]>
Commit | Line | Data |
---|---|---|
c9de560d AT |
1 | /* |
2 | * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com | |
3 | * Written by Alex Tomas <alex@clusterfs.com> | |
4 | * | |
5 | * This program is free software; you can redistribute it and/or modify | |
6 | * it under the terms of the GNU General Public License version 2 as | |
7 | * published by the Free Software Foundation. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public Licens | |
15 | * along with this program; if not, write to the Free Software | |
16 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111- | |
17 | */ | |
18 | ||
19 | ||
20 | /* | |
21 | * mballoc.c contains the multiblocks allocation routines | |
22 | */ | |
23 | ||
8f6e39a7 | 24 | #include "mballoc.h" |
6ba495e9 | 25 | #include <linux/debugfs.h> |
9bffad1e TT |
26 | #include <trace/events/ext4.h> |
27 | ||
c9de560d AT |
28 | /* |
29 | * MUSTDO: | |
30 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
31 | * - search for metadata in few groups | |
32 | * | |
33 | * TODO v4: | |
34 | * - normalization should take into account whether file is still open | |
35 | * - discard preallocations if no free space left (policy?) | |
36 | * - don't normalize tails | |
37 | * - quota | |
38 | * - reservation for superuser | |
39 | * | |
40 | * TODO v3: | |
41 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
42 | * - track min/max extents in each group for better group selection | |
43 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
44 | * - tree of groups sorted by number of free blocks | |
45 | * - error handling | |
46 | */ | |
47 | ||
48 | /* | |
49 | * The allocation request involve request for multiple number of blocks | |
50 | * near to the goal(block) value specified. | |
51 | * | |
b713a5ec TT |
52 | * During initialization phase of the allocator we decide to use the |
53 | * group preallocation or inode preallocation depending on the size of | |
54 | * the file. The size of the file could be the resulting file size we | |
55 | * would have after allocation, or the current file size, which ever | |
56 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
57 | * select to use the group preallocation. The default value of | |
58 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
59 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
60 | * terms of number of blocks. | |
c9de560d AT |
61 | * |
62 | * The main motivation for having small file use group preallocation is to | |
b713a5ec | 63 | * ensure that we have small files closer together on the disk. |
c9de560d | 64 | * |
b713a5ec TT |
65 | * First stage the allocator looks at the inode prealloc list, |
66 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
67 | * spaces for this particular inode. The inode prealloc space is | |
68 | * represented as: | |
c9de560d AT |
69 | * |
70 | * pa_lstart -> the logical start block for this prealloc space | |
71 | * pa_pstart -> the physical start block for this prealloc space | |
72 | * pa_len -> lenght for this prealloc space | |
73 | * pa_free -> free space available in this prealloc space | |
74 | * | |
75 | * The inode preallocation space is used looking at the _logical_ start | |
76 | * block. If only the logical file block falls within the range of prealloc | |
77 | * space we will consume the particular prealloc space. This make sure that | |
78 | * that the we have contiguous physical blocks representing the file blocks | |
79 | * | |
80 | * The important thing to be noted in case of inode prealloc space is that | |
81 | * we don't modify the values associated to inode prealloc space except | |
82 | * pa_free. | |
83 | * | |
84 | * If we are not able to find blocks in the inode prealloc space and if we | |
85 | * have the group allocation flag set then we look at the locality group | |
86 | * prealloc space. These are per CPU prealloc list repreasented as | |
87 | * | |
88 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
89 | * | |
90 | * The reason for having a per cpu locality group is to reduce the contention | |
91 | * between CPUs. It is possible to get scheduled at this point. | |
92 | * | |
93 | * The locality group prealloc space is used looking at whether we have | |
94 | * enough free space (pa_free) withing the prealloc space. | |
95 | * | |
96 | * If we can't allocate blocks via inode prealloc or/and locality group | |
97 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
98 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
99 | * mapped to the buddy and bitmap information regarding different | |
100 | * groups. The buddy information is attached to buddy cache inode so that | |
101 | * we can access them through the page cache. The information regarding | |
102 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
103 | * block bitmap and buddy information. The information are stored in the | |
104 | * inode as: | |
105 | * | |
106 | * { page } | |
c3a326a6 | 107 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
108 | * |
109 | * | |
110 | * one block each for bitmap and buddy information. So for each group we | |
111 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / | |
112 | * blocksize) blocks. So it can have information regarding groups_per_page | |
113 | * which is blocks_per_page/2 | |
114 | * | |
115 | * The buddy cache inode is not stored on disk. The inode is thrown | |
116 | * away when the filesystem is unmounted. | |
117 | * | |
118 | * We look for count number of blocks in the buddy cache. If we were able | |
119 | * to locate that many free blocks we return with additional information | |
120 | * regarding rest of the contiguous physical block available | |
121 | * | |
122 | * Before allocating blocks via buddy cache we normalize the request | |
123 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
124 | * blocks that we get after allocation is added to the respective prealloc | |
125 | * list. In case of inode preallocation we follow a list of heuristics | |
126 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
127 | * we are doing a group prealloc we try to normalize the request to | |
b713a5ec | 128 | * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is |
c9de560d | 129 | * 512 blocks. This can be tuned via |
b713a5ec | 130 | * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in |
c9de560d AT |
131 | * terms of number of blocks. If we have mounted the file system with -O |
132 | * stripe=<value> option the group prealloc request is normalized to the | |
133 | * stripe value (sbi->s_stripe) | |
134 | * | |
b713a5ec | 135 | * The regular allocator(using the buddy cache) supports few tunables. |
c9de560d | 136 | * |
b713a5ec TT |
137 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
138 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
139 | * /sys/fs/ext4/<partition>/mb_order2_req | |
c9de560d | 140 | * |
b713a5ec | 141 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
142 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
143 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 144 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
af901ca1 | 145 | * stripe size (sbi->s_stripe), we try to search for contiguous block in |
b713a5ec TT |
146 | * stripe size. This should result in better allocation on RAID setups. If |
147 | * not, we search in the specific group using bitmap for best extents. The | |
148 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 149 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 150 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
151 | * best extent in the found extents. Searching for the blocks starts with |
152 | * the group specified as the goal value in allocation context via | |
153 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
154 | * can used for allocation. ext4_mb_good_group explains how the groups are | |
155 | * checked. | |
156 | * | |
157 | * Both the prealloc space are getting populated as above. So for the first | |
158 | * request we will hit the buddy cache which will result in this prealloc | |
159 | * space getting filled. The prealloc space is then later used for the | |
160 | * subsequent request. | |
161 | */ | |
162 | ||
163 | /* | |
164 | * mballoc operates on the following data: | |
165 | * - on-disk bitmap | |
166 | * - in-core buddy (actually includes buddy and bitmap) | |
167 | * - preallocation descriptors (PAs) | |
168 | * | |
169 | * there are two types of preallocations: | |
170 | * - inode | |
171 | * assiged to specific inode and can be used for this inode only. | |
172 | * it describes part of inode's space preallocated to specific | |
173 | * physical blocks. any block from that preallocated can be used | |
174 | * independent. the descriptor just tracks number of blocks left | |
175 | * unused. so, before taking some block from descriptor, one must | |
176 | * make sure corresponded logical block isn't allocated yet. this | |
177 | * also means that freeing any block within descriptor's range | |
178 | * must discard all preallocated blocks. | |
179 | * - locality group | |
180 | * assigned to specific locality group which does not translate to | |
181 | * permanent set of inodes: inode can join and leave group. space | |
182 | * from this type of preallocation can be used for any inode. thus | |
183 | * it's consumed from the beginning to the end. | |
184 | * | |
185 | * relation between them can be expressed as: | |
186 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
187 | * | |
188 | * this mean blocks mballoc considers used are: | |
189 | * - allocated blocks (persistent) | |
190 | * - preallocated blocks (non-persistent) | |
191 | * | |
192 | * consistency in mballoc world means that at any time a block is either | |
193 | * free or used in ALL structures. notice: "any time" should not be read | |
194 | * literally -- time is discrete and delimited by locks. | |
195 | * | |
196 | * to keep it simple, we don't use block numbers, instead we count number of | |
197 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
198 | * | |
199 | * all operations can be expressed as: | |
200 | * - init buddy: buddy = on-disk + PAs | |
201 | * - new PA: buddy += N; PA = N | |
202 | * - use inode PA: on-disk += N; PA -= N | |
203 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
204 | * - use locality group PA on-disk += N; PA -= N | |
205 | * - discard locality group PA buddy -= PA; PA = 0 | |
206 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
207 | * is used in real operation because we can't know actual used | |
208 | * bits from PA, only from on-disk bitmap | |
209 | * | |
210 | * if we follow this strict logic, then all operations above should be atomic. | |
211 | * given some of them can block, we'd have to use something like semaphores | |
212 | * killing performance on high-end SMP hardware. let's try to relax it using | |
213 | * the following knowledge: | |
214 | * 1) if buddy is referenced, it's already initialized | |
215 | * 2) while block is used in buddy and the buddy is referenced, | |
216 | * nobody can re-allocate that block | |
217 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
218 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
219 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
220 | * block | |
221 | * | |
222 | * so, now we're building a concurrency table: | |
223 | * - init buddy vs. | |
224 | * - new PA | |
225 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
226 | * until PA is linked to allocation group to avoid concurrent buddy init | |
227 | * - use inode PA | |
228 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
229 | * given (3) we care that PA-=N operation doesn't interfere with init | |
230 | * - discard inode PA | |
231 | * the simplest way would be to have buddy initialized by the discard | |
232 | * - use locality group PA | |
233 | * again PA-=N must be serialized with init | |
234 | * - discard locality group PA | |
235 | * the simplest way would be to have buddy initialized by the discard | |
236 | * - new PA vs. | |
237 | * - use inode PA | |
238 | * i_data_sem serializes them | |
239 | * - discard inode PA | |
240 | * discard process must wait until PA isn't used by another process | |
241 | * - use locality group PA | |
242 | * some mutex should serialize them | |
243 | * - discard locality group PA | |
244 | * discard process must wait until PA isn't used by another process | |
245 | * - use inode PA | |
246 | * - use inode PA | |
247 | * i_data_sem or another mutex should serializes them | |
248 | * - discard inode PA | |
249 | * discard process must wait until PA isn't used by another process | |
250 | * - use locality group PA | |
251 | * nothing wrong here -- they're different PAs covering different blocks | |
252 | * - discard locality group PA | |
253 | * discard process must wait until PA isn't used by another process | |
254 | * | |
255 | * now we're ready to make few consequences: | |
256 | * - PA is referenced and while it is no discard is possible | |
257 | * - PA is referenced until block isn't marked in on-disk bitmap | |
258 | * - PA changes only after on-disk bitmap | |
259 | * - discard must not compete with init. either init is done before | |
260 | * any discard or they're serialized somehow | |
261 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
262 | * | |
263 | * a special case when we've used PA to emptiness. no need to modify buddy | |
264 | * in this case, but we should care about concurrent init | |
265 | * | |
266 | */ | |
267 | ||
268 | /* | |
269 | * Logic in few words: | |
270 | * | |
271 | * - allocation: | |
272 | * load group | |
273 | * find blocks | |
274 | * mark bits in on-disk bitmap | |
275 | * release group | |
276 | * | |
277 | * - use preallocation: | |
278 | * find proper PA (per-inode or group) | |
279 | * load group | |
280 | * mark bits in on-disk bitmap | |
281 | * release group | |
282 | * release PA | |
283 | * | |
284 | * - free: | |
285 | * load group | |
286 | * mark bits in on-disk bitmap | |
287 | * release group | |
288 | * | |
289 | * - discard preallocations in group: | |
290 | * mark PAs deleted | |
291 | * move them onto local list | |
292 | * load on-disk bitmap | |
293 | * load group | |
294 | * remove PA from object (inode or locality group) | |
295 | * mark free blocks in-core | |
296 | * | |
297 | * - discard inode's preallocations: | |
298 | */ | |
299 | ||
300 | /* | |
301 | * Locking rules | |
302 | * | |
303 | * Locks: | |
304 | * - bitlock on a group (group) | |
305 | * - object (inode/locality) (object) | |
306 | * - per-pa lock (pa) | |
307 | * | |
308 | * Paths: | |
309 | * - new pa | |
310 | * object | |
311 | * group | |
312 | * | |
313 | * - find and use pa: | |
314 | * pa | |
315 | * | |
316 | * - release consumed pa: | |
317 | * pa | |
318 | * group | |
319 | * object | |
320 | * | |
321 | * - generate in-core bitmap: | |
322 | * group | |
323 | * pa | |
324 | * | |
325 | * - discard all for given object (inode, locality group): | |
326 | * object | |
327 | * pa | |
328 | * group | |
329 | * | |
330 | * - discard all for given group: | |
331 | * group | |
332 | * pa | |
333 | * group | |
334 | * object | |
335 | * | |
336 | */ | |
c3a326a6 AK |
337 | static struct kmem_cache *ext4_pspace_cachep; |
338 | static struct kmem_cache *ext4_ac_cachep; | |
339 | static struct kmem_cache *ext4_free_ext_cachep; | |
340 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
341 | ext4_group_t group); | |
7a2fcbf7 AK |
342 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
343 | ext4_group_t group); | |
c3a326a6 AK |
344 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn); |
345 | ||
ffad0a44 AK |
346 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
347 | { | |
c9de560d | 348 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
349 | *bit += ((unsigned long) addr & 7UL) << 3; |
350 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 351 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
352 | *bit += ((unsigned long) addr & 3UL) << 3; |
353 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
354 | #else |
355 | #error "how many bits you are?!" | |
356 | #endif | |
ffad0a44 AK |
357 | return addr; |
358 | } | |
c9de560d AT |
359 | |
360 | static inline int mb_test_bit(int bit, void *addr) | |
361 | { | |
362 | /* | |
363 | * ext4_test_bit on architecture like powerpc | |
364 | * needs unsigned long aligned address | |
365 | */ | |
ffad0a44 | 366 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
367 | return ext4_test_bit(bit, addr); |
368 | } | |
369 | ||
370 | static inline void mb_set_bit(int bit, void *addr) | |
371 | { | |
ffad0a44 | 372 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
373 | ext4_set_bit(bit, addr); |
374 | } | |
375 | ||
c9de560d AT |
376 | static inline void mb_clear_bit(int bit, void *addr) |
377 | { | |
ffad0a44 | 378 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
379 | ext4_clear_bit(bit, addr); |
380 | } | |
381 | ||
ffad0a44 AK |
382 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
383 | { | |
e7dfb246 | 384 | int fix = 0, ret, tmpmax; |
ffad0a44 | 385 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 386 | tmpmax = max + fix; |
ffad0a44 AK |
387 | start += fix; |
388 | ||
e7dfb246 AK |
389 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
390 | if (ret > max) | |
391 | return max; | |
392 | return ret; | |
ffad0a44 AK |
393 | } |
394 | ||
395 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
396 | { | |
e7dfb246 | 397 | int fix = 0, ret, tmpmax; |
ffad0a44 | 398 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 399 | tmpmax = max + fix; |
ffad0a44 AK |
400 | start += fix; |
401 | ||
e7dfb246 AK |
402 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
403 | if (ret > max) | |
404 | return max; | |
405 | return ret; | |
ffad0a44 AK |
406 | } |
407 | ||
c9de560d AT |
408 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
409 | { | |
410 | char *bb; | |
411 | ||
c9de560d AT |
412 | BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); |
413 | BUG_ON(max == NULL); | |
414 | ||
415 | if (order > e4b->bd_blkbits + 1) { | |
416 | *max = 0; | |
417 | return NULL; | |
418 | } | |
419 | ||
420 | /* at order 0 we see each particular block */ | |
421 | *max = 1 << (e4b->bd_blkbits + 3); | |
422 | if (order == 0) | |
423 | return EXT4_MB_BITMAP(e4b); | |
424 | ||
425 | bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; | |
426 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; | |
427 | ||
428 | return bb; | |
429 | } | |
430 | ||
431 | #ifdef DOUBLE_CHECK | |
432 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
433 | int first, int count) | |
434 | { | |
435 | int i; | |
436 | struct super_block *sb = e4b->bd_sb; | |
437 | ||
438 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
439 | return; | |
bc8e6740 | 440 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
441 | for (i = 0; i < count; i++) { |
442 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
443 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
444 | |
445 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
c9de560d | 446 | blocknr += first + i; |
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, | |
a36b4498 | 625 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
626 | struct ext4_group_info *grp) |
627 | { | |
628 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
629 | ext4_grpblk_t min; |
630 | ext4_grpblk_t max; | |
631 | ext4_grpblk_t chunk; | |
c9de560d AT |
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 | ||
089ceecc ES |
660 | static noinline_for_stack |
661 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
662 | void *buddy, void *bitmap, ext4_group_t group) |
663 | { | |
664 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
a36b4498 ES |
665 | ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb); |
666 | ext4_grpblk_t i = 0; | |
667 | ext4_grpblk_t first; | |
668 | ext4_grpblk_t len; | |
c9de560d AT |
669 | unsigned free = 0; |
670 | unsigned fragments = 0; | |
671 | unsigned long long period = get_cycles(); | |
672 | ||
673 | /* initialize buddy from bitmap which is aggregation | |
674 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 675 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
676 | grp->bb_first_free = i; |
677 | while (i < max) { | |
678 | fragments++; | |
679 | first = i; | |
ffad0a44 | 680 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
681 | len = i - first; |
682 | free += len; | |
683 | if (len > 1) | |
684 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
685 | else | |
686 | grp->bb_counters[0]++; | |
687 | if (i < max) | |
ffad0a44 | 688 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
689 | } |
690 | grp->bb_fragments = fragments; | |
691 | ||
692 | if (free != grp->bb_free) { | |
5d1b1b3f | 693 | ext4_grp_locked_error(sb, group, __func__, |
a9df9a49 | 694 | "EXT4-fs: group %u: %u blocks in bitmap, %u in gd", |
c9de560d | 695 | group, free, grp->bb_free); |
e56eb659 AK |
696 | /* |
697 | * If we intent to continue, we consider group descritor | |
698 | * corrupt and update bb_free using bitmap value | |
699 | */ | |
c9de560d AT |
700 | grp->bb_free = free; |
701 | } | |
702 | ||
703 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
704 | ||
705 | period = get_cycles() - period; | |
706 | spin_lock(&EXT4_SB(sb)->s_bal_lock); | |
707 | EXT4_SB(sb)->s_mb_buddies_generated++; | |
708 | EXT4_SB(sb)->s_mb_generation_time += period; | |
709 | spin_unlock(&EXT4_SB(sb)->s_bal_lock); | |
710 | } | |
711 | ||
712 | /* The buddy information is attached the buddy cache inode | |
713 | * for convenience. The information regarding each group | |
714 | * is loaded via ext4_mb_load_buddy. The information involve | |
715 | * block bitmap and buddy information. The information are | |
716 | * stored in the inode as | |
717 | * | |
718 | * { page } | |
c3a326a6 | 719 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
720 | * |
721 | * | |
722 | * one block each for bitmap and buddy information. | |
723 | * So for each group we take up 2 blocks. A page can | |
724 | * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. | |
725 | * So it can have information regarding groups_per_page which | |
726 | * is blocks_per_page/2 | |
727 | */ | |
728 | ||
729 | static int ext4_mb_init_cache(struct page *page, char *incore) | |
730 | { | |
8df9675f | 731 | ext4_group_t ngroups; |
c9de560d AT |
732 | int blocksize; |
733 | int blocks_per_page; | |
734 | int groups_per_page; | |
735 | int err = 0; | |
736 | int i; | |
737 | ext4_group_t first_group; | |
738 | int first_block; | |
739 | struct super_block *sb; | |
740 | struct buffer_head *bhs; | |
741 | struct buffer_head **bh; | |
742 | struct inode *inode; | |
743 | char *data; | |
744 | char *bitmap; | |
745 | ||
6ba495e9 | 746 | mb_debug(1, "init page %lu\n", page->index); |
c9de560d AT |
747 | |
748 | inode = page->mapping->host; | |
749 | sb = inode->i_sb; | |
8df9675f | 750 | ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
751 | blocksize = 1 << inode->i_blkbits; |
752 | blocks_per_page = PAGE_CACHE_SIZE / blocksize; | |
753 | ||
754 | groups_per_page = blocks_per_page >> 1; | |
755 | if (groups_per_page == 0) | |
756 | groups_per_page = 1; | |
757 | ||
758 | /* allocate buffer_heads to read bitmaps */ | |
759 | if (groups_per_page > 1) { | |
760 | err = -ENOMEM; | |
761 | i = sizeof(struct buffer_head *) * groups_per_page; | |
762 | bh = kzalloc(i, GFP_NOFS); | |
763 | if (bh == NULL) | |
764 | goto out; | |
765 | } else | |
766 | bh = &bhs; | |
767 | ||
768 | first_group = page->index * blocks_per_page / 2; | |
769 | ||
770 | /* read all groups the page covers into the cache */ | |
771 | for (i = 0; i < groups_per_page; i++) { | |
772 | struct ext4_group_desc *desc; | |
773 | ||
8df9675f | 774 | if (first_group + i >= ngroups) |
c9de560d AT |
775 | break; |
776 | ||
777 | err = -EIO; | |
778 | desc = ext4_get_group_desc(sb, first_group + i, NULL); | |
779 | if (desc == NULL) | |
780 | goto out; | |
781 | ||
782 | err = -ENOMEM; | |
783 | bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc)); | |
784 | if (bh[i] == NULL) | |
785 | goto out; | |
786 | ||
2ccb5fb9 | 787 | if (bitmap_uptodate(bh[i])) |
c9de560d AT |
788 | continue; |
789 | ||
c806e68f | 790 | lock_buffer(bh[i]); |
2ccb5fb9 AK |
791 | if (bitmap_uptodate(bh[i])) { |
792 | unlock_buffer(bh[i]); | |
793 | continue; | |
794 | } | |
955ce5f5 | 795 | ext4_lock_group(sb, first_group + i); |
c9de560d AT |
796 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
797 | ext4_init_block_bitmap(sb, bh[i], | |
798 | first_group + i, desc); | |
2ccb5fb9 | 799 | set_bitmap_uptodate(bh[i]); |
c9de560d | 800 | set_buffer_uptodate(bh[i]); |
955ce5f5 | 801 | ext4_unlock_group(sb, first_group + i); |
3300beda | 802 | unlock_buffer(bh[i]); |
c9de560d AT |
803 | continue; |
804 | } | |
955ce5f5 | 805 | ext4_unlock_group(sb, first_group + i); |
2ccb5fb9 AK |
806 | if (buffer_uptodate(bh[i])) { |
807 | /* | |
808 | * if not uninit if bh is uptodate, | |
809 | * bitmap is also uptodate | |
810 | */ | |
811 | set_bitmap_uptodate(bh[i]); | |
812 | unlock_buffer(bh[i]); | |
813 | continue; | |
814 | } | |
c9de560d | 815 | get_bh(bh[i]); |
2ccb5fb9 AK |
816 | /* |
817 | * submit the buffer_head for read. We can | |
818 | * safely mark the bitmap as uptodate now. | |
819 | * We do it here so the bitmap uptodate bit | |
820 | * get set with buffer lock held. | |
821 | */ | |
822 | set_bitmap_uptodate(bh[i]); | |
c9de560d AT |
823 | bh[i]->b_end_io = end_buffer_read_sync; |
824 | submit_bh(READ, bh[i]); | |
6ba495e9 | 825 | mb_debug(1, "read bitmap for group %u\n", first_group + i); |
c9de560d AT |
826 | } |
827 | ||
828 | /* wait for I/O completion */ | |
829 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
830 | wait_on_buffer(bh[i]); | |
831 | ||
832 | err = -EIO; | |
833 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
834 | if (!buffer_uptodate(bh[i])) | |
835 | goto out; | |
836 | ||
31b481dc | 837 | err = 0; |
c9de560d | 838 | first_block = page->index * blocks_per_page; |
29eaf024 AK |
839 | /* init the page */ |
840 | memset(page_address(page), 0xff, PAGE_CACHE_SIZE); | |
c9de560d AT |
841 | for (i = 0; i < blocks_per_page; i++) { |
842 | int group; | |
843 | struct ext4_group_info *grinfo; | |
844 | ||
845 | group = (first_block + i) >> 1; | |
8df9675f | 846 | if (group >= ngroups) |
c9de560d AT |
847 | break; |
848 | ||
849 | /* | |
850 | * data carry information regarding this | |
851 | * particular group in the format specified | |
852 | * above | |
853 | * | |
854 | */ | |
855 | data = page_address(page) + (i * blocksize); | |
856 | bitmap = bh[group - first_group]->b_data; | |
857 | ||
858 | /* | |
859 | * We place the buddy block and bitmap block | |
860 | * close together | |
861 | */ | |
862 | if ((first_block + i) & 1) { | |
863 | /* this is block of buddy */ | |
864 | BUG_ON(incore == NULL); | |
6ba495e9 | 865 | mb_debug(1, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 866 | group, page->index, i * blocksize); |
c9de560d AT |
867 | grinfo = ext4_get_group_info(sb, group); |
868 | grinfo->bb_fragments = 0; | |
869 | memset(grinfo->bb_counters, 0, | |
1927805e ES |
870 | sizeof(*grinfo->bb_counters) * |
871 | (sb->s_blocksize_bits+2)); | |
c9de560d AT |
872 | /* |
873 | * incore got set to the group block bitmap below | |
874 | */ | |
7a2fcbf7 | 875 | ext4_lock_group(sb, group); |
c9de560d | 876 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 877 | ext4_unlock_group(sb, group); |
c9de560d AT |
878 | incore = NULL; |
879 | } else { | |
880 | /* this is block of bitmap */ | |
881 | BUG_ON(incore != NULL); | |
6ba495e9 | 882 | mb_debug(1, "put bitmap for group %u in page %lu/%x\n", |
c9de560d AT |
883 | group, page->index, i * blocksize); |
884 | ||
885 | /* see comments in ext4_mb_put_pa() */ | |
886 | ext4_lock_group(sb, group); | |
887 | memcpy(data, bitmap, blocksize); | |
888 | ||
889 | /* mark all preallocated blks used in in-core bitmap */ | |
890 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 891 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
892 | ext4_unlock_group(sb, group); |
893 | ||
894 | /* set incore so that the buddy information can be | |
895 | * generated using this | |
896 | */ | |
897 | incore = data; | |
898 | } | |
899 | } | |
900 | SetPageUptodate(page); | |
901 | ||
902 | out: | |
903 | if (bh) { | |
904 | for (i = 0; i < groups_per_page && bh[i]; i++) | |
905 | brelse(bh[i]); | |
906 | if (bh != &bhs) | |
907 | kfree(bh); | |
908 | } | |
909 | return err; | |
910 | } | |
911 | ||
b6a758ec AK |
912 | static noinline_for_stack |
913 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) | |
914 | { | |
915 | ||
916 | int ret = 0; | |
917 | void *bitmap; | |
918 | int blocks_per_page; | |
919 | int block, pnum, poff; | |
920 | int num_grp_locked = 0; | |
921 | struct ext4_group_info *this_grp; | |
922 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
923 | struct inode *inode = sbi->s_buddy_cache; | |
924 | struct page *page = NULL, *bitmap_page = NULL; | |
925 | ||
926 | mb_debug(1, "init group %u\n", group); | |
927 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
928 | this_grp = ext4_get_group_info(sb, group); | |
929 | /* | |
08c3a813 AK |
930 | * This ensures that we don't reinit the buddy cache |
931 | * page which map to the group from which we are already | |
932 | * allocating. If we are looking at the buddy cache we would | |
933 | * have taken a reference using ext4_mb_load_buddy and that | |
934 | * would have taken the alloc_sem lock. | |
b6a758ec AK |
935 | */ |
936 | num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group); | |
937 | if (!EXT4_MB_GRP_NEED_INIT(this_grp)) { | |
938 | /* | |
939 | * somebody initialized the group | |
940 | * return without doing anything | |
941 | */ | |
942 | ret = 0; | |
943 | goto err; | |
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 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
954 | if (page) { | |
955 | BUG_ON(page->mapping != inode->i_mapping); | |
956 | ret = ext4_mb_init_cache(page, NULL); | |
957 | if (ret) { | |
958 | unlock_page(page); | |
959 | goto err; | |
960 | } | |
961 | unlock_page(page); | |
962 | } | |
963 | if (page == NULL || !PageUptodate(page)) { | |
964 | ret = -EIO; | |
965 | goto err; | |
966 | } | |
967 | mark_page_accessed(page); | |
968 | bitmap_page = page; | |
969 | bitmap = page_address(page) + (poff * sb->s_blocksize); | |
970 | ||
971 | /* init buddy cache */ | |
972 | block++; | |
973 | pnum = block / blocks_per_page; | |
974 | poff = block % blocks_per_page; | |
975 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
976 | if (page == bitmap_page) { | |
977 | /* | |
978 | * If both the bitmap and buddy are in | |
979 | * the same page we don't need to force | |
980 | * init the buddy | |
981 | */ | |
982 | unlock_page(page); | |
983 | } else if (page) { | |
984 | BUG_ON(page->mapping != inode->i_mapping); | |
985 | ret = ext4_mb_init_cache(page, bitmap); | |
986 | if (ret) { | |
987 | unlock_page(page); | |
988 | goto err; | |
989 | } | |
990 | unlock_page(page); | |
991 | } | |
992 | if (page == NULL || !PageUptodate(page)) { | |
993 | ret = -EIO; | |
994 | goto err; | |
995 | } | |
996 | mark_page_accessed(page); | |
997 | err: | |
998 | ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked); | |
999 | if (bitmap_page) | |
1000 | page_cache_release(bitmap_page); | |
1001 | if (page) | |
1002 | page_cache_release(page); | |
1003 | return ret; | |
1004 | } | |
1005 | ||
4ddfef7b ES |
1006 | static noinline_for_stack int |
1007 | ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, | |
1008 | struct ext4_buddy *e4b) | |
c9de560d | 1009 | { |
c9de560d AT |
1010 | int blocks_per_page; |
1011 | int block; | |
1012 | int pnum; | |
1013 | int poff; | |
1014 | struct page *page; | |
fdf6c7a7 | 1015 | int ret; |
920313a7 AK |
1016 | struct ext4_group_info *grp; |
1017 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1018 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1019 | |
6ba495e9 | 1020 | mb_debug(1, "load group %u\n", group); |
c9de560d AT |
1021 | |
1022 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
920313a7 | 1023 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1024 | |
1025 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
1026 | e4b->bd_info = ext4_get_group_info(sb, group); | |
1027 | e4b->bd_sb = sb; | |
1028 | e4b->bd_group = group; | |
1029 | e4b->bd_buddy_page = NULL; | |
1030 | e4b->bd_bitmap_page = NULL; | |
920313a7 AK |
1031 | e4b->alloc_semp = &grp->alloc_sem; |
1032 | ||
1033 | /* Take the read lock on the group alloc | |
1034 | * sem. This would make sure a parallel | |
1035 | * ext4_mb_init_group happening on other | |
1036 | * groups mapped by the page is blocked | |
1037 | * till we are done with allocation | |
1038 | */ | |
f41c0750 | 1039 | repeat_load_buddy: |
920313a7 | 1040 | down_read(e4b->alloc_semp); |
c9de560d | 1041 | |
f41c0750 AK |
1042 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
1043 | /* we need to check for group need init flag | |
1044 | * with alloc_semp held so that we can be sure | |
1045 | * that new blocks didn't get added to the group | |
1046 | * when we are loading the buddy cache | |
1047 | */ | |
1048 | up_read(e4b->alloc_semp); | |
1049 | /* | |
1050 | * we need full data about the group | |
1051 | * to make a good selection | |
1052 | */ | |
1053 | ret = ext4_mb_init_group(sb, group); | |
1054 | if (ret) | |
1055 | return ret; | |
1056 | goto repeat_load_buddy; | |
1057 | } | |
1058 | ||
c9de560d AT |
1059 | /* |
1060 | * the buddy cache inode stores the block bitmap | |
1061 | * and buddy information in consecutive blocks. | |
1062 | * So for each group we need two blocks. | |
1063 | */ | |
1064 | block = group * 2; | |
1065 | pnum = block / blocks_per_page; | |
1066 | poff = block % blocks_per_page; | |
1067 | ||
1068 | /* we could use find_or_create_page(), but it locks page | |
1069 | * what we'd like to avoid in fast path ... */ | |
1070 | page = find_get_page(inode->i_mapping, pnum); | |
1071 | if (page == NULL || !PageUptodate(page)) { | |
1072 | if (page) | |
920313a7 AK |
1073 | /* |
1074 | * drop the page reference and try | |
1075 | * to get the page with lock. If we | |
1076 | * are not uptodate that implies | |
1077 | * somebody just created the page but | |
1078 | * is yet to initialize the same. So | |
1079 | * wait for it to initialize. | |
1080 | */ | |
c9de560d AT |
1081 | page_cache_release(page); |
1082 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1083 | if (page) { | |
1084 | BUG_ON(page->mapping != inode->i_mapping); | |
1085 | if (!PageUptodate(page)) { | |
fdf6c7a7 SF |
1086 | ret = ext4_mb_init_cache(page, NULL); |
1087 | if (ret) { | |
1088 | unlock_page(page); | |
1089 | goto err; | |
1090 | } | |
c9de560d AT |
1091 | mb_cmp_bitmaps(e4b, page_address(page) + |
1092 | (poff * sb->s_blocksize)); | |
1093 | } | |
1094 | unlock_page(page); | |
1095 | } | |
1096 | } | |
fdf6c7a7 SF |
1097 | if (page == NULL || !PageUptodate(page)) { |
1098 | ret = -EIO; | |
c9de560d | 1099 | goto err; |
fdf6c7a7 | 1100 | } |
c9de560d AT |
1101 | e4b->bd_bitmap_page = page; |
1102 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1103 | mark_page_accessed(page); | |
1104 | ||
1105 | block++; | |
1106 | pnum = block / blocks_per_page; | |
1107 | poff = block % blocks_per_page; | |
1108 | ||
1109 | page = find_get_page(inode->i_mapping, pnum); | |
1110 | if (page == NULL || !PageUptodate(page)) { | |
1111 | if (page) | |
1112 | page_cache_release(page); | |
1113 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1114 | if (page) { | |
1115 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 SF |
1116 | if (!PageUptodate(page)) { |
1117 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap); | |
1118 | if (ret) { | |
1119 | unlock_page(page); | |
1120 | goto err; | |
1121 | } | |
1122 | } | |
c9de560d AT |
1123 | unlock_page(page); |
1124 | } | |
1125 | } | |
fdf6c7a7 SF |
1126 | if (page == NULL || !PageUptodate(page)) { |
1127 | ret = -EIO; | |
c9de560d | 1128 | goto err; |
fdf6c7a7 | 1129 | } |
c9de560d AT |
1130 | e4b->bd_buddy_page = page; |
1131 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
1132 | mark_page_accessed(page); | |
1133 | ||
1134 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1135 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1136 | ||
1137 | return 0; | |
1138 | ||
1139 | err: | |
1140 | if (e4b->bd_bitmap_page) | |
1141 | page_cache_release(e4b->bd_bitmap_page); | |
1142 | if (e4b->bd_buddy_page) | |
1143 | page_cache_release(e4b->bd_buddy_page); | |
1144 | e4b->bd_buddy = NULL; | |
1145 | e4b->bd_bitmap = NULL; | |
920313a7 AK |
1146 | |
1147 | /* Done with the buddy cache */ | |
1148 | up_read(e4b->alloc_semp); | |
fdf6c7a7 | 1149 | return ret; |
c9de560d AT |
1150 | } |
1151 | ||
1152 | static void ext4_mb_release_desc(struct ext4_buddy *e4b) | |
1153 | { | |
1154 | if (e4b->bd_bitmap_page) | |
1155 | page_cache_release(e4b->bd_bitmap_page); | |
1156 | if (e4b->bd_buddy_page) | |
1157 | page_cache_release(e4b->bd_buddy_page); | |
920313a7 | 1158 | /* Done with the buddy cache */ |
8556e8f3 AK |
1159 | if (e4b->alloc_semp) |
1160 | up_read(e4b->alloc_semp); | |
c9de560d AT |
1161 | } |
1162 | ||
1163 | ||
1164 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1165 | { | |
1166 | int order = 1; | |
1167 | void *bb; | |
1168 | ||
1169 | BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); | |
1170 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); | |
1171 | ||
1172 | bb = EXT4_MB_BUDDY(e4b); | |
1173 | while (order <= e4b->bd_blkbits + 1) { | |
1174 | block = block >> 1; | |
1175 | if (!mb_test_bit(block, bb)) { | |
1176 | /* this block is part of buddy of order 'order' */ | |
1177 | return order; | |
1178 | } | |
1179 | bb += 1 << (e4b->bd_blkbits - order); | |
1180 | order++; | |
1181 | } | |
1182 | return 0; | |
1183 | } | |
1184 | ||
955ce5f5 | 1185 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1186 | { |
1187 | __u32 *addr; | |
1188 | ||
1189 | len = cur + len; | |
1190 | while (cur < len) { | |
1191 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1192 | /* fast path: clear whole word at once */ | |
1193 | addr = bm + (cur >> 3); | |
1194 | *addr = 0; | |
1195 | cur += 32; | |
1196 | continue; | |
1197 | } | |
955ce5f5 | 1198 | mb_clear_bit(cur, bm); |
c9de560d AT |
1199 | cur++; |
1200 | } | |
1201 | } | |
1202 | ||
955ce5f5 | 1203 | static void mb_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1204 | { |
1205 | __u32 *addr; | |
1206 | ||
1207 | len = cur + len; | |
1208 | while (cur < len) { | |
1209 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1210 | /* fast path: set whole word at once */ | |
1211 | addr = bm + (cur >> 3); | |
1212 | *addr = 0xffffffff; | |
1213 | cur += 32; | |
1214 | continue; | |
1215 | } | |
955ce5f5 | 1216 | mb_set_bit(cur, bm); |
c9de560d AT |
1217 | cur++; |
1218 | } | |
1219 | } | |
1220 | ||
7e5a8cdd | 1221 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
c9de560d AT |
1222 | int first, int count) |
1223 | { | |
1224 | int block = 0; | |
1225 | int max = 0; | |
1226 | int order; | |
1227 | void *buddy; | |
1228 | void *buddy2; | |
1229 | struct super_block *sb = e4b->bd_sb; | |
1230 | ||
1231 | BUG_ON(first + count > (sb->s_blocksize << 3)); | |
bc8e6740 | 1232 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
1233 | mb_check_buddy(e4b); |
1234 | mb_free_blocks_double(inode, e4b, first, count); | |
1235 | ||
1236 | e4b->bd_info->bb_free += count; | |
1237 | if (first < e4b->bd_info->bb_first_free) | |
1238 | e4b->bd_info->bb_first_free = first; | |
1239 | ||
1240 | /* let's maintain fragments counter */ | |
1241 | if (first != 0) | |
1242 | block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b)); | |
1243 | if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) | |
1244 | max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b)); | |
1245 | if (block && max) | |
1246 | e4b->bd_info->bb_fragments--; | |
1247 | else if (!block && !max) | |
1248 | e4b->bd_info->bb_fragments++; | |
1249 | ||
1250 | /* let's maintain buddy itself */ | |
1251 | while (count-- > 0) { | |
1252 | block = first++; | |
1253 | order = 0; | |
1254 | ||
1255 | if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) { | |
1256 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
1257 | |
1258 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
c9de560d | 1259 | blocknr += block; |
5d1b1b3f AK |
1260 | ext4_grp_locked_error(sb, e4b->bd_group, |
1261 | __func__, "double-free of inode" | |
a9df9a49 | 1262 | " %lu's block %llu(bit %u in group %u)", |
c9de560d AT |
1263 | inode ? inode->i_ino : 0, blocknr, block, |
1264 | e4b->bd_group); | |
1265 | } | |
1266 | mb_clear_bit(block, EXT4_MB_BITMAP(e4b)); | |
1267 | e4b->bd_info->bb_counters[order]++; | |
1268 | ||
1269 | /* start of the buddy */ | |
1270 | buddy = mb_find_buddy(e4b, order, &max); | |
1271 | ||
1272 | do { | |
1273 | block &= ~1UL; | |
1274 | if (mb_test_bit(block, buddy) || | |
1275 | mb_test_bit(block + 1, buddy)) | |
1276 | break; | |
1277 | ||
1278 | /* both the buddies are free, try to coalesce them */ | |
1279 | buddy2 = mb_find_buddy(e4b, order + 1, &max); | |
1280 | ||
1281 | if (!buddy2) | |
1282 | break; | |
1283 | ||
1284 | if (order > 0) { | |
1285 | /* for special purposes, we don't set | |
1286 | * free bits in bitmap */ | |
1287 | mb_set_bit(block, buddy); | |
1288 | mb_set_bit(block + 1, buddy); | |
1289 | } | |
1290 | e4b->bd_info->bb_counters[order]--; | |
1291 | e4b->bd_info->bb_counters[order]--; | |
1292 | ||
1293 | block = block >> 1; | |
1294 | order++; | |
1295 | e4b->bd_info->bb_counters[order]++; | |
1296 | ||
1297 | mb_clear_bit(block, buddy2); | |
1298 | buddy = buddy2; | |
1299 | } while (1); | |
1300 | } | |
1301 | mb_check_buddy(e4b); | |
c9de560d AT |
1302 | } |
1303 | ||
1304 | static int mb_find_extent(struct ext4_buddy *e4b, int order, int block, | |
1305 | int needed, struct ext4_free_extent *ex) | |
1306 | { | |
1307 | int next = block; | |
1308 | int max; | |
1309 | int ord; | |
1310 | void *buddy; | |
1311 | ||
bc8e6740 | 1312 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1313 | BUG_ON(ex == NULL); |
1314 | ||
1315 | buddy = mb_find_buddy(e4b, order, &max); | |
1316 | BUG_ON(buddy == NULL); | |
1317 | BUG_ON(block >= max); | |
1318 | if (mb_test_bit(block, buddy)) { | |
1319 | ex->fe_len = 0; | |
1320 | ex->fe_start = 0; | |
1321 | ex->fe_group = 0; | |
1322 | return 0; | |
1323 | } | |
1324 | ||
1325 | /* FIXME dorp order completely ? */ | |
1326 | if (likely(order == 0)) { | |
1327 | /* find actual order */ | |
1328 | order = mb_find_order_for_block(e4b, block); | |
1329 | block = block >> order; | |
1330 | } | |
1331 | ||
1332 | ex->fe_len = 1 << order; | |
1333 | ex->fe_start = block << order; | |
1334 | ex->fe_group = e4b->bd_group; | |
1335 | ||
1336 | /* calc difference from given start */ | |
1337 | next = next - ex->fe_start; | |
1338 | ex->fe_len -= next; | |
1339 | ex->fe_start += next; | |
1340 | ||
1341 | while (needed > ex->fe_len && | |
1342 | (buddy = mb_find_buddy(e4b, order, &max))) { | |
1343 | ||
1344 | if (block + 1 >= max) | |
1345 | break; | |
1346 | ||
1347 | next = (block + 1) * (1 << order); | |
1348 | if (mb_test_bit(next, EXT4_MB_BITMAP(e4b))) | |
1349 | break; | |
1350 | ||
1351 | ord = mb_find_order_for_block(e4b, next); | |
1352 | ||
1353 | order = ord; | |
1354 | block = next >> order; | |
1355 | ex->fe_len += 1 << order; | |
1356 | } | |
1357 | ||
1358 | BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); | |
1359 | return ex->fe_len; | |
1360 | } | |
1361 | ||
1362 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1363 | { | |
1364 | int ord; | |
1365 | int mlen = 0; | |
1366 | int max = 0; | |
1367 | int cur; | |
1368 | int start = ex->fe_start; | |
1369 | int len = ex->fe_len; | |
1370 | unsigned ret = 0; | |
1371 | int len0 = len; | |
1372 | void *buddy; | |
1373 | ||
1374 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1375 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1376 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1377 | mb_check_buddy(e4b); |
1378 | mb_mark_used_double(e4b, start, len); | |
1379 | ||
1380 | e4b->bd_info->bb_free -= len; | |
1381 | if (e4b->bd_info->bb_first_free == start) | |
1382 | e4b->bd_info->bb_first_free += len; | |
1383 | ||
1384 | /* let's maintain fragments counter */ | |
1385 | if (start != 0) | |
1386 | mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b)); | |
1387 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) | |
1388 | max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b)); | |
1389 | if (mlen && max) | |
1390 | e4b->bd_info->bb_fragments++; | |
1391 | else if (!mlen && !max) | |
1392 | e4b->bd_info->bb_fragments--; | |
1393 | ||
1394 | /* let's maintain buddy itself */ | |
1395 | while (len) { | |
1396 | ord = mb_find_order_for_block(e4b, start); | |
1397 | ||
1398 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1399 | /* the whole chunk may be allocated at once! */ | |
1400 | mlen = 1 << ord; | |
1401 | buddy = mb_find_buddy(e4b, ord, &max); | |
1402 | BUG_ON((start >> ord) >= max); | |
1403 | mb_set_bit(start >> ord, buddy); | |
1404 | e4b->bd_info->bb_counters[ord]--; | |
1405 | start += mlen; | |
1406 | len -= mlen; | |
1407 | BUG_ON(len < 0); | |
1408 | continue; | |
1409 | } | |
1410 | ||
1411 | /* store for history */ | |
1412 | if (ret == 0) | |
1413 | ret = len | (ord << 16); | |
1414 | ||
1415 | /* we have to split large buddy */ | |
1416 | BUG_ON(ord <= 0); | |
1417 | buddy = mb_find_buddy(e4b, ord, &max); | |
1418 | mb_set_bit(start >> ord, buddy); | |
1419 | e4b->bd_info->bb_counters[ord]--; | |
1420 | ||
1421 | ord--; | |
1422 | cur = (start >> ord) & ~1U; | |
1423 | buddy = mb_find_buddy(e4b, ord, &max); | |
1424 | mb_clear_bit(cur, buddy); | |
1425 | mb_clear_bit(cur + 1, buddy); | |
1426 | e4b->bd_info->bb_counters[ord]++; | |
1427 | e4b->bd_info->bb_counters[ord]++; | |
1428 | } | |
1429 | ||
955ce5f5 | 1430 | mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0); |
c9de560d AT |
1431 | mb_check_buddy(e4b); |
1432 | ||
1433 | return ret; | |
1434 | } | |
1435 | ||
1436 | /* | |
1437 | * Must be called under group lock! | |
1438 | */ | |
1439 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1440 | struct ext4_buddy *e4b) | |
1441 | { | |
1442 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1443 | int ret; | |
1444 | ||
1445 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1446 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1447 | ||
1448 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1449 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1450 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1451 | ||
1452 | /* preallocation can change ac_b_ex, thus we store actually | |
1453 | * allocated blocks for history */ | |
1454 | ac->ac_f_ex = ac->ac_b_ex; | |
1455 | ||
1456 | ac->ac_status = AC_STATUS_FOUND; | |
1457 | ac->ac_tail = ret & 0xffff; | |
1458 | ac->ac_buddy = ret >> 16; | |
1459 | ||
c3a326a6 AK |
1460 | /* |
1461 | * take the page reference. We want the page to be pinned | |
1462 | * so that we don't get a ext4_mb_init_cache_call for this | |
1463 | * group until we update the bitmap. That would mean we | |
1464 | * double allocate blocks. The reference is dropped | |
1465 | * in ext4_mb_release_context | |
1466 | */ | |
c9de560d AT |
1467 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1468 | get_page(ac->ac_bitmap_page); | |
1469 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1470 | get_page(ac->ac_buddy_page); | |
8556e8f3 AK |
1471 | /* on allocation we use ac to track the held semaphore */ |
1472 | ac->alloc_semp = e4b->alloc_semp; | |
1473 | e4b->alloc_semp = NULL; | |
c9de560d | 1474 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 1475 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
1476 | spin_lock(&sbi->s_md_lock); |
1477 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1478 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1479 | spin_unlock(&sbi->s_md_lock); | |
1480 | } | |
1481 | } | |
1482 | ||
1483 | /* | |
1484 | * regular allocator, for general purposes allocation | |
1485 | */ | |
1486 | ||
1487 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1488 | struct ext4_buddy *e4b, | |
1489 | int finish_group) | |
1490 | { | |
1491 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1492 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1493 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1494 | struct ext4_free_extent ex; | |
1495 | int max; | |
1496 | ||
032115fc AK |
1497 | if (ac->ac_status == AC_STATUS_FOUND) |
1498 | return; | |
c9de560d AT |
1499 | /* |
1500 | * We don't want to scan for a whole year | |
1501 | */ | |
1502 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1503 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1504 | ac->ac_status = AC_STATUS_BREAK; | |
1505 | return; | |
1506 | } | |
1507 | ||
1508 | /* | |
1509 | * Haven't found good chunk so far, let's continue | |
1510 | */ | |
1511 | if (bex->fe_len < gex->fe_len) | |
1512 | return; | |
1513 | ||
1514 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1515 | && bex->fe_group == e4b->bd_group) { | |
1516 | /* recheck chunk's availability - we don't know | |
1517 | * when it was found (within this lock-unlock | |
1518 | * period or not) */ | |
1519 | max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex); | |
1520 | if (max >= gex->fe_len) { | |
1521 | ext4_mb_use_best_found(ac, e4b); | |
1522 | return; | |
1523 | } | |
1524 | } | |
1525 | } | |
1526 | ||
1527 | /* | |
1528 | * The routine checks whether found extent is good enough. If it is, | |
1529 | * then the extent gets marked used and flag is set to the context | |
1530 | * to stop scanning. Otherwise, the extent is compared with the | |
1531 | * previous found extent and if new one is better, then it's stored | |
1532 | * in the context. Later, the best found extent will be used, if | |
1533 | * mballoc can't find good enough extent. | |
1534 | * | |
1535 | * FIXME: real allocation policy is to be designed yet! | |
1536 | */ | |
1537 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1538 | struct ext4_free_extent *ex, | |
1539 | struct ext4_buddy *e4b) | |
1540 | { | |
1541 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1542 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1543 | ||
1544 | BUG_ON(ex->fe_len <= 0); | |
8d03c7a0 | 1545 | BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
1546 | BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
1547 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); | |
1548 | ||
1549 | ac->ac_found++; | |
1550 | ||
1551 | /* | |
1552 | * The special case - take what you catch first | |
1553 | */ | |
1554 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1555 | *bex = *ex; | |
1556 | ext4_mb_use_best_found(ac, e4b); | |
1557 | return; | |
1558 | } | |
1559 | ||
1560 | /* | |
1561 | * Let's check whether the chuck is good enough | |
1562 | */ | |
1563 | if (ex->fe_len == gex->fe_len) { | |
1564 | *bex = *ex; | |
1565 | ext4_mb_use_best_found(ac, e4b); | |
1566 | return; | |
1567 | } | |
1568 | ||
1569 | /* | |
1570 | * If this is first found extent, just store it in the context | |
1571 | */ | |
1572 | if (bex->fe_len == 0) { | |
1573 | *bex = *ex; | |
1574 | return; | |
1575 | } | |
1576 | ||
1577 | /* | |
1578 | * If new found extent is better, store it in the context | |
1579 | */ | |
1580 | if (bex->fe_len < gex->fe_len) { | |
1581 | /* if the request isn't satisfied, any found extent | |
1582 | * larger than previous best one is better */ | |
1583 | if (ex->fe_len > bex->fe_len) | |
1584 | *bex = *ex; | |
1585 | } else if (ex->fe_len > gex->fe_len) { | |
1586 | /* if the request is satisfied, then we try to find | |
1587 | * an extent that still satisfy the request, but is | |
1588 | * smaller than previous one */ | |
1589 | if (ex->fe_len < bex->fe_len) | |
1590 | *bex = *ex; | |
1591 | } | |
1592 | ||
1593 | ext4_mb_check_limits(ac, e4b, 0); | |
1594 | } | |
1595 | ||
089ceecc ES |
1596 | static noinline_for_stack |
1597 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
1598 | struct ext4_buddy *e4b) |
1599 | { | |
1600 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1601 | ext4_group_t group = ex.fe_group; | |
1602 | int max; | |
1603 | int err; | |
1604 | ||
1605 | BUG_ON(ex.fe_len <= 0); | |
1606 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1607 | if (err) | |
1608 | return err; | |
1609 | ||
1610 | ext4_lock_group(ac->ac_sb, group); | |
1611 | max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex); | |
1612 | ||
1613 | if (max > 0) { | |
1614 | ac->ac_b_ex = ex; | |
1615 | ext4_mb_use_best_found(ac, e4b); | |
1616 | } | |
1617 | ||
1618 | ext4_unlock_group(ac->ac_sb, group); | |
1619 | ext4_mb_release_desc(e4b); | |
1620 | ||
1621 | return 0; | |
1622 | } | |
1623 | ||
089ceecc ES |
1624 | static noinline_for_stack |
1625 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
1626 | struct ext4_buddy *e4b) |
1627 | { | |
1628 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1629 | int max; | |
1630 | int err; | |
1631 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
c9de560d AT |
1632 | struct ext4_free_extent ex; |
1633 | ||
1634 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1635 | return 0; | |
1636 | ||
1637 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1638 | if (err) | |
1639 | return err; | |
1640 | ||
1641 | ext4_lock_group(ac->ac_sb, group); | |
1642 | max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start, | |
1643 | ac->ac_g_ex.fe_len, &ex); | |
1644 | ||
1645 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1646 | ext4_fsblk_t start; | |
1647 | ||
5661bd68 AM |
1648 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
1649 | ex.fe_start; | |
c9de560d AT |
1650 | /* use do_div to get remainder (would be 64-bit modulo) */ |
1651 | if (do_div(start, sbi->s_stripe) == 0) { | |
1652 | ac->ac_found++; | |
1653 | ac->ac_b_ex = ex; | |
1654 | ext4_mb_use_best_found(ac, e4b); | |
1655 | } | |
1656 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1657 | BUG_ON(ex.fe_len <= 0); | |
1658 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1659 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1660 | ac->ac_found++; | |
1661 | ac->ac_b_ex = ex; | |
1662 | ext4_mb_use_best_found(ac, e4b); | |
1663 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1664 | /* Sometimes, caller may want to merge even small | |
1665 | * number of blocks to an existing extent */ | |
1666 | BUG_ON(ex.fe_len <= 0); | |
1667 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1668 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1669 | ac->ac_found++; | |
1670 | ac->ac_b_ex = ex; | |
1671 | ext4_mb_use_best_found(ac, e4b); | |
1672 | } | |
1673 | ext4_unlock_group(ac->ac_sb, group); | |
1674 | ext4_mb_release_desc(e4b); | |
1675 | ||
1676 | return 0; | |
1677 | } | |
1678 | ||
1679 | /* | |
1680 | * The routine scans buddy structures (not bitmap!) from given order | |
1681 | * to max order and tries to find big enough chunk to satisfy the req | |
1682 | */ | |
089ceecc ES |
1683 | static noinline_for_stack |
1684 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1685 | struct ext4_buddy *e4b) |
1686 | { | |
1687 | struct super_block *sb = ac->ac_sb; | |
1688 | struct ext4_group_info *grp = e4b->bd_info; | |
1689 | void *buddy; | |
1690 | int i; | |
1691 | int k; | |
1692 | int max; | |
1693 | ||
1694 | BUG_ON(ac->ac_2order <= 0); | |
1695 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1696 | if (grp->bb_counters[i] == 0) | |
1697 | continue; | |
1698 | ||
1699 | buddy = mb_find_buddy(e4b, i, &max); | |
1700 | BUG_ON(buddy == NULL); | |
1701 | ||
ffad0a44 | 1702 | k = mb_find_next_zero_bit(buddy, max, 0); |
c9de560d AT |
1703 | BUG_ON(k >= max); |
1704 | ||
1705 | ac->ac_found++; | |
1706 | ||
1707 | ac->ac_b_ex.fe_len = 1 << i; | |
1708 | ac->ac_b_ex.fe_start = k << i; | |
1709 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1710 | ||
1711 | ext4_mb_use_best_found(ac, e4b); | |
1712 | ||
1713 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1714 | ||
1715 | if (EXT4_SB(sb)->s_mb_stats) | |
1716 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1717 | ||
1718 | break; | |
1719 | } | |
1720 | } | |
1721 | ||
1722 | /* | |
1723 | * The routine scans the group and measures all found extents. | |
1724 | * In order to optimize scanning, caller must pass number of | |
1725 | * free blocks in the group, so the routine can know upper limit. | |
1726 | */ | |
089ceecc ES |
1727 | static noinline_for_stack |
1728 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1729 | struct ext4_buddy *e4b) |
1730 | { | |
1731 | struct super_block *sb = ac->ac_sb; | |
1732 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1733 | struct ext4_free_extent ex; | |
1734 | int i; | |
1735 | int free; | |
1736 | ||
1737 | free = e4b->bd_info->bb_free; | |
1738 | BUG_ON(free <= 0); | |
1739 | ||
1740 | i = e4b->bd_info->bb_first_free; | |
1741 | ||
1742 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1743 | i = mb_find_next_zero_bit(bitmap, |
c9de560d AT |
1744 | EXT4_BLOCKS_PER_GROUP(sb), i); |
1745 | if (i >= EXT4_BLOCKS_PER_GROUP(sb)) { | |
26346ff6 | 1746 | /* |
e56eb659 | 1747 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1748 | * free blocks even though group info says we |
1749 | * we have free blocks | |
1750 | */ | |
5d1b1b3f AK |
1751 | ext4_grp_locked_error(sb, e4b->bd_group, |
1752 | __func__, "%d free blocks as per " | |
fde4d95a | 1753 | "group info. But bitmap says 0", |
26346ff6 | 1754 | free); |
c9de560d AT |
1755 | break; |
1756 | } | |
1757 | ||
1758 | mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex); | |
1759 | BUG_ON(ex.fe_len <= 0); | |
26346ff6 | 1760 | if (free < ex.fe_len) { |
5d1b1b3f AK |
1761 | ext4_grp_locked_error(sb, e4b->bd_group, |
1762 | __func__, "%d free blocks as per " | |
fde4d95a | 1763 | "group info. But got %d blocks", |
26346ff6 | 1764 | free, ex.fe_len); |
e56eb659 AK |
1765 | /* |
1766 | * The number of free blocks differs. This mostly | |
1767 | * indicate that the bitmap is corrupt. So exit | |
1768 | * without claiming the space. | |
1769 | */ | |
1770 | break; | |
26346ff6 | 1771 | } |
c9de560d AT |
1772 | |
1773 | ext4_mb_measure_extent(ac, &ex, e4b); | |
1774 | ||
1775 | i += ex.fe_len; | |
1776 | free -= ex.fe_len; | |
1777 | } | |
1778 | ||
1779 | ext4_mb_check_limits(ac, e4b, 1); | |
1780 | } | |
1781 | ||
1782 | /* | |
1783 | * This is a special case for storages like raid5 | |
1784 | * we try to find stripe-aligned chunks for stripe-size requests | |
1785 | * XXX should do so at least for multiples of stripe size as well | |
1786 | */ | |
089ceecc ES |
1787 | static noinline_for_stack |
1788 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
1789 | struct ext4_buddy *e4b) |
1790 | { | |
1791 | struct super_block *sb = ac->ac_sb; | |
1792 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1793 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1794 | struct ext4_free_extent ex; | |
1795 | ext4_fsblk_t first_group_block; | |
1796 | ext4_fsblk_t a; | |
1797 | ext4_grpblk_t i; | |
1798 | int max; | |
1799 | ||
1800 | BUG_ON(sbi->s_stripe == 0); | |
1801 | ||
1802 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
1803 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
1804 | ||
c9de560d AT |
1805 | a = first_group_block + sbi->s_stripe - 1; |
1806 | do_div(a, sbi->s_stripe); | |
1807 | i = (a * sbi->s_stripe) - first_group_block; | |
1808 | ||
1809 | while (i < EXT4_BLOCKS_PER_GROUP(sb)) { | |
1810 | if (!mb_test_bit(i, bitmap)) { | |
1811 | max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex); | |
1812 | if (max >= sbi->s_stripe) { | |
1813 | ac->ac_found++; | |
1814 | ac->ac_b_ex = ex; | |
1815 | ext4_mb_use_best_found(ac, e4b); | |
1816 | break; | |
1817 | } | |
1818 | } | |
1819 | i += sbi->s_stripe; | |
1820 | } | |
1821 | } | |
1822 | ||
1823 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, | |
1824 | ext4_group_t group, int cr) | |
1825 | { | |
1826 | unsigned free, fragments; | |
1827 | unsigned i, bits; | |
a4912123 | 1828 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
1829 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
1830 | ||
1831 | BUG_ON(cr < 0 || cr >= 4); | |
1832 | BUG_ON(EXT4_MB_GRP_NEED_INIT(grp)); | |
1833 | ||
1834 | free = grp->bb_free; | |
1835 | fragments = grp->bb_fragments; | |
1836 | if (free == 0) | |
1837 | return 0; | |
1838 | if (fragments == 0) | |
1839 | return 0; | |
1840 | ||
1841 | switch (cr) { | |
1842 | case 0: | |
1843 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 1844 | |
a4912123 TT |
1845 | /* Avoid using the first bg of a flexgroup for data files */ |
1846 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
1847 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
1848 | ((group % flex_size) == 0)) | |
1849 | return 0; | |
1850 | ||
c9de560d AT |
1851 | bits = ac->ac_sb->s_blocksize_bits + 1; |
1852 | for (i = ac->ac_2order; i <= bits; i++) | |
1853 | if (grp->bb_counters[i] > 0) | |
1854 | return 1; | |
1855 | break; | |
1856 | case 1: | |
1857 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
1858 | return 1; | |
1859 | break; | |
1860 | case 2: | |
1861 | if (free >= ac->ac_g_ex.fe_len) | |
1862 | return 1; | |
1863 | break; | |
1864 | case 3: | |
1865 | return 1; | |
1866 | default: | |
1867 | BUG(); | |
1868 | } | |
1869 | ||
1870 | return 0; | |
1871 | } | |
1872 | ||
920313a7 AK |
1873 | /* |
1874 | * lock the group_info alloc_sem of all the groups | |
1875 | * belonging to the same buddy cache page. This | |
1876 | * make sure other parallel operation on the buddy | |
1877 | * cache doesn't happen whild holding the buddy cache | |
1878 | * lock | |
1879 | */ | |
1880 | int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group) | |
1881 | { | |
1882 | int i; | |
1883 | int block, pnum; | |
1884 | int blocks_per_page; | |
1885 | int groups_per_page; | |
8df9675f | 1886 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
920313a7 AK |
1887 | ext4_group_t first_group; |
1888 | struct ext4_group_info *grp; | |
1889 | ||
1890 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1891 | /* | |
1892 | * the buddy cache inode stores the block bitmap | |
1893 | * and buddy information in consecutive blocks. | |
1894 | * So for each group we need two blocks. | |
1895 | */ | |
1896 | block = group * 2; | |
1897 | pnum = block / blocks_per_page; | |
1898 | first_group = pnum * blocks_per_page / 2; | |
1899 | ||
1900 | groups_per_page = blocks_per_page >> 1; | |
1901 | if (groups_per_page == 0) | |
1902 | groups_per_page = 1; | |
1903 | /* read all groups the page covers into the cache */ | |
1904 | for (i = 0; i < groups_per_page; i++) { | |
1905 | ||
8df9675f | 1906 | if ((first_group + i) >= ngroups) |
920313a7 AK |
1907 | break; |
1908 | grp = ext4_get_group_info(sb, first_group + i); | |
1909 | /* take all groups write allocation | |
1910 | * semaphore. This make sure there is | |
1911 | * no block allocation going on in any | |
1912 | * of that groups | |
1913 | */ | |
b7be019e | 1914 | down_write_nested(&grp->alloc_sem, i); |
920313a7 AK |
1915 | } |
1916 | return i; | |
1917 | } | |
1918 | ||
1919 | void ext4_mb_put_buddy_cache_lock(struct super_block *sb, | |
1920 | ext4_group_t group, int locked_group) | |
1921 | { | |
1922 | int i; | |
1923 | int block, pnum; | |
1924 | int blocks_per_page; | |
1925 | ext4_group_t first_group; | |
1926 | struct ext4_group_info *grp; | |
1927 | ||
1928 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1929 | /* | |
1930 | * the buddy cache inode stores the block bitmap | |
1931 | * and buddy information in consecutive blocks. | |
1932 | * So for each group we need two blocks. | |
1933 | */ | |
1934 | block = group * 2; | |
1935 | pnum = block / blocks_per_page; | |
1936 | first_group = pnum * blocks_per_page / 2; | |
1937 | /* release locks on all the groups */ | |
1938 | for (i = 0; i < locked_group; i++) { | |
1939 | ||
1940 | grp = ext4_get_group_info(sb, first_group + i); | |
1941 | /* take all groups write allocation | |
1942 | * semaphore. This make sure there is | |
1943 | * no block allocation going on in any | |
1944 | * of that groups | |
1945 | */ | |
1946 | up_write(&grp->alloc_sem); | |
1947 | } | |
1948 | ||
1949 | } | |
1950 | ||
4ddfef7b ES |
1951 | static noinline_for_stack int |
1952 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 1953 | { |
8df9675f | 1954 | ext4_group_t ngroups, group, i; |
c9de560d AT |
1955 | int cr; |
1956 | int err = 0; | |
1957 | int bsbits; | |
1958 | struct ext4_sb_info *sbi; | |
1959 | struct super_block *sb; | |
1960 | struct ext4_buddy e4b; | |
c9de560d AT |
1961 | |
1962 | sb = ac->ac_sb; | |
1963 | sbi = EXT4_SB(sb); | |
8df9675f | 1964 | ngroups = ext4_get_groups_count(sb); |
fb0a387d ES |
1965 | /* non-extent files are limited to low blocks/groups */ |
1966 | if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL)) | |
1967 | ngroups = sbi->s_blockfile_groups; | |
1968 | ||
c9de560d AT |
1969 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
1970 | ||
1971 | /* first, try the goal */ | |
1972 | err = ext4_mb_find_by_goal(ac, &e4b); | |
1973 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
1974 | goto out; | |
1975 | ||
1976 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
1977 | goto out; | |
1978 | ||
1979 | /* | |
1980 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
1981 | * if ac2_order is set we also set criteria to 0 so that we | |
1982 | * try exact allocation using buddy. | |
1983 | */ | |
1984 | i = fls(ac->ac_g_ex.fe_len); | |
1985 | ac->ac_2order = 0; | |
1986 | /* | |
1987 | * We search using buddy data only if the order of the request | |
1988 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 1989 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
c9de560d AT |
1990 | */ |
1991 | if (i >= sbi->s_mb_order2_reqs) { | |
1992 | /* | |
1993 | * This should tell if fe_len is exactly power of 2 | |
1994 | */ | |
1995 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1996 | ac->ac_2order = i - 1; | |
1997 | } | |
1998 | ||
1999 | bsbits = ac->ac_sb->s_blocksize_bits; | |
c9de560d | 2000 | |
4ba74d00 TT |
2001 | /* if stream allocation is enabled, use global goal */ |
2002 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2003 | /* TBD: may be hot point */ |
2004 | spin_lock(&sbi->s_md_lock); | |
2005 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2006 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2007 | spin_unlock(&sbi->s_md_lock); | |
2008 | } | |
4ba74d00 | 2009 | |
c9de560d AT |
2010 | /* Let's just scan groups to find more-less suitable blocks */ |
2011 | cr = ac->ac_2order ? 0 : 1; | |
2012 | /* | |
2013 | * cr == 0 try to get exact allocation, | |
2014 | * cr == 3 try to get anything | |
2015 | */ | |
2016 | repeat: | |
2017 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2018 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2019 | /* |
2020 | * searching for the right group start | |
2021 | * from the goal value specified | |
2022 | */ | |
2023 | group = ac->ac_g_ex.fe_group; | |
2024 | ||
8df9675f | 2025 | for (i = 0; i < ngroups; group++, i++) { |
c9de560d AT |
2026 | struct ext4_group_info *grp; |
2027 | struct ext4_group_desc *desc; | |
2028 | ||
8df9675f | 2029 | if (group == ngroups) |
c9de560d AT |
2030 | group = 0; |
2031 | ||
2032 | /* quick check to skip empty groups */ | |
920313a7 | 2033 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
2034 | if (grp->bb_free == 0) |
2035 | continue; | |
2036 | ||
c9de560d AT |
2037 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2038 | if (err) | |
2039 | goto out; | |
2040 | ||
2041 | ext4_lock_group(sb, group); | |
2042 | if (!ext4_mb_good_group(ac, group, cr)) { | |
2043 | /* someone did allocation from this group */ | |
2044 | ext4_unlock_group(sb, group); | |
2045 | ext4_mb_release_desc(&e4b); | |
2046 | continue; | |
2047 | } | |
2048 | ||
2049 | ac->ac_groups_scanned++; | |
2050 | desc = ext4_get_group_desc(sb, group, NULL); | |
75507efb | 2051 | if (cr == 0) |
c9de560d AT |
2052 | ext4_mb_simple_scan_group(ac, &e4b); |
2053 | else if (cr == 1 && | |
2054 | ac->ac_g_ex.fe_len == sbi->s_stripe) | |
2055 | ext4_mb_scan_aligned(ac, &e4b); | |
2056 | else | |
2057 | ext4_mb_complex_scan_group(ac, &e4b); | |
2058 | ||
2059 | ext4_unlock_group(sb, group); | |
2060 | ext4_mb_release_desc(&e4b); | |
2061 | ||
2062 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2063 | break; | |
2064 | } | |
2065 | } | |
2066 | ||
2067 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2068 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2069 | /* | |
2070 | * We've been searching too long. Let's try to allocate | |
2071 | * the best chunk we've found so far | |
2072 | */ | |
2073 | ||
2074 | ext4_mb_try_best_found(ac, &e4b); | |
2075 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2076 | /* | |
2077 | * Someone more lucky has already allocated it. | |
2078 | * The only thing we can do is just take first | |
2079 | * found block(s) | |
2080 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2081 | */ | |
2082 | ac->ac_b_ex.fe_group = 0; | |
2083 | ac->ac_b_ex.fe_start = 0; | |
2084 | ac->ac_b_ex.fe_len = 0; | |
2085 | ac->ac_status = AC_STATUS_CONTINUE; | |
2086 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2087 | cr = 3; | |
2088 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2089 | goto repeat; | |
2090 | } | |
2091 | } | |
2092 | out: | |
2093 | return err; | |
2094 | } | |
2095 | ||
c9de560d AT |
2096 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2097 | { | |
2098 | struct super_block *sb = seq->private; | |
c9de560d AT |
2099 | ext4_group_t group; |
2100 | ||
8df9675f | 2101 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2102 | return NULL; |
c9de560d | 2103 | group = *pos + 1; |
a9df9a49 | 2104 | return (void *) ((unsigned long) group); |
c9de560d AT |
2105 | } |
2106 | ||
2107 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2108 | { | |
2109 | struct super_block *sb = seq->private; | |
c9de560d AT |
2110 | ext4_group_t group; |
2111 | ||
2112 | ++*pos; | |
8df9675f | 2113 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2114 | return NULL; |
2115 | group = *pos + 1; | |
a9df9a49 | 2116 | return (void *) ((unsigned long) group); |
c9de560d AT |
2117 | } |
2118 | ||
2119 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2120 | { | |
2121 | struct super_block *sb = seq->private; | |
a9df9a49 | 2122 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d AT |
2123 | int i; |
2124 | int err; | |
2125 | struct ext4_buddy e4b; | |
2126 | struct sg { | |
2127 | struct ext4_group_info info; | |
a36b4498 | 2128 | ext4_grpblk_t counters[16]; |
c9de560d AT |
2129 | } sg; |
2130 | ||
2131 | group--; | |
2132 | if (group == 0) | |
2133 | seq_printf(seq, "#%-5s: %-5s %-5s %-5s " | |
2134 | "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " | |
2135 | "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", | |
2136 | "group", "free", "frags", "first", | |
2137 | "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", | |
2138 | "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); | |
2139 | ||
2140 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2141 | sizeof(struct ext4_group_info); | |
2142 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2143 | if (err) { | |
a9df9a49 | 2144 | seq_printf(seq, "#%-5u: I/O error\n", group); |
c9de560d AT |
2145 | return 0; |
2146 | } | |
2147 | ext4_lock_group(sb, group); | |
2148 | memcpy(&sg, ext4_get_group_info(sb, group), i); | |
2149 | ext4_unlock_group(sb, group); | |
2150 | ext4_mb_release_desc(&e4b); | |
2151 | ||
a9df9a49 | 2152 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2153 | sg.info.bb_fragments, sg.info.bb_first_free); |
2154 | for (i = 0; i <= 13; i++) | |
2155 | seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? | |
2156 | sg.info.bb_counters[i] : 0); | |
2157 | seq_printf(seq, " ]\n"); | |
2158 | ||
2159 | return 0; | |
2160 | } | |
2161 | ||
2162 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2163 | { | |
2164 | } | |
2165 | ||
7f1346a9 | 2166 | static const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2167 | .start = ext4_mb_seq_groups_start, |
2168 | .next = ext4_mb_seq_groups_next, | |
2169 | .stop = ext4_mb_seq_groups_stop, | |
2170 | .show = ext4_mb_seq_groups_show, | |
2171 | }; | |
2172 | ||
2173 | static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) | |
2174 | { | |
2175 | struct super_block *sb = PDE(inode)->data; | |
2176 | int rc; | |
2177 | ||
2178 | rc = seq_open(file, &ext4_mb_seq_groups_ops); | |
2179 | if (rc == 0) { | |
2180 | struct seq_file *m = (struct seq_file *)file->private_data; | |
2181 | m->private = sb; | |
2182 | } | |
2183 | return rc; | |
2184 | ||
2185 | } | |
2186 | ||
7f1346a9 | 2187 | static const struct file_operations ext4_mb_seq_groups_fops = { |
c9de560d AT |
2188 | .owner = THIS_MODULE, |
2189 | .open = ext4_mb_seq_groups_open, | |
2190 | .read = seq_read, | |
2191 | .llseek = seq_lseek, | |
2192 | .release = seq_release, | |
2193 | }; | |
2194 | ||
5f21b0e6 FB |
2195 | |
2196 | /* Create and initialize ext4_group_info data for the given group. */ | |
920313a7 | 2197 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2198 | struct ext4_group_desc *desc) |
2199 | { | |
2200 | int i, len; | |
2201 | int metalen = 0; | |
2202 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2203 | struct ext4_group_info **meta_group_info; | |
2204 | ||
2205 | /* | |
2206 | * First check if this group is the first of a reserved block. | |
2207 | * If it's true, we have to allocate a new table of pointers | |
2208 | * to ext4_group_info structures | |
2209 | */ | |
2210 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2211 | metalen = sizeof(*meta_group_info) << | |
2212 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2213 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2214 | if (meta_group_info == NULL) { | |
2215 | printk(KERN_ERR "EXT4-fs: can't allocate mem for a " | |
2216 | "buddy group\n"); | |
2217 | goto exit_meta_group_info; | |
2218 | } | |
2219 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = | |
2220 | meta_group_info; | |
2221 | } | |
2222 | ||
2223 | /* | |
2224 | * calculate needed size. if change bb_counters size, | |
2225 | * don't forget about ext4_mb_generate_buddy() | |
2226 | */ | |
2227 | len = offsetof(typeof(**meta_group_info), | |
2228 | bb_counters[sb->s_blocksize_bits + 2]); | |
2229 | ||
2230 | meta_group_info = | |
2231 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; | |
2232 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); | |
2233 | ||
2234 | meta_group_info[i] = kzalloc(len, GFP_KERNEL); | |
2235 | if (meta_group_info[i] == NULL) { | |
2236 | printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n"); | |
2237 | goto exit_group_info; | |
2238 | } | |
2239 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2240 | &(meta_group_info[i]->bb_state)); | |
2241 | ||
2242 | /* | |
2243 | * initialize bb_free to be able to skip | |
2244 | * empty groups without initialization | |
2245 | */ | |
2246 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2247 | meta_group_info[i]->bb_free = | |
2248 | ext4_free_blocks_after_init(sb, group, desc); | |
2249 | } else { | |
2250 | meta_group_info[i]->bb_free = | |
560671a0 | 2251 | ext4_free_blks_count(sb, desc); |
5f21b0e6 FB |
2252 | } |
2253 | ||
2254 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2255 | init_rwsem(&meta_group_info[i]->alloc_sem); |
5a4a7989 | 2256 | meta_group_info[i]->bb_free_root.rb_node = NULL; |
5f21b0e6 FB |
2257 | |
2258 | #ifdef DOUBLE_CHECK | |
2259 | { | |
2260 | struct buffer_head *bh; | |
2261 | meta_group_info[i]->bb_bitmap = | |
2262 | kmalloc(sb->s_blocksize, GFP_KERNEL); | |
2263 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); | |
2264 | bh = ext4_read_block_bitmap(sb, group); | |
2265 | BUG_ON(bh == NULL); | |
2266 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, | |
2267 | sb->s_blocksize); | |
2268 | put_bh(bh); | |
2269 | } | |
2270 | #endif | |
2271 | ||
2272 | return 0; | |
2273 | ||
2274 | exit_group_info: | |
2275 | /* If a meta_group_info table has been allocated, release it now */ | |
2276 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) | |
2277 | kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); | |
2278 | exit_meta_group_info: | |
2279 | return -ENOMEM; | |
2280 | } /* ext4_mb_add_groupinfo */ | |
2281 | ||
c9de560d AT |
2282 | static int ext4_mb_init_backend(struct super_block *sb) |
2283 | { | |
8df9675f | 2284 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2285 | ext4_group_t i; |
c9de560d | 2286 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
5f21b0e6 FB |
2287 | struct ext4_super_block *es = sbi->s_es; |
2288 | int num_meta_group_infos; | |
2289 | int num_meta_group_infos_max; | |
2290 | int array_size; | |
5f21b0e6 FB |
2291 | struct ext4_group_desc *desc; |
2292 | ||
2293 | /* This is the number of blocks used by GDT */ | |
8df9675f | 2294 | num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) - |
5f21b0e6 FB |
2295 | 1) >> EXT4_DESC_PER_BLOCK_BITS(sb); |
2296 | ||
2297 | /* | |
2298 | * This is the total number of blocks used by GDT including | |
2299 | * the number of reserved blocks for GDT. | |
2300 | * The s_group_info array is allocated with this value | |
2301 | * to allow a clean online resize without a complex | |
2302 | * manipulation of pointer. | |
2303 | * The drawback is the unused memory when no resize | |
2304 | * occurs but it's very low in terms of pages | |
2305 | * (see comments below) | |
2306 | * Need to handle this properly when META_BG resizing is allowed | |
2307 | */ | |
2308 | num_meta_group_infos_max = num_meta_group_infos + | |
2309 | le16_to_cpu(es->s_reserved_gdt_blocks); | |
c9de560d | 2310 | |
5f21b0e6 FB |
2311 | /* |
2312 | * array_size is the size of s_group_info array. We round it | |
2313 | * to the next power of two because this approximation is done | |
2314 | * internally by kmalloc so we can have some more memory | |
2315 | * for free here (e.g. may be used for META_BG resize). | |
2316 | */ | |
2317 | array_size = 1; | |
2318 | while (array_size < sizeof(*sbi->s_group_info) * | |
2319 | num_meta_group_infos_max) | |
2320 | array_size = array_size << 1; | |
c9de560d AT |
2321 | /* An 8TB filesystem with 64-bit pointers requires a 4096 byte |
2322 | * kmalloc. A 128kb malloc should suffice for a 256TB filesystem. | |
2323 | * So a two level scheme suffices for now. */ | |
5f21b0e6 | 2324 | sbi->s_group_info = kmalloc(array_size, GFP_KERNEL); |
c9de560d AT |
2325 | if (sbi->s_group_info == NULL) { |
2326 | printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n"); | |
2327 | return -ENOMEM; | |
2328 | } | |
2329 | sbi->s_buddy_cache = new_inode(sb); | |
2330 | if (sbi->s_buddy_cache == NULL) { | |
2331 | printk(KERN_ERR "EXT4-fs: can't get new inode\n"); | |
2332 | goto err_freesgi; | |
2333 | } | |
2334 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; | |
8df9675f | 2335 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2336 | desc = ext4_get_group_desc(sb, i, NULL); |
2337 | if (desc == NULL) { | |
2338 | printk(KERN_ERR | |
a9df9a49 | 2339 | "EXT4-fs: can't read descriptor %u\n", i); |
c9de560d AT |
2340 | goto err_freebuddy; |
2341 | } | |
5f21b0e6 FB |
2342 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2343 | goto err_freebuddy; | |
c9de560d AT |
2344 | } |
2345 | ||
2346 | return 0; | |
2347 | ||
2348 | err_freebuddy: | |
f1fa3342 | 2349 | while (i-- > 0) |
c9de560d | 2350 | kfree(ext4_get_group_info(sb, i)); |
c9de560d | 2351 | i = num_meta_group_infos; |
f1fa3342 | 2352 | while (i-- > 0) |
c9de560d AT |
2353 | kfree(sbi->s_group_info[i]); |
2354 | iput(sbi->s_buddy_cache); | |
2355 | err_freesgi: | |
2356 | kfree(sbi->s_group_info); | |
2357 | return -ENOMEM; | |
2358 | } | |
2359 | ||
2360 | int ext4_mb_init(struct super_block *sb, int needs_recovery) | |
2361 | { | |
2362 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2363 | unsigned i, j; |
c9de560d AT |
2364 | unsigned offset; |
2365 | unsigned max; | |
74767c5a | 2366 | int ret; |
c9de560d | 2367 | |
1927805e | 2368 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
2369 | |
2370 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2371 | if (sbi->s_mb_offsets == NULL) { | |
c9de560d AT |
2372 | return -ENOMEM; |
2373 | } | |
ff7ef329 | 2374 | |
1927805e | 2375 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
2376 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2377 | if (sbi->s_mb_maxs == NULL) { | |
a7b19448 | 2378 | kfree(sbi->s_mb_offsets); |
c9de560d AT |
2379 | return -ENOMEM; |
2380 | } | |
2381 | ||
2382 | /* order 0 is regular bitmap */ | |
2383 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2384 | sbi->s_mb_offsets[0] = 0; | |
2385 | ||
2386 | i = 1; | |
2387 | offset = 0; | |
2388 | max = sb->s_blocksize << 2; | |
2389 | do { | |
2390 | sbi->s_mb_offsets[i] = offset; | |
2391 | sbi->s_mb_maxs[i] = max; | |
2392 | offset += 1 << (sb->s_blocksize_bits - i); | |
2393 | max = max >> 1; | |
2394 | i++; | |
2395 | } while (i <= sb->s_blocksize_bits + 1); | |
2396 | ||
2397 | /* init file for buddy data */ | |
74767c5a SF |
2398 | ret = ext4_mb_init_backend(sb); |
2399 | if (ret != 0) { | |
c9de560d AT |
2400 | kfree(sbi->s_mb_offsets); |
2401 | kfree(sbi->s_mb_maxs); | |
74767c5a | 2402 | return ret; |
c9de560d AT |
2403 | } |
2404 | ||
2405 | spin_lock_init(&sbi->s_md_lock); | |
c9de560d AT |
2406 | spin_lock_init(&sbi->s_bal_lock); |
2407 | ||
2408 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2409 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2410 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2411 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2412 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
c9de560d AT |
2413 | sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC; |
2414 | ||
730c213c | 2415 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2416 | if (sbi->s_locality_groups == NULL) { |
c9de560d AT |
2417 | kfree(sbi->s_mb_offsets); |
2418 | kfree(sbi->s_mb_maxs); | |
2419 | return -ENOMEM; | |
2420 | } | |
730c213c | 2421 | for_each_possible_cpu(i) { |
c9de560d | 2422 | struct ext4_locality_group *lg; |
730c213c | 2423 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2424 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2425 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2426 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2427 | spin_lock_init(&lg->lg_prealloc_lock); |
2428 | } | |
2429 | ||
296c355c TT |
2430 | if (sbi->s_proc) |
2431 | proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, | |
2432 | &ext4_mb_seq_groups_fops, sb); | |
c9de560d | 2433 | |
0390131b FM |
2434 | if (sbi->s_journal) |
2435 | sbi->s_journal->j_commit_callback = release_blocks_on_commit; | |
c9de560d AT |
2436 | return 0; |
2437 | } | |
2438 | ||
955ce5f5 | 2439 | /* need to called with the ext4 group lock held */ |
c9de560d AT |
2440 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
2441 | { | |
2442 | struct ext4_prealloc_space *pa; | |
2443 | struct list_head *cur, *tmp; | |
2444 | int count = 0; | |
2445 | ||
2446 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2447 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2448 | list_del(&pa->pa_group_list); | |
2449 | count++; | |
688f05a0 | 2450 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2451 | } |
2452 | if (count) | |
6ba495e9 | 2453 | mb_debug(1, "mballoc: %u PAs left\n", count); |
c9de560d AT |
2454 | |
2455 | } | |
2456 | ||
2457 | int ext4_mb_release(struct super_block *sb) | |
2458 | { | |
8df9675f | 2459 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2460 | ext4_group_t i; |
2461 | int num_meta_group_infos; | |
2462 | struct ext4_group_info *grinfo; | |
2463 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2464 | ||
c9de560d | 2465 | if (sbi->s_group_info) { |
8df9675f | 2466 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2467 | grinfo = ext4_get_group_info(sb, i); |
2468 | #ifdef DOUBLE_CHECK | |
2469 | kfree(grinfo->bb_bitmap); | |
2470 | #endif | |
2471 | ext4_lock_group(sb, i); | |
2472 | ext4_mb_cleanup_pa(grinfo); | |
2473 | ext4_unlock_group(sb, i); | |
2474 | kfree(grinfo); | |
2475 | } | |
8df9675f | 2476 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2477 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2478 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2479 | for (i = 0; i < num_meta_group_infos; i++) | |
2480 | kfree(sbi->s_group_info[i]); | |
2481 | kfree(sbi->s_group_info); | |
2482 | } | |
2483 | kfree(sbi->s_mb_offsets); | |
2484 | kfree(sbi->s_mb_maxs); | |
2485 | if (sbi->s_buddy_cache) | |
2486 | iput(sbi->s_buddy_cache); | |
2487 | if (sbi->s_mb_stats) { | |
2488 | printk(KERN_INFO | |
2489 | "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n", | |
2490 | atomic_read(&sbi->s_bal_allocated), | |
2491 | atomic_read(&sbi->s_bal_reqs), | |
2492 | atomic_read(&sbi->s_bal_success)); | |
2493 | printk(KERN_INFO | |
2494 | "EXT4-fs: mballoc: %u extents scanned, %u goal hits, " | |
2495 | "%u 2^N hits, %u breaks, %u lost\n", | |
2496 | atomic_read(&sbi->s_bal_ex_scanned), | |
2497 | atomic_read(&sbi->s_bal_goals), | |
2498 | atomic_read(&sbi->s_bal_2orders), | |
2499 | atomic_read(&sbi->s_bal_breaks), | |
2500 | atomic_read(&sbi->s_mb_lost_chunks)); | |
2501 | printk(KERN_INFO | |
2502 | "EXT4-fs: mballoc: %lu generated and it took %Lu\n", | |
2503 | sbi->s_mb_buddies_generated++, | |
2504 | sbi->s_mb_generation_time); | |
2505 | printk(KERN_INFO | |
2506 | "EXT4-fs: mballoc: %u preallocated, %u discarded\n", | |
2507 | atomic_read(&sbi->s_mb_preallocated), | |
2508 | atomic_read(&sbi->s_mb_discarded)); | |
2509 | } | |
2510 | ||
730c213c | 2511 | free_percpu(sbi->s_locality_groups); |
296c355c TT |
2512 | if (sbi->s_proc) |
2513 | remove_proc_entry("mb_groups", sbi->s_proc); | |
c9de560d AT |
2514 | |
2515 | return 0; | |
2516 | } | |
2517 | ||
3e624fc7 TT |
2518 | /* |
2519 | * This function is called by the jbd2 layer once the commit has finished, | |
2520 | * so we know we can free the blocks that were released with that commit. | |
2521 | */ | |
2522 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn) | |
c9de560d | 2523 | { |
3e624fc7 | 2524 | struct super_block *sb = journal->j_private; |
c9de560d | 2525 | struct ext4_buddy e4b; |
c894058d | 2526 | struct ext4_group_info *db; |
c894058d AK |
2527 | int err, count = 0, count2 = 0; |
2528 | struct ext4_free_data *entry; | |
3e624fc7 | 2529 | struct list_head *l, *ltmp; |
c9de560d | 2530 | |
3e624fc7 TT |
2531 | list_for_each_safe(l, ltmp, &txn->t_private_list) { |
2532 | entry = list_entry(l, struct ext4_free_data, list); | |
c9de560d | 2533 | |
6ba495e9 | 2534 | mb_debug(1, "gonna free %u blocks in group %u (0x%p):", |
3e624fc7 | 2535 | entry->count, entry->group, entry); |
c9de560d | 2536 | |
c894058d | 2537 | err = ext4_mb_load_buddy(sb, entry->group, &e4b); |
c9de560d AT |
2538 | /* we expect to find existing buddy because it's pinned */ |
2539 | BUG_ON(err != 0); | |
2540 | ||
c894058d | 2541 | db = e4b.bd_info; |
c9de560d | 2542 | /* there are blocks to put in buddy to make them really free */ |
c894058d | 2543 | count += entry->count; |
c9de560d | 2544 | count2++; |
c894058d AK |
2545 | ext4_lock_group(sb, entry->group); |
2546 | /* Take it out of per group rb tree */ | |
2547 | rb_erase(&entry->node, &(db->bb_free_root)); | |
2548 | mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count); | |
2549 | ||
2550 | if (!db->bb_free_root.rb_node) { | |
2551 | /* No more items in the per group rb tree | |
2552 | * balance refcounts from ext4_mb_free_metadata() | |
2553 | */ | |
2554 | page_cache_release(e4b.bd_buddy_page); | |
2555 | page_cache_release(e4b.bd_bitmap_page); | |
c9de560d | 2556 | } |
c894058d | 2557 | ext4_unlock_group(sb, entry->group); |
5328e635 ES |
2558 | if (test_opt(sb, DISCARD)) { |
2559 | ext4_fsblk_t discard_block; | |
5328e635 | 2560 | |
5661bd68 AM |
2561 | discard_block = entry->start_blk + |
2562 | ext4_group_first_block_no(sb, entry->group); | |
5328e635 ES |
2563 | trace_ext4_discard_blocks(sb, |
2564 | (unsigned long long)discard_block, | |
2565 | entry->count); | |
2566 | sb_issue_discard(sb, discard_block, entry->count); | |
2567 | } | |
c894058d | 2568 | kmem_cache_free(ext4_free_ext_cachep, entry); |
c9de560d | 2569 | ext4_mb_release_desc(&e4b); |
3e624fc7 | 2570 | } |
c9de560d | 2571 | |
6ba495e9 | 2572 | mb_debug(1, "freed %u blocks in %u structures\n", count, count2); |
c9de560d AT |
2573 | } |
2574 | ||
6ba495e9 TT |
2575 | #ifdef CONFIG_EXT4_DEBUG |
2576 | u8 mb_enable_debug __read_mostly; | |
2577 | ||
2578 | static struct dentry *debugfs_dir; | |
2579 | static struct dentry *debugfs_debug; | |
2580 | ||
2581 | static void __init ext4_create_debugfs_entry(void) | |
2582 | { | |
2583 | debugfs_dir = debugfs_create_dir("ext4", NULL); | |
2584 | if (debugfs_dir) | |
2585 | debugfs_debug = debugfs_create_u8("mballoc-debug", | |
2586 | S_IRUGO | S_IWUSR, | |
2587 | debugfs_dir, | |
2588 | &mb_enable_debug); | |
2589 | } | |
2590 | ||
2591 | static void ext4_remove_debugfs_entry(void) | |
2592 | { | |
2593 | debugfs_remove(debugfs_debug); | |
2594 | debugfs_remove(debugfs_dir); | |
2595 | } | |
2596 | ||
2597 | #else | |
2598 | ||
2599 | static void __init ext4_create_debugfs_entry(void) | |
2600 | { | |
2601 | } | |
2602 | ||
2603 | static void ext4_remove_debugfs_entry(void) | |
2604 | { | |
2605 | } | |
2606 | ||
2607 | #endif | |
2608 | ||
c9de560d AT |
2609 | int __init init_ext4_mballoc(void) |
2610 | { | |
2611 | ext4_pspace_cachep = | |
2612 | kmem_cache_create("ext4_prealloc_space", | |
2613 | sizeof(struct ext4_prealloc_space), | |
2614 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2615 | if (ext4_pspace_cachep == NULL) | |
2616 | return -ENOMEM; | |
2617 | ||
256bdb49 ES |
2618 | ext4_ac_cachep = |
2619 | kmem_cache_create("ext4_alloc_context", | |
2620 | sizeof(struct ext4_allocation_context), | |
2621 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2622 | if (ext4_ac_cachep == NULL) { | |
2623 | kmem_cache_destroy(ext4_pspace_cachep); | |
2624 | return -ENOMEM; | |
2625 | } | |
c894058d AK |
2626 | |
2627 | ext4_free_ext_cachep = | |
2628 | kmem_cache_create("ext4_free_block_extents", | |
2629 | sizeof(struct ext4_free_data), | |
2630 | 0, SLAB_RECLAIM_ACCOUNT, NULL); | |
2631 | if (ext4_free_ext_cachep == NULL) { | |
2632 | kmem_cache_destroy(ext4_pspace_cachep); | |
2633 | kmem_cache_destroy(ext4_ac_cachep); | |
2634 | return -ENOMEM; | |
2635 | } | |
6ba495e9 | 2636 | ext4_create_debugfs_entry(); |
c9de560d AT |
2637 | return 0; |
2638 | } | |
2639 | ||
2640 | void exit_ext4_mballoc(void) | |
2641 | { | |
3e03f9ca JDB |
2642 | /* |
2643 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep | |
2644 | * before destroying the slab cache. | |
2645 | */ | |
2646 | rcu_barrier(); | |
c9de560d | 2647 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 2648 | kmem_cache_destroy(ext4_ac_cachep); |
c894058d | 2649 | kmem_cache_destroy(ext4_free_ext_cachep); |
6ba495e9 | 2650 | ext4_remove_debugfs_entry(); |
c9de560d AT |
2651 | } |
2652 | ||
2653 | ||
2654 | /* | |
2655 | * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps | |
2656 | * Returns 0 if success or error code | |
2657 | */ | |
4ddfef7b ES |
2658 | static noinline_for_stack int |
2659 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
498e5f24 | 2660 | handle_t *handle, unsigned int reserv_blks) |
c9de560d AT |
2661 | { |
2662 | struct buffer_head *bitmap_bh = NULL; | |
2663 | struct ext4_super_block *es; | |
2664 | struct ext4_group_desc *gdp; | |
2665 | struct buffer_head *gdp_bh; | |
2666 | struct ext4_sb_info *sbi; | |
2667 | struct super_block *sb; | |
2668 | ext4_fsblk_t block; | |
519deca0 | 2669 | int err, len; |
c9de560d AT |
2670 | |
2671 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2672 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2673 | ||
2674 | sb = ac->ac_sb; | |
2675 | sbi = EXT4_SB(sb); | |
2676 | es = sbi->s_es; | |
2677 | ||
c9de560d AT |
2678 | |
2679 | err = -EIO; | |
574ca174 | 2680 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
c9de560d AT |
2681 | if (!bitmap_bh) |
2682 | goto out_err; | |
2683 | ||
2684 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2685 | if (err) | |
2686 | goto out_err; | |
2687 | ||
2688 | err = -EIO; | |
2689 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
2690 | if (!gdp) | |
2691 | goto out_err; | |
2692 | ||
a9df9a49 | 2693 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
9fd9784c | 2694 | ext4_free_blks_count(sb, gdp)); |
03cddb80 | 2695 | |
c9de560d AT |
2696 | err = ext4_journal_get_write_access(handle, gdp_bh); |
2697 | if (err) | |
2698 | goto out_err; | |
2699 | ||
2700 | block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb) | |
2701 | + ac->ac_b_ex.fe_start | |
2702 | + le32_to_cpu(es->s_first_data_block); | |
2703 | ||
519deca0 | 2704 | len = ac->ac_b_ex.fe_len; |
6fd058f7 | 2705 | if (!ext4_data_block_valid(sbi, block, len)) { |
12062ddd | 2706 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
6fd058f7 | 2707 | "fs metadata\n", block, block+len); |
519deca0 AK |
2708 | /* File system mounted not to panic on error |
2709 | * Fix the bitmap and repeat the block allocation | |
2710 | * We leak some of the blocks here. | |
2711 | */ | |
955ce5f5 AK |
2712 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
2713 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, | |
2714 | ac->ac_b_ex.fe_len); | |
2715 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
0390131b | 2716 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 AK |
2717 | if (!err) |
2718 | err = -EAGAIN; | |
2719 | goto out_err; | |
c9de560d | 2720 | } |
955ce5f5 AK |
2721 | |
2722 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
2723 | #ifdef AGGRESSIVE_CHECK |
2724 | { | |
2725 | int i; | |
2726 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
2727 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
2728 | bitmap_bh->b_data)); | |
2729 | } | |
2730 | } | |
2731 | #endif | |
955ce5f5 | 2732 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len); |
c9de560d AT |
2733 | if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
2734 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
560671a0 AK |
2735 | ext4_free_blks_set(sb, gdp, |
2736 | ext4_free_blocks_after_init(sb, | |
2737 | ac->ac_b_ex.fe_group, gdp)); | |
c9de560d | 2738 | } |
560671a0 AK |
2739 | len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len; |
2740 | ext4_free_blks_set(sb, gdp, len); | |
c9de560d | 2741 | gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
2742 | |
2743 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
6bc6e63f | 2744 | percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 2745 | /* |
6bc6e63f | 2746 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 2747 | */ |
6bc6e63f AK |
2748 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
2749 | /* release all the reserved blocks if non delalloc */ | |
2750 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks); | |
c9de560d | 2751 | |
772cb7c8 JS |
2752 | if (sbi->s_log_groups_per_flex) { |
2753 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
2754 | ac->ac_b_ex.fe_group); | |
9f24e420 TT |
2755 | atomic_sub(ac->ac_b_ex.fe_len, |
2756 | &sbi->s_flex_groups[flex_group].free_blocks); | |
772cb7c8 JS |
2757 | } |
2758 | ||
0390131b | 2759 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
2760 | if (err) |
2761 | goto out_err; | |
0390131b | 2762 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
2763 | |
2764 | out_err: | |
2765 | sb->s_dirt = 1; | |
42a10add | 2766 | brelse(bitmap_bh); |
c9de560d AT |
2767 | return err; |
2768 | } | |
2769 | ||
2770 | /* | |
2771 | * here we normalize request for locality group | |
2772 | * Group request are normalized to s_strip size if we set the same via mount | |
2773 | * option. If not we set it to s_mb_group_prealloc which can be configured via | |
b713a5ec | 2774 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
2775 | * |
2776 | * XXX: should we try to preallocate more than the group has now? | |
2777 | */ | |
2778 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
2779 | { | |
2780 | struct super_block *sb = ac->ac_sb; | |
2781 | struct ext4_locality_group *lg = ac->ac_lg; | |
2782 | ||
2783 | BUG_ON(lg == NULL); | |
2784 | if (EXT4_SB(sb)->s_stripe) | |
2785 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe; | |
2786 | else | |
2787 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; | |
6ba495e9 | 2788 | mb_debug(1, "#%u: goal %u blocks for locality group\n", |
c9de560d AT |
2789 | current->pid, ac->ac_g_ex.fe_len); |
2790 | } | |
2791 | ||
2792 | /* | |
2793 | * Normalization means making request better in terms of | |
2794 | * size and alignment | |
2795 | */ | |
4ddfef7b ES |
2796 | static noinline_for_stack void |
2797 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
2798 | struct ext4_allocation_request *ar) |
2799 | { | |
2800 | int bsbits, max; | |
2801 | ext4_lblk_t end; | |
c9de560d AT |
2802 | loff_t size, orig_size, start_off; |
2803 | ext4_lblk_t start, orig_start; | |
2804 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); | |
9a0762c5 | 2805 | struct ext4_prealloc_space *pa; |
c9de560d AT |
2806 | |
2807 | /* do normalize only data requests, metadata requests | |
2808 | do not need preallocation */ | |
2809 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
2810 | return; | |
2811 | ||
2812 | /* sometime caller may want exact blocks */ | |
2813 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2814 | return; | |
2815 | ||
2816 | /* caller may indicate that preallocation isn't | |
2817 | * required (it's a tail, for example) */ | |
2818 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
2819 | return; | |
2820 | ||
2821 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
2822 | ext4_mb_normalize_group_request(ac); | |
2823 | return ; | |
2824 | } | |
2825 | ||
2826 | bsbits = ac->ac_sb->s_blocksize_bits; | |
2827 | ||
2828 | /* first, let's learn actual file size | |
2829 | * given current request is allocated */ | |
2830 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; | |
2831 | size = size << bsbits; | |
2832 | if (size < i_size_read(ac->ac_inode)) | |
2833 | size = i_size_read(ac->ac_inode); | |
2834 | ||
1930479c VC |
2835 | /* max size of free chunks */ |
2836 | max = 2 << bsbits; | |
c9de560d | 2837 | |
1930479c VC |
2838 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
2839 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
2840 | |
2841 | /* first, try to predict filesize */ | |
2842 | /* XXX: should this table be tunable? */ | |
2843 | start_off = 0; | |
2844 | if (size <= 16 * 1024) { | |
2845 | size = 16 * 1024; | |
2846 | } else if (size <= 32 * 1024) { | |
2847 | size = 32 * 1024; | |
2848 | } else if (size <= 64 * 1024) { | |
2849 | size = 64 * 1024; | |
2850 | } else if (size <= 128 * 1024) { | |
2851 | size = 128 * 1024; | |
2852 | } else if (size <= 256 * 1024) { | |
2853 | size = 256 * 1024; | |
2854 | } else if (size <= 512 * 1024) { | |
2855 | size = 512 * 1024; | |
2856 | } else if (size <= 1024 * 1024) { | |
2857 | size = 1024 * 1024; | |
1930479c | 2858 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 2859 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
2860 | (21 - bsbits)) << 21; |
2861 | size = 2 * 1024 * 1024; | |
2862 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
2863 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
2864 | (22 - bsbits)) << 22; | |
2865 | size = 4 * 1024 * 1024; | |
2866 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 2867 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
2868 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
2869 | (23 - bsbits)) << 23; | |
2870 | size = 8 * 1024 * 1024; | |
2871 | } else { | |
2872 | start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; | |
2873 | size = ac->ac_o_ex.fe_len << bsbits; | |
2874 | } | |
2875 | orig_size = size = size >> bsbits; | |
2876 | orig_start = start = start_off >> bsbits; | |
2877 | ||
2878 | /* don't cover already allocated blocks in selected range */ | |
2879 | if (ar->pleft && start <= ar->lleft) { | |
2880 | size -= ar->lleft + 1 - start; | |
2881 | start = ar->lleft + 1; | |
2882 | } | |
2883 | if (ar->pright && start + size - 1 >= ar->lright) | |
2884 | size -= start + size - ar->lright; | |
2885 | ||
2886 | end = start + size; | |
2887 | ||
2888 | /* check we don't cross already preallocated blocks */ | |
2889 | rcu_read_lock(); | |
9a0762c5 | 2890 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 2891 | ext4_lblk_t pa_end; |
c9de560d | 2892 | |
c9de560d AT |
2893 | if (pa->pa_deleted) |
2894 | continue; | |
2895 | spin_lock(&pa->pa_lock); | |
2896 | if (pa->pa_deleted) { | |
2897 | spin_unlock(&pa->pa_lock); | |
2898 | continue; | |
2899 | } | |
2900 | ||
2901 | pa_end = pa->pa_lstart + pa->pa_len; | |
2902 | ||
2903 | /* PA must not overlap original request */ | |
2904 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
2905 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
2906 | ||
38877f4e ES |
2907 | /* skip PAs this normalized request doesn't overlap with */ |
2908 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
2909 | spin_unlock(&pa->pa_lock); |
2910 | continue; | |
2911 | } | |
2912 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
2913 | ||
38877f4e | 2914 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
2915 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
2916 | BUG_ON(pa_end < start); | |
2917 | start = pa_end; | |
38877f4e | 2918 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
2919 | BUG_ON(pa->pa_lstart > end); |
2920 | end = pa->pa_lstart; | |
2921 | } | |
2922 | spin_unlock(&pa->pa_lock); | |
2923 | } | |
2924 | rcu_read_unlock(); | |
2925 | size = end - start; | |
2926 | ||
2927 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
2928 | rcu_read_lock(); | |
9a0762c5 | 2929 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 2930 | ext4_lblk_t pa_end; |
c9de560d AT |
2931 | spin_lock(&pa->pa_lock); |
2932 | if (pa->pa_deleted == 0) { | |
2933 | pa_end = pa->pa_lstart + pa->pa_len; | |
2934 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); | |
2935 | } | |
2936 | spin_unlock(&pa->pa_lock); | |
2937 | } | |
2938 | rcu_read_unlock(); | |
2939 | ||
2940 | if (start + size <= ac->ac_o_ex.fe_logical && | |
2941 | start > ac->ac_o_ex.fe_logical) { | |
2942 | printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n", | |
2943 | (unsigned long) start, (unsigned long) size, | |
2944 | (unsigned long) ac->ac_o_ex.fe_logical); | |
2945 | } | |
2946 | BUG_ON(start + size <= ac->ac_o_ex.fe_logical && | |
2947 | start > ac->ac_o_ex.fe_logical); | |
8d03c7a0 | 2948 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
2949 | |
2950 | /* now prepare goal request */ | |
2951 | ||
2952 | /* XXX: is it better to align blocks WRT to logical | |
2953 | * placement or satisfy big request as is */ | |
2954 | ac->ac_g_ex.fe_logical = start; | |
2955 | ac->ac_g_ex.fe_len = size; | |
2956 | ||
2957 | /* define goal start in order to merge */ | |
2958 | if (ar->pright && (ar->lright == (start + size))) { | |
2959 | /* merge to the right */ | |
2960 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
2961 | &ac->ac_f_ex.fe_group, | |
2962 | &ac->ac_f_ex.fe_start); | |
2963 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
2964 | } | |
2965 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
2966 | /* merge to the left */ | |
2967 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
2968 | &ac->ac_f_ex.fe_group, | |
2969 | &ac->ac_f_ex.fe_start); | |
2970 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
2971 | } | |
2972 | ||
6ba495e9 | 2973 | mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, |
c9de560d AT |
2974 | (unsigned) orig_size, (unsigned) start); |
2975 | } | |
2976 | ||
2977 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
2978 | { | |
2979 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2980 | ||
2981 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
2982 | atomic_inc(&sbi->s_bal_reqs); | |
2983 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
2984 | if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len) | |
2985 | atomic_inc(&sbi->s_bal_success); | |
2986 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
2987 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
2988 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
2989 | atomic_inc(&sbi->s_bal_goals); | |
2990 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
2991 | atomic_inc(&sbi->s_bal_breaks); | |
2992 | } | |
2993 | ||
296c355c TT |
2994 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
2995 | trace_ext4_mballoc_alloc(ac); | |
2996 | else | |
2997 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
2998 | } |
2999 | ||
b844167e CW |
3000 | /* |
3001 | * Called on failure; free up any blocks from the inode PA for this | |
3002 | * context. We don't need this for MB_GROUP_PA because we only change | |
3003 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
3004 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
3005 | */ | |
3006 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
3007 | { | |
3008 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
3009 | int len; | |
3010 | ||
3011 | if (pa && pa->pa_type == MB_INODE_PA) { | |
3012 | len = ac->ac_b_ex.fe_len; | |
3013 | pa->pa_free += len; | |
3014 | } | |
3015 | ||
3016 | } | |
3017 | ||
c9de560d AT |
3018 | /* |
3019 | * use blocks preallocated to inode | |
3020 | */ | |
3021 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3022 | struct ext4_prealloc_space *pa) | |
3023 | { | |
3024 | ext4_fsblk_t start; | |
3025 | ext4_fsblk_t end; | |
3026 | int len; | |
3027 | ||
3028 | /* found preallocated blocks, use them */ | |
3029 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
3030 | end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len); | |
3031 | len = end - start; | |
3032 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, | |
3033 | &ac->ac_b_ex.fe_start); | |
3034 | ac->ac_b_ex.fe_len = len; | |
3035 | ac->ac_status = AC_STATUS_FOUND; | |
3036 | ac->ac_pa = pa; | |
3037 | ||
3038 | BUG_ON(start < pa->pa_pstart); | |
3039 | BUG_ON(start + len > pa->pa_pstart + pa->pa_len); | |
3040 | BUG_ON(pa->pa_free < len); | |
3041 | pa->pa_free -= len; | |
3042 | ||
6ba495e9 | 3043 | mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); |
c9de560d AT |
3044 | } |
3045 | ||
3046 | /* | |
3047 | * use blocks preallocated to locality group | |
3048 | */ | |
3049 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3050 | struct ext4_prealloc_space *pa) | |
3051 | { | |
03cddb80 | 3052 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3053 | |
c9de560d AT |
3054 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3055 | &ac->ac_b_ex.fe_group, | |
3056 | &ac->ac_b_ex.fe_start); | |
3057 | ac->ac_b_ex.fe_len = len; | |
3058 | ac->ac_status = AC_STATUS_FOUND; | |
3059 | ac->ac_pa = pa; | |
3060 | ||
3061 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3062 | * possible race when the group is being loaded concurrently |
c9de560d | 3063 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3064 | * in on-disk bitmap -- see ext4_mb_release_context() |
3065 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 3066 | */ |
6ba495e9 | 3067 | mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); |
c9de560d AT |
3068 | } |
3069 | ||
5e745b04 AK |
3070 | /* |
3071 | * Return the prealloc space that have minimal distance | |
3072 | * from the goal block. @cpa is the prealloc | |
3073 | * space that is having currently known minimal distance | |
3074 | * from the goal block. | |
3075 | */ | |
3076 | static struct ext4_prealloc_space * | |
3077 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3078 | struct ext4_prealloc_space *pa, | |
3079 | struct ext4_prealloc_space *cpa) | |
3080 | { | |
3081 | ext4_fsblk_t cur_distance, new_distance; | |
3082 | ||
3083 | if (cpa == NULL) { | |
3084 | atomic_inc(&pa->pa_count); | |
3085 | return pa; | |
3086 | } | |
3087 | cur_distance = abs(goal_block - cpa->pa_pstart); | |
3088 | new_distance = abs(goal_block - pa->pa_pstart); | |
3089 | ||
3090 | if (cur_distance < new_distance) | |
3091 | return cpa; | |
3092 | ||
3093 | /* drop the previous reference */ | |
3094 | atomic_dec(&cpa->pa_count); | |
3095 | atomic_inc(&pa->pa_count); | |
3096 | return pa; | |
3097 | } | |
3098 | ||
c9de560d AT |
3099 | /* |
3100 | * search goal blocks in preallocated space | |
3101 | */ | |
4ddfef7b ES |
3102 | static noinline_for_stack int |
3103 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
c9de560d | 3104 | { |
6be2ded1 | 3105 | int order, i; |
c9de560d AT |
3106 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3107 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3108 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3109 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3110 | |
3111 | /* only data can be preallocated */ | |
3112 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3113 | return 0; | |
3114 | ||
3115 | /* first, try per-file preallocation */ | |
3116 | rcu_read_lock(); | |
9a0762c5 | 3117 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3118 | |
3119 | /* all fields in this condition don't change, | |
3120 | * so we can skip locking for them */ | |
3121 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
3122 | ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len) | |
3123 | continue; | |
3124 | ||
fb0a387d ES |
3125 | /* non-extent files can't have physical blocks past 2^32 */ |
3126 | if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL) && | |
3127 | pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS) | |
3128 | continue; | |
3129 | ||
c9de560d AT |
3130 | /* found preallocated blocks, use them */ |
3131 | spin_lock(&pa->pa_lock); | |
3132 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3133 | atomic_inc(&pa->pa_count); | |
3134 | ext4_mb_use_inode_pa(ac, pa); | |
3135 | spin_unlock(&pa->pa_lock); | |
3136 | ac->ac_criteria = 10; | |
3137 | rcu_read_unlock(); | |
3138 | return 1; | |
3139 | } | |
3140 | spin_unlock(&pa->pa_lock); | |
3141 | } | |
3142 | rcu_read_unlock(); | |
3143 | ||
3144 | /* can we use group allocation? */ | |
3145 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3146 | return 0; | |
3147 | ||
3148 | /* inode may have no locality group for some reason */ | |
3149 | lg = ac->ac_lg; | |
3150 | if (lg == NULL) | |
3151 | return 0; | |
6be2ded1 AK |
3152 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3153 | if (order > PREALLOC_TB_SIZE - 1) | |
3154 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3155 | order = PREALLOC_TB_SIZE - 1; | |
3156 | ||
5e745b04 AK |
3157 | goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) + |
3158 | ac->ac_g_ex.fe_start + | |
3159 | le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block); | |
3160 | /* | |
3161 | * search for the prealloc space that is having | |
3162 | * minimal distance from the goal block. | |
3163 | */ | |
6be2ded1 AK |
3164 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3165 | rcu_read_lock(); | |
3166 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3167 | pa_inode_list) { | |
3168 | spin_lock(&pa->pa_lock); | |
3169 | if (pa->pa_deleted == 0 && | |
3170 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3171 | |
3172 | cpa = ext4_mb_check_group_pa(goal_block, | |
3173 | pa, cpa); | |
6be2ded1 | 3174 | } |
c9de560d | 3175 | spin_unlock(&pa->pa_lock); |
c9de560d | 3176 | } |
6be2ded1 | 3177 | rcu_read_unlock(); |
c9de560d | 3178 | } |
5e745b04 AK |
3179 | if (cpa) { |
3180 | ext4_mb_use_group_pa(ac, cpa); | |
3181 | ac->ac_criteria = 20; | |
3182 | return 1; | |
3183 | } | |
c9de560d AT |
3184 | return 0; |
3185 | } | |
3186 | ||
7a2fcbf7 AK |
3187 | /* |
3188 | * the function goes through all block freed in the group | |
3189 | * but not yet committed and marks them used in in-core bitmap. | |
3190 | * buddy must be generated from this bitmap | |
955ce5f5 | 3191 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3192 | */ |
3193 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3194 | ext4_group_t group) | |
3195 | { | |
3196 | struct rb_node *n; | |
3197 | struct ext4_group_info *grp; | |
3198 | struct ext4_free_data *entry; | |
3199 | ||
3200 | grp = ext4_get_group_info(sb, group); | |
3201 | n = rb_first(&(grp->bb_free_root)); | |
3202 | ||
3203 | while (n) { | |
3204 | entry = rb_entry(n, struct ext4_free_data, node); | |
955ce5f5 | 3205 | mb_set_bits(bitmap, entry->start_blk, entry->count); |
7a2fcbf7 AK |
3206 | n = rb_next(n); |
3207 | } | |
3208 | return; | |
3209 | } | |
3210 | ||
c9de560d AT |
3211 | /* |
3212 | * the function goes through all preallocation in this group and marks them | |
3213 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3214 | * Need to be called with ext4 group lock held |
c9de560d | 3215 | */ |
089ceecc ES |
3216 | static noinline_for_stack |
3217 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
3218 | ext4_group_t group) |
3219 | { | |
3220 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3221 | struct ext4_prealloc_space *pa; | |
3222 | struct list_head *cur; | |
3223 | ext4_group_t groupnr; | |
3224 | ext4_grpblk_t start; | |
3225 | int preallocated = 0; | |
3226 | int count = 0; | |
3227 | int len; | |
3228 | ||
3229 | /* all form of preallocation discards first load group, | |
3230 | * so the only competing code is preallocation use. | |
3231 | * we don't need any locking here | |
3232 | * notice we do NOT ignore preallocations with pa_deleted | |
3233 | * otherwise we could leave used blocks available for | |
3234 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3235 | * is dropping preallocation | |
3236 | */ | |
3237 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3238 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3239 | spin_lock(&pa->pa_lock); | |
3240 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3241 | &groupnr, &start); | |
3242 | len = pa->pa_len; | |
3243 | spin_unlock(&pa->pa_lock); | |
3244 | if (unlikely(len == 0)) | |
3245 | continue; | |
3246 | BUG_ON(groupnr != group); | |
955ce5f5 | 3247 | mb_set_bits(bitmap, start, len); |
c9de560d AT |
3248 | preallocated += len; |
3249 | count++; | |
3250 | } | |
6ba495e9 | 3251 | mb_debug(1, "prellocated %u for group %u\n", preallocated, group); |
c9de560d AT |
3252 | } |
3253 | ||
3254 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3255 | { | |
3256 | struct ext4_prealloc_space *pa; | |
3257 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
3258 | kmem_cache_free(ext4_pspace_cachep, pa); | |
3259 | } | |
3260 | ||
3261 | /* | |
3262 | * drops a reference to preallocated space descriptor | |
3263 | * if this was the last reference and the space is consumed | |
3264 | */ | |
3265 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3266 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3267 | { | |
a9df9a49 | 3268 | ext4_group_t grp; |
d33a1976 | 3269 | ext4_fsblk_t grp_blk; |
c9de560d AT |
3270 | |
3271 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) | |
3272 | return; | |
3273 | ||
3274 | /* in this short window concurrent discard can set pa_deleted */ | |
3275 | spin_lock(&pa->pa_lock); | |
3276 | if (pa->pa_deleted == 1) { | |
3277 | spin_unlock(&pa->pa_lock); | |
3278 | return; | |
3279 | } | |
3280 | ||
3281 | pa->pa_deleted = 1; | |
3282 | spin_unlock(&pa->pa_lock); | |
3283 | ||
d33a1976 | 3284 | grp_blk = pa->pa_pstart; |
cc0fb9ad AK |
3285 | /* |
3286 | * If doing group-based preallocation, pa_pstart may be in the | |
3287 | * next group when pa is used up | |
3288 | */ | |
3289 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3290 | grp_blk--; |
3291 | ||
3292 | ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL); | |
c9de560d AT |
3293 | |
3294 | /* | |
3295 | * possible race: | |
3296 | * | |
3297 | * P1 (buddy init) P2 (regular allocation) | |
3298 | * find block B in PA | |
3299 | * copy on-disk bitmap to buddy | |
3300 | * mark B in on-disk bitmap | |
3301 | * drop PA from group | |
3302 | * mark all PAs in buddy | |
3303 | * | |
3304 | * thus, P1 initializes buddy with B available. to prevent this | |
3305 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3306 | * against that pair | |
3307 | */ | |
3308 | ext4_lock_group(sb, grp); | |
3309 | list_del(&pa->pa_group_list); | |
3310 | ext4_unlock_group(sb, grp); | |
3311 | ||
3312 | spin_lock(pa->pa_obj_lock); | |
3313 | list_del_rcu(&pa->pa_inode_list); | |
3314 | spin_unlock(pa->pa_obj_lock); | |
3315 | ||
3316 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3317 | } | |
3318 | ||
3319 | /* | |
3320 | * creates new preallocated space for given inode | |
3321 | */ | |
4ddfef7b ES |
3322 | static noinline_for_stack int |
3323 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3324 | { |
3325 | struct super_block *sb = ac->ac_sb; | |
3326 | struct ext4_prealloc_space *pa; | |
3327 | struct ext4_group_info *grp; | |
3328 | struct ext4_inode_info *ei; | |
3329 | ||
3330 | /* preallocate only when found space is larger then requested */ | |
3331 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3332 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3333 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3334 | ||
3335 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3336 | if (pa == NULL) | |
3337 | return -ENOMEM; | |
3338 | ||
3339 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3340 | int winl; | |
3341 | int wins; | |
3342 | int win; | |
3343 | int offs; | |
3344 | ||
3345 | /* we can't allocate as much as normalizer wants. | |
3346 | * so, found space must get proper lstart | |
3347 | * to cover original request */ | |
3348 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3349 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3350 | ||
3351 | /* we're limited by original request in that | |
3352 | * logical block must be covered any way | |
3353 | * winl is window we can move our chunk within */ | |
3354 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3355 | ||
3356 | /* also, we should cover whole original request */ | |
3357 | wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len; | |
3358 | ||
3359 | /* the smallest one defines real window */ | |
3360 | win = min(winl, wins); | |
3361 | ||
3362 | offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len; | |
3363 | if (offs && offs < win) | |
3364 | win = offs; | |
3365 | ||
3366 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win; | |
3367 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); | |
3368 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3369 | } | |
3370 | ||
3371 | /* preallocation can change ac_b_ex, thus we store actually | |
3372 | * allocated blocks for history */ | |
3373 | ac->ac_f_ex = ac->ac_b_ex; | |
3374 | ||
3375 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3376 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3377 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3378 | pa->pa_free = pa->pa_len; | |
3379 | atomic_set(&pa->pa_count, 1); | |
3380 | spin_lock_init(&pa->pa_lock); | |
d794bf8e AK |
3381 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3382 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3383 | pa->pa_deleted = 0; |
cc0fb9ad | 3384 | pa->pa_type = MB_INODE_PA; |
c9de560d | 3385 | |
6ba495e9 | 3386 | mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, |
c9de560d | 3387 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
9bffad1e | 3388 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3389 | |
3390 | ext4_mb_use_inode_pa(ac, pa); | |
3391 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3392 | ||
3393 | ei = EXT4_I(ac->ac_inode); | |
3394 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3395 | ||
3396 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3397 | pa->pa_inode = ac->ac_inode; | |
3398 | ||
3399 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3400 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3401 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3402 | ||
3403 | spin_lock(pa->pa_obj_lock); | |
3404 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3405 | spin_unlock(pa->pa_obj_lock); | |
3406 | ||
3407 | return 0; | |
3408 | } | |
3409 | ||
3410 | /* | |
3411 | * creates new preallocated space for locality group inodes belongs to | |
3412 | */ | |
4ddfef7b ES |
3413 | static noinline_for_stack int |
3414 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3415 | { |
3416 | struct super_block *sb = ac->ac_sb; | |
3417 | struct ext4_locality_group *lg; | |
3418 | struct ext4_prealloc_space *pa; | |
3419 | struct ext4_group_info *grp; | |
3420 | ||
3421 | /* preallocate only when found space is larger then requested */ | |
3422 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3423 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3424 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3425 | ||
3426 | BUG_ON(ext4_pspace_cachep == NULL); | |
3427 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3428 | if (pa == NULL) | |
3429 | return -ENOMEM; | |
3430 | ||
3431 | /* preallocation can change ac_b_ex, thus we store actually | |
3432 | * allocated blocks for history */ | |
3433 | ac->ac_f_ex = ac->ac_b_ex; | |
3434 | ||
3435 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3436 | pa->pa_lstart = pa->pa_pstart; | |
3437 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3438 | pa->pa_free = pa->pa_len; | |
3439 | atomic_set(&pa->pa_count, 1); | |
3440 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3441 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 3442 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 3443 | pa->pa_deleted = 0; |
cc0fb9ad | 3444 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 3445 | |
6ba495e9 | 3446 | mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, |
9bffad1e TT |
3447 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
3448 | trace_ext4_mb_new_group_pa(ac, pa); | |
c9de560d AT |
3449 | |
3450 | ext4_mb_use_group_pa(ac, pa); | |
3451 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3452 | ||
3453 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3454 | lg = ac->ac_lg; | |
3455 | BUG_ON(lg == NULL); | |
3456 | ||
3457 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3458 | pa->pa_inode = NULL; | |
3459 | ||
3460 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3461 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3462 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3463 | ||
6be2ded1 AK |
3464 | /* |
3465 | * We will later add the new pa to the right bucket | |
3466 | * after updating the pa_free in ext4_mb_release_context | |
3467 | */ | |
c9de560d AT |
3468 | return 0; |
3469 | } | |
3470 | ||
3471 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3472 | { | |
3473 | int err; | |
3474 | ||
3475 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3476 | err = ext4_mb_new_group_pa(ac); | |
3477 | else | |
3478 | err = ext4_mb_new_inode_pa(ac); | |
3479 | return err; | |
3480 | } | |
3481 | ||
3482 | /* | |
3483 | * finds all unused blocks in on-disk bitmap, frees them in | |
3484 | * in-core bitmap and buddy. | |
3485 | * @pa must be unlinked from inode and group lists, so that | |
3486 | * nobody else can find/use it. | |
3487 | * the caller MUST hold group/inode locks. | |
3488 | * TODO: optimize the case when there are no in-core structures yet | |
3489 | */ | |
4ddfef7b ES |
3490 | static noinline_for_stack int |
3491 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
c83617db AK |
3492 | struct ext4_prealloc_space *pa, |
3493 | struct ext4_allocation_context *ac) | |
c9de560d | 3494 | { |
c9de560d AT |
3495 | struct super_block *sb = e4b->bd_sb; |
3496 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
3497 | unsigned int end; |
3498 | unsigned int next; | |
c9de560d AT |
3499 | ext4_group_t group; |
3500 | ext4_grpblk_t bit; | |
ba80b101 | 3501 | unsigned long long grp_blk_start; |
c9de560d AT |
3502 | sector_t start; |
3503 | int err = 0; | |
3504 | int free = 0; | |
3505 | ||
3506 | BUG_ON(pa->pa_deleted == 0); | |
3507 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
ba80b101 | 3508 | grp_blk_start = pa->pa_pstart - bit; |
c9de560d AT |
3509 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
3510 | end = bit + pa->pa_len; | |
3511 | ||
256bdb49 ES |
3512 | if (ac) { |
3513 | ac->ac_sb = sb; | |
3514 | ac->ac_inode = pa->pa_inode; | |
256bdb49 | 3515 | } |
c9de560d AT |
3516 | |
3517 | while (bit < end) { | |
ffad0a44 | 3518 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3519 | if (bit >= end) |
3520 | break; | |
ffad0a44 | 3521 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
5661bd68 | 3522 | start = ext4_group_first_block_no(sb, group) + bit; |
6ba495e9 | 3523 | mb_debug(1, " free preallocated %u/%u in group %u\n", |
c9de560d AT |
3524 | (unsigned) start, (unsigned) next - bit, |
3525 | (unsigned) group); | |
3526 | free += next - bit; | |
3527 | ||
256bdb49 ES |
3528 | if (ac) { |
3529 | ac->ac_b_ex.fe_group = group; | |
3530 | ac->ac_b_ex.fe_start = bit; | |
3531 | ac->ac_b_ex.fe_len = next - bit; | |
3532 | ac->ac_b_ex.fe_logical = 0; | |
296c355c | 3533 | trace_ext4_mballoc_discard(ac); |
256bdb49 | 3534 | } |
c9de560d | 3535 | |
9bffad1e TT |
3536 | trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit, |
3537 | next - bit); | |
c9de560d AT |
3538 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
3539 | bit = next + 1; | |
3540 | } | |
3541 | if (free != pa->pa_free) { | |
26346ff6 | 3542 | printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n", |
c9de560d AT |
3543 | pa, (unsigned long) pa->pa_lstart, |
3544 | (unsigned long) pa->pa_pstart, | |
3545 | (unsigned long) pa->pa_len); | |
5d1b1b3f AK |
3546 | ext4_grp_locked_error(sb, group, |
3547 | __func__, "free %u, pa_free %u", | |
3548 | free, pa->pa_free); | |
e56eb659 AK |
3549 | /* |
3550 | * pa is already deleted so we use the value obtained | |
3551 | * from the bitmap and continue. | |
3552 | */ | |
c9de560d | 3553 | } |
c9de560d AT |
3554 | atomic_add(free, &sbi->s_mb_discarded); |
3555 | ||
3556 | return err; | |
3557 | } | |
3558 | ||
4ddfef7b ES |
3559 | static noinline_for_stack int |
3560 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
c83617db AK |
3561 | struct ext4_prealloc_space *pa, |
3562 | struct ext4_allocation_context *ac) | |
c9de560d | 3563 | { |
c9de560d AT |
3564 | struct super_block *sb = e4b->bd_sb; |
3565 | ext4_group_t group; | |
3566 | ext4_grpblk_t bit; | |
3567 | ||
9bffad1e | 3568 | trace_ext4_mb_release_group_pa(ac, pa); |
c9de560d AT |
3569 | BUG_ON(pa->pa_deleted == 0); |
3570 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3571 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3572 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3573 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3574 | ||
256bdb49 ES |
3575 | if (ac) { |
3576 | ac->ac_sb = sb; | |
3577 | ac->ac_inode = NULL; | |
3578 | ac->ac_b_ex.fe_group = group; | |
3579 | ac->ac_b_ex.fe_start = bit; | |
3580 | ac->ac_b_ex.fe_len = pa->pa_len; | |
3581 | ac->ac_b_ex.fe_logical = 0; | |
296c355c | 3582 | trace_ext4_mballoc_discard(ac); |
256bdb49 | 3583 | } |
c9de560d AT |
3584 | |
3585 | return 0; | |
3586 | } | |
3587 | ||
3588 | /* | |
3589 | * releases all preallocations in given group | |
3590 | * | |
3591 | * first, we need to decide discard policy: | |
3592 | * - when do we discard | |
3593 | * 1) ENOSPC | |
3594 | * - how many do we discard | |
3595 | * 1) how many requested | |
3596 | */ | |
4ddfef7b ES |
3597 | static noinline_for_stack int |
3598 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3599 | ext4_group_t group, int needed) |
3600 | { | |
3601 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3602 | struct buffer_head *bitmap_bh = NULL; | |
3603 | struct ext4_prealloc_space *pa, *tmp; | |
c83617db | 3604 | struct ext4_allocation_context *ac; |
c9de560d AT |
3605 | struct list_head list; |
3606 | struct ext4_buddy e4b; | |
3607 | int err; | |
3608 | int busy = 0; | |
3609 | int free = 0; | |
3610 | ||
6ba495e9 | 3611 | mb_debug(1, "discard preallocation for group %u\n", group); |
c9de560d AT |
3612 | |
3613 | if (list_empty(&grp->bb_prealloc_list)) | |
3614 | return 0; | |
3615 | ||
574ca174 | 3616 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3617 | if (bitmap_bh == NULL) { |
12062ddd | 3618 | ext4_error(sb, "Error reading block bitmap for %u", group); |
ce89f46c | 3619 | return 0; |
c9de560d AT |
3620 | } |
3621 | ||
3622 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3623 | if (err) { |
12062ddd | 3624 | ext4_error(sb, "Error loading buddy information for %u", group); |
ce89f46c AK |
3625 | put_bh(bitmap_bh); |
3626 | return 0; | |
3627 | } | |
c9de560d AT |
3628 | |
3629 | if (needed == 0) | |
3630 | needed = EXT4_BLOCKS_PER_GROUP(sb) + 1; | |
3631 | ||
c9de560d | 3632 | INIT_LIST_HEAD(&list); |
c83617db | 3633 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
9bffad1e TT |
3634 | if (ac) |
3635 | ac->ac_sb = sb; | |
c9de560d AT |
3636 | repeat: |
3637 | ext4_lock_group(sb, group); | |
3638 | list_for_each_entry_safe(pa, tmp, | |
3639 | &grp->bb_prealloc_list, pa_group_list) { | |
3640 | spin_lock(&pa->pa_lock); | |
3641 | if (atomic_read(&pa->pa_count)) { | |
3642 | spin_unlock(&pa->pa_lock); | |
3643 | busy = 1; | |
3644 | continue; | |
3645 | } | |
3646 | if (pa->pa_deleted) { | |
3647 | spin_unlock(&pa->pa_lock); | |
3648 | continue; | |
3649 | } | |
3650 | ||
3651 | /* seems this one can be freed ... */ | |
3652 | pa->pa_deleted = 1; | |
3653 | ||
3654 | /* we can trust pa_free ... */ | |
3655 | free += pa->pa_free; | |
3656 | ||
3657 | spin_unlock(&pa->pa_lock); | |
3658 | ||
3659 | list_del(&pa->pa_group_list); | |
3660 | list_add(&pa->u.pa_tmp_list, &list); | |
3661 | } | |
3662 | ||
3663 | /* if we still need more blocks and some PAs were used, try again */ | |
3664 | if (free < needed && busy) { | |
3665 | busy = 0; | |
3666 | ext4_unlock_group(sb, group); | |
3667 | /* | |
3668 | * Yield the CPU here so that we don't get soft lockup | |
3669 | * in non preempt case. | |
3670 | */ | |
3671 | yield(); | |
3672 | goto repeat; | |
3673 | } | |
3674 | ||
3675 | /* found anything to free? */ | |
3676 | if (list_empty(&list)) { | |
3677 | BUG_ON(free != 0); | |
3678 | goto out; | |
3679 | } | |
3680 | ||
3681 | /* now free all selected PAs */ | |
3682 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3683 | ||
3684 | /* remove from object (inode or locality group) */ | |
3685 | spin_lock(pa->pa_obj_lock); | |
3686 | list_del_rcu(&pa->pa_inode_list); | |
3687 | spin_unlock(pa->pa_obj_lock); | |
3688 | ||
cc0fb9ad | 3689 | if (pa->pa_type == MB_GROUP_PA) |
c83617db | 3690 | ext4_mb_release_group_pa(&e4b, pa, ac); |
c9de560d | 3691 | else |
c83617db | 3692 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); |
c9de560d AT |
3693 | |
3694 | list_del(&pa->u.pa_tmp_list); | |
3695 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3696 | } | |
3697 | ||
3698 | out: | |
3699 | ext4_unlock_group(sb, group); | |
c83617db AK |
3700 | if (ac) |
3701 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
3702 | ext4_mb_release_desc(&e4b); |
3703 | put_bh(bitmap_bh); | |
3704 | return free; | |
3705 | } | |
3706 | ||
3707 | /* | |
3708 | * releases all non-used preallocated blocks for given inode | |
3709 | * | |
3710 | * It's important to discard preallocations under i_data_sem | |
3711 | * We don't want another block to be served from the prealloc | |
3712 | * space when we are discarding the inode prealloc space. | |
3713 | * | |
3714 | * FIXME!! Make sure it is valid at all the call sites | |
3715 | */ | |
c2ea3fde | 3716 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
3717 | { |
3718 | struct ext4_inode_info *ei = EXT4_I(inode); | |
3719 | struct super_block *sb = inode->i_sb; | |
3720 | struct buffer_head *bitmap_bh = NULL; | |
3721 | struct ext4_prealloc_space *pa, *tmp; | |
c83617db | 3722 | struct ext4_allocation_context *ac; |
c9de560d AT |
3723 | ext4_group_t group = 0; |
3724 | struct list_head list; | |
3725 | struct ext4_buddy e4b; | |
3726 | int err; | |
3727 | ||
c2ea3fde | 3728 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
3729 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
3730 | return; | |
3731 | } | |
3732 | ||
6ba495e9 | 3733 | mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); |
9bffad1e | 3734 | trace_ext4_discard_preallocations(inode); |
c9de560d AT |
3735 | |
3736 | INIT_LIST_HEAD(&list); | |
3737 | ||
c83617db | 3738 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
9bffad1e TT |
3739 | if (ac) { |
3740 | ac->ac_sb = sb; | |
3741 | ac->ac_inode = inode; | |
3742 | } | |
c9de560d AT |
3743 | repeat: |
3744 | /* first, collect all pa's in the inode */ | |
3745 | spin_lock(&ei->i_prealloc_lock); | |
3746 | while (!list_empty(&ei->i_prealloc_list)) { | |
3747 | pa = list_entry(ei->i_prealloc_list.next, | |
3748 | struct ext4_prealloc_space, pa_inode_list); | |
3749 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
3750 | spin_lock(&pa->pa_lock); | |
3751 | if (atomic_read(&pa->pa_count)) { | |
3752 | /* this shouldn't happen often - nobody should | |
3753 | * use preallocation while we're discarding it */ | |
3754 | spin_unlock(&pa->pa_lock); | |
3755 | spin_unlock(&ei->i_prealloc_lock); | |
3756 | printk(KERN_ERR "uh-oh! used pa while discarding\n"); | |
3757 | WARN_ON(1); | |
3758 | schedule_timeout_uninterruptible(HZ); | |
3759 | goto repeat; | |
3760 | ||
3761 | } | |
3762 | if (pa->pa_deleted == 0) { | |
3763 | pa->pa_deleted = 1; | |
3764 | spin_unlock(&pa->pa_lock); | |
3765 | list_del_rcu(&pa->pa_inode_list); | |
3766 | list_add(&pa->u.pa_tmp_list, &list); | |
3767 | continue; | |
3768 | } | |
3769 | ||
3770 | /* someone is deleting pa right now */ | |
3771 | spin_unlock(&pa->pa_lock); | |
3772 | spin_unlock(&ei->i_prealloc_lock); | |
3773 | ||
3774 | /* we have to wait here because pa_deleted | |
3775 | * doesn't mean pa is already unlinked from | |
3776 | * the list. as we might be called from | |
3777 | * ->clear_inode() the inode will get freed | |
3778 | * and concurrent thread which is unlinking | |
3779 | * pa from inode's list may access already | |
3780 | * freed memory, bad-bad-bad */ | |
3781 | ||
3782 | /* XXX: if this happens too often, we can | |
3783 | * add a flag to force wait only in case | |
3784 | * of ->clear_inode(), but not in case of | |
3785 | * regular truncate */ | |
3786 | schedule_timeout_uninterruptible(HZ); | |
3787 | goto repeat; | |
3788 | } | |
3789 | spin_unlock(&ei->i_prealloc_lock); | |
3790 | ||
3791 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 3792 | BUG_ON(pa->pa_type != MB_INODE_PA); |
c9de560d AT |
3793 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); |
3794 | ||
3795 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3796 | if (err) { |
12062ddd ES |
3797 | ext4_error(sb, "Error loading buddy information for %u", |
3798 | group); | |
ce89f46c AK |
3799 | continue; |
3800 | } | |
c9de560d | 3801 | |
574ca174 | 3802 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3803 | if (bitmap_bh == NULL) { |
12062ddd ES |
3804 | ext4_error(sb, "Error reading block bitmap for %u", |
3805 | group); | |
c9de560d | 3806 | ext4_mb_release_desc(&e4b); |
ce89f46c | 3807 | continue; |
c9de560d AT |
3808 | } |
3809 | ||
3810 | ext4_lock_group(sb, group); | |
3811 | list_del(&pa->pa_group_list); | |
c83617db | 3812 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); |
c9de560d AT |
3813 | ext4_unlock_group(sb, group); |
3814 | ||
3815 | ext4_mb_release_desc(&e4b); | |
3816 | put_bh(bitmap_bh); | |
3817 | ||
3818 | list_del(&pa->u.pa_tmp_list); | |
3819 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3820 | } | |
c83617db AK |
3821 | if (ac) |
3822 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
3823 | } |
3824 | ||
3825 | /* | |
3826 | * finds all preallocated spaces and return blocks being freed to them | |
3827 | * if preallocated space becomes full (no block is used from the space) | |
3828 | * then the function frees space in buddy | |
3829 | * XXX: at the moment, truncate (which is the only way to free blocks) | |
3830 | * discards all preallocations | |
3831 | */ | |
3832 | static void ext4_mb_return_to_preallocation(struct inode *inode, | |
3833 | struct ext4_buddy *e4b, | |
3834 | sector_t block, int count) | |
3835 | { | |
3836 | BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list)); | |
3837 | } | |
6ba495e9 | 3838 | #ifdef CONFIG_EXT4_DEBUG |
c9de560d AT |
3839 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
3840 | { | |
3841 | struct super_block *sb = ac->ac_sb; | |
8df9675f | 3842 | ext4_group_t ngroups, i; |
c9de560d AT |
3843 | |
3844 | printk(KERN_ERR "EXT4-fs: Can't allocate:" | |
3845 | " Allocation context details:\n"); | |
3846 | printk(KERN_ERR "EXT4-fs: status %d flags %d\n", | |
3847 | ac->ac_status, ac->ac_flags); | |
3848 | printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, " | |
3849 | "best %lu/%lu/%lu@%lu cr %d\n", | |
3850 | (unsigned long)ac->ac_o_ex.fe_group, | |
3851 | (unsigned long)ac->ac_o_ex.fe_start, | |
3852 | (unsigned long)ac->ac_o_ex.fe_len, | |
3853 | (unsigned long)ac->ac_o_ex.fe_logical, | |
3854 | (unsigned long)ac->ac_g_ex.fe_group, | |
3855 | (unsigned long)ac->ac_g_ex.fe_start, | |
3856 | (unsigned long)ac->ac_g_ex.fe_len, | |
3857 | (unsigned long)ac->ac_g_ex.fe_logical, | |
3858 | (unsigned long)ac->ac_b_ex.fe_group, | |
3859 | (unsigned long)ac->ac_b_ex.fe_start, | |
3860 | (unsigned long)ac->ac_b_ex.fe_len, | |
3861 | (unsigned long)ac->ac_b_ex.fe_logical, | |
3862 | (int)ac->ac_criteria); | |
3863 | printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned, | |
3864 | ac->ac_found); | |
3865 | printk(KERN_ERR "EXT4-fs: groups: \n"); | |
8df9675f TT |
3866 | ngroups = ext4_get_groups_count(sb); |
3867 | for (i = 0; i < ngroups; i++) { | |
c9de560d AT |
3868 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
3869 | struct ext4_prealloc_space *pa; | |
3870 | ext4_grpblk_t start; | |
3871 | struct list_head *cur; | |
3872 | ext4_lock_group(sb, i); | |
3873 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3874 | pa = list_entry(cur, struct ext4_prealloc_space, | |
3875 | pa_group_list); | |
3876 | spin_lock(&pa->pa_lock); | |
3877 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3878 | NULL, &start); | |
3879 | spin_unlock(&pa->pa_lock); | |
1c718505 AF |
3880 | printk(KERN_ERR "PA:%u:%d:%u \n", i, |
3881 | start, pa->pa_len); | |
c9de560d | 3882 | } |
60bd63d1 | 3883 | ext4_unlock_group(sb, i); |
c9de560d AT |
3884 | |
3885 | if (grp->bb_free == 0) | |
3886 | continue; | |
1c718505 | 3887 | printk(KERN_ERR "%u: %d/%d \n", |
c9de560d AT |
3888 | i, grp->bb_free, grp->bb_fragments); |
3889 | } | |
3890 | printk(KERN_ERR "\n"); | |
3891 | } | |
3892 | #else | |
3893 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
3894 | { | |
3895 | return; | |
3896 | } | |
3897 | #endif | |
3898 | ||
3899 | /* | |
3900 | * We use locality group preallocation for small size file. The size of the | |
3901 | * file is determined by the current size or the resulting size after | |
3902 | * allocation which ever is larger | |
3903 | * | |
b713a5ec | 3904 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
3905 | */ |
3906 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
3907 | { | |
3908 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3909 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
3910 | loff_t size, isize; | |
3911 | ||
3912 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3913 | return; | |
3914 | ||
4ba74d00 TT |
3915 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
3916 | return; | |
3917 | ||
c9de560d | 3918 | size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; |
50797481 TT |
3919 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
3920 | >> bsbits; | |
c9de560d | 3921 | |
50797481 TT |
3922 | if ((size == isize) && |
3923 | !ext4_fs_is_busy(sbi) && | |
3924 | (atomic_read(&ac->ac_inode->i_writecount) == 0)) { | |
3925 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
3926 | return; | |
3927 | } | |
3928 | ||
c9de560d | 3929 | /* don't use group allocation for large files */ |
71780577 | 3930 | size = max(size, isize); |
cc483f10 | 3931 | if (size > sbi->s_mb_stream_request) { |
4ba74d00 | 3932 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
c9de560d | 3933 | return; |
4ba74d00 | 3934 | } |
c9de560d AT |
3935 | |
3936 | BUG_ON(ac->ac_lg != NULL); | |
3937 | /* | |
3938 | * locality group prealloc space are per cpu. The reason for having | |
3939 | * per cpu locality group is to reduce the contention between block | |
3940 | * request from multiple CPUs. | |
3941 | */ | |
ca0c9584 | 3942 | ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
3943 | |
3944 | /* we're going to use group allocation */ | |
3945 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
3946 | ||
3947 | /* serialize all allocations in the group */ | |
3948 | mutex_lock(&ac->ac_lg->lg_mutex); | |
3949 | } | |
3950 | ||
4ddfef7b ES |
3951 | static noinline_for_stack int |
3952 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
3953 | struct ext4_allocation_request *ar) |
3954 | { | |
3955 | struct super_block *sb = ar->inode->i_sb; | |
3956 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3957 | struct ext4_super_block *es = sbi->s_es; | |
3958 | ext4_group_t group; | |
498e5f24 TT |
3959 | unsigned int len; |
3960 | ext4_fsblk_t goal; | |
c9de560d AT |
3961 | ext4_grpblk_t block; |
3962 | ||
3963 | /* we can't allocate > group size */ | |
3964 | len = ar->len; | |
3965 | ||
3966 | /* just a dirty hack to filter too big requests */ | |
3967 | if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10) | |
3968 | len = EXT4_BLOCKS_PER_GROUP(sb) - 10; | |
3969 | ||
3970 | /* start searching from the goal */ | |
3971 | goal = ar->goal; | |
3972 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
3973 | goal >= ext4_blocks_count(es)) | |
3974 | goal = le32_to_cpu(es->s_first_data_block); | |
3975 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
3976 | ||
3977 | /* set up allocation goals */ | |
833576b3 | 3978 | memset(ac, 0, sizeof(struct ext4_allocation_context)); |
c9de560d | 3979 | ac->ac_b_ex.fe_logical = ar->logical; |
c9de560d | 3980 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
3981 | ac->ac_sb = sb; |
3982 | ac->ac_inode = ar->inode; | |
3983 | ac->ac_o_ex.fe_logical = ar->logical; | |
3984 | ac->ac_o_ex.fe_group = group; | |
3985 | ac->ac_o_ex.fe_start = block; | |
3986 | ac->ac_o_ex.fe_len = len; | |
3987 | ac->ac_g_ex.fe_logical = ar->logical; | |
3988 | ac->ac_g_ex.fe_group = group; | |
3989 | ac->ac_g_ex.fe_start = block; | |
3990 | ac->ac_g_ex.fe_len = len; | |
c9de560d | 3991 | ac->ac_flags = ar->flags; |
c9de560d AT |
3992 | |
3993 | /* we have to define context: we'll we work with a file or | |
3994 | * locality group. this is a policy, actually */ | |
3995 | ext4_mb_group_or_file(ac); | |
3996 | ||
6ba495e9 | 3997 | mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " |
c9de560d AT |
3998 | "left: %u/%u, right %u/%u to %swritable\n", |
3999 | (unsigned) ar->len, (unsigned) ar->logical, | |
4000 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4001 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4002 | (unsigned) ar->lright, (unsigned) ar->pright, | |
4003 | atomic_read(&ar->inode->i_writecount) ? "" : "non-"); | |
4004 | return 0; | |
4005 | ||
4006 | } | |
4007 | ||
6be2ded1 AK |
4008 | static noinline_for_stack void |
4009 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4010 | struct ext4_locality_group *lg, | |
4011 | int order, int total_entries) | |
4012 | { | |
4013 | ext4_group_t group = 0; | |
4014 | struct ext4_buddy e4b; | |
4015 | struct list_head discard_list; | |
4016 | struct ext4_prealloc_space *pa, *tmp; | |
4017 | struct ext4_allocation_context *ac; | |
4018 | ||
6ba495e9 | 4019 | mb_debug(1, "discard locality group preallocation\n"); |
6be2ded1 AK |
4020 | |
4021 | INIT_LIST_HEAD(&discard_list); | |
4022 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); | |
9bffad1e TT |
4023 | if (ac) |
4024 | ac->ac_sb = sb; | |
6be2ded1 AK |
4025 | |
4026 | spin_lock(&lg->lg_prealloc_lock); | |
4027 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
4028 | pa_inode_list) { | |
4029 | spin_lock(&pa->pa_lock); | |
4030 | if (atomic_read(&pa->pa_count)) { | |
4031 | /* | |
4032 | * This is the pa that we just used | |
4033 | * for block allocation. So don't | |
4034 | * free that | |
4035 | */ | |
4036 | spin_unlock(&pa->pa_lock); | |
4037 | continue; | |
4038 | } | |
4039 | if (pa->pa_deleted) { | |
4040 | spin_unlock(&pa->pa_lock); | |
4041 | continue; | |
4042 | } | |
4043 | /* only lg prealloc space */ | |
cc0fb9ad | 4044 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4045 | |
4046 | /* seems this one can be freed ... */ | |
4047 | pa->pa_deleted = 1; | |
4048 | spin_unlock(&pa->pa_lock); | |
4049 | ||
4050 | list_del_rcu(&pa->pa_inode_list); | |
4051 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4052 | ||
4053 | total_entries--; | |
4054 | if (total_entries <= 5) { | |
4055 | /* | |
4056 | * we want to keep only 5 entries | |
4057 | * allowing it to grow to 8. This | |
4058 | * mak sure we don't call discard | |
4059 | * soon for this list. | |
4060 | */ | |
4061 | break; | |
4062 | } | |
4063 | } | |
4064 | spin_unlock(&lg->lg_prealloc_lock); | |
4065 | ||
4066 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
4067 | ||
4068 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); | |
4069 | if (ext4_mb_load_buddy(sb, group, &e4b)) { | |
12062ddd ES |
4070 | ext4_error(sb, "Error loading buddy information for %u", |
4071 | group); | |
6be2ded1 AK |
4072 | continue; |
4073 | } | |
4074 | ext4_lock_group(sb, group); | |
4075 | list_del(&pa->pa_group_list); | |
4076 | ext4_mb_release_group_pa(&e4b, pa, ac); | |
4077 | ext4_unlock_group(sb, group); | |
4078 | ||
4079 | ext4_mb_release_desc(&e4b); | |
4080 | list_del(&pa->u.pa_tmp_list); | |
4081 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4082 | } | |
4083 | if (ac) | |
4084 | kmem_cache_free(ext4_ac_cachep, ac); | |
4085 | } | |
4086 | ||
4087 | /* | |
4088 | * We have incremented pa_count. So it cannot be freed at this | |
4089 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4090 | * possible from this lg. That means pa_free cannot be updated. | |
4091 | * | |
4092 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4093 | * which can cause the lg_prealloc_list to be updated. | |
4094 | */ | |
4095 | ||
4096 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4097 | { | |
4098 | int order, added = 0, lg_prealloc_count = 1; | |
4099 | struct super_block *sb = ac->ac_sb; | |
4100 | struct ext4_locality_group *lg = ac->ac_lg; | |
4101 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4102 | ||
4103 | order = fls(pa->pa_free) - 1; | |
4104 | if (order > PREALLOC_TB_SIZE - 1) | |
4105 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4106 | order = PREALLOC_TB_SIZE - 1; | |
4107 | /* Add the prealloc space to lg */ | |
4108 | rcu_read_lock(); | |
4109 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], | |
4110 | pa_inode_list) { | |
4111 | spin_lock(&tmp_pa->pa_lock); | |
4112 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4113 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4114 | continue; |
4115 | } | |
4116 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4117 | /* Add to the tail of the previous entry */ | |
4118 | list_add_tail_rcu(&pa->pa_inode_list, | |
4119 | &tmp_pa->pa_inode_list); | |
4120 | added = 1; | |
4121 | /* | |
4122 | * we want to count the total | |
4123 | * number of entries in the list | |
4124 | */ | |
4125 | } | |
4126 | spin_unlock(&tmp_pa->pa_lock); | |
4127 | lg_prealloc_count++; | |
4128 | } | |
4129 | if (!added) | |
4130 | list_add_tail_rcu(&pa->pa_inode_list, | |
4131 | &lg->lg_prealloc_list[order]); | |
4132 | rcu_read_unlock(); | |
4133 | ||
4134 | /* Now trim the list to be not more than 8 elements */ | |
4135 | if (lg_prealloc_count > 8) { | |
4136 | ext4_mb_discard_lg_preallocations(sb, lg, | |
4137 | order, lg_prealloc_count); | |
4138 | return; | |
4139 | } | |
4140 | return ; | |
4141 | } | |
4142 | ||
c9de560d AT |
4143 | /* |
4144 | * release all resource we used in allocation | |
4145 | */ | |
4146 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4147 | { | |
6be2ded1 AK |
4148 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4149 | if (pa) { | |
cc0fb9ad | 4150 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4151 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 AK |
4152 | spin_lock(&pa->pa_lock); |
4153 | pa->pa_pstart += ac->ac_b_ex.fe_len; | |
4154 | pa->pa_lstart += ac->ac_b_ex.fe_len; | |
4155 | pa->pa_free -= ac->ac_b_ex.fe_len; | |
4156 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4157 | spin_unlock(&pa->pa_lock); | |
c9de560d | 4158 | } |
c9de560d | 4159 | } |
8556e8f3 AK |
4160 | if (ac->alloc_semp) |
4161 | up_read(ac->alloc_semp); | |
ba443916 AK |
4162 | if (pa) { |
4163 | /* | |
4164 | * We want to add the pa to the right bucket. | |
4165 | * Remove it from the list and while adding | |
4166 | * make sure the list to which we are adding | |
4167 | * doesn't grow big. We need to release | |
4168 | * alloc_semp before calling ext4_mb_add_n_trim() | |
4169 | */ | |
cc0fb9ad | 4170 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { |
ba443916 AK |
4171 | spin_lock(pa->pa_obj_lock); |
4172 | list_del_rcu(&pa->pa_inode_list); | |
4173 | spin_unlock(pa->pa_obj_lock); | |
4174 | ext4_mb_add_n_trim(ac); | |
4175 | } | |
4176 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4177 | } | |
c9de560d AT |
4178 | if (ac->ac_bitmap_page) |
4179 | page_cache_release(ac->ac_bitmap_page); | |
4180 | if (ac->ac_buddy_page) | |
4181 | page_cache_release(ac->ac_buddy_page); | |
4182 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
4183 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4184 | ext4_mb_collect_stats(ac); | |
4185 | return 0; | |
4186 | } | |
4187 | ||
4188 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4189 | { | |
8df9675f | 4190 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4191 | int ret; |
4192 | int freed = 0; | |
4193 | ||
9bffad1e | 4194 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4195 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4196 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4197 | freed += ret; | |
4198 | needed -= ret; | |
4199 | } | |
4200 | ||
4201 | return freed; | |
4202 | } | |
4203 | ||
4204 | /* | |
4205 | * Main entry point into mballoc to allocate blocks | |
4206 | * it tries to use preallocation first, then falls back | |
4207 | * to usual allocation | |
4208 | */ | |
4209 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
4210 | struct ext4_allocation_request *ar, int *errp) | |
4211 | { | |
6bc6e63f | 4212 | int freed; |
256bdb49 | 4213 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4214 | struct ext4_sb_info *sbi; |
4215 | struct super_block *sb; | |
4216 | ext4_fsblk_t block = 0; | |
60e58e0f | 4217 | unsigned int inquota = 0; |
498e5f24 | 4218 | unsigned int reserv_blks = 0; |
c9de560d AT |
4219 | |
4220 | sb = ar->inode->i_sb; | |
4221 | sbi = EXT4_SB(sb); | |
4222 | ||
9bffad1e | 4223 | trace_ext4_request_blocks(ar); |
ba80b101 | 4224 | |
60e58e0f MC |
4225 | /* |
4226 | * For delayed allocation, we could skip the ENOSPC and | |
4227 | * EDQUOT check, as blocks and quotas have been already | |
4228 | * reserved when data being copied into pagecache. | |
4229 | */ | |
4230 | if (EXT4_I(ar->inode)->i_delalloc_reserved_flag) | |
4231 | ar->flags |= EXT4_MB_DELALLOC_RESERVED; | |
4232 | else { | |
4233 | /* Without delayed allocation we need to verify | |
4234 | * there is enough free blocks to do block allocation | |
4235 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4236 | */ |
030ba6bc AK |
4237 | while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) { |
4238 | /* let others to free the space */ | |
4239 | yield(); | |
4240 | ar->len = ar->len >> 1; | |
4241 | } | |
4242 | if (!ar->len) { | |
a30d542a AK |
4243 | *errp = -ENOSPC; |
4244 | return 0; | |
4245 | } | |
6bc6e63f | 4246 | reserv_blks = ar->len; |
a269eb18 | 4247 | while (ar->len && vfs_dq_alloc_block(ar->inode, ar->len)) { |
60e58e0f MC |
4248 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; |
4249 | ar->len--; | |
4250 | } | |
4251 | inquota = ar->len; | |
4252 | if (ar->len == 0) { | |
4253 | *errp = -EDQUOT; | |
4254 | goto out3; | |
4255 | } | |
07031431 | 4256 | } |
d2a17637 | 4257 | |
256bdb49 | 4258 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 4259 | if (!ac) { |
363d4251 | 4260 | ar->len = 0; |
256bdb49 | 4261 | *errp = -ENOMEM; |
363d4251 | 4262 | goto out1; |
256bdb49 ES |
4263 | } |
4264 | ||
256bdb49 | 4265 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4266 | if (*errp) { |
4267 | ar->len = 0; | |
363d4251 | 4268 | goto out2; |
c9de560d AT |
4269 | } |
4270 | ||
256bdb49 ES |
4271 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4272 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4273 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4274 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4275 | repeat: |
4276 | /* allocate space in core */ | |
256bdb49 | 4277 | ext4_mb_regular_allocator(ac); |
c9de560d AT |
4278 | |
4279 | /* as we've just preallocated more space than | |
4280 | * user requested orinally, we store allocated | |
4281 | * space in a special descriptor */ | |
256bdb49 ES |
4282 | if (ac->ac_status == AC_STATUS_FOUND && |
4283 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
4284 | ext4_mb_new_preallocation(ac); | |
c9de560d | 4285 | } |
256bdb49 | 4286 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
6bc6e63f | 4287 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks); |
519deca0 | 4288 | if (*errp == -EAGAIN) { |
8556e8f3 AK |
4289 | /* |
4290 | * drop the reference that we took | |
4291 | * in ext4_mb_use_best_found | |
4292 | */ | |
4293 | ext4_mb_release_context(ac); | |
519deca0 AK |
4294 | ac->ac_b_ex.fe_group = 0; |
4295 | ac->ac_b_ex.fe_start = 0; | |
4296 | ac->ac_b_ex.fe_len = 0; | |
4297 | ac->ac_status = AC_STATUS_CONTINUE; | |
4298 | goto repeat; | |
4299 | } else if (*errp) { | |
b844167e | 4300 | ext4_discard_allocated_blocks(ac); |
519deca0 AK |
4301 | ac->ac_b_ex.fe_len = 0; |
4302 | ar->len = 0; | |
4303 | ext4_mb_show_ac(ac); | |
4304 | } else { | |
4305 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4306 | ar->len = ac->ac_b_ex.fe_len; | |
4307 | } | |
c9de560d | 4308 | } else { |
256bdb49 | 4309 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4310 | if (freed) |
4311 | goto repeat; | |
4312 | *errp = -ENOSPC; | |
256bdb49 | 4313 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4314 | ar->len = 0; |
256bdb49 | 4315 | ext4_mb_show_ac(ac); |
c9de560d AT |
4316 | } |
4317 | ||
256bdb49 | 4318 | ext4_mb_release_context(ac); |
c9de560d | 4319 | |
363d4251 SF |
4320 | out2: |
4321 | kmem_cache_free(ext4_ac_cachep, ac); | |
4322 | out1: | |
60e58e0f | 4323 | if (inquota && ar->len < inquota) |
a269eb18 | 4324 | vfs_dq_free_block(ar->inode, inquota - ar->len); |
0087d9fb AK |
4325 | out3: |
4326 | if (!ar->len) { | |
4327 | if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) | |
4328 | /* release all the reserved blocks if non delalloc */ | |
4329 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, | |
4330 | reserv_blks); | |
4331 | } | |
c9de560d | 4332 | |
9bffad1e | 4333 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4334 | |
c9de560d AT |
4335 | return block; |
4336 | } | |
c9de560d | 4337 | |
c894058d AK |
4338 | /* |
4339 | * We can merge two free data extents only if the physical blocks | |
4340 | * are contiguous, AND the extents were freed by the same transaction, | |
4341 | * AND the blocks are associated with the same group. | |
4342 | */ | |
4343 | static int can_merge(struct ext4_free_data *entry1, | |
4344 | struct ext4_free_data *entry2) | |
4345 | { | |
4346 | if ((entry1->t_tid == entry2->t_tid) && | |
4347 | (entry1->group == entry2->group) && | |
4348 | ((entry1->start_blk + entry1->count) == entry2->start_blk)) | |
4349 | return 1; | |
4350 | return 0; | |
4351 | } | |
4352 | ||
4ddfef7b ES |
4353 | static noinline_for_stack int |
4354 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 4355 | struct ext4_free_data *new_entry) |
c9de560d | 4356 | { |
7a2fcbf7 AK |
4357 | ext4_grpblk_t block; |
4358 | struct ext4_free_data *entry; | |
c9de560d AT |
4359 | struct ext4_group_info *db = e4b->bd_info; |
4360 | struct super_block *sb = e4b->bd_sb; | |
4361 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4362 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
4363 | struct rb_node *parent = NULL, *new_node; | |
4364 | ||
0390131b | 4365 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4366 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4367 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4368 | ||
c894058d | 4369 | new_node = &new_entry->node; |
7a2fcbf7 | 4370 | block = new_entry->start_blk; |
c894058d | 4371 | |
c894058d AK |
4372 | if (!*n) { |
4373 | /* first free block exent. We need to | |
4374 | protect buddy cache from being freed, | |
4375 | * otherwise we'll refresh it from | |
4376 | * on-disk bitmap and lose not-yet-available | |
4377 | * blocks */ | |
4378 | page_cache_get(e4b->bd_buddy_page); | |
4379 | page_cache_get(e4b->bd_bitmap_page); | |
4380 | } | |
4381 | while (*n) { | |
4382 | parent = *n; | |
4383 | entry = rb_entry(parent, struct ext4_free_data, node); | |
4384 | if (block < entry->start_blk) | |
4385 | n = &(*n)->rb_left; | |
4386 | else if (block >= (entry->start_blk + entry->count)) | |
4387 | n = &(*n)->rb_right; | |
4388 | else { | |
5d1b1b3f AK |
4389 | ext4_grp_locked_error(sb, e4b->bd_group, __func__, |
4390 | "Double free of blocks %d (%d %d)", | |
4391 | block, entry->start_blk, entry->count); | |
c894058d | 4392 | return 0; |
c9de560d | 4393 | } |
c894058d | 4394 | } |
c9de560d | 4395 | |
c894058d AK |
4396 | rb_link_node(new_node, parent, n); |
4397 | rb_insert_color(new_node, &db->bb_free_root); | |
4398 | ||
4399 | /* Now try to see the extent can be merged to left and right */ | |
4400 | node = rb_prev(new_node); | |
4401 | if (node) { | |
4402 | entry = rb_entry(node, struct ext4_free_data, node); | |
4403 | if (can_merge(entry, new_entry)) { | |
4404 | new_entry->start_blk = entry->start_blk; | |
4405 | new_entry->count += entry->count; | |
4406 | rb_erase(node, &(db->bb_free_root)); | |
4407 | spin_lock(&sbi->s_md_lock); | |
4408 | list_del(&entry->list); | |
4409 | spin_unlock(&sbi->s_md_lock); | |
4410 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
c9de560d | 4411 | } |
c894058d | 4412 | } |
c9de560d | 4413 | |
c894058d AK |
4414 | node = rb_next(new_node); |
4415 | if (node) { | |
4416 | entry = rb_entry(node, struct ext4_free_data, node); | |
4417 | if (can_merge(new_entry, entry)) { | |
4418 | new_entry->count += entry->count; | |
4419 | rb_erase(node, &(db->bb_free_root)); | |
4420 | spin_lock(&sbi->s_md_lock); | |
4421 | list_del(&entry->list); | |
4422 | spin_unlock(&sbi->s_md_lock); | |
4423 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
c9de560d AT |
4424 | } |
4425 | } | |
3e624fc7 | 4426 | /* Add the extent to transaction's private list */ |
c894058d | 4427 | spin_lock(&sbi->s_md_lock); |
3e624fc7 | 4428 | list_add(&new_entry->list, &handle->h_transaction->t_private_list); |
c894058d | 4429 | spin_unlock(&sbi->s_md_lock); |
c9de560d AT |
4430 | return 0; |
4431 | } | |
4432 | ||
44338711 TT |
4433 | /** |
4434 | * ext4_free_blocks() -- Free given blocks and update quota | |
4435 | * @handle: handle for this transaction | |
4436 | * @inode: inode | |
4437 | * @block: start physical block to free | |
4438 | * @count: number of blocks to count | |
4439 | * @metadata: Are these metadata blocks | |
c9de560d | 4440 | */ |
44338711 | 4441 | void ext4_free_blocks(handle_t *handle, struct inode *inode, |
e6362609 TT |
4442 | struct buffer_head *bh, ext4_fsblk_t block, |
4443 | unsigned long count, int flags) | |
c9de560d | 4444 | { |
26346ff6 | 4445 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4446 | struct super_block *sb = inode->i_sb; |
256bdb49 | 4447 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4448 | struct ext4_group_desc *gdp; |
4449 | struct ext4_super_block *es; | |
44338711 | 4450 | unsigned long freed = 0; |
498e5f24 | 4451 | unsigned int overflow; |
c9de560d AT |
4452 | ext4_grpblk_t bit; |
4453 | struct buffer_head *gd_bh; | |
4454 | ext4_group_t block_group; | |
4455 | struct ext4_sb_info *sbi; | |
4456 | struct ext4_buddy e4b; | |
4457 | int err = 0; | |
4458 | int ret; | |
4459 | ||
e6362609 TT |
4460 | if (bh) { |
4461 | if (block) | |
4462 | BUG_ON(block != bh->b_blocknr); | |
4463 | else | |
4464 | block = bh->b_blocknr; | |
4465 | } | |
c9de560d | 4466 | |
c9de560d AT |
4467 | sbi = EXT4_SB(sb); |
4468 | es = EXT4_SB(sb)->s_es; | |
1f2acb60 TT |
4469 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
4470 | !ext4_data_block_valid(sbi, block, count)) { | |
12062ddd | 4471 | ext4_error(sb, "Freeing blocks not in datazone - " |
1f2acb60 | 4472 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
4473 | goto error_return; |
4474 | } | |
4475 | ||
0610b6e9 | 4476 | ext4_debug("freeing block %llu\n", block); |
e6362609 TT |
4477 | trace_ext4_free_blocks(inode, block, count, flags); |
4478 | ||
4479 | if (flags & EXT4_FREE_BLOCKS_FORGET) { | |
4480 | struct buffer_head *tbh = bh; | |
4481 | int i; | |
4482 | ||
4483 | BUG_ON(bh && (count > 1)); | |
4484 | ||
4485 | for (i = 0; i < count; i++) { | |
4486 | if (!bh) | |
4487 | tbh = sb_find_get_block(inode->i_sb, | |
4488 | block + i); | |
4489 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, | |
4490 | inode, tbh, block + i); | |
4491 | } | |
4492 | } | |
4493 | ||
4494 | /* | |
4495 | * We need to make sure we don't reuse the freed block until | |
4496 | * after the transaction is committed, which we can do by | |
4497 | * treating the block as metadata, below. We make an | |
4498 | * exception if the inode is to be written in writeback mode | |
4499 | * since writeback mode has weak data consistency guarantees. | |
4500 | */ | |
4501 | if (!ext4_should_writeback_data(inode)) | |
4502 | flags |= EXT4_FREE_BLOCKS_METADATA; | |
c9de560d | 4503 | |
256bdb49 ES |
4504 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
4505 | if (ac) { | |
256bdb49 ES |
4506 | ac->ac_inode = inode; |
4507 | ac->ac_sb = sb; | |
4508 | } | |
c9de560d AT |
4509 | |
4510 | do_more: | |
4511 | overflow = 0; | |
4512 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4513 | ||
4514 | /* | |
4515 | * Check to see if we are freeing blocks across a group | |
4516 | * boundary. | |
4517 | */ | |
4518 | if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { | |
4519 | overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb); | |
4520 | count -= overflow; | |
4521 | } | |
574ca174 | 4522 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
ce89f46c AK |
4523 | if (!bitmap_bh) { |
4524 | err = -EIO; | |
c9de560d | 4525 | goto error_return; |
ce89f46c | 4526 | } |
c9de560d | 4527 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4528 | if (!gdp) { |
4529 | err = -EIO; | |
c9de560d | 4530 | goto error_return; |
ce89f46c | 4531 | } |
c9de560d AT |
4532 | |
4533 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4534 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4535 | in_range(block, ext4_inode_table(sb, gdp), | |
4536 | EXT4_SB(sb)->s_itb_per_group) || | |
4537 | in_range(block + count - 1, ext4_inode_table(sb, gdp), | |
4538 | EXT4_SB(sb)->s_itb_per_group)) { | |
4539 | ||
12062ddd | 4540 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 4541 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
4542 | /* err = 0. ext4_std_error should be a no op */ |
4543 | goto error_return; | |
c9de560d AT |
4544 | } |
4545 | ||
4546 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4547 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4548 | if (err) | |
4549 | goto error_return; | |
4550 | ||
4551 | /* | |
4552 | * We are about to modify some metadata. Call the journal APIs | |
4553 | * to unshare ->b_data if a currently-committing transaction is | |
4554 | * using it | |
4555 | */ | |
4556 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4557 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4558 | if (err) | |
4559 | goto error_return; | |
c9de560d AT |
4560 | #ifdef AGGRESSIVE_CHECK |
4561 | { | |
4562 | int i; | |
4563 | for (i = 0; i < count; i++) | |
4564 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); | |
4565 | } | |
4566 | #endif | |
256bdb49 ES |
4567 | if (ac) { |
4568 | ac->ac_b_ex.fe_group = block_group; | |
4569 | ac->ac_b_ex.fe_start = bit; | |
4570 | ac->ac_b_ex.fe_len = count; | |
296c355c | 4571 | trace_ext4_mballoc_free(ac); |
256bdb49 | 4572 | } |
c9de560d | 4573 | |
920313a7 AK |
4574 | err = ext4_mb_load_buddy(sb, block_group, &e4b); |
4575 | if (err) | |
4576 | goto error_return; | |
e6362609 TT |
4577 | |
4578 | if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) { | |
7a2fcbf7 AK |
4579 | struct ext4_free_data *new_entry; |
4580 | /* | |
4581 | * blocks being freed are metadata. these blocks shouldn't | |
4582 | * be used until this transaction is committed | |
4583 | */ | |
4584 | new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS); | |
4585 | new_entry->start_blk = bit; | |
4586 | new_entry->group = block_group; | |
4587 | new_entry->count = count; | |
4588 | new_entry->t_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 4589 | |
7a2fcbf7 | 4590 | ext4_lock_group(sb, block_group); |
955ce5f5 | 4591 | mb_clear_bits(bitmap_bh->b_data, bit, count); |
7a2fcbf7 | 4592 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 4593 | } else { |
7a2fcbf7 AK |
4594 | /* need to update group_info->bb_free and bitmap |
4595 | * with group lock held. generate_buddy look at | |
4596 | * them with group lock_held | |
4597 | */ | |
955ce5f5 AK |
4598 | ext4_lock_group(sb, block_group); |
4599 | mb_clear_bits(bitmap_bh->b_data, bit, count); | |
7e5a8cdd | 4600 | mb_free_blocks(inode, &e4b, bit, count); |
c9de560d | 4601 | ext4_mb_return_to_preallocation(inode, &e4b, block, count); |
c9de560d AT |
4602 | } |
4603 | ||
560671a0 AK |
4604 | ret = ext4_free_blks_count(sb, gdp) + count; |
4605 | ext4_free_blks_set(sb, gdp, ret); | |
c9de560d | 4606 | gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); |
955ce5f5 | 4607 | ext4_unlock_group(sb, block_group); |
c9de560d AT |
4608 | percpu_counter_add(&sbi->s_freeblocks_counter, count); |
4609 | ||
772cb7c8 JS |
4610 | if (sbi->s_log_groups_per_flex) { |
4611 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
9f24e420 | 4612 | atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks); |
772cb7c8 JS |
4613 | } |
4614 | ||
c9de560d AT |
4615 | ext4_mb_release_desc(&e4b); |
4616 | ||
44338711 | 4617 | freed += count; |
c9de560d | 4618 | |
7a2fcbf7 AK |
4619 | /* We dirtied the bitmap block */ |
4620 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4621 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4622 | ||
c9de560d AT |
4623 | /* And the group descriptor block */ |
4624 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 4625 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
4626 | if (!err) |
4627 | err = ret; | |
4628 | ||
4629 | if (overflow && !err) { | |
4630 | block += count; | |
4631 | count = overflow; | |
4632 | put_bh(bitmap_bh); | |
4633 | goto do_more; | |
4634 | } | |
4635 | sb->s_dirt = 1; | |
4636 | error_return: | |
44338711 TT |
4637 | if (freed) |
4638 | vfs_dq_free_block(inode, freed); | |
c9de560d AT |
4639 | brelse(bitmap_bh); |
4640 | ext4_std_error(sb, err); | |
256bdb49 ES |
4641 | if (ac) |
4642 | kmem_cache_free(ext4_ac_cachep, ac); | |
c9de560d AT |
4643 | return; |
4644 | } |