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> |
5a0e3ad6 | 26 | #include <linux/slab.h> |
9bffad1e TT |
27 | #include <trace/events/ext4.h> |
28 | ||
c9de560d AT |
29 | /* |
30 | * MUSTDO: | |
31 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
32 | * - search for metadata in few groups | |
33 | * | |
34 | * TODO v4: | |
35 | * - normalization should take into account whether file is still open | |
36 | * - discard preallocations if no free space left (policy?) | |
37 | * - don't normalize tails | |
38 | * - quota | |
39 | * - reservation for superuser | |
40 | * | |
41 | * TODO v3: | |
42 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
43 | * - track min/max extents in each group for better group selection | |
44 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
45 | * - tree of groups sorted by number of free blocks | |
46 | * - error handling | |
47 | */ | |
48 | ||
49 | /* | |
50 | * The allocation request involve request for multiple number of blocks | |
51 | * near to the goal(block) value specified. | |
52 | * | |
b713a5ec TT |
53 | * During initialization phase of the allocator we decide to use the |
54 | * group preallocation or inode preallocation depending on the size of | |
55 | * the file. The size of the file could be the resulting file size we | |
56 | * would have after allocation, or the current file size, which ever | |
57 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
58 | * select to use the group preallocation. The default value of | |
59 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
60 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
61 | * terms of number of blocks. | |
c9de560d AT |
62 | * |
63 | * The main motivation for having small file use group preallocation is to | |
b713a5ec | 64 | * ensure that we have small files closer together on the disk. |
c9de560d | 65 | * |
b713a5ec TT |
66 | * First stage the allocator looks at the inode prealloc list, |
67 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
68 | * spaces for this particular inode. The inode prealloc space is | |
69 | * represented as: | |
c9de560d AT |
70 | * |
71 | * pa_lstart -> the logical start block for this prealloc space | |
72 | * pa_pstart -> the physical start block for this prealloc space | |
53accfa9 TT |
73 | * pa_len -> length for this prealloc space (in clusters) |
74 | * pa_free -> free space available in this prealloc space (in clusters) | |
c9de560d AT |
75 | * |
76 | * The inode preallocation space is used looking at the _logical_ start | |
77 | * block. If only the logical file block falls within the range of prealloc | |
caaf7a29 TM |
78 | * space we will consume the particular prealloc space. This makes sure that |
79 | * we have contiguous physical blocks representing the file blocks | |
c9de560d AT |
80 | * |
81 | * The important thing to be noted in case of inode prealloc space is that | |
82 | * we don't modify the values associated to inode prealloc space except | |
83 | * pa_free. | |
84 | * | |
85 | * If we are not able to find blocks in the inode prealloc space and if we | |
86 | * have the group allocation flag set then we look at the locality group | |
caaf7a29 | 87 | * prealloc space. These are per CPU prealloc list represented as |
c9de560d AT |
88 | * |
89 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
90 | * | |
91 | * The reason for having a per cpu locality group is to reduce the contention | |
92 | * between CPUs. It is possible to get scheduled at this point. | |
93 | * | |
94 | * The locality group prealloc space is used looking at whether we have | |
25985edc | 95 | * enough free space (pa_free) within the prealloc space. |
c9de560d AT |
96 | * |
97 | * If we can't allocate blocks via inode prealloc or/and locality group | |
98 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
99 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
100 | * mapped to the buddy and bitmap information regarding different | |
101 | * groups. The buddy information is attached to buddy cache inode so that | |
102 | * we can access them through the page cache. The information regarding | |
103 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
104 | * block bitmap and buddy information. The information are stored in the | |
105 | * inode as: | |
106 | * | |
107 | * { page } | |
c3a326a6 | 108 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
109 | * |
110 | * | |
111 | * one block each for bitmap and buddy information. So for each group we | |
112 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / | |
113 | * blocksize) blocks. So it can have information regarding groups_per_page | |
114 | * which is blocks_per_page/2 | |
115 | * | |
116 | * The buddy cache inode is not stored on disk. The inode is thrown | |
117 | * away when the filesystem is unmounted. | |
118 | * | |
119 | * We look for count number of blocks in the buddy cache. If we were able | |
120 | * to locate that many free blocks we return with additional information | |
121 | * regarding rest of the contiguous physical block available | |
122 | * | |
123 | * Before allocating blocks via buddy cache we normalize the request | |
124 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
125 | * blocks that we get after allocation is added to the respective prealloc | |
126 | * list. In case of inode preallocation we follow a list of heuristics | |
127 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
128 | * we are doing a group prealloc we try to normalize the request to | |
b713a5ec | 129 | * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is |
c9de560d | 130 | * 512 blocks. This can be tuned via |
d7a1fee1 | 131 | * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in |
c9de560d AT |
132 | * terms of number of blocks. If we have mounted the file system with -O |
133 | * stripe=<value> option the group prealloc request is normalized to the | |
d7a1fee1 DE |
134 | * the smallest multiple of the stripe value (sbi->s_stripe) which is |
135 | * greater than the default mb_group_prealloc. | |
c9de560d | 136 | * |
d7a1fee1 | 137 | * The regular allocator (using the buddy cache) supports a few tunables. |
c9de560d | 138 | * |
b713a5ec TT |
139 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
140 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
141 | * /sys/fs/ext4/<partition>/mb_order2_req | |
c9de560d | 142 | * |
b713a5ec | 143 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
144 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
145 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 146 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
af901ca1 | 147 | * stripe size (sbi->s_stripe), we try to search for contiguous block in |
b713a5ec TT |
148 | * stripe size. This should result in better allocation on RAID setups. If |
149 | * not, we search in the specific group using bitmap for best extents. The | |
150 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 151 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 152 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
153 | * best extent in the found extents. Searching for the blocks starts with |
154 | * the group specified as the goal value in allocation context via | |
155 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
caaf7a29 | 156 | * can be used for allocation. ext4_mb_good_group explains how the groups are |
c9de560d AT |
157 | * checked. |
158 | * | |
159 | * Both the prealloc space are getting populated as above. So for the first | |
160 | * request we will hit the buddy cache which will result in this prealloc | |
161 | * space getting filled. The prealloc space is then later used for the | |
162 | * subsequent request. | |
163 | */ | |
164 | ||
165 | /* | |
166 | * mballoc operates on the following data: | |
167 | * - on-disk bitmap | |
168 | * - in-core buddy (actually includes buddy and bitmap) | |
169 | * - preallocation descriptors (PAs) | |
170 | * | |
171 | * there are two types of preallocations: | |
172 | * - inode | |
173 | * assiged to specific inode and can be used for this inode only. | |
174 | * it describes part of inode's space preallocated to specific | |
175 | * physical blocks. any block from that preallocated can be used | |
176 | * independent. the descriptor just tracks number of blocks left | |
177 | * unused. so, before taking some block from descriptor, one must | |
178 | * make sure corresponded logical block isn't allocated yet. this | |
179 | * also means that freeing any block within descriptor's range | |
180 | * must discard all preallocated blocks. | |
181 | * - locality group | |
182 | * assigned to specific locality group which does not translate to | |
183 | * permanent set of inodes: inode can join and leave group. space | |
184 | * from this type of preallocation can be used for any inode. thus | |
185 | * it's consumed from the beginning to the end. | |
186 | * | |
187 | * relation between them can be expressed as: | |
188 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
189 | * | |
190 | * this mean blocks mballoc considers used are: | |
191 | * - allocated blocks (persistent) | |
192 | * - preallocated blocks (non-persistent) | |
193 | * | |
194 | * consistency in mballoc world means that at any time a block is either | |
195 | * free or used in ALL structures. notice: "any time" should not be read | |
196 | * literally -- time is discrete and delimited by locks. | |
197 | * | |
198 | * to keep it simple, we don't use block numbers, instead we count number of | |
199 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
200 | * | |
201 | * all operations can be expressed as: | |
202 | * - init buddy: buddy = on-disk + PAs | |
203 | * - new PA: buddy += N; PA = N | |
204 | * - use inode PA: on-disk += N; PA -= N | |
205 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
206 | * - use locality group PA on-disk += N; PA -= N | |
207 | * - discard locality group PA buddy -= PA; PA = 0 | |
208 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
209 | * is used in real operation because we can't know actual used | |
210 | * bits from PA, only from on-disk bitmap | |
211 | * | |
212 | * if we follow this strict logic, then all operations above should be atomic. | |
213 | * given some of them can block, we'd have to use something like semaphores | |
214 | * killing performance on high-end SMP hardware. let's try to relax it using | |
215 | * the following knowledge: | |
216 | * 1) if buddy is referenced, it's already initialized | |
217 | * 2) while block is used in buddy and the buddy is referenced, | |
218 | * nobody can re-allocate that block | |
219 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
220 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
221 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
222 | * block | |
223 | * | |
224 | * so, now we're building a concurrency table: | |
225 | * - init buddy vs. | |
226 | * - new PA | |
227 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
228 | * until PA is linked to allocation group to avoid concurrent buddy init | |
229 | * - use inode PA | |
230 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
231 | * given (3) we care that PA-=N operation doesn't interfere with init | |
232 | * - discard inode PA | |
233 | * the simplest way would be to have buddy initialized by the discard | |
234 | * - use locality group PA | |
235 | * again PA-=N must be serialized with init | |
236 | * - discard locality group PA | |
237 | * the simplest way would be to have buddy initialized by the discard | |
238 | * - new PA vs. | |
239 | * - use inode PA | |
240 | * i_data_sem serializes them | |
241 | * - discard inode PA | |
242 | * discard process must wait until PA isn't used by another process | |
243 | * - use locality group PA | |
244 | * some mutex should serialize them | |
245 | * - discard locality group PA | |
246 | * discard process must wait until PA isn't used by another process | |
247 | * - use inode PA | |
248 | * - use inode PA | |
249 | * i_data_sem or another mutex should serializes them | |
250 | * - discard inode PA | |
251 | * discard process must wait until PA isn't used by another process | |
252 | * - use locality group PA | |
253 | * nothing wrong here -- they're different PAs covering different blocks | |
254 | * - discard locality group PA | |
255 | * discard process must wait until PA isn't used by another process | |
256 | * | |
257 | * now we're ready to make few consequences: | |
258 | * - PA is referenced and while it is no discard is possible | |
259 | * - PA is referenced until block isn't marked in on-disk bitmap | |
260 | * - PA changes only after on-disk bitmap | |
261 | * - discard must not compete with init. either init is done before | |
262 | * any discard or they're serialized somehow | |
263 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
264 | * | |
265 | * a special case when we've used PA to emptiness. no need to modify buddy | |
266 | * in this case, but we should care about concurrent init | |
267 | * | |
268 | */ | |
269 | ||
270 | /* | |
271 | * Logic in few words: | |
272 | * | |
273 | * - allocation: | |
274 | * load group | |
275 | * find blocks | |
276 | * mark bits in on-disk bitmap | |
277 | * release group | |
278 | * | |
279 | * - use preallocation: | |
280 | * find proper PA (per-inode or group) | |
281 | * load group | |
282 | * mark bits in on-disk bitmap | |
283 | * release group | |
284 | * release PA | |
285 | * | |
286 | * - free: | |
287 | * load group | |
288 | * mark bits in on-disk bitmap | |
289 | * release group | |
290 | * | |
291 | * - discard preallocations in group: | |
292 | * mark PAs deleted | |
293 | * move them onto local list | |
294 | * load on-disk bitmap | |
295 | * load group | |
296 | * remove PA from object (inode or locality group) | |
297 | * mark free blocks in-core | |
298 | * | |
299 | * - discard inode's preallocations: | |
300 | */ | |
301 | ||
302 | /* | |
303 | * Locking rules | |
304 | * | |
305 | * Locks: | |
306 | * - bitlock on a group (group) | |
307 | * - object (inode/locality) (object) | |
308 | * - per-pa lock (pa) | |
309 | * | |
310 | * Paths: | |
311 | * - new pa | |
312 | * object | |
313 | * group | |
314 | * | |
315 | * - find and use pa: | |
316 | * pa | |
317 | * | |
318 | * - release consumed pa: | |
319 | * pa | |
320 | * group | |
321 | * object | |
322 | * | |
323 | * - generate in-core bitmap: | |
324 | * group | |
325 | * pa | |
326 | * | |
327 | * - discard all for given object (inode, locality group): | |
328 | * object | |
329 | * pa | |
330 | * group | |
331 | * | |
332 | * - discard all for given group: | |
333 | * group | |
334 | * pa | |
335 | * group | |
336 | * object | |
337 | * | |
338 | */ | |
c3a326a6 AK |
339 | static struct kmem_cache *ext4_pspace_cachep; |
340 | static struct kmem_cache *ext4_ac_cachep; | |
341 | static struct kmem_cache *ext4_free_ext_cachep; | |
fb1813f4 CW |
342 | |
343 | /* We create slab caches for groupinfo data structures based on the | |
344 | * superblock block size. There will be one per mounted filesystem for | |
345 | * each unique s_blocksize_bits */ | |
2892c15d | 346 | #define NR_GRPINFO_CACHES 8 |
fb1813f4 CW |
347 | static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; |
348 | ||
2892c15d ES |
349 | static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { |
350 | "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", | |
351 | "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", | |
352 | "ext4_groupinfo_64k", "ext4_groupinfo_128k" | |
353 | }; | |
354 | ||
c3a326a6 AK |
355 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
356 | ext4_group_t group); | |
7a2fcbf7 AK |
357 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
358 | ext4_group_t group); | |
c3a326a6 AK |
359 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn); |
360 | ||
ffad0a44 AK |
361 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
362 | { | |
c9de560d | 363 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
364 | *bit += ((unsigned long) addr & 7UL) << 3; |
365 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 366 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
367 | *bit += ((unsigned long) addr & 3UL) << 3; |
368 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
369 | #else |
370 | #error "how many bits you are?!" | |
371 | #endif | |
ffad0a44 AK |
372 | return addr; |
373 | } | |
c9de560d AT |
374 | |
375 | static inline int mb_test_bit(int bit, void *addr) | |
376 | { | |
377 | /* | |
378 | * ext4_test_bit on architecture like powerpc | |
379 | * needs unsigned long aligned address | |
380 | */ | |
ffad0a44 | 381 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
382 | return ext4_test_bit(bit, addr); |
383 | } | |
384 | ||
385 | static inline void mb_set_bit(int bit, void *addr) | |
386 | { | |
ffad0a44 | 387 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
388 | ext4_set_bit(bit, addr); |
389 | } | |
390 | ||
c9de560d AT |
391 | static inline void mb_clear_bit(int bit, void *addr) |
392 | { | |
ffad0a44 | 393 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
394 | ext4_clear_bit(bit, addr); |
395 | } | |
396 | ||
ffad0a44 AK |
397 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
398 | { | |
e7dfb246 | 399 | int fix = 0, ret, tmpmax; |
ffad0a44 | 400 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 401 | tmpmax = max + fix; |
ffad0a44 AK |
402 | start += fix; |
403 | ||
e7dfb246 AK |
404 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
405 | if (ret > max) | |
406 | return max; | |
407 | return ret; | |
ffad0a44 AK |
408 | } |
409 | ||
410 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
411 | { | |
e7dfb246 | 412 | int fix = 0, ret, tmpmax; |
ffad0a44 | 413 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 414 | tmpmax = max + fix; |
ffad0a44 AK |
415 | start += fix; |
416 | ||
e7dfb246 AK |
417 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
418 | if (ret > max) | |
419 | return max; | |
420 | return ret; | |
ffad0a44 AK |
421 | } |
422 | ||
c9de560d AT |
423 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
424 | { | |
425 | char *bb; | |
426 | ||
c9de560d AT |
427 | BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); |
428 | BUG_ON(max == NULL); | |
429 | ||
430 | if (order > e4b->bd_blkbits + 1) { | |
431 | *max = 0; | |
432 | return NULL; | |
433 | } | |
434 | ||
435 | /* at order 0 we see each particular block */ | |
84b775a3 CL |
436 | if (order == 0) { |
437 | *max = 1 << (e4b->bd_blkbits + 3); | |
c9de560d | 438 | return EXT4_MB_BITMAP(e4b); |
84b775a3 | 439 | } |
c9de560d AT |
440 | |
441 | bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; | |
442 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; | |
443 | ||
444 | return bb; | |
445 | } | |
446 | ||
447 | #ifdef DOUBLE_CHECK | |
448 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
449 | int first, int count) | |
450 | { | |
451 | int i; | |
452 | struct super_block *sb = e4b->bd_sb; | |
453 | ||
454 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
455 | return; | |
bc8e6740 | 456 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
457 | for (i = 0; i < count; i++) { |
458 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
459 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
460 | |
461 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 462 | blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
5d1b1b3f | 463 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
464 | inode ? inode->i_ino : 0, |
465 | blocknr, | |
466 | "freeing block already freed " | |
467 | "(bit %u)", | |
468 | first + i); | |
c9de560d AT |
469 | } |
470 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
471 | } | |
472 | } | |
473 | ||
474 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
475 | { | |
476 | int i; | |
477 | ||
478 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
479 | return; | |
bc8e6740 | 480 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
481 | for (i = 0; i < count; i++) { |
482 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
483 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
484 | } | |
485 | } | |
486 | ||
487 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
488 | { | |
489 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { | |
490 | unsigned char *b1, *b2; | |
491 | int i; | |
492 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
493 | b2 = (unsigned char *) bitmap; | |
494 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
495 | if (b1[i] != b2[i]) { | |
9d8b9ec4 TT |
496 | ext4_msg(e4b->bd_sb, KERN_ERR, |
497 | "corruption in group %u " | |
498 | "at byte %u(%u): %x in copy != %x " | |
499 | "on disk/prealloc", | |
500 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
501 | BUG(); |
502 | } | |
503 | } | |
504 | } | |
505 | } | |
506 | ||
507 | #else | |
508 | static inline void mb_free_blocks_double(struct inode *inode, | |
509 | struct ext4_buddy *e4b, int first, int count) | |
510 | { | |
511 | return; | |
512 | } | |
513 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
514 | int first, int count) | |
515 | { | |
516 | return; | |
517 | } | |
518 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
519 | { | |
520 | return; | |
521 | } | |
522 | #endif | |
523 | ||
524 | #ifdef AGGRESSIVE_CHECK | |
525 | ||
526 | #define MB_CHECK_ASSERT(assert) \ | |
527 | do { \ | |
528 | if (!(assert)) { \ | |
529 | printk(KERN_EMERG \ | |
530 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
531 | function, file, line, # assert); \ | |
532 | BUG(); \ | |
533 | } \ | |
534 | } while (0) | |
535 | ||
536 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
537 | const char *function, int line) | |
538 | { | |
539 | struct super_block *sb = e4b->bd_sb; | |
540 | int order = e4b->bd_blkbits + 1; | |
541 | int max; | |
542 | int max2; | |
543 | int i; | |
544 | int j; | |
545 | int k; | |
546 | int count; | |
547 | struct ext4_group_info *grp; | |
548 | int fragments = 0; | |
549 | int fstart; | |
550 | struct list_head *cur; | |
551 | void *buddy; | |
552 | void *buddy2; | |
553 | ||
c9de560d AT |
554 | { |
555 | static int mb_check_counter; | |
556 | if (mb_check_counter++ % 100 != 0) | |
557 | return 0; | |
558 | } | |
559 | ||
560 | while (order > 1) { | |
561 | buddy = mb_find_buddy(e4b, order, &max); | |
562 | MB_CHECK_ASSERT(buddy); | |
563 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
564 | MB_CHECK_ASSERT(buddy2); | |
565 | MB_CHECK_ASSERT(buddy != buddy2); | |
566 | MB_CHECK_ASSERT(max * 2 == max2); | |
567 | ||
568 | count = 0; | |
569 | for (i = 0; i < max; i++) { | |
570 | ||
571 | if (mb_test_bit(i, buddy)) { | |
572 | /* only single bit in buddy2 may be 1 */ | |
573 | if (!mb_test_bit(i << 1, buddy2)) { | |
574 | MB_CHECK_ASSERT( | |
575 | mb_test_bit((i<<1)+1, buddy2)); | |
576 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
577 | MB_CHECK_ASSERT( | |
578 | mb_test_bit(i << 1, buddy2)); | |
579 | } | |
580 | continue; | |
581 | } | |
582 | ||
583 | /* both bits in buddy2 must be 0 */ | |
584 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); | |
585 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
586 | ||
587 | for (j = 0; j < (1 << order); j++) { | |
588 | k = (i * (1 << order)) + j; | |
589 | MB_CHECK_ASSERT( | |
590 | !mb_test_bit(k, EXT4_MB_BITMAP(e4b))); | |
591 | } | |
592 | count++; | |
593 | } | |
594 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
595 | order--; | |
596 | } | |
597 | ||
598 | fstart = -1; | |
599 | buddy = mb_find_buddy(e4b, 0, &max); | |
600 | for (i = 0; i < max; i++) { | |
601 | if (!mb_test_bit(i, buddy)) { | |
602 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
603 | if (fstart == -1) { | |
604 | fragments++; | |
605 | fstart = i; | |
606 | } | |
607 | continue; | |
608 | } | |
609 | fstart = -1; | |
610 | /* check used bits only */ | |
611 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
612 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
613 | k = i >> j; | |
614 | MB_CHECK_ASSERT(k < max2); | |
615 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
616 | } | |
617 | } | |
618 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
619 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
620 | ||
621 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
622 | list_for_each(cur, &grp->bb_prealloc_list) { |
623 | ext4_group_t groupnr; | |
624 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
625 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
626 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 627 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 628 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
629 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
630 | } | |
631 | return 0; | |
632 | } | |
633 | #undef MB_CHECK_ASSERT | |
634 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 635 | __FILE__, __func__, __LINE__) |
c9de560d AT |
636 | #else |
637 | #define mb_check_buddy(e4b) | |
638 | #endif | |
639 | ||
7c786059 CL |
640 | /* |
641 | * Divide blocks started from @first with length @len into | |
642 | * smaller chunks with power of 2 blocks. | |
643 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
644 | * then increase bb_counters[] for corresponded chunk size. | |
645 | */ | |
c9de560d | 646 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 647 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
648 | struct ext4_group_info *grp) |
649 | { | |
650 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
651 | ext4_grpblk_t min; |
652 | ext4_grpblk_t max; | |
653 | ext4_grpblk_t chunk; | |
c9de560d AT |
654 | unsigned short border; |
655 | ||
7137d7a4 | 656 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
657 | |
658 | border = 2 << sb->s_blocksize_bits; | |
659 | ||
660 | while (len > 0) { | |
661 | /* find how many blocks can be covered since this position */ | |
662 | max = ffs(first | border) - 1; | |
663 | ||
664 | /* find how many blocks of power 2 we need to mark */ | |
665 | min = fls(len) - 1; | |
666 | ||
667 | if (max < min) | |
668 | min = max; | |
669 | chunk = 1 << min; | |
670 | ||
671 | /* mark multiblock chunks only */ | |
672 | grp->bb_counters[min]++; | |
673 | if (min > 0) | |
674 | mb_clear_bit(first >> min, | |
675 | buddy + sbi->s_mb_offsets[min]); | |
676 | ||
677 | len -= chunk; | |
678 | first += chunk; | |
679 | } | |
680 | } | |
681 | ||
8a57d9d6 CW |
682 | /* |
683 | * Cache the order of the largest free extent we have available in this block | |
684 | * group. | |
685 | */ | |
686 | static void | |
687 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
688 | { | |
689 | int i; | |
690 | int bits; | |
691 | ||
692 | grp->bb_largest_free_order = -1; /* uninit */ | |
693 | ||
694 | bits = sb->s_blocksize_bits + 1; | |
695 | for (i = bits; i >= 0; i--) { | |
696 | if (grp->bb_counters[i] > 0) { | |
697 | grp->bb_largest_free_order = i; | |
698 | break; | |
699 | } | |
700 | } | |
701 | } | |
702 | ||
089ceecc ES |
703 | static noinline_for_stack |
704 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
705 | void *buddy, void *bitmap, ext4_group_t group) |
706 | { | |
707 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
7137d7a4 | 708 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
709 | ext4_grpblk_t i = 0; |
710 | ext4_grpblk_t first; | |
711 | ext4_grpblk_t len; | |
c9de560d AT |
712 | unsigned free = 0; |
713 | unsigned fragments = 0; | |
714 | unsigned long long period = get_cycles(); | |
715 | ||
716 | /* initialize buddy from bitmap which is aggregation | |
717 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 718 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
719 | grp->bb_first_free = i; |
720 | while (i < max) { | |
721 | fragments++; | |
722 | first = i; | |
ffad0a44 | 723 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
724 | len = i - first; |
725 | free += len; | |
726 | if (len > 1) | |
727 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
728 | else | |
729 | grp->bb_counters[0]++; | |
730 | if (i < max) | |
ffad0a44 | 731 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
732 | } |
733 | grp->bb_fragments = fragments; | |
734 | ||
735 | if (free != grp->bb_free) { | |
e29136f8 | 736 | ext4_grp_locked_error(sb, group, 0, 0, |
53accfa9 | 737 | "%u clusters in bitmap, %u in gd", |
e29136f8 | 738 | free, grp->bb_free); |
e56eb659 AK |
739 | /* |
740 | * If we intent to continue, we consider group descritor | |
741 | * corrupt and update bb_free using bitmap value | |
742 | */ | |
c9de560d AT |
743 | grp->bb_free = free; |
744 | } | |
8a57d9d6 | 745 | mb_set_largest_free_order(sb, grp); |
c9de560d AT |
746 | |
747 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
748 | ||
749 | period = get_cycles() - period; | |
750 | spin_lock(&EXT4_SB(sb)->s_bal_lock); | |
751 | EXT4_SB(sb)->s_mb_buddies_generated++; | |
752 | EXT4_SB(sb)->s_mb_generation_time += period; | |
753 | spin_unlock(&EXT4_SB(sb)->s_bal_lock); | |
754 | } | |
755 | ||
756 | /* The buddy information is attached the buddy cache inode | |
757 | * for convenience. The information regarding each group | |
758 | * is loaded via ext4_mb_load_buddy. The information involve | |
759 | * block bitmap and buddy information. The information are | |
760 | * stored in the inode as | |
761 | * | |
762 | * { page } | |
c3a326a6 | 763 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
764 | * |
765 | * | |
766 | * one block each for bitmap and buddy information. | |
767 | * So for each group we take up 2 blocks. A page can | |
768 | * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. | |
769 | * So it can have information regarding groups_per_page which | |
770 | * is blocks_per_page/2 | |
8a57d9d6 CW |
771 | * |
772 | * Locking note: This routine takes the block group lock of all groups | |
773 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
774 | */ |
775 | ||
776 | static int ext4_mb_init_cache(struct page *page, char *incore) | |
777 | { | |
8df9675f | 778 | ext4_group_t ngroups; |
c9de560d AT |
779 | int blocksize; |
780 | int blocks_per_page; | |
781 | int groups_per_page; | |
782 | int err = 0; | |
783 | int i; | |
784 | ext4_group_t first_group; | |
785 | int first_block; | |
786 | struct super_block *sb; | |
787 | struct buffer_head *bhs; | |
788 | struct buffer_head **bh; | |
789 | struct inode *inode; | |
790 | char *data; | |
791 | char *bitmap; | |
9b8b7d35 | 792 | struct ext4_group_info *grinfo; |
c9de560d | 793 | |
6ba495e9 | 794 | mb_debug(1, "init page %lu\n", page->index); |
c9de560d AT |
795 | |
796 | inode = page->mapping->host; | |
797 | sb = inode->i_sb; | |
8df9675f | 798 | ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
799 | blocksize = 1 << inode->i_blkbits; |
800 | blocks_per_page = PAGE_CACHE_SIZE / blocksize; | |
801 | ||
802 | groups_per_page = blocks_per_page >> 1; | |
803 | if (groups_per_page == 0) | |
804 | groups_per_page = 1; | |
805 | ||
806 | /* allocate buffer_heads to read bitmaps */ | |
807 | if (groups_per_page > 1) { | |
808 | err = -ENOMEM; | |
809 | i = sizeof(struct buffer_head *) * groups_per_page; | |
810 | bh = kzalloc(i, GFP_NOFS); | |
811 | if (bh == NULL) | |
812 | goto out; | |
813 | } else | |
814 | bh = &bhs; | |
815 | ||
816 | first_group = page->index * blocks_per_page / 2; | |
817 | ||
818 | /* read all groups the page covers into the cache */ | |
819 | for (i = 0; i < groups_per_page; i++) { | |
820 | struct ext4_group_desc *desc; | |
821 | ||
8df9675f | 822 | if (first_group + i >= ngroups) |
c9de560d AT |
823 | break; |
824 | ||
9b8b7d35 AG |
825 | grinfo = ext4_get_group_info(sb, first_group + i); |
826 | /* | |
827 | * If page is uptodate then we came here after online resize | |
828 | * which added some new uninitialized group info structs, so | |
829 | * we must skip all initialized uptodate buddies on the page, | |
830 | * which may be currently in use by an allocating task. | |
831 | */ | |
832 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
833 | bh[i] = NULL; | |
834 | continue; | |
835 | } | |
836 | ||
c9de560d AT |
837 | err = -EIO; |
838 | desc = ext4_get_group_desc(sb, first_group + i, NULL); | |
839 | if (desc == NULL) | |
840 | goto out; | |
841 | ||
842 | err = -ENOMEM; | |
843 | bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc)); | |
844 | if (bh[i] == NULL) | |
845 | goto out; | |
846 | ||
2ccb5fb9 | 847 | if (bitmap_uptodate(bh[i])) |
c9de560d AT |
848 | continue; |
849 | ||
c806e68f | 850 | lock_buffer(bh[i]); |
2ccb5fb9 AK |
851 | if (bitmap_uptodate(bh[i])) { |
852 | unlock_buffer(bh[i]); | |
853 | continue; | |
854 | } | |
955ce5f5 | 855 | ext4_lock_group(sb, first_group + i); |
c9de560d AT |
856 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
857 | ext4_init_block_bitmap(sb, bh[i], | |
858 | first_group + i, desc); | |
2ccb5fb9 | 859 | set_bitmap_uptodate(bh[i]); |
c9de560d | 860 | set_buffer_uptodate(bh[i]); |
955ce5f5 | 861 | ext4_unlock_group(sb, first_group + i); |
3300beda | 862 | unlock_buffer(bh[i]); |
c9de560d AT |
863 | continue; |
864 | } | |
955ce5f5 | 865 | ext4_unlock_group(sb, first_group + i); |
2ccb5fb9 AK |
866 | if (buffer_uptodate(bh[i])) { |
867 | /* | |
868 | * if not uninit if bh is uptodate, | |
869 | * bitmap is also uptodate | |
870 | */ | |
871 | set_bitmap_uptodate(bh[i]); | |
872 | unlock_buffer(bh[i]); | |
873 | continue; | |
874 | } | |
c9de560d | 875 | get_bh(bh[i]); |
2ccb5fb9 AK |
876 | /* |
877 | * submit the buffer_head for read. We can | |
878 | * safely mark the bitmap as uptodate now. | |
879 | * We do it here so the bitmap uptodate bit | |
880 | * get set with buffer lock held. | |
881 | */ | |
882 | set_bitmap_uptodate(bh[i]); | |
c9de560d AT |
883 | bh[i]->b_end_io = end_buffer_read_sync; |
884 | submit_bh(READ, bh[i]); | |
6ba495e9 | 885 | mb_debug(1, "read bitmap for group %u\n", first_group + i); |
c9de560d AT |
886 | } |
887 | ||
888 | /* wait for I/O completion */ | |
9b8b7d35 AG |
889 | for (i = 0; i < groups_per_page; i++) |
890 | if (bh[i]) | |
891 | wait_on_buffer(bh[i]); | |
c9de560d AT |
892 | |
893 | err = -EIO; | |
9b8b7d35 AG |
894 | for (i = 0; i < groups_per_page; i++) |
895 | if (bh[i] && !buffer_uptodate(bh[i])) | |
c9de560d AT |
896 | goto out; |
897 | ||
31b481dc | 898 | err = 0; |
c9de560d AT |
899 | first_block = page->index * blocks_per_page; |
900 | for (i = 0; i < blocks_per_page; i++) { | |
901 | int group; | |
c9de560d AT |
902 | |
903 | group = (first_block + i) >> 1; | |
8df9675f | 904 | if (group >= ngroups) |
c9de560d AT |
905 | break; |
906 | ||
9b8b7d35 AG |
907 | if (!bh[group - first_group]) |
908 | /* skip initialized uptodate buddy */ | |
909 | continue; | |
910 | ||
c9de560d AT |
911 | /* |
912 | * data carry information regarding this | |
913 | * particular group in the format specified | |
914 | * above | |
915 | * | |
916 | */ | |
917 | data = page_address(page) + (i * blocksize); | |
918 | bitmap = bh[group - first_group]->b_data; | |
919 | ||
920 | /* | |
921 | * We place the buddy block and bitmap block | |
922 | * close together | |
923 | */ | |
924 | if ((first_block + i) & 1) { | |
925 | /* this is block of buddy */ | |
926 | BUG_ON(incore == NULL); | |
6ba495e9 | 927 | mb_debug(1, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 928 | group, page->index, i * blocksize); |
f307333e | 929 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
930 | grinfo = ext4_get_group_info(sb, group); |
931 | grinfo->bb_fragments = 0; | |
932 | memset(grinfo->bb_counters, 0, | |
1927805e ES |
933 | sizeof(*grinfo->bb_counters) * |
934 | (sb->s_blocksize_bits+2)); | |
c9de560d AT |
935 | /* |
936 | * incore got set to the group block bitmap below | |
937 | */ | |
7a2fcbf7 | 938 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
939 | /* init the buddy */ |
940 | memset(data, 0xff, blocksize); | |
c9de560d | 941 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 942 | ext4_unlock_group(sb, group); |
c9de560d AT |
943 | incore = NULL; |
944 | } else { | |
945 | /* this is block of bitmap */ | |
946 | BUG_ON(incore != NULL); | |
6ba495e9 | 947 | mb_debug(1, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 948 | group, page->index, i * blocksize); |
f307333e | 949 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
950 | |
951 | /* see comments in ext4_mb_put_pa() */ | |
952 | ext4_lock_group(sb, group); | |
953 | memcpy(data, bitmap, blocksize); | |
954 | ||
955 | /* mark all preallocated blks used in in-core bitmap */ | |
956 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 957 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
958 | ext4_unlock_group(sb, group); |
959 | ||
960 | /* set incore so that the buddy information can be | |
961 | * generated using this | |
962 | */ | |
963 | incore = data; | |
964 | } | |
965 | } | |
966 | SetPageUptodate(page); | |
967 | ||
968 | out: | |
969 | if (bh) { | |
9b8b7d35 | 970 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
971 | brelse(bh[i]); |
972 | if (bh != &bhs) | |
973 | kfree(bh); | |
974 | } | |
975 | return err; | |
976 | } | |
977 | ||
eee4adc7 | 978 | /* |
2de8807b AG |
979 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
980 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
981 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
982 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 983 | */ |
2de8807b AG |
984 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
985 | ext4_group_t group, struct ext4_buddy *e4b) | |
eee4adc7 | 986 | { |
2de8807b AG |
987 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
988 | int block, pnum, poff; | |
eee4adc7 | 989 | int blocks_per_page; |
2de8807b AG |
990 | struct page *page; |
991 | ||
992 | e4b->bd_buddy_page = NULL; | |
993 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 ES |
994 | |
995 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
996 | /* | |
997 | * the buddy cache inode stores the block bitmap | |
998 | * and buddy information in consecutive blocks. | |
999 | * So for each group we need two blocks. | |
1000 | */ | |
1001 | block = group * 2; | |
1002 | pnum = block / blocks_per_page; | |
2de8807b AG |
1003 | poff = block % blocks_per_page; |
1004 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1005 | if (!page) | |
1006 | return -EIO; | |
1007 | BUG_ON(page->mapping != inode->i_mapping); | |
1008 | e4b->bd_bitmap_page = page; | |
1009 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1010 | ||
1011 | if (blocks_per_page >= 2) { | |
1012 | /* buddy and bitmap are on the same page */ | |
1013 | return 0; | |
eee4adc7 | 1014 | } |
2de8807b AG |
1015 | |
1016 | block++; | |
1017 | pnum = block / blocks_per_page; | |
1018 | poff = block % blocks_per_page; | |
1019 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1020 | if (!page) | |
1021 | return -EIO; | |
1022 | BUG_ON(page->mapping != inode->i_mapping); | |
1023 | e4b->bd_buddy_page = page; | |
1024 | return 0; | |
eee4adc7 ES |
1025 | } |
1026 | ||
2de8807b | 1027 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1028 | { |
2de8807b AG |
1029 | if (e4b->bd_bitmap_page) { |
1030 | unlock_page(e4b->bd_bitmap_page); | |
1031 | page_cache_release(e4b->bd_bitmap_page); | |
1032 | } | |
1033 | if (e4b->bd_buddy_page) { | |
1034 | unlock_page(e4b->bd_buddy_page); | |
1035 | page_cache_release(e4b->bd_buddy_page); | |
eee4adc7 | 1036 | } |
eee4adc7 ES |
1037 | } |
1038 | ||
8a57d9d6 CW |
1039 | /* |
1040 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1041 | * block group lock of all groups for this page; do not hold the BG lock when | |
1042 | * calling this routine! | |
1043 | */ | |
b6a758ec AK |
1044 | static noinline_for_stack |
1045 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) | |
1046 | { | |
1047 | ||
b6a758ec | 1048 | struct ext4_group_info *this_grp; |
2de8807b AG |
1049 | struct ext4_buddy e4b; |
1050 | struct page *page; | |
1051 | int ret = 0; | |
b6a758ec AK |
1052 | |
1053 | mb_debug(1, "init group %u\n", group); | |
b6a758ec AK |
1054 | this_grp = ext4_get_group_info(sb, group); |
1055 | /* | |
08c3a813 AK |
1056 | * This ensures that we don't reinit the buddy cache |
1057 | * page which map to the group from which we are already | |
1058 | * allocating. If we are looking at the buddy cache we would | |
1059 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1060 | * would have pinned buddy page to page cache. |
b6a758ec | 1061 | */ |
2de8807b AG |
1062 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b); |
1063 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { | |
b6a758ec AK |
1064 | /* |
1065 | * somebody initialized the group | |
1066 | * return without doing anything | |
1067 | */ | |
b6a758ec AK |
1068 | goto err; |
1069 | } | |
2de8807b AG |
1070 | |
1071 | page = e4b.bd_bitmap_page; | |
1072 | ret = ext4_mb_init_cache(page, NULL); | |
1073 | if (ret) | |
1074 | goto err; | |
1075 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1076 | ret = -EIO; |
1077 | goto err; | |
1078 | } | |
1079 | mark_page_accessed(page); | |
b6a758ec | 1080 | |
2de8807b | 1081 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1082 | /* |
1083 | * If both the bitmap and buddy are in | |
1084 | * the same page we don't need to force | |
1085 | * init the buddy | |
1086 | */ | |
2de8807b AG |
1087 | ret = 0; |
1088 | goto err; | |
b6a758ec | 1089 | } |
2de8807b AG |
1090 | /* init buddy cache */ |
1091 | page = e4b.bd_buddy_page; | |
1092 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap); | |
1093 | if (ret) | |
1094 | goto err; | |
1095 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1096 | ret = -EIO; |
1097 | goto err; | |
1098 | } | |
1099 | mark_page_accessed(page); | |
1100 | err: | |
2de8807b | 1101 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1102 | return ret; |
1103 | } | |
1104 | ||
8a57d9d6 CW |
1105 | /* |
1106 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1107 | * block group lock of all groups for this page; do not hold the BG lock when | |
1108 | * calling this routine! | |
1109 | */ | |
4ddfef7b ES |
1110 | static noinline_for_stack int |
1111 | ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, | |
1112 | struct ext4_buddy *e4b) | |
c9de560d | 1113 | { |
c9de560d AT |
1114 | int blocks_per_page; |
1115 | int block; | |
1116 | int pnum; | |
1117 | int poff; | |
1118 | struct page *page; | |
fdf6c7a7 | 1119 | int ret; |
920313a7 AK |
1120 | struct ext4_group_info *grp; |
1121 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1122 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1123 | |
6ba495e9 | 1124 | mb_debug(1, "load group %u\n", group); |
c9de560d AT |
1125 | |
1126 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
920313a7 | 1127 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1128 | |
1129 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1130 | e4b->bd_info = grp; |
c9de560d AT |
1131 | e4b->bd_sb = sb; |
1132 | e4b->bd_group = group; | |
1133 | e4b->bd_buddy_page = NULL; | |
1134 | e4b->bd_bitmap_page = NULL; | |
1135 | ||
f41c0750 | 1136 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1137 | /* |
1138 | * we need full data about the group | |
1139 | * to make a good selection | |
1140 | */ | |
1141 | ret = ext4_mb_init_group(sb, group); | |
1142 | if (ret) | |
1143 | return ret; | |
f41c0750 AK |
1144 | } |
1145 | ||
c9de560d AT |
1146 | /* |
1147 | * the buddy cache inode stores the block bitmap | |
1148 | * and buddy information in consecutive blocks. | |
1149 | * So for each group we need two blocks. | |
1150 | */ | |
1151 | block = group * 2; | |
1152 | pnum = block / blocks_per_page; | |
1153 | poff = block % blocks_per_page; | |
1154 | ||
1155 | /* we could use find_or_create_page(), but it locks page | |
1156 | * what we'd like to avoid in fast path ... */ | |
1157 | page = find_get_page(inode->i_mapping, pnum); | |
1158 | if (page == NULL || !PageUptodate(page)) { | |
1159 | if (page) | |
920313a7 AK |
1160 | /* |
1161 | * drop the page reference and try | |
1162 | * to get the page with lock. If we | |
1163 | * are not uptodate that implies | |
1164 | * somebody just created the page but | |
1165 | * is yet to initialize the same. So | |
1166 | * wait for it to initialize. | |
1167 | */ | |
c9de560d AT |
1168 | page_cache_release(page); |
1169 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1170 | if (page) { | |
1171 | BUG_ON(page->mapping != inode->i_mapping); | |
1172 | if (!PageUptodate(page)) { | |
fdf6c7a7 SF |
1173 | ret = ext4_mb_init_cache(page, NULL); |
1174 | if (ret) { | |
1175 | unlock_page(page); | |
1176 | goto err; | |
1177 | } | |
c9de560d AT |
1178 | mb_cmp_bitmaps(e4b, page_address(page) + |
1179 | (poff * sb->s_blocksize)); | |
1180 | } | |
1181 | unlock_page(page); | |
1182 | } | |
1183 | } | |
fdf6c7a7 SF |
1184 | if (page == NULL || !PageUptodate(page)) { |
1185 | ret = -EIO; | |
c9de560d | 1186 | goto err; |
fdf6c7a7 | 1187 | } |
c9de560d AT |
1188 | e4b->bd_bitmap_page = page; |
1189 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1190 | mark_page_accessed(page); | |
1191 | ||
1192 | block++; | |
1193 | pnum = block / blocks_per_page; | |
1194 | poff = block % blocks_per_page; | |
1195 | ||
1196 | page = find_get_page(inode->i_mapping, pnum); | |
1197 | if (page == NULL || !PageUptodate(page)) { | |
1198 | if (page) | |
1199 | page_cache_release(page); | |
1200 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1201 | if (page) { | |
1202 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 SF |
1203 | if (!PageUptodate(page)) { |
1204 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap); | |
1205 | if (ret) { | |
1206 | unlock_page(page); | |
1207 | goto err; | |
1208 | } | |
1209 | } | |
c9de560d AT |
1210 | unlock_page(page); |
1211 | } | |
1212 | } | |
fdf6c7a7 SF |
1213 | if (page == NULL || !PageUptodate(page)) { |
1214 | ret = -EIO; | |
c9de560d | 1215 | goto err; |
fdf6c7a7 | 1216 | } |
c9de560d AT |
1217 | e4b->bd_buddy_page = page; |
1218 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
1219 | mark_page_accessed(page); | |
1220 | ||
1221 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1222 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1223 | ||
1224 | return 0; | |
1225 | ||
1226 | err: | |
26626f11 YR |
1227 | if (page) |
1228 | page_cache_release(page); | |
c9de560d AT |
1229 | if (e4b->bd_bitmap_page) |
1230 | page_cache_release(e4b->bd_bitmap_page); | |
1231 | if (e4b->bd_buddy_page) | |
1232 | page_cache_release(e4b->bd_buddy_page); | |
1233 | e4b->bd_buddy = NULL; | |
1234 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1235 | return ret; |
c9de560d AT |
1236 | } |
1237 | ||
e39e07fd | 1238 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1239 | { |
1240 | if (e4b->bd_bitmap_page) | |
1241 | page_cache_release(e4b->bd_bitmap_page); | |
1242 | if (e4b->bd_buddy_page) | |
1243 | page_cache_release(e4b->bd_buddy_page); | |
1244 | } | |
1245 | ||
1246 | ||
1247 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1248 | { | |
1249 | int order = 1; | |
1250 | void *bb; | |
1251 | ||
1252 | BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); | |
1253 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); | |
1254 | ||
1255 | bb = EXT4_MB_BUDDY(e4b); | |
1256 | while (order <= e4b->bd_blkbits + 1) { | |
1257 | block = block >> 1; | |
1258 | if (!mb_test_bit(block, bb)) { | |
1259 | /* this block is part of buddy of order 'order' */ | |
1260 | return order; | |
1261 | } | |
1262 | bb += 1 << (e4b->bd_blkbits - order); | |
1263 | order++; | |
1264 | } | |
1265 | return 0; | |
1266 | } | |
1267 | ||
955ce5f5 | 1268 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1269 | { |
1270 | __u32 *addr; | |
1271 | ||
1272 | len = cur + len; | |
1273 | while (cur < len) { | |
1274 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1275 | /* fast path: clear whole word at once */ | |
1276 | addr = bm + (cur >> 3); | |
1277 | *addr = 0; | |
1278 | cur += 32; | |
1279 | continue; | |
1280 | } | |
955ce5f5 | 1281 | mb_clear_bit(cur, bm); |
c9de560d AT |
1282 | cur++; |
1283 | } | |
1284 | } | |
1285 | ||
c3e94d1d | 1286 | void ext4_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1287 | { |
1288 | __u32 *addr; | |
1289 | ||
1290 | len = cur + len; | |
1291 | while (cur < len) { | |
1292 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1293 | /* fast path: set whole word at once */ | |
1294 | addr = bm + (cur >> 3); | |
1295 | *addr = 0xffffffff; | |
1296 | cur += 32; | |
1297 | continue; | |
1298 | } | |
955ce5f5 | 1299 | mb_set_bit(cur, bm); |
c9de560d AT |
1300 | cur++; |
1301 | } | |
1302 | } | |
1303 | ||
7e5a8cdd | 1304 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
c9de560d AT |
1305 | int first, int count) |
1306 | { | |
1307 | int block = 0; | |
1308 | int max = 0; | |
1309 | int order; | |
1310 | void *buddy; | |
1311 | void *buddy2; | |
1312 | struct super_block *sb = e4b->bd_sb; | |
1313 | ||
1314 | BUG_ON(first + count > (sb->s_blocksize << 3)); | |
bc8e6740 | 1315 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
1316 | mb_check_buddy(e4b); |
1317 | mb_free_blocks_double(inode, e4b, first, count); | |
1318 | ||
1319 | e4b->bd_info->bb_free += count; | |
1320 | if (first < e4b->bd_info->bb_first_free) | |
1321 | e4b->bd_info->bb_first_free = first; | |
1322 | ||
1323 | /* let's maintain fragments counter */ | |
1324 | if (first != 0) | |
1325 | block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b)); | |
1326 | if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) | |
1327 | max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b)); | |
1328 | if (block && max) | |
1329 | e4b->bd_info->bb_fragments--; | |
1330 | else if (!block && !max) | |
1331 | e4b->bd_info->bb_fragments++; | |
1332 | ||
1333 | /* let's maintain buddy itself */ | |
1334 | while (count-- > 0) { | |
1335 | block = first++; | |
1336 | order = 0; | |
1337 | ||
1338 | if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) { | |
1339 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
1340 | |
1341 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 1342 | blocknr += EXT4_C2B(EXT4_SB(sb), block); |
5d1b1b3f | 1343 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
1344 | inode ? inode->i_ino : 0, |
1345 | blocknr, | |
1346 | "freeing already freed block " | |
1347 | "(bit %u)", block); | |
c9de560d AT |
1348 | } |
1349 | mb_clear_bit(block, EXT4_MB_BITMAP(e4b)); | |
1350 | e4b->bd_info->bb_counters[order]++; | |
1351 | ||
1352 | /* start of the buddy */ | |
1353 | buddy = mb_find_buddy(e4b, order, &max); | |
1354 | ||
1355 | do { | |
1356 | block &= ~1UL; | |
1357 | if (mb_test_bit(block, buddy) || | |
1358 | mb_test_bit(block + 1, buddy)) | |
1359 | break; | |
1360 | ||
1361 | /* both the buddies are free, try to coalesce them */ | |
1362 | buddy2 = mb_find_buddy(e4b, order + 1, &max); | |
1363 | ||
1364 | if (!buddy2) | |
1365 | break; | |
1366 | ||
1367 | if (order > 0) { | |
1368 | /* for special purposes, we don't set | |
1369 | * free bits in bitmap */ | |
1370 | mb_set_bit(block, buddy); | |
1371 | mb_set_bit(block + 1, buddy); | |
1372 | } | |
1373 | e4b->bd_info->bb_counters[order]--; | |
1374 | e4b->bd_info->bb_counters[order]--; | |
1375 | ||
1376 | block = block >> 1; | |
1377 | order++; | |
1378 | e4b->bd_info->bb_counters[order]++; | |
1379 | ||
1380 | mb_clear_bit(block, buddy2); | |
1381 | buddy = buddy2; | |
1382 | } while (1); | |
1383 | } | |
8a57d9d6 | 1384 | mb_set_largest_free_order(sb, e4b->bd_info); |
c9de560d | 1385 | mb_check_buddy(e4b); |
c9de560d AT |
1386 | } |
1387 | ||
1388 | static int mb_find_extent(struct ext4_buddy *e4b, int order, int block, | |
1389 | int needed, struct ext4_free_extent *ex) | |
1390 | { | |
1391 | int next = block; | |
1392 | int max; | |
1393 | int ord; | |
1394 | void *buddy; | |
1395 | ||
bc8e6740 | 1396 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1397 | BUG_ON(ex == NULL); |
1398 | ||
1399 | buddy = mb_find_buddy(e4b, order, &max); | |
1400 | BUG_ON(buddy == NULL); | |
1401 | BUG_ON(block >= max); | |
1402 | if (mb_test_bit(block, buddy)) { | |
1403 | ex->fe_len = 0; | |
1404 | ex->fe_start = 0; | |
1405 | ex->fe_group = 0; | |
1406 | return 0; | |
1407 | } | |
1408 | ||
1409 | /* FIXME dorp order completely ? */ | |
1410 | if (likely(order == 0)) { | |
1411 | /* find actual order */ | |
1412 | order = mb_find_order_for_block(e4b, block); | |
1413 | block = block >> order; | |
1414 | } | |
1415 | ||
1416 | ex->fe_len = 1 << order; | |
1417 | ex->fe_start = block << order; | |
1418 | ex->fe_group = e4b->bd_group; | |
1419 | ||
1420 | /* calc difference from given start */ | |
1421 | next = next - ex->fe_start; | |
1422 | ex->fe_len -= next; | |
1423 | ex->fe_start += next; | |
1424 | ||
1425 | while (needed > ex->fe_len && | |
1426 | (buddy = mb_find_buddy(e4b, order, &max))) { | |
1427 | ||
1428 | if (block + 1 >= max) | |
1429 | break; | |
1430 | ||
1431 | next = (block + 1) * (1 << order); | |
1432 | if (mb_test_bit(next, EXT4_MB_BITMAP(e4b))) | |
1433 | break; | |
1434 | ||
1435 | ord = mb_find_order_for_block(e4b, next); | |
1436 | ||
1437 | order = ord; | |
1438 | block = next >> order; | |
1439 | ex->fe_len += 1 << order; | |
1440 | } | |
1441 | ||
1442 | BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); | |
1443 | return ex->fe_len; | |
1444 | } | |
1445 | ||
1446 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1447 | { | |
1448 | int ord; | |
1449 | int mlen = 0; | |
1450 | int max = 0; | |
1451 | int cur; | |
1452 | int start = ex->fe_start; | |
1453 | int len = ex->fe_len; | |
1454 | unsigned ret = 0; | |
1455 | int len0 = len; | |
1456 | void *buddy; | |
1457 | ||
1458 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1459 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1460 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1461 | mb_check_buddy(e4b); |
1462 | mb_mark_used_double(e4b, start, len); | |
1463 | ||
1464 | e4b->bd_info->bb_free -= len; | |
1465 | if (e4b->bd_info->bb_first_free == start) | |
1466 | e4b->bd_info->bb_first_free += len; | |
1467 | ||
1468 | /* let's maintain fragments counter */ | |
1469 | if (start != 0) | |
1470 | mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b)); | |
1471 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) | |
1472 | max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b)); | |
1473 | if (mlen && max) | |
1474 | e4b->bd_info->bb_fragments++; | |
1475 | else if (!mlen && !max) | |
1476 | e4b->bd_info->bb_fragments--; | |
1477 | ||
1478 | /* let's maintain buddy itself */ | |
1479 | while (len) { | |
1480 | ord = mb_find_order_for_block(e4b, start); | |
1481 | ||
1482 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1483 | /* the whole chunk may be allocated at once! */ | |
1484 | mlen = 1 << ord; | |
1485 | buddy = mb_find_buddy(e4b, ord, &max); | |
1486 | BUG_ON((start >> ord) >= max); | |
1487 | mb_set_bit(start >> ord, buddy); | |
1488 | e4b->bd_info->bb_counters[ord]--; | |
1489 | start += mlen; | |
1490 | len -= mlen; | |
1491 | BUG_ON(len < 0); | |
1492 | continue; | |
1493 | } | |
1494 | ||
1495 | /* store for history */ | |
1496 | if (ret == 0) | |
1497 | ret = len | (ord << 16); | |
1498 | ||
1499 | /* we have to split large buddy */ | |
1500 | BUG_ON(ord <= 0); | |
1501 | buddy = mb_find_buddy(e4b, ord, &max); | |
1502 | mb_set_bit(start >> ord, buddy); | |
1503 | e4b->bd_info->bb_counters[ord]--; | |
1504 | ||
1505 | ord--; | |
1506 | cur = (start >> ord) & ~1U; | |
1507 | buddy = mb_find_buddy(e4b, ord, &max); | |
1508 | mb_clear_bit(cur, buddy); | |
1509 | mb_clear_bit(cur + 1, buddy); | |
1510 | e4b->bd_info->bb_counters[ord]++; | |
1511 | e4b->bd_info->bb_counters[ord]++; | |
1512 | } | |
8a57d9d6 | 1513 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1514 | |
c3e94d1d | 1515 | ext4_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0); |
c9de560d AT |
1516 | mb_check_buddy(e4b); |
1517 | ||
1518 | return ret; | |
1519 | } | |
1520 | ||
1521 | /* | |
1522 | * Must be called under group lock! | |
1523 | */ | |
1524 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1525 | struct ext4_buddy *e4b) | |
1526 | { | |
1527 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1528 | int ret; | |
1529 | ||
1530 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1531 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1532 | ||
1533 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1534 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1535 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1536 | ||
1537 | /* preallocation can change ac_b_ex, thus we store actually | |
1538 | * allocated blocks for history */ | |
1539 | ac->ac_f_ex = ac->ac_b_ex; | |
1540 | ||
1541 | ac->ac_status = AC_STATUS_FOUND; | |
1542 | ac->ac_tail = ret & 0xffff; | |
1543 | ac->ac_buddy = ret >> 16; | |
1544 | ||
c3a326a6 AK |
1545 | /* |
1546 | * take the page reference. We want the page to be pinned | |
1547 | * so that we don't get a ext4_mb_init_cache_call for this | |
1548 | * group until we update the bitmap. That would mean we | |
1549 | * double allocate blocks. The reference is dropped | |
1550 | * in ext4_mb_release_context | |
1551 | */ | |
c9de560d AT |
1552 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1553 | get_page(ac->ac_bitmap_page); | |
1554 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1555 | get_page(ac->ac_buddy_page); | |
c9de560d | 1556 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 1557 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
1558 | spin_lock(&sbi->s_md_lock); |
1559 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1560 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1561 | spin_unlock(&sbi->s_md_lock); | |
1562 | } | |
1563 | } | |
1564 | ||
1565 | /* | |
1566 | * regular allocator, for general purposes allocation | |
1567 | */ | |
1568 | ||
1569 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1570 | struct ext4_buddy *e4b, | |
1571 | int finish_group) | |
1572 | { | |
1573 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1574 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1575 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1576 | struct ext4_free_extent ex; | |
1577 | int max; | |
1578 | ||
032115fc AK |
1579 | if (ac->ac_status == AC_STATUS_FOUND) |
1580 | return; | |
c9de560d AT |
1581 | /* |
1582 | * We don't want to scan for a whole year | |
1583 | */ | |
1584 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1585 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1586 | ac->ac_status = AC_STATUS_BREAK; | |
1587 | return; | |
1588 | } | |
1589 | ||
1590 | /* | |
1591 | * Haven't found good chunk so far, let's continue | |
1592 | */ | |
1593 | if (bex->fe_len < gex->fe_len) | |
1594 | return; | |
1595 | ||
1596 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1597 | && bex->fe_group == e4b->bd_group) { | |
1598 | /* recheck chunk's availability - we don't know | |
1599 | * when it was found (within this lock-unlock | |
1600 | * period or not) */ | |
1601 | max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex); | |
1602 | if (max >= gex->fe_len) { | |
1603 | ext4_mb_use_best_found(ac, e4b); | |
1604 | return; | |
1605 | } | |
1606 | } | |
1607 | } | |
1608 | ||
1609 | /* | |
1610 | * The routine checks whether found extent is good enough. If it is, | |
1611 | * then the extent gets marked used and flag is set to the context | |
1612 | * to stop scanning. Otherwise, the extent is compared with the | |
1613 | * previous found extent and if new one is better, then it's stored | |
1614 | * in the context. Later, the best found extent will be used, if | |
1615 | * mballoc can't find good enough extent. | |
1616 | * | |
1617 | * FIXME: real allocation policy is to be designed yet! | |
1618 | */ | |
1619 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1620 | struct ext4_free_extent *ex, | |
1621 | struct ext4_buddy *e4b) | |
1622 | { | |
1623 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1624 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1625 | ||
1626 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
1627 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
1628 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
1629 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
1630 | ||
1631 | ac->ac_found++; | |
1632 | ||
1633 | /* | |
1634 | * The special case - take what you catch first | |
1635 | */ | |
1636 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1637 | *bex = *ex; | |
1638 | ext4_mb_use_best_found(ac, e4b); | |
1639 | return; | |
1640 | } | |
1641 | ||
1642 | /* | |
1643 | * Let's check whether the chuck is good enough | |
1644 | */ | |
1645 | if (ex->fe_len == gex->fe_len) { | |
1646 | *bex = *ex; | |
1647 | ext4_mb_use_best_found(ac, e4b); | |
1648 | return; | |
1649 | } | |
1650 | ||
1651 | /* | |
1652 | * If this is first found extent, just store it in the context | |
1653 | */ | |
1654 | if (bex->fe_len == 0) { | |
1655 | *bex = *ex; | |
1656 | return; | |
1657 | } | |
1658 | ||
1659 | /* | |
1660 | * If new found extent is better, store it in the context | |
1661 | */ | |
1662 | if (bex->fe_len < gex->fe_len) { | |
1663 | /* if the request isn't satisfied, any found extent | |
1664 | * larger than previous best one is better */ | |
1665 | if (ex->fe_len > bex->fe_len) | |
1666 | *bex = *ex; | |
1667 | } else if (ex->fe_len > gex->fe_len) { | |
1668 | /* if the request is satisfied, then we try to find | |
1669 | * an extent that still satisfy the request, but is | |
1670 | * smaller than previous one */ | |
1671 | if (ex->fe_len < bex->fe_len) | |
1672 | *bex = *ex; | |
1673 | } | |
1674 | ||
1675 | ext4_mb_check_limits(ac, e4b, 0); | |
1676 | } | |
1677 | ||
089ceecc ES |
1678 | static noinline_for_stack |
1679 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
1680 | struct ext4_buddy *e4b) |
1681 | { | |
1682 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1683 | ext4_group_t group = ex.fe_group; | |
1684 | int max; | |
1685 | int err; | |
1686 | ||
1687 | BUG_ON(ex.fe_len <= 0); | |
1688 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1689 | if (err) | |
1690 | return err; | |
1691 | ||
1692 | ext4_lock_group(ac->ac_sb, group); | |
1693 | max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex); | |
1694 | ||
1695 | if (max > 0) { | |
1696 | ac->ac_b_ex = ex; | |
1697 | ext4_mb_use_best_found(ac, e4b); | |
1698 | } | |
1699 | ||
1700 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1701 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1702 | |
1703 | return 0; | |
1704 | } | |
1705 | ||
089ceecc ES |
1706 | static noinline_for_stack |
1707 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
1708 | struct ext4_buddy *e4b) |
1709 | { | |
1710 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1711 | int max; | |
1712 | int err; | |
1713 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
c9de560d AT |
1714 | struct ext4_free_extent ex; |
1715 | ||
1716 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1717 | return 0; | |
1718 | ||
1719 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1720 | if (err) | |
1721 | return err; | |
1722 | ||
1723 | ext4_lock_group(ac->ac_sb, group); | |
1724 | max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start, | |
1725 | ac->ac_g_ex.fe_len, &ex); | |
1726 | ||
1727 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1728 | ext4_fsblk_t start; | |
1729 | ||
5661bd68 AM |
1730 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
1731 | ex.fe_start; | |
c9de560d AT |
1732 | /* use do_div to get remainder (would be 64-bit modulo) */ |
1733 | if (do_div(start, sbi->s_stripe) == 0) { | |
1734 | ac->ac_found++; | |
1735 | ac->ac_b_ex = ex; | |
1736 | ext4_mb_use_best_found(ac, e4b); | |
1737 | } | |
1738 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1739 | BUG_ON(ex.fe_len <= 0); | |
1740 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1741 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1742 | ac->ac_found++; | |
1743 | ac->ac_b_ex = ex; | |
1744 | ext4_mb_use_best_found(ac, e4b); | |
1745 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1746 | /* Sometimes, caller may want to merge even small | |
1747 | * number of blocks to an existing extent */ | |
1748 | BUG_ON(ex.fe_len <= 0); | |
1749 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1750 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1751 | ac->ac_found++; | |
1752 | ac->ac_b_ex = ex; | |
1753 | ext4_mb_use_best_found(ac, e4b); | |
1754 | } | |
1755 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1756 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1757 | |
1758 | return 0; | |
1759 | } | |
1760 | ||
1761 | /* | |
1762 | * The routine scans buddy structures (not bitmap!) from given order | |
1763 | * to max order and tries to find big enough chunk to satisfy the req | |
1764 | */ | |
089ceecc ES |
1765 | static noinline_for_stack |
1766 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1767 | struct ext4_buddy *e4b) |
1768 | { | |
1769 | struct super_block *sb = ac->ac_sb; | |
1770 | struct ext4_group_info *grp = e4b->bd_info; | |
1771 | void *buddy; | |
1772 | int i; | |
1773 | int k; | |
1774 | int max; | |
1775 | ||
1776 | BUG_ON(ac->ac_2order <= 0); | |
1777 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1778 | if (grp->bb_counters[i] == 0) | |
1779 | continue; | |
1780 | ||
1781 | buddy = mb_find_buddy(e4b, i, &max); | |
1782 | BUG_ON(buddy == NULL); | |
1783 | ||
ffad0a44 | 1784 | k = mb_find_next_zero_bit(buddy, max, 0); |
c9de560d AT |
1785 | BUG_ON(k >= max); |
1786 | ||
1787 | ac->ac_found++; | |
1788 | ||
1789 | ac->ac_b_ex.fe_len = 1 << i; | |
1790 | ac->ac_b_ex.fe_start = k << i; | |
1791 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1792 | ||
1793 | ext4_mb_use_best_found(ac, e4b); | |
1794 | ||
1795 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1796 | ||
1797 | if (EXT4_SB(sb)->s_mb_stats) | |
1798 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1799 | ||
1800 | break; | |
1801 | } | |
1802 | } | |
1803 | ||
1804 | /* | |
1805 | * The routine scans the group and measures all found extents. | |
1806 | * In order to optimize scanning, caller must pass number of | |
1807 | * free blocks in the group, so the routine can know upper limit. | |
1808 | */ | |
089ceecc ES |
1809 | static noinline_for_stack |
1810 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1811 | struct ext4_buddy *e4b) |
1812 | { | |
1813 | struct super_block *sb = ac->ac_sb; | |
1814 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1815 | struct ext4_free_extent ex; | |
1816 | int i; | |
1817 | int free; | |
1818 | ||
1819 | free = e4b->bd_info->bb_free; | |
1820 | BUG_ON(free <= 0); | |
1821 | ||
1822 | i = e4b->bd_info->bb_first_free; | |
1823 | ||
1824 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1825 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
1826 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
1827 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 1828 | /* |
e56eb659 | 1829 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1830 | * free blocks even though group info says we |
1831 | * we have free blocks | |
1832 | */ | |
e29136f8 | 1833 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1834 | "%d free clusters as per " |
fde4d95a | 1835 | "group info. But bitmap says 0", |
26346ff6 | 1836 | free); |
c9de560d AT |
1837 | break; |
1838 | } | |
1839 | ||
1840 | mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex); | |
1841 | BUG_ON(ex.fe_len <= 0); | |
26346ff6 | 1842 | if (free < ex.fe_len) { |
e29136f8 | 1843 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1844 | "%d free clusters as per " |
fde4d95a | 1845 | "group info. But got %d blocks", |
26346ff6 | 1846 | free, ex.fe_len); |
e56eb659 AK |
1847 | /* |
1848 | * The number of free blocks differs. This mostly | |
1849 | * indicate that the bitmap is corrupt. So exit | |
1850 | * without claiming the space. | |
1851 | */ | |
1852 | break; | |
26346ff6 | 1853 | } |
c9de560d AT |
1854 | |
1855 | ext4_mb_measure_extent(ac, &ex, e4b); | |
1856 | ||
1857 | i += ex.fe_len; | |
1858 | free -= ex.fe_len; | |
1859 | } | |
1860 | ||
1861 | ext4_mb_check_limits(ac, e4b, 1); | |
1862 | } | |
1863 | ||
1864 | /* | |
1865 | * This is a special case for storages like raid5 | |
506bf2d8 | 1866 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 1867 | */ |
089ceecc ES |
1868 | static noinline_for_stack |
1869 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
1870 | struct ext4_buddy *e4b) |
1871 | { | |
1872 | struct super_block *sb = ac->ac_sb; | |
1873 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1874 | void *bitmap = EXT4_MB_BITMAP(e4b); | |
1875 | struct ext4_free_extent ex; | |
1876 | ext4_fsblk_t first_group_block; | |
1877 | ext4_fsblk_t a; | |
1878 | ext4_grpblk_t i; | |
1879 | int max; | |
1880 | ||
1881 | BUG_ON(sbi->s_stripe == 0); | |
1882 | ||
1883 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
1884 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
1885 | ||
c9de560d AT |
1886 | a = first_group_block + sbi->s_stripe - 1; |
1887 | do_div(a, sbi->s_stripe); | |
1888 | i = (a * sbi->s_stripe) - first_group_block; | |
1889 | ||
7137d7a4 | 1890 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d AT |
1891 | if (!mb_test_bit(i, bitmap)) { |
1892 | max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex); | |
1893 | if (max >= sbi->s_stripe) { | |
1894 | ac->ac_found++; | |
1895 | ac->ac_b_ex = ex; | |
1896 | ext4_mb_use_best_found(ac, e4b); | |
1897 | break; | |
1898 | } | |
1899 | } | |
1900 | i += sbi->s_stripe; | |
1901 | } | |
1902 | } | |
1903 | ||
8a57d9d6 | 1904 | /* This is now called BEFORE we load the buddy bitmap. */ |
c9de560d AT |
1905 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, |
1906 | ext4_group_t group, int cr) | |
1907 | { | |
1908 | unsigned free, fragments; | |
a4912123 | 1909 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
1910 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
1911 | ||
1912 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 CW |
1913 | |
1914 | /* We only do this if the grp has never been initialized */ | |
1915 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
1916 | int ret = ext4_mb_init_group(ac->ac_sb, group); | |
1917 | if (ret) | |
1918 | return 0; | |
1919 | } | |
c9de560d AT |
1920 | |
1921 | free = grp->bb_free; | |
1922 | fragments = grp->bb_fragments; | |
1923 | if (free == 0) | |
1924 | return 0; | |
1925 | if (fragments == 0) | |
1926 | return 0; | |
1927 | ||
1928 | switch (cr) { | |
1929 | case 0: | |
1930 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 1931 | |
8a57d9d6 CW |
1932 | if (grp->bb_largest_free_order < ac->ac_2order) |
1933 | return 0; | |
1934 | ||
a4912123 TT |
1935 | /* Avoid using the first bg of a flexgroup for data files */ |
1936 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
1937 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
1938 | ((group % flex_size) == 0)) | |
1939 | return 0; | |
1940 | ||
8a57d9d6 | 1941 | return 1; |
c9de560d AT |
1942 | case 1: |
1943 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
1944 | return 1; | |
1945 | break; | |
1946 | case 2: | |
1947 | if (free >= ac->ac_g_ex.fe_len) | |
1948 | return 1; | |
1949 | break; | |
1950 | case 3: | |
1951 | return 1; | |
1952 | default: | |
1953 | BUG(); | |
1954 | } | |
1955 | ||
1956 | return 0; | |
1957 | } | |
1958 | ||
4ddfef7b ES |
1959 | static noinline_for_stack int |
1960 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 1961 | { |
8df9675f | 1962 | ext4_group_t ngroups, group, i; |
c9de560d AT |
1963 | int cr; |
1964 | int err = 0; | |
c9de560d AT |
1965 | struct ext4_sb_info *sbi; |
1966 | struct super_block *sb; | |
1967 | struct ext4_buddy e4b; | |
c9de560d AT |
1968 | |
1969 | sb = ac->ac_sb; | |
1970 | sbi = EXT4_SB(sb); | |
8df9675f | 1971 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 1972 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 1973 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
1974 | ngroups = sbi->s_blockfile_groups; |
1975 | ||
c9de560d AT |
1976 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
1977 | ||
1978 | /* first, try the goal */ | |
1979 | err = ext4_mb_find_by_goal(ac, &e4b); | |
1980 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
1981 | goto out; | |
1982 | ||
1983 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
1984 | goto out; | |
1985 | ||
1986 | /* | |
1987 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
1988 | * if ac2_order is set we also set criteria to 0 so that we | |
1989 | * try exact allocation using buddy. | |
1990 | */ | |
1991 | i = fls(ac->ac_g_ex.fe_len); | |
1992 | ac->ac_2order = 0; | |
1993 | /* | |
1994 | * We search using buddy data only if the order of the request | |
1995 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 1996 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
c9de560d AT |
1997 | */ |
1998 | if (i >= sbi->s_mb_order2_reqs) { | |
1999 | /* | |
2000 | * This should tell if fe_len is exactly power of 2 | |
2001 | */ | |
2002 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
2003 | ac->ac_2order = i - 1; | |
2004 | } | |
2005 | ||
4ba74d00 TT |
2006 | /* if stream allocation is enabled, use global goal */ |
2007 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2008 | /* TBD: may be hot point */ |
2009 | spin_lock(&sbi->s_md_lock); | |
2010 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2011 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2012 | spin_unlock(&sbi->s_md_lock); | |
2013 | } | |
4ba74d00 | 2014 | |
c9de560d AT |
2015 | /* Let's just scan groups to find more-less suitable blocks */ |
2016 | cr = ac->ac_2order ? 0 : 1; | |
2017 | /* | |
2018 | * cr == 0 try to get exact allocation, | |
2019 | * cr == 3 try to get anything | |
2020 | */ | |
2021 | repeat: | |
2022 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2023 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2024 | /* |
2025 | * searching for the right group start | |
2026 | * from the goal value specified | |
2027 | */ | |
2028 | group = ac->ac_g_ex.fe_group; | |
2029 | ||
8df9675f | 2030 | for (i = 0; i < ngroups; group++, i++) { |
8df9675f | 2031 | if (group == ngroups) |
c9de560d AT |
2032 | group = 0; |
2033 | ||
8a57d9d6 CW |
2034 | /* This now checks without needing the buddy page */ |
2035 | if (!ext4_mb_good_group(ac, group, cr)) | |
c9de560d AT |
2036 | continue; |
2037 | ||
c9de560d AT |
2038 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2039 | if (err) | |
2040 | goto out; | |
2041 | ||
2042 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2043 | |
2044 | /* | |
2045 | * We need to check again after locking the | |
2046 | * block group | |
2047 | */ | |
c9de560d | 2048 | if (!ext4_mb_good_group(ac, group, cr)) { |
c9de560d | 2049 | ext4_unlock_group(sb, group); |
e39e07fd | 2050 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2051 | continue; |
2052 | } | |
2053 | ||
2054 | ac->ac_groups_scanned++; | |
75507efb | 2055 | if (cr == 0) |
c9de560d | 2056 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2057 | else if (cr == 1 && sbi->s_stripe && |
2058 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2059 | ext4_mb_scan_aligned(ac, &e4b); |
2060 | else | |
2061 | ext4_mb_complex_scan_group(ac, &e4b); | |
2062 | ||
2063 | ext4_unlock_group(sb, group); | |
e39e07fd | 2064 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2065 | |
2066 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2067 | break; | |
2068 | } | |
2069 | } | |
2070 | ||
2071 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2072 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2073 | /* | |
2074 | * We've been searching too long. Let's try to allocate | |
2075 | * the best chunk we've found so far | |
2076 | */ | |
2077 | ||
2078 | ext4_mb_try_best_found(ac, &e4b); | |
2079 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2080 | /* | |
2081 | * Someone more lucky has already allocated it. | |
2082 | * The only thing we can do is just take first | |
2083 | * found block(s) | |
2084 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2085 | */ | |
2086 | ac->ac_b_ex.fe_group = 0; | |
2087 | ac->ac_b_ex.fe_start = 0; | |
2088 | ac->ac_b_ex.fe_len = 0; | |
2089 | ac->ac_status = AC_STATUS_CONTINUE; | |
2090 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2091 | cr = 3; | |
2092 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2093 | goto repeat; | |
2094 | } | |
2095 | } | |
2096 | out: | |
2097 | return err; | |
2098 | } | |
2099 | ||
c9de560d AT |
2100 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2101 | { | |
2102 | struct super_block *sb = seq->private; | |
c9de560d AT |
2103 | ext4_group_t group; |
2104 | ||
8df9675f | 2105 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2106 | return NULL; |
c9de560d | 2107 | group = *pos + 1; |
a9df9a49 | 2108 | return (void *) ((unsigned long) group); |
c9de560d AT |
2109 | } |
2110 | ||
2111 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2112 | { | |
2113 | struct super_block *sb = seq->private; | |
c9de560d AT |
2114 | ext4_group_t group; |
2115 | ||
2116 | ++*pos; | |
8df9675f | 2117 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2118 | return NULL; |
2119 | group = *pos + 1; | |
a9df9a49 | 2120 | return (void *) ((unsigned long) group); |
c9de560d AT |
2121 | } |
2122 | ||
2123 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2124 | { | |
2125 | struct super_block *sb = seq->private; | |
a9df9a49 | 2126 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d AT |
2127 | int i; |
2128 | int err; | |
2129 | struct ext4_buddy e4b; | |
2130 | struct sg { | |
2131 | struct ext4_group_info info; | |
a36b4498 | 2132 | ext4_grpblk_t counters[16]; |
c9de560d AT |
2133 | } sg; |
2134 | ||
2135 | group--; | |
2136 | if (group == 0) | |
2137 | seq_printf(seq, "#%-5s: %-5s %-5s %-5s " | |
2138 | "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " | |
2139 | "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", | |
2140 | "group", "free", "frags", "first", | |
2141 | "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", | |
2142 | "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); | |
2143 | ||
2144 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2145 | sizeof(struct ext4_group_info); | |
2146 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2147 | if (err) { | |
a9df9a49 | 2148 | seq_printf(seq, "#%-5u: I/O error\n", group); |
c9de560d AT |
2149 | return 0; |
2150 | } | |
2151 | ext4_lock_group(sb, group); | |
2152 | memcpy(&sg, ext4_get_group_info(sb, group), i); | |
2153 | ext4_unlock_group(sb, group); | |
e39e07fd | 2154 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 2155 | |
a9df9a49 | 2156 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2157 | sg.info.bb_fragments, sg.info.bb_first_free); |
2158 | for (i = 0; i <= 13; i++) | |
2159 | seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? | |
2160 | sg.info.bb_counters[i] : 0); | |
2161 | seq_printf(seq, " ]\n"); | |
2162 | ||
2163 | return 0; | |
2164 | } | |
2165 | ||
2166 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2167 | { | |
2168 | } | |
2169 | ||
7f1346a9 | 2170 | static const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2171 | .start = ext4_mb_seq_groups_start, |
2172 | .next = ext4_mb_seq_groups_next, | |
2173 | .stop = ext4_mb_seq_groups_stop, | |
2174 | .show = ext4_mb_seq_groups_show, | |
2175 | }; | |
2176 | ||
2177 | static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) | |
2178 | { | |
2179 | struct super_block *sb = PDE(inode)->data; | |
2180 | int rc; | |
2181 | ||
2182 | rc = seq_open(file, &ext4_mb_seq_groups_ops); | |
2183 | if (rc == 0) { | |
a271fe85 | 2184 | struct seq_file *m = file->private_data; |
c9de560d AT |
2185 | m->private = sb; |
2186 | } | |
2187 | return rc; | |
2188 | ||
2189 | } | |
2190 | ||
7f1346a9 | 2191 | static const struct file_operations ext4_mb_seq_groups_fops = { |
c9de560d AT |
2192 | .owner = THIS_MODULE, |
2193 | .open = ext4_mb_seq_groups_open, | |
2194 | .read = seq_read, | |
2195 | .llseek = seq_lseek, | |
2196 | .release = seq_release, | |
2197 | }; | |
2198 | ||
fb1813f4 CW |
2199 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
2200 | { | |
2201 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2202 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
2203 | ||
2204 | BUG_ON(!cachep); | |
2205 | return cachep; | |
2206 | } | |
5f21b0e6 FB |
2207 | |
2208 | /* Create and initialize ext4_group_info data for the given group. */ | |
920313a7 | 2209 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2210 | struct ext4_group_desc *desc) |
2211 | { | |
fb1813f4 | 2212 | int i; |
5f21b0e6 FB |
2213 | int metalen = 0; |
2214 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2215 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 2216 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
2217 | |
2218 | /* | |
2219 | * First check if this group is the first of a reserved block. | |
2220 | * If it's true, we have to allocate a new table of pointers | |
2221 | * to ext4_group_info structures | |
2222 | */ | |
2223 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2224 | metalen = sizeof(*meta_group_info) << | |
2225 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2226 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2227 | if (meta_group_info == NULL) { | |
9d8b9ec4 TT |
2228 | ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate mem " |
2229 | "for a buddy group"); | |
5f21b0e6 FB |
2230 | goto exit_meta_group_info; |
2231 | } | |
2232 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = | |
2233 | meta_group_info; | |
2234 | } | |
2235 | ||
5f21b0e6 FB |
2236 | meta_group_info = |
2237 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; | |
2238 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); | |
2239 | ||
fb1813f4 | 2240 | meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL); |
5f21b0e6 | 2241 | if (meta_group_info[i] == NULL) { |
9d8b9ec4 | 2242 | ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate buddy mem"); |
5f21b0e6 FB |
2243 | goto exit_group_info; |
2244 | } | |
fb1813f4 | 2245 | memset(meta_group_info[i], 0, kmem_cache_size(cachep)); |
5f21b0e6 FB |
2246 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, |
2247 | &(meta_group_info[i]->bb_state)); | |
2248 | ||
2249 | /* | |
2250 | * initialize bb_free to be able to skip | |
2251 | * empty groups without initialization | |
2252 | */ | |
2253 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2254 | meta_group_info[i]->bb_free = | |
2255 | ext4_free_blocks_after_init(sb, group, desc); | |
2256 | } else { | |
2257 | meta_group_info[i]->bb_free = | |
560671a0 | 2258 | ext4_free_blks_count(sb, desc); |
5f21b0e6 FB |
2259 | } |
2260 | ||
2261 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2262 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 2263 | meta_group_info[i]->bb_free_root = RB_ROOT; |
8a57d9d6 | 2264 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
5f21b0e6 FB |
2265 | |
2266 | #ifdef DOUBLE_CHECK | |
2267 | { | |
2268 | struct buffer_head *bh; | |
2269 | meta_group_info[i]->bb_bitmap = | |
2270 | kmalloc(sb->s_blocksize, GFP_KERNEL); | |
2271 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); | |
2272 | bh = ext4_read_block_bitmap(sb, group); | |
2273 | BUG_ON(bh == NULL); | |
2274 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, | |
2275 | sb->s_blocksize); | |
2276 | put_bh(bh); | |
2277 | } | |
2278 | #endif | |
2279 | ||
2280 | return 0; | |
2281 | ||
2282 | exit_group_info: | |
2283 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 2284 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
5f21b0e6 | 2285 | kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); |
caaf7a29 TM |
2286 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL; |
2287 | } | |
5f21b0e6 FB |
2288 | exit_meta_group_info: |
2289 | return -ENOMEM; | |
2290 | } /* ext4_mb_add_groupinfo */ | |
2291 | ||
c9de560d AT |
2292 | static int ext4_mb_init_backend(struct super_block *sb) |
2293 | { | |
8df9675f | 2294 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2295 | ext4_group_t i; |
c9de560d | 2296 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
5f21b0e6 FB |
2297 | struct ext4_super_block *es = sbi->s_es; |
2298 | int num_meta_group_infos; | |
2299 | int num_meta_group_infos_max; | |
2300 | int array_size; | |
5f21b0e6 | 2301 | struct ext4_group_desc *desc; |
fb1813f4 | 2302 | struct kmem_cache *cachep; |
5f21b0e6 FB |
2303 | |
2304 | /* This is the number of blocks used by GDT */ | |
8df9675f | 2305 | num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) - |
5f21b0e6 FB |
2306 | 1) >> EXT4_DESC_PER_BLOCK_BITS(sb); |
2307 | ||
2308 | /* | |
2309 | * This is the total number of blocks used by GDT including | |
2310 | * the number of reserved blocks for GDT. | |
2311 | * The s_group_info array is allocated with this value | |
2312 | * to allow a clean online resize without a complex | |
2313 | * manipulation of pointer. | |
2314 | * The drawback is the unused memory when no resize | |
2315 | * occurs but it's very low in terms of pages | |
2316 | * (see comments below) | |
2317 | * Need to handle this properly when META_BG resizing is allowed | |
2318 | */ | |
2319 | num_meta_group_infos_max = num_meta_group_infos + | |
2320 | le16_to_cpu(es->s_reserved_gdt_blocks); | |
c9de560d | 2321 | |
5f21b0e6 FB |
2322 | /* |
2323 | * array_size is the size of s_group_info array. We round it | |
2324 | * to the next power of two because this approximation is done | |
2325 | * internally by kmalloc so we can have some more memory | |
2326 | * for free here (e.g. may be used for META_BG resize). | |
2327 | */ | |
2328 | array_size = 1; | |
2329 | while (array_size < sizeof(*sbi->s_group_info) * | |
2330 | num_meta_group_infos_max) | |
2331 | array_size = array_size << 1; | |
c9de560d AT |
2332 | /* An 8TB filesystem with 64-bit pointers requires a 4096 byte |
2333 | * kmalloc. A 128kb malloc should suffice for a 256TB filesystem. | |
2334 | * So a two level scheme suffices for now. */ | |
f18a5f21 | 2335 | sbi->s_group_info = ext4_kvzalloc(array_size, GFP_KERNEL); |
c9de560d | 2336 | if (sbi->s_group_info == NULL) { |
9d8b9ec4 | 2337 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); |
c9de560d AT |
2338 | return -ENOMEM; |
2339 | } | |
2340 | sbi->s_buddy_cache = new_inode(sb); | |
2341 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 2342 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
2343 | goto err_freesgi; |
2344 | } | |
48e6061b YJ |
2345 | /* To avoid potentially colliding with an valid on-disk inode number, |
2346 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
2347 | * not in the inode hash, so it should never be found by iget(), but | |
2348 | * this will avoid confusion if it ever shows up during debugging. */ | |
2349 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 2350 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 2351 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2352 | desc = ext4_get_group_desc(sb, i, NULL); |
2353 | if (desc == NULL) { | |
9d8b9ec4 | 2354 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
2355 | goto err_freebuddy; |
2356 | } | |
5f21b0e6 FB |
2357 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2358 | goto err_freebuddy; | |
c9de560d AT |
2359 | } |
2360 | ||
2361 | return 0; | |
2362 | ||
2363 | err_freebuddy: | |
fb1813f4 | 2364 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 2365 | while (i-- > 0) |
fb1813f4 | 2366 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
c9de560d | 2367 | i = num_meta_group_infos; |
f1fa3342 | 2368 | while (i-- > 0) |
c9de560d AT |
2369 | kfree(sbi->s_group_info[i]); |
2370 | iput(sbi->s_buddy_cache); | |
2371 | err_freesgi: | |
f18a5f21 | 2372 | ext4_kvfree(sbi->s_group_info); |
c9de560d AT |
2373 | return -ENOMEM; |
2374 | } | |
2375 | ||
2892c15d ES |
2376 | static void ext4_groupinfo_destroy_slabs(void) |
2377 | { | |
2378 | int i; | |
2379 | ||
2380 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
2381 | if (ext4_groupinfo_caches[i]) | |
2382 | kmem_cache_destroy(ext4_groupinfo_caches[i]); | |
2383 | ext4_groupinfo_caches[i] = NULL; | |
2384 | } | |
2385 | } | |
2386 | ||
2387 | static int ext4_groupinfo_create_slab(size_t size) | |
2388 | { | |
2389 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
2390 | int slab_size; | |
2391 | int blocksize_bits = order_base_2(size); | |
2392 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2393 | struct kmem_cache *cachep; | |
2394 | ||
2395 | if (cache_index >= NR_GRPINFO_CACHES) | |
2396 | return -EINVAL; | |
2397 | ||
2398 | if (unlikely(cache_index < 0)) | |
2399 | cache_index = 0; | |
2400 | ||
2401 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
2402 | if (ext4_groupinfo_caches[cache_index]) { | |
2403 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
2404 | return 0; /* Already created */ | |
2405 | } | |
2406 | ||
2407 | slab_size = offsetof(struct ext4_group_info, | |
2408 | bb_counters[blocksize_bits + 2]); | |
2409 | ||
2410 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
2411 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
2412 | NULL); | |
2413 | ||
823ba01f TM |
2414 | ext4_groupinfo_caches[cache_index] = cachep; |
2415 | ||
2892c15d ES |
2416 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
2417 | if (!cachep) { | |
9d8b9ec4 TT |
2418 | printk(KERN_EMERG |
2419 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
2420 | return -ENOMEM; |
2421 | } | |
2422 | ||
2892c15d ES |
2423 | return 0; |
2424 | } | |
2425 | ||
c9de560d AT |
2426 | int ext4_mb_init(struct super_block *sb, int needs_recovery) |
2427 | { | |
2428 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2429 | unsigned i, j; |
c9de560d AT |
2430 | unsigned offset; |
2431 | unsigned max; | |
74767c5a | 2432 | int ret; |
c9de560d | 2433 | |
1927805e | 2434 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
2435 | |
2436 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2437 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
2438 | ret = -ENOMEM; |
2439 | goto out; | |
c9de560d | 2440 | } |
ff7ef329 | 2441 | |
1927805e | 2442 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
2443 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2444 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
2445 | ret = -ENOMEM; |
2446 | goto out; | |
2447 | } | |
2448 | ||
2892c15d ES |
2449 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
2450 | if (ret < 0) | |
2451 | goto out; | |
c9de560d AT |
2452 | |
2453 | /* order 0 is regular bitmap */ | |
2454 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2455 | sbi->s_mb_offsets[0] = 0; | |
2456 | ||
2457 | i = 1; | |
2458 | offset = 0; | |
2459 | max = sb->s_blocksize << 2; | |
2460 | do { | |
2461 | sbi->s_mb_offsets[i] = offset; | |
2462 | sbi->s_mb_maxs[i] = max; | |
2463 | offset += 1 << (sb->s_blocksize_bits - i); | |
2464 | max = max >> 1; | |
2465 | i++; | |
2466 | } while (i <= sb->s_blocksize_bits + 1); | |
2467 | ||
c9de560d | 2468 | spin_lock_init(&sbi->s_md_lock); |
c9de560d AT |
2469 | spin_lock_init(&sbi->s_bal_lock); |
2470 | ||
2471 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2472 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2473 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2474 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2475 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
c9de560d | 2476 | sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC; |
d7a1fee1 DE |
2477 | /* |
2478 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
2479 | * to the lowest multiple of s_stripe which is bigger than | |
2480 | * the s_mb_group_prealloc as determined above. We want | |
2481 | * the preallocation size to be an exact multiple of the | |
2482 | * RAID stripe size so that preallocations don't fragment | |
2483 | * the stripes. | |
2484 | */ | |
2485 | if (sbi->s_stripe > 1) { | |
2486 | sbi->s_mb_group_prealloc = roundup( | |
2487 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
2488 | } | |
c9de560d | 2489 | |
730c213c | 2490 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2491 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 CW |
2492 | ret = -ENOMEM; |
2493 | goto out; | |
c9de560d | 2494 | } |
730c213c | 2495 | for_each_possible_cpu(i) { |
c9de560d | 2496 | struct ext4_locality_group *lg; |
730c213c | 2497 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2498 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2499 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2500 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2501 | spin_lock_init(&lg->lg_prealloc_lock); |
2502 | } | |
2503 | ||
79a77c5a YJ |
2504 | /* init file for buddy data */ |
2505 | ret = ext4_mb_init_backend(sb); | |
2506 | if (ret != 0) { | |
2507 | goto out; | |
2508 | } | |
2509 | ||
296c355c TT |
2510 | if (sbi->s_proc) |
2511 | proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, | |
2512 | &ext4_mb_seq_groups_fops, sb); | |
c9de560d | 2513 | |
0390131b FM |
2514 | if (sbi->s_journal) |
2515 | sbi->s_journal->j_commit_callback = release_blocks_on_commit; | |
fb1813f4 CW |
2516 | out: |
2517 | if (ret) { | |
2518 | kfree(sbi->s_mb_offsets); | |
2519 | kfree(sbi->s_mb_maxs); | |
fb1813f4 CW |
2520 | } |
2521 | return ret; | |
c9de560d AT |
2522 | } |
2523 | ||
955ce5f5 | 2524 | /* need to called with the ext4 group lock held */ |
c9de560d AT |
2525 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
2526 | { | |
2527 | struct ext4_prealloc_space *pa; | |
2528 | struct list_head *cur, *tmp; | |
2529 | int count = 0; | |
2530 | ||
2531 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2532 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2533 | list_del(&pa->pa_group_list); | |
2534 | count++; | |
688f05a0 | 2535 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2536 | } |
2537 | if (count) | |
6ba495e9 | 2538 | mb_debug(1, "mballoc: %u PAs left\n", count); |
c9de560d AT |
2539 | |
2540 | } | |
2541 | ||
2542 | int ext4_mb_release(struct super_block *sb) | |
2543 | { | |
8df9675f | 2544 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2545 | ext4_group_t i; |
2546 | int num_meta_group_infos; | |
2547 | struct ext4_group_info *grinfo; | |
2548 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
fb1813f4 | 2549 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
c9de560d | 2550 | |
c9de560d | 2551 | if (sbi->s_group_info) { |
8df9675f | 2552 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2553 | grinfo = ext4_get_group_info(sb, i); |
2554 | #ifdef DOUBLE_CHECK | |
2555 | kfree(grinfo->bb_bitmap); | |
2556 | #endif | |
2557 | ext4_lock_group(sb, i); | |
2558 | ext4_mb_cleanup_pa(grinfo); | |
2559 | ext4_unlock_group(sb, i); | |
fb1813f4 | 2560 | kmem_cache_free(cachep, grinfo); |
c9de560d | 2561 | } |
8df9675f | 2562 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2563 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2564 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2565 | for (i = 0; i < num_meta_group_infos; i++) | |
2566 | kfree(sbi->s_group_info[i]); | |
f18a5f21 | 2567 | ext4_kvfree(sbi->s_group_info); |
c9de560d AT |
2568 | } |
2569 | kfree(sbi->s_mb_offsets); | |
2570 | kfree(sbi->s_mb_maxs); | |
2571 | if (sbi->s_buddy_cache) | |
2572 | iput(sbi->s_buddy_cache); | |
2573 | if (sbi->s_mb_stats) { | |
9d8b9ec4 TT |
2574 | ext4_msg(sb, KERN_INFO, |
2575 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
2576 | atomic_read(&sbi->s_bal_allocated), |
2577 | atomic_read(&sbi->s_bal_reqs), | |
2578 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 TT |
2579 | ext4_msg(sb, KERN_INFO, |
2580 | "mballoc: %u extents scanned, %u goal hits, " | |
2581 | "%u 2^N hits, %u breaks, %u lost", | |
c9de560d AT |
2582 | atomic_read(&sbi->s_bal_ex_scanned), |
2583 | atomic_read(&sbi->s_bal_goals), | |
2584 | atomic_read(&sbi->s_bal_2orders), | |
2585 | atomic_read(&sbi->s_bal_breaks), | |
2586 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 TT |
2587 | ext4_msg(sb, KERN_INFO, |
2588 | "mballoc: %lu generated and it took %Lu", | |
ced156e4 | 2589 | sbi->s_mb_buddies_generated, |
c9de560d | 2590 | sbi->s_mb_generation_time); |
9d8b9ec4 TT |
2591 | ext4_msg(sb, KERN_INFO, |
2592 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
2593 | atomic_read(&sbi->s_mb_preallocated), |
2594 | atomic_read(&sbi->s_mb_discarded)); | |
2595 | } | |
2596 | ||
730c213c | 2597 | free_percpu(sbi->s_locality_groups); |
296c355c TT |
2598 | if (sbi->s_proc) |
2599 | remove_proc_entry("mb_groups", sbi->s_proc); | |
c9de560d AT |
2600 | |
2601 | return 0; | |
2602 | } | |
2603 | ||
77ca6cdf | 2604 | static inline int ext4_issue_discard(struct super_block *sb, |
5c521830 JZ |
2605 | ext4_group_t block_group, ext4_grpblk_t block, int count) |
2606 | { | |
5c521830 JZ |
2607 | ext4_fsblk_t discard_block; |
2608 | ||
2609 | discard_block = block + ext4_group_first_block_no(sb, block_group); | |
2610 | trace_ext4_discard_blocks(sb, | |
2611 | (unsigned long long) discard_block, count); | |
93259636 | 2612 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); |
5c521830 JZ |
2613 | } |
2614 | ||
3e624fc7 TT |
2615 | /* |
2616 | * This function is called by the jbd2 layer once the commit has finished, | |
2617 | * so we know we can free the blocks that were released with that commit. | |
2618 | */ | |
2619 | static void release_blocks_on_commit(journal_t *journal, transaction_t *txn) | |
c9de560d | 2620 | { |
3e624fc7 | 2621 | struct super_block *sb = journal->j_private; |
c9de560d | 2622 | struct ext4_buddy e4b; |
c894058d | 2623 | struct ext4_group_info *db; |
d9f34504 | 2624 | int err, count = 0, count2 = 0; |
c894058d | 2625 | struct ext4_free_data *entry; |
3e624fc7 | 2626 | struct list_head *l, *ltmp; |
c9de560d | 2627 | |
3e624fc7 TT |
2628 | list_for_each_safe(l, ltmp, &txn->t_private_list) { |
2629 | entry = list_entry(l, struct ext4_free_data, list); | |
c9de560d | 2630 | |
6ba495e9 | 2631 | mb_debug(1, "gonna free %u blocks in group %u (0x%p):", |
3e624fc7 | 2632 | entry->count, entry->group, entry); |
c9de560d | 2633 | |
d9f34504 TT |
2634 | if (test_opt(sb, DISCARD)) |
2635 | ext4_issue_discard(sb, entry->group, | |
2636 | entry->start_blk, entry->count); | |
b90f6870 | 2637 | |
c894058d | 2638 | err = ext4_mb_load_buddy(sb, entry->group, &e4b); |
c9de560d AT |
2639 | /* we expect to find existing buddy because it's pinned */ |
2640 | BUG_ON(err != 0); | |
2641 | ||
c894058d | 2642 | db = e4b.bd_info; |
c9de560d | 2643 | /* there are blocks to put in buddy to make them really free */ |
c894058d | 2644 | count += entry->count; |
c9de560d | 2645 | count2++; |
c894058d AK |
2646 | ext4_lock_group(sb, entry->group); |
2647 | /* Take it out of per group rb tree */ | |
2648 | rb_erase(&entry->node, &(db->bb_free_root)); | |
2649 | mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count); | |
2650 | ||
3d56b8d2 TM |
2651 | /* |
2652 | * Clear the trimmed flag for the group so that the next | |
2653 | * ext4_trim_fs can trim it. | |
2654 | * If the volume is mounted with -o discard, online discard | |
2655 | * is supported and the free blocks will be trimmed online. | |
2656 | */ | |
2657 | if (!test_opt(sb, DISCARD)) | |
2658 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
2659 | ||
c894058d AK |
2660 | if (!db->bb_free_root.rb_node) { |
2661 | /* No more items in the per group rb tree | |
2662 | * balance refcounts from ext4_mb_free_metadata() | |
2663 | */ | |
2664 | page_cache_release(e4b.bd_buddy_page); | |
2665 | page_cache_release(e4b.bd_bitmap_page); | |
c9de560d | 2666 | } |
c894058d | 2667 | ext4_unlock_group(sb, entry->group); |
c894058d | 2668 | kmem_cache_free(ext4_free_ext_cachep, entry); |
e39e07fd | 2669 | ext4_mb_unload_buddy(&e4b); |
3e624fc7 | 2670 | } |
c9de560d | 2671 | |
6ba495e9 | 2672 | mb_debug(1, "freed %u blocks in %u structures\n", count, count2); |
c9de560d AT |
2673 | } |
2674 | ||
6ba495e9 TT |
2675 | #ifdef CONFIG_EXT4_DEBUG |
2676 | u8 mb_enable_debug __read_mostly; | |
2677 | ||
2678 | static struct dentry *debugfs_dir; | |
2679 | static struct dentry *debugfs_debug; | |
2680 | ||
2681 | static void __init ext4_create_debugfs_entry(void) | |
2682 | { | |
2683 | debugfs_dir = debugfs_create_dir("ext4", NULL); | |
2684 | if (debugfs_dir) | |
2685 | debugfs_debug = debugfs_create_u8("mballoc-debug", | |
2686 | S_IRUGO | S_IWUSR, | |
2687 | debugfs_dir, | |
2688 | &mb_enable_debug); | |
2689 | } | |
2690 | ||
2691 | static void ext4_remove_debugfs_entry(void) | |
2692 | { | |
2693 | debugfs_remove(debugfs_debug); | |
2694 | debugfs_remove(debugfs_dir); | |
2695 | } | |
2696 | ||
2697 | #else | |
2698 | ||
2699 | static void __init ext4_create_debugfs_entry(void) | |
2700 | { | |
2701 | } | |
2702 | ||
2703 | static void ext4_remove_debugfs_entry(void) | |
2704 | { | |
2705 | } | |
2706 | ||
2707 | #endif | |
2708 | ||
5dabfc78 | 2709 | int __init ext4_init_mballoc(void) |
c9de560d | 2710 | { |
16828088 TT |
2711 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
2712 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d AT |
2713 | if (ext4_pspace_cachep == NULL) |
2714 | return -ENOMEM; | |
2715 | ||
16828088 TT |
2716 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
2717 | SLAB_RECLAIM_ACCOUNT); | |
256bdb49 ES |
2718 | if (ext4_ac_cachep == NULL) { |
2719 | kmem_cache_destroy(ext4_pspace_cachep); | |
2720 | return -ENOMEM; | |
2721 | } | |
c894058d | 2722 | |
16828088 TT |
2723 | ext4_free_ext_cachep = KMEM_CACHE(ext4_free_data, |
2724 | SLAB_RECLAIM_ACCOUNT); | |
c894058d AK |
2725 | if (ext4_free_ext_cachep == NULL) { |
2726 | kmem_cache_destroy(ext4_pspace_cachep); | |
2727 | kmem_cache_destroy(ext4_ac_cachep); | |
2728 | return -ENOMEM; | |
2729 | } | |
6ba495e9 | 2730 | ext4_create_debugfs_entry(); |
c9de560d AT |
2731 | return 0; |
2732 | } | |
2733 | ||
5dabfc78 | 2734 | void ext4_exit_mballoc(void) |
c9de560d | 2735 | { |
60e6679e | 2736 | /* |
3e03f9ca JDB |
2737 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
2738 | * before destroying the slab cache. | |
2739 | */ | |
2740 | rcu_barrier(); | |
c9de560d | 2741 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 2742 | kmem_cache_destroy(ext4_ac_cachep); |
c894058d | 2743 | kmem_cache_destroy(ext4_free_ext_cachep); |
2892c15d | 2744 | ext4_groupinfo_destroy_slabs(); |
6ba495e9 | 2745 | ext4_remove_debugfs_entry(); |
c9de560d AT |
2746 | } |
2747 | ||
2748 | ||
2749 | /* | |
73b2c716 | 2750 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
2751 | * Returns 0 if success or error code |
2752 | */ | |
4ddfef7b ES |
2753 | static noinline_for_stack int |
2754 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 2755 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
2756 | { |
2757 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
2758 | struct ext4_group_desc *gdp; |
2759 | struct buffer_head *gdp_bh; | |
2760 | struct ext4_sb_info *sbi; | |
2761 | struct super_block *sb; | |
2762 | ext4_fsblk_t block; | |
519deca0 | 2763 | int err, len; |
c9de560d AT |
2764 | |
2765 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2766 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2767 | ||
2768 | sb = ac->ac_sb; | |
2769 | sbi = EXT4_SB(sb); | |
c9de560d AT |
2770 | |
2771 | err = -EIO; | |
574ca174 | 2772 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
c9de560d AT |
2773 | if (!bitmap_bh) |
2774 | goto out_err; | |
2775 | ||
2776 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2777 | if (err) | |
2778 | goto out_err; | |
2779 | ||
2780 | err = -EIO; | |
2781 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
2782 | if (!gdp) | |
2783 | goto out_err; | |
2784 | ||
a9df9a49 | 2785 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
9fd9784c | 2786 | ext4_free_blks_count(sb, gdp)); |
03cddb80 | 2787 | |
c9de560d AT |
2788 | err = ext4_journal_get_write_access(handle, gdp_bh); |
2789 | if (err) | |
2790 | goto out_err; | |
2791 | ||
bda00de7 | 2792 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 2793 | |
53accfa9 | 2794 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
6fd058f7 | 2795 | if (!ext4_data_block_valid(sbi, block, len)) { |
12062ddd | 2796 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
6fd058f7 | 2797 | "fs metadata\n", block, block+len); |
519deca0 AK |
2798 | /* File system mounted not to panic on error |
2799 | * Fix the bitmap and repeat the block allocation | |
2800 | * We leak some of the blocks here. | |
2801 | */ | |
955ce5f5 | 2802 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
c3e94d1d YY |
2803 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
2804 | ac->ac_b_ex.fe_len); | |
955ce5f5 | 2805 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 2806 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 AK |
2807 | if (!err) |
2808 | err = -EAGAIN; | |
2809 | goto out_err; | |
c9de560d | 2810 | } |
955ce5f5 AK |
2811 | |
2812 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
2813 | #ifdef AGGRESSIVE_CHECK |
2814 | { | |
2815 | int i; | |
2816 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
2817 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
2818 | bitmap_bh->b_data)); | |
2819 | } | |
2820 | } | |
2821 | #endif | |
c3e94d1d YY |
2822 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
2823 | ac->ac_b_ex.fe_len); | |
c9de560d AT |
2824 | if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
2825 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
560671a0 AK |
2826 | ext4_free_blks_set(sb, gdp, |
2827 | ext4_free_blocks_after_init(sb, | |
2828 | ac->ac_b_ex.fe_group, gdp)); | |
c9de560d | 2829 | } |
560671a0 AK |
2830 | len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len; |
2831 | ext4_free_blks_set(sb, gdp, len); | |
c9de560d | 2832 | gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
2833 | |
2834 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
6bc6e63f | 2835 | percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 2836 | /* |
6bc6e63f | 2837 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 2838 | */ |
6bc6e63f AK |
2839 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
2840 | /* release all the reserved blocks if non delalloc */ | |
53accfa9 | 2841 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_clstrs); |
c9de560d | 2842 | |
772cb7c8 JS |
2843 | if (sbi->s_log_groups_per_flex) { |
2844 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
2845 | ac->ac_b_ex.fe_group); | |
9f24e420 TT |
2846 | atomic_sub(ac->ac_b_ex.fe_len, |
2847 | &sbi->s_flex_groups[flex_group].free_blocks); | |
772cb7c8 JS |
2848 | } |
2849 | ||
0390131b | 2850 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
2851 | if (err) |
2852 | goto out_err; | |
0390131b | 2853 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
2854 | |
2855 | out_err: | |
a0375156 | 2856 | ext4_mark_super_dirty(sb); |
42a10add | 2857 | brelse(bitmap_bh); |
c9de560d AT |
2858 | return err; |
2859 | } | |
2860 | ||
2861 | /* | |
2862 | * here we normalize request for locality group | |
d7a1fee1 DE |
2863 | * Group request are normalized to s_mb_group_prealloc, which goes to |
2864 | * s_strip if we set the same via mount option. | |
2865 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 2866 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
2867 | * |
2868 | * XXX: should we try to preallocate more than the group has now? | |
2869 | */ | |
2870 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
2871 | { | |
2872 | struct super_block *sb = ac->ac_sb; | |
2873 | struct ext4_locality_group *lg = ac->ac_lg; | |
2874 | ||
2875 | BUG_ON(lg == NULL); | |
d7a1fee1 | 2876 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
6ba495e9 | 2877 | mb_debug(1, "#%u: goal %u blocks for locality group\n", |
c9de560d AT |
2878 | current->pid, ac->ac_g_ex.fe_len); |
2879 | } | |
2880 | ||
2881 | /* | |
2882 | * Normalization means making request better in terms of | |
2883 | * size and alignment | |
2884 | */ | |
4ddfef7b ES |
2885 | static noinline_for_stack void |
2886 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
2887 | struct ext4_allocation_request *ar) |
2888 | { | |
53accfa9 | 2889 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
2890 | int bsbits, max; |
2891 | ext4_lblk_t end; | |
c9de560d | 2892 | loff_t size, orig_size, start_off; |
5a0790c2 | 2893 | ext4_lblk_t start; |
c9de560d | 2894 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 2895 | struct ext4_prealloc_space *pa; |
c9de560d AT |
2896 | |
2897 | /* do normalize only data requests, metadata requests | |
2898 | do not need preallocation */ | |
2899 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
2900 | return; | |
2901 | ||
2902 | /* sometime caller may want exact blocks */ | |
2903 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2904 | return; | |
2905 | ||
2906 | /* caller may indicate that preallocation isn't | |
2907 | * required (it's a tail, for example) */ | |
2908 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
2909 | return; | |
2910 | ||
2911 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
2912 | ext4_mb_normalize_group_request(ac); | |
2913 | return ; | |
2914 | } | |
2915 | ||
2916 | bsbits = ac->ac_sb->s_blocksize_bits; | |
2917 | ||
2918 | /* first, let's learn actual file size | |
2919 | * given current request is allocated */ | |
53accfa9 | 2920 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
2921 | size = size << bsbits; |
2922 | if (size < i_size_read(ac->ac_inode)) | |
2923 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 2924 | orig_size = size; |
c9de560d | 2925 | |
1930479c VC |
2926 | /* max size of free chunks */ |
2927 | max = 2 << bsbits; | |
c9de560d | 2928 | |
1930479c VC |
2929 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
2930 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
2931 | |
2932 | /* first, try to predict filesize */ | |
2933 | /* XXX: should this table be tunable? */ | |
2934 | start_off = 0; | |
2935 | if (size <= 16 * 1024) { | |
2936 | size = 16 * 1024; | |
2937 | } else if (size <= 32 * 1024) { | |
2938 | size = 32 * 1024; | |
2939 | } else if (size <= 64 * 1024) { | |
2940 | size = 64 * 1024; | |
2941 | } else if (size <= 128 * 1024) { | |
2942 | size = 128 * 1024; | |
2943 | } else if (size <= 256 * 1024) { | |
2944 | size = 256 * 1024; | |
2945 | } else if (size <= 512 * 1024) { | |
2946 | size = 512 * 1024; | |
2947 | } else if (size <= 1024 * 1024) { | |
2948 | size = 1024 * 1024; | |
1930479c | 2949 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 2950 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
2951 | (21 - bsbits)) << 21; |
2952 | size = 2 * 1024 * 1024; | |
2953 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
2954 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
2955 | (22 - bsbits)) << 22; | |
2956 | size = 4 * 1024 * 1024; | |
2957 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 2958 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
2959 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
2960 | (23 - bsbits)) << 23; | |
2961 | size = 8 * 1024 * 1024; | |
2962 | } else { | |
2963 | start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; | |
2964 | size = ac->ac_o_ex.fe_len << bsbits; | |
2965 | } | |
5a0790c2 AK |
2966 | size = size >> bsbits; |
2967 | start = start_off >> bsbits; | |
c9de560d AT |
2968 | |
2969 | /* don't cover already allocated blocks in selected range */ | |
2970 | if (ar->pleft && start <= ar->lleft) { | |
2971 | size -= ar->lleft + 1 - start; | |
2972 | start = ar->lleft + 1; | |
2973 | } | |
2974 | if (ar->pright && start + size - 1 >= ar->lright) | |
2975 | size -= start + size - ar->lright; | |
2976 | ||
2977 | end = start + size; | |
2978 | ||
2979 | /* check we don't cross already preallocated blocks */ | |
2980 | rcu_read_lock(); | |
9a0762c5 | 2981 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 2982 | ext4_lblk_t pa_end; |
c9de560d | 2983 | |
c9de560d AT |
2984 | if (pa->pa_deleted) |
2985 | continue; | |
2986 | spin_lock(&pa->pa_lock); | |
2987 | if (pa->pa_deleted) { | |
2988 | spin_unlock(&pa->pa_lock); | |
2989 | continue; | |
2990 | } | |
2991 | ||
53accfa9 TT |
2992 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
2993 | pa->pa_len); | |
c9de560d AT |
2994 | |
2995 | /* PA must not overlap original request */ | |
2996 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
2997 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
2998 | ||
38877f4e ES |
2999 | /* skip PAs this normalized request doesn't overlap with */ |
3000 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
3001 | spin_unlock(&pa->pa_lock); |
3002 | continue; | |
3003 | } | |
3004 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3005 | ||
38877f4e | 3006 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
3007 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
3008 | BUG_ON(pa_end < start); | |
3009 | start = pa_end; | |
38877f4e | 3010 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
3011 | BUG_ON(pa->pa_lstart > end); |
3012 | end = pa->pa_lstart; | |
3013 | } | |
3014 | spin_unlock(&pa->pa_lock); | |
3015 | } | |
3016 | rcu_read_unlock(); | |
3017 | size = end - start; | |
3018 | ||
3019 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3020 | rcu_read_lock(); | |
9a0762c5 | 3021 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3022 | ext4_lblk_t pa_end; |
53accfa9 | 3023 | |
c9de560d AT |
3024 | spin_lock(&pa->pa_lock); |
3025 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
3026 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3027 | pa->pa_len); | |
c9de560d AT |
3028 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
3029 | } | |
3030 | spin_unlock(&pa->pa_lock); | |
3031 | } | |
3032 | rcu_read_unlock(); | |
3033 | ||
3034 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3035 | start > ac->ac_o_ex.fe_logical) { | |
9d8b9ec4 TT |
3036 | ext4_msg(ac->ac_sb, KERN_ERR, |
3037 | "start %lu, size %lu, fe_logical %lu", | |
3038 | (unsigned long) start, (unsigned long) size, | |
3039 | (unsigned long) ac->ac_o_ex.fe_logical); | |
c9de560d AT |
3040 | } |
3041 | BUG_ON(start + size <= ac->ac_o_ex.fe_logical && | |
3042 | start > ac->ac_o_ex.fe_logical); | |
7137d7a4 | 3043 | BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
3044 | |
3045 | /* now prepare goal request */ | |
3046 | ||
3047 | /* XXX: is it better to align blocks WRT to logical | |
3048 | * placement or satisfy big request as is */ | |
3049 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 3050 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
3051 | |
3052 | /* define goal start in order to merge */ | |
3053 | if (ar->pright && (ar->lright == (start + size))) { | |
3054 | /* merge to the right */ | |
3055 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3056 | &ac->ac_f_ex.fe_group, | |
3057 | &ac->ac_f_ex.fe_start); | |
3058 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3059 | } | |
3060 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3061 | /* merge to the left */ | |
3062 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3063 | &ac->ac_f_ex.fe_group, | |
3064 | &ac->ac_f_ex.fe_start); | |
3065 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3066 | } | |
3067 | ||
6ba495e9 | 3068 | mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, |
c9de560d AT |
3069 | (unsigned) orig_size, (unsigned) start); |
3070 | } | |
3071 | ||
3072 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3073 | { | |
3074 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3075 | ||
3076 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3077 | atomic_inc(&sbi->s_bal_reqs); | |
3078 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 3079 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
3080 | atomic_inc(&sbi->s_bal_success); |
3081 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3082 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3083 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3084 | atomic_inc(&sbi->s_bal_goals); | |
3085 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3086 | atomic_inc(&sbi->s_bal_breaks); | |
3087 | } | |
3088 | ||
296c355c TT |
3089 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
3090 | trace_ext4_mballoc_alloc(ac); | |
3091 | else | |
3092 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
3093 | } |
3094 | ||
b844167e CW |
3095 | /* |
3096 | * Called on failure; free up any blocks from the inode PA for this | |
3097 | * context. We don't need this for MB_GROUP_PA because we only change | |
3098 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
3099 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
3100 | */ | |
3101 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
3102 | { | |
3103 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
3104 | int len; | |
3105 | ||
3106 | if (pa && pa->pa_type == MB_INODE_PA) { | |
3107 | len = ac->ac_b_ex.fe_len; | |
3108 | pa->pa_free += len; | |
3109 | } | |
3110 | ||
3111 | } | |
3112 | ||
c9de560d AT |
3113 | /* |
3114 | * use blocks preallocated to inode | |
3115 | */ | |
3116 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3117 | struct ext4_prealloc_space *pa) | |
3118 | { | |
53accfa9 | 3119 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3120 | ext4_fsblk_t start; |
3121 | ext4_fsblk_t end; | |
3122 | int len; | |
3123 | ||
3124 | /* found preallocated blocks, use them */ | |
3125 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
3126 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
3127 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
3128 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
3129 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
3130 | &ac->ac_b_ex.fe_start); | |
3131 | ac->ac_b_ex.fe_len = len; | |
3132 | ac->ac_status = AC_STATUS_FOUND; | |
3133 | ac->ac_pa = pa; | |
3134 | ||
3135 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 3136 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
3137 | BUG_ON(pa->pa_free < len); |
3138 | pa->pa_free -= len; | |
3139 | ||
6ba495e9 | 3140 | mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); |
c9de560d AT |
3141 | } |
3142 | ||
3143 | /* | |
3144 | * use blocks preallocated to locality group | |
3145 | */ | |
3146 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3147 | struct ext4_prealloc_space *pa) | |
3148 | { | |
03cddb80 | 3149 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3150 | |
c9de560d AT |
3151 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3152 | &ac->ac_b_ex.fe_group, | |
3153 | &ac->ac_b_ex.fe_start); | |
3154 | ac->ac_b_ex.fe_len = len; | |
3155 | ac->ac_status = AC_STATUS_FOUND; | |
3156 | ac->ac_pa = pa; | |
3157 | ||
3158 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3159 | * possible race when the group is being loaded concurrently |
c9de560d | 3160 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3161 | * in on-disk bitmap -- see ext4_mb_release_context() |
3162 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 3163 | */ |
6ba495e9 | 3164 | mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); |
c9de560d AT |
3165 | } |
3166 | ||
5e745b04 AK |
3167 | /* |
3168 | * Return the prealloc space that have minimal distance | |
3169 | * from the goal block. @cpa is the prealloc | |
3170 | * space that is having currently known minimal distance | |
3171 | * from the goal block. | |
3172 | */ | |
3173 | static struct ext4_prealloc_space * | |
3174 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3175 | struct ext4_prealloc_space *pa, | |
3176 | struct ext4_prealloc_space *cpa) | |
3177 | { | |
3178 | ext4_fsblk_t cur_distance, new_distance; | |
3179 | ||
3180 | if (cpa == NULL) { | |
3181 | atomic_inc(&pa->pa_count); | |
3182 | return pa; | |
3183 | } | |
3184 | cur_distance = abs(goal_block - cpa->pa_pstart); | |
3185 | new_distance = abs(goal_block - pa->pa_pstart); | |
3186 | ||
5a54b2f1 | 3187 | if (cur_distance <= new_distance) |
5e745b04 AK |
3188 | return cpa; |
3189 | ||
3190 | /* drop the previous reference */ | |
3191 | atomic_dec(&cpa->pa_count); | |
3192 | atomic_inc(&pa->pa_count); | |
3193 | return pa; | |
3194 | } | |
3195 | ||
c9de560d AT |
3196 | /* |
3197 | * search goal blocks in preallocated space | |
3198 | */ | |
4ddfef7b ES |
3199 | static noinline_for_stack int |
3200 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
c9de560d | 3201 | { |
53accfa9 | 3202 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 3203 | int order, i; |
c9de560d AT |
3204 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3205 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3206 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3207 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3208 | |
3209 | /* only data can be preallocated */ | |
3210 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3211 | return 0; | |
3212 | ||
3213 | /* first, try per-file preallocation */ | |
3214 | rcu_read_lock(); | |
9a0762c5 | 3215 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3216 | |
3217 | /* all fields in this condition don't change, | |
3218 | * so we can skip locking for them */ | |
3219 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
3220 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
3221 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
3222 | continue; |
3223 | ||
fb0a387d | 3224 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 3225 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
3226 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
3227 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
3228 | continue; |
3229 | ||
c9de560d AT |
3230 | /* found preallocated blocks, use them */ |
3231 | spin_lock(&pa->pa_lock); | |
3232 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3233 | atomic_inc(&pa->pa_count); | |
3234 | ext4_mb_use_inode_pa(ac, pa); | |
3235 | spin_unlock(&pa->pa_lock); | |
3236 | ac->ac_criteria = 10; | |
3237 | rcu_read_unlock(); | |
3238 | return 1; | |
3239 | } | |
3240 | spin_unlock(&pa->pa_lock); | |
3241 | } | |
3242 | rcu_read_unlock(); | |
3243 | ||
3244 | /* can we use group allocation? */ | |
3245 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3246 | return 0; | |
3247 | ||
3248 | /* inode may have no locality group for some reason */ | |
3249 | lg = ac->ac_lg; | |
3250 | if (lg == NULL) | |
3251 | return 0; | |
6be2ded1 AK |
3252 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3253 | if (order > PREALLOC_TB_SIZE - 1) | |
3254 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3255 | order = PREALLOC_TB_SIZE - 1; | |
3256 | ||
bda00de7 | 3257 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
3258 | /* |
3259 | * search for the prealloc space that is having | |
3260 | * minimal distance from the goal block. | |
3261 | */ | |
6be2ded1 AK |
3262 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3263 | rcu_read_lock(); | |
3264 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3265 | pa_inode_list) { | |
3266 | spin_lock(&pa->pa_lock); | |
3267 | if (pa->pa_deleted == 0 && | |
3268 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3269 | |
3270 | cpa = ext4_mb_check_group_pa(goal_block, | |
3271 | pa, cpa); | |
6be2ded1 | 3272 | } |
c9de560d | 3273 | spin_unlock(&pa->pa_lock); |
c9de560d | 3274 | } |
6be2ded1 | 3275 | rcu_read_unlock(); |
c9de560d | 3276 | } |
5e745b04 AK |
3277 | if (cpa) { |
3278 | ext4_mb_use_group_pa(ac, cpa); | |
3279 | ac->ac_criteria = 20; | |
3280 | return 1; | |
3281 | } | |
c9de560d AT |
3282 | return 0; |
3283 | } | |
3284 | ||
7a2fcbf7 AK |
3285 | /* |
3286 | * the function goes through all block freed in the group | |
3287 | * but not yet committed and marks them used in in-core bitmap. | |
3288 | * buddy must be generated from this bitmap | |
955ce5f5 | 3289 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3290 | */ |
3291 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3292 | ext4_group_t group) | |
3293 | { | |
3294 | struct rb_node *n; | |
3295 | struct ext4_group_info *grp; | |
3296 | struct ext4_free_data *entry; | |
3297 | ||
3298 | grp = ext4_get_group_info(sb, group); | |
3299 | n = rb_first(&(grp->bb_free_root)); | |
3300 | ||
3301 | while (n) { | |
3302 | entry = rb_entry(n, struct ext4_free_data, node); | |
c3e94d1d | 3303 | ext4_set_bits(bitmap, entry->start_blk, entry->count); |
7a2fcbf7 AK |
3304 | n = rb_next(n); |
3305 | } | |
3306 | return; | |
3307 | } | |
3308 | ||
c9de560d AT |
3309 | /* |
3310 | * the function goes through all preallocation in this group and marks them | |
3311 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3312 | * Need to be called with ext4 group lock held |
c9de560d | 3313 | */ |
089ceecc ES |
3314 | static noinline_for_stack |
3315 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
3316 | ext4_group_t group) |
3317 | { | |
3318 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3319 | struct ext4_prealloc_space *pa; | |
3320 | struct list_head *cur; | |
3321 | ext4_group_t groupnr; | |
3322 | ext4_grpblk_t start; | |
3323 | int preallocated = 0; | |
3324 | int count = 0; | |
3325 | int len; | |
3326 | ||
3327 | /* all form of preallocation discards first load group, | |
3328 | * so the only competing code is preallocation use. | |
3329 | * we don't need any locking here | |
3330 | * notice we do NOT ignore preallocations with pa_deleted | |
3331 | * otherwise we could leave used blocks available for | |
3332 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3333 | * is dropping preallocation | |
3334 | */ | |
3335 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3336 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3337 | spin_lock(&pa->pa_lock); | |
3338 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3339 | &groupnr, &start); | |
3340 | len = pa->pa_len; | |
3341 | spin_unlock(&pa->pa_lock); | |
3342 | if (unlikely(len == 0)) | |
3343 | continue; | |
3344 | BUG_ON(groupnr != group); | |
c3e94d1d | 3345 | ext4_set_bits(bitmap, start, len); |
c9de560d AT |
3346 | preallocated += len; |
3347 | count++; | |
3348 | } | |
6ba495e9 | 3349 | mb_debug(1, "prellocated %u for group %u\n", preallocated, group); |
c9de560d AT |
3350 | } |
3351 | ||
3352 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3353 | { | |
3354 | struct ext4_prealloc_space *pa; | |
3355 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
3356 | kmem_cache_free(ext4_pspace_cachep, pa); | |
3357 | } | |
3358 | ||
3359 | /* | |
3360 | * drops a reference to preallocated space descriptor | |
3361 | * if this was the last reference and the space is consumed | |
3362 | */ | |
3363 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3364 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3365 | { | |
a9df9a49 | 3366 | ext4_group_t grp; |
d33a1976 | 3367 | ext4_fsblk_t grp_blk; |
c9de560d AT |
3368 | |
3369 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) | |
3370 | return; | |
3371 | ||
3372 | /* in this short window concurrent discard can set pa_deleted */ | |
3373 | spin_lock(&pa->pa_lock); | |
3374 | if (pa->pa_deleted == 1) { | |
3375 | spin_unlock(&pa->pa_lock); | |
3376 | return; | |
3377 | } | |
3378 | ||
3379 | pa->pa_deleted = 1; | |
3380 | spin_unlock(&pa->pa_lock); | |
3381 | ||
d33a1976 | 3382 | grp_blk = pa->pa_pstart; |
60e6679e | 3383 | /* |
cc0fb9ad AK |
3384 | * If doing group-based preallocation, pa_pstart may be in the |
3385 | * next group when pa is used up | |
3386 | */ | |
3387 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3388 | grp_blk--; |
3389 | ||
3390 | ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL); | |
c9de560d AT |
3391 | |
3392 | /* | |
3393 | * possible race: | |
3394 | * | |
3395 | * P1 (buddy init) P2 (regular allocation) | |
3396 | * find block B in PA | |
3397 | * copy on-disk bitmap to buddy | |
3398 | * mark B in on-disk bitmap | |
3399 | * drop PA from group | |
3400 | * mark all PAs in buddy | |
3401 | * | |
3402 | * thus, P1 initializes buddy with B available. to prevent this | |
3403 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3404 | * against that pair | |
3405 | */ | |
3406 | ext4_lock_group(sb, grp); | |
3407 | list_del(&pa->pa_group_list); | |
3408 | ext4_unlock_group(sb, grp); | |
3409 | ||
3410 | spin_lock(pa->pa_obj_lock); | |
3411 | list_del_rcu(&pa->pa_inode_list); | |
3412 | spin_unlock(pa->pa_obj_lock); | |
3413 | ||
3414 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3415 | } | |
3416 | ||
3417 | /* | |
3418 | * creates new preallocated space for given inode | |
3419 | */ | |
4ddfef7b ES |
3420 | static noinline_for_stack int |
3421 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3422 | { |
3423 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 3424 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
3425 | struct ext4_prealloc_space *pa; |
3426 | struct ext4_group_info *grp; | |
3427 | struct ext4_inode_info *ei; | |
3428 | ||
3429 | /* preallocate only when found space is larger then requested */ | |
3430 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3431 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3432 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3433 | ||
3434 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3435 | if (pa == NULL) | |
3436 | return -ENOMEM; | |
3437 | ||
3438 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3439 | int winl; | |
3440 | int wins; | |
3441 | int win; | |
3442 | int offs; | |
3443 | ||
3444 | /* we can't allocate as much as normalizer wants. | |
3445 | * so, found space must get proper lstart | |
3446 | * to cover original request */ | |
3447 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3448 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3449 | ||
3450 | /* we're limited by original request in that | |
3451 | * logical block must be covered any way | |
3452 | * winl is window we can move our chunk within */ | |
3453 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3454 | ||
3455 | /* also, we should cover whole original request */ | |
53accfa9 | 3456 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
3457 | |
3458 | /* the smallest one defines real window */ | |
3459 | win = min(winl, wins); | |
3460 | ||
53accfa9 TT |
3461 | offs = ac->ac_o_ex.fe_logical % |
3462 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
3463 | if (offs && offs < win) |
3464 | win = offs; | |
3465 | ||
53accfa9 TT |
3466 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
3467 | EXT4_B2C(sbi, win); | |
c9de560d AT |
3468 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
3469 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3470 | } | |
3471 | ||
3472 | /* preallocation can change ac_b_ex, thus we store actually | |
3473 | * allocated blocks for history */ | |
3474 | ac->ac_f_ex = ac->ac_b_ex; | |
3475 | ||
3476 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3477 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3478 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3479 | pa->pa_free = pa->pa_len; | |
3480 | atomic_set(&pa->pa_count, 1); | |
3481 | spin_lock_init(&pa->pa_lock); | |
d794bf8e AK |
3482 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3483 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3484 | pa->pa_deleted = 0; |
cc0fb9ad | 3485 | pa->pa_type = MB_INODE_PA; |
c9de560d | 3486 | |
6ba495e9 | 3487 | mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, |
c9de560d | 3488 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
9bffad1e | 3489 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3490 | |
3491 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 3492 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
3493 | |
3494 | ei = EXT4_I(ac->ac_inode); | |
3495 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3496 | ||
3497 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3498 | pa->pa_inode = ac->ac_inode; | |
3499 | ||
3500 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3501 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3502 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3503 | ||
3504 | spin_lock(pa->pa_obj_lock); | |
3505 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3506 | spin_unlock(pa->pa_obj_lock); | |
3507 | ||
3508 | return 0; | |
3509 | } | |
3510 | ||
3511 | /* | |
3512 | * creates new preallocated space for locality group inodes belongs to | |
3513 | */ | |
4ddfef7b ES |
3514 | static noinline_for_stack int |
3515 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3516 | { |
3517 | struct super_block *sb = ac->ac_sb; | |
3518 | struct ext4_locality_group *lg; | |
3519 | struct ext4_prealloc_space *pa; | |
3520 | struct ext4_group_info *grp; | |
3521 | ||
3522 | /* preallocate only when found space is larger then requested */ | |
3523 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3524 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3525 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3526 | ||
3527 | BUG_ON(ext4_pspace_cachep == NULL); | |
3528 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3529 | if (pa == NULL) | |
3530 | return -ENOMEM; | |
3531 | ||
3532 | /* preallocation can change ac_b_ex, thus we store actually | |
3533 | * allocated blocks for history */ | |
3534 | ac->ac_f_ex = ac->ac_b_ex; | |
3535 | ||
3536 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3537 | pa->pa_lstart = pa->pa_pstart; | |
3538 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3539 | pa->pa_free = pa->pa_len; | |
3540 | atomic_set(&pa->pa_count, 1); | |
3541 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3542 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 3543 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 3544 | pa->pa_deleted = 0; |
cc0fb9ad | 3545 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 3546 | |
6ba495e9 | 3547 | mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, |
9bffad1e TT |
3548 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
3549 | trace_ext4_mb_new_group_pa(ac, pa); | |
c9de560d AT |
3550 | |
3551 | ext4_mb_use_group_pa(ac, pa); | |
3552 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3553 | ||
3554 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3555 | lg = ac->ac_lg; | |
3556 | BUG_ON(lg == NULL); | |
3557 | ||
3558 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3559 | pa->pa_inode = NULL; | |
3560 | ||
3561 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3562 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3563 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3564 | ||
6be2ded1 AK |
3565 | /* |
3566 | * We will later add the new pa to the right bucket | |
3567 | * after updating the pa_free in ext4_mb_release_context | |
3568 | */ | |
c9de560d AT |
3569 | return 0; |
3570 | } | |
3571 | ||
3572 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3573 | { | |
3574 | int err; | |
3575 | ||
3576 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3577 | err = ext4_mb_new_group_pa(ac); | |
3578 | else | |
3579 | err = ext4_mb_new_inode_pa(ac); | |
3580 | return err; | |
3581 | } | |
3582 | ||
3583 | /* | |
3584 | * finds all unused blocks in on-disk bitmap, frees them in | |
3585 | * in-core bitmap and buddy. | |
3586 | * @pa must be unlinked from inode and group lists, so that | |
3587 | * nobody else can find/use it. | |
3588 | * the caller MUST hold group/inode locks. | |
3589 | * TODO: optimize the case when there are no in-core structures yet | |
3590 | */ | |
4ddfef7b ES |
3591 | static noinline_for_stack int |
3592 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 3593 | struct ext4_prealloc_space *pa) |
c9de560d | 3594 | { |
c9de560d AT |
3595 | struct super_block *sb = e4b->bd_sb; |
3596 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
3597 | unsigned int end; |
3598 | unsigned int next; | |
c9de560d AT |
3599 | ext4_group_t group; |
3600 | ext4_grpblk_t bit; | |
ba80b101 | 3601 | unsigned long long grp_blk_start; |
c9de560d AT |
3602 | int err = 0; |
3603 | int free = 0; | |
3604 | ||
3605 | BUG_ON(pa->pa_deleted == 0); | |
3606 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 3607 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
3608 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
3609 | end = bit + pa->pa_len; | |
3610 | ||
c9de560d | 3611 | while (bit < end) { |
ffad0a44 | 3612 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3613 | if (bit >= end) |
3614 | break; | |
ffad0a44 | 3615 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
6ba495e9 | 3616 | mb_debug(1, " free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
3617 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
3618 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
3619 | free += next - bit; |
3620 | ||
3e1e5f50 | 3621 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
3622 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
3623 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 3624 | next - bit); |
c9de560d AT |
3625 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
3626 | bit = next + 1; | |
3627 | } | |
3628 | if (free != pa->pa_free) { | |
9d8b9ec4 TT |
3629 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
3630 | "pa %p: logic %lu, phys. %lu, len %lu", | |
3631 | pa, (unsigned long) pa->pa_lstart, | |
3632 | (unsigned long) pa->pa_pstart, | |
3633 | (unsigned long) pa->pa_len); | |
e29136f8 | 3634 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 3635 | free, pa->pa_free); |
e56eb659 AK |
3636 | /* |
3637 | * pa is already deleted so we use the value obtained | |
3638 | * from the bitmap and continue. | |
3639 | */ | |
c9de560d | 3640 | } |
c9de560d AT |
3641 | atomic_add(free, &sbi->s_mb_discarded); |
3642 | ||
3643 | return err; | |
3644 | } | |
3645 | ||
4ddfef7b ES |
3646 | static noinline_for_stack int |
3647 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 3648 | struct ext4_prealloc_space *pa) |
c9de560d | 3649 | { |
c9de560d AT |
3650 | struct super_block *sb = e4b->bd_sb; |
3651 | ext4_group_t group; | |
3652 | ext4_grpblk_t bit; | |
3653 | ||
a9c667f8 | 3654 | trace_ext4_mb_release_group_pa(pa); |
c9de560d AT |
3655 | BUG_ON(pa->pa_deleted == 0); |
3656 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3657 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3658 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3659 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 3660 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
3661 | |
3662 | return 0; | |
3663 | } | |
3664 | ||
3665 | /* | |
3666 | * releases all preallocations in given group | |
3667 | * | |
3668 | * first, we need to decide discard policy: | |
3669 | * - when do we discard | |
3670 | * 1) ENOSPC | |
3671 | * - how many do we discard | |
3672 | * 1) how many requested | |
3673 | */ | |
4ddfef7b ES |
3674 | static noinline_for_stack int |
3675 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3676 | ext4_group_t group, int needed) |
3677 | { | |
3678 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3679 | struct buffer_head *bitmap_bh = NULL; | |
3680 | struct ext4_prealloc_space *pa, *tmp; | |
3681 | struct list_head list; | |
3682 | struct ext4_buddy e4b; | |
3683 | int err; | |
3684 | int busy = 0; | |
3685 | int free = 0; | |
3686 | ||
6ba495e9 | 3687 | mb_debug(1, "discard preallocation for group %u\n", group); |
c9de560d AT |
3688 | |
3689 | if (list_empty(&grp->bb_prealloc_list)) | |
3690 | return 0; | |
3691 | ||
574ca174 | 3692 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3693 | if (bitmap_bh == NULL) { |
12062ddd | 3694 | ext4_error(sb, "Error reading block bitmap for %u", group); |
ce89f46c | 3695 | return 0; |
c9de560d AT |
3696 | } |
3697 | ||
3698 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3699 | if (err) { |
12062ddd | 3700 | ext4_error(sb, "Error loading buddy information for %u", group); |
ce89f46c AK |
3701 | put_bh(bitmap_bh); |
3702 | return 0; | |
3703 | } | |
c9de560d AT |
3704 | |
3705 | if (needed == 0) | |
7137d7a4 | 3706 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; |
c9de560d | 3707 | |
c9de560d | 3708 | INIT_LIST_HEAD(&list); |
c9de560d AT |
3709 | repeat: |
3710 | ext4_lock_group(sb, group); | |
3711 | list_for_each_entry_safe(pa, tmp, | |
3712 | &grp->bb_prealloc_list, pa_group_list) { | |
3713 | spin_lock(&pa->pa_lock); | |
3714 | if (atomic_read(&pa->pa_count)) { | |
3715 | spin_unlock(&pa->pa_lock); | |
3716 | busy = 1; | |
3717 | continue; | |
3718 | } | |
3719 | if (pa->pa_deleted) { | |
3720 | spin_unlock(&pa->pa_lock); | |
3721 | continue; | |
3722 | } | |
3723 | ||
3724 | /* seems this one can be freed ... */ | |
3725 | pa->pa_deleted = 1; | |
3726 | ||
3727 | /* we can trust pa_free ... */ | |
3728 | free += pa->pa_free; | |
3729 | ||
3730 | spin_unlock(&pa->pa_lock); | |
3731 | ||
3732 | list_del(&pa->pa_group_list); | |
3733 | list_add(&pa->u.pa_tmp_list, &list); | |
3734 | } | |
3735 | ||
3736 | /* if we still need more blocks and some PAs were used, try again */ | |
3737 | if (free < needed && busy) { | |
3738 | busy = 0; | |
3739 | ext4_unlock_group(sb, group); | |
3740 | /* | |
3741 | * Yield the CPU here so that we don't get soft lockup | |
3742 | * in non preempt case. | |
3743 | */ | |
3744 | yield(); | |
3745 | goto repeat; | |
3746 | } | |
3747 | ||
3748 | /* found anything to free? */ | |
3749 | if (list_empty(&list)) { | |
3750 | BUG_ON(free != 0); | |
3751 | goto out; | |
3752 | } | |
3753 | ||
3754 | /* now free all selected PAs */ | |
3755 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3756 | ||
3757 | /* remove from object (inode or locality group) */ | |
3758 | spin_lock(pa->pa_obj_lock); | |
3759 | list_del_rcu(&pa->pa_inode_list); | |
3760 | spin_unlock(pa->pa_obj_lock); | |
3761 | ||
cc0fb9ad | 3762 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 3763 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 3764 | else |
3e1e5f50 | 3765 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
3766 | |
3767 | list_del(&pa->u.pa_tmp_list); | |
3768 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3769 | } | |
3770 | ||
3771 | out: | |
3772 | ext4_unlock_group(sb, group); | |
e39e07fd | 3773 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
3774 | put_bh(bitmap_bh); |
3775 | return free; | |
3776 | } | |
3777 | ||
3778 | /* | |
3779 | * releases all non-used preallocated blocks for given inode | |
3780 | * | |
3781 | * It's important to discard preallocations under i_data_sem | |
3782 | * We don't want another block to be served from the prealloc | |
3783 | * space when we are discarding the inode prealloc space. | |
3784 | * | |
3785 | * FIXME!! Make sure it is valid at all the call sites | |
3786 | */ | |
c2ea3fde | 3787 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
3788 | { |
3789 | struct ext4_inode_info *ei = EXT4_I(inode); | |
3790 | struct super_block *sb = inode->i_sb; | |
3791 | struct buffer_head *bitmap_bh = NULL; | |
3792 | struct ext4_prealloc_space *pa, *tmp; | |
3793 | ext4_group_t group = 0; | |
3794 | struct list_head list; | |
3795 | struct ext4_buddy e4b; | |
3796 | int err; | |
3797 | ||
c2ea3fde | 3798 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
3799 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
3800 | return; | |
3801 | } | |
3802 | ||
6ba495e9 | 3803 | mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); |
9bffad1e | 3804 | trace_ext4_discard_preallocations(inode); |
c9de560d AT |
3805 | |
3806 | INIT_LIST_HEAD(&list); | |
3807 | ||
3808 | repeat: | |
3809 | /* first, collect all pa's in the inode */ | |
3810 | spin_lock(&ei->i_prealloc_lock); | |
3811 | while (!list_empty(&ei->i_prealloc_list)) { | |
3812 | pa = list_entry(ei->i_prealloc_list.next, | |
3813 | struct ext4_prealloc_space, pa_inode_list); | |
3814 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
3815 | spin_lock(&pa->pa_lock); | |
3816 | if (atomic_read(&pa->pa_count)) { | |
3817 | /* this shouldn't happen often - nobody should | |
3818 | * use preallocation while we're discarding it */ | |
3819 | spin_unlock(&pa->pa_lock); | |
3820 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
3821 | ext4_msg(sb, KERN_ERR, |
3822 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
3823 | WARN_ON(1); |
3824 | schedule_timeout_uninterruptible(HZ); | |
3825 | goto repeat; | |
3826 | ||
3827 | } | |
3828 | if (pa->pa_deleted == 0) { | |
3829 | pa->pa_deleted = 1; | |
3830 | spin_unlock(&pa->pa_lock); | |
3831 | list_del_rcu(&pa->pa_inode_list); | |
3832 | list_add(&pa->u.pa_tmp_list, &list); | |
3833 | continue; | |
3834 | } | |
3835 | ||
3836 | /* someone is deleting pa right now */ | |
3837 | spin_unlock(&pa->pa_lock); | |
3838 | spin_unlock(&ei->i_prealloc_lock); | |
3839 | ||
3840 | /* we have to wait here because pa_deleted | |
3841 | * doesn't mean pa is already unlinked from | |
3842 | * the list. as we might be called from | |
3843 | * ->clear_inode() the inode will get freed | |
3844 | * and concurrent thread which is unlinking | |
3845 | * pa from inode's list may access already | |
3846 | * freed memory, bad-bad-bad */ | |
3847 | ||
3848 | /* XXX: if this happens too often, we can | |
3849 | * add a flag to force wait only in case | |
3850 | * of ->clear_inode(), but not in case of | |
3851 | * regular truncate */ | |
3852 | schedule_timeout_uninterruptible(HZ); | |
3853 | goto repeat; | |
3854 | } | |
3855 | spin_unlock(&ei->i_prealloc_lock); | |
3856 | ||
3857 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 3858 | BUG_ON(pa->pa_type != MB_INODE_PA); |
c9de560d AT |
3859 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); |
3860 | ||
3861 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3862 | if (err) { |
12062ddd ES |
3863 | ext4_error(sb, "Error loading buddy information for %u", |
3864 | group); | |
ce89f46c AK |
3865 | continue; |
3866 | } | |
c9de560d | 3867 | |
574ca174 | 3868 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3869 | if (bitmap_bh == NULL) { |
12062ddd ES |
3870 | ext4_error(sb, "Error reading block bitmap for %u", |
3871 | group); | |
e39e07fd | 3872 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 3873 | continue; |
c9de560d AT |
3874 | } |
3875 | ||
3876 | ext4_lock_group(sb, group); | |
3877 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 3878 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
3879 | ext4_unlock_group(sb, group); |
3880 | ||
e39e07fd | 3881 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
3882 | put_bh(bitmap_bh); |
3883 | ||
3884 | list_del(&pa->u.pa_tmp_list); | |
3885 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3886 | } | |
3887 | } | |
3888 | ||
6ba495e9 | 3889 | #ifdef CONFIG_EXT4_DEBUG |
c9de560d AT |
3890 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
3891 | { | |
3892 | struct super_block *sb = ac->ac_sb; | |
8df9675f | 3893 | ext4_group_t ngroups, i; |
c9de560d | 3894 | |
4dd89fc6 TT |
3895 | if (!mb_enable_debug || |
3896 | (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) | |
e3570639 ES |
3897 | return; |
3898 | ||
9d8b9ec4 TT |
3899 | ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: Can't allocate:" |
3900 | " Allocation context details:"); | |
3901 | ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: status %d flags %d", | |
c9de560d | 3902 | ac->ac_status, ac->ac_flags); |
9d8b9ec4 TT |
3903 | ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: orig %lu/%lu/%lu@%lu, " |
3904 | "goal %lu/%lu/%lu@%lu, " | |
3905 | "best %lu/%lu/%lu@%lu cr %d", | |
c9de560d AT |
3906 | (unsigned long)ac->ac_o_ex.fe_group, |
3907 | (unsigned long)ac->ac_o_ex.fe_start, | |
3908 | (unsigned long)ac->ac_o_ex.fe_len, | |
3909 | (unsigned long)ac->ac_o_ex.fe_logical, | |
3910 | (unsigned long)ac->ac_g_ex.fe_group, | |
3911 | (unsigned long)ac->ac_g_ex.fe_start, | |
3912 | (unsigned long)ac->ac_g_ex.fe_len, | |
3913 | (unsigned long)ac->ac_g_ex.fe_logical, | |
3914 | (unsigned long)ac->ac_b_ex.fe_group, | |
3915 | (unsigned long)ac->ac_b_ex.fe_start, | |
3916 | (unsigned long)ac->ac_b_ex.fe_len, | |
3917 | (unsigned long)ac->ac_b_ex.fe_logical, | |
3918 | (int)ac->ac_criteria); | |
9d8b9ec4 TT |
3919 | ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: %lu scanned, %d found", |
3920 | ac->ac_ex_scanned, ac->ac_found); | |
3921 | ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: groups: "); | |
8df9675f TT |
3922 | ngroups = ext4_get_groups_count(sb); |
3923 | for (i = 0; i < ngroups; i++) { | |
c9de560d AT |
3924 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
3925 | struct ext4_prealloc_space *pa; | |
3926 | ext4_grpblk_t start; | |
3927 | struct list_head *cur; | |
3928 | ext4_lock_group(sb, i); | |
3929 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3930 | pa = list_entry(cur, struct ext4_prealloc_space, | |
3931 | pa_group_list); | |
3932 | spin_lock(&pa->pa_lock); | |
3933 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3934 | NULL, &start); | |
3935 | spin_unlock(&pa->pa_lock); | |
1c718505 AF |
3936 | printk(KERN_ERR "PA:%u:%d:%u \n", i, |
3937 | start, pa->pa_len); | |
c9de560d | 3938 | } |
60bd63d1 | 3939 | ext4_unlock_group(sb, i); |
c9de560d AT |
3940 | |
3941 | if (grp->bb_free == 0) | |
3942 | continue; | |
1c718505 | 3943 | printk(KERN_ERR "%u: %d/%d \n", |
c9de560d AT |
3944 | i, grp->bb_free, grp->bb_fragments); |
3945 | } | |
3946 | printk(KERN_ERR "\n"); | |
3947 | } | |
3948 | #else | |
3949 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
3950 | { | |
3951 | return; | |
3952 | } | |
3953 | #endif | |
3954 | ||
3955 | /* | |
3956 | * We use locality group preallocation for small size file. The size of the | |
3957 | * file is determined by the current size or the resulting size after | |
3958 | * allocation which ever is larger | |
3959 | * | |
b713a5ec | 3960 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
3961 | */ |
3962 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
3963 | { | |
3964 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3965 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
3966 | loff_t size, isize; | |
3967 | ||
3968 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3969 | return; | |
3970 | ||
4ba74d00 TT |
3971 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
3972 | return; | |
3973 | ||
53accfa9 | 3974 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
3975 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
3976 | >> bsbits; | |
c9de560d | 3977 | |
50797481 TT |
3978 | if ((size == isize) && |
3979 | !ext4_fs_is_busy(sbi) && | |
3980 | (atomic_read(&ac->ac_inode->i_writecount) == 0)) { | |
3981 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
3982 | return; | |
3983 | } | |
3984 | ||
c9de560d | 3985 | /* don't use group allocation for large files */ |
71780577 | 3986 | size = max(size, isize); |
cc483f10 | 3987 | if (size > sbi->s_mb_stream_request) { |
4ba74d00 | 3988 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
c9de560d | 3989 | return; |
4ba74d00 | 3990 | } |
c9de560d AT |
3991 | |
3992 | BUG_ON(ac->ac_lg != NULL); | |
3993 | /* | |
3994 | * locality group prealloc space are per cpu. The reason for having | |
3995 | * per cpu locality group is to reduce the contention between block | |
3996 | * request from multiple CPUs. | |
3997 | */ | |
ca0c9584 | 3998 | ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
3999 | |
4000 | /* we're going to use group allocation */ | |
4001 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4002 | ||
4003 | /* serialize all allocations in the group */ | |
4004 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4005 | } | |
4006 | ||
4ddfef7b ES |
4007 | static noinline_for_stack int |
4008 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
4009 | struct ext4_allocation_request *ar) |
4010 | { | |
4011 | struct super_block *sb = ar->inode->i_sb; | |
4012 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4013 | struct ext4_super_block *es = sbi->s_es; | |
4014 | ext4_group_t group; | |
498e5f24 TT |
4015 | unsigned int len; |
4016 | ext4_fsblk_t goal; | |
c9de560d AT |
4017 | ext4_grpblk_t block; |
4018 | ||
4019 | /* we can't allocate > group size */ | |
4020 | len = ar->len; | |
4021 | ||
4022 | /* just a dirty hack to filter too big requests */ | |
7137d7a4 TT |
4023 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb) - 10) |
4024 | len = EXT4_CLUSTERS_PER_GROUP(sb) - 10; | |
c9de560d AT |
4025 | |
4026 | /* start searching from the goal */ | |
4027 | goal = ar->goal; | |
4028 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4029 | goal >= ext4_blocks_count(es)) | |
4030 | goal = le32_to_cpu(es->s_first_data_block); | |
4031 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4032 | ||
4033 | /* set up allocation goals */ | |
833576b3 | 4034 | memset(ac, 0, sizeof(struct ext4_allocation_context)); |
53accfa9 | 4035 | ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1); |
c9de560d | 4036 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
4037 | ac->ac_sb = sb; |
4038 | ac->ac_inode = ar->inode; | |
53accfa9 | 4039 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
4040 | ac->ac_o_ex.fe_group = group; |
4041 | ac->ac_o_ex.fe_start = block; | |
4042 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 4043 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 4044 | ac->ac_flags = ar->flags; |
c9de560d AT |
4045 | |
4046 | /* we have to define context: we'll we work with a file or | |
4047 | * locality group. this is a policy, actually */ | |
4048 | ext4_mb_group_or_file(ac); | |
4049 | ||
6ba495e9 | 4050 | mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " |
c9de560d AT |
4051 | "left: %u/%u, right %u/%u to %swritable\n", |
4052 | (unsigned) ar->len, (unsigned) ar->logical, | |
4053 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4054 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4055 | (unsigned) ar->lright, (unsigned) ar->pright, | |
4056 | atomic_read(&ar->inode->i_writecount) ? "" : "non-"); | |
4057 | return 0; | |
4058 | ||
4059 | } | |
4060 | ||
6be2ded1 AK |
4061 | static noinline_for_stack void |
4062 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4063 | struct ext4_locality_group *lg, | |
4064 | int order, int total_entries) | |
4065 | { | |
4066 | ext4_group_t group = 0; | |
4067 | struct ext4_buddy e4b; | |
4068 | struct list_head discard_list; | |
4069 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 4070 | |
6ba495e9 | 4071 | mb_debug(1, "discard locality group preallocation\n"); |
6be2ded1 AK |
4072 | |
4073 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
4074 | |
4075 | spin_lock(&lg->lg_prealloc_lock); | |
4076 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
4077 | pa_inode_list) { | |
4078 | spin_lock(&pa->pa_lock); | |
4079 | if (atomic_read(&pa->pa_count)) { | |
4080 | /* | |
4081 | * This is the pa that we just used | |
4082 | * for block allocation. So don't | |
4083 | * free that | |
4084 | */ | |
4085 | spin_unlock(&pa->pa_lock); | |
4086 | continue; | |
4087 | } | |
4088 | if (pa->pa_deleted) { | |
4089 | spin_unlock(&pa->pa_lock); | |
4090 | continue; | |
4091 | } | |
4092 | /* only lg prealloc space */ | |
cc0fb9ad | 4093 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4094 | |
4095 | /* seems this one can be freed ... */ | |
4096 | pa->pa_deleted = 1; | |
4097 | spin_unlock(&pa->pa_lock); | |
4098 | ||
4099 | list_del_rcu(&pa->pa_inode_list); | |
4100 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4101 | ||
4102 | total_entries--; | |
4103 | if (total_entries <= 5) { | |
4104 | /* | |
4105 | * we want to keep only 5 entries | |
4106 | * allowing it to grow to 8. This | |
4107 | * mak sure we don't call discard | |
4108 | * soon for this list. | |
4109 | */ | |
4110 | break; | |
4111 | } | |
4112 | } | |
4113 | spin_unlock(&lg->lg_prealloc_lock); | |
4114 | ||
4115 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
4116 | ||
4117 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); | |
4118 | if (ext4_mb_load_buddy(sb, group, &e4b)) { | |
12062ddd ES |
4119 | ext4_error(sb, "Error loading buddy information for %u", |
4120 | group); | |
6be2ded1 AK |
4121 | continue; |
4122 | } | |
4123 | ext4_lock_group(sb, group); | |
4124 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4125 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
4126 | ext4_unlock_group(sb, group); |
4127 | ||
e39e07fd | 4128 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
4129 | list_del(&pa->u.pa_tmp_list); |
4130 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4131 | } | |
6be2ded1 AK |
4132 | } |
4133 | ||
4134 | /* | |
4135 | * We have incremented pa_count. So it cannot be freed at this | |
4136 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4137 | * possible from this lg. That means pa_free cannot be updated. | |
4138 | * | |
4139 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4140 | * which can cause the lg_prealloc_list to be updated. | |
4141 | */ | |
4142 | ||
4143 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4144 | { | |
4145 | int order, added = 0, lg_prealloc_count = 1; | |
4146 | struct super_block *sb = ac->ac_sb; | |
4147 | struct ext4_locality_group *lg = ac->ac_lg; | |
4148 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4149 | ||
4150 | order = fls(pa->pa_free) - 1; | |
4151 | if (order > PREALLOC_TB_SIZE - 1) | |
4152 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4153 | order = PREALLOC_TB_SIZE - 1; | |
4154 | /* Add the prealloc space to lg */ | |
4155 | rcu_read_lock(); | |
4156 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], | |
4157 | pa_inode_list) { | |
4158 | spin_lock(&tmp_pa->pa_lock); | |
4159 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4160 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4161 | continue; |
4162 | } | |
4163 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4164 | /* Add to the tail of the previous entry */ | |
4165 | list_add_tail_rcu(&pa->pa_inode_list, | |
4166 | &tmp_pa->pa_inode_list); | |
4167 | added = 1; | |
4168 | /* | |
4169 | * we want to count the total | |
4170 | * number of entries in the list | |
4171 | */ | |
4172 | } | |
4173 | spin_unlock(&tmp_pa->pa_lock); | |
4174 | lg_prealloc_count++; | |
4175 | } | |
4176 | if (!added) | |
4177 | list_add_tail_rcu(&pa->pa_inode_list, | |
4178 | &lg->lg_prealloc_list[order]); | |
4179 | rcu_read_unlock(); | |
4180 | ||
4181 | /* Now trim the list to be not more than 8 elements */ | |
4182 | if (lg_prealloc_count > 8) { | |
4183 | ext4_mb_discard_lg_preallocations(sb, lg, | |
4184 | order, lg_prealloc_count); | |
4185 | return; | |
4186 | } | |
4187 | return ; | |
4188 | } | |
4189 | ||
c9de560d AT |
4190 | /* |
4191 | * release all resource we used in allocation | |
4192 | */ | |
4193 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4194 | { | |
53accfa9 | 4195 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
4196 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4197 | if (pa) { | |
cc0fb9ad | 4198 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4199 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 4200 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
4201 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
4202 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
4203 | pa->pa_free -= ac->ac_b_ex.fe_len; |
4204 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4205 | spin_unlock(&pa->pa_lock); | |
c9de560d | 4206 | } |
c9de560d | 4207 | } |
ba443916 AK |
4208 | if (pa) { |
4209 | /* | |
4210 | * We want to add the pa to the right bucket. | |
4211 | * Remove it from the list and while adding | |
4212 | * make sure the list to which we are adding | |
44183d42 | 4213 | * doesn't grow big. |
ba443916 | 4214 | */ |
cc0fb9ad | 4215 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { |
ba443916 AK |
4216 | spin_lock(pa->pa_obj_lock); |
4217 | list_del_rcu(&pa->pa_inode_list); | |
4218 | spin_unlock(pa->pa_obj_lock); | |
4219 | ext4_mb_add_n_trim(ac); | |
4220 | } | |
4221 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4222 | } | |
c9de560d AT |
4223 | if (ac->ac_bitmap_page) |
4224 | page_cache_release(ac->ac_bitmap_page); | |
4225 | if (ac->ac_buddy_page) | |
4226 | page_cache_release(ac->ac_buddy_page); | |
4227 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
4228 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4229 | ext4_mb_collect_stats(ac); | |
4230 | return 0; | |
4231 | } | |
4232 | ||
4233 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4234 | { | |
8df9675f | 4235 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4236 | int ret; |
4237 | int freed = 0; | |
4238 | ||
9bffad1e | 4239 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4240 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4241 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4242 | freed += ret; | |
4243 | needed -= ret; | |
4244 | } | |
4245 | ||
4246 | return freed; | |
4247 | } | |
4248 | ||
4249 | /* | |
4250 | * Main entry point into mballoc to allocate blocks | |
4251 | * it tries to use preallocation first, then falls back | |
4252 | * to usual allocation | |
4253 | */ | |
4254 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 4255 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 4256 | { |
6bc6e63f | 4257 | int freed; |
256bdb49 | 4258 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4259 | struct ext4_sb_info *sbi; |
4260 | struct super_block *sb; | |
4261 | ext4_fsblk_t block = 0; | |
60e58e0f | 4262 | unsigned int inquota = 0; |
53accfa9 | 4263 | unsigned int reserv_clstrs = 0; |
c9de560d AT |
4264 | |
4265 | sb = ar->inode->i_sb; | |
4266 | sbi = EXT4_SB(sb); | |
4267 | ||
9bffad1e | 4268 | trace_ext4_request_blocks(ar); |
ba80b101 | 4269 | |
60e58e0f MC |
4270 | /* |
4271 | * For delayed allocation, we could skip the ENOSPC and | |
4272 | * EDQUOT check, as blocks and quotas have been already | |
4273 | * reserved when data being copied into pagecache. | |
4274 | */ | |
f2321097 | 4275 | if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED)) |
60e58e0f MC |
4276 | ar->flags |= EXT4_MB_DELALLOC_RESERVED; |
4277 | else { | |
4278 | /* Without delayed allocation we need to verify | |
4279 | * there is enough free blocks to do block allocation | |
4280 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4281 | */ |
55f020db AH |
4282 | while (ar->len && |
4283 | ext4_claim_free_blocks(sbi, ar->len, ar->flags)) { | |
4284 | ||
030ba6bc AK |
4285 | /* let others to free the space */ |
4286 | yield(); | |
4287 | ar->len = ar->len >> 1; | |
4288 | } | |
4289 | if (!ar->len) { | |
a30d542a AK |
4290 | *errp = -ENOSPC; |
4291 | return 0; | |
4292 | } | |
53accfa9 | 4293 | reserv_clstrs = ar->len; |
55f020db | 4294 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
4295 | dquot_alloc_block_nofail(ar->inode, |
4296 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
4297 | } else { |
4298 | while (ar->len && | |
53accfa9 TT |
4299 | dquot_alloc_block(ar->inode, |
4300 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
4301 | |
4302 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
4303 | ar->len--; | |
4304 | } | |
60e58e0f MC |
4305 | } |
4306 | inquota = ar->len; | |
4307 | if (ar->len == 0) { | |
4308 | *errp = -EDQUOT; | |
6c7a120a | 4309 | goto out; |
60e58e0f | 4310 | } |
07031431 | 4311 | } |
d2a17637 | 4312 | |
256bdb49 | 4313 | ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 4314 | if (!ac) { |
363d4251 | 4315 | ar->len = 0; |
256bdb49 | 4316 | *errp = -ENOMEM; |
6c7a120a | 4317 | goto out; |
256bdb49 ES |
4318 | } |
4319 | ||
256bdb49 | 4320 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4321 | if (*errp) { |
4322 | ar->len = 0; | |
6c7a120a | 4323 | goto out; |
c9de560d AT |
4324 | } |
4325 | ||
256bdb49 ES |
4326 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4327 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4328 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4329 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4330 | repeat: |
4331 | /* allocate space in core */ | |
6c7a120a AK |
4332 | *errp = ext4_mb_regular_allocator(ac); |
4333 | if (*errp) | |
4334 | goto errout; | |
c9de560d AT |
4335 | |
4336 | /* as we've just preallocated more space than | |
4337 | * user requested orinally, we store allocated | |
4338 | * space in a special descriptor */ | |
256bdb49 ES |
4339 | if (ac->ac_status == AC_STATUS_FOUND && |
4340 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
4341 | ext4_mb_new_preallocation(ac); | |
c9de560d | 4342 | } |
256bdb49 | 4343 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 4344 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
6c7a120a | 4345 | if (*errp == -EAGAIN) { |
8556e8f3 AK |
4346 | /* |
4347 | * drop the reference that we took | |
4348 | * in ext4_mb_use_best_found | |
4349 | */ | |
4350 | ext4_mb_release_context(ac); | |
519deca0 AK |
4351 | ac->ac_b_ex.fe_group = 0; |
4352 | ac->ac_b_ex.fe_start = 0; | |
4353 | ac->ac_b_ex.fe_len = 0; | |
4354 | ac->ac_status = AC_STATUS_CONTINUE; | |
4355 | goto repeat; | |
6c7a120a AK |
4356 | } else if (*errp) |
4357 | errout: | |
b844167e | 4358 | ext4_discard_allocated_blocks(ac); |
6c7a120a | 4359 | else { |
519deca0 AK |
4360 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
4361 | ar->len = ac->ac_b_ex.fe_len; | |
4362 | } | |
c9de560d | 4363 | } else { |
256bdb49 | 4364 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4365 | if (freed) |
4366 | goto repeat; | |
4367 | *errp = -ENOSPC; | |
6c7a120a AK |
4368 | } |
4369 | ||
4370 | if (*errp) { | |
256bdb49 | 4371 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4372 | ar->len = 0; |
256bdb49 | 4373 | ext4_mb_show_ac(ac); |
c9de560d | 4374 | } |
256bdb49 | 4375 | ext4_mb_release_context(ac); |
6c7a120a AK |
4376 | out: |
4377 | if (ac) | |
4378 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 4379 | if (inquota && ar->len < inquota) |
53accfa9 | 4380 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 4381 | if (!ar->len) { |
f2321097 TT |
4382 | if (!ext4_test_inode_state(ar->inode, |
4383 | EXT4_STATE_DELALLOC_RESERVED)) | |
0087d9fb AK |
4384 | /* release all the reserved blocks if non delalloc */ |
4385 | percpu_counter_sub(&sbi->s_dirtyblocks_counter, | |
53accfa9 | 4386 | reserv_clstrs); |
0087d9fb | 4387 | } |
c9de560d | 4388 | |
9bffad1e | 4389 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4390 | |
c9de560d AT |
4391 | return block; |
4392 | } | |
c9de560d | 4393 | |
c894058d AK |
4394 | /* |
4395 | * We can merge two free data extents only if the physical blocks | |
4396 | * are contiguous, AND the extents were freed by the same transaction, | |
4397 | * AND the blocks are associated with the same group. | |
4398 | */ | |
4399 | static int can_merge(struct ext4_free_data *entry1, | |
4400 | struct ext4_free_data *entry2) | |
4401 | { | |
4402 | if ((entry1->t_tid == entry2->t_tid) && | |
4403 | (entry1->group == entry2->group) && | |
4404 | ((entry1->start_blk + entry1->count) == entry2->start_blk)) | |
4405 | return 1; | |
4406 | return 0; | |
4407 | } | |
4408 | ||
4ddfef7b ES |
4409 | static noinline_for_stack int |
4410 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 4411 | struct ext4_free_data *new_entry) |
c9de560d | 4412 | { |
e29136f8 | 4413 | ext4_group_t group = e4b->bd_group; |
7a2fcbf7 AK |
4414 | ext4_grpblk_t block; |
4415 | struct ext4_free_data *entry; | |
c9de560d AT |
4416 | struct ext4_group_info *db = e4b->bd_info; |
4417 | struct super_block *sb = e4b->bd_sb; | |
4418 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4419 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
4420 | struct rb_node *parent = NULL, *new_node; | |
4421 | ||
0390131b | 4422 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4423 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4424 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4425 | ||
c894058d | 4426 | new_node = &new_entry->node; |
7a2fcbf7 | 4427 | block = new_entry->start_blk; |
c894058d | 4428 | |
c894058d AK |
4429 | if (!*n) { |
4430 | /* first free block exent. We need to | |
4431 | protect buddy cache from being freed, | |
4432 | * otherwise we'll refresh it from | |
4433 | * on-disk bitmap and lose not-yet-available | |
4434 | * blocks */ | |
4435 | page_cache_get(e4b->bd_buddy_page); | |
4436 | page_cache_get(e4b->bd_bitmap_page); | |
4437 | } | |
4438 | while (*n) { | |
4439 | parent = *n; | |
4440 | entry = rb_entry(parent, struct ext4_free_data, node); | |
4441 | if (block < entry->start_blk) | |
4442 | n = &(*n)->rb_left; | |
4443 | else if (block >= (entry->start_blk + entry->count)) | |
4444 | n = &(*n)->rb_right; | |
4445 | else { | |
e29136f8 TT |
4446 | ext4_grp_locked_error(sb, group, 0, |
4447 | ext4_group_first_block_no(sb, group) + block, | |
4448 | "Block already on to-be-freed list"); | |
c894058d | 4449 | return 0; |
c9de560d | 4450 | } |
c894058d | 4451 | } |
c9de560d | 4452 | |
c894058d AK |
4453 | rb_link_node(new_node, parent, n); |
4454 | rb_insert_color(new_node, &db->bb_free_root); | |
4455 | ||
4456 | /* Now try to see the extent can be merged to left and right */ | |
4457 | node = rb_prev(new_node); | |
4458 | if (node) { | |
4459 | entry = rb_entry(node, struct ext4_free_data, node); | |
4460 | if (can_merge(entry, new_entry)) { | |
4461 | new_entry->start_blk = entry->start_blk; | |
4462 | new_entry->count += entry->count; | |
4463 | rb_erase(node, &(db->bb_free_root)); | |
4464 | spin_lock(&sbi->s_md_lock); | |
4465 | list_del(&entry->list); | |
4466 | spin_unlock(&sbi->s_md_lock); | |
4467 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
c9de560d | 4468 | } |
c894058d | 4469 | } |
c9de560d | 4470 | |
c894058d AK |
4471 | node = rb_next(new_node); |
4472 | if (node) { | |
4473 | entry = rb_entry(node, struct ext4_free_data, node); | |
4474 | if (can_merge(new_entry, entry)) { | |
4475 | new_entry->count += entry->count; | |
4476 | rb_erase(node, &(db->bb_free_root)); | |
4477 | spin_lock(&sbi->s_md_lock); | |
4478 | list_del(&entry->list); | |
4479 | spin_unlock(&sbi->s_md_lock); | |
4480 | kmem_cache_free(ext4_free_ext_cachep, entry); | |
c9de560d AT |
4481 | } |
4482 | } | |
3e624fc7 | 4483 | /* Add the extent to transaction's private list */ |
c894058d | 4484 | spin_lock(&sbi->s_md_lock); |
3e624fc7 | 4485 | list_add(&new_entry->list, &handle->h_transaction->t_private_list); |
c894058d | 4486 | spin_unlock(&sbi->s_md_lock); |
c9de560d AT |
4487 | return 0; |
4488 | } | |
4489 | ||
44338711 TT |
4490 | /** |
4491 | * ext4_free_blocks() -- Free given blocks and update quota | |
4492 | * @handle: handle for this transaction | |
4493 | * @inode: inode | |
4494 | * @block: start physical block to free | |
4495 | * @count: number of blocks to count | |
5def1360 | 4496 | * @flags: flags used by ext4_free_blocks |
c9de560d | 4497 | */ |
44338711 | 4498 | void ext4_free_blocks(handle_t *handle, struct inode *inode, |
e6362609 TT |
4499 | struct buffer_head *bh, ext4_fsblk_t block, |
4500 | unsigned long count, int flags) | |
c9de560d | 4501 | { |
26346ff6 | 4502 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4503 | struct super_block *sb = inode->i_sb; |
c9de560d | 4504 | struct ext4_group_desc *gdp; |
44338711 | 4505 | unsigned long freed = 0; |
498e5f24 | 4506 | unsigned int overflow; |
c9de560d AT |
4507 | ext4_grpblk_t bit; |
4508 | struct buffer_head *gd_bh; | |
4509 | ext4_group_t block_group; | |
4510 | struct ext4_sb_info *sbi; | |
4511 | struct ext4_buddy e4b; | |
4512 | int err = 0; | |
4513 | int ret; | |
4514 | ||
e6362609 TT |
4515 | if (bh) { |
4516 | if (block) | |
4517 | BUG_ON(block != bh->b_blocknr); | |
4518 | else | |
4519 | block = bh->b_blocknr; | |
4520 | } | |
c9de560d | 4521 | |
c9de560d | 4522 | sbi = EXT4_SB(sb); |
1f2acb60 TT |
4523 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
4524 | !ext4_data_block_valid(sbi, block, count)) { | |
12062ddd | 4525 | ext4_error(sb, "Freeing blocks not in datazone - " |
1f2acb60 | 4526 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
4527 | goto error_return; |
4528 | } | |
4529 | ||
0610b6e9 | 4530 | ext4_debug("freeing block %llu\n", block); |
e6362609 TT |
4531 | trace_ext4_free_blocks(inode, block, count, flags); |
4532 | ||
4533 | if (flags & EXT4_FREE_BLOCKS_FORGET) { | |
4534 | struct buffer_head *tbh = bh; | |
4535 | int i; | |
4536 | ||
4537 | BUG_ON(bh && (count > 1)); | |
4538 | ||
4539 | for (i = 0; i < count; i++) { | |
4540 | if (!bh) | |
4541 | tbh = sb_find_get_block(inode->i_sb, | |
4542 | block + i); | |
87783690 NK |
4543 | if (unlikely(!tbh)) |
4544 | continue; | |
60e6679e | 4545 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, |
e6362609 TT |
4546 | inode, tbh, block + i); |
4547 | } | |
4548 | } | |
4549 | ||
60e6679e | 4550 | /* |
e6362609 TT |
4551 | * We need to make sure we don't reuse the freed block until |
4552 | * after the transaction is committed, which we can do by | |
4553 | * treating the block as metadata, below. We make an | |
4554 | * exception if the inode is to be written in writeback mode | |
4555 | * since writeback mode has weak data consistency guarantees. | |
4556 | */ | |
4557 | if (!ext4_should_writeback_data(inode)) | |
4558 | flags |= EXT4_FREE_BLOCKS_METADATA; | |
c9de560d | 4559 | |
c9de560d AT |
4560 | do_more: |
4561 | overflow = 0; | |
4562 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4563 | ||
4564 | /* | |
4565 | * Check to see if we are freeing blocks across a group | |
4566 | * boundary. | |
4567 | */ | |
7137d7a4 TT |
4568 | if (bit + count > EXT4_CLUSTERS_PER_GROUP(sb)) { |
4569 | overflow = bit + count - EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
4570 | count -= overflow; |
4571 | } | |
574ca174 | 4572 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
ce89f46c AK |
4573 | if (!bitmap_bh) { |
4574 | err = -EIO; | |
c9de560d | 4575 | goto error_return; |
ce89f46c | 4576 | } |
c9de560d | 4577 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4578 | if (!gdp) { |
4579 | err = -EIO; | |
c9de560d | 4580 | goto error_return; |
ce89f46c | 4581 | } |
c9de560d AT |
4582 | |
4583 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4584 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4585 | in_range(block, ext4_inode_table(sb, gdp), | |
4586 | EXT4_SB(sb)->s_itb_per_group) || | |
4587 | in_range(block + count - 1, ext4_inode_table(sb, gdp), | |
4588 | EXT4_SB(sb)->s_itb_per_group)) { | |
4589 | ||
12062ddd | 4590 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 4591 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
4592 | /* err = 0. ext4_std_error should be a no op */ |
4593 | goto error_return; | |
c9de560d AT |
4594 | } |
4595 | ||
4596 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4597 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4598 | if (err) | |
4599 | goto error_return; | |
4600 | ||
4601 | /* | |
4602 | * We are about to modify some metadata. Call the journal APIs | |
4603 | * to unshare ->b_data if a currently-committing transaction is | |
4604 | * using it | |
4605 | */ | |
4606 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4607 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4608 | if (err) | |
4609 | goto error_return; | |
c9de560d AT |
4610 | #ifdef AGGRESSIVE_CHECK |
4611 | { | |
4612 | int i; | |
4613 | for (i = 0; i < count; i++) | |
4614 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); | |
4615 | } | |
4616 | #endif | |
3e1e5f50 | 4617 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count); |
c9de560d | 4618 | |
920313a7 AK |
4619 | err = ext4_mb_load_buddy(sb, block_group, &e4b); |
4620 | if (err) | |
4621 | goto error_return; | |
e6362609 TT |
4622 | |
4623 | if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) { | |
7a2fcbf7 AK |
4624 | struct ext4_free_data *new_entry; |
4625 | /* | |
4626 | * blocks being freed are metadata. these blocks shouldn't | |
4627 | * be used until this transaction is committed | |
4628 | */ | |
b72143ab TT |
4629 | new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS); |
4630 | if (!new_entry) { | |
4631 | err = -ENOMEM; | |
4632 | goto error_return; | |
4633 | } | |
7a2fcbf7 AK |
4634 | new_entry->start_blk = bit; |
4635 | new_entry->group = block_group; | |
4636 | new_entry->count = count; | |
4637 | new_entry->t_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 4638 | |
7a2fcbf7 | 4639 | ext4_lock_group(sb, block_group); |
955ce5f5 | 4640 | mb_clear_bits(bitmap_bh->b_data, bit, count); |
7a2fcbf7 | 4641 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 4642 | } else { |
7a2fcbf7 AK |
4643 | /* need to update group_info->bb_free and bitmap |
4644 | * with group lock held. generate_buddy look at | |
4645 | * them with group lock_held | |
4646 | */ | |
955ce5f5 AK |
4647 | ext4_lock_group(sb, block_group); |
4648 | mb_clear_bits(bitmap_bh->b_data, bit, count); | |
7e5a8cdd | 4649 | mb_free_blocks(inode, &e4b, bit, count); |
c9de560d AT |
4650 | } |
4651 | ||
560671a0 AK |
4652 | ret = ext4_free_blks_count(sb, gdp) + count; |
4653 | ext4_free_blks_set(sb, gdp, ret); | |
c9de560d | 4654 | gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); |
955ce5f5 | 4655 | ext4_unlock_group(sb, block_group); |
c9de560d AT |
4656 | percpu_counter_add(&sbi->s_freeblocks_counter, count); |
4657 | ||
772cb7c8 JS |
4658 | if (sbi->s_log_groups_per_flex) { |
4659 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
9f24e420 | 4660 | atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks); |
772cb7c8 JS |
4661 | } |
4662 | ||
e39e07fd | 4663 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 4664 | |
44338711 | 4665 | freed += count; |
c9de560d | 4666 | |
7a2fcbf7 AK |
4667 | /* We dirtied the bitmap block */ |
4668 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4669 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4670 | ||
c9de560d AT |
4671 | /* And the group descriptor block */ |
4672 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 4673 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
4674 | if (!err) |
4675 | err = ret; | |
4676 | ||
4677 | if (overflow && !err) { | |
4678 | block += count; | |
4679 | count = overflow; | |
4680 | put_bh(bitmap_bh); | |
4681 | goto do_more; | |
4682 | } | |
a0375156 | 4683 | ext4_mark_super_dirty(sb); |
c9de560d | 4684 | error_return: |
7132de74 | 4685 | if (freed && !(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) |
5dd4056d | 4686 | dquot_free_block(inode, freed); |
c9de560d AT |
4687 | brelse(bitmap_bh); |
4688 | ext4_std_error(sb, err); | |
4689 | return; | |
4690 | } | |
7360d173 | 4691 | |
2846e820 | 4692 | /** |
0529155e | 4693 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
4694 | * @handle: handle to this transaction |
4695 | * @sb: super block | |
4696 | * @block: start physcial block to add to the block group | |
4697 | * @count: number of blocks to free | |
4698 | * | |
e73a347b | 4699 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 4700 | */ |
cc7365df | 4701 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
4702 | ext4_fsblk_t block, unsigned long count) |
4703 | { | |
4704 | struct buffer_head *bitmap_bh = NULL; | |
4705 | struct buffer_head *gd_bh; | |
4706 | ext4_group_t block_group; | |
4707 | ext4_grpblk_t bit; | |
4708 | unsigned int i; | |
4709 | struct ext4_group_desc *desc; | |
4710 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 4711 | struct ext4_buddy e4b; |
2846e820 AG |
4712 | int err = 0, ret, blk_free_count; |
4713 | ext4_grpblk_t blocks_freed; | |
2846e820 AG |
4714 | |
4715 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
4716 | ||
4740b830 YY |
4717 | if (count == 0) |
4718 | return 0; | |
4719 | ||
2846e820 | 4720 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
4721 | /* |
4722 | * Check to see if we are freeing blocks across a group | |
4723 | * boundary. | |
4724 | */ | |
cc7365df YY |
4725 | if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
4726 | ext4_warning(sb, "too much blocks added to group %u\n", | |
4727 | block_group); | |
4728 | err = -EINVAL; | |
2846e820 | 4729 | goto error_return; |
cc7365df | 4730 | } |
2cd05cc3 | 4731 | |
2846e820 | 4732 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
cc7365df YY |
4733 | if (!bitmap_bh) { |
4734 | err = -EIO; | |
2846e820 | 4735 | goto error_return; |
cc7365df YY |
4736 | } |
4737 | ||
2846e820 | 4738 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
4739 | if (!desc) { |
4740 | err = -EIO; | |
2846e820 | 4741 | goto error_return; |
cc7365df | 4742 | } |
2846e820 AG |
4743 | |
4744 | if (in_range(ext4_block_bitmap(sb, desc), block, count) || | |
4745 | in_range(ext4_inode_bitmap(sb, desc), block, count) || | |
4746 | in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || | |
4747 | in_range(block + count - 1, ext4_inode_table(sb, desc), | |
4748 | sbi->s_itb_per_group)) { | |
4749 | ext4_error(sb, "Adding blocks in system zones - " | |
4750 | "Block = %llu, count = %lu", | |
4751 | block, count); | |
cc7365df | 4752 | err = -EINVAL; |
2846e820 AG |
4753 | goto error_return; |
4754 | } | |
4755 | ||
2cd05cc3 TT |
4756 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
4757 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2846e820 AG |
4758 | if (err) |
4759 | goto error_return; | |
4760 | ||
4761 | /* | |
4762 | * We are about to modify some metadata. Call the journal APIs | |
4763 | * to unshare ->b_data if a currently-committing transaction is | |
4764 | * using it | |
4765 | */ | |
4766 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4767 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4768 | if (err) | |
4769 | goto error_return; | |
e73a347b | 4770 | |
2846e820 AG |
4771 | for (i = 0, blocks_freed = 0; i < count; i++) { |
4772 | BUFFER_TRACE(bitmap_bh, "clear bit"); | |
e73a347b | 4773 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
4774 | ext4_error(sb, "bit already cleared for block %llu", |
4775 | (ext4_fsblk_t)(block + i)); | |
4776 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
4777 | } else { | |
4778 | blocks_freed++; | |
4779 | } | |
4780 | } | |
e73a347b AG |
4781 | |
4782 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
4783 | if (err) | |
4784 | goto error_return; | |
4785 | ||
4786 | /* | |
4787 | * need to update group_info->bb_free and bitmap | |
4788 | * with group lock held. generate_buddy look at | |
4789 | * them with group lock_held | |
4790 | */ | |
2846e820 | 4791 | ext4_lock_group(sb, block_group); |
e73a347b AG |
4792 | mb_clear_bits(bitmap_bh->b_data, bit, count); |
4793 | mb_free_blocks(NULL, &e4b, bit, count); | |
2846e820 AG |
4794 | blk_free_count = blocks_freed + ext4_free_blks_count(sb, desc); |
4795 | ext4_free_blks_set(sb, desc, blk_free_count); | |
4796 | desc->bg_checksum = ext4_group_desc_csum(sbi, block_group, desc); | |
4797 | ext4_unlock_group(sb, block_group); | |
4798 | percpu_counter_add(&sbi->s_freeblocks_counter, blocks_freed); | |
4799 | ||
4800 | if (sbi->s_log_groups_per_flex) { | |
4801 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
4802 | atomic_add(blocks_freed, | |
4803 | &sbi->s_flex_groups[flex_group].free_blocks); | |
4804 | } | |
e73a347b AG |
4805 | |
4806 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
4807 | |
4808 | /* We dirtied the bitmap block */ | |
4809 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4810 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4811 | ||
4812 | /* And the group descriptor block */ | |
4813 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
4814 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
4815 | if (!err) | |
4816 | err = ret; | |
4817 | ||
4818 | error_return: | |
4819 | brelse(bitmap_bh); | |
4820 | ext4_std_error(sb, err); | |
cc7365df | 4821 | return err; |
2846e820 AG |
4822 | } |
4823 | ||
7360d173 LC |
4824 | /** |
4825 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
4826 | * @sb: super block for the file system | |
4827 | * @start: starting block of the free extent in the alloc. group | |
4828 | * @count: number of blocks to TRIM | |
4829 | * @group: alloc. group we are working with | |
4830 | * @e4b: ext4 buddy for the group | |
4831 | * | |
4832 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
4833 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
4834 | * be called with under the group lock. | |
4835 | */ | |
d9f34504 TT |
4836 | static void ext4_trim_extent(struct super_block *sb, int start, int count, |
4837 | ext4_group_t group, struct ext4_buddy *e4b) | |
7360d173 LC |
4838 | { |
4839 | struct ext4_free_extent ex; | |
7360d173 | 4840 | |
b3d4c2b1 TM |
4841 | trace_ext4_trim_extent(sb, group, start, count); |
4842 | ||
7360d173 LC |
4843 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
4844 | ||
4845 | ex.fe_start = start; | |
4846 | ex.fe_group = group; | |
4847 | ex.fe_len = count; | |
4848 | ||
4849 | /* | |
4850 | * Mark blocks used, so no one can reuse them while | |
4851 | * being trimmed. | |
4852 | */ | |
4853 | mb_mark_used(e4b, &ex); | |
4854 | ext4_unlock_group(sb, group); | |
d9f34504 | 4855 | ext4_issue_discard(sb, group, start, count); |
7360d173 LC |
4856 | ext4_lock_group(sb, group); |
4857 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
7360d173 LC |
4858 | } |
4859 | ||
4860 | /** | |
4861 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
4862 | * @sb: super block for file system | |
22612283 | 4863 | * @group: group to be trimmed |
7360d173 LC |
4864 | * @start: first group block to examine |
4865 | * @max: last group block to examine | |
4866 | * @minblocks: minimum extent block count | |
4867 | * | |
4868 | * ext4_trim_all_free walks through group's buddy bitmap searching for free | |
4869 | * extents. When the free block is found, ext4_trim_extent is called to TRIM | |
4870 | * the extent. | |
4871 | * | |
4872 | * | |
4873 | * ext4_trim_all_free walks through group's block bitmap searching for free | |
4874 | * extents. When the free extent is found, mark it as used in group buddy | |
4875 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
4876 | * the group buddy bitmap. This is done until whole group is scanned. | |
4877 | */ | |
0b75a840 | 4878 | static ext4_grpblk_t |
78944086 LC |
4879 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
4880 | ext4_grpblk_t start, ext4_grpblk_t max, | |
4881 | ext4_grpblk_t minblocks) | |
7360d173 LC |
4882 | { |
4883 | void *bitmap; | |
169ddc3e | 4884 | ext4_grpblk_t next, count = 0, free_count = 0; |
78944086 LC |
4885 | struct ext4_buddy e4b; |
4886 | int ret; | |
7360d173 | 4887 | |
b3d4c2b1 TM |
4888 | trace_ext4_trim_all_free(sb, group, start, max); |
4889 | ||
78944086 LC |
4890 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
4891 | if (ret) { | |
4892 | ext4_error(sb, "Error in loading buddy " | |
4893 | "information for %u", group); | |
4894 | return ret; | |
4895 | } | |
78944086 | 4896 | bitmap = e4b.bd_bitmap; |
28739eea LC |
4897 | |
4898 | ext4_lock_group(sb, group); | |
3d56b8d2 TM |
4899 | if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && |
4900 | minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) | |
4901 | goto out; | |
4902 | ||
78944086 LC |
4903 | start = (e4b.bd_info->bb_first_free > start) ? |
4904 | e4b.bd_info->bb_first_free : start; | |
7360d173 LC |
4905 | |
4906 | while (start < max) { | |
4907 | start = mb_find_next_zero_bit(bitmap, max, start); | |
4908 | if (start >= max) | |
4909 | break; | |
4910 | next = mb_find_next_bit(bitmap, max, start); | |
4911 | ||
4912 | if ((next - start) >= minblocks) { | |
d9f34504 | 4913 | ext4_trim_extent(sb, start, |
78944086 | 4914 | next - start, group, &e4b); |
7360d173 LC |
4915 | count += next - start; |
4916 | } | |
169ddc3e | 4917 | free_count += next - start; |
7360d173 LC |
4918 | start = next + 1; |
4919 | ||
4920 | if (fatal_signal_pending(current)) { | |
4921 | count = -ERESTARTSYS; | |
4922 | break; | |
4923 | } | |
4924 | ||
4925 | if (need_resched()) { | |
4926 | ext4_unlock_group(sb, group); | |
4927 | cond_resched(); | |
4928 | ext4_lock_group(sb, group); | |
4929 | } | |
4930 | ||
169ddc3e | 4931 | if ((e4b.bd_info->bb_free - free_count) < minblocks) |
7360d173 LC |
4932 | break; |
4933 | } | |
3d56b8d2 TM |
4934 | |
4935 | if (!ret) | |
4936 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); | |
4937 | out: | |
7360d173 | 4938 | ext4_unlock_group(sb, group); |
78944086 | 4939 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
4940 | |
4941 | ext4_debug("trimmed %d blocks in the group %d\n", | |
4942 | count, group); | |
4943 | ||
7360d173 LC |
4944 | return count; |
4945 | } | |
4946 | ||
4947 | /** | |
4948 | * ext4_trim_fs() -- trim ioctl handle function | |
4949 | * @sb: superblock for filesystem | |
4950 | * @range: fstrim_range structure | |
4951 | * | |
4952 | * start: First Byte to trim | |
4953 | * len: number of Bytes to trim from start | |
4954 | * minlen: minimum extent length in Bytes | |
4955 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
4956 | * start to start+len. For each such a group ext4_trim_all_free function | |
4957 | * is invoked to trim all free space. | |
4958 | */ | |
4959 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
4960 | { | |
78944086 | 4961 | struct ext4_group_info *grp; |
7360d173 LC |
4962 | ext4_group_t first_group, last_group; |
4963 | ext4_group_t group, ngroups = ext4_get_groups_count(sb); | |
7137d7a4 | 4964 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
78944086 | 4965 | uint64_t start, len, minlen, trimmed = 0; |
0f0a25bf JK |
4966 | ext4_fsblk_t first_data_blk = |
4967 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
7360d173 LC |
4968 | int ret = 0; |
4969 | ||
4970 | start = range->start >> sb->s_blocksize_bits; | |
4971 | len = range->len >> sb->s_blocksize_bits; | |
4972 | minlen = range->minlen >> sb->s_blocksize_bits; | |
7360d173 | 4973 | |
7137d7a4 | 4974 | if (unlikely(minlen > EXT4_CLUSTERS_PER_GROUP(sb))) |
7360d173 | 4975 | return -EINVAL; |
22f10457 TM |
4976 | if (start + len <= first_data_blk) |
4977 | goto out; | |
0f0a25bf JK |
4978 | if (start < first_data_blk) { |
4979 | len -= first_data_blk - start; | |
4980 | start = first_data_blk; | |
4981 | } | |
7360d173 LC |
4982 | |
4983 | /* Determine first and last group to examine based on start and len */ | |
4984 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, | |
7137d7a4 | 4985 | &first_group, &first_cluster); |
7360d173 | 4986 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) (start + len), |
7137d7a4 | 4987 | &last_group, &last_cluster); |
7360d173 | 4988 | last_group = (last_group > ngroups - 1) ? ngroups - 1 : last_group; |
7137d7a4 | 4989 | last_cluster = EXT4_CLUSTERS_PER_GROUP(sb); |
7360d173 LC |
4990 | |
4991 | if (first_group > last_group) | |
4992 | return -EINVAL; | |
4993 | ||
4994 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
4995 | grp = ext4_get_group_info(sb, group); |
4996 | /* We only do this if the grp has never been initialized */ | |
4997 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
4998 | ret = ext4_mb_init_group(sb, group); | |
4999 | if (ret) | |
5000 | break; | |
7360d173 LC |
5001 | } |
5002 | ||
0ba08517 TM |
5003 | /* |
5004 | * For all the groups except the last one, last block will | |
5005 | * always be EXT4_BLOCKS_PER_GROUP(sb), so we only need to | |
5006 | * change it for the last group in which case start + | |
5007 | * len < EXT4_BLOCKS_PER_GROUP(sb). | |
5008 | */ | |
7137d7a4 TT |
5009 | if (first_cluster + len < EXT4_CLUSTERS_PER_GROUP(sb)) |
5010 | last_cluster = first_cluster + len; | |
5011 | len -= last_cluster - first_cluster; | |
7360d173 | 5012 | |
78944086 | 5013 | if (grp->bb_free >= minlen) { |
7137d7a4 TT |
5014 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
5015 | last_cluster, minlen); | |
7360d173 LC |
5016 | if (cnt < 0) { |
5017 | ret = cnt; | |
7360d173 LC |
5018 | break; |
5019 | } | |
5020 | } | |
7360d173 | 5021 | trimmed += cnt; |
7137d7a4 | 5022 | first_cluster = 0; |
7360d173 LC |
5023 | } |
5024 | range->len = trimmed * sb->s_blocksize; | |
5025 | ||
3d56b8d2 TM |
5026 | if (!ret) |
5027 | atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); | |
5028 | ||
22f10457 | 5029 | out: |
7360d173 LC |
5030 | return ret; |
5031 | } |