritable\n",
4061 (unsigned) ar->len, (unsigned) ar->logical,
4062 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4063 (unsigned) ar->lleft, (unsigned) ar->pleft,
4064 (unsigned) ar->lright, (unsigned) ar->pright,
4065 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4066 return 0;
4070 static noinline_for_stack void
4071 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4072 struct ext4_locality_group *lg,
4073 int order, int total_entries)
4075 ext4_group_t group = 0;
4076 struct ext4_buddy e4b;
4077 struct list_head discard_list;
4078 struct ext4_prealloc_space *pa, *tmp;
4080 mb_debug(1, "discard locality group preallocation\n");
4082 INIT_LIST_HEAD(&discard_list);
4084 spin_lock(&lg->lg_prealloc_lock);
4085 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4086 pa_inode_list) {
4087 spin_lock(&pa->pa_lock);
4088 if (atomic_read(&pa->pa_count)) {
4090 * This is the pa that we just used
4091 * for block allocation. So don't
4092 * free that
4094 spin_unlock(&pa->pa_lock);
4095 continue;
4097 if (pa->pa_deleted) {
4098 spin_unlock(&pa->pa_lock);
4099 continue;
4101 /* only lg prealloc space */
4102 BUG_ON(pa->pa_type != MB_GROUP_PA);
4104 /* seems this one can be freed ... */
4105 pa->pa_deleted = 1;
4106 spin_unlock(&pa->pa_lock);
4108 list_del_rcu(&pa->pa_inode_list);
4109 list_add(&pa->u.pa_tmp_list, &discard_list);
4111 total_entries--;
4112 if (total_entries <= 5) {
4114 * we want to keep only 5 entries
4115 * allowing it to grow to 8. This
4116 * mak sure we don't call discard
4117 * soon for this list.
4119 break;
4122 spin_unlock(&lg->lg_prealloc_lock);
4124 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4126 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4127 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4128 ext4_error(sb, "Error loading buddy information for %u",
4129 group);
4130 continue;
4132 ext4_lock_group(sb, group);
4133 list_del(&pa->pa_group_list);
4134 ext4_mb_release_group_pa(&e4b, pa);
4135 ext4_unlock_group(sb, group);
4137 ext4_mb_unload_buddy(&e4b);
4138 list_del(&pa->u.pa_tmp_list);
4139 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4144 * We have incremented pa_count. So it cannot be freed at this
4145 * point. Also we hold lg_mutex. So no parallel allocation is
4146 * possible from this lg. That means pa_free cannot be updated.
4148 * A parallel ext4_mb_discard_group_preallocations is possible.
4149 * which can cause the lg_prealloc_list to be updated.
4152 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4154 int order, added = 0, lg_prealloc_count = 1;
4155 struct super_block *sb = ac->ac_sb;
4156 struct ext4_locality_group *lg = ac->ac_lg;
4157 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4159 order = fls(pa->pa_free) - 1;
4160 if (order > PREALLOC_TB_SIZE - 1)
4161 /* The max size of hash table is PREALLOC_TB_SIZE */
4162 order = PREALLOC_TB_SIZE - 1;
4163 /* Add the prealloc space to lg */
4164 rcu_read_lock();
4165 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4166 pa_inode_list) {
4167 spin_lock(&tmp_pa->pa_lock);
4168 if (tmp_pa->pa_deleted) {
4169 spin_unlock(&tmp_pa->pa_lock);
4170 continue;
4172 if (!added && pa->pa_free < tmp_pa->pa_free) {
4173 /* Add to the tail of the previous entry */
4174 list_add_tail_rcu(&pa->pa_inode_list,
4175 &tmp_pa->pa_inode_list);
4176 added = 1;
4178 * we want to count the total
4179 * number of entries in the list
4182 spin_unlock(&tmp_pa->pa_lock);
4183 lg_prealloc_count++;
4185 if (!added)
4186 list_add_tail_rcu(&pa->pa_inode_list,
4187 &lg->lg_prealloc_list[order]);
4188 rcu_read_unlock();
4190 /* Now trim the list to be not more than 8 elements */
4191 if (lg_prealloc_count > 8) {
4192 ext4_mb_discard_lg_preallocations(sb, lg,
4193 order, lg_prealloc_count);
4194 return;
4196 return ;
4200 * release all resource we used in allocation
4202 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4204 struct ext4_prealloc_space *pa = ac->ac_pa;
4205 if (pa) {
4206 if (pa->pa_type == MB_GROUP_PA) {
4207 /* see comment in ext4_mb_use_group_pa() */
4208 spin_lock(&pa->pa_lock);
4209 pa->pa_pstart += ac->ac_b_ex.fe_len;
4210 pa->pa_lstart += ac->ac_b_ex.fe_len;
4211 pa->pa_free -= ac->ac_b_ex.fe_len;
4212 pa->pa_len -= ac->ac_b_ex.fe_len;
4213 spin_unlock(&pa->pa_lock);
4216 if (ac->alloc_semp)
4217 up_read(ac->alloc_semp);
4218 if (pa) {
4220 * We want to add the pa to the right bucket.
4221 * Remove it from the list and while adding
4222 * make sure the list to which we are adding
4223 * doesn't grow big. We need to release
4224 * alloc_semp before calling ext4_mb_add_n_trim()
4226 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4227 spin_lock(pa->pa_obj_lock);
4228 list_del_rcu(&pa->pa_inode_list);
4229 spin_unlock(pa->pa_obj_lock);
4230 ext4_mb_add_n_trim(ac);
4232 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4234 if (ac->ac_bitmap_page)
4235 page_cache_release(ac->ac_bitmap_page);
4236 if (ac->ac_buddy_page)
4237 page_cache_release(ac->ac_buddy_page);
4238 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4239 mutex_unlock(&ac->ac_lg->lg_mutex);
4240 ext4_mb_collect_stats(ac);
4241 return 0;
4244 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4246 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4247 int ret;
4248 int freed = 0;
4250 trace_ext4_mb_discard_preallocations(sb, needed);
4251 for (i = 0; i < ngroups && needed > 0; i++) {
4252 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4253 freed += ret;
4254 needed -= ret;
4257 return freed;
4261 * Main entry point into mballoc to allocate blocks
4262 * it tries to use preallocation first, then falls back
4263 * to usual allocation
4265 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4266 struct ext4_allocation_request *ar, int *errp)
4268 int freed;
4269 struct ext4_allocation_context *ac = NULL;
4270 struct ext4_sb_info *sbi;
4271 struct super_block *sb;
4272 ext4_fsblk_t block = 0;
4273 unsigned int inquota = 0;
4274 unsigned int reserv_blks = 0;
4276 sb = ar->inode->i_sb;
4277 sbi = EXT4_SB(sb);
4279 trace_ext4_request_blocks(ar);
4282 * For delayed allocation, we could skip the ENOSPC and
4283 * EDQUOT check, as blocks and quotas have been already
4284 * reserved when data being copied into pagecache.
4286 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4287 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4288 else {
4289 /* Without delayed allocation we need to verify
4290 * there is enough free blocks to do block allocation
4291 * and verify allocation doesn't exceed the quota limits.
4293 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4294 /* let others to free the space */
4295 yield();
4296 ar->len = ar->len >> 1;
4298 if (!ar->len) {
4299 *errp = -ENOSPC;
4300 return 0;
4302 reserv_blks = ar->len;
4303 while (ar->len && dquot_alloc_block(ar->inode, ar->len)) {
4304 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4305 ar->len--;
4307 inquota = ar->len;
4308 if (ar->len == 0) {
4309 *errp = -EDQUOT;
4310 goto out;
4314 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4315 if (!ac) {
4316 ar->len = 0;
4317 *errp = -ENOMEM;
4318 goto out;
4321 *errp = ext4_mb_initialize_context(ac, ar);
4322 if (*errp) {
4323 ar->len = 0;
4324 goto out;
4327 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4328 if (!ext4_mb_use_preallocated(ac)) {
4329 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4330 ext4_mb_normalize_request(ac, ar);
4331 repeat:
4332 /* allocate space in core */
4333 *errp = ext4_mb_regular_allocator(ac);
4334 if (*errp)
4335 goto errout;
4337 /* as we've just preallocated more space than
4338 * user requested orinally, we store allocated
4339 * space in a special descriptor */
4340 if (ac->ac_status == AC_STATUS_FOUND &&
4341 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4342 ext4_mb_new_preallocation(ac);
4344 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4345 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4346 if (*errp == -EAGAIN) {
4348 * drop the reference that we took
4349 * in ext4_mb_use_best_found
4351 ext4_mb_release_context(ac);
4352 ac->ac_b_ex.fe_group = 0;
4353 ac->ac_b_ex.fe_start = 0;
4354 ac->ac_b_ex.fe_len = 0;
4355 ac->ac_status = AC_STATUS_CONTINUE;
4356 goto repeat;
4357 } else if (*errp)
4358 errout:
4359 ext4_discard_allocated_blocks(ac);
4360 else {
4361 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4362 ar->len = ac->ac_b_ex.fe_len;
4364 } else {
4365 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4366 if (freed)
4367 goto repeat;
4368 *errp = -ENOSPC;
4371 if (*errp) {
4372 ac->ac_b_ex.fe_len = 0;
4373 ar->len = 0;
4374 ext4_mb_show_ac(ac);
4376 ext4_mb_release_context(ac);
4377 out:
4378 if (ac)
4379 kmem_cache_free(ext4_ac_cachep, ac);
4380 if (inquota && ar->len < inquota)
4381 dquot_free_block(ar->inode, inquota - ar->len);
4382 if (!ar->len) {
4383 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4384 /* release all the reserved blocks if non delalloc */
4385 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4386 reserv_blks);
4389 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4391 return block;
4395 * We can merge two free data extents only if the physical blocks
4396 * are contiguous, AND the extents were freed by the same transaction,
4397 * AND the blocks are associated with the same group.
4399 static int can_merge(struct ext4_free_data *entry1,
4400 struct ext4_free_data *entry2)
4402 if ((entry1->t_tid == entry2->t_tid) &&
4403 (entry1->group == entry2->group) &&
4404 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4405 return 1;
4406 return 0;
4409 static noinline_for_stack int
4410 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4411 struct ext4_free_data *new_entry)
4413 ext4_group_t group = e4b->bd_group;
4414 ext4_grpblk_t block;
4415 struct ext4_free_data *entry;
4416 struct ext4_group_info *db = e4b->bd_info;
4417 struct super_block *sb = e4b->bd_sb;
4418 struct ext4_sb_info *sbi = EXT4_SB(sb);
4419 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4420 struct rb_node *parent = NULL, *new_node;
4422 BUG_ON(!ext4_handle_valid(handle));
4423 BUG_ON(e4b->bd_bitmap_page == NULL);
4424 BUG_ON(e4b->bd_buddy_page == NULL);
4426 new_node = &new_entry->node;
4427 block = new_entry->start_blk;
4429 if (!*n) {
4430 /* first free block exent. We need to
4431 protect buddy cache from being freed,
4432 * otherwise we'll refresh it from
4433 * on-disk bitmap and lose not-yet-available
4434 * blocks */
4435 page_cache_get(e4b->bd_buddy_page);
4436 page_cache_get(e4b->bd_bitmap_page);
4438 while (*n) {
4439 parent = *n;
4440 entry = rb_entry(parent, struct ext4_free_data, node);
4441 if (block < entry->start_blk)
4442 n = &(*n)->rb_left;
4443 else if (block >= (entry->start_blk + entry->count))
4444 n = &(*n)->rb_right;
4445 else {
4446 ext4_grp_locked_error(sb, group, 0,
4447 ext4_group_first_block_no(sb, group) + block,
4448 "Block already on to-be-freed list");
4449 return 0;
4453 rb_link_node(new_node, parent, n);
4454 rb_insert_color(new_node, &db->bb_free_root);
4456 /* Now try to see the extent can be merged to left and right */
4457 node = rb_prev(new_node);
4458 if (node) {
4459 entry = rb_entry(node, struct ext4_free_data, node);
4460 if (can_merge(entry, new_entry)) {
4461 new_entry->start_blk = entry->start_blk;
4462 new_entry->count += entry->count;
4463 rb_erase(node, &(db->bb_free_root));
4464 spin_lock(&sbi->s_md_lock);
4465 list_del(&entry->list);
4466 spin_unlock(&sbi->s_md_lock);
4467 kmem_cache_free(ext4_free_ext_cachep, entry);
4471 node = rb_next(new_node);
4472 if (node) {
4473 entry = rb_entry(node, struct ext4_free_data, node);
4474 if (can_merge(new_entry, entry)) {
4475 new_entry->count += entry->count;
4476 rb_erase(node, &(db->bb_free_root));
4477 spin_lock(&sbi->s_md_lock);
4478 list_del(&entry->list);
4479 spin_unlock(&sbi->s_md_lock);
4480 kmem_cache_free(ext4_free_ext_cachep, entry);
4483 /* Add the extent to transaction's private list */
4484 spin_lock(&sbi->s_md_lock);
4485 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4486 spin_unlock(&sbi->s_md_lock);
4487 return 0;
4491 * ext4_free_blocks() -- Free given blocks and update quota
4492 * @handle: handle for this transaction
4493 * @inode: inode
4494 * @block: start physical block to free
4495 * @count: number of blocks to count
4496 * @metadata: Are these metadata blocks
4498 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4499 struct buffer_head *bh, ext4_fsblk_t block,
4500 unsigned long count, int flags)
4502 struct buffer_head *bitmap_bh = NULL;
4503 struct super_block *sb = inode->i_sb;
4504 struct ext4_group_desc *gdp;
4505 unsigned long freed = 0;
4506 unsigned int overflow;
4507 ext4_grpblk_t bit;
4508 struct buffer_head *gd_bh;
4509 ext4_group_t block_group;
4510 struct ext4_sb_info *sbi;
4511 struct ext4_buddy e4b;
4512 int err = 0;
4513 int ret;
4515 if (bh) {
4516 if (block)
4517 BUG_ON(block != bh->b_blocknr);
4518 else
4519 block = bh->b_blocknr;
4522 sbi = EXT4_SB(sb);
4523 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4524 !ext4_data_block_valid(sbi, block, count)) {
4525 ext4_error(sb, "Freeing blocks not in datazone - "
4526 "block = %llu, count = %lu", block, count);
4527 goto error_return;
4530 ext4_debug("freeing block %llu\n", block);
4531 trace_ext4_free_blocks(inode, block, count, flags);
4533 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4534 struct buffer_head *tbh = bh;
4535 int i;
4537 BUG_ON(bh && (count > 1));
4539 for (i = 0; i < count; i++) {
4540 if (!bh)
4541 tbh = sb_find_get_block(inode->i_sb,
4542 block + i);
4543 if (unlikely(!tbh))
4544 continue;
4545 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4546 inode, tbh, block + i);
4551 * We need to make sure we don't reuse the freed block until
4552 * after the transaction is committed, which we can do by
4553 * treating the block as metadata, below. We make an
4554 * exception if the inode is to be written in writeback mode
4555 * since writeback mode has weak data consistency guarantees.
4557 if (!ext4_should_writeback_data(inode))
4558 flags |= EXT4_FREE_BLOCKS_METADATA;
4560 do_more:
4561 overflow = 0;
4562 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4565 * Check to see if we are freeing blocks across a group
4566 * boundary.
4568 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4569 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4570 count -= overflow;
4572 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4573 if (!bitmap_bh) {
4574 err = -EIO;
4575 goto error_return;
4577 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4578 if (!gdp) {
4579 err = -EIO;
4580 goto error_return;
4583 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4584 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4585 in_range(block, ext4_inode_table(sb, gdp),
4586 EXT4_SB(sb)->s_itb_per_group) ||
4587 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4588 EXT4_SB(sb)->s_itb_per_group)) {
4590 ext4_error(sb, "Freeing blocks in system zone - "
4591 "Block = %llu, count = %lu", block, count);
4592 /* err = 0. ext4_std_error should be a no op */
4593 goto error_return;
4596 BUFFER_TRACE(bitmap_bh, "getting write access");
4597 err = ext4_journal_get_write_access(handle, bitmap_bh);
4598 if (err)
4599 goto error_return;
4602 * We are about to modify some metadata. Call the journal APIs
4603 * to unshare ->b_data if a currently-committing transaction is
4604 * using it
4606 BUFFER_TRACE(gd_bh, "get_write_access");
4607 err = ext4_journal_get_write_access(handle, gd_bh);
4608 if (err)
4609 goto error_return;
4610 #ifdef AGGRESSIVE_CHECK
4612 int i;
4613 for (i = 0; i < count; i++)
4614 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4616 #endif
4617 trace_ext4_mballoc_free(sb, inode, block_group, bit, count);
4619 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4620 if (err)
4621 goto error_return;
4623 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4624 struct ext4_free_data *new_entry;
4626 * blocks being freed are metadata. these blocks shouldn't
4627 * be used until this transaction is committed
4629 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4630 new_entry->start_blk = bit;
4631 new_entry->group = block_group;
4632 new_entry->count = count;
4633 new_entry->t_tid = handle->h_transaction->t_tid;
4635 ext4_lock_group(sb, block_group);
4636 mb_clear_bits(bitmap_bh->b_data, bit, count);
4637 ext4_mb_free_metadata(handle, &e4b, new_entry);
4638 } else {
4639 /* need to update group_info->bb_free and bitmap
4640 * with group lock held. generate_buddy look at
4641 * them with group lock_held
4643 ext4_lock_group(sb, block_group);
4644 mb_clear_bits(bitmap_bh->b_data, bit, count);
4645 mb_free_blocks(inode, &e4b, bit, count);
4646 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4649 ret = ext4_free_blks_count(sb, gdp) + count;
4650 ext4_free_blks_set(sb, gdp, ret);
4651 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4652 ext4_unlock_group(sb, block_group);
4653 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4655 if (sbi->s_log_groups_per_flex) {
4656 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4657 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4660 ext4_mb_unload_buddy(&e4b);
4662 freed += count;
4664 /* We dirtied the bitmap block */
4665 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4666 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4668 /* And the group descriptor block */
4669 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4670 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4671 if (!err)
4672 err = ret;
4674 if (overflow && !err) {
4675 block += count;
4676 count = overflow;
4677 put_bh(bitmap_bh);
4678 goto do_more;
4680 ext4_mark_super_dirty(sb);
4681 error_return:
4682 if (freed)
4683 dquot_free_block(inode, freed);
4684 brelse(bitmap_bh);
4685 ext4_std_error(sb, err);
4686 return;
4690 * ext4_trim_extent -- function to TRIM one single free extent in the group
4691 * @sb: super block for the file system
4692 * @start: starting block of the free extent in the alloc. group
4693 * @count: number of blocks to TRIM
4694 * @group: alloc. group we are working with
4695 * @e4b: ext4 buddy for the group
4697 * Trim "count" blocks starting at "start" in the "group". To assure that no
4698 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4699 * be called with under the group lock.
4701 static int ext4_trim_extent(struct super_block *sb, int start, int count,
4702 ext4_group_t group, struct ext4_buddy *e4b)
4704 struct ext4_free_extent ex;
4705 int ret = 0;
4707 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4709 ex.fe_start = start;
4710 ex.fe_group = group;
4711 ex.fe_len = count;
4714 * Mark blocks used, so no one can reuse them while
4715 * being trimmed.
4717 mb_mark_used(e4b, &ex);
4718 ext4_unlock_group(sb, group);
4720 ret = ext4_issue_discard(sb, group, start, count);
4721 if (ret)
4722 ext4_std_error(sb, ret);
4724 ext4_lock_group(sb, group);
4725 mb_free_blocks(NULL, e4b, start, ex.fe_len);
4726 return ret;
4730 * ext4_trim_all_free -- function to trim all free space in alloc. group
4731 * @sb: super block for file system
4732 * @e4b: ext4 buddy
4733 * @start: first group block to examine
4734 * @max: last group block to examine
4735 * @minblocks: minimum extent block count
4737 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4738 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4739 * the extent.
4742 * ext4_trim_all_free walks through group's block bitmap searching for free
4743 * extents. When the free extent is found, mark it as used in group buddy
4744 * bitmap. Then issue a TRIM command on this extent and free the extent in
4745 * the group buddy bitmap. This is done until whole group is scanned.
4747 ext4_grpblk_t ext4_trim_all_free(struct super_block *sb, struct ext4_buddy *e4b,
4748 ext4_grpblk_t start, ext4_grpblk_t max, ext4_grpblk_t minblocks)
4750 void *bitmap;
4751 ext4_grpblk_t next, count = 0;
4752 ext4_group_t group;
4753 int ret = 0;
4755 BUG_ON(e4b == NULL);
4757 bitmap = e4b->bd_bitmap;
4758 group = e4b->bd_group;
4759 start = (e4b->bd_info->bb_first_free > start) ?
4760 e4b->bd_info->bb_first_free : start;
4761 ext4_lock_group(sb, group);
4763 while (start < max) {
4764 start = mb_find_next_zero_bit(bitmap, max, start);
4765 if (start >= max)
4766 break;
4767 next = mb_find_next_bit(bitmap, max, start);
4769 if ((next - start) >= minblocks) {
4770 ret = ext4_trim_extent(sb, start,
4771 next - start, group, e4b);
4772 if (ret < 0)
4773 break;
4774 count += next - start;
4776 start = next + 1;
4778 if (fatal_signal_pending(current)) {
4779 count = -ERESTARTSYS;
4780 break;
4783 if (need_resched()) {
4784 ext4_unlock_group(sb, group);
4785 cond_resched();
4786 ext4_lock_group(sb, group);
4789 if ((e4b->bd_info->bb_free - count) < minblocks)
4790 break;
4792 ext4_unlock_group(sb, group);
4794 ext4_debug("trimmed %d blocks in the group %d\n",
4795 count, group);
4797 if (ret < 0)
4798 count = ret;
4800 return count;
4804 * ext4_trim_fs() -- trim ioctl handle function
4805 * @sb: superblock for filesystem
4806 * @range: fstrim_range structure
4808 * start: First Byte to trim
4809 * len: number of Bytes to trim from start
4810 * minlen: minimum extent length in Bytes
4811 * ext4_trim_fs goes through all allocation groups containing Bytes from
4812 * start to start+len. For each such a group ext4_trim_all_free function
4813 * is invoked to trim all free space.
4815 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
4817 struct ext4_buddy e4b;
4818 ext4_group_t first_group, last_group;
4819 ext4_group_t group, ngroups = ext4_get_groups_count(sb);
4820 ext4_grpblk_t cnt = 0, first_block, last_block;
4821 uint64_t start, len, minlen, trimmed;
4822 int ret = 0;
4824 start = range->start >> sb->s_blocksize_bits;
4825 len = range->len >> sb->s_blocksize_bits;
4826 minlen = range->minlen >> sb->s_blocksize_bits;
4827 trimmed = 0;
4829 if (unlikely(minlen > EXT4_BLOCKS_PER_GROUP(sb)))
4830 return -EINVAL;
4832 /* Determine first and last group to examine based on start and len */
4833 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
4834 &first_group, &first_block);
4835 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) (start + len),
4836 &last_group, &last_block);
4837 last_group = (last_group > ngroups - 1) ? ngroups - 1 : last_group;
4838 last_block = EXT4_BLOCKS_PER_GROUP(sb);
4840 if (first_group > last_group)
4841 return -EINVAL;
4843 for (group = first_group; group <= last_group; group++) {
4844 ret = ext4_mb_load_buddy(sb, group, &e4b);
4845 if (ret) {
4846 ext4_error(sb, "Error in loading buddy "
4847 "information for %u", group);
4848 break;
4851 if (len >= EXT4_BLOCKS_PER_GROUP(sb))
4852 len -= (EXT4_BLOCKS_PER_GROUP(sb) - first_block);
4853 else
4854 last_block = len;
4856 if (e4b.bd_info->bb_free >= minlen) {
4857 cnt = ext4_trim_all_free(sb, &e4b, first_block,
4858 last_block, minlen);
4859 if (cnt < 0) {
4860 ret = cnt;
4861 ext4_mb_unload_buddy(&e4b);
4862 break;
4865 ext4_mb_unload_buddy(&e4b);
4866 trimmed += cnt;
4867 first_block = 0;
4869 range->len = trimmed * sb->s_blocksize;
4871 return ret;