]>
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 | ||
18aadd47 | 24 | #include "ext4_jbd2.h" |
8f6e39a7 | 25 | #include "mballoc.h" |
28623c2f | 26 | #include <linux/log2.h> |
a0b30c12 | 27 | #include <linux/module.h> |
5a0e3ad6 | 28 | #include <linux/slab.h> |
9bffad1e TT |
29 | #include <trace/events/ext4.h> |
30 | ||
a0b30c12 TT |
31 | #ifdef CONFIG_EXT4_DEBUG |
32 | ushort ext4_mballoc_debug __read_mostly; | |
33 | ||
34 | module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644); | |
35 | MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc"); | |
36 | #endif | |
37 | ||
c9de560d AT |
38 | /* |
39 | * MUSTDO: | |
40 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
41 | * - search for metadata in few groups | |
42 | * | |
43 | * TODO v4: | |
44 | * - normalization should take into account whether file is still open | |
45 | * - discard preallocations if no free space left (policy?) | |
46 | * - don't normalize tails | |
47 | * - quota | |
48 | * - reservation for superuser | |
49 | * | |
50 | * TODO v3: | |
51 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
52 | * - track min/max extents in each group for better group selection | |
53 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
54 | * - tree of groups sorted by number of free blocks | |
55 | * - error handling | |
56 | */ | |
57 | ||
58 | /* | |
59 | * The allocation request involve request for multiple number of blocks | |
60 | * near to the goal(block) value specified. | |
61 | * | |
b713a5ec TT |
62 | * During initialization phase of the allocator we decide to use the |
63 | * group preallocation or inode preallocation depending on the size of | |
64 | * the file. The size of the file could be the resulting file size we | |
65 | * would have after allocation, or the current file size, which ever | |
66 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
67 | * select to use the group preallocation. The default value of | |
68 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
69 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
70 | * terms of number of blocks. | |
c9de560d AT |
71 | * |
72 | * The main motivation for having small file use group preallocation is to | |
b713a5ec | 73 | * ensure that we have small files closer together on the disk. |
c9de560d | 74 | * |
b713a5ec TT |
75 | * First stage the allocator looks at the inode prealloc list, |
76 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
77 | * spaces for this particular inode. The inode prealloc space is | |
78 | * represented as: | |
c9de560d AT |
79 | * |
80 | * pa_lstart -> the logical start block for this prealloc space | |
81 | * pa_pstart -> the physical start block for this prealloc space | |
53accfa9 TT |
82 | * pa_len -> length for this prealloc space (in clusters) |
83 | * pa_free -> free space available in this prealloc space (in clusters) | |
c9de560d AT |
84 | * |
85 | * The inode preallocation space is used looking at the _logical_ start | |
86 | * block. If only the logical file block falls within the range of prealloc | |
caaf7a29 TM |
87 | * space we will consume the particular prealloc space. This makes sure that |
88 | * we have contiguous physical blocks representing the file blocks | |
c9de560d AT |
89 | * |
90 | * The important thing to be noted in case of inode prealloc space is that | |
91 | * we don't modify the values associated to inode prealloc space except | |
92 | * pa_free. | |
93 | * | |
94 | * If we are not able to find blocks in the inode prealloc space and if we | |
95 | * have the group allocation flag set then we look at the locality group | |
caaf7a29 | 96 | * prealloc space. These are per CPU prealloc list represented as |
c9de560d AT |
97 | * |
98 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
99 | * | |
100 | * The reason for having a per cpu locality group is to reduce the contention | |
101 | * between CPUs. It is possible to get scheduled at this point. | |
102 | * | |
103 | * The locality group prealloc space is used looking at whether we have | |
25985edc | 104 | * enough free space (pa_free) within the prealloc space. |
c9de560d AT |
105 | * |
106 | * If we can't allocate blocks via inode prealloc or/and locality group | |
107 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
108 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
109 | * mapped to the buddy and bitmap information regarding different | |
110 | * groups. The buddy information is attached to buddy cache inode so that | |
111 | * we can access them through the page cache. The information regarding | |
112 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
113 | * block bitmap and buddy information. The information are stored in the | |
114 | * inode as: | |
115 | * | |
116 | * { page } | |
c3a326a6 | 117 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
118 | * |
119 | * | |
120 | * one block each for bitmap and buddy information. So for each group we | |
121 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / | |
122 | * blocksize) blocks. So it can have information regarding groups_per_page | |
123 | * which is blocks_per_page/2 | |
124 | * | |
125 | * The buddy cache inode is not stored on disk. The inode is thrown | |
126 | * away when the filesystem is unmounted. | |
127 | * | |
128 | * We look for count number of blocks in the buddy cache. If we were able | |
129 | * to locate that many free blocks we return with additional information | |
130 | * regarding rest of the contiguous physical block available | |
131 | * | |
132 | * Before allocating blocks via buddy cache we normalize the request | |
133 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
134 | * blocks that we get after allocation is added to the respective prealloc | |
135 | * list. In case of inode preallocation we follow a list of heuristics | |
136 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
137 | * we are doing a group prealloc we try to normalize the request to | |
27baebb8 TT |
138 | * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is |
139 | * dependent on the cluster size; for non-bigalloc file systems, it is | |
c9de560d | 140 | * 512 blocks. This can be tuned via |
d7a1fee1 | 141 | * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in |
c9de560d AT |
142 | * terms of number of blocks. If we have mounted the file system with -O |
143 | * stripe=<value> option the group prealloc request is normalized to the | |
d7a1fee1 DE |
144 | * the smallest multiple of the stripe value (sbi->s_stripe) which is |
145 | * greater than the default mb_group_prealloc. | |
c9de560d | 146 | * |
d7a1fee1 | 147 | * The regular allocator (using the buddy cache) supports a few tunables. |
c9de560d | 148 | * |
b713a5ec TT |
149 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
150 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
151 | * /sys/fs/ext4/<partition>/mb_order2_req | |
c9de560d | 152 | * |
b713a5ec | 153 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
154 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
155 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 156 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
af901ca1 | 157 | * stripe size (sbi->s_stripe), we try to search for contiguous block in |
b713a5ec TT |
158 | * stripe size. This should result in better allocation on RAID setups. If |
159 | * not, we search in the specific group using bitmap for best extents. The | |
160 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 161 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 162 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
163 | * best extent in the found extents. Searching for the blocks starts with |
164 | * the group specified as the goal value in allocation context via | |
165 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
caaf7a29 | 166 | * can be used for allocation. ext4_mb_good_group explains how the groups are |
c9de560d AT |
167 | * checked. |
168 | * | |
169 | * Both the prealloc space are getting populated as above. So for the first | |
170 | * request we will hit the buddy cache which will result in this prealloc | |
171 | * space getting filled. The prealloc space is then later used for the | |
172 | * subsequent request. | |
173 | */ | |
174 | ||
175 | /* | |
176 | * mballoc operates on the following data: | |
177 | * - on-disk bitmap | |
178 | * - in-core buddy (actually includes buddy and bitmap) | |
179 | * - preallocation descriptors (PAs) | |
180 | * | |
181 | * there are two types of preallocations: | |
182 | * - inode | |
183 | * assiged to specific inode and can be used for this inode only. | |
184 | * it describes part of inode's space preallocated to specific | |
185 | * physical blocks. any block from that preallocated can be used | |
186 | * independent. the descriptor just tracks number of blocks left | |
187 | * unused. so, before taking some block from descriptor, one must | |
188 | * make sure corresponded logical block isn't allocated yet. this | |
189 | * also means that freeing any block within descriptor's range | |
190 | * must discard all preallocated blocks. | |
191 | * - locality group | |
192 | * assigned to specific locality group which does not translate to | |
193 | * permanent set of inodes: inode can join and leave group. space | |
194 | * from this type of preallocation can be used for any inode. thus | |
195 | * it's consumed from the beginning to the end. | |
196 | * | |
197 | * relation between them can be expressed as: | |
198 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
199 | * | |
200 | * this mean blocks mballoc considers used are: | |
201 | * - allocated blocks (persistent) | |
202 | * - preallocated blocks (non-persistent) | |
203 | * | |
204 | * consistency in mballoc world means that at any time a block is either | |
205 | * free or used in ALL structures. notice: "any time" should not be read | |
206 | * literally -- time is discrete and delimited by locks. | |
207 | * | |
208 | * to keep it simple, we don't use block numbers, instead we count number of | |
209 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
210 | * | |
211 | * all operations can be expressed as: | |
212 | * - init buddy: buddy = on-disk + PAs | |
213 | * - new PA: buddy += N; PA = N | |
214 | * - use inode PA: on-disk += N; PA -= N | |
215 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
216 | * - use locality group PA on-disk += N; PA -= N | |
217 | * - discard locality group PA buddy -= PA; PA = 0 | |
218 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
219 | * is used in real operation because we can't know actual used | |
220 | * bits from PA, only from on-disk bitmap | |
221 | * | |
222 | * if we follow this strict logic, then all operations above should be atomic. | |
223 | * given some of them can block, we'd have to use something like semaphores | |
224 | * killing performance on high-end SMP hardware. let's try to relax it using | |
225 | * the following knowledge: | |
226 | * 1) if buddy is referenced, it's already initialized | |
227 | * 2) while block is used in buddy and the buddy is referenced, | |
228 | * nobody can re-allocate that block | |
229 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
230 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
231 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
232 | * block | |
233 | * | |
234 | * so, now we're building a concurrency table: | |
235 | * - init buddy vs. | |
236 | * - new PA | |
237 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
238 | * until PA is linked to allocation group to avoid concurrent buddy init | |
239 | * - use inode PA | |
240 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
241 | * given (3) we care that PA-=N operation doesn't interfere with init | |
242 | * - discard inode PA | |
243 | * the simplest way would be to have buddy initialized by the discard | |
244 | * - use locality group PA | |
245 | * again PA-=N must be serialized with init | |
246 | * - discard locality group PA | |
247 | * the simplest way would be to have buddy initialized by the discard | |
248 | * - new PA vs. | |
249 | * - use inode PA | |
250 | * i_data_sem serializes them | |
251 | * - discard inode PA | |
252 | * discard process must wait until PA isn't used by another process | |
253 | * - use locality group PA | |
254 | * some mutex should serialize them | |
255 | * - discard locality group PA | |
256 | * discard process must wait until PA isn't used by another process | |
257 | * - use inode PA | |
258 | * - use inode PA | |
259 | * i_data_sem or another mutex should serializes them | |
260 | * - discard inode PA | |
261 | * discard process must wait until PA isn't used by another process | |
262 | * - use locality group PA | |
263 | * nothing wrong here -- they're different PAs covering different blocks | |
264 | * - discard locality group PA | |
265 | * discard process must wait until PA isn't used by another process | |
266 | * | |
267 | * now we're ready to make few consequences: | |
268 | * - PA is referenced and while it is no discard is possible | |
269 | * - PA is referenced until block isn't marked in on-disk bitmap | |
270 | * - PA changes only after on-disk bitmap | |
271 | * - discard must not compete with init. either init is done before | |
272 | * any discard or they're serialized somehow | |
273 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
274 | * | |
275 | * a special case when we've used PA to emptiness. no need to modify buddy | |
276 | * in this case, but we should care about concurrent init | |
277 | * | |
278 | */ | |
279 | ||
280 | /* | |
281 | * Logic in few words: | |
282 | * | |
283 | * - allocation: | |
284 | * load group | |
285 | * find blocks | |
286 | * mark bits in on-disk bitmap | |
287 | * release group | |
288 | * | |
289 | * - use preallocation: | |
290 | * find proper PA (per-inode or group) | |
291 | * load group | |
292 | * mark bits in on-disk bitmap | |
293 | * release group | |
294 | * release PA | |
295 | * | |
296 | * - free: | |
297 | * load group | |
298 | * mark bits in on-disk bitmap | |
299 | * release group | |
300 | * | |
301 | * - discard preallocations in group: | |
302 | * mark PAs deleted | |
303 | * move them onto local list | |
304 | * load on-disk bitmap | |
305 | * load group | |
306 | * remove PA from object (inode or locality group) | |
307 | * mark free blocks in-core | |
308 | * | |
309 | * - discard inode's preallocations: | |
310 | */ | |
311 | ||
312 | /* | |
313 | * Locking rules | |
314 | * | |
315 | * Locks: | |
316 | * - bitlock on a group (group) | |
317 | * - object (inode/locality) (object) | |
318 | * - per-pa lock (pa) | |
319 | * | |
320 | * Paths: | |
321 | * - new pa | |
322 | * object | |
323 | * group | |
324 | * | |
325 | * - find and use pa: | |
326 | * pa | |
327 | * | |
328 | * - release consumed pa: | |
329 | * pa | |
330 | * group | |
331 | * object | |
332 | * | |
333 | * - generate in-core bitmap: | |
334 | * group | |
335 | * pa | |
336 | * | |
337 | * - discard all for given object (inode, locality group): | |
338 | * object | |
339 | * pa | |
340 | * group | |
341 | * | |
342 | * - discard all for given group: | |
343 | * group | |
344 | * pa | |
345 | * group | |
346 | * object | |
347 | * | |
348 | */ | |
c3a326a6 AK |
349 | static struct kmem_cache *ext4_pspace_cachep; |
350 | static struct kmem_cache *ext4_ac_cachep; | |
18aadd47 | 351 | static struct kmem_cache *ext4_free_data_cachep; |
fb1813f4 CW |
352 | |
353 | /* We create slab caches for groupinfo data structures based on the | |
354 | * superblock block size. There will be one per mounted filesystem for | |
355 | * each unique s_blocksize_bits */ | |
2892c15d | 356 | #define NR_GRPINFO_CACHES 8 |
fb1813f4 CW |
357 | static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; |
358 | ||
2892c15d ES |
359 | static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { |
360 | "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", | |
361 | "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", | |
362 | "ext4_groupinfo_64k", "ext4_groupinfo_128k" | |
363 | }; | |
364 | ||
c3a326a6 AK |
365 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
366 | ext4_group_t group); | |
7a2fcbf7 AK |
367 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
368 | ext4_group_t group); | |
18aadd47 BJ |
369 | static void ext4_free_data_callback(struct super_block *sb, |
370 | struct ext4_journal_cb_entry *jce, int rc); | |
c3a326a6 | 371 | |
ffad0a44 AK |
372 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
373 | { | |
c9de560d | 374 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
375 | *bit += ((unsigned long) addr & 7UL) << 3; |
376 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 377 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
378 | *bit += ((unsigned long) addr & 3UL) << 3; |
379 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
380 | #else |
381 | #error "how many bits you are?!" | |
382 | #endif | |
ffad0a44 AK |
383 | return addr; |
384 | } | |
c9de560d AT |
385 | |
386 | static inline int mb_test_bit(int bit, void *addr) | |
387 | { | |
388 | /* | |
389 | * ext4_test_bit on architecture like powerpc | |
390 | * needs unsigned long aligned address | |
391 | */ | |
ffad0a44 | 392 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
393 | return ext4_test_bit(bit, addr); |
394 | } | |
395 | ||
396 | static inline void mb_set_bit(int bit, void *addr) | |
397 | { | |
ffad0a44 | 398 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
399 | ext4_set_bit(bit, addr); |
400 | } | |
401 | ||
c9de560d AT |
402 | static inline void mb_clear_bit(int bit, void *addr) |
403 | { | |
ffad0a44 | 404 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
405 | ext4_clear_bit(bit, addr); |
406 | } | |
407 | ||
eabe0444 AS |
408 | static inline int mb_test_and_clear_bit(int bit, void *addr) |
409 | { | |
410 | addr = mb_correct_addr_and_bit(&bit, addr); | |
411 | return ext4_test_and_clear_bit(bit, addr); | |
412 | } | |
413 | ||
ffad0a44 AK |
414 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
415 | { | |
e7dfb246 | 416 | int fix = 0, ret, tmpmax; |
ffad0a44 | 417 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 418 | tmpmax = max + fix; |
ffad0a44 AK |
419 | start += fix; |
420 | ||
e7dfb246 AK |
421 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
422 | if (ret > max) | |
423 | return max; | |
424 | return ret; | |
ffad0a44 AK |
425 | } |
426 | ||
427 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
428 | { | |
e7dfb246 | 429 | int fix = 0, ret, tmpmax; |
ffad0a44 | 430 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 431 | tmpmax = max + fix; |
ffad0a44 AK |
432 | start += fix; |
433 | ||
e7dfb246 AK |
434 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
435 | if (ret > max) | |
436 | return max; | |
437 | return ret; | |
ffad0a44 AK |
438 | } |
439 | ||
c9de560d AT |
440 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
441 | { | |
442 | char *bb; | |
443 | ||
c5e8f3f3 | 444 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
445 | BUG_ON(max == NULL); |
446 | ||
447 | if (order > e4b->bd_blkbits + 1) { | |
448 | *max = 0; | |
449 | return NULL; | |
450 | } | |
451 | ||
452 | /* at order 0 we see each particular block */ | |
84b775a3 CL |
453 | if (order == 0) { |
454 | *max = 1 << (e4b->bd_blkbits + 3); | |
c5e8f3f3 | 455 | return e4b->bd_bitmap; |
84b775a3 | 456 | } |
c9de560d | 457 | |
c5e8f3f3 | 458 | bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; |
c9de560d AT |
459 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; |
460 | ||
461 | return bb; | |
462 | } | |
463 | ||
464 | #ifdef DOUBLE_CHECK | |
465 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
466 | int first, int count) | |
467 | { | |
468 | int i; | |
469 | struct super_block *sb = e4b->bd_sb; | |
470 | ||
471 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
472 | return; | |
bc8e6740 | 473 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
474 | for (i = 0; i < count; i++) { |
475 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
476 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
477 | |
478 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 479 | blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
5d1b1b3f | 480 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
481 | inode ? inode->i_ino : 0, |
482 | blocknr, | |
483 | "freeing block already freed " | |
484 | "(bit %u)", | |
485 | first + i); | |
c9de560d AT |
486 | } |
487 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
488 | } | |
489 | } | |
490 | ||
491 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
492 | { | |
493 | int i; | |
494 | ||
495 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
496 | return; | |
bc8e6740 | 497 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
498 | for (i = 0; i < count; i++) { |
499 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
500 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
501 | } | |
502 | } | |
503 | ||
504 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
505 | { | |
506 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { | |
507 | unsigned char *b1, *b2; | |
508 | int i; | |
509 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
510 | b2 = (unsigned char *) bitmap; | |
511 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
512 | if (b1[i] != b2[i]) { | |
9d8b9ec4 TT |
513 | ext4_msg(e4b->bd_sb, KERN_ERR, |
514 | "corruption in group %u " | |
515 | "at byte %u(%u): %x in copy != %x " | |
516 | "on disk/prealloc", | |
517 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
518 | BUG(); |
519 | } | |
520 | } | |
521 | } | |
522 | } | |
523 | ||
524 | #else | |
525 | static inline void mb_free_blocks_double(struct inode *inode, | |
526 | struct ext4_buddy *e4b, int first, int count) | |
527 | { | |
528 | return; | |
529 | } | |
530 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
531 | int first, int count) | |
532 | { | |
533 | return; | |
534 | } | |
535 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
536 | { | |
537 | return; | |
538 | } | |
539 | #endif | |
540 | ||
541 | #ifdef AGGRESSIVE_CHECK | |
542 | ||
543 | #define MB_CHECK_ASSERT(assert) \ | |
544 | do { \ | |
545 | if (!(assert)) { \ | |
546 | printk(KERN_EMERG \ | |
547 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
548 | function, file, line, # assert); \ | |
549 | BUG(); \ | |
550 | } \ | |
551 | } while (0) | |
552 | ||
553 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
554 | const char *function, int line) | |
555 | { | |
556 | struct super_block *sb = e4b->bd_sb; | |
557 | int order = e4b->bd_blkbits + 1; | |
558 | int max; | |
559 | int max2; | |
560 | int i; | |
561 | int j; | |
562 | int k; | |
563 | int count; | |
564 | struct ext4_group_info *grp; | |
565 | int fragments = 0; | |
566 | int fstart; | |
567 | struct list_head *cur; | |
568 | void *buddy; | |
569 | void *buddy2; | |
570 | ||
c9de560d AT |
571 | { |
572 | static int mb_check_counter; | |
573 | if (mb_check_counter++ % 100 != 0) | |
574 | return 0; | |
575 | } | |
576 | ||
577 | while (order > 1) { | |
578 | buddy = mb_find_buddy(e4b, order, &max); | |
579 | MB_CHECK_ASSERT(buddy); | |
580 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
581 | MB_CHECK_ASSERT(buddy2); | |
582 | MB_CHECK_ASSERT(buddy != buddy2); | |
583 | MB_CHECK_ASSERT(max * 2 == max2); | |
584 | ||
585 | count = 0; | |
586 | for (i = 0; i < max; i++) { | |
587 | ||
588 | if (mb_test_bit(i, buddy)) { | |
589 | /* only single bit in buddy2 may be 1 */ | |
590 | if (!mb_test_bit(i << 1, buddy2)) { | |
591 | MB_CHECK_ASSERT( | |
592 | mb_test_bit((i<<1)+1, buddy2)); | |
593 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
594 | MB_CHECK_ASSERT( | |
595 | mb_test_bit(i << 1, buddy2)); | |
596 | } | |
597 | continue; | |
598 | } | |
599 | ||
0a10da73 | 600 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
601 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
602 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
603 | ||
604 | for (j = 0; j < (1 << order); j++) { | |
605 | k = (i * (1 << order)) + j; | |
606 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 607 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
608 | } |
609 | count++; | |
610 | } | |
611 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
612 | order--; | |
613 | } | |
614 | ||
615 | fstart = -1; | |
616 | buddy = mb_find_buddy(e4b, 0, &max); | |
617 | for (i = 0; i < max; i++) { | |
618 | if (!mb_test_bit(i, buddy)) { | |
619 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
620 | if (fstart == -1) { | |
621 | fragments++; | |
622 | fstart = i; | |
623 | } | |
624 | continue; | |
625 | } | |
626 | fstart = -1; | |
627 | /* check used bits only */ | |
628 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
629 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
630 | k = i >> j; | |
631 | MB_CHECK_ASSERT(k < max2); | |
632 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
633 | } | |
634 | } | |
635 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
636 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
637 | ||
638 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
639 | list_for_each(cur, &grp->bb_prealloc_list) { |
640 | ext4_group_t groupnr; | |
641 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
642 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
643 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 644 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 645 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
646 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
647 | } | |
648 | return 0; | |
649 | } | |
650 | #undef MB_CHECK_ASSERT | |
651 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 652 | __FILE__, __func__, __LINE__) |
c9de560d AT |
653 | #else |
654 | #define mb_check_buddy(e4b) | |
655 | #endif | |
656 | ||
7c786059 CL |
657 | /* |
658 | * Divide blocks started from @first with length @len into | |
659 | * smaller chunks with power of 2 blocks. | |
660 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
661 | * then increase bb_counters[] for corresponded chunk size. | |
662 | */ | |
c9de560d | 663 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 664 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
665 | struct ext4_group_info *grp) |
666 | { | |
667 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
668 | ext4_grpblk_t min; |
669 | ext4_grpblk_t max; | |
670 | ext4_grpblk_t chunk; | |
c9de560d AT |
671 | unsigned short border; |
672 | ||
7137d7a4 | 673 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
674 | |
675 | border = 2 << sb->s_blocksize_bits; | |
676 | ||
677 | while (len > 0) { | |
678 | /* find how many blocks can be covered since this position */ | |
679 | max = ffs(first | border) - 1; | |
680 | ||
681 | /* find how many blocks of power 2 we need to mark */ | |
682 | min = fls(len) - 1; | |
683 | ||
684 | if (max < min) | |
685 | min = max; | |
686 | chunk = 1 << min; | |
687 | ||
688 | /* mark multiblock chunks only */ | |
689 | grp->bb_counters[min]++; | |
690 | if (min > 0) | |
691 | mb_clear_bit(first >> min, | |
692 | buddy + sbi->s_mb_offsets[min]); | |
693 | ||
694 | len -= chunk; | |
695 | first += chunk; | |
696 | } | |
697 | } | |
698 | ||
8a57d9d6 CW |
699 | /* |
700 | * Cache the order of the largest free extent we have available in this block | |
701 | * group. | |
702 | */ | |
703 | static void | |
704 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
705 | { | |
706 | int i; | |
707 | int bits; | |
708 | ||
709 | grp->bb_largest_free_order = -1; /* uninit */ | |
710 | ||
711 | bits = sb->s_blocksize_bits + 1; | |
712 | for (i = bits; i >= 0; i--) { | |
713 | if (grp->bb_counters[i] > 0) { | |
714 | grp->bb_largest_free_order = i; | |
715 | break; | |
716 | } | |
717 | } | |
718 | } | |
719 | ||
089ceecc ES |
720 | static noinline_for_stack |
721 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
722 | void *buddy, void *bitmap, ext4_group_t group) |
723 | { | |
724 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
7137d7a4 | 725 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
726 | ext4_grpblk_t i = 0; |
727 | ext4_grpblk_t first; | |
728 | ext4_grpblk_t len; | |
c9de560d AT |
729 | unsigned free = 0; |
730 | unsigned fragments = 0; | |
731 | unsigned long long period = get_cycles(); | |
732 | ||
733 | /* initialize buddy from bitmap which is aggregation | |
734 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 735 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
736 | grp->bb_first_free = i; |
737 | while (i < max) { | |
738 | fragments++; | |
739 | first = i; | |
ffad0a44 | 740 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
741 | len = i - first; |
742 | free += len; | |
743 | if (len > 1) | |
744 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
745 | else | |
746 | grp->bb_counters[0]++; | |
747 | if (i < max) | |
ffad0a44 | 748 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
749 | } |
750 | grp->bb_fragments = fragments; | |
751 | ||
752 | if (free != grp->bb_free) { | |
e29136f8 | 753 | ext4_grp_locked_error(sb, group, 0, 0, |
163a203d DW |
754 | "%u clusters in bitmap, %u in gd; " |
755 | "block bitmap corrupt.", | |
e29136f8 | 756 | free, grp->bb_free); |
e56eb659 | 757 | /* |
163a203d | 758 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
759 | * corrupt and update bb_free using bitmap value |
760 | */ | |
c9de560d | 761 | grp->bb_free = free; |
163a203d | 762 | set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state); |
c9de560d | 763 | } |
8a57d9d6 | 764 | mb_set_largest_free_order(sb, grp); |
c9de560d AT |
765 | |
766 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
767 | ||
768 | period = get_cycles() - period; | |
769 | spin_lock(&EXT4_SB(sb)->s_bal_lock); | |
770 | EXT4_SB(sb)->s_mb_buddies_generated++; | |
771 | EXT4_SB(sb)->s_mb_generation_time += period; | |
772 | spin_unlock(&EXT4_SB(sb)->s_bal_lock); | |
773 | } | |
774 | ||
eabe0444 AS |
775 | static void mb_regenerate_buddy(struct ext4_buddy *e4b) |
776 | { | |
777 | int count; | |
778 | int order = 1; | |
779 | void *buddy; | |
780 | ||
781 | while ((buddy = mb_find_buddy(e4b, order++, &count))) { | |
782 | ext4_set_bits(buddy, 0, count); | |
783 | } | |
784 | e4b->bd_info->bb_fragments = 0; | |
785 | memset(e4b->bd_info->bb_counters, 0, | |
786 | sizeof(*e4b->bd_info->bb_counters) * | |
787 | (e4b->bd_sb->s_blocksize_bits + 2)); | |
788 | ||
789 | ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, | |
790 | e4b->bd_bitmap, e4b->bd_group); | |
791 | } | |
792 | ||
c9de560d AT |
793 | /* The buddy information is attached the buddy cache inode |
794 | * for convenience. The information regarding each group | |
795 | * is loaded via ext4_mb_load_buddy. The information involve | |
796 | * block bitmap and buddy information. The information are | |
797 | * stored in the inode as | |
798 | * | |
799 | * { page } | |
c3a326a6 | 800 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
801 | * |
802 | * | |
803 | * one block each for bitmap and buddy information. | |
804 | * So for each group we take up 2 blocks. A page can | |
805 | * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. | |
806 | * So it can have information regarding groups_per_page which | |
807 | * is blocks_per_page/2 | |
8a57d9d6 CW |
808 | * |
809 | * Locking note: This routine takes the block group lock of all groups | |
810 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
811 | */ |
812 | ||
813 | static int ext4_mb_init_cache(struct page *page, char *incore) | |
814 | { | |
8df9675f | 815 | ext4_group_t ngroups; |
c9de560d AT |
816 | int blocksize; |
817 | int blocks_per_page; | |
818 | int groups_per_page; | |
819 | int err = 0; | |
820 | int i; | |
813e5727 | 821 | ext4_group_t first_group, group; |
c9de560d AT |
822 | int first_block; |
823 | struct super_block *sb; | |
824 | struct buffer_head *bhs; | |
fa77dcfa | 825 | struct buffer_head **bh = NULL; |
c9de560d AT |
826 | struct inode *inode; |
827 | char *data; | |
828 | char *bitmap; | |
9b8b7d35 | 829 | struct ext4_group_info *grinfo; |
c9de560d | 830 | |
6ba495e9 | 831 | mb_debug(1, "init page %lu\n", page->index); |
c9de560d AT |
832 | |
833 | inode = page->mapping->host; | |
834 | sb = inode->i_sb; | |
8df9675f | 835 | ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
836 | blocksize = 1 << inode->i_blkbits; |
837 | blocks_per_page = PAGE_CACHE_SIZE / blocksize; | |
838 | ||
839 | groups_per_page = blocks_per_page >> 1; | |
840 | if (groups_per_page == 0) | |
841 | groups_per_page = 1; | |
842 | ||
843 | /* allocate buffer_heads to read bitmaps */ | |
844 | if (groups_per_page > 1) { | |
c9de560d AT |
845 | i = sizeof(struct buffer_head *) * groups_per_page; |
846 | bh = kzalloc(i, GFP_NOFS); | |
813e5727 TT |
847 | if (bh == NULL) { |
848 | err = -ENOMEM; | |
c9de560d | 849 | goto out; |
813e5727 | 850 | } |
c9de560d AT |
851 | } else |
852 | bh = &bhs; | |
853 | ||
854 | first_group = page->index * blocks_per_page / 2; | |
855 | ||
856 | /* read all groups the page covers into the cache */ | |
813e5727 TT |
857 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
858 | if (group >= ngroups) | |
c9de560d AT |
859 | break; |
860 | ||
813e5727 | 861 | grinfo = ext4_get_group_info(sb, group); |
9b8b7d35 AG |
862 | /* |
863 | * If page is uptodate then we came here after online resize | |
864 | * which added some new uninitialized group info structs, so | |
865 | * we must skip all initialized uptodate buddies on the page, | |
866 | * which may be currently in use by an allocating task. | |
867 | */ | |
868 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
869 | bh[i] = NULL; | |
870 | continue; | |
871 | } | |
813e5727 TT |
872 | if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) { |
873 | err = -ENOMEM; | |
c9de560d | 874 | goto out; |
2ccb5fb9 | 875 | } |
813e5727 | 876 | mb_debug(1, "read bitmap for group %u\n", group); |
c9de560d AT |
877 | } |
878 | ||
879 | /* wait for I/O completion */ | |
813e5727 TT |
880 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
881 | if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) { | |
882 | err = -EIO; | |
c9de560d | 883 | goto out; |
813e5727 TT |
884 | } |
885 | } | |
c9de560d AT |
886 | |
887 | first_block = page->index * blocks_per_page; | |
888 | for (i = 0; i < blocks_per_page; i++) { | |
c9de560d | 889 | group = (first_block + i) >> 1; |
8df9675f | 890 | if (group >= ngroups) |
c9de560d AT |
891 | break; |
892 | ||
9b8b7d35 AG |
893 | if (!bh[group - first_group]) |
894 | /* skip initialized uptodate buddy */ | |
895 | continue; | |
896 | ||
c9de560d AT |
897 | /* |
898 | * data carry information regarding this | |
899 | * particular group in the format specified | |
900 | * above | |
901 | * | |
902 | */ | |
903 | data = page_address(page) + (i * blocksize); | |
904 | bitmap = bh[group - first_group]->b_data; | |
905 | ||
906 | /* | |
907 | * We place the buddy block and bitmap block | |
908 | * close together | |
909 | */ | |
910 | if ((first_block + i) & 1) { | |
911 | /* this is block of buddy */ | |
912 | BUG_ON(incore == NULL); | |
6ba495e9 | 913 | mb_debug(1, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 914 | group, page->index, i * blocksize); |
f307333e | 915 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
916 | grinfo = ext4_get_group_info(sb, group); |
917 | grinfo->bb_fragments = 0; | |
918 | memset(grinfo->bb_counters, 0, | |
1927805e ES |
919 | sizeof(*grinfo->bb_counters) * |
920 | (sb->s_blocksize_bits+2)); | |
c9de560d AT |
921 | /* |
922 | * incore got set to the group block bitmap below | |
923 | */ | |
7a2fcbf7 | 924 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
925 | /* init the buddy */ |
926 | memset(data, 0xff, blocksize); | |
c9de560d | 927 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 928 | ext4_unlock_group(sb, group); |
c9de560d AT |
929 | incore = NULL; |
930 | } else { | |
931 | /* this is block of bitmap */ | |
932 | BUG_ON(incore != NULL); | |
6ba495e9 | 933 | mb_debug(1, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 934 | group, page->index, i * blocksize); |
f307333e | 935 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
936 | |
937 | /* see comments in ext4_mb_put_pa() */ | |
938 | ext4_lock_group(sb, group); | |
939 | memcpy(data, bitmap, blocksize); | |
940 | ||
941 | /* mark all preallocated blks used in in-core bitmap */ | |
942 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 943 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
944 | ext4_unlock_group(sb, group); |
945 | ||
946 | /* set incore so that the buddy information can be | |
947 | * generated using this | |
948 | */ | |
949 | incore = data; | |
950 | } | |
951 | } | |
952 | SetPageUptodate(page); | |
953 | ||
954 | out: | |
955 | if (bh) { | |
9b8b7d35 | 956 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
957 | brelse(bh[i]); |
958 | if (bh != &bhs) | |
959 | kfree(bh); | |
960 | } | |
961 | return err; | |
962 | } | |
963 | ||
eee4adc7 | 964 | /* |
2de8807b AG |
965 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
966 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
967 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
968 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 969 | */ |
2de8807b AG |
970 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
971 | ext4_group_t group, struct ext4_buddy *e4b) | |
eee4adc7 | 972 | { |
2de8807b AG |
973 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
974 | int block, pnum, poff; | |
eee4adc7 | 975 | int blocks_per_page; |
2de8807b AG |
976 | struct page *page; |
977 | ||
978 | e4b->bd_buddy_page = NULL; | |
979 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 ES |
980 | |
981 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
982 | /* | |
983 | * the buddy cache inode stores the block bitmap | |
984 | * and buddy information in consecutive blocks. | |
985 | * So for each group we need two blocks. | |
986 | */ | |
987 | block = group * 2; | |
988 | pnum = block / blocks_per_page; | |
2de8807b AG |
989 | poff = block % blocks_per_page; |
990 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
991 | if (!page) | |
c57ab39b | 992 | return -ENOMEM; |
2de8807b AG |
993 | BUG_ON(page->mapping != inode->i_mapping); |
994 | e4b->bd_bitmap_page = page; | |
995 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
996 | ||
997 | if (blocks_per_page >= 2) { | |
998 | /* buddy and bitmap are on the same page */ | |
999 | return 0; | |
eee4adc7 | 1000 | } |
2de8807b AG |
1001 | |
1002 | block++; | |
1003 | pnum = block / blocks_per_page; | |
2de8807b AG |
1004 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); |
1005 | if (!page) | |
c57ab39b | 1006 | return -ENOMEM; |
2de8807b AG |
1007 | BUG_ON(page->mapping != inode->i_mapping); |
1008 | e4b->bd_buddy_page = page; | |
1009 | return 0; | |
eee4adc7 ES |
1010 | } |
1011 | ||
2de8807b | 1012 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1013 | { |
2de8807b AG |
1014 | if (e4b->bd_bitmap_page) { |
1015 | unlock_page(e4b->bd_bitmap_page); | |
1016 | page_cache_release(e4b->bd_bitmap_page); | |
1017 | } | |
1018 | if (e4b->bd_buddy_page) { | |
1019 | unlock_page(e4b->bd_buddy_page); | |
1020 | page_cache_release(e4b->bd_buddy_page); | |
eee4adc7 | 1021 | } |
eee4adc7 ES |
1022 | } |
1023 | ||
8a57d9d6 CW |
1024 | /* |
1025 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1026 | * block group lock of all groups for this page; do not hold the BG lock when | |
1027 | * calling this routine! | |
1028 | */ | |
b6a758ec AK |
1029 | static noinline_for_stack |
1030 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) | |
1031 | { | |
1032 | ||
b6a758ec | 1033 | struct ext4_group_info *this_grp; |
2de8807b AG |
1034 | struct ext4_buddy e4b; |
1035 | struct page *page; | |
1036 | int ret = 0; | |
b6a758ec | 1037 | |
b10a44c3 | 1038 | might_sleep(); |
b6a758ec | 1039 | mb_debug(1, "init group %u\n", group); |
b6a758ec AK |
1040 | this_grp = ext4_get_group_info(sb, group); |
1041 | /* | |
08c3a813 AK |
1042 | * This ensures that we don't reinit the buddy cache |
1043 | * page which map to the group from which we are already | |
1044 | * allocating. If we are looking at the buddy cache we would | |
1045 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1046 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1047 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1048 | * page accessed. | |
b6a758ec | 1049 | */ |
2de8807b AG |
1050 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b); |
1051 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { | |
b6a758ec AK |
1052 | /* |
1053 | * somebody initialized the group | |
1054 | * return without doing anything | |
1055 | */ | |
b6a758ec AK |
1056 | goto err; |
1057 | } | |
2de8807b AG |
1058 | |
1059 | page = e4b.bd_bitmap_page; | |
1060 | ret = ext4_mb_init_cache(page, NULL); | |
1061 | if (ret) | |
1062 | goto err; | |
1063 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1064 | ret = -EIO; |
1065 | goto err; | |
1066 | } | |
b6a758ec | 1067 | |
2de8807b | 1068 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1069 | /* |
1070 | * If both the bitmap and buddy are in | |
1071 | * the same page we don't need to force | |
1072 | * init the buddy | |
1073 | */ | |
2de8807b AG |
1074 | ret = 0; |
1075 | goto err; | |
b6a758ec | 1076 | } |
2de8807b AG |
1077 | /* init buddy cache */ |
1078 | page = e4b.bd_buddy_page; | |
1079 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap); | |
1080 | if (ret) | |
1081 | goto err; | |
1082 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1083 | ret = -EIO; |
1084 | goto err; | |
1085 | } | |
b6a758ec | 1086 | err: |
2de8807b | 1087 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1088 | return ret; |
1089 | } | |
1090 | ||
8a57d9d6 CW |
1091 | /* |
1092 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1093 | * block group lock of all groups for this page; do not hold the BG lock when | |
1094 | * calling this routine! | |
1095 | */ | |
4ddfef7b ES |
1096 | static noinline_for_stack int |
1097 | ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, | |
1098 | struct ext4_buddy *e4b) | |
c9de560d | 1099 | { |
c9de560d AT |
1100 | int blocks_per_page; |
1101 | int block; | |
1102 | int pnum; | |
1103 | int poff; | |
1104 | struct page *page; | |
fdf6c7a7 | 1105 | int ret; |
920313a7 AK |
1106 | struct ext4_group_info *grp; |
1107 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1108 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1109 | |
b10a44c3 | 1110 | might_sleep(); |
6ba495e9 | 1111 | mb_debug(1, "load group %u\n", group); |
c9de560d AT |
1112 | |
1113 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
920313a7 | 1114 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1115 | |
1116 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1117 | e4b->bd_info = grp; |
c9de560d AT |
1118 | e4b->bd_sb = sb; |
1119 | e4b->bd_group = group; | |
1120 | e4b->bd_buddy_page = NULL; | |
1121 | e4b->bd_bitmap_page = NULL; | |
1122 | ||
f41c0750 | 1123 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1124 | /* |
1125 | * we need full data about the group | |
1126 | * to make a good selection | |
1127 | */ | |
1128 | ret = ext4_mb_init_group(sb, group); | |
1129 | if (ret) | |
1130 | return ret; | |
f41c0750 AK |
1131 | } |
1132 | ||
c9de560d AT |
1133 | /* |
1134 | * the buddy cache inode stores the block bitmap | |
1135 | * and buddy information in consecutive blocks. | |
1136 | * So for each group we need two blocks. | |
1137 | */ | |
1138 | block = group * 2; | |
1139 | pnum = block / blocks_per_page; | |
1140 | poff = block % blocks_per_page; | |
1141 | ||
1142 | /* we could use find_or_create_page(), but it locks page | |
1143 | * what we'd like to avoid in fast path ... */ | |
2457aec6 | 1144 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1145 | if (page == NULL || !PageUptodate(page)) { |
1146 | if (page) | |
920313a7 AK |
1147 | /* |
1148 | * drop the page reference and try | |
1149 | * to get the page with lock. If we | |
1150 | * are not uptodate that implies | |
1151 | * somebody just created the page but | |
1152 | * is yet to initialize the same. So | |
1153 | * wait for it to initialize. | |
1154 | */ | |
c9de560d AT |
1155 | page_cache_release(page); |
1156 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1157 | if (page) { | |
1158 | BUG_ON(page->mapping != inode->i_mapping); | |
1159 | if (!PageUptodate(page)) { | |
fdf6c7a7 SF |
1160 | ret = ext4_mb_init_cache(page, NULL); |
1161 | if (ret) { | |
1162 | unlock_page(page); | |
1163 | goto err; | |
1164 | } | |
c9de560d AT |
1165 | mb_cmp_bitmaps(e4b, page_address(page) + |
1166 | (poff * sb->s_blocksize)); | |
1167 | } | |
1168 | unlock_page(page); | |
1169 | } | |
1170 | } | |
c57ab39b YL |
1171 | if (page == NULL) { |
1172 | ret = -ENOMEM; | |
1173 | goto err; | |
1174 | } | |
1175 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1176 | ret = -EIO; |
c9de560d | 1177 | goto err; |
fdf6c7a7 | 1178 | } |
2457aec6 MG |
1179 | |
1180 | /* Pages marked accessed already */ | |
c9de560d AT |
1181 | e4b->bd_bitmap_page = page; |
1182 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1183 | |
1184 | block++; | |
1185 | pnum = block / blocks_per_page; | |
1186 | poff = block % blocks_per_page; | |
1187 | ||
2457aec6 | 1188 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1189 | if (page == NULL || !PageUptodate(page)) { |
1190 | if (page) | |
1191 | page_cache_release(page); | |
1192 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1193 | if (page) { | |
1194 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 SF |
1195 | if (!PageUptodate(page)) { |
1196 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap); | |
1197 | if (ret) { | |
1198 | unlock_page(page); | |
1199 | goto err; | |
1200 | } | |
1201 | } | |
c9de560d AT |
1202 | unlock_page(page); |
1203 | } | |
1204 | } | |
c57ab39b YL |
1205 | if (page == NULL) { |
1206 | ret = -ENOMEM; | |
1207 | goto err; | |
1208 | } | |
1209 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1210 | ret = -EIO; |
c9de560d | 1211 | goto err; |
fdf6c7a7 | 1212 | } |
2457aec6 MG |
1213 | |
1214 | /* Pages marked accessed already */ | |
c9de560d AT |
1215 | e4b->bd_buddy_page = page; |
1216 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1217 | |
1218 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1219 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1220 | ||
1221 | return 0; | |
1222 | ||
1223 | err: | |
26626f11 YR |
1224 | if (page) |
1225 | page_cache_release(page); | |
c9de560d AT |
1226 | if (e4b->bd_bitmap_page) |
1227 | page_cache_release(e4b->bd_bitmap_page); | |
1228 | if (e4b->bd_buddy_page) | |
1229 | page_cache_release(e4b->bd_buddy_page); | |
1230 | e4b->bd_buddy = NULL; | |
1231 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1232 | return ret; |
c9de560d AT |
1233 | } |
1234 | ||
e39e07fd | 1235 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1236 | { |
1237 | if (e4b->bd_bitmap_page) | |
1238 | page_cache_release(e4b->bd_bitmap_page); | |
1239 | if (e4b->bd_buddy_page) | |
1240 | page_cache_release(e4b->bd_buddy_page); | |
1241 | } | |
1242 | ||
1243 | ||
1244 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1245 | { | |
1246 | int order = 1; | |
1247 | void *bb; | |
1248 | ||
c5e8f3f3 | 1249 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1250 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1251 | ||
c5e8f3f3 | 1252 | bb = e4b->bd_buddy; |
c9de560d AT |
1253 | while (order <= e4b->bd_blkbits + 1) { |
1254 | block = block >> 1; | |
1255 | if (!mb_test_bit(block, bb)) { | |
1256 | /* this block is part of buddy of order 'order' */ | |
1257 | return order; | |
1258 | } | |
1259 | bb += 1 << (e4b->bd_blkbits - order); | |
1260 | order++; | |
1261 | } | |
1262 | return 0; | |
1263 | } | |
1264 | ||
955ce5f5 | 1265 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1266 | { |
1267 | __u32 *addr; | |
1268 | ||
1269 | len = cur + len; | |
1270 | while (cur < len) { | |
1271 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1272 | /* fast path: clear whole word at once */ | |
1273 | addr = bm + (cur >> 3); | |
1274 | *addr = 0; | |
1275 | cur += 32; | |
1276 | continue; | |
1277 | } | |
955ce5f5 | 1278 | mb_clear_bit(cur, bm); |
c9de560d AT |
1279 | cur++; |
1280 | } | |
1281 | } | |
1282 | ||
eabe0444 AS |
1283 | /* clear bits in given range |
1284 | * will return first found zero bit if any, -1 otherwise | |
1285 | */ | |
1286 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1287 | { | |
1288 | __u32 *addr; | |
1289 | int zero_bit = -1; | |
1290 | ||
1291 | len = cur + len; | |
1292 | while (cur < len) { | |
1293 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1294 | /* fast path: clear whole word at once */ | |
1295 | addr = bm + (cur >> 3); | |
1296 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1297 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1298 | *addr = 0; | |
1299 | cur += 32; | |
1300 | continue; | |
1301 | } | |
1302 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1303 | zero_bit = cur; | |
1304 | cur++; | |
1305 | } | |
1306 | ||
1307 | return zero_bit; | |
1308 | } | |
1309 | ||
c3e94d1d | 1310 | void ext4_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1311 | { |
1312 | __u32 *addr; | |
1313 | ||
1314 | len = cur + len; | |
1315 | while (cur < len) { | |
1316 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1317 | /* fast path: set whole word at once */ | |
1318 | addr = bm + (cur >> 3); | |
1319 | *addr = 0xffffffff; | |
1320 | cur += 32; | |
1321 | continue; | |
1322 | } | |
955ce5f5 | 1323 | mb_set_bit(cur, bm); |
c9de560d AT |
1324 | cur++; |
1325 | } | |
1326 | } | |
1327 | ||
eabe0444 AS |
1328 | /* |
1329 | * _________________________________________________________________ */ | |
1330 | ||
1331 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) | |
1332 | { | |
1333 | if (mb_test_bit(*bit + side, bitmap)) { | |
1334 | mb_clear_bit(*bit, bitmap); | |
1335 | (*bit) -= side; | |
1336 | return 1; | |
1337 | } | |
1338 | else { | |
1339 | (*bit) += side; | |
1340 | mb_set_bit(*bit, bitmap); | |
1341 | return -1; | |
1342 | } | |
1343 | } | |
1344 | ||
1345 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1346 | { | |
1347 | int max; | |
1348 | int order = 1; | |
1349 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1350 | ||
1351 | while (buddy) { | |
1352 | void *buddy2; | |
1353 | ||
1354 | /* Bits in range [first; last] are known to be set since | |
1355 | * corresponding blocks were allocated. Bits in range | |
1356 | * (first; last) will stay set because they form buddies on | |
1357 | * upper layer. We just deal with borders if they don't | |
1358 | * align with upper layer and then go up. | |
1359 | * Releasing entire group is all about clearing | |
1360 | * single bit of highest order buddy. | |
1361 | */ | |
1362 | ||
1363 | /* Example: | |
1364 | * --------------------------------- | |
1365 | * | 1 | 1 | 1 | 1 | | |
1366 | * --------------------------------- | |
1367 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1368 | * --------------------------------- | |
1369 | * 0 1 2 3 4 5 6 7 | |
1370 | * \_____________________/ | |
1371 | * | |
1372 | * Neither [1] nor [6] is aligned to above layer. | |
1373 | * Left neighbour [0] is free, so mark it busy, | |
1374 | * decrease bb_counters and extend range to | |
1375 | * [0; 6] | |
1376 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1377 | * mark [6] free, increase bb_counters and shrink range to | |
1378 | * [0; 5]. | |
1379 | * Then shift range to [0; 2], go up and do the same. | |
1380 | */ | |
1381 | ||
1382 | ||
1383 | if (first & 1) | |
1384 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1385 | if (!(last & 1)) | |
1386 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1387 | if (first > last) | |
1388 | break; | |
1389 | order++; | |
1390 | ||
1391 | if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { | |
1392 | mb_clear_bits(buddy, first, last - first + 1); | |
1393 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1394 | break; | |
1395 | } | |
1396 | first >>= 1; | |
1397 | last >>= 1; | |
1398 | buddy = buddy2; | |
1399 | } | |
1400 | } | |
1401 | ||
7e5a8cdd | 1402 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1403 | int first, int count) |
c9de560d | 1404 | { |
eabe0444 AS |
1405 | int left_is_free = 0; |
1406 | int right_is_free = 0; | |
1407 | int block; | |
1408 | int last = first + count - 1; | |
c9de560d AT |
1409 | struct super_block *sb = e4b->bd_sb; |
1410 | ||
eabe0444 | 1411 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1412 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1413 | /* Don't bother if the block group is corrupt. */ |
1414 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1415 | return; | |
1416 | ||
c9de560d AT |
1417 | mb_check_buddy(e4b); |
1418 | mb_free_blocks_double(inode, e4b, first, count); | |
1419 | ||
1420 | e4b->bd_info->bb_free += count; | |
1421 | if (first < e4b->bd_info->bb_first_free) | |
1422 | e4b->bd_info->bb_first_free = first; | |
1423 | ||
eabe0444 AS |
1424 | /* access memory sequentially: check left neighbour, |
1425 | * clear range and then check right neighbour | |
1426 | */ | |
c9de560d | 1427 | if (first != 0) |
eabe0444 AS |
1428 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1429 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1430 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1431 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1432 | ||
1433 | if (unlikely(block != -1)) { | |
1434 | ext4_fsblk_t blocknr; | |
1435 | ||
1436 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
1437 | blocknr += EXT4_C2B(EXT4_SB(sb), block); | |
1438 | ext4_grp_locked_error(sb, e4b->bd_group, | |
1439 | inode ? inode->i_ino : 0, | |
1440 | blocknr, | |
1441 | "freeing already freed block " | |
163a203d DW |
1442 | "(bit %u); block bitmap corrupt.", |
1443 | block); | |
1444 | /* Mark the block group as corrupt. */ | |
1445 | set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, | |
1446 | &e4b->bd_info->bb_state); | |
eabe0444 AS |
1447 | mb_regenerate_buddy(e4b); |
1448 | goto done; | |
1449 | } | |
1450 | ||
1451 | /* let's maintain fragments counter */ | |
1452 | if (left_is_free && right_is_free) | |
c9de560d | 1453 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1454 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1455 | e4b->bd_info->bb_fragments++; |
1456 | ||
eabe0444 AS |
1457 | /* buddy[0] == bd_bitmap is a special case, so handle |
1458 | * it right away and let mb_buddy_mark_free stay free of | |
1459 | * zero order checks. | |
1460 | * Check if neighbours are to be coaleasced, | |
1461 | * adjust bitmap bb_counters and borders appropriately. | |
1462 | */ | |
1463 | if (first & 1) { | |
1464 | first += !left_is_free; | |
1465 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1466 | } | |
1467 | if (!(last & 1)) { | |
1468 | last -= !right_is_free; | |
1469 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1470 | } | |
c9de560d | 1471 | |
eabe0444 AS |
1472 | if (first <= last) |
1473 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1474 | |
eabe0444 | 1475 | done: |
8a57d9d6 | 1476 | mb_set_largest_free_order(sb, e4b->bd_info); |
c9de560d | 1477 | mb_check_buddy(e4b); |
c9de560d AT |
1478 | } |
1479 | ||
15c006a2 | 1480 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1481 | int needed, struct ext4_free_extent *ex) |
1482 | { | |
1483 | int next = block; | |
15c006a2 | 1484 | int max, order; |
c9de560d AT |
1485 | void *buddy; |
1486 | ||
bc8e6740 | 1487 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1488 | BUG_ON(ex == NULL); |
1489 | ||
15c006a2 | 1490 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1491 | BUG_ON(buddy == NULL); |
1492 | BUG_ON(block >= max); | |
1493 | if (mb_test_bit(block, buddy)) { | |
1494 | ex->fe_len = 0; | |
1495 | ex->fe_start = 0; | |
1496 | ex->fe_group = 0; | |
1497 | return 0; | |
1498 | } | |
1499 | ||
15c006a2 RD |
1500 | /* find actual order */ |
1501 | order = mb_find_order_for_block(e4b, block); | |
1502 | block = block >> order; | |
c9de560d AT |
1503 | |
1504 | ex->fe_len = 1 << order; | |
1505 | ex->fe_start = block << order; | |
1506 | ex->fe_group = e4b->bd_group; | |
1507 | ||
1508 | /* calc difference from given start */ | |
1509 | next = next - ex->fe_start; | |
1510 | ex->fe_len -= next; | |
1511 | ex->fe_start += next; | |
1512 | ||
1513 | while (needed > ex->fe_len && | |
d8ec0c39 | 1514 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1515 | |
1516 | if (block + 1 >= max) | |
1517 | break; | |
1518 | ||
1519 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1520 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1521 | break; |
1522 | ||
b051d8dc | 1523 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1524 | |
c9de560d AT |
1525 | block = next >> order; |
1526 | ex->fe_len += 1 << order; | |
1527 | } | |
1528 | ||
1529 | BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); | |
1530 | return ex->fe_len; | |
1531 | } | |
1532 | ||
1533 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1534 | { | |
1535 | int ord; | |
1536 | int mlen = 0; | |
1537 | int max = 0; | |
1538 | int cur; | |
1539 | int start = ex->fe_start; | |
1540 | int len = ex->fe_len; | |
1541 | unsigned ret = 0; | |
1542 | int len0 = len; | |
1543 | void *buddy; | |
1544 | ||
1545 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1546 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1547 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1548 | mb_check_buddy(e4b); |
1549 | mb_mark_used_double(e4b, start, len); | |
1550 | ||
1551 | e4b->bd_info->bb_free -= len; | |
1552 | if (e4b->bd_info->bb_first_free == start) | |
1553 | e4b->bd_info->bb_first_free += len; | |
1554 | ||
1555 | /* let's maintain fragments counter */ | |
1556 | if (start != 0) | |
c5e8f3f3 | 1557 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1558 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1559 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1560 | if (mlen && max) |
1561 | e4b->bd_info->bb_fragments++; | |
1562 | else if (!mlen && !max) | |
1563 | e4b->bd_info->bb_fragments--; | |
1564 | ||
1565 | /* let's maintain buddy itself */ | |
1566 | while (len) { | |
1567 | ord = mb_find_order_for_block(e4b, start); | |
1568 | ||
1569 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1570 | /* the whole chunk may be allocated at once! */ | |
1571 | mlen = 1 << ord; | |
1572 | buddy = mb_find_buddy(e4b, ord, &max); | |
1573 | BUG_ON((start >> ord) >= max); | |
1574 | mb_set_bit(start >> ord, buddy); | |
1575 | e4b->bd_info->bb_counters[ord]--; | |
1576 | start += mlen; | |
1577 | len -= mlen; | |
1578 | BUG_ON(len < 0); | |
1579 | continue; | |
1580 | } | |
1581 | ||
1582 | /* store for history */ | |
1583 | if (ret == 0) | |
1584 | ret = len | (ord << 16); | |
1585 | ||
1586 | /* we have to split large buddy */ | |
1587 | BUG_ON(ord <= 0); | |
1588 | buddy = mb_find_buddy(e4b, ord, &max); | |
1589 | mb_set_bit(start >> ord, buddy); | |
1590 | e4b->bd_info->bb_counters[ord]--; | |
1591 | ||
1592 | ord--; | |
1593 | cur = (start >> ord) & ~1U; | |
1594 | buddy = mb_find_buddy(e4b, ord, &max); | |
1595 | mb_clear_bit(cur, buddy); | |
1596 | mb_clear_bit(cur + 1, buddy); | |
1597 | e4b->bd_info->bb_counters[ord]++; | |
1598 | e4b->bd_info->bb_counters[ord]++; | |
1599 | } | |
8a57d9d6 | 1600 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1601 | |
c5e8f3f3 | 1602 | ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
1603 | mb_check_buddy(e4b); |
1604 | ||
1605 | return ret; | |
1606 | } | |
1607 | ||
1608 | /* | |
1609 | * Must be called under group lock! | |
1610 | */ | |
1611 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1612 | struct ext4_buddy *e4b) | |
1613 | { | |
1614 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1615 | int ret; | |
1616 | ||
1617 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1618 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1619 | ||
1620 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1621 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1622 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1623 | ||
1624 | /* preallocation can change ac_b_ex, thus we store actually | |
1625 | * allocated blocks for history */ | |
1626 | ac->ac_f_ex = ac->ac_b_ex; | |
1627 | ||
1628 | ac->ac_status = AC_STATUS_FOUND; | |
1629 | ac->ac_tail = ret & 0xffff; | |
1630 | ac->ac_buddy = ret >> 16; | |
1631 | ||
c3a326a6 AK |
1632 | /* |
1633 | * take the page reference. We want the page to be pinned | |
1634 | * so that we don't get a ext4_mb_init_cache_call for this | |
1635 | * group until we update the bitmap. That would mean we | |
1636 | * double allocate blocks. The reference is dropped | |
1637 | * in ext4_mb_release_context | |
1638 | */ | |
c9de560d AT |
1639 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1640 | get_page(ac->ac_bitmap_page); | |
1641 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1642 | get_page(ac->ac_buddy_page); | |
c9de560d | 1643 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 1644 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
1645 | spin_lock(&sbi->s_md_lock); |
1646 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1647 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1648 | spin_unlock(&sbi->s_md_lock); | |
1649 | } | |
1650 | } | |
1651 | ||
1652 | /* | |
1653 | * regular allocator, for general purposes allocation | |
1654 | */ | |
1655 | ||
1656 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1657 | struct ext4_buddy *e4b, | |
1658 | int finish_group) | |
1659 | { | |
1660 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1661 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1662 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1663 | struct ext4_free_extent ex; | |
1664 | int max; | |
1665 | ||
032115fc AK |
1666 | if (ac->ac_status == AC_STATUS_FOUND) |
1667 | return; | |
c9de560d AT |
1668 | /* |
1669 | * We don't want to scan for a whole year | |
1670 | */ | |
1671 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1672 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1673 | ac->ac_status = AC_STATUS_BREAK; | |
1674 | return; | |
1675 | } | |
1676 | ||
1677 | /* | |
1678 | * Haven't found good chunk so far, let's continue | |
1679 | */ | |
1680 | if (bex->fe_len < gex->fe_len) | |
1681 | return; | |
1682 | ||
1683 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1684 | && bex->fe_group == e4b->bd_group) { | |
1685 | /* recheck chunk's availability - we don't know | |
1686 | * when it was found (within this lock-unlock | |
1687 | * period or not) */ | |
15c006a2 | 1688 | max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); |
c9de560d AT |
1689 | if (max >= gex->fe_len) { |
1690 | ext4_mb_use_best_found(ac, e4b); | |
1691 | return; | |
1692 | } | |
1693 | } | |
1694 | } | |
1695 | ||
1696 | /* | |
1697 | * The routine checks whether found extent is good enough. If it is, | |
1698 | * then the extent gets marked used and flag is set to the context | |
1699 | * to stop scanning. Otherwise, the extent is compared with the | |
1700 | * previous found extent and if new one is better, then it's stored | |
1701 | * in the context. Later, the best found extent will be used, if | |
1702 | * mballoc can't find good enough extent. | |
1703 | * | |
1704 | * FIXME: real allocation policy is to be designed yet! | |
1705 | */ | |
1706 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1707 | struct ext4_free_extent *ex, | |
1708 | struct ext4_buddy *e4b) | |
1709 | { | |
1710 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1711 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1712 | ||
1713 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
1714 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
1715 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
1716 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
1717 | ||
1718 | ac->ac_found++; | |
1719 | ||
1720 | /* | |
1721 | * The special case - take what you catch first | |
1722 | */ | |
1723 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1724 | *bex = *ex; | |
1725 | ext4_mb_use_best_found(ac, e4b); | |
1726 | return; | |
1727 | } | |
1728 | ||
1729 | /* | |
1730 | * Let's check whether the chuck is good enough | |
1731 | */ | |
1732 | if (ex->fe_len == gex->fe_len) { | |
1733 | *bex = *ex; | |
1734 | ext4_mb_use_best_found(ac, e4b); | |
1735 | return; | |
1736 | } | |
1737 | ||
1738 | /* | |
1739 | * If this is first found extent, just store it in the context | |
1740 | */ | |
1741 | if (bex->fe_len == 0) { | |
1742 | *bex = *ex; | |
1743 | return; | |
1744 | } | |
1745 | ||
1746 | /* | |
1747 | * If new found extent is better, store it in the context | |
1748 | */ | |
1749 | if (bex->fe_len < gex->fe_len) { | |
1750 | /* if the request isn't satisfied, any found extent | |
1751 | * larger than previous best one is better */ | |
1752 | if (ex->fe_len > bex->fe_len) | |
1753 | *bex = *ex; | |
1754 | } else if (ex->fe_len > gex->fe_len) { | |
1755 | /* if the request is satisfied, then we try to find | |
1756 | * an extent that still satisfy the request, but is | |
1757 | * smaller than previous one */ | |
1758 | if (ex->fe_len < bex->fe_len) | |
1759 | *bex = *ex; | |
1760 | } | |
1761 | ||
1762 | ext4_mb_check_limits(ac, e4b, 0); | |
1763 | } | |
1764 | ||
089ceecc ES |
1765 | static noinline_for_stack |
1766 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
1767 | struct ext4_buddy *e4b) |
1768 | { | |
1769 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1770 | ext4_group_t group = ex.fe_group; | |
1771 | int max; | |
1772 | int err; | |
1773 | ||
1774 | BUG_ON(ex.fe_len <= 0); | |
1775 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1776 | if (err) | |
1777 | return err; | |
1778 | ||
1779 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 1780 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
1781 | |
1782 | if (max > 0) { | |
1783 | ac->ac_b_ex = ex; | |
1784 | ext4_mb_use_best_found(ac, e4b); | |
1785 | } | |
1786 | ||
1787 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1788 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1789 | |
1790 | return 0; | |
1791 | } | |
1792 | ||
089ceecc ES |
1793 | static noinline_for_stack |
1794 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
1795 | struct ext4_buddy *e4b) |
1796 | { | |
1797 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1798 | int max; | |
1799 | int err; | |
1800 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 1801 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
1802 | struct ext4_free_extent ex; |
1803 | ||
1804 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1805 | return 0; | |
838cd0cf YY |
1806 | if (grp->bb_free == 0) |
1807 | return 0; | |
c9de560d AT |
1808 | |
1809 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1810 | if (err) | |
1811 | return err; | |
1812 | ||
163a203d DW |
1813 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { |
1814 | ext4_mb_unload_buddy(e4b); | |
1815 | return 0; | |
1816 | } | |
1817 | ||
c9de560d | 1818 | ext4_lock_group(ac->ac_sb, group); |
15c006a2 | 1819 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 1820 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 1821 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d AT |
1822 | |
1823 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1824 | ext4_fsblk_t start; | |
1825 | ||
5661bd68 AM |
1826 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
1827 | ex.fe_start; | |
c9de560d AT |
1828 | /* use do_div to get remainder (would be 64-bit modulo) */ |
1829 | if (do_div(start, sbi->s_stripe) == 0) { | |
1830 | ac->ac_found++; | |
1831 | ac->ac_b_ex = ex; | |
1832 | ext4_mb_use_best_found(ac, e4b); | |
1833 | } | |
1834 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1835 | BUG_ON(ex.fe_len <= 0); | |
1836 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1837 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1838 | ac->ac_found++; | |
1839 | ac->ac_b_ex = ex; | |
1840 | ext4_mb_use_best_found(ac, e4b); | |
1841 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1842 | /* Sometimes, caller may want to merge even small | |
1843 | * number of blocks to an existing extent */ | |
1844 | BUG_ON(ex.fe_len <= 0); | |
1845 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1846 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1847 | ac->ac_found++; | |
1848 | ac->ac_b_ex = ex; | |
1849 | ext4_mb_use_best_found(ac, e4b); | |
1850 | } | |
1851 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1852 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1853 | |
1854 | return 0; | |
1855 | } | |
1856 | ||
1857 | /* | |
1858 | * The routine scans buddy structures (not bitmap!) from given order | |
1859 | * to max order and tries to find big enough chunk to satisfy the req | |
1860 | */ | |
089ceecc ES |
1861 | static noinline_for_stack |
1862 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1863 | struct ext4_buddy *e4b) |
1864 | { | |
1865 | struct super_block *sb = ac->ac_sb; | |
1866 | struct ext4_group_info *grp = e4b->bd_info; | |
1867 | void *buddy; | |
1868 | int i; | |
1869 | int k; | |
1870 | int max; | |
1871 | ||
1872 | BUG_ON(ac->ac_2order <= 0); | |
1873 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1874 | if (grp->bb_counters[i] == 0) | |
1875 | continue; | |
1876 | ||
1877 | buddy = mb_find_buddy(e4b, i, &max); | |
1878 | BUG_ON(buddy == NULL); | |
1879 | ||
ffad0a44 | 1880 | k = mb_find_next_zero_bit(buddy, max, 0); |
c9de560d AT |
1881 | BUG_ON(k >= max); |
1882 | ||
1883 | ac->ac_found++; | |
1884 | ||
1885 | ac->ac_b_ex.fe_len = 1 << i; | |
1886 | ac->ac_b_ex.fe_start = k << i; | |
1887 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1888 | ||
1889 | ext4_mb_use_best_found(ac, e4b); | |
1890 | ||
1891 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1892 | ||
1893 | if (EXT4_SB(sb)->s_mb_stats) | |
1894 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1895 | ||
1896 | break; | |
1897 | } | |
1898 | } | |
1899 | ||
1900 | /* | |
1901 | * The routine scans the group and measures all found extents. | |
1902 | * In order to optimize scanning, caller must pass number of | |
1903 | * free blocks in the group, so the routine can know upper limit. | |
1904 | */ | |
089ceecc ES |
1905 | static noinline_for_stack |
1906 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1907 | struct ext4_buddy *e4b) |
1908 | { | |
1909 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 1910 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
1911 | struct ext4_free_extent ex; |
1912 | int i; | |
1913 | int free; | |
1914 | ||
1915 | free = e4b->bd_info->bb_free; | |
1916 | BUG_ON(free <= 0); | |
1917 | ||
1918 | i = e4b->bd_info->bb_first_free; | |
1919 | ||
1920 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1921 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
1922 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
1923 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 1924 | /* |
e56eb659 | 1925 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1926 | * free blocks even though group info says we |
1927 | * we have free blocks | |
1928 | */ | |
e29136f8 | 1929 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1930 | "%d free clusters as per " |
fde4d95a | 1931 | "group info. But bitmap says 0", |
26346ff6 | 1932 | free); |
c9de560d AT |
1933 | break; |
1934 | } | |
1935 | ||
15c006a2 | 1936 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
c9de560d | 1937 | BUG_ON(ex.fe_len <= 0); |
26346ff6 | 1938 | if (free < ex.fe_len) { |
e29136f8 | 1939 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1940 | "%d free clusters as per " |
fde4d95a | 1941 | "group info. But got %d blocks", |
26346ff6 | 1942 | free, ex.fe_len); |
e56eb659 AK |
1943 | /* |
1944 | * The number of free blocks differs. This mostly | |
1945 | * indicate that the bitmap is corrupt. So exit | |
1946 | * without claiming the space. | |
1947 | */ | |
1948 | break; | |
26346ff6 | 1949 | } |
ab0c00fc | 1950 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
1951 | ext4_mb_measure_extent(ac, &ex, e4b); |
1952 | ||
1953 | i += ex.fe_len; | |
1954 | free -= ex.fe_len; | |
1955 | } | |
1956 | ||
1957 | ext4_mb_check_limits(ac, e4b, 1); | |
1958 | } | |
1959 | ||
1960 | /* | |
1961 | * This is a special case for storages like raid5 | |
506bf2d8 | 1962 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 1963 | */ |
089ceecc ES |
1964 | static noinline_for_stack |
1965 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
1966 | struct ext4_buddy *e4b) |
1967 | { | |
1968 | struct super_block *sb = ac->ac_sb; | |
1969 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 1970 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
1971 | struct ext4_free_extent ex; |
1972 | ext4_fsblk_t first_group_block; | |
1973 | ext4_fsblk_t a; | |
1974 | ext4_grpblk_t i; | |
1975 | int max; | |
1976 | ||
1977 | BUG_ON(sbi->s_stripe == 0); | |
1978 | ||
1979 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
1980 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
1981 | ||
c9de560d AT |
1982 | a = first_group_block + sbi->s_stripe - 1; |
1983 | do_div(a, sbi->s_stripe); | |
1984 | i = (a * sbi->s_stripe) - first_group_block; | |
1985 | ||
7137d7a4 | 1986 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 1987 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 1988 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
1989 | if (max >= sbi->s_stripe) { |
1990 | ac->ac_found++; | |
ab0c00fc | 1991 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
1992 | ac->ac_b_ex = ex; |
1993 | ext4_mb_use_best_found(ac, e4b); | |
1994 | break; | |
1995 | } | |
1996 | } | |
1997 | i += sbi->s_stripe; | |
1998 | } | |
1999 | } | |
2000 | ||
8a57d9d6 | 2001 | /* This is now called BEFORE we load the buddy bitmap. */ |
c9de560d AT |
2002 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, |
2003 | ext4_group_t group, int cr) | |
2004 | { | |
2005 | unsigned free, fragments; | |
a4912123 | 2006 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2007 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2008 | ||
2009 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 2010 | |
01fc48e8 TT |
2011 | free = grp->bb_free; |
2012 | if (free == 0) | |
2013 | return 0; | |
2014 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
2015 | return 0; | |
2016 | ||
163a203d DW |
2017 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
2018 | return 0; | |
2019 | ||
8a57d9d6 CW |
2020 | /* We only do this if the grp has never been initialized */ |
2021 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
2022 | int ret = ext4_mb_init_group(ac->ac_sb, group); | |
2023 | if (ret) | |
2024 | return 0; | |
2025 | } | |
c9de560d | 2026 | |
c9de560d | 2027 | fragments = grp->bb_fragments; |
c9de560d AT |
2028 | if (fragments == 0) |
2029 | return 0; | |
2030 | ||
2031 | switch (cr) { | |
2032 | case 0: | |
2033 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 2034 | |
a4912123 TT |
2035 | /* Avoid using the first bg of a flexgroup for data files */ |
2036 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2037 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2038 | ((group % flex_size) == 0)) | |
2039 | return 0; | |
2040 | ||
40ae3487 TT |
2041 | if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) || |
2042 | (free / fragments) >= ac->ac_g_ex.fe_len) | |
2043 | return 1; | |
2044 | ||
2045 | if (grp->bb_largest_free_order < ac->ac_2order) | |
2046 | return 0; | |
2047 | ||
8a57d9d6 | 2048 | return 1; |
c9de560d AT |
2049 | case 1: |
2050 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
2051 | return 1; | |
2052 | break; | |
2053 | case 2: | |
2054 | if (free >= ac->ac_g_ex.fe_len) | |
2055 | return 1; | |
2056 | break; | |
2057 | case 3: | |
2058 | return 1; | |
2059 | default: | |
2060 | BUG(); | |
2061 | } | |
2062 | ||
2063 | return 0; | |
2064 | } | |
2065 | ||
4ddfef7b ES |
2066 | static noinline_for_stack int |
2067 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2068 | { |
8df9675f | 2069 | ext4_group_t ngroups, group, i; |
c9de560d AT |
2070 | int cr; |
2071 | int err = 0; | |
c9de560d AT |
2072 | struct ext4_sb_info *sbi; |
2073 | struct super_block *sb; | |
2074 | struct ext4_buddy e4b; | |
c9de560d AT |
2075 | |
2076 | sb = ac->ac_sb; | |
2077 | sbi = EXT4_SB(sb); | |
8df9675f | 2078 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2079 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2080 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2081 | ngroups = sbi->s_blockfile_groups; |
2082 | ||
c9de560d AT |
2083 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2084 | ||
2085 | /* first, try the goal */ | |
2086 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2087 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2088 | goto out; | |
2089 | ||
2090 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2091 | goto out; | |
2092 | ||
2093 | /* | |
2094 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
2095 | * if ac2_order is set we also set criteria to 0 so that we | |
2096 | * try exact allocation using buddy. | |
2097 | */ | |
2098 | i = fls(ac->ac_g_ex.fe_len); | |
2099 | ac->ac_2order = 0; | |
2100 | /* | |
2101 | * We search using buddy data only if the order of the request | |
2102 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2103 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
c9de560d AT |
2104 | */ |
2105 | if (i >= sbi->s_mb_order2_reqs) { | |
2106 | /* | |
2107 | * This should tell if fe_len is exactly power of 2 | |
2108 | */ | |
2109 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
2110 | ac->ac_2order = i - 1; | |
2111 | } | |
2112 | ||
4ba74d00 TT |
2113 | /* if stream allocation is enabled, use global goal */ |
2114 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2115 | /* TBD: may be hot point */ |
2116 | spin_lock(&sbi->s_md_lock); | |
2117 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2118 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2119 | spin_unlock(&sbi->s_md_lock); | |
2120 | } | |
4ba74d00 | 2121 | |
c9de560d AT |
2122 | /* Let's just scan groups to find more-less suitable blocks */ |
2123 | cr = ac->ac_2order ? 0 : 1; | |
2124 | /* | |
2125 | * cr == 0 try to get exact allocation, | |
2126 | * cr == 3 try to get anything | |
2127 | */ | |
2128 | repeat: | |
2129 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2130 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2131 | /* |
2132 | * searching for the right group start | |
2133 | * from the goal value specified | |
2134 | */ | |
2135 | group = ac->ac_g_ex.fe_group; | |
2136 | ||
8df9675f | 2137 | for (i = 0; i < ngroups; group++, i++) { |
2ed5724d | 2138 | cond_resched(); |
e6155736 LM |
2139 | /* |
2140 | * Artificially restricted ngroups for non-extent | |
2141 | * files makes group > ngroups possible on first loop. | |
2142 | */ | |
2143 | if (group >= ngroups) | |
c9de560d AT |
2144 | group = 0; |
2145 | ||
8a57d9d6 CW |
2146 | /* This now checks without needing the buddy page */ |
2147 | if (!ext4_mb_good_group(ac, group, cr)) | |
c9de560d AT |
2148 | continue; |
2149 | ||
c9de560d AT |
2150 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2151 | if (err) | |
2152 | goto out; | |
2153 | ||
2154 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2155 | |
2156 | /* | |
2157 | * We need to check again after locking the | |
2158 | * block group | |
2159 | */ | |
c9de560d | 2160 | if (!ext4_mb_good_group(ac, group, cr)) { |
c9de560d | 2161 | ext4_unlock_group(sb, group); |
e39e07fd | 2162 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2163 | continue; |
2164 | } | |
2165 | ||
2166 | ac->ac_groups_scanned++; | |
40ae3487 | 2167 | if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2) |
c9de560d | 2168 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2169 | else if (cr == 1 && sbi->s_stripe && |
2170 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2171 | ext4_mb_scan_aligned(ac, &e4b); |
2172 | else | |
2173 | ext4_mb_complex_scan_group(ac, &e4b); | |
2174 | ||
2175 | ext4_unlock_group(sb, group); | |
e39e07fd | 2176 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2177 | |
2178 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2179 | break; | |
2180 | } | |
2181 | } | |
2182 | ||
2183 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2184 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2185 | /* | |
2186 | * We've been searching too long. Let's try to allocate | |
2187 | * the best chunk we've found so far | |
2188 | */ | |
2189 | ||
2190 | ext4_mb_try_best_found(ac, &e4b); | |
2191 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2192 | /* | |
2193 | * Someone more lucky has already allocated it. | |
2194 | * The only thing we can do is just take first | |
2195 | * found block(s) | |
2196 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2197 | */ | |
2198 | ac->ac_b_ex.fe_group = 0; | |
2199 | ac->ac_b_ex.fe_start = 0; | |
2200 | ac->ac_b_ex.fe_len = 0; | |
2201 | ac->ac_status = AC_STATUS_CONTINUE; | |
2202 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2203 | cr = 3; | |
2204 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2205 | goto repeat; | |
2206 | } | |
2207 | } | |
2208 | out: | |
2209 | return err; | |
2210 | } | |
2211 | ||
c9de560d AT |
2212 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2213 | { | |
2214 | struct super_block *sb = seq->private; | |
c9de560d AT |
2215 | ext4_group_t group; |
2216 | ||
8df9675f | 2217 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2218 | return NULL; |
c9de560d | 2219 | group = *pos + 1; |
a9df9a49 | 2220 | return (void *) ((unsigned long) group); |
c9de560d AT |
2221 | } |
2222 | ||
2223 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2224 | { | |
2225 | struct super_block *sb = seq->private; | |
c9de560d AT |
2226 | ext4_group_t group; |
2227 | ||
2228 | ++*pos; | |
8df9675f | 2229 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2230 | return NULL; |
2231 | group = *pos + 1; | |
a9df9a49 | 2232 | return (void *) ((unsigned long) group); |
c9de560d AT |
2233 | } |
2234 | ||
2235 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2236 | { | |
2237 | struct super_block *sb = seq->private; | |
a9df9a49 | 2238 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2239 | int i; |
1c8457ca | 2240 | int err, buddy_loaded = 0; |
c9de560d | 2241 | struct ext4_buddy e4b; |
1c8457ca | 2242 | struct ext4_group_info *grinfo; |
c9de560d AT |
2243 | struct sg { |
2244 | struct ext4_group_info info; | |
a36b4498 | 2245 | ext4_grpblk_t counters[16]; |
c9de560d AT |
2246 | } sg; |
2247 | ||
2248 | group--; | |
2249 | if (group == 0) | |
2250 | seq_printf(seq, "#%-5s: %-5s %-5s %-5s " | |
2251 | "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " | |
2252 | "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", | |
2253 | "group", "free", "frags", "first", | |
2254 | "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", | |
2255 | "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); | |
2256 | ||
2257 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2258 | sizeof(struct ext4_group_info); | |
1c8457ca AK |
2259 | grinfo = ext4_get_group_info(sb, group); |
2260 | /* Load the group info in memory only if not already loaded. */ | |
2261 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2262 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2263 | if (err) { | |
2264 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2265 | return 0; | |
2266 | } | |
2267 | buddy_loaded = 1; | |
c9de560d | 2268 | } |
1c8457ca | 2269 | |
c9de560d | 2270 | memcpy(&sg, ext4_get_group_info(sb, group), i); |
1c8457ca AK |
2271 | |
2272 | if (buddy_loaded) | |
2273 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2274 | |
a9df9a49 | 2275 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2276 | sg.info.bb_fragments, sg.info.bb_first_free); |
2277 | for (i = 0; i <= 13; i++) | |
2278 | seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? | |
2279 | sg.info.bb_counters[i] : 0); | |
2280 | seq_printf(seq, " ]\n"); | |
2281 | ||
2282 | return 0; | |
2283 | } | |
2284 | ||
2285 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2286 | { | |
2287 | } | |
2288 | ||
7f1346a9 | 2289 | static const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2290 | .start = ext4_mb_seq_groups_start, |
2291 | .next = ext4_mb_seq_groups_next, | |
2292 | .stop = ext4_mb_seq_groups_stop, | |
2293 | .show = ext4_mb_seq_groups_show, | |
2294 | }; | |
2295 | ||
2296 | static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) | |
2297 | { | |
d9dda78b | 2298 | struct super_block *sb = PDE_DATA(inode); |
c9de560d AT |
2299 | int rc; |
2300 | ||
2301 | rc = seq_open(file, &ext4_mb_seq_groups_ops); | |
2302 | if (rc == 0) { | |
a271fe85 | 2303 | struct seq_file *m = file->private_data; |
c9de560d AT |
2304 | m->private = sb; |
2305 | } | |
2306 | return rc; | |
2307 | ||
2308 | } | |
2309 | ||
7f1346a9 | 2310 | static const struct file_operations ext4_mb_seq_groups_fops = { |
c9de560d AT |
2311 | .owner = THIS_MODULE, |
2312 | .open = ext4_mb_seq_groups_open, | |
2313 | .read = seq_read, | |
2314 | .llseek = seq_lseek, | |
2315 | .release = seq_release, | |
2316 | }; | |
2317 | ||
fb1813f4 CW |
2318 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
2319 | { | |
2320 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2321 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
2322 | ||
2323 | BUG_ON(!cachep); | |
2324 | return cachep; | |
2325 | } | |
5f21b0e6 | 2326 | |
28623c2f TT |
2327 | /* |
2328 | * Allocate the top-level s_group_info array for the specified number | |
2329 | * of groups | |
2330 | */ | |
2331 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
2332 | { | |
2333 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2334 | unsigned size; | |
2335 | struct ext4_group_info ***new_groupinfo; | |
2336 | ||
2337 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
2338 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2339 | if (size <= sbi->s_group_info_size) | |
2340 | return 0; | |
2341 | ||
2342 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
2343 | new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL); | |
2344 | if (!new_groupinfo) { | |
2345 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
2346 | return -ENOMEM; | |
2347 | } | |
2348 | if (sbi->s_group_info) { | |
2349 | memcpy(new_groupinfo, sbi->s_group_info, | |
2350 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); | |
2351 | ext4_kvfree(sbi->s_group_info); | |
2352 | } | |
2353 | sbi->s_group_info = new_groupinfo; | |
2354 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); | |
2355 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", | |
2356 | sbi->s_group_info_size); | |
2357 | return 0; | |
2358 | } | |
2359 | ||
5f21b0e6 | 2360 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 2361 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2362 | struct ext4_group_desc *desc) |
2363 | { | |
fb1813f4 | 2364 | int i; |
5f21b0e6 FB |
2365 | int metalen = 0; |
2366 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2367 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 2368 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
2369 | |
2370 | /* | |
2371 | * First check if this group is the first of a reserved block. | |
2372 | * If it's true, we have to allocate a new table of pointers | |
2373 | * to ext4_group_info structures | |
2374 | */ | |
2375 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2376 | metalen = sizeof(*meta_group_info) << | |
2377 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2378 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2379 | if (meta_group_info == NULL) { | |
7f6a11e7 | 2380 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 2381 | "for a buddy group"); |
5f21b0e6 FB |
2382 | goto exit_meta_group_info; |
2383 | } | |
2384 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = | |
2385 | meta_group_info; | |
2386 | } | |
2387 | ||
5f21b0e6 FB |
2388 | meta_group_info = |
2389 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; | |
2390 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); | |
2391 | ||
85556c9a | 2392 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL); |
5f21b0e6 | 2393 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 2394 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
2395 | goto exit_group_info; |
2396 | } | |
2397 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2398 | &(meta_group_info[i]->bb_state)); | |
2399 | ||
2400 | /* | |
2401 | * initialize bb_free to be able to skip | |
2402 | * empty groups without initialization | |
2403 | */ | |
2404 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2405 | meta_group_info[i]->bb_free = | |
cff1dfd7 | 2406 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
2407 | } else { |
2408 | meta_group_info[i]->bb_free = | |
021b65bb | 2409 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
2410 | } |
2411 | ||
2412 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2413 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 2414 | meta_group_info[i]->bb_free_root = RB_ROOT; |
8a57d9d6 | 2415 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
5f21b0e6 FB |
2416 | |
2417 | #ifdef DOUBLE_CHECK | |
2418 | { | |
2419 | struct buffer_head *bh; | |
2420 | meta_group_info[i]->bb_bitmap = | |
2421 | kmalloc(sb->s_blocksize, GFP_KERNEL); | |
2422 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); | |
2423 | bh = ext4_read_block_bitmap(sb, group); | |
2424 | BUG_ON(bh == NULL); | |
2425 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, | |
2426 | sb->s_blocksize); | |
2427 | put_bh(bh); | |
2428 | } | |
2429 | #endif | |
2430 | ||
2431 | return 0; | |
2432 | ||
2433 | exit_group_info: | |
2434 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 2435 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
5f21b0e6 | 2436 | kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); |
caaf7a29 TM |
2437 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL; |
2438 | } | |
5f21b0e6 FB |
2439 | exit_meta_group_info: |
2440 | return -ENOMEM; | |
2441 | } /* ext4_mb_add_groupinfo */ | |
2442 | ||
c9de560d AT |
2443 | static int ext4_mb_init_backend(struct super_block *sb) |
2444 | { | |
8df9675f | 2445 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2446 | ext4_group_t i; |
c9de560d | 2447 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 2448 | int err; |
5f21b0e6 | 2449 | struct ext4_group_desc *desc; |
fb1813f4 | 2450 | struct kmem_cache *cachep; |
5f21b0e6 | 2451 | |
28623c2f TT |
2452 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
2453 | if (err) | |
2454 | return err; | |
c9de560d | 2455 | |
c9de560d AT |
2456 | sbi->s_buddy_cache = new_inode(sb); |
2457 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 2458 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
2459 | goto err_freesgi; |
2460 | } | |
48e6061b YJ |
2461 | /* To avoid potentially colliding with an valid on-disk inode number, |
2462 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
2463 | * not in the inode hash, so it should never be found by iget(), but | |
2464 | * this will avoid confusion if it ever shows up during debugging. */ | |
2465 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 2466 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 2467 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2468 | desc = ext4_get_group_desc(sb, i, NULL); |
2469 | if (desc == NULL) { | |
9d8b9ec4 | 2470 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
2471 | goto err_freebuddy; |
2472 | } | |
5f21b0e6 FB |
2473 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2474 | goto err_freebuddy; | |
c9de560d AT |
2475 | } |
2476 | ||
2477 | return 0; | |
2478 | ||
2479 | err_freebuddy: | |
fb1813f4 | 2480 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 2481 | while (i-- > 0) |
fb1813f4 | 2482 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
28623c2f | 2483 | i = sbi->s_group_info_size; |
f1fa3342 | 2484 | while (i-- > 0) |
c9de560d AT |
2485 | kfree(sbi->s_group_info[i]); |
2486 | iput(sbi->s_buddy_cache); | |
2487 | err_freesgi: | |
f18a5f21 | 2488 | ext4_kvfree(sbi->s_group_info); |
c9de560d AT |
2489 | return -ENOMEM; |
2490 | } | |
2491 | ||
2892c15d ES |
2492 | static void ext4_groupinfo_destroy_slabs(void) |
2493 | { | |
2494 | int i; | |
2495 | ||
2496 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
2497 | if (ext4_groupinfo_caches[i]) | |
2498 | kmem_cache_destroy(ext4_groupinfo_caches[i]); | |
2499 | ext4_groupinfo_caches[i] = NULL; | |
2500 | } | |
2501 | } | |
2502 | ||
2503 | static int ext4_groupinfo_create_slab(size_t size) | |
2504 | { | |
2505 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
2506 | int slab_size; | |
2507 | int blocksize_bits = order_base_2(size); | |
2508 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2509 | struct kmem_cache *cachep; | |
2510 | ||
2511 | if (cache_index >= NR_GRPINFO_CACHES) | |
2512 | return -EINVAL; | |
2513 | ||
2514 | if (unlikely(cache_index < 0)) | |
2515 | cache_index = 0; | |
2516 | ||
2517 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
2518 | if (ext4_groupinfo_caches[cache_index]) { | |
2519 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
2520 | return 0; /* Already created */ | |
2521 | } | |
2522 | ||
2523 | slab_size = offsetof(struct ext4_group_info, | |
2524 | bb_counters[blocksize_bits + 2]); | |
2525 | ||
2526 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
2527 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
2528 | NULL); | |
2529 | ||
823ba01f TM |
2530 | ext4_groupinfo_caches[cache_index] = cachep; |
2531 | ||
2892c15d ES |
2532 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
2533 | if (!cachep) { | |
9d8b9ec4 TT |
2534 | printk(KERN_EMERG |
2535 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
2536 | return -ENOMEM; |
2537 | } | |
2538 | ||
2892c15d ES |
2539 | return 0; |
2540 | } | |
2541 | ||
9d99012f | 2542 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
2543 | { |
2544 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2545 | unsigned i, j; |
c9de560d AT |
2546 | unsigned offset; |
2547 | unsigned max; | |
74767c5a | 2548 | int ret; |
c9de560d | 2549 | |
1927805e | 2550 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
2551 | |
2552 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2553 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
2554 | ret = -ENOMEM; |
2555 | goto out; | |
c9de560d | 2556 | } |
ff7ef329 | 2557 | |
1927805e | 2558 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
2559 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2560 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
2561 | ret = -ENOMEM; |
2562 | goto out; | |
2563 | } | |
2564 | ||
2892c15d ES |
2565 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
2566 | if (ret < 0) | |
2567 | goto out; | |
c9de560d AT |
2568 | |
2569 | /* order 0 is regular bitmap */ | |
2570 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2571 | sbi->s_mb_offsets[0] = 0; | |
2572 | ||
2573 | i = 1; | |
2574 | offset = 0; | |
2575 | max = sb->s_blocksize << 2; | |
2576 | do { | |
2577 | sbi->s_mb_offsets[i] = offset; | |
2578 | sbi->s_mb_maxs[i] = max; | |
2579 | offset += 1 << (sb->s_blocksize_bits - i); | |
2580 | max = max >> 1; | |
2581 | i++; | |
2582 | } while (i <= sb->s_blocksize_bits + 1); | |
2583 | ||
c9de560d | 2584 | spin_lock_init(&sbi->s_md_lock); |
c9de560d AT |
2585 | spin_lock_init(&sbi->s_bal_lock); |
2586 | ||
2587 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2588 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2589 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2590 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2591 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27baebb8 TT |
2592 | /* |
2593 | * The default group preallocation is 512, which for 4k block | |
2594 | * sizes translates to 2 megabytes. However for bigalloc file | |
2595 | * systems, this is probably too big (i.e, if the cluster size | |
2596 | * is 1 megabyte, then group preallocation size becomes half a | |
2597 | * gigabyte!). As a default, we will keep a two megabyte | |
2598 | * group pralloc size for cluster sizes up to 64k, and after | |
2599 | * that, we will force a minimum group preallocation size of | |
2600 | * 32 clusters. This translates to 8 megs when the cluster | |
2601 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
2602 | * which seems reasonable as a default. | |
2603 | */ | |
2604 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
2605 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
2606 | /* |
2607 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
2608 | * to the lowest multiple of s_stripe which is bigger than | |
2609 | * the s_mb_group_prealloc as determined above. We want | |
2610 | * the preallocation size to be an exact multiple of the | |
2611 | * RAID stripe size so that preallocations don't fragment | |
2612 | * the stripes. | |
2613 | */ | |
2614 | if (sbi->s_stripe > 1) { | |
2615 | sbi->s_mb_group_prealloc = roundup( | |
2616 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
2617 | } | |
c9de560d | 2618 | |
730c213c | 2619 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2620 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 2621 | ret = -ENOMEM; |
029b10c5 | 2622 | goto out; |
c9de560d | 2623 | } |
730c213c | 2624 | for_each_possible_cpu(i) { |
c9de560d | 2625 | struct ext4_locality_group *lg; |
730c213c | 2626 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2627 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2628 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2629 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2630 | spin_lock_init(&lg->lg_prealloc_lock); |
2631 | } | |
2632 | ||
79a77c5a YJ |
2633 | /* init file for buddy data */ |
2634 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
2635 | if (ret != 0) |
2636 | goto out_free_locality_groups; | |
79a77c5a | 2637 | |
296c355c TT |
2638 | if (sbi->s_proc) |
2639 | proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, | |
2640 | &ext4_mb_seq_groups_fops, sb); | |
c9de560d | 2641 | |
7aa0baea TM |
2642 | return 0; |
2643 | ||
2644 | out_free_locality_groups: | |
2645 | free_percpu(sbi->s_locality_groups); | |
2646 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 2647 | out: |
7aa0baea TM |
2648 | kfree(sbi->s_mb_offsets); |
2649 | sbi->s_mb_offsets = NULL; | |
2650 | kfree(sbi->s_mb_maxs); | |
2651 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 2652 | return ret; |
c9de560d AT |
2653 | } |
2654 | ||
955ce5f5 | 2655 | /* need to called with the ext4 group lock held */ |
c9de560d AT |
2656 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
2657 | { | |
2658 | struct ext4_prealloc_space *pa; | |
2659 | struct list_head *cur, *tmp; | |
2660 | int count = 0; | |
2661 | ||
2662 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2663 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2664 | list_del(&pa->pa_group_list); | |
2665 | count++; | |
688f05a0 | 2666 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2667 | } |
2668 | if (count) | |
6ba495e9 | 2669 | mb_debug(1, "mballoc: %u PAs left\n", count); |
c9de560d AT |
2670 | |
2671 | } | |
2672 | ||
2673 | int ext4_mb_release(struct super_block *sb) | |
2674 | { | |
8df9675f | 2675 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2676 | ext4_group_t i; |
2677 | int num_meta_group_infos; | |
2678 | struct ext4_group_info *grinfo; | |
2679 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
fb1813f4 | 2680 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
c9de560d | 2681 | |
95599968 SQ |
2682 | if (sbi->s_proc) |
2683 | remove_proc_entry("mb_groups", sbi->s_proc); | |
2684 | ||
c9de560d | 2685 | if (sbi->s_group_info) { |
8df9675f | 2686 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2687 | grinfo = ext4_get_group_info(sb, i); |
2688 | #ifdef DOUBLE_CHECK | |
2689 | kfree(grinfo->bb_bitmap); | |
2690 | #endif | |
2691 | ext4_lock_group(sb, i); | |
2692 | ext4_mb_cleanup_pa(grinfo); | |
2693 | ext4_unlock_group(sb, i); | |
fb1813f4 | 2694 | kmem_cache_free(cachep, grinfo); |
c9de560d | 2695 | } |
8df9675f | 2696 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2697 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2698 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2699 | for (i = 0; i < num_meta_group_infos; i++) | |
2700 | kfree(sbi->s_group_info[i]); | |
f18a5f21 | 2701 | ext4_kvfree(sbi->s_group_info); |
c9de560d AT |
2702 | } |
2703 | kfree(sbi->s_mb_offsets); | |
2704 | kfree(sbi->s_mb_maxs); | |
2705 | if (sbi->s_buddy_cache) | |
2706 | iput(sbi->s_buddy_cache); | |
2707 | if (sbi->s_mb_stats) { | |
9d8b9ec4 TT |
2708 | ext4_msg(sb, KERN_INFO, |
2709 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
2710 | atomic_read(&sbi->s_bal_allocated), |
2711 | atomic_read(&sbi->s_bal_reqs), | |
2712 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 TT |
2713 | ext4_msg(sb, KERN_INFO, |
2714 | "mballoc: %u extents scanned, %u goal hits, " | |
2715 | "%u 2^N hits, %u breaks, %u lost", | |
c9de560d AT |
2716 | atomic_read(&sbi->s_bal_ex_scanned), |
2717 | atomic_read(&sbi->s_bal_goals), | |
2718 | atomic_read(&sbi->s_bal_2orders), | |
2719 | atomic_read(&sbi->s_bal_breaks), | |
2720 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 TT |
2721 | ext4_msg(sb, KERN_INFO, |
2722 | "mballoc: %lu generated and it took %Lu", | |
ced156e4 | 2723 | sbi->s_mb_buddies_generated, |
c9de560d | 2724 | sbi->s_mb_generation_time); |
9d8b9ec4 TT |
2725 | ext4_msg(sb, KERN_INFO, |
2726 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
2727 | atomic_read(&sbi->s_mb_preallocated), |
2728 | atomic_read(&sbi->s_mb_discarded)); | |
2729 | } | |
2730 | ||
730c213c | 2731 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
2732 | |
2733 | return 0; | |
2734 | } | |
2735 | ||
77ca6cdf | 2736 | static inline int ext4_issue_discard(struct super_block *sb, |
84130193 | 2737 | ext4_group_t block_group, ext4_grpblk_t cluster, int count) |
5c521830 | 2738 | { |
5c521830 JZ |
2739 | ext4_fsblk_t discard_block; |
2740 | ||
84130193 TT |
2741 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
2742 | ext4_group_first_block_no(sb, block_group)); | |
2743 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
2744 | trace_ext4_discard_blocks(sb, |
2745 | (unsigned long long) discard_block, count); | |
93259636 | 2746 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); |
5c521830 JZ |
2747 | } |
2748 | ||
3e624fc7 TT |
2749 | /* |
2750 | * This function is called by the jbd2 layer once the commit has finished, | |
2751 | * so we know we can free the blocks that were released with that commit. | |
2752 | */ | |
18aadd47 BJ |
2753 | static void ext4_free_data_callback(struct super_block *sb, |
2754 | struct ext4_journal_cb_entry *jce, | |
2755 | int rc) | |
c9de560d | 2756 | { |
18aadd47 | 2757 | struct ext4_free_data *entry = (struct ext4_free_data *)jce; |
c9de560d | 2758 | struct ext4_buddy e4b; |
c894058d | 2759 | struct ext4_group_info *db; |
d9f34504 | 2760 | int err, count = 0, count2 = 0; |
c9de560d | 2761 | |
18aadd47 BJ |
2762 | mb_debug(1, "gonna free %u blocks in group %u (0x%p):", |
2763 | entry->efd_count, entry->efd_group, entry); | |
c9de560d | 2764 | |
d71c1ae2 LC |
2765 | if (test_opt(sb, DISCARD)) { |
2766 | err = ext4_issue_discard(sb, entry->efd_group, | |
2767 | entry->efd_start_cluster, | |
2768 | entry->efd_count); | |
2769 | if (err && err != -EOPNOTSUPP) | |
2770 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
2771 | " group:%d block:%d count:%d failed" | |
2772 | " with %d", entry->efd_group, | |
2773 | entry->efd_start_cluster, | |
2774 | entry->efd_count, err); | |
2775 | } | |
c9de560d | 2776 | |
18aadd47 BJ |
2777 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
2778 | /* we expect to find existing buddy because it's pinned */ | |
2779 | BUG_ON(err != 0); | |
b90f6870 | 2780 | |
c9de560d | 2781 | |
18aadd47 BJ |
2782 | db = e4b.bd_info; |
2783 | /* there are blocks to put in buddy to make them really free */ | |
2784 | count += entry->efd_count; | |
2785 | count2++; | |
2786 | ext4_lock_group(sb, entry->efd_group); | |
2787 | /* Take it out of per group rb tree */ | |
2788 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
2789 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 2790 | |
18aadd47 BJ |
2791 | /* |
2792 | * Clear the trimmed flag for the group so that the next | |
2793 | * ext4_trim_fs can trim it. | |
2794 | * If the volume is mounted with -o discard, online discard | |
2795 | * is supported and the free blocks will be trimmed online. | |
2796 | */ | |
2797 | if (!test_opt(sb, DISCARD)) | |
2798 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 2799 | |
18aadd47 BJ |
2800 | if (!db->bb_free_root.rb_node) { |
2801 | /* No more items in the per group rb tree | |
2802 | * balance refcounts from ext4_mb_free_metadata() | |
2803 | */ | |
2804 | page_cache_release(e4b.bd_buddy_page); | |
2805 | page_cache_release(e4b.bd_bitmap_page); | |
3e624fc7 | 2806 | } |
18aadd47 BJ |
2807 | ext4_unlock_group(sb, entry->efd_group); |
2808 | kmem_cache_free(ext4_free_data_cachep, entry); | |
2809 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2810 | |
6ba495e9 | 2811 | mb_debug(1, "freed %u blocks in %u structures\n", count, count2); |
c9de560d AT |
2812 | } |
2813 | ||
5dabfc78 | 2814 | int __init ext4_init_mballoc(void) |
c9de560d | 2815 | { |
16828088 TT |
2816 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
2817 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d AT |
2818 | if (ext4_pspace_cachep == NULL) |
2819 | return -ENOMEM; | |
2820 | ||
16828088 TT |
2821 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
2822 | SLAB_RECLAIM_ACCOUNT); | |
256bdb49 ES |
2823 | if (ext4_ac_cachep == NULL) { |
2824 | kmem_cache_destroy(ext4_pspace_cachep); | |
2825 | return -ENOMEM; | |
2826 | } | |
c894058d | 2827 | |
18aadd47 BJ |
2828 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
2829 | SLAB_RECLAIM_ACCOUNT); | |
2830 | if (ext4_free_data_cachep == NULL) { | |
c894058d AK |
2831 | kmem_cache_destroy(ext4_pspace_cachep); |
2832 | kmem_cache_destroy(ext4_ac_cachep); | |
2833 | return -ENOMEM; | |
2834 | } | |
c9de560d AT |
2835 | return 0; |
2836 | } | |
2837 | ||
5dabfc78 | 2838 | void ext4_exit_mballoc(void) |
c9de560d | 2839 | { |
60e6679e | 2840 | /* |
3e03f9ca JDB |
2841 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
2842 | * before destroying the slab cache. | |
2843 | */ | |
2844 | rcu_barrier(); | |
c9de560d | 2845 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 2846 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 2847 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 2848 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
2849 | } |
2850 | ||
2851 | ||
2852 | /* | |
73b2c716 | 2853 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
2854 | * Returns 0 if success or error code |
2855 | */ | |
4ddfef7b ES |
2856 | static noinline_for_stack int |
2857 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 2858 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
2859 | { |
2860 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
2861 | struct ext4_group_desc *gdp; |
2862 | struct buffer_head *gdp_bh; | |
2863 | struct ext4_sb_info *sbi; | |
2864 | struct super_block *sb; | |
2865 | ext4_fsblk_t block; | |
519deca0 | 2866 | int err, len; |
c9de560d AT |
2867 | |
2868 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2869 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2870 | ||
2871 | sb = ac->ac_sb; | |
2872 | sbi = EXT4_SB(sb); | |
c9de560d AT |
2873 | |
2874 | err = -EIO; | |
574ca174 | 2875 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
c9de560d AT |
2876 | if (!bitmap_bh) |
2877 | goto out_err; | |
2878 | ||
5d601255 | 2879 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
c9de560d AT |
2880 | err = ext4_journal_get_write_access(handle, bitmap_bh); |
2881 | if (err) | |
2882 | goto out_err; | |
2883 | ||
2884 | err = -EIO; | |
2885 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
2886 | if (!gdp) | |
2887 | goto out_err; | |
2888 | ||
a9df9a49 | 2889 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 2890 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 2891 | |
5d601255 | 2892 | BUFFER_TRACE(gdp_bh, "get_write_access"); |
c9de560d AT |
2893 | err = ext4_journal_get_write_access(handle, gdp_bh); |
2894 | if (err) | |
2895 | goto out_err; | |
2896 | ||
bda00de7 | 2897 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 2898 | |
53accfa9 | 2899 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
6fd058f7 | 2900 | if (!ext4_data_block_valid(sbi, block, len)) { |
12062ddd | 2901 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 2902 | "fs metadata", block, block+len); |
519deca0 AK |
2903 | /* File system mounted not to panic on error |
2904 | * Fix the bitmap and repeat the block allocation | |
2905 | * We leak some of the blocks here. | |
2906 | */ | |
955ce5f5 | 2907 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
c3e94d1d YY |
2908 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
2909 | ac->ac_b_ex.fe_len); | |
955ce5f5 | 2910 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 2911 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 AK |
2912 | if (!err) |
2913 | err = -EAGAIN; | |
2914 | goto out_err; | |
c9de560d | 2915 | } |
955ce5f5 AK |
2916 | |
2917 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
2918 | #ifdef AGGRESSIVE_CHECK |
2919 | { | |
2920 | int i; | |
2921 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
2922 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
2923 | bitmap_bh->b_data)); | |
2924 | } | |
2925 | } | |
2926 | #endif | |
c3e94d1d YY |
2927 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
2928 | ac->ac_b_ex.fe_len); | |
c9de560d AT |
2929 | if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
2930 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
021b65bb | 2931 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 2932 | ext4_free_clusters_after_init(sb, |
021b65bb | 2933 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 2934 | } |
021b65bb TT |
2935 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
2936 | ext4_free_group_clusters_set(sb, gdp, len); | |
79f1ba49 | 2937 | ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); |
feb0ab32 | 2938 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
2939 | |
2940 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 2941 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 2942 | /* |
6bc6e63f | 2943 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 2944 | */ |
6bc6e63f AK |
2945 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
2946 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
2947 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
2948 | reserv_clstrs); | |
c9de560d | 2949 | |
772cb7c8 JS |
2950 | if (sbi->s_log_groups_per_flex) { |
2951 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
2952 | ac->ac_b_ex.fe_group); | |
90ba983f TT |
2953 | atomic64_sub(ac->ac_b_ex.fe_len, |
2954 | &sbi->s_flex_groups[flex_group].free_clusters); | |
772cb7c8 JS |
2955 | } |
2956 | ||
0390131b | 2957 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
2958 | if (err) |
2959 | goto out_err; | |
0390131b | 2960 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
2961 | |
2962 | out_err: | |
42a10add | 2963 | brelse(bitmap_bh); |
c9de560d AT |
2964 | return err; |
2965 | } | |
2966 | ||
2967 | /* | |
2968 | * here we normalize request for locality group | |
d7a1fee1 DE |
2969 | * Group request are normalized to s_mb_group_prealloc, which goes to |
2970 | * s_strip if we set the same via mount option. | |
2971 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 2972 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
2973 | * |
2974 | * XXX: should we try to preallocate more than the group has now? | |
2975 | */ | |
2976 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
2977 | { | |
2978 | struct super_block *sb = ac->ac_sb; | |
2979 | struct ext4_locality_group *lg = ac->ac_lg; | |
2980 | ||
2981 | BUG_ON(lg == NULL); | |
d7a1fee1 | 2982 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
6ba495e9 | 2983 | mb_debug(1, "#%u: goal %u blocks for locality group\n", |
c9de560d AT |
2984 | current->pid, ac->ac_g_ex.fe_len); |
2985 | } | |
2986 | ||
2987 | /* | |
2988 | * Normalization means making request better in terms of | |
2989 | * size and alignment | |
2990 | */ | |
4ddfef7b ES |
2991 | static noinline_for_stack void |
2992 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
2993 | struct ext4_allocation_request *ar) |
2994 | { | |
53accfa9 | 2995 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
2996 | int bsbits, max; |
2997 | ext4_lblk_t end; | |
1592d2c5 CW |
2998 | loff_t size, start_off; |
2999 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 3000 | ext4_lblk_t start; |
c9de560d | 3001 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 3002 | struct ext4_prealloc_space *pa; |
c9de560d AT |
3003 | |
3004 | /* do normalize only data requests, metadata requests | |
3005 | do not need preallocation */ | |
3006 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3007 | return; | |
3008 | ||
3009 | /* sometime caller may want exact blocks */ | |
3010 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
3011 | return; | |
3012 | ||
3013 | /* caller may indicate that preallocation isn't | |
3014 | * required (it's a tail, for example) */ | |
3015 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
3016 | return; | |
3017 | ||
3018 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
3019 | ext4_mb_normalize_group_request(ac); | |
3020 | return ; | |
3021 | } | |
3022 | ||
3023 | bsbits = ac->ac_sb->s_blocksize_bits; | |
3024 | ||
3025 | /* first, let's learn actual file size | |
3026 | * given current request is allocated */ | |
53accfa9 | 3027 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
3028 | size = size << bsbits; |
3029 | if (size < i_size_read(ac->ac_inode)) | |
3030 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 3031 | orig_size = size; |
c9de560d | 3032 | |
1930479c VC |
3033 | /* max size of free chunks */ |
3034 | max = 2 << bsbits; | |
c9de560d | 3035 | |
1930479c VC |
3036 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
3037 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
3038 | |
3039 | /* first, try to predict filesize */ | |
3040 | /* XXX: should this table be tunable? */ | |
3041 | start_off = 0; | |
3042 | if (size <= 16 * 1024) { | |
3043 | size = 16 * 1024; | |
3044 | } else if (size <= 32 * 1024) { | |
3045 | size = 32 * 1024; | |
3046 | } else if (size <= 64 * 1024) { | |
3047 | size = 64 * 1024; | |
3048 | } else if (size <= 128 * 1024) { | |
3049 | size = 128 * 1024; | |
3050 | } else if (size <= 256 * 1024) { | |
3051 | size = 256 * 1024; | |
3052 | } else if (size <= 512 * 1024) { | |
3053 | size = 512 * 1024; | |
3054 | } else if (size <= 1024 * 1024) { | |
3055 | size = 1024 * 1024; | |
1930479c | 3056 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 3057 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
3058 | (21 - bsbits)) << 21; |
3059 | size = 2 * 1024 * 1024; | |
3060 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
3061 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3062 | (22 - bsbits)) << 22; | |
3063 | size = 4 * 1024 * 1024; | |
3064 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 3065 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
3066 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3067 | (23 - bsbits)) << 23; | |
3068 | size = 8 * 1024 * 1024; | |
3069 | } else { | |
3070 | start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; | |
3071 | size = ac->ac_o_ex.fe_len << bsbits; | |
3072 | } | |
5a0790c2 AK |
3073 | size = size >> bsbits; |
3074 | start = start_off >> bsbits; | |
c9de560d AT |
3075 | |
3076 | /* don't cover already allocated blocks in selected range */ | |
3077 | if (ar->pleft && start <= ar->lleft) { | |
3078 | size -= ar->lleft + 1 - start; | |
3079 | start = ar->lleft + 1; | |
3080 | } | |
3081 | if (ar->pright && start + size - 1 >= ar->lright) | |
3082 | size -= start + size - ar->lright; | |
3083 | ||
3084 | end = start + size; | |
3085 | ||
3086 | /* check we don't cross already preallocated blocks */ | |
3087 | rcu_read_lock(); | |
9a0762c5 | 3088 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3089 | ext4_lblk_t pa_end; |
c9de560d | 3090 | |
c9de560d AT |
3091 | if (pa->pa_deleted) |
3092 | continue; | |
3093 | spin_lock(&pa->pa_lock); | |
3094 | if (pa->pa_deleted) { | |
3095 | spin_unlock(&pa->pa_lock); | |
3096 | continue; | |
3097 | } | |
3098 | ||
53accfa9 TT |
3099 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3100 | pa->pa_len); | |
c9de560d AT |
3101 | |
3102 | /* PA must not overlap original request */ | |
3103 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
3104 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
3105 | ||
38877f4e ES |
3106 | /* skip PAs this normalized request doesn't overlap with */ |
3107 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
3108 | spin_unlock(&pa->pa_lock); |
3109 | continue; | |
3110 | } | |
3111 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3112 | ||
38877f4e | 3113 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
3114 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
3115 | BUG_ON(pa_end < start); | |
3116 | start = pa_end; | |
38877f4e | 3117 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
3118 | BUG_ON(pa->pa_lstart > end); |
3119 | end = pa->pa_lstart; | |
3120 | } | |
3121 | spin_unlock(&pa->pa_lock); | |
3122 | } | |
3123 | rcu_read_unlock(); | |
3124 | size = end - start; | |
3125 | ||
3126 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3127 | rcu_read_lock(); | |
9a0762c5 | 3128 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3129 | ext4_lblk_t pa_end; |
53accfa9 | 3130 | |
c9de560d AT |
3131 | spin_lock(&pa->pa_lock); |
3132 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
3133 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3134 | pa->pa_len); | |
c9de560d AT |
3135 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
3136 | } | |
3137 | spin_unlock(&pa->pa_lock); | |
3138 | } | |
3139 | rcu_read_unlock(); | |
3140 | ||
3141 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3142 | start > ac->ac_o_ex.fe_logical) { | |
9d8b9ec4 TT |
3143 | ext4_msg(ac->ac_sb, KERN_ERR, |
3144 | "start %lu, size %lu, fe_logical %lu", | |
3145 | (unsigned long) start, (unsigned long) size, | |
3146 | (unsigned long) ac->ac_o_ex.fe_logical); | |
c9de560d AT |
3147 | } |
3148 | BUG_ON(start + size <= ac->ac_o_ex.fe_logical && | |
3149 | start > ac->ac_o_ex.fe_logical); | |
b5b60778 | 3150 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
3151 | |
3152 | /* now prepare goal request */ | |
3153 | ||
3154 | /* XXX: is it better to align blocks WRT to logical | |
3155 | * placement or satisfy big request as is */ | |
3156 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 3157 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
3158 | |
3159 | /* define goal start in order to merge */ | |
3160 | if (ar->pright && (ar->lright == (start + size))) { | |
3161 | /* merge to the right */ | |
3162 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3163 | &ac->ac_f_ex.fe_group, | |
3164 | &ac->ac_f_ex.fe_start); | |
3165 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3166 | } | |
3167 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3168 | /* merge to the left */ | |
3169 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3170 | &ac->ac_f_ex.fe_group, | |
3171 | &ac->ac_f_ex.fe_start); | |
3172 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3173 | } | |
3174 | ||
6ba495e9 | 3175 | mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, |
c9de560d AT |
3176 | (unsigned) orig_size, (unsigned) start); |
3177 | } | |
3178 | ||
3179 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3180 | { | |
3181 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3182 | ||
3183 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3184 | atomic_inc(&sbi->s_bal_reqs); | |
3185 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 3186 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
3187 | atomic_inc(&sbi->s_bal_success); |
3188 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3189 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3190 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3191 | atomic_inc(&sbi->s_bal_goals); | |
3192 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3193 | atomic_inc(&sbi->s_bal_breaks); | |
3194 | } | |
3195 | ||
296c355c TT |
3196 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
3197 | trace_ext4_mballoc_alloc(ac); | |
3198 | else | |
3199 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
3200 | } |
3201 | ||
b844167e CW |
3202 | /* |
3203 | * Called on failure; free up any blocks from the inode PA for this | |
3204 | * context. We don't need this for MB_GROUP_PA because we only change | |
3205 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
3206 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
3207 | */ | |
3208 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
3209 | { | |
3210 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
b844167e | 3211 | |
400db9d3 ZL |
3212 | if (pa && pa->pa_type == MB_INODE_PA) |
3213 | pa->pa_free += ac->ac_b_ex.fe_len; | |
b844167e CW |
3214 | } |
3215 | ||
c9de560d AT |
3216 | /* |
3217 | * use blocks preallocated to inode | |
3218 | */ | |
3219 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3220 | struct ext4_prealloc_space *pa) | |
3221 | { | |
53accfa9 | 3222 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3223 | ext4_fsblk_t start; |
3224 | ext4_fsblk_t end; | |
3225 | int len; | |
3226 | ||
3227 | /* found preallocated blocks, use them */ | |
3228 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
3229 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
3230 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
3231 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
3232 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
3233 | &ac->ac_b_ex.fe_start); | |
3234 | ac->ac_b_ex.fe_len = len; | |
3235 | ac->ac_status = AC_STATUS_FOUND; | |
3236 | ac->ac_pa = pa; | |
3237 | ||
3238 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 3239 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
3240 | BUG_ON(pa->pa_free < len); |
3241 | pa->pa_free -= len; | |
3242 | ||
6ba495e9 | 3243 | mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); |
c9de560d AT |
3244 | } |
3245 | ||
3246 | /* | |
3247 | * use blocks preallocated to locality group | |
3248 | */ | |
3249 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3250 | struct ext4_prealloc_space *pa) | |
3251 | { | |
03cddb80 | 3252 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3253 | |
c9de560d AT |
3254 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3255 | &ac->ac_b_ex.fe_group, | |
3256 | &ac->ac_b_ex.fe_start); | |
3257 | ac->ac_b_ex.fe_len = len; | |
3258 | ac->ac_status = AC_STATUS_FOUND; | |
3259 | ac->ac_pa = pa; | |
3260 | ||
3261 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3262 | * possible race when the group is being loaded concurrently |
c9de560d | 3263 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3264 | * in on-disk bitmap -- see ext4_mb_release_context() |
3265 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 3266 | */ |
6ba495e9 | 3267 | mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); |
c9de560d AT |
3268 | } |
3269 | ||
5e745b04 AK |
3270 | /* |
3271 | * Return the prealloc space that have minimal distance | |
3272 | * from the goal block. @cpa is the prealloc | |
3273 | * space that is having currently known minimal distance | |
3274 | * from the goal block. | |
3275 | */ | |
3276 | static struct ext4_prealloc_space * | |
3277 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3278 | struct ext4_prealloc_space *pa, | |
3279 | struct ext4_prealloc_space *cpa) | |
3280 | { | |
3281 | ext4_fsblk_t cur_distance, new_distance; | |
3282 | ||
3283 | if (cpa == NULL) { | |
3284 | atomic_inc(&pa->pa_count); | |
3285 | return pa; | |
3286 | } | |
3287 | cur_distance = abs(goal_block - cpa->pa_pstart); | |
3288 | new_distance = abs(goal_block - pa->pa_pstart); | |
3289 | ||
5a54b2f1 | 3290 | if (cur_distance <= new_distance) |
5e745b04 AK |
3291 | return cpa; |
3292 | ||
3293 | /* drop the previous reference */ | |
3294 | atomic_dec(&cpa->pa_count); | |
3295 | atomic_inc(&pa->pa_count); | |
3296 | return pa; | |
3297 | } | |
3298 | ||
c9de560d AT |
3299 | /* |
3300 | * search goal blocks in preallocated space | |
3301 | */ | |
4ddfef7b ES |
3302 | static noinline_for_stack int |
3303 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
c9de560d | 3304 | { |
53accfa9 | 3305 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 3306 | int order, i; |
c9de560d AT |
3307 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3308 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3309 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3310 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3311 | |
3312 | /* only data can be preallocated */ | |
3313 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3314 | return 0; | |
3315 | ||
3316 | /* first, try per-file preallocation */ | |
3317 | rcu_read_lock(); | |
9a0762c5 | 3318 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3319 | |
3320 | /* all fields in this condition don't change, | |
3321 | * so we can skip locking for them */ | |
3322 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
3323 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
3324 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
3325 | continue; |
3326 | ||
fb0a387d | 3327 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 3328 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
3329 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
3330 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
3331 | continue; |
3332 | ||
c9de560d AT |
3333 | /* found preallocated blocks, use them */ |
3334 | spin_lock(&pa->pa_lock); | |
3335 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3336 | atomic_inc(&pa->pa_count); | |
3337 | ext4_mb_use_inode_pa(ac, pa); | |
3338 | spin_unlock(&pa->pa_lock); | |
3339 | ac->ac_criteria = 10; | |
3340 | rcu_read_unlock(); | |
3341 | return 1; | |
3342 | } | |
3343 | spin_unlock(&pa->pa_lock); | |
3344 | } | |
3345 | rcu_read_unlock(); | |
3346 | ||
3347 | /* can we use group allocation? */ | |
3348 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3349 | return 0; | |
3350 | ||
3351 | /* inode may have no locality group for some reason */ | |
3352 | lg = ac->ac_lg; | |
3353 | if (lg == NULL) | |
3354 | return 0; | |
6be2ded1 AK |
3355 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3356 | if (order > PREALLOC_TB_SIZE - 1) | |
3357 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3358 | order = PREALLOC_TB_SIZE - 1; | |
3359 | ||
bda00de7 | 3360 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
3361 | /* |
3362 | * search for the prealloc space that is having | |
3363 | * minimal distance from the goal block. | |
3364 | */ | |
6be2ded1 AK |
3365 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3366 | rcu_read_lock(); | |
3367 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3368 | pa_inode_list) { | |
3369 | spin_lock(&pa->pa_lock); | |
3370 | if (pa->pa_deleted == 0 && | |
3371 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3372 | |
3373 | cpa = ext4_mb_check_group_pa(goal_block, | |
3374 | pa, cpa); | |
6be2ded1 | 3375 | } |
c9de560d | 3376 | spin_unlock(&pa->pa_lock); |
c9de560d | 3377 | } |
6be2ded1 | 3378 | rcu_read_unlock(); |
c9de560d | 3379 | } |
5e745b04 AK |
3380 | if (cpa) { |
3381 | ext4_mb_use_group_pa(ac, cpa); | |
3382 | ac->ac_criteria = 20; | |
3383 | return 1; | |
3384 | } | |
c9de560d AT |
3385 | return 0; |
3386 | } | |
3387 | ||
7a2fcbf7 AK |
3388 | /* |
3389 | * the function goes through all block freed in the group | |
3390 | * but not yet committed and marks them used in in-core bitmap. | |
3391 | * buddy must be generated from this bitmap | |
955ce5f5 | 3392 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3393 | */ |
3394 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3395 | ext4_group_t group) | |
3396 | { | |
3397 | struct rb_node *n; | |
3398 | struct ext4_group_info *grp; | |
3399 | struct ext4_free_data *entry; | |
3400 | ||
3401 | grp = ext4_get_group_info(sb, group); | |
3402 | n = rb_first(&(grp->bb_free_root)); | |
3403 | ||
3404 | while (n) { | |
18aadd47 BJ |
3405 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
3406 | ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); | |
7a2fcbf7 AK |
3407 | n = rb_next(n); |
3408 | } | |
3409 | return; | |
3410 | } | |
3411 | ||
c9de560d AT |
3412 | /* |
3413 | * the function goes through all preallocation in this group and marks them | |
3414 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3415 | * Need to be called with ext4 group lock held |
c9de560d | 3416 | */ |
089ceecc ES |
3417 | static noinline_for_stack |
3418 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
3419 | ext4_group_t group) |
3420 | { | |
3421 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3422 | struct ext4_prealloc_space *pa; | |
3423 | struct list_head *cur; | |
3424 | ext4_group_t groupnr; | |
3425 | ext4_grpblk_t start; | |
3426 | int preallocated = 0; | |
c9de560d AT |
3427 | int len; |
3428 | ||
3429 | /* all form of preallocation discards first load group, | |
3430 | * so the only competing code is preallocation use. | |
3431 | * we don't need any locking here | |
3432 | * notice we do NOT ignore preallocations with pa_deleted | |
3433 | * otherwise we could leave used blocks available for | |
3434 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3435 | * is dropping preallocation | |
3436 | */ | |
3437 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3438 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3439 | spin_lock(&pa->pa_lock); | |
3440 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3441 | &groupnr, &start); | |
3442 | len = pa->pa_len; | |
3443 | spin_unlock(&pa->pa_lock); | |
3444 | if (unlikely(len == 0)) | |
3445 | continue; | |
3446 | BUG_ON(groupnr != group); | |
c3e94d1d | 3447 | ext4_set_bits(bitmap, start, len); |
c9de560d | 3448 | preallocated += len; |
c9de560d | 3449 | } |
6ba495e9 | 3450 | mb_debug(1, "prellocated %u for group %u\n", preallocated, group); |
c9de560d AT |
3451 | } |
3452 | ||
3453 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3454 | { | |
3455 | struct ext4_prealloc_space *pa; | |
3456 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
4e8d2139 JR |
3457 | |
3458 | BUG_ON(atomic_read(&pa->pa_count)); | |
3459 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
3460 | kmem_cache_free(ext4_pspace_cachep, pa); |
3461 | } | |
3462 | ||
3463 | /* | |
3464 | * drops a reference to preallocated space descriptor | |
3465 | * if this was the last reference and the space is consumed | |
3466 | */ | |
3467 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3468 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3469 | { | |
a9df9a49 | 3470 | ext4_group_t grp; |
d33a1976 | 3471 | ext4_fsblk_t grp_blk; |
c9de560d | 3472 | |
c9de560d AT |
3473 | /* in this short window concurrent discard can set pa_deleted */ |
3474 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
3475 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
3476 | spin_unlock(&pa->pa_lock); | |
3477 | return; | |
3478 | } | |
3479 | ||
c9de560d AT |
3480 | if (pa->pa_deleted == 1) { |
3481 | spin_unlock(&pa->pa_lock); | |
3482 | return; | |
3483 | } | |
3484 | ||
3485 | pa->pa_deleted = 1; | |
3486 | spin_unlock(&pa->pa_lock); | |
3487 | ||
d33a1976 | 3488 | grp_blk = pa->pa_pstart; |
60e6679e | 3489 | /* |
cc0fb9ad AK |
3490 | * If doing group-based preallocation, pa_pstart may be in the |
3491 | * next group when pa is used up | |
3492 | */ | |
3493 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3494 | grp_blk--; |
3495 | ||
bd86298e | 3496 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
3497 | |
3498 | /* | |
3499 | * possible race: | |
3500 | * | |
3501 | * P1 (buddy init) P2 (regular allocation) | |
3502 | * find block B in PA | |
3503 | * copy on-disk bitmap to buddy | |
3504 | * mark B in on-disk bitmap | |
3505 | * drop PA from group | |
3506 | * mark all PAs in buddy | |
3507 | * | |
3508 | * thus, P1 initializes buddy with B available. to prevent this | |
3509 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3510 | * against that pair | |
3511 | */ | |
3512 | ext4_lock_group(sb, grp); | |
3513 | list_del(&pa->pa_group_list); | |
3514 | ext4_unlock_group(sb, grp); | |
3515 | ||
3516 | spin_lock(pa->pa_obj_lock); | |
3517 | list_del_rcu(&pa->pa_inode_list); | |
3518 | spin_unlock(pa->pa_obj_lock); | |
3519 | ||
3520 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3521 | } | |
3522 | ||
3523 | /* | |
3524 | * creates new preallocated space for given inode | |
3525 | */ | |
4ddfef7b ES |
3526 | static noinline_for_stack int |
3527 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3528 | { |
3529 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 3530 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
3531 | struct ext4_prealloc_space *pa; |
3532 | struct ext4_group_info *grp; | |
3533 | struct ext4_inode_info *ei; | |
3534 | ||
3535 | /* preallocate only when found space is larger then requested */ | |
3536 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3537 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3538 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3539 | ||
3540 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3541 | if (pa == NULL) | |
3542 | return -ENOMEM; | |
3543 | ||
3544 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3545 | int winl; | |
3546 | int wins; | |
3547 | int win; | |
3548 | int offs; | |
3549 | ||
3550 | /* we can't allocate as much as normalizer wants. | |
3551 | * so, found space must get proper lstart | |
3552 | * to cover original request */ | |
3553 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3554 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3555 | ||
3556 | /* we're limited by original request in that | |
3557 | * logical block must be covered any way | |
3558 | * winl is window we can move our chunk within */ | |
3559 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3560 | ||
3561 | /* also, we should cover whole original request */ | |
53accfa9 | 3562 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
3563 | |
3564 | /* the smallest one defines real window */ | |
3565 | win = min(winl, wins); | |
3566 | ||
53accfa9 TT |
3567 | offs = ac->ac_o_ex.fe_logical % |
3568 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
3569 | if (offs && offs < win) |
3570 | win = offs; | |
3571 | ||
53accfa9 | 3572 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
810da240 | 3573 | EXT4_NUM_B2C(sbi, win); |
c9de560d AT |
3574 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
3575 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3576 | } | |
3577 | ||
3578 | /* preallocation can change ac_b_ex, thus we store actually | |
3579 | * allocated blocks for history */ | |
3580 | ac->ac_f_ex = ac->ac_b_ex; | |
3581 | ||
3582 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3583 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3584 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3585 | pa->pa_free = pa->pa_len; | |
3586 | atomic_set(&pa->pa_count, 1); | |
3587 | spin_lock_init(&pa->pa_lock); | |
d794bf8e AK |
3588 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3589 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3590 | pa->pa_deleted = 0; |
cc0fb9ad | 3591 | pa->pa_type = MB_INODE_PA; |
c9de560d | 3592 | |
6ba495e9 | 3593 | mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, |
c9de560d | 3594 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
9bffad1e | 3595 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3596 | |
3597 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 3598 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
3599 | |
3600 | ei = EXT4_I(ac->ac_inode); | |
3601 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3602 | ||
3603 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3604 | pa->pa_inode = ac->ac_inode; | |
3605 | ||
3606 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3607 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3608 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3609 | ||
3610 | spin_lock(pa->pa_obj_lock); | |
3611 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3612 | spin_unlock(pa->pa_obj_lock); | |
3613 | ||
3614 | return 0; | |
3615 | } | |
3616 | ||
3617 | /* | |
3618 | * creates new preallocated space for locality group inodes belongs to | |
3619 | */ | |
4ddfef7b ES |
3620 | static noinline_for_stack int |
3621 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3622 | { |
3623 | struct super_block *sb = ac->ac_sb; | |
3624 | struct ext4_locality_group *lg; | |
3625 | struct ext4_prealloc_space *pa; | |
3626 | struct ext4_group_info *grp; | |
3627 | ||
3628 | /* preallocate only when found space is larger then requested */ | |
3629 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3630 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3631 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3632 | ||
3633 | BUG_ON(ext4_pspace_cachep == NULL); | |
3634 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3635 | if (pa == NULL) | |
3636 | return -ENOMEM; | |
3637 | ||
3638 | /* preallocation can change ac_b_ex, thus we store actually | |
3639 | * allocated blocks for history */ | |
3640 | ac->ac_f_ex = ac->ac_b_ex; | |
3641 | ||
3642 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3643 | pa->pa_lstart = pa->pa_pstart; | |
3644 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3645 | pa->pa_free = pa->pa_len; | |
3646 | atomic_set(&pa->pa_count, 1); | |
3647 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3648 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 3649 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 3650 | pa->pa_deleted = 0; |
cc0fb9ad | 3651 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 3652 | |
6ba495e9 | 3653 | mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, |
9bffad1e TT |
3654 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
3655 | trace_ext4_mb_new_group_pa(ac, pa); | |
c9de560d AT |
3656 | |
3657 | ext4_mb_use_group_pa(ac, pa); | |
3658 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3659 | ||
3660 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3661 | lg = ac->ac_lg; | |
3662 | BUG_ON(lg == NULL); | |
3663 | ||
3664 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3665 | pa->pa_inode = NULL; | |
3666 | ||
3667 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3668 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3669 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3670 | ||
6be2ded1 AK |
3671 | /* |
3672 | * We will later add the new pa to the right bucket | |
3673 | * after updating the pa_free in ext4_mb_release_context | |
3674 | */ | |
c9de560d AT |
3675 | return 0; |
3676 | } | |
3677 | ||
3678 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3679 | { | |
3680 | int err; | |
3681 | ||
3682 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3683 | err = ext4_mb_new_group_pa(ac); | |
3684 | else | |
3685 | err = ext4_mb_new_inode_pa(ac); | |
3686 | return err; | |
3687 | } | |
3688 | ||
3689 | /* | |
3690 | * finds all unused blocks in on-disk bitmap, frees them in | |
3691 | * in-core bitmap and buddy. | |
3692 | * @pa must be unlinked from inode and group lists, so that | |
3693 | * nobody else can find/use it. | |
3694 | * the caller MUST hold group/inode locks. | |
3695 | * TODO: optimize the case when there are no in-core structures yet | |
3696 | */ | |
4ddfef7b ES |
3697 | static noinline_for_stack int |
3698 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 3699 | struct ext4_prealloc_space *pa) |
c9de560d | 3700 | { |
c9de560d AT |
3701 | struct super_block *sb = e4b->bd_sb; |
3702 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
3703 | unsigned int end; |
3704 | unsigned int next; | |
c9de560d AT |
3705 | ext4_group_t group; |
3706 | ext4_grpblk_t bit; | |
ba80b101 | 3707 | unsigned long long grp_blk_start; |
c9de560d AT |
3708 | int err = 0; |
3709 | int free = 0; | |
3710 | ||
3711 | BUG_ON(pa->pa_deleted == 0); | |
3712 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 3713 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
3714 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
3715 | end = bit + pa->pa_len; | |
3716 | ||
c9de560d | 3717 | while (bit < end) { |
ffad0a44 | 3718 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3719 | if (bit >= end) |
3720 | break; | |
ffad0a44 | 3721 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
6ba495e9 | 3722 | mb_debug(1, " free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
3723 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
3724 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
3725 | free += next - bit; |
3726 | ||
3e1e5f50 | 3727 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
3728 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
3729 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 3730 | next - bit); |
c9de560d AT |
3731 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
3732 | bit = next + 1; | |
3733 | } | |
3734 | if (free != pa->pa_free) { | |
9d8b9ec4 TT |
3735 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
3736 | "pa %p: logic %lu, phys. %lu, len %lu", | |
3737 | pa, (unsigned long) pa->pa_lstart, | |
3738 | (unsigned long) pa->pa_pstart, | |
3739 | (unsigned long) pa->pa_len); | |
e29136f8 | 3740 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 3741 | free, pa->pa_free); |
e56eb659 AK |
3742 | /* |
3743 | * pa is already deleted so we use the value obtained | |
3744 | * from the bitmap and continue. | |
3745 | */ | |
c9de560d | 3746 | } |
c9de560d AT |
3747 | atomic_add(free, &sbi->s_mb_discarded); |
3748 | ||
3749 | return err; | |
3750 | } | |
3751 | ||
4ddfef7b ES |
3752 | static noinline_for_stack int |
3753 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 3754 | struct ext4_prealloc_space *pa) |
c9de560d | 3755 | { |
c9de560d AT |
3756 | struct super_block *sb = e4b->bd_sb; |
3757 | ext4_group_t group; | |
3758 | ext4_grpblk_t bit; | |
3759 | ||
60e07cf5 | 3760 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
3761 | BUG_ON(pa->pa_deleted == 0); |
3762 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3763 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3764 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3765 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 3766 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
3767 | |
3768 | return 0; | |
3769 | } | |
3770 | ||
3771 | /* | |
3772 | * releases all preallocations in given group | |
3773 | * | |
3774 | * first, we need to decide discard policy: | |
3775 | * - when do we discard | |
3776 | * 1) ENOSPC | |
3777 | * - how many do we discard | |
3778 | * 1) how many requested | |
3779 | */ | |
4ddfef7b ES |
3780 | static noinline_for_stack int |
3781 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3782 | ext4_group_t group, int needed) |
3783 | { | |
3784 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3785 | struct buffer_head *bitmap_bh = NULL; | |
3786 | struct ext4_prealloc_space *pa, *tmp; | |
3787 | struct list_head list; | |
3788 | struct ext4_buddy e4b; | |
3789 | int err; | |
3790 | int busy = 0; | |
3791 | int free = 0; | |
3792 | ||
6ba495e9 | 3793 | mb_debug(1, "discard preallocation for group %u\n", group); |
c9de560d AT |
3794 | |
3795 | if (list_empty(&grp->bb_prealloc_list)) | |
3796 | return 0; | |
3797 | ||
574ca174 | 3798 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3799 | if (bitmap_bh == NULL) { |
12062ddd | 3800 | ext4_error(sb, "Error reading block bitmap for %u", group); |
ce89f46c | 3801 | return 0; |
c9de560d AT |
3802 | } |
3803 | ||
3804 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3805 | if (err) { |
12062ddd | 3806 | ext4_error(sb, "Error loading buddy information for %u", group); |
ce89f46c AK |
3807 | put_bh(bitmap_bh); |
3808 | return 0; | |
3809 | } | |
c9de560d AT |
3810 | |
3811 | if (needed == 0) | |
7137d7a4 | 3812 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; |
c9de560d | 3813 | |
c9de560d | 3814 | INIT_LIST_HEAD(&list); |
c9de560d AT |
3815 | repeat: |
3816 | ext4_lock_group(sb, group); | |
3817 | list_for_each_entry_safe(pa, tmp, | |
3818 | &grp->bb_prealloc_list, pa_group_list) { | |
3819 | spin_lock(&pa->pa_lock); | |
3820 | if (atomic_read(&pa->pa_count)) { | |
3821 | spin_unlock(&pa->pa_lock); | |
3822 | busy = 1; | |
3823 | continue; | |
3824 | } | |
3825 | if (pa->pa_deleted) { | |
3826 | spin_unlock(&pa->pa_lock); | |
3827 | continue; | |
3828 | } | |
3829 | ||
3830 | /* seems this one can be freed ... */ | |
3831 | pa->pa_deleted = 1; | |
3832 | ||
3833 | /* we can trust pa_free ... */ | |
3834 | free += pa->pa_free; | |
3835 | ||
3836 | spin_unlock(&pa->pa_lock); | |
3837 | ||
3838 | list_del(&pa->pa_group_list); | |
3839 | list_add(&pa->u.pa_tmp_list, &list); | |
3840 | } | |
3841 | ||
3842 | /* if we still need more blocks and some PAs were used, try again */ | |
3843 | if (free < needed && busy) { | |
3844 | busy = 0; | |
3845 | ext4_unlock_group(sb, group); | |
bb8b20ed | 3846 | cond_resched(); |
c9de560d AT |
3847 | goto repeat; |
3848 | } | |
3849 | ||
3850 | /* found anything to free? */ | |
3851 | if (list_empty(&list)) { | |
3852 | BUG_ON(free != 0); | |
3853 | goto out; | |
3854 | } | |
3855 | ||
3856 | /* now free all selected PAs */ | |
3857 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3858 | ||
3859 | /* remove from object (inode or locality group) */ | |
3860 | spin_lock(pa->pa_obj_lock); | |
3861 | list_del_rcu(&pa->pa_inode_list); | |
3862 | spin_unlock(pa->pa_obj_lock); | |
3863 | ||
cc0fb9ad | 3864 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 3865 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 3866 | else |
3e1e5f50 | 3867 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
3868 | |
3869 | list_del(&pa->u.pa_tmp_list); | |
3870 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3871 | } | |
3872 | ||
3873 | out: | |
3874 | ext4_unlock_group(sb, group); | |
e39e07fd | 3875 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
3876 | put_bh(bitmap_bh); |
3877 | return free; | |
3878 | } | |
3879 | ||
3880 | /* | |
3881 | * releases all non-used preallocated blocks for given inode | |
3882 | * | |
3883 | * It's important to discard preallocations under i_data_sem | |
3884 | * We don't want another block to be served from the prealloc | |
3885 | * space when we are discarding the inode prealloc space. | |
3886 | * | |
3887 | * FIXME!! Make sure it is valid at all the call sites | |
3888 | */ | |
c2ea3fde | 3889 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
3890 | { |
3891 | struct ext4_inode_info *ei = EXT4_I(inode); | |
3892 | struct super_block *sb = inode->i_sb; | |
3893 | struct buffer_head *bitmap_bh = NULL; | |
3894 | struct ext4_prealloc_space *pa, *tmp; | |
3895 | ext4_group_t group = 0; | |
3896 | struct list_head list; | |
3897 | struct ext4_buddy e4b; | |
3898 | int err; | |
3899 | ||
c2ea3fde | 3900 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
3901 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
3902 | return; | |
3903 | } | |
3904 | ||
6ba495e9 | 3905 | mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); |
9bffad1e | 3906 | trace_ext4_discard_preallocations(inode); |
c9de560d AT |
3907 | |
3908 | INIT_LIST_HEAD(&list); | |
3909 | ||
3910 | repeat: | |
3911 | /* first, collect all pa's in the inode */ | |
3912 | spin_lock(&ei->i_prealloc_lock); | |
3913 | while (!list_empty(&ei->i_prealloc_list)) { | |
3914 | pa = list_entry(ei->i_prealloc_list.next, | |
3915 | struct ext4_prealloc_space, pa_inode_list); | |
3916 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
3917 | spin_lock(&pa->pa_lock); | |
3918 | if (atomic_read(&pa->pa_count)) { | |
3919 | /* this shouldn't happen often - nobody should | |
3920 | * use preallocation while we're discarding it */ | |
3921 | spin_unlock(&pa->pa_lock); | |
3922 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
3923 | ext4_msg(sb, KERN_ERR, |
3924 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
3925 | WARN_ON(1); |
3926 | schedule_timeout_uninterruptible(HZ); | |
3927 | goto repeat; | |
3928 | ||
3929 | } | |
3930 | if (pa->pa_deleted == 0) { | |
3931 | pa->pa_deleted = 1; | |
3932 | spin_unlock(&pa->pa_lock); | |
3933 | list_del_rcu(&pa->pa_inode_list); | |
3934 | list_add(&pa->u.pa_tmp_list, &list); | |
3935 | continue; | |
3936 | } | |
3937 | ||
3938 | /* someone is deleting pa right now */ | |
3939 | spin_unlock(&pa->pa_lock); | |
3940 | spin_unlock(&ei->i_prealloc_lock); | |
3941 | ||
3942 | /* we have to wait here because pa_deleted | |
3943 | * doesn't mean pa is already unlinked from | |
3944 | * the list. as we might be called from | |
3945 | * ->clear_inode() the inode will get freed | |
3946 | * and concurrent thread which is unlinking | |
3947 | * pa from inode's list may access already | |
3948 | * freed memory, bad-bad-bad */ | |
3949 | ||
3950 | /* XXX: if this happens too often, we can | |
3951 | * add a flag to force wait only in case | |
3952 | * of ->clear_inode(), but not in case of | |
3953 | * regular truncate */ | |
3954 | schedule_timeout_uninterruptible(HZ); | |
3955 | goto repeat; | |
3956 | } | |
3957 | spin_unlock(&ei->i_prealloc_lock); | |
3958 | ||
3959 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 3960 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 3961 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d AT |
3962 | |
3963 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3964 | if (err) { |
12062ddd ES |
3965 | ext4_error(sb, "Error loading buddy information for %u", |
3966 | group); | |
ce89f46c AK |
3967 | continue; |
3968 | } | |
c9de560d | 3969 | |
574ca174 | 3970 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3971 | if (bitmap_bh == NULL) { |
12062ddd ES |
3972 | ext4_error(sb, "Error reading block bitmap for %u", |
3973 | group); | |
e39e07fd | 3974 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 3975 | continue; |
c9de560d AT |
3976 | } |
3977 | ||
3978 | ext4_lock_group(sb, group); | |
3979 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 3980 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
3981 | ext4_unlock_group(sb, group); |
3982 | ||
e39e07fd | 3983 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
3984 | put_bh(bitmap_bh); |
3985 | ||
3986 | list_del(&pa->u.pa_tmp_list); | |
3987 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3988 | } | |
3989 | } | |
3990 | ||
6ba495e9 | 3991 | #ifdef CONFIG_EXT4_DEBUG |
c9de560d AT |
3992 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
3993 | { | |
3994 | struct super_block *sb = ac->ac_sb; | |
8df9675f | 3995 | ext4_group_t ngroups, i; |
c9de560d | 3996 | |
a0b30c12 | 3997 | if (!ext4_mballoc_debug || |
4dd89fc6 | 3998 | (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) |
e3570639 ES |
3999 | return; |
4000 | ||
7f6a11e7 | 4001 | ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:" |
9d8b9ec4 | 4002 | " Allocation context details:"); |
7f6a11e7 | 4003 | ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d", |
c9de560d | 4004 | ac->ac_status, ac->ac_flags); |
7f6a11e7 | 4005 | ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, " |
9d8b9ec4 TT |
4006 | "goal %lu/%lu/%lu@%lu, " |
4007 | "best %lu/%lu/%lu@%lu cr %d", | |
c9de560d AT |
4008 | (unsigned long)ac->ac_o_ex.fe_group, |
4009 | (unsigned long)ac->ac_o_ex.fe_start, | |
4010 | (unsigned long)ac->ac_o_ex.fe_len, | |
4011 | (unsigned long)ac->ac_o_ex.fe_logical, | |
4012 | (unsigned long)ac->ac_g_ex.fe_group, | |
4013 | (unsigned long)ac->ac_g_ex.fe_start, | |
4014 | (unsigned long)ac->ac_g_ex.fe_len, | |
4015 | (unsigned long)ac->ac_g_ex.fe_logical, | |
4016 | (unsigned long)ac->ac_b_ex.fe_group, | |
4017 | (unsigned long)ac->ac_b_ex.fe_start, | |
4018 | (unsigned long)ac->ac_b_ex.fe_len, | |
4019 | (unsigned long)ac->ac_b_ex.fe_logical, | |
4020 | (int)ac->ac_criteria); | |
dc9ddd98 | 4021 | ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found); |
7f6a11e7 | 4022 | ext4_msg(ac->ac_sb, KERN_ERR, "groups: "); |
8df9675f TT |
4023 | ngroups = ext4_get_groups_count(sb); |
4024 | for (i = 0; i < ngroups; i++) { | |
c9de560d AT |
4025 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
4026 | struct ext4_prealloc_space *pa; | |
4027 | ext4_grpblk_t start; | |
4028 | struct list_head *cur; | |
4029 | ext4_lock_group(sb, i); | |
4030 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4031 | pa = list_entry(cur, struct ext4_prealloc_space, | |
4032 | pa_group_list); | |
4033 | spin_lock(&pa->pa_lock); | |
4034 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4035 | NULL, &start); | |
4036 | spin_unlock(&pa->pa_lock); | |
1c718505 AF |
4037 | printk(KERN_ERR "PA:%u:%d:%u \n", i, |
4038 | start, pa->pa_len); | |
c9de560d | 4039 | } |
60bd63d1 | 4040 | ext4_unlock_group(sb, i); |
c9de560d AT |
4041 | |
4042 | if (grp->bb_free == 0) | |
4043 | continue; | |
1c718505 | 4044 | printk(KERN_ERR "%u: %d/%d \n", |
c9de560d AT |
4045 | i, grp->bb_free, grp->bb_fragments); |
4046 | } | |
4047 | printk(KERN_ERR "\n"); | |
4048 | } | |
4049 | #else | |
4050 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
4051 | { | |
4052 | return; | |
4053 | } | |
4054 | #endif | |
4055 | ||
4056 | /* | |
4057 | * We use locality group preallocation for small size file. The size of the | |
4058 | * file is determined by the current size or the resulting size after | |
4059 | * allocation which ever is larger | |
4060 | * | |
b713a5ec | 4061 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
4062 | */ |
4063 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
4064 | { | |
4065 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4066 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
4067 | loff_t size, isize; | |
4068 | ||
4069 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4070 | return; | |
4071 | ||
4ba74d00 TT |
4072 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
4073 | return; | |
4074 | ||
53accfa9 | 4075 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
4076 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
4077 | >> bsbits; | |
c9de560d | 4078 | |
50797481 TT |
4079 | if ((size == isize) && |
4080 | !ext4_fs_is_busy(sbi) && | |
4081 | (atomic_read(&ac->ac_inode->i_writecount) == 0)) { | |
4082 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
4083 | return; | |
4084 | } | |
4085 | ||
ebbe0277 RD |
4086 | if (sbi->s_mb_group_prealloc <= 0) { |
4087 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
4088 | return; | |
4089 | } | |
4090 | ||
c9de560d | 4091 | /* don't use group allocation for large files */ |
71780577 | 4092 | size = max(size, isize); |
cc483f10 | 4093 | if (size > sbi->s_mb_stream_request) { |
4ba74d00 | 4094 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
c9de560d | 4095 | return; |
4ba74d00 | 4096 | } |
c9de560d AT |
4097 | |
4098 | BUG_ON(ac->ac_lg != NULL); | |
4099 | /* | |
4100 | * locality group prealloc space are per cpu. The reason for having | |
4101 | * per cpu locality group is to reduce the contention between block | |
4102 | * request from multiple CPUs. | |
4103 | */ | |
ca0c9584 | 4104 | ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
4105 | |
4106 | /* we're going to use group allocation */ | |
4107 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4108 | ||
4109 | /* serialize all allocations in the group */ | |
4110 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4111 | } | |
4112 | ||
4ddfef7b ES |
4113 | static noinline_for_stack int |
4114 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
4115 | struct ext4_allocation_request *ar) |
4116 | { | |
4117 | struct super_block *sb = ar->inode->i_sb; | |
4118 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4119 | struct ext4_super_block *es = sbi->s_es; | |
4120 | ext4_group_t group; | |
498e5f24 TT |
4121 | unsigned int len; |
4122 | ext4_fsblk_t goal; | |
c9de560d AT |
4123 | ext4_grpblk_t block; |
4124 | ||
4125 | /* we can't allocate > group size */ | |
4126 | len = ar->len; | |
4127 | ||
4128 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
4129 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
4130 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
4131 | |
4132 | /* start searching from the goal */ | |
4133 | goal = ar->goal; | |
4134 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4135 | goal >= ext4_blocks_count(es)) | |
4136 | goal = le32_to_cpu(es->s_first_data_block); | |
4137 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4138 | ||
4139 | /* set up allocation goals */ | |
f5a44db5 | 4140 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 4141 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
4142 | ac->ac_sb = sb; |
4143 | ac->ac_inode = ar->inode; | |
53accfa9 | 4144 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
4145 | ac->ac_o_ex.fe_group = group; |
4146 | ac->ac_o_ex.fe_start = block; | |
4147 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 4148 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 4149 | ac->ac_flags = ar->flags; |
c9de560d AT |
4150 | |
4151 | /* we have to define context: we'll we work with a file or | |
4152 | * locality group. this is a policy, actually */ | |
4153 | ext4_mb_group_or_file(ac); | |
4154 | ||
6ba495e9 | 4155 | mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " |
c9de560d AT |
4156 | "left: %u/%u, right %u/%u to %swritable\n", |
4157 | (unsigned) ar->len, (unsigned) ar->logical, | |
4158 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4159 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4160 | (unsigned) ar->lright, (unsigned) ar->pright, | |
4161 | atomic_read(&ar->inode->i_writecount) ? "" : "non-"); | |
4162 | return 0; | |
4163 | ||
4164 | } | |
4165 | ||
6be2ded1 AK |
4166 | static noinline_for_stack void |
4167 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4168 | struct ext4_locality_group *lg, | |
4169 | int order, int total_entries) | |
4170 | { | |
4171 | ext4_group_t group = 0; | |
4172 | struct ext4_buddy e4b; | |
4173 | struct list_head discard_list; | |
4174 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 4175 | |
6ba495e9 | 4176 | mb_debug(1, "discard locality group preallocation\n"); |
6be2ded1 AK |
4177 | |
4178 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
4179 | |
4180 | spin_lock(&lg->lg_prealloc_lock); | |
4181 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
4182 | pa_inode_list) { | |
4183 | spin_lock(&pa->pa_lock); | |
4184 | if (atomic_read(&pa->pa_count)) { | |
4185 | /* | |
4186 | * This is the pa that we just used | |
4187 | * for block allocation. So don't | |
4188 | * free that | |
4189 | */ | |
4190 | spin_unlock(&pa->pa_lock); | |
4191 | continue; | |
4192 | } | |
4193 | if (pa->pa_deleted) { | |
4194 | spin_unlock(&pa->pa_lock); | |
4195 | continue; | |
4196 | } | |
4197 | /* only lg prealloc space */ | |
cc0fb9ad | 4198 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4199 | |
4200 | /* seems this one can be freed ... */ | |
4201 | pa->pa_deleted = 1; | |
4202 | spin_unlock(&pa->pa_lock); | |
4203 | ||
4204 | list_del_rcu(&pa->pa_inode_list); | |
4205 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4206 | ||
4207 | total_entries--; | |
4208 | if (total_entries <= 5) { | |
4209 | /* | |
4210 | * we want to keep only 5 entries | |
4211 | * allowing it to grow to 8. This | |
4212 | * mak sure we don't call discard | |
4213 | * soon for this list. | |
4214 | */ | |
4215 | break; | |
4216 | } | |
4217 | } | |
4218 | spin_unlock(&lg->lg_prealloc_lock); | |
4219 | ||
4220 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
4221 | ||
bd86298e | 4222 | group = ext4_get_group_number(sb, pa->pa_pstart); |
6be2ded1 | 4223 | if (ext4_mb_load_buddy(sb, group, &e4b)) { |
12062ddd ES |
4224 | ext4_error(sb, "Error loading buddy information for %u", |
4225 | group); | |
6be2ded1 AK |
4226 | continue; |
4227 | } | |
4228 | ext4_lock_group(sb, group); | |
4229 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4230 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
4231 | ext4_unlock_group(sb, group); |
4232 | ||
e39e07fd | 4233 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
4234 | list_del(&pa->u.pa_tmp_list); |
4235 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4236 | } | |
6be2ded1 AK |
4237 | } |
4238 | ||
4239 | /* | |
4240 | * We have incremented pa_count. So it cannot be freed at this | |
4241 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4242 | * possible from this lg. That means pa_free cannot be updated. | |
4243 | * | |
4244 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4245 | * which can cause the lg_prealloc_list to be updated. | |
4246 | */ | |
4247 | ||
4248 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4249 | { | |
4250 | int order, added = 0, lg_prealloc_count = 1; | |
4251 | struct super_block *sb = ac->ac_sb; | |
4252 | struct ext4_locality_group *lg = ac->ac_lg; | |
4253 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4254 | ||
4255 | order = fls(pa->pa_free) - 1; | |
4256 | if (order > PREALLOC_TB_SIZE - 1) | |
4257 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4258 | order = PREALLOC_TB_SIZE - 1; | |
4259 | /* Add the prealloc space to lg */ | |
f1167009 | 4260 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4261 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
4262 | pa_inode_list) { | |
4263 | spin_lock(&tmp_pa->pa_lock); | |
4264 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4265 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4266 | continue; |
4267 | } | |
4268 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4269 | /* Add to the tail of the previous entry */ | |
4270 | list_add_tail_rcu(&pa->pa_inode_list, | |
4271 | &tmp_pa->pa_inode_list); | |
4272 | added = 1; | |
4273 | /* | |
4274 | * we want to count the total | |
4275 | * number of entries in the list | |
4276 | */ | |
4277 | } | |
4278 | spin_unlock(&tmp_pa->pa_lock); | |
4279 | lg_prealloc_count++; | |
4280 | } | |
4281 | if (!added) | |
4282 | list_add_tail_rcu(&pa->pa_inode_list, | |
4283 | &lg->lg_prealloc_list[order]); | |
f1167009 | 4284 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4285 | |
4286 | /* Now trim the list to be not more than 8 elements */ | |
4287 | if (lg_prealloc_count > 8) { | |
4288 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 4289 | order, lg_prealloc_count); |
6be2ded1 AK |
4290 | return; |
4291 | } | |
4292 | return ; | |
4293 | } | |
4294 | ||
c9de560d AT |
4295 | /* |
4296 | * release all resource we used in allocation | |
4297 | */ | |
4298 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4299 | { | |
53accfa9 | 4300 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
4301 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4302 | if (pa) { | |
cc0fb9ad | 4303 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4304 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 4305 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
4306 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
4307 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
4308 | pa->pa_free -= ac->ac_b_ex.fe_len; |
4309 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4310 | spin_unlock(&pa->pa_lock); | |
c9de560d | 4311 | } |
c9de560d | 4312 | } |
ba443916 AK |
4313 | if (pa) { |
4314 | /* | |
4315 | * We want to add the pa to the right bucket. | |
4316 | * Remove it from the list and while adding | |
4317 | * make sure the list to which we are adding | |
44183d42 | 4318 | * doesn't grow big. |
ba443916 | 4319 | */ |
cc0fb9ad | 4320 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { |
ba443916 AK |
4321 | spin_lock(pa->pa_obj_lock); |
4322 | list_del_rcu(&pa->pa_inode_list); | |
4323 | spin_unlock(pa->pa_obj_lock); | |
4324 | ext4_mb_add_n_trim(ac); | |
4325 | } | |
4326 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4327 | } | |
c9de560d AT |
4328 | if (ac->ac_bitmap_page) |
4329 | page_cache_release(ac->ac_bitmap_page); | |
4330 | if (ac->ac_buddy_page) | |
4331 | page_cache_release(ac->ac_buddy_page); | |
4332 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
4333 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4334 | ext4_mb_collect_stats(ac); | |
4335 | return 0; | |
4336 | } | |
4337 | ||
4338 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4339 | { | |
8df9675f | 4340 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4341 | int ret; |
4342 | int freed = 0; | |
4343 | ||
9bffad1e | 4344 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4345 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4346 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4347 | freed += ret; | |
4348 | needed -= ret; | |
4349 | } | |
4350 | ||
4351 | return freed; | |
4352 | } | |
4353 | ||
4354 | /* | |
4355 | * Main entry point into mballoc to allocate blocks | |
4356 | * it tries to use preallocation first, then falls back | |
4357 | * to usual allocation | |
4358 | */ | |
4359 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 4360 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 4361 | { |
6bc6e63f | 4362 | int freed; |
256bdb49 | 4363 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4364 | struct ext4_sb_info *sbi; |
4365 | struct super_block *sb; | |
4366 | ext4_fsblk_t block = 0; | |
60e58e0f | 4367 | unsigned int inquota = 0; |
53accfa9 | 4368 | unsigned int reserv_clstrs = 0; |
c9de560d | 4369 | |
b10a44c3 | 4370 | might_sleep(); |
c9de560d AT |
4371 | sb = ar->inode->i_sb; |
4372 | sbi = EXT4_SB(sb); | |
4373 | ||
9bffad1e | 4374 | trace_ext4_request_blocks(ar); |
ba80b101 | 4375 | |
45dc63e7 DM |
4376 | /* Allow to use superuser reservation for quota file */ |
4377 | if (IS_NOQUOTA(ar->inode)) | |
4378 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; | |
4379 | ||
60e58e0f MC |
4380 | /* |
4381 | * For delayed allocation, we could skip the ENOSPC and | |
4382 | * EDQUOT check, as blocks and quotas have been already | |
4383 | * reserved when data being copied into pagecache. | |
4384 | */ | |
f2321097 | 4385 | if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED)) |
60e58e0f MC |
4386 | ar->flags |= EXT4_MB_DELALLOC_RESERVED; |
4387 | else { | |
4388 | /* Without delayed allocation we need to verify | |
4389 | * there is enough free blocks to do block allocation | |
4390 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4391 | */ |
55f020db | 4392 | while (ar->len && |
e7d5f315 | 4393 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 4394 | |
030ba6bc | 4395 | /* let others to free the space */ |
bb8b20ed | 4396 | cond_resched(); |
030ba6bc AK |
4397 | ar->len = ar->len >> 1; |
4398 | } | |
4399 | if (!ar->len) { | |
a30d542a AK |
4400 | *errp = -ENOSPC; |
4401 | return 0; | |
4402 | } | |
53accfa9 | 4403 | reserv_clstrs = ar->len; |
55f020db | 4404 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
4405 | dquot_alloc_block_nofail(ar->inode, |
4406 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
4407 | } else { |
4408 | while (ar->len && | |
53accfa9 TT |
4409 | dquot_alloc_block(ar->inode, |
4410 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
4411 | |
4412 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
4413 | ar->len--; | |
4414 | } | |
60e58e0f MC |
4415 | } |
4416 | inquota = ar->len; | |
4417 | if (ar->len == 0) { | |
4418 | *errp = -EDQUOT; | |
6c7a120a | 4419 | goto out; |
60e58e0f | 4420 | } |
07031431 | 4421 | } |
d2a17637 | 4422 | |
85556c9a | 4423 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 4424 | if (!ac) { |
363d4251 | 4425 | ar->len = 0; |
256bdb49 | 4426 | *errp = -ENOMEM; |
6c7a120a | 4427 | goto out; |
256bdb49 ES |
4428 | } |
4429 | ||
256bdb49 | 4430 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4431 | if (*errp) { |
4432 | ar->len = 0; | |
6c7a120a | 4433 | goto out; |
c9de560d AT |
4434 | } |
4435 | ||
256bdb49 ES |
4436 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4437 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4438 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4439 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4440 | repeat: |
4441 | /* allocate space in core */ | |
6c7a120a | 4442 | *errp = ext4_mb_regular_allocator(ac); |
2c00ef3e AK |
4443 | if (*errp) |
4444 | goto discard_and_exit; | |
c9de560d AT |
4445 | |
4446 | /* as we've just preallocated more space than | |
2c00ef3e | 4447 | * user requested originally, we store allocated |
c9de560d | 4448 | * space in a special descriptor */ |
256bdb49 | 4449 | if (ac->ac_status == AC_STATUS_FOUND && |
2c00ef3e AK |
4450 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) |
4451 | *errp = ext4_mb_new_preallocation(ac); | |
4452 | if (*errp) { | |
4453 | discard_and_exit: | |
4454 | ext4_discard_allocated_blocks(ac); | |
4455 | goto errout; | |
4456 | } | |
c9de560d | 4457 | } |
256bdb49 | 4458 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 4459 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
6c7a120a | 4460 | if (*errp == -EAGAIN) { |
8556e8f3 AK |
4461 | /* |
4462 | * drop the reference that we took | |
4463 | * in ext4_mb_use_best_found | |
4464 | */ | |
4465 | ext4_mb_release_context(ac); | |
519deca0 AK |
4466 | ac->ac_b_ex.fe_group = 0; |
4467 | ac->ac_b_ex.fe_start = 0; | |
4468 | ac->ac_b_ex.fe_len = 0; | |
4469 | ac->ac_status = AC_STATUS_CONTINUE; | |
4470 | goto repeat; | |
6d138ced | 4471 | } else if (*errp) { |
b844167e | 4472 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
4473 | goto errout; |
4474 | } else { | |
519deca0 AK |
4475 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
4476 | ar->len = ac->ac_b_ex.fe_len; | |
4477 | } | |
c9de560d | 4478 | } else { |
256bdb49 | 4479 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4480 | if (freed) |
4481 | goto repeat; | |
4482 | *errp = -ENOSPC; | |
6c7a120a AK |
4483 | } |
4484 | ||
6d138ced | 4485 | errout: |
6c7a120a | 4486 | if (*errp) { |
256bdb49 | 4487 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4488 | ar->len = 0; |
256bdb49 | 4489 | ext4_mb_show_ac(ac); |
c9de560d | 4490 | } |
256bdb49 | 4491 | ext4_mb_release_context(ac); |
6c7a120a AK |
4492 | out: |
4493 | if (ac) | |
4494 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 4495 | if (inquota && ar->len < inquota) |
53accfa9 | 4496 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 4497 | if (!ar->len) { |
f2321097 TT |
4498 | if (!ext4_test_inode_state(ar->inode, |
4499 | EXT4_STATE_DELALLOC_RESERVED)) | |
0087d9fb | 4500 | /* release all the reserved blocks if non delalloc */ |
57042651 | 4501 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 4502 | reserv_clstrs); |
0087d9fb | 4503 | } |
c9de560d | 4504 | |
9bffad1e | 4505 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4506 | |
c9de560d AT |
4507 | return block; |
4508 | } | |
c9de560d | 4509 | |
c894058d AK |
4510 | /* |
4511 | * We can merge two free data extents only if the physical blocks | |
4512 | * are contiguous, AND the extents were freed by the same transaction, | |
4513 | * AND the blocks are associated with the same group. | |
4514 | */ | |
4515 | static int can_merge(struct ext4_free_data *entry1, | |
4516 | struct ext4_free_data *entry2) | |
4517 | { | |
18aadd47 BJ |
4518 | if ((entry1->efd_tid == entry2->efd_tid) && |
4519 | (entry1->efd_group == entry2->efd_group) && | |
4520 | ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster)) | |
c894058d AK |
4521 | return 1; |
4522 | return 0; | |
4523 | } | |
4524 | ||
4ddfef7b ES |
4525 | static noinline_for_stack int |
4526 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 4527 | struct ext4_free_data *new_entry) |
c9de560d | 4528 | { |
e29136f8 | 4529 | ext4_group_t group = e4b->bd_group; |
84130193 | 4530 | ext4_grpblk_t cluster; |
7a2fcbf7 | 4531 | struct ext4_free_data *entry; |
c9de560d AT |
4532 | struct ext4_group_info *db = e4b->bd_info; |
4533 | struct super_block *sb = e4b->bd_sb; | |
4534 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4535 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
4536 | struct rb_node *parent = NULL, *new_node; | |
4537 | ||
0390131b | 4538 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4539 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4540 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4541 | ||
18aadd47 BJ |
4542 | new_node = &new_entry->efd_node; |
4543 | cluster = new_entry->efd_start_cluster; | |
c894058d | 4544 | |
c894058d AK |
4545 | if (!*n) { |
4546 | /* first free block exent. We need to | |
4547 | protect buddy cache from being freed, | |
4548 | * otherwise we'll refresh it from | |
4549 | * on-disk bitmap and lose not-yet-available | |
4550 | * blocks */ | |
4551 | page_cache_get(e4b->bd_buddy_page); | |
4552 | page_cache_get(e4b->bd_bitmap_page); | |
4553 | } | |
4554 | while (*n) { | |
4555 | parent = *n; | |
18aadd47 BJ |
4556 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
4557 | if (cluster < entry->efd_start_cluster) | |
c894058d | 4558 | n = &(*n)->rb_left; |
18aadd47 | 4559 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
4560 | n = &(*n)->rb_right; |
4561 | else { | |
e29136f8 | 4562 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
4563 | ext4_group_first_block_no(sb, group) + |
4564 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 4565 | "Block already on to-be-freed list"); |
c894058d | 4566 | return 0; |
c9de560d | 4567 | } |
c894058d | 4568 | } |
c9de560d | 4569 | |
c894058d AK |
4570 | rb_link_node(new_node, parent, n); |
4571 | rb_insert_color(new_node, &db->bb_free_root); | |
4572 | ||
4573 | /* Now try to see the extent can be merged to left and right */ | |
4574 | node = rb_prev(new_node); | |
4575 | if (node) { | |
18aadd47 | 4576 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
5d3ee208 DM |
4577 | if (can_merge(entry, new_entry) && |
4578 | ext4_journal_callback_try_del(handle, &entry->efd_jce)) { | |
18aadd47 BJ |
4579 | new_entry->efd_start_cluster = entry->efd_start_cluster; |
4580 | new_entry->efd_count += entry->efd_count; | |
c894058d | 4581 | rb_erase(node, &(db->bb_free_root)); |
18aadd47 | 4582 | kmem_cache_free(ext4_free_data_cachep, entry); |
c9de560d | 4583 | } |
c894058d | 4584 | } |
c9de560d | 4585 | |
c894058d AK |
4586 | node = rb_next(new_node); |
4587 | if (node) { | |
18aadd47 | 4588 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
5d3ee208 DM |
4589 | if (can_merge(new_entry, entry) && |
4590 | ext4_journal_callback_try_del(handle, &entry->efd_jce)) { | |
18aadd47 | 4591 | new_entry->efd_count += entry->efd_count; |
c894058d | 4592 | rb_erase(node, &(db->bb_free_root)); |
18aadd47 | 4593 | kmem_cache_free(ext4_free_data_cachep, entry); |
c9de560d AT |
4594 | } |
4595 | } | |
3e624fc7 | 4596 | /* Add the extent to transaction's private list */ |
18aadd47 BJ |
4597 | ext4_journal_callback_add(handle, ext4_free_data_callback, |
4598 | &new_entry->efd_jce); | |
c9de560d AT |
4599 | return 0; |
4600 | } | |
4601 | ||
44338711 TT |
4602 | /** |
4603 | * ext4_free_blocks() -- Free given blocks and update quota | |
4604 | * @handle: handle for this transaction | |
4605 | * @inode: inode | |
4606 | * @block: start physical block to free | |
4607 | * @count: number of blocks to count | |
5def1360 | 4608 | * @flags: flags used by ext4_free_blocks |
c9de560d | 4609 | */ |
44338711 | 4610 | void ext4_free_blocks(handle_t *handle, struct inode *inode, |
e6362609 TT |
4611 | struct buffer_head *bh, ext4_fsblk_t block, |
4612 | unsigned long count, int flags) | |
c9de560d | 4613 | { |
26346ff6 | 4614 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4615 | struct super_block *sb = inode->i_sb; |
c9de560d | 4616 | struct ext4_group_desc *gdp; |
498e5f24 | 4617 | unsigned int overflow; |
c9de560d AT |
4618 | ext4_grpblk_t bit; |
4619 | struct buffer_head *gd_bh; | |
4620 | ext4_group_t block_group; | |
4621 | struct ext4_sb_info *sbi; | |
7d734532 | 4622 | struct ext4_inode_info *ei = EXT4_I(inode); |
c9de560d | 4623 | struct ext4_buddy e4b; |
84130193 | 4624 | unsigned int count_clusters; |
c9de560d AT |
4625 | int err = 0; |
4626 | int ret; | |
4627 | ||
b10a44c3 | 4628 | might_sleep(); |
e6362609 TT |
4629 | if (bh) { |
4630 | if (block) | |
4631 | BUG_ON(block != bh->b_blocknr); | |
4632 | else | |
4633 | block = bh->b_blocknr; | |
4634 | } | |
c9de560d | 4635 | |
c9de560d | 4636 | sbi = EXT4_SB(sb); |
1f2acb60 TT |
4637 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
4638 | !ext4_data_block_valid(sbi, block, count)) { | |
12062ddd | 4639 | ext4_error(sb, "Freeing blocks not in datazone - " |
1f2acb60 | 4640 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
4641 | goto error_return; |
4642 | } | |
4643 | ||
0610b6e9 | 4644 | ext4_debug("freeing block %llu\n", block); |
e6362609 TT |
4645 | trace_ext4_free_blocks(inode, block, count, flags); |
4646 | ||
4647 | if (flags & EXT4_FREE_BLOCKS_FORGET) { | |
4648 | struct buffer_head *tbh = bh; | |
4649 | int i; | |
4650 | ||
4651 | BUG_ON(bh && (count > 1)); | |
4652 | ||
4653 | for (i = 0; i < count; i++) { | |
2ed5724d | 4654 | cond_resched(); |
e6362609 TT |
4655 | if (!bh) |
4656 | tbh = sb_find_get_block(inode->i_sb, | |
4657 | block + i); | |
2ed5724d | 4658 | if (!tbh) |
87783690 | 4659 | continue; |
60e6679e | 4660 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, |
e6362609 TT |
4661 | inode, tbh, block + i); |
4662 | } | |
4663 | } | |
4664 | ||
60e6679e | 4665 | /* |
e6362609 TT |
4666 | * We need to make sure we don't reuse the freed block until |
4667 | * after the transaction is committed, which we can do by | |
4668 | * treating the block as metadata, below. We make an | |
4669 | * exception if the inode is to be written in writeback mode | |
4670 | * since writeback mode has weak data consistency guarantees. | |
4671 | */ | |
4672 | if (!ext4_should_writeback_data(inode)) | |
4673 | flags |= EXT4_FREE_BLOCKS_METADATA; | |
c9de560d | 4674 | |
84130193 TT |
4675 | /* |
4676 | * If the extent to be freed does not begin on a cluster | |
4677 | * boundary, we need to deal with partial clusters at the | |
4678 | * beginning and end of the extent. Normally we will free | |
4679 | * blocks at the beginning or the end unless we are explicitly | |
4680 | * requested to avoid doing so. | |
4681 | */ | |
f5a44db5 | 4682 | overflow = EXT4_PBLK_COFF(sbi, block); |
84130193 TT |
4683 | if (overflow) { |
4684 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
4685 | overflow = sbi->s_cluster_ratio - overflow; | |
4686 | block += overflow; | |
4687 | if (count > overflow) | |
4688 | count -= overflow; | |
4689 | else | |
4690 | return; | |
4691 | } else { | |
4692 | block -= overflow; | |
4693 | count += overflow; | |
4694 | } | |
4695 | } | |
f5a44db5 | 4696 | overflow = EXT4_LBLK_COFF(sbi, count); |
84130193 TT |
4697 | if (overflow) { |
4698 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
4699 | if (count > overflow) | |
4700 | count -= overflow; | |
4701 | else | |
4702 | return; | |
4703 | } else | |
4704 | count += sbi->s_cluster_ratio - overflow; | |
4705 | } | |
4706 | ||
c9de560d AT |
4707 | do_more: |
4708 | overflow = 0; | |
4709 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4710 | ||
163a203d DW |
4711 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( |
4712 | ext4_get_group_info(sb, block_group)))) | |
4713 | return; | |
4714 | ||
c9de560d AT |
4715 | /* |
4716 | * Check to see if we are freeing blocks across a group | |
4717 | * boundary. | |
4718 | */ | |
84130193 TT |
4719 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
4720 | overflow = EXT4_C2B(sbi, bit) + count - | |
4721 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d AT |
4722 | count -= overflow; |
4723 | } | |
810da240 | 4724 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 4725 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
ce89f46c AK |
4726 | if (!bitmap_bh) { |
4727 | err = -EIO; | |
c9de560d | 4728 | goto error_return; |
ce89f46c | 4729 | } |
c9de560d | 4730 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4731 | if (!gdp) { |
4732 | err = -EIO; | |
c9de560d | 4733 | goto error_return; |
ce89f46c | 4734 | } |
c9de560d AT |
4735 | |
4736 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4737 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4738 | in_range(block, ext4_inode_table(sb, gdp), | |
84130193 | 4739 | EXT4_SB(sb)->s_itb_per_group) || |
c9de560d | 4740 | in_range(block + count - 1, ext4_inode_table(sb, gdp), |
84130193 | 4741 | EXT4_SB(sb)->s_itb_per_group)) { |
c9de560d | 4742 | |
12062ddd | 4743 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 4744 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
4745 | /* err = 0. ext4_std_error should be a no op */ |
4746 | goto error_return; | |
c9de560d AT |
4747 | } |
4748 | ||
4749 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4750 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4751 | if (err) | |
4752 | goto error_return; | |
4753 | ||
4754 | /* | |
4755 | * We are about to modify some metadata. Call the journal APIs | |
4756 | * to unshare ->b_data if a currently-committing transaction is | |
4757 | * using it | |
4758 | */ | |
4759 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4760 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4761 | if (err) | |
4762 | goto error_return; | |
c9de560d AT |
4763 | #ifdef AGGRESSIVE_CHECK |
4764 | { | |
4765 | int i; | |
84130193 | 4766 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
4767 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
4768 | } | |
4769 | #endif | |
84130193 | 4770 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 4771 | |
920313a7 AK |
4772 | err = ext4_mb_load_buddy(sb, block_group, &e4b); |
4773 | if (err) | |
4774 | goto error_return; | |
e6362609 TT |
4775 | |
4776 | if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) { | |
7a2fcbf7 AK |
4777 | struct ext4_free_data *new_entry; |
4778 | /* | |
4779 | * blocks being freed are metadata. these blocks shouldn't | |
4780 | * be used until this transaction is committed | |
4781 | */ | |
e7676a70 | 4782 | retry: |
18aadd47 | 4783 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS); |
b72143ab | 4784 | if (!new_entry) { |
e7676a70 TT |
4785 | /* |
4786 | * We use a retry loop because | |
4787 | * ext4_free_blocks() is not allowed to fail. | |
4788 | */ | |
4789 | cond_resched(); | |
4790 | congestion_wait(BLK_RW_ASYNC, HZ/50); | |
4791 | goto retry; | |
b72143ab | 4792 | } |
18aadd47 BJ |
4793 | new_entry->efd_start_cluster = bit; |
4794 | new_entry->efd_group = block_group; | |
4795 | new_entry->efd_count = count_clusters; | |
4796 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 4797 | |
7a2fcbf7 | 4798 | ext4_lock_group(sb, block_group); |
84130193 | 4799 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 4800 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 4801 | } else { |
7a2fcbf7 AK |
4802 | /* need to update group_info->bb_free and bitmap |
4803 | * with group lock held. generate_buddy look at | |
4804 | * them with group lock_held | |
4805 | */ | |
d71c1ae2 LC |
4806 | if (test_opt(sb, DISCARD)) { |
4807 | err = ext4_issue_discard(sb, block_group, bit, count); | |
4808 | if (err && err != -EOPNOTSUPP) | |
4809 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
4810 | " group:%d block:%d count:%lu failed" | |
4811 | " with %d", block_group, bit, count, | |
4812 | err); | |
8f9ff189 LC |
4813 | } else |
4814 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 4815 | |
955ce5f5 | 4816 | ext4_lock_group(sb, block_group); |
84130193 TT |
4817 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
4818 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
4819 | } |
4820 | ||
021b65bb TT |
4821 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
4822 | ext4_free_group_clusters_set(sb, gdp, ret); | |
79f1ba49 | 4823 | ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); |
feb0ab32 | 4824 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 4825 | ext4_unlock_group(sb, block_group); |
c9de560d | 4826 | |
772cb7c8 JS |
4827 | if (sbi->s_log_groups_per_flex) { |
4828 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f TT |
4829 | atomic64_add(count_clusters, |
4830 | &sbi->s_flex_groups[flex_group].free_clusters); | |
772cb7c8 JS |
4831 | } |
4832 | ||
7d734532 JK |
4833 | if (flags & EXT4_FREE_BLOCKS_RESERVE && ei->i_reserved_data_blocks) { |
4834 | percpu_counter_add(&sbi->s_dirtyclusters_counter, | |
4835 | count_clusters); | |
4836 | spin_lock(&ei->i_block_reservation_lock); | |
4837 | if (flags & EXT4_FREE_BLOCKS_METADATA) | |
4838 | ei->i_reserved_meta_blocks += count_clusters; | |
4839 | else | |
4840 | ei->i_reserved_data_blocks += count_clusters; | |
4841 | spin_unlock(&ei->i_block_reservation_lock); | |
4842 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) | |
4843 | dquot_reclaim_block(inode, | |
4844 | EXT4_C2B(sbi, count_clusters)); | |
4845 | } else if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) | |
7b415bf6 | 4846 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); |
7d734532 JK |
4847 | percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters); |
4848 | ||
4849 | ext4_mb_unload_buddy(&e4b); | |
7b415bf6 | 4850 | |
7a2fcbf7 AK |
4851 | /* We dirtied the bitmap block */ |
4852 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4853 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4854 | ||
c9de560d AT |
4855 | /* And the group descriptor block */ |
4856 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 4857 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
4858 | if (!err) |
4859 | err = ret; | |
4860 | ||
4861 | if (overflow && !err) { | |
4862 | block += count; | |
4863 | count = overflow; | |
4864 | put_bh(bitmap_bh); | |
4865 | goto do_more; | |
4866 | } | |
c9de560d AT |
4867 | error_return: |
4868 | brelse(bitmap_bh); | |
4869 | ext4_std_error(sb, err); | |
4870 | return; | |
4871 | } | |
7360d173 | 4872 | |
2846e820 | 4873 | /** |
0529155e | 4874 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
4875 | * @handle: handle to this transaction |
4876 | * @sb: super block | |
4907cb7b | 4877 | * @block: start physical block to add to the block group |
2846e820 AG |
4878 | * @count: number of blocks to free |
4879 | * | |
e73a347b | 4880 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 4881 | */ |
cc7365df | 4882 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
4883 | ext4_fsblk_t block, unsigned long count) |
4884 | { | |
4885 | struct buffer_head *bitmap_bh = NULL; | |
4886 | struct buffer_head *gd_bh; | |
4887 | ext4_group_t block_group; | |
4888 | ext4_grpblk_t bit; | |
4889 | unsigned int i; | |
4890 | struct ext4_group_desc *desc; | |
4891 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 4892 | struct ext4_buddy e4b; |
2846e820 AG |
4893 | int err = 0, ret, blk_free_count; |
4894 | ext4_grpblk_t blocks_freed; | |
2846e820 AG |
4895 | |
4896 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
4897 | ||
4740b830 YY |
4898 | if (count == 0) |
4899 | return 0; | |
4900 | ||
2846e820 | 4901 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
4902 | /* |
4903 | * Check to see if we are freeing blocks across a group | |
4904 | * boundary. | |
4905 | */ | |
cc7365df YY |
4906 | if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
4907 | ext4_warning(sb, "too much blocks added to group %u\n", | |
4908 | block_group); | |
4909 | err = -EINVAL; | |
2846e820 | 4910 | goto error_return; |
cc7365df | 4911 | } |
2cd05cc3 | 4912 | |
2846e820 | 4913 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
cc7365df YY |
4914 | if (!bitmap_bh) { |
4915 | err = -EIO; | |
2846e820 | 4916 | goto error_return; |
cc7365df YY |
4917 | } |
4918 | ||
2846e820 | 4919 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
4920 | if (!desc) { |
4921 | err = -EIO; | |
2846e820 | 4922 | goto error_return; |
cc7365df | 4923 | } |
2846e820 AG |
4924 | |
4925 | if (in_range(ext4_block_bitmap(sb, desc), block, count) || | |
4926 | in_range(ext4_inode_bitmap(sb, desc), block, count) || | |
4927 | in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || | |
4928 | in_range(block + count - 1, ext4_inode_table(sb, desc), | |
4929 | sbi->s_itb_per_group)) { | |
4930 | ext4_error(sb, "Adding blocks in system zones - " | |
4931 | "Block = %llu, count = %lu", | |
4932 | block, count); | |
cc7365df | 4933 | err = -EINVAL; |
2846e820 AG |
4934 | goto error_return; |
4935 | } | |
4936 | ||
2cd05cc3 TT |
4937 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
4938 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2846e820 AG |
4939 | if (err) |
4940 | goto error_return; | |
4941 | ||
4942 | /* | |
4943 | * We are about to modify some metadata. Call the journal APIs | |
4944 | * to unshare ->b_data if a currently-committing transaction is | |
4945 | * using it | |
4946 | */ | |
4947 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4948 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4949 | if (err) | |
4950 | goto error_return; | |
e73a347b | 4951 | |
2846e820 AG |
4952 | for (i = 0, blocks_freed = 0; i < count; i++) { |
4953 | BUFFER_TRACE(bitmap_bh, "clear bit"); | |
e73a347b | 4954 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
4955 | ext4_error(sb, "bit already cleared for block %llu", |
4956 | (ext4_fsblk_t)(block + i)); | |
4957 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
4958 | } else { | |
4959 | blocks_freed++; | |
4960 | } | |
4961 | } | |
e73a347b AG |
4962 | |
4963 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
4964 | if (err) | |
4965 | goto error_return; | |
4966 | ||
4967 | /* | |
4968 | * need to update group_info->bb_free and bitmap | |
4969 | * with group lock held. generate_buddy look at | |
4970 | * them with group lock_held | |
4971 | */ | |
2846e820 | 4972 | ext4_lock_group(sb, block_group); |
e73a347b AG |
4973 | mb_clear_bits(bitmap_bh->b_data, bit, count); |
4974 | mb_free_blocks(NULL, &e4b, bit, count); | |
021b65bb TT |
4975 | blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc); |
4976 | ext4_free_group_clusters_set(sb, desc, blk_free_count); | |
79f1ba49 | 4977 | ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); |
feb0ab32 | 4978 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 4979 | ext4_unlock_group(sb, block_group); |
57042651 | 4980 | percpu_counter_add(&sbi->s_freeclusters_counter, |
810da240 | 4981 | EXT4_NUM_B2C(sbi, blocks_freed)); |
2846e820 AG |
4982 | |
4983 | if (sbi->s_log_groups_per_flex) { | |
4984 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f TT |
4985 | atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed), |
4986 | &sbi->s_flex_groups[flex_group].free_clusters); | |
2846e820 | 4987 | } |
e73a347b AG |
4988 | |
4989 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
4990 | |
4991 | /* We dirtied the bitmap block */ | |
4992 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4993 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4994 | ||
4995 | /* And the group descriptor block */ | |
4996 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
4997 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
4998 | if (!err) | |
4999 | err = ret; | |
5000 | ||
5001 | error_return: | |
5002 | brelse(bitmap_bh); | |
5003 | ext4_std_error(sb, err); | |
cc7365df | 5004 | return err; |
2846e820 AG |
5005 | } |
5006 | ||
7360d173 LC |
5007 | /** |
5008 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
5009 | * @sb: super block for the file system | |
5010 | * @start: starting block of the free extent in the alloc. group | |
5011 | * @count: number of blocks to TRIM | |
5012 | * @group: alloc. group we are working with | |
5013 | * @e4b: ext4 buddy for the group | |
5014 | * | |
5015 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
5016 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
5017 | * be called with under the group lock. | |
5018 | */ | |
d71c1ae2 | 5019 | static int ext4_trim_extent(struct super_block *sb, int start, int count, |
d9f34504 | 5020 | ext4_group_t group, struct ext4_buddy *e4b) |
e2cbd587 | 5021 | __releases(bitlock) |
5022 | __acquires(bitlock) | |
7360d173 LC |
5023 | { |
5024 | struct ext4_free_extent ex; | |
d71c1ae2 | 5025 | int ret = 0; |
7360d173 | 5026 | |
b3d4c2b1 TM |
5027 | trace_ext4_trim_extent(sb, group, start, count); |
5028 | ||
7360d173 LC |
5029 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
5030 | ||
5031 | ex.fe_start = start; | |
5032 | ex.fe_group = group; | |
5033 | ex.fe_len = count; | |
5034 | ||
5035 | /* | |
5036 | * Mark blocks used, so no one can reuse them while | |
5037 | * being trimmed. | |
5038 | */ | |
5039 | mb_mark_used(e4b, &ex); | |
5040 | ext4_unlock_group(sb, group); | |
d71c1ae2 | 5041 | ret = ext4_issue_discard(sb, group, start, count); |
7360d173 LC |
5042 | ext4_lock_group(sb, group); |
5043 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 5044 | return ret; |
7360d173 LC |
5045 | } |
5046 | ||
5047 | /** | |
5048 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
5049 | * @sb: super block for file system | |
22612283 | 5050 | * @group: group to be trimmed |
7360d173 LC |
5051 | * @start: first group block to examine |
5052 | * @max: last group block to examine | |
5053 | * @minblocks: minimum extent block count | |
5054 | * | |
5055 | * ext4_trim_all_free walks through group's buddy bitmap searching for free | |
5056 | * extents. When the free block is found, ext4_trim_extent is called to TRIM | |
5057 | * the extent. | |
5058 | * | |
5059 | * | |
5060 | * ext4_trim_all_free walks through group's block bitmap searching for free | |
5061 | * extents. When the free extent is found, mark it as used in group buddy | |
5062 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
5063 | * the group buddy bitmap. This is done until whole group is scanned. | |
5064 | */ | |
0b75a840 | 5065 | static ext4_grpblk_t |
78944086 LC |
5066 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
5067 | ext4_grpblk_t start, ext4_grpblk_t max, | |
5068 | ext4_grpblk_t minblocks) | |
7360d173 LC |
5069 | { |
5070 | void *bitmap; | |
169ddc3e | 5071 | ext4_grpblk_t next, count = 0, free_count = 0; |
78944086 | 5072 | struct ext4_buddy e4b; |
d71c1ae2 | 5073 | int ret = 0; |
7360d173 | 5074 | |
b3d4c2b1 TM |
5075 | trace_ext4_trim_all_free(sb, group, start, max); |
5076 | ||
78944086 LC |
5077 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
5078 | if (ret) { | |
5079 | ext4_error(sb, "Error in loading buddy " | |
5080 | "information for %u", group); | |
5081 | return ret; | |
5082 | } | |
78944086 | 5083 | bitmap = e4b.bd_bitmap; |
28739eea LC |
5084 | |
5085 | ext4_lock_group(sb, group); | |
3d56b8d2 TM |
5086 | if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && |
5087 | minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) | |
5088 | goto out; | |
5089 | ||
78944086 LC |
5090 | start = (e4b.bd_info->bb_first_free > start) ? |
5091 | e4b.bd_info->bb_first_free : start; | |
7360d173 | 5092 | |
913eed83 LC |
5093 | while (start <= max) { |
5094 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
5095 | if (start > max) | |
7360d173 | 5096 | break; |
913eed83 | 5097 | next = mb_find_next_bit(bitmap, max + 1, start); |
7360d173 LC |
5098 | |
5099 | if ((next - start) >= minblocks) { | |
d71c1ae2 LC |
5100 | ret = ext4_trim_extent(sb, start, |
5101 | next - start, group, &e4b); | |
5102 | if (ret && ret != -EOPNOTSUPP) | |
5103 | break; | |
5104 | ret = 0; | |
7360d173 LC |
5105 | count += next - start; |
5106 | } | |
169ddc3e | 5107 | free_count += next - start; |
7360d173 LC |
5108 | start = next + 1; |
5109 | ||
5110 | if (fatal_signal_pending(current)) { | |
5111 | count = -ERESTARTSYS; | |
5112 | break; | |
5113 | } | |
5114 | ||
5115 | if (need_resched()) { | |
5116 | ext4_unlock_group(sb, group); | |
5117 | cond_resched(); | |
5118 | ext4_lock_group(sb, group); | |
5119 | } | |
5120 | ||
169ddc3e | 5121 | if ((e4b.bd_info->bb_free - free_count) < minblocks) |
7360d173 LC |
5122 | break; |
5123 | } | |
3d56b8d2 | 5124 | |
d71c1ae2 LC |
5125 | if (!ret) { |
5126 | ret = count; | |
3d56b8d2 | 5127 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); |
d71c1ae2 | 5128 | } |
3d56b8d2 | 5129 | out: |
7360d173 | 5130 | ext4_unlock_group(sb, group); |
78944086 | 5131 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
5132 | |
5133 | ext4_debug("trimmed %d blocks in the group %d\n", | |
5134 | count, group); | |
5135 | ||
d71c1ae2 | 5136 | return ret; |
7360d173 LC |
5137 | } |
5138 | ||
5139 | /** | |
5140 | * ext4_trim_fs() -- trim ioctl handle function | |
5141 | * @sb: superblock for filesystem | |
5142 | * @range: fstrim_range structure | |
5143 | * | |
5144 | * start: First Byte to trim | |
5145 | * len: number of Bytes to trim from start | |
5146 | * minlen: minimum extent length in Bytes | |
5147 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
5148 | * start to start+len. For each such a group ext4_trim_all_free function | |
5149 | * is invoked to trim all free space. | |
5150 | */ | |
5151 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
5152 | { | |
78944086 | 5153 | struct ext4_group_info *grp; |
913eed83 | 5154 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 5155 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 5156 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
5157 | ext4_fsblk_t first_data_blk = |
5158 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 5159 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
7360d173 LC |
5160 | int ret = 0; |
5161 | ||
5162 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 5163 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
5164 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
5165 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 5166 | |
5de35e8d LC |
5167 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
5168 | start >= max_blks || | |
5169 | range->len < sb->s_blocksize) | |
7360d173 | 5170 | return -EINVAL; |
913eed83 LC |
5171 | if (end >= max_blks) |
5172 | end = max_blks - 1; | |
5173 | if (end <= first_data_blk) | |
22f10457 | 5174 | goto out; |
913eed83 | 5175 | if (start < first_data_blk) |
0f0a25bf | 5176 | start = first_data_blk; |
7360d173 | 5177 | |
913eed83 | 5178 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 5179 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 5180 | &first_group, &first_cluster); |
913eed83 | 5181 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 5182 | &last_group, &last_cluster); |
7360d173 | 5183 | |
913eed83 LC |
5184 | /* end now represents the last cluster to discard in this group */ |
5185 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
7360d173 LC |
5186 | |
5187 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
5188 | grp = ext4_get_group_info(sb, group); |
5189 | /* We only do this if the grp has never been initialized */ | |
5190 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
5191 | ret = ext4_mb_init_group(sb, group); | |
5192 | if (ret) | |
5193 | break; | |
7360d173 LC |
5194 | } |
5195 | ||
0ba08517 | 5196 | /* |
913eed83 LC |
5197 | * For all the groups except the last one, last cluster will |
5198 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
5199 | * change it for the last group, note that last_cluster is | |
5200 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 5201 | */ |
913eed83 LC |
5202 | if (group == last_group) |
5203 | end = last_cluster; | |
7360d173 | 5204 | |
78944086 | 5205 | if (grp->bb_free >= minlen) { |
7137d7a4 | 5206 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
913eed83 | 5207 | end, minlen); |
7360d173 LC |
5208 | if (cnt < 0) { |
5209 | ret = cnt; | |
7360d173 LC |
5210 | break; |
5211 | } | |
21e7fd22 | 5212 | trimmed += cnt; |
7360d173 | 5213 | } |
913eed83 LC |
5214 | |
5215 | /* | |
5216 | * For every group except the first one, we are sure | |
5217 | * that the first cluster to discard will be cluster #0. | |
5218 | */ | |
7137d7a4 | 5219 | first_cluster = 0; |
7360d173 | 5220 | } |
7360d173 | 5221 | |
3d56b8d2 TM |
5222 | if (!ret) |
5223 | atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); | |
5224 | ||
22f10457 | 5225 | out: |
aaf7d73e | 5226 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
5227 | return ret; |
5228 | } |