]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com | |
3 | * Written by Alex Tomas <alex@clusterfs.com> | |
4 | * | |
5 | * This program is free software; you can redistribute it and/or modify | |
6 | * it under the terms of the GNU General Public License version 2 as | |
7 | * published by the Free Software Foundation. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public Licens | |
15 | * along with this program; if not, write to the Free Software | |
16 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111- | |
17 | */ | |
18 | ||
19 | ||
20 | /* | |
21 | * mballoc.c contains the multiblocks allocation routines | |
22 | */ | |
23 | ||
24 | #include "mballoc.h" | |
25 | #include <linux/debugfs.h> | |
26 | #include <linux/slab.h> | |
27 | #include <trace/events/ext4.h> | |
28 | ||
29 | /* | |
30 | * MUSTDO: | |
31 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
32 | * - search for metadata in few groups | |
33 | * | |
34 | * TODO v4: | |
35 | * - normalization should take into account whether file is still open | |
36 | * - discard preallocations if no free space left (policy?) | |
37 | * - don't normalize tails | |
38 | * - quota | |
39 | * - reservation for superuser | |
40 | * | |
41 | * TODO v3: | |
42 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
43 | * - track min/max extents in each group for better group selection | |
44 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
45 | * - tree of groups sorted by number of free blocks | |
46 | * - error handling | |
47 | */ | |
48 | ||
49 | /* | |
50 | * The allocation request involve request for multiple number of blocks | |
51 | * near to the goal(block) value specified. | |
52 | * | |
53 | * During initialization phase of the allocator we decide to use the | |
54 | * group preallocation or inode preallocation depending on the size of | |
55 | * the file. The size of the file could be the resulting file size we | |
56 | * would have after allocation, or the current file size, which ever | |
57 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
58 | * select to use the group preallocation. The default value of | |
59 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
60 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
61 | * terms of number of blocks. | |
62 | * | |
63 | * The main motivation for having small file use group preallocation is to | |
64 | * ensure that we have small files closer together on the disk. | |
65 | * | |
66 | * First stage the allocator looks at the inode prealloc list, | |
67 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
68 | * spaces for this particular inode. The inode prealloc space is | |
69 | * represented as: | |
70 | * | |
71 | * pa_lstart -> the logical start block for this prealloc space | |
72 | * pa_pstart -> the physical start block for this prealloc space | |
73 | * pa_len -> length for this prealloc space | |
74 | * pa_free -> free space available in this prealloc space | |
75 | * | |
76 | * The inode preallocation space is used looking at the _logical_ start | |
77 | * block. If only the logical file block falls within the range of prealloc | |
78 | * space we will consume the particular prealloc space. This make sure that | |
79 | * that the we have contiguous physical blocks representing the file blocks | |
80 | * | |
81 | * The important thing to be noted in case of inode prealloc space is that | |
82 | * we don't modify the values associated to inode prealloc space except | |
83 | * pa_free. | |
84 | * | |
85 | * If we are not able to find blocks in the inode prealloc space and if we | |
86 | * have the group allocation flag set then we look at the locality group | |
87 | * prealloc space. These are per CPU prealloc list repreasented as | |
88 | * | |
89 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
90 | * | |
91 | * The reason for having a per cpu locality group is to reduce the contention | |
92 | * between CPUs. It is possible to get scheduled at this point. | |
93 | * | |
94 | * The locality group prealloc space is used looking at whether we have | |
95 | * enough free space (pa_free) withing the prealloc space. | |
96 | * | |
97 | * If we can't allocate blocks via inode prealloc or/and locality group | |
98 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
99 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
100 | * mapped to the buddy and bitmap information regarding different | |
101 | * groups. The buddy information is attached to buddy cache inode so that | |
102 | * we can access them through the page cache. The information regarding | |
103 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
104 | * block bitmap and buddy information. The information are stored in the | |
105 | * inode as: | |
106 | * | |
107 | * { page } | |
108 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... | |
109 | * | |
110 | * | |
111 | * one block each for bitmap and buddy information. So for each group we | |
112 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / | |
113 | * blocksize) blocks. So it can have information regarding groups_per_page | |
114 | * which is blocks_per_page/2 | |
115 | * | |
116 | * The buddy cache inode is not stored on disk. The inode is thrown | |
117 | * away when the filesystem is unmounted. | |
118 | * | |
119 | * We look for count number of blocks in the buddy cache. If we were able | |
120 | * to locate that many free blocks we return with additional information | |
121 | * regarding rest of the contiguous physical block available | |
122 | * | |
123 | * Before allocating blocks via buddy cache we normalize the request | |
124 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
125 | * blocks that we get after allocation is added to the respective prealloc | |
126 | * list. In case of inode preallocation we follow a list of heuristics | |
127 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
128 | * we are doing a group prealloc we try to normalize the request to | |
129 | * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is | |
130 | * 512 blocks. This can be tuned via | |
131 | * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in | |
132 | * terms of number of blocks. If we have mounted the file system with -O | |
133 | * stripe=<value> option the group prealloc request is normalized to the | |
134 | * stripe value (sbi->s_stripe) | |
135 | * | |
136 | * The regular allocator(using the buddy cache) supports few tunables. | |
137 | * | |
138 | * /sys/fs/ext4/<partition>/mb_min_to_scan | |
139 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
140 | * /sys/fs/ext4/<partition>/mb_order2_req | |
141 | * | |
142 | * The regular allocator uses buddy scan only if the request len is power of | |
143 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The | |
144 | * value of s_mb_order2_reqs can be tuned via | |
145 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to | |
146 | * stripe size (sbi->s_stripe), we try to search for contiguous block in | |
147 | * stripe size. This should result in better allocation on RAID setups. If | |
148 | * not, we search in the specific group using bitmap for best extents. The | |
149 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
150 | * min_to_scan indicate how long the mballoc __must__ look for a best | |
151 | * extent and max_to_scan indicates how long the mballoc __can__ look for a | |
152 | * best extent in the found extents. Searching for the blocks starts with | |
153 | * the group specified as the goal value in allocation context via | |
154 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
155 | * can used for allocation. ext4_mb_good_group explains how the groups are | |
156 | * checked. | |
157 | * | |
158 | * Both the prealloc space are getting populated as above. So for the first | |
159 | * request we will hit the buddy cache which will result in this prealloc | |
160 | * space getting filled. The prealloc space is then later used for the | |
161 | * subsequent request. | |
162 | */ | |
163 | ||
164 | /* | |
165 | * mballoc operates on the following data: | |
166 | * - on-disk bitmap | |
167 | * - in-core buddy (actually includes buddy and bitmap) | |
168 | * - preallocation descriptors (PAs) | |
169 | * | |
170 | * there are two types of preallocations: | |
171 | * - inode | |
172 | * assiged to specific inode and can be used for this inode only. | |
173 | * it describes part of inode's space preallocated to specific | |
174 | * physical blocks. any block from that preallocated can be used | |
175 | * independent. the descriptor just tracks number of blocks left | |
176 | * unused. so, before taking some block from descriptor, one must | |
177 | * make sure corresponded logical block isn't allocated yet. this | |
178 | * also means that freeing any block within descriptor's range | |
179 | * must discard all preallocated blocks. | |
180 | * - locality group | |
181 | * assigned to specific locality group which does not translate to | |
182 | * permanent set of inodes: inode can join and leave group. space | |
183 | * from this type of preallocation can be used for any inode. thus | |
184 | * it's consumed from the beginning to the end. | |
185 | * | |
186 | * relation between them can be expressed as: | |
187 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
188 | * | |
189 | * this mean blocks mballoc considers used are: | |
190 | * - allocated blocks (persistent) | |
191 | * - preallocated blocks (non-persistent) | |
192 | * | |
193 | * consistency in mballoc world means that at any time a block is either | |
194 | * free or used in ALL structures. notice: "any time" should not be read | |
195 | * literally -- time is discrete and delimited by locks. | |
196 | * | |
197 | * to keep it simple, we don't use block numbers, instead we count number of | |
198 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
199 | * | |
200 | * all operations can be expressed as: | |
201 | * - init buddy: buddy = on-disk + PAs | |
202 | * - new PA: buddy += N; PA = N | |
203 | * - use inode PA: on-disk += N; PA -= N | |
204 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
205 | * - use locality group PA on-disk += N; PA -= N | |
206 | * - discard locality group PA buddy -= PA; PA = 0 | |
207 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
208 | * is used in real operation because we can't know actual used | |
209 | * bits from PA, only from on-disk bitmap | |
210 | * | |
211 | * if we follow this strict logic, then all operations above should be atomic. | |
212 | * given some of them can block, we'd have to use something like semaphores | |
213 | * killing performance on high-end SMP hardware. let's try to relax it using | |
214 | * the following knowledge: | |
215 | * 1) if buddy is referenced, it's already initialized | |
216 | * 2) while block is used in buddy and the buddy is referenced, | |
217 | * nobody can re-allocate that block | |
218 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
219 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
220 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
221 | * block | |
222 | * | |
223 | * so, now we're building a concurrency table: | |
224 | * - init buddy vs. | |
225 | * - new PA | |
226 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
227 | * until PA is linked to allocation group to avoid concurrent buddy init | |
228 | * - use inode PA | |
229 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
230 | * given (3) we care that PA-=N operation doesn't interfere with init | |
231 | * - discard inode PA | |
232 | * the simplest way would be to have buddy initialized by the discard | |
233 | * - use locality group PA | |
234 | * again PA-=N must be serialized with init | |
235 | * - discard locality group PA | |
236 | * the simplest way would be to have buddy initialized by the discard | |
237 | * - new PA vs. | |
238 | * - use inode PA | |
239 | * i_data_sem serializes them | |
240 | * - discard inode PA | |
241 | * discard process must wait until PA isn't used by another process | |
242 | * - use locality group PA | |
243 | * some mutex should serialize them | |
244 | * - discard locality group PA | |
245 | * discard process must wait until PA isn't used by another process | |
246 | * - use inode PA | |
247 | * - use inode PA | |
248 | * i_data_sem or another mutex should serializes them | |
249 | * - discard inode PA | |
250 | * discard process must wait until PA isn't used by another process | |
251 | * - use locality group PA | |
252 | * nothing wrong here -- they're different PAs covering different blocks | |
253 | * - discard locality group PA | |
254 | * discard process must wait until PA isn't used by another process | |
255 | * | |
256 | * now we're ready to make few consequences: | |
257 | * - PA is referenced and while it is no discard is possible | |
258 | * - PA is referenced until block isn't marked in on-disk bitmap | |
259 | * - PA changes only after on-disk bitmap | |
260 | * - discard must not compete with init. either init is done before | |
261 | * any discard or they're serialized somehow | |
262 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
263 | * | |
264 | * a special case when we've used PA to emptiness. no need to modify buddy | |
265 | * in this case, but we should care about concurrent init | |
266 | * | |
267 | */ | |
268 | ||
269 | /* | |
270 | * Logic in few words: | |
271 | * | |
272 | * - allocation: | |
273 | * load group | |
274 | * find blocks | |
275 | * mark bits in on-disk bitmap | |
276 | * release group | |
277 | * | |
278 | * - use preallocation: | |
279 | * find proper PA (per-inode or group) | |
280 | * load group | |
281 | * mark bits in on-disk bitmap | |
282 | * release group | |
283 | * release PA | |
284 | * | |
285 | * - free: | |
286 | * load group | |
287 | * mark bits in on-disk bitmap | |
288 | * release group | |
289 | * | |
290 | * - discard preallocations in group: | |
291 | * mark PAs deleted | |
292 | * move them onto local list | |
293 | * load on-disk bitmap | |
294 | * load group | |
295 | * remove PA from object (inode or locality group) | |
296 | * mark free blocks in-core | |
297 | * | |
298 | * - discard inode's preallocations: | |
299 | */ | |
300 | ||
301 | /* | |
302 | * Locking rules | |
303 | * | |
304 | * Locks: | |
305 | * - bitlock on a group (group) | |
306 | * - object (inode/locality) (object) | |
307 | * - per-pa lock (pa) | |
308 | * | |
309 | * Paths: | |
310 | * - new pa | |
311 | * object | |
312 | * group | |
313 | * | |
314 | * - find and use pa: | |
315 | * pa | |
316 | * | |
317 | * - release consumed pa: | |
318 | * pa | |
319 | * group | |
320 | * object | |
321 | * | |
322 | * - generate in-core bitmap: | |
323 | * group | |
324 | * pa | |
325 | * | |
326 | * - discard all for given object (inode, locality group): | |
327 | * object | |
328 | * pa | |
329 | * group | |
330 | * | |
331 | * - discard all for given group: | |
332 | * group | |
333 | * pa | |
334 | * group | |
335 | * object | |
336 | * | |
337 | */ | |
338 | static struct kmem_cache *ext4_pspace_cachep; | |
339 | static struct kmem_cache *ext4_ac_cachep; | |
340 | static struct kmem_cache *ext4_free_ext_cachep; | |
341 | ||
342 | /* We create slab caches for groupinfo data structures based on the | |
343 | * superblock block size. There will be one per mounted filesystem for | |
344 | * each unique s_blocksize_bits */ | |
345 | #define NR_GRPINFO_CACHES \ | |
346 | (EXT4_MAX_BLOCK_LOG_SIZE - EXT4_MIN_BLOCK_LOG_SIZE + 1) | |
347 | static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; | |
348 | ||
349 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
350 | ext4_group_t group); | |
351 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
352 | ext4_group_t group); | |
353 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn); | |
354 | ||
355 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) | |
356 | { | |
357 | #if BITS_PER_LONG == 64 | |
358 | *bit += ((unsigned long) addr & 7UL) << 3; | |
359 | addr = (void *) ((unsigned long) addr & ~7UL); | |
360 | #elif BITS_PER_LONG == 32 | |
361 | *bit += ((unsigned long) addr & 3UL) << 3; | |
362 | addr = (void *) ((unsigned long) addr & ~3UL); | |
363 | #else | |
364 | #error "how many bits you are?!" | |
365 | #endif | |
366 | return addr; | |
367 | } | |
368 | ||
369 | static inline int mb_test_bit(int bit, void *addr) | |
370 | { | |
371 | /* | |
372 | * ext4_test_bit on architecture like powerpc | |
373 | * needs unsigned long aligned address | |
374 | */ | |
375 | addr = mb_correct_addr_and_bit(&bit, addr); | |
376 | return ext4_test_bit(bit, addr); | |
377 | } | |
378 | ||
379 | static inline void mb_set_bit(int bit, void *addr) | |
380 | { | |
381 | addr = mb_correct_addr_and_bit(&bit, addr); | |
382 | ext4_set_bit(bit, addr); | |
383 | } | |
384 | ||
385 | static inline void mb_clear_bit(int bit, void *addr) | |
386 | { | |
387 | addr = mb_correct_addr_and_bit(&bit, addr); | |
388 | ext4_clear_bit(bit, addr); | |
389 | } | |
390 | ||
391 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) | |
392 | { | |
393 | int fix = 0, ret, tmpmax; | |
394 | addr = mb_correct_addr_and_bit(&fix, addr); | |
395 | tmpmax = max + fix; | |
396 | start += fix; | |
397 | ||
398 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; | |
399 | if (ret > max) | |
400 | return max; | |
401 | return ret; | |
402 | } | |
403 | ||
404 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
405 | { | |
406 | int fix = 0, ret, tmpmax; | |
407 | addr = mb_correct_addr_and_bit(&fix, addr); | |
408 | tmpmax = max + fix; | |
409 | start += fix; | |
410 | ||
411 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; | |
412 | if (ret > max) | |
413 | return max; | |
414 | return ret; | |
415 | } | |
416 | ||
417 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) | |
418 | { | |
419 | char *bb; | |
420 | ||
421 | BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); | |
422 | BUG_ON(max == NULL); | |
423 | ||
424 | if (order > e4b->bd_blkbits + 1) { | |
425 | *max = 0; | |
426 | return NULL; | |
427 | } | |
428 | ||
429 | /* at order 0 we see each particular block */ | |
430 | *max = 1 << (e4b->bd_blkbits + 3); | |
431 | if (order == 0) | |
432 | return EXT4_MB_BITMAP(e4b); | |
433 | ||
434 | bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; | |
435 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; | |
436 | ||
437 | return bb; | |
438 | } | |
439 | ||
440 | #ifdef DOUBLE_CHECK | |
441 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
442 | int first, int count) | |
443 | { | |
444 | int i; | |
445 | struct super_block *sb = e4b->bd_sb; | |
446 | ||
447 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
448 | return; | |
449 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); | |
450 | for (i = 0; i < count; i++) { | |
451 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
452 | ext4_fsblk_t blocknr; | |
453 | ||
454 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
455 | blocknr += first + i; | |
456 | ext4_grp_locked_error(sb, e4b->bd_group, | |
457 | inode ? inode->i_ino : 0, | |
458 | blocknr, | |
459 | "freeing block already freed " | |
460 | "(bit %u)", | |
461 | first + i); | |
462 | } | |
463 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
464 | } | |
465 | } | |
466 | ||
467 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
468 | { | |
469 | int i; | |
470 | ||
471 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
472 | return; | |
473 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); | |
474 | for (i = 0; i < count; i++) { | |
475 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
476 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
477 | } | |
478 | } | |
479 | ||
480 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
481 | { | |
482 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { | |
483 | unsigned char *b1, *b2; | |
484 | int i; | |
485 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
486 | b2 = (unsigned char *) bitmap; | |
487 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
488 | if (b1[i] != b2[i]) { | |
489 | printk(KERN_ERR "corruption in group %u " | |
490 | "at byte %u(%u): %x in copy != %x " | |
491 | "on disk/prealloc\n", | |
492 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
493 | BUG(); | |
494 | } | |
495 | } | |
496 | } | |
497 | } | |
498 | ||
499 | #else | |
500 | static inline void mb_free_blocks_double(struct inode *inode, | |
501 | struct ext4_buddy *e4b, int first, int count) | |
502 | { | |
503 | return; | |
504 | } | |
505 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
506 | int first, int count) | |
507 | { | |
508 | return; | |
509 | } | |
510 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
511 | { | |
512 | return; | |
513 | } | |
514 | #endif | |
515 | ||
516 | #ifdef AGGRESSIVE_CHECK | |
517 | ||
518 | #define MB_CHECK_ASSERT(assert) \ | |
519 | do { \ | |
520 | if (!(assert)) { \ | |
521 | printk(KERN_EMERG \ | |
522 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
523 | function, file, line, # assert); \ | |
524 | BUG(); \ | |
525 | } \ | |
526 | } while (0) | |
527 | ||
528 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
529 | const char *function, int line) | |
530 | { | |
531 | struct super_block *sb = e4b->bd_sb; | |
532 | int order = e4b->bd_blkbits + 1; | |
533 | int max; | |
534 | int max2; | |
535 | int i; | |
536 | int j; | |
537 | int k; | |
538 | int count; | |
539 | struct ext4_group_info *grp; | |
540 | int fragments = 0; | |
541 | int fstart; | |
542 | struct list_head *cur; | |
543 | void *buddy; | |
544 | void *buddy2; | |
545 | ||
546 | { | |
547 | static int mb_check_counter; | |
548 | if (mb_check_counter++ % 100 != 0) | |
549 | return 0; | |
550 | } | |
551 | ||
552 | while (order > 1) { | |
553 | buddy = mb_find_buddy(e4b, order, &max); | |
554 | MB_CHECK_ASSERT(buddy); | |
555 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
556 | MB_CHECK_ASSERT(buddy2); | |
557 | MB_CHECK_ASSERT(buddy != buddy2); | |
558 | MB_CHECK_ASSERT(max * 2 == max2); | |
559 | ||
560 | count = 0; | |
561 | for (i = 0; i < max; i++) { | |
562 | ||
563 | if (mb_test_bit(i, buddy)) { | |
564 | /* only single bit in buddy2 may be 1 */ | |
565 | if (!mb_test_bit(i << 1, buddy2)) { | |
566 | MB_CHECK_ASSERT( | |
567 | mb_test_bit((i<<1)+1, buddy2)); | |
568 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
569 | MB_CHECK_ASSERT( | |
570 | mb_test_bit(i << 1, buddy2)); | |
571 | } | |
572 | continue; | |
573 | } | |
574 | ||
575 | /* both bits in buddy2 must be 0 */ | |
576 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); | |
577 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
578 | ||
579 | for (j = 0; j < (1 << order); j++) { | |
580 | k = (i * (1 << order)) + j; | |
581 | MB_CHECK_ASSERT( | |
582 | !mb_test_bit(k, EXT4_MB_BITMAP(e4b))); | |
583 | } | |
584 | count++; | |
585 | } | |
586 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
587 | order--; | |
588 | } | |
589 | ||
590 | fstart = -1; | |
591 | buddy = mb_find_buddy(e4b, 0, &max); | |
592 | for (i = 0; i < max; i++) { | |
593 | if (!mb_test_bit(i, buddy)) { | |
594 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
595 | if (fstart == -1) { | |
596 | fragments++; | |
597 | fstart = i; | |
598 | } | |
599 | continue; | |
600 | } | |
601 | fstart = -1; | |
602 | /* check used bits only */ | |
603 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
604 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
605 | k = i >> j; | |
606 | MB_CHECK_ASSERT(k < max2); | |
607 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
608 | } | |
609 | } | |
610 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
611 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
612 | ||
613 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
614 | buddy = mb_find_buddy(e4b, 0, &max); | |
615 | list_for_each(cur, &grp->bb_prealloc_list) { | |
616 | ext4_group_t groupnr; | |
617 | struct ext4_prealloc_space *pa; | |
618 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
619 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
620 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); | |
621 | for (i = 0; i < pa->pa_len; i++) | |
622 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); | |
623 | } | |
624 | return 0; | |
625 | } | |
626 | #undef MB_CHECK_ASSERT | |
627 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
628 | __FILE__, __func__, __LINE__) | |
629 | #else | |
630 | #define mb_check_buddy(e4b) | |
631 | #endif | |
632 | ||
633 | /* FIXME!! need more doc */ | |
634 | static void ext4_mb_mark_free_simple(struct super_block *sb, | |
635 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, | |
636 | struct ext4_group_info *grp) | |
637 | { | |
638 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
639 | ext4_grpblk_t min; | |
640 | ext4_grpblk_t max; | |
641 | ext4_grpblk_t chunk; | |
642 | unsigned short border; | |
643 | ||
644 | BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb)); | |
645 | ||
646 | border = 2 << sb->s_blocksize_bits; | |
647 | ||
648 | while (len > 0) { | |
649 | /* find how many blocks can be covered since this position */ | |
650 | max = ffs(first | border) - 1; | |
651 | ||
652 | /* find how many blocks of power 2 we need to mark */ | |
653 | min = fls(len) - 1; | |
654 | ||
655 | if (max < min) | |
656 | min = max; | |
657 | chunk = 1 << min; | |
658 | ||
659 | /* mark multiblock chunks only */ | |
660 | grp->bb_counters[min]++; | |
661 | if (min > 0) | |
662 | mb_clear_bit(first >> min, | |
663 | buddy + sbi->s_mb_offsets[min]); | |
664 | ||
665 | len -= chunk; | |
666 | first += chunk; | |
667 | } | |
668 | } | |
669 | ||
670 | /* | |
671 | * Cache the order of the largest free extent we have available in this block | |
672 | * group. | |
673 | */ | |
674 | static void | |
675 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
676 | { | |
677 | int i; | |
678 | int bits; | |
679 | ||
680 | grp->bb_largest_free_order = -1; /* uninit */ | |
681 | ||
682 | bits = sb->s_blocksize_bits + 1; | |
683 | for (i = bits; i >= 0; i--) { | |
684 | if (grp->bb_counters[i] > 0) { | |
685 | grp->bb_largest_free_order = i; | |
686 | break; | |
687 | } | |
688 | } | |
689 | } | |
690 | ||
691 | static noinline_for_stack | |
692 | void ext4_mb_generate_buddy(struct super_block *sb, | |
693 | void *buddy, void *bitmap, ext4_group_t group) | |
694 | { | |
695 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
696 | ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb); | |
697 | ext4_grpblk_t i = 0; | |
698 | ext4_grpblk_t first; | |
699 | ext4_grpblk_t len; | |
700 | unsigned free = 0; | |
701 | unsigned fragments = 0; | |
702 | unsigned long long period = get_cycles(); | |
703 | ||
704 | /* initialize buddy from bitmap which is aggregation | |
705 | * of on-disk bitmap and preallocations */ | |
706 | i = mb_find_next_zero_bit(bitmap, max, 0); | |
707 | grp->bb_first_free = i; | |
708 | while (i < max) { | |
709 | fragments++; | |
710 | first = i; | |
711 | i = mb_find_next_bit(bitmap, max, i); | |
712 | len = i - first; | |
713 | free += len; | |
714 | if (len > 1) | |
715 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
716 | else | |
717 | grp->bb_counters[0]++; | |
718 | if (i < max) | |
719 | i = mb_find_next_zero_bit(bitmap, max, i); | |
720 | } | |
721 | grp->bb_fragments = fragments; | |
722 | ||
723 | if (free != grp->bb_free) { | |
724 | ext4_grp_locked_error(sb, group, 0, 0, | |
725 | "%u blocks in bitmap, %u in gd", | |
726 | free, grp->bb_free); | |
727 | /* | |
728 | * If we intent to continue, we consider group descritor | |
729 | * corrupt and update bb_free using bitmap value | |
730 | */ | |
731 | grp->bb_free = free; | |
732 | } | |
733 | mb_set_largest_free_order(sb, grp); | |
734 | ||
735 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
736 | ||
737 | period = get_cycles() - period; | |
738 | spin_lock(&EXT4_SB(sb)->s_bal_lock); | |
739 | EXT4_SB(sb)->s_mb_buddies_generated++; | |
740 | EXT4_SB(sb)->s_mb_generation_time += period; | |
741 | spin_unlock(&EXT4_SB(sb)->s_bal_lock); | |
742 | } | |
743 | ||
744 | /* The buddy information is attached the buddy cache inode | |
745 | * for convenience. The information regarding each group | |
746 | * is loaded via ext4_mb_load_buddy. The information involve | |
747 | * block bitmap and buddy information. The information are | |
748 | * stored in the inode as | |
749 | * | |
750 | * { page } | |
751 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... | |
752 | * | |
753 | * | |
754 | * one block each for bitmap and buddy information. | |
755 | * So for each group we take up 2 blocks. A page can | |
756 | * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. | |
757 | * So it can have information regarding groups_per_page which | |
758 | * is blocks_per_page/2 | |
759 | * | |
760 | * Locking note: This routine takes the block group lock of all groups | |
761 | * for this page; do not hold this lock when calling this routine! | |
762 | */ | |
763 | ||
764 | static int ext4_mb_init_cache(struct page *page, char *incore) | |
765 | { | |
766 | ext4_group_t ngroups; | |
767 | int blocksize; | |
768 | int blocks_per_page; | |
769 | int groups_per_page; | |
770 | int err = 0; | |
771 | int i; | |
772 | ext4_group_t first_group; | |
773 | int first_block; | |
774 | struct super_block *sb; | |
775 | struct buffer_head *bhs; | |
776 | struct buffer_head **bh; | |
777 | struct inode *inode; | |
778 | char *data; | |
779 | char *bitmap; | |
780 | ||
781 | mb_debug(1, "init page %lu\n", page->index); | |
782 | ||
783 | inode = page->mapping->host; | |
784 | sb = inode->i_sb; | |
785 | ngroups = ext4_get_groups_count(sb); | |
786 | blocksize = 1 << inode->i_blkbits; | |
787 | blocks_per_page = PAGE_CACHE_SIZE / blocksize; | |
788 | ||
789 | groups_per_page = blocks_per_page >> 1; | |
790 | if (groups_per_page == 0) | |
791 | groups_per_page = 1; | |
792 | ||
793 | /* allocate buffer_heads to read bitmaps */ | |
794 | if (groups_per_page > 1) { | |
795 | err = -ENOMEM; | |
796 | i = sizeof(struct buffer_head *) * groups_per_page; | |
797 | bh = kzalloc(i, GFP_NOFS); | |
798 | if (bh == NULL) | |
799 | goto out; | |
800 | } else | |
801 | bh = &bhs; | |
802 | ||
803 | first_group = page->index * blocks_per_page / 2; | |
804 | ||
805 | /* read all groups the page covers into the cache */ | |
806 | for (i = 0; i < groups_per_page; i++) { | |
807 | struct ext4_group_desc *desc; | |
808 | ||
809 | if (first_group + i >= ngroups) | |
810 | break; | |
811 | ||
812 | err = -EIO; | |
813 | desc = ext4_get_group_desc(sb, first_group + i, NULL); | |
814 | if (desc == NULL) | |
815 | goto out; | |
816 | ||
817 | err = -ENOMEM; | |
818 | bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc)); | |
819 | if (bh[i] == NULL) | |
820 | goto out; | |
821 | ||
822 | if (bitmap_uptodate(bh[i])) | |
823 | continue; | |
824 | ||
825 | lock_buffer(bh[i]); | |
826 | if (bitmap_uptodate(bh[i])) { | |
827 | unlock_buffer(bh[i]); | |
828 | continue; | |
829 | } | |
830 | ext4_lock_group(sb, first_group + i); | |
831 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
832 | ext4_init_block_bitmap(sb, bh[i], | |
833 | first_group + i, desc); | |
834 | set_bitmap_uptodate(bh[i]); | |
835 | set_buffer_uptodate(bh[i]); | |
836 | ext4_unlock_group(sb, first_group + i); | |
837 | unlock_buffer(bh[i]); | |
838 | continue; | |
839 | } | |
840 | ext4_unlock_group(sb, first_group + i); | |
841 | if (buffer_uptodate(bh[i])) { | |
842 | /* | |
843 | * if not uninit if bh is uptodate, | |
844 | * bitmap is also uptodate | |
845 | */ | |
846 | set_bitmap_uptodate(bh[i]); | |
847 | unlock_buffer(bh[i]); | |
848 | continue; | |
849 | } | |
850 | get_bh(bh[i]); | |
851 | /* | |
852 | * submit the buffer_head for read. We can | |
853 | * safely mark the bitmap as uptodate now. | |
854 | * We do it here so the bitmap uptodate bit | |
855 | * get set with buffer lock held. | |
856 | */ | |
857 | set_bitmap_uptodate(bh[i]); | |
858 | bh[i]->b_end_io = end_buffer_read_sync; | |
859 | submit_bh(READ, bh[i]); | |
860 | mb_debug(1, "read bitmap for group %u\n", first_group + i); | |
861 | } | |
862 | ||
863 | /* wait for I/O completion */ | |
864 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
865 | wait_on_buffer(bh[i]); | |
866 | ||
867 | err = -EIO; | |
868 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
869 | if (!buffer_uptodate(bh[i])) | |
870 | goto out; | |
871 | ||
872 | err = 0; | |
873 | first_block = page->index * blocks_per_page; | |
874 | /* init the page */ | |
875 | memset(page_address(page), 0xff, PAGE_CACHE_SIZE); | |
876 | for (i = 0; i < blocks_per_page; i++) { | |
877 | int group; | |
878 | struct ext4_group_info *grinfo; | |
879 | ||
880 | group = (first_block + i) >> 1; | |
881 | if (group >= ngroups) | |
882 | break; | |
883 | ||
884 | /* | |
885 | * data carry information regarding this | |
886 | * particular group in the format specified | |
887 | * above | |
888 | * | |
889 | */ | |
890 | data = page_address(page) + (i * blocksize); | |
891 | bitmap = bh[group - first_group]->b_data; | |
892 | ||
893 | /* | |
894 | * We place the buddy block and bitmap block | |
895 | * close together | |
896 | */ | |
897 | if ((first_block + i) & 1) { | |
898 | /* this is block of buddy */ | |
899 | BUG_ON(incore == NULL); | |
900 | mb_debug(1, "put buddy for group %u in page %lu/%x\n", | |
901 | group, page->index, i * blocksize); | |
902 | trace_ext4_mb_buddy_bitmap_load(sb, group); | |
903 | grinfo = ext4_get_group_info(sb, group); | |
904 | grinfo->bb_fragments = 0; | |
905 | memset(grinfo->bb_counters, 0, | |
906 | sizeof(*grinfo->bb_counters) * | |
907 | (sb->s_blocksize_bits+2)); | |
908 | /* | |
909 | * incore got set to the group block bitmap below | |
910 | */ | |
911 | ext4_lock_group(sb, group); | |
912 | ext4_mb_generate_buddy(sb, data, incore, group); | |
913 | ext4_unlock_group(sb, group); | |
914 | incore = NULL; | |
915 | } else { | |
916 | /* this is block of bitmap */ | |
917 | BUG_ON(incore != NULL); | |
918 | mb_debug(1, "put bitmap for group %u in page %lu/%x\n", | |
919 | group, page->index, i * blocksize); | |
920 | trace_ext4_mb_bitmap_load(sb, group); | |
921 | ||
922 | /* see comments in ext4_mb_put_pa() */ | |
923 | ext4_lock_group(sb, group); | |
924 | memcpy(data, bitmap, blocksize); | |
925 | ||
926 | /* mark all preallocated blks used in in-core bitmap */ | |
927 | ext4_mb_generate_from_pa(sb, data, group); | |
928 | ext4_mb_generate_from_freelist(sb, data, group); | |
929 | ext4_unlock_group(sb, group); | |
930 | ||
931 | /* set incore so that the buddy information can be | |
932 | * generated using this | |
933 | */ | |
934 | incore = data; | |
935 | } | |
936 | } | |
937 | SetPageUptodate(page); | |
938 | ||
939 | out: | |
940 | if (bh) { | |
941 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
942 | brelse(bh[i]); | |
943 | if (bh != &bhs) | |
944 | kfree(bh); | |
945 | } | |
946 | return err; | |
947 | } | |
948 | ||
949 | /* | |
950 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
951 | * block group lock of all groups for this page; do not hold the BG lock when | |
952 | * calling this routine! | |
953 | */ | |
954 | static noinline_for_stack | |
955 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) | |
956 | { | |
957 | ||
958 | int ret = 0; | |
959 | void *bitmap; | |
960 | int blocks_per_page; | |
961 | int block, pnum, poff; | |
962 | int num_grp_locked = 0; | |
963 | struct ext4_group_info *this_grp; | |
964 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
965 | struct inode *inode = sbi->s_buddy_cache; | |
966 | struct page *page = NULL, *bitmap_page = NULL; | |
967 | ||
968 | mb_debug(1, "init group %u\n", group); | |
969 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
970 | this_grp = ext4_get_group_info(sb, group); | |
971 | /* | |
972 | * This ensures that we don't reinit the buddy cache | |
973 | * page which map to the group from which we are already | |
974 | * allocating. If we are looking at the buddy cache we would | |
975 | * have taken a reference using ext4_mb_load_buddy and that | |
976 | * would have taken the alloc_sem lock. | |
977 | */ | |
978 | num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group); | |
979 | if (!EXT4_MB_GRP_NEED_INIT(this_grp)) { | |
980 | /* | |
981 | * somebody initialized the group | |
982 | * return without doing anything | |
983 | */ | |
984 | ret = 0; | |
985 | goto err; | |
986 | } | |
987 | /* | |
988 | * the buddy cache inode stores the block bitmap | |
989 | * and buddy information in consecutive blocks. | |
990 | * So for each group we need two blocks. | |
991 | */ | |
992 | block = group * 2; | |
993 | pnum = block / blocks_per_page; | |
994 | poff = block % blocks_per_page; | |
995 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
996 | if (page) { | |
997 | BUG_ON(page->mapping != inode->i_mapping); | |
998 | ret = ext4_mb_init_cache(page, NULL); | |
999 | if (ret) { | |
1000 | unlock_page(page); | |
1001 | goto err; | |
1002 | } | |
1003 | unlock_page(page); | |
1004 | } | |
1005 | if (page == NULL || !PageUptodate(page)) { | |
1006 | ret = -EIO; | |
1007 | goto err; | |
1008 | } | |
1009 | mark_page_accessed(page); | |
1010 | bitmap_page = page; | |
1011 | bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1012 | ||
1013 | /* init buddy cache */ | |
1014 | block++; | |
1015 | pnum = block / blocks_per_page; | |
1016 | poff = block % blocks_per_page; | |
1017 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1018 | if (page == bitmap_page) { | |
1019 | /* | |
1020 | * If both the bitmap and buddy are in | |
1021 | * the same page we don't need to force | |
1022 | * init the buddy | |
1023 | */ | |
1024 | unlock_page(page); | |
1025 | } else if (page) { | |
1026 | BUG_ON(page->mapping != inode->i_mapping); | |
1027 | ret = ext4_mb_init_cache(page, bitmap); | |
1028 | if (ret) { | |
1029 | unlock_page(page); | |
1030 | goto err; | |
1031 | } | |
1032 | unlock_page(page); | |
1033 | } | |
1034 | if (page == NULL || !PageUptodate(page)) { | |
1035 | ret = -EIO; | |
1036 | goto err; | |
1037 | } | |
1038 | mark_page_accessed(page); | |
1039 | err: | |
1040 | ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked); | |
1041 | if (bitmap_page) | |
1042 | page_cache_release(bitmap_page); | |
1043 | if (page) | |
1044 | page_cache_release(page); | |
1045 | return ret; | |
1046 | } | |
1047 | ||
1048 | /* | |
1049 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1050 | * block group lock of all groups for this page; do not hold the BG lock when | |
1051 | * calling this routine! | |
1052 | */ | |
1053 | static noinline_for_stack int | |
1054 | ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, | |
1055 | struct ext4_buddy *e4b) | |
1056 | { | |
1057 | int blocks_per_page; | |
1058 | int block; | |
1059 | int pnum; | |
1060 | int poff; | |
1061 | struct page *page; | |
1062 | int ret; | |
1063 | struct ext4_group_info *grp; | |
1064 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1065 | struct inode *inode = sbi->s_buddy_cache; | |
1066 | ||
1067 | mb_debug(1, "load group %u\n", group); | |
1068 | ||
1069 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1070 | grp = ext4_get_group_info(sb, group); | |
1071 | ||
1072 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
1073 | e4b->bd_info = ext4_get_group_info(sb, group); | |
1074 | e4b->bd_sb = sb; | |
1075 | e4b->bd_group = group; | |
1076 | e4b->bd_buddy_page = NULL; | |
1077 | e4b->bd_bitmap_page = NULL; | |
1078 | e4b->alloc_semp = &grp->alloc_sem; | |
1079 | ||
1080 | /* Take the read lock on the group alloc | |
1081 | * sem. This would make sure a parallel | |
1082 | * ext4_mb_init_group happening on other | |
1083 | * groups mapped by the page is blocked | |
1084 | * till we are done with allocation | |
1085 | */ | |
1086 | repeat_load_buddy: | |
1087 | down_read(e4b->alloc_semp); | |
1088 | ||
1089 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
1090 | /* we need to check for group need init flag | |
1091 | * with alloc_semp held so that we can be sure | |
1092 | * that new blocks didn't get added to the group | |
1093 | * when we are loading the buddy cache | |
1094 | */ | |
1095 | up_read(e4b->alloc_semp); | |
1096 | /* | |
1097 | * we need full data about the group | |
1098 | * to make a good selection | |
1099 | */ | |
1100 | ret = ext4_mb_init_group(sb, group); | |
1101 | if (ret) | |
1102 | return ret; | |
1103 | goto repeat_load_buddy; | |
1104 | } | |
1105 | ||
1106 | /* | |
1107 | * the buddy cache inode stores the block bitmap | |
1108 | * and buddy information in consecutive blocks. | |
1109 | * So for each group we need two blocks. | |
1110 | */ | |
1111 | block = group * 2; | |
1112 | pnum = block / blocks_per_page; | |
1113 | poff = block % blocks_per_page; | |
1114 | ||
1115 | /* we could use find_or_create_page(), but it locks page | |
1116 | * what we'd like to avoid in fast path ... */ | |
1117 | page = find_get_page(inode->i_mapping, pnum); | |
1118 | if (page == NULL || !PageUptodate(page)) { | |
1119 | if (page) | |
1120 | /* | |
1121 | * drop the page reference and try | |
1122 | * to get the page with lock. If we | |
1123 | * are not uptodate that implies | |
1124 | * somebody just created the page but | |
1125 | * is yet to initialize the same. So | |
1126 | * wait for it to initialize. | |
1127 | */ | |
1128 | page_cache_release(page); | |
1129 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1130 | if (page) { | |
1131 | BUG_ON(page->mapping != inode->i_mapping); | |
1132 | if (!PageUptodate(page)) { | |
1133 | ret = ext4_mb_init_cache(page, NULL); | |
1134 | if (ret) { | |
1135 | unlock_page(page); | |
1136 | goto err; | |
1137 | } | |
1138 | mb_cmp_bitmaps(e4b, page_address(page) + | |
1139 | (poff * sb->s_blocksize)); | |
1140 | } | |
1141 | unlock_page(page); | |
1142 | } | |
1143 | } | |
1144 | if (page == NULL || !PageUptodate(page)) { | |
1145 | ret = -EIO; | |
1146 | goto err; | |
1147 | } | |
1148 | e4b->bd_bitmap_page = page; | |
1149 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1150 | mark_page_accessed(page); | |
1151 | ||
1152 | block++; | |
1153 | pnum = block / blocks_per_page; | |
1154 | poff = block % blocks_per_page; | |
1155 | ||
1156 | page = find_get_page(inode->i_mapping, pnum); | |
1157 | if (page == NULL || !PageUptodate(page)) { | |
1158 | if (page) | |
1159 | page_cache_release(page); | |
1160 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1161 | if (page) { | |
1162 | BUG_ON(page->mapping != inode->i_mapping); | |
1163 | if (!PageUptodate(page)) { | |
1164 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap); | |
1165 | if (ret) { | |
1166 | unlock_page(page); | |
1167 | goto err; | |
1168 | } | |
1169 | } | |
1170 | unlock_page(page); | |
1171 | } | |
1172 | } | |
1173 | if (page == NULL || !PageUptodate(page)) { | |
1174 | ret = -EIO; | |
1175 | goto err; | |
1176 | } | |
1177 | e4b->bd_buddy_page = page; | |
1178 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
1179 | mark_page_accessed(page); | |
1180 | ||
1181 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1182 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1183 | ||
1184 | return 0; | |
1185 | ||
1186 | err: | |
1187 | if (e4b->bd_bitmap_page) | |
1188 | page_cache_release(e4b->bd_bitmap_page); | |
1189 | if (e4b->bd_buddy_page) | |
1190 | page_cache_release(e4b->bd_buddy_page); | |
1191 | e4b->bd_buddy = NULL; | |
1192 | e4b->bd_bitmap = NULL; | |
1193 | ||
1194 | /* Done with the buddy cache */ | |
1195 | up_read(e4b->alloc_semp); | |
1196 | return ret; | |
1197 | } | |
1198 | ||
1199 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) | |
1200 | { | |
1201 | if (e4b->bd_bitmap_page) | |
1202 | page_cache_release(e4b->bd_bitmap_page); | |
1203 | if (e4b->bd_buddy_page) | |
1204 | page_cache_release(e4b->bd_buddy_page); | |
1205 | /* Done with the buddy cache */ | |
1206 | if (e4b->alloc_semp) | |
1207 | up_read(e4b->alloc_semp); | |
1208 | } | |
1209 | ||
1210 | ||
1211 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1212 | { | |
1213 | int order = 1; | |
1214 | void *bb; | |
1215 | ||
1216 | BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); | |
1217 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); | |
1218 | ||
1219 | bb = EXT4_MB_BUDDY(e4b); | |
1220 | while (order <= e4b->bd_blkbits + 1) { | |
1221 | block = block >> 1; | |
1222 | if (!mb_test_bit(block, bb)) { | |
1223 | /* this block is part of buddy of order 'order' */ | |
1224 | return order; | |
1225 | } | |
1226 | bb += 1 << (e4b->bd_blkbits - order); | |
1227 | order++; | |
1228 | } | |
1229 | return 0; | |
1230 | } | |
1231 | ||
1232 | static void mb_clear_bits(void *bm, int cur, int len) | |
1233 | { | |
1234 | __u32 *addr; | |
1235 | ||
1236 | len = cur + len; | |
1237 | while (cur < len) { | |
1238 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1239 | /* fast path: clear whole word at once */ | |
1240 | addr = bm + (cur >> 3); | |
1241 | *addr = 0; | |
1242 | cur += 32; | |
1243 | continue; | |
1244 | } | |
1245 | mb_clear_bit(cur, bm); | |
1246 | cur++; | |
1247 | } | |
1248 | } | |
1249 | ||
1250 | static void mb_set_bits(void *bm, int cur, int len) | |
1251 | { | |
1252 | __u32 *addr; | |
1253 | ||
1254 | len = cur + len; | |
1255 | while (cur < len) { | |
1256 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1257 | /* fast path: set whole word at once */ | |
1258 | addr = bm + (cur >> 3); | |
1259 | *addr = 0xffffffff; | |
1260 | cur += 32; | |
1261 | continue; | |
1262 | } | |
1263 | mb_set_bit(cur, bm); | |
1264 | cur++; | |
1265 | } | |
1266 | } | |
1267 | ||
1268 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, | |
1269 | int first, int count) | |
1270 | { | |
1271 | int block = 0; | |
1272 | int max = 0; | |
1273 | int order; | |
1274 | void *buddy; | |
1275 | void *buddy2; | |
1276 | struct super_block *sb = e4b->bd_sb; | |
1277 | ||
1278 | BUG_ON(first + count > (sb->s_blocksize << 3)); | |
1279 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); | |
1280 | mb_check_buddy(e4b); | |
1281 | mb_free_blocks_double(inode, e4b, first, count); | |
1282 | ||
1283 | e4b->bd_info->bb_free += count; | |
1284 | if (first < e4b->bd_info->bb_first_free) | |
1285 | e4b->bd_info->bb_first_free = first; | |
1286 | ||
1287 | /* let's maintain fragments counter */ | |
1288 | if (first != 0) | |
1289 | block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b)); | |
1290 | if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) | |
1291 | max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b)); | |
1292 | if (block && max) | |
1293 | e4b->bd_info->bb_fragments--; | |
1294 | else if (!block && !max) | |
1295 | e4b->bd_info->bb_fragments++; | |
1296 | ||
1297 | /* let's maintain buddy itself */ | |
1298 | while (count-- > 0) { | |
1299 | block = first++; | |
1300 | order = 0; | |
1301 | ||
1302 | if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) { | |
1303 | ext4_fsblk_t blocknr; | |
1304 | ||
1305 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
1306 | blocknr += block; | |
1307 | ext4_grp_locked_error(sb, e4b->bd_group, | |
1308 | inode ? inode->i_ino : 0, | |
1309 | blocknr, | |
1310 | "freeing already freed block " | |
1311 | "(bit %u)", block); | |
1312 | } | |
1313 | mb_clear_bit(block, EXT4_MB_BITMAP(e4b)); | |
1314 | e4b->bd_info->bb_counters[order]++; | |
1315 | ||
1316 | /* start of the buddy */ | |
1317 | buddy = mb_find_buddy(e4b, order, &max); | |
1318 | ||
1319 | do { | |
1320 | block &= ~1UL; | |
1321 | if (mb_test_bit(block, buddy) || | |
1322 | mb_test_bit(block + 1, buddy)) | |
1323 | break; | |
1324 | ||
1325 | /* both the buddies are free, try to coalesce them */ | |
1326 | buddy2 = mb_find_buddy(e4b, order + 1, &max); | |
1327 | ||
1328 | if (!buddy2) | |
1329 | break; | |
1330 | ||
1331 | if (order > 0) { | |
1332 | /* for special purposes, we don't set | |
1333 | * free bits in bitmap */ | |
1334 | mb_set_bit(block, buddy); | |
1335 | mb_set_bit(block + 1, buddy); | |
1336 | } | |
1337 | e4b->bd_info->bb_counters[order]--; | |
1338 | e4b->bd_info->bb_counters[order]--; | |
1339 | ||
1340 | block = block >> 1; | |
1341 | order++; | |
1342 | e4b->bd_info->bb_counters[order]++; | |
1343 | ||
1344 | mb_clear_bit(block, buddy2); | |
1345 | buddy = buddy2; | |
1346 | } while (1); | |
1347 | } | |
1348 | mb_set_largest_free_order(sb, e4b->bd_info); | |
1349 | mb_check_buddy(e4b); | |
1350 | } | |
1351 | ||
1352 | static int mb_find_extent(struct ext4_buddy *e4b, int order, int block, | |
1353 | int needed, struct ext4_free_extent *ex) | |
1354 | { | |
1355 | int next = block; | |
1356 | int max; | |
1357 | int ord; | |
1358 | void *buddy; | |
1359 | ||
1360 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); | |
1361 | BUG_ON(ex == NULL); | |
1362 | ||
1363 | buddy = mb_find_buddy(e4b, order, &max); | |
1364 | BUG_ON(buddy == NULL); | |
1365 | BUG_ON(block >= max); | |
1366 | if (mb_test_bit(block, buddy)) { | |
1367 | ex->fe_len = 0; | |
1368 | ex->fe_start = 0; | |
1369 | ex->fe_group = 0; | |
1370 | return 0; | |
1371 | } | |
1372 | ||
1373 | /* FIXME dorp order completely ? */ | |
1374 | if (likely(order == 0)) { | |
1375 | /* find actual order */ | |
1376 | order = mb_find_order_for_block(e4b, block); | |
1377 | block = block >> order; | |
1378 | } | |
1379 | ||
1380 | ex->fe_len = 1 << order; | |
1381 | ex->fe_start = block << order; | |
1382 | ex->fe_group = e4b->bd_group; | |
1383 | ||
1384 | /* calc difference from given start */ | |
1385 | next = next - ex->fe_start; | |
1386 | ex->fe_len -= next; | |
1387 | ex->fe_start += next; | |
1388 | ||
1389 | while (needed > ex->fe_len && | |
1390 | (buddy = mb_find_buddy(e4b, order, &max))) { | |
1391 | ||
1392 | if (block + 1 >= max) | |
1393 | break; | |
1394 | ||
1395 | next = (block + 1) * (1 << order); | |
1396 | if (mb_test_bit(next, EXT4_MB_BITMAP(e4b))) | |
1397 | break; | |
1398 | ||
1399 | ord = mb_find_order_for_block(e4b, next); | |
1400 | ||
1401 | order = ord; | |
1402 | block = next >> order; | |
1403 | ex->fe_len += 1 << order; | |
1404 | } | |
1405 | ||
1406 | BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); | |
1407 | return ex->fe_len; | |
1408 | } | |
1409 | ||
1410 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1411 | { | |
1412 | int ord; | |
1413 | int mlen = 0; | |
1414 | int max = 0; | |
1415 | int cur; | |
1416 | int start = ex->fe_start; | |
1417 | int len = ex->fe_len; | |
1418 | unsigned ret = 0; | |
1419 | int len0 = len; | |
1420 | void *buddy; | |
1421 | ||
1422 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1423 | BUG_ON(e4b->bd_group != ex->fe_group); | |
1424 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); | |
1425 | mb_check_buddy(e4b); | |
1426 | mb_mark_used_double(e4b, start, len); | |
1427 | ||
1428 | e4b->bd_info->bb_free -= len; | |
1429 | if (e4b->bd_info->bb_first_free == start) | |
1430 | e4b->bd_info->bb_first_free += len; | |
1431 | ||
1432 | /* let's maintain fragments counter */ | |
1433 | if (start != 0) | |
1434 | mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b)); | |
1435 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) | |
1436 | max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b)); | |
1437 | if (mlen && max) | |
1438 | e4b->bd_info->bb_fragments++; | |
1439 | else if (!mlen && !max) | |
1440 | e4b->bd_info->bb_fragments--; | |
1441 | ||
1442 | /* let's maintain buddy itself */ | |
1443 | while (len) { | |
1444 | ord = mb_find_order_for_block(e4b, start); | |
1445 | ||
1446 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1447 | /* the whole chunk may be allocated at once! */ | |
1448 | mlen = 1 << ord; | |
1449 | buddy = mb_find_buddy(e4b, ord, &max); | |
1450 | BUG_ON((start >> ord) >= max); | |
1451 | mb_set_bit(start >> ord, buddy); | |
1452 | e4b->bd_info->bb_counters[ord]--; | |
1453 | start += mlen; | |
1454 | len -= mlen; | |
1455 | BUG_ON(len < 0); | |
1456 | continue; | |
1457 | } | |
1458 | ||
1459 | /* store for history */ | |
1460 | if (ret == 0) | |
1461 | ret = len | (ord << 16); | |
1462 | ||
1463 | /* we have to split large buddy */ | |
1464 | BUG_ON(ord <= 0); | |
1465 | buddy = mb_find_buddy(e4b, ord, &max); | |
1466 | mb_set_bit(start >> ord, buddy); | |
1467 | e4b->bd_info->bb_counters[ord]--; | |
1468 | ||
1469 | ord--; | |
1470 | cur = (start >> ord) & ~1U; | |
1471 | buddy = mb_find_buddy(e4b, ord, &max); | |
1472 | mb_clear_bit(cur, buddy); | |
1473 | mb_clear_bit(cur + 1, buddy); | |
1474 | e4b->bd_info->bb_counters[ord]++; | |
1475 | e4b->bd_info->bb_counters[ord]++; | |
1476 | } | |
1477 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); | |
1478 | ||
1479 | mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0); | |
1480 | mb_check_buddy(e4b); | |
1481 | ||
1482 | return ret; | |
1483 | } | |
1484 | ||
1485 | /* | |
1486 | * Must be called under group lock! | |
1487 | */ | |
1488 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1489 | struct ext4_buddy *e4b) | |
1490 | { | |
1491 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1492 | int ret; | |
1493 | ||
1494 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1495 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1496 | ||
1497 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1498 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1499 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1500 | ||
1501 | /* preallocation can change ac_b_ex, thus we store actually | |
1502 | * allocated blocks for history */ | |
1503 | ac->ac_f_ex = ac->ac_b_ex; | |
1504 | ||
1505 | ac->ac_status = AC_STATUS_FOUND; | |
1506 | ac->ac_tail = ret & 0xffff; | |
1507 | ac->ac_buddy = ret >> 16; | |
1508 | ||
1509 | /* | |
1510 | * take the page reference. We want the page to be pinned | |
1511 | * so that we don't get a ext4_mb_init_cache_call for this | |
1512 | * group until we update the bitmap. That would mean we | |
1513 | * double allocate blocks. The reference is dropped | |
1514 | * in ext4_mb_release_context | |
1515 | */ | |
1516 | ac->ac_bitmap_page = e4b->bd_bitmap_page; | |
1517 | get_page(ac->ac_bitmap_page); | |
1518 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1519 | get_page(ac->ac_buddy_page); | |
1520 | /* on allocation we use ac to track the held semaphore */ | |
1521 | ac->alloc_semp = e4b->alloc_semp; | |
1522 | e4b->alloc_semp = NULL; | |
1523 | /* store last allocated for subsequent stream allocation */ | |
1524 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
1525 | spin_lock(&sbi->s_md_lock); | |
1526 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1527 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1528 | spin_unlock(&sbi->s_md_lock); | |
1529 | } | |
1530 | } | |
1531 | ||
1532 | /* | |
1533 | * regular allocator, for general purposes allocation | |
1534 | */ | |
1535 | ||
1536 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1537 | struct ext4_buddy *e4b, | |
1538 | int finish_group) | |
1539 | { | |
1540 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1541 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1542 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1543 | struct ext4_free_extent ex; | |
1544 | int max; | |
1545 | ||
1546 | if (ac->ac_status == AC_STATUS_FOUND) | |
1547 | return; | |
1548 | /* | |
1549 | * We don't want to scan for a whole year | |
1550 | */ | |
1551 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1552 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1553 | ac->ac_status = AC_STATUS_BREAK; | |
1554 | return; | |
1555 | } | |
1556 | ||
1557 | /* | |
1558 | * Haven't found good chunk so far, let's continue | |
1559 | */ | |
1560 | if (bex->fe_len < gex->fe_len) | |
1561 | return; | |
1562 | ||
1563 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1564 | && bex->fe_group == e4b->bd_group) { | |
1565 | /* recheck chunk's availability - we don't know | |
1566 | * when it was found (within this lock-unlock | |
1567 | * period or not) */ | |
1568 | max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex); | |
1569 | if (max >= gex->fe_len) { | |
1570 | ext4_mb_use_best_found(ac, e4b); | |
1571 | return; | |
1572 | } | |
1573 | } | |
1574 | } | |
1575 | ||
1576 | /* | |
1577 | * The routine checks whether found extent is good enough. If it is, | |
1578 | * then the extent gets marked used and flag is set to the context | |
1579 | * to stop scanning. Otherwise, the extent is compared with the | |
1580 | * previous found extent and if new one is better, then it's stored | |
1581 | * in the context. Later, the best found extent will be used, if | |
1582 | * mballoc can't find good enough extent. | |
1583 | * | |
1584 | * FIXME: real allocation policy is to be designed yet! | |
1585 | */ | |
1586 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1587 | struct ext4_free_extent *ex, | |
1588 | struct ext4_buddy *e4b) | |
1589 | { | |
1590 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1591 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1592 | ||
1593 | BUG_ON(ex->fe_len <= 0); | |
1594 | BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); | |
1595 | BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); | |
1596 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); | |
1597 | ||
1598 | ac->ac_found++; | |
1599 | ||
1600 | /* | |
1601 | * The special case - take what you catch first | |
1602 | */ | |
1603 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1604 | *bex = *ex; | |
1605 | ext4_mb_use_best_found(ac, e4b); | |
1606 | return; | |
1607 | } | |
1608 | ||
1609 | /* | |
1610 | * Let's check whether the chuck is good enough | |
1611 | */ | |
1612 | if (ex->fe_len == gex->fe_len) { | |
1613 | *bex = *ex; | |
1614 | ext4_mb_use_best_found(ac, e4b); | |
1615 | return; | |
1616 | } | |
1617 | ||
1618 | /* | |
1619 | * If this is first found extent, just store it in the context | |
1620 | */ | |
1621 | if (bex->fe_len == 0) { | |
1622 | *bex = *ex; | |
1623 | return; | |
1624 | } | |
1625 | ||
1626 | /* | |
1627 | * If new found extent is better, store it in the context | |
1628 | */ | |
1629 | if (bex->fe_len < gex->fe_len) { | |
1630 | /* if the request isn't satisfied, any found extent | |
1631 | * larger than previous best one is better */ | |
1632 | if (ex->fe_len > bex->fe_len) | |
1633 | *bex = *ex; | |
1634 | } else if (ex->fe_len > gex->fe_len) { | |
1635 | /* if the request is satisfied, then we try to find | |
1636 | * an extent that still satisfy the request, but is | |
1637 | * smaller than previous one */ | |
1638 | if (ex->fe_len < bex->fe_len) | |
1639 | *bex = *ex; | |
1640 | } | |
1641 | ||
1642 | ext4_mb_check_limits(ac, e4b, 0); | |
1643 | } | |
1644 | ||
1645 | static noinline_for_stack | |
1646 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
1647 | struct ext4_buddy *e4b) | |
1648 | { | |
1649 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1650 | ext4_group_t group = ex.fe_group; | |
1651 | int max; | |
1652 | int err; | |
1653 | ||
1654 | BUG_ON(ex.fe_len <= 0); | |
1655 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1656 | if (err) | |
1657 | return err; | |
1658 | ||
1659 | ext4_lock_group(ac->ac_sb, group); | |
1660 | max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex); | |
1661 | ||
1662 | if (max > 0) { | |
1663 | ac->ac_b_ex = ex; | |
1664 | ext4_mb_use_best_found(ac, e4b); | |
1665 | } | |
1666 | ||
1667 | ext4_unlock_group(ac->ac_sb, group); | |
1668 | ext4_mb_unload_buddy(e4b); | |
1669 | ||
1670 | return 0; | |
1671 | } | |
1672 | ||
1673 | static noinline_for_stack | |
1674 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
1675 | struct ext4_buddy *e4b) | |
1676 | { | |
1677 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1678 | int max; | |
1679 | int err; | |
1680 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1681 | struct ext4_free_extent ex; | |
1682 | ||
1683 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1684 | return 0; | |
1685 | ||
1686 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1687 | if (err) | |
1688 | return err; | |
1689 | ||
1690 | ext4_lock_group(ac->ac_sb, group); | |
1691 | max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start, | |
1692 | ac->ac_g_ex.fe_len, &ex); | |
1693 | ||
1694 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1695 | ext4_fsblk_t start; | |
1696 | ||
1697 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + | |
1698 | ex.fe_start; | |
1699 | /* use do_div to get remainder (would be 64-bit modulo) */ | |
1700 | if (do_div(start, sbi->s_stripe) == 0) { | |
1701 | ac->ac_found++; | |
1702 | ac->ac_b_ex = ex; | |
1703 | ext4_mb_use_best_found(ac, e4b); | |
1704 | } | |
1705 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1706 | BUG_ON(ex.fe_len <= 0); | |
1707 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1708 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1709 | ac->ac_found++; | |
1710 | ac->ac_b_ex = ex; | |
1711 | ext4_mb_use_best_found(ac, e4b); | |
1712 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1713 | /* Sometimes, caller may want to merge even small | |
1714 | * number of blocks to an existing extent */ | |
1715 | BUG_ON(ex.fe_len <= 0); | |
1716 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1717 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1718 | ac->ac_found++; | |
1719 | ac->ac_b_ex = ex; | |
1720 | ext4_mb_use_best_found(ac, e4b); | |
1721 | } | |
1722 | ext4_unlock_group(ac->ac_sb, group); | |
1723 | ext4_mb_unload_buddy(e4b); | |
1724 | ||
1725 | return 0; | |
1726 | } | |
1727 | ||
1728 | /* | |
1729 | * The routine scans buddy structures (not bitmap!) from given order | |
1730 | * to max order and tries to find big enough chunk to satisfy the req | |
1731 | */ | |
1732 | static noinline_for_stack | |
1733 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
1734 | struct ext4_buddy *e4b) | |
1735 | { | |
1736 | struct super_block *sb = ac->ac_sb; | |
1737 | struct ext4_group_info *grp = e4b->bd_info; | |
1738 | void *buddy; | |
1739 | int i; | |
1740 | int k; | |
1741 | int max; | |
1742 | ||
1743 | BUG_ON(ac->ac_2order <= 0); | |
1744 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1745 | if (grp->bb_counters[i] == 0) | |
1746 | continue; | |
1747 | ||
1748 | buddy = mb_find_buddy(e4b, i, &max); | |
1749 | BUG_ON(buddy == NULL); | |
1750 | ||
1751 | k = mb_find_next_zero_bit(buddy, max, 0); | |
1752 | BUG_ON(k >= max); | |
1753 | ||
1754 | ac->ac_found++; | |
1755 | ||
1756 | ac->ac_b_ex.fe_len = 1 << i; | |
1757 | ac->ac_b_ex.fe_start = k << i; | |
1758 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1759 | ||
1760 | ext4_mb_use_best_found(ac, e4b); | |
1761 | ||
1762 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1763 | ||
1764 | if (EXT4_SB(sb)->s_mb_stats) | |
1765 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1766 | ||
1767 | break; | |
1768 | } | |
1769 | } | |
1770 | ||
1771 | /* | |
1772 | * The routine scans the group and measures all found extents. | |
1773 | * In order to optimize scanning, caller must pass number of | |
1774 | * free blocks in the group, so the routine can know upper limit. | |
1775 | */ | |
1776 | static noinline_for_stack | |
1777 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
1778 | struct ext4_buddy *e4b) | |
1779 | { | |
1780 | struct super_block *sb = ac->ac_sb; | |
1781 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1782 | struct ext4_free_extent ex; | |
1783 | int i; | |
1784 | int free; | |
1785 | ||
1786 | free = e4b->bd_info->bb_free; | |
1787 | BUG_ON(free <= 0); | |
1788 | ||
1789 | i = e4b->bd_info->bb_first_free; | |
1790 | ||
1791 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
1792 | i = mb_find_next_zero_bit(bitmap, | |
1793 | EXT4_BLOCKS_PER_GROUP(sb), i); | |
1794 | if (i >= EXT4_BLOCKS_PER_GROUP(sb)) { | |
1795 | /* | |
1796 | * IF we have corrupt bitmap, we won't find any | |
1797 | * free blocks even though group info says we | |
1798 | * we have free blocks | |
1799 | */ | |
1800 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, | |
1801 | "%d free blocks as per " | |
1802 | "group info. But bitmap says 0", | |
1803 | free); | |
1804 | break; | |
1805 | } | |
1806 | ||
1807 | mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex); | |
1808 | BUG_ON(ex.fe_len <= 0); | |
1809 | if (free < ex.fe_len) { | |
1810 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, | |
1811 | "%d free blocks as per " | |
1812 | "group info. But got %d blocks", | |
1813 | free, ex.fe_len); | |
1814 | /* | |
1815 | * The number of free blocks differs. This mostly | |
1816 | * indicate that the bitmap is corrupt. So exit | |
1817 | * without claiming the space. | |
1818 | */ | |
1819 | break; | |
1820 | } | |
1821 | ||
1822 | ext4_mb_measure_extent(ac, &ex, e4b); | |
1823 | ||
1824 | i += ex.fe_len; | |
1825 | free -= ex.fe_len; | |
1826 | } | |
1827 | ||
1828 | ext4_mb_check_limits(ac, e4b, 1); | |
1829 | } | |
1830 | ||
1831 | /* | |
1832 | * This is a special case for storages like raid5 | |
1833 | * we try to find stripe-aligned chunks for stripe-size-multiple requests | |
1834 | */ | |
1835 | static noinline_for_stack | |
1836 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
1837 | struct ext4_buddy *e4b) | |
1838 | { | |
1839 | struct super_block *sb = ac->ac_sb; | |
1840 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1841 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1842 | struct ext4_free_extent ex; | |
1843 | ext4_fsblk_t first_group_block; | |
1844 | ext4_fsblk_t a; | |
1845 | ext4_grpblk_t i; | |
1846 | int max; | |
1847 | ||
1848 | BUG_ON(sbi->s_stripe == 0); | |
1849 | ||
1850 | /* find first stripe-aligned block in group */ | |
1851 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); | |
1852 | ||
1853 | a = first_group_block + sbi->s_stripe - 1; | |
1854 | do_div(a, sbi->s_stripe); | |
1855 | i = (a * sbi->s_stripe) - first_group_block; | |
1856 | ||
1857 | while (i < EXT4_BLOCKS_PER_GROUP(sb)) { | |
1858 | if (!mb_test_bit(i, bitmap)) { | |
1859 | max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex); | |
1860 | if (max >= sbi->s_stripe) { | |
1861 | ac->ac_found++; | |
1862 | ac->ac_b_ex = ex; | |
1863 | ext4_mb_use_best_found(ac, e4b); | |
1864 | break; | |
1865 | } | |
1866 | } | |
1867 | i += sbi->s_stripe; | |
1868 | } | |
1869 | } | |
1870 | ||
1871 | /* This is now called BEFORE we load the buddy bitmap. */ | |
1872 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, | |
1873 | ext4_group_t group, int cr) | |
1874 | { | |
1875 | unsigned free, fragments; | |
1876 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); | |
1877 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); | |
1878 | ||
1879 | BUG_ON(cr < 0 || cr >= 4); | |
1880 | ||
1881 | /* We only do this if the grp has never been initialized */ | |
1882 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
1883 | int ret = ext4_mb_init_group(ac->ac_sb, group); | |
1884 | if (ret) | |
1885 | return 0; | |
1886 | } | |
1887 | ||
1888 | free = grp->bb_free; | |
1889 | fragments = grp->bb_fragments; | |
1890 | if (free == 0) | |
1891 | return 0; | |
1892 | if (fragments == 0) | |
1893 | return 0; | |
1894 | ||
1895 | switch (cr) { | |
1896 | case 0: | |
1897 | BUG_ON(ac->ac_2order == 0); | |
1898 | ||
1899 | if (grp->bb_largest_free_order < ac->ac_2order) | |
1900 | return 0; | |
1901 | ||
1902 | /* Avoid using the first bg of a flexgroup for data files */ | |
1903 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
1904 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
1905 | ((group % flex_size) == 0)) | |
1906 | return 0; | |
1907 | ||
1908 | return 1; | |
1909 | case 1: | |
1910 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
1911 | return 1; | |
1912 | break; | |
1913 | case 2: | |
1914 | if (free >= ac->ac_g_ex.fe_len) | |
1915 | return 1; | |
1916 | break; | |
1917 | case 3: | |
1918 | return 1; | |
1919 | default: | |
1920 | BUG(); | |
1921 | } | |
1922 | ||
1923 | return 0; | |
1924 | } | |
1925 | ||
1926 | /* | |
1927 | * lock the group_info alloc_sem of all the groups | |
1928 | * belonging to the same buddy cache page. This | |
1929 | * make sure other parallel operation on the buddy | |
1930 | * cache doesn't happen whild holding the buddy cache | |
1931 | * lock | |
1932 | */ | |
1933 | int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group) | |
1934 | { | |
1935 | int i; | |
1936 | int block, pnum; | |
1937 | int blocks_per_page; | |
1938 | int groups_per_page; | |
1939 | ext4_group_t ngroups = ext4_get_groups_count(sb); | |
1940 | ext4_group_t first_group; | |
1941 | struct ext4_group_info *grp; | |
1942 | ||
1943 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1944 | /* | |
1945 | * the buddy cache inode stores the block bitmap | |
1946 | * and buddy information in consecutive blocks. | |
1947 | * So for each group we need two blocks. | |
1948 | */ | |
1949 | block = group * 2; | |
1950 | pnum = block / blocks_per_page; | |
1951 | first_group = pnum * blocks_per_page / 2; | |
1952 | ||
1953 | groups_per_page = blocks_per_page >> 1; | |
1954 | if (groups_per_page == 0) | |
1955 | groups_per_page = 1; | |
1956 | /* read all groups the page covers into the cache */ | |
1957 | for (i = 0; i < groups_per_page; i++) { | |
1958 | ||
1959 | if ((first_group + i) >= ngroups) | |
1960 | break; | |
1961 | grp = ext4_get_group_info(sb, first_group + i); | |
1962 | /* take all groups write allocation | |
1963 | * semaphore. This make sure there is | |
1964 | * no block allocation going on in any | |
1965 | * of that groups | |
1966 | */ | |
1967 | down_write_nested(&grp->alloc_sem, i); | |
1968 | } | |
1969 | return i; | |
1970 | } | |
1971 | ||
1972 | void ext4_mb_put_buddy_cache_lock(struct super_block *sb, | |
1973 | ext4_group_t group, int locked_group) | |
1974 | { | |
1975 | int i; | |
1976 | int block, pnum; | |
1977 | int blocks_per_page; | |
1978 | ext4_group_t first_group; | |
1979 | struct ext4_group_info *grp; | |
1980 | ||
1981 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1982 | /* | |
1983 | * the buddy cache inode stores the block bitmap | |
1984 | * and buddy information in consecutive blocks. | |
1985 | * So for each group we need two blocks. | |
1986 | */ | |
1987 | block = group * 2; | |
1988 | pnum = block / blocks_per_page; | |
1989 | first_group = pnum * blocks_per_page / 2; | |
1990 | /* release locks on all the groups */ | |
1991 | for (i = 0; i < locked_group; i++) { | |
1992 | ||
1993 | grp = ext4_get_group_info(sb, first_group + i); | |
1994 | /* take all groups write allocation | |
1995 | * semaphore. This make sure there is | |
1996 | * no block allocation going on in any | |
1997 | * of that groups | |
1998 | */ | |
1999 | up_write(&grp->alloc_sem); | |
2000 | } | |
2001 | ||
2002 | } | |
2003 | ||
2004 | static noinline_for_stack int | |
2005 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
2006 | { | |
2007 | ext4_group_t ngroups, group, i; | |
2008 | int cr; | |
2009 | int err = 0; | |
2010 | struct ext4_sb_info *sbi; | |
2011 | struct super_block *sb; | |
2012 | struct ext4_buddy e4b; | |
2013 | ||
2014 | sb = ac->ac_sb; | |
2015 | sbi = EXT4_SB(sb); | |
2016 | ngroups = ext4_get_groups_count(sb); | |
2017 | /* non-extent files are limited to low blocks/groups */ | |
2018 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) | |
2019 | ngroups = sbi->s_blockfile_groups; | |
2020 | ||
2021 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
2022 | ||
2023 | /* first, try the goal */ | |
2024 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2025 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2026 | goto out; | |
2027 | ||
2028 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2029 | goto out; | |
2030 | ||
2031 | /* | |
2032 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
2033 | * if ac2_order is set we also set criteria to 0 so that we | |
2034 | * try exact allocation using buddy. | |
2035 | */ | |
2036 | i = fls(ac->ac_g_ex.fe_len); | |
2037 | ac->ac_2order = 0; | |
2038 | /* | |
2039 | * We search using buddy data only if the order of the request | |
2040 | * is greater than equal to the sbi_s_mb_order2_reqs | |
2041 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req | |
2042 | */ | |
2043 | if (i >= sbi->s_mb_order2_reqs) { | |
2044 | /* | |
2045 | * This should tell if fe_len is exactly power of 2 | |
2046 | */ | |
2047 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
2048 | ac->ac_2order = i - 1; | |
2049 | } | |
2050 | ||
2051 | /* if stream allocation is enabled, use global goal */ | |
2052 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
2053 | /* TBD: may be hot point */ | |
2054 | spin_lock(&sbi->s_md_lock); | |
2055 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2056 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2057 | spin_unlock(&sbi->s_md_lock); | |
2058 | } | |
2059 | ||
2060 | /* Let's just scan groups to find more-less suitable blocks */ | |
2061 | cr = ac->ac_2order ? 0 : 1; | |
2062 | /* | |
2063 | * cr == 0 try to get exact allocation, | |
2064 | * cr == 3 try to get anything | |
2065 | */ | |
2066 | repeat: | |
2067 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2068 | ac->ac_criteria = cr; | |
2069 | /* | |
2070 | * searching for the right group start | |
2071 | * from the goal value specified | |
2072 | */ | |
2073 | group = ac->ac_g_ex.fe_group; | |
2074 | ||
2075 | for (i = 0; i < ngroups; group++, i++) { | |
2076 | if (group == ngroups) | |
2077 | group = 0; | |
2078 | ||
2079 | /* This now checks without needing the buddy page */ | |
2080 | if (!ext4_mb_good_group(ac, group, cr)) | |
2081 | continue; | |
2082 | ||
2083 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2084 | if (err) | |
2085 | goto out; | |
2086 | ||
2087 | ext4_lock_group(sb, group); | |
2088 | ||
2089 | /* | |
2090 | * We need to check again after locking the | |
2091 | * block group | |
2092 | */ | |
2093 | if (!ext4_mb_good_group(ac, group, cr)) { | |
2094 | ext4_unlock_group(sb, group); | |
2095 | ext4_mb_unload_buddy(&e4b); | |
2096 | continue; | |
2097 | } | |
2098 | ||
2099 | ac->ac_groups_scanned++; | |
2100 | if (cr == 0) | |
2101 | ext4_mb_simple_scan_group(ac, &e4b); | |
2102 | else if (cr == 1 && sbi->s_stripe && | |
2103 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
2104 | ext4_mb_scan_aligned(ac, &e4b); | |
2105 | else | |
2106 | ext4_mb_complex_scan_group(ac, &e4b); | |
2107 | ||
2108 | ext4_unlock_group(sb, group); | |
2109 | ext4_mb_unload_buddy(&e4b); | |
2110 | ||
2111 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2112 | break; | |
2113 | } | |
2114 | } | |
2115 | ||
2116 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2117 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2118 | /* | |
2119 | * We've been searching too long. Let's try to allocate | |
2120 | * the best chunk we've found so far | |
2121 | */ | |
2122 | ||
2123 | ext4_mb_try_best_found(ac, &e4b); | |
2124 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2125 | /* | |
2126 | * Someone more lucky has already allocated it. | |
2127 | * The only thing we can do is just take first | |
2128 | * found block(s) | |
2129 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2130 | */ | |
2131 | ac->ac_b_ex.fe_group = 0; | |
2132 | ac->ac_b_ex.fe_start = 0; | |
2133 | ac->ac_b_ex.fe_len = 0; | |
2134 | ac->ac_status = AC_STATUS_CONTINUE; | |
2135 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2136 | cr = 3; | |
2137 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2138 | goto repeat; | |
2139 | } | |
2140 | } | |
2141 | out: | |
2142 | return err; | |
2143 | } | |
2144 | ||
2145 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) | |
2146 | { | |
2147 | struct super_block *sb = seq->private; | |
2148 | ext4_group_t group; | |
2149 | ||
2150 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) | |
2151 | return NULL; | |
2152 | group = *pos + 1; | |
2153 | return (void *) ((unsigned long) group); | |
2154 | } | |
2155 | ||
2156 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2157 | { | |
2158 | struct super_block *sb = seq->private; | |
2159 | ext4_group_t group; | |
2160 | ||
2161 | ++*pos; | |
2162 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) | |
2163 | return NULL; | |
2164 | group = *pos + 1; | |
2165 | return (void *) ((unsigned long) group); | |
2166 | } | |
2167 | ||
2168 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2169 | { | |
2170 | struct super_block *sb = seq->private; | |
2171 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); | |
2172 | int i; | |
2173 | int err; | |
2174 | struct ext4_buddy e4b; | |
2175 | struct sg { | |
2176 | struct ext4_group_info info; | |
2177 | ext4_grpblk_t counters[16]; | |
2178 | } sg; | |
2179 | ||
2180 | group--; | |
2181 | if (group == 0) | |
2182 | seq_printf(seq, "#%-5s: %-5s %-5s %-5s " | |
2183 | "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " | |
2184 | "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", | |
2185 | "group", "free", "frags", "first", | |
2186 | "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", | |
2187 | "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); | |
2188 | ||
2189 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2190 | sizeof(struct ext4_group_info); | |
2191 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2192 | if (err) { | |
2193 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2194 | return 0; | |
2195 | } | |
2196 | ext4_lock_group(sb, group); | |
2197 | memcpy(&sg, ext4_get_group_info(sb, group), i); | |
2198 | ext4_unlock_group(sb, group); | |
2199 | ext4_mb_unload_buddy(&e4b); | |
2200 | ||
2201 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, | |
2202 | sg.info.bb_fragments, sg.info.bb_first_free); | |
2203 | for (i = 0; i <= 13; i++) | |
2204 | seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? | |
2205 | sg.info.bb_counters[i] : 0); | |
2206 | seq_printf(seq, " ]\n"); | |
2207 | ||
2208 | return 0; | |
2209 | } | |
2210 | ||
2211 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2212 | { | |
2213 | } | |
2214 | ||
2215 | static const struct seq_operations ext4_mb_seq_groups_ops = { | |
2216 | .start = ext4_mb_seq_groups_start, | |
2217 | .next = ext4_mb_seq_groups_next, | |
2218 | .stop = ext4_mb_seq_groups_stop, | |
2219 | .show = ext4_mb_seq_groups_show, | |
2220 | }; | |
2221 | ||
2222 | static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) | |
2223 | { | |
2224 | struct super_block *sb = PDE(inode)->data; | |
2225 | int rc; | |
2226 | ||
2227 | rc = seq_open(file, &ext4_mb_seq_groups_ops); | |
2228 | if (rc == 0) { | |
2229 | struct seq_file *m = file->private_data; | |
2230 | m->private = sb; | |
2231 | } | |
2232 | return rc; | |
2233 | ||
2234 | } | |
2235 | ||
2236 | static const struct file_operations ext4_mb_seq_groups_fops = { | |
2237 | .owner = THIS_MODULE, | |
2238 | .open = ext4_mb_seq_groups_open, | |
2239 | .read = seq_read, | |
2240 | .llseek = seq_lseek, | |
2241 | .release = seq_release, | |
2242 | }; | |
2243 | ||
2244 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) | |
2245 | { | |
2246 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2247 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
2248 | ||
2249 | BUG_ON(!cachep); | |
2250 | return cachep; | |
2251 | } | |
2252 | ||
2253 | /* Create and initialize ext4_group_info data for the given group. */ | |
2254 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, | |
2255 | struct ext4_group_desc *desc) | |
2256 | { | |
2257 | int i; | |
2258 | int metalen = 0; | |
2259 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2260 | struct ext4_group_info **meta_group_info; | |
2261 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); | |
2262 | ||
2263 | /* | |
2264 | * First check if this group is the first of a reserved block. | |
2265 | * If it's true, we have to allocate a new table of pointers | |
2266 | * to ext4_group_info structures | |
2267 | */ | |
2268 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2269 | metalen = sizeof(*meta_group_info) << | |
2270 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2271 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2272 | if (meta_group_info == NULL) { | |
2273 | printk(KERN_ERR "EXT4-fs: can't allocate mem for a " | |
2274 | "buddy group\n"); | |
2275 | goto exit_meta_group_info; | |
2276 | } | |
2277 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = | |
2278 | meta_group_info; | |
2279 | } | |
2280 | ||
2281 | meta_group_info = | |
2282 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; | |
2283 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); | |
2284 | ||
2285 | meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL); | |
2286 | if (meta_group_info[i] == NULL) { | |
2287 | printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n"); | |
2288 | goto exit_group_info; | |
2289 | } | |
2290 | memset(meta_group_info[i], 0, kmem_cache_size(cachep)); | |
2291 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2292 | &(meta_group_info[i]->bb_state)); | |
2293 | ||
2294 | /* | |
2295 | * initialize bb_free to be able to skip | |
2296 | * empty groups without initialization | |
2297 | */ | |
2298 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2299 | meta_group_info[i]->bb_free = | |
2300 | ext4_free_blocks_after_init(sb, group, desc); | |
2301 | } else { | |
2302 | meta_group_info[i]->bb_free = | |
2303 | ext4_free_blks_count(sb, desc); | |
2304 | } | |
2305 | ||
2306 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
2307 | init_rwsem(&meta_group_info[i]->alloc_sem); | |
2308 | meta_group_info[i]->bb_free_root = RB_ROOT; | |
2309 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ | |
2310 | ||
2311 | #ifdef DOUBLE_CHECK | |
2312 | { | |
2313 | struct buffer_head *bh; | |
2314 | meta_group_info[i]->bb_bitmap = | |
2315 | kmalloc(sb->s_blocksize, GFP_KERNEL); | |
2316 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); | |
2317 | bh = ext4_read_block_bitmap(sb, group); | |
2318 | BUG_ON(bh == NULL); | |
2319 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, | |
2320 | sb->s_blocksize); | |
2321 | put_bh(bh); | |
2322 | } | |
2323 | #endif | |
2324 | ||
2325 | return 0; | |
2326 | ||
2327 | exit_group_info: | |
2328 | /* If a meta_group_info table has been allocated, release it now */ | |
2329 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) | |
2330 | kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); | |
2331 | exit_meta_group_info: | |
2332 | return -ENOMEM; | |
2333 | } /* ext4_mb_add_groupinfo */ | |
2334 | ||
2335 | static int ext4_mb_init_backend(struct super_block *sb) | |
2336 | { | |
2337 | ext4_group_t ngroups = ext4_get_groups_count(sb); | |
2338 | ext4_group_t i; | |
2339 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2340 | struct ext4_super_block *es = sbi->s_es; | |
2341 | int num_meta_group_infos; | |
2342 | int num_meta_group_infos_max; | |
2343 | int array_size; | |
2344 | struct ext4_group_desc *desc; | |
2345 | struct kmem_cache *cachep; | |
2346 | ||
2347 | /* This is the number of blocks used by GDT */ | |
2348 | num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) - | |
2349 | 1) >> EXT4_DESC_PER_BLOCK_BITS(sb); | |
2350 | ||
2351 | /* | |
2352 | * This is the total number of blocks used by GDT including | |
2353 | * the number of reserved blocks for GDT. | |
2354 | * The s_group_info array is allocated with this value | |
2355 | * to allow a clean online resize without a complex | |
2356 | * manipulation of pointer. | |
2357 | * The drawback is the unused memory when no resize | |
2358 | * occurs but it's very low in terms of pages | |
2359 | * (see comments below) | |
2360 | * Need to handle this properly when META_BG resizing is allowed | |
2361 | */ | |
2362 | num_meta_group_infos_max = num_meta_group_infos + | |
2363 | le16_to_cpu(es->s_reserved_gdt_blocks); | |
2364 | ||
2365 | /* | |
2366 | * array_size is the size of s_group_info array. We round it | |
2367 | * to the next power of two because this approximation is done | |
2368 | * internally by kmalloc so we can have some more memory | |
2369 | * for free here (e.g. may be used for META_BG resize). | |
2370 | */ | |
2371 | array_size = 1; | |
2372 | while (array_size < sizeof(*sbi->s_group_info) * | |
2373 | num_meta_group_infos_max) | |
2374 | array_size = array_size << 1; | |
2375 | /* An 8TB filesystem with 64-bit pointers requires a 4096 byte | |
2376 | * kmalloc. A 128kb malloc should suffice for a 256TB filesystem. | |
2377 | * So a two level scheme suffices for now. */ | |
2378 | sbi->s_group_info = kmalloc(array_size, GFP_KERNEL); | |
2379 | if (sbi->s_group_info == NULL) { | |
2380 | printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n"); | |
2381 | return -ENOMEM; | |
2382 | } | |
2383 | sbi->s_buddy_cache = new_inode(sb); | |
2384 | if (sbi->s_buddy_cache == NULL) { | |
2385 | printk(KERN_ERR "EXT4-fs: can't get new inode\n"); | |
2386 | goto err_freesgi; | |
2387 | } | |
2388 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; | |
2389 | for (i = 0; i < ngroups; i++) { | |
2390 | desc = ext4_get_group_desc(sb, i, NULL); | |
2391 | if (desc == NULL) { | |
2392 | printk(KERN_ERR | |
2393 | "EXT4-fs: can't read descriptor %u\n", i); | |
2394 | goto err_freebuddy; | |
2395 | } | |
2396 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) | |
2397 | goto err_freebuddy; | |
2398 | } | |
2399 | ||
2400 | return 0; | |
2401 | ||
2402 | err_freebuddy: | |
2403 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); | |
2404 | while (i-- > 0) | |
2405 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); | |
2406 | i = num_meta_group_infos; | |
2407 | while (i-- > 0) | |
2408 | kfree(sbi->s_group_info[i]); | |
2409 | iput(sbi->s_buddy_cache); | |
2410 | err_freesgi: | |
2411 | kfree(sbi->s_group_info); | |
2412 | return -ENOMEM; | |
2413 | } | |
2414 | ||
2415 | int ext4_mb_init(struct super_block *sb, int needs_recovery) | |
2416 | { | |
2417 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2418 | unsigned i, j; | |
2419 | unsigned offset; | |
2420 | unsigned max; | |
2421 | int ret; | |
2422 | int cache_index; | |
2423 | struct kmem_cache *cachep; | |
2424 | char *namep = NULL; | |
2425 | ||
2426 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); | |
2427 | ||
2428 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2429 | if (sbi->s_mb_offsets == NULL) { | |
2430 | ret = -ENOMEM; | |
2431 | goto out; | |
2432 | } | |
2433 | ||
2434 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); | |
2435 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); | |
2436 | if (sbi->s_mb_maxs == NULL) { | |
2437 | ret = -ENOMEM; | |
2438 | goto out; | |
2439 | } | |
2440 | ||
2441 | cache_index = sb->s_blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2442 | cachep = ext4_groupinfo_caches[cache_index]; | |
2443 | if (!cachep) { | |
2444 | char name[32]; | |
2445 | int len = offsetof(struct ext4_group_info, | |
2446 | bb_counters[sb->s_blocksize_bits + 2]); | |
2447 | ||
2448 | sprintf(name, "ext4_groupinfo_%d", sb->s_blocksize_bits); | |
2449 | namep = kstrdup(name, GFP_KERNEL); | |
2450 | if (!namep) { | |
2451 | ret = -ENOMEM; | |
2452 | goto out; | |
2453 | } | |
2454 | ||
2455 | /* Need to free the kmem_cache_name() when we | |
2456 | * destroy the slab */ | |
2457 | cachep = kmem_cache_create(namep, len, 0, | |
2458 | SLAB_RECLAIM_ACCOUNT, NULL); | |
2459 | if (!cachep) { | |
2460 | ret = -ENOMEM; | |
2461 | goto out; | |
2462 | } | |
2463 | ext4_groupinfo_caches[cache_index] = cachep; | |
2464 | } | |
2465 | ||
2466 | /* order 0 is regular bitmap */ | |
2467 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2468 | sbi->s_mb_offsets[0] = 0; | |
2469 | ||
2470 | i = 1; | |
2471 | offset = 0; | |
2472 | max = sb->s_blocksize << 2; | |
2473 | do { | |
2474 | sbi->s_mb_offsets[i] = offset; | |
2475 | sbi->s_mb_maxs[i] = max; | |
2476 | offset += 1 << (sb->s_blocksize_bits - i); | |
2477 | max = max >> 1; | |
2478 | i++; | |
2479 | } while (i <= sb->s_blocksize_bits + 1); | |
2480 | ||
2481 | /* init file for buddy data */ | |
2482 | ret = ext4_mb_init_backend(sb); | |
2483 | if (ret != 0) { | |
2484 | goto out; | |
2485 | } | |
2486 | ||
2487 | spin_lock_init(&sbi->s_md_lock); | |
2488 | spin_lock_init(&sbi->s_bal_lock); | |
2489 | ||
2490 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2491 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2492 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2493 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2494 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
2495 | sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC; | |
2496 | ||
2497 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); | |
2498 | if (sbi->s_locality_groups == NULL) { | |
2499 | ret = -ENOMEM; | |
2500 | goto out; | |
2501 | } | |
2502 | for_each_possible_cpu(i) { | |
2503 | struct ext4_locality_group *lg; | |
2504 | lg = per_cpu_ptr(sbi->s_locality_groups, i); | |
2505 | mutex_init(&lg->lg_mutex); | |
2506 | for (j = 0; j < PREALLOC_TB_SIZE; j++) | |
2507 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
2508 | spin_lock_init(&lg->lg_prealloc_lock); | |
2509 | } | |
2510 | ||
2511 | if (sbi->s_proc) | |
2512 | proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, | |
2513 | &ext4_mb_seq_groups_fops, sb); | |
2514 | ||
2515 | if (sbi->s_journal) | |
2516 | sbi->s_journal->j_commit_callback = release_blocks_on_commit; | |
2517 | out: | |
2518 | if (ret) { | |
2519 | kfree(sbi->s_mb_offsets); | |
2520 | kfree(sbi->s_mb_maxs); | |
2521 | kfree(namep); | |
2522 | } | |
2523 | return ret; | |
2524 | } | |
2525 | ||
2526 | /* need to called with the ext4 group lock held */ | |
2527 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) | |
2528 | { | |
2529 | struct ext4_prealloc_space *pa; | |
2530 | struct list_head *cur, *tmp; | |
2531 | int count = 0; | |
2532 | ||
2533 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2534 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2535 | list_del(&pa->pa_group_list); | |
2536 | count++; | |
2537 | kmem_cache_free(ext4_pspace_cachep, pa); | |
2538 | } | |
2539 | if (count) | |
2540 | mb_debug(1, "mballoc: %u PAs left\n", count); | |
2541 | ||
2542 | } | |
2543 | ||
2544 | int ext4_mb_release(struct super_block *sb) | |
2545 | { | |
2546 | ext4_group_t ngroups = ext4_get_groups_count(sb); | |
2547 | ext4_group_t i; | |
2548 | int num_meta_group_infos; | |
2549 | struct ext4_group_info *grinfo; | |
2550 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2551 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); | |
2552 | ||
2553 | if (sbi->s_group_info) { | |
2554 | for (i = 0; i < ngroups; i++) { | |
2555 | grinfo = ext4_get_group_info(sb, i); | |
2556 | #ifdef DOUBLE_CHECK | |
2557 | kfree(grinfo->bb_bitmap); | |
2558 | #endif | |
2559 | ext4_lock_group(sb, i); | |
2560 | ext4_mb_cleanup_pa(grinfo); | |
2561 | ext4_unlock_group(sb, i); | |
2562 | kmem_cache_free(cachep, grinfo); | |
2563 | } | |
2564 | num_meta_group_infos = (ngroups + | |
2565 | EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
2566 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2567 | for (i = 0; i < num_meta_group_infos; i++) | |
2568 | kfree(sbi->s_group_info[i]); | |
2569 | kfree(sbi->s_group_info); | |
2570 | } | |
2571 | kfree(sbi->s_mb_offsets); | |
2572 | kfree(sbi->s_mb_maxs); | |
2573 | if (sbi->s_buddy_cache) | |
2574 | iput(sbi->s_buddy_cache); | |
2575 | if (sbi->s_mb_stats) { | |
2576 | printk(KERN_INFO | |
2577 | "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n", | |
2578 | atomic_read(&sbi->s_bal_allocated), | |
2579 | atomic_read(&sbi->s_bal_reqs), | |
2580 | atomic_read(&sbi->s_bal_success)); | |
2581 | printk(KERN_INFO | |
2582 | "EXT4-fs: mballoc: %u extents scanned, %u goal hits, " | |
2583 | "%u 2^N hits, %u breaks, %u lost\n", | |
2584 | atomic_read(&sbi->s_bal_ex_scanned), | |
2585 | atomic_read(&sbi->s_bal_goals), | |
2586 | atomic_read(&sbi->s_bal_2orders), | |
2587 | atomic_read(&sbi->s_bal_breaks), | |
2588 | atomic_read(&sbi->s_mb_lost_chunks)); | |
2589 | printk(KERN_INFO | |
2590 | "EXT4-fs: mballoc: %lu generated and it took %Lu\n", | |
2591 | sbi->s_mb_buddies_generated++, | |
2592 | sbi->s_mb_generation_time); | |
2593 | printk(KERN_INFO | |
2594 | "EXT4-fs: mballoc: %u preallocated, %u discarded\n", | |
2595 | atomic_read(&sbi->s_mb_preallocated), | |
2596 | atomic_read(&sbi->s_mb_discarded)); | |
2597 | } | |
2598 | ||
2599 | free_percpu(sbi->s_locality_groups); | |
2600 | if (sbi->s_proc) | |
2601 | remove_proc_entry("mb_groups", sbi->s_proc); | |
2602 | ||
2603 | return 0; | |
2604 | } | |
2605 | ||
2606 | static inline void ext4_issue_discard(struct super_block *sb, | |
2607 | ext4_group_t block_group, ext4_grpblk_t block, int count) | |
2608 | { | |
2609 | int ret; | |
2610 | ext4_fsblk_t discard_block; | |
2611 | ||
2612 | discard_block = block + ext4_group_first_block_no(sb, block_group); | |
2613 | trace_ext4_discard_blocks(sb, | |
2614 | (unsigned long long) discard_block, count); | |
2615 | ret = sb_issue_discard(sb, discard_block, count); | |
2616 | if (ret == -EOPNOTSUPP) { | |
2617 | ext4_warning(sb, "discard not supported, disabling"); | |
2618 | clear_opt(EXT4_SB(sb)->s_mount_opt, DISCARD); | |
2619 | } | |
2620 | } | |
2621 | ||
2622 | /* | |
2623 | * This function is called by the jbd2 layer once the commit has finished, | |
2624 | * so we know we can free the blocks that were released with that commit. | |
2625 | */ | |
2626 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn) | |
2627 | { | |
2628 | struct super_block *sb = journal->j_private; | |
2629 | struct ext4_buddy e4b; | |
2630 | struct ext4_group_info *db; | |
2631 | int err, count = 0, count2 = 0; | |
2632 | struct ext4_free_data *entry; | |
2633 | struct list_head *l, *ltmp; | |
2634 | ||
2635 | list_for_each_safe(l, ltmp, &txn->t_private_list) { | |
2636 | entry = list_entry(l, struct ext4_free_data, list); | |
2637 | ||
2638 | mb_debug(1, "gonna free %u blocks in group %u (0x%p):", | |
2639 | entry->count, entry->group, entry); | |
2640 | ||
2641 | if (test_opt(sb, DISCARD)) | |
2642 | ext4_issue_discard(sb, entry->group, | |
2643 | entry->start_blk, entry->count); | |
2644 | ||
2645 | err = ext4_mb_load_buddy(sb, entry->group, &e4b); | |
2646 | /* we expect to find existing buddy because it's pinned */ | |
2647 | BUG_ON(err != 0); | |
2648 | ||
2649 | db = e4b.bd_info; | |
2650 | /* there are blocks to put in buddy to make them really free */ | |
2651 | count += entry->count; | |
2652 | count2++; | |
2653 | ext4_lock_group(sb, entry->group); | |
2654 | /* Take it out of per group rb tree */ | |
2655 | rb_erase(&entry->node, &(db->bb_free_root)); | |
2656 | mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count); | |
2657 | ||
2658 | if (!db->bb_free_root.rb_node) { | |
2659 | /* No more items in the per group rb tree | |
2660 | * balance refcounts from ext4_mb_free_metadata() | |
2661 | */ | |
2662 | page_cache_release(e4b.bd_buddy_page); | |
2663 | page_cache_release(e4b.bd_bitmap_page); | |
2664 | } | |
2665 | ext4_unlock_group(sb, entry->group); | |
2666 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
2667 | ext4_mb_unload_buddy(&e4b); | |
2668 | } | |
2669 | ||
2670 | mb_debug(1, "freed %u blocks in %u structures\n", count, count2); | |
2671 | } | |
2672 | ||
2673 | #ifdef CONFIG_EXT4_DEBUG | |
2674 | u8 mb_enable_debug __read_mostly; | |
2675 | ||
2676 | static struct dentry *debugfs_dir; | |
2677 | static struct dentry *debugfs_debug; | |
2678 | ||
2679 | static void __init ext4_create_debugfs_entry(void) | |
2680 | { | |
2681 | debugfs_dir = debugfs_create_dir("ext4", NULL); | |
2682 | if (debugfs_dir) | |
2683 | debugfs_debug = debugfs_create_u8("mballoc-debug", | |
2684 | S_IRUGO | S_IWUSR, | |
2685 | debugfs_dir, | |
2686 | &mb_enable_debug); | |
2687 | } | |
2688 | ||
2689 | static void ext4_remove_debugfs_entry(void) | |
2690 | { | |
2691 | debugfs_remove(debugfs_debug); | |
2692 | debugfs_remove(debugfs_dir); | |
2693 | } | |
2694 | ||
2695 | #else | |
2696 | ||
2697 | static void __init ext4_create_debugfs_entry(void) | |
2698 | { | |
2699 | } | |
2700 | ||
2701 | static void ext4_remove_debugfs_entry(void) | |
2702 | { | |
2703 | } | |
2704 | ||
2705 | #endif | |
2706 | ||
2707 | int __init init_ext4_mballoc(void) | |
2708 | { | |
2709 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, | |
2710 | SLAB_RECLAIM_ACCOUNT); | |
2711 | if (ext4_pspace_cachep == NULL) | |
2712 | return -ENOMEM; | |
2713 | ||
2714 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, | |
2715 | SLAB_RECLAIM_ACCOUNT); | |
2716 | if (ext4_ac_cachep == NULL) { | |
2717 | kmem_cache_destroy(ext4_pspace_cachep); | |
2718 | return -ENOMEM; | |
2719 | } | |
2720 | ||
2721 | ext4_free_ext_cachep = KMEM_CACHE(ext4_free_data, | |
2722 | SLAB_RECLAIM_ACCOUNT); | |
2723 | if (ext4_free_ext_cachep == NULL) { | |
2724 | kmem_cache_destroy(ext4_pspace_cachep); | |
2725 | kmem_cache_destroy(ext4_ac_cachep); | |
2726 | return -ENOMEM; | |
2727 | } | |
2728 | ext4_create_debugfs_entry(); | |
2729 | return 0; | |
2730 | } | |
2731 | ||
2732 | void exit_ext4_mballoc(void) | |
2733 | { | |
2734 | int i; | |
2735 | /* | |
2736 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep | |
2737 | * before destroying the slab cache. | |
2738 | */ | |
2739 | rcu_barrier(); | |
2740 | kmem_cache_destroy(ext4_pspace_cachep); | |
2741 | kmem_cache_destroy(ext4_ac_cachep); | |
2742 | kmem_cache_destroy(ext4_free_ext_cachep); | |
2743 | ||
2744 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
2745 | struct kmem_cache *cachep = ext4_groupinfo_caches[i]; | |
2746 | if (cachep) { | |
2747 | char *name = (char *)kmem_cache_name(cachep); | |
2748 | kmem_cache_destroy(cachep); | |
2749 | kfree(name); | |
2750 | } | |
2751 | } | |
2752 | ext4_remove_debugfs_entry(); | |
2753 | } | |
2754 | ||
2755 | ||
2756 | /* | |
2757 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps | |
2758 | * Returns 0 if success or error code | |
2759 | */ | |
2760 | static noinline_for_stack int | |
2761 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
2762 | handle_t *handle, unsigned int reserv_blks) | |
2763 | { | |
2764 | struct buffer_head *bitmap_bh = NULL; | |
2765 | struct ext4_group_desc *gdp; | |
2766 | struct buffer_head *gdp_bh; | |
2767 | struct ext4_sb_info *sbi; | |
2768 | struct super_block *sb; | |
2769 | ext4_fsblk_t block; | |
2770 | int err, len; | |
2771 | ||
2772 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2773 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2774 | ||
2775 | sb = ac->ac_sb; | |
2776 | sbi = EXT4_SB(sb); | |
2777 | ||
2778 | err = -EIO; | |
2779 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); | |
2780 | if (!bitmap_bh) | |
2781 | goto out_err; | |
2782 | ||
2783 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2784 | if (err) | |
2785 | goto out_err; | |
2786 | ||
2787 | err = -EIO; | |
2788 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
2789 | if (!gdp) | |
2790 | goto out_err; | |
2791 | ||
2792 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, | |
2793 | ext4_free_blks_count(sb, gdp)); | |
2794 | ||
2795 | err = ext4_journal_get_write_access(handle, gdp_bh); | |
2796 | if (err) | |
2797 | goto out_err; | |
2798 | ||
2799 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
2800 | ||
2801 | len = ac->ac_b_ex.fe_len; | |
2802 | if (!ext4_data_block_valid(sbi, block, len)) { | |
2803 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " | |
2804 | "fs metadata\n", block, block+len); | |
2805 | /* File system mounted not to panic on error | |
2806 | * Fix the bitmap and repeat the block allocation | |
2807 | * We leak some of the blocks here. | |
2808 | */ | |
2809 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
2810 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, | |
2811 | ac->ac_b_ex.fe_len); | |
2812 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
2813 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
2814 | if (!err) | |
2815 | err = -EAGAIN; | |
2816 | goto out_err; | |
2817 | } | |
2818 | ||
2819 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
2820 | #ifdef AGGRESSIVE_CHECK | |
2821 | { | |
2822 | int i; | |
2823 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
2824 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
2825 | bitmap_bh->b_data)); | |
2826 | } | |
2827 | } | |
2828 | #endif | |
2829 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len); | |
2830 | if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2831 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
2832 | ext4_free_blks_set(sb, gdp, | |
2833 | ext4_free_blocks_after_init(sb, | |
2834 | ac->ac_b_ex.fe_group, gdp)); | |
2835 | } | |
2836 | len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len; | |
2837 | ext4_free_blks_set(sb, gdp, len); | |
2838 | gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp); | |
2839 | ||
2840 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
2841 | percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len); | |
2842 | /* | |
2843 | * Now reduce the dirty block count also. Should not go negative | |
2844 | */ | |
2845 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) | |
2846 | /* release all the reserved blocks if non delalloc */ | |
2847 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks); | |
2848 | ||
2849 | if (sbi->s_log_groups_per_flex) { | |
2850 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
2851 | ac->ac_b_ex.fe_group); | |
2852 | atomic_sub(ac->ac_b_ex.fe_len, | |
2853 | &sbi->s_flex_groups[flex_group].free_blocks); | |
2854 | } | |
2855 | ||
2856 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
2857 | if (err) | |
2858 | goto out_err; | |
2859 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); | |
2860 | ||
2861 | out_err: | |
2862 | ext4_mark_super_dirty(sb); | |
2863 | brelse(bitmap_bh); | |
2864 | return err; | |
2865 | } | |
2866 | ||
2867 | /* | |
2868 | * here we normalize request for locality group | |
2869 | * Group request are normalized to s_strip size if we set the same via mount | |
2870 | * option. If not we set it to s_mb_group_prealloc which can be configured via | |
2871 | * /sys/fs/ext4/<partition>/mb_group_prealloc | |
2872 | * | |
2873 | * XXX: should we try to preallocate more than the group has now? | |
2874 | */ | |
2875 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
2876 | { | |
2877 | struct super_block *sb = ac->ac_sb; | |
2878 | struct ext4_locality_group *lg = ac->ac_lg; | |
2879 | ||
2880 | BUG_ON(lg == NULL); | |
2881 | if (EXT4_SB(sb)->s_stripe) | |
2882 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe; | |
2883 | else | |
2884 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; | |
2885 | mb_debug(1, "#%u: goal %u blocks for locality group\n", | |
2886 | current->pid, ac->ac_g_ex.fe_len); | |
2887 | } | |
2888 | ||
2889 | /* | |
2890 | * Normalization means making request better in terms of | |
2891 | * size and alignment | |
2892 | */ | |
2893 | static noinline_for_stack void | |
2894 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
2895 | struct ext4_allocation_request *ar) | |
2896 | { | |
2897 | int bsbits, max; | |
2898 | ext4_lblk_t end; | |
2899 | loff_t size, orig_size, start_off; | |
2900 | ext4_lblk_t start; | |
2901 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); | |
2902 | struct ext4_prealloc_space *pa; | |
2903 | ||
2904 | /* do normalize only data requests, metadata requests | |
2905 | do not need preallocation */ | |
2906 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
2907 | return; | |
2908 | ||
2909 | /* sometime caller may want exact blocks */ | |
2910 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2911 | return; | |
2912 | ||
2913 | /* caller may indicate that preallocation isn't | |
2914 | * required (it's a tail, for example) */ | |
2915 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
2916 | return; | |
2917 | ||
2918 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
2919 | ext4_mb_normalize_group_request(ac); | |
2920 | return ; | |
2921 | } | |
2922 | ||
2923 | bsbits = ac->ac_sb->s_blocksize_bits; | |
2924 | ||
2925 | /* first, let's learn actual file size | |
2926 | * given current request is allocated */ | |
2927 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; | |
2928 | size = size << bsbits; | |
2929 | if (size < i_size_read(ac->ac_inode)) | |
2930 | size = i_size_read(ac->ac_inode); | |
2931 | orig_size = size; | |
2932 | ||
2933 | /* max size of free chunks */ | |
2934 | max = 2 << bsbits; | |
2935 | ||
2936 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ | |
2937 | (req <= (size) || max <= (chunk_size)) | |
2938 | ||
2939 | /* first, try to predict filesize */ | |
2940 | /* XXX: should this table be tunable? */ | |
2941 | start_off = 0; | |
2942 | if (size <= 16 * 1024) { | |
2943 | size = 16 * 1024; | |
2944 | } else if (size <= 32 * 1024) { | |
2945 | size = 32 * 1024; | |
2946 | } else if (size <= 64 * 1024) { | |
2947 | size = 64 * 1024; | |
2948 | } else if (size <= 128 * 1024) { | |
2949 | size = 128 * 1024; | |
2950 | } else if (size <= 256 * 1024) { | |
2951 | size = 256 * 1024; | |
2952 | } else if (size <= 512 * 1024) { | |
2953 | size = 512 * 1024; | |
2954 | } else if (size <= 1024 * 1024) { | |
2955 | size = 1024 * 1024; | |
2956 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { | |
2957 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> | |
2958 | (21 - bsbits)) << 21; | |
2959 | size = 2 * 1024 * 1024; | |
2960 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
2961 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> | |
2962 | (22 - bsbits)) << 22; | |
2963 | size = 4 * 1024 * 1024; | |
2964 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
2965 | (8<<20)>>bsbits, max, 8 * 1024)) { | |
2966 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> | |
2967 | (23 - bsbits)) << 23; | |
2968 | size = 8 * 1024 * 1024; | |
2969 | } else { | |
2970 | start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; | |
2971 | size = ac->ac_o_ex.fe_len << bsbits; | |
2972 | } | |
2973 | size = size >> bsbits; | |
2974 | start = start_off >> bsbits; | |
2975 | ||
2976 | /* don't cover already allocated blocks in selected range */ | |
2977 | if (ar->pleft && start <= ar->lleft) { | |
2978 | size -= ar->lleft + 1 - start; | |
2979 | start = ar->lleft + 1; | |
2980 | } | |
2981 | if (ar->pright && start + size - 1 >= ar->lright) | |
2982 | size -= start + size - ar->lright; | |
2983 | ||
2984 | end = start + size; | |
2985 | ||
2986 | /* check we don't cross already preallocated blocks */ | |
2987 | rcu_read_lock(); | |
2988 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { | |
2989 | ext4_lblk_t pa_end; | |
2990 | ||
2991 | if (pa->pa_deleted) | |
2992 | continue; | |
2993 | spin_lock(&pa->pa_lock); | |
2994 | if (pa->pa_deleted) { | |
2995 | spin_unlock(&pa->pa_lock); | |
2996 | continue; | |
2997 | } | |
2998 | ||
2999 | pa_end = pa->pa_lstart + pa->pa_len; | |
3000 | ||
3001 | /* PA must not overlap original request */ | |
3002 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
3003 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
3004 | ||
3005 | /* skip PAs this normalized request doesn't overlap with */ | |
3006 | if (pa->pa_lstart >= end || pa_end <= start) { | |
3007 | spin_unlock(&pa->pa_lock); | |
3008 | continue; | |
3009 | } | |
3010 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3011 | ||
3012 | /* adjust start or end to be adjacent to this pa */ | |
3013 | if (pa_end <= ac->ac_o_ex.fe_logical) { | |
3014 | BUG_ON(pa_end < start); | |
3015 | start = pa_end; | |
3016 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { | |
3017 | BUG_ON(pa->pa_lstart > end); | |
3018 | end = pa->pa_lstart; | |
3019 | } | |
3020 | spin_unlock(&pa->pa_lock); | |
3021 | } | |
3022 | rcu_read_unlock(); | |
3023 | size = end - start; | |
3024 | ||
3025 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3026 | rcu_read_lock(); | |
3027 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { | |
3028 | ext4_lblk_t pa_end; | |
3029 | spin_lock(&pa->pa_lock); | |
3030 | if (pa->pa_deleted == 0) { | |
3031 | pa_end = pa->pa_lstart + pa->pa_len; | |
3032 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); | |
3033 | } | |
3034 | spin_unlock(&pa->pa_lock); | |
3035 | } | |
3036 | rcu_read_unlock(); | |
3037 | ||
3038 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3039 | start > ac->ac_o_ex.fe_logical) { | |
3040 | printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n", | |
3041 | (unsigned long) start, (unsigned long) size, | |
3042 | (unsigned long) ac->ac_o_ex.fe_logical); | |
3043 | } | |
3044 | BUG_ON(start + size <= ac->ac_o_ex.fe_logical && | |
3045 | start > ac->ac_o_ex.fe_logical); | |
3046 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); | |
3047 | ||
3048 | /* now prepare goal request */ | |
3049 | ||
3050 | /* XXX: is it better to align blocks WRT to logical | |
3051 | * placement or satisfy big request as is */ | |
3052 | ac->ac_g_ex.fe_logical = start; | |
3053 | ac->ac_g_ex.fe_len = size; | |
3054 | ||
3055 | /* define goal start in order to merge */ | |
3056 | if (ar->pright && (ar->lright == (start + size))) { | |
3057 | /* merge to the right */ | |
3058 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3059 | &ac->ac_f_ex.fe_group, | |
3060 | &ac->ac_f_ex.fe_start); | |
3061 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3062 | } | |
3063 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3064 | /* merge to the left */ | |
3065 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3066 | &ac->ac_f_ex.fe_group, | |
3067 | &ac->ac_f_ex.fe_start); | |
3068 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3069 | } | |
3070 | ||
3071 | mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, | |
3072 | (unsigned) orig_size, (unsigned) start); | |
3073 | } | |
3074 | ||
3075 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3076 | { | |
3077 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3078 | ||
3079 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3080 | atomic_inc(&sbi->s_bal_reqs); | |
3081 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
3082 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) | |
3083 | atomic_inc(&sbi->s_bal_success); | |
3084 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3085 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3086 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3087 | atomic_inc(&sbi->s_bal_goals); | |
3088 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3089 | atomic_inc(&sbi->s_bal_breaks); | |
3090 | } | |
3091 | ||
3092 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) | |
3093 | trace_ext4_mballoc_alloc(ac); | |
3094 | else | |
3095 | trace_ext4_mballoc_prealloc(ac); | |
3096 | } | |
3097 | ||
3098 | /* | |
3099 | * Called on failure; free up any blocks from the inode PA for this | |
3100 | * context. We don't need this for MB_GROUP_PA because we only change | |
3101 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
3102 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
3103 | */ | |
3104 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
3105 | { | |
3106 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
3107 | int len; | |
3108 | ||
3109 | if (pa && pa->pa_type == MB_INODE_PA) { | |
3110 | len = ac->ac_b_ex.fe_len; | |
3111 | pa->pa_free += len; | |
3112 | } | |
3113 | ||
3114 | } | |
3115 | ||
3116 | /* | |
3117 | * use blocks preallocated to inode | |
3118 | */ | |
3119 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3120 | struct ext4_prealloc_space *pa) | |
3121 | { | |
3122 | ext4_fsblk_t start; | |
3123 | ext4_fsblk_t end; | |
3124 | int len; | |
3125 | ||
3126 | /* found preallocated blocks, use them */ | |
3127 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
3128 | end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len); | |
3129 | len = end - start; | |
3130 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, | |
3131 | &ac->ac_b_ex.fe_start); | |
3132 | ac->ac_b_ex.fe_len = len; | |
3133 | ac->ac_status = AC_STATUS_FOUND; | |
3134 | ac->ac_pa = pa; | |
3135 | ||
3136 | BUG_ON(start < pa->pa_pstart); | |
3137 | BUG_ON(start + len > pa->pa_pstart + pa->pa_len); | |
3138 | BUG_ON(pa->pa_free < len); | |
3139 | pa->pa_free -= len; | |
3140 | ||
3141 | mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); | |
3142 | } | |
3143 | ||
3144 | /* | |
3145 | * use blocks preallocated to locality group | |
3146 | */ | |
3147 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3148 | struct ext4_prealloc_space *pa) | |
3149 | { | |
3150 | unsigned int len = ac->ac_o_ex.fe_len; | |
3151 | ||
3152 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, | |
3153 | &ac->ac_b_ex.fe_group, | |
3154 | &ac->ac_b_ex.fe_start); | |
3155 | ac->ac_b_ex.fe_len = len; | |
3156 | ac->ac_status = AC_STATUS_FOUND; | |
3157 | ac->ac_pa = pa; | |
3158 | ||
3159 | /* we don't correct pa_pstart or pa_plen here to avoid | |
3160 | * possible race when the group is being loaded concurrently | |
3161 | * instead we correct pa later, after blocks are marked | |
3162 | * in on-disk bitmap -- see ext4_mb_release_context() | |
3163 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
3164 | */ | |
3165 | mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); | |
3166 | } | |
3167 | ||
3168 | /* | |
3169 | * Return the prealloc space that have minimal distance | |
3170 | * from the goal block. @cpa is the prealloc | |
3171 | * space that is having currently known minimal distance | |
3172 | * from the goal block. | |
3173 | */ | |
3174 | static struct ext4_prealloc_space * | |
3175 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3176 | struct ext4_prealloc_space *pa, | |
3177 | struct ext4_prealloc_space *cpa) | |
3178 | { | |
3179 | ext4_fsblk_t cur_distance, new_distance; | |
3180 | ||
3181 | if (cpa == NULL) { | |
3182 | atomic_inc(&pa->pa_count); | |
3183 | return pa; | |
3184 | } | |
3185 | cur_distance = abs(goal_block - cpa->pa_pstart); | |
3186 | new_distance = abs(goal_block - pa->pa_pstart); | |
3187 | ||
3188 | if (cur_distance < new_distance) | |
3189 | return cpa; | |
3190 | ||
3191 | /* drop the previous reference */ | |
3192 | atomic_dec(&cpa->pa_count); | |
3193 | atomic_inc(&pa->pa_count); | |
3194 | return pa; | |
3195 | } | |
3196 | ||
3197 | /* | |
3198 | * search goal blocks in preallocated space | |
3199 | */ | |
3200 | static noinline_for_stack int | |
3201 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
3202 | { | |
3203 | int order, i; | |
3204 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); | |
3205 | struct ext4_locality_group *lg; | |
3206 | struct ext4_prealloc_space *pa, *cpa = NULL; | |
3207 | ext4_fsblk_t goal_block; | |
3208 | ||
3209 | /* only data can be preallocated */ | |
3210 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3211 | return 0; | |
3212 | ||
3213 | /* first, try per-file preallocation */ | |
3214 | rcu_read_lock(); | |
3215 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { | |
3216 | ||
3217 | /* all fields in this condition don't change, | |
3218 | * so we can skip locking for them */ | |
3219 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
3220 | ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len) | |
3221 | continue; | |
3222 | ||
3223 | /* non-extent files can't have physical blocks past 2^32 */ | |
3224 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && | |
3225 | pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS) | |
3226 | continue; | |
3227 | ||
3228 | /* found preallocated blocks, use them */ | |
3229 | spin_lock(&pa->pa_lock); | |
3230 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3231 | atomic_inc(&pa->pa_count); | |
3232 | ext4_mb_use_inode_pa(ac, pa); | |
3233 | spin_unlock(&pa->pa_lock); | |
3234 | ac->ac_criteria = 10; | |
3235 | rcu_read_unlock(); | |
3236 | return 1; | |
3237 | } | |
3238 | spin_unlock(&pa->pa_lock); | |
3239 | } | |
3240 | rcu_read_unlock(); | |
3241 | ||
3242 | /* can we use group allocation? */ | |
3243 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3244 | return 0; | |
3245 | ||
3246 | /* inode may have no locality group for some reason */ | |
3247 | lg = ac->ac_lg; | |
3248 | if (lg == NULL) | |
3249 | return 0; | |
3250 | order = fls(ac->ac_o_ex.fe_len) - 1; | |
3251 | if (order > PREALLOC_TB_SIZE - 1) | |
3252 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3253 | order = PREALLOC_TB_SIZE - 1; | |
3254 | ||
3255 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); | |
3256 | /* | |
3257 | * search for the prealloc space that is having | |
3258 | * minimal distance from the goal block. | |
3259 | */ | |
3260 | for (i = order; i < PREALLOC_TB_SIZE; i++) { | |
3261 | rcu_read_lock(); | |
3262 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3263 | pa_inode_list) { | |
3264 | spin_lock(&pa->pa_lock); | |
3265 | if (pa->pa_deleted == 0 && | |
3266 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
3267 | ||
3268 | cpa = ext4_mb_check_group_pa(goal_block, | |
3269 | pa, cpa); | |
3270 | } | |
3271 | spin_unlock(&pa->pa_lock); | |
3272 | } | |
3273 | rcu_read_unlock(); | |
3274 | } | |
3275 | if (cpa) { | |
3276 | ext4_mb_use_group_pa(ac, cpa); | |
3277 | ac->ac_criteria = 20; | |
3278 | return 1; | |
3279 | } | |
3280 | return 0; | |
3281 | } | |
3282 | ||
3283 | /* | |
3284 | * the function goes through all block freed in the group | |
3285 | * but not yet committed and marks them used in in-core bitmap. | |
3286 | * buddy must be generated from this bitmap | |
3287 | * Need to be called with the ext4 group lock held | |
3288 | */ | |
3289 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3290 | ext4_group_t group) | |
3291 | { | |
3292 | struct rb_node *n; | |
3293 | struct ext4_group_info *grp; | |
3294 | struct ext4_free_data *entry; | |
3295 | ||
3296 | grp = ext4_get_group_info(sb, group); | |
3297 | n = rb_first(&(grp->bb_free_root)); | |
3298 | ||
3299 | while (n) { | |
3300 | entry = rb_entry(n, struct ext4_free_data, node); | |
3301 | mb_set_bits(bitmap, entry->start_blk, entry->count); | |
3302 | n = rb_next(n); | |
3303 | } | |
3304 | return; | |
3305 | } | |
3306 | ||
3307 | /* | |
3308 | * the function goes through all preallocation in this group and marks them | |
3309 | * used in in-core bitmap. buddy must be generated from this bitmap | |
3310 | * Need to be called with ext4 group lock held | |
3311 | */ | |
3312 | static noinline_for_stack | |
3313 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
3314 | ext4_group_t group) | |
3315 | { | |
3316 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3317 | struct ext4_prealloc_space *pa; | |
3318 | struct list_head *cur; | |
3319 | ext4_group_t groupnr; | |
3320 | ext4_grpblk_t start; | |
3321 | int preallocated = 0; | |
3322 | int count = 0; | |
3323 | int len; | |
3324 | ||
3325 | /* all form of preallocation discards first load group, | |
3326 | * so the only competing code is preallocation use. | |
3327 | * we don't need any locking here | |
3328 | * notice we do NOT ignore preallocations with pa_deleted | |
3329 | * otherwise we could leave used blocks available for | |
3330 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3331 | * is dropping preallocation | |
3332 | */ | |
3333 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3334 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3335 | spin_lock(&pa->pa_lock); | |
3336 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3337 | &groupnr, &start); | |
3338 | len = pa->pa_len; | |
3339 | spin_unlock(&pa->pa_lock); | |
3340 | if (unlikely(len == 0)) | |
3341 | continue; | |
3342 | BUG_ON(groupnr != group); | |
3343 | mb_set_bits(bitmap, start, len); | |
3344 | preallocated += len; | |
3345 | count++; | |
3346 | } | |
3347 | mb_debug(1, "prellocated %u for group %u\n", preallocated, group); | |
3348 | } | |
3349 | ||
3350 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3351 | { | |
3352 | struct ext4_prealloc_space *pa; | |
3353 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
3354 | kmem_cache_free(ext4_pspace_cachep, pa); | |
3355 | } | |
3356 | ||
3357 | /* | |
3358 | * drops a reference to preallocated space descriptor | |
3359 | * if this was the last reference and the space is consumed | |
3360 | */ | |
3361 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3362 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3363 | { | |
3364 | ext4_group_t grp; | |
3365 | ext4_fsblk_t grp_blk; | |
3366 | ||
3367 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) | |
3368 | return; | |
3369 | ||
3370 | /* in this short window concurrent discard can set pa_deleted */ | |
3371 | spin_lock(&pa->pa_lock); | |
3372 | if (pa->pa_deleted == 1) { | |
3373 | spin_unlock(&pa->pa_lock); | |
3374 | return; | |
3375 | } | |
3376 | ||
3377 | pa->pa_deleted = 1; | |
3378 | spin_unlock(&pa->pa_lock); | |
3379 | ||
3380 | grp_blk = pa->pa_pstart; | |
3381 | /* | |
3382 | * If doing group-based preallocation, pa_pstart may be in the | |
3383 | * next group when pa is used up | |
3384 | */ | |
3385 | if (pa->pa_type == MB_GROUP_PA) | |
3386 | grp_blk--; | |
3387 | ||
3388 | ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL); | |
3389 | ||
3390 | /* | |
3391 | * possible race: | |
3392 | * | |
3393 | * P1 (buddy init) P2 (regular allocation) | |
3394 | * find block B in PA | |
3395 | * copy on-disk bitmap to buddy | |
3396 | * mark B in on-disk bitmap | |
3397 | * drop PA from group | |
3398 | * mark all PAs in buddy | |
3399 | * | |
3400 | * thus, P1 initializes buddy with B available. to prevent this | |
3401 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3402 | * against that pair | |
3403 | */ | |
3404 | ext4_lock_group(sb, grp); | |
3405 | list_del(&pa->pa_group_list); | |
3406 | ext4_unlock_group(sb, grp); | |
3407 | ||
3408 | spin_lock(pa->pa_obj_lock); | |
3409 | list_del_rcu(&pa->pa_inode_list); | |
3410 | spin_unlock(pa->pa_obj_lock); | |
3411 | ||
3412 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3413 | } | |
3414 | ||
3415 | /* | |
3416 | * creates new preallocated space for given inode | |
3417 | */ | |
3418 | static noinline_for_stack int | |
3419 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
3420 | { | |
3421 | struct super_block *sb = ac->ac_sb; | |
3422 | struct ext4_prealloc_space *pa; | |
3423 | struct ext4_group_info *grp; | |
3424 | struct ext4_inode_info *ei; | |
3425 | ||
3426 | /* preallocate only when found space is larger then requested */ | |
3427 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3428 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3429 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3430 | ||
3431 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3432 | if (pa == NULL) | |
3433 | return -ENOMEM; | |
3434 | ||
3435 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3436 | int winl; | |
3437 | int wins; | |
3438 | int win; | |
3439 | int offs; | |
3440 | ||
3441 | /* we can't allocate as much as normalizer wants. | |
3442 | * so, found space must get proper lstart | |
3443 | * to cover original request */ | |
3444 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3445 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3446 | ||
3447 | /* we're limited by original request in that | |
3448 | * logical block must be covered any way | |
3449 | * winl is window we can move our chunk within */ | |
3450 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3451 | ||
3452 | /* also, we should cover whole original request */ | |
3453 | wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len; | |
3454 | ||
3455 | /* the smallest one defines real window */ | |
3456 | win = min(winl, wins); | |
3457 | ||
3458 | offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len; | |
3459 | if (offs && offs < win) | |
3460 | win = offs; | |
3461 | ||
3462 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win; | |
3463 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); | |
3464 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3465 | } | |
3466 | ||
3467 | /* preallocation can change ac_b_ex, thus we store actually | |
3468 | * allocated blocks for history */ | |
3469 | ac->ac_f_ex = ac->ac_b_ex; | |
3470 | ||
3471 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3472 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3473 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3474 | pa->pa_free = pa->pa_len; | |
3475 | atomic_set(&pa->pa_count, 1); | |
3476 | spin_lock_init(&pa->pa_lock); | |
3477 | INIT_LIST_HEAD(&pa->pa_inode_list); | |
3478 | INIT_LIST_HEAD(&pa->pa_group_list); | |
3479 | pa->pa_deleted = 0; | |
3480 | pa->pa_type = MB_INODE_PA; | |
3481 | ||
3482 | mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, | |
3483 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); | |
3484 | trace_ext4_mb_new_inode_pa(ac, pa); | |
3485 | ||
3486 | ext4_mb_use_inode_pa(ac, pa); | |
3487 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3488 | ||
3489 | ei = EXT4_I(ac->ac_inode); | |
3490 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3491 | ||
3492 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3493 | pa->pa_inode = ac->ac_inode; | |
3494 | ||
3495 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3496 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3497 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3498 | ||
3499 | spin_lock(pa->pa_obj_lock); | |
3500 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3501 | spin_unlock(pa->pa_obj_lock); | |
3502 | ||
3503 | return 0; | |
3504 | } | |
3505 | ||
3506 | /* | |
3507 | * creates new preallocated space for locality group inodes belongs to | |
3508 | */ | |
3509 | static noinline_for_stack int | |
3510 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
3511 | { | |
3512 | struct super_block *sb = ac->ac_sb; | |
3513 | struct ext4_locality_group *lg; | |
3514 | struct ext4_prealloc_space *pa; | |
3515 | struct ext4_group_info *grp; | |
3516 | ||
3517 | /* preallocate only when found space is larger then requested */ | |
3518 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3519 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3520 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3521 | ||
3522 | BUG_ON(ext4_pspace_cachep == NULL); | |
3523 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3524 | if (pa == NULL) | |
3525 | return -ENOMEM; | |
3526 | ||
3527 | /* preallocation can change ac_b_ex, thus we store actually | |
3528 | * allocated blocks for history */ | |
3529 | ac->ac_f_ex = ac->ac_b_ex; | |
3530 | ||
3531 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3532 | pa->pa_lstart = pa->pa_pstart; | |
3533 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3534 | pa->pa_free = pa->pa_len; | |
3535 | atomic_set(&pa->pa_count, 1); | |
3536 | spin_lock_init(&pa->pa_lock); | |
3537 | INIT_LIST_HEAD(&pa->pa_inode_list); | |
3538 | INIT_LIST_HEAD(&pa->pa_group_list); | |
3539 | pa->pa_deleted = 0; | |
3540 | pa->pa_type = MB_GROUP_PA; | |
3541 | ||
3542 | mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, | |
3543 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); | |
3544 | trace_ext4_mb_new_group_pa(ac, pa); | |
3545 | ||
3546 | ext4_mb_use_group_pa(ac, pa); | |
3547 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3548 | ||
3549 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3550 | lg = ac->ac_lg; | |
3551 | BUG_ON(lg == NULL); | |
3552 | ||
3553 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3554 | pa->pa_inode = NULL; | |
3555 | ||
3556 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3557 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3558 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3559 | ||
3560 | /* | |
3561 | * We will later add the new pa to the right bucket | |
3562 | * after updating the pa_free in ext4_mb_release_context | |
3563 | */ | |
3564 | return 0; | |
3565 | } | |
3566 | ||
3567 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3568 | { | |
3569 | int err; | |
3570 | ||
3571 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3572 | err = ext4_mb_new_group_pa(ac); | |
3573 | else | |
3574 | err = ext4_mb_new_inode_pa(ac); | |
3575 | return err; | |
3576 | } | |
3577 | ||
3578 | /* | |
3579 | * finds all unused blocks in on-disk bitmap, frees them in | |
3580 | * in-core bitmap and buddy. | |
3581 | * @pa must be unlinked from inode and group lists, so that | |
3582 | * nobody else can find/use it. | |
3583 | * the caller MUST hold group/inode locks. | |
3584 | * TODO: optimize the case when there are no in-core structures yet | |
3585 | */ | |
3586 | static noinline_for_stack int | |
3587 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3588 | struct ext4_prealloc_space *pa) | |
3589 | { | |
3590 | struct super_block *sb = e4b->bd_sb; | |
3591 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3592 | unsigned int end; | |
3593 | unsigned int next; | |
3594 | ext4_group_t group; | |
3595 | ext4_grpblk_t bit; | |
3596 | unsigned long long grp_blk_start; | |
3597 | int err = 0; | |
3598 | int free = 0; | |
3599 | ||
3600 | BUG_ON(pa->pa_deleted == 0); | |
3601 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3602 | grp_blk_start = pa->pa_pstart - bit; | |
3603 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3604 | end = bit + pa->pa_len; | |
3605 | ||
3606 | while (bit < end) { | |
3607 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); | |
3608 | if (bit >= end) | |
3609 | break; | |
3610 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); | |
3611 | mb_debug(1, " free preallocated %u/%u in group %u\n", | |
3612 | (unsigned) ext4_group_first_block_no(sb, group) + bit, | |
3613 | (unsigned) next - bit, (unsigned) group); | |
3614 | free += next - bit; | |
3615 | ||
3616 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); | |
3617 | trace_ext4_mb_release_inode_pa(sb, pa->pa_inode, pa, | |
3618 | grp_blk_start + bit, next - bit); | |
3619 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); | |
3620 | bit = next + 1; | |
3621 | } | |
3622 | if (free != pa->pa_free) { | |
3623 | printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n", | |
3624 | pa, (unsigned long) pa->pa_lstart, | |
3625 | (unsigned long) pa->pa_pstart, | |
3626 | (unsigned long) pa->pa_len); | |
3627 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", | |
3628 | free, pa->pa_free); | |
3629 | /* | |
3630 | * pa is already deleted so we use the value obtained | |
3631 | * from the bitmap and continue. | |
3632 | */ | |
3633 | } | |
3634 | atomic_add(free, &sbi->s_mb_discarded); | |
3635 | ||
3636 | return err; | |
3637 | } | |
3638 | ||
3639 | static noinline_for_stack int | |
3640 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3641 | struct ext4_prealloc_space *pa) | |
3642 | { | |
3643 | struct super_block *sb = e4b->bd_sb; | |
3644 | ext4_group_t group; | |
3645 | ext4_grpblk_t bit; | |
3646 | ||
3647 | trace_ext4_mb_release_group_pa(sb, pa); | |
3648 | BUG_ON(pa->pa_deleted == 0); | |
3649 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3650 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3651 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3652 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3653 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); | |
3654 | ||
3655 | return 0; | |
3656 | } | |
3657 | ||
3658 | /* | |
3659 | * releases all preallocations in given group | |
3660 | * | |
3661 | * first, we need to decide discard policy: | |
3662 | * - when do we discard | |
3663 | * 1) ENOSPC | |
3664 | * - how many do we discard | |
3665 | * 1) how many requested | |
3666 | */ | |
3667 | static noinline_for_stack int | |
3668 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
3669 | ext4_group_t group, int needed) | |
3670 | { | |
3671 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3672 | struct buffer_head *bitmap_bh = NULL; | |
3673 | struct ext4_prealloc_space *pa, *tmp; | |
3674 | struct list_head list; | |
3675 | struct ext4_buddy e4b; | |
3676 | int err; | |
3677 | int busy = 0; | |
3678 | int free = 0; | |
3679 | ||
3680 | mb_debug(1, "discard preallocation for group %u\n", group); | |
3681 | ||
3682 | if (list_empty(&grp->bb_prealloc_list)) | |
3683 | return 0; | |
3684 | ||
3685 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
3686 | if (bitmap_bh == NULL) { | |
3687 | ext4_error(sb, "Error reading block bitmap for %u", group); | |
3688 | return 0; | |
3689 | } | |
3690 | ||
3691 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
3692 | if (err) { | |
3693 | ext4_error(sb, "Error loading buddy information for %u", group); | |
3694 | put_bh(bitmap_bh); | |
3695 | return 0; | |
3696 | } | |
3697 | ||
3698 | if (needed == 0) | |
3699 | needed = EXT4_BLOCKS_PER_GROUP(sb) + 1; | |
3700 | ||
3701 | INIT_LIST_HEAD(&list); | |
3702 | repeat: | |
3703 | ext4_lock_group(sb, group); | |
3704 | list_for_each_entry_safe(pa, tmp, | |
3705 | &grp->bb_prealloc_list, pa_group_list) { | |
3706 | spin_lock(&pa->pa_lock); | |
3707 | if (atomic_read(&pa->pa_count)) { | |
3708 | spin_unlock(&pa->pa_lock); | |
3709 | busy = 1; | |
3710 | continue; | |
3711 | } | |
3712 | if (pa->pa_deleted) { | |
3713 | spin_unlock(&pa->pa_lock); | |
3714 | continue; | |
3715 | } | |
3716 | ||
3717 | /* seems this one can be freed ... */ | |
3718 | pa->pa_deleted = 1; | |
3719 | ||
3720 | /* we can trust pa_free ... */ | |
3721 | free += pa->pa_free; | |
3722 | ||
3723 | spin_unlock(&pa->pa_lock); | |
3724 | ||
3725 | list_del(&pa->pa_group_list); | |
3726 | list_add(&pa->u.pa_tmp_list, &list); | |
3727 | } | |
3728 | ||
3729 | /* if we still need more blocks and some PAs were used, try again */ | |
3730 | if (free < needed && busy) { | |
3731 | busy = 0; | |
3732 | ext4_unlock_group(sb, group); | |
3733 | /* | |
3734 | * Yield the CPU here so that we don't get soft lockup | |
3735 | * in non preempt case. | |
3736 | */ | |
3737 | yield(); | |
3738 | goto repeat; | |
3739 | } | |
3740 | ||
3741 | /* found anything to free? */ | |
3742 | if (list_empty(&list)) { | |
3743 | BUG_ON(free != 0); | |
3744 | goto out; | |
3745 | } | |
3746 | ||
3747 | /* now free all selected PAs */ | |
3748 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3749 | ||
3750 | /* remove from object (inode or locality group) */ | |
3751 | spin_lock(pa->pa_obj_lock); | |
3752 | list_del_rcu(&pa->pa_inode_list); | |
3753 | spin_unlock(pa->pa_obj_lock); | |
3754 | ||
3755 | if (pa->pa_type == MB_GROUP_PA) | |
3756 | ext4_mb_release_group_pa(&e4b, pa); | |
3757 | else | |
3758 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); | |
3759 | ||
3760 | list_del(&pa->u.pa_tmp_list); | |
3761 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3762 | } | |
3763 | ||
3764 | out: | |
3765 | ext4_unlock_group(sb, group); | |
3766 | ext4_mb_unload_buddy(&e4b); | |
3767 | put_bh(bitmap_bh); | |
3768 | return free; | |
3769 | } | |
3770 | ||
3771 | /* | |
3772 | * releases all non-used preallocated blocks for given inode | |
3773 | * | |
3774 | * It's important to discard preallocations under i_data_sem | |
3775 | * We don't want another block to be served from the prealloc | |
3776 | * space when we are discarding the inode prealloc space. | |
3777 | * | |
3778 | * FIXME!! Make sure it is valid at all the call sites | |
3779 | */ | |
3780 | void ext4_discard_preallocations(struct inode *inode) | |
3781 | { | |
3782 | struct ext4_inode_info *ei = EXT4_I(inode); | |
3783 | struct super_block *sb = inode->i_sb; | |
3784 | struct buffer_head *bitmap_bh = NULL; | |
3785 | struct ext4_prealloc_space *pa, *tmp; | |
3786 | ext4_group_t group = 0; | |
3787 | struct list_head list; | |
3788 | struct ext4_buddy e4b; | |
3789 | int err; | |
3790 | ||
3791 | if (!S_ISREG(inode->i_mode)) { | |
3792 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ | |
3793 | return; | |
3794 | } | |
3795 | ||
3796 | mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); | |
3797 | trace_ext4_discard_preallocations(inode); | |
3798 | ||
3799 | INIT_LIST_HEAD(&list); | |
3800 | ||
3801 | repeat: | |
3802 | /* first, collect all pa's in the inode */ | |
3803 | spin_lock(&ei->i_prealloc_lock); | |
3804 | while (!list_empty(&ei->i_prealloc_list)) { | |
3805 | pa = list_entry(ei->i_prealloc_list.next, | |
3806 | struct ext4_prealloc_space, pa_inode_list); | |
3807 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
3808 | spin_lock(&pa->pa_lock); | |
3809 | if (atomic_read(&pa->pa_count)) { | |
3810 | /* this shouldn't happen often - nobody should | |
3811 | * use preallocation while we're discarding it */ | |
3812 | spin_unlock(&pa->pa_lock); | |
3813 | spin_unlock(&ei->i_prealloc_lock); | |
3814 | printk(KERN_ERR "uh-oh! used pa while discarding\n"); | |
3815 | WARN_ON(1); | |
3816 | schedule_timeout_uninterruptible(HZ); | |
3817 | goto repeat; | |
3818 | ||
3819 | } | |
3820 | if (pa->pa_deleted == 0) { | |
3821 | pa->pa_deleted = 1; | |
3822 | spin_unlock(&pa->pa_lock); | |
3823 | list_del_rcu(&pa->pa_inode_list); | |
3824 | list_add(&pa->u.pa_tmp_list, &list); | |
3825 | continue; | |
3826 | } | |
3827 | ||
3828 | /* someone is deleting pa right now */ | |
3829 | spin_unlock(&pa->pa_lock); | |
3830 | spin_unlock(&ei->i_prealloc_lock); | |
3831 | ||
3832 | /* we have to wait here because pa_deleted | |
3833 | * doesn't mean pa is already unlinked from | |
3834 | * the list. as we might be called from | |
3835 | * ->clear_inode() the inode will get freed | |
3836 | * and concurrent thread which is unlinking | |
3837 | * pa from inode's list may access already | |
3838 | * freed memory, bad-bad-bad */ | |
3839 | ||
3840 | /* XXX: if this happens too often, we can | |
3841 | * add a flag to force wait only in case | |
3842 | * of ->clear_inode(), but not in case of | |
3843 | * regular truncate */ | |
3844 | schedule_timeout_uninterruptible(HZ); | |
3845 | goto repeat; | |
3846 | } | |
3847 | spin_unlock(&ei->i_prealloc_lock); | |
3848 | ||
3849 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3850 | BUG_ON(pa->pa_type != MB_INODE_PA); | |
3851 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); | |
3852 | ||
3853 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
3854 | if (err) { | |
3855 | ext4_error(sb, "Error loading buddy information for %u", | |
3856 | group); | |
3857 | continue; | |
3858 | } | |
3859 | ||
3860 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
3861 | if (bitmap_bh == NULL) { | |
3862 | ext4_error(sb, "Error reading block bitmap for %u", | |
3863 | group); | |
3864 | ext4_mb_unload_buddy(&e4b); | |
3865 | continue; | |
3866 | } | |
3867 | ||
3868 | ext4_lock_group(sb, group); | |
3869 | list_del(&pa->pa_group_list); | |
3870 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); | |
3871 | ext4_unlock_group(sb, group); | |
3872 | ||
3873 | ext4_mb_unload_buddy(&e4b); | |
3874 | put_bh(bitmap_bh); | |
3875 | ||
3876 | list_del(&pa->u.pa_tmp_list); | |
3877 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3878 | } | |
3879 | } | |
3880 | ||
3881 | /* | |
3882 | * finds all preallocated spaces and return blocks being freed to them | |
3883 | * if preallocated space becomes full (no block is used from the space) | |
3884 | * then the function frees space in buddy | |
3885 | * XXX: at the moment, truncate (which is the only way to free blocks) | |
3886 | * discards all preallocations | |
3887 | */ | |
3888 | static void ext4_mb_return_to_preallocation(struct inode *inode, | |
3889 | struct ext4_buddy *e4b, | |
3890 | sector_t block, int count) | |
3891 | { | |
3892 | BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list)); | |
3893 | } | |
3894 | #ifdef CONFIG_EXT4_DEBUG | |
3895 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
3896 | { | |
3897 | struct super_block *sb = ac->ac_sb; | |
3898 | ext4_group_t ngroups, i; | |
3899 | ||
3900 | if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED) | |
3901 | return; | |
3902 | ||
3903 | printk(KERN_ERR "EXT4-fs: Can't allocate:" | |
3904 | " Allocation context details:\n"); | |
3905 | printk(KERN_ERR "EXT4-fs: status %d flags %d\n", | |
3906 | ac->ac_status, ac->ac_flags); | |
3907 | printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, " | |
3908 | "best %lu/%lu/%lu@%lu cr %d\n", | |
3909 | (unsigned long)ac->ac_o_ex.fe_group, | |
3910 | (unsigned long)ac->ac_o_ex.fe_start, | |
3911 | (unsigned long)ac->ac_o_ex.fe_len, | |
3912 | (unsigned long)ac->ac_o_ex.fe_logical, | |
3913 | (unsigned long)ac->ac_g_ex.fe_group, | |
3914 | (unsigned long)ac->ac_g_ex.fe_start, | |
3915 | (unsigned long)ac->ac_g_ex.fe_len, | |
3916 | (unsigned long)ac->ac_g_ex.fe_logical, | |
3917 | (unsigned long)ac->ac_b_ex.fe_group, | |
3918 | (unsigned long)ac->ac_b_ex.fe_start, | |
3919 | (unsigned long)ac->ac_b_ex.fe_len, | |
3920 | (unsigned long)ac->ac_b_ex.fe_logical, | |
3921 | (int)ac->ac_criteria); | |
3922 | printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned, | |
3923 | ac->ac_found); | |
3924 | printk(KERN_ERR "EXT4-fs: groups: \n"); | |
3925 | ngroups = ext4_get_groups_count(sb); | |
3926 | for (i = 0; i < ngroups; i++) { | |
3927 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); | |
3928 | struct ext4_prealloc_space *pa; | |
3929 | ext4_grpblk_t start; | |
3930 | struct list_head *cur; | |
3931 | ext4_lock_group(sb, i); | |
3932 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3933 | pa = list_entry(cur, struct ext4_prealloc_space, | |
3934 | pa_group_list); | |
3935 | spin_lock(&pa->pa_lock); | |
3936 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3937 | NULL, &start); | |
3938 | spin_unlock(&pa->pa_lock); | |
3939 | printk(KERN_ERR "PA:%u:%d:%u \n", i, | |
3940 | start, pa->pa_len); | |
3941 | } | |
3942 | ext4_unlock_group(sb, i); | |
3943 | ||
3944 | if (grp->bb_free == 0) | |
3945 | continue; | |
3946 | printk(KERN_ERR "%u: %d/%d \n", | |
3947 | i, grp->bb_free, grp->bb_fragments); | |
3948 | } | |
3949 | printk(KERN_ERR "\n"); | |
3950 | } | |
3951 | #else | |
3952 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
3953 | { | |
3954 | return; | |
3955 | } | |
3956 | #endif | |
3957 | ||
3958 | /* | |
3959 | * We use locality group preallocation for small size file. The size of the | |
3960 | * file is determined by the current size or the resulting size after | |
3961 | * allocation which ever is larger | |
3962 | * | |
3963 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req | |
3964 | */ | |
3965 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
3966 | { | |
3967 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3968 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
3969 | loff_t size, isize; | |
3970 | ||
3971 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3972 | return; | |
3973 | ||
3974 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
3975 | return; | |
3976 | ||
3977 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; | |
3978 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) | |
3979 | >> bsbits; | |
3980 | ||
3981 | if ((size == isize) && | |
3982 | !ext4_fs_is_busy(sbi) && | |
3983 | (atomic_read(&ac->ac_inode->i_writecount) == 0)) { | |
3984 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
3985 | return; | |
3986 | } | |
3987 | ||
3988 | /* don't use group allocation for large files */ | |
3989 | size = max(size, isize); | |
3990 | if (size > sbi->s_mb_stream_request) { | |
3991 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
3992 | return; | |
3993 | } | |
3994 | ||
3995 | BUG_ON(ac->ac_lg != NULL); | |
3996 | /* | |
3997 | * locality group prealloc space are per cpu. The reason for having | |
3998 | * per cpu locality group is to reduce the contention between block | |
3999 | * request from multiple CPUs. | |
4000 | */ | |
4001 | ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups); | |
4002 | ||
4003 | /* we're going to use group allocation */ | |
4004 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4005 | ||
4006 | /* serialize all allocations in the group */ | |
4007 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4008 | } | |
4009 | ||
4010 | static noinline_for_stack int | |
4011 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
4012 | struct ext4_allocation_request *ar) | |
4013 | { | |
4014 | struct super_block *sb = ar->inode->i_sb; | |
4015 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4016 | struct ext4_super_block *es = sbi->s_es; | |
4017 | ext4_group_t group; | |
4018 | unsigned int len; | |
4019 | ext4_fsblk_t goal; | |
4020 | ext4_grpblk_t block; | |
4021 | ||
4022 | /* we can't allocate > group size */ | |
4023 | len = ar->len; | |
4024 | ||
4025 | /* just a dirty hack to filter too big requests */ | |
4026 | if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10) | |
4027 | len = EXT4_BLOCKS_PER_GROUP(sb) - 10; | |
4028 | ||
4029 | /* start searching from the goal */ | |
4030 | goal = ar->goal; | |
4031 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4032 | goal >= ext4_blocks_count(es)) | |
4033 | goal = le32_to_cpu(es->s_first_data_block); | |
4034 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4035 | ||
4036 | /* set up allocation goals */ | |
4037 | memset(ac, 0, sizeof(struct ext4_allocation_context)); | |
4038 | ac->ac_b_ex.fe_logical = ar->logical; | |
4039 | ac->ac_status = AC_STATUS_CONTINUE; | |
4040 | ac->ac_sb = sb; | |
4041 | ac->ac_inode = ar->inode; | |
4042 | ac->ac_o_ex.fe_logical = ar->logical; | |
4043 | ac->ac_o_ex.fe_group = group; | |
4044 | ac->ac_o_ex.fe_start = block; | |
4045 | ac->ac_o_ex.fe_len = len; | |
4046 | ac->ac_g_ex.fe_logical = ar->logical; | |
4047 | ac->ac_g_ex.fe_group = group; | |
4048 | ac->ac_g_ex.fe_start = block; | |
4049 | ac->ac_g_ex.fe_len = len; | |
4050 | ac->ac_flags = ar->flags; | |
4051 | ||
4052 | /* we have to define context: we'll we work with a file or | |
4053 | * locality group. this is a policy, actually */ | |
4054 | ext4_mb_group_or_file(ac); | |
4055 | ||
4056 | mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " | |
4057 | "left: %u/%u, right %u/%u to %swritable\n", | |
4058 | (unsigned) ar->len, (unsigned) ar->logical, | |
4059 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4060 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4061 | (unsigned) ar->lright, (unsigned) ar->pright, | |
4062 | atomic_read(&ar->inode->i_writecount) ? "" : "non-"); | |
4063 | return 0; | |
4064 | ||
4065 | } | |
4066 | ||
4067 | static noinline_for_stack void | |
4068 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4069 | struct ext4_locality_group *lg, | |
4070 | int order, int total_entries) | |
4071 | { | |
4072 | ext4_group_t group = 0; | |
4073 | struct ext4_buddy e4b; | |
4074 | struct list_head discard_list; | |
4075 | struct ext4_prealloc_space *pa, *tmp; | |
4076 | ||
4077 | mb_debug(1, "discard locality group preallocation\n"); | |
4078 | ||
4079 | INIT_LIST_HEAD(&discard_list); | |
4080 | ||
4081 | spin_lock(&lg->lg_prealloc_lock); | |
4082 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
4083 | pa_inode_list) { | |
4084 | spin_lock(&pa->pa_lock); | |
4085 | if (atomic_read(&pa->pa_count)) { | |
4086 | /* | |
4087 | * This is the pa that we just used | |
4088 | * for block allocation. So don't | |
4089 | * free that | |
4090 | */ | |
4091 | spin_unlock(&pa->pa_lock); | |
4092 | continue; | |
4093 | } | |
4094 | if (pa->pa_deleted) { | |
4095 | spin_unlock(&pa->pa_lock); | |
4096 | continue; | |
4097 | } | |
4098 | /* only lg prealloc space */ | |
4099 | BUG_ON(pa->pa_type != MB_GROUP_PA); | |
4100 | ||
4101 | /* seems this one can be freed ... */ | |
4102 | pa->pa_deleted = 1; | |
4103 | spin_unlock(&pa->pa_lock); | |
4104 | ||
4105 | list_del_rcu(&pa->pa_inode_list); | |
4106 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4107 | ||
4108 | total_entries--; | |
4109 | if (total_entries <= 5) { | |
4110 | /* | |
4111 | * we want to keep only 5 entries | |
4112 | * allowing it to grow to 8. This | |
4113 | * mak sure we don't call discard | |
4114 | * soon for this list. | |
4115 | */ | |
4116 | break; | |
4117 | } | |
4118 | } | |
4119 | spin_unlock(&lg->lg_prealloc_lock); | |
4120 | ||
4121 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
4122 | ||
4123 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); | |
4124 | if (ext4_mb_load_buddy(sb, group, &e4b)) { | |
4125 | ext4_error(sb, "Error loading buddy information for %u", | |
4126 | group); | |
4127 | continue; | |
4128 | } | |
4129 | ext4_lock_group(sb, group); | |
4130 | list_del(&pa->pa_group_list); | |
4131 | ext4_mb_release_group_pa(&e4b, pa); | |
4132 | ext4_unlock_group(sb, group); | |
4133 | ||
4134 | ext4_mb_unload_buddy(&e4b); | |
4135 | list_del(&pa->u.pa_tmp_list); | |
4136 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4137 | } | |
4138 | } | |
4139 | ||
4140 | /* | |
4141 | * We have incremented pa_count. So it cannot be freed at this | |
4142 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4143 | * possible from this lg. That means pa_free cannot be updated. | |
4144 | * | |
4145 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4146 | * which can cause the lg_prealloc_list to be updated. | |
4147 | */ | |
4148 | ||
4149 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4150 | { | |
4151 | int order, added = 0, lg_prealloc_count = 1; | |
4152 | struct super_block *sb = ac->ac_sb; | |
4153 | struct ext4_locality_group *lg = ac->ac_lg; | |
4154 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4155 | ||
4156 | order = fls(pa->pa_free) - 1; | |
4157 | if (order > PREALLOC_TB_SIZE - 1) | |
4158 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4159 | order = PREALLOC_TB_SIZE - 1; | |
4160 | /* Add the prealloc space to lg */ | |
4161 | rcu_read_lock(); | |
4162 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], | |
4163 | pa_inode_list) { | |
4164 | spin_lock(&tmp_pa->pa_lock); | |
4165 | if (tmp_pa->pa_deleted) { | |
4166 | spin_unlock(&tmp_pa->pa_lock); | |
4167 | continue; | |
4168 | } | |
4169 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4170 | /* Add to the tail of the previous entry */ | |
4171 | list_add_tail_rcu(&pa->pa_inode_list, | |
4172 | &tmp_pa->pa_inode_list); | |
4173 | added = 1; | |
4174 | /* | |
4175 | * we want to count the total | |
4176 | * number of entries in the list | |
4177 | */ | |
4178 | } | |
4179 | spin_unlock(&tmp_pa->pa_lock); | |
4180 | lg_prealloc_count++; | |
4181 | } | |
4182 | if (!added) | |
4183 | list_add_tail_rcu(&pa->pa_inode_list, | |
4184 | &lg->lg_prealloc_list[order]); | |
4185 | rcu_read_unlock(); | |
4186 | ||
4187 | /* Now trim the list to be not more than 8 elements */ | |
4188 | if (lg_prealloc_count > 8) { | |
4189 | ext4_mb_discard_lg_preallocations(sb, lg, | |
4190 | order, lg_prealloc_count); | |
4191 | return; | |
4192 | } | |
4193 | return ; | |
4194 | } | |
4195 | ||
4196 | /* | |
4197 | * release all resource we used in allocation | |
4198 | */ | |
4199 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4200 | { | |
4201 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
4202 | if (pa) { | |
4203 | if (pa->pa_type == MB_GROUP_PA) { | |
4204 | /* see comment in ext4_mb_use_group_pa() */ | |
4205 | spin_lock(&pa->pa_lock); | |
4206 | pa->pa_pstart += ac->ac_b_ex.fe_len; | |
4207 | pa->pa_lstart += ac->ac_b_ex.fe_len; | |
4208 | pa->pa_free -= ac->ac_b_ex.fe_len; | |
4209 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4210 | spin_unlock(&pa->pa_lock); | |
4211 | } | |
4212 | } | |
4213 | if (ac->alloc_semp) | |
4214 | up_read(ac->alloc_semp); | |
4215 | if (pa) { | |
4216 | /* | |
4217 | * We want to add the pa to the right bucket. | |
4218 | * Remove it from the list and while adding | |
4219 | * make sure the list to which we are adding | |
4220 | * doesn't grow big. We need to release | |
4221 | * alloc_semp before calling ext4_mb_add_n_trim() | |
4222 | */ | |
4223 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { | |
4224 | spin_lock(pa->pa_obj_lock); | |
4225 | list_del_rcu(&pa->pa_inode_list); | |
4226 | spin_unlock(pa->pa_obj_lock); | |
4227 | ext4_mb_add_n_trim(ac); | |
4228 | } | |
4229 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4230 | } | |
4231 | if (ac->ac_bitmap_page) | |
4232 | page_cache_release(ac->ac_bitmap_page); | |
4233 | if (ac->ac_buddy_page) | |
4234 | page_cache_release(ac->ac_buddy_page); | |
4235 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
4236 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4237 | ext4_mb_collect_stats(ac); | |
4238 | return 0; | |
4239 | } | |
4240 | ||
4241 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4242 | { | |
4243 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); | |
4244 | int ret; | |
4245 | int freed = 0; | |
4246 | ||
4247 | trace_ext4_mb_discard_preallocations(sb, needed); | |
4248 | for (i = 0; i < ngroups && needed > 0; i++) { | |
4249 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); | |
4250 | freed += ret; | |
4251 | needed -= ret; | |
4252 | } | |
4253 | ||
4254 | return freed; | |
4255 | } | |
4256 | ||
4257 | /* | |
4258 | * Main entry point into mballoc to allocate blocks | |
4259 | * it tries to use preallocation first, then falls back | |
4260 | * to usual allocation | |
4261 | */ | |
4262 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
4263 | struct ext4_allocation_request *ar, int *errp) | |
4264 | { | |
4265 | int freed; | |
4266 | struct ext4_allocation_context *ac = NULL; | |
4267 | struct ext4_sb_info *sbi; | |
4268 | struct super_block *sb; | |
4269 | ext4_fsblk_t block = 0; | |
4270 | unsigned int inquota = 0; | |
4271 | unsigned int reserv_blks = 0; | |
4272 | ||
4273 | sb = ar->inode->i_sb; | |
4274 | sbi = EXT4_SB(sb); | |
4275 | ||
4276 | trace_ext4_request_blocks(ar); | |
4277 | ||
4278 | /* | |
4279 | * For delayed allocation, we could skip the ENOSPC and | |
4280 | * EDQUOT check, as blocks and quotas have been already | |
4281 | * reserved when data being copied into pagecache. | |
4282 | */ | |
4283 | if (EXT4_I(ar->inode)->i_delalloc_reserved_flag) | |
4284 | ar->flags |= EXT4_MB_DELALLOC_RESERVED; | |
4285 | else { | |
4286 | /* Without delayed allocation we need to verify | |
4287 | * there is enough free blocks to do block allocation | |
4288 | * and verify allocation doesn't exceed the quota limits. | |
4289 | */ | |
4290 | while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) { | |
4291 | /* let others to free the space */ | |
4292 | yield(); | |
4293 | ar->len = ar->len >> 1; | |
4294 | } | |
4295 | if (!ar->len) { | |
4296 | *errp = -ENOSPC; | |
4297 | return 0; | |
4298 | } | |
4299 | reserv_blks = ar->len; | |
4300 | while (ar->len && dquot_alloc_block(ar->inode, ar->len)) { | |
4301 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
4302 | ar->len--; | |
4303 | } | |
4304 | inquota = ar->len; | |
4305 | if (ar->len == 0) { | |
4306 | *errp = -EDQUOT; | |
4307 | goto out; | |
4308 | } | |
4309 | } | |
4310 | ||
4311 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); | |
4312 | if (!ac) { | |
4313 | ar->len = 0; | |
4314 | *errp = -ENOMEM; | |
4315 | goto out; | |
4316 | } | |
4317 | ||
4318 | *errp = ext4_mb_initialize_context(ac, ar); | |
4319 | if (*errp) { | |
4320 | ar->len = 0; | |
4321 | goto out; | |
4322 | } | |
4323 | ||
4324 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; | |
4325 | if (!ext4_mb_use_preallocated(ac)) { | |
4326 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; | |
4327 | ext4_mb_normalize_request(ac, ar); | |
4328 | repeat: | |
4329 | /* allocate space in core */ | |
4330 | *errp = ext4_mb_regular_allocator(ac); | |
4331 | if (*errp) | |
4332 | goto errout; | |
4333 | ||
4334 | /* as we've just preallocated more space than | |
4335 | * user requested orinally, we store allocated | |
4336 | * space in a special descriptor */ | |
4337 | if (ac->ac_status == AC_STATUS_FOUND && | |
4338 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
4339 | ext4_mb_new_preallocation(ac); | |
4340 | } | |
4341 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { | |
4342 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks); | |
4343 | if (*errp == -EAGAIN) { | |
4344 | /* | |
4345 | * drop the reference that we took | |
4346 | * in ext4_mb_use_best_found | |
4347 | */ | |
4348 | ext4_mb_release_context(ac); | |
4349 | ac->ac_b_ex.fe_group = 0; | |
4350 | ac->ac_b_ex.fe_start = 0; | |
4351 | ac->ac_b_ex.fe_len = 0; | |
4352 | ac->ac_status = AC_STATUS_CONTINUE; | |
4353 | goto repeat; | |
4354 | } else if (*errp) | |
4355 | errout: | |
4356 | ext4_discard_allocated_blocks(ac); | |
4357 | else { | |
4358 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4359 | ar->len = ac->ac_b_ex.fe_len; | |
4360 | } | |
4361 | } else { | |
4362 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); | |
4363 | if (freed) | |
4364 | goto repeat; | |
4365 | *errp = -ENOSPC; | |
4366 | } | |
4367 | ||
4368 | if (*errp) { | |
4369 | ac->ac_b_ex.fe_len = 0; | |
4370 | ar->len = 0; | |
4371 | ext4_mb_show_ac(ac); | |
4372 | } | |
4373 | ext4_mb_release_context(ac); | |
4374 | out: | |
4375 | if (ac) | |
4376 | kmem_cache_free(ext4_ac_cachep, ac); | |
4377 | if (inquota && ar->len < inquota) | |
4378 | dquot_free_block(ar->inode, inquota - ar->len); | |
4379 | if (!ar->len) { | |
4380 | if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) | |
4381 | /* release all the reserved blocks if non delalloc */ | |
4382 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, | |
4383 | reserv_blks); | |
4384 | } | |
4385 | ||
4386 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); | |
4387 | ||
4388 | return block; | |
4389 | } | |
4390 | ||
4391 | /* | |
4392 | * We can merge two free data extents only if the physical blocks | |
4393 | * are contiguous, AND the extents were freed by the same transaction, | |
4394 | * AND the blocks are associated with the same group. | |
4395 | */ | |
4396 | static int can_merge(struct ext4_free_data *entry1, | |
4397 | struct ext4_free_data *entry2) | |
4398 | { | |
4399 | if ((entry1->t_tid == entry2->t_tid) && | |
4400 | (entry1->group == entry2->group) && | |
4401 | ((entry1->start_blk + entry1->count) == entry2->start_blk)) | |
4402 | return 1; | |
4403 | return 0; | |
4404 | } | |
4405 | ||
4406 | static noinline_for_stack int | |
4407 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
4408 | struct ext4_free_data *new_entry) | |
4409 | { | |
4410 | ext4_group_t group = e4b->bd_group; | |
4411 | ext4_grpblk_t block; | |
4412 | struct ext4_free_data *entry; | |
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); | |
4416 | struct rb_node **n = &db->bb_free_root.rb_node, *node; | |
4417 | struct rb_node *parent = NULL, *new_node; | |
4418 | ||
4419 | BUG_ON(!ext4_handle_valid(handle)); | |
4420 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
4421 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4422 | ||
4423 | new_node = &new_entry->node; | |
4424 | block = new_entry->start_blk; | |
4425 | ||
4426 | if (!*n) { | |
4427 | /* first free block exent. We need to | |
4428 | protect buddy cache from being freed, | |
4429 | * otherwise we'll refresh it from | |
4430 | * on-disk bitmap and lose not-yet-available | |
4431 | * blocks */ | |
4432 | page_cache_get(e4b->bd_buddy_page); | |
4433 | page_cache_get(e4b->bd_bitmap_page); | |
4434 | } | |
4435 | while (*n) { | |
4436 | parent = *n; | |
4437 | entry = rb_entry(parent, struct ext4_free_data, node); | |
4438 | if (block < entry->start_blk) | |
4439 | n = &(*n)->rb_left; | |
4440 | else if (block >= (entry->start_blk + entry->count)) | |
4441 | n = &(*n)->rb_right; | |
4442 | else { | |
4443 | ext4_grp_locked_error(sb, group, 0, | |
4444 | ext4_group_first_block_no(sb, group) + block, | |
4445 | "Block already on to-be-freed list"); | |
4446 | return 0; | |
4447 | } | |
4448 | } | |
4449 | ||
4450 | rb_link_node(new_node, parent, n); | |
4451 | rb_insert_color(new_node, &db->bb_free_root); | |
4452 | ||
4453 | /* Now try to see the extent can be merged to left and right */ | |
4454 | node = rb_prev(new_node); | |
4455 | if (node) { | |
4456 | entry = rb_entry(node, struct ext4_free_data, node); | |
4457 | if (can_merge(entry, new_entry)) { | |
4458 | new_entry->start_blk = entry->start_blk; | |
4459 | new_entry->count += entry->count; | |
4460 | rb_erase(node, &(db->bb_free_root)); | |
4461 | spin_lock(&sbi->s_md_lock); | |
4462 | list_del(&entry->list); | |
4463 | spin_unlock(&sbi->s_md_lock); | |
4464 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
4465 | } | |
4466 | } | |
4467 | ||
4468 | node = rb_next(new_node); | |
4469 | if (node) { | |
4470 | entry = rb_entry(node, struct ext4_free_data, node); | |
4471 | if (can_merge(new_entry, entry)) { | |
4472 | new_entry->count += entry->count; | |
4473 | rb_erase(node, &(db->bb_free_root)); | |
4474 | spin_lock(&sbi->s_md_lock); | |
4475 | list_del(&entry->list); | |
4476 | spin_unlock(&sbi->s_md_lock); | |
4477 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
4478 | } | |
4479 | } | |
4480 | /* Add the extent to transaction's private list */ | |
4481 | spin_lock(&sbi->s_md_lock); | |
4482 | list_add(&new_entry->list, &handle->h_transaction->t_private_list); | |
4483 | spin_unlock(&sbi->s_md_lock); | |
4484 | return 0; | |
4485 | } | |
4486 | ||
4487 | /** | |
4488 | * ext4_free_blocks() -- Free given blocks and update quota | |
4489 | * @handle: handle for this transaction | |
4490 | * @inode: inode | |
4491 | * @block: start physical block to free | |
4492 | * @count: number of blocks to count | |
4493 | * @metadata: Are these metadata blocks | |
4494 | */ | |
4495 | void ext4_free_blocks(handle_t *handle, struct inode *inode, | |
4496 | struct buffer_head *bh, ext4_fsblk_t block, | |
4497 | unsigned long count, int flags) | |
4498 | { | |
4499 | struct buffer_head *bitmap_bh = NULL; | |
4500 | struct super_block *sb = inode->i_sb; | |
4501 | struct ext4_group_desc *gdp; | |
4502 | unsigned long freed = 0; | |
4503 | unsigned int overflow; | |
4504 | ext4_grpblk_t bit; | |
4505 | struct buffer_head *gd_bh; | |
4506 | ext4_group_t block_group; | |
4507 | struct ext4_sb_info *sbi; | |
4508 | struct ext4_buddy e4b; | |
4509 | int err = 0; | |
4510 | int ret; | |
4511 | ||
4512 | if (bh) { | |
4513 | if (block) | |
4514 | BUG_ON(block != bh->b_blocknr); | |
4515 | else | |
4516 | block = bh->b_blocknr; | |
4517 | } | |
4518 | ||
4519 | sbi = EXT4_SB(sb); | |
4520 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && | |
4521 | !ext4_data_block_valid(sbi, block, count)) { | |
4522 | ext4_error(sb, "Freeing blocks not in datazone - " | |
4523 | "block = %llu, count = %lu", block, count); | |
4524 | goto error_return; | |
4525 | } | |
4526 | ||
4527 | ext4_debug("freeing block %llu\n", block); | |
4528 | trace_ext4_free_blocks(inode, block, count, flags); | |
4529 | ||
4530 | if (flags & EXT4_FREE_BLOCKS_FORGET) { | |
4531 | struct buffer_head *tbh = bh; | |
4532 | int i; | |
4533 | ||
4534 | BUG_ON(bh && (count > 1)); | |
4535 | ||
4536 | for (i = 0; i < count; i++) { | |
4537 | if (!bh) | |
4538 | tbh = sb_find_get_block(inode->i_sb, | |
4539 | block + i); | |
4540 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, | |
4541 | inode, tbh, block + i); | |
4542 | } | |
4543 | } | |
4544 | ||
4545 | /* | |
4546 | * We need to make sure we don't reuse the freed block until | |
4547 | * after the transaction is committed, which we can do by | |
4548 | * treating the block as metadata, below. We make an | |
4549 | * exception if the inode is to be written in writeback mode | |
4550 | * since writeback mode has weak data consistency guarantees. | |
4551 | */ | |
4552 | if (!ext4_should_writeback_data(inode)) | |
4553 | flags |= EXT4_FREE_BLOCKS_METADATA; | |
4554 | ||
4555 | do_more: | |
4556 | overflow = 0; | |
4557 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4558 | ||
4559 | /* | |
4560 | * Check to see if we are freeing blocks across a group | |
4561 | * boundary. | |
4562 | */ | |
4563 | if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { | |
4564 | overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb); | |
4565 | count -= overflow; | |
4566 | } | |
4567 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); | |
4568 | if (!bitmap_bh) { | |
4569 | err = -EIO; | |
4570 | goto error_return; | |
4571 | } | |
4572 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); | |
4573 | if (!gdp) { | |
4574 | err = -EIO; | |
4575 | goto error_return; | |
4576 | } | |
4577 | ||
4578 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4579 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4580 | in_range(block, ext4_inode_table(sb, gdp), | |
4581 | EXT4_SB(sb)->s_itb_per_group) || | |
4582 | in_range(block + count - 1, ext4_inode_table(sb, gdp), | |
4583 | EXT4_SB(sb)->s_itb_per_group)) { | |
4584 | ||
4585 | ext4_error(sb, "Freeing blocks in system zone - " | |
4586 | "Block = %llu, count = %lu", block, count); | |
4587 | /* err = 0. ext4_std_error should be a no op */ | |
4588 | goto error_return; | |
4589 | } | |
4590 | ||
4591 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4592 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4593 | if (err) | |
4594 | goto error_return; | |
4595 | ||
4596 | /* | |
4597 | * We are about to modify some metadata. Call the journal APIs | |
4598 | * to unshare ->b_data if a currently-committing transaction is | |
4599 | * using it | |
4600 | */ | |
4601 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4602 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4603 | if (err) | |
4604 | goto error_return; | |
4605 | #ifdef AGGRESSIVE_CHECK | |
4606 | { | |
4607 | int i; | |
4608 | for (i = 0; i < count; i++) | |
4609 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); | |
4610 | } | |
4611 | #endif | |
4612 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count); | |
4613 | ||
4614 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
4615 | if (err) | |
4616 | goto error_return; | |
4617 | ||
4618 | if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) { | |
4619 | struct ext4_free_data *new_entry; | |
4620 | /* | |
4621 | * blocks being freed are metadata. these blocks shouldn't | |
4622 | * be used until this transaction is committed | |
4623 | */ | |
4624 | new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS); | |
4625 | new_entry->start_blk = bit; | |
4626 | new_entry->group = block_group; | |
4627 | new_entry->count = count; | |
4628 | new_entry->t_tid = handle->h_transaction->t_tid; | |
4629 | ||
4630 | ext4_lock_group(sb, block_group); | |
4631 | mb_clear_bits(bitmap_bh->b_data, bit, count); | |
4632 | ext4_mb_free_metadata(handle, &e4b, new_entry); | |
4633 | } else { | |
4634 | /* need to update group_info->bb_free and bitmap | |
4635 | * with group lock held. generate_buddy look at | |
4636 | * them with group lock_held | |
4637 | */ | |
4638 | if (test_opt(sb, DISCARD)) | |
4639 | ext4_issue_discard(sb, block_group, bit, count); | |
4640 | ext4_lock_group(sb, block_group); | |
4641 | mb_clear_bits(bitmap_bh->b_data, bit, count); | |
4642 | mb_free_blocks(inode, &e4b, bit, count); | |
4643 | ext4_mb_return_to_preallocation(inode, &e4b, block, count); | |
4644 | } | |
4645 | ||
4646 | ret = ext4_free_blks_count(sb, gdp) + count; | |
4647 | ext4_free_blks_set(sb, gdp, ret); | |
4648 | gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); | |
4649 | ext4_unlock_group(sb, block_group); | |
4650 | percpu_counter_add(&sbi->s_freeblocks_counter, count); | |
4651 | ||
4652 | if (sbi->s_log_groups_per_flex) { | |
4653 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
4654 | atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks); | |
4655 | } | |
4656 | ||
4657 | ext4_mb_unload_buddy(&e4b); | |
4658 | ||
4659 | freed += count; | |
4660 | ||
4661 | /* We dirtied the bitmap block */ | |
4662 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4663 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4664 | ||
4665 | /* And the group descriptor block */ | |
4666 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
4667 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
4668 | if (!err) | |
4669 | err = ret; | |
4670 | ||
4671 | if (overflow && !err) { | |
4672 | block += count; | |
4673 | count = overflow; | |
4674 | put_bh(bitmap_bh); | |
4675 | goto do_more; | |
4676 | } | |
4677 | ext4_mark_super_dirty(sb); | |
4678 | error_return: | |
4679 | if (freed) | |
4680 | dquot_free_block(inode, freed); | |
4681 | brelse(bitmap_bh); | |
4682 | ext4_std_error(sb, err); | |
4683 | return; | |
4684 | } |