ext4: Improve error handling in mballoc
[linux-2.6/mini2440.git] / fs / ext4 / mballoc.c
blob500d3920d41de0851939e203094b6f5741b097a1
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"
26 * MUSTDO:
27 * - test ext4_ext_search_left() and ext4_ext_search_right()
28 * - search for metadata in few groups
30 * TODO v4:
31 * - normalization should take into account whether file is still open
32 * - discard preallocations if no free space left (policy?)
33 * - don't normalize tails
34 * - quota
35 * - reservation for superuser
37 * TODO v3:
38 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
39 * - track min/max extents in each group for better group selection
40 * - mb_mark_used() may allocate chunk right after splitting buddy
41 * - tree of groups sorted by number of free blocks
42 * - error handling
46 * The allocation request involve request for multiple number of blocks
47 * near to the goal(block) value specified.
49 * During initialization phase of the allocator we decide to use the group
50 * preallocation or inode preallocation depending on the size file. The
51 * size of the file could be the resulting file size we would have after
52 * allocation or the current file size which ever is larger. If the size is
53 * less that sbi->s_mb_stream_request we select the group
54 * preallocation. The default value of s_mb_stream_request is 16
55 * blocks. This can also be tuned via
56 * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms
57 * of number of blocks.
59 * The main motivation for having small file use group preallocation is to
60 * ensure that we have small file closer in the disk.
62 * First stage the allocator looks at the inode prealloc list
63 * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for
64 * this particular inode. The inode prealloc space is represented as:
66 * pa_lstart -> the logical start block for this prealloc space
67 * pa_pstart -> the physical start block for this prealloc space
68 * pa_len -> lenght for this prealloc space
69 * pa_free -> free space available in this prealloc space
71 * The inode preallocation space is used looking at the _logical_ start
72 * block. If only the logical file block falls within the range of prealloc
73 * space we will consume the particular prealloc space. This make sure that
74 * that the we have contiguous physical blocks representing the file blocks
76 * The important thing to be noted in case of inode prealloc space is that
77 * we don't modify the values associated to inode prealloc space except
78 * pa_free.
80 * If we are not able to find blocks in the inode prealloc space and if we
81 * have the group allocation flag set then we look at the locality group
82 * prealloc space. These are per CPU prealloc list repreasented as
84 * ext4_sb_info.s_locality_groups[smp_processor_id()]
86 * The reason for having a per cpu locality group is to reduce the contention
87 * between CPUs. It is possible to get scheduled at this point.
89 * The locality group prealloc space is used looking at whether we have
90 * enough free space (pa_free) withing the prealloc space.
92 * If we can't allocate blocks via inode prealloc or/and locality group
93 * prealloc then we look at the buddy cache. The buddy cache is represented
94 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
95 * mapped to the buddy and bitmap information regarding different
96 * groups. The buddy information is attached to buddy cache inode so that
97 * we can access them through the page cache. The information regarding
98 * each group is loaded via ext4_mb_load_buddy. The information involve
99 * block bitmap and buddy information. The information are stored in the
100 * inode as:
102 * { page }
103 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
106 * one block each for bitmap and buddy information. So for each group we
107 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
108 * blocksize) blocks. So it can have information regarding groups_per_page
109 * which is blocks_per_page/2
111 * The buddy cache inode is not stored on disk. The inode is thrown
112 * away when the filesystem is unmounted.
114 * We look for count number of blocks in the buddy cache. If we were able
115 * to locate that many free blocks we return with additional information
116 * regarding rest of the contiguous physical block available
118 * Before allocating blocks via buddy cache we normalize the request
119 * blocks. This ensure we ask for more blocks that we needed. The extra
120 * blocks that we get after allocation is added to the respective prealloc
121 * list. In case of inode preallocation we follow a list of heuristics
122 * based on file size. This can be found in ext4_mb_normalize_request. If
123 * we are doing a group prealloc we try to normalize the request to
124 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is set to
125 * 512 blocks. This can be tuned via
126 * /proc/fs/ext4/<partition/group_prealloc. The value is represented in
127 * terms of number of blocks. If we have mounted the file system with -O
128 * stripe=<value> option the group prealloc request is normalized to the
129 * stripe value (sbi->s_stripe)
131 * The regular allocator(using the buddy cache) support few tunables.
133 * /proc/fs/ext4/<partition>/min_to_scan
134 * /proc/fs/ext4/<partition>/max_to_scan
135 * /proc/fs/ext4/<partition>/order2_req
137 * The regular allocator use buddy scan only if the request len is power of
138 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
139 * value of s_mb_order2_reqs can be tuned via
140 * /proc/fs/ext4/<partition>/order2_req. If the request len is equal to
141 * stripe size (sbi->s_stripe), we try to search for contigous block in
142 * stripe size. This should result in better allocation on RAID setup. If
143 * not we search in the specific group using bitmap for best extents. The
144 * tunable min_to_scan and max_to_scan controll the behaviour here.
145 * min_to_scan indicate how long the mballoc __must__ look for a best
146 * extent and max_to_scanindicate how long the mballoc __can__ look for a
147 * best extent in the found extents. Searching for the blocks starts with
148 * the group specified as the goal value in allocation context via
149 * ac_g_ex. Each group is first checked based on the criteria whether it
150 * can used for allocation. ext4_mb_good_group explains how the groups are
151 * checked.
153 * Both the prealloc space are getting populated as above. So for the first
154 * request we will hit the buddy cache which will result in this prealloc
155 * space getting filled. The prealloc space is then later used for the
156 * subsequent request.
160 * mballoc operates on the following data:
161 * - on-disk bitmap
162 * - in-core buddy (actually includes buddy and bitmap)
163 * - preallocation descriptors (PAs)
165 * there are two types of preallocations:
166 * - inode
167 * assiged to specific inode and can be used for this inode only.
168 * it describes part of inode's space preallocated to specific
169 * physical blocks. any block from that preallocated can be used
170 * independent. the descriptor just tracks number of blocks left
171 * unused. so, before taking some block from descriptor, one must
172 * make sure corresponded logical block isn't allocated yet. this
173 * also means that freeing any block within descriptor's range
174 * must discard all preallocated blocks.
175 * - locality group
176 * assigned to specific locality group which does not translate to
177 * permanent set of inodes: inode can join and leave group. space
178 * from this type of preallocation can be used for any inode. thus
179 * it's consumed from the beginning to the end.
181 * relation between them can be expressed as:
182 * in-core buddy = on-disk bitmap + preallocation descriptors
184 * this mean blocks mballoc considers used are:
185 * - allocated blocks (persistent)
186 * - preallocated blocks (non-persistent)
188 * consistency in mballoc world means that at any time a block is either
189 * free or used in ALL structures. notice: "any time" should not be read
190 * literally -- time is discrete and delimited by locks.
192 * to keep it simple, we don't use block numbers, instead we count number of
193 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
195 * all operations can be expressed as:
196 * - init buddy: buddy = on-disk + PAs
197 * - new PA: buddy += N; PA = N
198 * - use inode PA: on-disk += N; PA -= N
199 * - discard inode PA buddy -= on-disk - PA; PA = 0
200 * - use locality group PA on-disk += N; PA -= N
201 * - discard locality group PA buddy -= PA; PA = 0
202 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
203 * is used in real operation because we can't know actual used
204 * bits from PA, only from on-disk bitmap
206 * if we follow this strict logic, then all operations above should be atomic.
207 * given some of them can block, we'd have to use something like semaphores
208 * killing performance on high-end SMP hardware. let's try to relax it using
209 * the following knowledge:
210 * 1) if buddy is referenced, it's already initialized
211 * 2) while block is used in buddy and the buddy is referenced,
212 * nobody can re-allocate that block
213 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
214 * bit set and PA claims same block, it's OK. IOW, one can set bit in
215 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
216 * block
218 * so, now we're building a concurrency table:
219 * - init buddy vs.
220 * - new PA
221 * blocks for PA are allocated in the buddy, buddy must be referenced
222 * until PA is linked to allocation group to avoid concurrent buddy init
223 * - use inode PA
224 * we need to make sure that either on-disk bitmap or PA has uptodate data
225 * given (3) we care that PA-=N operation doesn't interfere with init
226 * - discard inode PA
227 * the simplest way would be to have buddy initialized by the discard
228 * - use locality group PA
229 * again PA-=N must be serialized with init
230 * - discard locality group PA
231 * the simplest way would be to have buddy initialized by the discard
232 * - new PA vs.
233 * - use inode PA
234 * i_data_sem serializes them
235 * - discard inode PA
236 * discard process must wait until PA isn't used by another process
237 * - use locality group PA
238 * some mutex should serialize them
239 * - discard locality group PA
240 * discard process must wait until PA isn't used by another process
241 * - use inode PA
242 * - use inode PA
243 * i_data_sem or another mutex should serializes them
244 * - discard inode PA
245 * discard process must wait until PA isn't used by another process
246 * - use locality group PA
247 * nothing wrong here -- they're different PAs covering different blocks
248 * - discard locality group PA
249 * discard process must wait until PA isn't used by another process
251 * now we're ready to make few consequences:
252 * - PA is referenced and while it is no discard is possible
253 * - PA is referenced until block isn't marked in on-disk bitmap
254 * - PA changes only after on-disk bitmap
255 * - discard must not compete with init. either init is done before
256 * any discard or they're serialized somehow
257 * - buddy init as sum of on-disk bitmap and PAs is done atomically
259 * a special case when we've used PA to emptiness. no need to modify buddy
260 * in this case, but we should care about concurrent init
265 * Logic in few words:
267 * - allocation:
268 * load group
269 * find blocks
270 * mark bits in on-disk bitmap
271 * release group
273 * - use preallocation:
274 * find proper PA (per-inode or group)
275 * load group
276 * mark bits in on-disk bitmap
277 * release group
278 * release PA
280 * - free:
281 * load group
282 * mark bits in on-disk bitmap
283 * release group
285 * - discard preallocations in group:
286 * mark PAs deleted
287 * move them onto local list
288 * load on-disk bitmap
289 * load group
290 * remove PA from object (inode or locality group)
291 * mark free blocks in-core
293 * - discard inode's preallocations:
297 * Locking rules
299 * Locks:
300 * - bitlock on a group (group)
301 * - object (inode/locality) (object)
302 * - per-pa lock (pa)
304 * Paths:
305 * - new pa
306 * object
307 * group
309 * - find and use pa:
310 * pa
312 * - release consumed pa:
313 * pa
314 * group
315 * object
317 * - generate in-core bitmap:
318 * group
319 * pa
321 * - discard all for given object (inode, locality group):
322 * object
323 * pa
324 * group
326 * - discard all for given group:
327 * group
328 * pa
329 * group
330 * object
334 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
336 #if BITS_PER_LONG == 64
337 *bit += ((unsigned long) addr & 7UL) << 3;
338 addr = (void *) ((unsigned long) addr & ~7UL);
339 #elif BITS_PER_LONG == 32
340 *bit += ((unsigned long) addr & 3UL) << 3;
341 addr = (void *) ((unsigned long) addr & ~3UL);
342 #else
343 #error "how many bits you are?!"
344 #endif
345 return addr;
348 static inline int mb_test_bit(int bit, void *addr)
351 * ext4_test_bit on architecture like powerpc
352 * needs unsigned long aligned address
354 addr = mb_correct_addr_and_bit(&bit, addr);
355 return ext4_test_bit(bit, addr);
358 static inline void mb_set_bit(int bit, void *addr)
360 addr = mb_correct_addr_and_bit(&bit, addr);
361 ext4_set_bit(bit, addr);
364 static inline void mb_set_bit_atomic(spinlock_t *lock, int bit, void *addr)
366 addr = mb_correct_addr_and_bit(&bit, addr);
367 ext4_set_bit_atomic(lock, bit, addr);
370 static inline void mb_clear_bit(int bit, void *addr)
372 addr = mb_correct_addr_and_bit(&bit, addr);
373 ext4_clear_bit(bit, addr);
376 static inline void mb_clear_bit_atomic(spinlock_t *lock, int bit, void *addr)
378 addr = mb_correct_addr_and_bit(&bit, addr);
379 ext4_clear_bit_atomic(lock, bit, addr);
382 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
384 int fix = 0, ret, tmpmax;
385 addr = mb_correct_addr_and_bit(&fix, addr);
386 tmpmax = max + fix;
387 start += fix;
389 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
390 if (ret > max)
391 return max;
392 return ret;
395 static inline int mb_find_next_bit(void *addr, int max, int start)
397 int fix = 0, ret, tmpmax;
398 addr = mb_correct_addr_and_bit(&fix, addr);
399 tmpmax = max + fix;
400 start += fix;
402 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
403 if (ret > max)
404 return max;
405 return ret;
408 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
410 char *bb;
412 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
413 BUG_ON(max == NULL);
415 if (order > e4b->bd_blkbits + 1) {
416 *max = 0;
417 return NULL;
420 /* at order 0 we see each particular block */
421 *max = 1 << (e4b->bd_blkbits + 3);
422 if (order == 0)
423 return EXT4_MB_BITMAP(e4b);
425 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
426 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
428 return bb;
431 #ifdef DOUBLE_CHECK
432 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
433 int first, int count)
435 int i;
436 struct super_block *sb = e4b->bd_sb;
438 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
439 return;
440 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
441 for (i = 0; i < count; i++) {
442 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
443 ext4_fsblk_t blocknr;
444 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
445 blocknr += first + i;
446 blocknr +=
447 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
449 ext4_error(sb, __func__, "double-free of inode"
450 " %lu's block %llu(bit %u in group %lu)\n",
451 inode ? inode->i_ino : 0, blocknr,
452 first + i, e4b->bd_group);
454 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
458 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
460 int i;
462 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
463 return;
464 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
465 for (i = 0; i < count; i++) {
466 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
467 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
471 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
473 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
474 unsigned char *b1, *b2;
475 int i;
476 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
477 b2 = (unsigned char *) bitmap;
478 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
479 if (b1[i] != b2[i]) {
480 printk("corruption in group %lu at byte %u(%u):"
481 " %x in copy != %x on disk/prealloc\n",
482 e4b->bd_group, i, i * 8, b1[i], b2[i]);
483 BUG();
489 #else
490 static inline void mb_free_blocks_double(struct inode *inode,
491 struct ext4_buddy *e4b, int first, int count)
493 return;
495 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
496 int first, int count)
498 return;
500 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
502 return;
504 #endif
506 #ifdef AGGRESSIVE_CHECK
508 #define MB_CHECK_ASSERT(assert) \
509 do { \
510 if (!(assert)) { \
511 printk(KERN_EMERG \
512 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
513 function, file, line, # assert); \
514 BUG(); \
516 } while (0)
518 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
519 const char *function, int line)
521 struct super_block *sb = e4b->bd_sb;
522 int order = e4b->bd_blkbits + 1;
523 int max;
524 int max2;
525 int i;
526 int j;
527 int k;
528 int count;
529 struct ext4_group_info *grp;
530 int fragments = 0;
531 int fstart;
532 struct list_head *cur;
533 void *buddy;
534 void *buddy2;
536 if (!test_opt(sb, MBALLOC))
537 return 0;
540 static int mb_check_counter;
541 if (mb_check_counter++ % 100 != 0)
542 return 0;
545 while (order > 1) {
546 buddy = mb_find_buddy(e4b, order, &max);
547 MB_CHECK_ASSERT(buddy);
548 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
549 MB_CHECK_ASSERT(buddy2);
550 MB_CHECK_ASSERT(buddy != buddy2);
551 MB_CHECK_ASSERT(max * 2 == max2);
553 count = 0;
554 for (i = 0; i < max; i++) {
556 if (mb_test_bit(i, buddy)) {
557 /* only single bit in buddy2 may be 1 */
558 if (!mb_test_bit(i << 1, buddy2)) {
559 MB_CHECK_ASSERT(
560 mb_test_bit((i<<1)+1, buddy2));
561 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
562 MB_CHECK_ASSERT(
563 mb_test_bit(i << 1, buddy2));
565 continue;
568 /* both bits in buddy2 must be 0 */
569 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
570 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
572 for (j = 0; j < (1 << order); j++) {
573 k = (i * (1 << order)) + j;
574 MB_CHECK_ASSERT(
575 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
577 count++;
579 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
580 order--;
583 fstart = -1;
584 buddy = mb_find_buddy(e4b, 0, &max);
585 for (i = 0; i < max; i++) {
586 if (!mb_test_bit(i, buddy)) {
587 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
588 if (fstart == -1) {
589 fragments++;
590 fstart = i;
592 continue;
594 fstart = -1;
595 /* check used bits only */
596 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
597 buddy2 = mb_find_buddy(e4b, j, &max2);
598 k = i >> j;
599 MB_CHECK_ASSERT(k < max2);
600 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
603 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
604 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
606 grp = ext4_get_group_info(sb, e4b->bd_group);
607 buddy = mb_find_buddy(e4b, 0, &max);
608 list_for_each(cur, &grp->bb_prealloc_list) {
609 ext4_group_t groupnr;
610 struct ext4_prealloc_space *pa;
611 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
612 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
613 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
614 for (i = 0; i < pa->pa_len; i++)
615 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
617 return 0;
619 #undef MB_CHECK_ASSERT
620 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
621 __FILE__, __func__, __LINE__)
622 #else
623 #define mb_check_buddy(e4b)
624 #endif
626 /* FIXME!! need more doc */
627 static void ext4_mb_mark_free_simple(struct super_block *sb,
628 void *buddy, unsigned first, int len,
629 struct ext4_group_info *grp)
631 struct ext4_sb_info *sbi = EXT4_SB(sb);
632 unsigned short min;
633 unsigned short max;
634 unsigned short chunk;
635 unsigned short border;
637 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
639 border = 2 << sb->s_blocksize_bits;
641 while (len > 0) {
642 /* find how many blocks can be covered since this position */
643 max = ffs(first | border) - 1;
645 /* find how many blocks of power 2 we need to mark */
646 min = fls(len) - 1;
648 if (max < min)
649 min = max;
650 chunk = 1 << min;
652 /* mark multiblock chunks only */
653 grp->bb_counters[min]++;
654 if (min > 0)
655 mb_clear_bit(first >> min,
656 buddy + sbi->s_mb_offsets[min]);
658 len -= chunk;
659 first += chunk;
663 static void ext4_mb_generate_buddy(struct super_block *sb,
664 void *buddy, void *bitmap, ext4_group_t group)
666 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
667 unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
668 unsigned short i = 0;
669 unsigned short first;
670 unsigned short len;
671 unsigned free = 0;
672 unsigned fragments = 0;
673 unsigned long long period = get_cycles();
675 /* initialize buddy from bitmap which is aggregation
676 * of on-disk bitmap and preallocations */
677 i = mb_find_next_zero_bit(bitmap, max, 0);
678 grp->bb_first_free = i;
679 while (i < max) {
680 fragments++;
681 first = i;
682 i = mb_find_next_bit(bitmap, max, i);
683 len = i - first;
684 free += len;
685 if (len > 1)
686 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
687 else
688 grp->bb_counters[0]++;
689 if (i < max)
690 i = mb_find_next_zero_bit(bitmap, max, i);
692 grp->bb_fragments = fragments;
694 if (free != grp->bb_free) {
695 ext4_error(sb, __func__,
696 "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
697 group, free, grp->bb_free);
699 * If we intent to continue, we consider group descritor
700 * corrupt and update bb_free using bitmap value
702 grp->bb_free = free;
705 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
707 period = get_cycles() - period;
708 spin_lock(&EXT4_SB(sb)->s_bal_lock);
709 EXT4_SB(sb)->s_mb_buddies_generated++;
710 EXT4_SB(sb)->s_mb_generation_time += period;
711 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
714 /* The buddy information is attached the buddy cache inode
715 * for convenience. The information regarding each group
716 * is loaded via ext4_mb_load_buddy. The information involve
717 * block bitmap and buddy information. The information are
718 * stored in the inode as
720 * { page }
721 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
724 * one block each for bitmap and buddy information.
725 * So for each group we take up 2 blocks. A page can
726 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
727 * So it can have information regarding groups_per_page which
728 * is blocks_per_page/2
731 static int ext4_mb_init_cache(struct page *page, char *incore)
733 int blocksize;
734 int blocks_per_page;
735 int groups_per_page;
736 int err = 0;
737 int i;
738 ext4_group_t first_group;
739 int first_block;
740 struct super_block *sb;
741 struct buffer_head *bhs;
742 struct buffer_head **bh;
743 struct inode *inode;
744 char *data;
745 char *bitmap;
747 mb_debug("init page %lu\n", page->index);
749 inode = page->mapping->host;
750 sb = inode->i_sb;
751 blocksize = 1 << inode->i_blkbits;
752 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
754 groups_per_page = blocks_per_page >> 1;
755 if (groups_per_page == 0)
756 groups_per_page = 1;
758 /* allocate buffer_heads to read bitmaps */
759 if (groups_per_page > 1) {
760 err = -ENOMEM;
761 i = sizeof(struct buffer_head *) * groups_per_page;
762 bh = kzalloc(i, GFP_NOFS);
763 if (bh == NULL)
764 goto out;
765 } else
766 bh = &bhs;
768 first_group = page->index * blocks_per_page / 2;
770 /* read all groups the page covers into the cache */
771 for (i = 0; i < groups_per_page; i++) {
772 struct ext4_group_desc *desc;
774 if (first_group + i >= EXT4_SB(sb)->s_groups_count)
775 break;
777 err = -EIO;
778 desc = ext4_get_group_desc(sb, first_group + i, NULL);
779 if (desc == NULL)
780 goto out;
782 err = -ENOMEM;
783 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
784 if (bh[i] == NULL)
785 goto out;
787 if (bh_uptodate_or_lock(bh[i]))
788 continue;
790 spin_lock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
791 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
792 ext4_init_block_bitmap(sb, bh[i],
793 first_group + i, desc);
794 set_buffer_uptodate(bh[i]);
795 unlock_buffer(bh[i]);
796 spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
797 continue;
799 spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
800 get_bh(bh[i]);
801 bh[i]->b_end_io = end_buffer_read_sync;
802 submit_bh(READ, bh[i]);
803 mb_debug("read bitmap for group %lu\n", first_group + i);
806 /* wait for I/O completion */
807 for (i = 0; i < groups_per_page && bh[i]; i++)
808 wait_on_buffer(bh[i]);
810 err = -EIO;
811 for (i = 0; i < groups_per_page && bh[i]; i++)
812 if (!buffer_uptodate(bh[i]))
813 goto out;
815 err = 0;
816 first_block = page->index * blocks_per_page;
817 for (i = 0; i < blocks_per_page; i++) {
818 int group;
819 struct ext4_group_info *grinfo;
821 group = (first_block + i) >> 1;
822 if (group >= EXT4_SB(sb)->s_groups_count)
823 break;
826 * data carry information regarding this
827 * particular group in the format specified
828 * above
831 data = page_address(page) + (i * blocksize);
832 bitmap = bh[group - first_group]->b_data;
835 * We place the buddy block and bitmap block
836 * close together
838 if ((first_block + i) & 1) {
839 /* this is block of buddy */
840 BUG_ON(incore == NULL);
841 mb_debug("put buddy for group %u in page %lu/%x\n",
842 group, page->index, i * blocksize);
843 memset(data, 0xff, blocksize);
844 grinfo = ext4_get_group_info(sb, group);
845 grinfo->bb_fragments = 0;
846 memset(grinfo->bb_counters, 0,
847 sizeof(unsigned short)*(sb->s_blocksize_bits+2));
849 * incore got set to the group block bitmap below
851 ext4_mb_generate_buddy(sb, data, incore, group);
852 incore = NULL;
853 } else {
854 /* this is block of bitmap */
855 BUG_ON(incore != NULL);
856 mb_debug("put bitmap for group %u in page %lu/%x\n",
857 group, page->index, i * blocksize);
859 /* see comments in ext4_mb_put_pa() */
860 ext4_lock_group(sb, group);
861 memcpy(data, bitmap, blocksize);
863 /* mark all preallocated blks used in in-core bitmap */
864 ext4_mb_generate_from_pa(sb, data, group);
865 ext4_unlock_group(sb, group);
867 /* set incore so that the buddy information can be
868 * generated using this
870 incore = data;
873 SetPageUptodate(page);
875 out:
876 if (bh) {
877 for (i = 0; i < groups_per_page && bh[i]; i++)
878 brelse(bh[i]);
879 if (bh != &bhs)
880 kfree(bh);
882 return err;
885 static noinline_for_stack int
886 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
887 struct ext4_buddy *e4b)
889 struct ext4_sb_info *sbi = EXT4_SB(sb);
890 struct inode *inode = sbi->s_buddy_cache;
891 int blocks_per_page;
892 int block;
893 int pnum;
894 int poff;
895 struct page *page;
896 int ret;
898 mb_debug("load group %lu\n", group);
900 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
902 e4b->bd_blkbits = sb->s_blocksize_bits;
903 e4b->bd_info = ext4_get_group_info(sb, group);
904 e4b->bd_sb = sb;
905 e4b->bd_group = group;
906 e4b->bd_buddy_page = NULL;
907 e4b->bd_bitmap_page = NULL;
910 * the buddy cache inode stores the block bitmap
911 * and buddy information in consecutive blocks.
912 * So for each group we need two blocks.
914 block = group * 2;
915 pnum = block / blocks_per_page;
916 poff = block % blocks_per_page;
918 /* we could use find_or_create_page(), but it locks page
919 * what we'd like to avoid in fast path ... */
920 page = find_get_page(inode->i_mapping, pnum);
921 if (page == NULL || !PageUptodate(page)) {
922 if (page)
923 page_cache_release(page);
924 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
925 if (page) {
926 BUG_ON(page->mapping != inode->i_mapping);
927 if (!PageUptodate(page)) {
928 ret = ext4_mb_init_cache(page, NULL);
929 if (ret) {
930 unlock_page(page);
931 goto err;
933 mb_cmp_bitmaps(e4b, page_address(page) +
934 (poff * sb->s_blocksize));
936 unlock_page(page);
939 if (page == NULL || !PageUptodate(page)) {
940 ret = -EIO;
941 goto err;
943 e4b->bd_bitmap_page = page;
944 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
945 mark_page_accessed(page);
947 block++;
948 pnum = block / blocks_per_page;
949 poff = block % blocks_per_page;
951 page = find_get_page(inode->i_mapping, pnum);
952 if (page == NULL || !PageUptodate(page)) {
953 if (page)
954 page_cache_release(page);
955 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
956 if (page) {
957 BUG_ON(page->mapping != inode->i_mapping);
958 if (!PageUptodate(page)) {
959 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
960 if (ret) {
961 unlock_page(page);
962 goto err;
965 unlock_page(page);
968 if (page == NULL || !PageUptodate(page)) {
969 ret = -EIO;
970 goto err;
972 e4b->bd_buddy_page = page;
973 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
974 mark_page_accessed(page);
976 BUG_ON(e4b->bd_bitmap_page == NULL);
977 BUG_ON(e4b->bd_buddy_page == NULL);
979 return 0;
981 err:
982 if (e4b->bd_bitmap_page)
983 page_cache_release(e4b->bd_bitmap_page);
984 if (e4b->bd_buddy_page)
985 page_cache_release(e4b->bd_buddy_page);
986 e4b->bd_buddy = NULL;
987 e4b->bd_bitmap = NULL;
988 return ret;
991 static void ext4_mb_release_desc(struct ext4_buddy *e4b)
993 if (e4b->bd_bitmap_page)
994 page_cache_release(e4b->bd_bitmap_page);
995 if (e4b->bd_buddy_page)
996 page_cache_release(e4b->bd_buddy_page);
1000 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1002 int order = 1;
1003 void *bb;
1005 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1006 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1008 bb = EXT4_MB_BUDDY(e4b);
1009 while (order <= e4b->bd_blkbits + 1) {
1010 block = block >> 1;
1011 if (!mb_test_bit(block, bb)) {
1012 /* this block is part of buddy of order 'order' */
1013 return order;
1015 bb += 1 << (e4b->bd_blkbits - order);
1016 order++;
1018 return 0;
1021 static void mb_clear_bits(spinlock_t *lock, void *bm, int cur, int len)
1023 __u32 *addr;
1025 len = cur + len;
1026 while (cur < len) {
1027 if ((cur & 31) == 0 && (len - cur) >= 32) {
1028 /* fast path: clear whole word at once */
1029 addr = bm + (cur >> 3);
1030 *addr = 0;
1031 cur += 32;
1032 continue;
1034 mb_clear_bit_atomic(lock, cur, bm);
1035 cur++;
1039 static void mb_set_bits(spinlock_t *lock, void *bm, int cur, int len)
1041 __u32 *addr;
1043 len = cur + len;
1044 while (cur < len) {
1045 if ((cur & 31) == 0 && (len - cur) >= 32) {
1046 /* fast path: set whole word at once */
1047 addr = bm + (cur >> 3);
1048 *addr = 0xffffffff;
1049 cur += 32;
1050 continue;
1052 mb_set_bit_atomic(lock, cur, bm);
1053 cur++;
1057 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1058 int first, int count)
1060 int block = 0;
1061 int max = 0;
1062 int order;
1063 void *buddy;
1064 void *buddy2;
1065 struct super_block *sb = e4b->bd_sb;
1067 BUG_ON(first + count > (sb->s_blocksize << 3));
1068 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
1069 mb_check_buddy(e4b);
1070 mb_free_blocks_double(inode, e4b, first, count);
1072 e4b->bd_info->bb_free += count;
1073 if (first < e4b->bd_info->bb_first_free)
1074 e4b->bd_info->bb_first_free = first;
1076 /* let's maintain fragments counter */
1077 if (first != 0)
1078 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1079 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1080 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1081 if (block && max)
1082 e4b->bd_info->bb_fragments--;
1083 else if (!block && !max)
1084 e4b->bd_info->bb_fragments++;
1086 /* let's maintain buddy itself */
1087 while (count-- > 0) {
1088 block = first++;
1089 order = 0;
1091 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1092 ext4_fsblk_t blocknr;
1093 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1094 blocknr += block;
1095 blocknr +=
1096 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
1097 ext4_unlock_group(sb, e4b->bd_group);
1098 ext4_error(sb, __func__, "double-free of inode"
1099 " %lu's block %llu(bit %u in group %lu)\n",
1100 inode ? inode->i_ino : 0, blocknr, block,
1101 e4b->bd_group);
1102 ext4_lock_group(sb, e4b->bd_group);
1104 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1105 e4b->bd_info->bb_counters[order]++;
1107 /* start of the buddy */
1108 buddy = mb_find_buddy(e4b, order, &max);
1110 do {
1111 block &= ~1UL;
1112 if (mb_test_bit(block, buddy) ||
1113 mb_test_bit(block + 1, buddy))
1114 break;
1116 /* both the buddies are free, try to coalesce them */
1117 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1119 if (!buddy2)
1120 break;
1122 if (order > 0) {
1123 /* for special purposes, we don't set
1124 * free bits in bitmap */
1125 mb_set_bit(block, buddy);
1126 mb_set_bit(block + 1, buddy);
1128 e4b->bd_info->bb_counters[order]--;
1129 e4b->bd_info->bb_counters[order]--;
1131 block = block >> 1;
1132 order++;
1133 e4b->bd_info->bb_counters[order]++;
1135 mb_clear_bit(block, buddy2);
1136 buddy = buddy2;
1137 } while (1);
1139 mb_check_buddy(e4b);
1142 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1143 int needed, struct ext4_free_extent *ex)
1145 int next = block;
1146 int max;
1147 int ord;
1148 void *buddy;
1150 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1151 BUG_ON(ex == NULL);
1153 buddy = mb_find_buddy(e4b, order, &max);
1154 BUG_ON(buddy == NULL);
1155 BUG_ON(block >= max);
1156 if (mb_test_bit(block, buddy)) {
1157 ex->fe_len = 0;
1158 ex->fe_start = 0;
1159 ex->fe_group = 0;
1160 return 0;
1163 /* FIXME dorp order completely ? */
1164 if (likely(order == 0)) {
1165 /* find actual order */
1166 order = mb_find_order_for_block(e4b, block);
1167 block = block >> order;
1170 ex->fe_len = 1 << order;
1171 ex->fe_start = block << order;
1172 ex->fe_group = e4b->bd_group;
1174 /* calc difference from given start */
1175 next = next - ex->fe_start;
1176 ex->fe_len -= next;
1177 ex->fe_start += next;
1179 while (needed > ex->fe_len &&
1180 (buddy = mb_find_buddy(e4b, order, &max))) {
1182 if (block + 1 >= max)
1183 break;
1185 next = (block + 1) * (1 << order);
1186 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1187 break;
1189 ord = mb_find_order_for_block(e4b, next);
1191 order = ord;
1192 block = next >> order;
1193 ex->fe_len += 1 << order;
1196 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1197 return ex->fe_len;
1200 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1202 int ord;
1203 int mlen = 0;
1204 int max = 0;
1205 int cur;
1206 int start = ex->fe_start;
1207 int len = ex->fe_len;
1208 unsigned ret = 0;
1209 int len0 = len;
1210 void *buddy;
1212 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1213 BUG_ON(e4b->bd_group != ex->fe_group);
1214 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1215 mb_check_buddy(e4b);
1216 mb_mark_used_double(e4b, start, len);
1218 e4b->bd_info->bb_free -= len;
1219 if (e4b->bd_info->bb_first_free == start)
1220 e4b->bd_info->bb_first_free += len;
1222 /* let's maintain fragments counter */
1223 if (start != 0)
1224 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1225 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1226 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1227 if (mlen && max)
1228 e4b->bd_info->bb_fragments++;
1229 else if (!mlen && !max)
1230 e4b->bd_info->bb_fragments--;
1232 /* let's maintain buddy itself */
1233 while (len) {
1234 ord = mb_find_order_for_block(e4b, start);
1236 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1237 /* the whole chunk may be allocated at once! */
1238 mlen = 1 << ord;
1239 buddy = mb_find_buddy(e4b, ord, &max);
1240 BUG_ON((start >> ord) >= max);
1241 mb_set_bit(start >> ord, buddy);
1242 e4b->bd_info->bb_counters[ord]--;
1243 start += mlen;
1244 len -= mlen;
1245 BUG_ON(len < 0);
1246 continue;
1249 /* store for history */
1250 if (ret == 0)
1251 ret = len | (ord << 16);
1253 /* we have to split large buddy */
1254 BUG_ON(ord <= 0);
1255 buddy = mb_find_buddy(e4b, ord, &max);
1256 mb_set_bit(start >> ord, buddy);
1257 e4b->bd_info->bb_counters[ord]--;
1259 ord--;
1260 cur = (start >> ord) & ~1U;
1261 buddy = mb_find_buddy(e4b, ord, &max);
1262 mb_clear_bit(cur, buddy);
1263 mb_clear_bit(cur + 1, buddy);
1264 e4b->bd_info->bb_counters[ord]++;
1265 e4b->bd_info->bb_counters[ord]++;
1268 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b->bd_sb), ex->fe_group),
1269 EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1270 mb_check_buddy(e4b);
1272 return ret;
1276 * Must be called under group lock!
1278 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1279 struct ext4_buddy *e4b)
1281 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1282 int ret;
1284 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1285 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1287 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1288 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1289 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1291 /* preallocation can change ac_b_ex, thus we store actually
1292 * allocated blocks for history */
1293 ac->ac_f_ex = ac->ac_b_ex;
1295 ac->ac_status = AC_STATUS_FOUND;
1296 ac->ac_tail = ret & 0xffff;
1297 ac->ac_buddy = ret >> 16;
1299 /* XXXXXXX: SUCH A HORRIBLE **CK */
1300 /*FIXME!! Why ? */
1301 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1302 get_page(ac->ac_bitmap_page);
1303 ac->ac_buddy_page = e4b->bd_buddy_page;
1304 get_page(ac->ac_buddy_page);
1306 /* store last allocated for subsequent stream allocation */
1307 if ((ac->ac_flags & EXT4_MB_HINT_DATA)) {
1308 spin_lock(&sbi->s_md_lock);
1309 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1310 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1311 spin_unlock(&sbi->s_md_lock);
1316 * regular allocator, for general purposes allocation
1319 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1320 struct ext4_buddy *e4b,
1321 int finish_group)
1323 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1324 struct ext4_free_extent *bex = &ac->ac_b_ex;
1325 struct ext4_free_extent *gex = &ac->ac_g_ex;
1326 struct ext4_free_extent ex;
1327 int max;
1330 * We don't want to scan for a whole year
1332 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1333 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1334 ac->ac_status = AC_STATUS_BREAK;
1335 return;
1339 * Haven't found good chunk so far, let's continue
1341 if (bex->fe_len < gex->fe_len)
1342 return;
1344 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1345 && bex->fe_group == e4b->bd_group) {
1346 /* recheck chunk's availability - we don't know
1347 * when it was found (within this lock-unlock
1348 * period or not) */
1349 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1350 if (max >= gex->fe_len) {
1351 ext4_mb_use_best_found(ac, e4b);
1352 return;
1358 * The routine checks whether found extent is good enough. If it is,
1359 * then the extent gets marked used and flag is set to the context
1360 * to stop scanning. Otherwise, the extent is compared with the
1361 * previous found extent and if new one is better, then it's stored
1362 * in the context. Later, the best found extent will be used, if
1363 * mballoc can't find good enough extent.
1365 * FIXME: real allocation policy is to be designed yet!
1367 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1368 struct ext4_free_extent *ex,
1369 struct ext4_buddy *e4b)
1371 struct ext4_free_extent *bex = &ac->ac_b_ex;
1372 struct ext4_free_extent *gex = &ac->ac_g_ex;
1374 BUG_ON(ex->fe_len <= 0);
1375 BUG_ON(ex->fe_len >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1376 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1377 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1379 ac->ac_found++;
1382 * The special case - take what you catch first
1384 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1385 *bex = *ex;
1386 ext4_mb_use_best_found(ac, e4b);
1387 return;
1391 * Let's check whether the chuck is good enough
1393 if (ex->fe_len == gex->fe_len) {
1394 *bex = *ex;
1395 ext4_mb_use_best_found(ac, e4b);
1396 return;
1400 * If this is first found extent, just store it in the context
1402 if (bex->fe_len == 0) {
1403 *bex = *ex;
1404 return;
1408 * If new found extent is better, store it in the context
1410 if (bex->fe_len < gex->fe_len) {
1411 /* if the request isn't satisfied, any found extent
1412 * larger than previous best one is better */
1413 if (ex->fe_len > bex->fe_len)
1414 *bex = *ex;
1415 } else if (ex->fe_len > gex->fe_len) {
1416 /* if the request is satisfied, then we try to find
1417 * an extent that still satisfy the request, but is
1418 * smaller than previous one */
1419 if (ex->fe_len < bex->fe_len)
1420 *bex = *ex;
1423 ext4_mb_check_limits(ac, e4b, 0);
1426 static int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1427 struct ext4_buddy *e4b)
1429 struct ext4_free_extent ex = ac->ac_b_ex;
1430 ext4_group_t group = ex.fe_group;
1431 int max;
1432 int err;
1434 BUG_ON(ex.fe_len <= 0);
1435 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1436 if (err)
1437 return err;
1439 ext4_lock_group(ac->ac_sb, group);
1440 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1442 if (max > 0) {
1443 ac->ac_b_ex = ex;
1444 ext4_mb_use_best_found(ac, e4b);
1447 ext4_unlock_group(ac->ac_sb, group);
1448 ext4_mb_release_desc(e4b);
1450 return 0;
1453 static int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1454 struct ext4_buddy *e4b)
1456 ext4_group_t group = ac->ac_g_ex.fe_group;
1457 int max;
1458 int err;
1459 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1460 struct ext4_super_block *es = sbi->s_es;
1461 struct ext4_free_extent ex;
1463 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1464 return 0;
1466 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1467 if (err)
1468 return err;
1470 ext4_lock_group(ac->ac_sb, group);
1471 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1472 ac->ac_g_ex.fe_len, &ex);
1474 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1475 ext4_fsblk_t start;
1477 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1478 ex.fe_start + le32_to_cpu(es->s_first_data_block);
1479 /* use do_div to get remainder (would be 64-bit modulo) */
1480 if (do_div(start, sbi->s_stripe) == 0) {
1481 ac->ac_found++;
1482 ac->ac_b_ex = ex;
1483 ext4_mb_use_best_found(ac, e4b);
1485 } else if (max >= ac->ac_g_ex.fe_len) {
1486 BUG_ON(ex.fe_len <= 0);
1487 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1488 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1489 ac->ac_found++;
1490 ac->ac_b_ex = ex;
1491 ext4_mb_use_best_found(ac, e4b);
1492 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1493 /* Sometimes, caller may want to merge even small
1494 * number of blocks to an existing extent */
1495 BUG_ON(ex.fe_len <= 0);
1496 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1497 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1498 ac->ac_found++;
1499 ac->ac_b_ex = ex;
1500 ext4_mb_use_best_found(ac, e4b);
1502 ext4_unlock_group(ac->ac_sb, group);
1503 ext4_mb_release_desc(e4b);
1505 return 0;
1509 * The routine scans buddy structures (not bitmap!) from given order
1510 * to max order and tries to find big enough chunk to satisfy the req
1512 static void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1513 struct ext4_buddy *e4b)
1515 struct super_block *sb = ac->ac_sb;
1516 struct ext4_group_info *grp = e4b->bd_info;
1517 void *buddy;
1518 int i;
1519 int k;
1520 int max;
1522 BUG_ON(ac->ac_2order <= 0);
1523 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1524 if (grp->bb_counters[i] == 0)
1525 continue;
1527 buddy = mb_find_buddy(e4b, i, &max);
1528 BUG_ON(buddy == NULL);
1530 k = mb_find_next_zero_bit(buddy, max, 0);
1531 BUG_ON(k >= max);
1533 ac->ac_found++;
1535 ac->ac_b_ex.fe_len = 1 << i;
1536 ac->ac_b_ex.fe_start = k << i;
1537 ac->ac_b_ex.fe_group = e4b->bd_group;
1539 ext4_mb_use_best_found(ac, e4b);
1541 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1543 if (EXT4_SB(sb)->s_mb_stats)
1544 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1546 break;
1551 * The routine scans the group and measures all found extents.
1552 * In order to optimize scanning, caller must pass number of
1553 * free blocks in the group, so the routine can know upper limit.
1555 static void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1556 struct ext4_buddy *e4b)
1558 struct super_block *sb = ac->ac_sb;
1559 void *bitmap = EXT4_MB_BITMAP(e4b);
1560 struct ext4_free_extent ex;
1561 int i;
1562 int free;
1564 free = e4b->bd_info->bb_free;
1565 BUG_ON(free <= 0);
1567 i = e4b->bd_info->bb_first_free;
1569 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1570 i = mb_find_next_zero_bit(bitmap,
1571 EXT4_BLOCKS_PER_GROUP(sb), i);
1572 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1574 * IF we have corrupt bitmap, we won't find any
1575 * free blocks even though group info says we
1576 * we have free blocks
1578 ext4_error(sb, __func__, "%d free blocks as per "
1579 "group info. But bitmap says 0\n",
1580 free);
1581 break;
1584 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1585 BUG_ON(ex.fe_len <= 0);
1586 if (free < ex.fe_len) {
1587 ext4_error(sb, __func__, "%d free blocks as per "
1588 "group info. But got %d blocks\n",
1589 free, ex.fe_len);
1591 * The number of free blocks differs. This mostly
1592 * indicate that the bitmap is corrupt. So exit
1593 * without claiming the space.
1595 break;
1598 ext4_mb_measure_extent(ac, &ex, e4b);
1600 i += ex.fe_len;
1601 free -= ex.fe_len;
1604 ext4_mb_check_limits(ac, e4b, 1);
1608 * This is a special case for storages like raid5
1609 * we try to find stripe-aligned chunks for stripe-size requests
1610 * XXX should do so at least for multiples of stripe size as well
1612 static void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1613 struct ext4_buddy *e4b)
1615 struct super_block *sb = ac->ac_sb;
1616 struct ext4_sb_info *sbi = EXT4_SB(sb);
1617 void *bitmap = EXT4_MB_BITMAP(e4b);
1618 struct ext4_free_extent ex;
1619 ext4_fsblk_t first_group_block;
1620 ext4_fsblk_t a;
1621 ext4_grpblk_t i;
1622 int max;
1624 BUG_ON(sbi->s_stripe == 0);
1626 /* find first stripe-aligned block in group */
1627 first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1628 + le32_to_cpu(sbi->s_es->s_first_data_block);
1629 a = first_group_block + sbi->s_stripe - 1;
1630 do_div(a, sbi->s_stripe);
1631 i = (a * sbi->s_stripe) - first_group_block;
1633 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1634 if (!mb_test_bit(i, bitmap)) {
1635 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1636 if (max >= sbi->s_stripe) {
1637 ac->ac_found++;
1638 ac->ac_b_ex = ex;
1639 ext4_mb_use_best_found(ac, e4b);
1640 break;
1643 i += sbi->s_stripe;
1647 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1648 ext4_group_t group, int cr)
1650 unsigned free, fragments;
1651 unsigned i, bits;
1652 struct ext4_group_desc *desc;
1653 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1655 BUG_ON(cr < 0 || cr >= 4);
1656 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1658 free = grp->bb_free;
1659 fragments = grp->bb_fragments;
1660 if (free == 0)
1661 return 0;
1662 if (fragments == 0)
1663 return 0;
1665 switch (cr) {
1666 case 0:
1667 BUG_ON(ac->ac_2order == 0);
1668 /* If this group is uninitialized, skip it initially */
1669 desc = ext4_get_group_desc(ac->ac_sb, group, NULL);
1670 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
1671 return 0;
1673 bits = ac->ac_sb->s_blocksize_bits + 1;
1674 for (i = ac->ac_2order; i <= bits; i++)
1675 if (grp->bb_counters[i] > 0)
1676 return 1;
1677 break;
1678 case 1:
1679 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1680 return 1;
1681 break;
1682 case 2:
1683 if (free >= ac->ac_g_ex.fe_len)
1684 return 1;
1685 break;
1686 case 3:
1687 return 1;
1688 default:
1689 BUG();
1692 return 0;
1695 static noinline_for_stack int
1696 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1698 ext4_group_t group;
1699 ext4_group_t i;
1700 int cr;
1701 int err = 0;
1702 int bsbits;
1703 struct ext4_sb_info *sbi;
1704 struct super_block *sb;
1705 struct ext4_buddy e4b;
1706 loff_t size, isize;
1708 sb = ac->ac_sb;
1709 sbi = EXT4_SB(sb);
1710 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1712 /* first, try the goal */
1713 err = ext4_mb_find_by_goal(ac, &e4b);
1714 if (err || ac->ac_status == AC_STATUS_FOUND)
1715 goto out;
1717 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1718 goto out;
1721 * ac->ac2_order is set only if the fe_len is a power of 2
1722 * if ac2_order is set we also set criteria to 0 so that we
1723 * try exact allocation using buddy.
1725 i = fls(ac->ac_g_ex.fe_len);
1726 ac->ac_2order = 0;
1728 * We search using buddy data only if the order of the request
1729 * is greater than equal to the sbi_s_mb_order2_reqs
1730 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1732 if (i >= sbi->s_mb_order2_reqs) {
1734 * This should tell if fe_len is exactly power of 2
1736 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1737 ac->ac_2order = i - 1;
1740 bsbits = ac->ac_sb->s_blocksize_bits;
1741 /* if stream allocation is enabled, use global goal */
1742 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
1743 isize = i_size_read(ac->ac_inode) >> bsbits;
1744 if (size < isize)
1745 size = isize;
1747 if (size < sbi->s_mb_stream_request &&
1748 (ac->ac_flags & EXT4_MB_HINT_DATA)) {
1749 /* TBD: may be hot point */
1750 spin_lock(&sbi->s_md_lock);
1751 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1752 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1753 spin_unlock(&sbi->s_md_lock);
1755 /* Let's just scan groups to find more-less suitable blocks */
1756 cr = ac->ac_2order ? 0 : 1;
1758 * cr == 0 try to get exact allocation,
1759 * cr == 3 try to get anything
1761 repeat:
1762 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1763 ac->ac_criteria = cr;
1765 * searching for the right group start
1766 * from the goal value specified
1768 group = ac->ac_g_ex.fe_group;
1770 for (i = 0; i < EXT4_SB(sb)->s_groups_count; group++, i++) {
1771 struct ext4_group_info *grp;
1772 struct ext4_group_desc *desc;
1774 if (group == EXT4_SB(sb)->s_groups_count)
1775 group = 0;
1777 /* quick check to skip empty groups */
1778 grp = ext4_get_group_info(ac->ac_sb, group);
1779 if (grp->bb_free == 0)
1780 continue;
1783 * if the group is already init we check whether it is
1784 * a good group and if not we don't load the buddy
1786 if (EXT4_MB_GRP_NEED_INIT(grp)) {
1788 * we need full data about the group
1789 * to make a good selection
1791 err = ext4_mb_load_buddy(sb, group, &e4b);
1792 if (err)
1793 goto out;
1794 ext4_mb_release_desc(&e4b);
1798 * If the particular group doesn't satisfy our
1799 * criteria we continue with the next group
1801 if (!ext4_mb_good_group(ac, group, cr))
1802 continue;
1804 err = ext4_mb_load_buddy(sb, group, &e4b);
1805 if (err)
1806 goto out;
1808 ext4_lock_group(sb, group);
1809 if (!ext4_mb_good_group(ac, group, cr)) {
1810 /* someone did allocation from this group */
1811 ext4_unlock_group(sb, group);
1812 ext4_mb_release_desc(&e4b);
1813 continue;
1816 ac->ac_groups_scanned++;
1817 desc = ext4_get_group_desc(sb, group, NULL);
1818 if (cr == 0 || (desc->bg_flags &
1819 cpu_to_le16(EXT4_BG_BLOCK_UNINIT) &&
1820 ac->ac_2order != 0))
1821 ext4_mb_simple_scan_group(ac, &e4b);
1822 else if (cr == 1 &&
1823 ac->ac_g_ex.fe_len == sbi->s_stripe)
1824 ext4_mb_scan_aligned(ac, &e4b);
1825 else
1826 ext4_mb_complex_scan_group(ac, &e4b);
1828 ext4_unlock_group(sb, group);
1829 ext4_mb_release_desc(&e4b);
1831 if (ac->ac_status != AC_STATUS_CONTINUE)
1832 break;
1836 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
1837 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1839 * We've been searching too long. Let's try to allocate
1840 * the best chunk we've found so far
1843 ext4_mb_try_best_found(ac, &e4b);
1844 if (ac->ac_status != AC_STATUS_FOUND) {
1846 * Someone more lucky has already allocated it.
1847 * The only thing we can do is just take first
1848 * found block(s)
1849 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
1851 ac->ac_b_ex.fe_group = 0;
1852 ac->ac_b_ex.fe_start = 0;
1853 ac->ac_b_ex.fe_len = 0;
1854 ac->ac_status = AC_STATUS_CONTINUE;
1855 ac->ac_flags |= EXT4_MB_HINT_FIRST;
1856 cr = 3;
1857 atomic_inc(&sbi->s_mb_lost_chunks);
1858 goto repeat;
1861 out:
1862 return err;
1865 #ifdef EXT4_MB_HISTORY
1866 struct ext4_mb_proc_session {
1867 struct ext4_mb_history *history;
1868 struct super_block *sb;
1869 int start;
1870 int max;
1873 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
1874 struct ext4_mb_history *hs,
1875 int first)
1877 if (hs == s->history + s->max)
1878 hs = s->history;
1879 if (!first && hs == s->history + s->start)
1880 return NULL;
1881 while (hs->orig.fe_len == 0) {
1882 hs++;
1883 if (hs == s->history + s->max)
1884 hs = s->history;
1885 if (hs == s->history + s->start)
1886 return NULL;
1888 return hs;
1891 static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
1893 struct ext4_mb_proc_session *s = seq->private;
1894 struct ext4_mb_history *hs;
1895 int l = *pos;
1897 if (l == 0)
1898 return SEQ_START_TOKEN;
1899 hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
1900 if (!hs)
1901 return NULL;
1902 while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
1903 return hs;
1906 static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
1907 loff_t *pos)
1909 struct ext4_mb_proc_session *s = seq->private;
1910 struct ext4_mb_history *hs = v;
1912 ++*pos;
1913 if (v == SEQ_START_TOKEN)
1914 return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
1915 else
1916 return ext4_mb_history_skip_empty(s, ++hs, 0);
1919 static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
1921 char buf[25], buf2[25], buf3[25], *fmt;
1922 struct ext4_mb_history *hs = v;
1924 if (v == SEQ_START_TOKEN) {
1925 seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
1926 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
1927 "pid", "inode", "original", "goal", "result", "found",
1928 "grps", "cr", "flags", "merge", "tail", "broken");
1929 return 0;
1932 if (hs->op == EXT4_MB_HISTORY_ALLOC) {
1933 fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
1934 "%-5u %-5s %-5u %-6u\n";
1935 sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
1936 hs->result.fe_start, hs->result.fe_len,
1937 hs->result.fe_logical);
1938 sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
1939 hs->orig.fe_start, hs->orig.fe_len,
1940 hs->orig.fe_logical);
1941 sprintf(buf3, "%lu/%d/%u@%u", hs->goal.fe_group,
1942 hs->goal.fe_start, hs->goal.fe_len,
1943 hs->goal.fe_logical);
1944 seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
1945 hs->found, hs->groups, hs->cr, hs->flags,
1946 hs->merged ? "M" : "", hs->tail,
1947 hs->buddy ? 1 << hs->buddy : 0);
1948 } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
1949 fmt = "%-5u %-8u %-23s %-23s %-23s\n";
1950 sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
1951 hs->result.fe_start, hs->result.fe_len,
1952 hs->result.fe_logical);
1953 sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
1954 hs->orig.fe_start, hs->orig.fe_len,
1955 hs->orig.fe_logical);
1956 seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
1957 } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
1958 sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
1959 hs->result.fe_start, hs->result.fe_len);
1960 seq_printf(seq, "%-5u %-8u %-23s discard\n",
1961 hs->pid, hs->ino, buf2);
1962 } else if (hs->op == EXT4_MB_HISTORY_FREE) {
1963 sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
1964 hs->result.fe_start, hs->result.fe_len);
1965 seq_printf(seq, "%-5u %-8u %-23s free\n",
1966 hs->pid, hs->ino, buf2);
1968 return 0;
1971 static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
1975 static struct seq_operations ext4_mb_seq_history_ops = {
1976 .start = ext4_mb_seq_history_start,
1977 .next = ext4_mb_seq_history_next,
1978 .stop = ext4_mb_seq_history_stop,
1979 .show = ext4_mb_seq_history_show,
1982 static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
1984 struct super_block *sb = PDE(inode)->data;
1985 struct ext4_sb_info *sbi = EXT4_SB(sb);
1986 struct ext4_mb_proc_session *s;
1987 int rc;
1988 int size;
1990 if (unlikely(sbi->s_mb_history == NULL))
1991 return -ENOMEM;
1992 s = kmalloc(sizeof(*s), GFP_KERNEL);
1993 if (s == NULL)
1994 return -ENOMEM;
1995 s->sb = sb;
1996 size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
1997 s->history = kmalloc(size, GFP_KERNEL);
1998 if (s->history == NULL) {
1999 kfree(s);
2000 return -ENOMEM;
2003 spin_lock(&sbi->s_mb_history_lock);
2004 memcpy(s->history, sbi->s_mb_history, size);
2005 s->max = sbi->s_mb_history_max;
2006 s->start = sbi->s_mb_history_cur % s->max;
2007 spin_unlock(&sbi->s_mb_history_lock);
2009 rc = seq_open(file, &ext4_mb_seq_history_ops);
2010 if (rc == 0) {
2011 struct seq_file *m = (struct seq_file *)file->private_data;
2012 m->private = s;
2013 } else {
2014 kfree(s->history);
2015 kfree(s);
2017 return rc;
2021 static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
2023 struct seq_file *seq = (struct seq_file *)file->private_data;
2024 struct ext4_mb_proc_session *s = seq->private;
2025 kfree(s->history);
2026 kfree(s);
2027 return seq_release(inode, file);
2030 static ssize_t ext4_mb_seq_history_write(struct file *file,
2031 const char __user *buffer,
2032 size_t count, loff_t *ppos)
2034 struct seq_file *seq = (struct seq_file *)file->private_data;
2035 struct ext4_mb_proc_session *s = seq->private;
2036 struct super_block *sb = s->sb;
2037 char str[32];
2038 int value;
2040 if (count >= sizeof(str)) {
2041 printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
2042 "mb_history", (int)sizeof(str));
2043 return -EOVERFLOW;
2046 if (copy_from_user(str, buffer, count))
2047 return -EFAULT;
2049 value = simple_strtol(str, NULL, 0);
2050 if (value < 0)
2051 return -ERANGE;
2052 EXT4_SB(sb)->s_mb_history_filter = value;
2054 return count;
2057 static struct file_operations ext4_mb_seq_history_fops = {
2058 .owner = THIS_MODULE,
2059 .open = ext4_mb_seq_history_open,
2060 .read = seq_read,
2061 .write = ext4_mb_seq_history_write,
2062 .llseek = seq_lseek,
2063 .release = ext4_mb_seq_history_release,
2066 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2068 struct super_block *sb = seq->private;
2069 struct ext4_sb_info *sbi = EXT4_SB(sb);
2070 ext4_group_t group;
2072 if (*pos < 0 || *pos >= sbi->s_groups_count)
2073 return NULL;
2075 group = *pos + 1;
2076 return (void *) group;
2079 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2081 struct super_block *sb = seq->private;
2082 struct ext4_sb_info *sbi = EXT4_SB(sb);
2083 ext4_group_t group;
2085 ++*pos;
2086 if (*pos < 0 || *pos >= sbi->s_groups_count)
2087 return NULL;
2088 group = *pos + 1;
2089 return (void *) group;;
2092 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2094 struct super_block *sb = seq->private;
2095 long group = (long) v;
2096 int i;
2097 int err;
2098 struct ext4_buddy e4b;
2099 struct sg {
2100 struct ext4_group_info info;
2101 unsigned short counters[16];
2102 } sg;
2104 group--;
2105 if (group == 0)
2106 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2107 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2108 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2109 "group", "free", "frags", "first",
2110 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2111 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2113 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2114 sizeof(struct ext4_group_info);
2115 err = ext4_mb_load_buddy(sb, group, &e4b);
2116 if (err) {
2117 seq_printf(seq, "#%-5lu: I/O error\n", group);
2118 return 0;
2120 ext4_lock_group(sb, group);
2121 memcpy(&sg, ext4_get_group_info(sb, group), i);
2122 ext4_unlock_group(sb, group);
2123 ext4_mb_release_desc(&e4b);
2125 seq_printf(seq, "#%-5lu: %-5u %-5u %-5u [", group, sg.info.bb_free,
2126 sg.info.bb_fragments, sg.info.bb_first_free);
2127 for (i = 0; i <= 13; i++)
2128 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2129 sg.info.bb_counters[i] : 0);
2130 seq_printf(seq, " ]\n");
2132 return 0;
2135 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2139 static struct seq_operations ext4_mb_seq_groups_ops = {
2140 .start = ext4_mb_seq_groups_start,
2141 .next = ext4_mb_seq_groups_next,
2142 .stop = ext4_mb_seq_groups_stop,
2143 .show = ext4_mb_seq_groups_show,
2146 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2148 struct super_block *sb = PDE(inode)->data;
2149 int rc;
2151 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2152 if (rc == 0) {
2153 struct seq_file *m = (struct seq_file *)file->private_data;
2154 m->private = sb;
2156 return rc;
2160 static struct file_operations ext4_mb_seq_groups_fops = {
2161 .owner = THIS_MODULE,
2162 .open = ext4_mb_seq_groups_open,
2163 .read = seq_read,
2164 .llseek = seq_lseek,
2165 .release = seq_release,
2168 static void ext4_mb_history_release(struct super_block *sb)
2170 struct ext4_sb_info *sbi = EXT4_SB(sb);
2172 remove_proc_entry("mb_groups", sbi->s_mb_proc);
2173 remove_proc_entry("mb_history", sbi->s_mb_proc);
2175 kfree(sbi->s_mb_history);
2178 static void ext4_mb_history_init(struct super_block *sb)
2180 struct ext4_sb_info *sbi = EXT4_SB(sb);
2181 int i;
2183 if (sbi->s_mb_proc != NULL) {
2184 proc_create_data("mb_history", S_IRUGO, sbi->s_mb_proc,
2185 &ext4_mb_seq_history_fops, sb);
2186 proc_create_data("mb_groups", S_IRUGO, sbi->s_mb_proc,
2187 &ext4_mb_seq_groups_fops, sb);
2190 sbi->s_mb_history_max = 1000;
2191 sbi->s_mb_history_cur = 0;
2192 spin_lock_init(&sbi->s_mb_history_lock);
2193 i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
2194 sbi->s_mb_history = kzalloc(i, GFP_KERNEL);
2195 /* if we can't allocate history, then we simple won't use it */
2198 static noinline_for_stack void
2199 ext4_mb_store_history(struct ext4_allocation_context *ac)
2201 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2202 struct ext4_mb_history h;
2204 if (unlikely(sbi->s_mb_history == NULL))
2205 return;
2207 if (!(ac->ac_op & sbi->s_mb_history_filter))
2208 return;
2210 h.op = ac->ac_op;
2211 h.pid = current->pid;
2212 h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2213 h.orig = ac->ac_o_ex;
2214 h.result = ac->ac_b_ex;
2215 h.flags = ac->ac_flags;
2216 h.found = ac->ac_found;
2217 h.groups = ac->ac_groups_scanned;
2218 h.cr = ac->ac_criteria;
2219 h.tail = ac->ac_tail;
2220 h.buddy = ac->ac_buddy;
2221 h.merged = 0;
2222 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
2223 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2224 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2225 h.merged = 1;
2226 h.goal = ac->ac_g_ex;
2227 h.result = ac->ac_f_ex;
2230 spin_lock(&sbi->s_mb_history_lock);
2231 memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2232 if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2233 sbi->s_mb_history_cur = 0;
2234 spin_unlock(&sbi->s_mb_history_lock);
2237 #else
2238 #define ext4_mb_history_release(sb)
2239 #define ext4_mb_history_init(sb)
2240 #endif
2243 /* Create and initialize ext4_group_info data for the given group. */
2244 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2245 struct ext4_group_desc *desc)
2247 int i, len;
2248 int metalen = 0;
2249 struct ext4_sb_info *sbi = EXT4_SB(sb);
2250 struct ext4_group_info **meta_group_info;
2253 * First check if this group is the first of a reserved block.
2254 * If it's true, we have to allocate a new table of pointers
2255 * to ext4_group_info structures
2257 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2258 metalen = sizeof(*meta_group_info) <<
2259 EXT4_DESC_PER_BLOCK_BITS(sb);
2260 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2261 if (meta_group_info == NULL) {
2262 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2263 "buddy group\n");
2264 goto exit_meta_group_info;
2266 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2267 meta_group_info;
2271 * calculate needed size. if change bb_counters size,
2272 * don't forget about ext4_mb_generate_buddy()
2274 len = offsetof(typeof(**meta_group_info),
2275 bb_counters[sb->s_blocksize_bits + 2]);
2277 meta_group_info =
2278 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2279 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2281 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2282 if (meta_group_info[i] == NULL) {
2283 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2284 goto exit_group_info;
2286 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2287 &(meta_group_info[i]->bb_state));
2290 * initialize bb_free to be able to skip
2291 * empty groups without initialization
2293 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2294 meta_group_info[i]->bb_free =
2295 ext4_free_blocks_after_init(sb, group, desc);
2296 } else {
2297 meta_group_info[i]->bb_free =
2298 le16_to_cpu(desc->bg_free_blocks_count);
2301 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2303 #ifdef DOUBLE_CHECK
2305 struct buffer_head *bh;
2306 meta_group_info[i]->bb_bitmap =
2307 kmalloc(sb->s_blocksize, GFP_KERNEL);
2308 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2309 bh = ext4_read_block_bitmap(sb, group);
2310 BUG_ON(bh == NULL);
2311 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2312 sb->s_blocksize);
2313 put_bh(bh);
2315 #endif
2317 return 0;
2319 exit_group_info:
2320 /* If a meta_group_info table has been allocated, release it now */
2321 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2322 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2323 exit_meta_group_info:
2324 return -ENOMEM;
2325 } /* ext4_mb_add_groupinfo */
2328 * Add a group to the existing groups.
2329 * This function is used for online resize
2331 int ext4_mb_add_more_groupinfo(struct super_block *sb, ext4_group_t group,
2332 struct ext4_group_desc *desc)
2334 struct ext4_sb_info *sbi = EXT4_SB(sb);
2335 struct inode *inode = sbi->s_buddy_cache;
2336 int blocks_per_page;
2337 int block;
2338 int pnum;
2339 struct page *page;
2340 int err;
2342 /* Add group based on group descriptor*/
2343 err = ext4_mb_add_groupinfo(sb, group, desc);
2344 if (err)
2345 return err;
2348 * Cache pages containing dynamic mb_alloc datas (buddy and bitmap
2349 * datas) are set not up to date so that they will be re-initilaized
2350 * during the next call to ext4_mb_load_buddy
2353 /* Set buddy page as not up to date */
2354 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
2355 block = group * 2;
2356 pnum = block / blocks_per_page;
2357 page = find_get_page(inode->i_mapping, pnum);
2358 if (page != NULL) {
2359 ClearPageUptodate(page);
2360 page_cache_release(page);
2363 /* Set bitmap page as not up to date */
2364 block++;
2365 pnum = block / blocks_per_page;
2366 page = find_get_page(inode->i_mapping, pnum);
2367 if (page != NULL) {
2368 ClearPageUptodate(page);
2369 page_cache_release(page);
2372 return 0;
2376 * Update an existing group.
2377 * This function is used for online resize
2379 void ext4_mb_update_group_info(struct ext4_group_info *grp, ext4_grpblk_t add)
2381 grp->bb_free += add;
2384 static int ext4_mb_init_backend(struct super_block *sb)
2386 ext4_group_t i;
2387 int metalen;
2388 struct ext4_sb_info *sbi = EXT4_SB(sb);
2389 struct ext4_super_block *es = sbi->s_es;
2390 int num_meta_group_infos;
2391 int num_meta_group_infos_max;
2392 int array_size;
2393 struct ext4_group_info **meta_group_info;
2394 struct ext4_group_desc *desc;
2396 /* This is the number of blocks used by GDT */
2397 num_meta_group_infos = (sbi->s_groups_count + EXT4_DESC_PER_BLOCK(sb) -
2398 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2401 * This is the total number of blocks used by GDT including
2402 * the number of reserved blocks for GDT.
2403 * The s_group_info array is allocated with this value
2404 * to allow a clean online resize without a complex
2405 * manipulation of pointer.
2406 * The drawback is the unused memory when no resize
2407 * occurs but it's very low in terms of pages
2408 * (see comments below)
2409 * Need to handle this properly when META_BG resizing is allowed
2411 num_meta_group_infos_max = num_meta_group_infos +
2412 le16_to_cpu(es->s_reserved_gdt_blocks);
2415 * array_size is the size of s_group_info array. We round it
2416 * to the next power of two because this approximation is done
2417 * internally by kmalloc so we can have some more memory
2418 * for free here (e.g. may be used for META_BG resize).
2420 array_size = 1;
2421 while (array_size < sizeof(*sbi->s_group_info) *
2422 num_meta_group_infos_max)
2423 array_size = array_size << 1;
2424 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2425 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2426 * So a two level scheme suffices for now. */
2427 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2428 if (sbi->s_group_info == NULL) {
2429 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2430 return -ENOMEM;
2432 sbi->s_buddy_cache = new_inode(sb);
2433 if (sbi->s_buddy_cache == NULL) {
2434 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2435 goto err_freesgi;
2437 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2439 metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb);
2440 for (i = 0; i < num_meta_group_infos; i++) {
2441 if ((i + 1) == num_meta_group_infos)
2442 metalen = sizeof(*meta_group_info) *
2443 (sbi->s_groups_count -
2444 (i << EXT4_DESC_PER_BLOCK_BITS(sb)));
2445 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2446 if (meta_group_info == NULL) {
2447 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2448 "buddy group\n");
2449 goto err_freemeta;
2451 sbi->s_group_info[i] = meta_group_info;
2454 for (i = 0; i < sbi->s_groups_count; i++) {
2455 desc = ext4_get_group_desc(sb, i, NULL);
2456 if (desc == NULL) {
2457 printk(KERN_ERR
2458 "EXT4-fs: can't read descriptor %lu\n", i);
2459 goto err_freebuddy;
2461 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2462 goto err_freebuddy;
2465 return 0;
2467 err_freebuddy:
2468 while (i-- > 0)
2469 kfree(ext4_get_group_info(sb, i));
2470 i = num_meta_group_infos;
2471 err_freemeta:
2472 while (i-- > 0)
2473 kfree(sbi->s_group_info[i]);
2474 iput(sbi->s_buddy_cache);
2475 err_freesgi:
2476 kfree(sbi->s_group_info);
2477 return -ENOMEM;
2480 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2482 struct ext4_sb_info *sbi = EXT4_SB(sb);
2483 unsigned i;
2484 unsigned offset;
2485 unsigned max;
2486 int ret;
2488 if (!test_opt(sb, MBALLOC))
2489 return 0;
2491 i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
2493 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2494 if (sbi->s_mb_offsets == NULL) {
2495 clear_opt(sbi->s_mount_opt, MBALLOC);
2496 return -ENOMEM;
2498 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2499 if (sbi->s_mb_maxs == NULL) {
2500 clear_opt(sbi->s_mount_opt, MBALLOC);
2501 kfree(sbi->s_mb_maxs);
2502 return -ENOMEM;
2505 /* order 0 is regular bitmap */
2506 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2507 sbi->s_mb_offsets[0] = 0;
2509 i = 1;
2510 offset = 0;
2511 max = sb->s_blocksize << 2;
2512 do {
2513 sbi->s_mb_offsets[i] = offset;
2514 sbi->s_mb_maxs[i] = max;
2515 offset += 1 << (sb->s_blocksize_bits - i);
2516 max = max >> 1;
2517 i++;
2518 } while (i <= sb->s_blocksize_bits + 1);
2520 /* init file for buddy data */
2521 ret = ext4_mb_init_backend(sb);
2522 if (ret != 0) {
2523 clear_opt(sbi->s_mount_opt, MBALLOC);
2524 kfree(sbi->s_mb_offsets);
2525 kfree(sbi->s_mb_maxs);
2526 return ret;
2529 spin_lock_init(&sbi->s_md_lock);
2530 INIT_LIST_HEAD(&sbi->s_active_transaction);
2531 INIT_LIST_HEAD(&sbi->s_closed_transaction);
2532 INIT_LIST_HEAD(&sbi->s_committed_transaction);
2533 spin_lock_init(&sbi->s_bal_lock);
2535 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2536 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2537 sbi->s_mb_stats = MB_DEFAULT_STATS;
2538 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2539 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2540 sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
2541 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2543 i = sizeof(struct ext4_locality_group) * NR_CPUS;
2544 sbi->s_locality_groups = kmalloc(i, GFP_KERNEL);
2545 if (sbi->s_locality_groups == NULL) {
2546 clear_opt(sbi->s_mount_opt, MBALLOC);
2547 kfree(sbi->s_mb_offsets);
2548 kfree(sbi->s_mb_maxs);
2549 return -ENOMEM;
2551 for (i = 0; i < NR_CPUS; i++) {
2552 struct ext4_locality_group *lg;
2553 lg = &sbi->s_locality_groups[i];
2554 mutex_init(&lg->lg_mutex);
2555 INIT_LIST_HEAD(&lg->lg_prealloc_list);
2556 spin_lock_init(&lg->lg_prealloc_lock);
2559 ext4_mb_init_per_dev_proc(sb);
2560 ext4_mb_history_init(sb);
2562 printk("EXT4-fs: mballoc enabled\n");
2563 return 0;
2566 /* need to called with ext4 group lock (ext4_lock_group) */
2567 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2569 struct ext4_prealloc_space *pa;
2570 struct list_head *cur, *tmp;
2571 int count = 0;
2573 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2574 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2575 list_del(&pa->pa_group_list);
2576 count++;
2577 kfree(pa);
2579 if (count)
2580 mb_debug("mballoc: %u PAs left\n", count);
2584 int ext4_mb_release(struct super_block *sb)
2586 ext4_group_t i;
2587 int num_meta_group_infos;
2588 struct ext4_group_info *grinfo;
2589 struct ext4_sb_info *sbi = EXT4_SB(sb);
2591 if (!test_opt(sb, MBALLOC))
2592 return 0;
2594 /* release freed, non-committed blocks */
2595 spin_lock(&sbi->s_md_lock);
2596 list_splice_init(&sbi->s_closed_transaction,
2597 &sbi->s_committed_transaction);
2598 list_splice_init(&sbi->s_active_transaction,
2599 &sbi->s_committed_transaction);
2600 spin_unlock(&sbi->s_md_lock);
2601 ext4_mb_free_committed_blocks(sb);
2603 if (sbi->s_group_info) {
2604 for (i = 0; i < sbi->s_groups_count; i++) {
2605 grinfo = ext4_get_group_info(sb, i);
2606 #ifdef DOUBLE_CHECK
2607 kfree(grinfo->bb_bitmap);
2608 #endif
2609 ext4_lock_group(sb, i);
2610 ext4_mb_cleanup_pa(grinfo);
2611 ext4_unlock_group(sb, i);
2612 kfree(grinfo);
2614 num_meta_group_infos = (sbi->s_groups_count +
2615 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2616 EXT4_DESC_PER_BLOCK_BITS(sb);
2617 for (i = 0; i < num_meta_group_infos; i++)
2618 kfree(sbi->s_group_info[i]);
2619 kfree(sbi->s_group_info);
2621 kfree(sbi->s_mb_offsets);
2622 kfree(sbi->s_mb_maxs);
2623 if (sbi->s_buddy_cache)
2624 iput(sbi->s_buddy_cache);
2625 if (sbi->s_mb_stats) {
2626 printk(KERN_INFO
2627 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2628 atomic_read(&sbi->s_bal_allocated),
2629 atomic_read(&sbi->s_bal_reqs),
2630 atomic_read(&sbi->s_bal_success));
2631 printk(KERN_INFO
2632 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2633 "%u 2^N hits, %u breaks, %u lost\n",
2634 atomic_read(&sbi->s_bal_ex_scanned),
2635 atomic_read(&sbi->s_bal_goals),
2636 atomic_read(&sbi->s_bal_2orders),
2637 atomic_read(&sbi->s_bal_breaks),
2638 atomic_read(&sbi->s_mb_lost_chunks));
2639 printk(KERN_INFO
2640 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2641 sbi->s_mb_buddies_generated++,
2642 sbi->s_mb_generation_time);
2643 printk(KERN_INFO
2644 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2645 atomic_read(&sbi->s_mb_preallocated),
2646 atomic_read(&sbi->s_mb_discarded));
2649 kfree(sbi->s_locality_groups);
2651 ext4_mb_history_release(sb);
2652 ext4_mb_destroy_per_dev_proc(sb);
2654 return 0;
2657 static noinline_for_stack void
2658 ext4_mb_free_committed_blocks(struct super_block *sb)
2660 struct ext4_sb_info *sbi = EXT4_SB(sb);
2661 int err;
2662 int i;
2663 int count = 0;
2664 int count2 = 0;
2665 struct ext4_free_metadata *md;
2666 struct ext4_buddy e4b;
2668 if (list_empty(&sbi->s_committed_transaction))
2669 return;
2671 /* there is committed blocks to be freed yet */
2672 do {
2673 /* get next array of blocks */
2674 md = NULL;
2675 spin_lock(&sbi->s_md_lock);
2676 if (!list_empty(&sbi->s_committed_transaction)) {
2677 md = list_entry(sbi->s_committed_transaction.next,
2678 struct ext4_free_metadata, list);
2679 list_del(&md->list);
2681 spin_unlock(&sbi->s_md_lock);
2683 if (md == NULL)
2684 break;
2686 mb_debug("gonna free %u blocks in group %lu (0x%p):",
2687 md->num, md->group, md);
2689 err = ext4_mb_load_buddy(sb, md->group, &e4b);
2690 /* we expect to find existing buddy because it's pinned */
2691 BUG_ON(err != 0);
2693 /* there are blocks to put in buddy to make them really free */
2694 count += md->num;
2695 count2++;
2696 ext4_lock_group(sb, md->group);
2697 for (i = 0; i < md->num; i++) {
2698 mb_debug(" %u", md->blocks[i]);
2699 mb_free_blocks(NULL, &e4b, md->blocks[i], 1);
2701 mb_debug("\n");
2702 ext4_unlock_group(sb, md->group);
2704 /* balance refcounts from ext4_mb_free_metadata() */
2705 page_cache_release(e4b.bd_buddy_page);
2706 page_cache_release(e4b.bd_bitmap_page);
2708 kfree(md);
2709 ext4_mb_release_desc(&e4b);
2711 } while (md);
2713 mb_debug("freed %u blocks in %u structures\n", count, count2);
2716 #define EXT4_MB_STATS_NAME "stats"
2717 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2718 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2719 #define EXT4_MB_ORDER2_REQ "order2_req"
2720 #define EXT4_MB_STREAM_REQ "stream_req"
2721 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2725 #define MB_PROC_FOPS(name) \
2726 static int ext4_mb_##name##_proc_show(struct seq_file *m, void *v) \
2728 struct ext4_sb_info *sbi = m->private; \
2730 seq_printf(m, "%ld\n", sbi->s_mb_##name); \
2731 return 0; \
2734 static int ext4_mb_##name##_proc_open(struct inode *inode, struct file *file)\
2736 return single_open(file, ext4_mb_##name##_proc_show, PDE(inode)->data);\
2739 static ssize_t ext4_mb_##name##_proc_write(struct file *file, \
2740 const char __user *buf, size_t cnt, loff_t *ppos) \
2742 struct ext4_sb_info *sbi = PDE(file->f_path.dentry->d_inode)->data;\
2743 char str[32]; \
2744 long value; \
2745 if (cnt >= sizeof(str)) \
2746 return -EINVAL; \
2747 if (copy_from_user(str, buf, cnt)) \
2748 return -EFAULT; \
2749 value = simple_strtol(str, NULL, 0); \
2750 if (value <= 0) \
2751 return -ERANGE; \
2752 sbi->s_mb_##name = value; \
2753 return cnt; \
2756 static const struct file_operations ext4_mb_##name##_proc_fops = { \
2757 .owner = THIS_MODULE, \
2758 .open = ext4_mb_##name##_proc_open, \
2759 .read = seq_read, \
2760 .llseek = seq_lseek, \
2761 .release = single_release, \
2762 .write = ext4_mb_##name##_proc_write, \
2765 MB_PROC_FOPS(stats);
2766 MB_PROC_FOPS(max_to_scan);
2767 MB_PROC_FOPS(min_to_scan);
2768 MB_PROC_FOPS(order2_reqs);
2769 MB_PROC_FOPS(stream_request);
2770 MB_PROC_FOPS(group_prealloc);
2772 #define MB_PROC_HANDLER(name, var) \
2773 do { \
2774 proc = proc_create_data(name, mode, sbi->s_mb_proc, \
2775 &ext4_mb_##var##_proc_fops, sbi); \
2776 if (proc == NULL) { \
2777 printk(KERN_ERR "EXT4-fs: can't to create %s\n", name); \
2778 goto err_out; \
2780 } while (0)
2782 static int ext4_mb_init_per_dev_proc(struct super_block *sb)
2784 mode_t mode = S_IFREG | S_IRUGO | S_IWUSR;
2785 struct ext4_sb_info *sbi = EXT4_SB(sb);
2786 struct proc_dir_entry *proc;
2787 char devname[64];
2789 if (proc_root_ext4 == NULL) {
2790 sbi->s_mb_proc = NULL;
2791 return -EINVAL;
2793 bdevname(sb->s_bdev, devname);
2794 sbi->s_mb_proc = proc_mkdir(devname, proc_root_ext4);
2796 MB_PROC_HANDLER(EXT4_MB_STATS_NAME, stats);
2797 MB_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME, max_to_scan);
2798 MB_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME, min_to_scan);
2799 MB_PROC_HANDLER(EXT4_MB_ORDER2_REQ, order2_reqs);
2800 MB_PROC_HANDLER(EXT4_MB_STREAM_REQ, stream_request);
2801 MB_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC, group_prealloc);
2803 return 0;
2805 err_out:
2806 printk(KERN_ERR "EXT4-fs: Unable to create %s\n", devname);
2807 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_mb_proc);
2808 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_mb_proc);
2809 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_mb_proc);
2810 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
2811 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
2812 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_mb_proc);
2813 remove_proc_entry(devname, proc_root_ext4);
2814 sbi->s_mb_proc = NULL;
2816 return -ENOMEM;
2819 static int ext4_mb_destroy_per_dev_proc(struct super_block *sb)
2821 struct ext4_sb_info *sbi = EXT4_SB(sb);
2822 char devname[64];
2824 if (sbi->s_mb_proc == NULL)
2825 return -EINVAL;
2827 bdevname(sb->s_bdev, devname);
2828 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_mb_proc);
2829 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_mb_proc);
2830 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_mb_proc);
2831 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
2832 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
2833 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_mb_proc);
2834 remove_proc_entry(devname, proc_root_ext4);
2836 return 0;
2839 int __init init_ext4_mballoc(void)
2841 ext4_pspace_cachep =
2842 kmem_cache_create("ext4_prealloc_space",
2843 sizeof(struct ext4_prealloc_space),
2844 0, SLAB_RECLAIM_ACCOUNT, NULL);
2845 if (ext4_pspace_cachep == NULL)
2846 return -ENOMEM;
2848 ext4_ac_cachep =
2849 kmem_cache_create("ext4_alloc_context",
2850 sizeof(struct ext4_allocation_context),
2851 0, SLAB_RECLAIM_ACCOUNT, NULL);
2852 if (ext4_ac_cachep == NULL) {
2853 kmem_cache_destroy(ext4_pspace_cachep);
2854 return -ENOMEM;
2856 #ifdef CONFIG_PROC_FS
2857 proc_root_ext4 = proc_mkdir("fs/ext4", NULL);
2858 if (proc_root_ext4 == NULL)
2859 printk(KERN_ERR "EXT4-fs: Unable to create fs/ext4\n");
2860 #endif
2861 return 0;
2864 void exit_ext4_mballoc(void)
2866 /* XXX: synchronize_rcu(); */
2867 kmem_cache_destroy(ext4_pspace_cachep);
2868 kmem_cache_destroy(ext4_ac_cachep);
2869 #ifdef CONFIG_PROC_FS
2870 remove_proc_entry("fs/ext4", NULL);
2871 #endif
2876 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2877 * Returns 0 if success or error code
2879 static noinline_for_stack int
2880 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2881 handle_t *handle)
2883 struct buffer_head *bitmap_bh = NULL;
2884 struct ext4_super_block *es;
2885 struct ext4_group_desc *gdp;
2886 struct buffer_head *gdp_bh;
2887 struct ext4_sb_info *sbi;
2888 struct super_block *sb;
2889 ext4_fsblk_t block;
2890 int err, len;
2892 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2893 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2895 sb = ac->ac_sb;
2896 sbi = EXT4_SB(sb);
2897 es = sbi->s_es;
2900 err = -EIO;
2901 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2902 if (!bitmap_bh)
2903 goto out_err;
2905 err = ext4_journal_get_write_access(handle, bitmap_bh);
2906 if (err)
2907 goto out_err;
2909 err = -EIO;
2910 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2911 if (!gdp)
2912 goto out_err;
2914 ext4_debug("using block group %lu(%d)\n", ac->ac_b_ex.fe_group,
2915 gdp->bg_free_blocks_count);
2917 err = ext4_journal_get_write_access(handle, gdp_bh);
2918 if (err)
2919 goto out_err;
2921 block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
2922 + ac->ac_b_ex.fe_start
2923 + le32_to_cpu(es->s_first_data_block);
2925 len = ac->ac_b_ex.fe_len;
2926 if (in_range(ext4_block_bitmap(sb, gdp), block, len) ||
2927 in_range(ext4_inode_bitmap(sb, gdp), block, len) ||
2928 in_range(block, ext4_inode_table(sb, gdp),
2929 EXT4_SB(sb)->s_itb_per_group) ||
2930 in_range(block + len - 1, ext4_inode_table(sb, gdp),
2931 EXT4_SB(sb)->s_itb_per_group)) {
2932 ext4_error(sb, __func__,
2933 "Allocating block in system zone - block = %llu",
2934 block);
2935 /* File system mounted not to panic on error
2936 * Fix the bitmap and repeat the block allocation
2937 * We leak some of the blocks here.
2939 mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group),
2940 bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2941 ac->ac_b_ex.fe_len);
2942 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
2943 if (!err)
2944 err = -EAGAIN;
2945 goto out_err;
2947 #ifdef AGGRESSIVE_CHECK
2949 int i;
2950 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2951 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2952 bitmap_bh->b_data));
2955 #endif
2956 mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), bitmap_bh->b_data,
2957 ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
2959 spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
2960 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2961 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2962 gdp->bg_free_blocks_count =
2963 cpu_to_le16(ext4_free_blocks_after_init(sb,
2964 ac->ac_b_ex.fe_group,
2965 gdp));
2967 le16_add_cpu(&gdp->bg_free_blocks_count, -ac->ac_b_ex.fe_len);
2968 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2969 spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
2972 * free blocks account has already be reduced/reserved
2973 * at write_begin() time for delayed allocation
2974 * do not double accounting
2976 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2977 percpu_counter_sub(&sbi->s_freeblocks_counter,
2978 ac->ac_b_ex.fe_len);
2980 if (sbi->s_log_groups_per_flex) {
2981 ext4_group_t flex_group = ext4_flex_group(sbi,
2982 ac->ac_b_ex.fe_group);
2983 spin_lock(sb_bgl_lock(sbi, flex_group));
2984 sbi->s_flex_groups[flex_group].free_blocks -= ac->ac_b_ex.fe_len;
2985 spin_unlock(sb_bgl_lock(sbi, flex_group));
2988 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
2989 if (err)
2990 goto out_err;
2991 err = ext4_journal_dirty_metadata(handle, gdp_bh);
2993 out_err:
2994 sb->s_dirt = 1;
2995 brelse(bitmap_bh);
2996 return err;
3000 * here we normalize request for locality group
3001 * Group request are normalized to s_strip size if we set the same via mount
3002 * option. If not we set it to s_mb_group_prealloc which can be configured via
3003 * /proc/fs/ext4/<partition>/group_prealloc
3005 * XXX: should we try to preallocate more than the group has now?
3007 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3009 struct super_block *sb = ac->ac_sb;
3010 struct ext4_locality_group *lg = ac->ac_lg;
3012 BUG_ON(lg == NULL);
3013 if (EXT4_SB(sb)->s_stripe)
3014 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
3015 else
3016 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3017 mb_debug("#%u: goal %u blocks for locality group\n",
3018 current->pid, ac->ac_g_ex.fe_len);
3022 * Normalization means making request better in terms of
3023 * size and alignment
3025 static noinline_for_stack void
3026 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3027 struct ext4_allocation_request *ar)
3029 int bsbits, max;
3030 ext4_lblk_t end;
3031 loff_t size, orig_size, start_off;
3032 ext4_lblk_t start, orig_start;
3033 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3034 struct ext4_prealloc_space *pa;
3036 /* do normalize only data requests, metadata requests
3037 do not need preallocation */
3038 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3039 return;
3041 /* sometime caller may want exact blocks */
3042 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3043 return;
3045 /* caller may indicate that preallocation isn't
3046 * required (it's a tail, for example) */
3047 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3048 return;
3050 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3051 ext4_mb_normalize_group_request(ac);
3052 return ;
3055 bsbits = ac->ac_sb->s_blocksize_bits;
3057 /* first, let's learn actual file size
3058 * given current request is allocated */
3059 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3060 size = size << bsbits;
3061 if (size < i_size_read(ac->ac_inode))
3062 size = i_size_read(ac->ac_inode);
3064 /* max size of free chunks */
3065 max = 2 << bsbits;
3067 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3068 (req <= (size) || max <= (chunk_size))
3070 /* first, try to predict filesize */
3071 /* XXX: should this table be tunable? */
3072 start_off = 0;
3073 if (size <= 16 * 1024) {
3074 size = 16 * 1024;
3075 } else if (size <= 32 * 1024) {
3076 size = 32 * 1024;
3077 } else if (size <= 64 * 1024) {
3078 size = 64 * 1024;
3079 } else if (size <= 128 * 1024) {
3080 size = 128 * 1024;
3081 } else if (size <= 256 * 1024) {
3082 size = 256 * 1024;
3083 } else if (size <= 512 * 1024) {
3084 size = 512 * 1024;
3085 } else if (size <= 1024 * 1024) {
3086 size = 1024 * 1024;
3087 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3088 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3089 (21 - bsbits)) << 21;
3090 size = 2 * 1024 * 1024;
3091 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3092 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3093 (22 - bsbits)) << 22;
3094 size = 4 * 1024 * 1024;
3095 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3096 (8<<20)>>bsbits, max, 8 * 1024)) {
3097 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3098 (23 - bsbits)) << 23;
3099 size = 8 * 1024 * 1024;
3100 } else {
3101 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3102 size = ac->ac_o_ex.fe_len << bsbits;
3104 orig_size = size = size >> bsbits;
3105 orig_start = start = start_off >> bsbits;
3107 /* don't cover already allocated blocks in selected range */
3108 if (ar->pleft && start <= ar->lleft) {
3109 size -= ar->lleft + 1 - start;
3110 start = ar->lleft + 1;
3112 if (ar->pright && start + size - 1 >= ar->lright)
3113 size -= start + size - ar->lright;
3115 end = start + size;
3117 /* check we don't cross already preallocated blocks */
3118 rcu_read_lock();
3119 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3120 unsigned long pa_end;
3122 if (pa->pa_deleted)
3123 continue;
3124 spin_lock(&pa->pa_lock);
3125 if (pa->pa_deleted) {
3126 spin_unlock(&pa->pa_lock);
3127 continue;
3130 pa_end = pa->pa_lstart + pa->pa_len;
3132 /* PA must not overlap original request */
3133 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3134 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3136 /* skip PA normalized request doesn't overlap with */
3137 if (pa->pa_lstart >= end) {
3138 spin_unlock(&pa->pa_lock);
3139 continue;
3141 if (pa_end <= start) {
3142 spin_unlock(&pa->pa_lock);
3143 continue;
3145 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3147 if (pa_end <= ac->ac_o_ex.fe_logical) {
3148 BUG_ON(pa_end < start);
3149 start = pa_end;
3152 if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3153 BUG_ON(pa->pa_lstart > end);
3154 end = pa->pa_lstart;
3156 spin_unlock(&pa->pa_lock);
3158 rcu_read_unlock();
3159 size = end - start;
3161 /* XXX: extra loop to check we really don't overlap preallocations */
3162 rcu_read_lock();
3163 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3164 unsigned long pa_end;
3165 spin_lock(&pa->pa_lock);
3166 if (pa->pa_deleted == 0) {
3167 pa_end = pa->pa_lstart + pa->pa_len;
3168 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3170 spin_unlock(&pa->pa_lock);
3172 rcu_read_unlock();
3174 if (start + size <= ac->ac_o_ex.fe_logical &&
3175 start > ac->ac_o_ex.fe_logical) {
3176 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3177 (unsigned long) start, (unsigned long) size,
3178 (unsigned long) ac->ac_o_ex.fe_logical);
3180 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3181 start > ac->ac_o_ex.fe_logical);
3182 BUG_ON(size <= 0 || size >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3184 /* now prepare goal request */
3186 /* XXX: is it better to align blocks WRT to logical
3187 * placement or satisfy big request as is */
3188 ac->ac_g_ex.fe_logical = start;
3189 ac->ac_g_ex.fe_len = size;
3191 /* define goal start in order to merge */
3192 if (ar->pright && (ar->lright == (start + size))) {
3193 /* merge to the right */
3194 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3195 &ac->ac_f_ex.fe_group,
3196 &ac->ac_f_ex.fe_start);
3197 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3199 if (ar->pleft && (ar->lleft + 1 == start)) {
3200 /* merge to the left */
3201 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3202 &ac->ac_f_ex.fe_group,
3203 &ac->ac_f_ex.fe_start);
3204 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3207 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
3208 (unsigned) orig_size, (unsigned) start);
3211 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3213 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3215 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3216 atomic_inc(&sbi->s_bal_reqs);
3217 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3218 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3219 atomic_inc(&sbi->s_bal_success);
3220 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3221 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3222 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3223 atomic_inc(&sbi->s_bal_goals);
3224 if (ac->ac_found > sbi->s_mb_max_to_scan)
3225 atomic_inc(&sbi->s_bal_breaks);
3228 ext4_mb_store_history(ac);
3232 * use blocks preallocated to inode
3234 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3235 struct ext4_prealloc_space *pa)
3237 ext4_fsblk_t start;
3238 ext4_fsblk_t end;
3239 int len;
3241 /* found preallocated blocks, use them */
3242 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3243 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3244 len = end - start;
3245 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3246 &ac->ac_b_ex.fe_start);
3247 ac->ac_b_ex.fe_len = len;
3248 ac->ac_status = AC_STATUS_FOUND;
3249 ac->ac_pa = pa;
3251 BUG_ON(start < pa->pa_pstart);
3252 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3253 BUG_ON(pa->pa_free < len);
3254 pa->pa_free -= len;
3256 mb_debug("use %llu/%u from inode pa %p\n", start, len, pa);
3260 * use blocks preallocated to locality group
3262 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3263 struct ext4_prealloc_space *pa)
3265 unsigned int len = ac->ac_o_ex.fe_len;
3266 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3267 &ac->ac_b_ex.fe_group,
3268 &ac->ac_b_ex.fe_start);
3269 ac->ac_b_ex.fe_len = len;
3270 ac->ac_status = AC_STATUS_FOUND;
3271 ac->ac_pa = pa;
3273 /* we don't correct pa_pstart or pa_plen here to avoid
3274 * possible race when the group is being loaded concurrently
3275 * instead we correct pa later, after blocks are marked
3276 * in on-disk bitmap -- see ext4_mb_release_context()
3277 * Other CPUs are prevented from allocating from this pa by lg_mutex
3279 mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3283 * search goal blocks in preallocated space
3285 static noinline_for_stack int
3286 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3288 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3289 struct ext4_locality_group *lg;
3290 struct ext4_prealloc_space *pa;
3292 /* only data can be preallocated */
3293 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3294 return 0;
3296 /* first, try per-file preallocation */
3297 rcu_read_lock();
3298 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3300 /* all fields in this condition don't change,
3301 * so we can skip locking for them */
3302 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3303 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3304 continue;
3306 /* found preallocated blocks, use them */
3307 spin_lock(&pa->pa_lock);
3308 if (pa->pa_deleted == 0 && pa->pa_free) {
3309 atomic_inc(&pa->pa_count);
3310 ext4_mb_use_inode_pa(ac, pa);
3311 spin_unlock(&pa->pa_lock);
3312 ac->ac_criteria = 10;
3313 rcu_read_unlock();
3314 return 1;
3316 spin_unlock(&pa->pa_lock);
3318 rcu_read_unlock();
3320 /* can we use group allocation? */
3321 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3322 return 0;
3324 /* inode may have no locality group for some reason */
3325 lg = ac->ac_lg;
3326 if (lg == NULL)
3327 return 0;
3329 rcu_read_lock();
3330 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list, pa_inode_list) {
3331 spin_lock(&pa->pa_lock);
3332 if (pa->pa_deleted == 0 && pa->pa_free >= ac->ac_o_ex.fe_len) {
3333 atomic_inc(&pa->pa_count);
3334 ext4_mb_use_group_pa(ac, pa);
3335 spin_unlock(&pa->pa_lock);
3336 ac->ac_criteria = 20;
3337 rcu_read_unlock();
3338 return 1;
3340 spin_unlock(&pa->pa_lock);
3342 rcu_read_unlock();
3344 return 0;
3348 * the function goes through all preallocation in this group and marks them
3349 * used in in-core bitmap. buddy must be generated from this bitmap
3350 * Need to be called with ext4 group lock (ext4_lock_group)
3352 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3353 ext4_group_t group)
3355 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3356 struct ext4_prealloc_space *pa;
3357 struct list_head *cur;
3358 ext4_group_t groupnr;
3359 ext4_grpblk_t start;
3360 int preallocated = 0;
3361 int count = 0;
3362 int len;
3364 /* all form of preallocation discards first load group,
3365 * so the only competing code is preallocation use.
3366 * we don't need any locking here
3367 * notice we do NOT ignore preallocations with pa_deleted
3368 * otherwise we could leave used blocks available for
3369 * allocation in buddy when concurrent ext4_mb_put_pa()
3370 * is dropping preallocation
3372 list_for_each(cur, &grp->bb_prealloc_list) {
3373 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3374 spin_lock(&pa->pa_lock);
3375 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3376 &groupnr, &start);
3377 len = pa->pa_len;
3378 spin_unlock(&pa->pa_lock);
3379 if (unlikely(len == 0))
3380 continue;
3381 BUG_ON(groupnr != group);
3382 mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
3383 bitmap, start, len);
3384 preallocated += len;
3385 count++;
3387 mb_debug("prellocated %u for group %lu\n", preallocated, group);
3390 static void ext4_mb_pa_callback(struct rcu_head *head)
3392 struct ext4_prealloc_space *pa;
3393 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3394 kmem_cache_free(ext4_pspace_cachep, pa);
3398 * drops a reference to preallocated space descriptor
3399 * if this was the last reference and the space is consumed
3401 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3402 struct super_block *sb, struct ext4_prealloc_space *pa)
3404 unsigned long grp;
3406 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3407 return;
3409 /* in this short window concurrent discard can set pa_deleted */
3410 spin_lock(&pa->pa_lock);
3411 if (pa->pa_deleted == 1) {
3412 spin_unlock(&pa->pa_lock);
3413 return;
3416 pa->pa_deleted = 1;
3417 spin_unlock(&pa->pa_lock);
3419 /* -1 is to protect from crossing allocation group */
3420 ext4_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL);
3423 * possible race:
3425 * P1 (buddy init) P2 (regular allocation)
3426 * find block B in PA
3427 * copy on-disk bitmap to buddy
3428 * mark B in on-disk bitmap
3429 * drop PA from group
3430 * mark all PAs in buddy
3432 * thus, P1 initializes buddy with B available. to prevent this
3433 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3434 * against that pair
3436 ext4_lock_group(sb, grp);
3437 list_del(&pa->pa_group_list);
3438 ext4_unlock_group(sb, grp);
3440 spin_lock(pa->pa_obj_lock);
3441 list_del_rcu(&pa->pa_inode_list);
3442 spin_unlock(pa->pa_obj_lock);
3444 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3448 * creates new preallocated space for given inode
3450 static noinline_for_stack int
3451 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3453 struct super_block *sb = ac->ac_sb;
3454 struct ext4_prealloc_space *pa;
3455 struct ext4_group_info *grp;
3456 struct ext4_inode_info *ei;
3458 /* preallocate only when found space is larger then requested */
3459 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3460 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3461 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3463 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3464 if (pa == NULL)
3465 return -ENOMEM;
3467 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3468 int winl;
3469 int wins;
3470 int win;
3471 int offs;
3473 /* we can't allocate as much as normalizer wants.
3474 * so, found space must get proper lstart
3475 * to cover original request */
3476 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3477 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3479 /* we're limited by original request in that
3480 * logical block must be covered any way
3481 * winl is window we can move our chunk within */
3482 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3484 /* also, we should cover whole original request */
3485 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3487 /* the smallest one defines real window */
3488 win = min(winl, wins);
3490 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3491 if (offs && offs < win)
3492 win = offs;
3494 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3495 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3496 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3499 /* preallocation can change ac_b_ex, thus we store actually
3500 * allocated blocks for history */
3501 ac->ac_f_ex = ac->ac_b_ex;
3503 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3504 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3505 pa->pa_len = ac->ac_b_ex.fe_len;
3506 pa->pa_free = pa->pa_len;
3507 atomic_set(&pa->pa_count, 1);
3508 spin_lock_init(&pa->pa_lock);
3509 pa->pa_deleted = 0;
3510 pa->pa_linear = 0;
3512 mb_debug("new inode pa %p: %llu/%u for %u\n", pa,
3513 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3515 ext4_mb_use_inode_pa(ac, pa);
3516 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3518 ei = EXT4_I(ac->ac_inode);
3519 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3521 pa->pa_obj_lock = &ei->i_prealloc_lock;
3522 pa->pa_inode = ac->ac_inode;
3524 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3525 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3526 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3528 spin_lock(pa->pa_obj_lock);
3529 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3530 spin_unlock(pa->pa_obj_lock);
3532 return 0;
3536 * creates new preallocated space for locality group inodes belongs to
3538 static noinline_for_stack int
3539 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3541 struct super_block *sb = ac->ac_sb;
3542 struct ext4_locality_group *lg;
3543 struct ext4_prealloc_space *pa;
3544 struct ext4_group_info *grp;
3546 /* preallocate only when found space is larger then requested */
3547 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3548 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3549 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3551 BUG_ON(ext4_pspace_cachep == NULL);
3552 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3553 if (pa == NULL)
3554 return -ENOMEM;
3556 /* preallocation can change ac_b_ex, thus we store actually
3557 * allocated blocks for history */
3558 ac->ac_f_ex = ac->ac_b_ex;
3560 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3561 pa->pa_lstart = pa->pa_pstart;
3562 pa->pa_len = ac->ac_b_ex.fe_len;
3563 pa->pa_free = pa->pa_len;
3564 atomic_set(&pa->pa_count, 1);
3565 spin_lock_init(&pa->pa_lock);
3566 pa->pa_deleted = 0;
3567 pa->pa_linear = 1;
3569 mb_debug("new group pa %p: %llu/%u for %u\n", pa,
3570 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3572 ext4_mb_use_group_pa(ac, pa);
3573 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3575 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3576 lg = ac->ac_lg;
3577 BUG_ON(lg == NULL);
3579 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3580 pa->pa_inode = NULL;
3582 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3583 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3584 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3586 spin_lock(pa->pa_obj_lock);
3587 list_add_tail_rcu(&pa->pa_inode_list, &lg->lg_prealloc_list);
3588 spin_unlock(pa->pa_obj_lock);
3590 return 0;
3593 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3595 int err;
3597 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3598 err = ext4_mb_new_group_pa(ac);
3599 else
3600 err = ext4_mb_new_inode_pa(ac);
3601 return err;
3605 * finds all unused blocks in on-disk bitmap, frees them in
3606 * in-core bitmap and buddy.
3607 * @pa must be unlinked from inode and group lists, so that
3608 * nobody else can find/use it.
3609 * the caller MUST hold group/inode locks.
3610 * TODO: optimize the case when there are no in-core structures yet
3612 static noinline_for_stack int
3613 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3614 struct ext4_prealloc_space *pa,
3615 struct ext4_allocation_context *ac)
3617 struct super_block *sb = e4b->bd_sb;
3618 struct ext4_sb_info *sbi = EXT4_SB(sb);
3619 unsigned long end;
3620 unsigned long next;
3621 ext4_group_t group;
3622 ext4_grpblk_t bit;
3623 sector_t start;
3624 int err = 0;
3625 int free = 0;
3627 BUG_ON(pa->pa_deleted == 0);
3628 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3629 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3630 end = bit + pa->pa_len;
3632 if (ac) {
3633 ac->ac_sb = sb;
3634 ac->ac_inode = pa->pa_inode;
3635 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3638 while (bit < end) {
3639 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3640 if (bit >= end)
3641 break;
3642 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3643 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3644 le32_to_cpu(sbi->s_es->s_first_data_block);
3645 mb_debug(" free preallocated %u/%u in group %u\n",
3646 (unsigned) start, (unsigned) next - bit,
3647 (unsigned) group);
3648 free += next - bit;
3650 if (ac) {
3651 ac->ac_b_ex.fe_group = group;
3652 ac->ac_b_ex.fe_start = bit;
3653 ac->ac_b_ex.fe_len = next - bit;
3654 ac->ac_b_ex.fe_logical = 0;
3655 ext4_mb_store_history(ac);
3658 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3659 bit = next + 1;
3661 if (free != pa->pa_free) {
3662 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3663 pa, (unsigned long) pa->pa_lstart,
3664 (unsigned long) pa->pa_pstart,
3665 (unsigned long) pa->pa_len);
3666 ext4_error(sb, __func__, "free %u, pa_free %u\n",
3667 free, pa->pa_free);
3669 * pa is already deleted so we use the value obtained
3670 * from the bitmap and continue.
3673 atomic_add(free, &sbi->s_mb_discarded);
3675 return err;
3678 static noinline_for_stack int
3679 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3680 struct ext4_prealloc_space *pa,
3681 struct ext4_allocation_context *ac)
3683 struct super_block *sb = e4b->bd_sb;
3684 ext4_group_t group;
3685 ext4_grpblk_t bit;
3687 if (ac)
3688 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3690 BUG_ON(pa->pa_deleted == 0);
3691 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3692 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3693 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3694 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3696 if (ac) {
3697 ac->ac_sb = sb;
3698 ac->ac_inode = NULL;
3699 ac->ac_b_ex.fe_group = group;
3700 ac->ac_b_ex.fe_start = bit;
3701 ac->ac_b_ex.fe_len = pa->pa_len;
3702 ac->ac_b_ex.fe_logical = 0;
3703 ext4_mb_store_history(ac);
3706 return 0;
3710 * releases all preallocations in given group
3712 * first, we need to decide discard policy:
3713 * - when do we discard
3714 * 1) ENOSPC
3715 * - how many do we discard
3716 * 1) how many requested
3718 static noinline_for_stack int
3719 ext4_mb_discard_group_preallocations(struct super_block *sb,
3720 ext4_group_t group, int needed)
3722 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3723 struct buffer_head *bitmap_bh = NULL;
3724 struct ext4_prealloc_space *pa, *tmp;
3725 struct ext4_allocation_context *ac;
3726 struct list_head list;
3727 struct ext4_buddy e4b;
3728 int err;
3729 int busy = 0;
3730 int free = 0;
3732 mb_debug("discard preallocation for group %lu\n", group);
3734 if (list_empty(&grp->bb_prealloc_list))
3735 return 0;
3737 bitmap_bh = ext4_read_block_bitmap(sb, group);
3738 if (bitmap_bh == NULL) {
3739 ext4_error(sb, __func__, "Error in reading block "
3740 "bitmap for %lu\n", group);
3741 return 0;
3744 err = ext4_mb_load_buddy(sb, group, &e4b);
3745 if (err) {
3746 ext4_error(sb, __func__, "Error in loading buddy "
3747 "information for %lu\n", group);
3748 put_bh(bitmap_bh);
3749 return 0;
3752 if (needed == 0)
3753 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3755 INIT_LIST_HEAD(&list);
3756 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3757 repeat:
3758 ext4_lock_group(sb, group);
3759 list_for_each_entry_safe(pa, tmp,
3760 &grp->bb_prealloc_list, pa_group_list) {
3761 spin_lock(&pa->pa_lock);
3762 if (atomic_read(&pa->pa_count)) {
3763 spin_unlock(&pa->pa_lock);
3764 busy = 1;
3765 continue;
3767 if (pa->pa_deleted) {
3768 spin_unlock(&pa->pa_lock);
3769 continue;
3772 /* seems this one can be freed ... */
3773 pa->pa_deleted = 1;
3775 /* we can trust pa_free ... */
3776 free += pa->pa_free;
3778 spin_unlock(&pa->pa_lock);
3780 list_del(&pa->pa_group_list);
3781 list_add(&pa->u.pa_tmp_list, &list);
3784 /* if we still need more blocks and some PAs were used, try again */
3785 if (free < needed && busy) {
3786 busy = 0;
3787 ext4_unlock_group(sb, group);
3789 * Yield the CPU here so that we don't get soft lockup
3790 * in non preempt case.
3792 yield();
3793 goto repeat;
3796 /* found anything to free? */
3797 if (list_empty(&list)) {
3798 BUG_ON(free != 0);
3799 goto out;
3802 /* now free all selected PAs */
3803 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3805 /* remove from object (inode or locality group) */
3806 spin_lock(pa->pa_obj_lock);
3807 list_del_rcu(&pa->pa_inode_list);
3808 spin_unlock(pa->pa_obj_lock);
3810 if (pa->pa_linear)
3811 ext4_mb_release_group_pa(&e4b, pa, ac);
3812 else
3813 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3815 list_del(&pa->u.pa_tmp_list);
3816 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3819 out:
3820 ext4_unlock_group(sb, group);
3821 if (ac)
3822 kmem_cache_free(ext4_ac_cachep, ac);
3823 ext4_mb_release_desc(&e4b);
3824 put_bh(bitmap_bh);
3825 return free;
3829 * releases all non-used preallocated blocks for given inode
3831 * It's important to discard preallocations under i_data_sem
3832 * We don't want another block to be served from the prealloc
3833 * space when we are discarding the inode prealloc space.
3835 * FIXME!! Make sure it is valid at all the call sites
3837 void ext4_mb_discard_inode_preallocations(struct inode *inode)
3839 struct ext4_inode_info *ei = EXT4_I(inode);
3840 struct super_block *sb = inode->i_sb;
3841 struct buffer_head *bitmap_bh = NULL;
3842 struct ext4_prealloc_space *pa, *tmp;
3843 struct ext4_allocation_context *ac;
3844 ext4_group_t group = 0;
3845 struct list_head list;
3846 struct ext4_buddy e4b;
3847 int err;
3849 if (!test_opt(sb, MBALLOC) || !S_ISREG(inode->i_mode)) {
3850 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3851 return;
3854 mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
3856 INIT_LIST_HEAD(&list);
3858 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3859 repeat:
3860 /* first, collect all pa's in the inode */
3861 spin_lock(&ei->i_prealloc_lock);
3862 while (!list_empty(&ei->i_prealloc_list)) {
3863 pa = list_entry(ei->i_prealloc_list.next,
3864 struct ext4_prealloc_space, pa_inode_list);
3865 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3866 spin_lock(&pa->pa_lock);
3867 if (atomic_read(&pa->pa_count)) {
3868 /* this shouldn't happen often - nobody should
3869 * use preallocation while we're discarding it */
3870 spin_unlock(&pa->pa_lock);
3871 spin_unlock(&ei->i_prealloc_lock);
3872 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3873 WARN_ON(1);
3874 schedule_timeout_uninterruptible(HZ);
3875 goto repeat;
3878 if (pa->pa_deleted == 0) {
3879 pa->pa_deleted = 1;
3880 spin_unlock(&pa->pa_lock);
3881 list_del_rcu(&pa->pa_inode_list);
3882 list_add(&pa->u.pa_tmp_list, &list);
3883 continue;
3886 /* someone is deleting pa right now */
3887 spin_unlock(&pa->pa_lock);
3888 spin_unlock(&ei->i_prealloc_lock);
3890 /* we have to wait here because pa_deleted
3891 * doesn't mean pa is already unlinked from
3892 * the list. as we might be called from
3893 * ->clear_inode() the inode will get freed
3894 * and concurrent thread which is unlinking
3895 * pa from inode's list may access already
3896 * freed memory, bad-bad-bad */
3898 /* XXX: if this happens too often, we can
3899 * add a flag to force wait only in case
3900 * of ->clear_inode(), but not in case of
3901 * regular truncate */
3902 schedule_timeout_uninterruptible(HZ);
3903 goto repeat;
3905 spin_unlock(&ei->i_prealloc_lock);
3907 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3908 BUG_ON(pa->pa_linear != 0);
3909 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3911 err = ext4_mb_load_buddy(sb, group, &e4b);
3912 if (err) {
3913 ext4_error(sb, __func__, "Error in loading buddy "
3914 "information for %lu\n", group);
3915 continue;
3918 bitmap_bh = ext4_read_block_bitmap(sb, group);
3919 if (bitmap_bh == NULL) {
3920 ext4_error(sb, __func__, "Error in reading block "
3921 "bitmap for %lu\n", group);
3922 ext4_mb_release_desc(&e4b);
3923 continue;
3926 ext4_lock_group(sb, group);
3927 list_del(&pa->pa_group_list);
3928 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3929 ext4_unlock_group(sb, group);
3931 ext4_mb_release_desc(&e4b);
3932 put_bh(bitmap_bh);
3934 list_del(&pa->u.pa_tmp_list);
3935 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3937 if (ac)
3938 kmem_cache_free(ext4_ac_cachep, ac);
3942 * finds all preallocated spaces and return blocks being freed to them
3943 * if preallocated space becomes full (no block is used from the space)
3944 * then the function frees space in buddy
3945 * XXX: at the moment, truncate (which is the only way to free blocks)
3946 * discards all preallocations
3948 static void ext4_mb_return_to_preallocation(struct inode *inode,
3949 struct ext4_buddy *e4b,
3950 sector_t block, int count)
3952 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
3954 #ifdef MB_DEBUG
3955 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3957 struct super_block *sb = ac->ac_sb;
3958 ext4_group_t i;
3960 printk(KERN_ERR "EXT4-fs: Can't allocate:"
3961 " Allocation context details:\n");
3962 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
3963 ac->ac_status, ac->ac_flags);
3964 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3965 "best %lu/%lu/%lu@%lu cr %d\n",
3966 (unsigned long)ac->ac_o_ex.fe_group,
3967 (unsigned long)ac->ac_o_ex.fe_start,
3968 (unsigned long)ac->ac_o_ex.fe_len,
3969 (unsigned long)ac->ac_o_ex.fe_logical,
3970 (unsigned long)ac->ac_g_ex.fe_group,
3971 (unsigned long)ac->ac_g_ex.fe_start,
3972 (unsigned long)ac->ac_g_ex.fe_len,
3973 (unsigned long)ac->ac_g_ex.fe_logical,
3974 (unsigned long)ac->ac_b_ex.fe_group,
3975 (unsigned long)ac->ac_b_ex.fe_start,
3976 (unsigned long)ac->ac_b_ex.fe_len,
3977 (unsigned long)ac->ac_b_ex.fe_logical,
3978 (int)ac->ac_criteria);
3979 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
3980 ac->ac_found);
3981 printk(KERN_ERR "EXT4-fs: groups: \n");
3982 for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
3983 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3984 struct ext4_prealloc_space *pa;
3985 ext4_grpblk_t start;
3986 struct list_head *cur;
3987 ext4_lock_group(sb, i);
3988 list_for_each(cur, &grp->bb_prealloc_list) {
3989 pa = list_entry(cur, struct ext4_prealloc_space,
3990 pa_group_list);
3991 spin_lock(&pa->pa_lock);
3992 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3993 NULL, &start);
3994 spin_unlock(&pa->pa_lock);
3995 printk(KERN_ERR "PA:%lu:%d:%u \n", i,
3996 start, pa->pa_len);
3998 ext4_unlock_group(sb, i);
4000 if (grp->bb_free == 0)
4001 continue;
4002 printk(KERN_ERR "%lu: %d/%d \n",
4003 i, grp->bb_free, grp->bb_fragments);
4005 printk(KERN_ERR "\n");
4007 #else
4008 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4010 return;
4012 #endif
4015 * We use locality group preallocation for small size file. The size of the
4016 * file is determined by the current size or the resulting size after
4017 * allocation which ever is larger
4019 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
4021 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4023 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4024 int bsbits = ac->ac_sb->s_blocksize_bits;
4025 loff_t size, isize;
4027 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4028 return;
4030 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4031 isize = i_size_read(ac->ac_inode) >> bsbits;
4032 size = max(size, isize);
4034 /* don't use group allocation for large files */
4035 if (size >= sbi->s_mb_stream_request)
4036 return;
4038 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4039 return;
4041 BUG_ON(ac->ac_lg != NULL);
4043 * locality group prealloc space are per cpu. The reason for having
4044 * per cpu locality group is to reduce the contention between block
4045 * request from multiple CPUs.
4047 ac->ac_lg = &sbi->s_locality_groups[get_cpu()];
4048 put_cpu();
4050 /* we're going to use group allocation */
4051 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4053 /* serialize all allocations in the group */
4054 mutex_lock(&ac->ac_lg->lg_mutex);
4057 static noinline_for_stack int
4058 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4059 struct ext4_allocation_request *ar)
4061 struct super_block *sb = ar->inode->i_sb;
4062 struct ext4_sb_info *sbi = EXT4_SB(sb);
4063 struct ext4_super_block *es = sbi->s_es;
4064 ext4_group_t group;
4065 unsigned long len;
4066 unsigned long goal;
4067 ext4_grpblk_t block;
4069 /* we can't allocate > group size */
4070 len = ar->len;
4072 /* just a dirty hack to filter too big requests */
4073 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4074 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4076 /* start searching from the goal */
4077 goal = ar->goal;
4078 if (goal < le32_to_cpu(es->s_first_data_block) ||
4079 goal >= ext4_blocks_count(es))
4080 goal = le32_to_cpu(es->s_first_data_block);
4081 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4083 /* set up allocation goals */
4084 ac->ac_b_ex.fe_logical = ar->logical;
4085 ac->ac_b_ex.fe_group = 0;
4086 ac->ac_b_ex.fe_start = 0;
4087 ac->ac_b_ex.fe_len = 0;
4088 ac->ac_status = AC_STATUS_CONTINUE;
4089 ac->ac_groups_scanned = 0;
4090 ac->ac_ex_scanned = 0;
4091 ac->ac_found = 0;
4092 ac->ac_sb = sb;
4093 ac->ac_inode = ar->inode;
4094 ac->ac_o_ex.fe_logical = ar->logical;
4095 ac->ac_o_ex.fe_group = group;
4096 ac->ac_o_ex.fe_start = block;
4097 ac->ac_o_ex.fe_len = len;
4098 ac->ac_g_ex.fe_logical = ar->logical;
4099 ac->ac_g_ex.fe_group = group;
4100 ac->ac_g_ex.fe_start = block;
4101 ac->ac_g_ex.fe_len = len;
4102 ac->ac_f_ex.fe_len = 0;
4103 ac->ac_flags = ar->flags;
4104 ac->ac_2order = 0;
4105 ac->ac_criteria = 0;
4106 ac->ac_pa = NULL;
4107 ac->ac_bitmap_page = NULL;
4108 ac->ac_buddy_page = NULL;
4109 ac->ac_lg = NULL;
4111 /* we have to define context: we'll we work with a file or
4112 * locality group. this is a policy, actually */
4113 ext4_mb_group_or_file(ac);
4115 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4116 "left: %u/%u, right %u/%u to %swritable\n",
4117 (unsigned) ar->len, (unsigned) ar->logical,
4118 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4119 (unsigned) ar->lleft, (unsigned) ar->pleft,
4120 (unsigned) ar->lright, (unsigned) ar->pright,
4121 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4122 return 0;
4127 * release all resource we used in allocation
4129 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4131 if (ac->ac_pa) {
4132 if (ac->ac_pa->pa_linear) {
4133 /* see comment in ext4_mb_use_group_pa() */
4134 spin_lock(&ac->ac_pa->pa_lock);
4135 ac->ac_pa->pa_pstart += ac->ac_b_ex.fe_len;
4136 ac->ac_pa->pa_lstart += ac->ac_b_ex.fe_len;
4137 ac->ac_pa->pa_free -= ac->ac_b_ex.fe_len;
4138 ac->ac_pa->pa_len -= ac->ac_b_ex.fe_len;
4139 spin_unlock(&ac->ac_pa->pa_lock);
4141 ext4_mb_put_pa(ac, ac->ac_sb, ac->ac_pa);
4143 if (ac->ac_bitmap_page)
4144 page_cache_release(ac->ac_bitmap_page);
4145 if (ac->ac_buddy_page)
4146 page_cache_release(ac->ac_buddy_page);
4147 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4148 mutex_unlock(&ac->ac_lg->lg_mutex);
4149 ext4_mb_collect_stats(ac);
4150 return 0;
4153 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4155 ext4_group_t i;
4156 int ret;
4157 int freed = 0;
4159 for (i = 0; i < EXT4_SB(sb)->s_groups_count && needed > 0; i++) {
4160 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4161 freed += ret;
4162 needed -= ret;
4165 return freed;
4169 * Main entry point into mballoc to allocate blocks
4170 * it tries to use preallocation first, then falls back
4171 * to usual allocation
4173 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4174 struct ext4_allocation_request *ar, int *errp)
4176 struct ext4_allocation_context *ac = NULL;
4177 struct ext4_sb_info *sbi;
4178 struct super_block *sb;
4179 ext4_fsblk_t block = 0;
4180 int freed;
4181 int inquota;
4183 sb = ar->inode->i_sb;
4184 sbi = EXT4_SB(sb);
4186 if (!test_opt(sb, MBALLOC)) {
4187 block = ext4_old_new_blocks(handle, ar->inode, ar->goal,
4188 &(ar->len), errp);
4189 return block;
4191 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) {
4193 * With delalloc we already reserved the blocks
4195 ar->len = ext4_has_free_blocks(sbi, ar->len);
4198 if (ar->len == 0) {
4199 *errp = -ENOSPC;
4200 return 0;
4203 while (ar->len && DQUOT_ALLOC_BLOCK(ar->inode, ar->len)) {
4204 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4205 ar->len--;
4207 if (ar->len == 0) {
4208 *errp = -EDQUOT;
4209 return 0;
4211 inquota = ar->len;
4213 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4214 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4216 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4217 if (!ac) {
4218 ar->len = 0;
4219 *errp = -ENOMEM;
4220 goto out1;
4223 ext4_mb_poll_new_transaction(sb, handle);
4225 *errp = ext4_mb_initialize_context(ac, ar);
4226 if (*errp) {
4227 ar->len = 0;
4228 goto out2;
4231 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4232 if (!ext4_mb_use_preallocated(ac)) {
4233 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4234 ext4_mb_normalize_request(ac, ar);
4235 repeat:
4236 /* allocate space in core */
4237 ext4_mb_regular_allocator(ac);
4239 /* as we've just preallocated more space than
4240 * user requested orinally, we store allocated
4241 * space in a special descriptor */
4242 if (ac->ac_status == AC_STATUS_FOUND &&
4243 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4244 ext4_mb_new_preallocation(ac);
4247 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4248 *errp = ext4_mb_mark_diskspace_used(ac, handle);
4249 if (*errp == -EAGAIN) {
4250 ac->ac_b_ex.fe_group = 0;
4251 ac->ac_b_ex.fe_start = 0;
4252 ac->ac_b_ex.fe_len = 0;
4253 ac->ac_status = AC_STATUS_CONTINUE;
4254 goto repeat;
4255 } else if (*errp) {
4256 ac->ac_b_ex.fe_len = 0;
4257 ar->len = 0;
4258 ext4_mb_show_ac(ac);
4259 } else {
4260 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4261 ar->len = ac->ac_b_ex.fe_len;
4263 } else {
4264 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4265 if (freed)
4266 goto repeat;
4267 *errp = -ENOSPC;
4268 ac->ac_b_ex.fe_len = 0;
4269 ar->len = 0;
4270 ext4_mb_show_ac(ac);
4273 ext4_mb_release_context(ac);
4275 out2:
4276 kmem_cache_free(ext4_ac_cachep, ac);
4277 out1:
4278 if (ar->len < inquota)
4279 DQUOT_FREE_BLOCK(ar->inode, inquota - ar->len);
4281 return block;
4283 static void ext4_mb_poll_new_transaction(struct super_block *sb,
4284 handle_t *handle)
4286 struct ext4_sb_info *sbi = EXT4_SB(sb);
4288 if (sbi->s_last_transaction == handle->h_transaction->t_tid)
4289 return;
4291 /* new transaction! time to close last one and free blocks for
4292 * committed transaction. we know that only transaction can be
4293 * active, so previos transaction can be being logged and we
4294 * know that transaction before previous is known to be already
4295 * logged. this means that now we may free blocks freed in all
4296 * transactions before previous one. hope I'm clear enough ... */
4298 spin_lock(&sbi->s_md_lock);
4299 if (sbi->s_last_transaction != handle->h_transaction->t_tid) {
4300 mb_debug("new transaction %lu, old %lu\n",
4301 (unsigned long) handle->h_transaction->t_tid,
4302 (unsigned long) sbi->s_last_transaction);
4303 list_splice_init(&sbi->s_closed_transaction,
4304 &sbi->s_committed_transaction);
4305 list_splice_init(&sbi->s_active_transaction,
4306 &sbi->s_closed_transaction);
4307 sbi->s_last_transaction = handle->h_transaction->t_tid;
4309 spin_unlock(&sbi->s_md_lock);
4311 ext4_mb_free_committed_blocks(sb);
4314 static noinline_for_stack int
4315 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4316 ext4_group_t group, ext4_grpblk_t block, int count)
4318 struct ext4_group_info *db = e4b->bd_info;
4319 struct super_block *sb = e4b->bd_sb;
4320 struct ext4_sb_info *sbi = EXT4_SB(sb);
4321 struct ext4_free_metadata *md;
4322 int i;
4324 BUG_ON(e4b->bd_bitmap_page == NULL);
4325 BUG_ON(e4b->bd_buddy_page == NULL);
4327 ext4_lock_group(sb, group);
4328 for (i = 0; i < count; i++) {
4329 md = db->bb_md_cur;
4330 if (md && db->bb_tid != handle->h_transaction->t_tid) {
4331 db->bb_md_cur = NULL;
4332 md = NULL;
4335 if (md == NULL) {
4336 ext4_unlock_group(sb, group);
4337 md = kmalloc(sizeof(*md), GFP_NOFS);
4338 if (md == NULL)
4339 return -ENOMEM;
4340 md->num = 0;
4341 md->group = group;
4343 ext4_lock_group(sb, group);
4344 if (db->bb_md_cur == NULL) {
4345 spin_lock(&sbi->s_md_lock);
4346 list_add(&md->list, &sbi->s_active_transaction);
4347 spin_unlock(&sbi->s_md_lock);
4348 /* protect buddy cache from being freed,
4349 * otherwise we'll refresh it from
4350 * on-disk bitmap and lose not-yet-available
4351 * blocks */
4352 page_cache_get(e4b->bd_buddy_page);
4353 page_cache_get(e4b->bd_bitmap_page);
4354 db->bb_md_cur = md;
4355 db->bb_tid = handle->h_transaction->t_tid;
4356 mb_debug("new md 0x%p for group %lu\n",
4357 md, md->group);
4358 } else {
4359 kfree(md);
4360 md = db->bb_md_cur;
4364 BUG_ON(md->num >= EXT4_BB_MAX_BLOCKS);
4365 md->blocks[md->num] = block + i;
4366 md->num++;
4367 if (md->num == EXT4_BB_MAX_BLOCKS) {
4368 /* no more space, put full container on a sb's list */
4369 db->bb_md_cur = NULL;
4372 ext4_unlock_group(sb, group);
4373 return 0;
4377 * Main entry point into mballoc to free blocks
4379 void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
4380 unsigned long block, unsigned long count,
4381 int metadata, unsigned long *freed)
4383 struct buffer_head *bitmap_bh = NULL;
4384 struct super_block *sb = inode->i_sb;
4385 struct ext4_allocation_context *ac = NULL;
4386 struct ext4_group_desc *gdp;
4387 struct ext4_super_block *es;
4388 unsigned long overflow;
4389 ext4_grpblk_t bit;
4390 struct buffer_head *gd_bh;
4391 ext4_group_t block_group;
4392 struct ext4_sb_info *sbi;
4393 struct ext4_buddy e4b;
4394 int err = 0;
4395 int ret;
4397 *freed = 0;
4399 ext4_mb_poll_new_transaction(sb, handle);
4401 sbi = EXT4_SB(sb);
4402 es = EXT4_SB(sb)->s_es;
4403 if (block < le32_to_cpu(es->s_first_data_block) ||
4404 block + count < block ||
4405 block + count > ext4_blocks_count(es)) {
4406 ext4_error(sb, __func__,
4407 "Freeing blocks not in datazone - "
4408 "block = %lu, count = %lu", block, count);
4409 goto error_return;
4412 ext4_debug("freeing block %lu\n", block);
4414 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4415 if (ac) {
4416 ac->ac_op = EXT4_MB_HISTORY_FREE;
4417 ac->ac_inode = inode;
4418 ac->ac_sb = sb;
4421 do_more:
4422 overflow = 0;
4423 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4426 * Check to see if we are freeing blocks across a group
4427 * boundary.
4429 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4430 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4431 count -= overflow;
4433 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4434 if (!bitmap_bh) {
4435 err = -EIO;
4436 goto error_return;
4438 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4439 if (!gdp) {
4440 err = -EIO;
4441 goto error_return;
4444 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4445 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4446 in_range(block, ext4_inode_table(sb, gdp),
4447 EXT4_SB(sb)->s_itb_per_group) ||
4448 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4449 EXT4_SB(sb)->s_itb_per_group)) {
4451 ext4_error(sb, __func__,
4452 "Freeing blocks in system zone - "
4453 "Block = %lu, count = %lu", block, count);
4454 /* err = 0. ext4_std_error should be a no op */
4455 goto error_return;
4458 BUFFER_TRACE(bitmap_bh, "getting write access");
4459 err = ext4_journal_get_write_access(handle, bitmap_bh);
4460 if (err)
4461 goto error_return;
4464 * We are about to modify some metadata. Call the journal APIs
4465 * to unshare ->b_data if a currently-committing transaction is
4466 * using it
4468 BUFFER_TRACE(gd_bh, "get_write_access");
4469 err = ext4_journal_get_write_access(handle, gd_bh);
4470 if (err)
4471 goto error_return;
4473 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4474 if (err)
4475 goto error_return;
4477 #ifdef AGGRESSIVE_CHECK
4479 int i;
4480 for (i = 0; i < count; i++)
4481 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4483 #endif
4484 mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data,
4485 bit, count);
4487 /* We dirtied the bitmap block */
4488 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4489 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
4491 if (ac) {
4492 ac->ac_b_ex.fe_group = block_group;
4493 ac->ac_b_ex.fe_start = bit;
4494 ac->ac_b_ex.fe_len = count;
4495 ext4_mb_store_history(ac);
4498 if (metadata) {
4499 /* blocks being freed are metadata. these blocks shouldn't
4500 * be used until this transaction is committed */
4501 ext4_mb_free_metadata(handle, &e4b, block_group, bit, count);
4502 } else {
4503 ext4_lock_group(sb, block_group);
4504 mb_free_blocks(inode, &e4b, bit, count);
4505 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4506 ext4_unlock_group(sb, block_group);
4509 spin_lock(sb_bgl_lock(sbi, block_group));
4510 le16_add_cpu(&gdp->bg_free_blocks_count, count);
4511 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4512 spin_unlock(sb_bgl_lock(sbi, block_group));
4513 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4515 if (sbi->s_log_groups_per_flex) {
4516 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4517 spin_lock(sb_bgl_lock(sbi, flex_group));
4518 sbi->s_flex_groups[flex_group].free_blocks += count;
4519 spin_unlock(sb_bgl_lock(sbi, flex_group));
4522 ext4_mb_release_desc(&e4b);
4524 *freed += count;
4526 /* And the group descriptor block */
4527 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4528 ret = ext4_journal_dirty_metadata(handle, gd_bh);
4529 if (!err)
4530 err = ret;
4532 if (overflow && !err) {
4533 block += count;
4534 count = overflow;
4535 put_bh(bitmap_bh);
4536 goto do_more;
4538 sb->s_dirt = 1;
4539 error_return:
4540 brelse(bitmap_bh);
4541 ext4_std_error(sb, err);
4542 if (ac)
4543 kmem_cache_free(ext4_ac_cachep, ac);
4544 return;