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" |
c9de560d AT |
25 | /* |
26 | * MUSTDO: | |
27 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
28 | * - search for metadata in few groups | |
29 | * | |
30 | * TODO v4: | |
31 | * - normalization should take into account whether file is still open | |
32 | * - discard preallocations if no free space left (policy?) | |
33 | * - don't normalize tails | |
34 | * - quota | |
35 | * - reservation for superuser | |
36 | * | |
37 | * TODO v3: | |
38 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
39 | * - track min/max extents in each group for better group selection | |
40 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
41 | * - tree of groups sorted by number of free blocks | |
42 | * - error handling | |
43 | */ | |
44 | ||
45 | /* | |
46 | * The allocation request involve request for multiple number of blocks | |
47 | * near to the goal(block) value specified. | |
48 | * | |
49 | * During initialization phase of the allocator we decide to use the group | |
50 | * preallocation or inode preallocation depending on the size file. The | |
51 | * size of the file could be the resulting file size we would have after | |
52 | * allocation or the current file size which ever is larger. If the size is | |
53 | * less that sbi->s_mb_stream_request we select the group | |
54 | * preallocation. The default value of s_mb_stream_request is 16 | |
55 | * blocks. This can also be tuned via | |
56 | * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms | |
57 | * of number of blocks. | |
58 | * | |
59 | * The main motivation for having small file use group preallocation is to | |
60 | * ensure that we have small file closer in the disk. | |
61 | * | |
62 | * First stage the allocator looks at the inode prealloc list | |
63 | * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for | |
64 | * this particular inode. The inode prealloc space is represented as: | |
65 | * | |
66 | * pa_lstart -> the logical start block for this prealloc space | |
67 | * pa_pstart -> the physical start block for this prealloc space | |
68 | * pa_len -> lenght for this prealloc space | |
69 | * pa_free -> free space available in this prealloc space | |
70 | * | |
71 | * The inode preallocation space is used looking at the _logical_ start | |
72 | * block. If only the logical file block falls within the range of prealloc | |
73 | * space we will consume the particular prealloc space. This make sure that | |
74 | * that the we have contiguous physical blocks representing the file blocks | |
75 | * | |
76 | * The important thing to be noted in case of inode prealloc space is that | |
77 | * we don't modify the values associated to inode prealloc space except | |
78 | * pa_free. | |
79 | * | |
80 | * If we are not able to find blocks in the inode prealloc space and if we | |
81 | * have the group allocation flag set then we look at the locality group | |
82 | * prealloc space. These are per CPU prealloc list repreasented as | |
83 | * | |
84 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
85 | * | |
86 | * The reason for having a per cpu locality group is to reduce the contention | |
87 | * between CPUs. It is possible to get scheduled at this point. | |
88 | * | |
89 | * The locality group prealloc space is used looking at whether we have | |
90 | * enough free space (pa_free) withing the prealloc space. | |
91 | * | |
92 | * If we can't allocate blocks via inode prealloc or/and locality group | |
93 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
94 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
95 | * mapped to the buddy and bitmap information regarding different | |
96 | * groups. The buddy information is attached to buddy cache inode so that | |
97 | * we can access them through the page cache. The information regarding | |
98 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
99 | * block bitmap and buddy information. The information are stored in the | |
100 | * inode as: | |
101 | * | |
102 | * { page } | |
103 | * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]... | |
104 | * | |
105 | * | |
106 | * one block each for bitmap and buddy information. So for each group we | |
107 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / | |
108 | * blocksize) blocks. So it can have information regarding groups_per_page | |
109 | * which is blocks_per_page/2 | |
110 | * | |
111 | * The buddy cache inode is not stored on disk. The inode is thrown | |
112 | * away when the filesystem is unmounted. | |
113 | * | |
114 | * We look for count number of blocks in the buddy cache. If we were able | |
115 | * to locate that many free blocks we return with additional information | |
116 | * regarding rest of the contiguous physical block available | |
117 | * | |
118 | * Before allocating blocks via buddy cache we normalize the request | |
119 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
120 | * blocks that we get after allocation is added to the respective prealloc | |
121 | * list. In case of inode preallocation we follow a list of heuristics | |
122 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
123 | * we are doing a group prealloc we try to normalize the request to | |
124 | * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is set to | |
125 | * 512 blocks. This can be tuned via | |
126 | * /proc/fs/ext4/<partition/group_prealloc. The value is represented in | |
127 | * terms of number of blocks. If we have mounted the file system with -O | |
128 | * stripe=<value> option the group prealloc request is normalized to the | |
129 | * stripe value (sbi->s_stripe) | |
130 | * | |
131 | * The regular allocator(using the buddy cache) support few tunables. | |
132 | * | |
133 | * /proc/fs/ext4/<partition>/min_to_scan | |
134 | * /proc/fs/ext4/<partition>/max_to_scan | |
135 | * /proc/fs/ext4/<partition>/order2_req | |
136 | * | |
137 | * The regular allocator use buddy scan only if the request len is power of | |
138 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The | |
139 | * value of s_mb_order2_reqs can be tuned via | |
140 | * /proc/fs/ext4/<partition>/order2_req. If the request len is equal to | |
141 | * stripe size (sbi->s_stripe), we try to search for contigous block in | |
142 | * stripe size. This should result in better allocation on RAID setup. If | |
143 | * not we search in the specific group using bitmap for best extents. The | |
144 | * tunable min_to_scan and max_to_scan controll the behaviour here. | |
145 | * min_to_scan indicate how long the mballoc __must__ look for a best | |
146 | * extent and max_to_scanindicate how long the mballoc __can__ look for a | |
147 | * best extent in the found extents. Searching for the blocks starts with | |
148 | * the group specified as the goal value in allocation context via | |
149 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
150 | * can used for allocation. ext4_mb_good_group explains how the groups are | |
151 | * checked. | |
152 | * | |
153 | * Both the prealloc space are getting populated as above. So for the first | |
154 | * request we will hit the buddy cache which will result in this prealloc | |
155 | * space getting filled. The prealloc space is then later used for the | |
156 | * subsequent request. | |
157 | */ | |
158 | ||
159 | /* | |
160 | * mballoc operates on the following data: | |
161 | * - on-disk bitmap | |
162 | * - in-core buddy (actually includes buddy and bitmap) | |
163 | * - preallocation descriptors (PAs) | |
164 | * | |
165 | * there are two types of preallocations: | |
166 | * - inode | |
167 | * assiged to specific inode and can be used for this inode only. | |
168 | * it describes part of inode's space preallocated to specific | |
169 | * physical blocks. any block from that preallocated can be used | |
170 | * independent. the descriptor just tracks number of blocks left | |
171 | * unused. so, before taking some block from descriptor, one must | |
172 | * make sure corresponded logical block isn't allocated yet. this | |
173 | * also means that freeing any block within descriptor's range | |
174 | * must discard all preallocated blocks. | |
175 | * - locality group | |
176 | * assigned to specific locality group which does not translate to | |
177 | * permanent set of inodes: inode can join and leave group. space | |
178 | * from this type of preallocation can be used for any inode. thus | |
179 | * it's consumed from the beginning to the end. | |
180 | * | |
181 | * relation between them can be expressed as: | |
182 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
183 | * | |
184 | * this mean blocks mballoc considers used are: | |
185 | * - allocated blocks (persistent) | |
186 | * - preallocated blocks (non-persistent) | |
187 | * | |
188 | * consistency in mballoc world means that at any time a block is either | |
189 | * free or used in ALL structures. notice: "any time" should not be read | |
190 | * literally -- time is discrete and delimited by locks. | |
191 | * | |
192 | * to keep it simple, we don't use block numbers, instead we count number of | |
193 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
194 | * | |
195 | * all operations can be expressed as: | |
196 | * - init buddy: buddy = on-disk + PAs | |
197 | * - new PA: buddy += N; PA = N | |
198 | * - use inode PA: on-disk += N; PA -= N | |
199 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
200 | * - use locality group PA on-disk += N; PA -= N | |
201 | * - discard locality group PA buddy -= PA; PA = 0 | |
202 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
203 | * is used in real operation because we can't know actual used | |
204 | * bits from PA, only from on-disk bitmap | |
205 | * | |
206 | * if we follow this strict logic, then all operations above should be atomic. | |
207 | * given some of them can block, we'd have to use something like semaphores | |
208 | * killing performance on high-end SMP hardware. let's try to relax it using | |
209 | * the following knowledge: | |
210 | * 1) if buddy is referenced, it's already initialized | |
211 | * 2) while block is used in buddy and the buddy is referenced, | |
212 | * nobody can re-allocate that block | |
213 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
214 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
215 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
216 | * block | |
217 | * | |
218 | * so, now we're building a concurrency table: | |
219 | * - init buddy vs. | |
220 | * - new PA | |
221 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
222 | * until PA is linked to allocation group to avoid concurrent buddy init | |
223 | * - use inode PA | |
224 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
225 | * given (3) we care that PA-=N operation doesn't interfere with init | |
226 | * - discard inode PA | |
227 | * the simplest way would be to have buddy initialized by the discard | |
228 | * - use locality group PA | |
229 | * again PA-=N must be serialized with init | |
230 | * - discard locality group PA | |
231 | * the simplest way would be to have buddy initialized by the discard | |
232 | * - new PA vs. | |
233 | * - use inode PA | |
234 | * i_data_sem serializes them | |
235 | * - discard inode PA | |
236 | * discard process must wait until PA isn't used by another process | |
237 | * - use locality group PA | |
238 | * some mutex should serialize them | |
239 | * - discard locality group PA | |
240 | * discard process must wait until PA isn't used by another process | |
241 | * - use inode PA | |
242 | * - use inode PA | |
243 | * i_data_sem or another mutex should serializes them | |
244 | * - discard inode PA | |
245 | * discard process must wait until PA isn't used by another process | |
246 | * - use locality group PA | |
247 | * nothing wrong here -- they're different PAs covering different blocks | |
248 | * - discard locality group PA | |
249 | * discard process must wait until PA isn't used by another process | |
250 | * | |
251 | * now we're ready to make few consequences: | |
252 | * - PA is referenced and while it is no discard is possible | |
253 | * - PA is referenced until block isn't marked in on-disk bitmap | |
254 | * - PA changes only after on-disk bitmap | |
255 | * - discard must not compete with init. either init is done before | |
256 | * any discard or they're serialized somehow | |
257 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
258 | * | |
259 | * a special case when we've used PA to emptiness. no need to modify buddy | |
260 | * in this case, but we should care about concurrent init | |
261 | * | |
262 | */ | |
263 | ||
264 | /* | |
265 | * Logic in few words: | |
266 | * | |
267 | * - allocation: | |
268 | * load group | |
269 | * find blocks | |
270 | * mark bits in on-disk bitmap | |
271 | * release group | |
272 | * | |
273 | * - use preallocation: | |
274 | * find proper PA (per-inode or group) | |
275 | * load group | |
276 | * mark bits in on-disk bitmap | |
277 | * release group | |
278 | * release PA | |
279 | * | |
280 | * - free: | |
281 | * load group | |
282 | * mark bits in on-disk bitmap | |
283 | * release group | |
284 | * | |
285 | * - discard preallocations in group: | |
286 | * mark PAs deleted | |
287 | * move them onto local list | |
288 | * load on-disk bitmap | |
289 | * load group | |
290 | * remove PA from object (inode or locality group) | |
291 | * mark free blocks in-core | |
292 | * | |
293 | * - discard inode's preallocations: | |
294 | */ | |
295 | ||
296 | /* | |
297 | * Locking rules | |
298 | * | |
299 | * Locks: | |
300 | * - bitlock on a group (group) | |
301 | * - object (inode/locality) (object) | |
302 | * - per-pa lock (pa) | |
303 | * | |
304 | * Paths: | |
305 | * - new pa | |
306 | * object | |
307 | * group | |
308 | * | |
309 | * - find and use pa: | |
310 | * pa | |
311 | * | |
312 | * - release consumed pa: | |
313 | * pa | |
314 | * group | |
315 | * object | |
316 | * | |
317 | * - generate in-core bitmap: | |
318 | * group | |
319 | * pa | |
320 | * | |
321 | * - discard all for given object (inode, locality group): | |
322 | * object | |
323 | * pa | |
324 | * group | |
325 | * | |
326 | * - discard all for given group: | |
327 | * group | |
328 | * pa | |
329 | * group | |
330 | * object | |
331 | * | |
332 | */ | |
333 | ||
ffad0a44 AK |
334 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
335 | { | |
c9de560d | 336 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
337 | *bit += ((unsigned long) addr & 7UL) << 3; |
338 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 339 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
340 | *bit += ((unsigned long) addr & 3UL) << 3; |
341 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
342 | #else |
343 | #error "how many bits you are?!" | |
344 | #endif | |
ffad0a44 AK |
345 | return addr; |
346 | } | |
c9de560d AT |
347 | |
348 | static inline int mb_test_bit(int bit, void *addr) | |
349 | { | |
350 | /* | |
351 | * ext4_test_bit on architecture like powerpc | |
352 | * needs unsigned long aligned address | |
353 | */ | |
ffad0a44 | 354 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
355 | return ext4_test_bit(bit, addr); |
356 | } | |
357 | ||
358 | static inline void mb_set_bit(int bit, void *addr) | |
359 | { | |
ffad0a44 | 360 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
361 | ext4_set_bit(bit, addr); |
362 | } | |
363 | ||
364 | static inline void mb_set_bit_atomic(spinlock_t *lock, int bit, void *addr) | |
365 | { | |
ffad0a44 | 366 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
367 | ext4_set_bit_atomic(lock, bit, addr); |
368 | } | |
369 | ||
370 | static inline void mb_clear_bit(int bit, void *addr) | |
371 | { | |
ffad0a44 | 372 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
373 | ext4_clear_bit(bit, addr); |
374 | } | |
375 | ||
376 | static inline void mb_clear_bit_atomic(spinlock_t *lock, int bit, void *addr) | |
377 | { | |
ffad0a44 | 378 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
379 | ext4_clear_bit_atomic(lock, bit, addr); |
380 | } | |
381 | ||
ffad0a44 AK |
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; | |
440 | BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group)); | |
441 | for (i = 0; i < count; i++) { | |
442 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
443 | ext4_fsblk_t blocknr; | |
444 | blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb); | |
445 | blocknr += first + i; | |
446 | blocknr += | |
447 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
448 | ||
46e665e9 | 449 | ext4_error(sb, __func__, "double-free of inode" |
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; | |
464 | BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group)); | |
465 | for (i = 0; i < count; i++) { | |
466 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
467 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
468 | } | |
469 | } | |
470 | ||
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 | ||
661 | static void ext4_mb_generate_buddy(struct super_block *sb, | |
662 | void *buddy, void *bitmap, ext4_group_t group) | |
663 | { | |
664 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
665 | unsigned short max = EXT4_BLOCKS_PER_GROUP(sb); | |
666 | unsigned short i = 0; | |
667 | unsigned short first; | |
668 | unsigned short len; | |
669 | unsigned free = 0; | |
670 | unsigned fragments = 0; | |
671 | unsigned long long period = get_cycles(); | |
672 | ||
673 | /* initialize buddy from bitmap which is aggregation | |
674 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 675 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
676 | grp->bb_first_free = i; |
677 | while (i < max) { | |
678 | fragments++; | |
679 | first = i; | |
ffad0a44 | 680 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
681 | len = i - first; |
682 | free += len; | |
683 | if (len > 1) | |
684 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
685 | else | |
686 | grp->bb_counters[0]++; | |
687 | if (i < max) | |
ffad0a44 | 688 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
689 | } |
690 | grp->bb_fragments = fragments; | |
691 | ||
692 | if (free != grp->bb_free) { | |
46e665e9 | 693 | ext4_error(sb, __func__, |
a9df9a49 | 694 | "EXT4-fs: group %u: %u blocks in bitmap, %u in gd", |
c9de560d | 695 | group, free, grp->bb_free); |
e56eb659 AK |
696 | /* |
697 | * If we intent to continue, we consider group descritor | |
698 | * corrupt and update bb_free using bitmap value | |
699 | */ | |
c9de560d AT |
700 | grp->bb_free = free; |
701 | } | |
702 | ||
703 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
704 | ||
705 | period = get_cycles() - period; | |
706 | spin_lock(&EXT4_SB(sb)->s_bal_lock); | |
707 | EXT4_SB(sb)->s_mb_buddies_generated++; | |
708 | EXT4_SB(sb)->s_mb_generation_time += period; | |
709 | spin_unlock(&EXT4_SB(sb)->s_bal_lock); | |
710 | } | |
711 | ||
712 | /* The buddy information is attached the buddy cache inode | |
713 | * for convenience. The information regarding each group | |
714 | * is loaded via ext4_mb_load_buddy. The information involve | |
715 | * block bitmap and buddy information. The information are | |
716 | * stored in the inode as | |
717 | * | |
718 | * { page } | |
719 | * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]... | |
720 | * | |
721 | * | |
722 | * one block each for bitmap and buddy information. | |
723 | * So for each group we take up 2 blocks. A page can | |
724 | * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. | |
725 | * So it can have information regarding groups_per_page which | |
726 | * is blocks_per_page/2 | |
727 | */ | |
728 | ||
729 | static int ext4_mb_init_cache(struct page *page, char *incore) | |
730 | { | |
731 | int blocksize; | |
732 | int blocks_per_page; | |
733 | int groups_per_page; | |
734 | int err = 0; | |
735 | int i; | |
736 | ext4_group_t first_group; | |
737 | int first_block; | |
738 | struct super_block *sb; | |
739 | struct buffer_head *bhs; | |
740 | struct buffer_head **bh; | |
741 | struct inode *inode; | |
742 | char *data; | |
743 | char *bitmap; | |
744 | ||
745 | mb_debug("init page %lu\n", page->index); | |
746 | ||
747 | inode = page->mapping->host; | |
748 | sb = inode->i_sb; | |
749 | blocksize = 1 << inode->i_blkbits; | |
750 | blocks_per_page = PAGE_CACHE_SIZE / blocksize; | |
751 | ||
752 | groups_per_page = blocks_per_page >> 1; | |
753 | if (groups_per_page == 0) | |
754 | groups_per_page = 1; | |
755 | ||
756 | /* allocate buffer_heads to read bitmaps */ | |
757 | if (groups_per_page > 1) { | |
758 | err = -ENOMEM; | |
759 | i = sizeof(struct buffer_head *) * groups_per_page; | |
760 | bh = kzalloc(i, GFP_NOFS); | |
761 | if (bh == NULL) | |
762 | goto out; | |
763 | } else | |
764 | bh = &bhs; | |
765 | ||
766 | first_group = page->index * blocks_per_page / 2; | |
767 | ||
768 | /* read all groups the page covers into the cache */ | |
769 | for (i = 0; i < groups_per_page; i++) { | |
770 | struct ext4_group_desc *desc; | |
771 | ||
772 | if (first_group + i >= EXT4_SB(sb)->s_groups_count) | |
773 | break; | |
774 | ||
775 | err = -EIO; | |
776 | desc = ext4_get_group_desc(sb, first_group + i, NULL); | |
777 | if (desc == NULL) | |
778 | goto out; | |
779 | ||
780 | err = -ENOMEM; | |
781 | bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc)); | |
782 | if (bh[i] == NULL) | |
783 | goto out; | |
784 | ||
c806e68f FB |
785 | if (buffer_uptodate(bh[i]) && |
786 | !(desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) | |
c9de560d AT |
787 | continue; |
788 | ||
c806e68f | 789 | lock_buffer(bh[i]); |
b5f10eed | 790 | spin_lock(sb_bgl_lock(EXT4_SB(sb), first_group + i)); |
c9de560d AT |
791 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
792 | ext4_init_block_bitmap(sb, bh[i], | |
793 | first_group + i, desc); | |
794 | set_buffer_uptodate(bh[i]); | |
795 | unlock_buffer(bh[i]); | |
b5f10eed | 796 | spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i)); |
c9de560d AT |
797 | continue; |
798 | } | |
b5f10eed | 799 | spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i)); |
c9de560d AT |
800 | get_bh(bh[i]); |
801 | bh[i]->b_end_io = end_buffer_read_sync; | |
802 | submit_bh(READ, bh[i]); | |
a9df9a49 | 803 | mb_debug("read bitmap for group %u\n", first_group + i); |
c9de560d AT |
804 | } |
805 | ||
806 | /* wait for I/O completion */ | |
807 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
808 | wait_on_buffer(bh[i]); | |
809 | ||
810 | err = -EIO; | |
811 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
812 | if (!buffer_uptodate(bh[i])) | |
813 | goto out; | |
814 | ||
31b481dc | 815 | err = 0; |
c9de560d AT |
816 | first_block = page->index * blocks_per_page; |
817 | for (i = 0; i < blocks_per_page; i++) { | |
818 | int group; | |
819 | struct ext4_group_info *grinfo; | |
820 | ||
821 | group = (first_block + i) >> 1; | |
822 | if (group >= EXT4_SB(sb)->s_groups_count) | |
823 | break; | |
824 | ||
825 | /* | |
826 | * data carry information regarding this | |
827 | * particular group in the format specified | |
828 | * above | |
829 | * | |
830 | */ | |
831 | data = page_address(page) + (i * blocksize); | |
832 | bitmap = bh[group - first_group]->b_data; | |
833 | ||
834 | /* | |
835 | * We place the buddy block and bitmap block | |
836 | * close together | |
837 | */ | |
838 | if ((first_block + i) & 1) { | |
839 | /* this is block of buddy */ | |
840 | BUG_ON(incore == NULL); | |
841 | mb_debug("put buddy for group %u in page %lu/%x\n", | |
842 | group, page->index, i * blocksize); | |
843 | memset(data, 0xff, blocksize); | |
844 | grinfo = ext4_get_group_info(sb, group); | |
845 | grinfo->bb_fragments = 0; | |
846 | memset(grinfo->bb_counters, 0, | |
847 | sizeof(unsigned short)*(sb->s_blocksize_bits+2)); | |
848 | /* | |
849 | * incore got set to the group block bitmap below | |
850 | */ | |
851 | ext4_mb_generate_buddy(sb, data, incore, group); | |
852 | incore = NULL; | |
853 | } else { | |
854 | /* this is block of bitmap */ | |
855 | BUG_ON(incore != NULL); | |
856 | mb_debug("put bitmap for group %u in page %lu/%x\n", | |
857 | group, page->index, i * blocksize); | |
858 | ||
859 | /* see comments in ext4_mb_put_pa() */ | |
860 | ext4_lock_group(sb, group); | |
861 | memcpy(data, bitmap, blocksize); | |
862 | ||
863 | /* mark all preallocated blks used in in-core bitmap */ | |
864 | ext4_mb_generate_from_pa(sb, data, group); | |
865 | ext4_unlock_group(sb, group); | |
866 | ||
867 | /* set incore so that the buddy information can be | |
868 | * generated using this | |
869 | */ | |
870 | incore = data; | |
871 | } | |
872 | } | |
873 | SetPageUptodate(page); | |
874 | ||
875 | out: | |
876 | if (bh) { | |
877 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
878 | brelse(bh[i]); | |
879 | if (bh != &bhs) | |
880 | kfree(bh); | |
881 | } | |
882 | return err; | |
883 | } | |
884 | ||
4ddfef7b ES |
885 | static noinline_for_stack int |
886 | ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, | |
887 | struct ext4_buddy *e4b) | |
c9de560d AT |
888 | { |
889 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
890 | struct inode *inode = sbi->s_buddy_cache; | |
891 | int blocks_per_page; | |
892 | int block; | |
893 | int pnum; | |
894 | int poff; | |
895 | struct page *page; | |
fdf6c7a7 | 896 | int ret; |
c9de560d | 897 | |
a9df9a49 | 898 | mb_debug("load group %u\n", group); |
c9de560d AT |
899 | |
900 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
901 | ||
902 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
903 | e4b->bd_info = ext4_get_group_info(sb, group); | |
904 | e4b->bd_sb = sb; | |
905 | e4b->bd_group = group; | |
906 | e4b->bd_buddy_page = NULL; | |
907 | e4b->bd_bitmap_page = NULL; | |
908 | ||
909 | /* | |
910 | * the buddy cache inode stores the block bitmap | |
911 | * and buddy information in consecutive blocks. | |
912 | * So for each group we need two blocks. | |
913 | */ | |
914 | block = group * 2; | |
915 | pnum = block / blocks_per_page; | |
916 | poff = block % blocks_per_page; | |
917 | ||
918 | /* we could use find_or_create_page(), but it locks page | |
919 | * what we'd like to avoid in fast path ... */ | |
920 | page = find_get_page(inode->i_mapping, pnum); | |
921 | if (page == NULL || !PageUptodate(page)) { | |
922 | if (page) | |
923 | page_cache_release(page); | |
924 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
925 | if (page) { | |
926 | BUG_ON(page->mapping != inode->i_mapping); | |
927 | if (!PageUptodate(page)) { | |
fdf6c7a7 SF |
928 | ret = ext4_mb_init_cache(page, NULL); |
929 | if (ret) { | |
930 | unlock_page(page); | |
931 | goto err; | |
932 | } | |
c9de560d AT |
933 | mb_cmp_bitmaps(e4b, page_address(page) + |
934 | (poff * sb->s_blocksize)); | |
935 | } | |
936 | unlock_page(page); | |
937 | } | |
938 | } | |
fdf6c7a7 SF |
939 | if (page == NULL || !PageUptodate(page)) { |
940 | ret = -EIO; | |
c9de560d | 941 | goto err; |
fdf6c7a7 | 942 | } |
c9de560d AT |
943 | e4b->bd_bitmap_page = page; |
944 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
945 | mark_page_accessed(page); | |
946 | ||
947 | block++; | |
948 | pnum = block / blocks_per_page; | |
949 | poff = block % blocks_per_page; | |
950 | ||
951 | page = find_get_page(inode->i_mapping, pnum); | |
952 | if (page == NULL || !PageUptodate(page)) { | |
953 | if (page) | |
954 | page_cache_release(page); | |
955 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
956 | if (page) { | |
957 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 SF |
958 | if (!PageUptodate(page)) { |
959 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap); | |
960 | if (ret) { | |
961 | unlock_page(page); | |
962 | goto err; | |
963 | } | |
964 | } | |
c9de560d AT |
965 | unlock_page(page); |
966 | } | |
967 | } | |
fdf6c7a7 SF |
968 | if (page == NULL || !PageUptodate(page)) { |
969 | ret = -EIO; | |
c9de560d | 970 | goto err; |
fdf6c7a7 | 971 | } |
c9de560d AT |
972 | e4b->bd_buddy_page = page; |
973 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
974 | mark_page_accessed(page); | |
975 | ||
976 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
977 | BUG_ON(e4b->bd_buddy_page == NULL); | |
978 | ||
979 | return 0; | |
980 | ||
981 | err: | |
982 | if (e4b->bd_bitmap_page) | |
983 | page_cache_release(e4b->bd_bitmap_page); | |
984 | if (e4b->bd_buddy_page) | |
985 | page_cache_release(e4b->bd_buddy_page); | |
986 | e4b->bd_buddy = NULL; | |
987 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 988 | return ret; |
c9de560d AT |
989 | } |
990 | ||
991 | static void ext4_mb_release_desc(struct ext4_buddy *e4b) | |
992 | { | |
993 | if (e4b->bd_bitmap_page) | |
994 | page_cache_release(e4b->bd_bitmap_page); | |
995 | if (e4b->bd_buddy_page) | |
996 | page_cache_release(e4b->bd_buddy_page); | |
997 | } | |
998 | ||
999 | ||
1000 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1001 | { | |
1002 | int order = 1; | |
1003 | void *bb; | |
1004 | ||
1005 | BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); | |
1006 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); | |
1007 | ||
1008 | bb = EXT4_MB_BUDDY(e4b); | |
1009 | while (order <= e4b->bd_blkbits + 1) { | |
1010 | block = block >> 1; | |
1011 | if (!mb_test_bit(block, bb)) { | |
1012 | /* this block is part of buddy of order 'order' */ | |
1013 | return order; | |
1014 | } | |
1015 | bb += 1 << (e4b->bd_blkbits - order); | |
1016 | order++; | |
1017 | } | |
1018 | return 0; | |
1019 | } | |
1020 | ||
1021 | static void mb_clear_bits(spinlock_t *lock, void *bm, int cur, int len) | |
1022 | { | |
1023 | __u32 *addr; | |
1024 | ||
1025 | len = cur + len; | |
1026 | while (cur < len) { | |
1027 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1028 | /* fast path: clear whole word at once */ | |
1029 | addr = bm + (cur >> 3); | |
1030 | *addr = 0; | |
1031 | cur += 32; | |
1032 | continue; | |
1033 | } | |
1034 | mb_clear_bit_atomic(lock, cur, bm); | |
1035 | cur++; | |
1036 | } | |
1037 | } | |
1038 | ||
1039 | static void mb_set_bits(spinlock_t *lock, void *bm, int cur, int len) | |
1040 | { | |
1041 | __u32 *addr; | |
1042 | ||
1043 | len = cur + len; | |
1044 | while (cur < len) { | |
1045 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1046 | /* fast path: set whole word at once */ | |
1047 | addr = bm + (cur >> 3); | |
1048 | *addr = 0xffffffff; | |
1049 | cur += 32; | |
1050 | continue; | |
1051 | } | |
1052 | mb_set_bit_atomic(lock, cur, bm); | |
1053 | cur++; | |
1054 | } | |
1055 | } | |
1056 | ||
7e5a8cdd | 1057 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
c9de560d AT |
1058 | int first, int count) |
1059 | { | |
1060 | int block = 0; | |
1061 | int max = 0; | |
1062 | int order; | |
1063 | void *buddy; | |
1064 | void *buddy2; | |
1065 | struct super_block *sb = e4b->bd_sb; | |
1066 | ||
1067 | BUG_ON(first + count > (sb->s_blocksize << 3)); | |
1068 | BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group)); | |
1069 | mb_check_buddy(e4b); | |
1070 | mb_free_blocks_double(inode, e4b, first, count); | |
1071 | ||
1072 | e4b->bd_info->bb_free += count; | |
1073 | if (first < e4b->bd_info->bb_first_free) | |
1074 | e4b->bd_info->bb_first_free = first; | |
1075 | ||
1076 | /* let's maintain fragments counter */ | |
1077 | if (first != 0) | |
1078 | block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b)); | |
1079 | if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) | |
1080 | max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b)); | |
1081 | if (block && max) | |
1082 | e4b->bd_info->bb_fragments--; | |
1083 | else if (!block && !max) | |
1084 | e4b->bd_info->bb_fragments++; | |
1085 | ||
1086 | /* let's maintain buddy itself */ | |
1087 | while (count-- > 0) { | |
1088 | block = first++; | |
1089 | order = 0; | |
1090 | ||
1091 | if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) { | |
1092 | ext4_fsblk_t blocknr; | |
1093 | blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb); | |
1094 | blocknr += block; | |
1095 | blocknr += | |
1096 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
7e5a8cdd | 1097 | ext4_unlock_group(sb, e4b->bd_group); |
46e665e9 | 1098 | ext4_error(sb, __func__, "double-free of inode" |
a9df9a49 | 1099 | " %lu's block %llu(bit %u in group %u)", |
c9de560d AT |
1100 | inode ? inode->i_ino : 0, blocknr, block, |
1101 | e4b->bd_group); | |
7e5a8cdd | 1102 | ext4_lock_group(sb, e4b->bd_group); |
c9de560d AT |
1103 | } |
1104 | mb_clear_bit(block, EXT4_MB_BITMAP(e4b)); | |
1105 | e4b->bd_info->bb_counters[order]++; | |
1106 | ||
1107 | /* start of the buddy */ | |
1108 | buddy = mb_find_buddy(e4b, order, &max); | |
1109 | ||
1110 | do { | |
1111 | block &= ~1UL; | |
1112 | if (mb_test_bit(block, buddy) || | |
1113 | mb_test_bit(block + 1, buddy)) | |
1114 | break; | |
1115 | ||
1116 | /* both the buddies are free, try to coalesce them */ | |
1117 | buddy2 = mb_find_buddy(e4b, order + 1, &max); | |
1118 | ||
1119 | if (!buddy2) | |
1120 | break; | |
1121 | ||
1122 | if (order > 0) { | |
1123 | /* for special purposes, we don't set | |
1124 | * free bits in bitmap */ | |
1125 | mb_set_bit(block, buddy); | |
1126 | mb_set_bit(block + 1, buddy); | |
1127 | } | |
1128 | e4b->bd_info->bb_counters[order]--; | |
1129 | e4b->bd_info->bb_counters[order]--; | |
1130 | ||
1131 | block = block >> 1; | |
1132 | order++; | |
1133 | e4b->bd_info->bb_counters[order]++; | |
1134 | ||
1135 | mb_clear_bit(block, buddy2); | |
1136 | buddy = buddy2; | |
1137 | } while (1); | |
1138 | } | |
1139 | mb_check_buddy(e4b); | |
c9de560d AT |
1140 | } |
1141 | ||
1142 | static int mb_find_extent(struct ext4_buddy *e4b, int order, int block, | |
1143 | int needed, struct ext4_free_extent *ex) | |
1144 | { | |
1145 | int next = block; | |
1146 | int max; | |
1147 | int ord; | |
1148 | void *buddy; | |
1149 | ||
1150 | BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group)); | |
1151 | BUG_ON(ex == NULL); | |
1152 | ||
1153 | buddy = mb_find_buddy(e4b, order, &max); | |
1154 | BUG_ON(buddy == NULL); | |
1155 | BUG_ON(block >= max); | |
1156 | if (mb_test_bit(block, buddy)) { | |
1157 | ex->fe_len = 0; | |
1158 | ex->fe_start = 0; | |
1159 | ex->fe_group = 0; | |
1160 | return 0; | |
1161 | } | |
1162 | ||
1163 | /* FIXME dorp order completely ? */ | |
1164 | if (likely(order == 0)) { | |
1165 | /* find actual order */ | |
1166 | order = mb_find_order_for_block(e4b, block); | |
1167 | block = block >> order; | |
1168 | } | |
1169 | ||
1170 | ex->fe_len = 1 << order; | |
1171 | ex->fe_start = block << order; | |
1172 | ex->fe_group = e4b->bd_group; | |
1173 | ||
1174 | /* calc difference from given start */ | |
1175 | next = next - ex->fe_start; | |
1176 | ex->fe_len -= next; | |
1177 | ex->fe_start += next; | |
1178 | ||
1179 | while (needed > ex->fe_len && | |
1180 | (buddy = mb_find_buddy(e4b, order, &max))) { | |
1181 | ||
1182 | if (block + 1 >= max) | |
1183 | break; | |
1184 | ||
1185 | next = (block + 1) * (1 << order); | |
1186 | if (mb_test_bit(next, EXT4_MB_BITMAP(e4b))) | |
1187 | break; | |
1188 | ||
1189 | ord = mb_find_order_for_block(e4b, next); | |
1190 | ||
1191 | order = ord; | |
1192 | block = next >> order; | |
1193 | ex->fe_len += 1 << order; | |
1194 | } | |
1195 | ||
1196 | BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); | |
1197 | return ex->fe_len; | |
1198 | } | |
1199 | ||
1200 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1201 | { | |
1202 | int ord; | |
1203 | int mlen = 0; | |
1204 | int max = 0; | |
1205 | int cur; | |
1206 | int start = ex->fe_start; | |
1207 | int len = ex->fe_len; | |
1208 | unsigned ret = 0; | |
1209 | int len0 = len; | |
1210 | void *buddy; | |
1211 | ||
1212 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1213 | BUG_ON(e4b->bd_group != ex->fe_group); | |
1214 | BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group)); | |
1215 | mb_check_buddy(e4b); | |
1216 | mb_mark_used_double(e4b, start, len); | |
1217 | ||
1218 | e4b->bd_info->bb_free -= len; | |
1219 | if (e4b->bd_info->bb_first_free == start) | |
1220 | e4b->bd_info->bb_first_free += len; | |
1221 | ||
1222 | /* let's maintain fragments counter */ | |
1223 | if (start != 0) | |
1224 | mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b)); | |
1225 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) | |
1226 | max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b)); | |
1227 | if (mlen && max) | |
1228 | e4b->bd_info->bb_fragments++; | |
1229 | else if (!mlen && !max) | |
1230 | e4b->bd_info->bb_fragments--; | |
1231 | ||
1232 | /* let's maintain buddy itself */ | |
1233 | while (len) { | |
1234 | ord = mb_find_order_for_block(e4b, start); | |
1235 | ||
1236 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1237 | /* the whole chunk may be allocated at once! */ | |
1238 | mlen = 1 << ord; | |
1239 | buddy = mb_find_buddy(e4b, ord, &max); | |
1240 | BUG_ON((start >> ord) >= max); | |
1241 | mb_set_bit(start >> ord, buddy); | |
1242 | e4b->bd_info->bb_counters[ord]--; | |
1243 | start += mlen; | |
1244 | len -= mlen; | |
1245 | BUG_ON(len < 0); | |
1246 | continue; | |
1247 | } | |
1248 | ||
1249 | /* store for history */ | |
1250 | if (ret == 0) | |
1251 | ret = len | (ord << 16); | |
1252 | ||
1253 | /* we have to split large buddy */ | |
1254 | BUG_ON(ord <= 0); | |
1255 | buddy = mb_find_buddy(e4b, ord, &max); | |
1256 | mb_set_bit(start >> ord, buddy); | |
1257 | e4b->bd_info->bb_counters[ord]--; | |
1258 | ||
1259 | ord--; | |
1260 | cur = (start >> ord) & ~1U; | |
1261 | buddy = mb_find_buddy(e4b, ord, &max); | |
1262 | mb_clear_bit(cur, buddy); | |
1263 | mb_clear_bit(cur + 1, buddy); | |
1264 | e4b->bd_info->bb_counters[ord]++; | |
1265 | e4b->bd_info->bb_counters[ord]++; | |
1266 | } | |
1267 | ||
1268 | mb_set_bits(sb_bgl_lock(EXT4_SB(e4b->bd_sb), ex->fe_group), | |
1269 | EXT4_MB_BITMAP(e4b), ex->fe_start, len0); | |
1270 | mb_check_buddy(e4b); | |
1271 | ||
1272 | return ret; | |
1273 | } | |
1274 | ||
1275 | /* | |
1276 | * Must be called under group lock! | |
1277 | */ | |
1278 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1279 | struct ext4_buddy *e4b) | |
1280 | { | |
1281 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1282 | int ret; | |
1283 | ||
1284 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1285 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1286 | ||
1287 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1288 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1289 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1290 | ||
1291 | /* preallocation can change ac_b_ex, thus we store actually | |
1292 | * allocated blocks for history */ | |
1293 | ac->ac_f_ex = ac->ac_b_ex; | |
1294 | ||
1295 | ac->ac_status = AC_STATUS_FOUND; | |
1296 | ac->ac_tail = ret & 0xffff; | |
1297 | ac->ac_buddy = ret >> 16; | |
1298 | ||
1299 | /* XXXXXXX: SUCH A HORRIBLE **CK */ | |
1300 | /*FIXME!! Why ? */ | |
1301 | ac->ac_bitmap_page = e4b->bd_bitmap_page; | |
1302 | get_page(ac->ac_bitmap_page); | |
1303 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1304 | get_page(ac->ac_buddy_page); | |
1305 | ||
1306 | /* store last allocated for subsequent stream allocation */ | |
1307 | if ((ac->ac_flags & EXT4_MB_HINT_DATA)) { | |
1308 | spin_lock(&sbi->s_md_lock); | |
1309 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1310 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1311 | spin_unlock(&sbi->s_md_lock); | |
1312 | } | |
1313 | } | |
1314 | ||
1315 | /* | |
1316 | * regular allocator, for general purposes allocation | |
1317 | */ | |
1318 | ||
1319 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1320 | struct ext4_buddy *e4b, | |
1321 | int finish_group) | |
1322 | { | |
1323 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1324 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1325 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1326 | struct ext4_free_extent ex; | |
1327 | int max; | |
1328 | ||
032115fc AK |
1329 | if (ac->ac_status == AC_STATUS_FOUND) |
1330 | return; | |
c9de560d AT |
1331 | /* |
1332 | * We don't want to scan for a whole year | |
1333 | */ | |
1334 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1335 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1336 | ac->ac_status = AC_STATUS_BREAK; | |
1337 | return; | |
1338 | } | |
1339 | ||
1340 | /* | |
1341 | * Haven't found good chunk so far, let's continue | |
1342 | */ | |
1343 | if (bex->fe_len < gex->fe_len) | |
1344 | return; | |
1345 | ||
1346 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1347 | && bex->fe_group == e4b->bd_group) { | |
1348 | /* recheck chunk's availability - we don't know | |
1349 | * when it was found (within this lock-unlock | |
1350 | * period or not) */ | |
1351 | max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex); | |
1352 | if (max >= gex->fe_len) { | |
1353 | ext4_mb_use_best_found(ac, e4b); | |
1354 | return; | |
1355 | } | |
1356 | } | |
1357 | } | |
1358 | ||
1359 | /* | |
1360 | * The routine checks whether found extent is good enough. If it is, | |
1361 | * then the extent gets marked used and flag is set to the context | |
1362 | * to stop scanning. Otherwise, the extent is compared with the | |
1363 | * previous found extent and if new one is better, then it's stored | |
1364 | * in the context. Later, the best found extent will be used, if | |
1365 | * mballoc can't find good enough extent. | |
1366 | * | |
1367 | * FIXME: real allocation policy is to be designed yet! | |
1368 | */ | |
1369 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1370 | struct ext4_free_extent *ex, | |
1371 | struct ext4_buddy *e4b) | |
1372 | { | |
1373 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1374 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1375 | ||
1376 | BUG_ON(ex->fe_len <= 0); | |
1377 | BUG_ON(ex->fe_len >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); | |
1378 | BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); | |
1379 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); | |
1380 | ||
1381 | ac->ac_found++; | |
1382 | ||
1383 | /* | |
1384 | * The special case - take what you catch first | |
1385 | */ | |
1386 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1387 | *bex = *ex; | |
1388 | ext4_mb_use_best_found(ac, e4b); | |
1389 | return; | |
1390 | } | |
1391 | ||
1392 | /* | |
1393 | * Let's check whether the chuck is good enough | |
1394 | */ | |
1395 | if (ex->fe_len == gex->fe_len) { | |
1396 | *bex = *ex; | |
1397 | ext4_mb_use_best_found(ac, e4b); | |
1398 | return; | |
1399 | } | |
1400 | ||
1401 | /* | |
1402 | * If this is first found extent, just store it in the context | |
1403 | */ | |
1404 | if (bex->fe_len == 0) { | |
1405 | *bex = *ex; | |
1406 | return; | |
1407 | } | |
1408 | ||
1409 | /* | |
1410 | * If new found extent is better, store it in the context | |
1411 | */ | |
1412 | if (bex->fe_len < gex->fe_len) { | |
1413 | /* if the request isn't satisfied, any found extent | |
1414 | * larger than previous best one is better */ | |
1415 | if (ex->fe_len > bex->fe_len) | |
1416 | *bex = *ex; | |
1417 | } else if (ex->fe_len > gex->fe_len) { | |
1418 | /* if the request is satisfied, then we try to find | |
1419 | * an extent that still satisfy the request, but is | |
1420 | * smaller than previous one */ | |
1421 | if (ex->fe_len < bex->fe_len) | |
1422 | *bex = *ex; | |
1423 | } | |
1424 | ||
1425 | ext4_mb_check_limits(ac, e4b, 0); | |
1426 | } | |
1427 | ||
1428 | static int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
1429 | struct ext4_buddy *e4b) | |
1430 | { | |
1431 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1432 | ext4_group_t group = ex.fe_group; | |
1433 | int max; | |
1434 | int err; | |
1435 | ||
1436 | BUG_ON(ex.fe_len <= 0); | |
1437 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1438 | if (err) | |
1439 | return err; | |
1440 | ||
1441 | ext4_lock_group(ac->ac_sb, group); | |
1442 | max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex); | |
1443 | ||
1444 | if (max > 0) { | |
1445 | ac->ac_b_ex = ex; | |
1446 | ext4_mb_use_best_found(ac, e4b); | |
1447 | } | |
1448 | ||
1449 | ext4_unlock_group(ac->ac_sb, group); | |
1450 | ext4_mb_release_desc(e4b); | |
1451 | ||
1452 | return 0; | |
1453 | } | |
1454 | ||
1455 | static int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
1456 | struct ext4_buddy *e4b) | |
1457 | { | |
1458 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1459 | int max; | |
1460 | int err; | |
1461 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1462 | struct ext4_super_block *es = sbi->s_es; | |
1463 | struct ext4_free_extent ex; | |
1464 | ||
1465 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1466 | return 0; | |
1467 | ||
1468 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1469 | if (err) | |
1470 | return err; | |
1471 | ||
1472 | ext4_lock_group(ac->ac_sb, group); | |
1473 | max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start, | |
1474 | ac->ac_g_ex.fe_len, &ex); | |
1475 | ||
1476 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1477 | ext4_fsblk_t start; | |
1478 | ||
1479 | start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) + | |
1480 | ex.fe_start + le32_to_cpu(es->s_first_data_block); | |
1481 | /* use do_div to get remainder (would be 64-bit modulo) */ | |
1482 | if (do_div(start, sbi->s_stripe) == 0) { | |
1483 | ac->ac_found++; | |
1484 | ac->ac_b_ex = ex; | |
1485 | ext4_mb_use_best_found(ac, e4b); | |
1486 | } | |
1487 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1488 | BUG_ON(ex.fe_len <= 0); | |
1489 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1490 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1491 | ac->ac_found++; | |
1492 | ac->ac_b_ex = ex; | |
1493 | ext4_mb_use_best_found(ac, e4b); | |
1494 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1495 | /* Sometimes, caller may want to merge even small | |
1496 | * number of blocks to an existing extent */ | |
1497 | BUG_ON(ex.fe_len <= 0); | |
1498 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1499 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1500 | ac->ac_found++; | |
1501 | ac->ac_b_ex = ex; | |
1502 | ext4_mb_use_best_found(ac, e4b); | |
1503 | } | |
1504 | ext4_unlock_group(ac->ac_sb, group); | |
1505 | ext4_mb_release_desc(e4b); | |
1506 | ||
1507 | return 0; | |
1508 | } | |
1509 | ||
1510 | /* | |
1511 | * The routine scans buddy structures (not bitmap!) from given order | |
1512 | * to max order and tries to find big enough chunk to satisfy the req | |
1513 | */ | |
1514 | static void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
1515 | struct ext4_buddy *e4b) | |
1516 | { | |
1517 | struct super_block *sb = ac->ac_sb; | |
1518 | struct ext4_group_info *grp = e4b->bd_info; | |
1519 | void *buddy; | |
1520 | int i; | |
1521 | int k; | |
1522 | int max; | |
1523 | ||
1524 | BUG_ON(ac->ac_2order <= 0); | |
1525 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1526 | if (grp->bb_counters[i] == 0) | |
1527 | continue; | |
1528 | ||
1529 | buddy = mb_find_buddy(e4b, i, &max); | |
1530 | BUG_ON(buddy == NULL); | |
1531 | ||
ffad0a44 | 1532 | k = mb_find_next_zero_bit(buddy, max, 0); |
c9de560d AT |
1533 | BUG_ON(k >= max); |
1534 | ||
1535 | ac->ac_found++; | |
1536 | ||
1537 | ac->ac_b_ex.fe_len = 1 << i; | |
1538 | ac->ac_b_ex.fe_start = k << i; | |
1539 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1540 | ||
1541 | ext4_mb_use_best_found(ac, e4b); | |
1542 | ||
1543 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1544 | ||
1545 | if (EXT4_SB(sb)->s_mb_stats) | |
1546 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1547 | ||
1548 | break; | |
1549 | } | |
1550 | } | |
1551 | ||
1552 | /* | |
1553 | * The routine scans the group and measures all found extents. | |
1554 | * In order to optimize scanning, caller must pass number of | |
1555 | * free blocks in the group, so the routine can know upper limit. | |
1556 | */ | |
1557 | static void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
1558 | struct ext4_buddy *e4b) | |
1559 | { | |
1560 | struct super_block *sb = ac->ac_sb; | |
1561 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1562 | struct ext4_free_extent ex; | |
1563 | int i; | |
1564 | int free; | |
1565 | ||
1566 | free = e4b->bd_info->bb_free; | |
1567 | BUG_ON(free <= 0); | |
1568 | ||
1569 | i = e4b->bd_info->bb_first_free; | |
1570 | ||
1571 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1572 | i = mb_find_next_zero_bit(bitmap, |
c9de560d AT |
1573 | EXT4_BLOCKS_PER_GROUP(sb), i); |
1574 | if (i >= EXT4_BLOCKS_PER_GROUP(sb)) { | |
26346ff6 | 1575 | /* |
e56eb659 | 1576 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1577 | * free blocks even though group info says we |
1578 | * we have free blocks | |
1579 | */ | |
46e665e9 | 1580 | ext4_error(sb, __func__, "%d free blocks as per " |
fde4d95a | 1581 | "group info. But bitmap says 0", |
26346ff6 | 1582 | free); |
c9de560d AT |
1583 | break; |
1584 | } | |
1585 | ||
1586 | mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex); | |
1587 | BUG_ON(ex.fe_len <= 0); | |
26346ff6 | 1588 | if (free < ex.fe_len) { |
46e665e9 | 1589 | ext4_error(sb, __func__, "%d free blocks as per " |
fde4d95a | 1590 | "group info. But got %d blocks", |
26346ff6 | 1591 | free, ex.fe_len); |
e56eb659 AK |
1592 | /* |
1593 | * The number of free blocks differs. This mostly | |
1594 | * indicate that the bitmap is corrupt. So exit | |
1595 | * without claiming the space. | |
1596 | */ | |
1597 | break; | |
26346ff6 | 1598 | } |
c9de560d AT |
1599 | |
1600 | ext4_mb_measure_extent(ac, &ex, e4b); | |
1601 | ||
1602 | i += ex.fe_len; | |
1603 | free -= ex.fe_len; | |
1604 | } | |
1605 | ||
1606 | ext4_mb_check_limits(ac, e4b, 1); | |
1607 | } | |
1608 | ||
1609 | /* | |
1610 | * This is a special case for storages like raid5 | |
1611 | * we try to find stripe-aligned chunks for stripe-size requests | |
1612 | * XXX should do so at least for multiples of stripe size as well | |
1613 | */ | |
1614 | static void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
1615 | struct ext4_buddy *e4b) | |
1616 | { | |
1617 | struct super_block *sb = ac->ac_sb; | |
1618 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1619 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1620 | struct ext4_free_extent ex; | |
1621 | ext4_fsblk_t first_group_block; | |
1622 | ext4_fsblk_t a; | |
1623 | ext4_grpblk_t i; | |
1624 | int max; | |
1625 | ||
1626 | BUG_ON(sbi->s_stripe == 0); | |
1627 | ||
1628 | /* find first stripe-aligned block in group */ | |
1629 | first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb) | |
1630 | + le32_to_cpu(sbi->s_es->s_first_data_block); | |
1631 | a = first_group_block + sbi->s_stripe - 1; | |
1632 | do_div(a, sbi->s_stripe); | |
1633 | i = (a * sbi->s_stripe) - first_group_block; | |
1634 | ||
1635 | while (i < EXT4_BLOCKS_PER_GROUP(sb)) { | |
1636 | if (!mb_test_bit(i, bitmap)) { | |
1637 | max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex); | |
1638 | if (max >= sbi->s_stripe) { | |
1639 | ac->ac_found++; | |
1640 | ac->ac_b_ex = ex; | |
1641 | ext4_mb_use_best_found(ac, e4b); | |
1642 | break; | |
1643 | } | |
1644 | } | |
1645 | i += sbi->s_stripe; | |
1646 | } | |
1647 | } | |
1648 | ||
1649 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, | |
1650 | ext4_group_t group, int cr) | |
1651 | { | |
1652 | unsigned free, fragments; | |
1653 | unsigned i, bits; | |
1654 | struct ext4_group_desc *desc; | |
1655 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); | |
1656 | ||
1657 | BUG_ON(cr < 0 || cr >= 4); | |
1658 | BUG_ON(EXT4_MB_GRP_NEED_INIT(grp)); | |
1659 | ||
1660 | free = grp->bb_free; | |
1661 | fragments = grp->bb_fragments; | |
1662 | if (free == 0) | |
1663 | return 0; | |
1664 | if (fragments == 0) | |
1665 | return 0; | |
1666 | ||
1667 | switch (cr) { | |
1668 | case 0: | |
1669 | BUG_ON(ac->ac_2order == 0); | |
1670 | /* If this group is uninitialized, skip it initially */ | |
1671 | desc = ext4_get_group_desc(ac->ac_sb, group, NULL); | |
1672 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) | |
1673 | return 0; | |
1674 | ||
1675 | bits = ac->ac_sb->s_blocksize_bits + 1; | |
1676 | for (i = ac->ac_2order; i <= bits; i++) | |
1677 | if (grp->bb_counters[i] > 0) | |
1678 | return 1; | |
1679 | break; | |
1680 | case 1: | |
1681 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
1682 | return 1; | |
1683 | break; | |
1684 | case 2: | |
1685 | if (free >= ac->ac_g_ex.fe_len) | |
1686 | return 1; | |
1687 | break; | |
1688 | case 3: | |
1689 | return 1; | |
1690 | default: | |
1691 | BUG(); | |
1692 | } | |
1693 | ||
1694 | return 0; | |
1695 | } | |
1696 | ||
4ddfef7b ES |
1697 | static noinline_for_stack int |
1698 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d AT |
1699 | { |
1700 | ext4_group_t group; | |
1701 | ext4_group_t i; | |
1702 | int cr; | |
1703 | int err = 0; | |
1704 | int bsbits; | |
1705 | struct ext4_sb_info *sbi; | |
1706 | struct super_block *sb; | |
1707 | struct ext4_buddy e4b; | |
1708 | loff_t size, isize; | |
1709 | ||
1710 | sb = ac->ac_sb; | |
1711 | sbi = EXT4_SB(sb); | |
1712 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1713 | ||
1714 | /* first, try the goal */ | |
1715 | err = ext4_mb_find_by_goal(ac, &e4b); | |
1716 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
1717 | goto out; | |
1718 | ||
1719 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
1720 | goto out; | |
1721 | ||
1722 | /* | |
1723 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
1724 | * if ac2_order is set we also set criteria to 0 so that we | |
1725 | * try exact allocation using buddy. | |
1726 | */ | |
1727 | i = fls(ac->ac_g_ex.fe_len); | |
1728 | ac->ac_2order = 0; | |
1729 | /* | |
1730 | * We search using buddy data only if the order of the request | |
1731 | * is greater than equal to the sbi_s_mb_order2_reqs | |
1732 | * You can tune it via /proc/fs/ext4/<partition>/order2_req | |
1733 | */ | |
1734 | if (i >= sbi->s_mb_order2_reqs) { | |
1735 | /* | |
1736 | * This should tell if fe_len is exactly power of 2 | |
1737 | */ | |
1738 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1739 | ac->ac_2order = i - 1; | |
1740 | } | |
1741 | ||
1742 | bsbits = ac->ac_sb->s_blocksize_bits; | |
1743 | /* if stream allocation is enabled, use global goal */ | |
1744 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; | |
1745 | isize = i_size_read(ac->ac_inode) >> bsbits; | |
1746 | if (size < isize) | |
1747 | size = isize; | |
1748 | ||
1749 | if (size < sbi->s_mb_stream_request && | |
1750 | (ac->ac_flags & EXT4_MB_HINT_DATA)) { | |
1751 | /* TBD: may be hot point */ | |
1752 | spin_lock(&sbi->s_md_lock); | |
1753 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
1754 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
1755 | spin_unlock(&sbi->s_md_lock); | |
1756 | } | |
c9de560d AT |
1757 | /* Let's just scan groups to find more-less suitable blocks */ |
1758 | cr = ac->ac_2order ? 0 : 1; | |
1759 | /* | |
1760 | * cr == 0 try to get exact allocation, | |
1761 | * cr == 3 try to get anything | |
1762 | */ | |
1763 | repeat: | |
1764 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
1765 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
1766 | /* |
1767 | * searching for the right group start | |
1768 | * from the goal value specified | |
1769 | */ | |
1770 | group = ac->ac_g_ex.fe_group; | |
1771 | ||
c9de560d AT |
1772 | for (i = 0; i < EXT4_SB(sb)->s_groups_count; group++, i++) { |
1773 | struct ext4_group_info *grp; | |
1774 | struct ext4_group_desc *desc; | |
1775 | ||
1776 | if (group == EXT4_SB(sb)->s_groups_count) | |
1777 | group = 0; | |
1778 | ||
1779 | /* quick check to skip empty groups */ | |
1780 | grp = ext4_get_group_info(ac->ac_sb, group); | |
1781 | if (grp->bb_free == 0) | |
1782 | continue; | |
1783 | ||
1784 | /* | |
1785 | * if the group is already init we check whether it is | |
1786 | * a good group and if not we don't load the buddy | |
1787 | */ | |
1788 | if (EXT4_MB_GRP_NEED_INIT(grp)) { | |
1789 | /* | |
1790 | * we need full data about the group | |
1791 | * to make a good selection | |
1792 | */ | |
1793 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
1794 | if (err) | |
1795 | goto out; | |
1796 | ext4_mb_release_desc(&e4b); | |
1797 | } | |
1798 | ||
1799 | /* | |
1800 | * If the particular group doesn't satisfy our | |
1801 | * criteria we continue with the next group | |
1802 | */ | |
1803 | if (!ext4_mb_good_group(ac, group, cr)) | |
1804 | continue; | |
1805 | ||
1806 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
1807 | if (err) | |
1808 | goto out; | |
1809 | ||
1810 | ext4_lock_group(sb, group); | |
1811 | if (!ext4_mb_good_group(ac, group, cr)) { | |
1812 | /* someone did allocation from this group */ | |
1813 | ext4_unlock_group(sb, group); | |
1814 | ext4_mb_release_desc(&e4b); | |
1815 | continue; | |
1816 | } | |
1817 | ||
1818 | ac->ac_groups_scanned++; | |
1819 | desc = ext4_get_group_desc(sb, group, NULL); | |
1820 | if (cr == 0 || (desc->bg_flags & | |
1821 | cpu_to_le16(EXT4_BG_BLOCK_UNINIT) && | |
1822 | ac->ac_2order != 0)) | |
1823 | ext4_mb_simple_scan_group(ac, &e4b); | |
1824 | else if (cr == 1 && | |
1825 | ac->ac_g_ex.fe_len == sbi->s_stripe) | |
1826 | ext4_mb_scan_aligned(ac, &e4b); | |
1827 | else | |
1828 | ext4_mb_complex_scan_group(ac, &e4b); | |
1829 | ||
1830 | ext4_unlock_group(sb, group); | |
1831 | ext4_mb_release_desc(&e4b); | |
1832 | ||
1833 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
1834 | break; | |
1835 | } | |
1836 | } | |
1837 | ||
1838 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
1839 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1840 | /* | |
1841 | * We've been searching too long. Let's try to allocate | |
1842 | * the best chunk we've found so far | |
1843 | */ | |
1844 | ||
1845 | ext4_mb_try_best_found(ac, &e4b); | |
1846 | if (ac->ac_status != AC_STATUS_FOUND) { | |
1847 | /* | |
1848 | * Someone more lucky has already allocated it. | |
1849 | * The only thing we can do is just take first | |
1850 | * found block(s) | |
1851 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
1852 | */ | |
1853 | ac->ac_b_ex.fe_group = 0; | |
1854 | ac->ac_b_ex.fe_start = 0; | |
1855 | ac->ac_b_ex.fe_len = 0; | |
1856 | ac->ac_status = AC_STATUS_CONTINUE; | |
1857 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
1858 | cr = 3; | |
1859 | atomic_inc(&sbi->s_mb_lost_chunks); | |
1860 | goto repeat; | |
1861 | } | |
1862 | } | |
1863 | out: | |
1864 | return err; | |
1865 | } | |
1866 | ||
1867 | #ifdef EXT4_MB_HISTORY | |
1868 | struct ext4_mb_proc_session { | |
1869 | struct ext4_mb_history *history; | |
1870 | struct super_block *sb; | |
1871 | int start; | |
1872 | int max; | |
1873 | }; | |
1874 | ||
1875 | static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s, | |
1876 | struct ext4_mb_history *hs, | |
1877 | int first) | |
1878 | { | |
1879 | if (hs == s->history + s->max) | |
1880 | hs = s->history; | |
1881 | if (!first && hs == s->history + s->start) | |
1882 | return NULL; | |
1883 | while (hs->orig.fe_len == 0) { | |
1884 | hs++; | |
1885 | if (hs == s->history + s->max) | |
1886 | hs = s->history; | |
1887 | if (hs == s->history + s->start) | |
1888 | return NULL; | |
1889 | } | |
1890 | return hs; | |
1891 | } | |
1892 | ||
1893 | static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos) | |
1894 | { | |
1895 | struct ext4_mb_proc_session *s = seq->private; | |
1896 | struct ext4_mb_history *hs; | |
1897 | int l = *pos; | |
1898 | ||
1899 | if (l == 0) | |
1900 | return SEQ_START_TOKEN; | |
1901 | hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1); | |
1902 | if (!hs) | |
1903 | return NULL; | |
1904 | while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL); | |
1905 | return hs; | |
1906 | } | |
1907 | ||
1908 | static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v, | |
1909 | loff_t *pos) | |
1910 | { | |
1911 | struct ext4_mb_proc_session *s = seq->private; | |
1912 | struct ext4_mb_history *hs = v; | |
1913 | ||
1914 | ++*pos; | |
1915 | if (v == SEQ_START_TOKEN) | |
1916 | return ext4_mb_history_skip_empty(s, s->history + s->start, 1); | |
1917 | else | |
1918 | return ext4_mb_history_skip_empty(s, ++hs, 0); | |
1919 | } | |
1920 | ||
1921 | static int ext4_mb_seq_history_show(struct seq_file *seq, void *v) | |
1922 | { | |
1923 | char buf[25], buf2[25], buf3[25], *fmt; | |
1924 | struct ext4_mb_history *hs = v; | |
1925 | ||
1926 | if (v == SEQ_START_TOKEN) { | |
1927 | seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s " | |
1928 | "%-5s %-2s %-5s %-5s %-5s %-6s\n", | |
1929 | "pid", "inode", "original", "goal", "result", "found", | |
1930 | "grps", "cr", "flags", "merge", "tail", "broken"); | |
1931 | return 0; | |
1932 | } | |
1933 | ||
1934 | if (hs->op == EXT4_MB_HISTORY_ALLOC) { | |
1935 | fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u " | |
1936 | "%-5u %-5s %-5u %-6u\n"; | |
a9df9a49 | 1937 | sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group, |
c9de560d AT |
1938 | hs->result.fe_start, hs->result.fe_len, |
1939 | hs->result.fe_logical); | |
a9df9a49 | 1940 | sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group, |
c9de560d AT |
1941 | hs->orig.fe_start, hs->orig.fe_len, |
1942 | hs->orig.fe_logical); | |
a9df9a49 | 1943 | sprintf(buf3, "%u/%d/%u@%u", hs->goal.fe_group, |
c9de560d AT |
1944 | hs->goal.fe_start, hs->goal.fe_len, |
1945 | hs->goal.fe_logical); | |
1946 | seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2, | |
1947 | hs->found, hs->groups, hs->cr, hs->flags, | |
1948 | hs->merged ? "M" : "", hs->tail, | |
1949 | hs->buddy ? 1 << hs->buddy : 0); | |
1950 | } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) { | |
1951 | fmt = "%-5u %-8u %-23s %-23s %-23s\n"; | |
a9df9a49 | 1952 | sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group, |
c9de560d AT |
1953 | hs->result.fe_start, hs->result.fe_len, |
1954 | hs->result.fe_logical); | |
a9df9a49 | 1955 | sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group, |
c9de560d AT |
1956 | hs->orig.fe_start, hs->orig.fe_len, |
1957 | hs->orig.fe_logical); | |
1958 | seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2); | |
1959 | } else if (hs->op == EXT4_MB_HISTORY_DISCARD) { | |
a9df9a49 | 1960 | sprintf(buf2, "%u/%d/%u", hs->result.fe_group, |
c9de560d AT |
1961 | hs->result.fe_start, hs->result.fe_len); |
1962 | seq_printf(seq, "%-5u %-8u %-23s discard\n", | |
1963 | hs->pid, hs->ino, buf2); | |
1964 | } else if (hs->op == EXT4_MB_HISTORY_FREE) { | |
a9df9a49 | 1965 | sprintf(buf2, "%u/%d/%u", hs->result.fe_group, |
c9de560d AT |
1966 | hs->result.fe_start, hs->result.fe_len); |
1967 | seq_printf(seq, "%-5u %-8u %-23s free\n", | |
1968 | hs->pid, hs->ino, buf2); | |
1969 | } | |
1970 | return 0; | |
1971 | } | |
1972 | ||
1973 | static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v) | |
1974 | { | |
1975 | } | |
1976 | ||
1977 | static struct seq_operations ext4_mb_seq_history_ops = { | |
1978 | .start = ext4_mb_seq_history_start, | |
1979 | .next = ext4_mb_seq_history_next, | |
1980 | .stop = ext4_mb_seq_history_stop, | |
1981 | .show = ext4_mb_seq_history_show, | |
1982 | }; | |
1983 | ||
1984 | static int ext4_mb_seq_history_open(struct inode *inode, struct file *file) | |
1985 | { | |
1986 | struct super_block *sb = PDE(inode)->data; | |
1987 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1988 | struct ext4_mb_proc_session *s; | |
1989 | int rc; | |
1990 | int size; | |
1991 | ||
74767c5a SF |
1992 | if (unlikely(sbi->s_mb_history == NULL)) |
1993 | return -ENOMEM; | |
c9de560d AT |
1994 | s = kmalloc(sizeof(*s), GFP_KERNEL); |
1995 | if (s == NULL) | |
1996 | return -ENOMEM; | |
1997 | s->sb = sb; | |
1998 | size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max; | |
1999 | s->history = kmalloc(size, GFP_KERNEL); | |
2000 | if (s->history == NULL) { | |
2001 | kfree(s); | |
2002 | return -ENOMEM; | |
2003 | } | |
2004 | ||
2005 | spin_lock(&sbi->s_mb_history_lock); | |
2006 | memcpy(s->history, sbi->s_mb_history, size); | |
2007 | s->max = sbi->s_mb_history_max; | |
2008 | s->start = sbi->s_mb_history_cur % s->max; | |
2009 | spin_unlock(&sbi->s_mb_history_lock); | |
2010 | ||
2011 | rc = seq_open(file, &ext4_mb_seq_history_ops); | |
2012 | if (rc == 0) { | |
2013 | struct seq_file *m = (struct seq_file *)file->private_data; | |
2014 | m->private = s; | |
2015 | } else { | |
2016 | kfree(s->history); | |
2017 | kfree(s); | |
2018 | } | |
2019 | return rc; | |
2020 | ||
2021 | } | |
2022 | ||
2023 | static int ext4_mb_seq_history_release(struct inode *inode, struct file *file) | |
2024 | { | |
2025 | struct seq_file *seq = (struct seq_file *)file->private_data; | |
2026 | struct ext4_mb_proc_session *s = seq->private; | |
2027 | kfree(s->history); | |
2028 | kfree(s); | |
2029 | return seq_release(inode, file); | |
2030 | } | |
2031 | ||
2032 | static ssize_t ext4_mb_seq_history_write(struct file *file, | |
2033 | const char __user *buffer, | |
2034 | size_t count, loff_t *ppos) | |
2035 | { | |
2036 | struct seq_file *seq = (struct seq_file *)file->private_data; | |
2037 | struct ext4_mb_proc_session *s = seq->private; | |
2038 | struct super_block *sb = s->sb; | |
2039 | char str[32]; | |
2040 | int value; | |
2041 | ||
2042 | if (count >= sizeof(str)) { | |
2043 | printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n", | |
2044 | "mb_history", (int)sizeof(str)); | |
2045 | return -EOVERFLOW; | |
2046 | } | |
2047 | ||
2048 | if (copy_from_user(str, buffer, count)) | |
2049 | return -EFAULT; | |
2050 | ||
2051 | value = simple_strtol(str, NULL, 0); | |
2052 | if (value < 0) | |
2053 | return -ERANGE; | |
2054 | EXT4_SB(sb)->s_mb_history_filter = value; | |
2055 | ||
2056 | return count; | |
2057 | } | |
2058 | ||
2059 | static struct file_operations ext4_mb_seq_history_fops = { | |
2060 | .owner = THIS_MODULE, | |
2061 | .open = ext4_mb_seq_history_open, | |
2062 | .read = seq_read, | |
2063 | .write = ext4_mb_seq_history_write, | |
2064 | .llseek = seq_lseek, | |
2065 | .release = ext4_mb_seq_history_release, | |
2066 | }; | |
2067 | ||
2068 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) | |
2069 | { | |
2070 | struct super_block *sb = seq->private; | |
2071 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2072 | ext4_group_t group; | |
2073 | ||
2074 | if (*pos < 0 || *pos >= sbi->s_groups_count) | |
2075 | return NULL; | |
2076 | ||
2077 | group = *pos + 1; | |
a9df9a49 | 2078 | return (void *) ((unsigned long) group); |
c9de560d AT |
2079 | } |
2080 | ||
2081 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2082 | { | |
2083 | struct super_block *sb = seq->private; | |
2084 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2085 | ext4_group_t group; | |
2086 | ||
2087 | ++*pos; | |
2088 | if (*pos < 0 || *pos >= sbi->s_groups_count) | |
2089 | return NULL; | |
2090 | group = *pos + 1; | |
a9df9a49 | 2091 | return (void *) ((unsigned long) group); |
c9de560d AT |
2092 | } |
2093 | ||
2094 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2095 | { | |
2096 | struct super_block *sb = seq->private; | |
a9df9a49 | 2097 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d AT |
2098 | int i; |
2099 | int err; | |
2100 | struct ext4_buddy e4b; | |
2101 | struct sg { | |
2102 | struct ext4_group_info info; | |
2103 | unsigned short counters[16]; | |
2104 | } sg; | |
2105 | ||
2106 | group--; | |
2107 | if (group == 0) | |
2108 | seq_printf(seq, "#%-5s: %-5s %-5s %-5s " | |
2109 | "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " | |
2110 | "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", | |
2111 | "group", "free", "frags", "first", | |
2112 | "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", | |
2113 | "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); | |
2114 | ||
2115 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2116 | sizeof(struct ext4_group_info); | |
2117 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2118 | if (err) { | |
a9df9a49 | 2119 | seq_printf(seq, "#%-5u: I/O error\n", group); |
c9de560d AT |
2120 | return 0; |
2121 | } | |
2122 | ext4_lock_group(sb, group); | |
2123 | memcpy(&sg, ext4_get_group_info(sb, group), i); | |
2124 | ext4_unlock_group(sb, group); | |
2125 | ext4_mb_release_desc(&e4b); | |
2126 | ||
a9df9a49 | 2127 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2128 | sg.info.bb_fragments, sg.info.bb_first_free); |
2129 | for (i = 0; i <= 13; i++) | |
2130 | seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? | |
2131 | sg.info.bb_counters[i] : 0); | |
2132 | seq_printf(seq, " ]\n"); | |
2133 | ||
2134 | return 0; | |
2135 | } | |
2136 | ||
2137 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2138 | { | |
2139 | } | |
2140 | ||
2141 | static struct seq_operations ext4_mb_seq_groups_ops = { | |
2142 | .start = ext4_mb_seq_groups_start, | |
2143 | .next = ext4_mb_seq_groups_next, | |
2144 | .stop = ext4_mb_seq_groups_stop, | |
2145 | .show = ext4_mb_seq_groups_show, | |
2146 | }; | |
2147 | ||
2148 | static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) | |
2149 | { | |
2150 | struct super_block *sb = PDE(inode)->data; | |
2151 | int rc; | |
2152 | ||
2153 | rc = seq_open(file, &ext4_mb_seq_groups_ops); | |
2154 | if (rc == 0) { | |
2155 | struct seq_file *m = (struct seq_file *)file->private_data; | |
2156 | m->private = sb; | |
2157 | } | |
2158 | return rc; | |
2159 | ||
2160 | } | |
2161 | ||
2162 | static struct file_operations ext4_mb_seq_groups_fops = { | |
2163 | .owner = THIS_MODULE, | |
2164 | .open = ext4_mb_seq_groups_open, | |
2165 | .read = seq_read, | |
2166 | .llseek = seq_lseek, | |
2167 | .release = seq_release, | |
2168 | }; | |
2169 | ||
2170 | static void ext4_mb_history_release(struct super_block *sb) | |
2171 | { | |
2172 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2173 | ||
9f6200bb TT |
2174 | if (sbi->s_proc != NULL) { |
2175 | remove_proc_entry("mb_groups", sbi->s_proc); | |
2176 | remove_proc_entry("mb_history", sbi->s_proc); | |
2177 | } | |
c9de560d AT |
2178 | kfree(sbi->s_mb_history); |
2179 | } | |
2180 | ||
2181 | static void ext4_mb_history_init(struct super_block *sb) | |
2182 | { | |
2183 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2184 | int i; | |
2185 | ||
9f6200bb TT |
2186 | if (sbi->s_proc != NULL) { |
2187 | proc_create_data("mb_history", S_IRUGO, sbi->s_proc, | |
46fe74f2 | 2188 | &ext4_mb_seq_history_fops, sb); |
9f6200bb | 2189 | proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, |
46fe74f2 | 2190 | &ext4_mb_seq_groups_fops, sb); |
c9de560d AT |
2191 | } |
2192 | ||
2193 | sbi->s_mb_history_max = 1000; | |
2194 | sbi->s_mb_history_cur = 0; | |
2195 | spin_lock_init(&sbi->s_mb_history_lock); | |
2196 | i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history); | |
74767c5a | 2197 | sbi->s_mb_history = kzalloc(i, GFP_KERNEL); |
c9de560d AT |
2198 | /* if we can't allocate history, then we simple won't use it */ |
2199 | } | |
2200 | ||
4ddfef7b ES |
2201 | static noinline_for_stack void |
2202 | ext4_mb_store_history(struct ext4_allocation_context *ac) | |
c9de560d AT |
2203 | { |
2204 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2205 | struct ext4_mb_history h; | |
2206 | ||
2207 | if (unlikely(sbi->s_mb_history == NULL)) | |
2208 | return; | |
2209 | ||
2210 | if (!(ac->ac_op & sbi->s_mb_history_filter)) | |
2211 | return; | |
2212 | ||
2213 | h.op = ac->ac_op; | |
2214 | h.pid = current->pid; | |
2215 | h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0; | |
2216 | h.orig = ac->ac_o_ex; | |
2217 | h.result = ac->ac_b_ex; | |
2218 | h.flags = ac->ac_flags; | |
2219 | h.found = ac->ac_found; | |
2220 | h.groups = ac->ac_groups_scanned; | |
2221 | h.cr = ac->ac_criteria; | |
2222 | h.tail = ac->ac_tail; | |
2223 | h.buddy = ac->ac_buddy; | |
2224 | h.merged = 0; | |
2225 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) { | |
2226 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
2227 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
2228 | h.merged = 1; | |
2229 | h.goal = ac->ac_g_ex; | |
2230 | h.result = ac->ac_f_ex; | |
2231 | } | |
2232 | ||
2233 | spin_lock(&sbi->s_mb_history_lock); | |
2234 | memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h)); | |
2235 | if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max) | |
2236 | sbi->s_mb_history_cur = 0; | |
2237 | spin_unlock(&sbi->s_mb_history_lock); | |
2238 | } | |
2239 | ||
2240 | #else | |
2241 | #define ext4_mb_history_release(sb) | |
2242 | #define ext4_mb_history_init(sb) | |
2243 | #endif | |
2244 | ||
5f21b0e6 FB |
2245 | |
2246 | /* Create and initialize ext4_group_info data for the given group. */ | |
2247 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, | |
2248 | struct ext4_group_desc *desc) | |
2249 | { | |
2250 | int i, len; | |
2251 | int metalen = 0; | |
2252 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2253 | struct ext4_group_info **meta_group_info; | |
2254 | ||
2255 | /* | |
2256 | * First check if this group is the first of a reserved block. | |
2257 | * If it's true, we have to allocate a new table of pointers | |
2258 | * to ext4_group_info structures | |
2259 | */ | |
2260 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2261 | metalen = sizeof(*meta_group_info) << | |
2262 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2263 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2264 | if (meta_group_info == NULL) { | |
2265 | printk(KERN_ERR "EXT4-fs: can't allocate mem for a " | |
2266 | "buddy group\n"); | |
2267 | goto exit_meta_group_info; | |
2268 | } | |
2269 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = | |
2270 | meta_group_info; | |
2271 | } | |
2272 | ||
2273 | /* | |
2274 | * calculate needed size. if change bb_counters size, | |
2275 | * don't forget about ext4_mb_generate_buddy() | |
2276 | */ | |
2277 | len = offsetof(typeof(**meta_group_info), | |
2278 | bb_counters[sb->s_blocksize_bits + 2]); | |
2279 | ||
2280 | meta_group_info = | |
2281 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; | |
2282 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); | |
2283 | ||
2284 | meta_group_info[i] = kzalloc(len, GFP_KERNEL); | |
2285 | if (meta_group_info[i] == NULL) { | |
2286 | printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n"); | |
2287 | goto exit_group_info; | |
2288 | } | |
2289 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2290 | &(meta_group_info[i]->bb_state)); | |
2291 | ||
2292 | /* | |
2293 | * initialize bb_free to be able to skip | |
2294 | * empty groups without initialization | |
2295 | */ | |
2296 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2297 | meta_group_info[i]->bb_free = | |
2298 | ext4_free_blocks_after_init(sb, group, desc); | |
2299 | } else { | |
2300 | meta_group_info[i]->bb_free = | |
2301 | le16_to_cpu(desc->bg_free_blocks_count); | |
2302 | } | |
2303 | ||
2304 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
c894058d | 2305 | meta_group_info[i]->bb_free_root.rb_node = NULL;; |
5f21b0e6 FB |
2306 | |
2307 | #ifdef DOUBLE_CHECK | |
2308 | { | |
2309 | struct buffer_head *bh; | |
2310 | meta_group_info[i]->bb_bitmap = | |
2311 | kmalloc(sb->s_blocksize, GFP_KERNEL); | |
2312 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); | |
2313 | bh = ext4_read_block_bitmap(sb, group); | |
2314 | BUG_ON(bh == NULL); | |
2315 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, | |
2316 | sb->s_blocksize); | |
2317 | put_bh(bh); | |
2318 | } | |
2319 | #endif | |
2320 | ||
2321 | return 0; | |
2322 | ||
2323 | exit_group_info: | |
2324 | /* If a meta_group_info table has been allocated, release it now */ | |
2325 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) | |
2326 | kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); | |
2327 | exit_meta_group_info: | |
2328 | return -ENOMEM; | |
2329 | } /* ext4_mb_add_groupinfo */ | |
2330 | ||
2331 | /* | |
2332 | * Add a group to the existing groups. | |
2333 | * This function is used for online resize | |
2334 | */ | |
2335 | int ext4_mb_add_more_groupinfo(struct super_block *sb, ext4_group_t group, | |
2336 | struct ext4_group_desc *desc) | |
2337 | { | |
2338 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2339 | struct inode *inode = sbi->s_buddy_cache; | |
2340 | int blocks_per_page; | |
2341 | int block; | |
2342 | int pnum; | |
2343 | struct page *page; | |
2344 | int err; | |
2345 | ||
2346 | /* Add group based on group descriptor*/ | |
2347 | err = ext4_mb_add_groupinfo(sb, group, desc); | |
2348 | if (err) | |
2349 | return err; | |
2350 | ||
2351 | /* | |
2352 | * Cache pages containing dynamic mb_alloc datas (buddy and bitmap | |
2353 | * datas) are set not up to date so that they will be re-initilaized | |
2354 | * during the next call to ext4_mb_load_buddy | |
2355 | */ | |
2356 | ||
2357 | /* Set buddy page as not up to date */ | |
2358 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
2359 | block = group * 2; | |
2360 | pnum = block / blocks_per_page; | |
2361 | page = find_get_page(inode->i_mapping, pnum); | |
2362 | if (page != NULL) { | |
2363 | ClearPageUptodate(page); | |
2364 | page_cache_release(page); | |
2365 | } | |
2366 | ||
2367 | /* Set bitmap page as not up to date */ | |
2368 | block++; | |
2369 | pnum = block / blocks_per_page; | |
2370 | page = find_get_page(inode->i_mapping, pnum); | |
2371 | if (page != NULL) { | |
2372 | ClearPageUptodate(page); | |
2373 | page_cache_release(page); | |
2374 | } | |
2375 | ||
2376 | return 0; | |
2377 | } | |
2378 | ||
2379 | /* | |
2380 | * Update an existing group. | |
2381 | * This function is used for online resize | |
2382 | */ | |
2383 | void ext4_mb_update_group_info(struct ext4_group_info *grp, ext4_grpblk_t add) | |
2384 | { | |
2385 | grp->bb_free += add; | |
2386 | } | |
2387 | ||
c9de560d AT |
2388 | static int ext4_mb_init_backend(struct super_block *sb) |
2389 | { | |
2390 | ext4_group_t i; | |
5f21b0e6 | 2391 | int metalen; |
c9de560d | 2392 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
5f21b0e6 FB |
2393 | struct ext4_super_block *es = sbi->s_es; |
2394 | int num_meta_group_infos; | |
2395 | int num_meta_group_infos_max; | |
2396 | int array_size; | |
c9de560d | 2397 | struct ext4_group_info **meta_group_info; |
5f21b0e6 FB |
2398 | struct ext4_group_desc *desc; |
2399 | ||
2400 | /* This is the number of blocks used by GDT */ | |
2401 | num_meta_group_infos = (sbi->s_groups_count + EXT4_DESC_PER_BLOCK(sb) - | |
2402 | 1) >> EXT4_DESC_PER_BLOCK_BITS(sb); | |
2403 | ||
2404 | /* | |
2405 | * This is the total number of blocks used by GDT including | |
2406 | * the number of reserved blocks for GDT. | |
2407 | * The s_group_info array is allocated with this value | |
2408 | * to allow a clean online resize without a complex | |
2409 | * manipulation of pointer. | |
2410 | * The drawback is the unused memory when no resize | |
2411 | * occurs but it's very low in terms of pages | |
2412 | * (see comments below) | |
2413 | * Need to handle this properly when META_BG resizing is allowed | |
2414 | */ | |
2415 | num_meta_group_infos_max = num_meta_group_infos + | |
2416 | le16_to_cpu(es->s_reserved_gdt_blocks); | |
c9de560d | 2417 | |
5f21b0e6 FB |
2418 | /* |
2419 | * array_size is the size of s_group_info array. We round it | |
2420 | * to the next power of two because this approximation is done | |
2421 | * internally by kmalloc so we can have some more memory | |
2422 | * for free here (e.g. may be used for META_BG resize). | |
2423 | */ | |
2424 | array_size = 1; | |
2425 | while (array_size < sizeof(*sbi->s_group_info) * | |
2426 | num_meta_group_infos_max) | |
2427 | array_size = array_size << 1; | |
c9de560d AT |
2428 | /* An 8TB filesystem with 64-bit pointers requires a 4096 byte |
2429 | * kmalloc. A 128kb malloc should suffice for a 256TB filesystem. | |
2430 | * So a two level scheme suffices for now. */ | |
5f21b0e6 | 2431 | sbi->s_group_info = kmalloc(array_size, GFP_KERNEL); |
c9de560d AT |
2432 | if (sbi->s_group_info == NULL) { |
2433 | printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n"); | |
2434 | return -ENOMEM; | |
2435 | } | |
2436 | sbi->s_buddy_cache = new_inode(sb); | |
2437 | if (sbi->s_buddy_cache == NULL) { | |
2438 | printk(KERN_ERR "EXT4-fs: can't get new inode\n"); | |
2439 | goto err_freesgi; | |
2440 | } | |
2441 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; | |
2442 | ||
2443 | metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb); | |
2444 | for (i = 0; i < num_meta_group_infos; i++) { | |
2445 | if ((i + 1) == num_meta_group_infos) | |
2446 | metalen = sizeof(*meta_group_info) * | |
2447 | (sbi->s_groups_count - | |
2448 | (i << EXT4_DESC_PER_BLOCK_BITS(sb))); | |
2449 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2450 | if (meta_group_info == NULL) { | |
2451 | printk(KERN_ERR "EXT4-fs: can't allocate mem for a " | |
2452 | "buddy group\n"); | |
2453 | goto err_freemeta; | |
2454 | } | |
2455 | sbi->s_group_info[i] = meta_group_info; | |
2456 | } | |
2457 | ||
c9de560d | 2458 | for (i = 0; i < sbi->s_groups_count; i++) { |
c9de560d AT |
2459 | desc = ext4_get_group_desc(sb, i, NULL); |
2460 | if (desc == NULL) { | |
2461 | printk(KERN_ERR | |
a9df9a49 | 2462 | "EXT4-fs: can't read descriptor %u\n", i); |
c9de560d AT |
2463 | goto err_freebuddy; |
2464 | } | |
5f21b0e6 FB |
2465 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2466 | goto err_freebuddy; | |
c9de560d AT |
2467 | } |
2468 | ||
2469 | return 0; | |
2470 | ||
2471 | err_freebuddy: | |
f1fa3342 | 2472 | while (i-- > 0) |
c9de560d | 2473 | kfree(ext4_get_group_info(sb, i)); |
c9de560d AT |
2474 | i = num_meta_group_infos; |
2475 | err_freemeta: | |
f1fa3342 | 2476 | while (i-- > 0) |
c9de560d AT |
2477 | kfree(sbi->s_group_info[i]); |
2478 | iput(sbi->s_buddy_cache); | |
2479 | err_freesgi: | |
2480 | kfree(sbi->s_group_info); | |
2481 | return -ENOMEM; | |
2482 | } | |
2483 | ||
2484 | int ext4_mb_init(struct super_block *sb, int needs_recovery) | |
2485 | { | |
2486 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2487 | unsigned i, j; |
c9de560d AT |
2488 | unsigned offset; |
2489 | unsigned max; | |
74767c5a | 2490 | int ret; |
c9de560d | 2491 | |
c9de560d AT |
2492 | i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short); |
2493 | ||
2494 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2495 | if (sbi->s_mb_offsets == NULL) { | |
c9de560d AT |
2496 | return -ENOMEM; |
2497 | } | |
ff7ef329 YG |
2498 | |
2499 | i = (sb->s_blocksize_bits + 2) * sizeof(unsigned int); | |
c9de560d AT |
2500 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2501 | if (sbi->s_mb_maxs == NULL) { | |
c9de560d AT |
2502 | kfree(sbi->s_mb_maxs); |
2503 | return -ENOMEM; | |
2504 | } | |
2505 | ||
2506 | /* order 0 is regular bitmap */ | |
2507 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2508 | sbi->s_mb_offsets[0] = 0; | |
2509 | ||
2510 | i = 1; | |
2511 | offset = 0; | |
2512 | max = sb->s_blocksize << 2; | |
2513 | do { | |
2514 | sbi->s_mb_offsets[i] = offset; | |
2515 | sbi->s_mb_maxs[i] = max; | |
2516 | offset += 1 << (sb->s_blocksize_bits - i); | |
2517 | max = max >> 1; | |
2518 | i++; | |
2519 | } while (i <= sb->s_blocksize_bits + 1); | |
2520 | ||
2521 | /* init file for buddy data */ | |
74767c5a SF |
2522 | ret = ext4_mb_init_backend(sb); |
2523 | if (ret != 0) { | |
c9de560d AT |
2524 | kfree(sbi->s_mb_offsets); |
2525 | kfree(sbi->s_mb_maxs); | |
74767c5a | 2526 | return ret; |
c9de560d AT |
2527 | } |
2528 | ||
2529 | spin_lock_init(&sbi->s_md_lock); | |
c9de560d AT |
2530 | spin_lock_init(&sbi->s_bal_lock); |
2531 | ||
2532 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2533 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2534 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2535 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2536 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
2537 | sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT; | |
2538 | sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC; | |
2539 | ||
730c213c | 2540 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2541 | if (sbi->s_locality_groups == NULL) { |
c9de560d AT |
2542 | kfree(sbi->s_mb_offsets); |
2543 | kfree(sbi->s_mb_maxs); | |
2544 | return -ENOMEM; | |
2545 | } | |
730c213c | 2546 | for_each_possible_cpu(i) { |
c9de560d | 2547 | struct ext4_locality_group *lg; |
730c213c | 2548 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2549 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2550 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2551 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2552 | spin_lock_init(&lg->lg_prealloc_lock); |
2553 | } | |
2554 | ||
2555 | ext4_mb_init_per_dev_proc(sb); | |
2556 | ext4_mb_history_init(sb); | |
2557 | ||
0390131b FM |
2558 | if (sbi->s_journal) |
2559 | sbi->s_journal->j_commit_callback = release_blocks_on_commit; | |
3e624fc7 | 2560 | |
4776004f | 2561 | printk(KERN_INFO "EXT4-fs: mballoc enabled\n"); |
c9de560d AT |
2562 | return 0; |
2563 | } | |
2564 | ||
2565 | /* need to called with ext4 group lock (ext4_lock_group) */ | |
2566 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) | |
2567 | { | |
2568 | struct ext4_prealloc_space *pa; | |
2569 | struct list_head *cur, *tmp; | |
2570 | int count = 0; | |
2571 | ||
2572 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2573 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2574 | list_del(&pa->pa_group_list); | |
2575 | count++; | |
688f05a0 | 2576 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2577 | } |
2578 | if (count) | |
2579 | mb_debug("mballoc: %u PAs left\n", count); | |
2580 | ||
2581 | } | |
2582 | ||
2583 | int ext4_mb_release(struct super_block *sb) | |
2584 | { | |
2585 | ext4_group_t i; | |
2586 | int num_meta_group_infos; | |
2587 | struct ext4_group_info *grinfo; | |
2588 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2589 | ||
c9de560d AT |
2590 | if (sbi->s_group_info) { |
2591 | for (i = 0; i < sbi->s_groups_count; i++) { | |
2592 | grinfo = ext4_get_group_info(sb, i); | |
2593 | #ifdef DOUBLE_CHECK | |
2594 | kfree(grinfo->bb_bitmap); | |
2595 | #endif | |
2596 | ext4_lock_group(sb, i); | |
2597 | ext4_mb_cleanup_pa(grinfo); | |
2598 | ext4_unlock_group(sb, i); | |
2599 | kfree(grinfo); | |
2600 | } | |
2601 | num_meta_group_infos = (sbi->s_groups_count + | |
2602 | EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
2603 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2604 | for (i = 0; i < num_meta_group_infos; i++) | |
2605 | kfree(sbi->s_group_info[i]); | |
2606 | kfree(sbi->s_group_info); | |
2607 | } | |
2608 | kfree(sbi->s_mb_offsets); | |
2609 | kfree(sbi->s_mb_maxs); | |
2610 | if (sbi->s_buddy_cache) | |
2611 | iput(sbi->s_buddy_cache); | |
2612 | if (sbi->s_mb_stats) { | |
2613 | printk(KERN_INFO | |
2614 | "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n", | |
2615 | atomic_read(&sbi->s_bal_allocated), | |
2616 | atomic_read(&sbi->s_bal_reqs), | |
2617 | atomic_read(&sbi->s_bal_success)); | |
2618 | printk(KERN_INFO | |
2619 | "EXT4-fs: mballoc: %u extents scanned, %u goal hits, " | |
2620 | "%u 2^N hits, %u breaks, %u lost\n", | |
2621 | atomic_read(&sbi->s_bal_ex_scanned), | |
2622 | atomic_read(&sbi->s_bal_goals), | |
2623 | atomic_read(&sbi->s_bal_2orders), | |
2624 | atomic_read(&sbi->s_bal_breaks), | |
2625 | atomic_read(&sbi->s_mb_lost_chunks)); | |
2626 | printk(KERN_INFO | |
2627 | "EXT4-fs: mballoc: %lu generated and it took %Lu\n", | |
2628 | sbi->s_mb_buddies_generated++, | |
2629 | sbi->s_mb_generation_time); | |
2630 | printk(KERN_INFO | |
2631 | "EXT4-fs: mballoc: %u preallocated, %u discarded\n", | |
2632 | atomic_read(&sbi->s_mb_preallocated), | |
2633 | atomic_read(&sbi->s_mb_discarded)); | |
2634 | } | |
2635 | ||
730c213c | 2636 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
2637 | ext4_mb_history_release(sb); |
2638 | ext4_mb_destroy_per_dev_proc(sb); | |
2639 | ||
2640 | return 0; | |
2641 | } | |
2642 | ||
3e624fc7 TT |
2643 | /* |
2644 | * This function is called by the jbd2 layer once the commit has finished, | |
2645 | * so we know we can free the blocks that were released with that commit. | |
2646 | */ | |
2647 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn) | |
c9de560d | 2648 | { |
3e624fc7 | 2649 | struct super_block *sb = journal->j_private; |
c9de560d | 2650 | struct ext4_buddy e4b; |
c894058d | 2651 | struct ext4_group_info *db; |
c894058d AK |
2652 | int err, count = 0, count2 = 0; |
2653 | struct ext4_free_data *entry; | |
8a0aba73 | 2654 | ext4_fsblk_t discard_block; |
3e624fc7 | 2655 | struct list_head *l, *ltmp; |
c9de560d | 2656 | |
3e624fc7 TT |
2657 | list_for_each_safe(l, ltmp, &txn->t_private_list) { |
2658 | entry = list_entry(l, struct ext4_free_data, list); | |
c9de560d | 2659 | |
a9df9a49 | 2660 | mb_debug("gonna free %u blocks in group %u (0x%p):", |
3e624fc7 | 2661 | entry->count, entry->group, entry); |
c9de560d | 2662 | |
c894058d | 2663 | err = ext4_mb_load_buddy(sb, entry->group, &e4b); |
c9de560d AT |
2664 | /* we expect to find existing buddy because it's pinned */ |
2665 | BUG_ON(err != 0); | |
2666 | ||
c894058d | 2667 | db = e4b.bd_info; |
c9de560d | 2668 | /* there are blocks to put in buddy to make them really free */ |
c894058d | 2669 | count += entry->count; |
c9de560d | 2670 | count2++; |
c894058d AK |
2671 | ext4_lock_group(sb, entry->group); |
2672 | /* Take it out of per group rb tree */ | |
2673 | rb_erase(&entry->node, &(db->bb_free_root)); | |
2674 | mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count); | |
2675 | ||
2676 | if (!db->bb_free_root.rb_node) { | |
2677 | /* No more items in the per group rb tree | |
2678 | * balance refcounts from ext4_mb_free_metadata() | |
2679 | */ | |
2680 | page_cache_release(e4b.bd_buddy_page); | |
2681 | page_cache_release(e4b.bd_bitmap_page); | |
c9de560d | 2682 | } |
c894058d | 2683 | ext4_unlock_group(sb, entry->group); |
8a0aba73 TT |
2684 | discard_block = (ext4_fsblk_t) entry->group * EXT4_BLOCKS_PER_GROUP(sb) |
2685 | + entry->start_blk | |
2686 | + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
2687 | trace_mark(ext4_discard_blocks, "dev %s blk %llu count %u", sb->s_id, | |
2688 | (unsigned long long) discard_block, entry->count); | |
2689 | sb_issue_discard(sb, discard_block, entry->count); | |
c9de560d | 2690 | |
c894058d | 2691 | kmem_cache_free(ext4_free_ext_cachep, entry); |
c9de560d | 2692 | ext4_mb_release_desc(&e4b); |
3e624fc7 | 2693 | } |
c9de560d AT |
2694 | |
2695 | mb_debug("freed %u blocks in %u structures\n", count, count2); | |
2696 | } | |
2697 | ||
c9de560d AT |
2698 | #define EXT4_MB_STATS_NAME "stats" |
2699 | #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan" | |
2700 | #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan" | |
2701 | #define EXT4_MB_ORDER2_REQ "order2_req" | |
2702 | #define EXT4_MB_STREAM_REQ "stream_req" | |
2703 | #define EXT4_MB_GROUP_PREALLOC "group_prealloc" | |
2704 | ||
c9de560d AT |
2705 | static int ext4_mb_init_per_dev_proc(struct super_block *sb) |
2706 | { | |
0b09923e | 2707 | #ifdef CONFIG_PROC_FS |
c9de560d AT |
2708 | mode_t mode = S_IFREG | S_IRUGO | S_IWUSR; |
2709 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2710 | struct proc_dir_entry *proc; | |
c9de560d | 2711 | |
9f6200bb | 2712 | if (sbi->s_proc == NULL) |
cfbe7e4f | 2713 | return -EINVAL; |
c9de560d | 2714 | |
5e8814f2 TT |
2715 | EXT4_PROC_HANDLER(EXT4_MB_STATS_NAME, mb_stats); |
2716 | EXT4_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME, mb_max_to_scan); | |
2717 | EXT4_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME, mb_min_to_scan); | |
2718 | EXT4_PROC_HANDLER(EXT4_MB_ORDER2_REQ, mb_order2_reqs); | |
2719 | EXT4_PROC_HANDLER(EXT4_MB_STREAM_REQ, mb_stream_request); | |
2720 | EXT4_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC, mb_group_prealloc); | |
c9de560d AT |
2721 | return 0; |
2722 | ||
2723 | err_out: | |
9f6200bb TT |
2724 | remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_proc); |
2725 | remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_proc); | |
2726 | remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_proc); | |
2727 | remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_proc); | |
2728 | remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_proc); | |
2729 | remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_proc); | |
c9de560d | 2730 | return -ENOMEM; |
0b09923e MK |
2731 | #else |
2732 | return 0; | |
2733 | #endif | |
c9de560d AT |
2734 | } |
2735 | ||
2736 | static int ext4_mb_destroy_per_dev_proc(struct super_block *sb) | |
2737 | { | |
0b09923e | 2738 | #ifdef CONFIG_PROC_FS |
c9de560d | 2739 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d | 2740 | |
9f6200bb | 2741 | if (sbi->s_proc == NULL) |
c9de560d AT |
2742 | return -EINVAL; |
2743 | ||
9f6200bb TT |
2744 | remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_proc); |
2745 | remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_proc); | |
2746 | remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_proc); | |
2747 | remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_proc); | |
2748 | remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_proc); | |
2749 | remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_proc); | |
0b09923e | 2750 | #endif |
c9de560d AT |
2751 | return 0; |
2752 | } | |
2753 | ||
2754 | int __init init_ext4_mballoc(void) | |
2755 | { | |
2756 | ext4_pspace_cachep = | |
2757 | kmem_cache_create("ext4_prealloc_space", | |
2758 | sizeof(struct ext4_prealloc_space), | |
2759 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2760 | if (ext4_pspace_cachep == NULL) | |
2761 | return -ENOMEM; | |
2762 | ||
256bdb49 ES |
2763 | ext4_ac_cachep = |
2764 | kmem_cache_create("ext4_alloc_context", | |
2765 | sizeof(struct ext4_allocation_context), | |
2766 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2767 | if (ext4_ac_cachep == NULL) { | |
2768 | kmem_cache_destroy(ext4_pspace_cachep); | |
2769 | return -ENOMEM; | |
2770 | } | |
c894058d AK |
2771 | |
2772 | ext4_free_ext_cachep = | |
2773 | kmem_cache_create("ext4_free_block_extents", | |
2774 | sizeof(struct ext4_free_data), | |
2775 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2776 | if (ext4_free_ext_cachep == NULL) { | |
2777 | kmem_cache_destroy(ext4_pspace_cachep); | |
2778 | kmem_cache_destroy(ext4_ac_cachep); | |
2779 | return -ENOMEM; | |
2780 | } | |
c9de560d AT |
2781 | return 0; |
2782 | } | |
2783 | ||
2784 | void exit_ext4_mballoc(void) | |
2785 | { | |
2786 | /* XXX: synchronize_rcu(); */ | |
2787 | kmem_cache_destroy(ext4_pspace_cachep); | |
256bdb49 | 2788 | kmem_cache_destroy(ext4_ac_cachep); |
c894058d | 2789 | kmem_cache_destroy(ext4_free_ext_cachep); |
c9de560d AT |
2790 | } |
2791 | ||
2792 | ||
2793 | /* | |
2794 | * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps | |
2795 | * Returns 0 if success or error code | |
2796 | */ | |
4ddfef7b ES |
2797 | static noinline_for_stack int |
2798 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
6bc6e63f | 2799 | handle_t *handle, unsigned long reserv_blks) |
c9de560d AT |
2800 | { |
2801 | struct buffer_head *bitmap_bh = NULL; | |
2802 | struct ext4_super_block *es; | |
2803 | struct ext4_group_desc *gdp; | |
2804 | struct buffer_head *gdp_bh; | |
2805 | struct ext4_sb_info *sbi; | |
2806 | struct super_block *sb; | |
2807 | ext4_fsblk_t block; | |
519deca0 | 2808 | int err, len; |
c9de560d AT |
2809 | |
2810 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2811 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2812 | ||
2813 | sb = ac->ac_sb; | |
2814 | sbi = EXT4_SB(sb); | |
2815 | es = sbi->s_es; | |
2816 | ||
c9de560d AT |
2817 | |
2818 | err = -EIO; | |
574ca174 | 2819 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
c9de560d AT |
2820 | if (!bitmap_bh) |
2821 | goto out_err; | |
2822 | ||
2823 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2824 | if (err) | |
2825 | goto out_err; | |
2826 | ||
2827 | err = -EIO; | |
2828 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
2829 | if (!gdp) | |
2830 | goto out_err; | |
2831 | ||
a9df9a49 | 2832 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
03cddb80 AK |
2833 | gdp->bg_free_blocks_count); |
2834 | ||
c9de560d AT |
2835 | err = ext4_journal_get_write_access(handle, gdp_bh); |
2836 | if (err) | |
2837 | goto out_err; | |
2838 | ||
2839 | block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb) | |
2840 | + ac->ac_b_ex.fe_start | |
2841 | + le32_to_cpu(es->s_first_data_block); | |
2842 | ||
519deca0 AK |
2843 | len = ac->ac_b_ex.fe_len; |
2844 | if (in_range(ext4_block_bitmap(sb, gdp), block, len) || | |
2845 | in_range(ext4_inode_bitmap(sb, gdp), block, len) || | |
2846 | in_range(block, ext4_inode_table(sb, gdp), | |
2847 | EXT4_SB(sb)->s_itb_per_group) || | |
2848 | in_range(block + len - 1, ext4_inode_table(sb, gdp), | |
2849 | EXT4_SB(sb)->s_itb_per_group)) { | |
46e665e9 | 2850 | ext4_error(sb, __func__, |
c9de560d AT |
2851 | "Allocating block in system zone - block = %llu", |
2852 | block); | |
519deca0 AK |
2853 | /* File system mounted not to panic on error |
2854 | * Fix the bitmap and repeat the block allocation | |
2855 | * We leak some of the blocks here. | |
2856 | */ | |
2857 | mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), | |
2858 | bitmap_bh->b_data, ac->ac_b_ex.fe_start, | |
2859 | ac->ac_b_ex.fe_len); | |
0390131b | 2860 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 AK |
2861 | if (!err) |
2862 | err = -EAGAIN; | |
2863 | goto out_err; | |
c9de560d AT |
2864 | } |
2865 | #ifdef AGGRESSIVE_CHECK | |
2866 | { | |
2867 | int i; | |
2868 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
2869 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
2870 | bitmap_bh->b_data)); | |
2871 | } | |
2872 | } | |
2873 | #endif | |
2874 | mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), bitmap_bh->b_data, | |
2875 | ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len); | |
2876 | ||
2877 | spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group)); | |
2878 | if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2879 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
2880 | gdp->bg_free_blocks_count = | |
2881 | cpu_to_le16(ext4_free_blocks_after_init(sb, | |
2882 | ac->ac_b_ex.fe_group, | |
2883 | gdp)); | |
2884 | } | |
e8546d06 | 2885 | le16_add_cpu(&gdp->bg_free_blocks_count, -ac->ac_b_ex.fe_len); |
c9de560d AT |
2886 | gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp); |
2887 | spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group)); | |
6bc6e63f | 2888 | percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 2889 | /* |
6bc6e63f | 2890 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 2891 | */ |
6bc6e63f AK |
2892 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
2893 | /* release all the reserved blocks if non delalloc */ | |
2894 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks); | |
2895 | else | |
2896 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, | |
2897 | ac->ac_b_ex.fe_len); | |
c9de560d | 2898 | |
772cb7c8 JS |
2899 | if (sbi->s_log_groups_per_flex) { |
2900 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
2901 | ac->ac_b_ex.fe_group); | |
2902 | spin_lock(sb_bgl_lock(sbi, flex_group)); | |
2903 | sbi->s_flex_groups[flex_group].free_blocks -= ac->ac_b_ex.fe_len; | |
2904 | spin_unlock(sb_bgl_lock(sbi, flex_group)); | |
2905 | } | |
2906 | ||
0390131b | 2907 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
2908 | if (err) |
2909 | goto out_err; | |
0390131b | 2910 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
2911 | |
2912 | out_err: | |
2913 | sb->s_dirt = 1; | |
42a10add | 2914 | brelse(bitmap_bh); |
c9de560d AT |
2915 | return err; |
2916 | } | |
2917 | ||
2918 | /* | |
2919 | * here we normalize request for locality group | |
2920 | * Group request are normalized to s_strip size if we set the same via mount | |
2921 | * option. If not we set it to s_mb_group_prealloc which can be configured via | |
2922 | * /proc/fs/ext4/<partition>/group_prealloc | |
2923 | * | |
2924 | * XXX: should we try to preallocate more than the group has now? | |
2925 | */ | |
2926 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
2927 | { | |
2928 | struct super_block *sb = ac->ac_sb; | |
2929 | struct ext4_locality_group *lg = ac->ac_lg; | |
2930 | ||
2931 | BUG_ON(lg == NULL); | |
2932 | if (EXT4_SB(sb)->s_stripe) | |
2933 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe; | |
2934 | else | |
2935 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; | |
60bd63d1 | 2936 | mb_debug("#%u: goal %u blocks for locality group\n", |
c9de560d AT |
2937 | current->pid, ac->ac_g_ex.fe_len); |
2938 | } | |
2939 | ||
2940 | /* | |
2941 | * Normalization means making request better in terms of | |
2942 | * size and alignment | |
2943 | */ | |
4ddfef7b ES |
2944 | static noinline_for_stack void |
2945 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
2946 | struct ext4_allocation_request *ar) |
2947 | { | |
2948 | int bsbits, max; | |
2949 | ext4_lblk_t end; | |
c9de560d AT |
2950 | loff_t size, orig_size, start_off; |
2951 | ext4_lblk_t start, orig_start; | |
2952 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); | |
9a0762c5 | 2953 | struct ext4_prealloc_space *pa; |
c9de560d AT |
2954 | |
2955 | /* do normalize only data requests, metadata requests | |
2956 | do not need preallocation */ | |
2957 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
2958 | return; | |
2959 | ||
2960 | /* sometime caller may want exact blocks */ | |
2961 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2962 | return; | |
2963 | ||
2964 | /* caller may indicate that preallocation isn't | |
2965 | * required (it's a tail, for example) */ | |
2966 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
2967 | return; | |
2968 | ||
2969 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
2970 | ext4_mb_normalize_group_request(ac); | |
2971 | return ; | |
2972 | } | |
2973 | ||
2974 | bsbits = ac->ac_sb->s_blocksize_bits; | |
2975 | ||
2976 | /* first, let's learn actual file size | |
2977 | * given current request is allocated */ | |
2978 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; | |
2979 | size = size << bsbits; | |
2980 | if (size < i_size_read(ac->ac_inode)) | |
2981 | size = i_size_read(ac->ac_inode); | |
2982 | ||
1930479c VC |
2983 | /* max size of free chunks */ |
2984 | max = 2 << bsbits; | |
c9de560d | 2985 | |
1930479c VC |
2986 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
2987 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
2988 | |
2989 | /* first, try to predict filesize */ | |
2990 | /* XXX: should this table be tunable? */ | |
2991 | start_off = 0; | |
2992 | if (size <= 16 * 1024) { | |
2993 | size = 16 * 1024; | |
2994 | } else if (size <= 32 * 1024) { | |
2995 | size = 32 * 1024; | |
2996 | } else if (size <= 64 * 1024) { | |
2997 | size = 64 * 1024; | |
2998 | } else if (size <= 128 * 1024) { | |
2999 | size = 128 * 1024; | |
3000 | } else if (size <= 256 * 1024) { | |
3001 | size = 256 * 1024; | |
3002 | } else if (size <= 512 * 1024) { | |
3003 | size = 512 * 1024; | |
3004 | } else if (size <= 1024 * 1024) { | |
3005 | size = 1024 * 1024; | |
1930479c | 3006 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 3007 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
3008 | (21 - bsbits)) << 21; |
3009 | size = 2 * 1024 * 1024; | |
3010 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
3011 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3012 | (22 - bsbits)) << 22; | |
3013 | size = 4 * 1024 * 1024; | |
3014 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 3015 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
3016 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3017 | (23 - bsbits)) << 23; | |
3018 | size = 8 * 1024 * 1024; | |
3019 | } else { | |
3020 | start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; | |
3021 | size = ac->ac_o_ex.fe_len << bsbits; | |
3022 | } | |
3023 | orig_size = size = size >> bsbits; | |
3024 | orig_start = start = start_off >> bsbits; | |
3025 | ||
3026 | /* don't cover already allocated blocks in selected range */ | |
3027 | if (ar->pleft && start <= ar->lleft) { | |
3028 | size -= ar->lleft + 1 - start; | |
3029 | start = ar->lleft + 1; | |
3030 | } | |
3031 | if (ar->pright && start + size - 1 >= ar->lright) | |
3032 | size -= start + size - ar->lright; | |
3033 | ||
3034 | end = start + size; | |
3035 | ||
3036 | /* check we don't cross already preallocated blocks */ | |
3037 | rcu_read_lock(); | |
9a0762c5 | 3038 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3039 | unsigned long pa_end; |
3040 | ||
c9de560d AT |
3041 | if (pa->pa_deleted) |
3042 | continue; | |
3043 | spin_lock(&pa->pa_lock); | |
3044 | if (pa->pa_deleted) { | |
3045 | spin_unlock(&pa->pa_lock); | |
3046 | continue; | |
3047 | } | |
3048 | ||
3049 | pa_end = pa->pa_lstart + pa->pa_len; | |
3050 | ||
3051 | /* PA must not overlap original request */ | |
3052 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
3053 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
3054 | ||
3055 | /* skip PA normalized request doesn't overlap with */ | |
3056 | if (pa->pa_lstart >= end) { | |
3057 | spin_unlock(&pa->pa_lock); | |
3058 | continue; | |
3059 | } | |
3060 | if (pa_end <= start) { | |
3061 | spin_unlock(&pa->pa_lock); | |
3062 | continue; | |
3063 | } | |
3064 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3065 | ||
3066 | if (pa_end <= ac->ac_o_ex.fe_logical) { | |
3067 | BUG_ON(pa_end < start); | |
3068 | start = pa_end; | |
3069 | } | |
3070 | ||
3071 | if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { | |
3072 | BUG_ON(pa->pa_lstart > end); | |
3073 | end = pa->pa_lstart; | |
3074 | } | |
3075 | spin_unlock(&pa->pa_lock); | |
3076 | } | |
3077 | rcu_read_unlock(); | |
3078 | size = end - start; | |
3079 | ||
3080 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3081 | rcu_read_lock(); | |
9a0762c5 | 3082 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d | 3083 | unsigned long pa_end; |
c9de560d AT |
3084 | spin_lock(&pa->pa_lock); |
3085 | if (pa->pa_deleted == 0) { | |
3086 | pa_end = pa->pa_lstart + pa->pa_len; | |
3087 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); | |
3088 | } | |
3089 | spin_unlock(&pa->pa_lock); | |
3090 | } | |
3091 | rcu_read_unlock(); | |
3092 | ||
3093 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3094 | start > ac->ac_o_ex.fe_logical) { | |
3095 | printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n", | |
3096 | (unsigned long) start, (unsigned long) size, | |
3097 | (unsigned long) ac->ac_o_ex.fe_logical); | |
3098 | } | |
3099 | BUG_ON(start + size <= ac->ac_o_ex.fe_logical && | |
3100 | start > ac->ac_o_ex.fe_logical); | |
3101 | BUG_ON(size <= 0 || size >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); | |
3102 | ||
3103 | /* now prepare goal request */ | |
3104 | ||
3105 | /* XXX: is it better to align blocks WRT to logical | |
3106 | * placement or satisfy big request as is */ | |
3107 | ac->ac_g_ex.fe_logical = start; | |
3108 | ac->ac_g_ex.fe_len = size; | |
3109 | ||
3110 | /* define goal start in order to merge */ | |
3111 | if (ar->pright && (ar->lright == (start + size))) { | |
3112 | /* merge to the right */ | |
3113 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3114 | &ac->ac_f_ex.fe_group, | |
3115 | &ac->ac_f_ex.fe_start); | |
3116 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3117 | } | |
3118 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3119 | /* merge to the left */ | |
3120 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3121 | &ac->ac_f_ex.fe_group, | |
3122 | &ac->ac_f_ex.fe_start); | |
3123 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3124 | } | |
3125 | ||
3126 | mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size, | |
3127 | (unsigned) orig_size, (unsigned) start); | |
3128 | } | |
3129 | ||
3130 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3131 | { | |
3132 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3133 | ||
3134 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3135 | atomic_inc(&sbi->s_bal_reqs); | |
3136 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
3137 | if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len) | |
3138 | atomic_inc(&sbi->s_bal_success); | |
3139 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3140 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3141 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3142 | atomic_inc(&sbi->s_bal_goals); | |
3143 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3144 | atomic_inc(&sbi->s_bal_breaks); | |
3145 | } | |
3146 | ||
3147 | ext4_mb_store_history(ac); | |
3148 | } | |
3149 | ||
3150 | /* | |
3151 | * use blocks preallocated to inode | |
3152 | */ | |
3153 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3154 | struct ext4_prealloc_space *pa) | |
3155 | { | |
3156 | ext4_fsblk_t start; | |
3157 | ext4_fsblk_t end; | |
3158 | int len; | |
3159 | ||
3160 | /* found preallocated blocks, use them */ | |
3161 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
3162 | end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len); | |
3163 | len = end - start; | |
3164 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, | |
3165 | &ac->ac_b_ex.fe_start); | |
3166 | ac->ac_b_ex.fe_len = len; | |
3167 | ac->ac_status = AC_STATUS_FOUND; | |
3168 | ac->ac_pa = pa; | |
3169 | ||
3170 | BUG_ON(start < pa->pa_pstart); | |
3171 | BUG_ON(start + len > pa->pa_pstart + pa->pa_len); | |
3172 | BUG_ON(pa->pa_free < len); | |
3173 | pa->pa_free -= len; | |
3174 | ||
60bd63d1 | 3175 | mb_debug("use %llu/%u from inode pa %p\n", start, len, pa); |
c9de560d AT |
3176 | } |
3177 | ||
3178 | /* | |
3179 | * use blocks preallocated to locality group | |
3180 | */ | |
3181 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3182 | struct ext4_prealloc_space *pa) | |
3183 | { | |
03cddb80 | 3184 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3185 | |
c9de560d AT |
3186 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3187 | &ac->ac_b_ex.fe_group, | |
3188 | &ac->ac_b_ex.fe_start); | |
3189 | ac->ac_b_ex.fe_len = len; | |
3190 | ac->ac_status = AC_STATUS_FOUND; | |
3191 | ac->ac_pa = pa; | |
3192 | ||
3193 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3194 | * possible race when the group is being loaded concurrently |
c9de560d | 3195 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3196 | * in on-disk bitmap -- see ext4_mb_release_context() |
3197 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d AT |
3198 | */ |
3199 | mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); | |
3200 | } | |
3201 | ||
5e745b04 AK |
3202 | /* |
3203 | * Return the prealloc space that have minimal distance | |
3204 | * from the goal block. @cpa is the prealloc | |
3205 | * space that is having currently known minimal distance | |
3206 | * from the goal block. | |
3207 | */ | |
3208 | static struct ext4_prealloc_space * | |
3209 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3210 | struct ext4_prealloc_space *pa, | |
3211 | struct ext4_prealloc_space *cpa) | |
3212 | { | |
3213 | ext4_fsblk_t cur_distance, new_distance; | |
3214 | ||
3215 | if (cpa == NULL) { | |
3216 | atomic_inc(&pa->pa_count); | |
3217 | return pa; | |
3218 | } | |
3219 | cur_distance = abs(goal_block - cpa->pa_pstart); | |
3220 | new_distance = abs(goal_block - pa->pa_pstart); | |
3221 | ||
3222 | if (cur_distance < new_distance) | |
3223 | return cpa; | |
3224 | ||
3225 | /* drop the previous reference */ | |
3226 | atomic_dec(&cpa->pa_count); | |
3227 | atomic_inc(&pa->pa_count); | |
3228 | return pa; | |
3229 | } | |
3230 | ||
c9de560d AT |
3231 | /* |
3232 | * search goal blocks in preallocated space | |
3233 | */ | |
4ddfef7b ES |
3234 | static noinline_for_stack int |
3235 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
c9de560d | 3236 | { |
6be2ded1 | 3237 | int order, i; |
c9de560d AT |
3238 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3239 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3240 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3241 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3242 | |
3243 | /* only data can be preallocated */ | |
3244 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3245 | return 0; | |
3246 | ||
3247 | /* first, try per-file preallocation */ | |
3248 | rcu_read_lock(); | |
9a0762c5 | 3249 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3250 | |
3251 | /* all fields in this condition don't change, | |
3252 | * so we can skip locking for them */ | |
3253 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
3254 | ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len) | |
3255 | continue; | |
3256 | ||
3257 | /* found preallocated blocks, use them */ | |
3258 | spin_lock(&pa->pa_lock); | |
3259 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3260 | atomic_inc(&pa->pa_count); | |
3261 | ext4_mb_use_inode_pa(ac, pa); | |
3262 | spin_unlock(&pa->pa_lock); | |
3263 | ac->ac_criteria = 10; | |
3264 | rcu_read_unlock(); | |
3265 | return 1; | |
3266 | } | |
3267 | spin_unlock(&pa->pa_lock); | |
3268 | } | |
3269 | rcu_read_unlock(); | |
3270 | ||
3271 | /* can we use group allocation? */ | |
3272 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3273 | return 0; | |
3274 | ||
3275 | /* inode may have no locality group for some reason */ | |
3276 | lg = ac->ac_lg; | |
3277 | if (lg == NULL) | |
3278 | return 0; | |
6be2ded1 AK |
3279 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3280 | if (order > PREALLOC_TB_SIZE - 1) | |
3281 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3282 | order = PREALLOC_TB_SIZE - 1; | |
3283 | ||
5e745b04 AK |
3284 | goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) + |
3285 | ac->ac_g_ex.fe_start + | |
3286 | le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block); | |
3287 | /* | |
3288 | * search for the prealloc space that is having | |
3289 | * minimal distance from the goal block. | |
3290 | */ | |
6be2ded1 AK |
3291 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3292 | rcu_read_lock(); | |
3293 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3294 | pa_inode_list) { | |
3295 | spin_lock(&pa->pa_lock); | |
3296 | if (pa->pa_deleted == 0 && | |
3297 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3298 | |
3299 | cpa = ext4_mb_check_group_pa(goal_block, | |
3300 | pa, cpa); | |
6be2ded1 | 3301 | } |
c9de560d | 3302 | spin_unlock(&pa->pa_lock); |
c9de560d | 3303 | } |
6be2ded1 | 3304 | rcu_read_unlock(); |
c9de560d | 3305 | } |
5e745b04 AK |
3306 | if (cpa) { |
3307 | ext4_mb_use_group_pa(ac, cpa); | |
3308 | ac->ac_criteria = 20; | |
3309 | return 1; | |
3310 | } | |
c9de560d AT |
3311 | return 0; |
3312 | } | |
3313 | ||
3314 | /* | |
3315 | * the function goes through all preallocation in this group and marks them | |
3316 | * used in in-core bitmap. buddy must be generated from this bitmap | |
3317 | * Need to be called with ext4 group lock (ext4_lock_group) | |
3318 | */ | |
3319 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
3320 | ext4_group_t group) | |
3321 | { | |
3322 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3323 | struct ext4_prealloc_space *pa; | |
3324 | struct list_head *cur; | |
3325 | ext4_group_t groupnr; | |
3326 | ext4_grpblk_t start; | |
3327 | int preallocated = 0; | |
3328 | int count = 0; | |
3329 | int len; | |
3330 | ||
3331 | /* all form of preallocation discards first load group, | |
3332 | * so the only competing code is preallocation use. | |
3333 | * we don't need any locking here | |
3334 | * notice we do NOT ignore preallocations with pa_deleted | |
3335 | * otherwise we could leave used blocks available for | |
3336 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3337 | * is dropping preallocation | |
3338 | */ | |
3339 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3340 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3341 | spin_lock(&pa->pa_lock); | |
3342 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3343 | &groupnr, &start); | |
3344 | len = pa->pa_len; | |
3345 | spin_unlock(&pa->pa_lock); | |
3346 | if (unlikely(len == 0)) | |
3347 | continue; | |
3348 | BUG_ON(groupnr != group); | |
3349 | mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group), | |
3350 | bitmap, start, len); | |
3351 | preallocated += len; | |
3352 | count++; | |
3353 | } | |
a9df9a49 | 3354 | mb_debug("prellocated %u for group %u\n", preallocated, group); |
c9de560d AT |
3355 | } |
3356 | ||
3357 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3358 | { | |
3359 | struct ext4_prealloc_space *pa; | |
3360 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
3361 | kmem_cache_free(ext4_pspace_cachep, pa); | |
3362 | } | |
3363 | ||
3364 | /* | |
3365 | * drops a reference to preallocated space descriptor | |
3366 | * if this was the last reference and the space is consumed | |
3367 | */ | |
3368 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3369 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3370 | { | |
a9df9a49 | 3371 | ext4_group_t grp; |
c9de560d AT |
3372 | |
3373 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) | |
3374 | return; | |
3375 | ||
3376 | /* in this short window concurrent discard can set pa_deleted */ | |
3377 | spin_lock(&pa->pa_lock); | |
3378 | if (pa->pa_deleted == 1) { | |
3379 | spin_unlock(&pa->pa_lock); | |
3380 | return; | |
3381 | } | |
3382 | ||
3383 | pa->pa_deleted = 1; | |
3384 | spin_unlock(&pa->pa_lock); | |
3385 | ||
3386 | /* -1 is to protect from crossing allocation group */ | |
3387 | ext4_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL); | |
3388 | ||
3389 | /* | |
3390 | * possible race: | |
3391 | * | |
3392 | * P1 (buddy init) P2 (regular allocation) | |
3393 | * find block B in PA | |
3394 | * copy on-disk bitmap to buddy | |
3395 | * mark B in on-disk bitmap | |
3396 | * drop PA from group | |
3397 | * mark all PAs in buddy | |
3398 | * | |
3399 | * thus, P1 initializes buddy with B available. to prevent this | |
3400 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3401 | * against that pair | |
3402 | */ | |
3403 | ext4_lock_group(sb, grp); | |
3404 | list_del(&pa->pa_group_list); | |
3405 | ext4_unlock_group(sb, grp); | |
3406 | ||
3407 | spin_lock(pa->pa_obj_lock); | |
3408 | list_del_rcu(&pa->pa_inode_list); | |
3409 | spin_unlock(pa->pa_obj_lock); | |
3410 | ||
3411 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3412 | } | |
3413 | ||
3414 | /* | |
3415 | * creates new preallocated space for given inode | |
3416 | */ | |
4ddfef7b ES |
3417 | static noinline_for_stack int |
3418 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3419 | { |
3420 | struct super_block *sb = ac->ac_sb; | |
3421 | struct ext4_prealloc_space *pa; | |
3422 | struct ext4_group_info *grp; | |
3423 | struct ext4_inode_info *ei; | |
3424 | ||
3425 | /* preallocate only when found space is larger then requested */ | |
3426 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3427 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3428 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3429 | ||
3430 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3431 | if (pa == NULL) | |
3432 | return -ENOMEM; | |
3433 | ||
3434 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3435 | int winl; | |
3436 | int wins; | |
3437 | int win; | |
3438 | int offs; | |
3439 | ||
3440 | /* we can't allocate as much as normalizer wants. | |
3441 | * so, found space must get proper lstart | |
3442 | * to cover original request */ | |
3443 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3444 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3445 | ||
3446 | /* we're limited by original request in that | |
3447 | * logical block must be covered any way | |
3448 | * winl is window we can move our chunk within */ | |
3449 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3450 | ||
3451 | /* also, we should cover whole original request */ | |
3452 | wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len; | |
3453 | ||
3454 | /* the smallest one defines real window */ | |
3455 | win = min(winl, wins); | |
3456 | ||
3457 | offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len; | |
3458 | if (offs && offs < win) | |
3459 | win = offs; | |
3460 | ||
3461 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win; | |
3462 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); | |
3463 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3464 | } | |
3465 | ||
3466 | /* preallocation can change ac_b_ex, thus we store actually | |
3467 | * allocated blocks for history */ | |
3468 | ac->ac_f_ex = ac->ac_b_ex; | |
3469 | ||
3470 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3471 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3472 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3473 | pa->pa_free = pa->pa_len; | |
3474 | atomic_set(&pa->pa_count, 1); | |
3475 | spin_lock_init(&pa->pa_lock); | |
3476 | pa->pa_deleted = 0; | |
3477 | pa->pa_linear = 0; | |
3478 | ||
3479 | mb_debug("new inode pa %p: %llu/%u for %u\n", pa, | |
3480 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); | |
3481 | ||
3482 | ext4_mb_use_inode_pa(ac, pa); | |
3483 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3484 | ||
3485 | ei = EXT4_I(ac->ac_inode); | |
3486 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3487 | ||
3488 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3489 | pa->pa_inode = ac->ac_inode; | |
3490 | ||
3491 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3492 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3493 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3494 | ||
3495 | spin_lock(pa->pa_obj_lock); | |
3496 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3497 | spin_unlock(pa->pa_obj_lock); | |
3498 | ||
3499 | return 0; | |
3500 | } | |
3501 | ||
3502 | /* | |
3503 | * creates new preallocated space for locality group inodes belongs to | |
3504 | */ | |
4ddfef7b ES |
3505 | static noinline_for_stack int |
3506 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3507 | { |
3508 | struct super_block *sb = ac->ac_sb; | |
3509 | struct ext4_locality_group *lg; | |
3510 | struct ext4_prealloc_space *pa; | |
3511 | struct ext4_group_info *grp; | |
3512 | ||
3513 | /* preallocate only when found space is larger then requested */ | |
3514 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3515 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3516 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3517 | ||
3518 | BUG_ON(ext4_pspace_cachep == NULL); | |
3519 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3520 | if (pa == NULL) | |
3521 | return -ENOMEM; | |
3522 | ||
3523 | /* preallocation can change ac_b_ex, thus we store actually | |
3524 | * allocated blocks for history */ | |
3525 | ac->ac_f_ex = ac->ac_b_ex; | |
3526 | ||
3527 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3528 | pa->pa_lstart = pa->pa_pstart; | |
3529 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3530 | pa->pa_free = pa->pa_len; | |
3531 | atomic_set(&pa->pa_count, 1); | |
3532 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3533 | INIT_LIST_HEAD(&pa->pa_inode_list); |
c9de560d AT |
3534 | pa->pa_deleted = 0; |
3535 | pa->pa_linear = 1; | |
3536 | ||
3537 | mb_debug("new group pa %p: %llu/%u for %u\n", pa, | |
3538 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); | |
3539 | ||
3540 | ext4_mb_use_group_pa(ac, pa); | |
3541 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3542 | ||
3543 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3544 | lg = ac->ac_lg; | |
3545 | BUG_ON(lg == NULL); | |
3546 | ||
3547 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3548 | pa->pa_inode = NULL; | |
3549 | ||
3550 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3551 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3552 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3553 | ||
6be2ded1 AK |
3554 | /* |
3555 | * We will later add the new pa to the right bucket | |
3556 | * after updating the pa_free in ext4_mb_release_context | |
3557 | */ | |
c9de560d AT |
3558 | return 0; |
3559 | } | |
3560 | ||
3561 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3562 | { | |
3563 | int err; | |
3564 | ||
3565 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3566 | err = ext4_mb_new_group_pa(ac); | |
3567 | else | |
3568 | err = ext4_mb_new_inode_pa(ac); | |
3569 | return err; | |
3570 | } | |
3571 | ||
3572 | /* | |
3573 | * finds all unused blocks in on-disk bitmap, frees them in | |
3574 | * in-core bitmap and buddy. | |
3575 | * @pa must be unlinked from inode and group lists, so that | |
3576 | * nobody else can find/use it. | |
3577 | * the caller MUST hold group/inode locks. | |
3578 | * TODO: optimize the case when there are no in-core structures yet | |
3579 | */ | |
4ddfef7b ES |
3580 | static noinline_for_stack int |
3581 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
c83617db AK |
3582 | struct ext4_prealloc_space *pa, |
3583 | struct ext4_allocation_context *ac) | |
c9de560d | 3584 | { |
c9de560d AT |
3585 | struct super_block *sb = e4b->bd_sb; |
3586 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3587 | unsigned long end; | |
3588 | unsigned long next; | |
3589 | ext4_group_t group; | |
3590 | ext4_grpblk_t bit; | |
3591 | sector_t start; | |
3592 | int err = 0; | |
3593 | int free = 0; | |
3594 | ||
3595 | BUG_ON(pa->pa_deleted == 0); | |
3596 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3597 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3598 | end = bit + pa->pa_len; | |
3599 | ||
256bdb49 ES |
3600 | if (ac) { |
3601 | ac->ac_sb = sb; | |
3602 | ac->ac_inode = pa->pa_inode; | |
3603 | ac->ac_op = EXT4_MB_HISTORY_DISCARD; | |
3604 | } | |
c9de560d AT |
3605 | |
3606 | while (bit < end) { | |
ffad0a44 | 3607 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3608 | if (bit >= end) |
3609 | break; | |
ffad0a44 | 3610 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3611 | start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit + |
3612 | le32_to_cpu(sbi->s_es->s_first_data_block); | |
3613 | mb_debug(" free preallocated %u/%u in group %u\n", | |
3614 | (unsigned) start, (unsigned) next - bit, | |
3615 | (unsigned) group); | |
3616 | free += next - bit; | |
3617 | ||
256bdb49 ES |
3618 | if (ac) { |
3619 | ac->ac_b_ex.fe_group = group; | |
3620 | ac->ac_b_ex.fe_start = bit; | |
3621 | ac->ac_b_ex.fe_len = next - bit; | |
3622 | ac->ac_b_ex.fe_logical = 0; | |
3623 | ext4_mb_store_history(ac); | |
3624 | } | |
c9de560d AT |
3625 | |
3626 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); | |
3627 | bit = next + 1; | |
3628 | } | |
3629 | if (free != pa->pa_free) { | |
26346ff6 | 3630 | printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n", |
c9de560d AT |
3631 | pa, (unsigned long) pa->pa_lstart, |
3632 | (unsigned long) pa->pa_pstart, | |
3633 | (unsigned long) pa->pa_len); | |
fde4d95a | 3634 | ext4_error(sb, __func__, "free %u, pa_free %u", |
26346ff6 | 3635 | free, pa->pa_free); |
e56eb659 AK |
3636 | /* |
3637 | * pa is already deleted so we use the value obtained | |
3638 | * from the bitmap and continue. | |
3639 | */ | |
c9de560d | 3640 | } |
c9de560d AT |
3641 | atomic_add(free, &sbi->s_mb_discarded); |
3642 | ||
3643 | return err; | |
3644 | } | |
3645 | ||
4ddfef7b ES |
3646 | static noinline_for_stack int |
3647 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
c83617db AK |
3648 | struct ext4_prealloc_space *pa, |
3649 | struct ext4_allocation_context *ac) | |
c9de560d | 3650 | { |
c9de560d AT |
3651 | struct super_block *sb = e4b->bd_sb; |
3652 | ext4_group_t group; | |
3653 | ext4_grpblk_t bit; | |
3654 | ||
256bdb49 ES |
3655 | if (ac) |
3656 | ac->ac_op = EXT4_MB_HISTORY_DISCARD; | |
c9de560d AT |
3657 | |
3658 | BUG_ON(pa->pa_deleted == 0); | |
3659 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3660 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3661 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3662 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3663 | ||
256bdb49 ES |
3664 | if (ac) { |
3665 | ac->ac_sb = sb; | |
3666 | ac->ac_inode = NULL; | |
3667 | ac->ac_b_ex.fe_group = group; | |
3668 | ac->ac_b_ex.fe_start = bit; | |
3669 | ac->ac_b_ex.fe_len = pa->pa_len; | |
3670 | ac->ac_b_ex.fe_logical = 0; | |
3671 | ext4_mb_store_history(ac); | |
256bdb49 | 3672 | } |
c9de560d AT |
3673 | |
3674 | return 0; | |
3675 | } | |
3676 | ||
3677 | /* | |
3678 | * releases all preallocations in given group | |
3679 | * | |
3680 | * first, we need to decide discard policy: | |
3681 | * - when do we discard | |
3682 | * 1) ENOSPC | |
3683 | * - how many do we discard | |
3684 | * 1) how many requested | |
3685 | */ | |
4ddfef7b ES |
3686 | static noinline_for_stack int |
3687 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3688 | ext4_group_t group, int needed) |
3689 | { | |
3690 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3691 | struct buffer_head *bitmap_bh = NULL; | |
3692 | struct ext4_prealloc_space *pa, *tmp; | |
c83617db | 3693 | struct ext4_allocation_context *ac; |
c9de560d AT |
3694 | struct list_head list; |
3695 | struct ext4_buddy e4b; | |
3696 | int err; | |
3697 | int busy = 0; | |
3698 | int free = 0; | |
3699 | ||
a9df9a49 | 3700 | mb_debug("discard preallocation for group %u\n", group); |
c9de560d AT |
3701 | |
3702 | if (list_empty(&grp->bb_prealloc_list)) | |
3703 | return 0; | |
3704 | ||
574ca174 | 3705 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3706 | if (bitmap_bh == NULL) { |
ce89f46c | 3707 | ext4_error(sb, __func__, "Error in reading block " |
a9df9a49 | 3708 | "bitmap for %u", group); |
ce89f46c | 3709 | return 0; |
c9de560d AT |
3710 | } |
3711 | ||
3712 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c AK |
3713 | if (err) { |
3714 | ext4_error(sb, __func__, "Error in loading buddy " | |
a9df9a49 | 3715 | "information for %u", group); |
ce89f46c AK |
3716 | put_bh(bitmap_bh); |
3717 | return 0; | |
3718 | } | |
c9de560d AT |
3719 | |
3720 | if (needed == 0) | |
3721 | needed = EXT4_BLOCKS_PER_GROUP(sb) + 1; | |
3722 | ||
c9de560d | 3723 | INIT_LIST_HEAD(&list); |
c83617db | 3724 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
c9de560d AT |
3725 | repeat: |
3726 | ext4_lock_group(sb, group); | |
3727 | list_for_each_entry_safe(pa, tmp, | |
3728 | &grp->bb_prealloc_list, pa_group_list) { | |
3729 | spin_lock(&pa->pa_lock); | |
3730 | if (atomic_read(&pa->pa_count)) { | |
3731 | spin_unlock(&pa->pa_lock); | |
3732 | busy = 1; | |
3733 | continue; | |
3734 | } | |
3735 | if (pa->pa_deleted) { | |
3736 | spin_unlock(&pa->pa_lock); | |
3737 | continue; | |
3738 | } | |
3739 | ||
3740 | /* seems this one can be freed ... */ | |
3741 | pa->pa_deleted = 1; | |
3742 | ||
3743 | /* we can trust pa_free ... */ | |
3744 | free += pa->pa_free; | |
3745 | ||
3746 | spin_unlock(&pa->pa_lock); | |
3747 | ||
3748 | list_del(&pa->pa_group_list); | |
3749 | list_add(&pa->u.pa_tmp_list, &list); | |
3750 | } | |
3751 | ||
3752 | /* if we still need more blocks and some PAs were used, try again */ | |
3753 | if (free < needed && busy) { | |
3754 | busy = 0; | |
3755 | ext4_unlock_group(sb, group); | |
3756 | /* | |
3757 | * Yield the CPU here so that we don't get soft lockup | |
3758 | * in non preempt case. | |
3759 | */ | |
3760 | yield(); | |
3761 | goto repeat; | |
3762 | } | |
3763 | ||
3764 | /* found anything to free? */ | |
3765 | if (list_empty(&list)) { | |
3766 | BUG_ON(free != 0); | |
3767 | goto out; | |
3768 | } | |
3769 | ||
3770 | /* now free all selected PAs */ | |
3771 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3772 | ||
3773 | /* remove from object (inode or locality group) */ | |
3774 | spin_lock(pa->pa_obj_lock); | |
3775 | list_del_rcu(&pa->pa_inode_list); | |
3776 | spin_unlock(pa->pa_obj_lock); | |
3777 | ||
3778 | if (pa->pa_linear) | |
c83617db | 3779 | ext4_mb_release_group_pa(&e4b, pa, ac); |
c9de560d | 3780 | else |
c83617db | 3781 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); |
c9de560d AT |
3782 | |
3783 | list_del(&pa->u.pa_tmp_list); | |
3784 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3785 | } | |
3786 | ||
3787 | out: | |
3788 | ext4_unlock_group(sb, group); | |
c83617db AK |
3789 | if (ac) |
3790 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
3791 | ext4_mb_release_desc(&e4b); |
3792 | put_bh(bitmap_bh); | |
3793 | return free; | |
3794 | } | |
3795 | ||
3796 | /* | |
3797 | * releases all non-used preallocated blocks for given inode | |
3798 | * | |
3799 | * It's important to discard preallocations under i_data_sem | |
3800 | * We don't want another block to be served from the prealloc | |
3801 | * space when we are discarding the inode prealloc space. | |
3802 | * | |
3803 | * FIXME!! Make sure it is valid at all the call sites | |
3804 | */ | |
c2ea3fde | 3805 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
3806 | { |
3807 | struct ext4_inode_info *ei = EXT4_I(inode); | |
3808 | struct super_block *sb = inode->i_sb; | |
3809 | struct buffer_head *bitmap_bh = NULL; | |
3810 | struct ext4_prealloc_space *pa, *tmp; | |
c83617db | 3811 | struct ext4_allocation_context *ac; |
c9de560d AT |
3812 | ext4_group_t group = 0; |
3813 | struct list_head list; | |
3814 | struct ext4_buddy e4b; | |
3815 | int err; | |
3816 | ||
c2ea3fde | 3817 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
3818 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
3819 | return; | |
3820 | } | |
3821 | ||
3822 | mb_debug("discard preallocation for inode %lu\n", inode->i_ino); | |
3823 | ||
3824 | INIT_LIST_HEAD(&list); | |
3825 | ||
c83617db | 3826 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
c9de560d AT |
3827 | repeat: |
3828 | /* first, collect all pa's in the inode */ | |
3829 | spin_lock(&ei->i_prealloc_lock); | |
3830 | while (!list_empty(&ei->i_prealloc_list)) { | |
3831 | pa = list_entry(ei->i_prealloc_list.next, | |
3832 | struct ext4_prealloc_space, pa_inode_list); | |
3833 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
3834 | spin_lock(&pa->pa_lock); | |
3835 | if (atomic_read(&pa->pa_count)) { | |
3836 | /* this shouldn't happen often - nobody should | |
3837 | * use preallocation while we're discarding it */ | |
3838 | spin_unlock(&pa->pa_lock); | |
3839 | spin_unlock(&ei->i_prealloc_lock); | |
3840 | printk(KERN_ERR "uh-oh! used pa while discarding\n"); | |
3841 | WARN_ON(1); | |
3842 | schedule_timeout_uninterruptible(HZ); | |
3843 | goto repeat; | |
3844 | ||
3845 | } | |
3846 | if (pa->pa_deleted == 0) { | |
3847 | pa->pa_deleted = 1; | |
3848 | spin_unlock(&pa->pa_lock); | |
3849 | list_del_rcu(&pa->pa_inode_list); | |
3850 | list_add(&pa->u.pa_tmp_list, &list); | |
3851 | continue; | |
3852 | } | |
3853 | ||
3854 | /* someone is deleting pa right now */ | |
3855 | spin_unlock(&pa->pa_lock); | |
3856 | spin_unlock(&ei->i_prealloc_lock); | |
3857 | ||
3858 | /* we have to wait here because pa_deleted | |
3859 | * doesn't mean pa is already unlinked from | |
3860 | * the list. as we might be called from | |
3861 | * ->clear_inode() the inode will get freed | |
3862 | * and concurrent thread which is unlinking | |
3863 | * pa from inode's list may access already | |
3864 | * freed memory, bad-bad-bad */ | |
3865 | ||
3866 | /* XXX: if this happens too often, we can | |
3867 | * add a flag to force wait only in case | |
3868 | * of ->clear_inode(), but not in case of | |
3869 | * regular truncate */ | |
3870 | schedule_timeout_uninterruptible(HZ); | |
3871 | goto repeat; | |
3872 | } | |
3873 | spin_unlock(&ei->i_prealloc_lock); | |
3874 | ||
3875 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3876 | BUG_ON(pa->pa_linear != 0); | |
3877 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); | |
3878 | ||
3879 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c AK |
3880 | if (err) { |
3881 | ext4_error(sb, __func__, "Error in loading buddy " | |
a9df9a49 | 3882 | "information for %u", group); |
ce89f46c AK |
3883 | continue; |
3884 | } | |
c9de560d | 3885 | |
574ca174 | 3886 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3887 | if (bitmap_bh == NULL) { |
ce89f46c | 3888 | ext4_error(sb, __func__, "Error in reading block " |
a9df9a49 | 3889 | "bitmap for %u", group); |
c9de560d | 3890 | ext4_mb_release_desc(&e4b); |
ce89f46c | 3891 | continue; |
c9de560d AT |
3892 | } |
3893 | ||
3894 | ext4_lock_group(sb, group); | |
3895 | list_del(&pa->pa_group_list); | |
c83617db | 3896 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); |
c9de560d AT |
3897 | ext4_unlock_group(sb, group); |
3898 | ||
3899 | ext4_mb_release_desc(&e4b); | |
3900 | put_bh(bitmap_bh); | |
3901 | ||
3902 | list_del(&pa->u.pa_tmp_list); | |
3903 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3904 | } | |
c83617db AK |
3905 | if (ac) |
3906 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
3907 | } |
3908 | ||
3909 | /* | |
3910 | * finds all preallocated spaces and return blocks being freed to them | |
3911 | * if preallocated space becomes full (no block is used from the space) | |
3912 | * then the function frees space in buddy | |
3913 | * XXX: at the moment, truncate (which is the only way to free blocks) | |
3914 | * discards all preallocations | |
3915 | */ | |
3916 | static void ext4_mb_return_to_preallocation(struct inode *inode, | |
3917 | struct ext4_buddy *e4b, | |
3918 | sector_t block, int count) | |
3919 | { | |
3920 | BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list)); | |
3921 | } | |
3922 | #ifdef MB_DEBUG | |
3923 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
3924 | { | |
3925 | struct super_block *sb = ac->ac_sb; | |
3926 | ext4_group_t i; | |
3927 | ||
3928 | printk(KERN_ERR "EXT4-fs: Can't allocate:" | |
3929 | " Allocation context details:\n"); | |
3930 | printk(KERN_ERR "EXT4-fs: status %d flags %d\n", | |
3931 | ac->ac_status, ac->ac_flags); | |
3932 | printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, " | |
3933 | "best %lu/%lu/%lu@%lu cr %d\n", | |
3934 | (unsigned long)ac->ac_o_ex.fe_group, | |
3935 | (unsigned long)ac->ac_o_ex.fe_start, | |
3936 | (unsigned long)ac->ac_o_ex.fe_len, | |
3937 | (unsigned long)ac->ac_o_ex.fe_logical, | |
3938 | (unsigned long)ac->ac_g_ex.fe_group, | |
3939 | (unsigned long)ac->ac_g_ex.fe_start, | |
3940 | (unsigned long)ac->ac_g_ex.fe_len, | |
3941 | (unsigned long)ac->ac_g_ex.fe_logical, | |
3942 | (unsigned long)ac->ac_b_ex.fe_group, | |
3943 | (unsigned long)ac->ac_b_ex.fe_start, | |
3944 | (unsigned long)ac->ac_b_ex.fe_len, | |
3945 | (unsigned long)ac->ac_b_ex.fe_logical, | |
3946 | (int)ac->ac_criteria); | |
3947 | printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned, | |
3948 | ac->ac_found); | |
3949 | printk(KERN_ERR "EXT4-fs: groups: \n"); | |
3950 | for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) { | |
3951 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); | |
3952 | struct ext4_prealloc_space *pa; | |
3953 | ext4_grpblk_t start; | |
3954 | struct list_head *cur; | |
3955 | ext4_lock_group(sb, i); | |
3956 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3957 | pa = list_entry(cur, struct ext4_prealloc_space, | |
3958 | pa_group_list); | |
3959 | spin_lock(&pa->pa_lock); | |
3960 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3961 | NULL, &start); | |
3962 | spin_unlock(&pa->pa_lock); | |
3963 | printk(KERN_ERR "PA:%lu:%d:%u \n", i, | |
3964 | start, pa->pa_len); | |
3965 | } | |
60bd63d1 | 3966 | ext4_unlock_group(sb, i); |
c9de560d AT |
3967 | |
3968 | if (grp->bb_free == 0) | |
3969 | continue; | |
3970 | printk(KERN_ERR "%lu: %d/%d \n", | |
3971 | i, grp->bb_free, grp->bb_fragments); | |
3972 | } | |
3973 | printk(KERN_ERR "\n"); | |
3974 | } | |
3975 | #else | |
3976 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
3977 | { | |
3978 | return; | |
3979 | } | |
3980 | #endif | |
3981 | ||
3982 | /* | |
3983 | * We use locality group preallocation for small size file. The size of the | |
3984 | * file is determined by the current size or the resulting size after | |
3985 | * allocation which ever is larger | |
3986 | * | |
3987 | * One can tune this size via /proc/fs/ext4/<partition>/stream_req | |
3988 | */ | |
3989 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
3990 | { | |
3991 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3992 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
3993 | loff_t size, isize; | |
3994 | ||
3995 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3996 | return; | |
3997 | ||
3998 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; | |
3999 | isize = i_size_read(ac->ac_inode) >> bsbits; | |
4000 | size = max(size, isize); | |
4001 | ||
4002 | /* don't use group allocation for large files */ | |
4003 | if (size >= sbi->s_mb_stream_request) | |
4004 | return; | |
4005 | ||
4006 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
4007 | return; | |
4008 | ||
4009 | BUG_ON(ac->ac_lg != NULL); | |
4010 | /* | |
4011 | * locality group prealloc space are per cpu. The reason for having | |
4012 | * per cpu locality group is to reduce the contention between block | |
4013 | * request from multiple CPUs. | |
4014 | */ | |
730c213c | 4015 | ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id()); |
c9de560d AT |
4016 | |
4017 | /* we're going to use group allocation */ | |
4018 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4019 | ||
4020 | /* serialize all allocations in the group */ | |
4021 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4022 | } | |
4023 | ||
4ddfef7b ES |
4024 | static noinline_for_stack int |
4025 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
4026 | struct ext4_allocation_request *ar) |
4027 | { | |
4028 | struct super_block *sb = ar->inode->i_sb; | |
4029 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4030 | struct ext4_super_block *es = sbi->s_es; | |
4031 | ext4_group_t group; | |
4032 | unsigned long len; | |
4033 | unsigned long goal; | |
4034 | ext4_grpblk_t block; | |
4035 | ||
4036 | /* we can't allocate > group size */ | |
4037 | len = ar->len; | |
4038 | ||
4039 | /* just a dirty hack to filter too big requests */ | |
4040 | if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10) | |
4041 | len = EXT4_BLOCKS_PER_GROUP(sb) - 10; | |
4042 | ||
4043 | /* start searching from the goal */ | |
4044 | goal = ar->goal; | |
4045 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4046 | goal >= ext4_blocks_count(es)) | |
4047 | goal = le32_to_cpu(es->s_first_data_block); | |
4048 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4049 | ||
4050 | /* set up allocation goals */ | |
4051 | ac->ac_b_ex.fe_logical = ar->logical; | |
4052 | ac->ac_b_ex.fe_group = 0; | |
4053 | ac->ac_b_ex.fe_start = 0; | |
4054 | ac->ac_b_ex.fe_len = 0; | |
4055 | ac->ac_status = AC_STATUS_CONTINUE; | |
4056 | ac->ac_groups_scanned = 0; | |
4057 | ac->ac_ex_scanned = 0; | |
4058 | ac->ac_found = 0; | |
4059 | ac->ac_sb = sb; | |
4060 | ac->ac_inode = ar->inode; | |
4061 | ac->ac_o_ex.fe_logical = ar->logical; | |
4062 | ac->ac_o_ex.fe_group = group; | |
4063 | ac->ac_o_ex.fe_start = block; | |
4064 | ac->ac_o_ex.fe_len = len; | |
4065 | ac->ac_g_ex.fe_logical = ar->logical; | |
4066 | ac->ac_g_ex.fe_group = group; | |
4067 | ac->ac_g_ex.fe_start = block; | |
4068 | ac->ac_g_ex.fe_len = len; | |
4069 | ac->ac_f_ex.fe_len = 0; | |
4070 | ac->ac_flags = ar->flags; | |
4071 | ac->ac_2order = 0; | |
4072 | ac->ac_criteria = 0; | |
4073 | ac->ac_pa = NULL; | |
4074 | ac->ac_bitmap_page = NULL; | |
4075 | ac->ac_buddy_page = NULL; | |
4076 | ac->ac_lg = NULL; | |
4077 | ||
4078 | /* we have to define context: we'll we work with a file or | |
4079 | * locality group. this is a policy, actually */ | |
4080 | ext4_mb_group_or_file(ac); | |
4081 | ||
4082 | mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " | |
4083 | "left: %u/%u, right %u/%u to %swritable\n", | |
4084 | (unsigned) ar->len, (unsigned) ar->logical, | |
4085 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4086 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4087 | (unsigned) ar->lright, (unsigned) ar->pright, | |
4088 | atomic_read(&ar->inode->i_writecount) ? "" : "non-"); | |
4089 | return 0; | |
4090 | ||
4091 | } | |
4092 | ||
6be2ded1 AK |
4093 | static noinline_for_stack void |
4094 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4095 | struct ext4_locality_group *lg, | |
4096 | int order, int total_entries) | |
4097 | { | |
4098 | ext4_group_t group = 0; | |
4099 | struct ext4_buddy e4b; | |
4100 | struct list_head discard_list; | |
4101 | struct ext4_prealloc_space *pa, *tmp; | |
4102 | struct ext4_allocation_context *ac; | |
4103 | ||
4104 | mb_debug("discard locality group preallocation\n"); | |
4105 | ||
4106 | INIT_LIST_HEAD(&discard_list); | |
4107 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); | |
4108 | ||
4109 | spin_lock(&lg->lg_prealloc_lock); | |
4110 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
4111 | pa_inode_list) { | |
4112 | spin_lock(&pa->pa_lock); | |
4113 | if (atomic_read(&pa->pa_count)) { | |
4114 | /* | |
4115 | * This is the pa that we just used | |
4116 | * for block allocation. So don't | |
4117 | * free that | |
4118 | */ | |
4119 | spin_unlock(&pa->pa_lock); | |
4120 | continue; | |
4121 | } | |
4122 | if (pa->pa_deleted) { | |
4123 | spin_unlock(&pa->pa_lock); | |
4124 | continue; | |
4125 | } | |
4126 | /* only lg prealloc space */ | |
4127 | BUG_ON(!pa->pa_linear); | |
4128 | ||
4129 | /* seems this one can be freed ... */ | |
4130 | pa->pa_deleted = 1; | |
4131 | spin_unlock(&pa->pa_lock); | |
4132 | ||
4133 | list_del_rcu(&pa->pa_inode_list); | |
4134 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4135 | ||
4136 | total_entries--; | |
4137 | if (total_entries <= 5) { | |
4138 | /* | |
4139 | * we want to keep only 5 entries | |
4140 | * allowing it to grow to 8. This | |
4141 | * mak sure we don't call discard | |
4142 | * soon for this list. | |
4143 | */ | |
4144 | break; | |
4145 | } | |
4146 | } | |
4147 | spin_unlock(&lg->lg_prealloc_lock); | |
4148 | ||
4149 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
4150 | ||
4151 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); | |
4152 | if (ext4_mb_load_buddy(sb, group, &e4b)) { | |
4153 | ext4_error(sb, __func__, "Error in loading buddy " | |
a9df9a49 | 4154 | "information for %u", group); |
6be2ded1 AK |
4155 | continue; |
4156 | } | |
4157 | ext4_lock_group(sb, group); | |
4158 | list_del(&pa->pa_group_list); | |
4159 | ext4_mb_release_group_pa(&e4b, pa, ac); | |
4160 | ext4_unlock_group(sb, group); | |
4161 | ||
4162 | ext4_mb_release_desc(&e4b); | |
4163 | list_del(&pa->u.pa_tmp_list); | |
4164 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4165 | } | |
4166 | if (ac) | |
4167 | kmem_cache_free(ext4_ac_cachep, ac); | |
4168 | } | |
4169 | ||
4170 | /* | |
4171 | * We have incremented pa_count. So it cannot be freed at this | |
4172 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4173 | * possible from this lg. That means pa_free cannot be updated. | |
4174 | * | |
4175 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4176 | * which can cause the lg_prealloc_list to be updated. | |
4177 | */ | |
4178 | ||
4179 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4180 | { | |
4181 | int order, added = 0, lg_prealloc_count = 1; | |
4182 | struct super_block *sb = ac->ac_sb; | |
4183 | struct ext4_locality_group *lg = ac->ac_lg; | |
4184 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4185 | ||
4186 | order = fls(pa->pa_free) - 1; | |
4187 | if (order > PREALLOC_TB_SIZE - 1) | |
4188 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4189 | order = PREALLOC_TB_SIZE - 1; | |
4190 | /* Add the prealloc space to lg */ | |
4191 | rcu_read_lock(); | |
4192 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], | |
4193 | pa_inode_list) { | |
4194 | spin_lock(&tmp_pa->pa_lock); | |
4195 | if (tmp_pa->pa_deleted) { | |
4196 | spin_unlock(&pa->pa_lock); | |
4197 | continue; | |
4198 | } | |
4199 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4200 | /* Add to the tail of the previous entry */ | |
4201 | list_add_tail_rcu(&pa->pa_inode_list, | |
4202 | &tmp_pa->pa_inode_list); | |
4203 | added = 1; | |
4204 | /* | |
4205 | * we want to count the total | |
4206 | * number of entries in the list | |
4207 | */ | |
4208 | } | |
4209 | spin_unlock(&tmp_pa->pa_lock); | |
4210 | lg_prealloc_count++; | |
4211 | } | |
4212 | if (!added) | |
4213 | list_add_tail_rcu(&pa->pa_inode_list, | |
4214 | &lg->lg_prealloc_list[order]); | |
4215 | rcu_read_unlock(); | |
4216 | ||
4217 | /* Now trim the list to be not more than 8 elements */ | |
4218 | if (lg_prealloc_count > 8) { | |
4219 | ext4_mb_discard_lg_preallocations(sb, lg, | |
4220 | order, lg_prealloc_count); | |
4221 | return; | |
4222 | } | |
4223 | return ; | |
4224 | } | |
4225 | ||
c9de560d AT |
4226 | /* |
4227 | * release all resource we used in allocation | |
4228 | */ | |
4229 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4230 | { | |
6be2ded1 AK |
4231 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4232 | if (pa) { | |
4233 | if (pa->pa_linear) { | |
c9de560d | 4234 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 AK |
4235 | spin_lock(&pa->pa_lock); |
4236 | pa->pa_pstart += ac->ac_b_ex.fe_len; | |
4237 | pa->pa_lstart += ac->ac_b_ex.fe_len; | |
4238 | pa->pa_free -= ac->ac_b_ex.fe_len; | |
4239 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4240 | spin_unlock(&pa->pa_lock); | |
4241 | /* | |
4242 | * We want to add the pa to the right bucket. | |
4243 | * Remove it from the list and while adding | |
4244 | * make sure the list to which we are adding | |
4245 | * doesn't grow big. | |
4246 | */ | |
4247 | if (likely(pa->pa_free)) { | |
4248 | spin_lock(pa->pa_obj_lock); | |
4249 | list_del_rcu(&pa->pa_inode_list); | |
4250 | spin_unlock(pa->pa_obj_lock); | |
4251 | ext4_mb_add_n_trim(ac); | |
4252 | } | |
c9de560d | 4253 | } |
6be2ded1 | 4254 | ext4_mb_put_pa(ac, ac->ac_sb, pa); |
c9de560d AT |
4255 | } |
4256 | if (ac->ac_bitmap_page) | |
4257 | page_cache_release(ac->ac_bitmap_page); | |
4258 | if (ac->ac_buddy_page) | |
4259 | page_cache_release(ac->ac_buddy_page); | |
4260 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
4261 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4262 | ext4_mb_collect_stats(ac); | |
4263 | return 0; | |
4264 | } | |
4265 | ||
4266 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4267 | { | |
4268 | ext4_group_t i; | |
4269 | int ret; | |
4270 | int freed = 0; | |
4271 | ||
4272 | for (i = 0; i < EXT4_SB(sb)->s_groups_count && needed > 0; i++) { | |
4273 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); | |
4274 | freed += ret; | |
4275 | needed -= ret; | |
4276 | } | |
4277 | ||
4278 | return freed; | |
4279 | } | |
4280 | ||
4281 | /* | |
4282 | * Main entry point into mballoc to allocate blocks | |
4283 | * it tries to use preallocation first, then falls back | |
4284 | * to usual allocation | |
4285 | */ | |
4286 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
4287 | struct ext4_allocation_request *ar, int *errp) | |
4288 | { | |
6bc6e63f | 4289 | int freed; |
256bdb49 | 4290 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4291 | struct ext4_sb_info *sbi; |
4292 | struct super_block *sb; | |
4293 | ext4_fsblk_t block = 0; | |
6bc6e63f AK |
4294 | unsigned long inquota; |
4295 | unsigned long reserv_blks = 0; | |
c9de560d AT |
4296 | |
4297 | sb = ar->inode->i_sb; | |
4298 | sbi = EXT4_SB(sb); | |
4299 | ||
d2a17637 MC |
4300 | if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) { |
4301 | /* | |
4302 | * With delalloc we already reserved the blocks | |
4303 | */ | |
030ba6bc AK |
4304 | while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) { |
4305 | /* let others to free the space */ | |
4306 | yield(); | |
4307 | ar->len = ar->len >> 1; | |
4308 | } | |
4309 | if (!ar->len) { | |
a30d542a AK |
4310 | *errp = -ENOSPC; |
4311 | return 0; | |
4312 | } | |
6bc6e63f | 4313 | reserv_blks = ar->len; |
07031431 | 4314 | } |
c9de560d AT |
4315 | while (ar->len && DQUOT_ALLOC_BLOCK(ar->inode, ar->len)) { |
4316 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
4317 | ar->len--; | |
4318 | } | |
4319 | if (ar->len == 0) { | |
4320 | *errp = -EDQUOT; | |
4321 | return 0; | |
4322 | } | |
4323 | inquota = ar->len; | |
4324 | ||
d2a17637 MC |
4325 | if (EXT4_I(ar->inode)->i_delalloc_reserved_flag) |
4326 | ar->flags |= EXT4_MB_DELALLOC_RESERVED; | |
4327 | ||
256bdb49 ES |
4328 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
4329 | if (!ac) { | |
363d4251 | 4330 | ar->len = 0; |
256bdb49 | 4331 | *errp = -ENOMEM; |
363d4251 | 4332 | goto out1; |
256bdb49 ES |
4333 | } |
4334 | ||
256bdb49 | 4335 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4336 | if (*errp) { |
4337 | ar->len = 0; | |
363d4251 | 4338 | goto out2; |
c9de560d AT |
4339 | } |
4340 | ||
256bdb49 ES |
4341 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4342 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4343 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4344 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4345 | repeat: |
4346 | /* allocate space in core */ | |
256bdb49 | 4347 | ext4_mb_regular_allocator(ac); |
c9de560d AT |
4348 | |
4349 | /* as we've just preallocated more space than | |
4350 | * user requested orinally, we store allocated | |
4351 | * space in a special descriptor */ | |
256bdb49 ES |
4352 | if (ac->ac_status == AC_STATUS_FOUND && |
4353 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
4354 | ext4_mb_new_preallocation(ac); | |
c9de560d AT |
4355 | } |
4356 | ||
256bdb49 | 4357 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
6bc6e63f | 4358 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks); |
519deca0 AK |
4359 | if (*errp == -EAGAIN) { |
4360 | ac->ac_b_ex.fe_group = 0; | |
4361 | ac->ac_b_ex.fe_start = 0; | |
4362 | ac->ac_b_ex.fe_len = 0; | |
4363 | ac->ac_status = AC_STATUS_CONTINUE; | |
4364 | goto repeat; | |
4365 | } else if (*errp) { | |
4366 | ac->ac_b_ex.fe_len = 0; | |
4367 | ar->len = 0; | |
4368 | ext4_mb_show_ac(ac); | |
4369 | } else { | |
4370 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4371 | ar->len = ac->ac_b_ex.fe_len; | |
4372 | } | |
c9de560d | 4373 | } else { |
256bdb49 | 4374 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4375 | if (freed) |
4376 | goto repeat; | |
4377 | *errp = -ENOSPC; | |
256bdb49 | 4378 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4379 | ar->len = 0; |
256bdb49 | 4380 | ext4_mb_show_ac(ac); |
c9de560d AT |
4381 | } |
4382 | ||
256bdb49 | 4383 | ext4_mb_release_context(ac); |
c9de560d | 4384 | |
363d4251 SF |
4385 | out2: |
4386 | kmem_cache_free(ext4_ac_cachep, ac); | |
4387 | out1: | |
c9de560d AT |
4388 | if (ar->len < inquota) |
4389 | DQUOT_FREE_BLOCK(ar->inode, inquota - ar->len); | |
4390 | ||
4391 | return block; | |
4392 | } | |
c9de560d | 4393 | |
c894058d AK |
4394 | /* |
4395 | * We can merge two free data extents only if the physical blocks | |
4396 | * are contiguous, AND the extents were freed by the same transaction, | |
4397 | * AND the blocks are associated with the same group. | |
4398 | */ | |
4399 | static int can_merge(struct ext4_free_data *entry1, | |
4400 | struct ext4_free_data *entry2) | |
4401 | { | |
4402 | if ((entry1->t_tid == entry2->t_tid) && | |
4403 | (entry1->group == entry2->group) && | |
4404 | ((entry1->start_blk + entry1->count) == entry2->start_blk)) | |
4405 | return 1; | |
4406 | return 0; | |
4407 | } | |
4408 | ||
4ddfef7b ES |
4409 | static noinline_for_stack int |
4410 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
c9de560d AT |
4411 | ext4_group_t group, ext4_grpblk_t block, int count) |
4412 | { | |
4413 | struct ext4_group_info *db = e4b->bd_info; | |
4414 | struct super_block *sb = e4b->bd_sb; | |
4415 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4416 | struct ext4_free_data *entry, *new_entry; |
4417 | struct rb_node **n = &db->bb_free_root.rb_node, *node; | |
4418 | struct rb_node *parent = NULL, *new_node; | |
4419 | ||
0390131b | 4420 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4421 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4422 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4423 | ||
c894058d AK |
4424 | new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS); |
4425 | new_entry->start_blk = block; | |
4426 | new_entry->group = group; | |
4427 | new_entry->count = count; | |
4428 | new_entry->t_tid = handle->h_transaction->t_tid; | |
4429 | new_node = &new_entry->node; | |
4430 | ||
c9de560d | 4431 | ext4_lock_group(sb, group); |
c894058d AK |
4432 | if (!*n) { |
4433 | /* first free block exent. We need to | |
4434 | protect buddy cache from being freed, | |
4435 | * otherwise we'll refresh it from | |
4436 | * on-disk bitmap and lose not-yet-available | |
4437 | * blocks */ | |
4438 | page_cache_get(e4b->bd_buddy_page); | |
4439 | page_cache_get(e4b->bd_bitmap_page); | |
4440 | } | |
4441 | while (*n) { | |
4442 | parent = *n; | |
4443 | entry = rb_entry(parent, struct ext4_free_data, node); | |
4444 | if (block < entry->start_blk) | |
4445 | n = &(*n)->rb_left; | |
4446 | else if (block >= (entry->start_blk + entry->count)) | |
4447 | n = &(*n)->rb_right; | |
4448 | else { | |
ae2d9fb1 | 4449 | ext4_unlock_group(sb, group); |
c894058d | 4450 | ext4_error(sb, __func__, |
fde4d95a | 4451 | "Double free of blocks %d (%d %d)", |
c894058d AK |
4452 | block, entry->start_blk, entry->count); |
4453 | return 0; | |
c9de560d | 4454 | } |
c894058d | 4455 | } |
c9de560d | 4456 | |
c894058d AK |
4457 | rb_link_node(new_node, parent, n); |
4458 | rb_insert_color(new_node, &db->bb_free_root); | |
4459 | ||
4460 | /* Now try to see the extent can be merged to left and right */ | |
4461 | node = rb_prev(new_node); | |
4462 | if (node) { | |
4463 | entry = rb_entry(node, struct ext4_free_data, node); | |
4464 | if (can_merge(entry, new_entry)) { | |
4465 | new_entry->start_blk = entry->start_blk; | |
4466 | new_entry->count += entry->count; | |
4467 | rb_erase(node, &(db->bb_free_root)); | |
4468 | spin_lock(&sbi->s_md_lock); | |
4469 | list_del(&entry->list); | |
4470 | spin_unlock(&sbi->s_md_lock); | |
4471 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
c9de560d | 4472 | } |
c894058d | 4473 | } |
c9de560d | 4474 | |
c894058d AK |
4475 | node = rb_next(new_node); |
4476 | if (node) { | |
4477 | entry = rb_entry(node, struct ext4_free_data, node); | |
4478 | if (can_merge(new_entry, entry)) { | |
4479 | new_entry->count += entry->count; | |
4480 | rb_erase(node, &(db->bb_free_root)); | |
4481 | spin_lock(&sbi->s_md_lock); | |
4482 | list_del(&entry->list); | |
4483 | spin_unlock(&sbi->s_md_lock); | |
4484 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
c9de560d AT |
4485 | } |
4486 | } | |
3e624fc7 | 4487 | /* Add the extent to transaction's private list */ |
c894058d | 4488 | spin_lock(&sbi->s_md_lock); |
3e624fc7 | 4489 | list_add(&new_entry->list, &handle->h_transaction->t_private_list); |
c894058d | 4490 | spin_unlock(&sbi->s_md_lock); |
c9de560d AT |
4491 | ext4_unlock_group(sb, group); |
4492 | return 0; | |
4493 | } | |
4494 | ||
4495 | /* | |
4496 | * Main entry point into mballoc to free blocks | |
4497 | */ | |
4498 | void ext4_mb_free_blocks(handle_t *handle, struct inode *inode, | |
4499 | unsigned long block, unsigned long count, | |
4500 | int metadata, unsigned long *freed) | |
4501 | { | |
26346ff6 | 4502 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4503 | struct super_block *sb = inode->i_sb; |
256bdb49 | 4504 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4505 | struct ext4_group_desc *gdp; |
4506 | struct ext4_super_block *es; | |
4507 | unsigned long overflow; | |
4508 | ext4_grpblk_t bit; | |
4509 | struct buffer_head *gd_bh; | |
4510 | ext4_group_t block_group; | |
4511 | struct ext4_sb_info *sbi; | |
4512 | struct ext4_buddy e4b; | |
4513 | int err = 0; | |
4514 | int ret; | |
4515 | ||
4516 | *freed = 0; | |
4517 | ||
c9de560d AT |
4518 | sbi = EXT4_SB(sb); |
4519 | es = EXT4_SB(sb)->s_es; | |
4520 | if (block < le32_to_cpu(es->s_first_data_block) || | |
4521 | block + count < block || | |
4522 | block + count > ext4_blocks_count(es)) { | |
46e665e9 | 4523 | ext4_error(sb, __func__, |
c9de560d AT |
4524 | "Freeing blocks not in datazone - " |
4525 | "block = %lu, count = %lu", block, count); | |
4526 | goto error_return; | |
4527 | } | |
4528 | ||
4529 | ext4_debug("freeing block %lu\n", block); | |
4530 | ||
256bdb49 ES |
4531 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
4532 | if (ac) { | |
4533 | ac->ac_op = EXT4_MB_HISTORY_FREE; | |
4534 | ac->ac_inode = inode; | |
4535 | ac->ac_sb = sb; | |
4536 | } | |
c9de560d AT |
4537 | |
4538 | do_more: | |
4539 | overflow = 0; | |
4540 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4541 | ||
4542 | /* | |
4543 | * Check to see if we are freeing blocks across a group | |
4544 | * boundary. | |
4545 | */ | |
4546 | if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { | |
4547 | overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb); | |
4548 | count -= overflow; | |
4549 | } | |
574ca174 | 4550 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
ce89f46c AK |
4551 | if (!bitmap_bh) { |
4552 | err = -EIO; | |
c9de560d | 4553 | goto error_return; |
ce89f46c | 4554 | } |
c9de560d | 4555 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4556 | if (!gdp) { |
4557 | err = -EIO; | |
c9de560d | 4558 | goto error_return; |
ce89f46c | 4559 | } |
c9de560d AT |
4560 | |
4561 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4562 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4563 | in_range(block, ext4_inode_table(sb, gdp), | |
4564 | EXT4_SB(sb)->s_itb_per_group) || | |
4565 | in_range(block + count - 1, ext4_inode_table(sb, gdp), | |
4566 | EXT4_SB(sb)->s_itb_per_group)) { | |
4567 | ||
46e665e9 | 4568 | ext4_error(sb, __func__, |
c9de560d AT |
4569 | "Freeing blocks in system zone - " |
4570 | "Block = %lu, count = %lu", block, count); | |
519deca0 AK |
4571 | /* err = 0. ext4_std_error should be a no op */ |
4572 | goto error_return; | |
c9de560d AT |
4573 | } |
4574 | ||
4575 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4576 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4577 | if (err) | |
4578 | goto error_return; | |
4579 | ||
4580 | /* | |
4581 | * We are about to modify some metadata. Call the journal APIs | |
4582 | * to unshare ->b_data if a currently-committing transaction is | |
4583 | * using it | |
4584 | */ | |
4585 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4586 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4587 | if (err) | |
4588 | goto error_return; | |
4589 | ||
4590 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
4591 | if (err) | |
4592 | goto error_return; | |
4593 | ||
4594 | #ifdef AGGRESSIVE_CHECK | |
4595 | { | |
4596 | int i; | |
4597 | for (i = 0; i < count; i++) | |
4598 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); | |
4599 | } | |
4600 | #endif | |
4601 | mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data, | |
4602 | bit, count); | |
4603 | ||
4604 | /* We dirtied the bitmap block */ | |
4605 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
0390131b | 4606 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d | 4607 | |
256bdb49 ES |
4608 | if (ac) { |
4609 | ac->ac_b_ex.fe_group = block_group; | |
4610 | ac->ac_b_ex.fe_start = bit; | |
4611 | ac->ac_b_ex.fe_len = count; | |
4612 | ext4_mb_store_history(ac); | |
4613 | } | |
c9de560d | 4614 | |
0390131b | 4615 | if (metadata && ext4_handle_valid(handle)) { |
c9de560d AT |
4616 | /* blocks being freed are metadata. these blocks shouldn't |
4617 | * be used until this transaction is committed */ | |
4618 | ext4_mb_free_metadata(handle, &e4b, block_group, bit, count); | |
4619 | } else { | |
4620 | ext4_lock_group(sb, block_group); | |
7e5a8cdd | 4621 | mb_free_blocks(inode, &e4b, bit, count); |
c9de560d AT |
4622 | ext4_mb_return_to_preallocation(inode, &e4b, block, count); |
4623 | ext4_unlock_group(sb, block_group); | |
c9de560d AT |
4624 | } |
4625 | ||
4626 | spin_lock(sb_bgl_lock(sbi, block_group)); | |
e8546d06 | 4627 | le16_add_cpu(&gdp->bg_free_blocks_count, count); |
c9de560d AT |
4628 | gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); |
4629 | spin_unlock(sb_bgl_lock(sbi, block_group)); | |
4630 | percpu_counter_add(&sbi->s_freeblocks_counter, count); | |
4631 | ||
772cb7c8 JS |
4632 | if (sbi->s_log_groups_per_flex) { |
4633 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
4634 | spin_lock(sb_bgl_lock(sbi, flex_group)); | |
4635 | sbi->s_flex_groups[flex_group].free_blocks += count; | |
4636 | spin_unlock(sb_bgl_lock(sbi, flex_group)); | |
4637 | } | |
4638 | ||
c9de560d AT |
4639 | ext4_mb_release_desc(&e4b); |
4640 | ||
4641 | *freed += count; | |
4642 | ||
4643 | /* And the group descriptor block */ | |
4644 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 4645 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
4646 | if (!err) |
4647 | err = ret; | |
4648 | ||
4649 | if (overflow && !err) { | |
4650 | block += count; | |
4651 | count = overflow; | |
4652 | put_bh(bitmap_bh); | |
4653 | goto do_more; | |
4654 | } | |
4655 | sb->s_dirt = 1; | |
4656 | error_return: | |
4657 | brelse(bitmap_bh); | |
4658 | ext4_std_error(sb, err); | |
256bdb49 ES |
4659 | if (ac) |
4660 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
4661 | return; |
4662 | } |