Commit | Line | Data |
---|---|---|
c9de560d AT |
1 | /* |
2 | * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com | |
3 | * Written by Alex Tomas <alex@clusterfs.com> | |
4 | * | |
5 | * This program is free software; you can redistribute it and/or modify | |
6 | * it under the terms of the GNU General Public License version 2 as | |
7 | * published by the Free Software Foundation. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public Licens | |
15 | * along with this program; if not, write to the Free Software | |
16 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111- | |
17 | */ | |
18 | ||
19 | ||
20 | /* | |
21 | * mballoc.c contains the multiblocks allocation routines | |
22 | */ | |
23 | ||
8f6e39a7 | 24 | #include "mballoc.h" |
9bffad1e TT |
25 | #include <trace/events/ext4.h> |
26 | ||
c9de560d AT |
27 | /* |
28 | * MUSTDO: | |
29 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
30 | * - search for metadata in few groups | |
31 | * | |
32 | * TODO v4: | |
33 | * - normalization should take into account whether file is still open | |
34 | * - discard preallocations if no free space left (policy?) | |
35 | * - don't normalize tails | |
36 | * - quota | |
37 | * - reservation for superuser | |
38 | * | |
39 | * TODO v3: | |
40 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
41 | * - track min/max extents in each group for better group selection | |
42 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
43 | * - tree of groups sorted by number of free blocks | |
44 | * - error handling | |
45 | */ | |
46 | ||
47 | /* | |
48 | * The allocation request involve request for multiple number of blocks | |
49 | * near to the goal(block) value specified. | |
50 | * | |
b713a5ec TT |
51 | * During initialization phase of the allocator we decide to use the |
52 | * group preallocation or inode preallocation depending on the size of | |
53 | * the file. The size of the file could be the resulting file size we | |
54 | * would have after allocation, or the current file size, which ever | |
55 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
56 | * select to use the group preallocation. The default value of | |
57 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
58 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
59 | * terms of number of blocks. | |
c9de560d AT |
60 | * |
61 | * The main motivation for having small file use group preallocation is to | |
b713a5ec | 62 | * ensure that we have small files closer together on the disk. |
c9de560d | 63 | * |
b713a5ec TT |
64 | * First stage the allocator looks at the inode prealloc list, |
65 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
66 | * spaces for this particular inode. The inode prealloc space is | |
67 | * represented as: | |
c9de560d AT |
68 | * |
69 | * pa_lstart -> the logical start block for this prealloc space | |
70 | * pa_pstart -> the physical start block for this prealloc space | |
71 | * pa_len -> lenght for this prealloc space | |
72 | * pa_free -> free space available in this prealloc space | |
73 | * | |
74 | * The inode preallocation space is used looking at the _logical_ start | |
75 | * block. If only the logical file block falls within the range of prealloc | |
76 | * space we will consume the particular prealloc space. This make sure that | |
77 | * that the we have contiguous physical blocks representing the file blocks | |
78 | * | |
79 | * The important thing to be noted in case of inode prealloc space is that | |
80 | * we don't modify the values associated to inode prealloc space except | |
81 | * pa_free. | |
82 | * | |
83 | * If we are not able to find blocks in the inode prealloc space and if we | |
84 | * have the group allocation flag set then we look at the locality group | |
85 | * prealloc space. These are per CPU prealloc list repreasented as | |
86 | * | |
87 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
88 | * | |
89 | * The reason for having a per cpu locality group is to reduce the contention | |
90 | * between CPUs. It is possible to get scheduled at this point. | |
91 | * | |
92 | * The locality group prealloc space is used looking at whether we have | |
93 | * enough free space (pa_free) withing the prealloc space. | |
94 | * | |
95 | * If we can't allocate blocks via inode prealloc or/and locality group | |
96 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
97 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
98 | * mapped to the buddy and bitmap information regarding different | |
99 | * groups. The buddy information is attached to buddy cache inode so that | |
100 | * we can access them through the page cache. The information regarding | |
101 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
102 | * block bitmap and buddy information. The information are stored in the | |
103 | * inode as: | |
104 | * | |
105 | * { page } | |
c3a326a6 | 106 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
107 | * |
108 | * | |
109 | * one block each for bitmap and buddy information. So for each group we | |
110 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / | |
111 | * blocksize) blocks. So it can have information regarding groups_per_page | |
112 | * which is blocks_per_page/2 | |
113 | * | |
114 | * The buddy cache inode is not stored on disk. The inode is thrown | |
115 | * away when the filesystem is unmounted. | |
116 | * | |
117 | * We look for count number of blocks in the buddy cache. If we were able | |
118 | * to locate that many free blocks we return with additional information | |
119 | * regarding rest of the contiguous physical block available | |
120 | * | |
121 | * Before allocating blocks via buddy cache we normalize the request | |
122 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
123 | * blocks that we get after allocation is added to the respective prealloc | |
124 | * list. In case of inode preallocation we follow a list of heuristics | |
125 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
126 | * we are doing a group prealloc we try to normalize the request to | |
b713a5ec | 127 | * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is |
c9de560d | 128 | * 512 blocks. This can be tuned via |
b713a5ec | 129 | * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in |
c9de560d AT |
130 | * terms of number of blocks. If we have mounted the file system with -O |
131 | * stripe=<value> option the group prealloc request is normalized to the | |
132 | * stripe value (sbi->s_stripe) | |
133 | * | |
b713a5ec | 134 | * The regular allocator(using the buddy cache) supports few tunables. |
c9de560d | 135 | * |
b713a5ec TT |
136 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
137 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
138 | * /sys/fs/ext4/<partition>/mb_order2_req | |
c9de560d | 139 | * |
b713a5ec | 140 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
141 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
142 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 143 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
c9de560d | 144 | * stripe size (sbi->s_stripe), we try to search for contigous block in |
b713a5ec TT |
145 | * stripe size. This should result in better allocation on RAID setups. If |
146 | * not, we search in the specific group using bitmap for best extents. The | |
147 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 148 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 149 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
150 | * best extent in the found extents. Searching for the blocks starts with |
151 | * the group specified as the goal value in allocation context via | |
152 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
153 | * can used for allocation. ext4_mb_good_group explains how the groups are | |
154 | * checked. | |
155 | * | |
156 | * Both the prealloc space are getting populated as above. So for the first | |
157 | * request we will hit the buddy cache which will result in this prealloc | |
158 | * space getting filled. The prealloc space is then later used for the | |
159 | * subsequent request. | |
160 | */ | |
161 | ||
162 | /* | |
163 | * mballoc operates on the following data: | |
164 | * - on-disk bitmap | |
165 | * - in-core buddy (actually includes buddy and bitmap) | |
166 | * - preallocation descriptors (PAs) | |
167 | * | |
168 | * there are two types of preallocations: | |
169 | * - inode | |
170 | * assiged to specific inode and can be used for this inode only. | |
171 | * it describes part of inode's space preallocated to specific | |
172 | * physical blocks. any block from that preallocated can be used | |
173 | * independent. the descriptor just tracks number of blocks left | |
174 | * unused. so, before taking some block from descriptor, one must | |
175 | * make sure corresponded logical block isn't allocated yet. this | |
176 | * also means that freeing any block within descriptor's range | |
177 | * must discard all preallocated blocks. | |
178 | * - locality group | |
179 | * assigned to specific locality group which does not translate to | |
180 | * permanent set of inodes: inode can join and leave group. space | |
181 | * from this type of preallocation can be used for any inode. thus | |
182 | * it's consumed from the beginning to the end. | |
183 | * | |
184 | * relation between them can be expressed as: | |
185 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
186 | * | |
187 | * this mean blocks mballoc considers used are: | |
188 | * - allocated blocks (persistent) | |
189 | * - preallocated blocks (non-persistent) | |
190 | * | |
191 | * consistency in mballoc world means that at any time a block is either | |
192 | * free or used in ALL structures. notice: "any time" should not be read | |
193 | * literally -- time is discrete and delimited by locks. | |
194 | * | |
195 | * to keep it simple, we don't use block numbers, instead we count number of | |
196 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
197 | * | |
198 | * all operations can be expressed as: | |
199 | * - init buddy: buddy = on-disk + PAs | |
200 | * - new PA: buddy += N; PA = N | |
201 | * - use inode PA: on-disk += N; PA -= N | |
202 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
203 | * - use locality group PA on-disk += N; PA -= N | |
204 | * - discard locality group PA buddy -= PA; PA = 0 | |
205 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
206 | * is used in real operation because we can't know actual used | |
207 | * bits from PA, only from on-disk bitmap | |
208 | * | |
209 | * if we follow this strict logic, then all operations above should be atomic. | |
210 | * given some of them can block, we'd have to use something like semaphores | |
211 | * killing performance on high-end SMP hardware. let's try to relax it using | |
212 | * the following knowledge: | |
213 | * 1) if buddy is referenced, it's already initialized | |
214 | * 2) while block is used in buddy and the buddy is referenced, | |
215 | * nobody can re-allocate that block | |
216 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
217 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
218 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
219 | * block | |
220 | * | |
221 | * so, now we're building a concurrency table: | |
222 | * - init buddy vs. | |
223 | * - new PA | |
224 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
225 | * until PA is linked to allocation group to avoid concurrent buddy init | |
226 | * - use inode PA | |
227 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
228 | * given (3) we care that PA-=N operation doesn't interfere with init | |
229 | * - discard inode PA | |
230 | * the simplest way would be to have buddy initialized by the discard | |
231 | * - use locality group PA | |
232 | * again PA-=N must be serialized with init | |
233 | * - discard locality group PA | |
234 | * the simplest way would be to have buddy initialized by the discard | |
235 | * - new PA vs. | |
236 | * - use inode PA | |
237 | * i_data_sem serializes them | |
238 | * - discard inode PA | |
239 | * discard process must wait until PA isn't used by another process | |
240 | * - use locality group PA | |
241 | * some mutex should serialize them | |
242 | * - discard locality group PA | |
243 | * discard process must wait until PA isn't used by another process | |
244 | * - use inode PA | |
245 | * - use inode PA | |
246 | * i_data_sem or another mutex should serializes them | |
247 | * - discard inode PA | |
248 | * discard process must wait until PA isn't used by another process | |
249 | * - use locality group PA | |
250 | * nothing wrong here -- they're different PAs covering different blocks | |
251 | * - discard locality group PA | |
252 | * discard process must wait until PA isn't used by another process | |
253 | * | |
254 | * now we're ready to make few consequences: | |
255 | * - PA is referenced and while it is no discard is possible | |
256 | * - PA is referenced until block isn't marked in on-disk bitmap | |
257 | * - PA changes only after on-disk bitmap | |
258 | * - discard must not compete with init. either init is done before | |
259 | * any discard or they're serialized somehow | |
260 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
261 | * | |
262 | * a special case when we've used PA to emptiness. no need to modify buddy | |
263 | * in this case, but we should care about concurrent init | |
264 | * | |
265 | */ | |
266 | ||
267 | /* | |
268 | * Logic in few words: | |
269 | * | |
270 | * - allocation: | |
271 | * load group | |
272 | * find blocks | |
273 | * mark bits in on-disk bitmap | |
274 | * release group | |
275 | * | |
276 | * - use preallocation: | |
277 | * find proper PA (per-inode or group) | |
278 | * load group | |
279 | * mark bits in on-disk bitmap | |
280 | * release group | |
281 | * release PA | |
282 | * | |
283 | * - free: | |
284 | * load group | |
285 | * mark bits in on-disk bitmap | |
286 | * release group | |
287 | * | |
288 | * - discard preallocations in group: | |
289 | * mark PAs deleted | |
290 | * move them onto local list | |
291 | * load on-disk bitmap | |
292 | * load group | |
293 | * remove PA from object (inode or locality group) | |
294 | * mark free blocks in-core | |
295 | * | |
296 | * - discard inode's preallocations: | |
297 | */ | |
298 | ||
299 | /* | |
300 | * Locking rules | |
301 | * | |
302 | * Locks: | |
303 | * - bitlock on a group (group) | |
304 | * - object (inode/locality) (object) | |
305 | * - per-pa lock (pa) | |
306 | * | |
307 | * Paths: | |
308 | * - new pa | |
309 | * object | |
310 | * group | |
311 | * | |
312 | * - find and use pa: | |
313 | * pa | |
314 | * | |
315 | * - release consumed pa: | |
316 | * pa | |
317 | * group | |
318 | * object | |
319 | * | |
320 | * - generate in-core bitmap: | |
321 | * group | |
322 | * pa | |
323 | * | |
324 | * - discard all for given object (inode, locality group): | |
325 | * object | |
326 | * pa | |
327 | * group | |
328 | * | |
329 | * - discard all for given group: | |
330 | * group | |
331 | * pa | |
332 | * group | |
333 | * object | |
334 | * | |
335 | */ | |
c3a326a6 AK |
336 | static struct kmem_cache *ext4_pspace_cachep; |
337 | static struct kmem_cache *ext4_ac_cachep; | |
338 | static struct kmem_cache *ext4_free_ext_cachep; | |
339 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
340 | ext4_group_t group); | |
7a2fcbf7 AK |
341 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
342 | ext4_group_t group); | |
c3a326a6 AK |
343 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn); |
344 | ||
ffad0a44 AK |
345 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
346 | { | |
c9de560d | 347 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
348 | *bit += ((unsigned long) addr & 7UL) << 3; |
349 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 350 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
351 | *bit += ((unsigned long) addr & 3UL) << 3; |
352 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
353 | #else |
354 | #error "how many bits you are?!" | |
355 | #endif | |
ffad0a44 AK |
356 | return addr; |
357 | } | |
c9de560d AT |
358 | |
359 | static inline int mb_test_bit(int bit, void *addr) | |
360 | { | |
361 | /* | |
362 | * ext4_test_bit on architecture like powerpc | |
363 | * needs unsigned long aligned address | |
364 | */ | |
ffad0a44 | 365 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
366 | return ext4_test_bit(bit, addr); |
367 | } | |
368 | ||
369 | static inline void mb_set_bit(int bit, void *addr) | |
370 | { | |
ffad0a44 | 371 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
372 | ext4_set_bit(bit, addr); |
373 | } | |
374 | ||
c9de560d AT |
375 | static inline void mb_clear_bit(int bit, void *addr) |
376 | { | |
ffad0a44 | 377 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
378 | ext4_clear_bit(bit, addr); |
379 | } | |
380 | ||
ffad0a44 AK |
381 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
382 | { | |
e7dfb246 | 383 | int fix = 0, ret, tmpmax; |
ffad0a44 | 384 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 385 | tmpmax = max + fix; |
ffad0a44 AK |
386 | start += fix; |
387 | ||
e7dfb246 AK |
388 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
389 | if (ret > max) | |
390 | return max; | |
391 | return ret; | |
ffad0a44 AK |
392 | } |
393 | ||
394 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
395 | { | |
e7dfb246 | 396 | int fix = 0, ret, tmpmax; |
ffad0a44 | 397 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 398 | tmpmax = max + fix; |
ffad0a44 AK |
399 | start += fix; |
400 | ||
e7dfb246 AK |
401 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
402 | if (ret > max) | |
403 | return max; | |
404 | return ret; | |
ffad0a44 AK |
405 | } |
406 | ||
c9de560d AT |
407 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
408 | { | |
409 | char *bb; | |
410 | ||
c9de560d AT |
411 | BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); |
412 | BUG_ON(max == NULL); | |
413 | ||
414 | if (order > e4b->bd_blkbits + 1) { | |
415 | *max = 0; | |
416 | return NULL; | |
417 | } | |
418 | ||
419 | /* at order 0 we see each particular block */ | |
420 | *max = 1 << (e4b->bd_blkbits + 3); | |
421 | if (order == 0) | |
422 | return EXT4_MB_BITMAP(e4b); | |
423 | ||
424 | bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; | |
425 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; | |
426 | ||
427 | return bb; | |
428 | } | |
429 | ||
430 | #ifdef DOUBLE_CHECK | |
431 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
432 | int first, int count) | |
433 | { | |
434 | int i; | |
435 | struct super_block *sb = e4b->bd_sb; | |
436 | ||
437 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
438 | return; | |
bc8e6740 | 439 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
440 | for (i = 0; i < count; i++) { |
441 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
442 | ext4_fsblk_t blocknr; | |
443 | blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb); | |
444 | blocknr += first + i; | |
445 | blocknr += | |
446 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
5d1b1b3f AK |
447 | ext4_grp_locked_error(sb, e4b->bd_group, |
448 | __func__, "double-free of inode" | |
a9df9a49 | 449 | " %lu's block %llu(bit %u in group %u)", |
c9de560d AT |
450 | inode ? inode->i_ino : 0, blocknr, |
451 | first + i, e4b->bd_group); | |
452 | } | |
453 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
454 | } | |
455 | } | |
456 | ||
457 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
458 | { | |
459 | int i; | |
460 | ||
461 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
462 | return; | |
bc8e6740 | 463 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
464 | for (i = 0; i < count; i++) { |
465 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
466 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
467 | } | |
468 | } | |
469 | ||
470 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
471 | { | |
472 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { | |
473 | unsigned char *b1, *b2; | |
474 | int i; | |
475 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
476 | b2 = (unsigned char *) bitmap; | |
477 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
478 | if (b1[i] != b2[i]) { | |
a9df9a49 | 479 | printk(KERN_ERR "corruption in group %u " |
4776004f TT |
480 | "at byte %u(%u): %x in copy != %x " |
481 | "on disk/prealloc\n", | |
482 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
483 | BUG(); |
484 | } | |
485 | } | |
486 | } | |
487 | } | |
488 | ||
489 | #else | |
490 | static inline void mb_free_blocks_double(struct inode *inode, | |
491 | struct ext4_buddy *e4b, int first, int count) | |
492 | { | |
493 | return; | |
494 | } | |
495 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
496 | int first, int count) | |
497 | { | |
498 | return; | |
499 | } | |
500 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
501 | { | |
502 | return; | |
503 | } | |
504 | #endif | |
505 | ||
506 | #ifdef AGGRESSIVE_CHECK | |
507 | ||
508 | #define MB_CHECK_ASSERT(assert) \ | |
509 | do { \ | |
510 | if (!(assert)) { \ | |
511 | printk(KERN_EMERG \ | |
512 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
513 | function, file, line, # assert); \ | |
514 | BUG(); \ | |
515 | } \ | |
516 | } while (0) | |
517 | ||
518 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
519 | const char *function, int line) | |
520 | { | |
521 | struct super_block *sb = e4b->bd_sb; | |
522 | int order = e4b->bd_blkbits + 1; | |
523 | int max; | |
524 | int max2; | |
525 | int i; | |
526 | int j; | |
527 | int k; | |
528 | int count; | |
529 | struct ext4_group_info *grp; | |
530 | int fragments = 0; | |
531 | int fstart; | |
532 | struct list_head *cur; | |
533 | void *buddy; | |
534 | void *buddy2; | |
535 | ||
c9de560d AT |
536 | { |
537 | static int mb_check_counter; | |
538 | if (mb_check_counter++ % 100 != 0) | |
539 | return 0; | |
540 | } | |
541 | ||
542 | while (order > 1) { | |
543 | buddy = mb_find_buddy(e4b, order, &max); | |
544 | MB_CHECK_ASSERT(buddy); | |
545 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
546 | MB_CHECK_ASSERT(buddy2); | |
547 | MB_CHECK_ASSERT(buddy != buddy2); | |
548 | MB_CHECK_ASSERT(max * 2 == max2); | |
549 | ||
550 | count = 0; | |
551 | for (i = 0; i < max; i++) { | |
552 | ||
553 | if (mb_test_bit(i, buddy)) { | |
554 | /* only single bit in buddy2 may be 1 */ | |
555 | if (!mb_test_bit(i << 1, buddy2)) { | |
556 | MB_CHECK_ASSERT( | |
557 | mb_test_bit((i<<1)+1, buddy2)); | |
558 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
559 | MB_CHECK_ASSERT( | |
560 | mb_test_bit(i << 1, buddy2)); | |
561 | } | |
562 | continue; | |
563 | } | |
564 | ||
565 | /* both bits in buddy2 must be 0 */ | |
566 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); | |
567 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
568 | ||
569 | for (j = 0; j < (1 << order); j++) { | |
570 | k = (i * (1 << order)) + j; | |
571 | MB_CHECK_ASSERT( | |
572 | !mb_test_bit(k, EXT4_MB_BITMAP(e4b))); | |
573 | } | |
574 | count++; | |
575 | } | |
576 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
577 | order--; | |
578 | } | |
579 | ||
580 | fstart = -1; | |
581 | buddy = mb_find_buddy(e4b, 0, &max); | |
582 | for (i = 0; i < max; i++) { | |
583 | if (!mb_test_bit(i, buddy)) { | |
584 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
585 | if (fstart == -1) { | |
586 | fragments++; | |
587 | fstart = i; | |
588 | } | |
589 | continue; | |
590 | } | |
591 | fstart = -1; | |
592 | /* check used bits only */ | |
593 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
594 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
595 | k = i >> j; | |
596 | MB_CHECK_ASSERT(k < max2); | |
597 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
598 | } | |
599 | } | |
600 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
601 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
602 | ||
603 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
604 | buddy = mb_find_buddy(e4b, 0, &max); | |
605 | list_for_each(cur, &grp->bb_prealloc_list) { | |
606 | ext4_group_t groupnr; | |
607 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
608 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
609 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 610 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 611 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
612 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
613 | } | |
614 | return 0; | |
615 | } | |
616 | #undef MB_CHECK_ASSERT | |
617 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 618 | __FILE__, __func__, __LINE__) |
c9de560d AT |
619 | #else |
620 | #define mb_check_buddy(e4b) | |
621 | #endif | |
622 | ||
623 | /* FIXME!! need more doc */ | |
624 | static void ext4_mb_mark_free_simple(struct super_block *sb, | |
625 | void *buddy, unsigned first, int len, | |
626 | struct ext4_group_info *grp) | |
627 | { | |
628 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
629 | unsigned short min; | |
630 | unsigned short max; | |
631 | unsigned short chunk; | |
632 | unsigned short border; | |
633 | ||
b73fce69 | 634 | BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb)); |
c9de560d AT |
635 | |
636 | border = 2 << sb->s_blocksize_bits; | |
637 | ||
638 | while (len > 0) { | |
639 | /* find how many blocks can be covered since this position */ | |
640 | max = ffs(first | border) - 1; | |
641 | ||
642 | /* find how many blocks of power 2 we need to mark */ | |
643 | min = fls(len) - 1; | |
644 | ||
645 | if (max < min) | |
646 | min = max; | |
647 | chunk = 1 << min; | |
648 | ||
649 | /* mark multiblock chunks only */ | |
650 | grp->bb_counters[min]++; | |
651 | if (min > 0) | |
652 | mb_clear_bit(first >> min, | |
653 | buddy + sbi->s_mb_offsets[min]); | |
654 | ||
655 | len -= chunk; | |
656 | first += chunk; | |
657 | } | |
658 | } | |
659 | ||
660 | static void ext4_mb_generate_buddy(struct super_block *sb, | |
661 | void *buddy, void *bitmap, ext4_group_t group) | |
662 | { | |
663 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
664 | unsigned short max = EXT4_BLOCKS_PER_GROUP(sb); | |
665 | unsigned short i = 0; | |
666 | unsigned short first; | |
667 | unsigned short len; | |
668 | unsigned free = 0; | |
669 | unsigned fragments = 0; | |
670 | unsigned long long period = get_cycles(); | |
671 | ||
672 | /* initialize buddy from bitmap which is aggregation | |
673 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 674 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
675 | grp->bb_first_free = i; |
676 | while (i < max) { | |
677 | fragments++; | |
678 | first = i; | |
ffad0a44 | 679 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
680 | len = i - first; |
681 | free += len; | |
682 | if (len > 1) | |
683 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
684 | else | |
685 | grp->bb_counters[0]++; | |
686 | if (i < max) | |
ffad0a44 | 687 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
688 | } |
689 | grp->bb_fragments = fragments; | |
690 | ||
691 | if (free != grp->bb_free) { | |
5d1b1b3f | 692 | ext4_grp_locked_error(sb, group, __func__, |
a9df9a49 | 693 | "EXT4-fs: group %u: %u blocks in bitmap, %u in gd", |
c9de560d | 694 | group, free, grp->bb_free); |
e56eb659 AK |
695 | /* |
696 | * If we intent to continue, we consider group descritor | |
697 | * corrupt and update bb_free using bitmap value | |
698 | */ | |
c9de560d AT |
699 | grp->bb_free = free; |
700 | } | |
701 | ||
702 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
703 | ||
704 | period = get_cycles() - period; | |
705 | spin_lock(&EXT4_SB(sb)->s_bal_lock); | |
706 | EXT4_SB(sb)->s_mb_buddies_generated++; | |
707 | EXT4_SB(sb)->s_mb_generation_time += period; | |
708 | spin_unlock(&EXT4_SB(sb)->s_bal_lock); | |
709 | } | |
710 | ||
711 | /* The buddy information is attached the buddy cache inode | |
712 | * for convenience. The information regarding each group | |
713 | * is loaded via ext4_mb_load_buddy. The information involve | |
714 | * block bitmap and buddy information. The information are | |
715 | * stored in the inode as | |
716 | * | |
717 | * { page } | |
c3a326a6 | 718 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
719 | * |
720 | * | |
721 | * one block each for bitmap and buddy information. | |
722 | * So for each group we take up 2 blocks. A page can | |
723 | * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. | |
724 | * So it can have information regarding groups_per_page which | |
725 | * is blocks_per_page/2 | |
726 | */ | |
727 | ||
728 | static int ext4_mb_init_cache(struct page *page, char *incore) | |
729 | { | |
8df9675f | 730 | ext4_group_t ngroups; |
c9de560d AT |
731 | int blocksize; |
732 | int blocks_per_page; | |
733 | int groups_per_page; | |
734 | int err = 0; | |
735 | int i; | |
736 | ext4_group_t first_group; | |
737 | int first_block; | |
738 | struct super_block *sb; | |
739 | struct buffer_head *bhs; | |
740 | struct buffer_head **bh; | |
741 | struct inode *inode; | |
742 | char *data; | |
743 | char *bitmap; | |
744 | ||
745 | mb_debug("init page %lu\n", page->index); | |
746 | ||
747 | inode = page->mapping->host; | |
748 | sb = inode->i_sb; | |
8df9675f | 749 | ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
750 | blocksize = 1 << inode->i_blkbits; |
751 | blocks_per_page = PAGE_CACHE_SIZE / blocksize; | |
752 | ||
753 | groups_per_page = blocks_per_page >> 1; | |
754 | if (groups_per_page == 0) | |
755 | groups_per_page = 1; | |
756 | ||
757 | /* allocate buffer_heads to read bitmaps */ | |
758 | if (groups_per_page > 1) { | |
759 | err = -ENOMEM; | |
760 | i = sizeof(struct buffer_head *) * groups_per_page; | |
761 | bh = kzalloc(i, GFP_NOFS); | |
762 | if (bh == NULL) | |
763 | goto out; | |
764 | } else | |
765 | bh = &bhs; | |
766 | ||
767 | first_group = page->index * blocks_per_page / 2; | |
768 | ||
769 | /* read all groups the page covers into the cache */ | |
770 | for (i = 0; i < groups_per_page; i++) { | |
771 | struct ext4_group_desc *desc; | |
772 | ||
8df9675f | 773 | if (first_group + i >= ngroups) |
c9de560d AT |
774 | break; |
775 | ||
776 | err = -EIO; | |
777 | desc = ext4_get_group_desc(sb, first_group + i, NULL); | |
778 | if (desc == NULL) | |
779 | goto out; | |
780 | ||
781 | err = -ENOMEM; | |
782 | bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc)); | |
783 | if (bh[i] == NULL) | |
784 | goto out; | |
785 | ||
2ccb5fb9 | 786 | if (bitmap_uptodate(bh[i])) |
c9de560d AT |
787 | continue; |
788 | ||
c806e68f | 789 | lock_buffer(bh[i]); |
2ccb5fb9 AK |
790 | if (bitmap_uptodate(bh[i])) { |
791 | unlock_buffer(bh[i]); | |
792 | continue; | |
793 | } | |
955ce5f5 | 794 | ext4_lock_group(sb, first_group + i); |
c9de560d AT |
795 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
796 | ext4_init_block_bitmap(sb, bh[i], | |
797 | first_group + i, desc); | |
2ccb5fb9 | 798 | set_bitmap_uptodate(bh[i]); |
c9de560d | 799 | set_buffer_uptodate(bh[i]); |
955ce5f5 | 800 | ext4_unlock_group(sb, first_group + i); |
3300beda | 801 | unlock_buffer(bh[i]); |
c9de560d AT |
802 | continue; |
803 | } | |
955ce5f5 | 804 | ext4_unlock_group(sb, first_group + i); |
2ccb5fb9 AK |
805 | if (buffer_uptodate(bh[i])) { |
806 | /* | |
807 | * if not uninit if bh is uptodate, | |
808 | * bitmap is also uptodate | |
809 | */ | |
810 | set_bitmap_uptodate(bh[i]); | |
811 | unlock_buffer(bh[i]); | |
812 | continue; | |
813 | } | |
c9de560d | 814 | get_bh(bh[i]); |
2ccb5fb9 AK |
815 | /* |
816 | * submit the buffer_head for read. We can | |
817 | * safely mark the bitmap as uptodate now. | |
818 | * We do it here so the bitmap uptodate bit | |
819 | * get set with buffer lock held. | |
820 | */ | |
821 | set_bitmap_uptodate(bh[i]); | |
c9de560d AT |
822 | bh[i]->b_end_io = end_buffer_read_sync; |
823 | submit_bh(READ, bh[i]); | |
a9df9a49 | 824 | mb_debug("read bitmap for group %u\n", first_group + i); |
c9de560d AT |
825 | } |
826 | ||
827 | /* wait for I/O completion */ | |
828 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
829 | wait_on_buffer(bh[i]); | |
830 | ||
831 | err = -EIO; | |
832 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
833 | if (!buffer_uptodate(bh[i])) | |
834 | goto out; | |
835 | ||
31b481dc | 836 | err = 0; |
c9de560d | 837 | first_block = page->index * blocks_per_page; |
29eaf024 AK |
838 | /* init the page */ |
839 | memset(page_address(page), 0xff, PAGE_CACHE_SIZE); | |
c9de560d AT |
840 | for (i = 0; i < blocks_per_page; i++) { |
841 | int group; | |
842 | struct ext4_group_info *grinfo; | |
843 | ||
844 | group = (first_block + i) >> 1; | |
8df9675f | 845 | if (group >= ngroups) |
c9de560d AT |
846 | break; |
847 | ||
848 | /* | |
849 | * data carry information regarding this | |
850 | * particular group in the format specified | |
851 | * above | |
852 | * | |
853 | */ | |
854 | data = page_address(page) + (i * blocksize); | |
855 | bitmap = bh[group - first_group]->b_data; | |
856 | ||
857 | /* | |
858 | * We place the buddy block and bitmap block | |
859 | * close together | |
860 | */ | |
861 | if ((first_block + i) & 1) { | |
862 | /* this is block of buddy */ | |
863 | BUG_ON(incore == NULL); | |
864 | mb_debug("put buddy for group %u in page %lu/%x\n", | |
865 | group, page->index, i * blocksize); | |
c9de560d AT |
866 | grinfo = ext4_get_group_info(sb, group); |
867 | grinfo->bb_fragments = 0; | |
868 | memset(grinfo->bb_counters, 0, | |
869 | sizeof(unsigned short)*(sb->s_blocksize_bits+2)); | |
870 | /* | |
871 | * incore got set to the group block bitmap below | |
872 | */ | |
7a2fcbf7 | 873 | ext4_lock_group(sb, group); |
c9de560d | 874 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 875 | ext4_unlock_group(sb, group); |
c9de560d AT |
876 | incore = NULL; |
877 | } else { | |
878 | /* this is block of bitmap */ | |
879 | BUG_ON(incore != NULL); | |
880 | mb_debug("put bitmap for group %u in page %lu/%x\n", | |
881 | group, page->index, i * blocksize); | |
882 | ||
883 | /* see comments in ext4_mb_put_pa() */ | |
884 | ext4_lock_group(sb, group); | |
885 | memcpy(data, bitmap, blocksize); | |
886 | ||
887 | /* mark all preallocated blks used in in-core bitmap */ | |
888 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 889 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
890 | ext4_unlock_group(sb, group); |
891 | ||
892 | /* set incore so that the buddy information can be | |
893 | * generated using this | |
894 | */ | |
895 | incore = data; | |
896 | } | |
897 | } | |
898 | SetPageUptodate(page); | |
899 | ||
900 | out: | |
901 | if (bh) { | |
902 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
903 | brelse(bh[i]); | |
904 | if (bh != &bhs) | |
905 | kfree(bh); | |
906 | } | |
907 | return err; | |
908 | } | |
909 | ||
4ddfef7b ES |
910 | static noinline_for_stack int |
911 | ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, | |
912 | struct ext4_buddy *e4b) | |
c9de560d | 913 | { |
c9de560d AT |
914 | int blocks_per_page; |
915 | int block; | |
916 | int pnum; | |
917 | int poff; | |
918 | struct page *page; | |
fdf6c7a7 | 919 | int ret; |
920313a7 AK |
920 | struct ext4_group_info *grp; |
921 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
922 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 923 | |
a9df9a49 | 924 | mb_debug("load group %u\n", group); |
c9de560d AT |
925 | |
926 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
920313a7 | 927 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
928 | |
929 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
930 | e4b->bd_info = ext4_get_group_info(sb, group); | |
931 | e4b->bd_sb = sb; | |
932 | e4b->bd_group = group; | |
933 | e4b->bd_buddy_page = NULL; | |
934 | e4b->bd_bitmap_page = NULL; | |
920313a7 AK |
935 | e4b->alloc_semp = &grp->alloc_sem; |
936 | ||
937 | /* Take the read lock on the group alloc | |
938 | * sem. This would make sure a parallel | |
939 | * ext4_mb_init_group happening on other | |
940 | * groups mapped by the page is blocked | |
941 | * till we are done with allocation | |
942 | */ | |
943 | down_read(e4b->alloc_semp); | |
c9de560d AT |
944 | |
945 | /* | |
946 | * the buddy cache inode stores the block bitmap | |
947 | * and buddy information in consecutive blocks. | |
948 | * So for each group we need two blocks. | |
949 | */ | |
950 | block = group * 2; | |
951 | pnum = block / blocks_per_page; | |
952 | poff = block % blocks_per_page; | |
953 | ||
954 | /* we could use find_or_create_page(), but it locks page | |
955 | * what we'd like to avoid in fast path ... */ | |
956 | page = find_get_page(inode->i_mapping, pnum); | |
957 | if (page == NULL || !PageUptodate(page)) { | |
958 | if (page) | |
920313a7 AK |
959 | /* |
960 | * drop the page reference and try | |
961 | * to get the page with lock. If we | |
962 | * are not uptodate that implies | |
963 | * somebody just created the page but | |
964 | * is yet to initialize the same. So | |
965 | * wait for it to initialize. | |
966 | */ | |
c9de560d AT |
967 | page_cache_release(page); |
968 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
969 | if (page) { | |
970 | BUG_ON(page->mapping != inode->i_mapping); | |
971 | if (!PageUptodate(page)) { | |
fdf6c7a7 SF |
972 | ret = ext4_mb_init_cache(page, NULL); |
973 | if (ret) { | |
974 | unlock_page(page); | |
975 | goto err; | |
976 | } | |
c9de560d AT |
977 | mb_cmp_bitmaps(e4b, page_address(page) + |
978 | (poff * sb->s_blocksize)); | |
979 | } | |
980 | unlock_page(page); | |
981 | } | |
982 | } | |
fdf6c7a7 SF |
983 | if (page == NULL || !PageUptodate(page)) { |
984 | ret = -EIO; | |
c9de560d | 985 | goto err; |
fdf6c7a7 | 986 | } |
c9de560d AT |
987 | e4b->bd_bitmap_page = page; |
988 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
989 | mark_page_accessed(page); | |
990 | ||
991 | block++; | |
992 | pnum = block / blocks_per_page; | |
993 | poff = block % blocks_per_page; | |
994 | ||
995 | page = find_get_page(inode->i_mapping, pnum); | |
996 | if (page == NULL || !PageUptodate(page)) { | |
997 | if (page) | |
998 | page_cache_release(page); | |
999 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1000 | if (page) { | |
1001 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 SF |
1002 | if (!PageUptodate(page)) { |
1003 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap); | |
1004 | if (ret) { | |
1005 | unlock_page(page); | |
1006 | goto err; | |
1007 | } | |
1008 | } | |
c9de560d AT |
1009 | unlock_page(page); |
1010 | } | |
1011 | } | |
fdf6c7a7 SF |
1012 | if (page == NULL || !PageUptodate(page)) { |
1013 | ret = -EIO; | |
c9de560d | 1014 | goto err; |
fdf6c7a7 | 1015 | } |
c9de560d AT |
1016 | e4b->bd_buddy_page = page; |
1017 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
1018 | mark_page_accessed(page); | |
1019 | ||
1020 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1021 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1022 | ||
1023 | return 0; | |
1024 | ||
1025 | err: | |
1026 | if (e4b->bd_bitmap_page) | |
1027 | page_cache_release(e4b->bd_bitmap_page); | |
1028 | if (e4b->bd_buddy_page) | |
1029 | page_cache_release(e4b->bd_buddy_page); | |
1030 | e4b->bd_buddy = NULL; | |
1031 | e4b->bd_bitmap = NULL; | |
920313a7 AK |
1032 | |
1033 | /* Done with the buddy cache */ | |
1034 | up_read(e4b->alloc_semp); | |
fdf6c7a7 | 1035 | return ret; |
c9de560d AT |
1036 | } |
1037 | ||
1038 | static void ext4_mb_release_desc(struct ext4_buddy *e4b) | |
1039 | { | |
1040 | if (e4b->bd_bitmap_page) | |
1041 | page_cache_release(e4b->bd_bitmap_page); | |
1042 | if (e4b->bd_buddy_page) | |
1043 | page_cache_release(e4b->bd_buddy_page); | |
920313a7 | 1044 | /* Done with the buddy cache */ |
8556e8f3 AK |
1045 | if (e4b->alloc_semp) |
1046 | up_read(e4b->alloc_semp); | |
c9de560d AT |
1047 | } |
1048 | ||
1049 | ||
1050 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1051 | { | |
1052 | int order = 1; | |
1053 | void *bb; | |
1054 | ||
1055 | BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); | |
1056 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); | |
1057 | ||
1058 | bb = EXT4_MB_BUDDY(e4b); | |
1059 | while (order <= e4b->bd_blkbits + 1) { | |
1060 | block = block >> 1; | |
1061 | if (!mb_test_bit(block, bb)) { | |
1062 | /* this block is part of buddy of order 'order' */ | |
1063 | return order; | |
1064 | } | |
1065 | bb += 1 << (e4b->bd_blkbits - order); | |
1066 | order++; | |
1067 | } | |
1068 | return 0; | |
1069 | } | |
1070 | ||
955ce5f5 | 1071 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1072 | { |
1073 | __u32 *addr; | |
1074 | ||
1075 | len = cur + len; | |
1076 | while (cur < len) { | |
1077 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1078 | /* fast path: clear whole word at once */ | |
1079 | addr = bm + (cur >> 3); | |
1080 | *addr = 0; | |
1081 | cur += 32; | |
1082 | continue; | |
1083 | } | |
955ce5f5 | 1084 | mb_clear_bit(cur, bm); |
c9de560d AT |
1085 | cur++; |
1086 | } | |
1087 | } | |
1088 | ||
955ce5f5 | 1089 | static void mb_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1090 | { |
1091 | __u32 *addr; | |
1092 | ||
1093 | len = cur + len; | |
1094 | while (cur < len) { | |
1095 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1096 | /* fast path: set whole word at once */ | |
1097 | addr = bm + (cur >> 3); | |
1098 | *addr = 0xffffffff; | |
1099 | cur += 32; | |
1100 | continue; | |
1101 | } | |
955ce5f5 | 1102 | mb_set_bit(cur, bm); |
c9de560d AT |
1103 | cur++; |
1104 | } | |
1105 | } | |
1106 | ||
7e5a8cdd | 1107 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
c9de560d AT |
1108 | int first, int count) |
1109 | { | |
1110 | int block = 0; | |
1111 | int max = 0; | |
1112 | int order; | |
1113 | void *buddy; | |
1114 | void *buddy2; | |
1115 | struct super_block *sb = e4b->bd_sb; | |
1116 | ||
1117 | BUG_ON(first + count > (sb->s_blocksize << 3)); | |
bc8e6740 | 1118 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
1119 | mb_check_buddy(e4b); |
1120 | mb_free_blocks_double(inode, e4b, first, count); | |
1121 | ||
1122 | e4b->bd_info->bb_free += count; | |
1123 | if (first < e4b->bd_info->bb_first_free) | |
1124 | e4b->bd_info->bb_first_free = first; | |
1125 | ||
1126 | /* let's maintain fragments counter */ | |
1127 | if (first != 0) | |
1128 | block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b)); | |
1129 | if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) | |
1130 | max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b)); | |
1131 | if (block && max) | |
1132 | e4b->bd_info->bb_fragments--; | |
1133 | else if (!block && !max) | |
1134 | e4b->bd_info->bb_fragments++; | |
1135 | ||
1136 | /* let's maintain buddy itself */ | |
1137 | while (count-- > 0) { | |
1138 | block = first++; | |
1139 | order = 0; | |
1140 | ||
1141 | if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) { | |
1142 | ext4_fsblk_t blocknr; | |
1143 | blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb); | |
1144 | blocknr += block; | |
1145 | blocknr += | |
1146 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
5d1b1b3f AK |
1147 | ext4_grp_locked_error(sb, e4b->bd_group, |
1148 | __func__, "double-free of inode" | |
a9df9a49 | 1149 | " %lu's block %llu(bit %u in group %u)", |
c9de560d AT |
1150 | inode ? inode->i_ino : 0, blocknr, block, |
1151 | e4b->bd_group); | |
1152 | } | |
1153 | mb_clear_bit(block, EXT4_MB_BITMAP(e4b)); | |
1154 | e4b->bd_info->bb_counters[order]++; | |
1155 | ||
1156 | /* start of the buddy */ | |
1157 | buddy = mb_find_buddy(e4b, order, &max); | |
1158 | ||
1159 | do { | |
1160 | block &= ~1UL; | |
1161 | if (mb_test_bit(block, buddy) || | |
1162 | mb_test_bit(block + 1, buddy)) | |
1163 | break; | |
1164 | ||
1165 | /* both the buddies are free, try to coalesce them */ | |
1166 | buddy2 = mb_find_buddy(e4b, order + 1, &max); | |
1167 | ||
1168 | if (!buddy2) | |
1169 | break; | |
1170 | ||
1171 | if (order > 0) { | |
1172 | /* for special purposes, we don't set | |
1173 | * free bits in bitmap */ | |
1174 | mb_set_bit(block, buddy); | |
1175 | mb_set_bit(block + 1, buddy); | |
1176 | } | |
1177 | e4b->bd_info->bb_counters[order]--; | |
1178 | e4b->bd_info->bb_counters[order]--; | |
1179 | ||
1180 | block = block >> 1; | |
1181 | order++; | |
1182 | e4b->bd_info->bb_counters[order]++; | |
1183 | ||
1184 | mb_clear_bit(block, buddy2); | |
1185 | buddy = buddy2; | |
1186 | } while (1); | |
1187 | } | |
1188 | mb_check_buddy(e4b); | |
c9de560d AT |
1189 | } |
1190 | ||
1191 | static int mb_find_extent(struct ext4_buddy *e4b, int order, int block, | |
1192 | int needed, struct ext4_free_extent *ex) | |
1193 | { | |
1194 | int next = block; | |
1195 | int max; | |
1196 | int ord; | |
1197 | void *buddy; | |
1198 | ||
bc8e6740 | 1199 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1200 | BUG_ON(ex == NULL); |
1201 | ||
1202 | buddy = mb_find_buddy(e4b, order, &max); | |
1203 | BUG_ON(buddy == NULL); | |
1204 | BUG_ON(block >= max); | |
1205 | if (mb_test_bit(block, buddy)) { | |
1206 | ex->fe_len = 0; | |
1207 | ex->fe_start = 0; | |
1208 | ex->fe_group = 0; | |
1209 | return 0; | |
1210 | } | |
1211 | ||
1212 | /* FIXME dorp order completely ? */ | |
1213 | if (likely(order == 0)) { | |
1214 | /* find actual order */ | |
1215 | order = mb_find_order_for_block(e4b, block); | |
1216 | block = block >> order; | |
1217 | } | |
1218 | ||
1219 | ex->fe_len = 1 << order; | |
1220 | ex->fe_start = block << order; | |
1221 | ex->fe_group = e4b->bd_group; | |
1222 | ||
1223 | /* calc difference from given start */ | |
1224 | next = next - ex->fe_start; | |
1225 | ex->fe_len -= next; | |
1226 | ex->fe_start += next; | |
1227 | ||
1228 | while (needed > ex->fe_len && | |
1229 | (buddy = mb_find_buddy(e4b, order, &max))) { | |
1230 | ||
1231 | if (block + 1 >= max) | |
1232 | break; | |
1233 | ||
1234 | next = (block + 1) * (1 << order); | |
1235 | if (mb_test_bit(next, EXT4_MB_BITMAP(e4b))) | |
1236 | break; | |
1237 | ||
1238 | ord = mb_find_order_for_block(e4b, next); | |
1239 | ||
1240 | order = ord; | |
1241 | block = next >> order; | |
1242 | ex->fe_len += 1 << order; | |
1243 | } | |
1244 | ||
1245 | BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); | |
1246 | return ex->fe_len; | |
1247 | } | |
1248 | ||
1249 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1250 | { | |
1251 | int ord; | |
1252 | int mlen = 0; | |
1253 | int max = 0; | |
1254 | int cur; | |
1255 | int start = ex->fe_start; | |
1256 | int len = ex->fe_len; | |
1257 | unsigned ret = 0; | |
1258 | int len0 = len; | |
1259 | void *buddy; | |
1260 | ||
1261 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1262 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1263 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1264 | mb_check_buddy(e4b); |
1265 | mb_mark_used_double(e4b, start, len); | |
1266 | ||
1267 | e4b->bd_info->bb_free -= len; | |
1268 | if (e4b->bd_info->bb_first_free == start) | |
1269 | e4b->bd_info->bb_first_free += len; | |
1270 | ||
1271 | /* let's maintain fragments counter */ | |
1272 | if (start != 0) | |
1273 | mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b)); | |
1274 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) | |
1275 | max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b)); | |
1276 | if (mlen && max) | |
1277 | e4b->bd_info->bb_fragments++; | |
1278 | else if (!mlen && !max) | |
1279 | e4b->bd_info->bb_fragments--; | |
1280 | ||
1281 | /* let's maintain buddy itself */ | |
1282 | while (len) { | |
1283 | ord = mb_find_order_for_block(e4b, start); | |
1284 | ||
1285 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1286 | /* the whole chunk may be allocated at once! */ | |
1287 | mlen = 1 << ord; | |
1288 | buddy = mb_find_buddy(e4b, ord, &max); | |
1289 | BUG_ON((start >> ord) >= max); | |
1290 | mb_set_bit(start >> ord, buddy); | |
1291 | e4b->bd_info->bb_counters[ord]--; | |
1292 | start += mlen; | |
1293 | len -= mlen; | |
1294 | BUG_ON(len < 0); | |
1295 | continue; | |
1296 | } | |
1297 | ||
1298 | /* store for history */ | |
1299 | if (ret == 0) | |
1300 | ret = len | (ord << 16); | |
1301 | ||
1302 | /* we have to split large buddy */ | |
1303 | BUG_ON(ord <= 0); | |
1304 | buddy = mb_find_buddy(e4b, ord, &max); | |
1305 | mb_set_bit(start >> ord, buddy); | |
1306 | e4b->bd_info->bb_counters[ord]--; | |
1307 | ||
1308 | ord--; | |
1309 | cur = (start >> ord) & ~1U; | |
1310 | buddy = mb_find_buddy(e4b, ord, &max); | |
1311 | mb_clear_bit(cur, buddy); | |
1312 | mb_clear_bit(cur + 1, buddy); | |
1313 | e4b->bd_info->bb_counters[ord]++; | |
1314 | e4b->bd_info->bb_counters[ord]++; | |
1315 | } | |
1316 | ||
955ce5f5 | 1317 | mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0); |
c9de560d AT |
1318 | mb_check_buddy(e4b); |
1319 | ||
1320 | return ret; | |
1321 | } | |
1322 | ||
1323 | /* | |
1324 | * Must be called under group lock! | |
1325 | */ | |
1326 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1327 | struct ext4_buddy *e4b) | |
1328 | { | |
1329 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1330 | int ret; | |
1331 | ||
1332 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1333 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1334 | ||
1335 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1336 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1337 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1338 | ||
1339 | /* preallocation can change ac_b_ex, thus we store actually | |
1340 | * allocated blocks for history */ | |
1341 | ac->ac_f_ex = ac->ac_b_ex; | |
1342 | ||
1343 | ac->ac_status = AC_STATUS_FOUND; | |
1344 | ac->ac_tail = ret & 0xffff; | |
1345 | ac->ac_buddy = ret >> 16; | |
1346 | ||
c3a326a6 AK |
1347 | /* |
1348 | * take the page reference. We want the page to be pinned | |
1349 | * so that we don't get a ext4_mb_init_cache_call for this | |
1350 | * group until we update the bitmap. That would mean we | |
1351 | * double allocate blocks. The reference is dropped | |
1352 | * in ext4_mb_release_context | |
1353 | */ | |
c9de560d AT |
1354 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1355 | get_page(ac->ac_bitmap_page); | |
1356 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1357 | get_page(ac->ac_buddy_page); | |
8556e8f3 AK |
1358 | /* on allocation we use ac to track the held semaphore */ |
1359 | ac->alloc_semp = e4b->alloc_semp; | |
1360 | e4b->alloc_semp = NULL; | |
c9de560d AT |
1361 | /* store last allocated for subsequent stream allocation */ |
1362 | if ((ac->ac_flags & EXT4_MB_HINT_DATA)) { | |
1363 | spin_lock(&sbi->s_md_lock); | |
1364 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1365 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1366 | spin_unlock(&sbi->s_md_lock); | |
1367 | } | |
1368 | } | |
1369 | ||
1370 | /* | |
1371 | * regular allocator, for general purposes allocation | |
1372 | */ | |
1373 | ||
1374 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1375 | struct ext4_buddy *e4b, | |
1376 | int finish_group) | |
1377 | { | |
1378 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1379 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1380 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1381 | struct ext4_free_extent ex; | |
1382 | int max; | |
1383 | ||
032115fc AK |
1384 | if (ac->ac_status == AC_STATUS_FOUND) |
1385 | return; | |
c9de560d AT |
1386 | /* |
1387 | * We don't want to scan for a whole year | |
1388 | */ | |
1389 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1390 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1391 | ac->ac_status = AC_STATUS_BREAK; | |
1392 | return; | |
1393 | } | |
1394 | ||
1395 | /* | |
1396 | * Haven't found good chunk so far, let's continue | |
1397 | */ | |
1398 | if (bex->fe_len < gex->fe_len) | |
1399 | return; | |
1400 | ||
1401 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1402 | && bex->fe_group == e4b->bd_group) { | |
1403 | /* recheck chunk's availability - we don't know | |
1404 | * when it was found (within this lock-unlock | |
1405 | * period or not) */ | |
1406 | max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex); | |
1407 | if (max >= gex->fe_len) { | |
1408 | ext4_mb_use_best_found(ac, e4b); | |
1409 | return; | |
1410 | } | |
1411 | } | |
1412 | } | |
1413 | ||
1414 | /* | |
1415 | * The routine checks whether found extent is good enough. If it is, | |
1416 | * then the extent gets marked used and flag is set to the context | |
1417 | * to stop scanning. Otherwise, the extent is compared with the | |
1418 | * previous found extent and if new one is better, then it's stored | |
1419 | * in the context. Later, the best found extent will be used, if | |
1420 | * mballoc can't find good enough extent. | |
1421 | * | |
1422 | * FIXME: real allocation policy is to be designed yet! | |
1423 | */ | |
1424 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1425 | struct ext4_free_extent *ex, | |
1426 | struct ext4_buddy *e4b) | |
1427 | { | |
1428 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1429 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1430 | ||
1431 | BUG_ON(ex->fe_len <= 0); | |
8d03c7a0 | 1432 | BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
1433 | BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
1434 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); | |
1435 | ||
1436 | ac->ac_found++; | |
1437 | ||
1438 | /* | |
1439 | * The special case - take what you catch first | |
1440 | */ | |
1441 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1442 | *bex = *ex; | |
1443 | ext4_mb_use_best_found(ac, e4b); | |
1444 | return; | |
1445 | } | |
1446 | ||
1447 | /* | |
1448 | * Let's check whether the chuck is good enough | |
1449 | */ | |
1450 | if (ex->fe_len == gex->fe_len) { | |
1451 | *bex = *ex; | |
1452 | ext4_mb_use_best_found(ac, e4b); | |
1453 | return; | |
1454 | } | |
1455 | ||
1456 | /* | |
1457 | * If this is first found extent, just store it in the context | |
1458 | */ | |
1459 | if (bex->fe_len == 0) { | |
1460 | *bex = *ex; | |
1461 | return; | |
1462 | } | |
1463 | ||
1464 | /* | |
1465 | * If new found extent is better, store it in the context | |
1466 | */ | |
1467 | if (bex->fe_len < gex->fe_len) { | |
1468 | /* if the request isn't satisfied, any found extent | |
1469 | * larger than previous best one is better */ | |
1470 | if (ex->fe_len > bex->fe_len) | |
1471 | *bex = *ex; | |
1472 | } else if (ex->fe_len > gex->fe_len) { | |
1473 | /* if the request is satisfied, then we try to find | |
1474 | * an extent that still satisfy the request, but is | |
1475 | * smaller than previous one */ | |
1476 | if (ex->fe_len < bex->fe_len) | |
1477 | *bex = *ex; | |
1478 | } | |
1479 | ||
1480 | ext4_mb_check_limits(ac, e4b, 0); | |
1481 | } | |
1482 | ||
1483 | static int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
1484 | struct ext4_buddy *e4b) | |
1485 | { | |
1486 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1487 | ext4_group_t group = ex.fe_group; | |
1488 | int max; | |
1489 | int err; | |
1490 | ||
1491 | BUG_ON(ex.fe_len <= 0); | |
1492 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1493 | if (err) | |
1494 | return err; | |
1495 | ||
1496 | ext4_lock_group(ac->ac_sb, group); | |
1497 | max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex); | |
1498 | ||
1499 | if (max > 0) { | |
1500 | ac->ac_b_ex = ex; | |
1501 | ext4_mb_use_best_found(ac, e4b); | |
1502 | } | |
1503 | ||
1504 | ext4_unlock_group(ac->ac_sb, group); | |
1505 | ext4_mb_release_desc(e4b); | |
1506 | ||
1507 | return 0; | |
1508 | } | |
1509 | ||
1510 | static int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
1511 | struct ext4_buddy *e4b) | |
1512 | { | |
1513 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1514 | int max; | |
1515 | int err; | |
1516 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1517 | struct ext4_super_block *es = sbi->s_es; | |
1518 | struct ext4_free_extent ex; | |
1519 | ||
1520 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1521 | return 0; | |
1522 | ||
1523 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1524 | if (err) | |
1525 | return err; | |
1526 | ||
1527 | ext4_lock_group(ac->ac_sb, group); | |
1528 | max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start, | |
1529 | ac->ac_g_ex.fe_len, &ex); | |
1530 | ||
1531 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1532 | ext4_fsblk_t start; | |
1533 | ||
1534 | start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) + | |
1535 | ex.fe_start + le32_to_cpu(es->s_first_data_block); | |
1536 | /* use do_div to get remainder (would be 64-bit modulo) */ | |
1537 | if (do_div(start, sbi->s_stripe) == 0) { | |
1538 | ac->ac_found++; | |
1539 | ac->ac_b_ex = ex; | |
1540 | ext4_mb_use_best_found(ac, e4b); | |
1541 | } | |
1542 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1543 | BUG_ON(ex.fe_len <= 0); | |
1544 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1545 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1546 | ac->ac_found++; | |
1547 | ac->ac_b_ex = ex; | |
1548 | ext4_mb_use_best_found(ac, e4b); | |
1549 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1550 | /* Sometimes, caller may want to merge even small | |
1551 | * number of blocks to an existing extent */ | |
1552 | BUG_ON(ex.fe_len <= 0); | |
1553 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1554 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1555 | ac->ac_found++; | |
1556 | ac->ac_b_ex = ex; | |
1557 | ext4_mb_use_best_found(ac, e4b); | |
1558 | } | |
1559 | ext4_unlock_group(ac->ac_sb, group); | |
1560 | ext4_mb_release_desc(e4b); | |
1561 | ||
1562 | return 0; | |
1563 | } | |
1564 | ||
1565 | /* | |
1566 | * The routine scans buddy structures (not bitmap!) from given order | |
1567 | * to max order and tries to find big enough chunk to satisfy the req | |
1568 | */ | |
1569 | static void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
1570 | struct ext4_buddy *e4b) | |
1571 | { | |
1572 | struct super_block *sb = ac->ac_sb; | |
1573 | struct ext4_group_info *grp = e4b->bd_info; | |
1574 | void *buddy; | |
1575 | int i; | |
1576 | int k; | |
1577 | int max; | |
1578 | ||
1579 | BUG_ON(ac->ac_2order <= 0); | |
1580 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1581 | if (grp->bb_counters[i] == 0) | |
1582 | continue; | |
1583 | ||
1584 | buddy = mb_find_buddy(e4b, i, &max); | |
1585 | BUG_ON(buddy == NULL); | |
1586 | ||
ffad0a44 | 1587 | k = mb_find_next_zero_bit(buddy, max, 0); |
c9de560d AT |
1588 | BUG_ON(k >= max); |
1589 | ||
1590 | ac->ac_found++; | |
1591 | ||
1592 | ac->ac_b_ex.fe_len = 1 << i; | |
1593 | ac->ac_b_ex.fe_start = k << i; | |
1594 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1595 | ||
1596 | ext4_mb_use_best_found(ac, e4b); | |
1597 | ||
1598 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1599 | ||
1600 | if (EXT4_SB(sb)->s_mb_stats) | |
1601 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1602 | ||
1603 | break; | |
1604 | } | |
1605 | } | |
1606 | ||
1607 | /* | |
1608 | * The routine scans the group and measures all found extents. | |
1609 | * In order to optimize scanning, caller must pass number of | |
1610 | * free blocks in the group, so the routine can know upper limit. | |
1611 | */ | |
1612 | static void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
1613 | struct ext4_buddy *e4b) | |
1614 | { | |
1615 | struct super_block *sb = ac->ac_sb; | |
1616 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1617 | struct ext4_free_extent ex; | |
1618 | int i; | |
1619 | int free; | |
1620 | ||
1621 | free = e4b->bd_info->bb_free; | |
1622 | BUG_ON(free <= 0); | |
1623 | ||
1624 | i = e4b->bd_info->bb_first_free; | |
1625 | ||
1626 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1627 | i = mb_find_next_zero_bit(bitmap, |
c9de560d AT |
1628 | EXT4_BLOCKS_PER_GROUP(sb), i); |
1629 | if (i >= EXT4_BLOCKS_PER_GROUP(sb)) { | |
26346ff6 | 1630 | /* |
e56eb659 | 1631 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1632 | * free blocks even though group info says we |
1633 | * we have free blocks | |
1634 | */ | |
5d1b1b3f AK |
1635 | ext4_grp_locked_error(sb, e4b->bd_group, |
1636 | __func__, "%d free blocks as per " | |
fde4d95a | 1637 | "group info. But bitmap says 0", |
26346ff6 | 1638 | free); |
c9de560d AT |
1639 | break; |
1640 | } | |
1641 | ||
1642 | mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex); | |
1643 | BUG_ON(ex.fe_len <= 0); | |
26346ff6 | 1644 | if (free < ex.fe_len) { |
5d1b1b3f AK |
1645 | ext4_grp_locked_error(sb, e4b->bd_group, |
1646 | __func__, "%d free blocks as per " | |
fde4d95a | 1647 | "group info. But got %d blocks", |
26346ff6 | 1648 | free, ex.fe_len); |
e56eb659 AK |
1649 | /* |
1650 | * The number of free blocks differs. This mostly | |
1651 | * indicate that the bitmap is corrupt. So exit | |
1652 | * without claiming the space. | |
1653 | */ | |
1654 | break; | |
26346ff6 | 1655 | } |
c9de560d AT |
1656 | |
1657 | ext4_mb_measure_extent(ac, &ex, e4b); | |
1658 | ||
1659 | i += ex.fe_len; | |
1660 | free -= ex.fe_len; | |
1661 | } | |
1662 | ||
1663 | ext4_mb_check_limits(ac, e4b, 1); | |
1664 | } | |
1665 | ||
1666 | /* | |
1667 | * This is a special case for storages like raid5 | |
1668 | * we try to find stripe-aligned chunks for stripe-size requests | |
1669 | * XXX should do so at least for multiples of stripe size as well | |
1670 | */ | |
1671 | static void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
1672 | struct ext4_buddy *e4b) | |
1673 | { | |
1674 | struct super_block *sb = ac->ac_sb; | |
1675 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1676 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1677 | struct ext4_free_extent ex; | |
1678 | ext4_fsblk_t first_group_block; | |
1679 | ext4_fsblk_t a; | |
1680 | ext4_grpblk_t i; | |
1681 | int max; | |
1682 | ||
1683 | BUG_ON(sbi->s_stripe == 0); | |
1684 | ||
1685 | /* find first stripe-aligned block in group */ | |
1686 | first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb) | |
1687 | + le32_to_cpu(sbi->s_es->s_first_data_block); | |
1688 | a = first_group_block + sbi->s_stripe - 1; | |
1689 | do_div(a, sbi->s_stripe); | |
1690 | i = (a * sbi->s_stripe) - first_group_block; | |
1691 | ||
1692 | while (i < EXT4_BLOCKS_PER_GROUP(sb)) { | |
1693 | if (!mb_test_bit(i, bitmap)) { | |
1694 | max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex); | |
1695 | if (max >= sbi->s_stripe) { | |
1696 | ac->ac_found++; | |
1697 | ac->ac_b_ex = ex; | |
1698 | ext4_mb_use_best_found(ac, e4b); | |
1699 | break; | |
1700 | } | |
1701 | } | |
1702 | i += sbi->s_stripe; | |
1703 | } | |
1704 | } | |
1705 | ||
1706 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, | |
1707 | ext4_group_t group, int cr) | |
1708 | { | |
1709 | unsigned free, fragments; | |
1710 | unsigned i, bits; | |
a4912123 | 1711 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
1712 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
1713 | ||
1714 | BUG_ON(cr < 0 || cr >= 4); | |
1715 | BUG_ON(EXT4_MB_GRP_NEED_INIT(grp)); | |
1716 | ||
1717 | free = grp->bb_free; | |
1718 | fragments = grp->bb_fragments; | |
1719 | if (free == 0) | |
1720 | return 0; | |
1721 | if (fragments == 0) | |
1722 | return 0; | |
1723 | ||
1724 | switch (cr) { | |
1725 | case 0: | |
1726 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 1727 | |
a4912123 TT |
1728 | /* Avoid using the first bg of a flexgroup for data files */ |
1729 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
1730 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
1731 | ((group % flex_size) == 0)) | |
1732 | return 0; | |
1733 | ||
c9de560d AT |
1734 | bits = ac->ac_sb->s_blocksize_bits + 1; |
1735 | for (i = ac->ac_2order; i <= bits; i++) | |
1736 | if (grp->bb_counters[i] > 0) | |
1737 | return 1; | |
1738 | break; | |
1739 | case 1: | |
1740 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
1741 | return 1; | |
1742 | break; | |
1743 | case 2: | |
1744 | if (free >= ac->ac_g_ex.fe_len) | |
1745 | return 1; | |
1746 | break; | |
1747 | case 3: | |
1748 | return 1; | |
1749 | default: | |
1750 | BUG(); | |
1751 | } | |
1752 | ||
1753 | return 0; | |
1754 | } | |
1755 | ||
920313a7 AK |
1756 | /* |
1757 | * lock the group_info alloc_sem of all the groups | |
1758 | * belonging to the same buddy cache page. This | |
1759 | * make sure other parallel operation on the buddy | |
1760 | * cache doesn't happen whild holding the buddy cache | |
1761 | * lock | |
1762 | */ | |
1763 | int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group) | |
1764 | { | |
1765 | int i; | |
1766 | int block, pnum; | |
1767 | int blocks_per_page; | |
1768 | int groups_per_page; | |
8df9675f | 1769 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
920313a7 AK |
1770 | ext4_group_t first_group; |
1771 | struct ext4_group_info *grp; | |
1772 | ||
1773 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1774 | /* | |
1775 | * the buddy cache inode stores the block bitmap | |
1776 | * and buddy information in consecutive blocks. | |
1777 | * So for each group we need two blocks. | |
1778 | */ | |
1779 | block = group * 2; | |
1780 | pnum = block / blocks_per_page; | |
1781 | first_group = pnum * blocks_per_page / 2; | |
1782 | ||
1783 | groups_per_page = blocks_per_page >> 1; | |
1784 | if (groups_per_page == 0) | |
1785 | groups_per_page = 1; | |
1786 | /* read all groups the page covers into the cache */ | |
1787 | for (i = 0; i < groups_per_page; i++) { | |
1788 | ||
8df9675f | 1789 | if ((first_group + i) >= ngroups) |
920313a7 AK |
1790 | break; |
1791 | grp = ext4_get_group_info(sb, first_group + i); | |
1792 | /* take all groups write allocation | |
1793 | * semaphore. This make sure there is | |
1794 | * no block allocation going on in any | |
1795 | * of that groups | |
1796 | */ | |
b7be019e | 1797 | down_write_nested(&grp->alloc_sem, i); |
920313a7 AK |
1798 | } |
1799 | return i; | |
1800 | } | |
1801 | ||
1802 | void ext4_mb_put_buddy_cache_lock(struct super_block *sb, | |
1803 | ext4_group_t group, int locked_group) | |
1804 | { | |
1805 | int i; | |
1806 | int block, pnum; | |
1807 | int blocks_per_page; | |
1808 | ext4_group_t first_group; | |
1809 | struct ext4_group_info *grp; | |
1810 | ||
1811 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1812 | /* | |
1813 | * the buddy cache inode stores the block bitmap | |
1814 | * and buddy information in consecutive blocks. | |
1815 | * So for each group we need two blocks. | |
1816 | */ | |
1817 | block = group * 2; | |
1818 | pnum = block / blocks_per_page; | |
1819 | first_group = pnum * blocks_per_page / 2; | |
1820 | /* release locks on all the groups */ | |
1821 | for (i = 0; i < locked_group; i++) { | |
1822 | ||
1823 | grp = ext4_get_group_info(sb, first_group + i); | |
1824 | /* take all groups write allocation | |
1825 | * semaphore. This make sure there is | |
1826 | * no block allocation going on in any | |
1827 | * of that groups | |
1828 | */ | |
1829 | up_write(&grp->alloc_sem); | |
1830 | } | |
1831 | ||
1832 | } | |
1833 | ||
1834 | static int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) | |
1835 | { | |
1836 | ||
1837 | int ret; | |
1838 | void *bitmap; | |
1839 | int blocks_per_page; | |
1840 | int block, pnum, poff; | |
1841 | int num_grp_locked = 0; | |
1842 | struct ext4_group_info *this_grp; | |
1843 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1844 | struct inode *inode = sbi->s_buddy_cache; | |
1845 | struct page *page = NULL, *bitmap_page = NULL; | |
1846 | ||
1847 | mb_debug("init group %lu\n", group); | |
1848 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1849 | this_grp = ext4_get_group_info(sb, group); | |
1850 | /* | |
1851 | * This ensures we don't add group | |
1852 | * to this buddy cache via resize | |
1853 | */ | |
1854 | num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group); | |
1855 | if (!EXT4_MB_GRP_NEED_INIT(this_grp)) { | |
1856 | /* | |
1857 | * somebody initialized the group | |
1858 | * return without doing anything | |
1859 | */ | |
1860 | ret = 0; | |
1861 | goto err; | |
1862 | } | |
1863 | /* | |
1864 | * the buddy cache inode stores the block bitmap | |
1865 | * and buddy information in consecutive blocks. | |
1866 | * So for each group we need two blocks. | |
1867 | */ | |
1868 | block = group * 2; | |
1869 | pnum = block / blocks_per_page; | |
1870 | poff = block % blocks_per_page; | |
1871 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1872 | if (page) { | |
1873 | BUG_ON(page->mapping != inode->i_mapping); | |
1874 | ret = ext4_mb_init_cache(page, NULL); | |
1875 | if (ret) { | |
1876 | unlock_page(page); | |
1877 | goto err; | |
1878 | } | |
1879 | unlock_page(page); | |
1880 | } | |
1881 | if (page == NULL || !PageUptodate(page)) { | |
1882 | ret = -EIO; | |
1883 | goto err; | |
1884 | } | |
1885 | mark_page_accessed(page); | |
1886 | bitmap_page = page; | |
1887 | bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1888 | ||
1889 | /* init buddy cache */ | |
1890 | block++; | |
1891 | pnum = block / blocks_per_page; | |
1892 | poff = block % blocks_per_page; | |
1893 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1894 | if (page == bitmap_page) { | |
1895 | /* | |
1896 | * If both the bitmap and buddy are in | |
1897 | * the same page we don't need to force | |
1898 | * init the buddy | |
1899 | */ | |
1900 | unlock_page(page); | |
1901 | } else if (page) { | |
1902 | BUG_ON(page->mapping != inode->i_mapping); | |
1903 | ret = ext4_mb_init_cache(page, bitmap); | |
1904 | if (ret) { | |
1905 | unlock_page(page); | |
1906 | goto err; | |
1907 | } | |
1908 | unlock_page(page); | |
1909 | } | |
1910 | if (page == NULL || !PageUptodate(page)) { | |
1911 | ret = -EIO; | |
1912 | goto err; | |
1913 | } | |
1914 | mark_page_accessed(page); | |
1915 | err: | |
1916 | ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked); | |
1917 | if (bitmap_page) | |
1918 | page_cache_release(bitmap_page); | |
1919 | if (page) | |
1920 | page_cache_release(page); | |
1921 | return ret; | |
1922 | } | |
1923 | ||
4ddfef7b ES |
1924 | static noinline_for_stack int |
1925 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 1926 | { |
8df9675f | 1927 | ext4_group_t ngroups, group, i; |
c9de560d AT |
1928 | int cr; |
1929 | int err = 0; | |
1930 | int bsbits; | |
1931 | struct ext4_sb_info *sbi; | |
1932 | struct super_block *sb; | |
1933 | struct ext4_buddy e4b; | |
1934 | loff_t size, isize; | |
1935 | ||
1936 | sb = ac->ac_sb; | |
1937 | sbi = EXT4_SB(sb); | |
8df9675f | 1938 | ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
1939 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
1940 | ||
1941 | /* first, try the goal */ | |
1942 | err = ext4_mb_find_by_goal(ac, &e4b); | |
1943 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
1944 | goto out; | |
1945 | ||
1946 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
1947 | goto out; | |
1948 | ||
1949 | /* | |
1950 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
1951 | * if ac2_order is set we also set criteria to 0 so that we | |
1952 | * try exact allocation using buddy. | |
1953 | */ | |
1954 | i = fls(ac->ac_g_ex.fe_len); | |
1955 | ac->ac_2order = 0; | |
1956 | /* | |
1957 | * We search using buddy data only if the order of the request | |
1958 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 1959 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
c9de560d AT |
1960 | */ |
1961 | if (i >= sbi->s_mb_order2_reqs) { | |
1962 | /* | |
1963 | * This should tell if fe_len is exactly power of 2 | |
1964 | */ | |
1965 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1966 | ac->ac_2order = i - 1; | |
1967 | } | |
1968 | ||
1969 | bsbits = ac->ac_sb->s_blocksize_bits; | |
1970 | /* if stream allocation is enabled, use global goal */ | |
1971 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; | |
1972 | isize = i_size_read(ac->ac_inode) >> bsbits; | |
1973 | if (size < isize) | |
1974 | size = isize; | |
1975 | ||
1976 | if (size < sbi->s_mb_stream_request && | |
1977 | (ac->ac_flags & EXT4_MB_HINT_DATA)) { | |
1978 | /* TBD: may be hot point */ | |
1979 | spin_lock(&sbi->s_md_lock); | |
1980 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
1981 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
1982 | spin_unlock(&sbi->s_md_lock); | |
1983 | } | |
c9de560d AT |
1984 | /* Let's just scan groups to find more-less suitable blocks */ |
1985 | cr = ac->ac_2order ? 0 : 1; | |
1986 | /* | |
1987 | * cr == 0 try to get exact allocation, | |
1988 | * cr == 3 try to get anything | |
1989 | */ | |
1990 | repeat: | |
1991 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
1992 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
1993 | /* |
1994 | * searching for the right group start | |
1995 | * from the goal value specified | |
1996 | */ | |
1997 | group = ac->ac_g_ex.fe_group; | |
1998 | ||
8df9675f | 1999 | for (i = 0; i < ngroups; group++, i++) { |
c9de560d AT |
2000 | struct ext4_group_info *grp; |
2001 | struct ext4_group_desc *desc; | |
2002 | ||
8df9675f | 2003 | if (group == ngroups) |
c9de560d AT |
2004 | group = 0; |
2005 | ||
2006 | /* quick check to skip empty groups */ | |
920313a7 | 2007 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
2008 | if (grp->bb_free == 0) |
2009 | continue; | |
2010 | ||
2011 | /* | |
2012 | * if the group is already init we check whether it is | |
2013 | * a good group and if not we don't load the buddy | |
2014 | */ | |
2015 | if (EXT4_MB_GRP_NEED_INIT(grp)) { | |
2016 | /* | |
2017 | * we need full data about the group | |
2018 | * to make a good selection | |
2019 | */ | |
920313a7 | 2020 | err = ext4_mb_init_group(sb, group); |
c9de560d AT |
2021 | if (err) |
2022 | goto out; | |
c9de560d AT |
2023 | } |
2024 | ||
2025 | /* | |
2026 | * If the particular group doesn't satisfy our | |
2027 | * criteria we continue with the next group | |
2028 | */ | |
2029 | if (!ext4_mb_good_group(ac, group, cr)) | |
2030 | continue; | |
2031 | ||
2032 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2033 | if (err) | |
2034 | goto out; | |
2035 | ||
2036 | ext4_lock_group(sb, group); | |
2037 | if (!ext4_mb_good_group(ac, group, cr)) { | |
2038 | /* someone did allocation from this group */ | |
2039 | ext4_unlock_group(sb, group); | |
2040 | ext4_mb_release_desc(&e4b); | |
2041 | continue; | |
2042 | } | |
2043 | ||
2044 | ac->ac_groups_scanned++; | |
2045 | desc = ext4_get_group_desc(sb, group, NULL); | |
75507efb | 2046 | if (cr == 0) |
c9de560d AT |
2047 | ext4_mb_simple_scan_group(ac, &e4b); |
2048 | else if (cr == 1 && | |
2049 | ac->ac_g_ex.fe_len == sbi->s_stripe) | |
2050 | ext4_mb_scan_aligned(ac, &e4b); | |
2051 | else | |
2052 | ext4_mb_complex_scan_group(ac, &e4b); | |
2053 | ||
2054 | ext4_unlock_group(sb, group); | |
2055 | ext4_mb_release_desc(&e4b); | |
2056 | ||
2057 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2058 | break; | |
2059 | } | |
2060 | } | |
2061 | ||
2062 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2063 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2064 | /* | |
2065 | * We've been searching too long. Let's try to allocate | |
2066 | * the best chunk we've found so far | |
2067 | */ | |
2068 | ||
2069 | ext4_mb_try_best_found(ac, &e4b); | |
2070 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2071 | /* | |
2072 | * Someone more lucky has already allocated it. | |
2073 | * The only thing we can do is just take first | |
2074 | * found block(s) | |
2075 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2076 | */ | |
2077 | ac->ac_b_ex.fe_group = 0; | |
2078 | ac->ac_b_ex.fe_start = 0; | |
2079 | ac->ac_b_ex.fe_len = 0; | |
2080 | ac->ac_status = AC_STATUS_CONTINUE; | |
2081 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2082 | cr = 3; | |
2083 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2084 | goto repeat; | |
2085 | } | |
2086 | } | |
2087 | out: | |
2088 | return err; | |
2089 | } | |
2090 | ||
2091 | #ifdef EXT4_MB_HISTORY | |
2092 | struct ext4_mb_proc_session { | |
2093 | struct ext4_mb_history *history; | |
2094 | struct super_block *sb; | |
2095 | int start; | |
2096 | int max; | |
2097 | }; | |
2098 | ||
2099 | static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s, | |
2100 | struct ext4_mb_history *hs, | |
2101 | int first) | |
2102 | { | |
2103 | if (hs == s->history + s->max) | |
2104 | hs = s->history; | |
2105 | if (!first && hs == s->history + s->start) | |
2106 | return NULL; | |
2107 | while (hs->orig.fe_len == 0) { | |
2108 | hs++; | |
2109 | if (hs == s->history + s->max) | |
2110 | hs = s->history; | |
2111 | if (hs == s->history + s->start) | |
2112 | return NULL; | |
2113 | } | |
2114 | return hs; | |
2115 | } | |
2116 | ||
2117 | static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos) | |
2118 | { | |
2119 | struct ext4_mb_proc_session *s = seq->private; | |
2120 | struct ext4_mb_history *hs; | |
2121 | int l = *pos; | |
2122 | ||
2123 | if (l == 0) | |
2124 | return SEQ_START_TOKEN; | |
2125 | hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1); | |
2126 | if (!hs) | |
2127 | return NULL; | |
2128 | while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL); | |
2129 | return hs; | |
2130 | } | |
2131 | ||
2132 | static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v, | |
2133 | loff_t *pos) | |
2134 | { | |
2135 | struct ext4_mb_proc_session *s = seq->private; | |
2136 | struct ext4_mb_history *hs = v; | |
2137 | ||
2138 | ++*pos; | |
2139 | if (v == SEQ_START_TOKEN) | |
2140 | return ext4_mb_history_skip_empty(s, s->history + s->start, 1); | |
2141 | else | |
2142 | return ext4_mb_history_skip_empty(s, ++hs, 0); | |
2143 | } | |
2144 | ||
2145 | static int ext4_mb_seq_history_show(struct seq_file *seq, void *v) | |
2146 | { | |
2147 | char buf[25], buf2[25], buf3[25], *fmt; | |
2148 | struct ext4_mb_history *hs = v; | |
2149 | ||
2150 | if (v == SEQ_START_TOKEN) { | |
2151 | seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s " | |
2152 | "%-5s %-2s %-5s %-5s %-5s %-6s\n", | |
2153 | "pid", "inode", "original", "goal", "result", "found", | |
2154 | "grps", "cr", "flags", "merge", "tail", "broken"); | |
2155 | return 0; | |
2156 | } | |
2157 | ||
2158 | if (hs->op == EXT4_MB_HISTORY_ALLOC) { | |
2159 | fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u " | |
2160 | "%-5u %-5s %-5u %-6u\n"; | |
a9df9a49 | 2161 | sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group, |
c9de560d AT |
2162 | hs->result.fe_start, hs->result.fe_len, |
2163 | hs->result.fe_logical); | |
a9df9a49 | 2164 | sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group, |
c9de560d AT |
2165 | hs->orig.fe_start, hs->orig.fe_len, |
2166 | hs->orig.fe_logical); | |
a9df9a49 | 2167 | sprintf(buf3, "%u/%d/%u@%u", hs->goal.fe_group, |
c9de560d AT |
2168 | hs->goal.fe_start, hs->goal.fe_len, |
2169 | hs->goal.fe_logical); | |
2170 | seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2, | |
2171 | hs->found, hs->groups, hs->cr, hs->flags, | |
2172 | hs->merged ? "M" : "", hs->tail, | |
2173 | hs->buddy ? 1 << hs->buddy : 0); | |
2174 | } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) { | |
2175 | fmt = "%-5u %-8u %-23s %-23s %-23s\n"; | |
a9df9a49 | 2176 | sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group, |
c9de560d AT |
2177 | hs->result.fe_start, hs->result.fe_len, |
2178 | hs->result.fe_logical); | |
a9df9a49 | 2179 | sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group, |
c9de560d AT |
2180 | hs->orig.fe_start, hs->orig.fe_len, |
2181 | hs->orig.fe_logical); | |
2182 | seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2); | |
2183 | } else if (hs->op == EXT4_MB_HISTORY_DISCARD) { | |
a9df9a49 | 2184 | sprintf(buf2, "%u/%d/%u", hs->result.fe_group, |
c9de560d AT |
2185 | hs->result.fe_start, hs->result.fe_len); |
2186 | seq_printf(seq, "%-5u %-8u %-23s discard\n", | |
2187 | hs->pid, hs->ino, buf2); | |
2188 | } else if (hs->op == EXT4_MB_HISTORY_FREE) { | |
a9df9a49 | 2189 | sprintf(buf2, "%u/%d/%u", hs->result.fe_group, |
c9de560d AT |
2190 | hs->result.fe_start, hs->result.fe_len); |
2191 | seq_printf(seq, "%-5u %-8u %-23s free\n", | |
2192 | hs->pid, hs->ino, buf2); | |
2193 | } | |
2194 | return 0; | |
2195 | } | |
2196 | ||
2197 | static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v) | |
2198 | { | |
2199 | } | |
2200 | ||
2201 | static struct seq_operations ext4_mb_seq_history_ops = { | |
2202 | .start = ext4_mb_seq_history_start, | |
2203 | .next = ext4_mb_seq_history_next, | |
2204 | .stop = ext4_mb_seq_history_stop, | |
2205 | .show = ext4_mb_seq_history_show, | |
2206 | }; | |
2207 | ||
2208 | static int ext4_mb_seq_history_open(struct inode *inode, struct file *file) | |
2209 | { | |
2210 | struct super_block *sb = PDE(inode)->data; | |
2211 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2212 | struct ext4_mb_proc_session *s; | |
2213 | int rc; | |
2214 | int size; | |
2215 | ||
74767c5a SF |
2216 | if (unlikely(sbi->s_mb_history == NULL)) |
2217 | return -ENOMEM; | |
c9de560d AT |
2218 | s = kmalloc(sizeof(*s), GFP_KERNEL); |
2219 | if (s == NULL) | |
2220 | return -ENOMEM; | |
2221 | s->sb = sb; | |
2222 | size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max; | |
2223 | s->history = kmalloc(size, GFP_KERNEL); | |
2224 | if (s->history == NULL) { | |
2225 | kfree(s); | |
2226 | return -ENOMEM; | |
2227 | } | |
2228 | ||
2229 | spin_lock(&sbi->s_mb_history_lock); | |
2230 | memcpy(s->history, sbi->s_mb_history, size); | |
2231 | s->max = sbi->s_mb_history_max; | |
2232 | s->start = sbi->s_mb_history_cur % s->max; | |
2233 | spin_unlock(&sbi->s_mb_history_lock); | |
2234 | ||
2235 | rc = seq_open(file, &ext4_mb_seq_history_ops); | |
2236 | if (rc == 0) { | |
2237 | struct seq_file *m = (struct seq_file *)file->private_data; | |
2238 | m->private = s; | |
2239 | } else { | |
2240 | kfree(s->history); | |
2241 | kfree(s); | |
2242 | } | |
2243 | return rc; | |
2244 | ||
2245 | } | |
2246 | ||
2247 | static int ext4_mb_seq_history_release(struct inode *inode, struct file *file) | |
2248 | { | |
2249 | struct seq_file *seq = (struct seq_file *)file->private_data; | |
2250 | struct ext4_mb_proc_session *s = seq->private; | |
2251 | kfree(s->history); | |
2252 | kfree(s); | |
2253 | return seq_release(inode, file); | |
2254 | } | |
2255 | ||
2256 | static ssize_t ext4_mb_seq_history_write(struct file *file, | |
2257 | const char __user *buffer, | |
2258 | size_t count, loff_t *ppos) | |
2259 | { | |
2260 | struct seq_file *seq = (struct seq_file *)file->private_data; | |
2261 | struct ext4_mb_proc_session *s = seq->private; | |
2262 | struct super_block *sb = s->sb; | |
2263 | char str[32]; | |
2264 | int value; | |
2265 | ||
2266 | if (count >= sizeof(str)) { | |
2267 | printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n", | |
2268 | "mb_history", (int)sizeof(str)); | |
2269 | return -EOVERFLOW; | |
2270 | } | |
2271 | ||
2272 | if (copy_from_user(str, buffer, count)) | |
2273 | return -EFAULT; | |
2274 | ||
2275 | value = simple_strtol(str, NULL, 0); | |
2276 | if (value < 0) | |
2277 | return -ERANGE; | |
2278 | EXT4_SB(sb)->s_mb_history_filter = value; | |
2279 | ||
2280 | return count; | |
2281 | } | |
2282 | ||
2283 | static struct file_operations ext4_mb_seq_history_fops = { | |
2284 | .owner = THIS_MODULE, | |
2285 | .open = ext4_mb_seq_history_open, | |
2286 | .read = seq_read, | |
2287 | .write = ext4_mb_seq_history_write, | |
2288 | .llseek = seq_lseek, | |
2289 | .release = ext4_mb_seq_history_release, | |
2290 | }; | |
2291 | ||
2292 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) | |
2293 | { | |
2294 | struct super_block *sb = seq->private; | |
c9de560d AT |
2295 | ext4_group_t group; |
2296 | ||
8df9675f | 2297 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2298 | return NULL; |
c9de560d | 2299 | group = *pos + 1; |
a9df9a49 | 2300 | return (void *) ((unsigned long) group); |
c9de560d AT |
2301 | } |
2302 | ||
2303 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2304 | { | |
2305 | struct super_block *sb = seq->private; | |
c9de560d AT |
2306 | ext4_group_t group; |
2307 | ||
2308 | ++*pos; | |
8df9675f | 2309 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2310 | return NULL; |
2311 | group = *pos + 1; | |
a9df9a49 | 2312 | return (void *) ((unsigned long) group); |
c9de560d AT |
2313 | } |
2314 | ||
2315 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2316 | { | |
2317 | struct super_block *sb = seq->private; | |
a9df9a49 | 2318 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d AT |
2319 | int i; |
2320 | int err; | |
2321 | struct ext4_buddy e4b; | |
2322 | struct sg { | |
2323 | struct ext4_group_info info; | |
2324 | unsigned short counters[16]; | |
2325 | } sg; | |
2326 | ||
2327 | group--; | |
2328 | if (group == 0) | |
2329 | seq_printf(seq, "#%-5s: %-5s %-5s %-5s " | |
2330 | "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " | |
2331 | "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", | |
2332 | "group", "free", "frags", "first", | |
2333 | "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", | |
2334 | "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); | |
2335 | ||
2336 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2337 | sizeof(struct ext4_group_info); | |
2338 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2339 | if (err) { | |
a9df9a49 | 2340 | seq_printf(seq, "#%-5u: I/O error\n", group); |
c9de560d AT |
2341 | return 0; |
2342 | } | |
2343 | ext4_lock_group(sb, group); | |
2344 | memcpy(&sg, ext4_get_group_info(sb, group), i); | |
2345 | ext4_unlock_group(sb, group); | |
2346 | ext4_mb_release_desc(&e4b); | |
2347 | ||
a9df9a49 | 2348 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2349 | sg.info.bb_fragments, sg.info.bb_first_free); |
2350 | for (i = 0; i <= 13; i++) | |
2351 | seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? | |
2352 | sg.info.bb_counters[i] : 0); | |
2353 | seq_printf(seq, " ]\n"); | |
2354 | ||
2355 | return 0; | |
2356 | } | |
2357 | ||
2358 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2359 | { | |
2360 | } | |
2361 | ||
2362 | static struct seq_operations ext4_mb_seq_groups_ops = { | |
2363 | .start = ext4_mb_seq_groups_start, | |
2364 | .next = ext4_mb_seq_groups_next, | |
2365 | .stop = ext4_mb_seq_groups_stop, | |
2366 | .show = ext4_mb_seq_groups_show, | |
2367 | }; | |
2368 | ||
2369 | static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) | |
2370 | { | |
2371 | struct super_block *sb = PDE(inode)->data; | |
2372 | int rc; | |
2373 | ||
2374 | rc = seq_open(file, &ext4_mb_seq_groups_ops); | |
2375 | if (rc == 0) { | |
2376 | struct seq_file *m = (struct seq_file *)file->private_data; | |
2377 | m->private = sb; | |
2378 | } | |
2379 | return rc; | |
2380 | ||
2381 | } | |
2382 | ||
2383 | static struct file_operations ext4_mb_seq_groups_fops = { | |
2384 | .owner = THIS_MODULE, | |
2385 | .open = ext4_mb_seq_groups_open, | |
2386 | .read = seq_read, | |
2387 | .llseek = seq_lseek, | |
2388 | .release = seq_release, | |
2389 | }; | |
2390 | ||
2391 | static void ext4_mb_history_release(struct super_block *sb) | |
2392 | { | |
2393 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2394 | ||
9f6200bb TT |
2395 | if (sbi->s_proc != NULL) { |
2396 | remove_proc_entry("mb_groups", sbi->s_proc); | |
f4033903 CW |
2397 | if (sbi->s_mb_history_max) |
2398 | remove_proc_entry("mb_history", sbi->s_proc); | |
9f6200bb | 2399 | } |
c9de560d AT |
2400 | kfree(sbi->s_mb_history); |
2401 | } | |
2402 | ||
2403 | static void ext4_mb_history_init(struct super_block *sb) | |
2404 | { | |
2405 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2406 | int i; | |
2407 | ||
9f6200bb | 2408 | if (sbi->s_proc != NULL) { |
f4033903 CW |
2409 | if (sbi->s_mb_history_max) |
2410 | proc_create_data("mb_history", S_IRUGO, sbi->s_proc, | |
2411 | &ext4_mb_seq_history_fops, sb); | |
9f6200bb | 2412 | proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, |
46fe74f2 | 2413 | &ext4_mb_seq_groups_fops, sb); |
c9de560d AT |
2414 | } |
2415 | ||
c9de560d AT |
2416 | sbi->s_mb_history_cur = 0; |
2417 | spin_lock_init(&sbi->s_mb_history_lock); | |
2418 | i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history); | |
f4033903 | 2419 | sbi->s_mb_history = i ? kzalloc(i, GFP_KERNEL) : NULL; |
c9de560d AT |
2420 | /* if we can't allocate history, then we simple won't use it */ |
2421 | } | |
2422 | ||
4ddfef7b ES |
2423 | static noinline_for_stack void |
2424 | ext4_mb_store_history(struct ext4_allocation_context *ac) | |
c9de560d AT |
2425 | { |
2426 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2427 | struct ext4_mb_history h; | |
2428 | ||
f4033903 | 2429 | if (sbi->s_mb_history == NULL) |
c9de560d AT |
2430 | return; |
2431 | ||
2432 | if (!(ac->ac_op & sbi->s_mb_history_filter)) | |
2433 | return; | |
2434 | ||
2435 | h.op = ac->ac_op; | |
2436 | h.pid = current->pid; | |
2437 | h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0; | |
2438 | h.orig = ac->ac_o_ex; | |
2439 | h.result = ac->ac_b_ex; | |
2440 | h.flags = ac->ac_flags; | |
2441 | h.found = ac->ac_found; | |
2442 | h.groups = ac->ac_groups_scanned; | |
2443 | h.cr = ac->ac_criteria; | |
2444 | h.tail = ac->ac_tail; | |
2445 | h.buddy = ac->ac_buddy; | |
2446 | h.merged = 0; | |
2447 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) { | |
2448 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
2449 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
2450 | h.merged = 1; | |
2451 | h.goal = ac->ac_g_ex; | |
2452 | h.result = ac->ac_f_ex; | |
2453 | } | |
2454 | ||
2455 | spin_lock(&sbi->s_mb_history_lock); | |
2456 | memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h)); | |
2457 | if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max) | |
2458 | sbi->s_mb_history_cur = 0; | |
2459 | spin_unlock(&sbi->s_mb_history_lock); | |
2460 | } | |
2461 | ||
2462 | #else | |
2463 | #define ext4_mb_history_release(sb) | |
2464 | #define ext4_mb_history_init(sb) | |
2465 | #endif | |
2466 | ||
5f21b0e6 FB |
2467 | |
2468 | /* Create and initialize ext4_group_info data for the given group. */ | |
920313a7 | 2469 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2470 | struct ext4_group_desc *desc) |
2471 | { | |
2472 | int i, len; | |
2473 | int metalen = 0; | |
2474 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2475 | struct ext4_group_info **meta_group_info; | |
2476 | ||
2477 | /* | |
2478 | * First check if this group is the first of a reserved block. | |
2479 | * If it's true, we have to allocate a new table of pointers | |
2480 | * to ext4_group_info structures | |
2481 | */ | |
2482 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2483 | metalen = sizeof(*meta_group_info) << | |
2484 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2485 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2486 | if (meta_group_info == NULL) { | |
2487 | printk(KERN_ERR "EXT4-fs: can't allocate mem for a " | |
2488 | "buddy group\n"); | |
2489 | goto exit_meta_group_info; | |
2490 | } | |
2491 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = | |
2492 | meta_group_info; | |
2493 | } | |
2494 | ||
2495 | /* | |
2496 | * calculate needed size. if change bb_counters size, | |
2497 | * don't forget about ext4_mb_generate_buddy() | |
2498 | */ | |
2499 | len = offsetof(typeof(**meta_group_info), | |
2500 | bb_counters[sb->s_blocksize_bits + 2]); | |
2501 | ||
2502 | meta_group_info = | |
2503 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; | |
2504 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); | |
2505 | ||
2506 | meta_group_info[i] = kzalloc(len, GFP_KERNEL); | |
2507 | if (meta_group_info[i] == NULL) { | |
2508 | printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n"); | |
2509 | goto exit_group_info; | |
2510 | } | |
2511 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2512 | &(meta_group_info[i]->bb_state)); | |
2513 | ||
2514 | /* | |
2515 | * initialize bb_free to be able to skip | |
2516 | * empty groups without initialization | |
2517 | */ | |
2518 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2519 | meta_group_info[i]->bb_free = | |
2520 | ext4_free_blocks_after_init(sb, group, desc); | |
2521 | } else { | |
2522 | meta_group_info[i]->bb_free = | |
560671a0 | 2523 | ext4_free_blks_count(sb, desc); |
5f21b0e6 FB |
2524 | } |
2525 | ||
2526 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2527 | init_rwsem(&meta_group_info[i]->alloc_sem); |
c894058d | 2528 | meta_group_info[i]->bb_free_root.rb_node = NULL;; |
5f21b0e6 FB |
2529 | |
2530 | #ifdef DOUBLE_CHECK | |
2531 | { | |
2532 | struct buffer_head *bh; | |
2533 | meta_group_info[i]->bb_bitmap = | |
2534 | kmalloc(sb->s_blocksize, GFP_KERNEL); | |
2535 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); | |
2536 | bh = ext4_read_block_bitmap(sb, group); | |
2537 | BUG_ON(bh == NULL); | |
2538 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, | |
2539 | sb->s_blocksize); | |
2540 | put_bh(bh); | |
2541 | } | |
2542 | #endif | |
2543 | ||
2544 | return 0; | |
2545 | ||
2546 | exit_group_info: | |
2547 | /* If a meta_group_info table has been allocated, release it now */ | |
2548 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) | |
2549 | kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); | |
2550 | exit_meta_group_info: | |
2551 | return -ENOMEM; | |
2552 | } /* ext4_mb_add_groupinfo */ | |
2553 | ||
5f21b0e6 FB |
2554 | /* |
2555 | * Update an existing group. | |
2556 | * This function is used for online resize | |
2557 | */ | |
2558 | void ext4_mb_update_group_info(struct ext4_group_info *grp, ext4_grpblk_t add) | |
2559 | { | |
2560 | grp->bb_free += add; | |
2561 | } | |
2562 | ||
c9de560d AT |
2563 | static int ext4_mb_init_backend(struct super_block *sb) |
2564 | { | |
8df9675f | 2565 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2566 | ext4_group_t i; |
5f21b0e6 | 2567 | int metalen; |
c9de560d | 2568 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
5f21b0e6 FB |
2569 | struct ext4_super_block *es = sbi->s_es; |
2570 | int num_meta_group_infos; | |
2571 | int num_meta_group_infos_max; | |
2572 | int array_size; | |
c9de560d | 2573 | struct ext4_group_info **meta_group_info; |
5f21b0e6 FB |
2574 | struct ext4_group_desc *desc; |
2575 | ||
2576 | /* This is the number of blocks used by GDT */ | |
8df9675f | 2577 | num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) - |
5f21b0e6 FB |
2578 | 1) >> EXT4_DESC_PER_BLOCK_BITS(sb); |
2579 | ||
2580 | /* | |
2581 | * This is the total number of blocks used by GDT including | |
2582 | * the number of reserved blocks for GDT. | |
2583 | * The s_group_info array is allocated with this value | |
2584 | * to allow a clean online resize without a complex | |
2585 | * manipulation of pointer. | |
2586 | * The drawback is the unused memory when no resize | |
2587 | * occurs but it's very low in terms of pages | |
2588 | * (see comments below) | |
2589 | * Need to handle this properly when META_BG resizing is allowed | |
2590 | */ | |
2591 | num_meta_group_infos_max = num_meta_group_infos + | |
2592 | le16_to_cpu(es->s_reserved_gdt_blocks); | |
c9de560d | 2593 | |
5f21b0e6 FB |
2594 | /* |
2595 | * array_size is the size of s_group_info array. We round it | |
2596 | * to the next power of two because this approximation is done | |
2597 | * internally by kmalloc so we can have some more memory | |
2598 | * for free here (e.g. may be used for META_BG resize). | |
2599 | */ | |
2600 | array_size = 1; | |
2601 | while (array_size < sizeof(*sbi->s_group_info) * | |
2602 | num_meta_group_infos_max) | |
2603 | array_size = array_size << 1; | |
c9de560d AT |
2604 | /* An 8TB filesystem with 64-bit pointers requires a 4096 byte |
2605 | * kmalloc. A 128kb malloc should suffice for a 256TB filesystem. | |
2606 | * So a two level scheme suffices for now. */ | |
5f21b0e6 | 2607 | sbi->s_group_info = kmalloc(array_size, GFP_KERNEL); |
c9de560d AT |
2608 | if (sbi->s_group_info == NULL) { |
2609 | printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n"); | |
2610 | return -ENOMEM; | |
2611 | } | |
2612 | sbi->s_buddy_cache = new_inode(sb); | |
2613 | if (sbi->s_buddy_cache == NULL) { | |
2614 | printk(KERN_ERR "EXT4-fs: can't get new inode\n"); | |
2615 | goto err_freesgi; | |
2616 | } | |
2617 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; | |
2618 | ||
2619 | metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb); | |
2620 | for (i = 0; i < num_meta_group_infos; i++) { | |
2621 | if ((i + 1) == num_meta_group_infos) | |
2622 | metalen = sizeof(*meta_group_info) * | |
8df9675f | 2623 | (ngroups - |
c9de560d AT |
2624 | (i << EXT4_DESC_PER_BLOCK_BITS(sb))); |
2625 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2626 | if (meta_group_info == NULL) { | |
2627 | printk(KERN_ERR "EXT4-fs: can't allocate mem for a " | |
2628 | "buddy group\n"); | |
2629 | goto err_freemeta; | |
2630 | } | |
2631 | sbi->s_group_info[i] = meta_group_info; | |
2632 | } | |
2633 | ||
8df9675f | 2634 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2635 | desc = ext4_get_group_desc(sb, i, NULL); |
2636 | if (desc == NULL) { | |
2637 | printk(KERN_ERR | |
a9df9a49 | 2638 | "EXT4-fs: can't read descriptor %u\n", i); |
c9de560d AT |
2639 | goto err_freebuddy; |
2640 | } | |
5f21b0e6 FB |
2641 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2642 | goto err_freebuddy; | |
c9de560d AT |
2643 | } |
2644 | ||
2645 | return 0; | |
2646 | ||
2647 | err_freebuddy: | |
f1fa3342 | 2648 | while (i-- > 0) |
c9de560d | 2649 | kfree(ext4_get_group_info(sb, i)); |
c9de560d AT |
2650 | i = num_meta_group_infos; |
2651 | err_freemeta: | |
f1fa3342 | 2652 | while (i-- > 0) |
c9de560d AT |
2653 | kfree(sbi->s_group_info[i]); |
2654 | iput(sbi->s_buddy_cache); | |
2655 | err_freesgi: | |
2656 | kfree(sbi->s_group_info); | |
2657 | return -ENOMEM; | |
2658 | } | |
2659 | ||
2660 | int ext4_mb_init(struct super_block *sb, int needs_recovery) | |
2661 | { | |
2662 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2663 | unsigned i, j; |
c9de560d AT |
2664 | unsigned offset; |
2665 | unsigned max; | |
74767c5a | 2666 | int ret; |
c9de560d | 2667 | |
c9de560d AT |
2668 | i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short); |
2669 | ||
2670 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2671 | if (sbi->s_mb_offsets == NULL) { | |
c9de560d AT |
2672 | return -ENOMEM; |
2673 | } | |
ff7ef329 YG |
2674 | |
2675 | i = (sb->s_blocksize_bits + 2) * sizeof(unsigned int); | |
c9de560d AT |
2676 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2677 | if (sbi->s_mb_maxs == NULL) { | |
a7b19448 | 2678 | kfree(sbi->s_mb_offsets); |
c9de560d AT |
2679 | return -ENOMEM; |
2680 | } | |
2681 | ||
2682 | /* order 0 is regular bitmap */ | |
2683 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2684 | sbi->s_mb_offsets[0] = 0; | |
2685 | ||
2686 | i = 1; | |
2687 | offset = 0; | |
2688 | max = sb->s_blocksize << 2; | |
2689 | do { | |
2690 | sbi->s_mb_offsets[i] = offset; | |
2691 | sbi->s_mb_maxs[i] = max; | |
2692 | offset += 1 << (sb->s_blocksize_bits - i); | |
2693 | max = max >> 1; | |
2694 | i++; | |
2695 | } while (i <= sb->s_blocksize_bits + 1); | |
2696 | ||
2697 | /* init file for buddy data */ | |
74767c5a SF |
2698 | ret = ext4_mb_init_backend(sb); |
2699 | if (ret != 0) { | |
c9de560d AT |
2700 | kfree(sbi->s_mb_offsets); |
2701 | kfree(sbi->s_mb_maxs); | |
74767c5a | 2702 | return ret; |
c9de560d AT |
2703 | } |
2704 | ||
2705 | spin_lock_init(&sbi->s_md_lock); | |
c9de560d AT |
2706 | spin_lock_init(&sbi->s_bal_lock); |
2707 | ||
2708 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2709 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2710 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2711 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2712 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
2713 | sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT; | |
2714 | sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC; | |
2715 | ||
730c213c | 2716 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2717 | if (sbi->s_locality_groups == NULL) { |
c9de560d AT |
2718 | kfree(sbi->s_mb_offsets); |
2719 | kfree(sbi->s_mb_maxs); | |
2720 | return -ENOMEM; | |
2721 | } | |
730c213c | 2722 | for_each_possible_cpu(i) { |
c9de560d | 2723 | struct ext4_locality_group *lg; |
730c213c | 2724 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2725 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2726 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2727 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2728 | spin_lock_init(&lg->lg_prealloc_lock); |
2729 | } | |
2730 | ||
c9de560d AT |
2731 | ext4_mb_history_init(sb); |
2732 | ||
0390131b FM |
2733 | if (sbi->s_journal) |
2734 | sbi->s_journal->j_commit_callback = release_blocks_on_commit; | |
3e624fc7 | 2735 | |
4776004f | 2736 | printk(KERN_INFO "EXT4-fs: mballoc enabled\n"); |
c9de560d AT |
2737 | return 0; |
2738 | } | |
2739 | ||
955ce5f5 | 2740 | /* need to called with the ext4 group lock held */ |
c9de560d AT |
2741 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
2742 | { | |
2743 | struct ext4_prealloc_space *pa; | |
2744 | struct list_head *cur, *tmp; | |
2745 | int count = 0; | |
2746 | ||
2747 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2748 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2749 | list_del(&pa->pa_group_list); | |
2750 | count++; | |
688f05a0 | 2751 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2752 | } |
2753 | if (count) | |
2754 | mb_debug("mballoc: %u PAs left\n", count); | |
2755 | ||
2756 | } | |
2757 | ||
2758 | int ext4_mb_release(struct super_block *sb) | |
2759 | { | |
8df9675f | 2760 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2761 | ext4_group_t i; |
2762 | int num_meta_group_infos; | |
2763 | struct ext4_group_info *grinfo; | |
2764 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2765 | ||
c9de560d | 2766 | if (sbi->s_group_info) { |
8df9675f | 2767 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2768 | grinfo = ext4_get_group_info(sb, i); |
2769 | #ifdef DOUBLE_CHECK | |
2770 | kfree(grinfo->bb_bitmap); | |
2771 | #endif | |
2772 | ext4_lock_group(sb, i); | |
2773 | ext4_mb_cleanup_pa(grinfo); | |
2774 | ext4_unlock_group(sb, i); | |
2775 | kfree(grinfo); | |
2776 | } | |
8df9675f | 2777 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2778 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2779 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2780 | for (i = 0; i < num_meta_group_infos; i++) | |
2781 | kfree(sbi->s_group_info[i]); | |
2782 | kfree(sbi->s_group_info); | |
2783 | } | |
2784 | kfree(sbi->s_mb_offsets); | |
2785 | kfree(sbi->s_mb_maxs); | |
2786 | if (sbi->s_buddy_cache) | |
2787 | iput(sbi->s_buddy_cache); | |
2788 | if (sbi->s_mb_stats) { | |
2789 | printk(KERN_INFO | |
2790 | "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n", | |
2791 | atomic_read(&sbi->s_bal_allocated), | |
2792 | atomic_read(&sbi->s_bal_reqs), | |
2793 | atomic_read(&sbi->s_bal_success)); | |
2794 | printk(KERN_INFO | |
2795 | "EXT4-fs: mballoc: %u extents scanned, %u goal hits, " | |
2796 | "%u 2^N hits, %u breaks, %u lost\n", | |
2797 | atomic_read(&sbi->s_bal_ex_scanned), | |
2798 | atomic_read(&sbi->s_bal_goals), | |
2799 | atomic_read(&sbi->s_bal_2orders), | |
2800 | atomic_read(&sbi->s_bal_breaks), | |
2801 | atomic_read(&sbi->s_mb_lost_chunks)); | |
2802 | printk(KERN_INFO | |
2803 | "EXT4-fs: mballoc: %lu generated and it took %Lu\n", | |
2804 | sbi->s_mb_buddies_generated++, | |
2805 | sbi->s_mb_generation_time); | |
2806 | printk(KERN_INFO | |
2807 | "EXT4-fs: mballoc: %u preallocated, %u discarded\n", | |
2808 | atomic_read(&sbi->s_mb_preallocated), | |
2809 | atomic_read(&sbi->s_mb_discarded)); | |
2810 | } | |
2811 | ||
730c213c | 2812 | free_percpu(sbi->s_locality_groups); |
c9de560d | 2813 | ext4_mb_history_release(sb); |
c9de560d AT |
2814 | |
2815 | return 0; | |
2816 | } | |
2817 | ||
3e624fc7 TT |
2818 | /* |
2819 | * This function is called by the jbd2 layer once the commit has finished, | |
2820 | * so we know we can free the blocks that were released with that commit. | |
2821 | */ | |
2822 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn) | |
c9de560d | 2823 | { |
3e624fc7 | 2824 | struct super_block *sb = journal->j_private; |
c9de560d | 2825 | struct ext4_buddy e4b; |
c894058d | 2826 | struct ext4_group_info *db; |
c894058d AK |
2827 | int err, count = 0, count2 = 0; |
2828 | struct ext4_free_data *entry; | |
8a0aba73 | 2829 | ext4_fsblk_t discard_block; |
3e624fc7 | 2830 | struct list_head *l, *ltmp; |
c9de560d | 2831 | |
3e624fc7 TT |
2832 | list_for_each_safe(l, ltmp, &txn->t_private_list) { |
2833 | entry = list_entry(l, struct ext4_free_data, list); | |
c9de560d | 2834 | |
a9df9a49 | 2835 | mb_debug("gonna free %u blocks in group %u (0x%p):", |
3e624fc7 | 2836 | entry->count, entry->group, entry); |
c9de560d | 2837 | |
c894058d | 2838 | err = ext4_mb_load_buddy(sb, entry->group, &e4b); |
c9de560d AT |
2839 | /* we expect to find existing buddy because it's pinned */ |
2840 | BUG_ON(err != 0); | |
2841 | ||
c894058d | 2842 | db = e4b.bd_info; |
c9de560d | 2843 | /* there are blocks to put in buddy to make them really free */ |
c894058d | 2844 | count += entry->count; |
c9de560d | 2845 | count2++; |
c894058d AK |
2846 | ext4_lock_group(sb, entry->group); |
2847 | /* Take it out of per group rb tree */ | |
2848 | rb_erase(&entry->node, &(db->bb_free_root)); | |
2849 | mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count); | |
2850 | ||
2851 | if (!db->bb_free_root.rb_node) { | |
2852 | /* No more items in the per group rb tree | |
2853 | * balance refcounts from ext4_mb_free_metadata() | |
2854 | */ | |
2855 | page_cache_release(e4b.bd_buddy_page); | |
2856 | page_cache_release(e4b.bd_bitmap_page); | |
c9de560d | 2857 | } |
c894058d | 2858 | ext4_unlock_group(sb, entry->group); |
8a0aba73 TT |
2859 | discard_block = (ext4_fsblk_t) entry->group * EXT4_BLOCKS_PER_GROUP(sb) |
2860 | + entry->start_blk | |
2861 | + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
9bffad1e TT |
2862 | trace_ext4_discard_blocks(sb, (unsigned long long)discard_block, |
2863 | entry->count); | |
8a0aba73 | 2864 | sb_issue_discard(sb, discard_block, entry->count); |
c9de560d | 2865 | |
c894058d | 2866 | kmem_cache_free(ext4_free_ext_cachep, entry); |
c9de560d | 2867 | ext4_mb_release_desc(&e4b); |
3e624fc7 | 2868 | } |
c9de560d AT |
2869 | |
2870 | mb_debug("freed %u blocks in %u structures\n", count, count2); | |
2871 | } | |
2872 | ||
c9de560d AT |
2873 | int __init init_ext4_mballoc(void) |
2874 | { | |
2875 | ext4_pspace_cachep = | |
2876 | kmem_cache_create("ext4_prealloc_space", | |
2877 | sizeof(struct ext4_prealloc_space), | |
2878 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2879 | if (ext4_pspace_cachep == NULL) | |
2880 | return -ENOMEM; | |
2881 | ||
256bdb49 ES |
2882 | ext4_ac_cachep = |
2883 | kmem_cache_create("ext4_alloc_context", | |
2884 | sizeof(struct ext4_allocation_context), | |
2885 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2886 | if (ext4_ac_cachep == NULL) { | |
2887 | kmem_cache_destroy(ext4_pspace_cachep); | |
2888 | return -ENOMEM; | |
2889 | } | |
c894058d AK |
2890 | |
2891 | ext4_free_ext_cachep = | |
2892 | kmem_cache_create("ext4_free_block_extents", | |
2893 | sizeof(struct ext4_free_data), | |
2894 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2895 | if (ext4_free_ext_cachep == NULL) { | |
2896 | kmem_cache_destroy(ext4_pspace_cachep); | |
2897 | kmem_cache_destroy(ext4_ac_cachep); | |
2898 | return -ENOMEM; | |
2899 | } | |
c9de560d AT |
2900 | return 0; |
2901 | } | |
2902 | ||
2903 | void exit_ext4_mballoc(void) | |
2904 | { | |
2905 | /* XXX: synchronize_rcu(); */ | |
2906 | kmem_cache_destroy(ext4_pspace_cachep); | |
256bdb49 | 2907 | kmem_cache_destroy(ext4_ac_cachep); |
c894058d | 2908 | kmem_cache_destroy(ext4_free_ext_cachep); |
c9de560d AT |
2909 | } |
2910 | ||
2911 | ||
2912 | /* | |
2913 | * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps | |
2914 | * Returns 0 if success or error code | |
2915 | */ | |
4ddfef7b ES |
2916 | static noinline_for_stack int |
2917 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
498e5f24 | 2918 | handle_t *handle, unsigned int reserv_blks) |
c9de560d AT |
2919 | { |
2920 | struct buffer_head *bitmap_bh = NULL; | |
2921 | struct ext4_super_block *es; | |
2922 | struct ext4_group_desc *gdp; | |
2923 | struct buffer_head *gdp_bh; | |
2924 | struct ext4_sb_info *sbi; | |
2925 | struct super_block *sb; | |
2926 | ext4_fsblk_t block; | |
519deca0 | 2927 | int err, len; |
c9de560d AT |
2928 | |
2929 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2930 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2931 | ||
2932 | sb = ac->ac_sb; | |
2933 | sbi = EXT4_SB(sb); | |
2934 | es = sbi->s_es; | |
2935 | ||
c9de560d AT |
2936 | |
2937 | err = -EIO; | |
574ca174 | 2938 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
c9de560d AT |
2939 | if (!bitmap_bh) |
2940 | goto out_err; | |
2941 | ||
2942 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2943 | if (err) | |
2944 | goto out_err; | |
2945 | ||
2946 | err = -EIO; | |
2947 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
2948 | if (!gdp) | |
2949 | goto out_err; | |
2950 | ||
a9df9a49 | 2951 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
9fd9784c | 2952 | ext4_free_blks_count(sb, gdp)); |
03cddb80 | 2953 | |
c9de560d AT |
2954 | err = ext4_journal_get_write_access(handle, gdp_bh); |
2955 | if (err) | |
2956 | goto out_err; | |
2957 | ||
2958 | block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb) | |
2959 | + ac->ac_b_ex.fe_start | |
2960 | + le32_to_cpu(es->s_first_data_block); | |
2961 | ||
519deca0 | 2962 | len = ac->ac_b_ex.fe_len; |
6fd058f7 | 2963 | if (!ext4_data_block_valid(sbi, block, len)) { |
46e665e9 | 2964 | ext4_error(sb, __func__, |
6fd058f7 TT |
2965 | "Allocating blocks %llu-%llu which overlap " |
2966 | "fs metadata\n", block, block+len); | |
519deca0 AK |
2967 | /* File system mounted not to panic on error |
2968 | * Fix the bitmap and repeat the block allocation | |
2969 | * We leak some of the blocks here. | |
2970 | */ | |
955ce5f5 AK |
2971 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
2972 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, | |
2973 | ac->ac_b_ex.fe_len); | |
2974 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
0390131b | 2975 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 AK |
2976 | if (!err) |
2977 | err = -EAGAIN; | |
2978 | goto out_err; | |
c9de560d | 2979 | } |
955ce5f5 AK |
2980 | |
2981 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
2982 | #ifdef AGGRESSIVE_CHECK |
2983 | { | |
2984 | int i; | |
2985 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
2986 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
2987 | bitmap_bh->b_data)); | |
2988 | } | |
2989 | } | |
2990 | #endif | |
955ce5f5 | 2991 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len); |
c9de560d AT |
2992 | if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
2993 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
560671a0 AK |
2994 | ext4_free_blks_set(sb, gdp, |
2995 | ext4_free_blocks_after_init(sb, | |
2996 | ac->ac_b_ex.fe_group, gdp)); | |
c9de560d | 2997 | } |
560671a0 AK |
2998 | len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len; |
2999 | ext4_free_blks_set(sb, gdp, len); | |
c9de560d | 3000 | gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
3001 | |
3002 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
6bc6e63f | 3003 | percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 3004 | /* |
6bc6e63f | 3005 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 3006 | */ |
6bc6e63f AK |
3007 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
3008 | /* release all the reserved blocks if non delalloc */ | |
3009 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks); | |
60e58e0f | 3010 | else { |
6bc6e63f AK |
3011 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, |
3012 | ac->ac_b_ex.fe_len); | |
60e58e0f MC |
3013 | /* convert reserved quota blocks to real quota blocks */ |
3014 | vfs_dq_claim_block(ac->ac_inode, ac->ac_b_ex.fe_len); | |
3015 | } | |
c9de560d | 3016 | |
772cb7c8 JS |
3017 | if (sbi->s_log_groups_per_flex) { |
3018 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
3019 | ac->ac_b_ex.fe_group); | |
9f24e420 TT |
3020 | atomic_sub(ac->ac_b_ex.fe_len, |
3021 | &sbi->s_flex_groups[flex_group].free_blocks); | |
772cb7c8 JS |
3022 | } |
3023 | ||
0390131b | 3024 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
3025 | if (err) |
3026 | goto out_err; | |
0390131b | 3027 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
3028 | |
3029 | out_err: | |
3030 | sb->s_dirt = 1; | |
42a10add | 3031 | brelse(bitmap_bh); |
c9de560d AT |
3032 | return err; |
3033 | } | |
3034 | ||
3035 | /* | |
3036 | * here we normalize request for locality group | |
3037 | * Group request are normalized to s_strip size if we set the same via mount | |
3038 | * option. If not we set it to s_mb_group_prealloc which can be configured via | |
b713a5ec | 3039 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
3040 | * |
3041 | * XXX: should we try to preallocate more than the group has now? | |
3042 | */ | |
3043 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
3044 | { | |
3045 | struct super_block *sb = ac->ac_sb; | |
3046 | struct ext4_locality_group *lg = ac->ac_lg; | |
3047 | ||
3048 | BUG_ON(lg == NULL); | |
3049 | if (EXT4_SB(sb)->s_stripe) | |
3050 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe; | |
3051 | else | |
3052 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; | |
60bd63d1 | 3053 | mb_debug("#%u: goal %u blocks for locality group\n", |
c9de560d AT |
3054 | current->pid, ac->ac_g_ex.fe_len); |
3055 | } | |
3056 | ||
3057 | /* | |
3058 | * Normalization means making request better in terms of | |
3059 | * size and alignment | |
3060 | */ | |
4ddfef7b ES |
3061 | static noinline_for_stack void |
3062 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
3063 | struct ext4_allocation_request *ar) |
3064 | { | |
3065 | int bsbits, max; | |
3066 | ext4_lblk_t end; | |
c9de560d AT |
3067 | loff_t size, orig_size, start_off; |
3068 | ext4_lblk_t start, orig_start; | |
3069 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); | |
9a0762c5 | 3070 | struct ext4_prealloc_space *pa; |
c9de560d AT |
3071 | |
3072 | /* do normalize only data requests, metadata requests | |
3073 | do not need preallocation */ | |
3074 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3075 | return; | |
3076 | ||
3077 | /* sometime caller may want exact blocks */ | |
3078 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
3079 | return; | |
3080 | ||
3081 | /* caller may indicate that preallocation isn't | |
3082 | * required (it's a tail, for example) */ | |
3083 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
3084 | return; | |
3085 | ||
3086 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
3087 | ext4_mb_normalize_group_request(ac); | |
3088 | return ; | |
3089 | } | |
3090 | ||
3091 | bsbits = ac->ac_sb->s_blocksize_bits; | |
3092 | ||
3093 | /* first, let's learn actual file size | |
3094 | * given current request is allocated */ | |
3095 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; | |
3096 | size = size << bsbits; | |
3097 | if (size < i_size_read(ac->ac_inode)) | |
3098 | size = i_size_read(ac->ac_inode); | |
3099 | ||
1930479c VC |
3100 | /* max size of free chunks */ |
3101 | max = 2 << bsbits; | |
c9de560d | 3102 | |
1930479c VC |
3103 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
3104 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
3105 | |
3106 | /* first, try to predict filesize */ | |
3107 | /* XXX: should this table be tunable? */ | |
3108 | start_off = 0; | |
3109 | if (size <= 16 * 1024) { | |
3110 | size = 16 * 1024; | |
3111 | } else if (size <= 32 * 1024) { | |
3112 | size = 32 * 1024; | |
3113 | } else if (size <= 64 * 1024) { | |
3114 | size = 64 * 1024; | |
3115 | } else if (size <= 128 * 1024) { | |
3116 | size = 128 * 1024; | |
3117 | } else if (size <= 256 * 1024) { | |
3118 | size = 256 * 1024; | |
3119 | } else if (size <= 512 * 1024) { | |
3120 | size = 512 * 1024; | |
3121 | } else if (size <= 1024 * 1024) { | |
3122 | size = 1024 * 1024; | |
1930479c | 3123 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 3124 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
3125 | (21 - bsbits)) << 21; |
3126 | size = 2 * 1024 * 1024; | |
3127 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
3128 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3129 | (22 - bsbits)) << 22; | |
3130 | size = 4 * 1024 * 1024; | |
3131 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 3132 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
3133 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3134 | (23 - bsbits)) << 23; | |
3135 | size = 8 * 1024 * 1024; | |
3136 | } else { | |
3137 | start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; | |
3138 | size = ac->ac_o_ex.fe_len << bsbits; | |
3139 | } | |
3140 | orig_size = size = size >> bsbits; | |
3141 | orig_start = start = start_off >> bsbits; | |
3142 | ||
3143 | /* don't cover already allocated blocks in selected range */ | |
3144 | if (ar->pleft && start <= ar->lleft) { | |
3145 | size -= ar->lleft + 1 - start; | |
3146 | start = ar->lleft + 1; | |
3147 | } | |
3148 | if (ar->pright && start + size - 1 >= ar->lright) | |
3149 | size -= start + size - ar->lright; | |
3150 | ||
3151 | end = start + size; | |
3152 | ||
3153 | /* check we don't cross already preallocated blocks */ | |
3154 | rcu_read_lock(); | |
9a0762c5 | 3155 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3156 | ext4_lblk_t pa_end; |
c9de560d | 3157 | |
c9de560d AT |
3158 | if (pa->pa_deleted) |
3159 | continue; | |
3160 | spin_lock(&pa->pa_lock); | |
3161 | if (pa->pa_deleted) { | |
3162 | spin_unlock(&pa->pa_lock); | |
3163 | continue; | |
3164 | } | |
3165 | ||
3166 | pa_end = pa->pa_lstart + pa->pa_len; | |
3167 | ||
3168 | /* PA must not overlap original request */ | |
3169 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
3170 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
3171 | ||
3172 | /* skip PA normalized request doesn't overlap with */ | |
3173 | if (pa->pa_lstart >= end) { | |
3174 | spin_unlock(&pa->pa_lock); | |
3175 | continue; | |
3176 | } | |
3177 | if (pa_end <= start) { | |
3178 | spin_unlock(&pa->pa_lock); | |
3179 | continue; | |
3180 | } | |
3181 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3182 | ||
3183 | if (pa_end <= ac->ac_o_ex.fe_logical) { | |
3184 | BUG_ON(pa_end < start); | |
3185 | start = pa_end; | |
3186 | } | |
3187 | ||
3188 | if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { | |
3189 | BUG_ON(pa->pa_lstart > end); | |
3190 | end = pa->pa_lstart; | |
3191 | } | |
3192 | spin_unlock(&pa->pa_lock); | |
3193 | } | |
3194 | rcu_read_unlock(); | |
3195 | size = end - start; | |
3196 | ||
3197 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3198 | rcu_read_lock(); | |
9a0762c5 | 3199 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3200 | ext4_lblk_t pa_end; |
c9de560d AT |
3201 | spin_lock(&pa->pa_lock); |
3202 | if (pa->pa_deleted == 0) { | |
3203 | pa_end = pa->pa_lstart + pa->pa_len; | |
3204 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); | |
3205 | } | |
3206 | spin_unlock(&pa->pa_lock); | |
3207 | } | |
3208 | rcu_read_unlock(); | |
3209 | ||
3210 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3211 | start > ac->ac_o_ex.fe_logical) { | |
3212 | printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n", | |
3213 | (unsigned long) start, (unsigned long) size, | |
3214 | (unsigned long) ac->ac_o_ex.fe_logical); | |
3215 | } | |
3216 | BUG_ON(start + size <= ac->ac_o_ex.fe_logical && | |
3217 | start > ac->ac_o_ex.fe_logical); | |
8d03c7a0 | 3218 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
3219 | |
3220 | /* now prepare goal request */ | |
3221 | ||
3222 | /* XXX: is it better to align blocks WRT to logical | |
3223 | * placement or satisfy big request as is */ | |
3224 | ac->ac_g_ex.fe_logical = start; | |
3225 | ac->ac_g_ex.fe_len = size; | |
3226 | ||
3227 | /* define goal start in order to merge */ | |
3228 | if (ar->pright && (ar->lright == (start + size))) { | |
3229 | /* merge to the right */ | |
3230 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3231 | &ac->ac_f_ex.fe_group, | |
3232 | &ac->ac_f_ex.fe_start); | |
3233 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3234 | } | |
3235 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3236 | /* merge to the left */ | |
3237 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3238 | &ac->ac_f_ex.fe_group, | |
3239 | &ac->ac_f_ex.fe_start); | |
3240 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3241 | } | |
3242 | ||
3243 | mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size, | |
3244 | (unsigned) orig_size, (unsigned) start); | |
3245 | } | |
3246 | ||
3247 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3248 | { | |
3249 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3250 | ||
3251 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3252 | atomic_inc(&sbi->s_bal_reqs); | |
3253 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
3254 | if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len) | |
3255 | atomic_inc(&sbi->s_bal_success); | |
3256 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3257 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3258 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3259 | atomic_inc(&sbi->s_bal_goals); | |
3260 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3261 | atomic_inc(&sbi->s_bal_breaks); | |
3262 | } | |
3263 | ||
3264 | ext4_mb_store_history(ac); | |
3265 | } | |
3266 | ||
3267 | /* | |
3268 | * use blocks preallocated to inode | |
3269 | */ | |
3270 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3271 | struct ext4_prealloc_space *pa) | |
3272 | { | |
3273 | ext4_fsblk_t start; | |
3274 | ext4_fsblk_t end; | |
3275 | int len; | |
3276 | ||
3277 | /* found preallocated blocks, use them */ | |
3278 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
3279 | end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len); | |
3280 | len = end - start; | |
3281 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, | |
3282 | &ac->ac_b_ex.fe_start); | |
3283 | ac->ac_b_ex.fe_len = len; | |
3284 | ac->ac_status = AC_STATUS_FOUND; | |
3285 | ac->ac_pa = pa; | |
3286 | ||
3287 | BUG_ON(start < pa->pa_pstart); | |
3288 | BUG_ON(start + len > pa->pa_pstart + pa->pa_len); | |
3289 | BUG_ON(pa->pa_free < len); | |
3290 | pa->pa_free -= len; | |
3291 | ||
60bd63d1 | 3292 | mb_debug("use %llu/%u from inode pa %p\n", start, len, pa); |
c9de560d AT |
3293 | } |
3294 | ||
3295 | /* | |
3296 | * use blocks preallocated to locality group | |
3297 | */ | |
3298 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3299 | struct ext4_prealloc_space *pa) | |
3300 | { | |
03cddb80 | 3301 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3302 | |
c9de560d AT |
3303 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3304 | &ac->ac_b_ex.fe_group, | |
3305 | &ac->ac_b_ex.fe_start); | |
3306 | ac->ac_b_ex.fe_len = len; | |
3307 | ac->ac_status = AC_STATUS_FOUND; | |
3308 | ac->ac_pa = pa; | |
3309 | ||
3310 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3311 | * possible race when the group is being loaded concurrently |
c9de560d | 3312 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3313 | * in on-disk bitmap -- see ext4_mb_release_context() |
3314 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d AT |
3315 | */ |
3316 | mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); | |
3317 | } | |
3318 | ||
5e745b04 AK |
3319 | /* |
3320 | * Return the prealloc space that have minimal distance | |
3321 | * from the goal block. @cpa is the prealloc | |
3322 | * space that is having currently known minimal distance | |
3323 | * from the goal block. | |
3324 | */ | |
3325 | static struct ext4_prealloc_space * | |
3326 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3327 | struct ext4_prealloc_space *pa, | |
3328 | struct ext4_prealloc_space *cpa) | |
3329 | { | |
3330 | ext4_fsblk_t cur_distance, new_distance; | |
3331 | ||
3332 | if (cpa == NULL) { | |
3333 | atomic_inc(&pa->pa_count); | |
3334 | return pa; | |
3335 | } | |
3336 | cur_distance = abs(goal_block - cpa->pa_pstart); | |
3337 | new_distance = abs(goal_block - pa->pa_pstart); | |
3338 | ||
3339 | if (cur_distance < new_distance) | |
3340 | return cpa; | |
3341 | ||
3342 | /* drop the previous reference */ | |
3343 | atomic_dec(&cpa->pa_count); | |
3344 | atomic_inc(&pa->pa_count); | |
3345 | return pa; | |
3346 | } | |
3347 | ||
c9de560d AT |
3348 | /* |
3349 | * search goal blocks in preallocated space | |
3350 | */ | |
4ddfef7b ES |
3351 | static noinline_for_stack int |
3352 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
c9de560d | 3353 | { |
6be2ded1 | 3354 | int order, i; |
c9de560d AT |
3355 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3356 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3357 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3358 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3359 | |
3360 | /* only data can be preallocated */ | |
3361 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3362 | return 0; | |
3363 | ||
3364 | /* first, try per-file preallocation */ | |
3365 | rcu_read_lock(); | |
9a0762c5 | 3366 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3367 | |
3368 | /* all fields in this condition don't change, | |
3369 | * so we can skip locking for them */ | |
3370 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
3371 | ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len) | |
3372 | continue; | |
3373 | ||
3374 | /* found preallocated blocks, use them */ | |
3375 | spin_lock(&pa->pa_lock); | |
3376 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3377 | atomic_inc(&pa->pa_count); | |
3378 | ext4_mb_use_inode_pa(ac, pa); | |
3379 | spin_unlock(&pa->pa_lock); | |
3380 | ac->ac_criteria = 10; | |
3381 | rcu_read_unlock(); | |
3382 | return 1; | |
3383 | } | |
3384 | spin_unlock(&pa->pa_lock); | |
3385 | } | |
3386 | rcu_read_unlock(); | |
3387 | ||
3388 | /* can we use group allocation? */ | |
3389 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3390 | return 0; | |
3391 | ||
3392 | /* inode may have no locality group for some reason */ | |
3393 | lg = ac->ac_lg; | |
3394 | if (lg == NULL) | |
3395 | return 0; | |
6be2ded1 AK |
3396 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3397 | if (order > PREALLOC_TB_SIZE - 1) | |
3398 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3399 | order = PREALLOC_TB_SIZE - 1; | |
3400 | ||
5e745b04 AK |
3401 | goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) + |
3402 | ac->ac_g_ex.fe_start + | |
3403 | le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block); | |
3404 | /* | |
3405 | * search for the prealloc space that is having | |
3406 | * minimal distance from the goal block. | |
3407 | */ | |
6be2ded1 AK |
3408 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3409 | rcu_read_lock(); | |
3410 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3411 | pa_inode_list) { | |
3412 | spin_lock(&pa->pa_lock); | |
3413 | if (pa->pa_deleted == 0 && | |
3414 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3415 | |
3416 | cpa = ext4_mb_check_group_pa(goal_block, | |
3417 | pa, cpa); | |
6be2ded1 | 3418 | } |
c9de560d | 3419 | spin_unlock(&pa->pa_lock); |
c9de560d | 3420 | } |
6be2ded1 | 3421 | rcu_read_unlock(); |
c9de560d | 3422 | } |
5e745b04 AK |
3423 | if (cpa) { |
3424 | ext4_mb_use_group_pa(ac, cpa); | |
3425 | ac->ac_criteria = 20; | |
3426 | return 1; | |
3427 | } | |
c9de560d AT |
3428 | return 0; |
3429 | } | |
3430 | ||
7a2fcbf7 AK |
3431 | /* |
3432 | * the function goes through all block freed in the group | |
3433 | * but not yet committed and marks them used in in-core bitmap. | |
3434 | * buddy must be generated from this bitmap | |
955ce5f5 | 3435 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3436 | */ |
3437 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3438 | ext4_group_t group) | |
3439 | { | |
3440 | struct rb_node *n; | |
3441 | struct ext4_group_info *grp; | |
3442 | struct ext4_free_data *entry; | |
3443 | ||
3444 | grp = ext4_get_group_info(sb, group); | |
3445 | n = rb_first(&(grp->bb_free_root)); | |
3446 | ||
3447 | while (n) { | |
3448 | entry = rb_entry(n, struct ext4_free_data, node); | |
955ce5f5 | 3449 | mb_set_bits(bitmap, entry->start_blk, entry->count); |
7a2fcbf7 AK |
3450 | n = rb_next(n); |
3451 | } | |
3452 | return; | |
3453 | } | |
3454 | ||
c9de560d AT |
3455 | /* |
3456 | * the function goes through all preallocation in this group and marks them | |
3457 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3458 | * Need to be called with ext4 group lock held |
c9de560d AT |
3459 | */ |
3460 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
3461 | ext4_group_t group) | |
3462 | { | |
3463 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3464 | struct ext4_prealloc_space *pa; | |
3465 | struct list_head *cur; | |
3466 | ext4_group_t groupnr; | |
3467 | ext4_grpblk_t start; | |
3468 | int preallocated = 0; | |
3469 | int count = 0; | |
3470 | int len; | |
3471 | ||
3472 | /* all form of preallocation discards first load group, | |
3473 | * so the only competing code is preallocation use. | |
3474 | * we don't need any locking here | |
3475 | * notice we do NOT ignore preallocations with pa_deleted | |
3476 | * otherwise we could leave used blocks available for | |
3477 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3478 | * is dropping preallocation | |
3479 | */ | |
3480 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3481 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3482 | spin_lock(&pa->pa_lock); | |
3483 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3484 | &groupnr, &start); | |
3485 | len = pa->pa_len; | |
3486 | spin_unlock(&pa->pa_lock); | |
3487 | if (unlikely(len == 0)) | |
3488 | continue; | |
3489 | BUG_ON(groupnr != group); | |
955ce5f5 | 3490 | mb_set_bits(bitmap, start, len); |
c9de560d AT |
3491 | preallocated += len; |
3492 | count++; | |
3493 | } | |
a9df9a49 | 3494 | mb_debug("prellocated %u for group %u\n", preallocated, group); |
c9de560d AT |
3495 | } |
3496 | ||
3497 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3498 | { | |
3499 | struct ext4_prealloc_space *pa; | |
3500 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
3501 | kmem_cache_free(ext4_pspace_cachep, pa); | |
3502 | } | |
3503 | ||
3504 | /* | |
3505 | * drops a reference to preallocated space descriptor | |
3506 | * if this was the last reference and the space is consumed | |
3507 | */ | |
3508 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3509 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3510 | { | |
a9df9a49 | 3511 | ext4_group_t grp; |
d33a1976 | 3512 | ext4_fsblk_t grp_blk; |
c9de560d AT |
3513 | |
3514 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) | |
3515 | return; | |
3516 | ||
3517 | /* in this short window concurrent discard can set pa_deleted */ | |
3518 | spin_lock(&pa->pa_lock); | |
3519 | if (pa->pa_deleted == 1) { | |
3520 | spin_unlock(&pa->pa_lock); | |
3521 | return; | |
3522 | } | |
3523 | ||
3524 | pa->pa_deleted = 1; | |
3525 | spin_unlock(&pa->pa_lock); | |
3526 | ||
d33a1976 | 3527 | grp_blk = pa->pa_pstart; |
cc0fb9ad AK |
3528 | /* |
3529 | * If doing group-based preallocation, pa_pstart may be in the | |
3530 | * next group when pa is used up | |
3531 | */ | |
3532 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3533 | grp_blk--; |
3534 | ||
3535 | ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL); | |
c9de560d AT |
3536 | |
3537 | /* | |
3538 | * possible race: | |
3539 | * | |
3540 | * P1 (buddy init) P2 (regular allocation) | |
3541 | * find block B in PA | |
3542 | * copy on-disk bitmap to buddy | |
3543 | * mark B in on-disk bitmap | |
3544 | * drop PA from group | |
3545 | * mark all PAs in buddy | |
3546 | * | |
3547 | * thus, P1 initializes buddy with B available. to prevent this | |
3548 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3549 | * against that pair | |
3550 | */ | |
3551 | ext4_lock_group(sb, grp); | |
3552 | list_del(&pa->pa_group_list); | |
3553 | ext4_unlock_group(sb, grp); | |
3554 | ||
3555 | spin_lock(pa->pa_obj_lock); | |
3556 | list_del_rcu(&pa->pa_inode_list); | |
3557 | spin_unlock(pa->pa_obj_lock); | |
3558 | ||
3559 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3560 | } | |
3561 | ||
3562 | /* | |
3563 | * creates new preallocated space for given inode | |
3564 | */ | |
4ddfef7b ES |
3565 | static noinline_for_stack int |
3566 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3567 | { |
3568 | struct super_block *sb = ac->ac_sb; | |
3569 | struct ext4_prealloc_space *pa; | |
3570 | struct ext4_group_info *grp; | |
3571 | struct ext4_inode_info *ei; | |
3572 | ||
3573 | /* preallocate only when found space is larger then requested */ | |
3574 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3575 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3576 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3577 | ||
3578 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3579 | if (pa == NULL) | |
3580 | return -ENOMEM; | |
3581 | ||
3582 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3583 | int winl; | |
3584 | int wins; | |
3585 | int win; | |
3586 | int offs; | |
3587 | ||
3588 | /* we can't allocate as much as normalizer wants. | |
3589 | * so, found space must get proper lstart | |
3590 | * to cover original request */ | |
3591 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3592 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3593 | ||
3594 | /* we're limited by original request in that | |
3595 | * logical block must be covered any way | |
3596 | * winl is window we can move our chunk within */ | |
3597 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3598 | ||
3599 | /* also, we should cover whole original request */ | |
3600 | wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len; | |
3601 | ||
3602 | /* the smallest one defines real window */ | |
3603 | win = min(winl, wins); | |
3604 | ||
3605 | offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len; | |
3606 | if (offs && offs < win) | |
3607 | win = offs; | |
3608 | ||
3609 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win; | |
3610 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); | |
3611 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3612 | } | |
3613 | ||
3614 | /* preallocation can change ac_b_ex, thus we store actually | |
3615 | * allocated blocks for history */ | |
3616 | ac->ac_f_ex = ac->ac_b_ex; | |
3617 | ||
3618 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3619 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3620 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3621 | pa->pa_free = pa->pa_len; | |
3622 | atomic_set(&pa->pa_count, 1); | |
3623 | spin_lock_init(&pa->pa_lock); | |
d794bf8e AK |
3624 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3625 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3626 | pa->pa_deleted = 0; |
cc0fb9ad | 3627 | pa->pa_type = MB_INODE_PA; |
c9de560d AT |
3628 | |
3629 | mb_debug("new inode pa %p: %llu/%u for %u\n", pa, | |
3630 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); | |
9bffad1e | 3631 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3632 | |
3633 | ext4_mb_use_inode_pa(ac, pa); | |
3634 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3635 | ||
3636 | ei = EXT4_I(ac->ac_inode); | |
3637 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3638 | ||
3639 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3640 | pa->pa_inode = ac->ac_inode; | |
3641 | ||
3642 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3643 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3644 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3645 | ||
3646 | spin_lock(pa->pa_obj_lock); | |
3647 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3648 | spin_unlock(pa->pa_obj_lock); | |
3649 | ||
3650 | return 0; | |
3651 | } | |
3652 | ||
3653 | /* | |
3654 | * creates new preallocated space for locality group inodes belongs to | |
3655 | */ | |
4ddfef7b ES |
3656 | static noinline_for_stack int |
3657 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3658 | { |
3659 | struct super_block *sb = ac->ac_sb; | |
3660 | struct ext4_locality_group *lg; | |
3661 | struct ext4_prealloc_space *pa; | |
3662 | struct ext4_group_info *grp; | |
3663 | ||
3664 | /* preallocate only when found space is larger then requested */ | |
3665 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3666 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3667 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3668 | ||
3669 | BUG_ON(ext4_pspace_cachep == NULL); | |
3670 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3671 | if (pa == NULL) | |
3672 | return -ENOMEM; | |
3673 | ||
3674 | /* preallocation can change ac_b_ex, thus we store actually | |
3675 | * allocated blocks for history */ | |
3676 | ac->ac_f_ex = ac->ac_b_ex; | |
3677 | ||
3678 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3679 | pa->pa_lstart = pa->pa_pstart; | |
3680 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3681 | pa->pa_free = pa->pa_len; | |
3682 | atomic_set(&pa->pa_count, 1); | |
3683 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3684 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 3685 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 3686 | pa->pa_deleted = 0; |
cc0fb9ad | 3687 | pa->pa_type = MB_GROUP_PA; |
c9de560d AT |
3688 | |
3689 | mb_debug("new group pa %p: %llu/%u for %u\n", pa, | |
9bffad1e TT |
3690 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
3691 | trace_ext4_mb_new_group_pa(ac, pa); | |
c9de560d AT |
3692 | |
3693 | ext4_mb_use_group_pa(ac, pa); | |
3694 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3695 | ||
3696 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3697 | lg = ac->ac_lg; | |
3698 | BUG_ON(lg == NULL); | |
3699 | ||
3700 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3701 | pa->pa_inode = NULL; | |
3702 | ||
3703 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3704 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3705 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3706 | ||
6be2ded1 AK |
3707 | /* |
3708 | * We will later add the new pa to the right bucket | |
3709 | * after updating the pa_free in ext4_mb_release_context | |
3710 | */ | |
c9de560d AT |
3711 | return 0; |
3712 | } | |
3713 | ||
3714 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3715 | { | |
3716 | int err; | |
3717 | ||
3718 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3719 | err = ext4_mb_new_group_pa(ac); | |
3720 | else | |
3721 | err = ext4_mb_new_inode_pa(ac); | |
3722 | return err; | |
3723 | } | |
3724 | ||
3725 | /* | |
3726 | * finds all unused blocks in on-disk bitmap, frees them in | |
3727 | * in-core bitmap and buddy. | |
3728 | * @pa must be unlinked from inode and group lists, so that | |
3729 | * nobody else can find/use it. | |
3730 | * the caller MUST hold group/inode locks. | |
3731 | * TODO: optimize the case when there are no in-core structures yet | |
3732 | */ | |
4ddfef7b ES |
3733 | static noinline_for_stack int |
3734 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
c83617db AK |
3735 | struct ext4_prealloc_space *pa, |
3736 | struct ext4_allocation_context *ac) | |
c9de560d | 3737 | { |
c9de560d AT |
3738 | struct super_block *sb = e4b->bd_sb; |
3739 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
3740 | unsigned int end; |
3741 | unsigned int next; | |
c9de560d AT |
3742 | ext4_group_t group; |
3743 | ext4_grpblk_t bit; | |
ba80b101 | 3744 | unsigned long long grp_blk_start; |
c9de560d AT |
3745 | sector_t start; |
3746 | int err = 0; | |
3747 | int free = 0; | |
3748 | ||
3749 | BUG_ON(pa->pa_deleted == 0); | |
3750 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
ba80b101 | 3751 | grp_blk_start = pa->pa_pstart - bit; |
c9de560d AT |
3752 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
3753 | end = bit + pa->pa_len; | |
3754 | ||
256bdb49 ES |
3755 | if (ac) { |
3756 | ac->ac_sb = sb; | |
3757 | ac->ac_inode = pa->pa_inode; | |
3758 | ac->ac_op = EXT4_MB_HISTORY_DISCARD; | |
3759 | } | |
c9de560d AT |
3760 | |
3761 | while (bit < end) { | |
ffad0a44 | 3762 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3763 | if (bit >= end) |
3764 | break; | |
ffad0a44 | 3765 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3766 | start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit + |
3767 | le32_to_cpu(sbi->s_es->s_first_data_block); | |
3768 | mb_debug(" free preallocated %u/%u in group %u\n", | |
3769 | (unsigned) start, (unsigned) next - bit, | |
3770 | (unsigned) group); | |
3771 | free += next - bit; | |
3772 | ||
256bdb49 ES |
3773 | if (ac) { |
3774 | ac->ac_b_ex.fe_group = group; | |
3775 | ac->ac_b_ex.fe_start = bit; | |
3776 | ac->ac_b_ex.fe_len = next - bit; | |
3777 | ac->ac_b_ex.fe_logical = 0; | |
3778 | ext4_mb_store_history(ac); | |
3779 | } | |
c9de560d | 3780 | |
9bffad1e TT |
3781 | trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit, |
3782 | next - bit); | |
c9de560d AT |
3783 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
3784 | bit = next + 1; | |
3785 | } | |
3786 | if (free != pa->pa_free) { | |
26346ff6 | 3787 | printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n", |
c9de560d AT |
3788 | pa, (unsigned long) pa->pa_lstart, |
3789 | (unsigned long) pa->pa_pstart, | |
3790 | (unsigned long) pa->pa_len); | |
5d1b1b3f AK |
3791 | ext4_grp_locked_error(sb, group, |
3792 | __func__, "free %u, pa_free %u", | |
3793 | free, pa->pa_free); | |
e56eb659 AK |
3794 | /* |
3795 | * pa is already deleted so we use the value obtained | |
3796 | * from the bitmap and continue. | |
3797 | */ | |
c9de560d | 3798 | } |
c9de560d AT |
3799 | atomic_add(free, &sbi->s_mb_discarded); |
3800 | ||
3801 | return err; | |
3802 | } | |
3803 | ||
4ddfef7b ES |
3804 | static noinline_for_stack int |
3805 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
c83617db AK |
3806 | struct ext4_prealloc_space *pa, |
3807 | struct ext4_allocation_context *ac) | |
c9de560d | 3808 | { |
c9de560d AT |
3809 | struct super_block *sb = e4b->bd_sb; |
3810 | ext4_group_t group; | |
3811 | ext4_grpblk_t bit; | |
3812 | ||
256bdb49 ES |
3813 | if (ac) |
3814 | ac->ac_op = EXT4_MB_HISTORY_DISCARD; | |
c9de560d | 3815 | |
9bffad1e | 3816 | trace_ext4_mb_release_group_pa(ac, pa); |
c9de560d AT |
3817 | BUG_ON(pa->pa_deleted == 0); |
3818 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3819 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3820 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3821 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3822 | ||
256bdb49 ES |
3823 | if (ac) { |
3824 | ac->ac_sb = sb; | |
3825 | ac->ac_inode = NULL; | |
3826 | ac->ac_b_ex.fe_group = group; | |
3827 | ac->ac_b_ex.fe_start = bit; | |
3828 | ac->ac_b_ex.fe_len = pa->pa_len; | |
3829 | ac->ac_b_ex.fe_logical = 0; | |
3830 | ext4_mb_store_history(ac); | |
256bdb49 | 3831 | } |
c9de560d AT |
3832 | |
3833 | return 0; | |
3834 | } | |
3835 | ||
3836 | /* | |
3837 | * releases all preallocations in given group | |
3838 | * | |
3839 | * first, we need to decide discard policy: | |
3840 | * - when do we discard | |
3841 | * 1) ENOSPC | |
3842 | * - how many do we discard | |
3843 | * 1) how many requested | |
3844 | */ | |
4ddfef7b ES |
3845 | static noinline_for_stack int |
3846 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3847 | ext4_group_t group, int needed) |
3848 | { | |
3849 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3850 | struct buffer_head *bitmap_bh = NULL; | |
3851 | struct ext4_prealloc_space *pa, *tmp; | |
c83617db | 3852 | struct ext4_allocation_context *ac; |
c9de560d AT |
3853 | struct list_head list; |
3854 | struct ext4_buddy e4b; | |
3855 | int err; | |
3856 | int busy = 0; | |
3857 | int free = 0; | |
3858 | ||
a9df9a49 | 3859 | mb_debug("discard preallocation for group %u\n", group); |
c9de560d AT |
3860 | |
3861 | if (list_empty(&grp->bb_prealloc_list)) | |
3862 | return 0; | |
3863 | ||
574ca174 | 3864 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3865 | if (bitmap_bh == NULL) { |
ce89f46c | 3866 | ext4_error(sb, __func__, "Error in reading block " |
a9df9a49 | 3867 | "bitmap for %u", group); |
ce89f46c | 3868 | return 0; |
c9de560d AT |
3869 | } |
3870 | ||
3871 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c AK |
3872 | if (err) { |
3873 | ext4_error(sb, __func__, "Error in loading buddy " | |
a9df9a49 | 3874 | "information for %u", group); |
ce89f46c AK |
3875 | put_bh(bitmap_bh); |
3876 | return 0; | |
3877 | } | |
c9de560d AT |
3878 | |
3879 | if (needed == 0) | |
3880 | needed = EXT4_BLOCKS_PER_GROUP(sb) + 1; | |
3881 | ||
c9de560d | 3882 | INIT_LIST_HEAD(&list); |
c83617db | 3883 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
9bffad1e TT |
3884 | if (ac) |
3885 | ac->ac_sb = sb; | |
c9de560d AT |
3886 | repeat: |
3887 | ext4_lock_group(sb, group); | |
3888 | list_for_each_entry_safe(pa, tmp, | |
3889 | &grp->bb_prealloc_list, pa_group_list) { | |
3890 | spin_lock(&pa->pa_lock); | |
3891 | if (atomic_read(&pa->pa_count)) { | |
3892 | spin_unlock(&pa->pa_lock); | |
3893 | busy = 1; | |
3894 | continue; | |
3895 | } | |
3896 | if (pa->pa_deleted) { | |
3897 | spin_unlock(&pa->pa_lock); | |
3898 | continue; | |
3899 | } | |
3900 | ||
3901 | /* seems this one can be freed ... */ | |
3902 | pa->pa_deleted = 1; | |
3903 | ||
3904 | /* we can trust pa_free ... */ | |
3905 | free += pa->pa_free; | |
3906 | ||
3907 | spin_unlock(&pa->pa_lock); | |
3908 | ||
3909 | list_del(&pa->pa_group_list); | |
3910 | list_add(&pa->u.pa_tmp_list, &list); | |
3911 | } | |
3912 | ||
3913 | /* if we still need more blocks and some PAs were used, try again */ | |
3914 | if (free < needed && busy) { | |
3915 | busy = 0; | |
3916 | ext4_unlock_group(sb, group); | |
3917 | /* | |
3918 | * Yield the CPU here so that we don't get soft lockup | |
3919 | * in non preempt case. | |
3920 | */ | |
3921 | yield(); | |
3922 | goto repeat; | |
3923 | } | |
3924 | ||
3925 | /* found anything to free? */ | |
3926 | if (list_empty(&list)) { | |
3927 | BUG_ON(free != 0); | |
3928 | goto out; | |
3929 | } | |
3930 | ||
3931 | /* now free all selected PAs */ | |
3932 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3933 | ||
3934 | /* remove from object (inode or locality group) */ | |
3935 | spin_lock(pa->pa_obj_lock); | |
3936 | list_del_rcu(&pa->pa_inode_list); | |
3937 | spin_unlock(pa->pa_obj_lock); | |
3938 | ||
cc0fb9ad | 3939 | if (pa->pa_type == MB_GROUP_PA) |
c83617db | 3940 | ext4_mb_release_group_pa(&e4b, pa, ac); |
c9de560d | 3941 | else |
c83617db | 3942 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); |
c9de560d AT |
3943 | |
3944 | list_del(&pa->u.pa_tmp_list); | |
3945 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3946 | } | |
3947 | ||
3948 | out: | |
3949 | ext4_unlock_group(sb, group); | |
c83617db AK |
3950 | if (ac) |
3951 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
3952 | ext4_mb_release_desc(&e4b); |
3953 | put_bh(bitmap_bh); | |
3954 | return free; | |
3955 | } | |
3956 | ||
3957 | /* | |
3958 | * releases all non-used preallocated blocks for given inode | |
3959 | * | |
3960 | * It's important to discard preallocations under i_data_sem | |
3961 | * We don't want another block to be served from the prealloc | |
3962 | * space when we are discarding the inode prealloc space. | |
3963 | * | |
3964 | * FIXME!! Make sure it is valid at all the call sites | |
3965 | */ | |
c2ea3fde | 3966 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
3967 | { |
3968 | struct ext4_inode_info *ei = EXT4_I(inode); | |
3969 | struct super_block *sb = inode->i_sb; | |
3970 | struct buffer_head *bitmap_bh = NULL; | |
3971 | struct ext4_prealloc_space *pa, *tmp; | |
c83617db | 3972 | struct ext4_allocation_context *ac; |
c9de560d AT |
3973 | ext4_group_t group = 0; |
3974 | struct list_head list; | |
3975 | struct ext4_buddy e4b; | |
3976 | int err; | |
3977 | ||
c2ea3fde | 3978 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
3979 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
3980 | return; | |
3981 | } | |
3982 | ||
3983 | mb_debug("discard preallocation for inode %lu\n", inode->i_ino); | |
9bffad1e | 3984 | trace_ext4_discard_preallocations(inode); |
c9de560d AT |
3985 | |
3986 | INIT_LIST_HEAD(&list); | |
3987 | ||
c83617db | 3988 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
9bffad1e TT |
3989 | if (ac) { |
3990 | ac->ac_sb = sb; | |
3991 | ac->ac_inode = inode; | |
3992 | } | |
c9de560d AT |
3993 | repeat: |
3994 | /* first, collect all pa's in the inode */ | |
3995 | spin_lock(&ei->i_prealloc_lock); | |
3996 | while (!list_empty(&ei->i_prealloc_list)) { | |
3997 | pa = list_entry(ei->i_prealloc_list.next, | |
3998 | struct ext4_prealloc_space, pa_inode_list); | |
3999 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
4000 | spin_lock(&pa->pa_lock); | |
4001 | if (atomic_read(&pa->pa_count)) { | |
4002 | /* this shouldn't happen often - nobody should | |
4003 | * use preallocation while we're discarding it */ | |
4004 | spin_unlock(&pa->pa_lock); | |
4005 | spin_unlock(&ei->i_prealloc_lock); | |
4006 | printk(KERN_ERR "uh-oh! used pa while discarding\n"); | |
4007 | WARN_ON(1); | |
4008 | schedule_timeout_uninterruptible(HZ); | |
4009 | goto repeat; | |
4010 | ||
4011 | } | |
4012 | if (pa->pa_deleted == 0) { | |
4013 | pa->pa_deleted = 1; | |
4014 | spin_unlock(&pa->pa_lock); | |
4015 | list_del_rcu(&pa->pa_inode_list); | |
4016 | list_add(&pa->u.pa_tmp_list, &list); | |
4017 | continue; | |
4018 | } | |
4019 | ||
4020 | /* someone is deleting pa right now */ | |
4021 | spin_unlock(&pa->pa_lock); | |
4022 | spin_unlock(&ei->i_prealloc_lock); | |
4023 | ||
4024 | /* we have to wait here because pa_deleted | |
4025 | * doesn't mean pa is already unlinked from | |
4026 | * the list. as we might be called from | |
4027 | * ->clear_inode() the inode will get freed | |
4028 | * and concurrent thread which is unlinking | |
4029 | * pa from inode's list may access already | |
4030 | * freed memory, bad-bad-bad */ | |
4031 | ||
4032 | /* XXX: if this happens too often, we can | |
4033 | * add a flag to force wait only in case | |
4034 | * of ->clear_inode(), but not in case of | |
4035 | * regular truncate */ | |
4036 | schedule_timeout_uninterruptible(HZ); | |
4037 | goto repeat; | |
4038 | } | |
4039 | spin_unlock(&ei->i_prealloc_lock); | |
4040 | ||
4041 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 4042 | BUG_ON(pa->pa_type != MB_INODE_PA); |
c9de560d AT |
4043 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); |
4044 | ||
4045 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c AK |
4046 | if (err) { |
4047 | ext4_error(sb, __func__, "Error in loading buddy " | |
a9df9a49 | 4048 | "information for %u", group); |
ce89f46c AK |
4049 | continue; |
4050 | } | |
c9de560d | 4051 | |
574ca174 | 4052 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 4053 | if (bitmap_bh == NULL) { |
ce89f46c | 4054 | ext4_error(sb, __func__, "Error in reading block " |
a9df9a49 | 4055 | "bitmap for %u", group); |
c9de560d | 4056 | ext4_mb_release_desc(&e4b); |
ce89f46c | 4057 | continue; |
c9de560d AT |
4058 | } |
4059 | ||
4060 | ext4_lock_group(sb, group); | |
4061 | list_del(&pa->pa_group_list); | |
c83617db | 4062 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); |
c9de560d AT |
4063 | ext4_unlock_group(sb, group); |
4064 | ||
4065 | ext4_mb_release_desc(&e4b); | |
4066 | put_bh(bitmap_bh); | |
4067 | ||
4068 | list_del(&pa->u.pa_tmp_list); | |
4069 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4070 | } | |
c83617db AK |
4071 | if (ac) |
4072 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
4073 | } |
4074 | ||
4075 | /* | |
4076 | * finds all preallocated spaces and return blocks being freed to them | |
4077 | * if preallocated space becomes full (no block is used from the space) | |
4078 | * then the function frees space in buddy | |
4079 | * XXX: at the moment, truncate (which is the only way to free blocks) | |
4080 | * discards all preallocations | |
4081 | */ | |
4082 | static void ext4_mb_return_to_preallocation(struct inode *inode, | |
4083 | struct ext4_buddy *e4b, | |
4084 | sector_t block, int count) | |
4085 | { | |
4086 | BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list)); | |
4087 | } | |
4088 | #ifdef MB_DEBUG | |
4089 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
4090 | { | |
4091 | struct super_block *sb = ac->ac_sb; | |
8df9675f | 4092 | ext4_group_t ngroups, i; |
c9de560d AT |
4093 | |
4094 | printk(KERN_ERR "EXT4-fs: Can't allocate:" | |
4095 | " Allocation context details:\n"); | |
4096 | printk(KERN_ERR "EXT4-fs: status %d flags %d\n", | |
4097 | ac->ac_status, ac->ac_flags); | |
4098 | printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, " | |
4099 | "best %lu/%lu/%lu@%lu cr %d\n", | |
4100 | (unsigned long)ac->ac_o_ex.fe_group, | |
4101 | (unsigned long)ac->ac_o_ex.fe_start, | |
4102 | (unsigned long)ac->ac_o_ex.fe_len, | |
4103 | (unsigned long)ac->ac_o_ex.fe_logical, | |
4104 | (unsigned long)ac->ac_g_ex.fe_group, | |
4105 | (unsigned long)ac->ac_g_ex.fe_start, | |
4106 | (unsigned long)ac->ac_g_ex.fe_len, | |
4107 | (unsigned long)ac->ac_g_ex.fe_logical, | |
4108 | (unsigned long)ac->ac_b_ex.fe_group, | |
4109 | (unsigned long)ac->ac_b_ex.fe_start, | |
4110 | (unsigned long)ac->ac_b_ex.fe_len, | |
4111 | (unsigned long)ac->ac_b_ex.fe_logical, | |
4112 | (int)ac->ac_criteria); | |
4113 | printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned, | |
4114 | ac->ac_found); | |
4115 | printk(KERN_ERR "EXT4-fs: groups: \n"); | |
8df9675f TT |
4116 | ngroups = ext4_get_groups_count(sb); |
4117 | for (i = 0; i < ngroups; i++) { | |
c9de560d AT |
4118 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
4119 | struct ext4_prealloc_space *pa; | |
4120 | ext4_grpblk_t start; | |
4121 | struct list_head *cur; | |
4122 | ext4_lock_group(sb, i); | |
4123 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4124 | pa = list_entry(cur, struct ext4_prealloc_space, | |
4125 | pa_group_list); | |
4126 | spin_lock(&pa->pa_lock); | |
4127 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4128 | NULL, &start); | |
4129 | spin_unlock(&pa->pa_lock); | |
4130 | printk(KERN_ERR "PA:%lu:%d:%u \n", i, | |
4131 | start, pa->pa_len); | |
4132 | } | |
60bd63d1 | 4133 | ext4_unlock_group(sb, i); |
c9de560d AT |
4134 | |
4135 | if (grp->bb_free == 0) | |
4136 | continue; | |
4137 | printk(KERN_ERR "%lu: %d/%d \n", | |
4138 | i, grp->bb_free, grp->bb_fragments); | |
4139 | } | |
4140 | printk(KERN_ERR "\n"); | |
4141 | } | |
4142 | #else | |
4143 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
4144 | { | |
4145 | return; | |
4146 | } | |
4147 | #endif | |
4148 | ||
4149 | /* | |
4150 | * We use locality group preallocation for small size file. The size of the | |
4151 | * file is determined by the current size or the resulting size after | |
4152 | * allocation which ever is larger | |
4153 | * | |
b713a5ec | 4154 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
4155 | */ |
4156 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
4157 | { | |
4158 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4159 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
4160 | loff_t size, isize; | |
4161 | ||
4162 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4163 | return; | |
4164 | ||
4165 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; | |
4166 | isize = i_size_read(ac->ac_inode) >> bsbits; | |
4167 | size = max(size, isize); | |
4168 | ||
4169 | /* don't use group allocation for large files */ | |
4170 | if (size >= sbi->s_mb_stream_request) | |
4171 | return; | |
4172 | ||
4173 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
4174 | return; | |
4175 | ||
4176 | BUG_ON(ac->ac_lg != NULL); | |
4177 | /* | |
4178 | * locality group prealloc space are per cpu. The reason for having | |
4179 | * per cpu locality group is to reduce the contention between block | |
4180 | * request from multiple CPUs. | |
4181 | */ | |
730c213c | 4182 | ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id()); |
c9de560d AT |
4183 | |
4184 | /* we're going to use group allocation */ | |
4185 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4186 | ||
4187 | /* serialize all allocations in the group */ | |
4188 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4189 | } | |
4190 | ||
4ddfef7b ES |
4191 | static noinline_for_stack int |
4192 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
4193 | struct ext4_allocation_request *ar) |
4194 | { | |
4195 | struct super_block *sb = ar->inode->i_sb; | |
4196 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4197 | struct ext4_super_block *es = sbi->s_es; | |
4198 | ext4_group_t group; | |
498e5f24 TT |
4199 | unsigned int len; |
4200 | ext4_fsblk_t goal; | |
c9de560d AT |
4201 | ext4_grpblk_t block; |
4202 | ||
4203 | /* we can't allocate > group size */ | |
4204 | len = ar->len; | |
4205 | ||
4206 | /* just a dirty hack to filter too big requests */ | |
4207 | if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10) | |
4208 | len = EXT4_BLOCKS_PER_GROUP(sb) - 10; | |
4209 | ||
4210 | /* start searching from the goal */ | |
4211 | goal = ar->goal; | |
4212 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4213 | goal >= ext4_blocks_count(es)) | |
4214 | goal = le32_to_cpu(es->s_first_data_block); | |
4215 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4216 | ||
4217 | /* set up allocation goals */ | |
4218 | ac->ac_b_ex.fe_logical = ar->logical; | |
4219 | ac->ac_b_ex.fe_group = 0; | |
4220 | ac->ac_b_ex.fe_start = 0; | |
4221 | ac->ac_b_ex.fe_len = 0; | |
4222 | ac->ac_status = AC_STATUS_CONTINUE; | |
4223 | ac->ac_groups_scanned = 0; | |
4224 | ac->ac_ex_scanned = 0; | |
4225 | ac->ac_found = 0; | |
4226 | ac->ac_sb = sb; | |
4227 | ac->ac_inode = ar->inode; | |
4228 | ac->ac_o_ex.fe_logical = ar->logical; | |
4229 | ac->ac_o_ex.fe_group = group; | |
4230 | ac->ac_o_ex.fe_start = block; | |
4231 | ac->ac_o_ex.fe_len = len; | |
4232 | ac->ac_g_ex.fe_logical = ar->logical; | |
4233 | ac->ac_g_ex.fe_group = group; | |
4234 | ac->ac_g_ex.fe_start = block; | |
4235 | ac->ac_g_ex.fe_len = len; | |
4236 | ac->ac_f_ex.fe_len = 0; | |
4237 | ac->ac_flags = ar->flags; | |
4238 | ac->ac_2order = 0; | |
4239 | ac->ac_criteria = 0; | |
4240 | ac->ac_pa = NULL; | |
4241 | ac->ac_bitmap_page = NULL; | |
4242 | ac->ac_buddy_page = NULL; | |
8556e8f3 | 4243 | ac->alloc_semp = NULL; |
c9de560d AT |
4244 | ac->ac_lg = NULL; |
4245 | ||
4246 | /* we have to define context: we'll we work with a file or | |
4247 | * locality group. this is a policy, actually */ | |
4248 | ext4_mb_group_or_file(ac); | |
4249 | ||
4250 | mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " | |
4251 | "left: %u/%u, right %u/%u to %swritable\n", | |
4252 | (unsigned) ar->len, (unsigned) ar->logical, | |
4253 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4254 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4255 | (unsigned) ar->lright, (unsigned) ar->pright, | |
4256 | atomic_read(&ar->inode->i_writecount) ? "" : "non-"); | |
4257 | return 0; | |
4258 | ||
4259 | } | |
4260 | ||
6be2ded1 AK |
4261 | static noinline_for_stack void |
4262 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4263 | struct ext4_locality_group *lg, | |
4264 | int order, int total_entries) | |
4265 | { | |
4266 | ext4_group_t group = 0; | |
4267 | struct ext4_buddy e4b; | |
4268 | struct list_head discard_list; | |
4269 | struct ext4_prealloc_space *pa, *tmp; | |
4270 | struct ext4_allocation_context *ac; | |
4271 | ||
4272 | mb_debug("discard locality group preallocation\n"); | |
4273 | ||
4274 | INIT_LIST_HEAD(&discard_list); | |
4275 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); | |
9bffad1e TT |
4276 | if (ac) |
4277 | ac->ac_sb = sb; | |
6be2ded1 AK |
4278 | |
4279 | spin_lock(&lg->lg_prealloc_lock); | |
4280 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
4281 | pa_inode_list) { | |
4282 | spin_lock(&pa->pa_lock); | |
4283 | if (atomic_read(&pa->pa_count)) { | |
4284 | /* | |
4285 | * This is the pa that we just used | |
4286 | * for block allocation. So don't | |
4287 | * free that | |
4288 | */ | |
4289 | spin_unlock(&pa->pa_lock); | |
4290 | continue; | |
4291 | } | |
4292 | if (pa->pa_deleted) { | |
4293 | spin_unlock(&pa->pa_lock); | |
4294 | continue; | |
4295 | } | |
4296 | /* only lg prealloc space */ | |
cc0fb9ad | 4297 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4298 | |
4299 | /* seems this one can be freed ... */ | |
4300 | pa->pa_deleted = 1; | |
4301 | spin_unlock(&pa->pa_lock); | |
4302 | ||
4303 | list_del_rcu(&pa->pa_inode_list); | |
4304 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4305 | ||
4306 | total_entries--; | |
4307 | if (total_entries <= 5) { | |
4308 | /* | |
4309 | * we want to keep only 5 entries | |
4310 | * allowing it to grow to 8. This | |
4311 | * mak sure we don't call discard | |
4312 | * soon for this list. | |
4313 | */ | |
4314 | break; | |
4315 | } | |
4316 | } | |
4317 | spin_unlock(&lg->lg_prealloc_lock); | |
4318 | ||
4319 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
4320 | ||
4321 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); | |
4322 | if (ext4_mb_load_buddy(sb, group, &e4b)) { | |
4323 | ext4_error(sb, __func__, "Error in loading buddy " | |
a9df9a49 | 4324 | "information for %u", group); |
6be2ded1 AK |
4325 | continue; |
4326 | } | |
4327 | ext4_lock_group(sb, group); | |
4328 | list_del(&pa->pa_group_list); | |
4329 | ext4_mb_release_group_pa(&e4b, pa, ac); | |
4330 | ext4_unlock_group(sb, group); | |
4331 | ||
4332 | ext4_mb_release_desc(&e4b); | |
4333 | list_del(&pa->u.pa_tmp_list); | |
4334 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4335 | } | |
4336 | if (ac) | |
4337 | kmem_cache_free(ext4_ac_cachep, ac); | |
4338 | } | |
4339 | ||
4340 | /* | |
4341 | * We have incremented pa_count. So it cannot be freed at this | |
4342 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4343 | * possible from this lg. That means pa_free cannot be updated. | |
4344 | * | |
4345 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4346 | * which can cause the lg_prealloc_list to be updated. | |
4347 | */ | |
4348 | ||
4349 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4350 | { | |
4351 | int order, added = 0, lg_prealloc_count = 1; | |
4352 | struct super_block *sb = ac->ac_sb; | |
4353 | struct ext4_locality_group *lg = ac->ac_lg; | |
4354 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4355 | ||
4356 | order = fls(pa->pa_free) - 1; | |
4357 | if (order > PREALLOC_TB_SIZE - 1) | |
4358 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4359 | order = PREALLOC_TB_SIZE - 1; | |
4360 | /* Add the prealloc space to lg */ | |
4361 | rcu_read_lock(); | |
4362 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], | |
4363 | pa_inode_list) { | |
4364 | spin_lock(&tmp_pa->pa_lock); | |
4365 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4366 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4367 | continue; |
4368 | } | |
4369 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4370 | /* Add to the tail of the previous entry */ | |
4371 | list_add_tail_rcu(&pa->pa_inode_list, | |
4372 | &tmp_pa->pa_inode_list); | |
4373 | added = 1; | |
4374 | /* | |
4375 | * we want to count the total | |
4376 | * number of entries in the list | |
4377 | */ | |
4378 | } | |
4379 | spin_unlock(&tmp_pa->pa_lock); | |
4380 | lg_prealloc_count++; | |
4381 | } | |
4382 | if (!added) | |
4383 | list_add_tail_rcu(&pa->pa_inode_list, | |
4384 | &lg->lg_prealloc_list[order]); | |
4385 | rcu_read_unlock(); | |
4386 | ||
4387 | /* Now trim the list to be not more than 8 elements */ | |
4388 | if (lg_prealloc_count > 8) { | |
4389 | ext4_mb_discard_lg_preallocations(sb, lg, | |
4390 | order, lg_prealloc_count); | |
4391 | return; | |
4392 | } | |
4393 | return ; | |
4394 | } | |
4395 | ||
c9de560d AT |
4396 | /* |
4397 | * release all resource we used in allocation | |
4398 | */ | |
4399 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4400 | { | |
6be2ded1 AK |
4401 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4402 | if (pa) { | |
cc0fb9ad | 4403 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4404 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 AK |
4405 | spin_lock(&pa->pa_lock); |
4406 | pa->pa_pstart += ac->ac_b_ex.fe_len; | |
4407 | pa->pa_lstart += ac->ac_b_ex.fe_len; | |
4408 | pa->pa_free -= ac->ac_b_ex.fe_len; | |
4409 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4410 | spin_unlock(&pa->pa_lock); | |
c9de560d | 4411 | } |
c9de560d | 4412 | } |
8556e8f3 AK |
4413 | if (ac->alloc_semp) |
4414 | up_read(ac->alloc_semp); | |
ba443916 AK |
4415 | if (pa) { |
4416 | /* | |
4417 | * We want to add the pa to the right bucket. | |
4418 | * Remove it from the list and while adding | |
4419 | * make sure the list to which we are adding | |
4420 | * doesn't grow big. We need to release | |
4421 | * alloc_semp before calling ext4_mb_add_n_trim() | |
4422 | */ | |
cc0fb9ad | 4423 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { |
ba443916 AK |
4424 | spin_lock(pa->pa_obj_lock); |
4425 | list_del_rcu(&pa->pa_inode_list); | |
4426 | spin_unlock(pa->pa_obj_lock); | |
4427 | ext4_mb_add_n_trim(ac); | |
4428 | } | |
4429 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4430 | } | |
c9de560d AT |
4431 | if (ac->ac_bitmap_page) |
4432 | page_cache_release(ac->ac_bitmap_page); | |
4433 | if (ac->ac_buddy_page) | |
4434 | page_cache_release(ac->ac_buddy_page); | |
4435 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
4436 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4437 | ext4_mb_collect_stats(ac); | |
4438 | return 0; | |
4439 | } | |
4440 | ||
4441 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4442 | { | |
8df9675f | 4443 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4444 | int ret; |
4445 | int freed = 0; | |
4446 | ||
9bffad1e | 4447 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4448 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4449 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4450 | freed += ret; | |
4451 | needed -= ret; | |
4452 | } | |
4453 | ||
4454 | return freed; | |
4455 | } | |
4456 | ||
4457 | /* | |
4458 | * Main entry point into mballoc to allocate blocks | |
4459 | * it tries to use preallocation first, then falls back | |
4460 | * to usual allocation | |
4461 | */ | |
4462 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
4463 | struct ext4_allocation_request *ar, int *errp) | |
4464 | { | |
6bc6e63f | 4465 | int freed; |
256bdb49 | 4466 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4467 | struct ext4_sb_info *sbi; |
4468 | struct super_block *sb; | |
4469 | ext4_fsblk_t block = 0; | |
60e58e0f | 4470 | unsigned int inquota = 0; |
498e5f24 | 4471 | unsigned int reserv_blks = 0; |
c9de560d AT |
4472 | |
4473 | sb = ar->inode->i_sb; | |
4474 | sbi = EXT4_SB(sb); | |
4475 | ||
9bffad1e | 4476 | trace_ext4_request_blocks(ar); |
ba80b101 | 4477 | |
60e58e0f MC |
4478 | /* |
4479 | * For delayed allocation, we could skip the ENOSPC and | |
4480 | * EDQUOT check, as blocks and quotas have been already | |
4481 | * reserved when data being copied into pagecache. | |
4482 | */ | |
4483 | if (EXT4_I(ar->inode)->i_delalloc_reserved_flag) | |
4484 | ar->flags |= EXT4_MB_DELALLOC_RESERVED; | |
4485 | else { | |
4486 | /* Without delayed allocation we need to verify | |
4487 | * there is enough free blocks to do block allocation | |
4488 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4489 | */ |
030ba6bc AK |
4490 | while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) { |
4491 | /* let others to free the space */ | |
4492 | yield(); | |
4493 | ar->len = ar->len >> 1; | |
4494 | } | |
4495 | if (!ar->len) { | |
a30d542a AK |
4496 | *errp = -ENOSPC; |
4497 | return 0; | |
4498 | } | |
6bc6e63f | 4499 | reserv_blks = ar->len; |
a269eb18 | 4500 | while (ar->len && vfs_dq_alloc_block(ar->inode, ar->len)) { |
60e58e0f MC |
4501 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; |
4502 | ar->len--; | |
4503 | } | |
4504 | inquota = ar->len; | |
4505 | if (ar->len == 0) { | |
4506 | *errp = -EDQUOT; | |
4507 | goto out3; | |
4508 | } | |
07031431 | 4509 | } |
d2a17637 | 4510 | |
256bdb49 | 4511 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
9bffad1e TT |
4512 | if (ac) { |
4513 | ac->ac_sb = sb; | |
4514 | ac->ac_inode = ar->inode; | |
4515 | } else { | |
363d4251 | 4516 | ar->len = 0; |
256bdb49 | 4517 | *errp = -ENOMEM; |
363d4251 | 4518 | goto out1; |
256bdb49 ES |
4519 | } |
4520 | ||
256bdb49 | 4521 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4522 | if (*errp) { |
4523 | ar->len = 0; | |
363d4251 | 4524 | goto out2; |
c9de560d AT |
4525 | } |
4526 | ||
256bdb49 ES |
4527 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4528 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4529 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4530 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4531 | repeat: |
4532 | /* allocate space in core */ | |
256bdb49 | 4533 | ext4_mb_regular_allocator(ac); |
c9de560d AT |
4534 | |
4535 | /* as we've just preallocated more space than | |
4536 | * user requested orinally, we store allocated | |
4537 | * space in a special descriptor */ | |
256bdb49 ES |
4538 | if (ac->ac_status == AC_STATUS_FOUND && |
4539 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
4540 | ext4_mb_new_preallocation(ac); | |
c9de560d | 4541 | } |
256bdb49 | 4542 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
6bc6e63f | 4543 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks); |
519deca0 | 4544 | if (*errp == -EAGAIN) { |
8556e8f3 AK |
4545 | /* |
4546 | * drop the reference that we took | |
4547 | * in ext4_mb_use_best_found | |
4548 | */ | |
4549 | ext4_mb_release_context(ac); | |
519deca0 AK |
4550 | ac->ac_b_ex.fe_group = 0; |
4551 | ac->ac_b_ex.fe_start = 0; | |
4552 | ac->ac_b_ex.fe_len = 0; | |
4553 | ac->ac_status = AC_STATUS_CONTINUE; | |
4554 | goto repeat; | |
4555 | } else if (*errp) { | |
4556 | ac->ac_b_ex.fe_len = 0; | |
4557 | ar->len = 0; | |
4558 | ext4_mb_show_ac(ac); | |
4559 | } else { | |
4560 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4561 | ar->len = ac->ac_b_ex.fe_len; | |
4562 | } | |
c9de560d | 4563 | } else { |
256bdb49 | 4564 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4565 | if (freed) |
4566 | goto repeat; | |
4567 | *errp = -ENOSPC; | |
256bdb49 | 4568 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4569 | ar->len = 0; |
256bdb49 | 4570 | ext4_mb_show_ac(ac); |
c9de560d AT |
4571 | } |
4572 | ||
256bdb49 | 4573 | ext4_mb_release_context(ac); |
c9de560d | 4574 | |
363d4251 SF |
4575 | out2: |
4576 | kmem_cache_free(ext4_ac_cachep, ac); | |
4577 | out1: | |
60e58e0f | 4578 | if (inquota && ar->len < inquota) |
a269eb18 | 4579 | vfs_dq_free_block(ar->inode, inquota - ar->len); |
0087d9fb AK |
4580 | out3: |
4581 | if (!ar->len) { | |
4582 | if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) | |
4583 | /* release all the reserved blocks if non delalloc */ | |
4584 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, | |
4585 | reserv_blks); | |
4586 | } | |
c9de560d | 4587 | |
9bffad1e | 4588 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4589 | |
c9de560d AT |
4590 | return block; |
4591 | } | |
c9de560d | 4592 | |
c894058d AK |
4593 | /* |
4594 | * We can merge two free data extents only if the physical blocks | |
4595 | * are contiguous, AND the extents were freed by the same transaction, | |
4596 | * AND the blocks are associated with the same group. | |
4597 | */ | |
4598 | static int can_merge(struct ext4_free_data *entry1, | |
4599 | struct ext4_free_data *entry2) | |
4600 | { | |
4601 | if ((entry1->t_tid == entry2->t_tid) && | |
4602 | (entry1->group == entry2->group) && | |
4603 | ((entry1->start_blk + entry1->count) == entry2->start_blk)) | |
4604 | return 1; | |
4605 | return 0; | |
4606 | } | |
4607 | ||
4ddfef7b ES |
4608 | static noinline_for_stack int |
4609 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 4610 | struct ext4_free_data *new_entry) |
c9de560d | 4611 | { |
7a2fcbf7 AK |
4612 | ext4_grpblk_t block; |
4613 | struct ext4_free_data *entry; | |
c9de560d AT |
4614 | struct ext4_group_info *db = e4b->bd_info; |
4615 | struct super_block *sb = e4b->bd_sb; | |
4616 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4617 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
4618 | struct rb_node *parent = NULL, *new_node; | |
4619 | ||
0390131b | 4620 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4621 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4622 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4623 | ||
c894058d | 4624 | new_node = &new_entry->node; |
7a2fcbf7 | 4625 | block = new_entry->start_blk; |
c894058d | 4626 | |
c894058d AK |
4627 | if (!*n) { |
4628 | /* first free block exent. We need to | |
4629 | protect buddy cache from being freed, | |
4630 | * otherwise we'll refresh it from | |
4631 | * on-disk bitmap and lose not-yet-available | |
4632 | * blocks */ | |
4633 | page_cache_get(e4b->bd_buddy_page); | |
4634 | page_cache_get(e4b->bd_bitmap_page); | |
4635 | } | |
4636 | while (*n) { | |
4637 | parent = *n; | |
4638 | entry = rb_entry(parent, struct ext4_free_data, node); | |
4639 | if (block < entry->start_blk) | |
4640 | n = &(*n)->rb_left; | |
4641 | else if (block >= (entry->start_blk + entry->count)) | |
4642 | n = &(*n)->rb_right; | |
4643 | else { | |
5d1b1b3f AK |
4644 | ext4_grp_locked_error(sb, e4b->bd_group, __func__, |
4645 | "Double free of blocks %d (%d %d)", | |
4646 | block, entry->start_blk, entry->count); | |
c894058d | 4647 | return 0; |
c9de560d | 4648 | } |
c894058d | 4649 | } |
c9de560d | 4650 | |
c894058d AK |
4651 | rb_link_node(new_node, parent, n); |
4652 | rb_insert_color(new_node, &db->bb_free_root); | |
4653 | ||
4654 | /* Now try to see the extent can be merged to left and right */ | |
4655 | node = rb_prev(new_node); | |
4656 | if (node) { | |
4657 | entry = rb_entry(node, struct ext4_free_data, node); | |
4658 | if (can_merge(entry, new_entry)) { | |
4659 | new_entry->start_blk = entry->start_blk; | |
4660 | new_entry->count += entry->count; | |
4661 | rb_erase(node, &(db->bb_free_root)); | |
4662 | spin_lock(&sbi->s_md_lock); | |
4663 | list_del(&entry->list); | |
4664 | spin_unlock(&sbi->s_md_lock); | |
4665 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
c9de560d | 4666 | } |
c894058d | 4667 | } |
c9de560d | 4668 | |
c894058d AK |
4669 | node = rb_next(new_node); |
4670 | if (node) { | |
4671 | entry = rb_entry(node, struct ext4_free_data, node); | |
4672 | if (can_merge(new_entry, entry)) { | |
4673 | new_entry->count += entry->count; | |
4674 | rb_erase(node, &(db->bb_free_root)); | |
4675 | spin_lock(&sbi->s_md_lock); | |
4676 | list_del(&entry->list); | |
4677 | spin_unlock(&sbi->s_md_lock); | |
4678 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
c9de560d AT |
4679 | } |
4680 | } | |
3e624fc7 | 4681 | /* Add the extent to transaction's private list */ |
c894058d | 4682 | spin_lock(&sbi->s_md_lock); |
3e624fc7 | 4683 | list_add(&new_entry->list, &handle->h_transaction->t_private_list); |
c894058d | 4684 | spin_unlock(&sbi->s_md_lock); |
c9de560d AT |
4685 | return 0; |
4686 | } | |
4687 | ||
4688 | /* | |
4689 | * Main entry point into mballoc to free blocks | |
4690 | */ | |
4691 | void ext4_mb_free_blocks(handle_t *handle, struct inode *inode, | |
0610b6e9 | 4692 | ext4_fsblk_t block, unsigned long count, |
c9de560d AT |
4693 | int metadata, unsigned long *freed) |
4694 | { | |
26346ff6 | 4695 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4696 | struct super_block *sb = inode->i_sb; |
256bdb49 | 4697 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4698 | struct ext4_group_desc *gdp; |
4699 | struct ext4_super_block *es; | |
498e5f24 | 4700 | unsigned int overflow; |
c9de560d AT |
4701 | ext4_grpblk_t bit; |
4702 | struct buffer_head *gd_bh; | |
4703 | ext4_group_t block_group; | |
4704 | struct ext4_sb_info *sbi; | |
4705 | struct ext4_buddy e4b; | |
4706 | int err = 0; | |
4707 | int ret; | |
4708 | ||
4709 | *freed = 0; | |
4710 | ||
c9de560d AT |
4711 | sbi = EXT4_SB(sb); |
4712 | es = EXT4_SB(sb)->s_es; | |
4713 | if (block < le32_to_cpu(es->s_first_data_block) || | |
4714 | block + count < block || | |
4715 | block + count > ext4_blocks_count(es)) { | |
46e665e9 | 4716 | ext4_error(sb, __func__, |
c9de560d | 4717 | "Freeing blocks not in datazone - " |
0610b6e9 | 4718 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
4719 | goto error_return; |
4720 | } | |
4721 | ||
0610b6e9 | 4722 | ext4_debug("freeing block %llu\n", block); |
9bffad1e | 4723 | trace_ext4_free_blocks(inode, block, count, metadata); |
c9de560d | 4724 | |
256bdb49 ES |
4725 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
4726 | if (ac) { | |
4727 | ac->ac_op = EXT4_MB_HISTORY_FREE; | |
4728 | ac->ac_inode = inode; | |
4729 | ac->ac_sb = sb; | |
4730 | } | |
c9de560d AT |
4731 | |
4732 | do_more: | |
4733 | overflow = 0; | |
4734 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4735 | ||
4736 | /* | |
4737 | * Check to see if we are freeing blocks across a group | |
4738 | * boundary. | |
4739 | */ | |
4740 | if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { | |
4741 | overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb); | |
4742 | count -= overflow; | |
4743 | } | |
574ca174 | 4744 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
ce89f46c AK |
4745 | if (!bitmap_bh) { |
4746 | err = -EIO; | |
c9de560d | 4747 | goto error_return; |
ce89f46c | 4748 | } |
c9de560d | 4749 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4750 | if (!gdp) { |
4751 | err = -EIO; | |
c9de560d | 4752 | goto error_return; |
ce89f46c | 4753 | } |
c9de560d AT |
4754 | |
4755 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4756 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4757 | in_range(block, ext4_inode_table(sb, gdp), | |
4758 | EXT4_SB(sb)->s_itb_per_group) || | |
4759 | in_range(block + count - 1, ext4_inode_table(sb, gdp), | |
4760 | EXT4_SB(sb)->s_itb_per_group)) { | |
4761 | ||
46e665e9 | 4762 | ext4_error(sb, __func__, |
c9de560d | 4763 | "Freeing blocks in system zone - " |
0610b6e9 | 4764 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
4765 | /* err = 0. ext4_std_error should be a no op */ |
4766 | goto error_return; | |
c9de560d AT |
4767 | } |
4768 | ||
4769 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4770 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4771 | if (err) | |
4772 | goto error_return; | |
4773 | ||
4774 | /* | |
4775 | * We are about to modify some metadata. Call the journal APIs | |
4776 | * to unshare ->b_data if a currently-committing transaction is | |
4777 | * using it | |
4778 | */ | |
4779 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4780 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4781 | if (err) | |
4782 | goto error_return; | |
c9de560d AT |
4783 | #ifdef AGGRESSIVE_CHECK |
4784 | { | |
4785 | int i; | |
4786 | for (i = 0; i < count; i++) | |
4787 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); | |
4788 | } | |
4789 | #endif | |
256bdb49 ES |
4790 | if (ac) { |
4791 | ac->ac_b_ex.fe_group = block_group; | |
4792 | ac->ac_b_ex.fe_start = bit; | |
4793 | ac->ac_b_ex.fe_len = count; | |
4794 | ext4_mb_store_history(ac); | |
4795 | } | |
c9de560d | 4796 | |
920313a7 AK |
4797 | err = ext4_mb_load_buddy(sb, block_group, &e4b); |
4798 | if (err) | |
4799 | goto error_return; | |
0390131b | 4800 | if (metadata && ext4_handle_valid(handle)) { |
7a2fcbf7 AK |
4801 | struct ext4_free_data *new_entry; |
4802 | /* | |
4803 | * blocks being freed are metadata. these blocks shouldn't | |
4804 | * be used until this transaction is committed | |
4805 | */ | |
4806 | new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS); | |
4807 | new_entry->start_blk = bit; | |
4808 | new_entry->group = block_group; | |
4809 | new_entry->count = count; | |
4810 | new_entry->t_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 4811 | |
7a2fcbf7 | 4812 | ext4_lock_group(sb, block_group); |
955ce5f5 | 4813 | mb_clear_bits(bitmap_bh->b_data, bit, count); |
7a2fcbf7 | 4814 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 4815 | } else { |
7a2fcbf7 AK |
4816 | /* need to update group_info->bb_free and bitmap |
4817 | * with group lock held. generate_buddy look at | |
4818 | * them with group lock_held | |
4819 | */ | |
955ce5f5 AK |
4820 | ext4_lock_group(sb, block_group); |
4821 | mb_clear_bits(bitmap_bh->b_data, bit, count); | |
7e5a8cdd | 4822 | mb_free_blocks(inode, &e4b, bit, count); |
c9de560d | 4823 | ext4_mb_return_to_preallocation(inode, &e4b, block, count); |
c9de560d AT |
4824 | } |
4825 | ||
560671a0 AK |
4826 | ret = ext4_free_blks_count(sb, gdp) + count; |
4827 | ext4_free_blks_set(sb, gdp, ret); | |
c9de560d | 4828 | gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); |
955ce5f5 | 4829 | ext4_unlock_group(sb, block_group); |
c9de560d AT |
4830 | percpu_counter_add(&sbi->s_freeblocks_counter, count); |
4831 | ||
772cb7c8 JS |
4832 | if (sbi->s_log_groups_per_flex) { |
4833 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
9f24e420 | 4834 | atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks); |
772cb7c8 JS |
4835 | } |
4836 | ||
c9de560d AT |
4837 | ext4_mb_release_desc(&e4b); |
4838 | ||
4839 | *freed += count; | |
4840 | ||
7a2fcbf7 AK |
4841 | /* We dirtied the bitmap block */ |
4842 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4843 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4844 | ||
c9de560d AT |
4845 | /* And the group descriptor block */ |
4846 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 4847 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
4848 | if (!err) |
4849 | err = ret; | |
4850 | ||
4851 | if (overflow && !err) { | |
4852 | block += count; | |
4853 | count = overflow; | |
4854 | put_bh(bitmap_bh); | |
4855 | goto do_more; | |
4856 | } | |
4857 | sb->s_dirt = 1; | |
4858 | error_return: | |
4859 | brelse(bitmap_bh); | |
4860 | ext4_std_error(sb, err); | |
256bdb49 ES |
4861 | if (ac) |
4862 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
4863 | return; |
4864 | } |