]>
Commit | Line | Data |
---|---|---|
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" |
6ba495e9 | 25 | #include <linux/debugfs.h> |
9bffad1e TT |
26 | #include <trace/events/ext4.h> |
27 | ||
c9de560d AT |
28 | /* |
29 | * MUSTDO: | |
30 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
31 | * - search for metadata in few groups | |
32 | * | |
33 | * TODO v4: | |
34 | * - normalization should take into account whether file is still open | |
35 | * - discard preallocations if no free space left (policy?) | |
36 | * - don't normalize tails | |
37 | * - quota | |
38 | * - reservation for superuser | |
39 | * | |
40 | * TODO v3: | |
41 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
42 | * - track min/max extents in each group for better group selection | |
43 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
44 | * - tree of groups sorted by number of free blocks | |
45 | * - error handling | |
46 | */ | |
47 | ||
48 | /* | |
49 | * The allocation request involve request for multiple number of blocks | |
50 | * near to the goal(block) value specified. | |
51 | * | |
b713a5ec TT |
52 | * During initialization phase of the allocator we decide to use the |
53 | * group preallocation or inode preallocation depending on the size of | |
54 | * the file. The size of the file could be the resulting file size we | |
55 | * would have after allocation, or the current file size, which ever | |
56 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
57 | * select to use the group preallocation. The default value of | |
58 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
59 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
60 | * terms of number of blocks. | |
c9de560d AT |
61 | * |
62 | * The main motivation for having small file use group preallocation is to | |
b713a5ec | 63 | * ensure that we have small files closer together on the disk. |
c9de560d | 64 | * |
b713a5ec TT |
65 | * First stage the allocator looks at the inode prealloc list, |
66 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
67 | * spaces for this particular inode. The inode prealloc space is | |
68 | * represented as: | |
c9de560d AT |
69 | * |
70 | * pa_lstart -> the logical start block for this prealloc space | |
71 | * pa_pstart -> the physical start block for this prealloc space | |
72 | * pa_len -> lenght for this prealloc space | |
73 | * pa_free -> free space available in this prealloc space | |
74 | * | |
75 | * The inode preallocation space is used looking at the _logical_ start | |
76 | * block. If only the logical file block falls within the range of prealloc | |
77 | * space we will consume the particular prealloc space. This make sure that | |
78 | * that the we have contiguous physical blocks representing the file blocks | |
79 | * | |
80 | * The important thing to be noted in case of inode prealloc space is that | |
81 | * we don't modify the values associated to inode prealloc space except | |
82 | * pa_free. | |
83 | * | |
84 | * If we are not able to find blocks in the inode prealloc space and if we | |
85 | * have the group allocation flag set then we look at the locality group | |
86 | * prealloc space. These are per CPU prealloc list repreasented as | |
87 | * | |
88 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
89 | * | |
90 | * The reason for having a per cpu locality group is to reduce the contention | |
91 | * between CPUs. It is possible to get scheduled at this point. | |
92 | * | |
93 | * The locality group prealloc space is used looking at whether we have | |
94 | * enough free space (pa_free) withing the prealloc space. | |
95 | * | |
96 | * If we can't allocate blocks via inode prealloc or/and locality group | |
97 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
98 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
99 | * mapped to the buddy and bitmap information regarding different | |
100 | * groups. The buddy information is attached to buddy cache inode so that | |
101 | * we can access them through the page cache. The information regarding | |
102 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
103 | * block bitmap and buddy information. The information are stored in the | |
104 | * inode as: | |
105 | * | |
106 | * { page } | |
c3a326a6 | 107 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
108 | * |
109 | * | |
110 | * one block each for bitmap and buddy information. So for each group we | |
111 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / | |
112 | * blocksize) blocks. So it can have information regarding groups_per_page | |
113 | * which is blocks_per_page/2 | |
114 | * | |
115 | * The buddy cache inode is not stored on disk. The inode is thrown | |
116 | * away when the filesystem is unmounted. | |
117 | * | |
118 | * We look for count number of blocks in the buddy cache. If we were able | |
119 | * to locate that many free blocks we return with additional information | |
120 | * regarding rest of the contiguous physical block available | |
121 | * | |
122 | * Before allocating blocks via buddy cache we normalize the request | |
123 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
124 | * blocks that we get after allocation is added to the respective prealloc | |
125 | * list. In case of inode preallocation we follow a list of heuristics | |
126 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
127 | * we are doing a group prealloc we try to normalize the request to | |
b713a5ec | 128 | * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is |
c9de560d | 129 | * 512 blocks. This can be tuned via |
b713a5ec | 130 | * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in |
c9de560d AT |
131 | * terms of number of blocks. If we have mounted the file system with -O |
132 | * stripe=<value> option the group prealloc request is normalized to the | |
133 | * stripe value (sbi->s_stripe) | |
134 | * | |
b713a5ec | 135 | * The regular allocator(using the buddy cache) supports few tunables. |
c9de560d | 136 | * |
b713a5ec TT |
137 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
138 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
139 | * /sys/fs/ext4/<partition>/mb_order2_req | |
c9de560d | 140 | * |
b713a5ec | 141 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
142 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
143 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 144 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
c9de560d | 145 | * stripe size (sbi->s_stripe), we try to search for contigous block in |
b713a5ec TT |
146 | * stripe size. This should result in better allocation on RAID setups. If |
147 | * not, we search in the specific group using bitmap for best extents. The | |
148 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 149 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 150 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
151 | * best extent in the found extents. Searching for the blocks starts with |
152 | * the group specified as the goal value in allocation context via | |
153 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
154 | * can used for allocation. ext4_mb_good_group explains how the groups are | |
155 | * checked. | |
156 | * | |
157 | * Both the prealloc space are getting populated as above. So for the first | |
158 | * request we will hit the buddy cache which will result in this prealloc | |
159 | * space getting filled. The prealloc space is then later used for the | |
160 | * subsequent request. | |
161 | */ | |
162 | ||
163 | /* | |
164 | * mballoc operates on the following data: | |
165 | * - on-disk bitmap | |
166 | * - in-core buddy (actually includes buddy and bitmap) | |
167 | * - preallocation descriptors (PAs) | |
168 | * | |
169 | * there are two types of preallocations: | |
170 | * - inode | |
171 | * assiged to specific inode and can be used for this inode only. | |
172 | * it describes part of inode's space preallocated to specific | |
173 | * physical blocks. any block from that preallocated can be used | |
174 | * independent. the descriptor just tracks number of blocks left | |
175 | * unused. so, before taking some block from descriptor, one must | |
176 | * make sure corresponded logical block isn't allocated yet. this | |
177 | * also means that freeing any block within descriptor's range | |
178 | * must discard all preallocated blocks. | |
179 | * - locality group | |
180 | * assigned to specific locality group which does not translate to | |
181 | * permanent set of inodes: inode can join and leave group. space | |
182 | * from this type of preallocation can be used for any inode. thus | |
183 | * it's consumed from the beginning to the end. | |
184 | * | |
185 | * relation between them can be expressed as: | |
186 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
187 | * | |
188 | * this mean blocks mballoc considers used are: | |
189 | * - allocated blocks (persistent) | |
190 | * - preallocated blocks (non-persistent) | |
191 | * | |
192 | * consistency in mballoc world means that at any time a block is either | |
193 | * free or used in ALL structures. notice: "any time" should not be read | |
194 | * literally -- time is discrete and delimited by locks. | |
195 | * | |
196 | * to keep it simple, we don't use block numbers, instead we count number of | |
197 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
198 | * | |
199 | * all operations can be expressed as: | |
200 | * - init buddy: buddy = on-disk + PAs | |
201 | * - new PA: buddy += N; PA = N | |
202 | * - use inode PA: on-disk += N; PA -= N | |
203 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
204 | * - use locality group PA on-disk += N; PA -= N | |
205 | * - discard locality group PA buddy -= PA; PA = 0 | |
206 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
207 | * is used in real operation because we can't know actual used | |
208 | * bits from PA, only from on-disk bitmap | |
209 | * | |
210 | * if we follow this strict logic, then all operations above should be atomic. | |
211 | * given some of them can block, we'd have to use something like semaphores | |
212 | * killing performance on high-end SMP hardware. let's try to relax it using | |
213 | * the following knowledge: | |
214 | * 1) if buddy is referenced, it's already initialized | |
215 | * 2) while block is used in buddy and the buddy is referenced, | |
216 | * nobody can re-allocate that block | |
217 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
218 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
219 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
220 | * block | |
221 | * | |
222 | * so, now we're building a concurrency table: | |
223 | * - init buddy vs. | |
224 | * - new PA | |
225 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
226 | * until PA is linked to allocation group to avoid concurrent buddy init | |
227 | * - use inode PA | |
228 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
229 | * given (3) we care that PA-=N operation doesn't interfere with init | |
230 | * - discard inode PA | |
231 | * the simplest way would be to have buddy initialized by the discard | |
232 | * - use locality group PA | |
233 | * again PA-=N must be serialized with init | |
234 | * - discard locality group PA | |
235 | * the simplest way would be to have buddy initialized by the discard | |
236 | * - new PA vs. | |
237 | * - use inode PA | |
238 | * i_data_sem serializes them | |
239 | * - discard inode PA | |
240 | * discard process must wait until PA isn't used by another process | |
241 | * - use locality group PA | |
242 | * some mutex should serialize them | |
243 | * - discard locality group PA | |
244 | * discard process must wait until PA isn't used by another process | |
245 | * - use inode PA | |
246 | * - use inode PA | |
247 | * i_data_sem or another mutex should serializes them | |
248 | * - discard inode PA | |
249 | * discard process must wait until PA isn't used by another process | |
250 | * - use locality group PA | |
251 | * nothing wrong here -- they're different PAs covering different blocks | |
252 | * - discard locality group PA | |
253 | * discard process must wait until PA isn't used by another process | |
254 | * | |
255 | * now we're ready to make few consequences: | |
256 | * - PA is referenced and while it is no discard is possible | |
257 | * - PA is referenced until block isn't marked in on-disk bitmap | |
258 | * - PA changes only after on-disk bitmap | |
259 | * - discard must not compete with init. either init is done before | |
260 | * any discard or they're serialized somehow | |
261 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
262 | * | |
263 | * a special case when we've used PA to emptiness. no need to modify buddy | |
264 | * in this case, but we should care about concurrent init | |
265 | * | |
266 | */ | |
267 | ||
268 | /* | |
269 | * Logic in few words: | |
270 | * | |
271 | * - allocation: | |
272 | * load group | |
273 | * find blocks | |
274 | * mark bits in on-disk bitmap | |
275 | * release group | |
276 | * | |
277 | * - use preallocation: | |
278 | * find proper PA (per-inode or group) | |
279 | * load group | |
280 | * mark bits in on-disk bitmap | |
281 | * release group | |
282 | * release PA | |
283 | * | |
284 | * - free: | |
285 | * load group | |
286 | * mark bits in on-disk bitmap | |
287 | * release group | |
288 | * | |
289 | * - discard preallocations in group: | |
290 | * mark PAs deleted | |
291 | * move them onto local list | |
292 | * load on-disk bitmap | |
293 | * load group | |
294 | * remove PA from object (inode or locality group) | |
295 | * mark free blocks in-core | |
296 | * | |
297 | * - discard inode's preallocations: | |
298 | */ | |
299 | ||
300 | /* | |
301 | * Locking rules | |
302 | * | |
303 | * Locks: | |
304 | * - bitlock on a group (group) | |
305 | * - object (inode/locality) (object) | |
306 | * - per-pa lock (pa) | |
307 | * | |
308 | * Paths: | |
309 | * - new pa | |
310 | * object | |
311 | * group | |
312 | * | |
313 | * - find and use pa: | |
314 | * pa | |
315 | * | |
316 | * - release consumed pa: | |
317 | * pa | |
318 | * group | |
319 | * object | |
320 | * | |
321 | * - generate in-core bitmap: | |
322 | * group | |
323 | * pa | |
324 | * | |
325 | * - discard all for given object (inode, locality group): | |
326 | * object | |
327 | * pa | |
328 | * group | |
329 | * | |
330 | * - discard all for given group: | |
331 | * group | |
332 | * pa | |
333 | * group | |
334 | * object | |
335 | * | |
336 | */ | |
c3a326a6 AK |
337 | static struct kmem_cache *ext4_pspace_cachep; |
338 | static struct kmem_cache *ext4_ac_cachep; | |
339 | static struct kmem_cache *ext4_free_ext_cachep; | |
340 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
341 | ext4_group_t group); | |
7a2fcbf7 AK |
342 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
343 | ext4_group_t group); | |
c3a326a6 AK |
344 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn); |
345 | ||
ffad0a44 AK |
346 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
347 | { | |
c9de560d | 348 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
349 | *bit += ((unsigned long) addr & 7UL) << 3; |
350 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 351 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
352 | *bit += ((unsigned long) addr & 3UL) << 3; |
353 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
354 | #else |
355 | #error "how many bits you are?!" | |
356 | #endif | |
ffad0a44 AK |
357 | return addr; |
358 | } | |
c9de560d AT |
359 | |
360 | static inline int mb_test_bit(int bit, void *addr) | |
361 | { | |
362 | /* | |
363 | * ext4_test_bit on architecture like powerpc | |
364 | * needs unsigned long aligned address | |
365 | */ | |
ffad0a44 | 366 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
367 | return ext4_test_bit(bit, addr); |
368 | } | |
369 | ||
370 | static inline void mb_set_bit(int bit, void *addr) | |
371 | { | |
ffad0a44 | 372 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
373 | ext4_set_bit(bit, addr); |
374 | } | |
375 | ||
c9de560d AT |
376 | static inline void mb_clear_bit(int bit, void *addr) |
377 | { | |
ffad0a44 | 378 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
379 | ext4_clear_bit(bit, addr); |
380 | } | |
381 | ||
ffad0a44 AK |
382 | static 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 | ||
395 | static 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 |
408 | static 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 | |
432 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
433 | int first, int count) | |
434 | { | |
435 | int i; | |
436 | struct super_block *sb = e4b->bd_sb; | |
437 | ||
438 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
439 | return; | |
bc8e6740 | 440 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
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); | |
5d1b1b3f AK |
448 | ext4_grp_locked_error(sb, e4b->bd_group, |
449 | __func__, "double-free of inode" | |
a9df9a49 | 450 | " %lu's block %llu(bit %u in group %u)", |
c9de560d AT |
451 | inode ? inode->i_ino : 0, blocknr, |
452 | first + i, e4b->bd_group); | |
453 | } | |
454 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
455 | } | |
456 | } | |
457 | ||
458 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
459 | { | |
460 | int i; | |
461 | ||
462 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
463 | return; | |
bc8e6740 | 464 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
465 | for (i = 0; i < count; i++) { |
466 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
467 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
468 | } | |
469 | } | |
470 | ||
471 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
472 | { | |
473 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { | |
474 | unsigned char *b1, *b2; | |
475 | int i; | |
476 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
477 | b2 = (unsigned char *) bitmap; | |
478 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
479 | if (b1[i] != b2[i]) { | |
a9df9a49 | 480 | printk(KERN_ERR "corruption in group %u " |
4776004f TT |
481 | "at byte %u(%u): %x in copy != %x " |
482 | "on disk/prealloc\n", | |
483 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
484 | BUG(); |
485 | } | |
486 | } | |
487 | } | |
488 | } | |
489 | ||
490 | #else | |
491 | static inline void mb_free_blocks_double(struct inode *inode, | |
492 | struct ext4_buddy *e4b, int first, int count) | |
493 | { | |
494 | return; | |
495 | } | |
496 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
497 | int first, int count) | |
498 | { | |
499 | return; | |
500 | } | |
501 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
502 | { | |
503 | return; | |
504 | } | |
505 | #endif | |
506 | ||
507 | #ifdef AGGRESSIVE_CHECK | |
508 | ||
509 | #define MB_CHECK_ASSERT(assert) \ | |
510 | do { \ | |
511 | if (!(assert)) { \ | |
512 | printk(KERN_EMERG \ | |
513 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
514 | function, file, line, # assert); \ | |
515 | BUG(); \ | |
516 | } \ | |
517 | } while (0) | |
518 | ||
519 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
520 | const char *function, int line) | |
521 | { | |
522 | struct super_block *sb = e4b->bd_sb; | |
523 | int order = e4b->bd_blkbits + 1; | |
524 | int max; | |
525 | int max2; | |
526 | int i; | |
527 | int j; | |
528 | int k; | |
529 | int count; | |
530 | struct ext4_group_info *grp; | |
531 | int fragments = 0; | |
532 | int fstart; | |
533 | struct list_head *cur; | |
534 | void *buddy; | |
535 | void *buddy2; | |
536 | ||
c9de560d AT |
537 | { |
538 | static int mb_check_counter; | |
539 | if (mb_check_counter++ % 100 != 0) | |
540 | return 0; | |
541 | } | |
542 | ||
543 | while (order > 1) { | |
544 | buddy = mb_find_buddy(e4b, order, &max); | |
545 | MB_CHECK_ASSERT(buddy); | |
546 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
547 | MB_CHECK_ASSERT(buddy2); | |
548 | MB_CHECK_ASSERT(buddy != buddy2); | |
549 | MB_CHECK_ASSERT(max * 2 == max2); | |
550 | ||
551 | count = 0; | |
552 | for (i = 0; i < max; i++) { | |
553 | ||
554 | if (mb_test_bit(i, buddy)) { | |
555 | /* only single bit in buddy2 may be 1 */ | |
556 | if (!mb_test_bit(i << 1, buddy2)) { | |
557 | MB_CHECK_ASSERT( | |
558 | mb_test_bit((i<<1)+1, buddy2)); | |
559 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
560 | MB_CHECK_ASSERT( | |
561 | mb_test_bit(i << 1, buddy2)); | |
562 | } | |
563 | continue; | |
564 | } | |
565 | ||
566 | /* both bits in buddy2 must be 0 */ | |
567 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); | |
568 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
569 | ||
570 | for (j = 0; j < (1 << order); j++) { | |
571 | k = (i * (1 << order)) + j; | |
572 | MB_CHECK_ASSERT( | |
573 | !mb_test_bit(k, EXT4_MB_BITMAP(e4b))); | |
574 | } | |
575 | count++; | |
576 | } | |
577 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
578 | order--; | |
579 | } | |
580 | ||
581 | fstart = -1; | |
582 | buddy = mb_find_buddy(e4b, 0, &max); | |
583 | for (i = 0; i < max; i++) { | |
584 | if (!mb_test_bit(i, buddy)) { | |
585 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
586 | if (fstart == -1) { | |
587 | fragments++; | |
588 | fstart = i; | |
589 | } | |
590 | continue; | |
591 | } | |
592 | fstart = -1; | |
593 | /* check used bits only */ | |
594 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
595 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
596 | k = i >> j; | |
597 | MB_CHECK_ASSERT(k < max2); | |
598 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
599 | } | |
600 | } | |
601 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
602 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
603 | ||
604 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
605 | buddy = mb_find_buddy(e4b, 0, &max); | |
606 | list_for_each(cur, &grp->bb_prealloc_list) { | |
607 | ext4_group_t groupnr; | |
608 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
609 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
610 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 611 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 612 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
613 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
614 | } | |
615 | return 0; | |
616 | } | |
617 | #undef MB_CHECK_ASSERT | |
618 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 619 | __FILE__, __func__, __LINE__) |
c9de560d AT |
620 | #else |
621 | #define mb_check_buddy(e4b) | |
622 | #endif | |
623 | ||
624 | /* FIXME!! need more doc */ | |
625 | static void ext4_mb_mark_free_simple(struct super_block *sb, | |
626 | void *buddy, unsigned first, int len, | |
627 | struct ext4_group_info *grp) | |
628 | { | |
629 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
630 | unsigned short min; | |
631 | unsigned short max; | |
632 | unsigned short chunk; | |
633 | unsigned short border; | |
634 | ||
b73fce69 | 635 | BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb)); |
c9de560d AT |
636 | |
637 | border = 2 << sb->s_blocksize_bits; | |
638 | ||
639 | while (len > 0) { | |
640 | /* find how many blocks can be covered since this position */ | |
641 | max = ffs(first | border) - 1; | |
642 | ||
643 | /* find how many blocks of power 2 we need to mark */ | |
644 | min = fls(len) - 1; | |
645 | ||
646 | if (max < min) | |
647 | min = max; | |
648 | chunk = 1 << min; | |
649 | ||
650 | /* mark multiblock chunks only */ | |
651 | grp->bb_counters[min]++; | |
652 | if (min > 0) | |
653 | mb_clear_bit(first >> min, | |
654 | buddy + sbi->s_mb_offsets[min]); | |
655 | ||
656 | len -= chunk; | |
657 | first += chunk; | |
658 | } | |
659 | } | |
660 | ||
089ceecc ES |
661 | static noinline_for_stack |
662 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
663 | void *buddy, void *bitmap, ext4_group_t group) |
664 | { | |
665 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
666 | unsigned short max = EXT4_BLOCKS_PER_GROUP(sb); | |
667 | unsigned short i = 0; | |
668 | unsigned short first; | |
669 | unsigned short len; | |
670 | unsigned free = 0; | |
671 | unsigned fragments = 0; | |
672 | unsigned long long period = get_cycles(); | |
673 | ||
674 | /* initialize buddy from bitmap which is aggregation | |
675 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 676 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
677 | grp->bb_first_free = i; |
678 | while (i < max) { | |
679 | fragments++; | |
680 | first = i; | |
ffad0a44 | 681 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
682 | len = i - first; |
683 | free += len; | |
684 | if (len > 1) | |
685 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
686 | else | |
687 | grp->bb_counters[0]++; | |
688 | if (i < max) | |
ffad0a44 | 689 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
690 | } |
691 | grp->bb_fragments = fragments; | |
692 | ||
693 | if (free != grp->bb_free) { | |
5d1b1b3f | 694 | ext4_grp_locked_error(sb, group, __func__, |
a9df9a49 | 695 | "EXT4-fs: group %u: %u blocks in bitmap, %u in gd", |
c9de560d | 696 | group, free, grp->bb_free); |
e56eb659 AK |
697 | /* |
698 | * If we intent to continue, we consider group descritor | |
699 | * corrupt and update bb_free using bitmap value | |
700 | */ | |
c9de560d AT |
701 | grp->bb_free = free; |
702 | } | |
703 | ||
704 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
705 | ||
706 | period = get_cycles() - period; | |
707 | spin_lock(&EXT4_SB(sb)->s_bal_lock); | |
708 | EXT4_SB(sb)->s_mb_buddies_generated++; | |
709 | EXT4_SB(sb)->s_mb_generation_time += period; | |
710 | spin_unlock(&EXT4_SB(sb)->s_bal_lock); | |
711 | } | |
712 | ||
713 | /* The buddy information is attached the buddy cache inode | |
714 | * for convenience. The information regarding each group | |
715 | * is loaded via ext4_mb_load_buddy. The information involve | |
716 | * block bitmap and buddy information. The information are | |
717 | * stored in the inode as | |
718 | * | |
719 | * { page } | |
c3a326a6 | 720 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
721 | * |
722 | * | |
723 | * one block each for bitmap and buddy information. | |
724 | * So for each group we take up 2 blocks. A page can | |
725 | * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. | |
726 | * So it can have information regarding groups_per_page which | |
727 | * is blocks_per_page/2 | |
728 | */ | |
729 | ||
730 | static int ext4_mb_init_cache(struct page *page, char *incore) | |
731 | { | |
8df9675f | 732 | ext4_group_t ngroups; |
c9de560d AT |
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 | ||
6ba495e9 | 747 | mb_debug(1, "init page %lu\n", page->index); |
c9de560d AT |
748 | |
749 | inode = page->mapping->host; | |
750 | sb = inode->i_sb; | |
8df9675f | 751 | ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
752 | blocksize = 1 << inode->i_blkbits; |
753 | blocks_per_page = PAGE_CACHE_SIZE / blocksize; | |
754 | ||
755 | groups_per_page = blocks_per_page >> 1; | |
756 | if (groups_per_page == 0) | |
757 | groups_per_page = 1; | |
758 | ||
759 | /* allocate buffer_heads to read bitmaps */ | |
760 | if (groups_per_page > 1) { | |
761 | err = -ENOMEM; | |
762 | i = sizeof(struct buffer_head *) * groups_per_page; | |
763 | bh = kzalloc(i, GFP_NOFS); | |
764 | if (bh == NULL) | |
765 | goto out; | |
766 | } else | |
767 | bh = &bhs; | |
768 | ||
769 | first_group = page->index * blocks_per_page / 2; | |
770 | ||
771 | /* read all groups the page covers into the cache */ | |
772 | for (i = 0; i < groups_per_page; i++) { | |
773 | struct ext4_group_desc *desc; | |
774 | ||
8df9675f | 775 | if (first_group + i >= ngroups) |
c9de560d AT |
776 | break; |
777 | ||
778 | err = -EIO; | |
779 | desc = ext4_get_group_desc(sb, first_group + i, NULL); | |
780 | if (desc == NULL) | |
781 | goto out; | |
782 | ||
783 | err = -ENOMEM; | |
784 | bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc)); | |
785 | if (bh[i] == NULL) | |
786 | goto out; | |
787 | ||
2ccb5fb9 | 788 | if (bitmap_uptodate(bh[i])) |
c9de560d AT |
789 | continue; |
790 | ||
c806e68f | 791 | lock_buffer(bh[i]); |
2ccb5fb9 AK |
792 | if (bitmap_uptodate(bh[i])) { |
793 | unlock_buffer(bh[i]); | |
794 | continue; | |
795 | } | |
955ce5f5 | 796 | ext4_lock_group(sb, first_group + i); |
c9de560d AT |
797 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
798 | ext4_init_block_bitmap(sb, bh[i], | |
799 | first_group + i, desc); | |
2ccb5fb9 | 800 | set_bitmap_uptodate(bh[i]); |
c9de560d | 801 | set_buffer_uptodate(bh[i]); |
955ce5f5 | 802 | ext4_unlock_group(sb, first_group + i); |
3300beda | 803 | unlock_buffer(bh[i]); |
c9de560d AT |
804 | continue; |
805 | } | |
955ce5f5 | 806 | ext4_unlock_group(sb, first_group + i); |
2ccb5fb9 AK |
807 | if (buffer_uptodate(bh[i])) { |
808 | /* | |
809 | * if not uninit if bh is uptodate, | |
810 | * bitmap is also uptodate | |
811 | */ | |
812 | set_bitmap_uptodate(bh[i]); | |
813 | unlock_buffer(bh[i]); | |
814 | continue; | |
815 | } | |
c9de560d | 816 | get_bh(bh[i]); |
2ccb5fb9 AK |
817 | /* |
818 | * submit the buffer_head for read. We can | |
819 | * safely mark the bitmap as uptodate now. | |
820 | * We do it here so the bitmap uptodate bit | |
821 | * get set with buffer lock held. | |
822 | */ | |
823 | set_bitmap_uptodate(bh[i]); | |
c9de560d AT |
824 | bh[i]->b_end_io = end_buffer_read_sync; |
825 | submit_bh(READ, bh[i]); | |
6ba495e9 | 826 | mb_debug(1, "read bitmap for group %u\n", first_group + i); |
c9de560d AT |
827 | } |
828 | ||
829 | /* wait for I/O completion */ | |
830 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
831 | wait_on_buffer(bh[i]); | |
832 | ||
833 | err = -EIO; | |
834 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
835 | if (!buffer_uptodate(bh[i])) | |
836 | goto out; | |
837 | ||
31b481dc | 838 | err = 0; |
c9de560d | 839 | first_block = page->index * blocks_per_page; |
29eaf024 AK |
840 | /* init the page */ |
841 | memset(page_address(page), 0xff, PAGE_CACHE_SIZE); | |
c9de560d AT |
842 | for (i = 0; i < blocks_per_page; i++) { |
843 | int group; | |
844 | struct ext4_group_info *grinfo; | |
845 | ||
846 | group = (first_block + i) >> 1; | |
8df9675f | 847 | if (group >= ngroups) |
c9de560d AT |
848 | break; |
849 | ||
850 | /* | |
851 | * data carry information regarding this | |
852 | * particular group in the format specified | |
853 | * above | |
854 | * | |
855 | */ | |
856 | data = page_address(page) + (i * blocksize); | |
857 | bitmap = bh[group - first_group]->b_data; | |
858 | ||
859 | /* | |
860 | * We place the buddy block and bitmap block | |
861 | * close together | |
862 | */ | |
863 | if ((first_block + i) & 1) { | |
864 | /* this is block of buddy */ | |
865 | BUG_ON(incore == NULL); | |
6ba495e9 | 866 | mb_debug(1, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 867 | group, page->index, i * blocksize); |
c9de560d AT |
868 | grinfo = ext4_get_group_info(sb, group); |
869 | grinfo->bb_fragments = 0; | |
870 | memset(grinfo->bb_counters, 0, | |
871 | sizeof(unsigned short)*(sb->s_blocksize_bits+2)); | |
872 | /* | |
873 | * incore got set to the group block bitmap below | |
874 | */ | |
7a2fcbf7 | 875 | ext4_lock_group(sb, group); |
c9de560d | 876 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 877 | ext4_unlock_group(sb, group); |
c9de560d AT |
878 | incore = NULL; |
879 | } else { | |
880 | /* this is block of bitmap */ | |
881 | BUG_ON(incore != NULL); | |
6ba495e9 | 882 | mb_debug(1, "put bitmap for group %u in page %lu/%x\n", |
c9de560d AT |
883 | group, page->index, i * blocksize); |
884 | ||
885 | /* see comments in ext4_mb_put_pa() */ | |
886 | ext4_lock_group(sb, group); | |
887 | memcpy(data, bitmap, blocksize); | |
888 | ||
889 | /* mark all preallocated blks used in in-core bitmap */ | |
890 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 891 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
892 | ext4_unlock_group(sb, group); |
893 | ||
894 | /* set incore so that the buddy information can be | |
895 | * generated using this | |
896 | */ | |
897 | incore = data; | |
898 | } | |
899 | } | |
900 | SetPageUptodate(page); | |
901 | ||
902 | out: | |
903 | if (bh) { | |
904 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
905 | brelse(bh[i]); | |
906 | if (bh != &bhs) | |
907 | kfree(bh); | |
908 | } | |
909 | return err; | |
910 | } | |
911 | ||
4ddfef7b ES |
912 | static noinline_for_stack int |
913 | ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, | |
914 | struct ext4_buddy *e4b) | |
c9de560d | 915 | { |
c9de560d AT |
916 | int blocks_per_page; |
917 | int block; | |
918 | int pnum; | |
919 | int poff; | |
920 | struct page *page; | |
fdf6c7a7 | 921 | int ret; |
920313a7 AK |
922 | struct ext4_group_info *grp; |
923 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
924 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 925 | |
6ba495e9 | 926 | mb_debug(1, "load group %u\n", group); |
c9de560d AT |
927 | |
928 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
920313a7 | 929 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
930 | |
931 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
932 | e4b->bd_info = ext4_get_group_info(sb, group); | |
933 | e4b->bd_sb = sb; | |
934 | e4b->bd_group = group; | |
935 | e4b->bd_buddy_page = NULL; | |
936 | e4b->bd_bitmap_page = NULL; | |
920313a7 AK |
937 | e4b->alloc_semp = &grp->alloc_sem; |
938 | ||
939 | /* Take the read lock on the group alloc | |
940 | * sem. This would make sure a parallel | |
941 | * ext4_mb_init_group happening on other | |
942 | * groups mapped by the page is blocked | |
943 | * till we are done with allocation | |
944 | */ | |
945 | down_read(e4b->alloc_semp); | |
c9de560d AT |
946 | |
947 | /* | |
948 | * the buddy cache inode stores the block bitmap | |
949 | * and buddy information in consecutive blocks. | |
950 | * So for each group we need two blocks. | |
951 | */ | |
952 | block = group * 2; | |
953 | pnum = block / blocks_per_page; | |
954 | poff = block % blocks_per_page; | |
955 | ||
956 | /* we could use find_or_create_page(), but it locks page | |
957 | * what we'd like to avoid in fast path ... */ | |
958 | page = find_get_page(inode->i_mapping, pnum); | |
959 | if (page == NULL || !PageUptodate(page)) { | |
960 | if (page) | |
920313a7 AK |
961 | /* |
962 | * drop the page reference and try | |
963 | * to get the page with lock. If we | |
964 | * are not uptodate that implies | |
965 | * somebody just created the page but | |
966 | * is yet to initialize the same. So | |
967 | * wait for it to initialize. | |
968 | */ | |
c9de560d AT |
969 | page_cache_release(page); |
970 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
971 | if (page) { | |
972 | BUG_ON(page->mapping != inode->i_mapping); | |
973 | if (!PageUptodate(page)) { | |
fdf6c7a7 SF |
974 | ret = ext4_mb_init_cache(page, NULL); |
975 | if (ret) { | |
976 | unlock_page(page); | |
977 | goto err; | |
978 | } | |
c9de560d AT |
979 | mb_cmp_bitmaps(e4b, page_address(page) + |
980 | (poff * sb->s_blocksize)); | |
981 | } | |
982 | unlock_page(page); | |
983 | } | |
984 | } | |
fdf6c7a7 SF |
985 | if (page == NULL || !PageUptodate(page)) { |
986 | ret = -EIO; | |
c9de560d | 987 | goto err; |
fdf6c7a7 | 988 | } |
c9de560d AT |
989 | e4b->bd_bitmap_page = page; |
990 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
991 | mark_page_accessed(page); | |
992 | ||
993 | block++; | |
994 | pnum = block / blocks_per_page; | |
995 | poff = block % blocks_per_page; | |
996 | ||
997 | page = find_get_page(inode->i_mapping, pnum); | |
998 | if (page == NULL || !PageUptodate(page)) { | |
999 | if (page) | |
1000 | page_cache_release(page); | |
1001 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1002 | if (page) { | |
1003 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 SF |
1004 | if (!PageUptodate(page)) { |
1005 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap); | |
1006 | if (ret) { | |
1007 | unlock_page(page); | |
1008 | goto err; | |
1009 | } | |
1010 | } | |
c9de560d AT |
1011 | unlock_page(page); |
1012 | } | |
1013 | } | |
fdf6c7a7 SF |
1014 | if (page == NULL || !PageUptodate(page)) { |
1015 | ret = -EIO; | |
c9de560d | 1016 | goto err; |
fdf6c7a7 | 1017 | } |
c9de560d AT |
1018 | e4b->bd_buddy_page = page; |
1019 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
1020 | mark_page_accessed(page); | |
1021 | ||
1022 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1023 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1024 | ||
1025 | return 0; | |
1026 | ||
1027 | err: | |
1028 | if (e4b->bd_bitmap_page) | |
1029 | page_cache_release(e4b->bd_bitmap_page); | |
1030 | if (e4b->bd_buddy_page) | |
1031 | page_cache_release(e4b->bd_buddy_page); | |
1032 | e4b->bd_buddy = NULL; | |
1033 | e4b->bd_bitmap = NULL; | |
920313a7 AK |
1034 | |
1035 | /* Done with the buddy cache */ | |
1036 | up_read(e4b->alloc_semp); | |
fdf6c7a7 | 1037 | return ret; |
c9de560d AT |
1038 | } |
1039 | ||
1040 | static void ext4_mb_release_desc(struct ext4_buddy *e4b) | |
1041 | { | |
1042 | if (e4b->bd_bitmap_page) | |
1043 | page_cache_release(e4b->bd_bitmap_page); | |
1044 | if (e4b->bd_buddy_page) | |
1045 | page_cache_release(e4b->bd_buddy_page); | |
920313a7 | 1046 | /* Done with the buddy cache */ |
8556e8f3 AK |
1047 | if (e4b->alloc_semp) |
1048 | up_read(e4b->alloc_semp); | |
c9de560d AT |
1049 | } |
1050 | ||
1051 | ||
1052 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1053 | { | |
1054 | int order = 1; | |
1055 | void *bb; | |
1056 | ||
1057 | BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); | |
1058 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); | |
1059 | ||
1060 | bb = EXT4_MB_BUDDY(e4b); | |
1061 | while (order <= e4b->bd_blkbits + 1) { | |
1062 | block = block >> 1; | |
1063 | if (!mb_test_bit(block, bb)) { | |
1064 | /* this block is part of buddy of order 'order' */ | |
1065 | return order; | |
1066 | } | |
1067 | bb += 1 << (e4b->bd_blkbits - order); | |
1068 | order++; | |
1069 | } | |
1070 | return 0; | |
1071 | } | |
1072 | ||
955ce5f5 | 1073 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1074 | { |
1075 | __u32 *addr; | |
1076 | ||
1077 | len = cur + len; | |
1078 | while (cur < len) { | |
1079 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1080 | /* fast path: clear whole word at once */ | |
1081 | addr = bm + (cur >> 3); | |
1082 | *addr = 0; | |
1083 | cur += 32; | |
1084 | continue; | |
1085 | } | |
955ce5f5 | 1086 | mb_clear_bit(cur, bm); |
c9de560d AT |
1087 | cur++; |
1088 | } | |
1089 | } | |
1090 | ||
955ce5f5 | 1091 | static void mb_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1092 | { |
1093 | __u32 *addr; | |
1094 | ||
1095 | len = cur + len; | |
1096 | while (cur < len) { | |
1097 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1098 | /* fast path: set whole word at once */ | |
1099 | addr = bm + (cur >> 3); | |
1100 | *addr = 0xffffffff; | |
1101 | cur += 32; | |
1102 | continue; | |
1103 | } | |
955ce5f5 | 1104 | mb_set_bit(cur, bm); |
c9de560d AT |
1105 | cur++; |
1106 | } | |
1107 | } | |
1108 | ||
7e5a8cdd | 1109 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
c9de560d AT |
1110 | int first, int count) |
1111 | { | |
1112 | int block = 0; | |
1113 | int max = 0; | |
1114 | int order; | |
1115 | void *buddy; | |
1116 | void *buddy2; | |
1117 | struct super_block *sb = e4b->bd_sb; | |
1118 | ||
1119 | BUG_ON(first + count > (sb->s_blocksize << 3)); | |
bc8e6740 | 1120 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
1121 | mb_check_buddy(e4b); |
1122 | mb_free_blocks_double(inode, e4b, first, count); | |
1123 | ||
1124 | e4b->bd_info->bb_free += count; | |
1125 | if (first < e4b->bd_info->bb_first_free) | |
1126 | e4b->bd_info->bb_first_free = first; | |
1127 | ||
1128 | /* let's maintain fragments counter */ | |
1129 | if (first != 0) | |
1130 | block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b)); | |
1131 | if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) | |
1132 | max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b)); | |
1133 | if (block && max) | |
1134 | e4b->bd_info->bb_fragments--; | |
1135 | else if (!block && !max) | |
1136 | e4b->bd_info->bb_fragments++; | |
1137 | ||
1138 | /* let's maintain buddy itself */ | |
1139 | while (count-- > 0) { | |
1140 | block = first++; | |
1141 | order = 0; | |
1142 | ||
1143 | if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) { | |
1144 | ext4_fsblk_t blocknr; | |
1145 | blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb); | |
1146 | blocknr += block; | |
1147 | blocknr += | |
1148 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
5d1b1b3f AK |
1149 | ext4_grp_locked_error(sb, e4b->bd_group, |
1150 | __func__, "double-free of inode" | |
a9df9a49 | 1151 | " %lu's block %llu(bit %u in group %u)", |
c9de560d AT |
1152 | inode ? inode->i_ino : 0, blocknr, block, |
1153 | e4b->bd_group); | |
1154 | } | |
1155 | mb_clear_bit(block, EXT4_MB_BITMAP(e4b)); | |
1156 | e4b->bd_info->bb_counters[order]++; | |
1157 | ||
1158 | /* start of the buddy */ | |
1159 | buddy = mb_find_buddy(e4b, order, &max); | |
1160 | ||
1161 | do { | |
1162 | block &= ~1UL; | |
1163 | if (mb_test_bit(block, buddy) || | |
1164 | mb_test_bit(block + 1, buddy)) | |
1165 | break; | |
1166 | ||
1167 | /* both the buddies are free, try to coalesce them */ | |
1168 | buddy2 = mb_find_buddy(e4b, order + 1, &max); | |
1169 | ||
1170 | if (!buddy2) | |
1171 | break; | |
1172 | ||
1173 | if (order > 0) { | |
1174 | /* for special purposes, we don't set | |
1175 | * free bits in bitmap */ | |
1176 | mb_set_bit(block, buddy); | |
1177 | mb_set_bit(block + 1, buddy); | |
1178 | } | |
1179 | e4b->bd_info->bb_counters[order]--; | |
1180 | e4b->bd_info->bb_counters[order]--; | |
1181 | ||
1182 | block = block >> 1; | |
1183 | order++; | |
1184 | e4b->bd_info->bb_counters[order]++; | |
1185 | ||
1186 | mb_clear_bit(block, buddy2); | |
1187 | buddy = buddy2; | |
1188 | } while (1); | |
1189 | } | |
1190 | mb_check_buddy(e4b); | |
c9de560d AT |
1191 | } |
1192 | ||
1193 | static int mb_find_extent(struct ext4_buddy *e4b, int order, int block, | |
1194 | int needed, struct ext4_free_extent *ex) | |
1195 | { | |
1196 | int next = block; | |
1197 | int max; | |
1198 | int ord; | |
1199 | void *buddy; | |
1200 | ||
bc8e6740 | 1201 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1202 | BUG_ON(ex == NULL); |
1203 | ||
1204 | buddy = mb_find_buddy(e4b, order, &max); | |
1205 | BUG_ON(buddy == NULL); | |
1206 | BUG_ON(block >= max); | |
1207 | if (mb_test_bit(block, buddy)) { | |
1208 | ex->fe_len = 0; | |
1209 | ex->fe_start = 0; | |
1210 | ex->fe_group = 0; | |
1211 | return 0; | |
1212 | } | |
1213 | ||
1214 | /* FIXME dorp order completely ? */ | |
1215 | if (likely(order == 0)) { | |
1216 | /* find actual order */ | |
1217 | order = mb_find_order_for_block(e4b, block); | |
1218 | block = block >> order; | |
1219 | } | |
1220 | ||
1221 | ex->fe_len = 1 << order; | |
1222 | ex->fe_start = block << order; | |
1223 | ex->fe_group = e4b->bd_group; | |
1224 | ||
1225 | /* calc difference from given start */ | |
1226 | next = next - ex->fe_start; | |
1227 | ex->fe_len -= next; | |
1228 | ex->fe_start += next; | |
1229 | ||
1230 | while (needed > ex->fe_len && | |
1231 | (buddy = mb_find_buddy(e4b, order, &max))) { | |
1232 | ||
1233 | if (block + 1 >= max) | |
1234 | break; | |
1235 | ||
1236 | next = (block + 1) * (1 << order); | |
1237 | if (mb_test_bit(next, EXT4_MB_BITMAP(e4b))) | |
1238 | break; | |
1239 | ||
1240 | ord = mb_find_order_for_block(e4b, next); | |
1241 | ||
1242 | order = ord; | |
1243 | block = next >> order; | |
1244 | ex->fe_len += 1 << order; | |
1245 | } | |
1246 | ||
1247 | BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); | |
1248 | return ex->fe_len; | |
1249 | } | |
1250 | ||
1251 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1252 | { | |
1253 | int ord; | |
1254 | int mlen = 0; | |
1255 | int max = 0; | |
1256 | int cur; | |
1257 | int start = ex->fe_start; | |
1258 | int len = ex->fe_len; | |
1259 | unsigned ret = 0; | |
1260 | int len0 = len; | |
1261 | void *buddy; | |
1262 | ||
1263 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1264 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1265 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1266 | mb_check_buddy(e4b); |
1267 | mb_mark_used_double(e4b, start, len); | |
1268 | ||
1269 | e4b->bd_info->bb_free -= len; | |
1270 | if (e4b->bd_info->bb_first_free == start) | |
1271 | e4b->bd_info->bb_first_free += len; | |
1272 | ||
1273 | /* let's maintain fragments counter */ | |
1274 | if (start != 0) | |
1275 | mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b)); | |
1276 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) | |
1277 | max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b)); | |
1278 | if (mlen && max) | |
1279 | e4b->bd_info->bb_fragments++; | |
1280 | else if (!mlen && !max) | |
1281 | e4b->bd_info->bb_fragments--; | |
1282 | ||
1283 | /* let's maintain buddy itself */ | |
1284 | while (len) { | |
1285 | ord = mb_find_order_for_block(e4b, start); | |
1286 | ||
1287 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1288 | /* the whole chunk may be allocated at once! */ | |
1289 | mlen = 1 << ord; | |
1290 | buddy = mb_find_buddy(e4b, ord, &max); | |
1291 | BUG_ON((start >> ord) >= max); | |
1292 | mb_set_bit(start >> ord, buddy); | |
1293 | e4b->bd_info->bb_counters[ord]--; | |
1294 | start += mlen; | |
1295 | len -= mlen; | |
1296 | BUG_ON(len < 0); | |
1297 | continue; | |
1298 | } | |
1299 | ||
1300 | /* store for history */ | |
1301 | if (ret == 0) | |
1302 | ret = len | (ord << 16); | |
1303 | ||
1304 | /* we have to split large buddy */ | |
1305 | BUG_ON(ord <= 0); | |
1306 | buddy = mb_find_buddy(e4b, ord, &max); | |
1307 | mb_set_bit(start >> ord, buddy); | |
1308 | e4b->bd_info->bb_counters[ord]--; | |
1309 | ||
1310 | ord--; | |
1311 | cur = (start >> ord) & ~1U; | |
1312 | buddy = mb_find_buddy(e4b, ord, &max); | |
1313 | mb_clear_bit(cur, buddy); | |
1314 | mb_clear_bit(cur + 1, buddy); | |
1315 | e4b->bd_info->bb_counters[ord]++; | |
1316 | e4b->bd_info->bb_counters[ord]++; | |
1317 | } | |
1318 | ||
955ce5f5 | 1319 | mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0); |
c9de560d AT |
1320 | mb_check_buddy(e4b); |
1321 | ||
1322 | return ret; | |
1323 | } | |
1324 | ||
1325 | /* | |
1326 | * Must be called under group lock! | |
1327 | */ | |
1328 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1329 | struct ext4_buddy *e4b) | |
1330 | { | |
1331 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1332 | int ret; | |
1333 | ||
1334 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1335 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1336 | ||
1337 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1338 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1339 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1340 | ||
1341 | /* preallocation can change ac_b_ex, thus we store actually | |
1342 | * allocated blocks for history */ | |
1343 | ac->ac_f_ex = ac->ac_b_ex; | |
1344 | ||
1345 | ac->ac_status = AC_STATUS_FOUND; | |
1346 | ac->ac_tail = ret & 0xffff; | |
1347 | ac->ac_buddy = ret >> 16; | |
1348 | ||
c3a326a6 AK |
1349 | /* |
1350 | * take the page reference. We want the page to be pinned | |
1351 | * so that we don't get a ext4_mb_init_cache_call for this | |
1352 | * group until we update the bitmap. That would mean we | |
1353 | * double allocate blocks. The reference is dropped | |
1354 | * in ext4_mb_release_context | |
1355 | */ | |
c9de560d AT |
1356 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1357 | get_page(ac->ac_bitmap_page); | |
1358 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1359 | get_page(ac->ac_buddy_page); | |
8556e8f3 AK |
1360 | /* on allocation we use ac to track the held semaphore */ |
1361 | ac->alloc_semp = e4b->alloc_semp; | |
1362 | e4b->alloc_semp = NULL; | |
c9de560d | 1363 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 1364 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
1365 | spin_lock(&sbi->s_md_lock); |
1366 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1367 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1368 | spin_unlock(&sbi->s_md_lock); | |
1369 | } | |
1370 | } | |
1371 | ||
1372 | /* | |
1373 | * regular allocator, for general purposes allocation | |
1374 | */ | |
1375 | ||
1376 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1377 | struct ext4_buddy *e4b, | |
1378 | int finish_group) | |
1379 | { | |
1380 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1381 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1382 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1383 | struct ext4_free_extent ex; | |
1384 | int max; | |
1385 | ||
032115fc AK |
1386 | if (ac->ac_status == AC_STATUS_FOUND) |
1387 | return; | |
c9de560d AT |
1388 | /* |
1389 | * We don't want to scan for a whole year | |
1390 | */ | |
1391 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1392 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1393 | ac->ac_status = AC_STATUS_BREAK; | |
1394 | return; | |
1395 | } | |
1396 | ||
1397 | /* | |
1398 | * Haven't found good chunk so far, let's continue | |
1399 | */ | |
1400 | if (bex->fe_len < gex->fe_len) | |
1401 | return; | |
1402 | ||
1403 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1404 | && bex->fe_group == e4b->bd_group) { | |
1405 | /* recheck chunk's availability - we don't know | |
1406 | * when it was found (within this lock-unlock | |
1407 | * period or not) */ | |
1408 | max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex); | |
1409 | if (max >= gex->fe_len) { | |
1410 | ext4_mb_use_best_found(ac, e4b); | |
1411 | return; | |
1412 | } | |
1413 | } | |
1414 | } | |
1415 | ||
1416 | /* | |
1417 | * The routine checks whether found extent is good enough. If it is, | |
1418 | * then the extent gets marked used and flag is set to the context | |
1419 | * to stop scanning. Otherwise, the extent is compared with the | |
1420 | * previous found extent and if new one is better, then it's stored | |
1421 | * in the context. Later, the best found extent will be used, if | |
1422 | * mballoc can't find good enough extent. | |
1423 | * | |
1424 | * FIXME: real allocation policy is to be designed yet! | |
1425 | */ | |
1426 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1427 | struct ext4_free_extent *ex, | |
1428 | struct ext4_buddy *e4b) | |
1429 | { | |
1430 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1431 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1432 | ||
1433 | BUG_ON(ex->fe_len <= 0); | |
8d03c7a0 | 1434 | BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
1435 | BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
1436 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); | |
1437 | ||
1438 | ac->ac_found++; | |
1439 | ||
1440 | /* | |
1441 | * The special case - take what you catch first | |
1442 | */ | |
1443 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1444 | *bex = *ex; | |
1445 | ext4_mb_use_best_found(ac, e4b); | |
1446 | return; | |
1447 | } | |
1448 | ||
1449 | /* | |
1450 | * Let's check whether the chuck is good enough | |
1451 | */ | |
1452 | if (ex->fe_len == gex->fe_len) { | |
1453 | *bex = *ex; | |
1454 | ext4_mb_use_best_found(ac, e4b); | |
1455 | return; | |
1456 | } | |
1457 | ||
1458 | /* | |
1459 | * If this is first found extent, just store it in the context | |
1460 | */ | |
1461 | if (bex->fe_len == 0) { | |
1462 | *bex = *ex; | |
1463 | return; | |
1464 | } | |
1465 | ||
1466 | /* | |
1467 | * If new found extent is better, store it in the context | |
1468 | */ | |
1469 | if (bex->fe_len < gex->fe_len) { | |
1470 | /* if the request isn't satisfied, any found extent | |
1471 | * larger than previous best one is better */ | |
1472 | if (ex->fe_len > bex->fe_len) | |
1473 | *bex = *ex; | |
1474 | } else if (ex->fe_len > gex->fe_len) { | |
1475 | /* if the request is satisfied, then we try to find | |
1476 | * an extent that still satisfy the request, but is | |
1477 | * smaller than previous one */ | |
1478 | if (ex->fe_len < bex->fe_len) | |
1479 | *bex = *ex; | |
1480 | } | |
1481 | ||
1482 | ext4_mb_check_limits(ac, e4b, 0); | |
1483 | } | |
1484 | ||
089ceecc ES |
1485 | static noinline_for_stack |
1486 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
1487 | struct ext4_buddy *e4b) |
1488 | { | |
1489 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1490 | ext4_group_t group = ex.fe_group; | |
1491 | int max; | |
1492 | int err; | |
1493 | ||
1494 | BUG_ON(ex.fe_len <= 0); | |
1495 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1496 | if (err) | |
1497 | return err; | |
1498 | ||
1499 | ext4_lock_group(ac->ac_sb, group); | |
1500 | max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex); | |
1501 | ||
1502 | if (max > 0) { | |
1503 | ac->ac_b_ex = ex; | |
1504 | ext4_mb_use_best_found(ac, e4b); | |
1505 | } | |
1506 | ||
1507 | ext4_unlock_group(ac->ac_sb, group); | |
1508 | ext4_mb_release_desc(e4b); | |
1509 | ||
1510 | return 0; | |
1511 | } | |
1512 | ||
089ceecc ES |
1513 | static noinline_for_stack |
1514 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
1515 | struct ext4_buddy *e4b) |
1516 | { | |
1517 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1518 | int max; | |
1519 | int err; | |
1520 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1521 | struct ext4_super_block *es = sbi->s_es; | |
1522 | struct ext4_free_extent ex; | |
1523 | ||
1524 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1525 | return 0; | |
1526 | ||
1527 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1528 | if (err) | |
1529 | return err; | |
1530 | ||
1531 | ext4_lock_group(ac->ac_sb, group); | |
1532 | max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start, | |
1533 | ac->ac_g_ex.fe_len, &ex); | |
1534 | ||
1535 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1536 | ext4_fsblk_t start; | |
1537 | ||
1538 | start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) + | |
1539 | ex.fe_start + le32_to_cpu(es->s_first_data_block); | |
1540 | /* use do_div to get remainder (would be 64-bit modulo) */ | |
1541 | if (do_div(start, sbi->s_stripe) == 0) { | |
1542 | ac->ac_found++; | |
1543 | ac->ac_b_ex = ex; | |
1544 | ext4_mb_use_best_found(ac, e4b); | |
1545 | } | |
1546 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1547 | BUG_ON(ex.fe_len <= 0); | |
1548 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1549 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1550 | ac->ac_found++; | |
1551 | ac->ac_b_ex = ex; | |
1552 | ext4_mb_use_best_found(ac, e4b); | |
1553 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1554 | /* Sometimes, caller may want to merge even small | |
1555 | * number of blocks to an existing extent */ | |
1556 | BUG_ON(ex.fe_len <= 0); | |
1557 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1558 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1559 | ac->ac_found++; | |
1560 | ac->ac_b_ex = ex; | |
1561 | ext4_mb_use_best_found(ac, e4b); | |
1562 | } | |
1563 | ext4_unlock_group(ac->ac_sb, group); | |
1564 | ext4_mb_release_desc(e4b); | |
1565 | ||
1566 | return 0; | |
1567 | } | |
1568 | ||
1569 | /* | |
1570 | * The routine scans buddy structures (not bitmap!) from given order | |
1571 | * to max order and tries to find big enough chunk to satisfy the req | |
1572 | */ | |
089ceecc ES |
1573 | static noinline_for_stack |
1574 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1575 | struct ext4_buddy *e4b) |
1576 | { | |
1577 | struct super_block *sb = ac->ac_sb; | |
1578 | struct ext4_group_info *grp = e4b->bd_info; | |
1579 | void *buddy; | |
1580 | int i; | |
1581 | int k; | |
1582 | int max; | |
1583 | ||
1584 | BUG_ON(ac->ac_2order <= 0); | |
1585 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1586 | if (grp->bb_counters[i] == 0) | |
1587 | continue; | |
1588 | ||
1589 | buddy = mb_find_buddy(e4b, i, &max); | |
1590 | BUG_ON(buddy == NULL); | |
1591 | ||
ffad0a44 | 1592 | k = mb_find_next_zero_bit(buddy, max, 0); |
c9de560d AT |
1593 | BUG_ON(k >= max); |
1594 | ||
1595 | ac->ac_found++; | |
1596 | ||
1597 | ac->ac_b_ex.fe_len = 1 << i; | |
1598 | ac->ac_b_ex.fe_start = k << i; | |
1599 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1600 | ||
1601 | ext4_mb_use_best_found(ac, e4b); | |
1602 | ||
1603 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1604 | ||
1605 | if (EXT4_SB(sb)->s_mb_stats) | |
1606 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1607 | ||
1608 | break; | |
1609 | } | |
1610 | } | |
1611 | ||
1612 | /* | |
1613 | * The routine scans the group and measures all found extents. | |
1614 | * In order to optimize scanning, caller must pass number of | |
1615 | * free blocks in the group, so the routine can know upper limit. | |
1616 | */ | |
089ceecc ES |
1617 | static noinline_for_stack |
1618 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1619 | struct ext4_buddy *e4b) |
1620 | { | |
1621 | struct super_block *sb = ac->ac_sb; | |
1622 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1623 | struct ext4_free_extent ex; | |
1624 | int i; | |
1625 | int free; | |
1626 | ||
1627 | free = e4b->bd_info->bb_free; | |
1628 | BUG_ON(free <= 0); | |
1629 | ||
1630 | i = e4b->bd_info->bb_first_free; | |
1631 | ||
1632 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1633 | i = mb_find_next_zero_bit(bitmap, |
c9de560d AT |
1634 | EXT4_BLOCKS_PER_GROUP(sb), i); |
1635 | if (i >= EXT4_BLOCKS_PER_GROUP(sb)) { | |
26346ff6 | 1636 | /* |
e56eb659 | 1637 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1638 | * free blocks even though group info says we |
1639 | * we have free blocks | |
1640 | */ | |
5d1b1b3f AK |
1641 | ext4_grp_locked_error(sb, e4b->bd_group, |
1642 | __func__, "%d free blocks as per " | |
fde4d95a | 1643 | "group info. But bitmap says 0", |
26346ff6 | 1644 | free); |
c9de560d AT |
1645 | break; |
1646 | } | |
1647 | ||
1648 | mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex); | |
1649 | BUG_ON(ex.fe_len <= 0); | |
26346ff6 | 1650 | if (free < ex.fe_len) { |
5d1b1b3f AK |
1651 | ext4_grp_locked_error(sb, e4b->bd_group, |
1652 | __func__, "%d free blocks as per " | |
fde4d95a | 1653 | "group info. But got %d blocks", |
26346ff6 | 1654 | free, ex.fe_len); |
e56eb659 AK |
1655 | /* |
1656 | * The number of free blocks differs. This mostly | |
1657 | * indicate that the bitmap is corrupt. So exit | |
1658 | * without claiming the space. | |
1659 | */ | |
1660 | break; | |
26346ff6 | 1661 | } |
c9de560d AT |
1662 | |
1663 | ext4_mb_measure_extent(ac, &ex, e4b); | |
1664 | ||
1665 | i += ex.fe_len; | |
1666 | free -= ex.fe_len; | |
1667 | } | |
1668 | ||
1669 | ext4_mb_check_limits(ac, e4b, 1); | |
1670 | } | |
1671 | ||
1672 | /* | |
1673 | * This is a special case for storages like raid5 | |
1674 | * we try to find stripe-aligned chunks for stripe-size requests | |
1675 | * XXX should do so at least for multiples of stripe size as well | |
1676 | */ | |
089ceecc ES |
1677 | static noinline_for_stack |
1678 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
1679 | struct ext4_buddy *e4b) |
1680 | { | |
1681 | struct super_block *sb = ac->ac_sb; | |
1682 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1683 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1684 | struct ext4_free_extent ex; | |
1685 | ext4_fsblk_t first_group_block; | |
1686 | ext4_fsblk_t a; | |
1687 | ext4_grpblk_t i; | |
1688 | int max; | |
1689 | ||
1690 | BUG_ON(sbi->s_stripe == 0); | |
1691 | ||
1692 | /* find first stripe-aligned block in group */ | |
1693 | first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb) | |
1694 | + le32_to_cpu(sbi->s_es->s_first_data_block); | |
1695 | a = first_group_block + sbi->s_stripe - 1; | |
1696 | do_div(a, sbi->s_stripe); | |
1697 | i = (a * sbi->s_stripe) - first_group_block; | |
1698 | ||
1699 | while (i < EXT4_BLOCKS_PER_GROUP(sb)) { | |
1700 | if (!mb_test_bit(i, bitmap)) { | |
1701 | max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex); | |
1702 | if (max >= sbi->s_stripe) { | |
1703 | ac->ac_found++; | |
1704 | ac->ac_b_ex = ex; | |
1705 | ext4_mb_use_best_found(ac, e4b); | |
1706 | break; | |
1707 | } | |
1708 | } | |
1709 | i += sbi->s_stripe; | |
1710 | } | |
1711 | } | |
1712 | ||
1713 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, | |
1714 | ext4_group_t group, int cr) | |
1715 | { | |
1716 | unsigned free, fragments; | |
1717 | unsigned i, bits; | |
a4912123 | 1718 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
1719 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
1720 | ||
1721 | BUG_ON(cr < 0 || cr >= 4); | |
1722 | BUG_ON(EXT4_MB_GRP_NEED_INIT(grp)); | |
1723 | ||
1724 | free = grp->bb_free; | |
1725 | fragments = grp->bb_fragments; | |
1726 | if (free == 0) | |
1727 | return 0; | |
1728 | if (fragments == 0) | |
1729 | return 0; | |
1730 | ||
1731 | switch (cr) { | |
1732 | case 0: | |
1733 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 1734 | |
a4912123 TT |
1735 | /* Avoid using the first bg of a flexgroup for data files */ |
1736 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
1737 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
1738 | ((group % flex_size) == 0)) | |
1739 | return 0; | |
1740 | ||
c9de560d AT |
1741 | bits = ac->ac_sb->s_blocksize_bits + 1; |
1742 | for (i = ac->ac_2order; i <= bits; i++) | |
1743 | if (grp->bb_counters[i] > 0) | |
1744 | return 1; | |
1745 | break; | |
1746 | case 1: | |
1747 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
1748 | return 1; | |
1749 | break; | |
1750 | case 2: | |
1751 | if (free >= ac->ac_g_ex.fe_len) | |
1752 | return 1; | |
1753 | break; | |
1754 | case 3: | |
1755 | return 1; | |
1756 | default: | |
1757 | BUG(); | |
1758 | } | |
1759 | ||
1760 | return 0; | |
1761 | } | |
1762 | ||
920313a7 AK |
1763 | /* |
1764 | * lock the group_info alloc_sem of all the groups | |
1765 | * belonging to the same buddy cache page. This | |
1766 | * make sure other parallel operation on the buddy | |
1767 | * cache doesn't happen whild holding the buddy cache | |
1768 | * lock | |
1769 | */ | |
1770 | int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group) | |
1771 | { | |
1772 | int i; | |
1773 | int block, pnum; | |
1774 | int blocks_per_page; | |
1775 | int groups_per_page; | |
8df9675f | 1776 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
920313a7 AK |
1777 | ext4_group_t first_group; |
1778 | struct ext4_group_info *grp; | |
1779 | ||
1780 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1781 | /* | |
1782 | * the buddy cache inode stores the block bitmap | |
1783 | * and buddy information in consecutive blocks. | |
1784 | * So for each group we need two blocks. | |
1785 | */ | |
1786 | block = group * 2; | |
1787 | pnum = block / blocks_per_page; | |
1788 | first_group = pnum * blocks_per_page / 2; | |
1789 | ||
1790 | groups_per_page = blocks_per_page >> 1; | |
1791 | if (groups_per_page == 0) | |
1792 | groups_per_page = 1; | |
1793 | /* read all groups the page covers into the cache */ | |
1794 | for (i = 0; i < groups_per_page; i++) { | |
1795 | ||
8df9675f | 1796 | if ((first_group + i) >= ngroups) |
920313a7 AK |
1797 | break; |
1798 | grp = ext4_get_group_info(sb, first_group + i); | |
1799 | /* take all groups write allocation | |
1800 | * semaphore. This make sure there is | |
1801 | * no block allocation going on in any | |
1802 | * of that groups | |
1803 | */ | |
b7be019e | 1804 | down_write_nested(&grp->alloc_sem, i); |
920313a7 AK |
1805 | } |
1806 | return i; | |
1807 | } | |
1808 | ||
1809 | void ext4_mb_put_buddy_cache_lock(struct super_block *sb, | |
1810 | ext4_group_t group, int locked_group) | |
1811 | { | |
1812 | int i; | |
1813 | int block, pnum; | |
1814 | int blocks_per_page; | |
1815 | ext4_group_t first_group; | |
1816 | struct ext4_group_info *grp; | |
1817 | ||
1818 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1819 | /* | |
1820 | * the buddy cache inode stores the block bitmap | |
1821 | * and buddy information in consecutive blocks. | |
1822 | * So for each group we need two blocks. | |
1823 | */ | |
1824 | block = group * 2; | |
1825 | pnum = block / blocks_per_page; | |
1826 | first_group = pnum * blocks_per_page / 2; | |
1827 | /* release locks on all the groups */ | |
1828 | for (i = 0; i < locked_group; i++) { | |
1829 | ||
1830 | grp = ext4_get_group_info(sb, first_group + i); | |
1831 | /* take all groups write allocation | |
1832 | * semaphore. This make sure there is | |
1833 | * no block allocation going on in any | |
1834 | * of that groups | |
1835 | */ | |
1836 | up_write(&grp->alloc_sem); | |
1837 | } | |
1838 | ||
1839 | } | |
1840 | ||
089ceecc ES |
1841 | static noinline_for_stack |
1842 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) | |
920313a7 AK |
1843 | { |
1844 | ||
1845 | int ret; | |
1846 | void *bitmap; | |
1847 | int blocks_per_page; | |
1848 | int block, pnum, poff; | |
1849 | int num_grp_locked = 0; | |
1850 | struct ext4_group_info *this_grp; | |
1851 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1852 | struct inode *inode = sbi->s_buddy_cache; | |
1853 | struct page *page = NULL, *bitmap_page = NULL; | |
1854 | ||
6ba495e9 | 1855 | mb_debug(1, "init group %u\n", group); |
920313a7 AK |
1856 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; |
1857 | this_grp = ext4_get_group_info(sb, group); | |
1858 | /* | |
1859 | * This ensures we don't add group | |
1860 | * to this buddy cache via resize | |
1861 | */ | |
1862 | num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group); | |
1863 | if (!EXT4_MB_GRP_NEED_INIT(this_grp)) { | |
1864 | /* | |
1865 | * somebody initialized the group | |
1866 | * return without doing anything | |
1867 | */ | |
1868 | ret = 0; | |
1869 | goto err; | |
1870 | } | |
1871 | /* | |
1872 | * the buddy cache inode stores the block bitmap | |
1873 | * and buddy information in consecutive blocks. | |
1874 | * So for each group we need two blocks. | |
1875 | */ | |
1876 | block = group * 2; | |
1877 | pnum = block / blocks_per_page; | |
1878 | poff = block % blocks_per_page; | |
1879 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1880 | if (page) { | |
1881 | BUG_ON(page->mapping != inode->i_mapping); | |
1882 | ret = ext4_mb_init_cache(page, NULL); | |
1883 | if (ret) { | |
1884 | unlock_page(page); | |
1885 | goto err; | |
1886 | } | |
1887 | unlock_page(page); | |
1888 | } | |
1889 | if (page == NULL || !PageUptodate(page)) { | |
1890 | ret = -EIO; | |
1891 | goto err; | |
1892 | } | |
1893 | mark_page_accessed(page); | |
1894 | bitmap_page = page; | |
1895 | bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1896 | ||
1897 | /* init buddy cache */ | |
1898 | block++; | |
1899 | pnum = block / blocks_per_page; | |
1900 | poff = block % blocks_per_page; | |
1901 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1902 | if (page == bitmap_page) { | |
1903 | /* | |
1904 | * If both the bitmap and buddy are in | |
1905 | * the same page we don't need to force | |
1906 | * init the buddy | |
1907 | */ | |
1908 | unlock_page(page); | |
1909 | } else if (page) { | |
1910 | BUG_ON(page->mapping != inode->i_mapping); | |
1911 | ret = ext4_mb_init_cache(page, bitmap); | |
1912 | if (ret) { | |
1913 | unlock_page(page); | |
1914 | goto err; | |
1915 | } | |
1916 | unlock_page(page); | |
1917 | } | |
1918 | if (page == NULL || !PageUptodate(page)) { | |
1919 | ret = -EIO; | |
1920 | goto err; | |
1921 | } | |
1922 | mark_page_accessed(page); | |
1923 | err: | |
1924 | ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked); | |
1925 | if (bitmap_page) | |
1926 | page_cache_release(bitmap_page); | |
1927 | if (page) | |
1928 | page_cache_release(page); | |
1929 | return ret; | |
1930 | } | |
1931 | ||
4ddfef7b ES |
1932 | static noinline_for_stack int |
1933 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 1934 | { |
8df9675f | 1935 | ext4_group_t ngroups, group, i; |
c9de560d AT |
1936 | int cr; |
1937 | int err = 0; | |
1938 | int bsbits; | |
1939 | struct ext4_sb_info *sbi; | |
1940 | struct super_block *sb; | |
1941 | struct ext4_buddy e4b; | |
c9de560d AT |
1942 | |
1943 | sb = ac->ac_sb; | |
1944 | sbi = EXT4_SB(sb); | |
8df9675f | 1945 | ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
1946 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
1947 | ||
1948 | /* first, try the goal */ | |
1949 | err = ext4_mb_find_by_goal(ac, &e4b); | |
1950 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
1951 | goto out; | |
1952 | ||
1953 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
1954 | goto out; | |
1955 | ||
1956 | /* | |
1957 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
1958 | * if ac2_order is set we also set criteria to 0 so that we | |
1959 | * try exact allocation using buddy. | |
1960 | */ | |
1961 | i = fls(ac->ac_g_ex.fe_len); | |
1962 | ac->ac_2order = 0; | |
1963 | /* | |
1964 | * We search using buddy data only if the order of the request | |
1965 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 1966 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
c9de560d AT |
1967 | */ |
1968 | if (i >= sbi->s_mb_order2_reqs) { | |
1969 | /* | |
1970 | * This should tell if fe_len is exactly power of 2 | |
1971 | */ | |
1972 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1973 | ac->ac_2order = i - 1; | |
1974 | } | |
1975 | ||
1976 | bsbits = ac->ac_sb->s_blocksize_bits; | |
c9de560d | 1977 | |
4ba74d00 TT |
1978 | /* if stream allocation is enabled, use global goal */ |
1979 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
1980 | /* TBD: may be hot point */ |
1981 | spin_lock(&sbi->s_md_lock); | |
1982 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
1983 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
1984 | spin_unlock(&sbi->s_md_lock); | |
1985 | } | |
4ba74d00 | 1986 | |
c9de560d AT |
1987 | /* Let's just scan groups to find more-less suitable blocks */ |
1988 | cr = ac->ac_2order ? 0 : 1; | |
1989 | /* | |
1990 | * cr == 0 try to get exact allocation, | |
1991 | * cr == 3 try to get anything | |
1992 | */ | |
1993 | repeat: | |
1994 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
1995 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
1996 | /* |
1997 | * searching for the right group start | |
1998 | * from the goal value specified | |
1999 | */ | |
2000 | group = ac->ac_g_ex.fe_group; | |
2001 | ||
8df9675f | 2002 | for (i = 0; i < ngroups; group++, i++) { |
c9de560d AT |
2003 | struct ext4_group_info *grp; |
2004 | struct ext4_group_desc *desc; | |
2005 | ||
8df9675f | 2006 | if (group == ngroups) |
c9de560d AT |
2007 | group = 0; |
2008 | ||
2009 | /* quick check to skip empty groups */ | |
920313a7 | 2010 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
2011 | if (grp->bb_free == 0) |
2012 | continue; | |
2013 | ||
2014 | /* | |
2015 | * if the group is already init we check whether it is | |
2016 | * a good group and if not we don't load the buddy | |
2017 | */ | |
2018 | if (EXT4_MB_GRP_NEED_INIT(grp)) { | |
2019 | /* | |
2020 | * we need full data about the group | |
2021 | * to make a good selection | |
2022 | */ | |
920313a7 | 2023 | err = ext4_mb_init_group(sb, group); |
c9de560d AT |
2024 | if (err) |
2025 | goto out; | |
c9de560d AT |
2026 | } |
2027 | ||
2028 | /* | |
2029 | * If the particular group doesn't satisfy our | |
2030 | * criteria we continue with the next group | |
2031 | */ | |
2032 | if (!ext4_mb_good_group(ac, group, cr)) | |
2033 | continue; | |
2034 | ||
2035 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2036 | if (err) | |
2037 | goto out; | |
2038 | ||
2039 | ext4_lock_group(sb, group); | |
2040 | if (!ext4_mb_good_group(ac, group, cr)) { | |
2041 | /* someone did allocation from this group */ | |
2042 | ext4_unlock_group(sb, group); | |
2043 | ext4_mb_release_desc(&e4b); | |
2044 | continue; | |
2045 | } | |
2046 | ||
2047 | ac->ac_groups_scanned++; | |
2048 | desc = ext4_get_group_desc(sb, group, NULL); | |
75507efb | 2049 | if (cr == 0) |
c9de560d AT |
2050 | ext4_mb_simple_scan_group(ac, &e4b); |
2051 | else if (cr == 1 && | |
2052 | ac->ac_g_ex.fe_len == sbi->s_stripe) | |
2053 | ext4_mb_scan_aligned(ac, &e4b); | |
2054 | else | |
2055 | ext4_mb_complex_scan_group(ac, &e4b); | |
2056 | ||
2057 | ext4_unlock_group(sb, group); | |
2058 | ext4_mb_release_desc(&e4b); | |
2059 | ||
2060 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2061 | break; | |
2062 | } | |
2063 | } | |
2064 | ||
2065 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2066 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2067 | /* | |
2068 | * We've been searching too long. Let's try to allocate | |
2069 | * the best chunk we've found so far | |
2070 | */ | |
2071 | ||
2072 | ext4_mb_try_best_found(ac, &e4b); | |
2073 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2074 | /* | |
2075 | * Someone more lucky has already allocated it. | |
2076 | * The only thing we can do is just take first | |
2077 | * found block(s) | |
2078 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2079 | */ | |
2080 | ac->ac_b_ex.fe_group = 0; | |
2081 | ac->ac_b_ex.fe_start = 0; | |
2082 | ac->ac_b_ex.fe_len = 0; | |
2083 | ac->ac_status = AC_STATUS_CONTINUE; | |
2084 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2085 | cr = 3; | |
2086 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2087 | goto repeat; | |
2088 | } | |
2089 | } | |
2090 | out: | |
2091 | return err; | |
2092 | } | |
2093 | ||
2094 | #ifdef EXT4_MB_HISTORY | |
2095 | struct ext4_mb_proc_session { | |
2096 | struct ext4_mb_history *history; | |
2097 | struct super_block *sb; | |
2098 | int start; | |
2099 | int max; | |
2100 | }; | |
2101 | ||
2102 | static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s, | |
2103 | struct ext4_mb_history *hs, | |
2104 | int first) | |
2105 | { | |
2106 | if (hs == s->history + s->max) | |
2107 | hs = s->history; | |
2108 | if (!first && hs == s->history + s->start) | |
2109 | return NULL; | |
2110 | while (hs->orig.fe_len == 0) { | |
2111 | hs++; | |
2112 | if (hs == s->history + s->max) | |
2113 | hs = s->history; | |
2114 | if (hs == s->history + s->start) | |
2115 | return NULL; | |
2116 | } | |
2117 | return hs; | |
2118 | } | |
2119 | ||
2120 | static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos) | |
2121 | { | |
2122 | struct ext4_mb_proc_session *s = seq->private; | |
2123 | struct ext4_mb_history *hs; | |
2124 | int l = *pos; | |
2125 | ||
2126 | if (l == 0) | |
2127 | return SEQ_START_TOKEN; | |
2128 | hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1); | |
2129 | if (!hs) | |
2130 | return NULL; | |
2131 | while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL); | |
2132 | return hs; | |
2133 | } | |
2134 | ||
2135 | static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v, | |
2136 | loff_t *pos) | |
2137 | { | |
2138 | struct ext4_mb_proc_session *s = seq->private; | |
2139 | struct ext4_mb_history *hs = v; | |
2140 | ||
2141 | ++*pos; | |
2142 | if (v == SEQ_START_TOKEN) | |
2143 | return ext4_mb_history_skip_empty(s, s->history + s->start, 1); | |
2144 | else | |
2145 | return ext4_mb_history_skip_empty(s, ++hs, 0); | |
2146 | } | |
2147 | ||
2148 | static int ext4_mb_seq_history_show(struct seq_file *seq, void *v) | |
2149 | { | |
2150 | char buf[25], buf2[25], buf3[25], *fmt; | |
2151 | struct ext4_mb_history *hs = v; | |
2152 | ||
2153 | if (v == SEQ_START_TOKEN) { | |
2154 | seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s " | |
0ef90db9 | 2155 | "%-5s %-2s %-6s %-5s %-5s %-6s\n", |
c9de560d AT |
2156 | "pid", "inode", "original", "goal", "result", "found", |
2157 | "grps", "cr", "flags", "merge", "tail", "broken"); | |
2158 | return 0; | |
2159 | } | |
2160 | ||
2161 | if (hs->op == EXT4_MB_HISTORY_ALLOC) { | |
2162 | fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u " | |
0ef90db9 | 2163 | "0x%04x %-5s %-5u %-6u\n"; |
a9df9a49 | 2164 | sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group, |
c9de560d AT |
2165 | hs->result.fe_start, hs->result.fe_len, |
2166 | hs->result.fe_logical); | |
a9df9a49 | 2167 | sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group, |
c9de560d AT |
2168 | hs->orig.fe_start, hs->orig.fe_len, |
2169 | hs->orig.fe_logical); | |
a9df9a49 | 2170 | sprintf(buf3, "%u/%d/%u@%u", hs->goal.fe_group, |
c9de560d AT |
2171 | hs->goal.fe_start, hs->goal.fe_len, |
2172 | hs->goal.fe_logical); | |
2173 | seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2, | |
2174 | hs->found, hs->groups, hs->cr, hs->flags, | |
2175 | hs->merged ? "M" : "", hs->tail, | |
2176 | hs->buddy ? 1 << hs->buddy : 0); | |
2177 | } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) { | |
2178 | fmt = "%-5u %-8u %-23s %-23s %-23s\n"; | |
a9df9a49 | 2179 | sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group, |
c9de560d AT |
2180 | hs->result.fe_start, hs->result.fe_len, |
2181 | hs->result.fe_logical); | |
a9df9a49 | 2182 | sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group, |
c9de560d AT |
2183 | hs->orig.fe_start, hs->orig.fe_len, |
2184 | hs->orig.fe_logical); | |
2185 | seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2); | |
2186 | } else if (hs->op == EXT4_MB_HISTORY_DISCARD) { | |
a9df9a49 | 2187 | sprintf(buf2, "%u/%d/%u", hs->result.fe_group, |
c9de560d AT |
2188 | hs->result.fe_start, hs->result.fe_len); |
2189 | seq_printf(seq, "%-5u %-8u %-23s discard\n", | |
2190 | hs->pid, hs->ino, buf2); | |
2191 | } else if (hs->op == EXT4_MB_HISTORY_FREE) { | |
a9df9a49 | 2192 | sprintf(buf2, "%u/%d/%u", hs->result.fe_group, |
c9de560d AT |
2193 | hs->result.fe_start, hs->result.fe_len); |
2194 | seq_printf(seq, "%-5u %-8u %-23s free\n", | |
2195 | hs->pid, hs->ino, buf2); | |
2196 | } | |
2197 | return 0; | |
2198 | } | |
2199 | ||
2200 | static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v) | |
2201 | { | |
2202 | } | |
2203 | ||
2204 | static struct seq_operations ext4_mb_seq_history_ops = { | |
2205 | .start = ext4_mb_seq_history_start, | |
2206 | .next = ext4_mb_seq_history_next, | |
2207 | .stop = ext4_mb_seq_history_stop, | |
2208 | .show = ext4_mb_seq_history_show, | |
2209 | }; | |
2210 | ||
2211 | static int ext4_mb_seq_history_open(struct inode *inode, struct file *file) | |
2212 | { | |
2213 | struct super_block *sb = PDE(inode)->data; | |
2214 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2215 | struct ext4_mb_proc_session *s; | |
2216 | int rc; | |
2217 | int size; | |
2218 | ||
74767c5a SF |
2219 | if (unlikely(sbi->s_mb_history == NULL)) |
2220 | return -ENOMEM; | |
c9de560d AT |
2221 | s = kmalloc(sizeof(*s), GFP_KERNEL); |
2222 | if (s == NULL) | |
2223 | return -ENOMEM; | |
2224 | s->sb = sb; | |
2225 | size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max; | |
2226 | s->history = kmalloc(size, GFP_KERNEL); | |
2227 | if (s->history == NULL) { | |
2228 | kfree(s); | |
2229 | return -ENOMEM; | |
2230 | } | |
2231 | ||
2232 | spin_lock(&sbi->s_mb_history_lock); | |
2233 | memcpy(s->history, sbi->s_mb_history, size); | |
2234 | s->max = sbi->s_mb_history_max; | |
2235 | s->start = sbi->s_mb_history_cur % s->max; | |
2236 | spin_unlock(&sbi->s_mb_history_lock); | |
2237 | ||
2238 | rc = seq_open(file, &ext4_mb_seq_history_ops); | |
2239 | if (rc == 0) { | |
2240 | struct seq_file *m = (struct seq_file *)file->private_data; | |
2241 | m->private = s; | |
2242 | } else { | |
2243 | kfree(s->history); | |
2244 | kfree(s); | |
2245 | } | |
2246 | return rc; | |
2247 | ||
2248 | } | |
2249 | ||
2250 | static int ext4_mb_seq_history_release(struct inode *inode, struct file *file) | |
2251 | { | |
2252 | struct seq_file *seq = (struct seq_file *)file->private_data; | |
2253 | struct ext4_mb_proc_session *s = seq->private; | |
2254 | kfree(s->history); | |
2255 | kfree(s); | |
2256 | return seq_release(inode, file); | |
2257 | } | |
2258 | ||
2259 | static ssize_t ext4_mb_seq_history_write(struct file *file, | |
2260 | const char __user *buffer, | |
2261 | size_t count, loff_t *ppos) | |
2262 | { | |
2263 | struct seq_file *seq = (struct seq_file *)file->private_data; | |
2264 | struct ext4_mb_proc_session *s = seq->private; | |
2265 | struct super_block *sb = s->sb; | |
2266 | char str[32]; | |
2267 | int value; | |
2268 | ||
2269 | if (count >= sizeof(str)) { | |
2270 | printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n", | |
2271 | "mb_history", (int)sizeof(str)); | |
2272 | return -EOVERFLOW; | |
2273 | } | |
2274 | ||
2275 | if (copy_from_user(str, buffer, count)) | |
2276 | return -EFAULT; | |
2277 | ||
2278 | value = simple_strtol(str, NULL, 0); | |
2279 | if (value < 0) | |
2280 | return -ERANGE; | |
2281 | EXT4_SB(sb)->s_mb_history_filter = value; | |
2282 | ||
2283 | return count; | |
2284 | } | |
2285 | ||
2286 | static struct file_operations ext4_mb_seq_history_fops = { | |
2287 | .owner = THIS_MODULE, | |
2288 | .open = ext4_mb_seq_history_open, | |
2289 | .read = seq_read, | |
2290 | .write = ext4_mb_seq_history_write, | |
2291 | .llseek = seq_lseek, | |
2292 | .release = ext4_mb_seq_history_release, | |
2293 | }; | |
2294 | ||
2295 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) | |
2296 | { | |
2297 | struct super_block *sb = seq->private; | |
c9de560d AT |
2298 | ext4_group_t group; |
2299 | ||
8df9675f | 2300 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2301 | return NULL; |
c9de560d | 2302 | group = *pos + 1; |
a9df9a49 | 2303 | return (void *) ((unsigned long) group); |
c9de560d AT |
2304 | } |
2305 | ||
2306 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2307 | { | |
2308 | struct super_block *sb = seq->private; | |
c9de560d AT |
2309 | ext4_group_t group; |
2310 | ||
2311 | ++*pos; | |
8df9675f | 2312 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2313 | return NULL; |
2314 | group = *pos + 1; | |
a9df9a49 | 2315 | return (void *) ((unsigned long) group); |
c9de560d AT |
2316 | } |
2317 | ||
2318 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2319 | { | |
2320 | struct super_block *sb = seq->private; | |
a9df9a49 | 2321 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d AT |
2322 | int i; |
2323 | int err; | |
2324 | struct ext4_buddy e4b; | |
2325 | struct sg { | |
2326 | struct ext4_group_info info; | |
2327 | unsigned short counters[16]; | |
2328 | } sg; | |
2329 | ||
2330 | group--; | |
2331 | if (group == 0) | |
2332 | seq_printf(seq, "#%-5s: %-5s %-5s %-5s " | |
2333 | "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " | |
2334 | "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", | |
2335 | "group", "free", "frags", "first", | |
2336 | "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", | |
2337 | "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); | |
2338 | ||
2339 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2340 | sizeof(struct ext4_group_info); | |
2341 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2342 | if (err) { | |
a9df9a49 | 2343 | seq_printf(seq, "#%-5u: I/O error\n", group); |
c9de560d AT |
2344 | return 0; |
2345 | } | |
2346 | ext4_lock_group(sb, group); | |
2347 | memcpy(&sg, ext4_get_group_info(sb, group), i); | |
2348 | ext4_unlock_group(sb, group); | |
2349 | ext4_mb_release_desc(&e4b); | |
2350 | ||
a9df9a49 | 2351 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2352 | sg.info.bb_fragments, sg.info.bb_first_free); |
2353 | for (i = 0; i <= 13; i++) | |
2354 | seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? | |
2355 | sg.info.bb_counters[i] : 0); | |
2356 | seq_printf(seq, " ]\n"); | |
2357 | ||
2358 | return 0; | |
2359 | } | |
2360 | ||
2361 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2362 | { | |
2363 | } | |
2364 | ||
2365 | static struct seq_operations ext4_mb_seq_groups_ops = { | |
2366 | .start = ext4_mb_seq_groups_start, | |
2367 | .next = ext4_mb_seq_groups_next, | |
2368 | .stop = ext4_mb_seq_groups_stop, | |
2369 | .show = ext4_mb_seq_groups_show, | |
2370 | }; | |
2371 | ||
2372 | static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) | |
2373 | { | |
2374 | struct super_block *sb = PDE(inode)->data; | |
2375 | int rc; | |
2376 | ||
2377 | rc = seq_open(file, &ext4_mb_seq_groups_ops); | |
2378 | if (rc == 0) { | |
2379 | struct seq_file *m = (struct seq_file *)file->private_data; | |
2380 | m->private = sb; | |
2381 | } | |
2382 | return rc; | |
2383 | ||
2384 | } | |
2385 | ||
2386 | static struct file_operations ext4_mb_seq_groups_fops = { | |
2387 | .owner = THIS_MODULE, | |
2388 | .open = ext4_mb_seq_groups_open, | |
2389 | .read = seq_read, | |
2390 | .llseek = seq_lseek, | |
2391 | .release = seq_release, | |
2392 | }; | |
2393 | ||
2394 | static void ext4_mb_history_release(struct super_block *sb) | |
2395 | { | |
2396 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2397 | ||
9f6200bb TT |
2398 | if (sbi->s_proc != NULL) { |
2399 | remove_proc_entry("mb_groups", sbi->s_proc); | |
f4033903 CW |
2400 | if (sbi->s_mb_history_max) |
2401 | remove_proc_entry("mb_history", sbi->s_proc); | |
9f6200bb | 2402 | } |
c9de560d AT |
2403 | kfree(sbi->s_mb_history); |
2404 | } | |
2405 | ||
2406 | static void ext4_mb_history_init(struct super_block *sb) | |
2407 | { | |
2408 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2409 | int i; | |
2410 | ||
9f6200bb | 2411 | if (sbi->s_proc != NULL) { |
f4033903 CW |
2412 | if (sbi->s_mb_history_max) |
2413 | proc_create_data("mb_history", S_IRUGO, sbi->s_proc, | |
2414 | &ext4_mb_seq_history_fops, sb); | |
9f6200bb | 2415 | proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, |
46fe74f2 | 2416 | &ext4_mb_seq_groups_fops, sb); |
c9de560d AT |
2417 | } |
2418 | ||
c9de560d AT |
2419 | sbi->s_mb_history_cur = 0; |
2420 | spin_lock_init(&sbi->s_mb_history_lock); | |
2421 | i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history); | |
f4033903 | 2422 | sbi->s_mb_history = i ? kzalloc(i, GFP_KERNEL) : NULL; |
c9de560d AT |
2423 | /* if we can't allocate history, then we simple won't use it */ |
2424 | } | |
2425 | ||
4ddfef7b ES |
2426 | static noinline_for_stack void |
2427 | ext4_mb_store_history(struct ext4_allocation_context *ac) | |
c9de560d AT |
2428 | { |
2429 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2430 | struct ext4_mb_history h; | |
2431 | ||
f4033903 | 2432 | if (sbi->s_mb_history == NULL) |
c9de560d AT |
2433 | return; |
2434 | ||
2435 | if (!(ac->ac_op & sbi->s_mb_history_filter)) | |
2436 | return; | |
2437 | ||
2438 | h.op = ac->ac_op; | |
2439 | h.pid = current->pid; | |
2440 | h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0; | |
2441 | h.orig = ac->ac_o_ex; | |
2442 | h.result = ac->ac_b_ex; | |
2443 | h.flags = ac->ac_flags; | |
2444 | h.found = ac->ac_found; | |
2445 | h.groups = ac->ac_groups_scanned; | |
2446 | h.cr = ac->ac_criteria; | |
2447 | h.tail = ac->ac_tail; | |
2448 | h.buddy = ac->ac_buddy; | |
2449 | h.merged = 0; | |
2450 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) { | |
2451 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
2452 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
2453 | h.merged = 1; | |
2454 | h.goal = ac->ac_g_ex; | |
2455 | h.result = ac->ac_f_ex; | |
2456 | } | |
2457 | ||
2458 | spin_lock(&sbi->s_mb_history_lock); | |
2459 | memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h)); | |
2460 | if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max) | |
2461 | sbi->s_mb_history_cur = 0; | |
2462 | spin_unlock(&sbi->s_mb_history_lock); | |
2463 | } | |
2464 | ||
2465 | #else | |
2466 | #define ext4_mb_history_release(sb) | |
2467 | #define ext4_mb_history_init(sb) | |
2468 | #endif | |
2469 | ||
5f21b0e6 FB |
2470 | |
2471 | /* Create and initialize ext4_group_info data for the given group. */ | |
920313a7 | 2472 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2473 | struct ext4_group_desc *desc) |
2474 | { | |
2475 | int i, len; | |
2476 | int metalen = 0; | |
2477 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2478 | struct ext4_group_info **meta_group_info; | |
2479 | ||
2480 | /* | |
2481 | * First check if this group is the first of a reserved block. | |
2482 | * If it's true, we have to allocate a new table of pointers | |
2483 | * to ext4_group_info structures | |
2484 | */ | |
2485 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2486 | metalen = sizeof(*meta_group_info) << | |
2487 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2488 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2489 | if (meta_group_info == NULL) { | |
2490 | printk(KERN_ERR "EXT4-fs: can't allocate mem for a " | |
2491 | "buddy group\n"); | |
2492 | goto exit_meta_group_info; | |
2493 | } | |
2494 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = | |
2495 | meta_group_info; | |
2496 | } | |
2497 | ||
2498 | /* | |
2499 | * calculate needed size. if change bb_counters size, | |
2500 | * don't forget about ext4_mb_generate_buddy() | |
2501 | */ | |
2502 | len = offsetof(typeof(**meta_group_info), | |
2503 | bb_counters[sb->s_blocksize_bits + 2]); | |
2504 | ||
2505 | meta_group_info = | |
2506 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; | |
2507 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); | |
2508 | ||
2509 | meta_group_info[i] = kzalloc(len, GFP_KERNEL); | |
2510 | if (meta_group_info[i] == NULL) { | |
2511 | printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n"); | |
2512 | goto exit_group_info; | |
2513 | } | |
2514 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2515 | &(meta_group_info[i]->bb_state)); | |
2516 | ||
2517 | /* | |
2518 | * initialize bb_free to be able to skip | |
2519 | * empty groups without initialization | |
2520 | */ | |
2521 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2522 | meta_group_info[i]->bb_free = | |
2523 | ext4_free_blocks_after_init(sb, group, desc); | |
2524 | } else { | |
2525 | meta_group_info[i]->bb_free = | |
560671a0 | 2526 | ext4_free_blks_count(sb, desc); |
5f21b0e6 FB |
2527 | } |
2528 | ||
2529 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2530 | init_rwsem(&meta_group_info[i]->alloc_sem); |
5a4a7989 | 2531 | meta_group_info[i]->bb_free_root.rb_node = NULL; |
5f21b0e6 FB |
2532 | |
2533 | #ifdef DOUBLE_CHECK | |
2534 | { | |
2535 | struct buffer_head *bh; | |
2536 | meta_group_info[i]->bb_bitmap = | |
2537 | kmalloc(sb->s_blocksize, GFP_KERNEL); | |
2538 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); | |
2539 | bh = ext4_read_block_bitmap(sb, group); | |
2540 | BUG_ON(bh == NULL); | |
2541 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, | |
2542 | sb->s_blocksize); | |
2543 | put_bh(bh); | |
2544 | } | |
2545 | #endif | |
2546 | ||
2547 | return 0; | |
2548 | ||
2549 | exit_group_info: | |
2550 | /* If a meta_group_info table has been allocated, release it now */ | |
2551 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) | |
2552 | kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); | |
2553 | exit_meta_group_info: | |
2554 | return -ENOMEM; | |
2555 | } /* ext4_mb_add_groupinfo */ | |
2556 | ||
c9de560d AT |
2557 | static int ext4_mb_init_backend(struct super_block *sb) |
2558 | { | |
8df9675f | 2559 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2560 | ext4_group_t i; |
c9de560d | 2561 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
5f21b0e6 FB |
2562 | struct ext4_super_block *es = sbi->s_es; |
2563 | int num_meta_group_infos; | |
2564 | int num_meta_group_infos_max; | |
2565 | int array_size; | |
5f21b0e6 FB |
2566 | struct ext4_group_desc *desc; |
2567 | ||
2568 | /* This is the number of blocks used by GDT */ | |
8df9675f | 2569 | num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) - |
5f21b0e6 FB |
2570 | 1) >> EXT4_DESC_PER_BLOCK_BITS(sb); |
2571 | ||
2572 | /* | |
2573 | * This is the total number of blocks used by GDT including | |
2574 | * the number of reserved blocks for GDT. | |
2575 | * The s_group_info array is allocated with this value | |
2576 | * to allow a clean online resize without a complex | |
2577 | * manipulation of pointer. | |
2578 | * The drawback is the unused memory when no resize | |
2579 | * occurs but it's very low in terms of pages | |
2580 | * (see comments below) | |
2581 | * Need to handle this properly when META_BG resizing is allowed | |
2582 | */ | |
2583 | num_meta_group_infos_max = num_meta_group_infos + | |
2584 | le16_to_cpu(es->s_reserved_gdt_blocks); | |
c9de560d | 2585 | |
5f21b0e6 FB |
2586 | /* |
2587 | * array_size is the size of s_group_info array. We round it | |
2588 | * to the next power of two because this approximation is done | |
2589 | * internally by kmalloc so we can have some more memory | |
2590 | * for free here (e.g. may be used for META_BG resize). | |
2591 | */ | |
2592 | array_size = 1; | |
2593 | while (array_size < sizeof(*sbi->s_group_info) * | |
2594 | num_meta_group_infos_max) | |
2595 | array_size = array_size << 1; | |
c9de560d AT |
2596 | /* An 8TB filesystem with 64-bit pointers requires a 4096 byte |
2597 | * kmalloc. A 128kb malloc should suffice for a 256TB filesystem. | |
2598 | * So a two level scheme suffices for now. */ | |
5f21b0e6 | 2599 | sbi->s_group_info = kmalloc(array_size, GFP_KERNEL); |
c9de560d AT |
2600 | if (sbi->s_group_info == NULL) { |
2601 | printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n"); | |
2602 | return -ENOMEM; | |
2603 | } | |
2604 | sbi->s_buddy_cache = new_inode(sb); | |
2605 | if (sbi->s_buddy_cache == NULL) { | |
2606 | printk(KERN_ERR "EXT4-fs: can't get new inode\n"); | |
2607 | goto err_freesgi; | |
2608 | } | |
2609 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; | |
8df9675f | 2610 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2611 | desc = ext4_get_group_desc(sb, i, NULL); |
2612 | if (desc == NULL) { | |
2613 | printk(KERN_ERR | |
a9df9a49 | 2614 | "EXT4-fs: can't read descriptor %u\n", i); |
c9de560d AT |
2615 | goto err_freebuddy; |
2616 | } | |
5f21b0e6 FB |
2617 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2618 | goto err_freebuddy; | |
c9de560d AT |
2619 | } |
2620 | ||
2621 | return 0; | |
2622 | ||
2623 | err_freebuddy: | |
f1fa3342 | 2624 | while (i-- > 0) |
c9de560d | 2625 | kfree(ext4_get_group_info(sb, i)); |
c9de560d | 2626 | i = num_meta_group_infos; |
f1fa3342 | 2627 | while (i-- > 0) |
c9de560d AT |
2628 | kfree(sbi->s_group_info[i]); |
2629 | iput(sbi->s_buddy_cache); | |
2630 | err_freesgi: | |
2631 | kfree(sbi->s_group_info); | |
2632 | return -ENOMEM; | |
2633 | } | |
2634 | ||
2635 | int ext4_mb_init(struct super_block *sb, int needs_recovery) | |
2636 | { | |
2637 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2638 | unsigned i, j; |
c9de560d AT |
2639 | unsigned offset; |
2640 | unsigned max; | |
74767c5a | 2641 | int ret; |
c9de560d | 2642 | |
c9de560d AT |
2643 | i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short); |
2644 | ||
2645 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2646 | if (sbi->s_mb_offsets == NULL) { | |
c9de560d AT |
2647 | return -ENOMEM; |
2648 | } | |
ff7ef329 YG |
2649 | |
2650 | i = (sb->s_blocksize_bits + 2) * sizeof(unsigned int); | |
c9de560d AT |
2651 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2652 | if (sbi->s_mb_maxs == NULL) { | |
a7b19448 | 2653 | kfree(sbi->s_mb_offsets); |
c9de560d AT |
2654 | return -ENOMEM; |
2655 | } | |
2656 | ||
2657 | /* order 0 is regular bitmap */ | |
2658 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2659 | sbi->s_mb_offsets[0] = 0; | |
2660 | ||
2661 | i = 1; | |
2662 | offset = 0; | |
2663 | max = sb->s_blocksize << 2; | |
2664 | do { | |
2665 | sbi->s_mb_offsets[i] = offset; | |
2666 | sbi->s_mb_maxs[i] = max; | |
2667 | offset += 1 << (sb->s_blocksize_bits - i); | |
2668 | max = max >> 1; | |
2669 | i++; | |
2670 | } while (i <= sb->s_blocksize_bits + 1); | |
2671 | ||
2672 | /* init file for buddy data */ | |
74767c5a SF |
2673 | ret = ext4_mb_init_backend(sb); |
2674 | if (ret != 0) { | |
c9de560d AT |
2675 | kfree(sbi->s_mb_offsets); |
2676 | kfree(sbi->s_mb_maxs); | |
74767c5a | 2677 | return ret; |
c9de560d AT |
2678 | } |
2679 | ||
2680 | spin_lock_init(&sbi->s_md_lock); | |
c9de560d AT |
2681 | spin_lock_init(&sbi->s_bal_lock); |
2682 | ||
2683 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2684 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2685 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2686 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2687 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
2688 | sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT; | |
2689 | sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC; | |
2690 | ||
730c213c | 2691 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2692 | if (sbi->s_locality_groups == NULL) { |
c9de560d AT |
2693 | kfree(sbi->s_mb_offsets); |
2694 | kfree(sbi->s_mb_maxs); | |
2695 | return -ENOMEM; | |
2696 | } | |
730c213c | 2697 | for_each_possible_cpu(i) { |
c9de560d | 2698 | struct ext4_locality_group *lg; |
730c213c | 2699 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2700 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2701 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2702 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2703 | spin_lock_init(&lg->lg_prealloc_lock); |
2704 | } | |
2705 | ||
c9de560d AT |
2706 | ext4_mb_history_init(sb); |
2707 | ||
0390131b FM |
2708 | if (sbi->s_journal) |
2709 | sbi->s_journal->j_commit_callback = release_blocks_on_commit; | |
3e624fc7 | 2710 | |
4776004f | 2711 | printk(KERN_INFO "EXT4-fs: mballoc enabled\n"); |
c9de560d AT |
2712 | return 0; |
2713 | } | |
2714 | ||
955ce5f5 | 2715 | /* need to called with the ext4 group lock held */ |
c9de560d AT |
2716 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
2717 | { | |
2718 | struct ext4_prealloc_space *pa; | |
2719 | struct list_head *cur, *tmp; | |
2720 | int count = 0; | |
2721 | ||
2722 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2723 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2724 | list_del(&pa->pa_group_list); | |
2725 | count++; | |
688f05a0 | 2726 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2727 | } |
2728 | if (count) | |
6ba495e9 | 2729 | mb_debug(1, "mballoc: %u PAs left\n", count); |
c9de560d AT |
2730 | |
2731 | } | |
2732 | ||
2733 | int ext4_mb_release(struct super_block *sb) | |
2734 | { | |
8df9675f | 2735 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2736 | ext4_group_t i; |
2737 | int num_meta_group_infos; | |
2738 | struct ext4_group_info *grinfo; | |
2739 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2740 | ||
c9de560d | 2741 | if (sbi->s_group_info) { |
8df9675f | 2742 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2743 | grinfo = ext4_get_group_info(sb, i); |
2744 | #ifdef DOUBLE_CHECK | |
2745 | kfree(grinfo->bb_bitmap); | |
2746 | #endif | |
2747 | ext4_lock_group(sb, i); | |
2748 | ext4_mb_cleanup_pa(grinfo); | |
2749 | ext4_unlock_group(sb, i); | |
2750 | kfree(grinfo); | |
2751 | } | |
8df9675f | 2752 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2753 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2754 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2755 | for (i = 0; i < num_meta_group_infos; i++) | |
2756 | kfree(sbi->s_group_info[i]); | |
2757 | kfree(sbi->s_group_info); | |
2758 | } | |
2759 | kfree(sbi->s_mb_offsets); | |
2760 | kfree(sbi->s_mb_maxs); | |
2761 | if (sbi->s_buddy_cache) | |
2762 | iput(sbi->s_buddy_cache); | |
2763 | if (sbi->s_mb_stats) { | |
2764 | printk(KERN_INFO | |
2765 | "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n", | |
2766 | atomic_read(&sbi->s_bal_allocated), | |
2767 | atomic_read(&sbi->s_bal_reqs), | |
2768 | atomic_read(&sbi->s_bal_success)); | |
2769 | printk(KERN_INFO | |
2770 | "EXT4-fs: mballoc: %u extents scanned, %u goal hits, " | |
2771 | "%u 2^N hits, %u breaks, %u lost\n", | |
2772 | atomic_read(&sbi->s_bal_ex_scanned), | |
2773 | atomic_read(&sbi->s_bal_goals), | |
2774 | atomic_read(&sbi->s_bal_2orders), | |
2775 | atomic_read(&sbi->s_bal_breaks), | |
2776 | atomic_read(&sbi->s_mb_lost_chunks)); | |
2777 | printk(KERN_INFO | |
2778 | "EXT4-fs: mballoc: %lu generated and it took %Lu\n", | |
2779 | sbi->s_mb_buddies_generated++, | |
2780 | sbi->s_mb_generation_time); | |
2781 | printk(KERN_INFO | |
2782 | "EXT4-fs: mballoc: %u preallocated, %u discarded\n", | |
2783 | atomic_read(&sbi->s_mb_preallocated), | |
2784 | atomic_read(&sbi->s_mb_discarded)); | |
2785 | } | |
2786 | ||
730c213c | 2787 | free_percpu(sbi->s_locality_groups); |
c9de560d | 2788 | ext4_mb_history_release(sb); |
c9de560d AT |
2789 | |
2790 | return 0; | |
2791 | } | |
2792 | ||
3e624fc7 TT |
2793 | /* |
2794 | * This function is called by the jbd2 layer once the commit has finished, | |
2795 | * so we know we can free the blocks that were released with that commit. | |
2796 | */ | |
2797 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn) | |
c9de560d | 2798 | { |
3e624fc7 | 2799 | struct super_block *sb = journal->j_private; |
c9de560d | 2800 | struct ext4_buddy e4b; |
c894058d | 2801 | struct ext4_group_info *db; |
c894058d AK |
2802 | int err, count = 0, count2 = 0; |
2803 | struct ext4_free_data *entry; | |
8a0aba73 | 2804 | ext4_fsblk_t discard_block; |
3e624fc7 | 2805 | struct list_head *l, *ltmp; |
c9de560d | 2806 | |
3e624fc7 TT |
2807 | list_for_each_safe(l, ltmp, &txn->t_private_list) { |
2808 | entry = list_entry(l, struct ext4_free_data, list); | |
c9de560d | 2809 | |
6ba495e9 | 2810 | mb_debug(1, "gonna free %u blocks in group %u (0x%p):", |
3e624fc7 | 2811 | entry->count, entry->group, entry); |
c9de560d | 2812 | |
c894058d | 2813 | err = ext4_mb_load_buddy(sb, entry->group, &e4b); |
c9de560d AT |
2814 | /* we expect to find existing buddy because it's pinned */ |
2815 | BUG_ON(err != 0); | |
2816 | ||
c894058d | 2817 | db = e4b.bd_info; |
c9de560d | 2818 | /* there are blocks to put in buddy to make them really free */ |
c894058d | 2819 | count += entry->count; |
c9de560d | 2820 | count2++; |
c894058d AK |
2821 | ext4_lock_group(sb, entry->group); |
2822 | /* Take it out of per group rb tree */ | |
2823 | rb_erase(&entry->node, &(db->bb_free_root)); | |
2824 | mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count); | |
2825 | ||
2826 | if (!db->bb_free_root.rb_node) { | |
2827 | /* No more items in the per group rb tree | |
2828 | * balance refcounts from ext4_mb_free_metadata() | |
2829 | */ | |
2830 | page_cache_release(e4b.bd_buddy_page); | |
2831 | page_cache_release(e4b.bd_bitmap_page); | |
c9de560d | 2832 | } |
c894058d | 2833 | ext4_unlock_group(sb, entry->group); |
8a0aba73 TT |
2834 | discard_block = (ext4_fsblk_t) entry->group * EXT4_BLOCKS_PER_GROUP(sb) |
2835 | + entry->start_blk | |
2836 | + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
9bffad1e TT |
2837 | trace_ext4_discard_blocks(sb, (unsigned long long)discard_block, |
2838 | entry->count); | |
8a0aba73 | 2839 | sb_issue_discard(sb, discard_block, entry->count); |
c9de560d | 2840 | |
c894058d | 2841 | kmem_cache_free(ext4_free_ext_cachep, entry); |
c9de560d | 2842 | ext4_mb_release_desc(&e4b); |
3e624fc7 | 2843 | } |
c9de560d | 2844 | |
6ba495e9 | 2845 | mb_debug(1, "freed %u blocks in %u structures\n", count, count2); |
c9de560d AT |
2846 | } |
2847 | ||
6ba495e9 TT |
2848 | #ifdef CONFIG_EXT4_DEBUG |
2849 | u8 mb_enable_debug __read_mostly; | |
2850 | ||
2851 | static struct dentry *debugfs_dir; | |
2852 | static struct dentry *debugfs_debug; | |
2853 | ||
2854 | static void __init ext4_create_debugfs_entry(void) | |
2855 | { | |
2856 | debugfs_dir = debugfs_create_dir("ext4", NULL); | |
2857 | if (debugfs_dir) | |
2858 | debugfs_debug = debugfs_create_u8("mballoc-debug", | |
2859 | S_IRUGO | S_IWUSR, | |
2860 | debugfs_dir, | |
2861 | &mb_enable_debug); | |
2862 | } | |
2863 | ||
2864 | static void ext4_remove_debugfs_entry(void) | |
2865 | { | |
2866 | debugfs_remove(debugfs_debug); | |
2867 | debugfs_remove(debugfs_dir); | |
2868 | } | |
2869 | ||
2870 | #else | |
2871 | ||
2872 | static void __init ext4_create_debugfs_entry(void) | |
2873 | { | |
2874 | } | |
2875 | ||
2876 | static void ext4_remove_debugfs_entry(void) | |
2877 | { | |
2878 | } | |
2879 | ||
2880 | #endif | |
2881 | ||
c9de560d AT |
2882 | int __init init_ext4_mballoc(void) |
2883 | { | |
2884 | ext4_pspace_cachep = | |
2885 | kmem_cache_create("ext4_prealloc_space", | |
2886 | sizeof(struct ext4_prealloc_space), | |
2887 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2888 | if (ext4_pspace_cachep == NULL) | |
2889 | return -ENOMEM; | |
2890 | ||
256bdb49 ES |
2891 | ext4_ac_cachep = |
2892 | kmem_cache_create("ext4_alloc_context", | |
2893 | sizeof(struct ext4_allocation_context), | |
2894 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2895 | if (ext4_ac_cachep == NULL) { | |
2896 | kmem_cache_destroy(ext4_pspace_cachep); | |
2897 | return -ENOMEM; | |
2898 | } | |
c894058d AK |
2899 | |
2900 | ext4_free_ext_cachep = | |
2901 | kmem_cache_create("ext4_free_block_extents", | |
2902 | sizeof(struct ext4_free_data), | |
2903 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2904 | if (ext4_free_ext_cachep == NULL) { | |
2905 | kmem_cache_destroy(ext4_pspace_cachep); | |
2906 | kmem_cache_destroy(ext4_ac_cachep); | |
2907 | return -ENOMEM; | |
2908 | } | |
6ba495e9 | 2909 | ext4_create_debugfs_entry(); |
c9de560d AT |
2910 | return 0; |
2911 | } | |
2912 | ||
2913 | void exit_ext4_mballoc(void) | |
2914 | { | |
3e03f9ca JDB |
2915 | /* |
2916 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep | |
2917 | * before destroying the slab cache. | |
2918 | */ | |
2919 | rcu_barrier(); | |
c9de560d | 2920 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 2921 | kmem_cache_destroy(ext4_ac_cachep); |
c894058d | 2922 | kmem_cache_destroy(ext4_free_ext_cachep); |
6ba495e9 | 2923 | ext4_remove_debugfs_entry(); |
c9de560d AT |
2924 | } |
2925 | ||
2926 | ||
2927 | /* | |
2928 | * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps | |
2929 | * Returns 0 if success or error code | |
2930 | */ | |
4ddfef7b ES |
2931 | static noinline_for_stack int |
2932 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
498e5f24 | 2933 | handle_t *handle, unsigned int reserv_blks) |
c9de560d AT |
2934 | { |
2935 | struct buffer_head *bitmap_bh = NULL; | |
2936 | struct ext4_super_block *es; | |
2937 | struct ext4_group_desc *gdp; | |
2938 | struct buffer_head *gdp_bh; | |
2939 | struct ext4_sb_info *sbi; | |
2940 | struct super_block *sb; | |
2941 | ext4_fsblk_t block; | |
519deca0 | 2942 | int err, len; |
c9de560d AT |
2943 | |
2944 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2945 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2946 | ||
2947 | sb = ac->ac_sb; | |
2948 | sbi = EXT4_SB(sb); | |
2949 | es = sbi->s_es; | |
2950 | ||
c9de560d AT |
2951 | |
2952 | err = -EIO; | |
574ca174 | 2953 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
c9de560d AT |
2954 | if (!bitmap_bh) |
2955 | goto out_err; | |
2956 | ||
2957 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2958 | if (err) | |
2959 | goto out_err; | |
2960 | ||
2961 | err = -EIO; | |
2962 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
2963 | if (!gdp) | |
2964 | goto out_err; | |
2965 | ||
a9df9a49 | 2966 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
9fd9784c | 2967 | ext4_free_blks_count(sb, gdp)); |
03cddb80 | 2968 | |
c9de560d AT |
2969 | err = ext4_journal_get_write_access(handle, gdp_bh); |
2970 | if (err) | |
2971 | goto out_err; | |
2972 | ||
2973 | block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb) | |
2974 | + ac->ac_b_ex.fe_start | |
2975 | + le32_to_cpu(es->s_first_data_block); | |
2976 | ||
519deca0 | 2977 | len = ac->ac_b_ex.fe_len; |
6fd058f7 | 2978 | if (!ext4_data_block_valid(sbi, block, len)) { |
46e665e9 | 2979 | ext4_error(sb, __func__, |
6fd058f7 TT |
2980 | "Allocating blocks %llu-%llu which overlap " |
2981 | "fs metadata\n", block, block+len); | |
519deca0 AK |
2982 | /* File system mounted not to panic on error |
2983 | * Fix the bitmap and repeat the block allocation | |
2984 | * We leak some of the blocks here. | |
2985 | */ | |
955ce5f5 AK |
2986 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
2987 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, | |
2988 | ac->ac_b_ex.fe_len); | |
2989 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
0390131b | 2990 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 AK |
2991 | if (!err) |
2992 | err = -EAGAIN; | |
2993 | goto out_err; | |
c9de560d | 2994 | } |
955ce5f5 AK |
2995 | |
2996 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
2997 | #ifdef AGGRESSIVE_CHECK |
2998 | { | |
2999 | int i; | |
3000 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
3001 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
3002 | bitmap_bh->b_data)); | |
3003 | } | |
3004 | } | |
3005 | #endif | |
955ce5f5 | 3006 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len); |
c9de560d AT |
3007 | if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
3008 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
560671a0 AK |
3009 | ext4_free_blks_set(sb, gdp, |
3010 | ext4_free_blocks_after_init(sb, | |
3011 | ac->ac_b_ex.fe_group, gdp)); | |
c9de560d | 3012 | } |
560671a0 AK |
3013 | len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len; |
3014 | ext4_free_blks_set(sb, gdp, len); | |
c9de560d | 3015 | gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
3016 | |
3017 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
6bc6e63f | 3018 | percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 3019 | /* |
6bc6e63f | 3020 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 3021 | */ |
6bc6e63f AK |
3022 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
3023 | /* release all the reserved blocks if non delalloc */ | |
3024 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks); | |
60e58e0f | 3025 | else { |
6bc6e63f AK |
3026 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, |
3027 | ac->ac_b_ex.fe_len); | |
60e58e0f MC |
3028 | /* convert reserved quota blocks to real quota blocks */ |
3029 | vfs_dq_claim_block(ac->ac_inode, ac->ac_b_ex.fe_len); | |
3030 | } | |
c9de560d | 3031 | |
772cb7c8 JS |
3032 | if (sbi->s_log_groups_per_flex) { |
3033 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
3034 | ac->ac_b_ex.fe_group); | |
9f24e420 TT |
3035 | atomic_sub(ac->ac_b_ex.fe_len, |
3036 | &sbi->s_flex_groups[flex_group].free_blocks); | |
772cb7c8 JS |
3037 | } |
3038 | ||
0390131b | 3039 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
3040 | if (err) |
3041 | goto out_err; | |
0390131b | 3042 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
3043 | |
3044 | out_err: | |
3045 | sb->s_dirt = 1; | |
42a10add | 3046 | brelse(bitmap_bh); |
c9de560d AT |
3047 | return err; |
3048 | } | |
3049 | ||
3050 | /* | |
3051 | * here we normalize request for locality group | |
3052 | * Group request are normalized to s_strip size if we set the same via mount | |
3053 | * option. If not we set it to s_mb_group_prealloc which can be configured via | |
b713a5ec | 3054 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
3055 | * |
3056 | * XXX: should we try to preallocate more than the group has now? | |
3057 | */ | |
3058 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
3059 | { | |
3060 | struct super_block *sb = ac->ac_sb; | |
3061 | struct ext4_locality_group *lg = ac->ac_lg; | |
3062 | ||
3063 | BUG_ON(lg == NULL); | |
3064 | if (EXT4_SB(sb)->s_stripe) | |
3065 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe; | |
3066 | else | |
3067 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; | |
6ba495e9 | 3068 | mb_debug(1, "#%u: goal %u blocks for locality group\n", |
c9de560d AT |
3069 | current->pid, ac->ac_g_ex.fe_len); |
3070 | } | |
3071 | ||
3072 | /* | |
3073 | * Normalization means making request better in terms of | |
3074 | * size and alignment | |
3075 | */ | |
4ddfef7b ES |
3076 | static noinline_for_stack void |
3077 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
3078 | struct ext4_allocation_request *ar) |
3079 | { | |
3080 | int bsbits, max; | |
3081 | ext4_lblk_t end; | |
c9de560d AT |
3082 | loff_t size, orig_size, start_off; |
3083 | ext4_lblk_t start, orig_start; | |
3084 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); | |
9a0762c5 | 3085 | struct ext4_prealloc_space *pa; |
c9de560d AT |
3086 | |
3087 | /* do normalize only data requests, metadata requests | |
3088 | do not need preallocation */ | |
3089 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3090 | return; | |
3091 | ||
3092 | /* sometime caller may want exact blocks */ | |
3093 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
3094 | return; | |
3095 | ||
3096 | /* caller may indicate that preallocation isn't | |
3097 | * required (it's a tail, for example) */ | |
3098 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
3099 | return; | |
3100 | ||
3101 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
3102 | ext4_mb_normalize_group_request(ac); | |
3103 | return ; | |
3104 | } | |
3105 | ||
3106 | bsbits = ac->ac_sb->s_blocksize_bits; | |
3107 | ||
3108 | /* first, let's learn actual file size | |
3109 | * given current request is allocated */ | |
3110 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; | |
3111 | size = size << bsbits; | |
3112 | if (size < i_size_read(ac->ac_inode)) | |
3113 | size = i_size_read(ac->ac_inode); | |
3114 | ||
1930479c VC |
3115 | /* max size of free chunks */ |
3116 | max = 2 << bsbits; | |
c9de560d | 3117 | |
1930479c VC |
3118 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
3119 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
3120 | |
3121 | /* first, try to predict filesize */ | |
3122 | /* XXX: should this table be tunable? */ | |
3123 | start_off = 0; | |
3124 | if (size <= 16 * 1024) { | |
3125 | size = 16 * 1024; | |
3126 | } else if (size <= 32 * 1024) { | |
3127 | size = 32 * 1024; | |
3128 | } else if (size <= 64 * 1024) { | |
3129 | size = 64 * 1024; | |
3130 | } else if (size <= 128 * 1024) { | |
3131 | size = 128 * 1024; | |
3132 | } else if (size <= 256 * 1024) { | |
3133 | size = 256 * 1024; | |
3134 | } else if (size <= 512 * 1024) { | |
3135 | size = 512 * 1024; | |
3136 | } else if (size <= 1024 * 1024) { | |
3137 | size = 1024 * 1024; | |
1930479c | 3138 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 3139 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
3140 | (21 - bsbits)) << 21; |
3141 | size = 2 * 1024 * 1024; | |
3142 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
3143 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3144 | (22 - bsbits)) << 22; | |
3145 | size = 4 * 1024 * 1024; | |
3146 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 3147 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
3148 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3149 | (23 - bsbits)) << 23; | |
3150 | size = 8 * 1024 * 1024; | |
3151 | } else { | |
3152 | start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; | |
3153 | size = ac->ac_o_ex.fe_len << bsbits; | |
3154 | } | |
3155 | orig_size = size = size >> bsbits; | |
3156 | orig_start = start = start_off >> bsbits; | |
3157 | ||
3158 | /* don't cover already allocated blocks in selected range */ | |
3159 | if (ar->pleft && start <= ar->lleft) { | |
3160 | size -= ar->lleft + 1 - start; | |
3161 | start = ar->lleft + 1; | |
3162 | } | |
3163 | if (ar->pright && start + size - 1 >= ar->lright) | |
3164 | size -= start + size - ar->lright; | |
3165 | ||
3166 | end = start + size; | |
3167 | ||
3168 | /* check we don't cross already preallocated blocks */ | |
3169 | rcu_read_lock(); | |
9a0762c5 | 3170 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3171 | ext4_lblk_t pa_end; |
c9de560d | 3172 | |
c9de560d AT |
3173 | if (pa->pa_deleted) |
3174 | continue; | |
3175 | spin_lock(&pa->pa_lock); | |
3176 | if (pa->pa_deleted) { | |
3177 | spin_unlock(&pa->pa_lock); | |
3178 | continue; | |
3179 | } | |
3180 | ||
3181 | pa_end = pa->pa_lstart + pa->pa_len; | |
3182 | ||
3183 | /* PA must not overlap original request */ | |
3184 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
3185 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
3186 | ||
3187 | /* skip PA normalized request doesn't overlap with */ | |
3188 | if (pa->pa_lstart >= end) { | |
3189 | spin_unlock(&pa->pa_lock); | |
3190 | continue; | |
3191 | } | |
3192 | if (pa_end <= start) { | |
3193 | spin_unlock(&pa->pa_lock); | |
3194 | continue; | |
3195 | } | |
3196 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3197 | ||
3198 | if (pa_end <= ac->ac_o_ex.fe_logical) { | |
3199 | BUG_ON(pa_end < start); | |
3200 | start = pa_end; | |
3201 | } | |
3202 | ||
3203 | if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { | |
3204 | BUG_ON(pa->pa_lstart > end); | |
3205 | end = pa->pa_lstart; | |
3206 | } | |
3207 | spin_unlock(&pa->pa_lock); | |
3208 | } | |
3209 | rcu_read_unlock(); | |
3210 | size = end - start; | |
3211 | ||
3212 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3213 | rcu_read_lock(); | |
9a0762c5 | 3214 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3215 | ext4_lblk_t pa_end; |
c9de560d AT |
3216 | spin_lock(&pa->pa_lock); |
3217 | if (pa->pa_deleted == 0) { | |
3218 | pa_end = pa->pa_lstart + pa->pa_len; | |
3219 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); | |
3220 | } | |
3221 | spin_unlock(&pa->pa_lock); | |
3222 | } | |
3223 | rcu_read_unlock(); | |
3224 | ||
3225 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3226 | start > ac->ac_o_ex.fe_logical) { | |
3227 | printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n", | |
3228 | (unsigned long) start, (unsigned long) size, | |
3229 | (unsigned long) ac->ac_o_ex.fe_logical); | |
3230 | } | |
3231 | BUG_ON(start + size <= ac->ac_o_ex.fe_logical && | |
3232 | start > ac->ac_o_ex.fe_logical); | |
8d03c7a0 | 3233 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
3234 | |
3235 | /* now prepare goal request */ | |
3236 | ||
3237 | /* XXX: is it better to align blocks WRT to logical | |
3238 | * placement or satisfy big request as is */ | |
3239 | ac->ac_g_ex.fe_logical = start; | |
3240 | ac->ac_g_ex.fe_len = size; | |
3241 | ||
3242 | /* define goal start in order to merge */ | |
3243 | if (ar->pright && (ar->lright == (start + size))) { | |
3244 | /* merge to the right */ | |
3245 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3246 | &ac->ac_f_ex.fe_group, | |
3247 | &ac->ac_f_ex.fe_start); | |
3248 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3249 | } | |
3250 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3251 | /* merge to the left */ | |
3252 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3253 | &ac->ac_f_ex.fe_group, | |
3254 | &ac->ac_f_ex.fe_start); | |
3255 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3256 | } | |
3257 | ||
6ba495e9 | 3258 | mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, |
c9de560d AT |
3259 | (unsigned) orig_size, (unsigned) start); |
3260 | } | |
3261 | ||
3262 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3263 | { | |
3264 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3265 | ||
3266 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3267 | atomic_inc(&sbi->s_bal_reqs); | |
3268 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
3269 | if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len) | |
3270 | atomic_inc(&sbi->s_bal_success); | |
3271 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3272 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3273 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3274 | atomic_inc(&sbi->s_bal_goals); | |
3275 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3276 | atomic_inc(&sbi->s_bal_breaks); | |
3277 | } | |
3278 | ||
3279 | ext4_mb_store_history(ac); | |
3280 | } | |
3281 | ||
3282 | /* | |
3283 | * use blocks preallocated to inode | |
3284 | */ | |
3285 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3286 | struct ext4_prealloc_space *pa) | |
3287 | { | |
3288 | ext4_fsblk_t start; | |
3289 | ext4_fsblk_t end; | |
3290 | int len; | |
3291 | ||
3292 | /* found preallocated blocks, use them */ | |
3293 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
3294 | end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len); | |
3295 | len = end - start; | |
3296 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, | |
3297 | &ac->ac_b_ex.fe_start); | |
3298 | ac->ac_b_ex.fe_len = len; | |
3299 | ac->ac_status = AC_STATUS_FOUND; | |
3300 | ac->ac_pa = pa; | |
3301 | ||
3302 | BUG_ON(start < pa->pa_pstart); | |
3303 | BUG_ON(start + len > pa->pa_pstart + pa->pa_len); | |
3304 | BUG_ON(pa->pa_free < len); | |
3305 | pa->pa_free -= len; | |
3306 | ||
6ba495e9 | 3307 | mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); |
c9de560d AT |
3308 | } |
3309 | ||
3310 | /* | |
3311 | * use blocks preallocated to locality group | |
3312 | */ | |
3313 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3314 | struct ext4_prealloc_space *pa) | |
3315 | { | |
03cddb80 | 3316 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3317 | |
c9de560d AT |
3318 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3319 | &ac->ac_b_ex.fe_group, | |
3320 | &ac->ac_b_ex.fe_start); | |
3321 | ac->ac_b_ex.fe_len = len; | |
3322 | ac->ac_status = AC_STATUS_FOUND; | |
3323 | ac->ac_pa = pa; | |
3324 | ||
3325 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3326 | * possible race when the group is being loaded concurrently |
c9de560d | 3327 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3328 | * in on-disk bitmap -- see ext4_mb_release_context() |
3329 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 3330 | */ |
6ba495e9 | 3331 | mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); |
c9de560d AT |
3332 | } |
3333 | ||
5e745b04 AK |
3334 | /* |
3335 | * Return the prealloc space that have minimal distance | |
3336 | * from the goal block. @cpa is the prealloc | |
3337 | * space that is having currently known minimal distance | |
3338 | * from the goal block. | |
3339 | */ | |
3340 | static struct ext4_prealloc_space * | |
3341 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3342 | struct ext4_prealloc_space *pa, | |
3343 | struct ext4_prealloc_space *cpa) | |
3344 | { | |
3345 | ext4_fsblk_t cur_distance, new_distance; | |
3346 | ||
3347 | if (cpa == NULL) { | |
3348 | atomic_inc(&pa->pa_count); | |
3349 | return pa; | |
3350 | } | |
3351 | cur_distance = abs(goal_block - cpa->pa_pstart); | |
3352 | new_distance = abs(goal_block - pa->pa_pstart); | |
3353 | ||
3354 | if (cur_distance < new_distance) | |
3355 | return cpa; | |
3356 | ||
3357 | /* drop the previous reference */ | |
3358 | atomic_dec(&cpa->pa_count); | |
3359 | atomic_inc(&pa->pa_count); | |
3360 | return pa; | |
3361 | } | |
3362 | ||
c9de560d AT |
3363 | /* |
3364 | * search goal blocks in preallocated space | |
3365 | */ | |
4ddfef7b ES |
3366 | static noinline_for_stack int |
3367 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
c9de560d | 3368 | { |
6be2ded1 | 3369 | int order, i; |
c9de560d AT |
3370 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3371 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3372 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3373 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3374 | |
3375 | /* only data can be preallocated */ | |
3376 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3377 | return 0; | |
3378 | ||
3379 | /* first, try per-file preallocation */ | |
3380 | rcu_read_lock(); | |
9a0762c5 | 3381 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3382 | |
3383 | /* all fields in this condition don't change, | |
3384 | * so we can skip locking for them */ | |
3385 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
3386 | ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len) | |
3387 | continue; | |
3388 | ||
3389 | /* found preallocated blocks, use them */ | |
3390 | spin_lock(&pa->pa_lock); | |
3391 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3392 | atomic_inc(&pa->pa_count); | |
3393 | ext4_mb_use_inode_pa(ac, pa); | |
3394 | spin_unlock(&pa->pa_lock); | |
3395 | ac->ac_criteria = 10; | |
3396 | rcu_read_unlock(); | |
3397 | return 1; | |
3398 | } | |
3399 | spin_unlock(&pa->pa_lock); | |
3400 | } | |
3401 | rcu_read_unlock(); | |
3402 | ||
3403 | /* can we use group allocation? */ | |
3404 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3405 | return 0; | |
3406 | ||
3407 | /* inode may have no locality group for some reason */ | |
3408 | lg = ac->ac_lg; | |
3409 | if (lg == NULL) | |
3410 | return 0; | |
6be2ded1 AK |
3411 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3412 | if (order > PREALLOC_TB_SIZE - 1) | |
3413 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3414 | order = PREALLOC_TB_SIZE - 1; | |
3415 | ||
5e745b04 AK |
3416 | goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) + |
3417 | ac->ac_g_ex.fe_start + | |
3418 | le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block); | |
3419 | /* | |
3420 | * search for the prealloc space that is having | |
3421 | * minimal distance from the goal block. | |
3422 | */ | |
6be2ded1 AK |
3423 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3424 | rcu_read_lock(); | |
3425 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3426 | pa_inode_list) { | |
3427 | spin_lock(&pa->pa_lock); | |
3428 | if (pa->pa_deleted == 0 && | |
3429 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3430 | |
3431 | cpa = ext4_mb_check_group_pa(goal_block, | |
3432 | pa, cpa); | |
6be2ded1 | 3433 | } |
c9de560d | 3434 | spin_unlock(&pa->pa_lock); |
c9de560d | 3435 | } |
6be2ded1 | 3436 | rcu_read_unlock(); |
c9de560d | 3437 | } |
5e745b04 AK |
3438 | if (cpa) { |
3439 | ext4_mb_use_group_pa(ac, cpa); | |
3440 | ac->ac_criteria = 20; | |
3441 | return 1; | |
3442 | } | |
c9de560d AT |
3443 | return 0; |
3444 | } | |
3445 | ||
7a2fcbf7 AK |
3446 | /* |
3447 | * the function goes through all block freed in the group | |
3448 | * but not yet committed and marks them used in in-core bitmap. | |
3449 | * buddy must be generated from this bitmap | |
955ce5f5 | 3450 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3451 | */ |
3452 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3453 | ext4_group_t group) | |
3454 | { | |
3455 | struct rb_node *n; | |
3456 | struct ext4_group_info *grp; | |
3457 | struct ext4_free_data *entry; | |
3458 | ||
3459 | grp = ext4_get_group_info(sb, group); | |
3460 | n = rb_first(&(grp->bb_free_root)); | |
3461 | ||
3462 | while (n) { | |
3463 | entry = rb_entry(n, struct ext4_free_data, node); | |
955ce5f5 | 3464 | mb_set_bits(bitmap, entry->start_blk, entry->count); |
7a2fcbf7 AK |
3465 | n = rb_next(n); |
3466 | } | |
3467 | return; | |
3468 | } | |
3469 | ||
c9de560d AT |
3470 | /* |
3471 | * the function goes through all preallocation in this group and marks them | |
3472 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3473 | * Need to be called with ext4 group lock held |
c9de560d | 3474 | */ |
089ceecc ES |
3475 | static noinline_for_stack |
3476 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
3477 | ext4_group_t group) |
3478 | { | |
3479 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3480 | struct ext4_prealloc_space *pa; | |
3481 | struct list_head *cur; | |
3482 | ext4_group_t groupnr; | |
3483 | ext4_grpblk_t start; | |
3484 | int preallocated = 0; | |
3485 | int count = 0; | |
3486 | int len; | |
3487 | ||
3488 | /* all form of preallocation discards first load group, | |
3489 | * so the only competing code is preallocation use. | |
3490 | * we don't need any locking here | |
3491 | * notice we do NOT ignore preallocations with pa_deleted | |
3492 | * otherwise we could leave used blocks available for | |
3493 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3494 | * is dropping preallocation | |
3495 | */ | |
3496 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3497 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3498 | spin_lock(&pa->pa_lock); | |
3499 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3500 | &groupnr, &start); | |
3501 | len = pa->pa_len; | |
3502 | spin_unlock(&pa->pa_lock); | |
3503 | if (unlikely(len == 0)) | |
3504 | continue; | |
3505 | BUG_ON(groupnr != group); | |
955ce5f5 | 3506 | mb_set_bits(bitmap, start, len); |
c9de560d AT |
3507 | preallocated += len; |
3508 | count++; | |
3509 | } | |
6ba495e9 | 3510 | mb_debug(1, "prellocated %u for group %u\n", preallocated, group); |
c9de560d AT |
3511 | } |
3512 | ||
3513 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3514 | { | |
3515 | struct ext4_prealloc_space *pa; | |
3516 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
3517 | kmem_cache_free(ext4_pspace_cachep, pa); | |
3518 | } | |
3519 | ||
3520 | /* | |
3521 | * drops a reference to preallocated space descriptor | |
3522 | * if this was the last reference and the space is consumed | |
3523 | */ | |
3524 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3525 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3526 | { | |
a9df9a49 | 3527 | ext4_group_t grp; |
d33a1976 | 3528 | ext4_fsblk_t grp_blk; |
c9de560d AT |
3529 | |
3530 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) | |
3531 | return; | |
3532 | ||
3533 | /* in this short window concurrent discard can set pa_deleted */ | |
3534 | spin_lock(&pa->pa_lock); | |
3535 | if (pa->pa_deleted == 1) { | |
3536 | spin_unlock(&pa->pa_lock); | |
3537 | return; | |
3538 | } | |
3539 | ||
3540 | pa->pa_deleted = 1; | |
3541 | spin_unlock(&pa->pa_lock); | |
3542 | ||
d33a1976 | 3543 | grp_blk = pa->pa_pstart; |
cc0fb9ad AK |
3544 | /* |
3545 | * If doing group-based preallocation, pa_pstart may be in the | |
3546 | * next group when pa is used up | |
3547 | */ | |
3548 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3549 | grp_blk--; |
3550 | ||
3551 | ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL); | |
c9de560d AT |
3552 | |
3553 | /* | |
3554 | * possible race: | |
3555 | * | |
3556 | * P1 (buddy init) P2 (regular allocation) | |
3557 | * find block B in PA | |
3558 | * copy on-disk bitmap to buddy | |
3559 | * mark B in on-disk bitmap | |
3560 | * drop PA from group | |
3561 | * mark all PAs in buddy | |
3562 | * | |
3563 | * thus, P1 initializes buddy with B available. to prevent this | |
3564 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3565 | * against that pair | |
3566 | */ | |
3567 | ext4_lock_group(sb, grp); | |
3568 | list_del(&pa->pa_group_list); | |
3569 | ext4_unlock_group(sb, grp); | |
3570 | ||
3571 | spin_lock(pa->pa_obj_lock); | |
3572 | list_del_rcu(&pa->pa_inode_list); | |
3573 | spin_unlock(pa->pa_obj_lock); | |
3574 | ||
3575 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3576 | } | |
3577 | ||
3578 | /* | |
3579 | * creates new preallocated space for given inode | |
3580 | */ | |
4ddfef7b ES |
3581 | static noinline_for_stack int |
3582 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3583 | { |
3584 | struct super_block *sb = ac->ac_sb; | |
3585 | struct ext4_prealloc_space *pa; | |
3586 | struct ext4_group_info *grp; | |
3587 | struct ext4_inode_info *ei; | |
3588 | ||
3589 | /* preallocate only when found space is larger then requested */ | |
3590 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3591 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3592 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3593 | ||
3594 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3595 | if (pa == NULL) | |
3596 | return -ENOMEM; | |
3597 | ||
3598 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3599 | int winl; | |
3600 | int wins; | |
3601 | int win; | |
3602 | int offs; | |
3603 | ||
3604 | /* we can't allocate as much as normalizer wants. | |
3605 | * so, found space must get proper lstart | |
3606 | * to cover original request */ | |
3607 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3608 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3609 | ||
3610 | /* we're limited by original request in that | |
3611 | * logical block must be covered any way | |
3612 | * winl is window we can move our chunk within */ | |
3613 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3614 | ||
3615 | /* also, we should cover whole original request */ | |
3616 | wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len; | |
3617 | ||
3618 | /* the smallest one defines real window */ | |
3619 | win = min(winl, wins); | |
3620 | ||
3621 | offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len; | |
3622 | if (offs && offs < win) | |
3623 | win = offs; | |
3624 | ||
3625 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win; | |
3626 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); | |
3627 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3628 | } | |
3629 | ||
3630 | /* preallocation can change ac_b_ex, thus we store actually | |
3631 | * allocated blocks for history */ | |
3632 | ac->ac_f_ex = ac->ac_b_ex; | |
3633 | ||
3634 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3635 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3636 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3637 | pa->pa_free = pa->pa_len; | |
3638 | atomic_set(&pa->pa_count, 1); | |
3639 | spin_lock_init(&pa->pa_lock); | |
d794bf8e AK |
3640 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3641 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3642 | pa->pa_deleted = 0; |
cc0fb9ad | 3643 | pa->pa_type = MB_INODE_PA; |
c9de560d | 3644 | |
6ba495e9 | 3645 | mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, |
c9de560d | 3646 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
9bffad1e | 3647 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3648 | |
3649 | ext4_mb_use_inode_pa(ac, pa); | |
3650 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3651 | ||
3652 | ei = EXT4_I(ac->ac_inode); | |
3653 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3654 | ||
3655 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3656 | pa->pa_inode = ac->ac_inode; | |
3657 | ||
3658 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3659 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3660 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3661 | ||
3662 | spin_lock(pa->pa_obj_lock); | |
3663 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3664 | spin_unlock(pa->pa_obj_lock); | |
3665 | ||
3666 | return 0; | |
3667 | } | |
3668 | ||
3669 | /* | |
3670 | * creates new preallocated space for locality group inodes belongs to | |
3671 | */ | |
4ddfef7b ES |
3672 | static noinline_for_stack int |
3673 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3674 | { |
3675 | struct super_block *sb = ac->ac_sb; | |
3676 | struct ext4_locality_group *lg; | |
3677 | struct ext4_prealloc_space *pa; | |
3678 | struct ext4_group_info *grp; | |
3679 | ||
3680 | /* preallocate only when found space is larger then requested */ | |
3681 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3682 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3683 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3684 | ||
3685 | BUG_ON(ext4_pspace_cachep == NULL); | |
3686 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3687 | if (pa == NULL) | |
3688 | return -ENOMEM; | |
3689 | ||
3690 | /* preallocation can change ac_b_ex, thus we store actually | |
3691 | * allocated blocks for history */ | |
3692 | ac->ac_f_ex = ac->ac_b_ex; | |
3693 | ||
3694 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3695 | pa->pa_lstart = pa->pa_pstart; | |
3696 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3697 | pa->pa_free = pa->pa_len; | |
3698 | atomic_set(&pa->pa_count, 1); | |
3699 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3700 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 3701 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 3702 | pa->pa_deleted = 0; |
cc0fb9ad | 3703 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 3704 | |
6ba495e9 | 3705 | mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, |
9bffad1e TT |
3706 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
3707 | trace_ext4_mb_new_group_pa(ac, pa); | |
c9de560d AT |
3708 | |
3709 | ext4_mb_use_group_pa(ac, pa); | |
3710 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3711 | ||
3712 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3713 | lg = ac->ac_lg; | |
3714 | BUG_ON(lg == NULL); | |
3715 | ||
3716 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3717 | pa->pa_inode = NULL; | |
3718 | ||
3719 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3720 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3721 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3722 | ||
6be2ded1 AK |
3723 | /* |
3724 | * We will later add the new pa to the right bucket | |
3725 | * after updating the pa_free in ext4_mb_release_context | |
3726 | */ | |
c9de560d AT |
3727 | return 0; |
3728 | } | |
3729 | ||
3730 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3731 | { | |
3732 | int err; | |
3733 | ||
3734 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3735 | err = ext4_mb_new_group_pa(ac); | |
3736 | else | |
3737 | err = ext4_mb_new_inode_pa(ac); | |
3738 | return err; | |
3739 | } | |
3740 | ||
3741 | /* | |
3742 | * finds all unused blocks in on-disk bitmap, frees them in | |
3743 | * in-core bitmap and buddy. | |
3744 | * @pa must be unlinked from inode and group lists, so that | |
3745 | * nobody else can find/use it. | |
3746 | * the caller MUST hold group/inode locks. | |
3747 | * TODO: optimize the case when there are no in-core structures yet | |
3748 | */ | |
4ddfef7b ES |
3749 | static noinline_for_stack int |
3750 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
c83617db AK |
3751 | struct ext4_prealloc_space *pa, |
3752 | struct ext4_allocation_context *ac) | |
c9de560d | 3753 | { |
c9de560d AT |
3754 | struct super_block *sb = e4b->bd_sb; |
3755 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
3756 | unsigned int end; |
3757 | unsigned int next; | |
c9de560d AT |
3758 | ext4_group_t group; |
3759 | ext4_grpblk_t bit; | |
ba80b101 | 3760 | unsigned long long grp_blk_start; |
c9de560d AT |
3761 | sector_t start; |
3762 | int err = 0; | |
3763 | int free = 0; | |
3764 | ||
3765 | BUG_ON(pa->pa_deleted == 0); | |
3766 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
ba80b101 | 3767 | grp_blk_start = pa->pa_pstart - bit; |
c9de560d AT |
3768 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
3769 | end = bit + pa->pa_len; | |
3770 | ||
256bdb49 ES |
3771 | if (ac) { |
3772 | ac->ac_sb = sb; | |
3773 | ac->ac_inode = pa->pa_inode; | |
3774 | ac->ac_op = EXT4_MB_HISTORY_DISCARD; | |
3775 | } | |
c9de560d AT |
3776 | |
3777 | while (bit < end) { | |
ffad0a44 | 3778 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3779 | if (bit >= end) |
3780 | break; | |
ffad0a44 | 3781 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3782 | start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit + |
3783 | le32_to_cpu(sbi->s_es->s_first_data_block); | |
6ba495e9 | 3784 | mb_debug(1, " free preallocated %u/%u in group %u\n", |
c9de560d AT |
3785 | (unsigned) start, (unsigned) next - bit, |
3786 | (unsigned) group); | |
3787 | free += next - bit; | |
3788 | ||
256bdb49 ES |
3789 | if (ac) { |
3790 | ac->ac_b_ex.fe_group = group; | |
3791 | ac->ac_b_ex.fe_start = bit; | |
3792 | ac->ac_b_ex.fe_len = next - bit; | |
3793 | ac->ac_b_ex.fe_logical = 0; | |
3794 | ext4_mb_store_history(ac); | |
3795 | } | |
c9de560d | 3796 | |
9bffad1e TT |
3797 | trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit, |
3798 | next - bit); | |
c9de560d AT |
3799 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
3800 | bit = next + 1; | |
3801 | } | |
3802 | if (free != pa->pa_free) { | |
26346ff6 | 3803 | printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n", |
c9de560d AT |
3804 | pa, (unsigned long) pa->pa_lstart, |
3805 | (unsigned long) pa->pa_pstart, | |
3806 | (unsigned long) pa->pa_len); | |
5d1b1b3f AK |
3807 | ext4_grp_locked_error(sb, group, |
3808 | __func__, "free %u, pa_free %u", | |
3809 | free, pa->pa_free); | |
e56eb659 AK |
3810 | /* |
3811 | * pa is already deleted so we use the value obtained | |
3812 | * from the bitmap and continue. | |
3813 | */ | |
c9de560d | 3814 | } |
c9de560d AT |
3815 | atomic_add(free, &sbi->s_mb_discarded); |
3816 | ||
3817 | return err; | |
3818 | } | |
3819 | ||
4ddfef7b ES |
3820 | static noinline_for_stack int |
3821 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
c83617db AK |
3822 | struct ext4_prealloc_space *pa, |
3823 | struct ext4_allocation_context *ac) | |
c9de560d | 3824 | { |
c9de560d AT |
3825 | struct super_block *sb = e4b->bd_sb; |
3826 | ext4_group_t group; | |
3827 | ext4_grpblk_t bit; | |
3828 | ||
256bdb49 ES |
3829 | if (ac) |
3830 | ac->ac_op = EXT4_MB_HISTORY_DISCARD; | |
c9de560d | 3831 | |
9bffad1e | 3832 | trace_ext4_mb_release_group_pa(ac, pa); |
c9de560d AT |
3833 | BUG_ON(pa->pa_deleted == 0); |
3834 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3835 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3836 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3837 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3838 | ||
256bdb49 ES |
3839 | if (ac) { |
3840 | ac->ac_sb = sb; | |
3841 | ac->ac_inode = NULL; | |
3842 | ac->ac_b_ex.fe_group = group; | |
3843 | ac->ac_b_ex.fe_start = bit; | |
3844 | ac->ac_b_ex.fe_len = pa->pa_len; | |
3845 | ac->ac_b_ex.fe_logical = 0; | |
3846 | ext4_mb_store_history(ac); | |
256bdb49 | 3847 | } |
c9de560d AT |
3848 | |
3849 | return 0; | |
3850 | } | |
3851 | ||
3852 | /* | |
3853 | * releases all preallocations in given group | |
3854 | * | |
3855 | * first, we need to decide discard policy: | |
3856 | * - when do we discard | |
3857 | * 1) ENOSPC | |
3858 | * - how many do we discard | |
3859 | * 1) how many requested | |
3860 | */ | |
4ddfef7b ES |
3861 | static noinline_for_stack int |
3862 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3863 | ext4_group_t group, int needed) |
3864 | { | |
3865 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3866 | struct buffer_head *bitmap_bh = NULL; | |
3867 | struct ext4_prealloc_space *pa, *tmp; | |
c83617db | 3868 | struct ext4_allocation_context *ac; |
c9de560d AT |
3869 | struct list_head list; |
3870 | struct ext4_buddy e4b; | |
3871 | int err; | |
3872 | int busy = 0; | |
3873 | int free = 0; | |
3874 | ||
6ba495e9 | 3875 | mb_debug(1, "discard preallocation for group %u\n", group); |
c9de560d AT |
3876 | |
3877 | if (list_empty(&grp->bb_prealloc_list)) | |
3878 | return 0; | |
3879 | ||
574ca174 | 3880 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3881 | if (bitmap_bh == NULL) { |
ce89f46c | 3882 | ext4_error(sb, __func__, "Error in reading block " |
a9df9a49 | 3883 | "bitmap for %u", group); |
ce89f46c | 3884 | return 0; |
c9de560d AT |
3885 | } |
3886 | ||
3887 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c AK |
3888 | if (err) { |
3889 | ext4_error(sb, __func__, "Error in loading buddy " | |
a9df9a49 | 3890 | "information for %u", group); |
ce89f46c AK |
3891 | put_bh(bitmap_bh); |
3892 | return 0; | |
3893 | } | |
c9de560d AT |
3894 | |
3895 | if (needed == 0) | |
3896 | needed = EXT4_BLOCKS_PER_GROUP(sb) + 1; | |
3897 | ||
c9de560d | 3898 | INIT_LIST_HEAD(&list); |
c83617db | 3899 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
9bffad1e TT |
3900 | if (ac) |
3901 | ac->ac_sb = sb; | |
c9de560d AT |
3902 | repeat: |
3903 | ext4_lock_group(sb, group); | |
3904 | list_for_each_entry_safe(pa, tmp, | |
3905 | &grp->bb_prealloc_list, pa_group_list) { | |
3906 | spin_lock(&pa->pa_lock); | |
3907 | if (atomic_read(&pa->pa_count)) { | |
3908 | spin_unlock(&pa->pa_lock); | |
3909 | busy = 1; | |
3910 | continue; | |
3911 | } | |
3912 | if (pa->pa_deleted) { | |
3913 | spin_unlock(&pa->pa_lock); | |
3914 | continue; | |
3915 | } | |
3916 | ||
3917 | /* seems this one can be freed ... */ | |
3918 | pa->pa_deleted = 1; | |
3919 | ||
3920 | /* we can trust pa_free ... */ | |
3921 | free += pa->pa_free; | |
3922 | ||
3923 | spin_unlock(&pa->pa_lock); | |
3924 | ||
3925 | list_del(&pa->pa_group_list); | |
3926 | list_add(&pa->u.pa_tmp_list, &list); | |
3927 | } | |
3928 | ||
3929 | /* if we still need more blocks and some PAs were used, try again */ | |
3930 | if (free < needed && busy) { | |
3931 | busy = 0; | |
3932 | ext4_unlock_group(sb, group); | |
3933 | /* | |
3934 | * Yield the CPU here so that we don't get soft lockup | |
3935 | * in non preempt case. | |
3936 | */ | |
3937 | yield(); | |
3938 | goto repeat; | |
3939 | } | |
3940 | ||
3941 | /* found anything to free? */ | |
3942 | if (list_empty(&list)) { | |
3943 | BUG_ON(free != 0); | |
3944 | goto out; | |
3945 | } | |
3946 | ||
3947 | /* now free all selected PAs */ | |
3948 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3949 | ||
3950 | /* remove from object (inode or locality group) */ | |
3951 | spin_lock(pa->pa_obj_lock); | |
3952 | list_del_rcu(&pa->pa_inode_list); | |
3953 | spin_unlock(pa->pa_obj_lock); | |
3954 | ||
cc0fb9ad | 3955 | if (pa->pa_type == MB_GROUP_PA) |
c83617db | 3956 | ext4_mb_release_group_pa(&e4b, pa, ac); |
c9de560d | 3957 | else |
c83617db | 3958 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); |
c9de560d AT |
3959 | |
3960 | list_del(&pa->u.pa_tmp_list); | |
3961 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3962 | } | |
3963 | ||
3964 | out: | |
3965 | ext4_unlock_group(sb, group); | |
c83617db AK |
3966 | if (ac) |
3967 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
3968 | ext4_mb_release_desc(&e4b); |
3969 | put_bh(bitmap_bh); | |
3970 | return free; | |
3971 | } | |
3972 | ||
3973 | /* | |
3974 | * releases all non-used preallocated blocks for given inode | |
3975 | * | |
3976 | * It's important to discard preallocations under i_data_sem | |
3977 | * We don't want another block to be served from the prealloc | |
3978 | * space when we are discarding the inode prealloc space. | |
3979 | * | |
3980 | * FIXME!! Make sure it is valid at all the call sites | |
3981 | */ | |
c2ea3fde | 3982 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
3983 | { |
3984 | struct ext4_inode_info *ei = EXT4_I(inode); | |
3985 | struct super_block *sb = inode->i_sb; | |
3986 | struct buffer_head *bitmap_bh = NULL; | |
3987 | struct ext4_prealloc_space *pa, *tmp; | |
c83617db | 3988 | struct ext4_allocation_context *ac; |
c9de560d AT |
3989 | ext4_group_t group = 0; |
3990 | struct list_head list; | |
3991 | struct ext4_buddy e4b; | |
3992 | int err; | |
3993 | ||
c2ea3fde | 3994 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
3995 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
3996 | return; | |
3997 | } | |
3998 | ||
6ba495e9 | 3999 | mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); |
9bffad1e | 4000 | trace_ext4_discard_preallocations(inode); |
c9de560d AT |
4001 | |
4002 | INIT_LIST_HEAD(&list); | |
4003 | ||
c83617db | 4004 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
9bffad1e TT |
4005 | if (ac) { |
4006 | ac->ac_sb = sb; | |
4007 | ac->ac_inode = inode; | |
4008 | } | |
c9de560d AT |
4009 | repeat: |
4010 | /* first, collect all pa's in the inode */ | |
4011 | spin_lock(&ei->i_prealloc_lock); | |
4012 | while (!list_empty(&ei->i_prealloc_list)) { | |
4013 | pa = list_entry(ei->i_prealloc_list.next, | |
4014 | struct ext4_prealloc_space, pa_inode_list); | |
4015 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
4016 | spin_lock(&pa->pa_lock); | |
4017 | if (atomic_read(&pa->pa_count)) { | |
4018 | /* this shouldn't happen often - nobody should | |
4019 | * use preallocation while we're discarding it */ | |
4020 | spin_unlock(&pa->pa_lock); | |
4021 | spin_unlock(&ei->i_prealloc_lock); | |
4022 | printk(KERN_ERR "uh-oh! used pa while discarding\n"); | |
4023 | WARN_ON(1); | |
4024 | schedule_timeout_uninterruptible(HZ); | |
4025 | goto repeat; | |
4026 | ||
4027 | } | |
4028 | if (pa->pa_deleted == 0) { | |
4029 | pa->pa_deleted = 1; | |
4030 | spin_unlock(&pa->pa_lock); | |
4031 | list_del_rcu(&pa->pa_inode_list); | |
4032 | list_add(&pa->u.pa_tmp_list, &list); | |
4033 | continue; | |
4034 | } | |
4035 | ||
4036 | /* someone is deleting pa right now */ | |
4037 | spin_unlock(&pa->pa_lock); | |
4038 | spin_unlock(&ei->i_prealloc_lock); | |
4039 | ||
4040 | /* we have to wait here because pa_deleted | |
4041 | * doesn't mean pa is already unlinked from | |
4042 | * the list. as we might be called from | |
4043 | * ->clear_inode() the inode will get freed | |
4044 | * and concurrent thread which is unlinking | |
4045 | * pa from inode's list may access already | |
4046 | * freed memory, bad-bad-bad */ | |
4047 | ||
4048 | /* XXX: if this happens too often, we can | |
4049 | * add a flag to force wait only in case | |
4050 | * of ->clear_inode(), but not in case of | |
4051 | * regular truncate */ | |
4052 | schedule_timeout_uninterruptible(HZ); | |
4053 | goto repeat; | |
4054 | } | |
4055 | spin_unlock(&ei->i_prealloc_lock); | |
4056 | ||
4057 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 4058 | BUG_ON(pa->pa_type != MB_INODE_PA); |
c9de560d AT |
4059 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); |
4060 | ||
4061 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c AK |
4062 | if (err) { |
4063 | ext4_error(sb, __func__, "Error in loading buddy " | |
a9df9a49 | 4064 | "information for %u", group); |
ce89f46c AK |
4065 | continue; |
4066 | } | |
c9de560d | 4067 | |
574ca174 | 4068 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 4069 | if (bitmap_bh == NULL) { |
ce89f46c | 4070 | ext4_error(sb, __func__, "Error in reading block " |
a9df9a49 | 4071 | "bitmap for %u", group); |
c9de560d | 4072 | ext4_mb_release_desc(&e4b); |
ce89f46c | 4073 | continue; |
c9de560d AT |
4074 | } |
4075 | ||
4076 | ext4_lock_group(sb, group); | |
4077 | list_del(&pa->pa_group_list); | |
c83617db | 4078 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); |
c9de560d AT |
4079 | ext4_unlock_group(sb, group); |
4080 | ||
4081 | ext4_mb_release_desc(&e4b); | |
4082 | put_bh(bitmap_bh); | |
4083 | ||
4084 | list_del(&pa->u.pa_tmp_list); | |
4085 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4086 | } | |
c83617db AK |
4087 | if (ac) |
4088 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
4089 | } |
4090 | ||
4091 | /* | |
4092 | * finds all preallocated spaces and return blocks being freed to them | |
4093 | * if preallocated space becomes full (no block is used from the space) | |
4094 | * then the function frees space in buddy | |
4095 | * XXX: at the moment, truncate (which is the only way to free blocks) | |
4096 | * discards all preallocations | |
4097 | */ | |
4098 | static void ext4_mb_return_to_preallocation(struct inode *inode, | |
4099 | struct ext4_buddy *e4b, | |
4100 | sector_t block, int count) | |
4101 | { | |
4102 | BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list)); | |
4103 | } | |
6ba495e9 | 4104 | #ifdef CONFIG_EXT4_DEBUG |
c9de560d AT |
4105 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
4106 | { | |
4107 | struct super_block *sb = ac->ac_sb; | |
8df9675f | 4108 | ext4_group_t ngroups, i; |
c9de560d AT |
4109 | |
4110 | printk(KERN_ERR "EXT4-fs: Can't allocate:" | |
4111 | " Allocation context details:\n"); | |
4112 | printk(KERN_ERR "EXT4-fs: status %d flags %d\n", | |
4113 | ac->ac_status, ac->ac_flags); | |
4114 | printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, " | |
4115 | "best %lu/%lu/%lu@%lu cr %d\n", | |
4116 | (unsigned long)ac->ac_o_ex.fe_group, | |
4117 | (unsigned long)ac->ac_o_ex.fe_start, | |
4118 | (unsigned long)ac->ac_o_ex.fe_len, | |
4119 | (unsigned long)ac->ac_o_ex.fe_logical, | |
4120 | (unsigned long)ac->ac_g_ex.fe_group, | |
4121 | (unsigned long)ac->ac_g_ex.fe_start, | |
4122 | (unsigned long)ac->ac_g_ex.fe_len, | |
4123 | (unsigned long)ac->ac_g_ex.fe_logical, | |
4124 | (unsigned long)ac->ac_b_ex.fe_group, | |
4125 | (unsigned long)ac->ac_b_ex.fe_start, | |
4126 | (unsigned long)ac->ac_b_ex.fe_len, | |
4127 | (unsigned long)ac->ac_b_ex.fe_logical, | |
4128 | (int)ac->ac_criteria); | |
4129 | printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned, | |
4130 | ac->ac_found); | |
4131 | printk(KERN_ERR "EXT4-fs: groups: \n"); | |
8df9675f TT |
4132 | ngroups = ext4_get_groups_count(sb); |
4133 | for (i = 0; i < ngroups; i++) { | |
c9de560d AT |
4134 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
4135 | struct ext4_prealloc_space *pa; | |
4136 | ext4_grpblk_t start; | |
4137 | struct list_head *cur; | |
4138 | ext4_lock_group(sb, i); | |
4139 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4140 | pa = list_entry(cur, struct ext4_prealloc_space, | |
4141 | pa_group_list); | |
4142 | spin_lock(&pa->pa_lock); | |
4143 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4144 | NULL, &start); | |
4145 | spin_unlock(&pa->pa_lock); | |
1c718505 AF |
4146 | printk(KERN_ERR "PA:%u:%d:%u \n", i, |
4147 | start, pa->pa_len); | |
c9de560d | 4148 | } |
60bd63d1 | 4149 | ext4_unlock_group(sb, i); |
c9de560d AT |
4150 | |
4151 | if (grp->bb_free == 0) | |
4152 | continue; | |
1c718505 | 4153 | printk(KERN_ERR "%u: %d/%d \n", |
c9de560d AT |
4154 | i, grp->bb_free, grp->bb_fragments); |
4155 | } | |
4156 | printk(KERN_ERR "\n"); | |
4157 | } | |
4158 | #else | |
4159 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
4160 | { | |
4161 | return; | |
4162 | } | |
4163 | #endif | |
4164 | ||
4165 | /* | |
4166 | * We use locality group preallocation for small size file. The size of the | |
4167 | * file is determined by the current size or the resulting size after | |
4168 | * allocation which ever is larger | |
4169 | * | |
b713a5ec | 4170 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
4171 | */ |
4172 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
4173 | { | |
4174 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4175 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
4176 | loff_t size, isize; | |
4177 | ||
4178 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4179 | return; | |
4180 | ||
4ba74d00 TT |
4181 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
4182 | return; | |
4183 | ||
c9de560d | 4184 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; |
50797481 TT |
4185 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
4186 | >> bsbits; | |
c9de560d AT |
4187 | size = max(size, isize); |
4188 | ||
50797481 TT |
4189 | if ((size == isize) && |
4190 | !ext4_fs_is_busy(sbi) && | |
4191 | (atomic_read(&ac->ac_inode->i_writecount) == 0)) { | |
4192 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
4193 | return; | |
4194 | } | |
4195 | ||
c9de560d | 4196 | /* don't use group allocation for large files */ |
4ba74d00 TT |
4197 | if (size >= sbi->s_mb_stream_request) { |
4198 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
c9de560d | 4199 | return; |
4ba74d00 | 4200 | } |
c9de560d AT |
4201 | |
4202 | BUG_ON(ac->ac_lg != NULL); | |
4203 | /* | |
4204 | * locality group prealloc space are per cpu. The reason for having | |
4205 | * per cpu locality group is to reduce the contention between block | |
4206 | * request from multiple CPUs. | |
4207 | */ | |
730c213c | 4208 | ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id()); |
c9de560d AT |
4209 | |
4210 | /* we're going to use group allocation */ | |
4211 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4212 | ||
4213 | /* serialize all allocations in the group */ | |
4214 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4215 | } | |
4216 | ||
4ddfef7b ES |
4217 | static noinline_for_stack int |
4218 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
4219 | struct ext4_allocation_request *ar) |
4220 | { | |
4221 | struct super_block *sb = ar->inode->i_sb; | |
4222 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4223 | struct ext4_super_block *es = sbi->s_es; | |
4224 | ext4_group_t group; | |
498e5f24 TT |
4225 | unsigned int len; |
4226 | ext4_fsblk_t goal; | |
c9de560d AT |
4227 | ext4_grpblk_t block; |
4228 | ||
4229 | /* we can't allocate > group size */ | |
4230 | len = ar->len; | |
4231 | ||
4232 | /* just a dirty hack to filter too big requests */ | |
4233 | if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10) | |
4234 | len = EXT4_BLOCKS_PER_GROUP(sb) - 10; | |
4235 | ||
4236 | /* start searching from the goal */ | |
4237 | goal = ar->goal; | |
4238 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4239 | goal >= ext4_blocks_count(es)) | |
4240 | goal = le32_to_cpu(es->s_first_data_block); | |
4241 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4242 | ||
4243 | /* set up allocation goals */ | |
833576b3 | 4244 | memset(ac, 0, sizeof(struct ext4_allocation_context)); |
c9de560d | 4245 | ac->ac_b_ex.fe_logical = ar->logical; |
c9de560d | 4246 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
4247 | ac->ac_sb = sb; |
4248 | ac->ac_inode = ar->inode; | |
4249 | ac->ac_o_ex.fe_logical = ar->logical; | |
4250 | ac->ac_o_ex.fe_group = group; | |
4251 | ac->ac_o_ex.fe_start = block; | |
4252 | ac->ac_o_ex.fe_len = len; | |
4253 | ac->ac_g_ex.fe_logical = ar->logical; | |
4254 | ac->ac_g_ex.fe_group = group; | |
4255 | ac->ac_g_ex.fe_start = block; | |
4256 | ac->ac_g_ex.fe_len = len; | |
c9de560d | 4257 | ac->ac_flags = ar->flags; |
c9de560d AT |
4258 | |
4259 | /* we have to define context: we'll we work with a file or | |
4260 | * locality group. this is a policy, actually */ | |
4261 | ext4_mb_group_or_file(ac); | |
4262 | ||
6ba495e9 | 4263 | mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " |
c9de560d AT |
4264 | "left: %u/%u, right %u/%u to %swritable\n", |
4265 | (unsigned) ar->len, (unsigned) ar->logical, | |
4266 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4267 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4268 | (unsigned) ar->lright, (unsigned) ar->pright, | |
4269 | atomic_read(&ar->inode->i_writecount) ? "" : "non-"); | |
4270 | return 0; | |
4271 | ||
4272 | } | |
4273 | ||
6be2ded1 AK |
4274 | static noinline_for_stack void |
4275 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4276 | struct ext4_locality_group *lg, | |
4277 | int order, int total_entries) | |
4278 | { | |
4279 | ext4_group_t group = 0; | |
4280 | struct ext4_buddy e4b; | |
4281 | struct list_head discard_list; | |
4282 | struct ext4_prealloc_space *pa, *tmp; | |
4283 | struct ext4_allocation_context *ac; | |
4284 | ||
6ba495e9 | 4285 | mb_debug(1, "discard locality group preallocation\n"); |
6be2ded1 AK |
4286 | |
4287 | INIT_LIST_HEAD(&discard_list); | |
4288 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); | |
9bffad1e TT |
4289 | if (ac) |
4290 | ac->ac_sb = sb; | |
6be2ded1 AK |
4291 | |
4292 | spin_lock(&lg->lg_prealloc_lock); | |
4293 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
4294 | pa_inode_list) { | |
4295 | spin_lock(&pa->pa_lock); | |
4296 | if (atomic_read(&pa->pa_count)) { | |
4297 | /* | |
4298 | * This is the pa that we just used | |
4299 | * for block allocation. So don't | |
4300 | * free that | |
4301 | */ | |
4302 | spin_unlock(&pa->pa_lock); | |
4303 | continue; | |
4304 | } | |
4305 | if (pa->pa_deleted) { | |
4306 | spin_unlock(&pa->pa_lock); | |
4307 | continue; | |
4308 | } | |
4309 | /* only lg prealloc space */ | |
cc0fb9ad | 4310 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4311 | |
4312 | /* seems this one can be freed ... */ | |
4313 | pa->pa_deleted = 1; | |
4314 | spin_unlock(&pa->pa_lock); | |
4315 | ||
4316 | list_del_rcu(&pa->pa_inode_list); | |
4317 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4318 | ||
4319 | total_entries--; | |
4320 | if (total_entries <= 5) { | |
4321 | /* | |
4322 | * we want to keep only 5 entries | |
4323 | * allowing it to grow to 8. This | |
4324 | * mak sure we don't call discard | |
4325 | * soon for this list. | |
4326 | */ | |
4327 | break; | |
4328 | } | |
4329 | } | |
4330 | spin_unlock(&lg->lg_prealloc_lock); | |
4331 | ||
4332 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
4333 | ||
4334 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); | |
4335 | if (ext4_mb_load_buddy(sb, group, &e4b)) { | |
4336 | ext4_error(sb, __func__, "Error in loading buddy " | |
a9df9a49 | 4337 | "information for %u", group); |
6be2ded1 AK |
4338 | continue; |
4339 | } | |
4340 | ext4_lock_group(sb, group); | |
4341 | list_del(&pa->pa_group_list); | |
4342 | ext4_mb_release_group_pa(&e4b, pa, ac); | |
4343 | ext4_unlock_group(sb, group); | |
4344 | ||
4345 | ext4_mb_release_desc(&e4b); | |
4346 | list_del(&pa->u.pa_tmp_list); | |
4347 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4348 | } | |
4349 | if (ac) | |
4350 | kmem_cache_free(ext4_ac_cachep, ac); | |
4351 | } | |
4352 | ||
4353 | /* | |
4354 | * We have incremented pa_count. So it cannot be freed at this | |
4355 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4356 | * possible from this lg. That means pa_free cannot be updated. | |
4357 | * | |
4358 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4359 | * which can cause the lg_prealloc_list to be updated. | |
4360 | */ | |
4361 | ||
4362 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4363 | { | |
4364 | int order, added = 0, lg_prealloc_count = 1; | |
4365 | struct super_block *sb = ac->ac_sb; | |
4366 | struct ext4_locality_group *lg = ac->ac_lg; | |
4367 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4368 | ||
4369 | order = fls(pa->pa_free) - 1; | |
4370 | if (order > PREALLOC_TB_SIZE - 1) | |
4371 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4372 | order = PREALLOC_TB_SIZE - 1; | |
4373 | /* Add the prealloc space to lg */ | |
4374 | rcu_read_lock(); | |
4375 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], | |
4376 | pa_inode_list) { | |
4377 | spin_lock(&tmp_pa->pa_lock); | |
4378 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4379 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4380 | continue; |
4381 | } | |
4382 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4383 | /* Add to the tail of the previous entry */ | |
4384 | list_add_tail_rcu(&pa->pa_inode_list, | |
4385 | &tmp_pa->pa_inode_list); | |
4386 | added = 1; | |
4387 | /* | |
4388 | * we want to count the total | |
4389 | * number of entries in the list | |
4390 | */ | |
4391 | } | |
4392 | spin_unlock(&tmp_pa->pa_lock); | |
4393 | lg_prealloc_count++; | |
4394 | } | |
4395 | if (!added) | |
4396 | list_add_tail_rcu(&pa->pa_inode_list, | |
4397 | &lg->lg_prealloc_list[order]); | |
4398 | rcu_read_unlock(); | |
4399 | ||
4400 | /* Now trim the list to be not more than 8 elements */ | |
4401 | if (lg_prealloc_count > 8) { | |
4402 | ext4_mb_discard_lg_preallocations(sb, lg, | |
4403 | order, lg_prealloc_count); | |
4404 | return; | |
4405 | } | |
4406 | return ; | |
4407 | } | |
4408 | ||
c9de560d AT |
4409 | /* |
4410 | * release all resource we used in allocation | |
4411 | */ | |
4412 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4413 | { | |
6be2ded1 AK |
4414 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4415 | if (pa) { | |
cc0fb9ad | 4416 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4417 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 AK |
4418 | spin_lock(&pa->pa_lock); |
4419 | pa->pa_pstart += ac->ac_b_ex.fe_len; | |
4420 | pa->pa_lstart += ac->ac_b_ex.fe_len; | |
4421 | pa->pa_free -= ac->ac_b_ex.fe_len; | |
4422 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4423 | spin_unlock(&pa->pa_lock); | |
c9de560d | 4424 | } |
c9de560d | 4425 | } |
8556e8f3 AK |
4426 | if (ac->alloc_semp) |
4427 | up_read(ac->alloc_semp); | |
ba443916 AK |
4428 | if (pa) { |
4429 | /* | |
4430 | * We want to add the pa to the right bucket. | |
4431 | * Remove it from the list and while adding | |
4432 | * make sure the list to which we are adding | |
4433 | * doesn't grow big. We need to release | |
4434 | * alloc_semp before calling ext4_mb_add_n_trim() | |
4435 | */ | |
cc0fb9ad | 4436 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { |
ba443916 AK |
4437 | spin_lock(pa->pa_obj_lock); |
4438 | list_del_rcu(&pa->pa_inode_list); | |
4439 | spin_unlock(pa->pa_obj_lock); | |
4440 | ext4_mb_add_n_trim(ac); | |
4441 | } | |
4442 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4443 | } | |
c9de560d AT |
4444 | if (ac->ac_bitmap_page) |
4445 | page_cache_release(ac->ac_bitmap_page); | |
4446 | if (ac->ac_buddy_page) | |
4447 | page_cache_release(ac->ac_buddy_page); | |
4448 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
4449 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4450 | ext4_mb_collect_stats(ac); | |
4451 | return 0; | |
4452 | } | |
4453 | ||
4454 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4455 | { | |
8df9675f | 4456 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4457 | int ret; |
4458 | int freed = 0; | |
4459 | ||
9bffad1e | 4460 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4461 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4462 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4463 | freed += ret; | |
4464 | needed -= ret; | |
4465 | } | |
4466 | ||
4467 | return freed; | |
4468 | } | |
4469 | ||
4470 | /* | |
4471 | * Main entry point into mballoc to allocate blocks | |
4472 | * it tries to use preallocation first, then falls back | |
4473 | * to usual allocation | |
4474 | */ | |
4475 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
4476 | struct ext4_allocation_request *ar, int *errp) | |
4477 | { | |
6bc6e63f | 4478 | int freed; |
256bdb49 | 4479 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4480 | struct ext4_sb_info *sbi; |
4481 | struct super_block *sb; | |
4482 | ext4_fsblk_t block = 0; | |
60e58e0f | 4483 | unsigned int inquota = 0; |
498e5f24 | 4484 | unsigned int reserv_blks = 0; |
c9de560d AT |
4485 | |
4486 | sb = ar->inode->i_sb; | |
4487 | sbi = EXT4_SB(sb); | |
4488 | ||
9bffad1e | 4489 | trace_ext4_request_blocks(ar); |
ba80b101 | 4490 | |
60e58e0f MC |
4491 | /* |
4492 | * For delayed allocation, we could skip the ENOSPC and | |
4493 | * EDQUOT check, as blocks and quotas have been already | |
4494 | * reserved when data being copied into pagecache. | |
4495 | */ | |
4496 | if (EXT4_I(ar->inode)->i_delalloc_reserved_flag) | |
4497 | ar->flags |= EXT4_MB_DELALLOC_RESERVED; | |
4498 | else { | |
4499 | /* Without delayed allocation we need to verify | |
4500 | * there is enough free blocks to do block allocation | |
4501 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4502 | */ |
030ba6bc AK |
4503 | while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) { |
4504 | /* let others to free the space */ | |
4505 | yield(); | |
4506 | ar->len = ar->len >> 1; | |
4507 | } | |
4508 | if (!ar->len) { | |
a30d542a AK |
4509 | *errp = -ENOSPC; |
4510 | return 0; | |
4511 | } | |
6bc6e63f | 4512 | reserv_blks = ar->len; |
a269eb18 | 4513 | while (ar->len && vfs_dq_alloc_block(ar->inode, ar->len)) { |
60e58e0f MC |
4514 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; |
4515 | ar->len--; | |
4516 | } | |
4517 | inquota = ar->len; | |
4518 | if (ar->len == 0) { | |
4519 | *errp = -EDQUOT; | |
4520 | goto out3; | |
4521 | } | |
07031431 | 4522 | } |
d2a17637 | 4523 | |
256bdb49 | 4524 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 4525 | if (!ac) { |
363d4251 | 4526 | ar->len = 0; |
256bdb49 | 4527 | *errp = -ENOMEM; |
363d4251 | 4528 | goto out1; |
256bdb49 ES |
4529 | } |
4530 | ||
256bdb49 | 4531 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4532 | if (*errp) { |
4533 | ar->len = 0; | |
363d4251 | 4534 | goto out2; |
c9de560d AT |
4535 | } |
4536 | ||
256bdb49 ES |
4537 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4538 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4539 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4540 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4541 | repeat: |
4542 | /* allocate space in core */ | |
256bdb49 | 4543 | ext4_mb_regular_allocator(ac); |
c9de560d AT |
4544 | |
4545 | /* as we've just preallocated more space than | |
4546 | * user requested orinally, we store allocated | |
4547 | * space in a special descriptor */ | |
256bdb49 ES |
4548 | if (ac->ac_status == AC_STATUS_FOUND && |
4549 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
4550 | ext4_mb_new_preallocation(ac); | |
c9de560d | 4551 | } |
256bdb49 | 4552 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
6bc6e63f | 4553 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks); |
519deca0 | 4554 | if (*errp == -EAGAIN) { |
8556e8f3 AK |
4555 | /* |
4556 | * drop the reference that we took | |
4557 | * in ext4_mb_use_best_found | |
4558 | */ | |
4559 | ext4_mb_release_context(ac); | |
519deca0 AK |
4560 | ac->ac_b_ex.fe_group = 0; |
4561 | ac->ac_b_ex.fe_start = 0; | |
4562 | ac->ac_b_ex.fe_len = 0; | |
4563 | ac->ac_status = AC_STATUS_CONTINUE; | |
4564 | goto repeat; | |
4565 | } else if (*errp) { | |
4566 | ac->ac_b_ex.fe_len = 0; | |
4567 | ar->len = 0; | |
4568 | ext4_mb_show_ac(ac); | |
4569 | } else { | |
4570 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4571 | ar->len = ac->ac_b_ex.fe_len; | |
4572 | } | |
c9de560d | 4573 | } else { |
256bdb49 | 4574 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4575 | if (freed) |
4576 | goto repeat; | |
4577 | *errp = -ENOSPC; | |
256bdb49 | 4578 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4579 | ar->len = 0; |
256bdb49 | 4580 | ext4_mb_show_ac(ac); |
c9de560d AT |
4581 | } |
4582 | ||
256bdb49 | 4583 | ext4_mb_release_context(ac); |
c9de560d | 4584 | |
363d4251 SF |
4585 | out2: |
4586 | kmem_cache_free(ext4_ac_cachep, ac); | |
4587 | out1: | |
60e58e0f | 4588 | if (inquota && ar->len < inquota) |
a269eb18 | 4589 | vfs_dq_free_block(ar->inode, inquota - ar->len); |
0087d9fb AK |
4590 | out3: |
4591 | if (!ar->len) { | |
4592 | if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) | |
4593 | /* release all the reserved blocks if non delalloc */ | |
4594 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, | |
4595 | reserv_blks); | |
4596 | } | |
c9de560d | 4597 | |
9bffad1e | 4598 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4599 | |
c9de560d AT |
4600 | return block; |
4601 | } | |
c9de560d | 4602 | |
c894058d AK |
4603 | /* |
4604 | * We can merge two free data extents only if the physical blocks | |
4605 | * are contiguous, AND the extents were freed by the same transaction, | |
4606 | * AND the blocks are associated with the same group. | |
4607 | */ | |
4608 | static int can_merge(struct ext4_free_data *entry1, | |
4609 | struct ext4_free_data *entry2) | |
4610 | { | |
4611 | if ((entry1->t_tid == entry2->t_tid) && | |
4612 | (entry1->group == entry2->group) && | |
4613 | ((entry1->start_blk + entry1->count) == entry2->start_blk)) | |
4614 | return 1; | |
4615 | return 0; | |
4616 | } | |
4617 | ||
4ddfef7b ES |
4618 | static noinline_for_stack int |
4619 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 4620 | struct ext4_free_data *new_entry) |
c9de560d | 4621 | { |
7a2fcbf7 AK |
4622 | ext4_grpblk_t block; |
4623 | struct ext4_free_data *entry; | |
c9de560d AT |
4624 | struct ext4_group_info *db = e4b->bd_info; |
4625 | struct super_block *sb = e4b->bd_sb; | |
4626 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4627 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
4628 | struct rb_node *parent = NULL, *new_node; | |
4629 | ||
0390131b | 4630 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4631 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4632 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4633 | ||
c894058d | 4634 | new_node = &new_entry->node; |
7a2fcbf7 | 4635 | block = new_entry->start_blk; |
c894058d | 4636 | |
c894058d AK |
4637 | if (!*n) { |
4638 | /* first free block exent. We need to | |
4639 | protect buddy cache from being freed, | |
4640 | * otherwise we'll refresh it from | |
4641 | * on-disk bitmap and lose not-yet-available | |
4642 | * blocks */ | |
4643 | page_cache_get(e4b->bd_buddy_page); | |
4644 | page_cache_get(e4b->bd_bitmap_page); | |
4645 | } | |
4646 | while (*n) { | |
4647 | parent = *n; | |
4648 | entry = rb_entry(parent, struct ext4_free_data, node); | |
4649 | if (block < entry->start_blk) | |
4650 | n = &(*n)->rb_left; | |
4651 | else if (block >= (entry->start_blk + entry->count)) | |
4652 | n = &(*n)->rb_right; | |
4653 | else { | |
5d1b1b3f AK |
4654 | ext4_grp_locked_error(sb, e4b->bd_group, __func__, |
4655 | "Double free of blocks %d (%d %d)", | |
4656 | block, entry->start_blk, entry->count); | |
c894058d | 4657 | return 0; |
c9de560d | 4658 | } |
c894058d | 4659 | } |
c9de560d | 4660 | |
c894058d AK |
4661 | rb_link_node(new_node, parent, n); |
4662 | rb_insert_color(new_node, &db->bb_free_root); | |
4663 | ||
4664 | /* Now try to see the extent can be merged to left and right */ | |
4665 | node = rb_prev(new_node); | |
4666 | if (node) { | |
4667 | entry = rb_entry(node, struct ext4_free_data, node); | |
4668 | if (can_merge(entry, new_entry)) { | |
4669 | new_entry->start_blk = entry->start_blk; | |
4670 | new_entry->count += entry->count; | |
4671 | rb_erase(node, &(db->bb_free_root)); | |
4672 | spin_lock(&sbi->s_md_lock); | |
4673 | list_del(&entry->list); | |
4674 | spin_unlock(&sbi->s_md_lock); | |
4675 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
c9de560d | 4676 | } |
c894058d | 4677 | } |
c9de560d | 4678 | |
c894058d AK |
4679 | node = rb_next(new_node); |
4680 | if (node) { | |
4681 | entry = rb_entry(node, struct ext4_free_data, node); | |
4682 | if (can_merge(new_entry, entry)) { | |
4683 | new_entry->count += entry->count; | |
4684 | rb_erase(node, &(db->bb_free_root)); | |
4685 | spin_lock(&sbi->s_md_lock); | |
4686 | list_del(&entry->list); | |
4687 | spin_unlock(&sbi->s_md_lock); | |
4688 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
c9de560d AT |
4689 | } |
4690 | } | |
3e624fc7 | 4691 | /* Add the extent to transaction's private list */ |
c894058d | 4692 | spin_lock(&sbi->s_md_lock); |
3e624fc7 | 4693 | list_add(&new_entry->list, &handle->h_transaction->t_private_list); |
c894058d | 4694 | spin_unlock(&sbi->s_md_lock); |
c9de560d AT |
4695 | return 0; |
4696 | } | |
4697 | ||
4698 | /* | |
4699 | * Main entry point into mballoc to free blocks | |
4700 | */ | |
4701 | void ext4_mb_free_blocks(handle_t *handle, struct inode *inode, | |
0610b6e9 | 4702 | ext4_fsblk_t block, unsigned long count, |
c9de560d AT |
4703 | int metadata, unsigned long *freed) |
4704 | { | |
26346ff6 | 4705 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4706 | struct super_block *sb = inode->i_sb; |
256bdb49 | 4707 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4708 | struct ext4_group_desc *gdp; |
4709 | struct ext4_super_block *es; | |
498e5f24 | 4710 | unsigned int overflow; |
c9de560d AT |
4711 | ext4_grpblk_t bit; |
4712 | struct buffer_head *gd_bh; | |
4713 | ext4_group_t block_group; | |
4714 | struct ext4_sb_info *sbi; | |
4715 | struct ext4_buddy e4b; | |
4716 | int err = 0; | |
4717 | int ret; | |
4718 | ||
4719 | *freed = 0; | |
4720 | ||
c9de560d AT |
4721 | sbi = EXT4_SB(sb); |
4722 | es = EXT4_SB(sb)->s_es; | |
4723 | if (block < le32_to_cpu(es->s_first_data_block) || | |
4724 | block + count < block || | |
4725 | block + count > ext4_blocks_count(es)) { | |
46e665e9 | 4726 | ext4_error(sb, __func__, |
c9de560d | 4727 | "Freeing blocks not in datazone - " |
0610b6e9 | 4728 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
4729 | goto error_return; |
4730 | } | |
4731 | ||
0610b6e9 | 4732 | ext4_debug("freeing block %llu\n", block); |
9bffad1e | 4733 | trace_ext4_free_blocks(inode, block, count, metadata); |
c9de560d | 4734 | |
256bdb49 ES |
4735 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
4736 | if (ac) { | |
4737 | ac->ac_op = EXT4_MB_HISTORY_FREE; | |
4738 | ac->ac_inode = inode; | |
4739 | ac->ac_sb = sb; | |
4740 | } | |
c9de560d AT |
4741 | |
4742 | do_more: | |
4743 | overflow = 0; | |
4744 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4745 | ||
4746 | /* | |
4747 | * Check to see if we are freeing blocks across a group | |
4748 | * boundary. | |
4749 | */ | |
4750 | if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { | |
4751 | overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb); | |
4752 | count -= overflow; | |
4753 | } | |
574ca174 | 4754 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
ce89f46c AK |
4755 | if (!bitmap_bh) { |
4756 | err = -EIO; | |
c9de560d | 4757 | goto error_return; |
ce89f46c | 4758 | } |
c9de560d | 4759 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4760 | if (!gdp) { |
4761 | err = -EIO; | |
c9de560d | 4762 | goto error_return; |
ce89f46c | 4763 | } |
c9de560d AT |
4764 | |
4765 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4766 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4767 | in_range(block, ext4_inode_table(sb, gdp), | |
4768 | EXT4_SB(sb)->s_itb_per_group) || | |
4769 | in_range(block + count - 1, ext4_inode_table(sb, gdp), | |
4770 | EXT4_SB(sb)->s_itb_per_group)) { | |
4771 | ||
46e665e9 | 4772 | ext4_error(sb, __func__, |
c9de560d | 4773 | "Freeing blocks in system zone - " |
0610b6e9 | 4774 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
4775 | /* err = 0. ext4_std_error should be a no op */ |
4776 | goto error_return; | |
c9de560d AT |
4777 | } |
4778 | ||
4779 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4780 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4781 | if (err) | |
4782 | goto error_return; | |
4783 | ||
4784 | /* | |
4785 | * We are about to modify some metadata. Call the journal APIs | |
4786 | * to unshare ->b_data if a currently-committing transaction is | |
4787 | * using it | |
4788 | */ | |
4789 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4790 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4791 | if (err) | |
4792 | goto error_return; | |
c9de560d AT |
4793 | #ifdef AGGRESSIVE_CHECK |
4794 | { | |
4795 | int i; | |
4796 | for (i = 0; i < count; i++) | |
4797 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); | |
4798 | } | |
4799 | #endif | |
256bdb49 ES |
4800 | if (ac) { |
4801 | ac->ac_b_ex.fe_group = block_group; | |
4802 | ac->ac_b_ex.fe_start = bit; | |
4803 | ac->ac_b_ex.fe_len = count; | |
4804 | ext4_mb_store_history(ac); | |
4805 | } | |
c9de560d | 4806 | |
920313a7 AK |
4807 | err = ext4_mb_load_buddy(sb, block_group, &e4b); |
4808 | if (err) | |
4809 | goto error_return; | |
0390131b | 4810 | if (metadata && ext4_handle_valid(handle)) { |
7a2fcbf7 AK |
4811 | struct ext4_free_data *new_entry; |
4812 | /* | |
4813 | * blocks being freed are metadata. these blocks shouldn't | |
4814 | * be used until this transaction is committed | |
4815 | */ | |
4816 | new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS); | |
4817 | new_entry->start_blk = bit; | |
4818 | new_entry->group = block_group; | |
4819 | new_entry->count = count; | |
4820 | new_entry->t_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 4821 | |
7a2fcbf7 | 4822 | ext4_lock_group(sb, block_group); |
955ce5f5 | 4823 | mb_clear_bits(bitmap_bh->b_data, bit, count); |
7a2fcbf7 | 4824 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 4825 | } else { |
7a2fcbf7 AK |
4826 | /* need to update group_info->bb_free and bitmap |
4827 | * with group lock held. generate_buddy look at | |
4828 | * them with group lock_held | |
4829 | */ | |
955ce5f5 AK |
4830 | ext4_lock_group(sb, block_group); |
4831 | mb_clear_bits(bitmap_bh->b_data, bit, count); | |
7e5a8cdd | 4832 | mb_free_blocks(inode, &e4b, bit, count); |
c9de560d | 4833 | ext4_mb_return_to_preallocation(inode, &e4b, block, count); |
c9de560d AT |
4834 | } |
4835 | ||
560671a0 AK |
4836 | ret = ext4_free_blks_count(sb, gdp) + count; |
4837 | ext4_free_blks_set(sb, gdp, ret); | |
c9de560d | 4838 | gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); |
955ce5f5 | 4839 | ext4_unlock_group(sb, block_group); |
c9de560d AT |
4840 | percpu_counter_add(&sbi->s_freeblocks_counter, count); |
4841 | ||
772cb7c8 JS |
4842 | if (sbi->s_log_groups_per_flex) { |
4843 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
9f24e420 | 4844 | atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks); |
772cb7c8 JS |
4845 | } |
4846 | ||
c9de560d AT |
4847 | ext4_mb_release_desc(&e4b); |
4848 | ||
4849 | *freed += count; | |
4850 | ||
7a2fcbf7 AK |
4851 | /* We dirtied the bitmap block */ |
4852 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4853 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4854 | ||
c9de560d AT |
4855 | /* And the group descriptor block */ |
4856 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 4857 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
4858 | if (!err) |
4859 | err = ret; | |
4860 | ||
4861 | if (overflow && !err) { | |
4862 | block += count; | |
4863 | count = overflow; | |
4864 | put_bh(bitmap_bh); | |
4865 | goto do_more; | |
4866 | } | |
4867 | sb->s_dirt = 1; | |
4868 | error_return: | |
4869 | brelse(bitmap_bh); | |
4870 | ext4_std_error(sb, err); | |
256bdb49 ES |
4871 | if (ac) |
4872 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
4873 | return; |
4874 | } |