projects / GitHub / mt8127 / android_kernel_alcatel_ttab.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
percpu counter: clean up percpu_counter_sum_and_set()
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / fs / ext4 / mballoc.c
CommitLineData
c9de560d
AT		 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
8f6e39a7 24#include "mballoc.h"
c9de560d
AT		 25/*
26 * MUSTDO:
27 * - test ext4_ext_search_left() and ext4_ext_search_right()
28 * - search for metadata in few groups
29 *
30 * TODO v4:
31 * - normalization should take into account whether file is still open
32 * - discard preallocations if no free space left (policy?)
33 * - don't normalize tails
34 * - quota
35 * - reservation for superuser
36 *
37 * TODO v3:
38 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
39 * - track min/max extents in each group for better group selection
40 * - mb_mark_used() may allocate chunk right after splitting buddy
41 * - tree of groups sorted by number of free blocks
42 * - error handling
43 */
44
45/*
46 * The allocation request involve request for multiple number of blocks
47 * near to the goal(block) value specified.
48 *
49 * During initialization phase of the allocator we decide to use the group
50 * preallocation or inode preallocation depending on the size file. The
51 * size of the file could be the resulting file size we would have after
52 * allocation or the current file size which ever is larger. If the size is
53 * less that sbi->s_mb_stream_request we select the group
54 * preallocation. The default value of s_mb_stream_request is 16
55 * blocks. This can also be tuned via
56 * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms
57 * of number of blocks.
58 *
59 * The main motivation for having small file use group preallocation is to
60 * ensure that we have small file closer in the disk.
61 *
62 * First stage the allocator looks at the inode prealloc list
63 * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for
64 * this particular inode. The inode prealloc space is represented as:
65 *
66 * pa_lstart -> the logical start block for this prealloc space
67 * pa_pstart -> the physical start block for this prealloc space
68 * pa_len -> lenght for this prealloc space
69 * pa_free -> free space available in this prealloc space
70 *
71 * The inode preallocation space is used looking at the _logical_ start
72 * block. If only the logical file block falls within the range of prealloc
73 * space we will consume the particular prealloc space. This make sure that
74 * that the we have contiguous physical blocks representing the file blocks
75 *
76 * The important thing to be noted in case of inode prealloc space is that
77 * we don't modify the values associated to inode prealloc space except
78 * pa_free.
79 *
80 * If we are not able to find blocks in the inode prealloc space and if we
81 * have the group allocation flag set then we look at the locality group
82 * prealloc space. These are per CPU prealloc list repreasented as
83 *
84 * ext4_sb_info.s_locality_groups[smp_processor_id()]
85 *
86 * The reason for having a per cpu locality group is to reduce the contention
87 * between CPUs. It is possible to get scheduled at this point.
88 *
89 * The locality group prealloc space is used looking at whether we have
90 * enough free space (pa_free) withing the prealloc space.
91 *
92 * If we can't allocate blocks via inode prealloc or/and locality group
93 * prealloc then we look at the buddy cache. The buddy cache is represented
94 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
95 * mapped to the buddy and bitmap information regarding different
96 * groups. The buddy information is attached to buddy cache inode so that
97 * we can access them through the page cache. The information regarding
98 * each group is loaded via ext4_mb_load_buddy. The information involve
99 * block bitmap and buddy information. The information are stored in the
100 * inode as:
101 *
102 * { page }
103 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
104 *
105 *
106 * one block each for bitmap and buddy information. So for each group we
107 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
108 * blocksize) blocks. So it can have information regarding groups_per_page
109 * which is blocks_per_page/2
110 *
111 * The buddy cache inode is not stored on disk. The inode is thrown
112 * away when the filesystem is unmounted.
113 *
114 * We look for count number of blocks in the buddy cache. If we were able
115 * to locate that many free blocks we return with additional information
116 * regarding rest of the contiguous physical block available
117 *
118 * Before allocating blocks via buddy cache we normalize the request
119 * blocks. This ensure we ask for more blocks that we needed. The extra
120 * blocks that we get after allocation is added to the respective prealloc
121 * list. In case of inode preallocation we follow a list of heuristics
122 * based on file size. This can be found in ext4_mb_normalize_request. If
123 * we are doing a group prealloc we try to normalize the request to
124 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is set to
125 * 512 blocks. This can be tuned via
126 * /proc/fs/ext4/<partition/group_prealloc. The value is represented in
127 * terms of number of blocks. If we have mounted the file system with -O
128 * stripe=<value> option the group prealloc request is normalized to the
129 * stripe value (sbi->s_stripe)
130 *
131 * The regular allocator(using the buddy cache) support few tunables.
132 *
133 * /proc/fs/ext4/<partition>/min_to_scan
134 * /proc/fs/ext4/<partition>/max_to_scan
135 * /proc/fs/ext4/<partition>/order2_req
136 *
137 * The regular allocator use buddy scan only if the request len is power of
138 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
139 * value of s_mb_order2_reqs can be tuned via
140 * /proc/fs/ext4/<partition>/order2_req. If the request len is equal to
141 * stripe size (sbi->s_stripe), we try to search for contigous block in
142 * stripe size. This should result in better allocation on RAID setup. If
143 * not we search in the specific group using bitmap for best extents. The
144 * tunable min_to_scan and max_to_scan controll the behaviour here.
145 * min_to_scan indicate how long the mballoc __must__ look for a best
146 * extent and max_to_scanindicate how long the mballoc __can__ look for a
147 * best extent in the found extents. Searching for the blocks starts with
148 * the group specified as the goal value in allocation context via
149 * ac_g_ex. Each group is first checked based on the criteria whether it
150 * can used for allocation. ext4_mb_good_group explains how the groups are
151 * checked.
152 *
153 * Both the prealloc space are getting populated as above. So for the first
154 * request we will hit the buddy cache which will result in this prealloc
155 * space getting filled. The prealloc space is then later used for the
156 * subsequent request.
157 */
158
159/*
160 * mballoc operates on the following data:
161 * - on-disk bitmap
162 * - in-core buddy (actually includes buddy and bitmap)
163 * - preallocation descriptors (PAs)
164 *
165 * there are two types of preallocations:
166 * - inode
167 * assiged to specific inode and can be used for this inode only.
168 * it describes part of inode's space preallocated to specific
169 * physical blocks. any block from that preallocated can be used
170 * independent. the descriptor just tracks number of blocks left
171 * unused. so, before taking some block from descriptor, one must
172 * make sure corresponded logical block isn't allocated yet. this
173 * also means that freeing any block within descriptor's range
174 * must discard all preallocated blocks.
175 * - locality group
176 * assigned to specific locality group which does not translate to
177 * permanent set of inodes: inode can join and leave group. space
178 * from this type of preallocation can be used for any inode. thus
179 * it's consumed from the beginning to the end.
180 *
181 * relation between them can be expressed as:
182 * in-core buddy = on-disk bitmap + preallocation descriptors
183 *
184 * this mean blocks mballoc considers used are:
185 * - allocated blocks (persistent)
186 * - preallocated blocks (non-persistent)
187 *
188 * consistency in mballoc world means that at any time a block is either
189 * free or used in ALL structures. notice: "any time" should not be read
190 * literally -- time is discrete and delimited by locks.
191 *
192 * to keep it simple, we don't use block numbers, instead we count number of
193 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
194 *
195 * all operations can be expressed as:
196 * - init buddy: buddy = on-disk + PAs
197 * - new PA: buddy += N; PA = N
198 * - use inode PA: on-disk += N; PA -= N
199 * - discard inode PA buddy -= on-disk - PA; PA = 0
200 * - use locality group PA on-disk += N; PA -= N
201 * - discard locality group PA buddy -= PA; PA = 0
202 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
203 * is used in real operation because we can't know actual used
204 * bits from PA, only from on-disk bitmap
205 *
206 * if we follow this strict logic, then all operations above should be atomic.
207 * given some of them can block, we'd have to use something like semaphores
208 * killing performance on high-end SMP hardware. let's try to relax it using
209 * the following knowledge:
210 * 1) if buddy is referenced, it's already initialized
211 * 2) while block is used in buddy and the buddy is referenced,
212 * nobody can re-allocate that block
213 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
214 * bit set and PA claims same block, it's OK. IOW, one can set bit in
215 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
216 * block
217 *
218 * so, now we're building a concurrency table:
219 * - init buddy vs.
220 * - new PA
221 * blocks for PA are allocated in the buddy, buddy must be referenced
222 * until PA is linked to allocation group to avoid concurrent buddy init
223 * - use inode PA
224 * we need to make sure that either on-disk bitmap or PA has uptodate data
225 * given (3) we care that PA-=N operation doesn't interfere with init
226 * - discard inode PA
227 * the simplest way would be to have buddy initialized by the discard
228 * - use locality group PA
229 * again PA-=N must be serialized with init
230 * - discard locality group PA
231 * the simplest way would be to have buddy initialized by the discard
232 * - new PA vs.
233 * - use inode PA
234 * i_data_sem serializes them
235 * - discard inode PA
236 * discard process must wait until PA isn't used by another process
237 * - use locality group PA
238 * some mutex should serialize them
239 * - discard locality group PA
240 * discard process must wait until PA isn't used by another process
241 * - use inode PA
242 * - use inode PA
243 * i_data_sem or another mutex should serializes them
244 * - discard inode PA
245 * discard process must wait until PA isn't used by another process
246 * - use locality group PA
247 * nothing wrong here -- they're different PAs covering different blocks
248 * - discard locality group PA
249 * discard process must wait until PA isn't used by another process
250 *
251 * now we're ready to make few consequences:
252 * - PA is referenced and while it is no discard is possible
253 * - PA is referenced until block isn't marked in on-disk bitmap
254 * - PA changes only after on-disk bitmap
255 * - discard must not compete with init. either init is done before
256 * any discard or they're serialized somehow
257 * - buddy init as sum of on-disk bitmap and PAs is done atomically
258 *
259 * a special case when we've used PA to emptiness. no need to modify buddy
260 * in this case, but we should care about concurrent init
261 *
262 */
263
264 /*
265 * Logic in few words:
266 *
267 * - allocation:
268 * load group
269 * find blocks
270 * mark bits in on-disk bitmap
271 * release group
272 *
273 * - use preallocation:
274 * find proper PA (per-inode or group)
275 * load group
276 * mark bits in on-disk bitmap
277 * release group
278 * release PA
279 *
280 * - free:
281 * load group
282 * mark bits in on-disk bitmap
283 * release group
284 *
285 * - discard preallocations in group:
286 * mark PAs deleted
287 * move them onto local list
288 * load on-disk bitmap
289 * load group
290 * remove PA from object (inode or locality group)
291 * mark free blocks in-core
292 *
293 * - discard inode's preallocations:
294 */
295
296/*
297 * Locking rules
298 *
299 * Locks:
300 * - bitlock on a group (group)
301 * - object (inode/locality) (object)
302 * - per-pa lock (pa)
303 *
304 * Paths:
305 * - new pa
306 * object
307 * group
308 *
309 * - find and use pa:
310 * pa
311 *
312 * - release consumed pa:
313 * pa
314 * group
315 * object
316 *
317 * - generate in-core bitmap:
318 * group
319 * pa
320 *
321 * - discard all for given object (inode, locality group):
322 * object
323 * pa
324 * group
325 *
326 * - discard all for given group:
327 * group
328 * pa
329 * group
330 * object
331 *
332 */
333
ffad0a44
AK		 334static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
335{
c9de560d 336#if BITS_PER_LONG == 64
ffad0a44
AK		 337 *bit += ((unsigned long) addr & 7UL) << 3;
338 addr = (void *) ((unsigned long) addr & ~7UL);
c9de560d 339#elif BITS_PER_LONG == 32
ffad0a44
AK		 340 *bit += ((unsigned long) addr & 3UL) << 3;
341 addr = (void *) ((unsigned long) addr & ~3UL);
c9de560d
AT		 342#else
343#error "how many bits you are?!"
344#endif
ffad0a44
AK		 345 return addr;
346}
c9de560d
AT		 347
348static inline int mb_test_bit(int bit, void *addr)
349{
350 /*
351 * ext4_test_bit on architecture like powerpc
352 * needs unsigned long aligned address
353 */
ffad0a44 354 addr = mb_correct_addr_and_bit(&bit, addr);
c9de560d
AT		 355 return ext4_test_bit(bit, addr);
356}
357
358static inline void mb_set_bit(int bit, void *addr)
359{
ffad0a44 360 addr = mb_correct_addr_and_bit(&bit, addr);
c9de560d
AT		 361 ext4_set_bit(bit, addr);
362}
363
364static inline void mb_set_bit_atomic(spinlock_t *lock, int bit, void *addr)
365{
ffad0a44 366 addr = mb_correct_addr_and_bit(&bit, addr);
c9de560d
AT		 367 ext4_set_bit_atomic(lock, bit, addr);
368}
369
370static inline void mb_clear_bit(int bit, void *addr)
371{
ffad0a44 372 addr = mb_correct_addr_and_bit(&bit, addr);
c9de560d
AT		 373 ext4_clear_bit(bit, addr);
374}
375
376static inline void mb_clear_bit_atomic(spinlock_t *lock, int bit, void *addr)
377{
ffad0a44 378 addr = mb_correct_addr_and_bit(&bit, addr);
c9de560d
AT		 379 ext4_clear_bit_atomic(lock, bit, addr);
380}
381
ffad0a44
AK		 382static inline int mb_find_next_zero_bit(void *addr, int max, int start)
383{
e7dfb246 384 int fix = 0, ret, tmpmax;
ffad0a44 385 addr = mb_correct_addr_and_bit(&fix, addr);
e7dfb246 386 tmpmax = max + fix;
ffad0a44
AK		 387 start += fix;
388
e7dfb246
AK		 389 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
390 if (ret > max)
391 return max;
392 return ret;
ffad0a44
AK		 393}
394
395static inline int mb_find_next_bit(void *addr, int max, int start)
396{
e7dfb246 397 int fix = 0, ret, tmpmax;
ffad0a44 398 addr = mb_correct_addr_and_bit(&fix, addr);
e7dfb246 399 tmpmax = max + fix;
ffad0a44
AK		 400 start += fix;
401
e7dfb246
AK		 402 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
403 if (ret > max)
404 return max;
405 return ret;
ffad0a44
AK		 406}
407
c9de560d
AT		 408static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
409{
410 char *bb;
411
c9de560d
AT		 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
432static 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 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
441 for (i = 0; i < count; i++) {
442 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
443 ext4_fsblk_t blocknr;
444 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
445 blocknr += first + i;
446 blocknr +=
447 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
448
46e665e9 449 ext4_error(sb, __func__, "double-free of inode"
c9de560d
AT		 450 " %lu's block %llu(bit %u in group %lu)\n",
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
458static 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 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
465 for (i = 0; i < count; i++) {
466 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
467 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
468 }
469}
470
471static 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("corruption in group %lu at byte %u(%u):"
481 " %x in copy != %x on disk/prealloc\n",
482 e4b->bd_group, i, i * 8, b1[i], b2[i]);
483 BUG();
484 }
485 }
486 }
487}
488
489#else
490static inline void mb_free_blocks_double(struct inode *inode,
491 struct ext4_buddy *e4b, int first, int count)
492{
493 return;
494}
495static inline void mb_mark_used_double(struct ext4_buddy *e4b,
496 int first, int count)
497{
498 return;
499}
500static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
501{
502 return;
503}
504#endif
505
506#ifdef AGGRESSIVE_CHECK
507
508#define MB_CHECK_ASSERT(assert) \
509do { \
510 if (!(assert)) { \
511 printk(KERN_EMERG \
512 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
513 function, file, line, # assert); \
514 BUG(); \
515 } \
516} while (0)
517
518static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
519 const char *function, int line)
520{
521 struct super_block *sb = e4b->bd_sb;
522 int order = e4b->bd_blkbits + 1;
523 int max;
524 int max2;
525 int i;
526 int j;
527 int k;
528 int count;
529 struct ext4_group_info *grp;
530 int fragments = 0;
531 int fstart;
532 struct list_head *cur;
533 void *buddy;
534 void *buddy2;
535
536 if (!test_opt(sb, MBALLOC))
537 return 0;
538
539 {
540 static int mb_check_counter;
541 if (mb_check_counter++ % 100 != 0)
542 return 0;
543 }
544
545 while (order > 1) {
546 buddy = mb_find_buddy(e4b, order, &max);
547 MB_CHECK_ASSERT(buddy);
548 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
549 MB_CHECK_ASSERT(buddy2);
550 MB_CHECK_ASSERT(buddy != buddy2);
551 MB_CHECK_ASSERT(max * 2 == max2);
552
553 count = 0;
554 for (i = 0; i < max; i++) {
555
556 if (mb_test_bit(i, buddy)) {
557 /* only single bit in buddy2 may be 1 */
558 if (!mb_test_bit(i << 1, buddy2)) {
559 MB_CHECK_ASSERT(
560 mb_test_bit((i<<1)+1, buddy2));
561 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
562 MB_CHECK_ASSERT(
563 mb_test_bit(i << 1, buddy2));
564 }
565 continue;
566 }
567
568 /* both bits in buddy2 must be 0 */
569 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
570 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
571
572 for (j = 0; j < (1 << order); j++) {
573 k = (i * (1 << order)) + j;
574 MB_CHECK_ASSERT(
575 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
576 }
577 count++;
578 }
579 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
580 order--;
581 }
582
583 fstart = -1;
584 buddy = mb_find_buddy(e4b, 0, &max);
585 for (i = 0; i < max; i++) {
586 if (!mb_test_bit(i, buddy)) {
587 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
588 if (fstart == -1) {
589 fragments++;
590 fstart = i;
591 }
592 continue;
593 }
594 fstart = -1;
595 /* check used bits only */
596 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
597 buddy2 = mb_find_buddy(e4b, j, &max2);
598 k = i >> j;
599 MB_CHECK_ASSERT(k < max2);
600 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
601 }
602 }
603 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
604 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
605
606 grp = ext4_get_group_info(sb, e4b->bd_group);
607 buddy = mb_find_buddy(e4b, 0, &max);
608 list_for_each(cur, &grp->bb_prealloc_list) {
609 ext4_group_t groupnr;
610 struct ext4_prealloc_space *pa;
60bd63d1
SR		 611 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
612 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
c9de560d 613 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
60bd63d1 614 for (i = 0; i < pa->pa_len; i++)
c9de560d
AT		 615 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
616 }
617 return 0;
618}
619#undef MB_CHECK_ASSERT
620#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
46e665e9 621 __FILE__, __func__, __LINE__)
c9de560d
AT		 622#else
623#define mb_check_buddy(e4b)
624#endif
625
626/* FIXME!! need more doc */
627static void ext4_mb_mark_free_simple(struct super_block *sb,
628 void *buddy, unsigned first, int len,
629 struct ext4_group_info *grp)
630{
631 struct ext4_sb_info *sbi = EXT4_SB(sb);
632 unsigned short min;
633 unsigned short max;
634 unsigned short chunk;
635 unsigned short border;
636
b73fce69 637 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
c9de560d
AT		 638
639 border = 2 << sb->s_blocksize_bits;
640
641 while (len > 0) {
642 /* find how many blocks can be covered since this position */
643 max = ffs(first | border) - 1;
644
645 /* find how many blocks of power 2 we need to mark */
646 min = fls(len) - 1;
647
648 if (max < min)
649 min = max;
650 chunk = 1 << min;
651
652 /* mark multiblock chunks only */
653 grp->bb_counters[min]++;
654 if (min > 0)
655 mb_clear_bit(first >> min,
656 buddy + sbi->s_mb_offsets[min]);
657
658 len -= chunk;
659 first += chunk;
660 }
661}
662
663static void ext4_mb_generate_buddy(struct super_block *sb,
664 void *buddy, void *bitmap, ext4_group_t group)
665{
666 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
667 unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
668 unsigned short i = 0;
669 unsigned short first;
670 unsigned short len;
671 unsigned free = 0;
672 unsigned fragments = 0;
673 unsigned long long period = get_cycles();
674
675 /* initialize buddy from bitmap which is aggregation
676 * of on-disk bitmap and preallocations */
ffad0a44 677 i = mb_find_next_zero_bit(bitmap, max, 0);
c9de560d
AT		 678 grp->bb_first_free = i;
679 while (i < max) {
680 fragments++;
681 first = i;
ffad0a44 682 i = mb_find_next_bit(bitmap, max, i);
c9de560d
AT		 683 len = i - first;
684 free += len;
685 if (len > 1)
686 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
687 else
688 grp->bb_counters[0]++;
689 if (i < max)
ffad0a44 690 i = mb_find_next_zero_bit(bitmap, max, i);
c9de560d
AT		 691 }
692 grp->bb_fragments = fragments;
693
694 if (free != grp->bb_free) {
46e665e9 695 ext4_error(sb, __func__,
c9de560d
AT		 696 "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
697 group, free, grp->bb_free);
e56eb659
AK		 698 /*
699 * If we intent to continue, we consider group descritor
700 * corrupt and update bb_free using bitmap value
701 */
c9de560d
AT		 702 grp->bb_free = free;
703 }
704
705 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
706
707 period = get_cycles() - period;
708 spin_lock(&EXT4_SB(sb)->s_bal_lock);
709 EXT4_SB(sb)->s_mb_buddies_generated++;
710 EXT4_SB(sb)->s_mb_generation_time += period;
711 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
712}
713
714/* The buddy information is attached the buddy cache inode
715 * for convenience. The information regarding each group
716 * is loaded via ext4_mb_load_buddy. The information involve
717 * block bitmap and buddy information. The information are
718 * stored in the inode as
719 *
720 * { page }
721 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
722 *
723 *
724 * one block each for bitmap and buddy information.
725 * So for each group we take up 2 blocks. A page can
726 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
727 * So it can have information regarding groups_per_page which
728 * is blocks_per_page/2
729 */
730
731static int ext4_mb_init_cache(struct page *page, char *incore)
732{
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("init page %lu\n", page->index);
748
749 inode = page->mapping->host;
750 sb = inode->i_sb;
751 blocksize = 1 << inode->i_blkbits;
752 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
753
754 groups_per_page = blocks_per_page >> 1;
755 if (groups_per_page == 0)
756 groups_per_page = 1;
757
758 /* allocate buffer_heads to read bitmaps */
759 if (groups_per_page > 1) {
760 err = -ENOMEM;
761 i = sizeof(struct buffer_head *) * groups_per_page;
762 bh = kzalloc(i, GFP_NOFS);
763 if (bh == NULL)
764 goto out;
765 } else
766 bh = &bhs;
767
768 first_group = page->index * blocks_per_page / 2;
769
770 /* read all groups the page covers into the cache */
771 for (i = 0; i < groups_per_page; i++) {
772 struct ext4_group_desc *desc;
773
774 if (first_group + i >= EXT4_SB(sb)->s_groups_count)
775 break;
776
777 err = -EIO;
778 desc = ext4_get_group_desc(sb, first_group + i, NULL);
779 if (desc == NULL)
780 goto out;
781
782 err = -ENOMEM;
783 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
784 if (bh[i] == NULL)
785 goto out;
786
787 if (bh_uptodate_or_lock(bh[i]))
788 continue;
789
b5f10eed 790 spin_lock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
c9de560d
AT		 791 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
792 ext4_init_block_bitmap(sb, bh[i],
793 first_group + i, desc);
794 set_buffer_uptodate(bh[i]);
795 unlock_buffer(bh[i]);
b5f10eed 796 spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
c9de560d
AT		 797 continue;
798 }
b5f10eed 799 spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
c9de560d
AT		 800 get_bh(bh[i]);
801 bh[i]->b_end_io = end_buffer_read_sync;
802 submit_bh(READ, bh[i]);
803 mb_debug("read bitmap for group %lu\n", first_group + i);
804 }
805
806 /* wait for I/O completion */
807 for (i = 0; i < groups_per_page && bh[i]; i++)
808 wait_on_buffer(bh[i]);
809
810 err = -EIO;
811 for (i = 0; i < groups_per_page && bh[i]; i++)
812 if (!buffer_uptodate(bh[i]))
813 goto out;
814
31b481dc 815 err = 0;
c9de560d
AT		 816 first_block = page->index * blocks_per_page;
817 for (i = 0; i < blocks_per_page; i++) {
818 int group;
819 struct ext4_group_info *grinfo;
820
821 group = (first_block + i) >> 1;
822 if (group >= EXT4_SB(sb)->s_groups_count)
823 break;
824
825 /*
826 * data carry information regarding this
827 * particular group in the format specified
828 * above
829 *
830 */
831 data = page_address(page) + (i * blocksize);
832 bitmap = bh[group - first_group]->b_data;
833
834 /*
835 * We place the buddy block and bitmap block
836 * close together
837 */
838 if ((first_block + i) & 1) {
839 /* this is block of buddy */
840 BUG_ON(incore == NULL);
841 mb_debug("put buddy for group %u in page %lu/%x\n",
842 group, page->index, i * blocksize);
843 memset(data, 0xff, blocksize);
844 grinfo = ext4_get_group_info(sb, group);
845 grinfo->bb_fragments = 0;
846 memset(grinfo->bb_counters, 0,
847 sizeof(unsigned short)*(sb->s_blocksize_bits+2));
848 /*
849 * incore got set to the group block bitmap below
850 */
851 ext4_mb_generate_buddy(sb, data, incore, group);
852 incore = NULL;
853 } else {
854 /* this is block of bitmap */
855 BUG_ON(incore != NULL);
856 mb_debug("put bitmap for group %u in page %lu/%x\n",
857 group, page->index, i * blocksize);
858
859 /* see comments in ext4_mb_put_pa() */
860 ext4_lock_group(sb, group);
861 memcpy(data, bitmap, blocksize);
862
863 /* mark all preallocated blks used in in-core bitmap */
864 ext4_mb_generate_from_pa(sb, data, group);
865 ext4_unlock_group(sb, group);
866
867 /* set incore so that the buddy information can be
868 * generated using this
869 */
870 incore = data;
871 }
872 }
873 SetPageUptodate(page);
874
875out:
876 if (bh) {
877 for (i = 0; i < groups_per_page && bh[i]; i++)
878 brelse(bh[i]);
879 if (bh != &bhs)
880 kfree(bh);
881 }
882 return err;
883}
884
4ddfef7b
ES		 885static noinline_for_stack int
886ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
887 struct ext4_buddy *e4b)
c9de560d
AT		 888{
889 struct ext4_sb_info *sbi = EXT4_SB(sb);
890 struct inode *inode = sbi->s_buddy_cache;
891 int blocks_per_page;
892 int block;
893 int pnum;
894 int poff;
895 struct page *page;
fdf6c7a7 896 int ret;
c9de560d
AT		 897
898 mb_debug("load group %lu\n", group);
899
900 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
901
902 e4b->bd_blkbits = sb->s_blocksize_bits;
903 e4b->bd_info = ext4_get_group_info(sb, group);
904 e4b->bd_sb = sb;
905 e4b->bd_group = group;
906 e4b->bd_buddy_page = NULL;
907 e4b->bd_bitmap_page = NULL;
908
909 /*
910 * the buddy cache inode stores the block bitmap
911 * and buddy information in consecutive blocks.
912 * So for each group we need two blocks.
913 */
914 block = group * 2;
915 pnum = block / blocks_per_page;
916 poff = block % blocks_per_page;
917
918 /* we could use find_or_create_page(), but it locks page
919 * what we'd like to avoid in fast path ... */
920 page = find_get_page(inode->i_mapping, pnum);
921 if (page == NULL || !PageUptodate(page)) {
922 if (page)
923 page_cache_release(page);
924 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
925 if (page) {
926 BUG_ON(page->mapping != inode->i_mapping);
927 if (!PageUptodate(page)) {
fdf6c7a7
SF		 928 ret = ext4_mb_init_cache(page, NULL);
929 if (ret) {
930 unlock_page(page);
931 goto err;
932 }
c9de560d
AT		 933 mb_cmp_bitmaps(e4b, page_address(page) +
934 (poff * sb->s_blocksize));
935 }
936 unlock_page(page);
937 }
938 }
fdf6c7a7
SF		 939 if (page == NULL || !PageUptodate(page)) {
940 ret = -EIO;
c9de560d 941 goto err;
fdf6c7a7 942 }
c9de560d
AT		 943 e4b->bd_bitmap_page = page;
944 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
945 mark_page_accessed(page);
946
947 block++;
948 pnum = block / blocks_per_page;
949 poff = block % blocks_per_page;
950
951 page = find_get_page(inode->i_mapping, pnum);
952 if (page == NULL || !PageUptodate(page)) {
953 if (page)
954 page_cache_release(page);
955 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
956 if (page) {
957 BUG_ON(page->mapping != inode->i_mapping);
fdf6c7a7
SF		 958 if (!PageUptodate(page)) {
959 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
960 if (ret) {
961 unlock_page(page);
962 goto err;
963 }
964 }
c9de560d
AT		 965 unlock_page(page);
966 }
967 }
fdf6c7a7
SF		 968 if (page == NULL || !PageUptodate(page)) {
969 ret = -EIO;
c9de560d 970 goto err;
fdf6c7a7 971 }
c9de560d
AT		 972 e4b->bd_buddy_page = page;
973 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
974 mark_page_accessed(page);
975
976 BUG_ON(e4b->bd_bitmap_page == NULL);
977 BUG_ON(e4b->bd_buddy_page == NULL);
978
979 return 0;
980
981err:
982 if (e4b->bd_bitmap_page)
983 page_cache_release(e4b->bd_bitmap_page);
984 if (e4b->bd_buddy_page)
985 page_cache_release(e4b->bd_buddy_page);
986 e4b->bd_buddy = NULL;
987 e4b->bd_bitmap = NULL;
fdf6c7a7 988 return ret;
c9de560d
AT		 989}
990
991static void ext4_mb_release_desc(struct ext4_buddy *e4b)
992{
993 if (e4b->bd_bitmap_page)
994 page_cache_release(e4b->bd_bitmap_page);
995 if (e4b->bd_buddy_page)
996 page_cache_release(e4b->bd_buddy_page);
997}
998
999
1000static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1001{
1002 int order = 1;
1003 void *bb;
1004
1005 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1006 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1007
1008 bb = EXT4_MB_BUDDY(e4b);
1009 while (order <= e4b->bd_blkbits + 1) {
1010 block = block >> 1;
1011 if (!mb_test_bit(block, bb)) {
1012 /* this block is part of buddy of order 'order' */
1013 return order;
1014 }
1015 bb += 1 << (e4b->bd_blkbits - order);
1016 order++;
1017 }
1018 return 0;
1019}
1020
1021static void mb_clear_bits(spinlock_t *lock, void *bm, int cur, int len)
1022{
1023 __u32 *addr;
1024
1025 len = cur + len;
1026 while (cur < len) {
1027 if ((cur & 31) == 0 && (len - cur) >= 32) {
1028 /* fast path: clear whole word at once */
1029 addr = bm + (cur >> 3);
1030 *addr = 0;
1031 cur += 32;
1032 continue;
1033 }
1034 mb_clear_bit_atomic(lock, cur, bm);
1035 cur++;
1036 }
1037}
1038
1039static void mb_set_bits(spinlock_t *lock, void *bm, int cur, int len)
1040{
1041 __u32 *addr;
1042
1043 len = cur + len;
1044 while (cur < len) {
1045 if ((cur & 31) == 0 && (len - cur) >= 32) {
1046 /* fast path: set whole word at once */
1047 addr = bm + (cur >> 3);
1048 *addr = 0xffffffff;
1049 cur += 32;
1050 continue;
1051 }
1052 mb_set_bit_atomic(lock, cur, bm);
1053 cur++;
1054 }
1055}
1056
7e5a8cdd 1057static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
c9de560d
AT		 1058 int first, int count)
1059{
1060 int block = 0;
1061 int max = 0;
1062 int order;
1063 void *buddy;
1064 void *buddy2;
1065 struct super_block *sb = e4b->bd_sb;
1066
1067 BUG_ON(first + count > (sb->s_blocksize << 3));
1068 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
1069 mb_check_buddy(e4b);
1070 mb_free_blocks_double(inode, e4b, first, count);
1071
1072 e4b->bd_info->bb_free += count;
1073 if (first < e4b->bd_info->bb_first_free)
1074 e4b->bd_info->bb_first_free = first;
1075
1076 /* let's maintain fragments counter */
1077 if (first != 0)
1078 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1079 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1080 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1081 if (block && max)
1082 e4b->bd_info->bb_fragments--;
1083 else if (!block && !max)
1084 e4b->bd_info->bb_fragments++;
1085
1086 /* let's maintain buddy itself */
1087 while (count-- > 0) {
1088 block = first++;
1089 order = 0;
1090
1091 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1092 ext4_fsblk_t blocknr;
1093 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1094 blocknr += block;
1095 blocknr +=
1096 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
7e5a8cdd 1097 ext4_unlock_group(sb, e4b->bd_group);
46e665e9 1098 ext4_error(sb, __func__, "double-free of inode"
c9de560d
AT		 1099 " %lu's block %llu(bit %u in group %lu)\n",
1100 inode ? inode->i_ino : 0, blocknr, block,
1101 e4b->bd_group);
7e5a8cdd 1102 ext4_lock_group(sb, e4b->bd_group);
c9de560d
AT		 1103 }
1104 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1105 e4b->bd_info->bb_counters[order]++;
1106
1107 /* start of the buddy */
1108 buddy = mb_find_buddy(e4b, order, &max);
1109
1110 do {
1111 block &= ~1UL;
1112 if (mb_test_bit(block, buddy) ||
1113 mb_test_bit(block + 1, buddy))
1114 break;
1115
1116 /* both the buddies are free, try to coalesce them */
1117 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1118
1119 if (!buddy2)
1120 break;
1121
1122 if (order > 0) {
1123 /* for special purposes, we don't set
1124 * free bits in bitmap */
1125 mb_set_bit(block, buddy);
1126 mb_set_bit(block + 1, buddy);
1127 }
1128 e4b->bd_info->bb_counters[order]--;
1129 e4b->bd_info->bb_counters[order]--;
1130
1131 block = block >> 1;
1132 order++;
1133 e4b->bd_info->bb_counters[order]++;
1134
1135 mb_clear_bit(block, buddy2);
1136 buddy = buddy2;
1137 } while (1);
1138 }
1139 mb_check_buddy(e4b);
c9de560d
AT		 1140}
1141
1142static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1143 int needed, struct ext4_free_extent *ex)
1144{
1145 int next = block;
1146 int max;
1147 int ord;
1148 void *buddy;
1149
1150 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1151 BUG_ON(ex == NULL);
1152
1153 buddy = mb_find_buddy(e4b, order, &max);
1154 BUG_ON(buddy == NULL);
1155 BUG_ON(block >= max);
1156 if (mb_test_bit(block, buddy)) {
1157 ex->fe_len = 0;
1158 ex->fe_start = 0;
1159 ex->fe_group = 0;
1160 return 0;
1161 }
1162
1163 /* FIXME dorp order completely ? */
1164 if (likely(order == 0)) {
1165 /* find actual order */
1166 order = mb_find_order_for_block(e4b, block);
1167 block = block >> order;
1168 }
1169
1170 ex->fe_len = 1 << order;
1171 ex->fe_start = block << order;
1172 ex->fe_group = e4b->bd_group;
1173
1174 /* calc difference from given start */
1175 next = next - ex->fe_start;
1176 ex->fe_len -= next;
1177 ex->fe_start += next;
1178
1179 while (needed > ex->fe_len &&
1180 (buddy = mb_find_buddy(e4b, order, &max))) {
1181
1182 if (block + 1 >= max)
1183 break;
1184
1185 next = (block + 1) * (1 << order);
1186 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1187 break;
1188
1189 ord = mb_find_order_for_block(e4b, next);
1190
1191 order = ord;
1192 block = next >> order;
1193 ex->fe_len += 1 << order;
1194 }
1195
1196 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1197 return ex->fe_len;
1198}
1199
1200static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1201{
1202 int ord;
1203 int mlen = 0;
1204 int max = 0;
1205 int cur;
1206 int start = ex->fe_start;
1207 int len = ex->fe_len;
1208 unsigned ret = 0;
1209 int len0 = len;
1210 void *buddy;
1211
1212 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1213 BUG_ON(e4b->bd_group != ex->fe_group);
1214 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1215 mb_check_buddy(e4b);
1216 mb_mark_used_double(e4b, start, len);
1217
1218 e4b->bd_info->bb_free -= len;
1219 if (e4b->bd_info->bb_first_free == start)
1220 e4b->bd_info->bb_first_free += len;
1221
1222 /* let's maintain fragments counter */
1223 if (start != 0)
1224 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1225 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1226 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1227 if (mlen && max)
1228 e4b->bd_info->bb_fragments++;
1229 else if (!mlen && !max)
1230 e4b->bd_info->bb_fragments--;
1231
1232 /* let's maintain buddy itself */
1233 while (len) {
1234 ord = mb_find_order_for_block(e4b, start);
1235
1236 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1237 /* the whole chunk may be allocated at once! */
1238 mlen = 1 << ord;
1239 buddy = mb_find_buddy(e4b, ord, &max);
1240 BUG_ON((start >> ord) >= max);
1241 mb_set_bit(start >> ord, buddy);
1242 e4b->bd_info->bb_counters[ord]--;
1243 start += mlen;
1244 len -= mlen;
1245 BUG_ON(len < 0);
1246 continue;
1247 }
1248
1249 /* store for history */
1250 if (ret == 0)
1251 ret = len | (ord << 16);
1252
1253 /* we have to split large buddy */
1254 BUG_ON(ord <= 0);
1255 buddy = mb_find_buddy(e4b, ord, &max);
1256 mb_set_bit(start >> ord, buddy);
1257 e4b->bd_info->bb_counters[ord]--;
1258
1259 ord--;
1260 cur = (start >> ord) & ~1U;
1261 buddy = mb_find_buddy(e4b, ord, &max);
1262 mb_clear_bit(cur, buddy);
1263 mb_clear_bit(cur + 1, buddy);
1264 e4b->bd_info->bb_counters[ord]++;
1265 e4b->bd_info->bb_counters[ord]++;
1266 }
1267
1268 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b->bd_sb), ex->fe_group),
1269 EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1270 mb_check_buddy(e4b);
1271
1272 return ret;
1273}
1274
1275/*
1276 * Must be called under group lock!
1277 */
1278static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1279 struct ext4_buddy *e4b)
1280{
1281 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1282 int ret;
1283
1284 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1285 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1286
1287 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1288 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1289 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1290
1291 /* preallocation can change ac_b_ex, thus we store actually
1292 * allocated blocks for history */
1293 ac->ac_f_ex = ac->ac_b_ex;
1294
1295 ac->ac_status = AC_STATUS_FOUND;
1296 ac->ac_tail = ret & 0xffff;
1297 ac->ac_buddy = ret >> 16;
1298
1299 /* XXXXXXX: SUCH A HORRIBLE **CK */
1300 /*FIXME!! Why ? */
1301 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1302 get_page(ac->ac_bitmap_page);
1303 ac->ac_buddy_page = e4b->bd_buddy_page;
1304 get_page(ac->ac_buddy_page);
1305
1306 /* store last allocated for subsequent stream allocation */
1307 if ((ac->ac_flags & EXT4_MB_HINT_DATA)) {
1308 spin_lock(&sbi->s_md_lock);
1309 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1310 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1311 spin_unlock(&sbi->s_md_lock);
1312 }
1313}
1314
1315/*
1316 * regular allocator, for general purposes allocation
1317 */
1318
1319static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1320 struct ext4_buddy *e4b,
1321 int finish_group)
1322{
1323 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1324 struct ext4_free_extent *bex = &ac->ac_b_ex;
1325 struct ext4_free_extent *gex = &ac->ac_g_ex;
1326 struct ext4_free_extent ex;
1327 int max;
1328
1329 /*
1330 * We don't want to scan for a whole year
1331 */
1332 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1333 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1334 ac->ac_status = AC_STATUS_BREAK;
1335 return;
1336 }
1337
1338 /*
1339 * Haven't found good chunk so far, let's continue
1340 */
1341 if (bex->fe_len < gex->fe_len)
1342 return;
1343
1344 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1345 && bex->fe_group == e4b->bd_group) {
1346 /* recheck chunk's availability - we don't know
1347 * when it was found (within this lock-unlock
1348 * period or not) */
1349 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1350 if (max >= gex->fe_len) {
1351 ext4_mb_use_best_found(ac, e4b);
1352 return;
1353 }
1354 }
1355}
1356
1357/*
1358 * The routine checks whether found extent is good enough. If it is,
1359 * then the extent gets marked used and flag is set to the context
1360 * to stop scanning. Otherwise, the extent is compared with the
1361 * previous found extent and if new one is better, then it's stored
1362 * in the context. Later, the best found extent will be used, if
1363 * mballoc can't find good enough extent.
1364 *
1365 * FIXME: real allocation policy is to be designed yet!
1366 */
1367static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1368 struct ext4_free_extent *ex,
1369 struct ext4_buddy *e4b)
1370{
1371 struct ext4_free_extent *bex = &ac->ac_b_ex;
1372 struct ext4_free_extent *gex = &ac->ac_g_ex;
1373
1374 BUG_ON(ex->fe_len <= 0);
1375 BUG_ON(ex->fe_len >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1376 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1377 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1378
1379 ac->ac_found++;
1380
1381 /*
1382 * The special case - take what you catch first
1383 */
1384 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1385 *bex = *ex;
1386 ext4_mb_use_best_found(ac, e4b);
1387 return;
1388 }
1389
1390 /*
1391 * Let's check whether the chuck is good enough
1392 */
1393 if (ex->fe_len == gex->fe_len) {
1394 *bex = *ex;
1395 ext4_mb_use_best_found(ac, e4b);
1396 return;
1397 }
1398
1399 /*
1400 * If this is first found extent, just store it in the context
1401 */
1402 if (bex->fe_len == 0) {
1403 *bex = *ex;
1404 return;
1405 }
1406
1407 /*
1408 * If new found extent is better, store it in the context
1409 */
1410 if (bex->fe_len < gex->fe_len) {
1411 /* if the request isn't satisfied, any found extent
1412 * larger than previous best one is better */
1413 if (ex->fe_len > bex->fe_len)
1414 *bex = *ex;
1415 } else if (ex->fe_len > gex->fe_len) {
1416 /* if the request is satisfied, then we try to find
1417 * an extent that still satisfy the request, but is
1418 * smaller than previous one */
1419 if (ex->fe_len < bex->fe_len)
1420 *bex = *ex;
1421 }
1422
1423 ext4_mb_check_limits(ac, e4b, 0);
1424}
1425
1426static int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1427 struct ext4_buddy *e4b)
1428{
1429 struct ext4_free_extent ex = ac->ac_b_ex;
1430 ext4_group_t group = ex.fe_group;
1431 int max;
1432 int err;
1433
1434 BUG_ON(ex.fe_len <= 0);
1435 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1436 if (err)
1437 return err;
1438
1439 ext4_lock_group(ac->ac_sb, group);
1440 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1441
1442 if (max > 0) {
1443 ac->ac_b_ex = ex;
1444 ext4_mb_use_best_found(ac, e4b);
1445 }
1446
1447 ext4_unlock_group(ac->ac_sb, group);
1448 ext4_mb_release_desc(e4b);
1449
1450 return 0;
1451}
1452
1453static int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1454 struct ext4_buddy *e4b)
1455{
1456 ext4_group_t group = ac->ac_g_ex.fe_group;
1457 int max;
1458 int err;
1459 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1460 struct ext4_super_block *es = sbi->s_es;
1461 struct ext4_free_extent ex;
1462
1463 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1464 return 0;
1465
1466 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1467 if (err)
1468 return err;
1469
1470 ext4_lock_group(ac->ac_sb, group);
1471 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1472 ac->ac_g_ex.fe_len, &ex);
1473
1474 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1475 ext4_fsblk_t start;
1476
1477 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1478 ex.fe_start + le32_to_cpu(es->s_first_data_block);
1479 /* use do_div to get remainder (would be 64-bit modulo) */
1480 if (do_div(start, sbi->s_stripe) == 0) {
1481 ac->ac_found++;
1482 ac->ac_b_ex = ex;
1483 ext4_mb_use_best_found(ac, e4b);
1484 }
1485 } else if (max >= ac->ac_g_ex.fe_len) {
1486 BUG_ON(ex.fe_len <= 0);
1487 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1488 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1489 ac->ac_found++;
1490 ac->ac_b_ex = ex;
1491 ext4_mb_use_best_found(ac, e4b);
1492 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1493 /* Sometimes, caller may want to merge even small
1494 * number of blocks to an existing extent */
1495 BUG_ON(ex.fe_len <= 0);
1496 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1497 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1498 ac->ac_found++;
1499 ac->ac_b_ex = ex;
1500 ext4_mb_use_best_found(ac, e4b);
1501 }
1502 ext4_unlock_group(ac->ac_sb, group);
1503 ext4_mb_release_desc(e4b);
1504
1505 return 0;
1506}
1507
1508/*
1509 * The routine scans buddy structures (not bitmap!) from given order
1510 * to max order and tries to find big enough chunk to satisfy the req
1511 */
1512static void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1513 struct ext4_buddy *e4b)
1514{
1515 struct super_block *sb = ac->ac_sb;
1516 struct ext4_group_info *grp = e4b->bd_info;
1517 void *buddy;
1518 int i;
1519 int k;
1520 int max;
1521
1522 BUG_ON(ac->ac_2order <= 0);
1523 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1524 if (grp->bb_counters[i] == 0)
1525 continue;
1526
1527 buddy = mb_find_buddy(e4b, i, &max);
1528 BUG_ON(buddy == NULL);
1529
ffad0a44 1530 k = mb_find_next_zero_bit(buddy, max, 0);
c9de560d
AT		 1531 BUG_ON(k >= max);
1532
1533 ac->ac_found++;
1534
1535 ac->ac_b_ex.fe_len = 1 << i;
1536 ac->ac_b_ex.fe_start = k << i;
1537 ac->ac_b_ex.fe_group = e4b->bd_group;
1538
1539 ext4_mb_use_best_found(ac, e4b);
1540
1541 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1542
1543 if (EXT4_SB(sb)->s_mb_stats)
1544 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1545
1546 break;
1547 }
1548}
1549
1550/*
1551 * The routine scans the group and measures all found extents.
1552 * In order to optimize scanning, caller must pass number of
1553 * free blocks in the group, so the routine can know upper limit.
1554 */
1555static void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1556 struct ext4_buddy *e4b)
1557{
1558 struct super_block *sb = ac->ac_sb;
1559 void *bitmap = EXT4_MB_BITMAP(e4b);
1560 struct ext4_free_extent ex;
1561 int i;
1562 int free;
1563
1564 free = e4b->bd_info->bb_free;
1565 BUG_ON(free <= 0);
1566
1567 i = e4b->bd_info->bb_first_free;
1568
1569 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
ffad0a44 1570 i = mb_find_next_zero_bit(bitmap,
c9de560d
AT		 1571 EXT4_BLOCKS_PER_GROUP(sb), i);
1572 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
26346ff6 1573 /*
e56eb659 1574 * IF we have corrupt bitmap, we won't find any
26346ff6
AK		 1575 * free blocks even though group info says we
1576 * we have free blocks
1577 */
46e665e9 1578 ext4_error(sb, __func__, "%d free blocks as per "
26346ff6
AK		 1579 "group info. But bitmap says 0\n",
1580 free);
c9de560d
AT		 1581 break;
1582 }
1583
1584 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1585 BUG_ON(ex.fe_len <= 0);
26346ff6 1586 if (free < ex.fe_len) {
46e665e9 1587 ext4_error(sb, __func__, "%d free blocks as per "
26346ff6
AK		 1588 "group info. But got %d blocks\n",
1589 free, ex.fe_len);
e56eb659
AK		 1590 /*
1591 * The number of free blocks differs. This mostly
1592 * indicate that the bitmap is corrupt. So exit
1593 * without claiming the space.
1594 */
1595 break;
26346ff6 1596 }
c9de560d
AT		 1597
1598 ext4_mb_measure_extent(ac, &ex, e4b);
1599
1600 i += ex.fe_len;
1601 free -= ex.fe_len;
1602 }
1603
1604 ext4_mb_check_limits(ac, e4b, 1);
1605}
1606
1607/*
1608 * This is a special case for storages like raid5
1609 * we try to find stripe-aligned chunks for stripe-size requests
1610 * XXX should do so at least for multiples of stripe size as well
1611 */
1612static void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1613 struct ext4_buddy *e4b)
1614{
1615 struct super_block *sb = ac->ac_sb;
1616 struct ext4_sb_info *sbi = EXT4_SB(sb);
1617 void *bitmap = EXT4_MB_BITMAP(e4b);
1618 struct ext4_free_extent ex;
1619 ext4_fsblk_t first_group_block;
1620 ext4_fsblk_t a;
1621 ext4_grpblk_t i;
1622 int max;
1623
1624 BUG_ON(sbi->s_stripe == 0);
1625
1626 /* find first stripe-aligned block in group */
1627 first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1628 + le32_to_cpu(sbi->s_es->s_first_data_block);
1629 a = first_group_block + sbi->s_stripe - 1;
1630 do_div(a, sbi->s_stripe);
1631 i = (a * sbi->s_stripe) - first_group_block;
1632
1633 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1634 if (!mb_test_bit(i, bitmap)) {
1635 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1636 if (max >= sbi->s_stripe) {
1637 ac->ac_found++;
1638 ac->ac_b_ex = ex;
1639 ext4_mb_use_best_found(ac, e4b);
1640 break;
1641 }
1642 }
1643 i += sbi->s_stripe;
1644 }
1645}
1646
1647static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1648 ext4_group_t group, int cr)
1649{
1650 unsigned free, fragments;
1651 unsigned i, bits;
1652 struct ext4_group_desc *desc;
1653 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1654
1655 BUG_ON(cr < 0 || cr >= 4);
1656 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1657
1658 free = grp->bb_free;
1659 fragments = grp->bb_fragments;
1660 if (free == 0)
1661 return 0;
1662 if (fragments == 0)
1663 return 0;
1664
1665 switch (cr) {
1666 case 0:
1667 BUG_ON(ac->ac_2order == 0);
1668 /* If this group is uninitialized, skip it initially */
1669 desc = ext4_get_group_desc(ac->ac_sb, group, NULL);
1670 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
1671 return 0;
1672
1673 bits = ac->ac_sb->s_blocksize_bits + 1;
1674 for (i = ac->ac_2order; i <= bits; i++)
1675 if (grp->bb_counters[i] > 0)
1676 return 1;
1677 break;
1678 case 1:
1679 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1680 return 1;
1681 break;
1682 case 2:
1683 if (free >= ac->ac_g_ex.fe_len)
1684 return 1;
1685 break;
1686 case 3:
1687 return 1;
1688 default:
1689 BUG();
1690 }
1691
1692 return 0;
1693}
1694
4ddfef7b
ES		 1695static noinline_for_stack int
1696ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
c9de560d
AT		 1697{
1698 ext4_group_t group;
1699 ext4_group_t i;
1700 int cr;
1701 int err = 0;
1702 int bsbits;
1703 struct ext4_sb_info *sbi;
1704 struct super_block *sb;
1705 struct ext4_buddy e4b;
1706 loff_t size, isize;
1707
1708 sb = ac->ac_sb;
1709 sbi = EXT4_SB(sb);
1710 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1711
1712 /* first, try the goal */
1713 err = ext4_mb_find_by_goal(ac, &e4b);
1714 if (err || ac->ac_status == AC_STATUS_FOUND)
1715 goto out;
1716
1717 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1718 goto out;
1719
1720 /*
1721 * ac->ac2_order is set only if the fe_len is a power of 2
1722 * if ac2_order is set we also set criteria to 0 so that we
1723 * try exact allocation using buddy.
1724 */
1725 i = fls(ac->ac_g_ex.fe_len);
1726 ac->ac_2order = 0;
1727 /*
1728 * We search using buddy data only if the order of the request
1729 * is greater than equal to the sbi_s_mb_order2_reqs
1730 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1731 */
1732 if (i >= sbi->s_mb_order2_reqs) {
1733 /*
1734 * This should tell if fe_len is exactly power of 2
1735 */
1736 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1737 ac->ac_2order = i - 1;
1738 }
1739
1740 bsbits = ac->ac_sb->s_blocksize_bits;
1741 /* if stream allocation is enabled, use global goal */
1742 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
1743 isize = i_size_read(ac->ac_inode) >> bsbits;
1744 if (size < isize)
1745 size = isize;
1746
1747 if (size < sbi->s_mb_stream_request &&
1748 (ac->ac_flags & EXT4_MB_HINT_DATA)) {
1749 /* TBD: may be hot point */
1750 spin_lock(&sbi->s_md_lock);
1751 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1752 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1753 spin_unlock(&sbi->s_md_lock);
1754 }
c9de560d
AT		 1755 /* Let's just scan groups to find more-less suitable blocks */
1756 cr = ac->ac_2order ? 0 : 1;
1757 /*
1758 * cr == 0 try to get exact allocation,
1759 * cr == 3 try to get anything
1760 */
1761repeat:
1762 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1763 ac->ac_criteria = cr;
ed8f9c75
AK		 1764 /*
1765 * searching for the right group start
1766 * from the goal value specified
1767 */
1768 group = ac->ac_g_ex.fe_group;
1769
c9de560d
AT		 1770 for (i = 0; i < EXT4_SB(sb)->s_groups_count; group++, i++) {
1771 struct ext4_group_info *grp;
1772 struct ext4_group_desc *desc;
1773
1774 if (group == EXT4_SB(sb)->s_groups_count)
1775 group = 0;
1776
1777 /* quick check to skip empty groups */
1778 grp = ext4_get_group_info(ac->ac_sb, group);
1779 if (grp->bb_free == 0)
1780 continue;
1781
1782 /*
1783 * if the group is already init we check whether it is
1784 * a good group and if not we don't load the buddy
1785 */
1786 if (EXT4_MB_GRP_NEED_INIT(grp)) {
1787 /*
1788 * we need full data about the group
1789 * to make a good selection
1790 */
1791 err = ext4_mb_load_buddy(sb, group, &e4b);
1792 if (err)
1793 goto out;
1794 ext4_mb_release_desc(&e4b);
1795 }
1796
1797 /*
1798 * If the particular group doesn't satisfy our
1799 * criteria we continue with the next group
1800 */
1801 if (!ext4_mb_good_group(ac, group, cr))
1802 continue;
1803
1804 err = ext4_mb_load_buddy(sb, group, &e4b);
1805 if (err)
1806 goto out;
1807
1808 ext4_lock_group(sb, group);
1809 if (!ext4_mb_good_group(ac, group, cr)) {
1810 /* someone did allocation from this group */
1811 ext4_unlock_group(sb, group);
1812 ext4_mb_release_desc(&e4b);
1813 continue;
1814 }
1815
1816 ac->ac_groups_scanned++;
1817 desc = ext4_get_group_desc(sb, group, NULL);
1818 if (cr == 0 || (desc->bg_flags &
1819 cpu_to_le16(EXT4_BG_BLOCK_UNINIT) &&
1820 ac->ac_2order != 0))
1821 ext4_mb_simple_scan_group(ac, &e4b);
1822 else if (cr == 1 &&
1823 ac->ac_g_ex.fe_len == sbi->s_stripe)
1824 ext4_mb_scan_aligned(ac, &e4b);
1825 else
1826 ext4_mb_complex_scan_group(ac, &e4b);
1827
1828 ext4_unlock_group(sb, group);
1829 ext4_mb_release_desc(&e4b);
1830
1831 if (ac->ac_status != AC_STATUS_CONTINUE)
1832 break;
1833 }
1834 }
1835
1836 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
1837 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1838 /*
1839 * We've been searching too long. Let's try to allocate
1840 * the best chunk we've found so far
1841 */
1842
1843 ext4_mb_try_best_found(ac, &e4b);
1844 if (ac->ac_status != AC_STATUS_FOUND) {
1845 /*
1846 * Someone more lucky has already allocated it.
1847 * The only thing we can do is just take first
1848 * found block(s)
1849 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
1850 */
1851 ac->ac_b_ex.fe_group = 0;
1852 ac->ac_b_ex.fe_start = 0;
1853 ac->ac_b_ex.fe_len = 0;
1854 ac->ac_status = AC_STATUS_CONTINUE;
1855 ac->ac_flags |= EXT4_MB_HINT_FIRST;
1856 cr = 3;
1857 atomic_inc(&sbi->s_mb_lost_chunks);
1858 goto repeat;
1859 }
1860 }
1861out:
1862 return err;
1863}
1864
1865#ifdef EXT4_MB_HISTORY
1866struct ext4_mb_proc_session {
1867 struct ext4_mb_history *history;
1868 struct super_block *sb;
1869 int start;
1870 int max;
1871};
1872
1873static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
1874 struct ext4_mb_history *hs,
1875 int first)
1876{
1877 if (hs == s->history + s->max)
1878 hs = s->history;
1879 if (!first && hs == s->history + s->start)
1880 return NULL;
1881 while (hs->orig.fe_len == 0) {
1882 hs++;
1883 if (hs == s->history + s->max)
1884 hs = s->history;
1885 if (hs == s->history + s->start)
1886 return NULL;
1887 }
1888 return hs;
1889}
1890
1891static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
1892{
1893 struct ext4_mb_proc_session *s = seq->private;
1894 struct ext4_mb_history *hs;
1895 int l = *pos;
1896
1897 if (l == 0)
1898 return SEQ_START_TOKEN;
1899 hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
1900 if (!hs)
1901 return NULL;
1902 while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
1903 return hs;
1904}
1905
1906static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
1907 loff_t *pos)
1908{
1909 struct ext4_mb_proc_session *s = seq->private;
1910 struct ext4_mb_history *hs = v;
1911
1912 ++*pos;
1913 if (v == SEQ_START_TOKEN)
1914 return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
1915 else
1916 return ext4_mb_history_skip_empty(s, ++hs, 0);
1917}
1918
1919static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
1920{
1921 char buf[25], buf2[25], buf3[25], *fmt;
1922 struct ext4_mb_history *hs = v;
1923
1924 if (v == SEQ_START_TOKEN) {
1925 seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
1926 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
1927 "pid", "inode", "original", "goal", "result", "found",
1928 "grps", "cr", "flags", "merge", "tail", "broken");
1929 return 0;
1930 }
1931
1932 if (hs->op == EXT4_MB_HISTORY_ALLOC) {
1933 fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
1934 "%-5u %-5s %-5u %-6u\n";
1935 sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
1936 hs->result.fe_start, hs->result.fe_len,
1937 hs->result.fe_logical);
1938 sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
1939 hs->orig.fe_start, hs->orig.fe_len,
1940 hs->orig.fe_logical);
1941 sprintf(buf3, "%lu/%d/%u@%u", hs->goal.fe_group,
1942 hs->goal.fe_start, hs->goal.fe_len,
1943 hs->goal.fe_logical);
1944 seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
1945 hs->found, hs->groups, hs->cr, hs->flags,
1946 hs->merged ? "M" : "", hs->tail,
1947 hs->buddy ? 1 << hs->buddy : 0);
1948 } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
1949 fmt = "%-5u %-8u %-23s %-23s %-23s\n";
1950 sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
1951 hs->result.fe_start, hs->result.fe_len,
1952 hs->result.fe_logical);
1953 sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
1954 hs->orig.fe_start, hs->orig.fe_len,
1955 hs->orig.fe_logical);
1956 seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
1957 } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
1958 sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
1959 hs->result.fe_start, hs->result.fe_len);
1960 seq_printf(seq, "%-5u %-8u %-23s discard\n",
1961 hs->pid, hs->ino, buf2);
1962 } else if (hs->op == EXT4_MB_HISTORY_FREE) {
1963 sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
1964 hs->result.fe_start, hs->result.fe_len);
1965 seq_printf(seq, "%-5u %-8u %-23s free\n",
1966 hs->pid, hs->ino, buf2);
1967 }
1968 return 0;
1969}
1970
1971static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
1972{
1973}
1974
1975static struct seq_operations ext4_mb_seq_history_ops = {
1976 .start = ext4_mb_seq_history_start,
1977 .next = ext4_mb_seq_history_next,
1978 .stop = ext4_mb_seq_history_stop,
1979 .show = ext4_mb_seq_history_show,
1980};
1981
1982static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
1983{
1984 struct super_block *sb = PDE(inode)->data;
1985 struct ext4_sb_info *sbi = EXT4_SB(sb);
1986 struct ext4_mb_proc_session *s;
1987 int rc;
1988 int size;
1989
74767c5a
SF		 1990 if (unlikely(sbi->s_mb_history == NULL))
1991 return -ENOMEM;
c9de560d
AT		 1992 s = kmalloc(sizeof(*s), GFP_KERNEL);
1993 if (s == NULL)
1994 return -ENOMEM;
1995 s->sb = sb;
1996 size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
1997 s->history = kmalloc(size, GFP_KERNEL);
1998 if (s->history == NULL) {
1999 kfree(s);
2000 return -ENOMEM;
2001 }
2002
2003 spin_lock(&sbi->s_mb_history_lock);
2004 memcpy(s->history, sbi->s_mb_history, size);
2005 s->max = sbi->s_mb_history_max;
2006 s->start = sbi->s_mb_history_cur % s->max;
2007 spin_unlock(&sbi->s_mb_history_lock);
2008
2009 rc = seq_open(file, &ext4_mb_seq_history_ops);
2010 if (rc == 0) {
2011 struct seq_file *m = (struct seq_file *)file->private_data;
2012 m->private = s;
2013 } else {
2014 kfree(s->history);
2015 kfree(s);
2016 }
2017 return rc;
2018
2019}
2020
2021static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
2022{
2023 struct seq_file *seq = (struct seq_file *)file->private_data;
2024 struct ext4_mb_proc_session *s = seq->private;
2025 kfree(s->history);
2026 kfree(s);
2027 return seq_release(inode, file);
2028}
2029
2030static ssize_t ext4_mb_seq_history_write(struct file *file,
2031 const char __user *buffer,
2032 size_t count, loff_t *ppos)
2033{
2034 struct seq_file *seq = (struct seq_file *)file->private_data;
2035 struct ext4_mb_proc_session *s = seq->private;
2036 struct super_block *sb = s->sb;
2037 char str[32];
2038 int value;
2039
2040 if (count >= sizeof(str)) {
2041 printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
2042 "mb_history", (int)sizeof(str));
2043 return -EOVERFLOW;
2044 }
2045
2046 if (copy_from_user(str, buffer, count))
2047 return -EFAULT;
2048
2049 value = simple_strtol(str, NULL, 0);
2050 if (value < 0)
2051 return -ERANGE;
2052 EXT4_SB(sb)->s_mb_history_filter = value;
2053
2054 return count;
2055}
2056
2057static struct file_operations ext4_mb_seq_history_fops = {
2058 .owner = THIS_MODULE,
2059 .open = ext4_mb_seq_history_open,
2060 .read = seq_read,
2061 .write = ext4_mb_seq_history_write,
2062 .llseek = seq_lseek,
2063 .release = ext4_mb_seq_history_release,
2064};
2065
2066static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2067{
2068 struct super_block *sb = seq->private;
2069 struct ext4_sb_info *sbi = EXT4_SB(sb);
2070 ext4_group_t group;
2071
2072 if (*pos < 0 || *pos >= sbi->s_groups_count)
2073 return NULL;
2074
2075 group = *pos + 1;
2076 return (void *) group;
2077}
2078
2079static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2080{
2081 struct super_block *sb = seq->private;
2082 struct ext4_sb_info *sbi = EXT4_SB(sb);
2083 ext4_group_t group;
2084
2085 ++*pos;
2086 if (*pos < 0 || *pos >= sbi->s_groups_count)
2087 return NULL;
2088 group = *pos + 1;
2089 return (void *) group;;
2090}
2091
2092static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2093{
2094 struct super_block *sb = seq->private;
2095 long group = (long) v;
2096 int i;
2097 int err;
2098 struct ext4_buddy e4b;
2099 struct sg {
2100 struct ext4_group_info info;
2101 unsigned short counters[16];
2102 } sg;
2103
2104 group--;
2105 if (group == 0)
2106 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2107 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2108 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2109 "group", "free", "frags", "first",
2110 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2111 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2112
2113 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2114 sizeof(struct ext4_group_info);
2115 err = ext4_mb_load_buddy(sb, group, &e4b);
2116 if (err) {
2117 seq_printf(seq, "#%-5lu: I/O error\n", group);
2118 return 0;
2119 }
2120 ext4_lock_group(sb, group);
2121 memcpy(&sg, ext4_get_group_info(sb, group), i);
2122 ext4_unlock_group(sb, group);
2123 ext4_mb_release_desc(&e4b);
2124
2125 seq_printf(seq, "#%-5lu: %-5u %-5u %-5u [", group, sg.info.bb_free,
2126 sg.info.bb_fragments, sg.info.bb_first_free);
2127 for (i = 0; i <= 13; i++)
2128 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2129 sg.info.bb_counters[i] : 0);
2130 seq_printf(seq, " ]\n");
2131
2132 return 0;
2133}
2134
2135static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2136{
2137}
2138
2139static struct seq_operations ext4_mb_seq_groups_ops = {
2140 .start = ext4_mb_seq_groups_start,
2141 .next = ext4_mb_seq_groups_next,
2142 .stop = ext4_mb_seq_groups_stop,
2143 .show = ext4_mb_seq_groups_show,
2144};
2145
2146static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2147{
2148 struct super_block *sb = PDE(inode)->data;
2149 int rc;
2150
2151 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2152 if (rc == 0) {
2153 struct seq_file *m = (struct seq_file *)file->private_data;
2154 m->private = sb;
2155 }
2156 return rc;
2157
2158}
2159
2160static struct file_operations ext4_mb_seq_groups_fops = {
2161 .owner = THIS_MODULE,
2162 .open = ext4_mb_seq_groups_open,
2163 .read = seq_read,
2164 .llseek = seq_lseek,
2165 .release = seq_release,
2166};
2167
2168static void ext4_mb_history_release(struct super_block *sb)
2169{
2170 struct ext4_sb_info *sbi = EXT4_SB(sb);
2171
2172 remove_proc_entry("mb_groups", sbi->s_mb_proc);
2173 remove_proc_entry("mb_history", sbi->s_mb_proc);
2174
2175 kfree(sbi->s_mb_history);
2176}
2177
2178static void ext4_mb_history_init(struct super_block *sb)
2179{
2180 struct ext4_sb_info *sbi = EXT4_SB(sb);
2181 int i;
2182
2183 if (sbi->s_mb_proc != NULL) {
46fe74f2
DL		 2184 proc_create_data("mb_history", S_IRUGO, sbi->s_mb_proc,
2185 &ext4_mb_seq_history_fops, sb);
2186 proc_create_data("mb_groups", S_IRUGO, sbi->s_mb_proc,
2187 &ext4_mb_seq_groups_fops, sb);
c9de560d
AT		 2188 }
2189
2190 sbi->s_mb_history_max = 1000;
2191 sbi->s_mb_history_cur = 0;
2192 spin_lock_init(&sbi->s_mb_history_lock);
2193 i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
74767c5a 2194 sbi->s_mb_history = kzalloc(i, GFP_KERNEL);
c9de560d
AT		 2195 /* if we can't allocate history, then we simple won't use it */
2196}
2197
4ddfef7b
ES		 2198static noinline_for_stack void
2199ext4_mb_store_history(struct ext4_allocation_context *ac)
c9de560d
AT		 2200{
2201 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2202 struct ext4_mb_history h;
2203
2204 if (unlikely(sbi->s_mb_history == NULL))
2205 return;
2206
2207 if (!(ac->ac_op & sbi->s_mb_history_filter))
2208 return;
2209
2210 h.op = ac->ac_op;
2211 h.pid = current->pid;
2212 h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2213 h.orig = ac->ac_o_ex;
2214 h.result = ac->ac_b_ex;
2215 h.flags = ac->ac_flags;
2216 h.found = ac->ac_found;
2217 h.groups = ac->ac_groups_scanned;
2218 h.cr = ac->ac_criteria;
2219 h.tail = ac->ac_tail;
2220 h.buddy = ac->ac_buddy;
2221 h.merged = 0;
2222 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
2223 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2224 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2225 h.merged = 1;
2226 h.goal = ac->ac_g_ex;
2227 h.result = ac->ac_f_ex;
2228 }
2229
2230 spin_lock(&sbi->s_mb_history_lock);
2231 memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2232 if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2233 sbi->s_mb_history_cur = 0;
2234 spin_unlock(&sbi->s_mb_history_lock);
2235}
2236
2237#else
2238#define ext4_mb_history_release(sb)
2239#define ext4_mb_history_init(sb)
2240#endif
2241
5f21b0e6
FB		 2242
2243/* Create and initialize ext4_group_info data for the given group. */
2244int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2245 struct ext4_group_desc *desc)
2246{
2247 int i, len;
2248 int metalen = 0;
2249 struct ext4_sb_info *sbi = EXT4_SB(sb);
2250 struct ext4_group_info **meta_group_info;
2251
2252 /*
2253 * First check if this group is the first of a reserved block.
2254 * If it's true, we have to allocate a new table of pointers
2255 * to ext4_group_info structures
2256 */
2257 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2258 metalen = sizeof(*meta_group_info) <<
2259 EXT4_DESC_PER_BLOCK_BITS(sb);
2260 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2261 if (meta_group_info == NULL) {
2262 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2263 "buddy group\n");
2264 goto exit_meta_group_info;
2265 }
2266 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2267 meta_group_info;
2268 }
2269
2270 /*
2271 * calculate needed size. if change bb_counters size,
2272 * don't forget about ext4_mb_generate_buddy()
2273 */
2274 len = offsetof(typeof(**meta_group_info),
2275 bb_counters[sb->s_blocksize_bits + 2]);
2276
2277 meta_group_info =
2278 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2279 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2280
2281 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2282 if (meta_group_info[i] == NULL) {
2283 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2284 goto exit_group_info;
2285 }
2286 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2287 &(meta_group_info[i]->bb_state));
2288
2289 /*
2290 * initialize bb_free to be able to skip
2291 * empty groups without initialization
2292 */
2293 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2294 meta_group_info[i]->bb_free =
2295 ext4_free_blocks_after_init(sb, group, desc);
2296 } else {
2297 meta_group_info[i]->bb_free =
2298 le16_to_cpu(desc->bg_free_blocks_count);
2299 }
2300
2301 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2302
2303#ifdef DOUBLE_CHECK
2304 {
2305 struct buffer_head *bh;
2306 meta_group_info[i]->bb_bitmap =
2307 kmalloc(sb->s_blocksize, GFP_KERNEL);
2308 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2309 bh = ext4_read_block_bitmap(sb, group);
2310 BUG_ON(bh == NULL);
2311 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2312 sb->s_blocksize);
2313 put_bh(bh);
2314 }
2315#endif
2316
2317 return 0;
2318
2319exit_group_info:
2320 /* If a meta_group_info table has been allocated, release it now */
2321 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2322 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2323exit_meta_group_info:
2324 return -ENOMEM;
2325} /* ext4_mb_add_groupinfo */
2326
2327/*
2328 * Add a group to the existing groups.
2329 * This function is used for online resize
2330 */
2331int ext4_mb_add_more_groupinfo(struct super_block *sb, ext4_group_t group,
2332 struct ext4_group_desc *desc)
2333{
2334 struct ext4_sb_info *sbi = EXT4_SB(sb);
2335 struct inode *inode = sbi->s_buddy_cache;
2336 int blocks_per_page;
2337 int block;
2338 int pnum;
2339 struct page *page;
2340 int err;
2341
2342 /* Add group based on group descriptor*/
2343 err = ext4_mb_add_groupinfo(sb, group, desc);
2344 if (err)
2345 return err;
2346
2347 /*
2348 * Cache pages containing dynamic mb_alloc datas (buddy and bitmap
2349 * datas) are set not up to date so that they will be re-initilaized
2350 * during the next call to ext4_mb_load_buddy
2351 */
2352
2353 /* Set buddy page as not up to date */
2354 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
2355 block = group * 2;
2356 pnum = block / blocks_per_page;
2357 page = find_get_page(inode->i_mapping, pnum);
2358 if (page != NULL) {
2359 ClearPageUptodate(page);
2360 page_cache_release(page);
2361 }
2362
2363 /* Set bitmap page as not up to date */
2364 block++;
2365 pnum = block / blocks_per_page;
2366 page = find_get_page(inode->i_mapping, pnum);
2367 if (page != NULL) {
2368 ClearPageUptodate(page);
2369 page_cache_release(page);
2370 }
2371
2372 return 0;
2373}
2374
2375/*
2376 * Update an existing group.
2377 * This function is used for online resize
2378 */
2379void ext4_mb_update_group_info(struct ext4_group_info *grp, ext4_grpblk_t add)
2380{
2381 grp->bb_free += add;
2382}
2383
c9de560d
AT		 2384static int ext4_mb_init_backend(struct super_block *sb)
2385{
2386 ext4_group_t i;
5f21b0e6 2387 int metalen;
c9de560d 2388 struct ext4_sb_info *sbi = EXT4_SB(sb);
5f21b0e6
FB		 2389 struct ext4_super_block *es = sbi->s_es;
2390 int num_meta_group_infos;
2391 int num_meta_group_infos_max;
2392 int array_size;
c9de560d 2393 struct ext4_group_info **meta_group_info;
5f21b0e6
FB		 2394 struct ext4_group_desc *desc;
2395
2396 /* This is the number of blocks used by GDT */
2397 num_meta_group_infos = (sbi->s_groups_count + EXT4_DESC_PER_BLOCK(sb) -
2398 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2399
2400 /*
2401 * This is the total number of blocks used by GDT including
2402 * the number of reserved blocks for GDT.
2403 * The s_group_info array is allocated with this value
2404 * to allow a clean online resize without a complex
2405 * manipulation of pointer.
2406 * The drawback is the unused memory when no resize
2407 * occurs but it's very low in terms of pages
2408 * (see comments below)
2409 * Need to handle this properly when META_BG resizing is allowed
2410 */
2411 num_meta_group_infos_max = num_meta_group_infos +
2412 le16_to_cpu(es->s_reserved_gdt_blocks);
c9de560d 2413
5f21b0e6
FB		 2414 /*
2415 * array_size is the size of s_group_info array. We round it
2416 * to the next power of two because this approximation is done
2417 * internally by kmalloc so we can have some more memory
2418 * for free here (e.g. may be used for META_BG resize).
2419 */
2420 array_size = 1;
2421 while (array_size < sizeof(*sbi->s_group_info) *
2422 num_meta_group_infos_max)
2423 array_size = array_size << 1;
c9de560d
AT		 2424 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2425 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2426 * So a two level scheme suffices for now. */
5f21b0e6 2427 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
c9de560d
AT		 2428 if (sbi->s_group_info == NULL) {
2429 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2430 return -ENOMEM;
2431 }
2432 sbi->s_buddy_cache = new_inode(sb);
2433 if (sbi->s_buddy_cache == NULL) {
2434 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2435 goto err_freesgi;
2436 }
2437 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2438
2439 metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb);
2440 for (i = 0; i < num_meta_group_infos; i++) {
2441 if ((i + 1) == num_meta_group_infos)
2442 metalen = sizeof(*meta_group_info) *
2443 (sbi->s_groups_count -
2444 (i << EXT4_DESC_PER_BLOCK_BITS(sb)));
2445 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2446 if (meta_group_info == NULL) {
2447 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2448 "buddy group\n");
2449 goto err_freemeta;
2450 }
2451 sbi->s_group_info[i] = meta_group_info;
2452 }
2453
c9de560d 2454 for (i = 0; i < sbi->s_groups_count; i++) {
c9de560d
AT		 2455 desc = ext4_get_group_desc(sb, i, NULL);
2456 if (desc == NULL) {
2457 printk(KERN_ERR
2458 "EXT4-fs: can't read descriptor %lu\n", i);
2459 goto err_freebuddy;
2460 }
5f21b0e6
FB		 2461 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2462 goto err_freebuddy;
c9de560d
AT		 2463 }
2464
2465 return 0;
2466
2467err_freebuddy:
f1fa3342 2468 while (i-- > 0)
c9de560d 2469 kfree(ext4_get_group_info(sb, i));
c9de560d
AT		 2470 i = num_meta_group_infos;
2471err_freemeta:
f1fa3342 2472 while (i-- > 0)
c9de560d
AT		 2473 kfree(sbi->s_group_info[i]);
2474 iput(sbi->s_buddy_cache);
2475err_freesgi:
2476 kfree(sbi->s_group_info);
2477 return -ENOMEM;
2478}
2479
2480int ext4_mb_init(struct super_block *sb, int needs_recovery)
2481{
2482 struct ext4_sb_info *sbi = EXT4_SB(sb);
6be2ded1 2483 unsigned i, j;
c9de560d
AT		 2484 unsigned offset;
2485 unsigned max;
74767c5a 2486 int ret;
c9de560d
AT		 2487
2488 if (!test_opt(sb, MBALLOC))
2489 return 0;
2490
2491 i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
2492
2493 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2494 if (sbi->s_mb_offsets == NULL) {
2495 clear_opt(sbi->s_mount_opt, MBALLOC);
2496 return -ENOMEM;
2497 }
2498 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2499 if (sbi->s_mb_maxs == NULL) {
2500 clear_opt(sbi->s_mount_opt, MBALLOC);
2501 kfree(sbi->s_mb_maxs);
2502 return -ENOMEM;
2503 }
2504
2505 /* order 0 is regular bitmap */
2506 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2507 sbi->s_mb_offsets[0] = 0;
2508
2509 i = 1;
2510 offset = 0;
2511 max = sb->s_blocksize << 2;
2512 do {
2513 sbi->s_mb_offsets[i] = offset;
2514 sbi->s_mb_maxs[i] = max;
2515 offset += 1 << (sb->s_blocksize_bits - i);
2516 max = max >> 1;
2517 i++;
2518 } while (i <= sb->s_blocksize_bits + 1);
2519
2520 /* init file for buddy data */
74767c5a
SF		 2521 ret = ext4_mb_init_backend(sb);
2522 if (ret != 0) {
c9de560d
AT		 2523 clear_opt(sbi->s_mount_opt, MBALLOC);
2524 kfree(sbi->s_mb_offsets);
2525 kfree(sbi->s_mb_maxs);
74767c5a 2526 return ret;
c9de560d
AT		 2527 }
2528
2529 spin_lock_init(&sbi->s_md_lock);
2530 INIT_LIST_HEAD(&sbi->s_active_transaction);
2531 INIT_LIST_HEAD(&sbi->s_closed_transaction);
2532 INIT_LIST_HEAD(&sbi->s_committed_transaction);
2533 spin_lock_init(&sbi->s_bal_lock);
2534
2535 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2536 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2537 sbi->s_mb_stats = MB_DEFAULT_STATS;
2538 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2539 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2540 sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
2541 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2542
1320cbcf 2543 i = sizeof(struct ext4_locality_group) * nr_cpu_ids;
c9de560d
AT		 2544 sbi->s_locality_groups = kmalloc(i, GFP_KERNEL);
2545 if (sbi->s_locality_groups == NULL) {
2546 clear_opt(sbi->s_mount_opt, MBALLOC);
2547 kfree(sbi->s_mb_offsets);
2548 kfree(sbi->s_mb_maxs);
2549 return -ENOMEM;
2550 }
1320cbcf 2551 for (i = 0; i < nr_cpu_ids; i++) {
c9de560d
AT		 2552 struct ext4_locality_group *lg;
2553 lg = &sbi->s_locality_groups[i];
2554 mutex_init(&lg->lg_mutex);
6be2ded1
AK		 2555 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2556 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
c9de560d
AT		 2557 spin_lock_init(&lg->lg_prealloc_lock);
2558 }
2559
2560 ext4_mb_init_per_dev_proc(sb);
2561 ext4_mb_history_init(sb);
2562
2563 printk("EXT4-fs: mballoc enabled\n");
2564 return 0;
2565}
2566
2567/* need to called with ext4 group lock (ext4_lock_group) */
2568static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2569{
2570 struct ext4_prealloc_space *pa;
2571 struct list_head *cur, *tmp;
2572 int count = 0;
2573
2574 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2575 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2576 list_del(&pa->pa_group_list);
2577 count++;
2578 kfree(pa);
2579 }
2580 if (count)
2581 mb_debug("mballoc: %u PAs left\n", count);
2582
2583}
2584
2585int ext4_mb_release(struct super_block *sb)
2586{
2587 ext4_group_t i;
2588 int num_meta_group_infos;
2589 struct ext4_group_info *grinfo;
2590 struct ext4_sb_info *sbi = EXT4_SB(sb);
2591
2592 if (!test_opt(sb, MBALLOC))
2593 return 0;
2594
2595 /* release freed, non-committed blocks */
2596 spin_lock(&sbi->s_md_lock);
2597 list_splice_init(&sbi->s_closed_transaction,
2598 &sbi->s_committed_transaction);
2599 list_splice_init(&sbi->s_active_transaction,
2600 &sbi->s_committed_transaction);
2601 spin_unlock(&sbi->s_md_lock);
2602 ext4_mb_free_committed_blocks(sb);
2603
2604 if (sbi->s_group_info) {
2605 for (i = 0; i < sbi->s_groups_count; i++) {
2606 grinfo = ext4_get_group_info(sb, i);
2607#ifdef DOUBLE_CHECK
2608 kfree(grinfo->bb_bitmap);
2609#endif
2610 ext4_lock_group(sb, i);
2611 ext4_mb_cleanup_pa(grinfo);
2612 ext4_unlock_group(sb, i);
2613 kfree(grinfo);
2614 }
2615 num_meta_group_infos = (sbi->s_groups_count +
2616 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2617 EXT4_DESC_PER_BLOCK_BITS(sb);
2618 for (i = 0; i < num_meta_group_infos; i++)
2619 kfree(sbi->s_group_info[i]);
2620 kfree(sbi->s_group_info);
2621 }
2622 kfree(sbi->s_mb_offsets);
2623 kfree(sbi->s_mb_maxs);
2624 if (sbi->s_buddy_cache)
2625 iput(sbi->s_buddy_cache);
2626 if (sbi->s_mb_stats) {
2627 printk(KERN_INFO
2628 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2629 atomic_read(&sbi->s_bal_allocated),
2630 atomic_read(&sbi->s_bal_reqs),
2631 atomic_read(&sbi->s_bal_success));
2632 printk(KERN_INFO
2633 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2634 "%u 2^N hits, %u breaks, %u lost\n",
2635 atomic_read(&sbi->s_bal_ex_scanned),
2636 atomic_read(&sbi->s_bal_goals),
2637 atomic_read(&sbi->s_bal_2orders),
2638 atomic_read(&sbi->s_bal_breaks),
2639 atomic_read(&sbi->s_mb_lost_chunks));
2640 printk(KERN_INFO
2641 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2642 sbi->s_mb_buddies_generated++,
2643 sbi->s_mb_generation_time);
2644 printk(KERN_INFO
2645 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2646 atomic_read(&sbi->s_mb_preallocated),
2647 atomic_read(&sbi->s_mb_discarded));
2648 }
2649
2650 kfree(sbi->s_locality_groups);
2651
2652 ext4_mb_history_release(sb);
2653 ext4_mb_destroy_per_dev_proc(sb);
2654
2655 return 0;
2656}
2657
4ddfef7b
ES		 2658static noinline_for_stack void
2659ext4_mb_free_committed_blocks(struct super_block *sb)
c9de560d
AT		 2660{
2661 struct ext4_sb_info *sbi = EXT4_SB(sb);
2662 int err;
2663 int i;
2664 int count = 0;
2665 int count2 = 0;
2666 struct ext4_free_metadata *md;
2667 struct ext4_buddy e4b;
2668
2669 if (list_empty(&sbi->s_committed_transaction))
2670 return;
2671
2672 /* there is committed blocks to be freed yet */
2673 do {
2674 /* get next array of blocks */
2675 md = NULL;
2676 spin_lock(&sbi->s_md_lock);
2677 if (!list_empty(&sbi->s_committed_transaction)) {
2678 md = list_entry(sbi->s_committed_transaction.next,
2679 struct ext4_free_metadata, list);
2680 list_del(&md->list);
2681 }
2682 spin_unlock(&sbi->s_md_lock);
2683
2684 if (md == NULL)
2685 break;
2686
2687 mb_debug("gonna free %u blocks in group %lu (0x%p):",
2688 md->num, md->group, md);
2689
2690 err = ext4_mb_load_buddy(sb, md->group, &e4b);
2691 /* we expect to find existing buddy because it's pinned */
2692 BUG_ON(err != 0);
2693
2694 /* there are blocks to put in buddy to make them really free */
2695 count += md->num;
2696 count2++;
2697 ext4_lock_group(sb, md->group);
2698 for (i = 0; i < md->num; i++) {
2699 mb_debug(" %u", md->blocks[i]);
7e5a8cdd 2700 mb_free_blocks(NULL, &e4b, md->blocks[i], 1);
c9de560d
AT		 2701 }
2702 mb_debug("\n");
2703 ext4_unlock_group(sb, md->group);
2704
2705 /* balance refcounts from ext4_mb_free_metadata() */
2706 page_cache_release(e4b.bd_buddy_page);
2707 page_cache_release(e4b.bd_bitmap_page);
2708
2709 kfree(md);
2710 ext4_mb_release_desc(&e4b);
2711
2712 } while (md);
2713
2714 mb_debug("freed %u blocks in %u structures\n", count, count2);
2715}
2716
c9de560d
AT		 2717#define EXT4_MB_STATS_NAME "stats"
2718#define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2719#define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2720#define EXT4_MB_ORDER2_REQ "order2_req"
2721#define EXT4_MB_STREAM_REQ "stream_req"
2722#define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2723
2724
2725
91d99827
AD		 2726#define MB_PROC_FOPS(name) \
2727static int ext4_mb_##name##_proc_show(struct seq_file *m, void *v) \
c9de560d 2728{ \
91d99827
AD		 2729 struct ext4_sb_info *sbi = m->private; \
2730 \
2731 seq_printf(m, "%ld\n", sbi->s_mb_##name); \
2732 return 0; \
2733} \
2734 \
2735static int ext4_mb_##name##_proc_open(struct inode *inode, struct file *file)\
2736{ \
2737 return single_open(file, ext4_mb_##name##_proc_show, PDE(inode)->data);\
2738} \
2739 \
2740static ssize_t ext4_mb_##name##_proc_write(struct file *file, \
2741 const char __user *buf, size_t cnt, loff_t *ppos) \
c9de560d 2742{ \
91d99827 2743 struct ext4_sb_info *sbi = PDE(file->f_path.dentry->d_inode)->data;\
c9de560d
AT		 2744 char str[32]; \
2745 long value; \
2746 if (cnt >= sizeof(str)) \
2747 return -EINVAL; \
2748 if (copy_from_user(str, buf, cnt)) \
2749 return -EFAULT; \
2750 value = simple_strtol(str, NULL, 0); \
2751 if (value <= 0) \
2752 return -ERANGE; \
2753 sbi->s_mb_##name = value; \
2754 return cnt; \
91d99827
AD		 2755} \
2756 \
2757static const struct file_operations ext4_mb_##name##_proc_fops = { \
2758 .owner = THIS_MODULE, \
2759 .open = ext4_mb_##name##_proc_open, \
2760 .read = seq_read, \
2761 .llseek = seq_lseek, \
2762 .release = single_release, \
2763 .write = ext4_mb_##name##_proc_write, \
2764};
c9de560d 2765
91d99827
AD		 2766MB_PROC_FOPS(stats);
2767MB_PROC_FOPS(max_to_scan);
2768MB_PROC_FOPS(min_to_scan);
2769MB_PROC_FOPS(order2_reqs);
2770MB_PROC_FOPS(stream_request);
2771MB_PROC_FOPS(group_prealloc);
c9de560d
AT		 2772
2773#define MB_PROC_HANDLER(name, var) \
2774do { \
91d99827
AD		 2775 proc = proc_create_data(name, mode, sbi->s_mb_proc, \
2776 &ext4_mb_##var##_proc_fops, sbi); \
c9de560d
AT		 2777 if (proc == NULL) { \
2778 printk(KERN_ERR "EXT4-fs: can't to create %s\n", name); \
2779 goto err_out; \
2780 } \
c9de560d
AT		 2781} while (0)
2782
2783static int ext4_mb_init_per_dev_proc(struct super_block *sb)
2784{
2785 mode_t mode = S_IFREG | S_IRUGO | S_IWUSR;
2786 struct ext4_sb_info *sbi = EXT4_SB(sb);
2787 struct proc_dir_entry *proc;
2788 char devname[64];
2789
cfbe7e4f
SF		 2790 if (proc_root_ext4 == NULL) {
2791 sbi->s_mb_proc = NULL;
2792 return -EINVAL;
2793 }
f36f21ec 2794 bdevname(sb->s_bdev, devname);
c9de560d
AT		 2795 sbi->s_mb_proc = proc_mkdir(devname, proc_root_ext4);
2796
2797 MB_PROC_HANDLER(EXT4_MB_STATS_NAME, stats);
2798 MB_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME, max_to_scan);
2799 MB_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME, min_to_scan);
2800 MB_PROC_HANDLER(EXT4_MB_ORDER2_REQ, order2_reqs);
2801 MB_PROC_HANDLER(EXT4_MB_STREAM_REQ, stream_request);
2802 MB_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC, group_prealloc);
2803
2804 return 0;
2805
2806err_out:
2807 printk(KERN_ERR "EXT4-fs: Unable to create %s\n", devname);
2808 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_mb_proc);
2809 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_mb_proc);
2810 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_mb_proc);
2811 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
2812 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
2813 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_mb_proc);
2814 remove_proc_entry(devname, proc_root_ext4);
2815 sbi->s_mb_proc = NULL;
2816
2817 return -ENOMEM;
2818}
2819
2820static int ext4_mb_destroy_per_dev_proc(struct super_block *sb)
2821{
2822 struct ext4_sb_info *sbi = EXT4_SB(sb);
2823 char devname[64];
2824
2825 if (sbi->s_mb_proc == NULL)
2826 return -EINVAL;
2827
f36f21ec 2828 bdevname(sb->s_bdev, devname);
c9de560d
AT		 2829 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_mb_proc);
2830 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_mb_proc);
2831 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_mb_proc);
2832 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
2833 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
2834 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_mb_proc);
2835 remove_proc_entry(devname, proc_root_ext4);
2836
2837 return 0;
2838}
2839
2840int __init init_ext4_mballoc(void)
2841{
2842 ext4_pspace_cachep =
2843 kmem_cache_create("ext4_prealloc_space",
2844 sizeof(struct ext4_prealloc_space),
2845 0, SLAB_RECLAIM_ACCOUNT, NULL);
2846 if (ext4_pspace_cachep == NULL)
2847 return -ENOMEM;
2848
256bdb49
ES		 2849 ext4_ac_cachep =
2850 kmem_cache_create("ext4_alloc_context",
2851 sizeof(struct ext4_allocation_context),
2852 0, SLAB_RECLAIM_ACCOUNT, NULL);
2853 if (ext4_ac_cachep == NULL) {
2854 kmem_cache_destroy(ext4_pspace_cachep);
2855 return -ENOMEM;
2856 }
c9de560d 2857#ifdef CONFIG_PROC_FS
36a5aeb8 2858 proc_root_ext4 = proc_mkdir("fs/ext4", NULL);
c9de560d 2859 if (proc_root_ext4 == NULL)
36a5aeb8 2860 printk(KERN_ERR "EXT4-fs: Unable to create fs/ext4\n");
c9de560d 2861#endif
c9de560d
AT		 2862 return 0;
2863}
2864
2865void exit_ext4_mballoc(void)
2866{
2867 /* XXX: synchronize_rcu(); */
2868 kmem_cache_destroy(ext4_pspace_cachep);
256bdb49 2869 kmem_cache_destroy(ext4_ac_cachep);
c9de560d 2870#ifdef CONFIG_PROC_FS
36a5aeb8 2871 remove_proc_entry("fs/ext4", NULL);
c9de560d
AT		 2872#endif
2873}
2874
2875
2876/*
2877 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2878 * Returns 0 if success or error code
2879 */
4ddfef7b
ES		 2880static noinline_for_stack int
2881ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
c9de560d
AT		 2882 handle_t *handle)
2883{
2884 struct buffer_head *bitmap_bh = NULL;
2885 struct ext4_super_block *es;
2886 struct ext4_group_desc *gdp;
2887 struct buffer_head *gdp_bh;
2888 struct ext4_sb_info *sbi;
2889 struct super_block *sb;
2890 ext4_fsblk_t block;
519deca0 2891 int err, len;
c9de560d
AT		 2892
2893 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2894 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2895
2896 sb = ac->ac_sb;
2897 sbi = EXT4_SB(sb);
2898 es = sbi->s_es;
2899
c9de560d
AT		 2900
2901 err = -EIO;
574ca174 2902 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
c9de560d
AT		 2903 if (!bitmap_bh)
2904 goto out_err;
2905
2906 err = ext4_journal_get_write_access(handle, bitmap_bh);
2907 if (err)
2908 goto out_err;
2909
2910 err = -EIO;
2911 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2912 if (!gdp)
2913 goto out_err;
2914
03cddb80
AK		 2915 ext4_debug("using block group %lu(%d)\n", ac->ac_b_ex.fe_group,
2916 gdp->bg_free_blocks_count);
2917
c9de560d
AT		 2918 err = ext4_journal_get_write_access(handle, gdp_bh);
2919 if (err)
2920 goto out_err;
2921
2922 block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
2923 + ac->ac_b_ex.fe_start
2924 + le32_to_cpu(es->s_first_data_block);
2925
519deca0
AK		 2926 len = ac->ac_b_ex.fe_len;
2927 if (in_range(ext4_block_bitmap(sb, gdp), block, len) ||
2928 in_range(ext4_inode_bitmap(sb, gdp), block, len) ||
2929 in_range(block, ext4_inode_table(sb, gdp),
2930 EXT4_SB(sb)->s_itb_per_group) ||
2931 in_range(block + len - 1, ext4_inode_table(sb, gdp),
2932 EXT4_SB(sb)->s_itb_per_group)) {
46e665e9 2933 ext4_error(sb, __func__,
c9de560d
AT		 2934 "Allocating block in system zone - block = %llu",
2935 block);
519deca0
AK		 2936 /* File system mounted not to panic on error
2937 * Fix the bitmap and repeat the block allocation
2938 * We leak some of the blocks here.
2939 */
2940 mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group),
2941 bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2942 ac->ac_b_ex.fe_len);
2943 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
2944 if (!err)
2945 err = -EAGAIN;
2946 goto out_err;
c9de560d
AT		 2947 }
2948#ifdef AGGRESSIVE_CHECK
2949 {
2950 int i;
2951 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2952 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2953 bitmap_bh->b_data));
2954 }
2955 }
2956#endif
2957 mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), bitmap_bh->b_data,
2958 ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
2959
2960 spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
2961 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2962 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2963 gdp->bg_free_blocks_count =
2964 cpu_to_le16(ext4_free_blocks_after_init(sb,
2965 ac->ac_b_ex.fe_group,
2966 gdp));
2967 }
e8546d06 2968 le16_add_cpu(&gdp->bg_free_blocks_count, -ac->ac_b_ex.fe_len);
c9de560d
AT		 2969 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2970 spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
d2a17637
MC		 2971
2972 /*
2973 * free blocks account has already be reduced/reserved
2974 * at write_begin() time for delayed allocation
2975 * do not double accounting
2976 */
2977 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2978 percpu_counter_sub(&sbi->s_freeblocks_counter,
2979 ac->ac_b_ex.fe_len);
c9de560d 2980
772cb7c8
JS		 2981 if (sbi->s_log_groups_per_flex) {
2982 ext4_group_t flex_group = ext4_flex_group(sbi,
2983 ac->ac_b_ex.fe_group);
2984 spin_lock(sb_bgl_lock(sbi, flex_group));
2985 sbi->s_flex_groups[flex_group].free_blocks -= ac->ac_b_ex.fe_len;
2986 spin_unlock(sb_bgl_lock(sbi, flex_group));
2987 }
2988
c9de560d
AT		 2989 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
2990 if (err)
2991 goto out_err;
2992 err = ext4_journal_dirty_metadata(handle, gdp_bh);
2993
2994out_err:
2995 sb->s_dirt = 1;
42a10add 2996 brelse(bitmap_bh);
c9de560d
AT		 2997 return err;
2998}
2999
3000/*
3001 * here we normalize request for locality group
3002 * Group request are normalized to s_strip size if we set the same via mount
3003 * option. If not we set it to s_mb_group_prealloc which can be configured via
3004 * /proc/fs/ext4/<partition>/group_prealloc
3005 *
3006 * XXX: should we try to preallocate more than the group has now?
3007 */
3008static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3009{
3010 struct super_block *sb = ac->ac_sb;
3011 struct ext4_locality_group *lg = ac->ac_lg;
3012
3013 BUG_ON(lg == NULL);
3014 if (EXT4_SB(sb)->s_stripe)
3015 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
3016 else
3017 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
60bd63d1 3018 mb_debug("#%u: goal %u blocks for locality group\n",
c9de560d
AT		 3019 current->pid, ac->ac_g_ex.fe_len);
3020}
3021
3022/*
3023 * Normalization means making request better in terms of
3024 * size and alignment
3025 */
4ddfef7b
ES		 3026static noinline_for_stack void
3027ext4_mb_normalize_request(struct ext4_allocation_context *ac,
c9de560d
AT		 3028 struct ext4_allocation_request *ar)
3029{
3030 int bsbits, max;
3031 ext4_lblk_t end;
c9de560d
AT		 3032 loff_t size, orig_size, start_off;
3033 ext4_lblk_t start, orig_start;
3034 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
9a0762c5 3035 struct ext4_prealloc_space *pa;
c9de560d
AT		 3036
3037 /* do normalize only data requests, metadata requests
3038 do not need preallocation */
3039 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3040 return;
3041
3042 /* sometime caller may want exact blocks */
3043 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3044 return;
3045
3046 /* caller may indicate that preallocation isn't
3047 * required (it's a tail, for example) */
3048 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3049 return;
3050
3051 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3052 ext4_mb_normalize_group_request(ac);
3053 return ;
3054 }
3055
3056 bsbits = ac->ac_sb->s_blocksize_bits;
3057
3058 /* first, let's learn actual file size
3059 * given current request is allocated */
3060 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3061 size = size << bsbits;
3062 if (size < i_size_read(ac->ac_inode))
3063 size = i_size_read(ac->ac_inode);
3064
1930479c
VC		 3065 /* max size of free chunks */
3066 max = 2 << bsbits;
c9de560d 3067
1930479c
VC		 3068#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3069 (req <= (size) || max <= (chunk_size))
c9de560d
AT		 3070
3071 /* first, try to predict filesize */
3072 /* XXX: should this table be tunable? */
3073 start_off = 0;
3074 if (size <= 16 * 1024) {
3075 size = 16 * 1024;
3076 } else if (size <= 32 * 1024) {
3077 size = 32 * 1024;
3078 } else if (size <= 64 * 1024) {
3079 size = 64 * 1024;
3080 } else if (size <= 128 * 1024) {
3081 size = 128 * 1024;
3082 } else if (size <= 256 * 1024) {
3083 size = 256 * 1024;
3084 } else if (size <= 512 * 1024) {
3085 size = 512 * 1024;
3086 } else if (size <= 1024 * 1024) {
3087 size = 1024 * 1024;
1930479c 3088 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
c9de560d 3089 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
1930479c
VC		 3090 (21 - bsbits)) << 21;
3091 size = 2 * 1024 * 1024;
3092 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
c9de560d
AT		 3093 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3094 (22 - bsbits)) << 22;
3095 size = 4 * 1024 * 1024;
3096 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
1930479c 3097 (8<<20)>>bsbits, max, 8 * 1024)) {
c9de560d
AT		 3098 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3099 (23 - bsbits)) << 23;
3100 size = 8 * 1024 * 1024;
3101 } else {
3102 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3103 size = ac->ac_o_ex.fe_len << bsbits;
3104 }
3105 orig_size = size = size >> bsbits;
3106 orig_start = start = start_off >> bsbits;
3107
3108 /* don't cover already allocated blocks in selected range */
3109 if (ar->pleft && start <= ar->lleft) {
3110 size -= ar->lleft + 1 - start;
3111 start = ar->lleft + 1;
3112 }
3113 if (ar->pright && start + size - 1 >= ar->lright)
3114 size -= start + size - ar->lright;
3115
3116 end = start + size;
3117
3118 /* check we don't cross already preallocated blocks */
3119 rcu_read_lock();
9a0762c5 3120 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
c9de560d
AT		 3121 unsigned long pa_end;
3122
c9de560d
AT		 3123 if (pa->pa_deleted)
3124 continue;
3125 spin_lock(&pa->pa_lock);
3126 if (pa->pa_deleted) {
3127 spin_unlock(&pa->pa_lock);
3128 continue;
3129 }
3130
3131 pa_end = pa->pa_lstart + pa->pa_len;
3132
3133 /* PA must not overlap original request */
3134 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3135 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3136
3137 /* skip PA normalized request doesn't overlap with */
3138 if (pa->pa_lstart >= end) {
3139 spin_unlock(&pa->pa_lock);
3140 continue;
3141 }
3142 if (pa_end <= start) {
3143 spin_unlock(&pa->pa_lock);
3144 continue;
3145 }
3146 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3147
3148 if (pa_end <= ac->ac_o_ex.fe_logical) {
3149 BUG_ON(pa_end < start);
3150 start = pa_end;
3151 }
3152
3153 if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3154 BUG_ON(pa->pa_lstart > end);
3155 end = pa->pa_lstart;
3156 }
3157 spin_unlock(&pa->pa_lock);
3158 }
3159 rcu_read_unlock();
3160 size = end - start;
3161
3162 /* XXX: extra loop to check we really don't overlap preallocations */
3163 rcu_read_lock();
9a0762c5 3164 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
c9de560d 3165 unsigned long pa_end;
c9de560d
AT		 3166 spin_lock(&pa->pa_lock);
3167 if (pa->pa_deleted == 0) {
3168 pa_end = pa->pa_lstart + pa->pa_len;
3169 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3170 }
3171 spin_unlock(&pa->pa_lock);
3172 }
3173 rcu_read_unlock();
3174
3175 if (start + size <= ac->ac_o_ex.fe_logical &&
3176 start > ac->ac_o_ex.fe_logical) {
3177 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3178 (unsigned long) start, (unsigned long) size,
3179 (unsigned long) ac->ac_o_ex.fe_logical);
3180 }
3181 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3182 start > ac->ac_o_ex.fe_logical);
3183 BUG_ON(size <= 0 || size >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3184
3185 /* now prepare goal request */
3186
3187 /* XXX: is it better to align blocks WRT to logical
3188 * placement or satisfy big request as is */
3189 ac->ac_g_ex.fe_logical = start;
3190 ac->ac_g_ex.fe_len = size;
3191
3192 /* define goal start in order to merge */
3193 if (ar->pright && (ar->lright == (start + size))) {
3194 /* merge to the right */
3195 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3196 &ac->ac_f_ex.fe_group,
3197 &ac->ac_f_ex.fe_start);
3198 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3199 }
3200 if (ar->pleft && (ar->lleft + 1 == start)) {
3201 /* merge to the left */
3202 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3203 &ac->ac_f_ex.fe_group,
3204 &ac->ac_f_ex.fe_start);
3205 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3206 }
3207
3208 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
3209 (unsigned) orig_size, (unsigned) start);
3210}
3211
3212static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3213{
3214 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3215
3216 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3217 atomic_inc(&sbi->s_bal_reqs);
3218 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3219 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3220 atomic_inc(&sbi->s_bal_success);
3221 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3222 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3223 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3224 atomic_inc(&sbi->s_bal_goals);
3225 if (ac->ac_found > sbi->s_mb_max_to_scan)
3226 atomic_inc(&sbi->s_bal_breaks);
3227 }
3228
3229 ext4_mb_store_history(ac);
3230}
3231
3232/*
3233 * use blocks preallocated to inode
3234 */
3235static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3236 struct ext4_prealloc_space *pa)
3237{
3238 ext4_fsblk_t start;
3239 ext4_fsblk_t end;
3240 int len;
3241
3242 /* found preallocated blocks, use them */
3243 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3244 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3245 len = end - start;
3246 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3247 &ac->ac_b_ex.fe_start);
3248 ac->ac_b_ex.fe_len = len;
3249 ac->ac_status = AC_STATUS_FOUND;
3250 ac->ac_pa = pa;
3251
3252 BUG_ON(start < pa->pa_pstart);
3253 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3254 BUG_ON(pa->pa_free < len);
3255 pa->pa_free -= len;
3256
60bd63d1 3257 mb_debug("use %llu/%u from inode pa %p\n", start, len, pa);
c9de560d
AT		 3258}
3259
3260/*
3261 * use blocks preallocated to locality group
3262 */
3263static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3264 struct ext4_prealloc_space *pa)
3265{
03cddb80 3266 unsigned int len = ac->ac_o_ex.fe_len;
6be2ded1 3267
c9de560d
AT		 3268 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3269 &ac->ac_b_ex.fe_group,
3270 &ac->ac_b_ex.fe_start);
3271 ac->ac_b_ex.fe_len = len;
3272 ac->ac_status = AC_STATUS_FOUND;
3273 ac->ac_pa = pa;
3274
3275 /* we don't correct pa_pstart or pa_plen here to avoid
26346ff6 3276 * possible race when the group is being loaded concurrently
c9de560d 3277 * instead we correct pa later, after blocks are marked
26346ff6
AK		 3278 * in on-disk bitmap -- see ext4_mb_release_context()
3279 * Other CPUs are prevented from allocating from this pa by lg_mutex
c9de560d
AT		 3280 */
3281 mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3282}
3283
5e745b04
AK		 3284/*
3285 * Return the prealloc space that have minimal distance
3286 * from the goal block. @cpa is the prealloc
3287 * space that is having currently known minimal distance
3288 * from the goal block.
3289 */
3290static struct ext4_prealloc_space *
3291ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3292 struct ext4_prealloc_space *pa,
3293 struct ext4_prealloc_space *cpa)
3294{
3295 ext4_fsblk_t cur_distance, new_distance;
3296
3297 if (cpa == NULL) {
3298 atomic_inc(&pa->pa_count);
3299 return pa;
3300 }
3301 cur_distance = abs(goal_block - cpa->pa_pstart);
3302 new_distance = abs(goal_block - pa->pa_pstart);
3303
3304 if (cur_distance < new_distance)
3305 return cpa;
3306
3307 /* drop the previous reference */
3308 atomic_dec(&cpa->pa_count);
3309 atomic_inc(&pa->pa_count);
3310 return pa;
3311}
3312
c9de560d
AT		 3313/*
3314 * search goal blocks in preallocated space
3315 */
4ddfef7b
ES		 3316static noinline_for_stack int
3317ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
c9de560d 3318{
6be2ded1 3319 int order, i;
c9de560d
AT		 3320 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3321 struct ext4_locality_group *lg;
5e745b04
AK		 3322 struct ext4_prealloc_space *pa, *cpa = NULL;
3323 ext4_fsblk_t goal_block;
c9de560d
AT		 3324
3325 /* only data can be preallocated */
3326 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3327 return 0;
3328
3329 /* first, try per-file preallocation */
3330 rcu_read_lock();
9a0762c5 3331 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
c9de560d
AT		 3332
3333 /* all fields in this condition don't change,
3334 * so we can skip locking for them */
3335 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3336 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3337 continue;
3338
3339 /* found preallocated blocks, use them */
3340 spin_lock(&pa->pa_lock);
3341 if (pa->pa_deleted == 0 && pa->pa_free) {
3342 atomic_inc(&pa->pa_count);
3343 ext4_mb_use_inode_pa(ac, pa);
3344 spin_unlock(&pa->pa_lock);
3345 ac->ac_criteria = 10;
3346 rcu_read_unlock();
3347 return 1;
3348 }
3349 spin_unlock(&pa->pa_lock);
3350 }
3351 rcu_read_unlock();
3352
3353 /* can we use group allocation? */
3354 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3355 return 0;
3356
3357 /* inode may have no locality group for some reason */
3358 lg = ac->ac_lg;
3359 if (lg == NULL)
3360 return 0;
6be2ded1
AK		 3361 order = fls(ac->ac_o_ex.fe_len) - 1;
3362 if (order > PREALLOC_TB_SIZE - 1)
3363 /* The max size of hash table is PREALLOC_TB_SIZE */
3364 order = PREALLOC_TB_SIZE - 1;
3365
5e745b04
AK		 3366 goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) +
3367 ac->ac_g_ex.fe_start +
3368 le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block);
3369 /*
3370 * search for the prealloc space that is having
3371 * minimal distance from the goal block.
3372 */
6be2ded1
AK		 3373 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3374 rcu_read_lock();
3375 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3376 pa_inode_list) {
3377 spin_lock(&pa->pa_lock);
3378 if (pa->pa_deleted == 0 &&
3379 pa->pa_free >= ac->ac_o_ex.fe_len) {
5e745b04
AK		 3380
3381 cpa = ext4_mb_check_group_pa(goal_block,
3382 pa, cpa);
6be2ded1 3383 }
c9de560d 3384 spin_unlock(&pa->pa_lock);
c9de560d 3385 }
6be2ded1 3386 rcu_read_unlock();
c9de560d 3387 }
5e745b04
AK		 3388 if (cpa) {
3389 ext4_mb_use_group_pa(ac, cpa);
3390 ac->ac_criteria = 20;
3391 return 1;
3392 }
c9de560d
AT		 3393 return 0;
3394}
3395
3396/*
3397 * the function goes through all preallocation in this group and marks them
3398 * used in in-core bitmap. buddy must be generated from this bitmap
3399 * Need to be called with ext4 group lock (ext4_lock_group)
3400 */
3401static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3402 ext4_group_t group)
3403{
3404 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3405 struct ext4_prealloc_space *pa;
3406 struct list_head *cur;
3407 ext4_group_t groupnr;
3408 ext4_grpblk_t start;
3409 int preallocated = 0;
3410 int count = 0;
3411 int len;
3412
3413 /* all form of preallocation discards first load group,
3414 * so the only competing code is preallocation use.
3415 * we don't need any locking here
3416 * notice we do NOT ignore preallocations with pa_deleted
3417 * otherwise we could leave used blocks available for
3418 * allocation in buddy when concurrent ext4_mb_put_pa()
3419 * is dropping preallocation
3420 */
3421 list_for_each(cur, &grp->bb_prealloc_list) {
3422 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3423 spin_lock(&pa->pa_lock);
3424 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3425 &groupnr, &start);
3426 len = pa->pa_len;
3427 spin_unlock(&pa->pa_lock);
3428 if (unlikely(len == 0))
3429 continue;
3430 BUG_ON(groupnr != group);
3431 mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
3432 bitmap, start, len);
3433 preallocated += len;
3434 count++;
3435 }
3436 mb_debug("prellocated %u for group %lu\n", preallocated, group);
3437}
3438
3439static void ext4_mb_pa_callback(struct rcu_head *head)
3440{
3441 struct ext4_prealloc_space *pa;
3442 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3443 kmem_cache_free(ext4_pspace_cachep, pa);
3444}
3445
3446/*
3447 * drops a reference to preallocated space descriptor
3448 * if this was the last reference and the space is consumed
3449 */
3450static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3451 struct super_block *sb, struct ext4_prealloc_space *pa)
3452{
3453 unsigned long grp;
3454
3455 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3456 return;
3457
3458 /* in this short window concurrent discard can set pa_deleted */
3459 spin_lock(&pa->pa_lock);
3460 if (pa->pa_deleted == 1) {
3461 spin_unlock(&pa->pa_lock);
3462 return;
3463 }
3464
3465 pa->pa_deleted = 1;
3466 spin_unlock(&pa->pa_lock);
3467
3468 /* -1 is to protect from crossing allocation group */
3469 ext4_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL);
3470
3471 /*
3472 * possible race:
3473 *
3474 * P1 (buddy init) P2 (regular allocation)
3475 * find block B in PA
3476 * copy on-disk bitmap to buddy
3477 * mark B in on-disk bitmap
3478 * drop PA from group
3479 * mark all PAs in buddy
3480 *
3481 * thus, P1 initializes buddy with B available. to prevent this
3482 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3483 * against that pair
3484 */
3485 ext4_lock_group(sb, grp);
3486 list_del(&pa->pa_group_list);
3487 ext4_unlock_group(sb, grp);
3488
3489 spin_lock(pa->pa_obj_lock);
3490 list_del_rcu(&pa->pa_inode_list);
3491 spin_unlock(pa->pa_obj_lock);
3492
3493 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3494}
3495
3496/*
3497 * creates new preallocated space for given inode
3498 */
4ddfef7b
ES		 3499static noinline_for_stack int
3500ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
c9de560d
AT		 3501{
3502 struct super_block *sb = ac->ac_sb;
3503 struct ext4_prealloc_space *pa;
3504 struct ext4_group_info *grp;
3505 struct ext4_inode_info *ei;
3506
3507 /* preallocate only when found space is larger then requested */
3508 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3509 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3510 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3511
3512 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3513 if (pa == NULL)
3514 return -ENOMEM;
3515
3516 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3517 int winl;
3518 int wins;
3519 int win;
3520 int offs;
3521
3522 /* we can't allocate as much as normalizer wants.
3523 * so, found space must get proper lstart
3524 * to cover original request */
3525 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3526 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3527
3528 /* we're limited by original request in that
3529 * logical block must be covered any way
3530 * winl is window we can move our chunk within */
3531 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3532
3533 /* also, we should cover whole original request */
3534 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3535
3536 /* the smallest one defines real window */
3537 win = min(winl, wins);
3538
3539 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3540 if (offs && offs < win)
3541 win = offs;
3542
3543 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3544 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3545 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3546 }
3547
3548 /* preallocation can change ac_b_ex, thus we store actually
3549 * allocated blocks for history */
3550 ac->ac_f_ex = ac->ac_b_ex;
3551
3552 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3553 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3554 pa->pa_len = ac->ac_b_ex.fe_len;
3555 pa->pa_free = pa->pa_len;
3556 atomic_set(&pa->pa_count, 1);
3557 spin_lock_init(&pa->pa_lock);
3558 pa->pa_deleted = 0;
3559 pa->pa_linear = 0;
3560
3561 mb_debug("new inode pa %p: %llu/%u for %u\n", pa,
3562 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3563
3564 ext4_mb_use_inode_pa(ac, pa);
3565 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3566
3567 ei = EXT4_I(ac->ac_inode);
3568 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3569
3570 pa->pa_obj_lock = &ei->i_prealloc_lock;
3571 pa->pa_inode = ac->ac_inode;
3572
3573 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3574 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3575 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3576
3577 spin_lock(pa->pa_obj_lock);
3578 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3579 spin_unlock(pa->pa_obj_lock);
3580
3581 return 0;
3582}
3583
3584/*
3585 * creates new preallocated space for locality group inodes belongs to
3586 */
4ddfef7b
ES		 3587static noinline_for_stack int
3588ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
c9de560d
AT		 3589{
3590 struct super_block *sb = ac->ac_sb;
3591 struct ext4_locality_group *lg;
3592 struct ext4_prealloc_space *pa;
3593 struct ext4_group_info *grp;
3594
3595 /* preallocate only when found space is larger then requested */
3596 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3597 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3598 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3599
3600 BUG_ON(ext4_pspace_cachep == NULL);
3601 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3602 if (pa == NULL)
3603 return -ENOMEM;
3604
3605 /* preallocation can change ac_b_ex, thus we store actually
3606 * allocated blocks for history */
3607 ac->ac_f_ex = ac->ac_b_ex;
3608
3609 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3610 pa->pa_lstart = pa->pa_pstart;
3611 pa->pa_len = ac->ac_b_ex.fe_len;
3612 pa->pa_free = pa->pa_len;
3613 atomic_set(&pa->pa_count, 1);
3614 spin_lock_init(&pa->pa_lock);
6be2ded1 3615 INIT_LIST_HEAD(&pa->pa_inode_list);
c9de560d
AT		 3616 pa->pa_deleted = 0;
3617 pa->pa_linear = 1;
3618
3619 mb_debug("new group pa %p: %llu/%u for %u\n", pa,
3620 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3621
3622 ext4_mb_use_group_pa(ac, pa);
3623 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3624
3625 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3626 lg = ac->ac_lg;
3627 BUG_ON(lg == NULL);
3628
3629 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3630 pa->pa_inode = NULL;
3631
3632 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3633 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3634 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3635
6be2ded1
AK		 3636 /*
3637 * We will later add the new pa to the right bucket
3638 * after updating the pa_free in ext4_mb_release_context
3639 */
c9de560d
AT		 3640 return 0;
3641}
3642
3643static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3644{
3645 int err;
3646
3647 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3648 err = ext4_mb_new_group_pa(ac);
3649 else
3650 err = ext4_mb_new_inode_pa(ac);
3651 return err;
3652}
3653
3654/*
3655 * finds all unused blocks in on-disk bitmap, frees them in
3656 * in-core bitmap and buddy.
3657 * @pa must be unlinked from inode and group lists, so that
3658 * nobody else can find/use it.
3659 * the caller MUST hold group/inode locks.
3660 * TODO: optimize the case when there are no in-core structures yet
3661 */
4ddfef7b
ES		 3662static noinline_for_stack int
3663ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
c83617db
AK		 3664 struct ext4_prealloc_space *pa,
3665 struct ext4_allocation_context *ac)
c9de560d 3666{
c9de560d
AT		 3667 struct super_block *sb = e4b->bd_sb;
3668 struct ext4_sb_info *sbi = EXT4_SB(sb);
3669 unsigned long end;
3670 unsigned long next;
3671 ext4_group_t group;
3672 ext4_grpblk_t bit;
3673 sector_t start;
3674 int err = 0;
3675 int free = 0;
3676
3677 BUG_ON(pa->pa_deleted == 0);
3678 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3679 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3680 end = bit + pa->pa_len;
3681
256bdb49
ES		 3682 if (ac) {
3683 ac->ac_sb = sb;
3684 ac->ac_inode = pa->pa_inode;
3685 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
3686 }
c9de560d
AT		 3687
3688 while (bit < end) {
ffad0a44 3689 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
c9de560d
AT		 3690 if (bit >= end)
3691 break;
ffad0a44 3692 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
c9de560d
AT		 3693 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3694 le32_to_cpu(sbi->s_es->s_first_data_block);
3695 mb_debug(" free preallocated %u/%u in group %u\n",
3696 (unsigned) start, (unsigned) next - bit,
3697 (unsigned) group);
3698 free += next - bit;
3699
256bdb49
ES		 3700 if (ac) {
3701 ac->ac_b_ex.fe_group = group;
3702 ac->ac_b_ex.fe_start = bit;
3703 ac->ac_b_ex.fe_len = next - bit;
3704 ac->ac_b_ex.fe_logical = 0;
3705 ext4_mb_store_history(ac);
3706 }
c9de560d
AT		 3707
3708 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3709 bit = next + 1;
3710 }
3711 if (free != pa->pa_free) {
26346ff6 3712 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
c9de560d
AT		 3713 pa, (unsigned long) pa->pa_lstart,
3714 (unsigned long) pa->pa_pstart,
3715 (unsigned long) pa->pa_len);
46e665e9 3716 ext4_error(sb, __func__, "free %u, pa_free %u\n",
26346ff6 3717 free, pa->pa_free);
e56eb659
AK		 3718 /*
3719 * pa is already deleted so we use the value obtained
3720 * from the bitmap and continue.
3721 */
c9de560d 3722 }
c9de560d
AT		 3723 atomic_add(free, &sbi->s_mb_discarded);
3724
3725 return err;
3726}
3727
4ddfef7b
ES		 3728static noinline_for_stack int
3729ext4_mb_release_group_pa(struct ext4_buddy *e4b,
c83617db
AK		 3730 struct ext4_prealloc_space *pa,
3731 struct ext4_allocation_context *ac)
c9de560d 3732{
c9de560d
AT		 3733 struct super_block *sb = e4b->bd_sb;
3734 ext4_group_t group;
3735 ext4_grpblk_t bit;
3736
256bdb49
ES		 3737 if (ac)
3738 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
c9de560d
AT		 3739
3740 BUG_ON(pa->pa_deleted == 0);
3741 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3742 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3743 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3744 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3745
256bdb49
ES		 3746 if (ac) {
3747 ac->ac_sb = sb;
3748 ac->ac_inode = NULL;
3749 ac->ac_b_ex.fe_group = group;
3750 ac->ac_b_ex.fe_start = bit;
3751 ac->ac_b_ex.fe_len = pa->pa_len;
3752 ac->ac_b_ex.fe_logical = 0;
3753 ext4_mb_store_history(ac);
256bdb49 3754 }
c9de560d
AT		 3755
3756 return 0;
3757}
3758
3759/*
3760 * releases all preallocations in given group
3761 *
3762 * first, we need to decide discard policy:
3763 * - when do we discard
3764 * 1) ENOSPC
3765 * - how many do we discard
3766 * 1) how many requested
3767 */
4ddfef7b
ES		 3768static noinline_for_stack int
3769ext4_mb_discard_group_preallocations(struct super_block *sb,
c9de560d
AT		 3770 ext4_group_t group, int needed)
3771{
3772 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3773 struct buffer_head *bitmap_bh = NULL;
3774 struct ext4_prealloc_space *pa, *tmp;
c83617db 3775 struct ext4_allocation_context *ac;
c9de560d
AT		 3776 struct list_head list;
3777 struct ext4_buddy e4b;
3778 int err;
3779 int busy = 0;
3780 int free = 0;
3781
3782 mb_debug("discard preallocation for group %lu\n", group);
3783
3784 if (list_empty(&grp->bb_prealloc_list))
3785 return 0;
3786
574ca174 3787 bitmap_bh = ext4_read_block_bitmap(sb, group);
c9de560d 3788 if (bitmap_bh == NULL) {
ce89f46c
AK		 3789 ext4_error(sb, __func__, "Error in reading block "
3790 "bitmap for %lu\n", group);
3791 return 0;
c9de560d
AT		 3792 }
3793
3794 err = ext4_mb_load_buddy(sb, group, &e4b);
ce89f46c
AK		 3795 if (err) {
3796 ext4_error(sb, __func__, "Error in loading buddy "
3797 "information for %lu\n", group);
3798 put_bh(bitmap_bh);
3799 return 0;
3800 }
c9de560d
AT		 3801
3802 if (needed == 0)
3803 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3804
c9de560d 3805 INIT_LIST_HEAD(&list);
c83617db 3806 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
c9de560d
AT		 3807repeat:
3808 ext4_lock_group(sb, group);
3809 list_for_each_entry_safe(pa, tmp,
3810 &grp->bb_prealloc_list, pa_group_list) {
3811 spin_lock(&pa->pa_lock);
3812 if (atomic_read(&pa->pa_count)) {
3813 spin_unlock(&pa->pa_lock);
3814 busy = 1;
3815 continue;
3816 }
3817 if (pa->pa_deleted) {
3818 spin_unlock(&pa->pa_lock);
3819 continue;
3820 }
3821
3822 /* seems this one can be freed ... */
3823 pa->pa_deleted = 1;
3824
3825 /* we can trust pa_free ... */
3826 free += pa->pa_free;
3827
3828 spin_unlock(&pa->pa_lock);
3829
3830 list_del(&pa->pa_group_list);
3831 list_add(&pa->u.pa_tmp_list, &list);
3832 }
3833
3834 /* if we still need more blocks and some PAs were used, try again */
3835 if (free < needed && busy) {
3836 busy = 0;
3837 ext4_unlock_group(sb, group);
3838 /*
3839 * Yield the CPU here so that we don't get soft lockup
3840 * in non preempt case.
3841 */
3842 yield();
3843 goto repeat;
3844 }
3845
3846 /* found anything to free? */
3847 if (list_empty(&list)) {
3848 BUG_ON(free != 0);
3849 goto out;
3850 }
3851
3852 /* now free all selected PAs */
3853 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3854
3855 /* remove from object (inode or locality group) */
3856 spin_lock(pa->pa_obj_lock);
3857 list_del_rcu(&pa->pa_inode_list);
3858 spin_unlock(pa->pa_obj_lock);
3859
3860 if (pa->pa_linear)
c83617db 3861 ext4_mb_release_group_pa(&e4b, pa, ac);
c9de560d 3862 else
c83617db 3863 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
c9de560d
AT		 3864
3865 list_del(&pa->u.pa_tmp_list);
3866 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3867 }
3868
3869out:
3870 ext4_unlock_group(sb, group);
c83617db
AK		 3871 if (ac)
3872 kmem_cache_free(ext4_ac_cachep, ac);
c9de560d
AT		 3873 ext4_mb_release_desc(&e4b);
3874 put_bh(bitmap_bh);
3875 return free;
3876}
3877
3878/*
3879 * releases all non-used preallocated blocks for given inode
3880 *
3881 * It's important to discard preallocations under i_data_sem
3882 * We don't want another block to be served from the prealloc
3883 * space when we are discarding the inode prealloc space.
3884 *
3885 * FIXME!! Make sure it is valid at all the call sites
3886 */
3887void ext4_mb_discard_inode_preallocations(struct inode *inode)
3888{
3889 struct ext4_inode_info *ei = EXT4_I(inode);
3890 struct super_block *sb = inode->i_sb;
3891 struct buffer_head *bitmap_bh = NULL;
3892 struct ext4_prealloc_space *pa, *tmp;
c83617db 3893 struct ext4_allocation_context *ac;
c9de560d
AT		 3894 ext4_group_t group = 0;
3895 struct list_head list;
3896 struct ext4_buddy e4b;
3897 int err;
3898
3899 if (!test_opt(sb, MBALLOC) || !S_ISREG(inode->i_mode)) {
3900 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3901 return;
3902 }
3903
3904 mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
3905
3906 INIT_LIST_HEAD(&list);
3907
c83617db 3908 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
c9de560d
AT		 3909repeat:
3910 /* first, collect all pa's in the inode */
3911 spin_lock(&ei->i_prealloc_lock);
3912 while (!list_empty(&ei->i_prealloc_list)) {
3913 pa = list_entry(ei->i_prealloc_list.next,
3914 struct ext4_prealloc_space, pa_inode_list);
3915 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3916 spin_lock(&pa->pa_lock);
3917 if (atomic_read(&pa->pa_count)) {
3918 /* this shouldn't happen often - nobody should
3919 * use preallocation while we're discarding it */
3920 spin_unlock(&pa->pa_lock);
3921 spin_unlock(&ei->i_prealloc_lock);
3922 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3923 WARN_ON(1);
3924 schedule_timeout_uninterruptible(HZ);
3925 goto repeat;
3926
3927 }
3928 if (pa->pa_deleted == 0) {
3929 pa->pa_deleted = 1;
3930 spin_unlock(&pa->pa_lock);
3931 list_del_rcu(&pa->pa_inode_list);
3932 list_add(&pa->u.pa_tmp_list, &list);
3933 continue;
3934 }
3935
3936 /* someone is deleting pa right now */
3937 spin_unlock(&pa->pa_lock);
3938 spin_unlock(&ei->i_prealloc_lock);
3939
3940 /* we have to wait here because pa_deleted
3941 * doesn't mean pa is already unlinked from
3942 * the list. as we might be called from
3943 * ->clear_inode() the inode will get freed
3944 * and concurrent thread which is unlinking
3945 * pa from inode's list may access already
3946 * freed memory, bad-bad-bad */
3947
3948 /* XXX: if this happens too often, we can
3949 * add a flag to force wait only in case
3950 * of ->clear_inode(), but not in case of
3951 * regular truncate */
3952 schedule_timeout_uninterruptible(HZ);
3953 goto repeat;
3954 }
3955 spin_unlock(&ei->i_prealloc_lock);
3956
3957 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3958 BUG_ON(pa->pa_linear != 0);
3959 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3960
3961 err = ext4_mb_load_buddy(sb, group, &e4b);
ce89f46c
AK		 3962 if (err) {
3963 ext4_error(sb, __func__, "Error in loading buddy "
3964 "information for %lu\n", group);
3965 continue;
3966 }
c9de560d 3967
574ca174 3968 bitmap_bh = ext4_read_block_bitmap(sb, group);
c9de560d 3969 if (bitmap_bh == NULL) {
ce89f46c
AK		 3970 ext4_error(sb, __func__, "Error in reading block "
3971 "bitmap for %lu\n", group);
c9de560d 3972 ext4_mb_release_desc(&e4b);
ce89f46c 3973 continue;
c9de560d
AT		 3974 }
3975
3976 ext4_lock_group(sb, group);
3977 list_del(&pa->pa_group_list);
c83617db 3978 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
c9de560d
AT		 3979 ext4_unlock_group(sb, group);
3980
3981 ext4_mb_release_desc(&e4b);
3982 put_bh(bitmap_bh);
3983
3984 list_del(&pa->u.pa_tmp_list);
3985 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3986 }
c83617db
AK		 3987 if (ac)
3988 kmem_cache_free(ext4_ac_cachep, ac);
c9de560d
AT		 3989}
3990
3991/*
3992 * finds all preallocated spaces and return blocks being freed to them
3993 * if preallocated space becomes full (no block is used from the space)
3994 * then the function frees space in buddy
3995 * XXX: at the moment, truncate (which is the only way to free blocks)
3996 * discards all preallocations
3997 */
3998static void ext4_mb_return_to_preallocation(struct inode *inode,
3999 struct ext4_buddy *e4b,
4000 sector_t block, int count)
4001{
4002 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
4003}
4004#ifdef MB_DEBUG
4005static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4006{
4007 struct super_block *sb = ac->ac_sb;
4008 ext4_group_t i;
4009
4010 printk(KERN_ERR "EXT4-fs: Can't allocate:"
4011 " Allocation context details:\n");
4012 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
4013 ac->ac_status, ac->ac_flags);
4014 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4015 "best %lu/%lu/%lu@%lu cr %d\n",
4016 (unsigned long)ac->ac_o_ex.fe_group,
4017 (unsigned long)ac->ac_o_ex.fe_start,
4018 (unsigned long)ac->ac_o_ex.fe_len,
4019 (unsigned long)ac->ac_o_ex.fe_logical,
4020 (unsigned long)ac->ac_g_ex.fe_group,
4021 (unsigned long)ac->ac_g_ex.fe_start,
4022 (unsigned long)ac->ac_g_ex.fe_len,
4023 (unsigned long)ac->ac_g_ex.fe_logical,
4024 (unsigned long)ac->ac_b_ex.fe_group,
4025 (unsigned long)ac->ac_b_ex.fe_start,
4026 (unsigned long)ac->ac_b_ex.fe_len,
4027 (unsigned long)ac->ac_b_ex.fe_logical,
4028 (int)ac->ac_criteria);
4029 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
4030 ac->ac_found);
4031 printk(KERN_ERR "EXT4-fs: groups: \n");
4032 for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
4033 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4034 struct ext4_prealloc_space *pa;
4035 ext4_grpblk_t start;
4036 struct list_head *cur;
4037 ext4_lock_group(sb, i);
4038 list_for_each(cur, &grp->bb_prealloc_list) {
4039 pa = list_entry(cur, struct ext4_prealloc_space,
4040 pa_group_list);
4041 spin_lock(&pa->pa_lock);
4042 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4043 NULL, &start);
4044 spin_unlock(&pa->pa_lock);
4045 printk(KERN_ERR "PA:%lu:%d:%u \n", i,
4046 start, pa->pa_len);
4047 }
60bd63d1 4048 ext4_unlock_group(sb, i);
c9de560d
AT		 4049
4050 if (grp->bb_free == 0)
4051 continue;
4052 printk(KERN_ERR "%lu: %d/%d \n",
4053 i, grp->bb_free, grp->bb_fragments);
4054 }
4055 printk(KERN_ERR "\n");
4056}
4057#else
4058static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4059{
4060 return;
4061}
4062#endif
4063
4064/*
4065 * We use locality group preallocation for small size file. The size of the
4066 * file is determined by the current size or the resulting size after
4067 * allocation which ever is larger
4068 *
4069 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
4070 */
4071static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4072{
4073 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4074 int bsbits = ac->ac_sb->s_blocksize_bits;
4075 loff_t size, isize;
4076
4077 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4078 return;
4079
4080 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4081 isize = i_size_read(ac->ac_inode) >> bsbits;
4082 size = max(size, isize);
4083
4084 /* don't use group allocation for large files */
4085 if (size >= sbi->s_mb_stream_request)
4086 return;
4087
4088 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4089 return;
4090
4091 BUG_ON(ac->ac_lg != NULL);
4092 /*
4093 * locality group prealloc space are per cpu. The reason for having
4094 * per cpu locality group is to reduce the contention between block
4095 * request from multiple CPUs.
4096 */
4097 ac->ac_lg = &sbi->s_locality_groups[get_cpu()];
4098 put_cpu();
4099
4100 /* we're going to use group allocation */
4101 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4102
4103 /* serialize all allocations in the group */
4104 mutex_lock(&ac->ac_lg->lg_mutex);
4105}
4106
4ddfef7b
ES		 4107static noinline_for_stack int
4108ext4_mb_initialize_context(struct ext4_allocation_context *ac,
c9de560d
AT		 4109 struct ext4_allocation_request *ar)
4110{
4111 struct super_block *sb = ar->inode->i_sb;
4112 struct ext4_sb_info *sbi = EXT4_SB(sb);
4113 struct ext4_super_block *es = sbi->s_es;
4114 ext4_group_t group;
4115 unsigned long len;
4116 unsigned long goal;
4117 ext4_grpblk_t block;
4118
4119 /* we can't allocate > group size */
4120 len = ar->len;
4121
4122 /* just a dirty hack to filter too big requests */
4123 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4124 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4125
4126 /* start searching from the goal */
4127 goal = ar->goal;
4128 if (goal < le32_to_cpu(es->s_first_data_block) ||
4129 goal >= ext4_blocks_count(es))
4130 goal = le32_to_cpu(es->s_first_data_block);
4131 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4132
4133 /* set up allocation goals */
4134 ac->ac_b_ex.fe_logical = ar->logical;
4135 ac->ac_b_ex.fe_group = 0;
4136 ac->ac_b_ex.fe_start = 0;
4137 ac->ac_b_ex.fe_len = 0;
4138 ac->ac_status = AC_STATUS_CONTINUE;
4139 ac->ac_groups_scanned = 0;
4140 ac->ac_ex_scanned = 0;
4141 ac->ac_found = 0;
4142 ac->ac_sb = sb;
4143 ac->ac_inode = ar->inode;
4144 ac->ac_o_ex.fe_logical = ar->logical;
4145 ac->ac_o_ex.fe_group = group;
4146 ac->ac_o_ex.fe_start = block;
4147 ac->ac_o_ex.fe_len = len;
4148 ac->ac_g_ex.fe_logical = ar->logical;
4149 ac->ac_g_ex.fe_group = group;
4150 ac->ac_g_ex.fe_start = block;
4151 ac->ac_g_ex.fe_len = len;
4152 ac->ac_f_ex.fe_len = 0;
4153 ac->ac_flags = ar->flags;
4154 ac->ac_2order = 0;
4155 ac->ac_criteria = 0;
4156 ac->ac_pa = NULL;
4157 ac->ac_bitmap_page = NULL;
4158 ac->ac_buddy_page = NULL;
4159 ac->ac_lg = NULL;
4160
4161 /* we have to define context: we'll we work with a file or
4162 * locality group. this is a policy, actually */
4163 ext4_mb_group_or_file(ac);
4164
4165 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4166 "left: %u/%u, right %u/%u to %swritable\n",
4167 (unsigned) ar->len, (unsigned) ar->logical,
4168 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4169 (unsigned) ar->lleft, (unsigned) ar->pleft,
4170 (unsigned) ar->lright, (unsigned) ar->pright,
4171 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4172 return 0;
4173
4174}
4175
6be2ded1
AK		 4176static noinline_for_stack void
4177ext4_mb_discard_lg_preallocations(struct super_block *sb,
4178 struct ext4_locality_group *lg,
4179 int order, int total_entries)
4180{
4181 ext4_group_t group = 0;
4182 struct ext4_buddy e4b;
4183 struct list_head discard_list;
4184 struct ext4_prealloc_space *pa, *tmp;
4185 struct ext4_allocation_context *ac;
4186
4187 mb_debug("discard locality group preallocation\n");
4188
4189 INIT_LIST_HEAD(&discard_list);
4190 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4191
4192 spin_lock(&lg->lg_prealloc_lock);
4193 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4194 pa_inode_list) {
4195 spin_lock(&pa->pa_lock);
4196 if (atomic_read(&pa->pa_count)) {
4197 /*
4198 * This is the pa that we just used
4199 * for block allocation. So don't
4200 * free that
4201 */
4202 spin_unlock(&pa->pa_lock);
4203 continue;
4204 }
4205 if (pa->pa_deleted) {
4206 spin_unlock(&pa->pa_lock);
4207 continue;
4208 }
4209 /* only lg prealloc space */
4210 BUG_ON(!pa->pa_linear);
4211
4212 /* seems this one can be freed ... */
4213 pa->pa_deleted = 1;
4214 spin_unlock(&pa->pa_lock);
4215
4216 list_del_rcu(&pa->pa_inode_list);
4217 list_add(&pa->u.pa_tmp_list, &discard_list);
4218
4219 total_entries--;
4220 if (total_entries <= 5) {
4221 /*
4222 * we want to keep only 5 entries
4223 * allowing it to grow to 8. This
4224 * mak sure we don't call discard
4225 * soon for this list.
4226 */
4227 break;
4228 }
4229 }
4230 spin_unlock(&lg->lg_prealloc_lock);
4231
4232 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4233
4234 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4235 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4236 ext4_error(sb, __func__, "Error in loading buddy "
4237 "information for %lu\n", group);
4238 continue;
4239 }
4240 ext4_lock_group(sb, group);
4241 list_del(&pa->pa_group_list);
4242 ext4_mb_release_group_pa(&e4b, pa, ac);
4243 ext4_unlock_group(sb, group);
4244
4245 ext4_mb_release_desc(&e4b);
4246 list_del(&pa->u.pa_tmp_list);
4247 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4248 }
4249 if (ac)
4250 kmem_cache_free(ext4_ac_cachep, ac);
4251}
4252
4253/*
4254 * We have incremented pa_count. So it cannot be freed at this
4255 * point. Also we hold lg_mutex. So no parallel allocation is
4256 * possible from this lg. That means pa_free cannot be updated.
4257 *
4258 * A parallel ext4_mb_discard_group_preallocations is possible.
4259 * which can cause the lg_prealloc_list to be updated.
4260 */
4261
4262static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4263{
4264 int order, added = 0, lg_prealloc_count = 1;
4265 struct super_block *sb = ac->ac_sb;
4266 struct ext4_locality_group *lg = ac->ac_lg;
4267 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4268
4269 order = fls(pa->pa_free) - 1;
4270 if (order > PREALLOC_TB_SIZE - 1)
4271 /* The max size of hash table is PREALLOC_TB_SIZE */
4272 order = PREALLOC_TB_SIZE - 1;
4273 /* Add the prealloc space to lg */
4274 rcu_read_lock();
4275 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4276 pa_inode_list) {
4277 spin_lock(&tmp_pa->pa_lock);
4278 if (tmp_pa->pa_deleted) {
4279 spin_unlock(&pa->pa_lock);
4280 continue;
4281 }
4282 if (!added && pa->pa_free < tmp_pa->pa_free) {
4283 /* Add to the tail of the previous entry */
4284 list_add_tail_rcu(&pa->pa_inode_list,
4285 &tmp_pa->pa_inode_list);
4286 added = 1;
4287 /*
4288 * we want to count the total
4289 * number of entries in the list
4290 */
4291 }
4292 spin_unlock(&tmp_pa->pa_lock);
4293 lg_prealloc_count++;
4294 }
4295 if (!added)
4296 list_add_tail_rcu(&pa->pa_inode_list,
4297 &lg->lg_prealloc_list[order]);
4298 rcu_read_unlock();
4299
4300 /* Now trim the list to be not more than 8 elements */
4301 if (lg_prealloc_count > 8) {
4302 ext4_mb_discard_lg_preallocations(sb, lg,
4303 order, lg_prealloc_count);
4304 return;
4305 }
4306 return ;
4307}
4308
c9de560d
AT		 4309/*
4310 * release all resource we used in allocation
4311 */
4312static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4313{
6be2ded1
AK		 4314 struct ext4_prealloc_space *pa = ac->ac_pa;
4315 if (pa) {
4316 if (pa->pa_linear) {
c9de560d 4317 /* see comment in ext4_mb_use_group_pa() */
6be2ded1
AK		 4318 spin_lock(&pa->pa_lock);
4319 pa->pa_pstart += ac->ac_b_ex.fe_len;
4320 pa->pa_lstart += ac->ac_b_ex.fe_len;
4321 pa->pa_free -= ac->ac_b_ex.fe_len;
4322 pa->pa_len -= ac->ac_b_ex.fe_len;
4323 spin_unlock(&pa->pa_lock);
4324 /*
4325 * We want to add the pa to the right bucket.
4326 * Remove it from the list and while adding
4327 * make sure the list to which we are adding
4328 * doesn't grow big.
4329 */
4330 if (likely(pa->pa_free)) {
4331 spin_lock(pa->pa_obj_lock);
4332 list_del_rcu(&pa->pa_inode_list);
4333 spin_unlock(pa->pa_obj_lock);
4334 ext4_mb_add_n_trim(ac);
4335 }
c9de560d 4336 }
6be2ded1 4337 ext4_mb_put_pa(ac, ac->ac_sb, pa);
c9de560d
AT		 4338 }
4339 if (ac->ac_bitmap_page)
4340 page_cache_release(ac->ac_bitmap_page);
4341 if (ac->ac_buddy_page)
4342 page_cache_release(ac->ac_buddy_page);
4343 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4344 mutex_unlock(&ac->ac_lg->lg_mutex);
4345 ext4_mb_collect_stats(ac);
4346 return 0;
4347}
4348
4349static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4350{
4351 ext4_group_t i;
4352 int ret;
4353 int freed = 0;
4354
4355 for (i = 0; i < EXT4_SB(sb)->s_groups_count && needed > 0; i++) {
4356 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4357 freed += ret;
4358 needed -= ret;
4359 }
4360
4361 return freed;
4362}
4363
4364/*
4365 * Main entry point into mballoc to allocate blocks
4366 * it tries to use preallocation first, then falls back
4367 * to usual allocation
4368 */
4369ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4370 struct ext4_allocation_request *ar, int *errp)
4371{
256bdb49 4372 struct ext4_allocation_context *ac = NULL;
c9de560d
AT		 4373 struct ext4_sb_info *sbi;
4374 struct super_block *sb;
4375 ext4_fsblk_t block = 0;
4376 int freed;
4377 int inquota;
4378
4379 sb = ar->inode->i_sb;
4380 sbi = EXT4_SB(sb);
4381
4382 if (!test_opt(sb, MBALLOC)) {
654b4908 4383 block = ext4_old_new_blocks(handle, ar->inode, ar->goal,
c9de560d
AT		 4384 &(ar->len), errp);
4385 return block;
4386 }
d2a17637
MC		 4387 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) {
4388 /*
4389 * With delalloc we already reserved the blocks
4390 */
4391 ar->len = ext4_has_free_blocks(sbi, ar->len);
4392 }
07031431
MC		 4393
4394 if (ar->len == 0) {
4395 *errp = -ENOSPC;
4396 return 0;
4397 }
c9de560d
AT		 4398
4399 while (ar->len && DQUOT_ALLOC_BLOCK(ar->inode, ar->len)) {
4400 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4401 ar->len--;
4402 }
4403 if (ar->len == 0) {
4404 *errp = -EDQUOT;
4405 return 0;
4406 }
4407 inquota = ar->len;
4408
d2a17637
MC		 4409 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4410 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4411
256bdb49
ES		 4412 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4413 if (!ac) {
363d4251 4414 ar->len = 0;
256bdb49 4415 *errp = -ENOMEM;
363d4251 4416 goto out1;
256bdb49
ES		 4417 }
4418
c9de560d
AT		 4419 ext4_mb_poll_new_transaction(sb, handle);
4420
256bdb49 4421 *errp = ext4_mb_initialize_context(ac, ar);
c9de560d
AT		 4422 if (*errp) {
4423 ar->len = 0;
363d4251 4424 goto out2;
c9de560d
AT		 4425 }
4426
256bdb49
ES		 4427 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4428 if (!ext4_mb_use_preallocated(ac)) {
256bdb49
ES		 4429 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4430 ext4_mb_normalize_request(ac, ar);
c9de560d
AT		 4431repeat:
4432 /* allocate space in core */
256bdb49 4433 ext4_mb_regular_allocator(ac);
c9de560d
AT		 4434
4435 /* as we've just preallocated more space than
4436 * user requested orinally, we store allocated
4437 * space in a special descriptor */
256bdb49
ES		 4438 if (ac->ac_status == AC_STATUS_FOUND &&
4439 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4440 ext4_mb_new_preallocation(ac);
c9de560d
AT		 4441 }
4442
256bdb49 4443 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
519deca0
AK		 4444 *errp = ext4_mb_mark_diskspace_used(ac, handle);
4445 if (*errp == -EAGAIN) {
4446 ac->ac_b_ex.fe_group = 0;
4447 ac->ac_b_ex.fe_start = 0;
4448 ac->ac_b_ex.fe_len = 0;
4449 ac->ac_status = AC_STATUS_CONTINUE;
4450 goto repeat;
4451 } else if (*errp) {
4452 ac->ac_b_ex.fe_len = 0;
4453 ar->len = 0;
4454 ext4_mb_show_ac(ac);
4455 } else {
4456 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4457 ar->len = ac->ac_b_ex.fe_len;
4458 }
c9de560d 4459 } else {
256bdb49 4460 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
c9de560d
AT		 4461 if (freed)
4462 goto repeat;
4463 *errp = -ENOSPC;
256bdb49 4464 ac->ac_b_ex.fe_len = 0;
c9de560d 4465 ar->len = 0;
256bdb49 4466 ext4_mb_show_ac(ac);
c9de560d
AT		 4467 }
4468
256bdb49 4469 ext4_mb_release_context(ac);
c9de560d 4470
363d4251
SF		 4471out2:
4472 kmem_cache_free(ext4_ac_cachep, ac);
4473out1:
c9de560d
AT		 4474 if (ar->len < inquota)
4475 DQUOT_FREE_BLOCK(ar->inode, inquota - ar->len);
4476
4477 return block;
4478}
4479static void ext4_mb_poll_new_transaction(struct super_block *sb,
4480 handle_t *handle)
4481{
4482 struct ext4_sb_info *sbi = EXT4_SB(sb);
4483
4484 if (sbi->s_last_transaction == handle->h_transaction->t_tid)
4485 return;
4486
4487 /* new transaction! time to close last one and free blocks for
4488 * committed transaction. we know that only transaction can be
4489 * active, so previos transaction can be being logged and we
4490 * know that transaction before previous is known to be already
4491 * logged. this means that now we may free blocks freed in all
4492 * transactions before previous one. hope I'm clear enough ... */
4493
4494 spin_lock(&sbi->s_md_lock);
4495 if (sbi->s_last_transaction != handle->h_transaction->t_tid) {
4496 mb_debug("new transaction %lu, old %lu\n",
4497 (unsigned long) handle->h_transaction->t_tid,
4498 (unsigned long) sbi->s_last_transaction);
4499 list_splice_init(&sbi->s_closed_transaction,
4500 &sbi->s_committed_transaction);
4501 list_splice_init(&sbi->s_active_transaction,
4502 &sbi->s_closed_transaction);
4503 sbi->s_last_transaction = handle->h_transaction->t_tid;
4504 }
4505 spin_unlock(&sbi->s_md_lock);
4506
4507 ext4_mb_free_committed_blocks(sb);
4508}
4509
4ddfef7b
ES		 4510static noinline_for_stack int
4511ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
c9de560d
AT		 4512 ext4_group_t group, ext4_grpblk_t block, int count)
4513{
4514 struct ext4_group_info *db = e4b->bd_info;
4515 struct super_block *sb = e4b->bd_sb;
4516 struct ext4_sb_info *sbi = EXT4_SB(sb);
4517 struct ext4_free_metadata *md;
4518 int i;
4519
4520 BUG_ON(e4b->bd_bitmap_page == NULL);
4521 BUG_ON(e4b->bd_buddy_page == NULL);
4522
4523 ext4_lock_group(sb, group);
4524 for (i = 0; i < count; i++) {
4525 md = db->bb_md_cur;
4526 if (md && db->bb_tid != handle->h_transaction->t_tid) {
4527 db->bb_md_cur = NULL;
4528 md = NULL;
4529 }
4530
4531 if (md == NULL) {
4532 ext4_unlock_group(sb, group);
4533 md = kmalloc(sizeof(*md), GFP_NOFS);
4534 if (md == NULL)
4535 return -ENOMEM;
4536 md->num = 0;
4537 md->group = group;
4538
4539 ext4_lock_group(sb, group);
4540 if (db->bb_md_cur == NULL) {
4541 spin_lock(&sbi->s_md_lock);
4542 list_add(&md->list, &sbi->s_active_transaction);
4543 spin_unlock(&sbi->s_md_lock);
4544 /* protect buddy cache from being freed,
4545 * otherwise we'll refresh it from
4546 * on-disk bitmap and lose not-yet-available
4547 * blocks */
4548 page_cache_get(e4b->bd_buddy_page);
4549 page_cache_get(e4b->bd_bitmap_page);
4550 db->bb_md_cur = md;
4551 db->bb_tid = handle->h_transaction->t_tid;
4552 mb_debug("new md 0x%p for group %lu\n",
4553 md, md->group);
4554 } else {
4555 kfree(md);
4556 md = db->bb_md_cur;
4557 }
4558 }
4559
4560 BUG_ON(md->num >= EXT4_BB_MAX_BLOCKS);
4561 md->blocks[md->num] = block + i;
4562 md->num++;
4563 if (md->num == EXT4_BB_MAX_BLOCKS) {
4564 /* no more space, put full container on a sb's list */
4565 db->bb_md_cur = NULL;
4566 }
4567 }
4568 ext4_unlock_group(sb, group);
4569 return 0;
4570}
4571
4572/*
4573 * Main entry point into mballoc to free blocks
4574 */
4575void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
4576 unsigned long block, unsigned long count,
4577 int metadata, unsigned long *freed)
4578{
26346ff6 4579 struct buffer_head *bitmap_bh = NULL;
c9de560d 4580 struct super_block *sb = inode->i_sb;
256bdb49 4581 struct ext4_allocation_context *ac = NULL;
c9de560d
AT		 4582 struct ext4_group_desc *gdp;
4583 struct ext4_super_block *es;
4584 unsigned long overflow;
4585 ext4_grpblk_t bit;
4586 struct buffer_head *gd_bh;
4587 ext4_group_t block_group;
4588 struct ext4_sb_info *sbi;
4589 struct ext4_buddy e4b;
4590 int err = 0;
4591 int ret;
4592
4593 *freed = 0;
4594
4595 ext4_mb_poll_new_transaction(sb, handle);
4596
4597 sbi = EXT4_SB(sb);
4598 es = EXT4_SB(sb)->s_es;
4599 if (block < le32_to_cpu(es->s_first_data_block) ||
4600 block + count < block ||
4601 block + count > ext4_blocks_count(es)) {
46e665e9 4602 ext4_error(sb, __func__,
c9de560d
AT		 4603 "Freeing blocks not in datazone - "
4604 "block = %lu, count = %lu", block, count);
4605 goto error_return;
4606 }
4607
4608 ext4_debug("freeing block %lu\n", block);
4609
256bdb49
ES		 4610 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4611 if (ac) {
4612 ac->ac_op = EXT4_MB_HISTORY_FREE;
4613 ac->ac_inode = inode;
4614 ac->ac_sb = sb;
4615 }
c9de560d
AT		 4616
4617do_more:
4618 overflow = 0;
4619 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4620
4621 /*
4622 * Check to see if we are freeing blocks across a group
4623 * boundary.
4624 */
4625 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4626 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4627 count -= overflow;
4628 }
574ca174 4629 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
ce89f46c
AK		 4630 if (!bitmap_bh) {
4631 err = -EIO;
c9de560d 4632 goto error_return;
ce89f46c 4633 }
c9de560d 4634 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
ce89f46c
AK		 4635 if (!gdp) {
4636 err = -EIO;
c9de560d 4637 goto error_return;
ce89f46c 4638 }
c9de560d
AT		 4639
4640 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4641 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4642 in_range(block, ext4_inode_table(sb, gdp),
4643 EXT4_SB(sb)->s_itb_per_group) ||
4644 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4645 EXT4_SB(sb)->s_itb_per_group)) {
4646
46e665e9 4647 ext4_error(sb, __func__,
c9de560d
AT		 4648 "Freeing blocks in system zone - "
4649 "Block = %lu, count = %lu", block, count);
519deca0
AK		 4650 /* err = 0. ext4_std_error should be a no op */
4651 goto error_return;
c9de560d
AT		 4652 }
4653
4654 BUFFER_TRACE(bitmap_bh, "getting write access");
4655 err = ext4_journal_get_write_access(handle, bitmap_bh);
4656 if (err)
4657 goto error_return;
4658
4659 /*
4660 * We are about to modify some metadata. Call the journal APIs
4661 * to unshare ->b_data if a currently-committing transaction is
4662 * using it
4663 */
4664 BUFFER_TRACE(gd_bh, "get_write_access");
4665 err = ext4_journal_get_write_access(handle, gd_bh);
4666 if (err)
4667 goto error_return;
4668
4669 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4670 if (err)
4671 goto error_return;
4672
4673#ifdef AGGRESSIVE_CHECK
4674 {
4675 int i;
4676 for (i = 0; i < count; i++)
4677 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4678 }
4679#endif
4680 mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data,
4681 bit, count);
4682
4683 /* We dirtied the bitmap block */
4684 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4685 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
4686
256bdb49
ES		 4687 if (ac) {
4688 ac->ac_b_ex.fe_group = block_group;
4689 ac->ac_b_ex.fe_start = bit;
4690 ac->ac_b_ex.fe_len = count;
4691 ext4_mb_store_history(ac);
4692 }
c9de560d
AT		 4693
4694 if (metadata) {
4695 /* blocks being freed are metadata. these blocks shouldn't
4696 * be used until this transaction is committed */
4697 ext4_mb_free_metadata(handle, &e4b, block_group, bit, count);
4698 } else {
4699 ext4_lock_group(sb, block_group);
7e5a8cdd 4700 mb_free_blocks(inode, &e4b, bit, count);
c9de560d
AT		 4701 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4702 ext4_unlock_group(sb, block_group);
c9de560d
AT		 4703 }
4704
4705 spin_lock(sb_bgl_lock(sbi, block_group));
e8546d06 4706 le16_add_cpu(&gdp->bg_free_blocks_count, count);
c9de560d
AT		 4707 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4708 spin_unlock(sb_bgl_lock(sbi, block_group));
4709 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4710
772cb7c8
JS		 4711 if (sbi->s_log_groups_per_flex) {
4712 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4713 spin_lock(sb_bgl_lock(sbi, flex_group));
4714 sbi->s_flex_groups[flex_group].free_blocks += count;
4715 spin_unlock(sb_bgl_lock(sbi, flex_group));
4716 }
4717
c9de560d
AT		 4718 ext4_mb_release_desc(&e4b);
4719
4720 *freed += count;
4721
4722 /* And the group descriptor block */
4723 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4724 ret = ext4_journal_dirty_metadata(handle, gd_bh);
4725 if (!err)
4726 err = ret;
4727
4728 if (overflow && !err) {
4729 block += count;
4730 count = overflow;
4731 put_bh(bitmap_bh);
4732 goto do_more;
4733 }
4734 sb->s_dirt = 1;
4735error_return:
4736 brelse(bitmap_bh);
4737 ext4_std_error(sb, err);
256bdb49
ES		 4738 if (ac)
4739 kmem_cache_free(ext4_ac_cachep, ac);
c9de560d
AT		 4740 return;
4741}
kernel source for telekom puls (alcatel/mediatek)