Merge branch 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4
[linux-2.6.git] / fs / ext4 / mballoc.c
1 /*
2  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3  * Written by Alex Tomas <alex@clusterfs.com>
4  *
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.
8  *
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.
13  *
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-
17  */
18
19
20 /*
21  * mballoc.c contains the multiblocks allocation routines
22  */
23
24 #include "mballoc.h"
25 #include <linux/debugfs.h>
26 #include <trace/events/ext4.h>
27
28 /*
29  * MUSTDO:
30  *   - test ext4_ext_search_left() and ext4_ext_search_right()
31  *   - search for metadata in few groups
32  *
33  * TODO v4:
34  *   - normalization should take into account whether file is still open
35  *   - discard preallocations if no free space left (policy?)
36  *   - don't normalize tails
37  *   - quota
38  *   - reservation for superuser
39  *
40  * TODO v3:
41  *   - bitmap read-ahead (proposed by Oleg Drokin aka green)
42  *   - track min/max extents in each group for better group selection
43  *   - mb_mark_used() may allocate chunk right after splitting buddy
44  *   - tree of groups sorted by number of free blocks
45  *   - error handling
46  */
47
48 /*
49  * The allocation request involve request for multiple number of blocks
50  * near to the goal(block) value specified.
51  *
52  * During initialization phase of the allocator we decide to use the
53  * group preallocation or inode preallocation depending on the size of
54  * the file. The size of the file could be the resulting file size we
55  * would have after allocation, or the current file size, which ever
56  * is larger. If the size is less than sbi->s_mb_stream_request we
57  * select to use the group preallocation. The default value of
58  * s_mb_stream_request is 16 blocks. This can also be tuned via
59  * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
60  * terms of number of blocks.
61  *
62  * The main motivation for having small file use group preallocation is to
63  * ensure that we have small files closer together on the disk.
64  *
65  * First stage the allocator looks at the inode prealloc list,
66  * ext4_inode_info->i_prealloc_list, which contains list of prealloc
67  * spaces for this particular inode. The inode prealloc space is
68  * represented as:
69  *
70  * pa_lstart -> the logical start block for this prealloc space
71  * pa_pstart -> the physical start block for this prealloc space
72  * pa_len    -> lenght for this prealloc space
73  * pa_free   ->  free space available in this prealloc space
74  *
75  * The inode preallocation space is used looking at the _logical_ start
76  * block. If only the logical file block falls within the range of prealloc
77  * space we will consume the particular prealloc space. This make sure that
78  * that the we have contiguous physical blocks representing the file blocks
79  *
80  * The important thing to be noted in case of inode prealloc space is that
81  * we don't modify the values associated to inode prealloc space except
82  * pa_free.
83  *
84  * If we are not able to find blocks in the inode prealloc space and if we
85  * have the group allocation flag set then we look at the locality group
86  * prealloc space. These are per CPU prealloc list repreasented as
87  *
88  * ext4_sb_info.s_locality_groups[smp_processor_id()]
89  *
90  * The reason for having a per cpu locality group is to reduce the contention
91  * between CPUs. It is possible to get scheduled at this point.
92  *
93  * The locality group prealloc space is used looking at whether we have
94  * enough free space (pa_free) withing the prealloc space.
95  *
96  * If we can't allocate blocks via inode prealloc or/and locality group
97  * prealloc then we look at the buddy cache. The buddy cache is represented
98  * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
99  * mapped to the buddy and bitmap information regarding different
100  * groups. The buddy information is attached to buddy cache inode so that
101  * we can access them through the page cache. The information regarding
102  * each group is loaded via ext4_mb_load_buddy.  The information involve
103  * block bitmap and buddy information. The information are stored in the
104  * inode as:
105  *
106  *  {                        page                        }
107  *  [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
108  *
109  *
110  * one block each for bitmap and buddy information.  So for each group we
111  * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
112  * blocksize) blocks.  So it can have information regarding groups_per_page
113  * which is blocks_per_page/2
114  *
115  * The buddy cache inode is not stored on disk. The inode is thrown
116  * away when the filesystem is unmounted.
117  *
118  * We look for count number of blocks in the buddy cache. If we were able
119  * to locate that many free blocks we return with additional information
120  * regarding rest of the contiguous physical block available
121  *
122  * Before allocating blocks via buddy cache we normalize the request
123  * blocks. This ensure we ask for more blocks that we needed. The extra
124  * blocks that we get after allocation is added to the respective prealloc
125  * list. In case of inode preallocation we follow a list of heuristics
126  * based on file size. This can be found in ext4_mb_normalize_request. If
127  * we are doing a group prealloc we try to normalize the request to
128  * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
129  * 512 blocks. This can be tuned via
130  * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
131  * terms of number of blocks. If we have mounted the file system with -O
132  * stripe=<value> option the group prealloc request is normalized to the
133  * stripe value (sbi->s_stripe)
134  *
135  * The regular allocator(using the buddy cache) supports few tunables.
136  *
137  * /sys/fs/ext4/<partition>/mb_min_to_scan
138  * /sys/fs/ext4/<partition>/mb_max_to_scan
139  * /sys/fs/ext4/<partition>/mb_order2_req
140  *
141  * The regular allocator uses buddy scan only if the request len is power of
142  * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
143  * value of s_mb_order2_reqs can be tuned via
144  * /sys/fs/ext4/<partition>/mb_order2_req.  If the request len is equal to
145  * stripe size (sbi->s_stripe), we try to search for contigous block in
146  * stripe size. This should result in better allocation on RAID setups. If
147  * not, we search in the specific group using bitmap for best extents. The
148  * tunable min_to_scan and max_to_scan control the behaviour here.
149  * min_to_scan indicate how long the mballoc __must__ look for a best
150  * extent and max_to_scan indicates how long the mballoc __can__ look for a
151  * best extent in the found extents. Searching for the blocks starts with
152  * the group specified as the goal value in allocation context via
153  * ac_g_ex. Each group is first checked based on the criteria whether it
154  * can used for allocation. ext4_mb_good_group explains how the groups are
155  * checked.
156  *
157  * Both the prealloc space are getting populated as above. So for the first
158  * request we will hit the buddy cache which will result in this prealloc
159  * space getting filled. The prealloc space is then later used for the
160  * subsequent request.
161  */
162
163 /*
164  * mballoc operates on the following data:
165  *  - on-disk bitmap
166  *  - in-core buddy (actually includes buddy and bitmap)
167  *  - preallocation descriptors (PAs)
168  *
169  * there are two types of preallocations:
170  *  - inode
171  *    assiged to specific inode and can be used for this inode only.
172  *    it describes part of inode's space preallocated to specific
173  *    physical blocks. any block from that preallocated can be used
174  *    independent. the descriptor just tracks number of blocks left
175  *    unused. so, before taking some block from descriptor, one must
176  *    make sure corresponded logical block isn't allocated yet. this
177  *    also means that freeing any block within descriptor's range
178  *    must discard all preallocated blocks.
179  *  - locality group
180  *    assigned to specific locality group which does not translate to
181  *    permanent set of inodes: inode can join and leave group. space
182  *    from this type of preallocation can be used for any inode. thus
183  *    it's consumed from the beginning to the end.
184  *
185  * relation between them can be expressed as:
186  *    in-core buddy = on-disk bitmap + preallocation descriptors
187  *
188  * this mean blocks mballoc considers used are:
189  *  - allocated blocks (persistent)
190  *  - preallocated blocks (non-persistent)
191  *
192  * consistency in mballoc world means that at any time a block is either
193  * free or used in ALL structures. notice: "any time" should not be read
194  * literally -- time is discrete and delimited by locks.
195  *
196  *  to keep it simple, we don't use block numbers, instead we count number of
197  *  blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
198  *
199  * all operations can be expressed as:
200  *  - init buddy:                       buddy = on-disk + PAs
201  *  - new PA:                           buddy += N; PA = N
202  *  - use inode PA:                     on-disk += N; PA -= N
203  *  - discard inode PA                  buddy -= on-disk - PA; PA = 0
204  *  - use locality group PA             on-disk += N; PA -= N
205  *  - discard locality group PA         buddy -= PA; PA = 0
206  *  note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
207  *        is used in real operation because we can't know actual used
208  *        bits from PA, only from on-disk bitmap
209  *
210  * if we follow this strict logic, then all operations above should be atomic.
211  * given some of them can block, we'd have to use something like semaphores
212  * killing performance on high-end SMP hardware. let's try to relax it using
213  * the following knowledge:
214  *  1) if buddy is referenced, it's already initialized
215  *  2) while block is used in buddy and the buddy is referenced,
216  *     nobody can re-allocate that block
217  *  3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
218  *     bit set and PA claims same block, it's OK. IOW, one can set bit in
219  *     on-disk bitmap if buddy has same bit set or/and PA covers corresponded
220  *     block
221  *
222  * so, now we're building a concurrency table:
223  *  - init buddy vs.
224  *    - new PA
225  *      blocks for PA are allocated in the buddy, buddy must be referenced
226  *      until PA is linked to allocation group to avoid concurrent buddy init
227  *    - use inode PA
228  *      we need to make sure that either on-disk bitmap or PA has uptodate data
229  *      given (3) we care that PA-=N operation doesn't interfere with init
230  *    - discard inode PA
231  *      the simplest way would be to have buddy initialized by the discard
232  *    - use locality group PA
233  *      again PA-=N must be serialized with init
234  *    - discard locality group PA
235  *      the simplest way would be to have buddy initialized by the discard
236  *  - new PA vs.
237  *    - use inode PA
238  *      i_data_sem serializes them
239  *    - discard inode PA
240  *      discard process must wait until PA isn't used by another process
241  *    - use locality group PA
242  *      some mutex should serialize them
243  *    - discard locality group PA
244  *      discard process must wait until PA isn't used by another process
245  *  - use inode PA
246  *    - use inode PA
247  *      i_data_sem or another mutex should serializes them
248  *    - discard inode PA
249  *      discard process must wait until PA isn't used by another process
250  *    - use locality group PA
251  *      nothing wrong here -- they're different PAs covering different blocks
252  *    - discard locality group PA
253  *      discard process must wait until PA isn't used by another process
254  *
255  * now we're ready to make few consequences:
256  *  - PA is referenced and while it is no discard is possible
257  *  - PA is referenced until block isn't marked in on-disk bitmap
258  *  - PA changes only after on-disk bitmap
259  *  - discard must not compete with init. either init is done before
260  *    any discard or they're serialized somehow
261  *  - buddy init as sum of on-disk bitmap and PAs is done atomically
262  *
263  * a special case when we've used PA to emptiness. no need to modify buddy
264  * in this case, but we should care about concurrent init
265  *
266  */
267
268  /*
269  * Logic in few words:
270  *
271  *  - allocation:
272  *    load group
273  *    find blocks
274  *    mark bits in on-disk bitmap
275  *    release group
276  *
277  *  - use preallocation:
278  *    find proper PA (per-inode or group)
279  *    load group
280  *    mark bits in on-disk bitmap
281  *    release group
282  *    release PA
283  *
284  *  - free:
285  *    load group
286  *    mark bits in on-disk bitmap
287  *    release group
288  *
289  *  - discard preallocations in group:
290  *    mark PAs deleted
291  *    move them onto local list
292  *    load on-disk bitmap
293  *    load group
294  *    remove PA from object (inode or locality group)
295  *    mark free blocks in-core
296  *
297  *  - discard inode's preallocations:
298  */
299
300 /*
301  * Locking rules
302  *
303  * Locks:
304  *  - bitlock on a group        (group)
305  *  - object (inode/locality)   (object)
306  *  - per-pa lock               (pa)
307  *
308  * Paths:
309  *  - new pa
310  *    object
311  *    group
312  *
313  *  - find and use pa:
314  *    pa
315  *
316  *  - release consumed pa:
317  *    pa
318  *    group
319  *    object
320  *
321  *  - generate in-core bitmap:
322  *    group
323  *        pa
324  *
325  *  - discard all for given object (inode, locality group):
326  *    object
327  *        pa
328  *    group
329  *
330  *  - discard all for given group:
331  *    group
332  *        pa
333  *    group
334  *        object
335  *
336  */
337 static struct kmem_cache *ext4_pspace_cachep;
338 static struct kmem_cache *ext4_ac_cachep;
339 static struct kmem_cache *ext4_free_ext_cachep;
340 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
341                                         ext4_group_t group);
342 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
343                                                 ext4_group_t group);
344 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
345
346 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
347 {
348 #if BITS_PER_LONG == 64
349         *bit += ((unsigned long) addr & 7UL) << 3;
350         addr = (void *) ((unsigned long) addr & ~7UL);
351 #elif BITS_PER_LONG == 32
352         *bit += ((unsigned long) addr & 3UL) << 3;
353         addr = (void *) ((unsigned long) addr & ~3UL);
354 #else
355 #error "how many bits you are?!"
356 #endif
357         return addr;
358 }
359
360 static inline int mb_test_bit(int bit, void *addr)
361 {
362         /*
363          * ext4_test_bit on architecture like powerpc
364          * needs unsigned long aligned address
365          */
366         addr = mb_correct_addr_and_bit(&bit, addr);
367         return ext4_test_bit(bit, addr);
368 }
369
370 static inline void mb_set_bit(int bit, void *addr)
371 {
372         addr = mb_correct_addr_and_bit(&bit, addr);
373         ext4_set_bit(bit, addr);
374 }
375
376 static inline void mb_clear_bit(int bit, void *addr)
377 {
378         addr = mb_correct_addr_and_bit(&bit, addr);
379         ext4_clear_bit(bit, addr);
380 }
381
382 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
383 {
384         int fix = 0, ret, tmpmax;
385         addr = mb_correct_addr_and_bit(&fix, addr);
386         tmpmax = max + fix;
387         start += fix;
388
389         ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
390         if (ret > max)
391                 return max;
392         return ret;
393 }
394
395 static inline int mb_find_next_bit(void *addr, int max, int start)
396 {
397         int fix = 0, ret, tmpmax;
398         addr = mb_correct_addr_and_bit(&fix, addr);
399         tmpmax = max + fix;
400         start += fix;
401
402         ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
403         if (ret > max)
404                 return max;
405         return ret;
406 }
407
408 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
409 {
410         char *bb;
411
412         BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
413         BUG_ON(max == NULL);
414
415         if (order > e4b->bd_blkbits + 1) {
416                 *max = 0;
417                 return NULL;
418         }
419
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);
424
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];
427
428         return bb;
429 }
430
431 #ifdef DOUBLE_CHECK
432 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
433                            int first, int count)
434 {
435         int i;
436         struct super_block *sb = e4b->bd_sb;
437
438         if (unlikely(e4b->bd_info->bb_bitmap == NULL))
439                 return;
440         assert_spin_locked(ext4_group_lock_ptr(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);
448                         ext4_grp_locked_error(sb, e4b->bd_group,
449                                    __func__, "double-free of inode"
450                                    " %lu's block %llu(bit %u in group %u)",
451                                    inode ? inode->i_ino : 0, blocknr,
452                                    first + i, e4b->bd_group);
453                 }
454                 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
455         }
456 }
457
458 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
459 {
460         int i;
461
462         if (unlikely(e4b->bd_info->bb_bitmap == NULL))
463                 return;
464         assert_spin_locked(ext4_group_lock_ptr(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);
468         }
469 }
470
471 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
472 {
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(KERN_ERR "corruption in group %u "
481                                        "at byte %u(%u): %x in copy != %x "
482                                        "on disk/prealloc\n",
483                                        e4b->bd_group, i, i * 8, b1[i], b2[i]);
484                                 BUG();
485                         }
486                 }
487         }
488 }
489
490 #else
491 static inline void mb_free_blocks_double(struct inode *inode,
492                                 struct ext4_buddy *e4b, int first, int count)
493 {
494         return;
495 }
496 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
497                                                 int first, int count)
498 {
499         return;
500 }
501 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
502 {
503         return;
504 }
505 #endif
506
507 #ifdef AGGRESSIVE_CHECK
508
509 #define MB_CHECK_ASSERT(assert)                                         \
510 do {                                                                    \
511         if (!(assert)) {                                                \
512                 printk(KERN_EMERG                                       \
513                         "Assertion failure in %s() at %s:%d: \"%s\"\n", \
514                         function, file, line, # assert);                \
515                 BUG();                                                  \
516         }                                                               \
517 } while (0)
518
519 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
520                                 const char *function, int line)
521 {
522         struct super_block *sb = e4b->bd_sb;
523         int order = e4b->bd_blkbits + 1;
524         int max;
525         int max2;
526         int i;
527         int j;
528         int k;
529         int count;
530         struct ext4_group_info *grp;
531         int fragments = 0;
532         int fstart;
533         struct list_head *cur;
534         void *buddy;
535         void *buddy2;
536
537         {
538                 static int mb_check_counter;
539                 if (mb_check_counter++ % 100 != 0)
540                         return 0;
541         }
542
543         while (order > 1) {
544                 buddy = mb_find_buddy(e4b, order, &max);
545                 MB_CHECK_ASSERT(buddy);
546                 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
547                 MB_CHECK_ASSERT(buddy2);
548                 MB_CHECK_ASSERT(buddy != buddy2);
549                 MB_CHECK_ASSERT(max * 2 == max2);
550
551                 count = 0;
552                 for (i = 0; i < max; i++) {
553
554                         if (mb_test_bit(i, buddy)) {
555                                 /* only single bit in buddy2 may be 1 */
556                                 if (!mb_test_bit(i << 1, buddy2)) {
557                                         MB_CHECK_ASSERT(
558                                                 mb_test_bit((i<<1)+1, buddy2));
559                                 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
560                                         MB_CHECK_ASSERT(
561                                                 mb_test_bit(i << 1, buddy2));
562                                 }
563                                 continue;
564                         }
565
566                         /* both bits in buddy2 must be 0 */
567                         MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
568                         MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
569
570                         for (j = 0; j < (1 << order); j++) {
571                                 k = (i * (1 << order)) + j;
572                                 MB_CHECK_ASSERT(
573                                         !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
574                         }
575                         count++;
576                 }
577                 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
578                 order--;
579         }
580
581         fstart = -1;
582         buddy = mb_find_buddy(e4b, 0, &max);
583         for (i = 0; i < max; i++) {
584                 if (!mb_test_bit(i, buddy)) {
585                         MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
586                         if (fstart == -1) {
587                                 fragments++;
588                                 fstart = i;
589                         }
590                         continue;
591                 }
592                 fstart = -1;
593                 /* check used bits only */
594                 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
595                         buddy2 = mb_find_buddy(e4b, j, &max2);
596                         k = i >> j;
597                         MB_CHECK_ASSERT(k < max2);
598                         MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
599                 }
600         }
601         MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
602         MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
603
604         grp = ext4_get_group_info(sb, e4b->bd_group);
605         buddy = mb_find_buddy(e4b, 0, &max);
606         list_for_each(cur, &grp->bb_prealloc_list) {
607                 ext4_group_t groupnr;
608                 struct ext4_prealloc_space *pa;
609                 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
610                 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
611                 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
612                 for (i = 0; i < pa->pa_len; i++)
613                         MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
614         }
615         return 0;
616 }
617 #undef MB_CHECK_ASSERT
618 #define mb_check_buddy(e4b) __mb_check_buddy(e4b,       \
619                                         __FILE__, __func__, __LINE__)
620 #else
621 #define mb_check_buddy(e4b)
622 #endif
623
624 /* FIXME!! need more doc */
625 static void ext4_mb_mark_free_simple(struct super_block *sb,
626                                 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
627                                         struct ext4_group_info *grp)
628 {
629         struct ext4_sb_info *sbi = EXT4_SB(sb);
630         ext4_grpblk_t min;
631         ext4_grpblk_t max;
632         ext4_grpblk_t chunk;
633         unsigned short border;
634
635         BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
636
637         border = 2 << sb->s_blocksize_bits;
638
639         while (len > 0) {
640                 /* find how many blocks can be covered since this position */
641                 max = ffs(first | border) - 1;
642
643                 /* find how many blocks of power 2 we need to mark */
644                 min = fls(len) - 1;
645
646                 if (max < min)
647                         min = max;
648                 chunk = 1 << min;
649
650                 /* mark multiblock chunks only */
651                 grp->bb_counters[min]++;
652                 if (min > 0)
653                         mb_clear_bit(first >> min,
654                                      buddy + sbi->s_mb_offsets[min]);
655
656                 len -= chunk;
657                 first += chunk;
658         }
659 }
660
661 static noinline_for_stack
662 void ext4_mb_generate_buddy(struct super_block *sb,
663                                 void *buddy, void *bitmap, ext4_group_t group)
664 {
665         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
666         ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb);
667         ext4_grpblk_t i = 0;
668         ext4_grpblk_t first;
669         ext4_grpblk_t len;
670         unsigned free = 0;
671         unsigned fragments = 0;
672         unsigned long long period = get_cycles();
673
674         /* initialize buddy from bitmap which is aggregation
675          * of on-disk bitmap and preallocations */
676         i = mb_find_next_zero_bit(bitmap, max, 0);
677         grp->bb_first_free = i;
678         while (i < max) {
679                 fragments++;
680                 first = i;
681                 i = mb_find_next_bit(bitmap, max, i);
682                 len = i - first;
683                 free += len;
684                 if (len > 1)
685                         ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
686                 else
687                         grp->bb_counters[0]++;
688                 if (i < max)
689                         i = mb_find_next_zero_bit(bitmap, max, i);
690         }
691         grp->bb_fragments = fragments;
692
693         if (free != grp->bb_free) {
694                 ext4_grp_locked_error(sb, group,  __func__,
695                         "EXT4-fs: group %u: %u blocks in bitmap, %u in gd",
696                         group, free, grp->bb_free);
697                 /*
698                  * If we intent to continue, we consider group descritor
699                  * corrupt and update bb_free using bitmap value
700                  */
701                 grp->bb_free = free;
702         }
703
704         clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
705
706         period = get_cycles() - period;
707         spin_lock(&EXT4_SB(sb)->s_bal_lock);
708         EXT4_SB(sb)->s_mb_buddies_generated++;
709         EXT4_SB(sb)->s_mb_generation_time += period;
710         spin_unlock(&EXT4_SB(sb)->s_bal_lock);
711 }
712
713 /* The buddy information is attached the buddy cache inode
714  * for convenience. The information regarding each group
715  * is loaded via ext4_mb_load_buddy. The information involve
716  * block bitmap and buddy information. The information are
717  * stored in the inode as
718  *
719  * {                        page                        }
720  * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
721  *
722  *
723  * one block each for bitmap and buddy information.
724  * So for each group we take up 2 blocks. A page can
725  * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize)  blocks.
726  * So it can have information regarding groups_per_page which
727  * is blocks_per_page/2
728  */
729
730 static int ext4_mb_init_cache(struct page *page, char *incore)
731 {
732         ext4_group_t ngroups;
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;
746
747         mb_debug(1, "init page %lu\n", page->index);
748
749         inode = page->mapping->host;
750         sb = inode->i_sb;
751         ngroups = ext4_get_groups_count(sb);
752         blocksize = 1 << inode->i_blkbits;
753         blocks_per_page = PAGE_CACHE_SIZE / blocksize;
754
755         groups_per_page = blocks_per_page >> 1;
756         if (groups_per_page == 0)
757                 groups_per_page = 1;
758
759         /* allocate buffer_heads to read bitmaps */
760         if (groups_per_page > 1) {
761                 err = -ENOMEM;
762                 i = sizeof(struct buffer_head *) * groups_per_page;
763                 bh = kzalloc(i, GFP_NOFS);
764                 if (bh == NULL)
765                         goto out;
766         } else
767                 bh = &bhs;
768
769         first_group = page->index * blocks_per_page / 2;
770
771         /* read all groups the page covers into the cache */
772         for (i = 0; i < groups_per_page; i++) {
773                 struct ext4_group_desc *desc;
774
775                 if (first_group + i >= ngroups)
776                         break;
777
778                 err = -EIO;
779                 desc = ext4_get_group_desc(sb, first_group + i, NULL);
780                 if (desc == NULL)
781                         goto out;
782
783                 err = -ENOMEM;
784                 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
785                 if (bh[i] == NULL)
786                         goto out;
787
788                 if (bitmap_uptodate(bh[i]))
789                         continue;
790
791                 lock_buffer(bh[i]);
792                 if (bitmap_uptodate(bh[i])) {
793                         unlock_buffer(bh[i]);
794                         continue;
795                 }
796                 ext4_lock_group(sb, first_group + i);
797                 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
798                         ext4_init_block_bitmap(sb, bh[i],
799                                                 first_group + i, desc);
800                         set_bitmap_uptodate(bh[i]);
801                         set_buffer_uptodate(bh[i]);
802                         ext4_unlock_group(sb, first_group + i);
803                         unlock_buffer(bh[i]);
804                         continue;
805                 }
806                 ext4_unlock_group(sb, first_group + i);
807                 if (buffer_uptodate(bh[i])) {
808                         /*
809                          * if not uninit if bh is uptodate,
810                          * bitmap is also uptodate
811                          */
812                         set_bitmap_uptodate(bh[i]);
813                         unlock_buffer(bh[i]);
814                         continue;
815                 }
816                 get_bh(bh[i]);
817                 /*
818                  * submit the buffer_head for read. We can
819                  * safely mark the bitmap as uptodate now.
820                  * We do it here so the bitmap uptodate bit
821                  * get set with buffer lock held.
822                  */
823                 set_bitmap_uptodate(bh[i]);
824                 bh[i]->b_end_io = end_buffer_read_sync;
825                 submit_bh(READ, bh[i]);
826                 mb_debug(1, "read bitmap for group %u\n", first_group + i);
827         }
828
829         /* wait for I/O completion */
830         for (i = 0; i < groups_per_page && bh[i]; i++)
831                 wait_on_buffer(bh[i]);
832
833         err = -EIO;
834         for (i = 0; i < groups_per_page && bh[i]; i++)
835                 if (!buffer_uptodate(bh[i]))
836                         goto out;
837
838         err = 0;
839         first_block = page->index * blocks_per_page;
840         /* init the page  */
841         memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
842         for (i = 0; i < blocks_per_page; i++) {
843                 int group;
844                 struct ext4_group_info *grinfo;
845
846                 group = (first_block + i) >> 1;
847                 if (group >= ngroups)
848                         break;
849
850                 /*
851                  * data carry information regarding this
852                  * particular group in the format specified
853                  * above
854                  *
855                  */
856                 data = page_address(page) + (i * blocksize);
857                 bitmap = bh[group - first_group]->b_data;
858
859                 /*
860                  * We place the buddy block and bitmap block
861                  * close together
862                  */
863                 if ((first_block + i) & 1) {
864                         /* this is block of buddy */
865                         BUG_ON(incore == NULL);
866                         mb_debug(1, "put buddy for group %u in page %lu/%x\n",
867                                 group, page->index, i * blocksize);
868                         grinfo = ext4_get_group_info(sb, group);
869                         grinfo->bb_fragments = 0;
870                         memset(grinfo->bb_counters, 0,
871                                sizeof(*grinfo->bb_counters) *
872                                 (sb->s_blocksize_bits+2));
873                         /*
874                          * incore got set to the group block bitmap below
875                          */
876                         ext4_lock_group(sb, group);
877                         ext4_mb_generate_buddy(sb, data, incore, group);
878                         ext4_unlock_group(sb, group);
879                         incore = NULL;
880                 } else {
881                         /* this is block of bitmap */
882                         BUG_ON(incore != NULL);
883                         mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
884                                 group, page->index, i * blocksize);
885
886                         /* see comments in ext4_mb_put_pa() */
887                         ext4_lock_group(sb, group);
888                         memcpy(data, bitmap, blocksize);
889
890                         /* mark all preallocated blks used in in-core bitmap */
891                         ext4_mb_generate_from_pa(sb, data, group);
892                         ext4_mb_generate_from_freelist(sb, data, group);
893                         ext4_unlock_group(sb, group);
894
895                         /* set incore so that the buddy information can be
896                          * generated using this
897                          */
898                         incore = data;
899                 }
900         }
901         SetPageUptodate(page);
902
903 out:
904         if (bh) {
905                 for (i = 0; i < groups_per_page && bh[i]; i++)
906                         brelse(bh[i]);
907                 if (bh != &bhs)
908                         kfree(bh);
909         }
910         return err;
911 }
912
913 static noinline_for_stack
914 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
915 {
916
917         int ret = 0;
918         void *bitmap;
919         int blocks_per_page;
920         int block, pnum, poff;
921         int num_grp_locked = 0;
922         struct ext4_group_info *this_grp;
923         struct ext4_sb_info *sbi = EXT4_SB(sb);
924         struct inode *inode = sbi->s_buddy_cache;
925         struct page *page = NULL, *bitmap_page = NULL;
926
927         mb_debug(1, "init group %u\n", group);
928         blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
929         this_grp = ext4_get_group_info(sb, group);
930         /*
931          * This ensures that we don't reinit the buddy cache
932          * page which map to the group from which we are already
933          * allocating. If we are looking at the buddy cache we would
934          * have taken a reference using ext4_mb_load_buddy and that
935          * would have taken the alloc_sem lock.
936          */
937         num_grp_locked =  ext4_mb_get_buddy_cache_lock(sb, group);
938         if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
939                 /*
940                  * somebody initialized the group
941                  * return without doing anything
942                  */
943                 ret = 0;
944                 goto err;
945         }
946         /*
947          * the buddy cache inode stores the block bitmap
948          * and buddy information in consecutive blocks.
949          * So for each group we need two blocks.
950          */
951         block = group * 2;
952         pnum = block / blocks_per_page;
953         poff = block % blocks_per_page;
954         page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
955         if (page) {
956                 BUG_ON(page->mapping != inode->i_mapping);
957                 ret = ext4_mb_init_cache(page, NULL);
958                 if (ret) {
959                         unlock_page(page);
960                         goto err;
961                 }
962                 unlock_page(page);
963         }
964         if (page == NULL || !PageUptodate(page)) {
965                 ret = -EIO;
966                 goto err;
967         }
968         mark_page_accessed(page);
969         bitmap_page = page;
970         bitmap = page_address(page) + (poff * sb->s_blocksize);
971
972         /* init buddy cache */
973         block++;
974         pnum = block / blocks_per_page;
975         poff = block % blocks_per_page;
976         page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
977         if (page == bitmap_page) {
978                 /*
979                  * If both the bitmap and buddy are in
980                  * the same page we don't need to force
981                  * init the buddy
982                  */
983                 unlock_page(page);
984         } else if (page) {
985                 BUG_ON(page->mapping != inode->i_mapping);
986                 ret = ext4_mb_init_cache(page, bitmap);
987                 if (ret) {
988                         unlock_page(page);
989                         goto err;
990                 }
991                 unlock_page(page);
992         }
993         if (page == NULL || !PageUptodate(page)) {
994                 ret = -EIO;
995                 goto err;
996         }
997         mark_page_accessed(page);
998 err:
999         ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
1000         if (bitmap_page)
1001                 page_cache_release(bitmap_page);
1002         if (page)
1003                 page_cache_release(page);
1004         return ret;
1005 }
1006
1007 static noinline_for_stack int
1008 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1009                                         struct ext4_buddy *e4b)
1010 {
1011         int blocks_per_page;
1012         int block;
1013         int pnum;
1014         int poff;
1015         struct page *page;
1016         int ret;
1017         struct ext4_group_info *grp;
1018         struct ext4_sb_info *sbi = EXT4_SB(sb);
1019         struct inode *inode = sbi->s_buddy_cache;
1020
1021         mb_debug(1, "load group %u\n", group);
1022
1023         blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1024         grp = ext4_get_group_info(sb, group);
1025
1026         e4b->bd_blkbits = sb->s_blocksize_bits;
1027         e4b->bd_info = ext4_get_group_info(sb, group);
1028         e4b->bd_sb = sb;
1029         e4b->bd_group = group;
1030         e4b->bd_buddy_page = NULL;
1031         e4b->bd_bitmap_page = NULL;
1032         e4b->alloc_semp = &grp->alloc_sem;
1033
1034         /* Take the read lock on the group alloc
1035          * sem. This would make sure a parallel
1036          * ext4_mb_init_group happening on other
1037          * groups mapped by the page is blocked
1038          * till we are done with allocation
1039          */
1040 repeat_load_buddy:
1041         down_read(e4b->alloc_semp);
1042
1043         if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1044                 /* we need to check for group need init flag
1045                  * with alloc_semp held so that we can be sure
1046                  * that new blocks didn't get added to the group
1047                  * when we are loading the buddy cache
1048                  */
1049                 up_read(e4b->alloc_semp);
1050                 /*
1051                  * we need full data about the group
1052                  * to make a good selection
1053                  */
1054                 ret = ext4_mb_init_group(sb, group);
1055                 if (ret)
1056                         return ret;
1057                 goto repeat_load_buddy;
1058         }
1059
1060         /*
1061          * the buddy cache inode stores the block bitmap
1062          * and buddy information in consecutive blocks.
1063          * So for each group we need two blocks.
1064          */
1065         block = group * 2;
1066         pnum = block / blocks_per_page;
1067         poff = block % blocks_per_page;
1068
1069         /* we could use find_or_create_page(), but it locks page
1070          * what we'd like to avoid in fast path ... */
1071         page = find_get_page(inode->i_mapping, pnum);
1072         if (page == NULL || !PageUptodate(page)) {
1073                 if (page)
1074                         /*
1075                          * drop the page reference and try
1076                          * to get the page with lock. If we
1077                          * are not uptodate that implies
1078                          * somebody just created the page but
1079                          * is yet to initialize the same. So
1080                          * wait for it to initialize.
1081                          */
1082                         page_cache_release(page);
1083                 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1084                 if (page) {
1085                         BUG_ON(page->mapping != inode->i_mapping);
1086                         if (!PageUptodate(page)) {
1087                                 ret = ext4_mb_init_cache(page, NULL);
1088                                 if (ret) {
1089                                         unlock_page(page);
1090                                         goto err;
1091                                 }
1092                                 mb_cmp_bitmaps(e4b, page_address(page) +
1093                                                (poff * sb->s_blocksize));
1094                         }
1095                         unlock_page(page);
1096                 }
1097         }
1098         if (page == NULL || !PageUptodate(page)) {
1099                 ret = -EIO;
1100                 goto err;
1101         }
1102         e4b->bd_bitmap_page = page;
1103         e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1104         mark_page_accessed(page);
1105
1106         block++;
1107         pnum = block / blocks_per_page;
1108         poff = block % blocks_per_page;
1109
1110         page = find_get_page(inode->i_mapping, pnum);
1111         if (page == NULL || !PageUptodate(page)) {
1112                 if (page)
1113                         page_cache_release(page);
1114                 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1115                 if (page) {
1116                         BUG_ON(page->mapping != inode->i_mapping);
1117                         if (!PageUptodate(page)) {
1118                                 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1119                                 if (ret) {
1120                                         unlock_page(page);
1121                                         goto err;
1122                                 }
1123                         }
1124                         unlock_page(page);
1125                 }
1126         }
1127         if (page == NULL || !PageUptodate(page)) {
1128                 ret = -EIO;
1129                 goto err;
1130         }
1131         e4b->bd_buddy_page = page;
1132         e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1133         mark_page_accessed(page);
1134
1135         BUG_ON(e4b->bd_bitmap_page == NULL);
1136         BUG_ON(e4b->bd_buddy_page == NULL);
1137
1138         return 0;
1139
1140 err:
1141         if (e4b->bd_bitmap_page)
1142                 page_cache_release(e4b->bd_bitmap_page);
1143         if (e4b->bd_buddy_page)
1144                 page_cache_release(e4b->bd_buddy_page);
1145         e4b->bd_buddy = NULL;
1146         e4b->bd_bitmap = NULL;
1147
1148         /* Done with the buddy cache */
1149         up_read(e4b->alloc_semp);
1150         return ret;
1151 }
1152
1153 static void ext4_mb_release_desc(struct ext4_buddy *e4b)
1154 {
1155         if (e4b->bd_bitmap_page)
1156                 page_cache_release(e4b->bd_bitmap_page);
1157         if (e4b->bd_buddy_page)
1158                 page_cache_release(e4b->bd_buddy_page);
1159         /* Done with the buddy cache */
1160         if (e4b->alloc_semp)
1161                 up_read(e4b->alloc_semp);
1162 }
1163
1164
1165 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1166 {
1167         int order = 1;
1168         void *bb;
1169
1170         BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1171         BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1172
1173         bb = EXT4_MB_BUDDY(e4b);
1174         while (order <= e4b->bd_blkbits + 1) {
1175                 block = block >> 1;
1176                 if (!mb_test_bit(block, bb)) {
1177                         /* this block is part of buddy of order 'order' */
1178                         return order;
1179                 }
1180                 bb += 1 << (e4b->bd_blkbits - order);
1181                 order++;
1182         }
1183         return 0;
1184 }
1185
1186 static void mb_clear_bits(void *bm, int cur, int len)
1187 {
1188         __u32 *addr;
1189
1190         len = cur + len;
1191         while (cur < len) {
1192                 if ((cur & 31) == 0 && (len - cur) >= 32) {
1193                         /* fast path: clear whole word at once */
1194                         addr = bm + (cur >> 3);
1195                         *addr = 0;
1196                         cur += 32;
1197                         continue;
1198                 }
1199                 mb_clear_bit(cur, bm);
1200                 cur++;
1201         }
1202 }
1203
1204 static void mb_set_bits(void *bm, int cur, int len)
1205 {
1206         __u32 *addr;
1207
1208         len = cur + len;
1209         while (cur < len) {
1210                 if ((cur & 31) == 0 && (len - cur) >= 32) {
1211                         /* fast path: set whole word at once */
1212                         addr = bm + (cur >> 3);
1213                         *addr = 0xffffffff;
1214                         cur += 32;
1215                         continue;
1216                 }
1217                 mb_set_bit(cur, bm);
1218                 cur++;
1219         }
1220 }
1221
1222 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1223                           int first, int count)
1224 {
1225         int block = 0;
1226         int max = 0;
1227         int order;
1228         void *buddy;
1229         void *buddy2;
1230         struct super_block *sb = e4b->bd_sb;
1231
1232         BUG_ON(first + count > (sb->s_blocksize << 3));
1233         assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1234         mb_check_buddy(e4b);
1235         mb_free_blocks_double(inode, e4b, first, count);
1236
1237         e4b->bd_info->bb_free += count;
1238         if (first < e4b->bd_info->bb_first_free)
1239                 e4b->bd_info->bb_first_free = first;
1240
1241         /* let's maintain fragments counter */
1242         if (first != 0)
1243                 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1244         if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1245                 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1246         if (block && max)
1247                 e4b->bd_info->bb_fragments--;
1248         else if (!block && !max)
1249                 e4b->bd_info->bb_fragments++;
1250
1251         /* let's maintain buddy itself */
1252         while (count-- > 0) {
1253                 block = first++;
1254                 order = 0;
1255
1256                 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1257                         ext4_fsblk_t blocknr;
1258                         blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1259                         blocknr += block;
1260                         blocknr +=
1261                             le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
1262                         ext4_grp_locked_error(sb, e4b->bd_group,
1263                                    __func__, "double-free of inode"
1264                                    " %lu's block %llu(bit %u in group %u)",
1265                                    inode ? inode->i_ino : 0, blocknr, block,
1266                                    e4b->bd_group);
1267                 }
1268                 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1269                 e4b->bd_info->bb_counters[order]++;
1270
1271                 /* start of the buddy */
1272                 buddy = mb_find_buddy(e4b, order, &max);
1273
1274                 do {
1275                         block &= ~1UL;
1276                         if (mb_test_bit(block, buddy) ||
1277                                         mb_test_bit(block + 1, buddy))
1278                                 break;
1279
1280                         /* both the buddies are free, try to coalesce them */
1281                         buddy2 = mb_find_buddy(e4b, order + 1, &max);
1282
1283                         if (!buddy2)
1284                                 break;
1285
1286                         if (order > 0) {
1287                                 /* for special purposes, we don't set
1288                                  * free bits in bitmap */
1289                                 mb_set_bit(block, buddy);
1290                                 mb_set_bit(block + 1, buddy);
1291                         }
1292                         e4b->bd_info->bb_counters[order]--;
1293                         e4b->bd_info->bb_counters[order]--;
1294
1295                         block = block >> 1;
1296                         order++;
1297                         e4b->bd_info->bb_counters[order]++;
1298
1299                         mb_clear_bit(block, buddy2);
1300                         buddy = buddy2;
1301                 } while (1);
1302         }
1303         mb_check_buddy(e4b);
1304 }
1305
1306 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1307                                 int needed, struct ext4_free_extent *ex)
1308 {
1309         int next = block;
1310         int max;
1311         int ord;
1312         void *buddy;
1313
1314         assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1315         BUG_ON(ex == NULL);
1316
1317         buddy = mb_find_buddy(e4b, order, &max);
1318         BUG_ON(buddy == NULL);
1319         BUG_ON(block >= max);
1320         if (mb_test_bit(block, buddy)) {
1321                 ex->fe_len = 0;
1322                 ex->fe_start = 0;
1323                 ex->fe_group = 0;
1324                 return 0;
1325         }
1326
1327         /* FIXME dorp order completely ? */
1328         if (likely(order == 0)) {
1329                 /* find actual order */
1330                 order = mb_find_order_for_block(e4b, block);
1331                 block = block >> order;
1332         }
1333
1334         ex->fe_len = 1 << order;
1335         ex->fe_start = block << order;
1336         ex->fe_group = e4b->bd_group;
1337
1338         /* calc difference from given start */
1339         next = next - ex->fe_start;
1340         ex->fe_len -= next;
1341         ex->fe_start += next;
1342
1343         while (needed > ex->fe_len &&
1344                (buddy = mb_find_buddy(e4b, order, &max))) {
1345
1346                 if (block + 1 >= max)
1347                         break;
1348
1349                 next = (block + 1) * (1 << order);
1350                 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1351                         break;
1352
1353                 ord = mb_find_order_for_block(e4b, next);
1354
1355                 order = ord;
1356                 block = next >> order;
1357                 ex->fe_len += 1 << order;
1358         }
1359
1360         BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1361         return ex->fe_len;
1362 }
1363
1364 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1365 {
1366         int ord;
1367         int mlen = 0;
1368         int max = 0;
1369         int cur;
1370         int start = ex->fe_start;
1371         int len = ex->fe_len;
1372         unsigned ret = 0;
1373         int len0 = len;
1374         void *buddy;
1375
1376         BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1377         BUG_ON(e4b->bd_group != ex->fe_group);
1378         assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1379         mb_check_buddy(e4b);
1380         mb_mark_used_double(e4b, start, len);
1381
1382         e4b->bd_info->bb_free -= len;
1383         if (e4b->bd_info->bb_first_free == start)
1384                 e4b->bd_info->bb_first_free += len;
1385
1386         /* let's maintain fragments counter */
1387         if (start != 0)
1388                 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1389         if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1390                 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1391         if (mlen && max)
1392                 e4b->bd_info->bb_fragments++;
1393         else if (!mlen && !max)
1394                 e4b->bd_info->bb_fragments--;
1395
1396         /* let's maintain buddy itself */
1397         while (len) {
1398                 ord = mb_find_order_for_block(e4b, start);
1399
1400                 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1401                         /* the whole chunk may be allocated at once! */
1402                         mlen = 1 << ord;
1403                         buddy = mb_find_buddy(e4b, ord, &max);
1404                         BUG_ON((start >> ord) >= max);
1405                         mb_set_bit(start >> ord, buddy);
1406                         e4b->bd_info->bb_counters[ord]--;
1407                         start += mlen;
1408                         len -= mlen;
1409                         BUG_ON(len < 0);
1410                         continue;
1411                 }
1412
1413                 /* store for history */
1414                 if (ret == 0)
1415                         ret = len | (ord << 16);
1416
1417                 /* we have to split large buddy */
1418                 BUG_ON(ord <= 0);
1419                 buddy = mb_find_buddy(e4b, ord, &max);
1420                 mb_set_bit(start >> ord, buddy);
1421                 e4b->bd_info->bb_counters[ord]--;
1422
1423                 ord--;
1424                 cur = (start >> ord) & ~1U;
1425                 buddy = mb_find_buddy(e4b, ord, &max);
1426                 mb_clear_bit(cur, buddy);
1427                 mb_clear_bit(cur + 1, buddy);
1428                 e4b->bd_info->bb_counters[ord]++;
1429                 e4b->bd_info->bb_counters[ord]++;
1430         }
1431
1432         mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1433         mb_check_buddy(e4b);
1434
1435         return ret;
1436 }
1437
1438 /*
1439  * Must be called under group lock!
1440  */
1441 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1442                                         struct ext4_buddy *e4b)
1443 {
1444         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1445         int ret;
1446
1447         BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1448         BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1449
1450         ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1451         ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1452         ret = mb_mark_used(e4b, &ac->ac_b_ex);
1453
1454         /* preallocation can change ac_b_ex, thus we store actually
1455          * allocated blocks for history */
1456         ac->ac_f_ex = ac->ac_b_ex;
1457
1458         ac->ac_status = AC_STATUS_FOUND;
1459         ac->ac_tail = ret & 0xffff;
1460         ac->ac_buddy = ret >> 16;
1461
1462         /*
1463          * take the page reference. We want the page to be pinned
1464          * so that we don't get a ext4_mb_init_cache_call for this
1465          * group until we update the bitmap. That would mean we
1466          * double allocate blocks. The reference is dropped
1467          * in ext4_mb_release_context
1468          */
1469         ac->ac_bitmap_page = e4b->bd_bitmap_page;
1470         get_page(ac->ac_bitmap_page);
1471         ac->ac_buddy_page = e4b->bd_buddy_page;
1472         get_page(ac->ac_buddy_page);
1473         /* on allocation we use ac to track the held semaphore */
1474         ac->alloc_semp =  e4b->alloc_semp;
1475         e4b->alloc_semp = NULL;
1476         /* store last allocated for subsequent stream allocation */
1477         if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1478                 spin_lock(&sbi->s_md_lock);
1479                 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1480                 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1481                 spin_unlock(&sbi->s_md_lock);
1482         }
1483 }
1484
1485 /*
1486  * regular allocator, for general purposes allocation
1487  */
1488
1489 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1490                                         struct ext4_buddy *e4b,
1491                                         int finish_group)
1492 {
1493         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1494         struct ext4_free_extent *bex = &ac->ac_b_ex;
1495         struct ext4_free_extent *gex = &ac->ac_g_ex;
1496         struct ext4_free_extent ex;
1497         int max;
1498
1499         if (ac->ac_status == AC_STATUS_FOUND)
1500                 return;
1501         /*
1502          * We don't want to scan for a whole year
1503          */
1504         if (ac->ac_found > sbi->s_mb_max_to_scan &&
1505                         !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1506                 ac->ac_status = AC_STATUS_BREAK;
1507                 return;
1508         }
1509
1510         /*
1511          * Haven't found good chunk so far, let's continue
1512          */
1513         if (bex->fe_len < gex->fe_len)
1514                 return;
1515
1516         if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1517                         && bex->fe_group == e4b->bd_group) {
1518                 /* recheck chunk's availability - we don't know
1519                  * when it was found (within this lock-unlock
1520                  * period or not) */
1521                 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1522                 if (max >= gex->fe_len) {
1523                         ext4_mb_use_best_found(ac, e4b);
1524                         return;
1525                 }
1526         }
1527 }
1528
1529 /*
1530  * The routine checks whether found extent is good enough. If it is,
1531  * then the extent gets marked used and flag is set to the context
1532  * to stop scanning. Otherwise, the extent is compared with the
1533  * previous found extent and if new one is better, then it's stored
1534  * in the context. Later, the best found extent will be used, if
1535  * mballoc can't find good enough extent.
1536  *
1537  * FIXME: real allocation policy is to be designed yet!
1538  */
1539 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1540                                         struct ext4_free_extent *ex,
1541                                         struct ext4_buddy *e4b)
1542 {
1543         struct ext4_free_extent *bex = &ac->ac_b_ex;
1544         struct ext4_free_extent *gex = &ac->ac_g_ex;
1545
1546         BUG_ON(ex->fe_len <= 0);
1547         BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1548         BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1549         BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1550
1551         ac->ac_found++;
1552
1553         /*
1554          * The special case - take what you catch first
1555          */
1556         if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1557                 *bex = *ex;
1558                 ext4_mb_use_best_found(ac, e4b);
1559                 return;
1560         }
1561
1562         /*
1563          * Let's check whether the chuck is good enough
1564          */
1565         if (ex->fe_len == gex->fe_len) {
1566                 *bex = *ex;
1567                 ext4_mb_use_best_found(ac, e4b);
1568                 return;
1569         }
1570
1571         /*
1572          * If this is first found extent, just store it in the context
1573          */
1574         if (bex->fe_len == 0) {
1575                 *bex = *ex;
1576                 return;
1577         }
1578
1579         /*
1580          * If new found extent is better, store it in the context
1581          */
1582         if (bex->fe_len < gex->fe_len) {
1583                 /* if the request isn't satisfied, any found extent
1584                  * larger than previous best one is better */
1585                 if (ex->fe_len > bex->fe_len)
1586                         *bex = *ex;
1587         } else if (ex->fe_len > gex->fe_len) {
1588                 /* if the request is satisfied, then we try to find
1589                  * an extent that still satisfy the request, but is
1590                  * smaller than previous one */
1591                 if (ex->fe_len < bex->fe_len)
1592                         *bex = *ex;
1593         }
1594
1595         ext4_mb_check_limits(ac, e4b, 0);
1596 }
1597
1598 static noinline_for_stack
1599 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1600                                         struct ext4_buddy *e4b)
1601 {
1602         struct ext4_free_extent ex = ac->ac_b_ex;
1603         ext4_group_t group = ex.fe_group;
1604         int max;
1605         int err;
1606
1607         BUG_ON(ex.fe_len <= 0);
1608         err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1609         if (err)
1610                 return err;
1611
1612         ext4_lock_group(ac->ac_sb, group);
1613         max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1614
1615         if (max > 0) {
1616                 ac->ac_b_ex = ex;
1617                 ext4_mb_use_best_found(ac, e4b);
1618         }
1619
1620         ext4_unlock_group(ac->ac_sb, group);
1621         ext4_mb_release_desc(e4b);
1622
1623         return 0;
1624 }
1625
1626 static noinline_for_stack
1627 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1628                                 struct ext4_buddy *e4b)
1629 {
1630         ext4_group_t group = ac->ac_g_ex.fe_group;
1631         int max;
1632         int err;
1633         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1634         struct ext4_super_block *es = sbi->s_es;
1635         struct ext4_free_extent ex;
1636
1637         if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1638                 return 0;
1639
1640         err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1641         if (err)
1642                 return err;
1643
1644         ext4_lock_group(ac->ac_sb, group);
1645         max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1646                              ac->ac_g_ex.fe_len, &ex);
1647
1648         if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1649                 ext4_fsblk_t start;
1650
1651                 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1652                         ex.fe_start + le32_to_cpu(es->s_first_data_block);
1653                 /* use do_div to get remainder (would be 64-bit modulo) */
1654                 if (do_div(start, sbi->s_stripe) == 0) {
1655                         ac->ac_found++;
1656                         ac->ac_b_ex = ex;
1657                         ext4_mb_use_best_found(ac, e4b);
1658                 }
1659         } else if (max >= ac->ac_g_ex.fe_len) {
1660                 BUG_ON(ex.fe_len <= 0);
1661                 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1662                 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1663                 ac->ac_found++;
1664                 ac->ac_b_ex = ex;
1665                 ext4_mb_use_best_found(ac, e4b);
1666         } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1667                 /* Sometimes, caller may want to merge even small
1668                  * number of blocks to an existing extent */
1669                 BUG_ON(ex.fe_len <= 0);
1670                 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1671                 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1672                 ac->ac_found++;
1673                 ac->ac_b_ex = ex;
1674                 ext4_mb_use_best_found(ac, e4b);
1675         }
1676         ext4_unlock_group(ac->ac_sb, group);
1677         ext4_mb_release_desc(e4b);
1678
1679         return 0;
1680 }
1681
1682 /*
1683  * The routine scans buddy structures (not bitmap!) from given order
1684  * to max order and tries to find big enough chunk to satisfy the req
1685  */
1686 static noinline_for_stack
1687 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1688                                         struct ext4_buddy *e4b)
1689 {
1690         struct super_block *sb = ac->ac_sb;
1691         struct ext4_group_info *grp = e4b->bd_info;
1692         void *buddy;
1693         int i;
1694         int k;
1695         int max;
1696
1697         BUG_ON(ac->ac_2order <= 0);
1698         for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1699                 if (grp->bb_counters[i] == 0)
1700                         continue;
1701
1702                 buddy = mb_find_buddy(e4b, i, &max);
1703                 BUG_ON(buddy == NULL);
1704
1705                 k = mb_find_next_zero_bit(buddy, max, 0);
1706                 BUG_ON(k >= max);
1707
1708                 ac->ac_found++;
1709
1710                 ac->ac_b_ex.fe_len = 1 << i;
1711                 ac->ac_b_ex.fe_start = k << i;
1712                 ac->ac_b_ex.fe_group = e4b->bd_group;
1713
1714                 ext4_mb_use_best_found(ac, e4b);
1715
1716                 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1717
1718                 if (EXT4_SB(sb)->s_mb_stats)
1719                         atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1720
1721                 break;
1722         }
1723 }
1724
1725 /*
1726  * The routine scans the group and measures all found extents.
1727  * In order to optimize scanning, caller must pass number of
1728  * free blocks in the group, so the routine can know upper limit.
1729  */
1730 static noinline_for_stack
1731 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1732                                         struct ext4_buddy *e4b)
1733 {
1734         struct super_block *sb = ac->ac_sb;
1735         void *bitmap = EXT4_MB_BITMAP(e4b);
1736         struct ext4_free_extent ex;
1737         int i;
1738         int free;
1739
1740         free = e4b->bd_info->bb_free;
1741         BUG_ON(free <= 0);
1742
1743         i = e4b->bd_info->bb_first_free;
1744
1745         while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1746                 i = mb_find_next_zero_bit(bitmap,
1747                                                 EXT4_BLOCKS_PER_GROUP(sb), i);
1748                 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1749                         /*
1750                          * IF we have corrupt bitmap, we won't find any
1751                          * free blocks even though group info says we
1752                          * we have free blocks
1753                          */
1754                         ext4_grp_locked_error(sb, e4b->bd_group,
1755                                         __func__, "%d free blocks as per "
1756                                         "group info. But bitmap says 0",
1757                                         free);
1758                         break;
1759                 }
1760
1761                 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1762                 BUG_ON(ex.fe_len <= 0);
1763                 if (free < ex.fe_len) {
1764                         ext4_grp_locked_error(sb, e4b->bd_group,
1765                                         __func__, "%d free blocks as per "
1766                                         "group info. But got %d blocks",
1767                                         free, ex.fe_len);
1768                         /*
1769                          * The number of free blocks differs. This mostly
1770                          * indicate that the bitmap is corrupt. So exit
1771                          * without claiming the space.
1772                          */
1773                         break;
1774                 }
1775
1776                 ext4_mb_measure_extent(ac, &ex, e4b);
1777
1778                 i += ex.fe_len;
1779                 free -= ex.fe_len;
1780         }
1781
1782         ext4_mb_check_limits(ac, e4b, 1);
1783 }
1784
1785 /*
1786  * This is a special case for storages like raid5
1787  * we try to find stripe-aligned chunks for stripe-size requests
1788  * XXX should do so at least for multiples of stripe size as well
1789  */
1790 static noinline_for_stack
1791 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1792                                  struct ext4_buddy *e4b)
1793 {
1794         struct super_block *sb = ac->ac_sb;
1795         struct ext4_sb_info *sbi = EXT4_SB(sb);
1796         void *bitmap = EXT4_MB_BITMAP(e4b);
1797         struct ext4_free_extent ex;
1798         ext4_fsblk_t first_group_block;
1799         ext4_fsblk_t a;
1800         ext4_grpblk_t i;
1801         int max;
1802
1803         BUG_ON(sbi->s_stripe == 0);
1804
1805         /* find first stripe-aligned block in group */
1806         first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1807                 + le32_to_cpu(sbi->s_es->s_first_data_block);
1808         a = first_group_block + sbi->s_stripe - 1;
1809         do_div(a, sbi->s_stripe);
1810         i = (a * sbi->s_stripe) - first_group_block;
1811
1812         while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1813                 if (!mb_test_bit(i, bitmap)) {
1814                         max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1815                         if (max >= sbi->s_stripe) {
1816                                 ac->ac_found++;
1817                                 ac->ac_b_ex = ex;
1818                                 ext4_mb_use_best_found(ac, e4b);
1819                                 break;
1820                         }
1821                 }
1822                 i += sbi->s_stripe;
1823         }
1824 }
1825
1826 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1827                                 ext4_group_t group, int cr)
1828 {
1829         unsigned free, fragments;
1830         unsigned i, bits;
1831         int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1832         struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1833
1834         BUG_ON(cr < 0 || cr >= 4);
1835         BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1836
1837         free = grp->bb_free;
1838         fragments = grp->bb_fragments;
1839         if (free == 0)
1840                 return 0;
1841         if (fragments == 0)
1842                 return 0;
1843
1844         switch (cr) {
1845         case 0:
1846                 BUG_ON(ac->ac_2order == 0);
1847
1848                 /* Avoid using the first bg of a flexgroup for data files */
1849                 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1850                     (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1851                     ((group % flex_size) == 0))
1852                         return 0;
1853
1854                 bits = ac->ac_sb->s_blocksize_bits + 1;
1855                 for (i = ac->ac_2order; i <= bits; i++)
1856                         if (grp->bb_counters[i] > 0)
1857                                 return 1;
1858                 break;
1859         case 1:
1860                 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1861                         return 1;
1862                 break;
1863         case 2:
1864                 if (free >= ac->ac_g_ex.fe_len)
1865                         return 1;
1866                 break;
1867         case 3:
1868                 return 1;
1869         default:
1870                 BUG();
1871         }
1872
1873         return 0;
1874 }
1875
1876 /*
1877  * lock the group_info alloc_sem of all the groups
1878  * belonging to the same buddy cache page. This
1879  * make sure other parallel operation on the buddy
1880  * cache doesn't happen  whild holding the buddy cache
1881  * lock
1882  */
1883 int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1884 {
1885         int i;
1886         int block, pnum;
1887         int blocks_per_page;
1888         int groups_per_page;
1889         ext4_group_t ngroups = ext4_get_groups_count(sb);
1890         ext4_group_t first_group;
1891         struct ext4_group_info *grp;
1892
1893         blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1894         /*
1895          * the buddy cache inode stores the block bitmap
1896          * and buddy information in consecutive blocks.
1897          * So for each group we need two blocks.
1898          */
1899         block = group * 2;
1900         pnum = block / blocks_per_page;
1901         first_group = pnum * blocks_per_page / 2;
1902
1903         groups_per_page = blocks_per_page >> 1;
1904         if (groups_per_page == 0)
1905                 groups_per_page = 1;
1906         /* read all groups the page covers into the cache */
1907         for (i = 0; i < groups_per_page; i++) {
1908
1909                 if ((first_group + i) >= ngroups)
1910                         break;
1911                 grp = ext4_get_group_info(sb, first_group + i);
1912                 /* take all groups write allocation
1913                  * semaphore. This make sure there is
1914                  * no block allocation going on in any
1915                  * of that groups
1916                  */
1917                 down_write_nested(&grp->alloc_sem, i);
1918         }
1919         return i;
1920 }
1921
1922 void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1923                                         ext4_group_t group, int locked_group)
1924 {
1925         int i;
1926         int block, pnum;
1927         int blocks_per_page;
1928         ext4_group_t first_group;
1929         struct ext4_group_info *grp;
1930
1931         blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1932         /*
1933          * the buddy cache inode stores the block bitmap
1934          * and buddy information in consecutive blocks.
1935          * So for each group we need two blocks.
1936          */
1937         block = group * 2;
1938         pnum = block / blocks_per_page;
1939         first_group = pnum * blocks_per_page / 2;
1940         /* release locks on all the groups */
1941         for (i = 0; i < locked_group; i++) {
1942
1943                 grp = ext4_get_group_info(sb, first_group + i);
1944                 /* take all groups write allocation
1945                  * semaphore. This make sure there is
1946                  * no block allocation going on in any
1947                  * of that groups
1948                  */
1949                 up_write(&grp->alloc_sem);
1950         }
1951
1952 }
1953
1954 static noinline_for_stack int
1955 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1956 {
1957         ext4_group_t ngroups, group, i;
1958         int cr;
1959         int err = 0;
1960         int bsbits;
1961         struct ext4_sb_info *sbi;
1962         struct super_block *sb;
1963         struct ext4_buddy e4b;
1964
1965         sb = ac->ac_sb;
1966         sbi = EXT4_SB(sb);
1967         ngroups = ext4_get_groups_count(sb);
1968         /* non-extent files are limited to low blocks/groups */
1969         if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL))
1970                 ngroups = sbi->s_blockfile_groups;
1971
1972         BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1973
1974         /* first, try the goal */
1975         err = ext4_mb_find_by_goal(ac, &e4b);
1976         if (err || ac->ac_status == AC_STATUS_FOUND)
1977                 goto out;
1978
1979         if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1980                 goto out;
1981
1982         /*
1983          * ac->ac2_order is set only if the fe_len is a power of 2
1984          * if ac2_order is set we also set criteria to 0 so that we
1985          * try exact allocation using buddy.
1986          */
1987         i = fls(ac->ac_g_ex.fe_len);
1988         ac->ac_2order = 0;
1989         /*
1990          * We search using buddy data only if the order of the request
1991          * is greater than equal to the sbi_s_mb_order2_reqs
1992          * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1993          */
1994         if (i >= sbi->s_mb_order2_reqs) {
1995                 /*
1996                  * This should tell if fe_len is exactly power of 2
1997                  */
1998                 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1999                         ac->ac_2order = i - 1;
2000         }
2001
2002         bsbits = ac->ac_sb->s_blocksize_bits;
2003
2004         /* if stream allocation is enabled, use global goal */
2005         if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2006                 /* TBD: may be hot point */
2007                 spin_lock(&sbi->s_md_lock);
2008                 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2009                 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2010                 spin_unlock(&sbi->s_md_lock);
2011         }
2012
2013         /* Let's just scan groups to find more-less suitable blocks */
2014         cr = ac->ac_2order ? 0 : 1;
2015         /*
2016          * cr == 0 try to get exact allocation,
2017          * cr == 3  try to get anything
2018          */
2019 repeat:
2020         for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2021                 ac->ac_criteria = cr;
2022                 /*
2023                  * searching for the right group start
2024                  * from the goal value specified
2025                  */
2026                 group = ac->ac_g_ex.fe_group;
2027
2028                 for (i = 0; i < ngroups; group++, i++) {
2029                         struct ext4_group_info *grp;
2030                         struct ext4_group_desc *desc;
2031
2032                         if (group == ngroups)
2033                                 group = 0;
2034
2035                         /* quick check to skip empty groups */
2036                         grp = ext4_get_group_info(sb, group);
2037                         if (grp->bb_free == 0)
2038                                 continue;
2039
2040                         err = ext4_mb_load_buddy(sb, group, &e4b);
2041                         if (err)
2042                                 goto out;
2043
2044                         ext4_lock_group(sb, group);
2045                         if (!ext4_mb_good_group(ac, group, cr)) {
2046                                 /* someone did allocation from this group */
2047                                 ext4_unlock_group(sb, group);
2048                                 ext4_mb_release_desc(&e4b);
2049                                 continue;
2050                         }
2051
2052                         ac->ac_groups_scanned++;
2053                         desc = ext4_get_group_desc(sb, group, NULL);
2054                         if (cr == 0)
2055                                 ext4_mb_simple_scan_group(ac, &e4b);
2056                         else if (cr == 1 &&
2057                                         ac->ac_g_ex.fe_len == sbi->s_stripe)
2058                                 ext4_mb_scan_aligned(ac, &e4b);
2059                         else
2060                                 ext4_mb_complex_scan_group(ac, &e4b);
2061
2062                         ext4_unlock_group(sb, group);
2063                         ext4_mb_release_desc(&e4b);
2064
2065                         if (ac->ac_status != AC_STATUS_CONTINUE)
2066                                 break;
2067                 }
2068         }
2069
2070         if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2071             !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2072                 /*
2073                  * We've been searching too long. Let's try to allocate
2074                  * the best chunk we've found so far
2075                  */
2076
2077                 ext4_mb_try_best_found(ac, &e4b);
2078                 if (ac->ac_status != AC_STATUS_FOUND) {
2079                         /*
2080                          * Someone more lucky has already allocated it.
2081                          * The only thing we can do is just take first
2082                          * found block(s)
2083                         printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2084                          */
2085                         ac->ac_b_ex.fe_group = 0;
2086                         ac->ac_b_ex.fe_start = 0;
2087                         ac->ac_b_ex.fe_len = 0;
2088                         ac->ac_status = AC_STATUS_CONTINUE;
2089                         ac->ac_flags |= EXT4_MB_HINT_FIRST;
2090                         cr = 3;
2091                         atomic_inc(&sbi->s_mb_lost_chunks);
2092                         goto repeat;
2093                 }
2094         }
2095 out:
2096         return err;
2097 }
2098
2099 #ifdef EXT4_MB_HISTORY
2100 struct ext4_mb_proc_session {
2101         struct ext4_mb_history *history;
2102         struct super_block *sb;
2103         int start;
2104         int max;
2105 };
2106
2107 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
2108                                         struct ext4_mb_history *hs,
2109                                         int first)
2110 {
2111         if (hs == s->history + s->max)
2112                 hs = s->history;
2113         if (!first && hs == s->history + s->start)
2114                 return NULL;
2115         while (hs->orig.fe_len == 0) {
2116                 hs++;
2117                 if (hs == s->history + s->max)
2118                         hs = s->history;
2119                 if (hs == s->history + s->start)
2120                         return NULL;
2121         }
2122         return hs;
2123 }
2124
2125 static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
2126 {
2127         struct ext4_mb_proc_session *s = seq->private;
2128         struct ext4_mb_history *hs;
2129         int l = *pos;
2130
2131         if (l == 0)
2132                 return SEQ_START_TOKEN;
2133         hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2134         if (!hs)
2135                 return NULL;
2136         while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
2137         return hs;
2138 }
2139
2140 static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
2141                                       loff_t *pos)
2142 {
2143         struct ext4_mb_proc_session *s = seq->private;
2144         struct ext4_mb_history *hs = v;
2145
2146         ++*pos;
2147         if (v == SEQ_START_TOKEN)
2148                 return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2149         else
2150                 return ext4_mb_history_skip_empty(s, ++hs, 0);
2151 }
2152
2153 static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
2154 {
2155         char buf[25], buf2[25], buf3[25], *fmt;
2156         struct ext4_mb_history *hs = v;
2157
2158         if (v == SEQ_START_TOKEN) {
2159                 seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
2160                                 "%-5s %-2s %-6s %-5s %-5s %-6s\n",
2161                           "pid", "inode", "original", "goal", "result", "found",
2162                            "grps", "cr", "flags", "merge", "tail", "broken");
2163                 return 0;
2164         }
2165
2166         if (hs->op == EXT4_MB_HISTORY_ALLOC) {
2167                 fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2168                         "0x%04x %-5s %-5u %-6u\n";
2169                 sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group,
2170                         hs->result.fe_start, hs->result.fe_len,
2171                         hs->result.fe_logical);
2172                 sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group,
2173                         hs->orig.fe_start, hs->orig.fe_len,
2174                         hs->orig.fe_logical);
2175                 sprintf(buf3, "%u/%d/%u@%u", hs->goal.fe_group,
2176                         hs->goal.fe_start, hs->goal.fe_len,
2177                         hs->goal.fe_logical);
2178                 seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
2179                                 hs->found, hs->groups, hs->cr, hs->flags,
2180                                 hs->merged ? "M" : "", hs->tail,
2181                                 hs->buddy ? 1 << hs->buddy : 0);
2182         } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
2183                 fmt = "%-5u %-8u %-23s %-23s %-23s\n";
2184                 sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group,
2185                         hs->result.fe_start, hs->result.fe_len,
2186                         hs->result.fe_logical);
2187                 sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group,
2188                         hs->orig.fe_start, hs->orig.fe_len,
2189                         hs->orig.fe_logical);
2190                 seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
2191         } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
2192                 sprintf(buf2, "%u/%d/%u", hs->result.fe_group,
2193                         hs->result.fe_start, hs->result.fe_len);
2194                 seq_printf(seq, "%-5u %-8u %-23s discard\n",
2195                                 hs->pid, hs->ino, buf2);
2196         } else if (hs->op == EXT4_MB_HISTORY_FREE) {
2197                 sprintf(buf2, "%u/%d/%u", hs->result.fe_group,
2198                         hs->result.fe_start, hs->result.fe_len);
2199                 seq_printf(seq, "%-5u %-8u %-23s free\n",
2200                                 hs->pid, hs->ino, buf2);
2201         }
2202         return 0;
2203 }
2204
2205 static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
2206 {
2207 }
2208
2209 static const struct seq_operations ext4_mb_seq_history_ops = {
2210         .start  = ext4_mb_seq_history_start,
2211         .next   = ext4_mb_seq_history_next,
2212         .stop   = ext4_mb_seq_history_stop,
2213         .show   = ext4_mb_seq_history_show,
2214 };
2215
2216 static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
2217 {
2218         struct super_block *sb = PDE(inode)->data;
2219         struct ext4_sb_info *sbi = EXT4_SB(sb);
2220         struct ext4_mb_proc_session *s;
2221         int rc;
2222         int size;
2223
2224         if (unlikely(sbi->s_mb_history == NULL))
2225                 return -ENOMEM;
2226         s = kmalloc(sizeof(*s), GFP_KERNEL);
2227         if (s == NULL)
2228                 return -ENOMEM;
2229         s->sb = sb;
2230         size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
2231         s->history = kmalloc(size, GFP_KERNEL);
2232         if (s->history == NULL) {
2233                 kfree(s);
2234                 return -ENOMEM;
2235         }
2236
2237         spin_lock(&sbi->s_mb_history_lock);
2238         memcpy(s->history, sbi->s_mb_history, size);
2239         s->max = sbi->s_mb_history_max;
2240         s->start = sbi->s_mb_history_cur % s->max;
2241         spin_unlock(&sbi->s_mb_history_lock);
2242
2243         rc = seq_open(file, &ext4_mb_seq_history_ops);
2244         if (rc == 0) {
2245                 struct seq_file *m = (struct seq_file *)file->private_data;
2246                 m->private = s;
2247         } else {
2248                 kfree(s->history);
2249                 kfree(s);
2250         }
2251         return rc;
2252
2253 }
2254
2255 static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
2256 {
2257         struct seq_file *seq = (struct seq_file *)file->private_data;
2258         struct ext4_mb_proc_session *s = seq->private;
2259         kfree(s->history);
2260         kfree(s);
2261         return seq_release(inode, file);
2262 }
2263
2264 static ssize_t ext4_mb_seq_history_write(struct file *file,
2265                                 const char __user *buffer,
2266                                 size_t count, loff_t *ppos)
2267 {
2268         struct seq_file *seq = (struct seq_file *)file->private_data;
2269         struct ext4_mb_proc_session *s = seq->private;
2270         struct super_block *sb = s->sb;
2271         char str[32];
2272         int value;
2273
2274         if (count >= sizeof(str)) {
2275                 printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
2276                                 "mb_history", (int)sizeof(str));
2277                 return -EOVERFLOW;
2278         }
2279
2280         if (copy_from_user(str, buffer, count))
2281                 return -EFAULT;
2282
2283         value = simple_strtol(str, NULL, 0);
2284         if (value < 0)
2285                 return -ERANGE;
2286         EXT4_SB(sb)->s_mb_history_filter = value;
2287
2288         return count;
2289 }
2290
2291 static const struct file_operations ext4_mb_seq_history_fops = {
2292         .owner          = THIS_MODULE,
2293         .open           = ext4_mb_seq_history_open,
2294         .read           = seq_read,
2295         .write          = ext4_mb_seq_history_write,
2296         .llseek         = seq_lseek,
2297         .release        = ext4_mb_seq_history_release,
2298 };
2299
2300 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2301 {
2302         struct super_block *sb = seq->private;
2303         ext4_group_t group;
2304
2305         if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2306                 return NULL;
2307         group = *pos + 1;
2308         return (void *) ((unsigned long) group);
2309 }
2310
2311 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2312 {
2313         struct super_block *sb = seq->private;
2314         ext4_group_t group;
2315
2316         ++*pos;
2317         if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2318                 return NULL;
2319         group = *pos + 1;
2320         return (void *) ((unsigned long) group);
2321 }
2322
2323 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2324 {
2325         struct super_block *sb = seq->private;
2326         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2327         int i;
2328         int err;
2329         struct ext4_buddy e4b;
2330         struct sg {
2331                 struct ext4_group_info info;
2332                 ext4_grpblk_t counters[16];
2333         } sg;
2334
2335         group--;
2336         if (group == 0)
2337                 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2338                                 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2339                                   "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2340                            "group", "free", "frags", "first",
2341                            "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2342                            "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2343
2344         i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2345                 sizeof(struct ext4_group_info);
2346         err = ext4_mb_load_buddy(sb, group, &e4b);
2347         if (err) {
2348                 seq_printf(seq, "#%-5u: I/O error\n", group);
2349                 return 0;
2350         }
2351         ext4_lock_group(sb, group);
2352         memcpy(&sg, ext4_get_group_info(sb, group), i);
2353         ext4_unlock_group(sb, group);
2354         ext4_mb_release_desc(&e4b);
2355
2356         seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2357                         sg.info.bb_fragments, sg.info.bb_first_free);
2358         for (i = 0; i <= 13; i++)
2359                 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2360                                 sg.info.bb_counters[i] : 0);
2361         seq_printf(seq, " ]\n");
2362
2363         return 0;
2364 }
2365
2366 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2367 {
2368 }
2369
2370 static const struct seq_operations ext4_mb_seq_groups_ops = {
2371         .start  = ext4_mb_seq_groups_start,
2372         .next   = ext4_mb_seq_groups_next,
2373         .stop   = ext4_mb_seq_groups_stop,
2374         .show   = ext4_mb_seq_groups_show,
2375 };
2376
2377 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2378 {
2379         struct super_block *sb = PDE(inode)->data;
2380         int rc;
2381
2382         rc = seq_open(file, &ext4_mb_seq_groups_ops);
2383         if (rc == 0) {
2384                 struct seq_file *m = (struct seq_file *)file->private_data;
2385                 m->private = sb;
2386         }
2387         return rc;
2388
2389 }
2390
2391 static const struct file_operations ext4_mb_seq_groups_fops = {
2392         .owner          = THIS_MODULE,
2393         .open           = ext4_mb_seq_groups_open,
2394         .read           = seq_read,
2395         .llseek         = seq_lseek,
2396         .release        = seq_release,
2397 };
2398
2399 static void ext4_mb_history_release(struct super_block *sb)
2400 {
2401         struct ext4_sb_info *sbi = EXT4_SB(sb);
2402
2403         if (sbi->s_proc != NULL) {
2404                 remove_proc_entry("mb_groups", sbi->s_proc);
2405                 if (sbi->s_mb_history_max)
2406                         remove_proc_entry("mb_history", sbi->s_proc);
2407         }
2408         kfree(sbi->s_mb_history);
2409 }
2410
2411 static void ext4_mb_history_init(struct super_block *sb)
2412 {
2413         struct ext4_sb_info *sbi = EXT4_SB(sb);
2414         int i;
2415
2416         if (sbi->s_proc != NULL) {
2417                 if (sbi->s_mb_history_max)
2418                         proc_create_data("mb_history", S_IRUGO, sbi->s_proc,
2419                                          &ext4_mb_seq_history_fops, sb);
2420                 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2421                                  &ext4_mb_seq_groups_fops, sb);
2422         }
2423
2424         sbi->s_mb_history_cur = 0;
2425         spin_lock_init(&sbi->s_mb_history_lock);
2426         i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
2427         sbi->s_mb_history = i ? kzalloc(i, GFP_KERNEL) : NULL;
2428         /* if we can't allocate history, then we simple won't use it */
2429 }
2430
2431 static noinline_for_stack void
2432 ext4_mb_store_history(struct ext4_allocation_context *ac)
2433 {
2434         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2435         struct ext4_mb_history h;
2436
2437         if (sbi->s_mb_history == NULL)
2438                 return;
2439
2440         if (!(ac->ac_op & sbi->s_mb_history_filter))
2441                 return;
2442
2443         h.op = ac->ac_op;
2444         h.pid = current->pid;
2445         h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2446         h.orig = ac->ac_o_ex;
2447         h.result = ac->ac_b_ex;
2448         h.flags = ac->ac_flags;
2449         h.found = ac->ac_found;
2450         h.groups = ac->ac_groups_scanned;
2451         h.cr = ac->ac_criteria;
2452         h.tail = ac->ac_tail;
2453         h.buddy = ac->ac_buddy;
2454         h.merged = 0;
2455         if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
2456                 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2457                                 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2458                         h.merged = 1;
2459                 h.goal = ac->ac_g_ex;
2460                 h.result = ac->ac_f_ex;
2461         }
2462
2463         spin_lock(&sbi->s_mb_history_lock);
2464         memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2465         if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2466                 sbi->s_mb_history_cur = 0;
2467         spin_unlock(&sbi->s_mb_history_lock);
2468 }
2469
2470 #else
2471 #define ext4_mb_history_release(sb)
2472 #define ext4_mb_history_init(sb)
2473 #endif
2474
2475
2476 /* Create and initialize ext4_group_info data for the given group. */
2477 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2478                           struct ext4_group_desc *desc)
2479 {
2480         int i, len;
2481         int metalen = 0;
2482         struct ext4_sb_info *sbi = EXT4_SB(sb);
2483         struct ext4_group_info **meta_group_info;
2484
2485         /*
2486          * First check if this group is the first of a reserved block.
2487          * If it's true, we have to allocate a new table of pointers
2488          * to ext4_group_info structures
2489          */
2490         if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2491                 metalen = sizeof(*meta_group_info) <<
2492                         EXT4_DESC_PER_BLOCK_BITS(sb);
2493                 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2494                 if (meta_group_info == NULL) {
2495                         printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2496                                "buddy group\n");
2497                         goto exit_meta_group_info;
2498                 }
2499                 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2500                         meta_group_info;
2501         }
2502
2503         /*
2504          * calculate needed size. if change bb_counters size,
2505          * don't forget about ext4_mb_generate_buddy()
2506          */
2507         len = offsetof(typeof(**meta_group_info),
2508                        bb_counters[sb->s_blocksize_bits + 2]);
2509
2510         meta_group_info =
2511                 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2512         i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2513
2514         meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2515         if (meta_group_info[i] == NULL) {
2516                 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2517                 goto exit_group_info;
2518         }
2519         set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2520                 &(meta_group_info[i]->bb_state));
2521
2522         /*
2523          * initialize bb_free to be able to skip
2524          * empty groups without initialization
2525          */
2526         if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2527                 meta_group_info[i]->bb_free =
2528                         ext4_free_blocks_after_init(sb, group, desc);
2529         } else {
2530                 meta_group_info[i]->bb_free =
2531                         ext4_free_blks_count(sb, desc);
2532         }
2533
2534         INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2535         init_rwsem(&meta_group_info[i]->alloc_sem);
2536         meta_group_info[i]->bb_free_root.rb_node = NULL;
2537
2538 #ifdef DOUBLE_CHECK
2539         {
2540                 struct buffer_head *bh;
2541                 meta_group_info[i]->bb_bitmap =
2542                         kmalloc(sb->s_blocksize, GFP_KERNEL);
2543                 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2544                 bh = ext4_read_block_bitmap(sb, group);
2545                 BUG_ON(bh == NULL);
2546                 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2547                         sb->s_blocksize);
2548                 put_bh(bh);
2549         }
2550 #endif
2551
2552         return 0;
2553
2554 exit_group_info:
2555         /* If a meta_group_info table has been allocated, release it now */
2556         if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2557                 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2558 exit_meta_group_info:
2559         return -ENOMEM;
2560 } /* ext4_mb_add_groupinfo */
2561
2562 static int ext4_mb_init_backend(struct super_block *sb)
2563 {
2564         ext4_group_t ngroups = ext4_get_groups_count(sb);
2565         ext4_group_t i;
2566         struct ext4_sb_info *sbi = EXT4_SB(sb);
2567         struct ext4_super_block *es = sbi->s_es;
2568         int num_meta_group_infos;
2569         int num_meta_group_infos_max;
2570         int array_size;
2571         struct ext4_group_desc *desc;
2572
2573         /* This is the number of blocks used by GDT */
2574         num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2575                                 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2576
2577         /*
2578          * This is the total number of blocks used by GDT including
2579          * the number of reserved blocks for GDT.
2580          * The s_group_info array is allocated with this value
2581          * to allow a clean online resize without a complex
2582          * manipulation of pointer.
2583          * The drawback is the unused memory when no resize
2584          * occurs but it's very low in terms of pages
2585          * (see comments below)
2586          * Need to handle this properly when META_BG resizing is allowed
2587          */
2588         num_meta_group_infos_max = num_meta_group_infos +
2589                                 le16_to_cpu(es->s_reserved_gdt_blocks);
2590
2591         /*
2592          * array_size is the size of s_group_info array. We round it
2593          * to the next power of two because this approximation is done
2594          * internally by kmalloc so we can have some more memory
2595          * for free here (e.g. may be used for META_BG resize).
2596          */
2597         array_size = 1;
2598         while (array_size < sizeof(*sbi->s_group_info) *
2599                num_meta_group_infos_max)
2600                 array_size = array_size << 1;
2601         /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2602          * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2603          * So a two level scheme suffices for now. */
2604         sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2605         if (sbi->s_group_info == NULL) {
2606                 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2607                 return -ENOMEM;
2608         }
2609         sbi->s_buddy_cache = new_inode(sb);
2610         if (sbi->s_buddy_cache == NULL) {
2611                 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2612                 goto err_freesgi;
2613         }
2614         EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2615         for (i = 0; i < ngroups; i++) {
2616                 desc = ext4_get_group_desc(sb, i, NULL);
2617                 if (desc == NULL) {
2618                         printk(KERN_ERR
2619                                 "EXT4-fs: can't read descriptor %u\n", i);
2620                         goto err_freebuddy;
2621                 }
2622                 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2623                         goto err_freebuddy;
2624         }
2625
2626         return 0;
2627
2628 err_freebuddy:
2629         while (i-- > 0)
2630                 kfree(ext4_get_group_info(sb, i));
2631         i = num_meta_group_infos;
2632         while (i-- > 0)
2633                 kfree(sbi->s_group_info[i]);
2634         iput(sbi->s_buddy_cache);
2635 err_freesgi:
2636         kfree(sbi->s_group_info);
2637         return -ENOMEM;
2638 }
2639
2640 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2641 {
2642         struct ext4_sb_info *sbi = EXT4_SB(sb);
2643         unsigned i, j;
2644         unsigned offset;
2645         unsigned max;
2646         int ret;
2647
2648         i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2649
2650         sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2651         if (sbi->s_mb_offsets == NULL) {
2652                 return -ENOMEM;
2653         }
2654
2655         i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2656         sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2657         if (sbi->s_mb_maxs == NULL) {
2658                 kfree(sbi->s_mb_offsets);
2659                 return -ENOMEM;
2660         }
2661
2662         /* order 0 is regular bitmap */
2663         sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2664         sbi->s_mb_offsets[0] = 0;
2665
2666         i = 1;
2667         offset = 0;
2668         max = sb->s_blocksize << 2;
2669         do {
2670                 sbi->s_mb_offsets[i] = offset;
2671                 sbi->s_mb_maxs[i] = max;
2672                 offset += 1 << (sb->s_blocksize_bits - i);
2673                 max = max >> 1;
2674                 i++;
2675         } while (i <= sb->s_blocksize_bits + 1);
2676
2677         /* init file for buddy data */
2678         ret = ext4_mb_init_backend(sb);
2679         if (ret != 0) {
2680                 kfree(sbi->s_mb_offsets);
2681                 kfree(sbi->s_mb_maxs);
2682                 return ret;
2683         }
2684
2685         spin_lock_init(&sbi->s_md_lock);
2686         spin_lock_init(&sbi->s_bal_lock);
2687
2688         sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2689         sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2690         sbi->s_mb_stats = MB_DEFAULT_STATS;
2691         sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2692         sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2693         sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
2694         sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2695
2696         sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2697         if (sbi->s_locality_groups == NULL) {
2698                 kfree(sbi->s_mb_offsets);
2699                 kfree(sbi->s_mb_maxs);
2700                 return -ENOMEM;
2701         }
2702         for_each_possible_cpu(i) {
2703                 struct ext4_locality_group *lg;
2704                 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2705                 mutex_init(&lg->lg_mutex);
2706                 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2707                         INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2708                 spin_lock_init(&lg->lg_prealloc_lock);
2709         }
2710
2711         ext4_mb_history_init(sb);
2712
2713         if (sbi->s_journal)
2714                 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2715
2716         printk(KERN_INFO "EXT4-fs: mballoc enabled\n");
2717         return 0;
2718 }
2719
2720 /* need to called with the ext4 group lock held */
2721 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2722 {
2723         struct ext4_prealloc_space *pa;
2724         struct list_head *cur, *tmp;
2725         int count = 0;
2726
2727         list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2728                 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2729                 list_del(&pa->pa_group_list);
2730                 count++;
2731                 kmem_cache_free(ext4_pspace_cachep, pa);
2732         }
2733         if (count)
2734                 mb_debug(1, "mballoc: %u PAs left\n", count);
2735
2736 }
2737
2738 int ext4_mb_release(struct super_block *sb)
2739 {
2740         ext4_group_t ngroups = ext4_get_groups_count(sb);
2741         ext4_group_t i;
2742         int num_meta_group_infos;
2743         struct ext4_group_info *grinfo;
2744         struct ext4_sb_info *sbi = EXT4_SB(sb);
2745
2746         if (sbi->s_group_info) {
2747                 for (i = 0; i < ngroups; i++) {
2748                         grinfo = ext4_get_group_info(sb, i);
2749 #ifdef DOUBLE_CHECK
2750                         kfree(grinfo->bb_bitmap);
2751 #endif
2752                         ext4_lock_group(sb, i);
2753                         ext4_mb_cleanup_pa(grinfo);
2754                         ext4_unlock_group(sb, i);
2755                         kfree(grinfo);
2756                 }
2757                 num_meta_group_infos = (ngroups +
2758                                 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2759                         EXT4_DESC_PER_BLOCK_BITS(sb);
2760                 for (i = 0; i < num_meta_group_infos; i++)
2761                         kfree(sbi->s_group_info[i]);
2762                 kfree(sbi->s_group_info);
2763         }
2764         kfree(sbi->s_mb_offsets);
2765         kfree(sbi->s_mb_maxs);
2766         if (sbi->s_buddy_cache)
2767                 iput(sbi->s_buddy_cache);
2768         if (sbi->s_mb_stats) {
2769                 printk(KERN_INFO
2770                        "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2771                                 atomic_read(&sbi->s_bal_allocated),
2772                                 atomic_read(&sbi->s_bal_reqs),
2773                                 atomic_read(&sbi->s_bal_success));
2774                 printk(KERN_INFO
2775                       "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2776                                 "%u 2^N hits, %u breaks, %u lost\n",
2777                                 atomic_read(&sbi->s_bal_ex_scanned),
2778                                 atomic_read(&sbi->s_bal_goals),
2779                                 atomic_read(&sbi->s_bal_2orders),
2780                                 atomic_read(&sbi->s_bal_breaks),
2781                                 atomic_read(&sbi->s_mb_lost_chunks));
2782                 printk(KERN_INFO
2783                        "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2784                                 sbi->s_mb_buddies_generated++,
2785                                 sbi->s_mb_generation_time);
2786                 printk(KERN_INFO
2787                        "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2788                                 atomic_read(&sbi->s_mb_preallocated),
2789                                 atomic_read(&sbi->s_mb_discarded));
2790         }
2791
2792         free_percpu(sbi->s_locality_groups);
2793         ext4_mb_history_release(sb);
2794
2795         return 0;
2796 }
2797
2798 /*
2799  * This function is called by the jbd2 layer once the commit has finished,
2800  * so we know we can free the blocks that were released with that commit.
2801  */
2802 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2803 {
2804         struct super_block *sb = journal->j_private;
2805         struct ext4_buddy e4b;
2806         struct ext4_group_info *db;
2807         int err, count = 0, count2 = 0;
2808         struct ext4_free_data *entry;
2809         ext4_fsblk_t discard_block;
2810         struct list_head *l, *ltmp;
2811
2812         list_for_each_safe(l, ltmp, &txn->t_private_list) {
2813                 entry = list_entry(l, struct ext4_free_data, list);
2814
2815                 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2816                          entry->count, entry->group, entry);
2817
2818                 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2819                 /* we expect to find existing buddy because it's pinned */
2820                 BUG_ON(err != 0);
2821
2822                 db = e4b.bd_info;
2823                 /* there are blocks to put in buddy to make them really free */
2824                 count += entry->count;
2825                 count2++;
2826                 ext4_lock_group(sb, entry->group);
2827                 /* Take it out of per group rb tree */
2828                 rb_erase(&entry->node, &(db->bb_free_root));
2829                 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2830
2831                 if (!db->bb_free_root.rb_node) {
2832                         /* No more items in the per group rb tree
2833                          * balance refcounts from ext4_mb_free_metadata()
2834                          */
2835                         page_cache_release(e4b.bd_buddy_page);
2836                         page_cache_release(e4b.bd_bitmap_page);
2837                 }
2838                 ext4_unlock_group(sb, entry->group);
2839                 discard_block = (ext4_fsblk_t) entry->group * EXT4_BLOCKS_PER_GROUP(sb)
2840                         + entry->start_blk
2841                         + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
2842                 trace_ext4_discard_blocks(sb, (unsigned long long)discard_block,
2843                                           entry->count);
2844                 sb_issue_discard(sb, discard_block, entry->count);
2845
2846                 kmem_cache_free(ext4_free_ext_cachep, entry);
2847                 ext4_mb_release_desc(&e4b);
2848         }
2849
2850         mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2851 }
2852
2853 #ifdef CONFIG_EXT4_DEBUG
2854 u8 mb_enable_debug __read_mostly;
2855
2856 static struct dentry *debugfs_dir;
2857 static struct dentry *debugfs_debug;
2858
2859 static void __init ext4_create_debugfs_entry(void)
2860 {
2861         debugfs_dir = debugfs_create_dir("ext4", NULL);
2862         if (debugfs_dir)
2863                 debugfs_debug = debugfs_create_u8("mballoc-debug",
2864                                                   S_IRUGO | S_IWUSR,
2865                                                   debugfs_dir,
2866                                                   &mb_enable_debug);
2867 }
2868
2869 static void ext4_remove_debugfs_entry(void)
2870 {
2871         debugfs_remove(debugfs_debug);
2872         debugfs_remove(debugfs_dir);
2873 }
2874
2875 #else
2876
2877 static void __init ext4_create_debugfs_entry(void)
2878 {
2879 }
2880
2881 static void ext4_remove_debugfs_entry(void)
2882 {
2883 }
2884
2885 #endif
2886
2887 int __init init_ext4_mballoc(void)
2888 {
2889         ext4_pspace_cachep =
2890                 kmem_cache_create("ext4_prealloc_space",
2891                                      sizeof(struct ext4_prealloc_space),
2892                                      0, SLAB_RECLAIM_ACCOUNT, NULL);
2893         if (ext4_pspace_cachep == NULL)
2894                 return -ENOMEM;
2895
2896         ext4_ac_cachep =
2897                 kmem_cache_create("ext4_alloc_context",
2898                                      sizeof(struct ext4_allocation_context),
2899                                      0, SLAB_RECLAIM_ACCOUNT, NULL);
2900         if (ext4_ac_cachep == NULL) {
2901                 kmem_cache_destroy(ext4_pspace_cachep);
2902                 return -ENOMEM;
2903         }
2904
2905         ext4_free_ext_cachep =
2906                 kmem_cache_create("ext4_free_block_extents",
2907                                      sizeof(struct ext4_free_data),
2908                                      0, SLAB_RECLAIM_ACCOUNT, NULL);
2909         if (ext4_free_ext_cachep == NULL) {
2910                 kmem_cache_destroy(ext4_pspace_cachep);
2911                 kmem_cache_destroy(ext4_ac_cachep);
2912                 return -ENOMEM;
2913         }
2914         ext4_create_debugfs_entry();
2915         return 0;
2916 }
2917
2918 void exit_ext4_mballoc(void)
2919 {
2920         /* 
2921          * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2922          * before destroying the slab cache.
2923          */
2924         rcu_barrier();
2925         kmem_cache_destroy(ext4_pspace_cachep);
2926         kmem_cache_destroy(ext4_ac_cachep);
2927         kmem_cache_destroy(ext4_free_ext_cachep);
2928         ext4_remove_debugfs_entry();
2929 }
2930
2931
2932 /*
2933  * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2934  * Returns 0 if success or error code
2935  */
2936 static noinline_for_stack int
2937 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2938                                 handle_t *handle, unsigned int reserv_blks)
2939 {
2940         struct buffer_head *bitmap_bh = NULL;
2941         struct ext4_super_block *es;
2942         struct ext4_group_desc *gdp;
2943         struct buffer_head *gdp_bh;
2944         struct ext4_sb_info *sbi;
2945         struct super_block *sb;
2946         ext4_fsblk_t block;
2947         int err, len;
2948
2949         BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2950         BUG_ON(ac->ac_b_ex.fe_len <= 0);
2951
2952         sb = ac->ac_sb;
2953         sbi = EXT4_SB(sb);
2954         es = sbi->s_es;
2955
2956
2957         err = -EIO;
2958         bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2959         if (!bitmap_bh)
2960                 goto out_err;
2961
2962         err = ext4_journal_get_write_access(handle, bitmap_bh);
2963         if (err)
2964                 goto out_err;
2965
2966         err = -EIO;
2967         gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2968         if (!gdp)
2969                 goto out_err;
2970
2971         ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2972                         ext4_free_blks_count(sb, gdp));
2973
2974         err = ext4_journal_get_write_access(handle, gdp_bh);
2975         if (err)
2976                 goto out_err;
2977
2978         block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
2979                 + ac->ac_b_ex.fe_start
2980                 + le32_to_cpu(es->s_first_data_block);
2981
2982         len = ac->ac_b_ex.fe_len;
2983         if (!ext4_data_block_valid(sbi, block, len)) {
2984                 ext4_error(sb, __func__,
2985                            "Allocating blocks %llu-%llu which overlap "
2986                            "fs metadata\n", block, block+len);
2987                 /* File system mounted not to panic on error
2988                  * Fix the bitmap and repeat the block allocation
2989                  * We leak some of the blocks here.
2990                  */
2991                 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2992                 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2993                             ac->ac_b_ex.fe_len);
2994                 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2995                 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2996                 if (!err)
2997                         err = -EAGAIN;
2998                 goto out_err;
2999         }
3000
3001         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3002 #ifdef AGGRESSIVE_CHECK
3003         {
3004                 int i;
3005                 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3006                         BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3007                                                 bitmap_bh->b_data));
3008                 }
3009         }
3010 #endif
3011         mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
3012         if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
3013                 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3014                 ext4_free_blks_set(sb, gdp,
3015                                         ext4_free_blocks_after_init(sb,
3016                                         ac->ac_b_ex.fe_group, gdp));
3017         }
3018         len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
3019         ext4_free_blks_set(sb, gdp, len);
3020         gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
3021
3022         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3023         percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
3024         /*
3025          * Now reduce the dirty block count also. Should not go negative
3026          */
3027         if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3028                 /* release all the reserved blocks if non delalloc */
3029                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
3030         else {
3031                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
3032                                                 ac->ac_b_ex.fe_len);
3033                 /* convert reserved quota blocks to real quota blocks */
3034                 vfs_dq_claim_block(ac->ac_inode, ac->ac_b_ex.fe_len);
3035         }
3036
3037         if (sbi->s_log_groups_per_flex) {
3038                 ext4_group_t flex_group = ext4_flex_group(sbi,
3039                                                           ac->ac_b_ex.fe_group);
3040                 atomic_sub(ac->ac_b_ex.fe_len,
3041                            &sbi->s_flex_groups[flex_group].free_blocks);
3042         }
3043
3044         err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3045         if (err)
3046                 goto out_err;
3047         err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3048
3049 out_err:
3050         sb->s_dirt = 1;
3051         brelse(bitmap_bh);
3052         return err;
3053 }
3054
3055 /*
3056  * here we normalize request for locality group
3057  * Group request are normalized to s_strip size if we set the same via mount
3058  * option. If not we set it to s_mb_group_prealloc which can be configured via
3059  * /sys/fs/ext4/<partition>/mb_group_prealloc
3060  *
3061  * XXX: should we try to preallocate more than the group has now?
3062  */
3063 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3064 {
3065         struct super_block *sb = ac->ac_sb;
3066         struct ext4_locality_group *lg = ac->ac_lg;
3067
3068         BUG_ON(lg == NULL);
3069         if (EXT4_SB(sb)->s_stripe)
3070                 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
3071         else
3072                 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3073         mb_debug(1, "#%u: goal %u blocks for locality group\n",
3074                 current->pid, ac->ac_g_ex.fe_len);
3075 }
3076
3077 /*
3078  * Normalization means making request better in terms of
3079  * size and alignment
3080  */
3081 static noinline_for_stack void
3082 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3083                                 struct ext4_allocation_request *ar)
3084 {
3085         int bsbits, max;
3086         ext4_lblk_t end;
3087         loff_t size, orig_size, start_off;
3088         ext4_lblk_t start, orig_start;
3089         struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3090         struct ext4_prealloc_space *pa;
3091
3092         /* do normalize only data requests, metadata requests
3093            do not need preallocation */
3094         if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3095                 return;
3096
3097         /* sometime caller may want exact blocks */
3098         if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3099                 return;
3100
3101         /* caller may indicate that preallocation isn't
3102          * required (it's a tail, for example) */
3103         if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3104                 return;
3105
3106         if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3107                 ext4_mb_normalize_group_request(ac);
3108                 return ;
3109         }
3110
3111         bsbits = ac->ac_sb->s_blocksize_bits;
3112
3113         /* first, let's learn actual file size
3114          * given current request is allocated */
3115         size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3116         size = size << bsbits;
3117         if (size < i_size_read(ac->ac_inode))
3118                 size = i_size_read(ac->ac_inode);
3119
3120         /* max size of free chunks */
3121         max = 2 << bsbits;
3122
3123 #define NRL_CHECK_SIZE(req, size, max, chunk_size)      \
3124                 (req <= (size) || max <= (chunk_size))
3125
3126         /* first, try to predict filesize */
3127         /* XXX: should this table be tunable? */
3128         start_off = 0;
3129         if (size <= 16 * 1024) {
3130                 size = 16 * 1024;
3131         } else if (size <= 32 * 1024) {
3132                 size = 32 * 1024;
3133         } else if (size <= 64 * 1024) {
3134                 size = 64 * 1024;
3135         } else if (size <= 128 * 1024) {
3136                 size = 128 * 1024;
3137         } else if (size <= 256 * 1024) {
3138                 size = 256 * 1024;
3139         } else if (size <= 512 * 1024) {
3140                 size = 512 * 1024;
3141         } else if (size <= 1024 * 1024) {
3142                 size = 1024 * 1024;
3143         } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3144                 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3145                                                 (21 - bsbits)) << 21;
3146                 size = 2 * 1024 * 1024;
3147         } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3148                 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3149                                                         (22 - bsbits)) << 22;
3150                 size = 4 * 1024 * 1024;
3151         } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3152                                         (8<<20)>>bsbits, max, 8 * 1024)) {
3153                 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3154                                                         (23 - bsbits)) << 23;
3155                 size = 8 * 1024 * 1024;
3156         } else {
3157                 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3158                 size      = ac->ac_o_ex.fe_len << bsbits;
3159         }
3160         orig_size = size = size >> bsbits;
3161         orig_start = start = start_off >> bsbits;
3162
3163         /* don't cover already allocated blocks in selected range */
3164         if (ar->pleft && start <= ar->lleft) {
3165                 size -= ar->lleft + 1 - start;
3166                 start = ar->lleft + 1;
3167         }
3168         if (ar->pright && start + size - 1 >= ar->lright)
3169                 size -= start + size - ar->lright;
3170
3171         end = start + size;
3172
3173         /* check we don't cross already preallocated blocks */
3174         rcu_read_lock();
3175         list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3176                 ext4_lblk_t pa_end;
3177
3178                 if (pa->pa_deleted)
3179                         continue;
3180                 spin_lock(&pa->pa_lock);
3181                 if (pa->pa_deleted) {
3182                         spin_unlock(&pa->pa_lock);
3183                         continue;
3184                 }
3185
3186                 pa_end = pa->pa_lstart + pa->pa_len;
3187
3188                 /* PA must not overlap original request */
3189                 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3190                         ac->ac_o_ex.fe_logical < pa->pa_lstart));
3191
3192                 /* skip PAs this normalized request doesn't overlap with */
3193                 if (pa->pa_lstart >= end || pa_end <= start) {
3194                         spin_unlock(&pa->pa_lock);
3195                         continue;
3196                 }
3197                 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3198
3199                 /* adjust start or end to be adjacent to this pa */
3200                 if (pa_end <= ac->ac_o_ex.fe_logical) {
3201                         BUG_ON(pa_end < start);
3202                         start = pa_end;
3203                 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3204                         BUG_ON(pa->pa_lstart > end);
3205                         end = pa->pa_lstart;
3206                 }
3207                 spin_unlock(&pa->pa_lock);
3208         }
3209         rcu_read_unlock();
3210         size = end - start;
3211
3212         /* XXX: extra loop to check we really don't overlap preallocations */
3213         rcu_read_lock();
3214         list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3215                 ext4_lblk_t pa_end;
3216                 spin_lock(&pa->pa_lock);
3217                 if (pa->pa_deleted == 0) {
3218                         pa_end = pa->pa_lstart + pa->pa_len;
3219                         BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3220                 }
3221                 spin_unlock(&pa->pa_lock);
3222         }
3223         rcu_read_unlock();
3224
3225         if (start + size <= ac->ac_o_ex.fe_logical &&
3226                         start > ac->ac_o_ex.fe_logical) {
3227                 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3228                         (unsigned long) start, (unsigned long) size,
3229                         (unsigned long) ac->ac_o_ex.fe_logical);
3230         }
3231         BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3232                         start > ac->ac_o_ex.fe_logical);
3233         BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3234
3235         /* now prepare goal request */
3236
3237         /* XXX: is it better to align blocks WRT to logical
3238          * placement or satisfy big request as is */
3239         ac->ac_g_ex.fe_logical = start;
3240         ac->ac_g_ex.fe_len = size;
3241
3242         /* define goal start in order to merge */
3243         if (ar->pright && (ar->lright == (start + size))) {
3244                 /* merge to the right */
3245                 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3246                                                 &ac->ac_f_ex.fe_group,
3247                                                 &ac->ac_f_ex.fe_start);
3248                 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3249         }
3250         if (ar->pleft && (ar->lleft + 1 == start)) {
3251                 /* merge to the left */
3252                 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3253                                                 &ac->ac_f_ex.fe_group,
3254                                                 &ac->ac_f_ex.fe_start);
3255                 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3256         }
3257
3258         mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3259                 (unsigned) orig_size, (unsigned) start);
3260 }
3261
3262 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3263 {
3264         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3265
3266         if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3267                 atomic_inc(&sbi->s_bal_reqs);
3268                 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3269                 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3270                         atomic_inc(&sbi->s_bal_success);
3271                 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3272                 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3273                                 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3274                         atomic_inc(&sbi->s_bal_goals);
3275                 if (ac->ac_found > sbi->s_mb_max_to_scan)
3276                         atomic_inc(&sbi->s_bal_breaks);
3277         }
3278
3279         ext4_mb_store_history(ac);
3280 }
3281
3282 /*
3283  * use blocks preallocated to inode
3284  */
3285 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3286                                 struct ext4_prealloc_space *pa)
3287 {
3288         ext4_fsblk_t start;
3289         ext4_fsblk_t end;
3290         int len;
3291
3292         /* found preallocated blocks, use them */
3293         start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3294         end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3295         len = end - start;
3296         ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3297                                         &ac->ac_b_ex.fe_start);
3298         ac->ac_b_ex.fe_len = len;
3299         ac->ac_status = AC_STATUS_FOUND;
3300         ac->ac_pa = pa;
3301
3302         BUG_ON(start < pa->pa_pstart);
3303         BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3304         BUG_ON(pa->pa_free < len);
3305         pa->pa_free -= len;
3306
3307         mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3308 }
3309
3310 /*
3311  * use blocks preallocated to locality group
3312  */
3313 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3314                                 struct ext4_prealloc_space *pa)
3315 {
3316         unsigned int len = ac->ac_o_ex.fe_len;
3317
3318         ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3319                                         &ac->ac_b_ex.fe_group,
3320                                         &ac->ac_b_ex.fe_start);
3321         ac->ac_b_ex.fe_len = len;
3322         ac->ac_status = AC_STATUS_FOUND;
3323         ac->ac_pa = pa;
3324
3325         /* we don't correct pa_pstart or pa_plen here to avoid
3326          * possible race when the group is being loaded concurrently
3327          * instead we correct pa later, after blocks are marked
3328          * in on-disk bitmap -- see ext4_mb_release_context()
3329          * Other CPUs are prevented from allocating from this pa by lg_mutex
3330          */
3331         mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3332 }
3333
3334 /*
3335  * Return the prealloc space that have minimal distance
3336  * from the goal block. @cpa is the prealloc
3337  * space that is having currently known minimal distance
3338  * from the goal block.
3339  */
3340 static struct ext4_prealloc_space *
3341 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3342                         struct ext4_prealloc_space *pa,
3343                         struct ext4_prealloc_space *cpa)
3344 {
3345         ext4_fsblk_t cur_distance, new_distance;
3346
3347         if (cpa == NULL) {
3348                 atomic_inc(&pa->pa_count);
3349                 return pa;
3350         }
3351         cur_distance = abs(goal_block - cpa->pa_pstart);
3352         new_distance = abs(goal_block - pa->pa_pstart);
3353
3354         if (cur_distance < new_distance)
3355                 return cpa;
3356
3357         /* drop the previous reference */
3358         atomic_dec(&cpa->pa_count);
3359         atomic_inc(&pa->pa_count);
3360         return pa;
3361 }
3362
3363 /*
3364  * search goal blocks in preallocated space
3365  */
3366 static noinline_for_stack int
3367 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3368 {
3369         int order, i;
3370         struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3371         struct ext4_locality_group *lg;
3372         struct ext4_prealloc_space *pa, *cpa = NULL;
3373         ext4_fsblk_t goal_block;
3374
3375         /* only data can be preallocated */
3376         if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3377                 return 0;
3378
3379         /* first, try per-file preallocation */
3380         rcu_read_lock();
3381         list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3382
3383                 /* all fields in this condition don't change,
3384                  * so we can skip locking for them */
3385                 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3386                         ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3387                         continue;
3388
3389                 /* non-extent files can't have physical blocks past 2^32 */
3390                 if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL) &&
3391                         pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3392                         continue;
3393
3394                 /* found preallocated blocks, use them */
3395                 spin_lock(&pa->pa_lock);
3396                 if (pa->pa_deleted == 0 && pa->pa_free) {
3397                         atomic_inc(&pa->pa_count);
3398                         ext4_mb_use_inode_pa(ac, pa);
3399                         spin_unlock(&pa->pa_lock);
3400                         ac->ac_criteria = 10;
3401                         rcu_read_unlock();
3402                         return 1;
3403                 }
3404                 spin_unlock(&pa->pa_lock);
3405         }
3406         rcu_read_unlock();
3407
3408         /* can we use group allocation? */
3409         if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3410                 return 0;
3411
3412         /* inode may have no locality group for some reason */
3413         lg = ac->ac_lg;
3414         if (lg == NULL)
3415                 return 0;
3416         order  = fls(ac->ac_o_ex.fe_len) - 1;
3417         if (order > PREALLOC_TB_SIZE - 1)
3418                 /* The max size of hash table is PREALLOC_TB_SIZE */
3419                 order = PREALLOC_TB_SIZE - 1;
3420
3421         goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) +
3422                      ac->ac_g_ex.fe_start +
3423                      le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block);
3424         /*
3425          * search for the prealloc space that is having
3426          * minimal distance from the goal block.
3427          */
3428         for (i = order; i < PREALLOC_TB_SIZE; i++) {
3429                 rcu_read_lock();
3430                 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3431                                         pa_inode_list) {
3432                         spin_lock(&pa->pa_lock);
3433                         if (pa->pa_deleted == 0 &&
3434                                         pa->pa_free >= ac->ac_o_ex.fe_len) {
3435
3436                                 cpa = ext4_mb_check_group_pa(goal_block,
3437                                                                 pa, cpa);
3438                         }
3439                         spin_unlock(&pa->pa_lock);
3440                 }
3441                 rcu_read_unlock();
3442         }
3443         if (cpa) {
3444                 ext4_mb_use_group_pa(ac, cpa);
3445                 ac->ac_criteria = 20;
3446                 return 1;
3447         }
3448         return 0;
3449 }
3450
3451 /*
3452  * the function goes through all block freed in the group
3453  * but not yet committed and marks them used in in-core bitmap.
3454  * buddy must be generated from this bitmap
3455  * Need to be called with the ext4 group lock held
3456  */
3457 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3458                                                 ext4_group_t group)
3459 {
3460         struct rb_node *n;
3461         struct ext4_group_info *grp;
3462         struct ext4_free_data *entry;
3463
3464         grp = ext4_get_group_info(sb, group);
3465         n = rb_first(&(grp->bb_free_root));
3466
3467         while (n) {
3468                 entry = rb_entry(n, struct ext4_free_data, node);
3469                 mb_set_bits(bitmap, entry->start_blk, entry->count);
3470                 n = rb_next(n);
3471         }
3472         return;
3473 }
3474
3475 /*
3476  * the function goes through all preallocation in this group and marks them
3477  * used in in-core bitmap. buddy must be generated from this bitmap
3478  * Need to be called with ext4 group lock held
3479  */
3480 static noinline_for_stack
3481 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3482                                         ext4_group_t group)
3483 {
3484         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3485         struct ext4_prealloc_space *pa;
3486         struct list_head *cur;
3487         ext4_group_t groupnr;
3488         ext4_grpblk_t start;
3489         int preallocated = 0;
3490         int count = 0;
3491         int len;
3492
3493         /* all form of preallocation discards first load group,
3494          * so the only competing code is preallocation use.
3495          * we don't need any locking here
3496          * notice we do NOT ignore preallocations with pa_deleted
3497          * otherwise we could leave used blocks available for
3498          * allocation in buddy when concurrent ext4_mb_put_pa()
3499          * is dropping preallocation
3500          */
3501         list_for_each(cur, &grp->bb_prealloc_list) {
3502                 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3503                 spin_lock(&pa->pa_lock);
3504                 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3505                                              &groupnr, &start);
3506                 len = pa->pa_len;
3507                 spin_unlock(&pa->pa_lock);
3508                 if (unlikely(len == 0))
3509                         continue;
3510                 BUG_ON(groupnr != group);
3511                 mb_set_bits(bitmap, start, len);
3512                 preallocated += len;
3513                 count++;
3514         }
3515         mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3516 }
3517
3518 static void ext4_mb_pa_callback(struct rcu_head *head)
3519 {
3520         struct ext4_prealloc_space *pa;
3521         pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3522         kmem_cache_free(ext4_pspace_cachep, pa);
3523 }
3524
3525 /*
3526  * drops a reference to preallocated space descriptor
3527  * if this was the last reference and the space is consumed
3528  */
3529 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3530                         struct super_block *sb, struct ext4_prealloc_space *pa)
3531 {
3532         ext4_group_t grp;
3533         ext4_fsblk_t grp_blk;
3534
3535         if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3536                 return;
3537
3538         /* in this short window concurrent discard can set pa_deleted */
3539         spin_lock(&pa->pa_lock);
3540         if (pa->pa_deleted == 1) {
3541                 spin_unlock(&pa->pa_lock);
3542                 return;
3543         }
3544
3545         pa->pa_deleted = 1;
3546         spin_unlock(&pa->pa_lock);
3547
3548         grp_blk = pa->pa_pstart;
3549         /* 
3550          * If doing group-based preallocation, pa_pstart may be in the
3551          * next group when pa is used up
3552          */
3553         if (pa->pa_type == MB_GROUP_PA)
3554                 grp_blk--;
3555
3556         ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3557
3558         /*
3559          * possible race:
3560          *
3561          *  P1 (buddy init)                     P2 (regular allocation)
3562          *                                      find block B in PA
3563          *  copy on-disk bitmap to buddy
3564          *                                      mark B in on-disk bitmap
3565          *                                      drop PA from group
3566          *  mark all PAs in buddy
3567          *
3568          * thus, P1 initializes buddy with B available. to prevent this
3569          * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3570          * against that pair
3571          */
3572         ext4_lock_group(sb, grp);
3573         list_del(&pa->pa_group_list);
3574         ext4_unlock_group(sb, grp);
3575
3576         spin_lock(pa->pa_obj_lock);
3577         list_del_rcu(&pa->pa_inode_list);
3578         spin_unlock(pa->pa_obj_lock);
3579
3580         call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3581 }
3582
3583 /*
3584  * creates new preallocated space for given inode
3585  */
3586 static noinline_for_stack int
3587 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3588 {
3589         struct super_block *sb = ac->ac_sb;
3590         struct ext4_prealloc_space *pa;
3591         struct ext4_group_info *grp;
3592         struct ext4_inode_info *ei;
3593
3594         /* preallocate only when found space is larger then requested */
3595         BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3596         BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3597         BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3598
3599         pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3600         if (pa == NULL)
3601                 return -ENOMEM;
3602
3603         if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3604                 int winl;
3605                 int wins;
3606                 int win;
3607                 int offs;
3608
3609                 /* we can't allocate as much as normalizer wants.
3610                  * so, found space must get proper lstart
3611                  * to cover original request */
3612                 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3613                 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3614
3615                 /* we're limited by original request in that
3616                  * logical block must be covered any way
3617                  * winl is window we can move our chunk within */
3618                 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3619
3620                 /* also, we should cover whole original request */
3621                 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3622
3623                 /* the smallest one defines real window */
3624                 win = min(winl, wins);
3625
3626                 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3627                 if (offs && offs < win)
3628                         win = offs;
3629
3630                 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3631                 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3632                 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3633         }
3634
3635         /* preallocation can change ac_b_ex, thus we store actually
3636          * allocated blocks for history */
3637         ac->ac_f_ex = ac->ac_b_ex;
3638
3639         pa->pa_lstart = ac->ac_b_ex.fe_logical;
3640         pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3641         pa->pa_len = ac->ac_b_ex.fe_len;
3642         pa->pa_free = pa->pa_len;
3643         atomic_set(&pa->pa_count, 1);
3644         spin_lock_init(&pa->pa_lock);
3645         INIT_LIST_HEAD(&pa->pa_inode_list);
3646         INIT_LIST_HEAD(&pa->pa_group_list);
3647         pa->pa_deleted = 0;
3648         pa->pa_type = MB_INODE_PA;
3649
3650         mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3651                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3652         trace_ext4_mb_new_inode_pa(ac, pa);
3653
3654         ext4_mb_use_inode_pa(ac, pa);
3655         atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3656
3657         ei = EXT4_I(ac->ac_inode);
3658         grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3659
3660         pa->pa_obj_lock = &ei->i_prealloc_lock;
3661         pa->pa_inode = ac->ac_inode;
3662
3663         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3664         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3665         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3666
3667         spin_lock(pa->pa_obj_lock);
3668         list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3669         spin_unlock(pa->pa_obj_lock);
3670
3671         return 0;
3672 }
3673
3674 /*
3675  * creates new preallocated space for locality group inodes belongs to
3676  */
3677 static noinline_for_stack int
3678 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3679 {
3680         struct super_block *sb = ac->ac_sb;
3681         struct ext4_locality_group *lg;
3682         struct ext4_prealloc_space *pa;
3683         struct ext4_group_info *grp;
3684
3685         /* preallocate only when found space is larger then requested */
3686         BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3687         BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3688         BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3689
3690         BUG_ON(ext4_pspace_cachep == NULL);
3691         pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3692         if (pa == NULL)
3693                 return -ENOMEM;
3694
3695         /* preallocation can change ac_b_ex, thus we store actually
3696          * allocated blocks for history */
3697         ac->ac_f_ex = ac->ac_b_ex;
3698
3699         pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3700         pa->pa_lstart = pa->pa_pstart;
3701         pa->pa_len = ac->ac_b_ex.fe_len;
3702         pa->pa_free = pa->pa_len;
3703         atomic_set(&pa->pa_count, 1);
3704         spin_lock_init(&pa->pa_lock);
3705         INIT_LIST_HEAD(&pa->pa_inode_list);
3706         INIT_LIST_HEAD(&pa->pa_group_list);
3707         pa->pa_deleted = 0;
3708         pa->pa_type = MB_GROUP_PA;
3709
3710         mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3711                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3712         trace_ext4_mb_new_group_pa(ac, pa);
3713
3714         ext4_mb_use_group_pa(ac, pa);
3715         atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3716
3717         grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3718         lg = ac->ac_lg;
3719         BUG_ON(lg == NULL);
3720
3721         pa->pa_obj_lock = &lg->lg_prealloc_lock;
3722         pa->pa_inode = NULL;
3723
3724         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3725         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3726         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3727
3728         /*
3729          * We will later add the new pa to the right bucket
3730          * after updating the pa_free in ext4_mb_release_context
3731          */
3732         return 0;
3733 }
3734
3735 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3736 {
3737         int err;
3738
3739         if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3740                 err = ext4_mb_new_group_pa(ac);
3741         else
3742                 err = ext4_mb_new_inode_pa(ac);
3743         return err;
3744 }
3745
3746 /*
3747  * finds all unused blocks in on-disk bitmap, frees them in
3748  * in-core bitmap and buddy.
3749  * @pa must be unlinked from inode and group lists, so that
3750  * nobody else can find/use it.
3751  * the caller MUST hold group/inode locks.
3752  * TODO: optimize the case when there are no in-core structures yet
3753  */
3754 static noinline_for_stack int
3755 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3756                         struct ext4_prealloc_space *pa,
3757                         struct ext4_allocation_context *ac)
3758 {
3759         struct super_block *sb = e4b->bd_sb;
3760         struct ext4_sb_info *sbi = EXT4_SB(sb);
3761         unsigned int end;
3762         unsigned int next;
3763         ext4_group_t group;
3764         ext4_grpblk_t bit;
3765         unsigned long long grp_blk_start;
3766         sector_t start;
3767         int err = 0;
3768         int free = 0;
3769
3770         BUG_ON(pa->pa_deleted == 0);
3771         ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3772         grp_blk_start = pa->pa_pstart - bit;
3773         BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3774         end = bit + pa->pa_len;
3775
3776         if (ac) {
3777                 ac->ac_sb = sb;
3778                 ac->ac_inode = pa->pa_inode;
3779                 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3780         }
3781
3782         while (bit < end) {
3783                 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3784                 if (bit >= end)
3785                         break;
3786                 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3787                 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3788                                 le32_to_cpu(sbi->s_es->s_first_data_block);
3789                 mb_debug(1, "    free preallocated %u/%u in group %u\n",
3790                                 (unsigned) start, (unsigned) next - bit,
3791                                 (unsigned) group);
3792                 free += next - bit;
3793
3794                 if (ac) {
3795                         ac->ac_b_ex.fe_group = group;
3796                         ac->ac_b_ex.fe_start = bit;
3797                         ac->ac_b_ex.fe_len = next - bit;
3798                         ac->ac_b_ex.fe_logical = 0;
3799                         ext4_mb_store_history(ac);
3800                 }
3801
3802                 trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit,
3803                                                next - bit);
3804                 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3805                 bit = next + 1;
3806         }
3807         if (free != pa->pa_free) {
3808                 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3809                         pa, (unsigned long) pa->pa_lstart,
3810                         (unsigned long) pa->pa_pstart,
3811                         (unsigned long) pa->pa_len);
3812                 ext4_grp_locked_error(sb, group,
3813                                         __func__, "free %u, pa_free %u",
3814                                         free, pa->pa_free);
3815                 /*
3816                  * pa is already deleted so we use the value obtained
3817                  * from the bitmap and continue.
3818                  */
3819         }
3820         atomic_add(free, &sbi->s_mb_discarded);
3821
3822         return err;
3823 }
3824
3825 static noinline_for_stack int
3826 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3827                                 struct ext4_prealloc_space *pa,
3828                                 struct ext4_allocation_context *ac)
3829 {
3830         struct super_block *sb = e4b->bd_sb;
3831         ext4_group_t group;
3832         ext4_grpblk_t bit;
3833
3834         if (ac)
3835                 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3836
3837         trace_ext4_mb_release_group_pa(ac, pa);
3838         BUG_ON(pa->pa_deleted == 0);
3839         ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3840         BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3841         mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3842         atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3843
3844         if (ac) {
3845                 ac->ac_sb = sb;
3846                 ac->ac_inode = NULL;
3847                 ac->ac_b_ex.fe_group = group;
3848                 ac->ac_b_ex.fe_start = bit;
3849                 ac->ac_b_ex.fe_len = pa->pa_len;
3850                 ac->ac_b_ex.fe_logical = 0;
3851                 ext4_mb_store_history(ac);
3852         }
3853
3854         return 0;
3855 }
3856
3857 /*
3858  * releases all preallocations in given group
3859  *
3860  * first, we need to decide discard policy:
3861  * - when do we discard
3862  *   1) ENOSPC
3863  * - how many do we discard
3864  *   1) how many requested
3865  */
3866 static noinline_for_stack int
3867 ext4_mb_discard_group_preallocations(struct super_block *sb,
3868                                         ext4_group_t group, int needed)
3869 {
3870         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3871         struct buffer_head *bitmap_bh = NULL;
3872         struct ext4_prealloc_space *pa, *tmp;
3873         struct ext4_allocation_context *ac;
3874         struct list_head list;
3875         struct ext4_buddy e4b;
3876         int err;
3877         int busy = 0;
3878         int free = 0;
3879
3880         mb_debug(1, "discard preallocation for group %u\n", group);
3881
3882         if (list_empty(&grp->bb_prealloc_list))
3883                 return 0;
3884
3885         bitmap_bh = ext4_read_block_bitmap(sb, group);
3886         if (bitmap_bh == NULL) {
3887                 ext4_error(sb, __func__, "Error in reading block "
3888                                 "bitmap for %u", group);
3889                 return 0;
3890         }
3891
3892         err = ext4_mb_load_buddy(sb, group, &e4b);
3893         if (err) {
3894                 ext4_error(sb, __func__, "Error in loading buddy "
3895                                 "information for %u", group);
3896                 put_bh(bitmap_bh);
3897                 return 0;
3898         }
3899
3900         if (needed == 0)
3901                 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3902
3903         INIT_LIST_HEAD(&list);
3904         ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3905         if (ac)
3906                 ac->ac_sb = sb;
3907 repeat:
3908         ext4_lock_group(sb, group);
3909         list_for_each_entry_safe(pa, tmp,
3910                                 &grp->bb_prealloc_list, pa_group_list) {
3911                 spin_lock(&pa->pa_lock);
3912                 if (atomic_read(&pa->pa_count)) {
3913                         spin_unlock(&pa->pa_lock);
3914                         busy = 1;
3915                         continue;
3916                 }
3917                 if (pa->pa_deleted) {
3918                         spin_unlock(&pa->pa_lock);
3919                         continue;
3920                 }
3921
3922                 /* seems this one can be freed ... */
3923                 pa->pa_deleted = 1;
3924
3925                 /* we can trust pa_free ... */
3926                 free += pa->pa_free;
3927
3928                 spin_unlock(&pa->pa_lock);
3929
3930                 list_del(&pa->pa_group_list);
3931                 list_add(&pa->u.pa_tmp_list, &list);
3932         }
3933
3934         /* if we still need more blocks and some PAs were used, try again */
3935         if (free < needed && busy) {
3936                 busy = 0;
3937                 ext4_unlock_group(sb, group);
3938                 /*
3939                  * Yield the CPU here so that we don't get soft lockup
3940                  * in non preempt case.
3941                  */
3942                 yield();
3943                 goto repeat;
3944         }
3945
3946         /* found anything to free? */
3947         if (list_empty(&list)) {
3948                 BUG_ON(free != 0);
3949                 goto out;
3950         }
3951
3952         /* now free all selected PAs */
3953         list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3954
3955                 /* remove from object (inode or locality group) */
3956                 spin_lock(pa->pa_obj_lock);
3957                 list_del_rcu(&pa->pa_inode_list);
3958                 spin_unlock(pa->pa_obj_lock);
3959
3960                 if (pa->pa_type == MB_GROUP_PA)
3961                         ext4_mb_release_group_pa(&e4b, pa, ac);
3962                 else
3963                         ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3964
3965                 list_del(&pa->u.pa_tmp_list);
3966                 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3967         }
3968
3969 out:
3970         ext4_unlock_group(sb, group);
3971         if (ac)
3972                 kmem_cache_free(ext4_ac_cachep, ac);
3973         ext4_mb_release_desc(&e4b);
3974         put_bh(bitmap_bh);
3975         return free;
3976 }
3977
3978 /*
3979  * releases all non-used preallocated blocks for given inode
3980  *
3981  * It's important to discard preallocations under i_data_sem
3982  * We don't want another block to be served from the prealloc
3983  * space when we are discarding the inode prealloc space.
3984  *
3985  * FIXME!! Make sure it is valid at all the call sites
3986  */
3987 void ext4_discard_preallocations(struct inode *inode)
3988 {
3989         struct ext4_inode_info *ei = EXT4_I(inode);
3990         struct super_block *sb = inode->i_sb;
3991         struct buffer_head *bitmap_bh = NULL;
3992         struct ext4_prealloc_space *pa, *tmp;
3993         struct ext4_allocation_context *ac;
3994         ext4_group_t group = 0;
3995         struct list_head list;
3996         struct ext4_buddy e4b;
3997         int err;
3998
3999         if (!S_ISREG(inode->i_mode)) {
4000                 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4001                 return;
4002         }
4003
4004         mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
4005         trace_ext4_discard_preallocations(inode);
4006
4007         INIT_LIST_HEAD(&list);
4008
4009         ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4010         if (ac) {
4011                 ac->ac_sb = sb;
4012                 ac->ac_inode = inode;
4013         }
4014 repeat:
4015         /* first, collect all pa's in the inode */
4016         spin_lock(&ei->i_prealloc_lock);
4017         while (!list_empty(&ei->i_prealloc_list)) {
4018                 pa = list_entry(ei->i_prealloc_list.next,
4019                                 struct ext4_prealloc_space, pa_inode_list);
4020                 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4021                 spin_lock(&pa->pa_lock);
4022                 if (atomic_read(&pa->pa_count)) {
4023                         /* this shouldn't happen often - nobody should
4024                          * use preallocation while we're discarding it */
4025                         spin_unlock(&pa->pa_lock);
4026                         spin_unlock(&ei->i_prealloc_lock);
4027                         printk(KERN_ERR "uh-oh! used pa while discarding\n");
4028                         WARN_ON(1);
4029                         schedule_timeout_uninterruptible(HZ);
4030                         goto repeat;
4031
4032                 }
4033                 if (pa->pa_deleted == 0) {
4034                         pa->pa_deleted = 1;
4035                         spin_unlock(&pa->pa_lock);
4036                         list_del_rcu(&pa->pa_inode_list);
4037                         list_add(&pa->u.pa_tmp_list, &list);
4038                         continue;
4039                 }
4040
4041                 /* someone is deleting pa right now */
4042                 spin_unlock(&pa->pa_lock);
4043                 spin_unlock(&ei->i_prealloc_lock);
4044
4045                 /* we have to wait here because pa_deleted
4046                  * doesn't mean pa is already unlinked from
4047                  * the list. as we might be called from
4048                  * ->clear_inode() the inode will get freed
4049                  * and concurrent thread which is unlinking
4050                  * pa from inode's list may access already
4051                  * freed memory, bad-bad-bad */
4052
4053                 /* XXX: if this happens too often, we can
4054                  * add a flag to force wait only in case
4055                  * of ->clear_inode(), but not in case of
4056                  * regular truncate */
4057                 schedule_timeout_uninterruptible(HZ);
4058                 goto repeat;
4059         }
4060         spin_unlock(&ei->i_prealloc_lock);
4061
4062         list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4063                 BUG_ON(pa->pa_type != MB_INODE_PA);
4064                 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4065
4066                 err = ext4_mb_load_buddy(sb, group, &e4b);
4067                 if (err) {
4068                         ext4_error(sb, __func__, "Error in loading buddy "
4069                                         "information for %u", group);
4070                         continue;
4071                 }
4072
4073                 bitmap_bh = ext4_read_block_bitmap(sb, group);
4074                 if (bitmap_bh == NULL) {
4075                         ext4_error(sb, __func__, "Error in reading block "
4076                                         "bitmap for %u", group);
4077                         ext4_mb_release_desc(&e4b);
4078                         continue;
4079                 }
4080
4081                 ext4_lock_group(sb, group);
4082                 list_del(&pa->pa_group_list);
4083                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
4084                 ext4_unlock_group(sb, group);
4085
4086                 ext4_mb_release_desc(&e4b);
4087                 put_bh(bitmap_bh);
4088
4089                 list_del(&pa->u.pa_tmp_list);
4090                 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4091         }
4092         if (ac)
4093                 kmem_cache_free(ext4_ac_cachep, ac);
4094 }
4095
4096 /*
4097  * finds all preallocated spaces and return blocks being freed to them
4098  * if preallocated space becomes full (no block is used from the space)
4099  * then the function frees space in buddy
4100  * XXX: at the moment, truncate (which is the only way to free blocks)
4101  * discards all preallocations
4102  */
4103 static void ext4_mb_return_to_preallocation(struct inode *inode,
4104                                         struct ext4_buddy *e4b,
4105                                         sector_t block, int count)
4106 {
4107         BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
4108 }
4109 #ifdef CONFIG_EXT4_DEBUG
4110 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4111 {
4112         struct super_block *sb = ac->ac_sb;
4113         ext4_group_t ngroups, i;
4114
4115         printk(KERN_ERR "EXT4-fs: Can't allocate:"
4116                         " Allocation context details:\n");
4117         printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
4118                         ac->ac_status, ac->ac_flags);
4119         printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4120                         "best %lu/%lu/%lu@%lu cr %d\n",
4121                         (unsigned long)ac->ac_o_ex.fe_group,
4122                         (unsigned long)ac->ac_o_ex.fe_start,
4123                         (unsigned long)ac->ac_o_ex.fe_len,
4124                         (unsigned long)ac->ac_o_ex.fe_logical,
4125                         (unsigned long)ac->ac_g_ex.fe_group,
4126                         (unsigned long)ac->ac_g_ex.fe_start,
4127                         (unsigned long)ac->ac_g_ex.fe_len,
4128                         (unsigned long)ac->ac_g_ex.fe_logical,
4129                         (unsigned long)ac->ac_b_ex.fe_group,
4130                         (unsigned long)ac->ac_b_ex.fe_start,
4131                         (unsigned long)ac->ac_b_ex.fe_len,
4132                         (unsigned long)ac->ac_b_ex.fe_logical,
4133                         (int)ac->ac_criteria);
4134         printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
4135                 ac->ac_found);
4136         printk(KERN_ERR "EXT4-fs: groups: \n");
4137         ngroups = ext4_get_groups_count(sb);
4138         for (i = 0; i < ngroups; i++) {
4139                 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4140                 struct ext4_prealloc_space *pa;
4141                 ext4_grpblk_t start;
4142                 struct list_head *cur;
4143                 ext4_lock_group(sb, i);
4144                 list_for_each(cur, &grp->bb_prealloc_list) {
4145                         pa = list_entry(cur, struct ext4_prealloc_space,
4146                                         pa_group_list);
4147                         spin_lock(&pa->pa_lock);
4148                         ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4149                                                      NULL, &start);
4150                         spin_unlock(&pa->pa_lock);
4151                         printk(KERN_ERR "PA:%u:%d:%u \n", i,
4152                                start, pa->pa_len);
4153                 }
4154                 ext4_unlock_group(sb, i);
4155
4156                 if (grp->bb_free == 0)
4157                         continue;
4158                 printk(KERN_ERR "%u: %d/%d \n",
4159                        i, grp->bb_free, grp->bb_fragments);
4160         }
4161         printk(KERN_ERR "\n");
4162 }
4163 #else
4164 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4165 {
4166         return;
4167 }
4168 #endif
4169
4170 /*
4171  * We use locality group preallocation for small size file. The size of the
4172  * file is determined by the current size or the resulting size after
4173  * allocation which ever is larger
4174  *
4175  * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4176  */
4177 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4178 {
4179         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4180         int bsbits = ac->ac_sb->s_blocksize_bits;
4181         loff_t size, isize;
4182
4183         if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4184                 return;
4185
4186         if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4187                 return;
4188
4189         size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4190         isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4191                 >> bsbits;
4192         size = max(size, isize);
4193
4194         if ((size == isize) &&
4195             !ext4_fs_is_busy(sbi) &&
4196             (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4197                 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4198                 return;
4199         }
4200
4201         /* don't use group allocation for large files */
4202         if (size >= sbi->s_mb_stream_request) {
4203                 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4204                 return;
4205         }
4206
4207         BUG_ON(ac->ac_lg != NULL);
4208         /*
4209          * locality group prealloc space are per cpu. The reason for having
4210          * per cpu locality group is to reduce the contention between block
4211          * request from multiple CPUs.
4212          */
4213         ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id());
4214
4215         /* we're going to use group allocation */
4216         ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4217
4218         /* serialize all allocations in the group */
4219         mutex_lock(&ac->ac_lg->lg_mutex);
4220 }
4221
4222 static noinline_for_stack int
4223 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4224                                 struct ext4_allocation_request *ar)
4225 {
4226         struct super_block *sb = ar->inode->i_sb;
4227         struct ext4_sb_info *sbi = EXT4_SB(sb);
4228         struct ext4_super_block *es = sbi->s_es;
4229         ext4_group_t group;
4230         unsigned int len;
4231         ext4_fsblk_t goal;
4232         ext4_grpblk_t block;
4233
4234         /* we can't allocate > group size */
4235         len = ar->len;
4236
4237         /* just a dirty hack to filter too big requests  */
4238         if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4239                 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4240
4241         /* start searching from the goal */
4242         goal = ar->goal;
4243         if (goal < le32_to_cpu(es->s_first_data_block) ||
4244                         goal >= ext4_blocks_count(es))
4245                 goal = le32_to_cpu(es->s_first_data_block);
4246         ext4_get_group_no_and_offset(sb, goal, &group, &block);
4247
4248         /* set up allocation goals */
4249         memset(ac, 0, sizeof(struct ext4_allocation_context));
4250         ac->ac_b_ex.fe_logical = ar->logical;
4251         ac->ac_status = AC_STATUS_CONTINUE;
4252         ac->ac_sb = sb;
4253         ac->ac_inode = ar->inode;
4254         ac->ac_o_ex.fe_logical = ar->logical;
4255         ac->ac_o_ex.fe_group = group;
4256         ac->ac_o_ex.fe_start = block;
4257         ac->ac_o_ex.fe_len = len;
4258         ac->ac_g_ex.fe_logical = ar->logical;
4259         ac->ac_g_ex.fe_group = group;
4260         ac->ac_g_ex.fe_start = block;
4261         ac->ac_g_ex.fe_len = len;
4262         ac->ac_flags = ar->flags;
4263
4264         /* we have to define context: we'll we work with a file or
4265          * locality group. this is a policy, actually */
4266         ext4_mb_group_or_file(ac);
4267
4268         mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4269                         "left: %u/%u, right %u/%u to %swritable\n",
4270                         (unsigned) ar->len, (unsigned) ar->logical,
4271                         (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4272                         (unsigned) ar->lleft, (unsigned) ar->pleft,
4273                         (unsigned) ar->lright, (unsigned) ar->pright,
4274                         atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4275         return 0;
4276
4277 }
4278
4279 static noinline_for_stack void
4280 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4281                                         struct ext4_locality_group *lg,
4282                                         int order, int total_entries)
4283 {
4284         ext4_group_t group = 0;
4285         struct ext4_buddy e4b;
4286         struct list_head discard_list;
4287         struct ext4_prealloc_space *pa, *tmp;
4288         struct ext4_allocation_context *ac;
4289
4290         mb_debug(1, "discard locality group preallocation\n");
4291
4292         INIT_LIST_HEAD(&discard_list);
4293         ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4294         if (ac)
4295                 ac->ac_sb = sb;
4296
4297         spin_lock(&lg->lg_prealloc_lock);
4298         list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4299                                                 pa_inode_list) {
4300                 spin_lock(&pa->pa_lock);
4301                 if (atomic_read(&pa->pa_count)) {
4302                         /*
4303                          * This is the pa that we just used
4304                          * for block allocation. So don't
4305                          * free that
4306                          */
4307                         spin_unlock(&pa->pa_lock);
4308                         continue;
4309                 }
4310                 if (pa->pa_deleted) {
4311                         spin_unlock(&pa->pa_lock);
4312                         continue;
4313                 }
4314                 /* only lg prealloc space */
4315                 BUG_ON(pa->pa_type != MB_GROUP_PA);
4316
4317                 /* seems this one can be freed ... */
4318                 pa->pa_deleted = 1;
4319                 spin_unlock(&pa->pa_lock);
4320
4321                 list_del_rcu(&pa->pa_inode_list);
4322                 list_add(&pa->u.pa_tmp_list, &discard_list);
4323
4324                 total_entries--;
4325                 if (total_entries <= 5) {
4326                         /*
4327                          * we want to keep only 5 entries
4328                          * allowing it to grow to 8. This
4329                          * mak sure we don't call discard
4330                          * soon for this list.
4331                          */
4332                         break;
4333                 }
4334         }
4335         spin_unlock(&lg->lg_prealloc_lock);
4336
4337         list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4338
4339                 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4340                 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4341                         ext4_error(sb, __func__, "Error in loading buddy "
4342                                         "information for %u", group);
4343                         continue;
4344                 }
4345                 ext4_lock_group(sb, group);
4346                 list_del(&pa->pa_group_list);
4347                 ext4_mb_release_group_pa(&e4b, pa, ac);
4348                 ext4_unlock_group(sb, group);
4349
4350                 ext4_mb_release_desc(&e4b);
4351                 list_del(&pa->u.pa_tmp_list);
4352                 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4353         }
4354         if (ac)
4355                 kmem_cache_free(ext4_ac_cachep, ac);
4356 }
4357
4358 /*
4359  * We have incremented pa_count. So it cannot be freed at this
4360  * point. Also we hold lg_mutex. So no parallel allocation is
4361  * possible from this lg. That means pa_free cannot be updated.
4362  *
4363  * A parallel ext4_mb_discard_group_preallocations is possible.
4364  * which can cause the lg_prealloc_list to be updated.
4365  */
4366
4367 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4368 {
4369         int order, added = 0, lg_prealloc_count = 1;
4370         struct super_block *sb = ac->ac_sb;
4371         struct ext4_locality_group *lg = ac->ac_lg;
4372         struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4373
4374         order = fls(pa->pa_free) - 1;
4375         if (order > PREALLOC_TB_SIZE - 1)
4376                 /* The max size of hash table is PREALLOC_TB_SIZE */
4377                 order = PREALLOC_TB_SIZE - 1;
4378         /* Add the prealloc space to lg */
4379         rcu_read_lock();
4380         list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4381                                                 pa_inode_list) {
4382                 spin_lock(&tmp_pa->pa_lock);
4383                 if (tmp_pa->pa_deleted) {
4384                         spin_unlock(&tmp_pa->pa_lock);
4385                         continue;
4386                 }
4387                 if (!added && pa->pa_free < tmp_pa->pa_free) {
4388                         /* Add to the tail of the previous entry */
4389                         list_add_tail_rcu(&pa->pa_inode_list,
4390                                                 &tmp_pa->pa_inode_list);
4391                         added = 1;
4392                         /*
4393                          * we want to count the total
4394                          * number of entries in the list
4395                          */
4396                 }
4397                 spin_unlock(&tmp_pa->pa_lock);
4398                 lg_prealloc_count++;
4399         }
4400         if (!added)
4401                 list_add_tail_rcu(&pa->pa_inode_list,
4402                                         &lg->lg_prealloc_list[order]);
4403         rcu_read_unlock();
4404
4405         /* Now trim the list to be not more than 8 elements */
4406         if (lg_prealloc_count > 8) {
4407                 ext4_mb_discard_lg_preallocations(sb, lg,
4408                                                 order, lg_prealloc_count);
4409                 return;
4410         }
4411         return ;
4412 }
4413
4414 /*
4415  * release all resource we used in allocation
4416  */
4417 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4418 {
4419         struct ext4_prealloc_space *pa = ac->ac_pa;
4420         if (pa) {
4421                 if (pa->pa_type == MB_GROUP_PA) {
4422                         /* see comment in ext4_mb_use_group_pa() */
4423                         spin_lock(&pa->pa_lock);
4424                         pa->pa_pstart += ac->ac_b_ex.fe_len;
4425                         pa->pa_lstart += ac->ac_b_ex.fe_len;
4426                         pa->pa_free -= ac->ac_b_ex.fe_len;
4427                         pa->pa_len -= ac->ac_b_ex.fe_len;
4428                         spin_unlock(&pa->pa_lock);
4429                 }
4430         }
4431         if (ac->alloc_semp)
4432                 up_read(ac->alloc_semp);
4433         if (pa) {
4434                 /*
4435                  * We want to add the pa to the right bucket.
4436                  * Remove it from the list and while adding
4437                  * make sure the list to which we are adding
4438                  * doesn't grow big.  We need to release
4439                  * alloc_semp before calling ext4_mb_add_n_trim()
4440                  */
4441                 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4442                         spin_lock(pa->pa_obj_lock);
4443                         list_del_rcu(&pa->pa_inode_list);
4444                         spin_unlock(pa->pa_obj_lock);
4445                         ext4_mb_add_n_trim(ac);
4446                 }
4447                 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4448         }
4449         if (ac->ac_bitmap_page)
4450                 page_cache_release(ac->ac_bitmap_page);
4451         if (ac->ac_buddy_page)
4452                 page_cache_release(ac->ac_buddy_page);
4453         if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4454                 mutex_unlock(&ac->ac_lg->lg_mutex);
4455         ext4_mb_collect_stats(ac);
4456         return 0;
4457 }
4458
4459 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4460 {
4461         ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4462         int ret;
4463         int freed = 0;
4464
4465         trace_ext4_mb_discard_preallocations(sb, needed);
4466         for (i = 0; i < ngroups && needed > 0; i++) {
4467                 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4468                 freed += ret;
4469                 needed -= ret;
4470         }
4471
4472         return freed;
4473 }
4474
4475 /*
4476  * Main entry point into mballoc to allocate blocks
4477  * it tries to use preallocation first, then falls back
4478  * to usual allocation
4479  */
4480 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4481                                  struct ext4_allocation_request *ar, int *errp)
4482 {
4483         int freed;
4484         struct ext4_allocation_context *ac = NULL;
4485         struct ext4_sb_info *sbi;
4486         struct super_block *sb;
4487         ext4_fsblk_t block = 0;
4488         unsigned int inquota = 0;
4489         unsigned int reserv_blks = 0;
4490
4491         sb = ar->inode->i_sb;
4492         sbi = EXT4_SB(sb);
4493
4494         trace_ext4_request_blocks(ar);
4495
4496         /*
4497          * For delayed allocation, we could skip the ENOSPC and
4498          * EDQUOT check, as blocks and quotas have been already
4499          * reserved when data being copied into pagecache.
4500          */
4501         if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4502                 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4503         else {
4504                 /* Without delayed allocation we need to verify
4505                  * there is enough free blocks to do block allocation
4506                  * and verify allocation doesn't exceed the quota limits.
4507                  */
4508                 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4509                         /* let others to free the space */
4510                         yield();
4511                         ar->len = ar->len >> 1;
4512                 }
4513                 if (!ar->len) {
4514                         *errp = -ENOSPC;
4515                         return 0;
4516                 }
4517                 reserv_blks = ar->len;
4518                 while (ar->len && vfs_dq_alloc_block(ar->inode, ar->len)) {
4519                         ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4520                         ar->len--;
4521                 }
4522                 inquota = ar->len;
4523                 if (ar->len == 0) {
4524                         *errp = -EDQUOT;
4525                         goto out3;
4526                 }
4527         }
4528
4529         ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4530         if (!ac) {
4531                 ar->len = 0;
4532                 *errp = -ENOMEM;
4533                 goto out1;
4534         }
4535
4536         *errp = ext4_mb_initialize_context(ac, ar);
4537         if (*errp) {
4538                 ar->len = 0;
4539                 goto out2;
4540         }
4541
4542         ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4543         if (!ext4_mb_use_preallocated(ac)) {
4544                 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4545                 ext4_mb_normalize_request(ac, ar);
4546 repeat:
4547                 /* allocate space in core */
4548                 ext4_mb_regular_allocator(ac);
4549
4550                 /* as we've just preallocated more space than
4551                  * user requested orinally, we store allocated
4552                  * space in a special descriptor */
4553                 if (ac->ac_status == AC_STATUS_FOUND &&
4554                                 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4555                         ext4_mb_new_preallocation(ac);
4556         }
4557         if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4558                 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4559                 if (*errp ==  -EAGAIN) {
4560                         /*
4561                          * drop the reference that we took
4562                          * in ext4_mb_use_best_found
4563                          */
4564                         ext4_mb_release_context(ac);
4565                         ac->ac_b_ex.fe_group = 0;
4566                         ac->ac_b_ex.fe_start = 0;
4567                         ac->ac_b_ex.fe_len = 0;
4568                         ac->ac_status = AC_STATUS_CONTINUE;
4569                         goto repeat;
4570                 } else if (*errp) {
4571                         ac->ac_b_ex.fe_len = 0;
4572                         ar->len = 0;
4573                         ext4_mb_show_ac(ac);
4574                 } else {
4575                         block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4576                         ar->len = ac->ac_b_ex.fe_len;
4577                 }
4578         } else {
4579                 freed  = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4580                 if (freed)
4581                         goto repeat;
4582                 *errp = -ENOSPC;
4583                 ac->ac_b_ex.fe_len = 0;
4584                 ar->len = 0;
4585                 ext4_mb_show_ac(ac);
4586         }
4587
4588         ext4_mb_release_context(ac);
4589
4590 out2:
4591         kmem_cache_free(ext4_ac_cachep, ac);
4592 out1:
4593         if (inquota && ar->len < inquota)
4594                 vfs_dq_free_block(ar->inode, inquota - ar->len);
4595 out3:
4596         if (!ar->len) {
4597                 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4598                         /* release all the reserved blocks if non delalloc */
4599                         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4600                                                 reserv_blks);
4601         }
4602
4603         trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4604
4605         return block;
4606 }
4607
4608 /*
4609  * We can merge two free data extents only if the physical blocks
4610  * are contiguous, AND the extents were freed by the same transaction,
4611  * AND the blocks are associated with the same group.
4612  */
4613 static int can_merge(struct ext4_free_data *entry1,
4614                         struct ext4_free_data *entry2)
4615 {
4616         if ((entry1->t_tid == entry2->t_tid) &&
4617             (entry1->group == entry2->group) &&
4618             ((entry1->start_blk + entry1->count) == entry2->start_blk))
4619                 return 1;
4620         return 0;
4621 }
4622
4623 static noinline_for_stack int
4624 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4625                       struct ext4_free_data *new_entry)
4626 {
4627         ext4_grpblk_t block;
4628         struct ext4_free_data *entry;
4629         struct ext4_group_info *db = e4b->bd_info;
4630         struct super_block *sb = e4b->bd_sb;
4631         struct ext4_sb_info *sbi = EXT4_SB(sb);
4632         struct rb_node **n = &db->bb_free_root.rb_node, *node;
4633         struct rb_node *parent = NULL, *new_node;
4634
4635         BUG_ON(!ext4_handle_valid(handle));
4636         BUG_ON(e4b->bd_bitmap_page == NULL);
4637         BUG_ON(e4b->bd_buddy_page == NULL);
4638
4639         new_node = &new_entry->node;
4640         block = new_entry->start_blk;
4641
4642         if (!*n) {
4643                 /* first free block exent. We need to
4644                    protect buddy cache from being freed,
4645                  * otherwise we'll refresh it from
4646                  * on-disk bitmap and lose not-yet-available
4647                  * blocks */
4648                 page_cache_get(e4b->bd_buddy_page);
4649                 page_cache_get(e4b->bd_bitmap_page);
4650         }
4651         while (*n) {
4652                 parent = *n;
4653                 entry = rb_entry(parent, struct ext4_free_data, node);
4654                 if (block < entry->start_blk)
4655                         n = &(*n)->rb_left;
4656                 else if (block >= (entry->start_blk + entry->count))
4657                         n = &(*n)->rb_right;
4658                 else {
4659                         ext4_grp_locked_error(sb, e4b->bd_group, __func__,
4660                                         "Double free of blocks %d (%d %d)",
4661                                         block, entry->start_blk, entry->count);
4662                         return 0;
4663                 }
4664         }
4665
4666         rb_link_node(new_node, parent, n);
4667         rb_insert_color(new_node, &db->bb_free_root);
4668
4669         /* Now try to see the extent can be merged to left and right */
4670         node = rb_prev(new_node);
4671         if (node) {
4672                 entry = rb_entry(node, struct ext4_free_data, node);
4673                 if (can_merge(entry, new_entry)) {
4674                         new_entry->start_blk = entry->start_blk;
4675                         new_entry->count += entry->count;
4676                         rb_erase(node, &(db->bb_free_root));
4677                         spin_lock(&sbi->s_md_lock);
4678                         list_del(&entry->list);
4679                         spin_unlock(&sbi->s_md_lock);
4680                         kmem_cache_free(ext4_free_ext_cachep, entry);
4681                 }
4682         }
4683
4684         node = rb_next(new_node);
4685         if (node) {
4686                 entry = rb_entry(node, struct ext4_free_data, node);
4687                 if (can_merge(new_entry, entry)) {
4688                         new_entry->count += entry->count;
4689                         rb_erase(node, &(db->bb_free_root));
4690                         spin_lock(&sbi->s_md_lock);
4691                         list_del(&entry->list);
4692                         spin_unlock(&sbi->s_md_lock);
4693                         kmem_cache_free(ext4_free_ext_cachep, entry);
4694                 }
4695         }
4696         /* Add the extent to transaction's private list */
4697         spin_lock(&sbi->s_md_lock);
4698         list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4699         spin_unlock(&sbi->s_md_lock);
4700         return 0;
4701 }
4702
4703 /*
4704  * Main entry point into mballoc to free blocks
4705  */
4706 void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
4707                         ext4_fsblk_t block, unsigned long count,
4708                         int metadata, unsigned long *freed)
4709 {
4710         struct buffer_head *bitmap_bh = NULL;
4711         struct super_block *sb = inode->i_sb;
4712         struct ext4_allocation_context *ac = NULL;
4713         struct ext4_group_desc *gdp;
4714         struct ext4_super_block *es;
4715         unsigned int overflow;
4716         ext4_grpblk_t bit;
4717         struct buffer_head *gd_bh;
4718         ext4_group_t block_group;
4719         struct ext4_sb_info *sbi;
4720         struct ext4_buddy e4b;
4721         int err = 0;
4722         int ret;
4723
4724         *freed = 0;
4725
4726         sbi = EXT4_SB(sb);
4727         es = EXT4_SB(sb)->s_es;
4728         if (block < le32_to_cpu(es->s_first_data_block) ||
4729             block + count < block ||
4730             block + count > ext4_blocks_count(es)) {
4731                 ext4_error(sb, __func__,
4732                             "Freeing blocks not in datazone - "
4733                             "block = %llu, count = %lu", block, count);
4734                 goto error_return;
4735         }
4736
4737         ext4_debug("freeing block %llu\n", block);
4738         trace_ext4_free_blocks(inode, block, count, metadata);
4739
4740         ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4741         if (ac) {
4742                 ac->ac_op = EXT4_MB_HISTORY_FREE;
4743                 ac->ac_inode = inode;
4744                 ac->ac_sb = sb;
4745         }
4746
4747 do_more:
4748         overflow = 0;
4749         ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4750
4751         /*
4752          * Check to see if we are freeing blocks across a group
4753          * boundary.
4754          */
4755         if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4756                 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4757                 count -= overflow;
4758         }
4759         bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4760         if (!bitmap_bh) {
4761                 err = -EIO;
4762                 goto error_return;
4763         }
4764         gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4765         if (!gdp) {
4766                 err = -EIO;
4767                 goto error_return;
4768         }
4769
4770         if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4771             in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4772             in_range(block, ext4_inode_table(sb, gdp),
4773                       EXT4_SB(sb)->s_itb_per_group) ||
4774             in_range(block + count - 1, ext4_inode_table(sb, gdp),
4775                       EXT4_SB(sb)->s_itb_per_group)) {
4776
4777                 ext4_error(sb, __func__,
4778                            "Freeing blocks in system zone - "
4779                            "Block = %llu, count = %lu", block, count);
4780                 /* err = 0. ext4_std_error should be a no op */
4781                 goto error_return;
4782         }
4783
4784         BUFFER_TRACE(bitmap_bh, "getting write access");
4785         err = ext4_journal_get_write_access(handle, bitmap_bh);
4786         if (err)
4787                 goto error_return;
4788
4789         /*
4790          * We are about to modify some metadata.  Call the journal APIs
4791          * to unshare ->b_data if a currently-committing transaction is
4792          * using it
4793          */
4794         BUFFER_TRACE(gd_bh, "get_write_access");
4795         err = ext4_journal_get_write_access(handle, gd_bh);
4796         if (err)
4797                 goto error_return;
4798 #ifdef AGGRESSIVE_CHECK
4799         {
4800                 int i;
4801                 for (i = 0; i < count; i++)
4802                         BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4803         }
4804 #endif
4805         if (ac) {
4806                 ac->ac_b_ex.fe_group = block_group;
4807                 ac->ac_b_ex.fe_start = bit;
4808                 ac->ac_b_ex.fe_len = count;
4809                 ext4_mb_store_history(ac);
4810         }
4811
4812         err = ext4_mb_load_buddy(sb, block_group, &e4b);
4813         if (err)
4814                 goto error_return;
4815         if (metadata && ext4_handle_valid(handle)) {
4816                 struct ext4_free_data *new_entry;
4817                 /*
4818                  * blocks being freed are metadata. these blocks shouldn't
4819                  * be used until this transaction is committed
4820                  */
4821                 new_entry  = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4822                 new_entry->start_blk = bit;
4823                 new_entry->group  = block_group;
4824                 new_entry->count = count;
4825                 new_entry->t_tid = handle->h_transaction->t_tid;
4826
4827                 ext4_lock_group(sb, block_group);
4828                 mb_clear_bits(bitmap_bh->b_data, bit, count);
4829                 ext4_mb_free_metadata(handle, &e4b, new_entry);
4830         } else {
4831                 /* need to update group_info->bb_free and bitmap
4832                  * with group lock held. generate_buddy look at
4833                  * them with group lock_held
4834                  */
4835                 ext4_lock_group(sb, block_group);
4836                 mb_clear_bits(bitmap_bh->b_data, bit, count);
4837                 mb_free_blocks(inode, &e4b, bit, count);
4838                 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4839         }
4840
4841         ret = ext4_free_blks_count(sb, gdp) + count;
4842         ext4_free_blks_set(sb, gdp, ret);
4843         gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4844         ext4_unlock_group(sb, block_group);
4845         percpu_counter_add(&sbi->s_freeblocks_counter, count);
4846
4847         if (sbi->s_log_groups_per_flex) {
4848                 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4849                 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4850         }
4851
4852         ext4_mb_release_desc(&e4b);
4853
4854         *freed += count;
4855
4856         /* We dirtied the bitmap block */
4857         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4858         err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4859
4860         /* And the group descriptor block */
4861         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4862         ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4863         if (!err)
4864                 err = ret;
4865
4866         if (overflow && !err) {
4867                 block += count;
4868                 count = overflow;
4869                 put_bh(bitmap_bh);
4870                 goto do_more;
4871         }
4872         sb->s_dirt = 1;
4873 error_return:
4874         brelse(bitmap_bh);
4875         ext4_std_error(sb, err);
4876         if (ac)
4877                 kmem_cache_free(ext4_ac_cachep, ac);
4878         return;
4879 }