]>
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 | ||
ffad0a44 AK |
408 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
409 | { | |
e7dfb246 | 410 | int fix = 0, ret, tmpmax; |
ffad0a44 | 411 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 412 | tmpmax = max + fix; |
ffad0a44 AK |
413 | start += fix; |
414 | ||
e7dfb246 AK |
415 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
416 | if (ret > max) | |
417 | return max; | |
418 | return ret; | |
ffad0a44 AK |
419 | } |
420 | ||
421 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
422 | { | |
e7dfb246 | 423 | int fix = 0, ret, tmpmax; |
ffad0a44 | 424 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 425 | tmpmax = max + fix; |
ffad0a44 AK |
426 | start += fix; |
427 | ||
e7dfb246 AK |
428 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
429 | if (ret > max) | |
430 | return max; | |
431 | return ret; | |
ffad0a44 AK |
432 | } |
433 | ||
c9de560d AT |
434 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
435 | { | |
436 | char *bb; | |
437 | ||
c5e8f3f3 | 438 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
439 | BUG_ON(max == NULL); |
440 | ||
441 | if (order > e4b->bd_blkbits + 1) { | |
442 | *max = 0; | |
443 | return NULL; | |
444 | } | |
445 | ||
446 | /* at order 0 we see each particular block */ | |
84b775a3 CL |
447 | if (order == 0) { |
448 | *max = 1 << (e4b->bd_blkbits + 3); | |
c5e8f3f3 | 449 | return e4b->bd_bitmap; |
84b775a3 | 450 | } |
c9de560d | 451 | |
c5e8f3f3 | 452 | bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; |
c9de560d AT |
453 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; |
454 | ||
455 | return bb; | |
456 | } | |
457 | ||
458 | #ifdef DOUBLE_CHECK | |
459 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
460 | int first, int count) | |
461 | { | |
462 | int i; | |
463 | struct super_block *sb = e4b->bd_sb; | |
464 | ||
465 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
466 | return; | |
bc8e6740 | 467 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
468 | for (i = 0; i < count; i++) { |
469 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
470 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
471 | |
472 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 473 | blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
5d1b1b3f | 474 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
475 | inode ? inode->i_ino : 0, |
476 | blocknr, | |
477 | "freeing block already freed " | |
478 | "(bit %u)", | |
479 | first + i); | |
c9de560d AT |
480 | } |
481 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
482 | } | |
483 | } | |
484 | ||
485 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
486 | { | |
487 | int i; | |
488 | ||
489 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
490 | return; | |
bc8e6740 | 491 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
492 | for (i = 0; i < count; i++) { |
493 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
494 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
495 | } | |
496 | } | |
497 | ||
498 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
499 | { | |
500 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { | |
501 | unsigned char *b1, *b2; | |
502 | int i; | |
503 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
504 | b2 = (unsigned char *) bitmap; | |
505 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
506 | if (b1[i] != b2[i]) { | |
9d8b9ec4 TT |
507 | ext4_msg(e4b->bd_sb, KERN_ERR, |
508 | "corruption in group %u " | |
509 | "at byte %u(%u): %x in copy != %x " | |
510 | "on disk/prealloc", | |
511 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
512 | BUG(); |
513 | } | |
514 | } | |
515 | } | |
516 | } | |
517 | ||
518 | #else | |
519 | static inline void mb_free_blocks_double(struct inode *inode, | |
520 | struct ext4_buddy *e4b, int first, int count) | |
521 | { | |
522 | return; | |
523 | } | |
524 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
525 | int first, int count) | |
526 | { | |
527 | return; | |
528 | } | |
529 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
530 | { | |
531 | return; | |
532 | } | |
533 | #endif | |
534 | ||
535 | #ifdef AGGRESSIVE_CHECK | |
536 | ||
537 | #define MB_CHECK_ASSERT(assert) \ | |
538 | do { \ | |
539 | if (!(assert)) { \ | |
540 | printk(KERN_EMERG \ | |
541 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
542 | function, file, line, # assert); \ | |
543 | BUG(); \ | |
544 | } \ | |
545 | } while (0) | |
546 | ||
547 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
548 | const char *function, int line) | |
549 | { | |
550 | struct super_block *sb = e4b->bd_sb; | |
551 | int order = e4b->bd_blkbits + 1; | |
552 | int max; | |
553 | int max2; | |
554 | int i; | |
555 | int j; | |
556 | int k; | |
557 | int count; | |
558 | struct ext4_group_info *grp; | |
559 | int fragments = 0; | |
560 | int fstart; | |
561 | struct list_head *cur; | |
562 | void *buddy; | |
563 | void *buddy2; | |
564 | ||
c9de560d AT |
565 | { |
566 | static int mb_check_counter; | |
567 | if (mb_check_counter++ % 100 != 0) | |
568 | return 0; | |
569 | } | |
570 | ||
571 | while (order > 1) { | |
572 | buddy = mb_find_buddy(e4b, order, &max); | |
573 | MB_CHECK_ASSERT(buddy); | |
574 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
575 | MB_CHECK_ASSERT(buddy2); | |
576 | MB_CHECK_ASSERT(buddy != buddy2); | |
577 | MB_CHECK_ASSERT(max * 2 == max2); | |
578 | ||
579 | count = 0; | |
580 | for (i = 0; i < max; i++) { | |
581 | ||
582 | if (mb_test_bit(i, buddy)) { | |
583 | /* only single bit in buddy2 may be 1 */ | |
584 | if (!mb_test_bit(i << 1, buddy2)) { | |
585 | MB_CHECK_ASSERT( | |
586 | mb_test_bit((i<<1)+1, buddy2)); | |
587 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
588 | MB_CHECK_ASSERT( | |
589 | mb_test_bit(i << 1, buddy2)); | |
590 | } | |
591 | continue; | |
592 | } | |
593 | ||
0a10da73 | 594 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
595 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
596 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
597 | ||
598 | for (j = 0; j < (1 << order); j++) { | |
599 | k = (i * (1 << order)) + j; | |
600 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 601 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
602 | } |
603 | count++; | |
604 | } | |
605 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
606 | order--; | |
607 | } | |
608 | ||
609 | fstart = -1; | |
610 | buddy = mb_find_buddy(e4b, 0, &max); | |
611 | for (i = 0; i < max; i++) { | |
612 | if (!mb_test_bit(i, buddy)) { | |
613 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
614 | if (fstart == -1) { | |
615 | fragments++; | |
616 | fstart = i; | |
617 | } | |
618 | continue; | |
619 | } | |
620 | fstart = -1; | |
621 | /* check used bits only */ | |
622 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
623 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
624 | k = i >> j; | |
625 | MB_CHECK_ASSERT(k < max2); | |
626 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
627 | } | |
628 | } | |
629 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
630 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
631 | ||
632 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
633 | list_for_each(cur, &grp->bb_prealloc_list) { |
634 | ext4_group_t groupnr; | |
635 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
636 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
637 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 638 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 639 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
640 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
641 | } | |
642 | return 0; | |
643 | } | |
644 | #undef MB_CHECK_ASSERT | |
645 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 646 | __FILE__, __func__, __LINE__) |
c9de560d AT |
647 | #else |
648 | #define mb_check_buddy(e4b) | |
649 | #endif | |
650 | ||
7c786059 CL |
651 | /* |
652 | * Divide blocks started from @first with length @len into | |
653 | * smaller chunks with power of 2 blocks. | |
654 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
655 | * then increase bb_counters[] for corresponded chunk size. | |
656 | */ | |
c9de560d | 657 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 658 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
659 | struct ext4_group_info *grp) |
660 | { | |
661 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
662 | ext4_grpblk_t min; |
663 | ext4_grpblk_t max; | |
664 | ext4_grpblk_t chunk; | |
c9de560d AT |
665 | unsigned short border; |
666 | ||
7137d7a4 | 667 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
668 | |
669 | border = 2 << sb->s_blocksize_bits; | |
670 | ||
671 | while (len > 0) { | |
672 | /* find how many blocks can be covered since this position */ | |
673 | max = ffs(first | border) - 1; | |
674 | ||
675 | /* find how many blocks of power 2 we need to mark */ | |
676 | min = fls(len) - 1; | |
677 | ||
678 | if (max < min) | |
679 | min = max; | |
680 | chunk = 1 << min; | |
681 | ||
682 | /* mark multiblock chunks only */ | |
683 | grp->bb_counters[min]++; | |
684 | if (min > 0) | |
685 | mb_clear_bit(first >> min, | |
686 | buddy + sbi->s_mb_offsets[min]); | |
687 | ||
688 | len -= chunk; | |
689 | first += chunk; | |
690 | } | |
691 | } | |
692 | ||
8a57d9d6 CW |
693 | /* |
694 | * Cache the order of the largest free extent we have available in this block | |
695 | * group. | |
696 | */ | |
697 | static void | |
698 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
699 | { | |
700 | int i; | |
701 | int bits; | |
702 | ||
703 | grp->bb_largest_free_order = -1; /* uninit */ | |
704 | ||
705 | bits = sb->s_blocksize_bits + 1; | |
706 | for (i = bits; i >= 0; i--) { | |
707 | if (grp->bb_counters[i] > 0) { | |
708 | grp->bb_largest_free_order = i; | |
709 | break; | |
710 | } | |
711 | } | |
712 | } | |
713 | ||
089ceecc ES |
714 | static noinline_for_stack |
715 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
716 | void *buddy, void *bitmap, ext4_group_t group) |
717 | { | |
718 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
7137d7a4 | 719 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
720 | ext4_grpblk_t i = 0; |
721 | ext4_grpblk_t first; | |
722 | ext4_grpblk_t len; | |
c9de560d AT |
723 | unsigned free = 0; |
724 | unsigned fragments = 0; | |
725 | unsigned long long period = get_cycles(); | |
726 | ||
727 | /* initialize buddy from bitmap which is aggregation | |
728 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 729 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
730 | grp->bb_first_free = i; |
731 | while (i < max) { | |
732 | fragments++; | |
733 | first = i; | |
ffad0a44 | 734 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
735 | len = i - first; |
736 | free += len; | |
737 | if (len > 1) | |
738 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
739 | else | |
740 | grp->bb_counters[0]++; | |
741 | if (i < max) | |
ffad0a44 | 742 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
743 | } |
744 | grp->bb_fragments = fragments; | |
745 | ||
746 | if (free != grp->bb_free) { | |
e29136f8 | 747 | ext4_grp_locked_error(sb, group, 0, 0, |
53accfa9 | 748 | "%u clusters in bitmap, %u in gd", |
e29136f8 | 749 | free, grp->bb_free); |
e56eb659 AK |
750 | /* |
751 | * If we intent to continue, we consider group descritor | |
752 | * corrupt and update bb_free using bitmap value | |
753 | */ | |
c9de560d AT |
754 | grp->bb_free = free; |
755 | } | |
8a57d9d6 | 756 | mb_set_largest_free_order(sb, grp); |
c9de560d AT |
757 | |
758 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
759 | ||
760 | period = get_cycles() - period; | |
761 | spin_lock(&EXT4_SB(sb)->s_bal_lock); | |
762 | EXT4_SB(sb)->s_mb_buddies_generated++; | |
763 | EXT4_SB(sb)->s_mb_generation_time += period; | |
764 | spin_unlock(&EXT4_SB(sb)->s_bal_lock); | |
765 | } | |
766 | ||
767 | /* The buddy information is attached the buddy cache inode | |
768 | * for convenience. The information regarding each group | |
769 | * is loaded via ext4_mb_load_buddy. The information involve | |
770 | * block bitmap and buddy information. The information are | |
771 | * stored in the inode as | |
772 | * | |
773 | * { page } | |
c3a326a6 | 774 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
775 | * |
776 | * | |
777 | * one block each for bitmap and buddy information. | |
778 | * So for each group we take up 2 blocks. A page can | |
779 | * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. | |
780 | * So it can have information regarding groups_per_page which | |
781 | * is blocks_per_page/2 | |
8a57d9d6 CW |
782 | * |
783 | * Locking note: This routine takes the block group lock of all groups | |
784 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
785 | */ |
786 | ||
787 | static int ext4_mb_init_cache(struct page *page, char *incore) | |
788 | { | |
8df9675f | 789 | ext4_group_t ngroups; |
c9de560d AT |
790 | int blocksize; |
791 | int blocks_per_page; | |
792 | int groups_per_page; | |
793 | int err = 0; | |
794 | int i; | |
813e5727 | 795 | ext4_group_t first_group, group; |
c9de560d AT |
796 | int first_block; |
797 | struct super_block *sb; | |
798 | struct buffer_head *bhs; | |
fa77dcfa | 799 | struct buffer_head **bh = NULL; |
c9de560d AT |
800 | struct inode *inode; |
801 | char *data; | |
802 | char *bitmap; | |
9b8b7d35 | 803 | struct ext4_group_info *grinfo; |
c9de560d | 804 | |
6ba495e9 | 805 | mb_debug(1, "init page %lu\n", page->index); |
c9de560d AT |
806 | |
807 | inode = page->mapping->host; | |
808 | sb = inode->i_sb; | |
8df9675f | 809 | ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
810 | blocksize = 1 << inode->i_blkbits; |
811 | blocks_per_page = PAGE_CACHE_SIZE / blocksize; | |
812 | ||
813 | groups_per_page = blocks_per_page >> 1; | |
814 | if (groups_per_page == 0) | |
815 | groups_per_page = 1; | |
816 | ||
817 | /* allocate buffer_heads to read bitmaps */ | |
818 | if (groups_per_page > 1) { | |
c9de560d AT |
819 | i = sizeof(struct buffer_head *) * groups_per_page; |
820 | bh = kzalloc(i, GFP_NOFS); | |
813e5727 TT |
821 | if (bh == NULL) { |
822 | err = -ENOMEM; | |
c9de560d | 823 | goto out; |
813e5727 | 824 | } |
c9de560d AT |
825 | } else |
826 | bh = &bhs; | |
827 | ||
828 | first_group = page->index * blocks_per_page / 2; | |
829 | ||
830 | /* read all groups the page covers into the cache */ | |
813e5727 TT |
831 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
832 | if (group >= ngroups) | |
c9de560d AT |
833 | break; |
834 | ||
813e5727 | 835 | grinfo = ext4_get_group_info(sb, group); |
9b8b7d35 AG |
836 | /* |
837 | * If page is uptodate then we came here after online resize | |
838 | * which added some new uninitialized group info structs, so | |
839 | * we must skip all initialized uptodate buddies on the page, | |
840 | * which may be currently in use by an allocating task. | |
841 | */ | |
842 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
843 | bh[i] = NULL; | |
844 | continue; | |
845 | } | |
813e5727 TT |
846 | if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) { |
847 | err = -ENOMEM; | |
c9de560d | 848 | goto out; |
2ccb5fb9 | 849 | } |
813e5727 | 850 | mb_debug(1, "read bitmap for group %u\n", group); |
c9de560d AT |
851 | } |
852 | ||
853 | /* wait for I/O completion */ | |
813e5727 TT |
854 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
855 | if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) { | |
856 | err = -EIO; | |
c9de560d | 857 | goto out; |
813e5727 TT |
858 | } |
859 | } | |
c9de560d AT |
860 | |
861 | first_block = page->index * blocks_per_page; | |
862 | for (i = 0; i < blocks_per_page; i++) { | |
863 | int group; | |
c9de560d AT |
864 | |
865 | group = (first_block + i) >> 1; | |
8df9675f | 866 | if (group >= ngroups) |
c9de560d AT |
867 | break; |
868 | ||
9b8b7d35 AG |
869 | if (!bh[group - first_group]) |
870 | /* skip initialized uptodate buddy */ | |
871 | continue; | |
872 | ||
c9de560d AT |
873 | /* |
874 | * data carry information regarding this | |
875 | * particular group in the format specified | |
876 | * above | |
877 | * | |
878 | */ | |
879 | data = page_address(page) + (i * blocksize); | |
880 | bitmap = bh[group - first_group]->b_data; | |
881 | ||
882 | /* | |
883 | * We place the buddy block and bitmap block | |
884 | * close together | |
885 | */ | |
886 | if ((first_block + i) & 1) { | |
887 | /* this is block of buddy */ | |
888 | BUG_ON(incore == NULL); | |
6ba495e9 | 889 | mb_debug(1, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 890 | group, page->index, i * blocksize); |
f307333e | 891 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
892 | grinfo = ext4_get_group_info(sb, group); |
893 | grinfo->bb_fragments = 0; | |
894 | memset(grinfo->bb_counters, 0, | |
1927805e ES |
895 | sizeof(*grinfo->bb_counters) * |
896 | (sb->s_blocksize_bits+2)); | |
c9de560d AT |
897 | /* |
898 | * incore got set to the group block bitmap below | |
899 | */ | |
7a2fcbf7 | 900 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
901 | /* init the buddy */ |
902 | memset(data, 0xff, blocksize); | |
c9de560d | 903 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 904 | ext4_unlock_group(sb, group); |
c9de560d AT |
905 | incore = NULL; |
906 | } else { | |
907 | /* this is block of bitmap */ | |
908 | BUG_ON(incore != NULL); | |
6ba495e9 | 909 | mb_debug(1, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 910 | group, page->index, i * blocksize); |
f307333e | 911 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
912 | |
913 | /* see comments in ext4_mb_put_pa() */ | |
914 | ext4_lock_group(sb, group); | |
915 | memcpy(data, bitmap, blocksize); | |
916 | ||
917 | /* mark all preallocated blks used in in-core bitmap */ | |
918 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 919 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
920 | ext4_unlock_group(sb, group); |
921 | ||
922 | /* set incore so that the buddy information can be | |
923 | * generated using this | |
924 | */ | |
925 | incore = data; | |
926 | } | |
927 | } | |
928 | SetPageUptodate(page); | |
929 | ||
930 | out: | |
931 | if (bh) { | |
9b8b7d35 | 932 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
933 | brelse(bh[i]); |
934 | if (bh != &bhs) | |
935 | kfree(bh); | |
936 | } | |
937 | return err; | |
938 | } | |
939 | ||
eee4adc7 | 940 | /* |
2de8807b AG |
941 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
942 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
943 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
944 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 945 | */ |
2de8807b AG |
946 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
947 | ext4_group_t group, struct ext4_buddy *e4b) | |
eee4adc7 | 948 | { |
2de8807b AG |
949 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
950 | int block, pnum, poff; | |
eee4adc7 | 951 | int blocks_per_page; |
2de8807b AG |
952 | struct page *page; |
953 | ||
954 | e4b->bd_buddy_page = NULL; | |
955 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 ES |
956 | |
957 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
958 | /* | |
959 | * the buddy cache inode stores the block bitmap | |
960 | * and buddy information in consecutive blocks. | |
961 | * So for each group we need two blocks. | |
962 | */ | |
963 | block = group * 2; | |
964 | pnum = block / blocks_per_page; | |
2de8807b AG |
965 | poff = block % blocks_per_page; |
966 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
967 | if (!page) | |
968 | return -EIO; | |
969 | BUG_ON(page->mapping != inode->i_mapping); | |
970 | e4b->bd_bitmap_page = page; | |
971 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
972 | ||
973 | if (blocks_per_page >= 2) { | |
974 | /* buddy and bitmap are on the same page */ | |
975 | return 0; | |
eee4adc7 | 976 | } |
2de8807b AG |
977 | |
978 | block++; | |
979 | pnum = block / blocks_per_page; | |
2de8807b AG |
980 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); |
981 | if (!page) | |
982 | return -EIO; | |
983 | BUG_ON(page->mapping != inode->i_mapping); | |
984 | e4b->bd_buddy_page = page; | |
985 | return 0; | |
eee4adc7 ES |
986 | } |
987 | ||
2de8807b | 988 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 989 | { |
2de8807b AG |
990 | if (e4b->bd_bitmap_page) { |
991 | unlock_page(e4b->bd_bitmap_page); | |
992 | page_cache_release(e4b->bd_bitmap_page); | |
993 | } | |
994 | if (e4b->bd_buddy_page) { | |
995 | unlock_page(e4b->bd_buddy_page); | |
996 | page_cache_release(e4b->bd_buddy_page); | |
eee4adc7 | 997 | } |
eee4adc7 ES |
998 | } |
999 | ||
8a57d9d6 CW |
1000 | /* |
1001 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1002 | * block group lock of all groups for this page; do not hold the BG lock when | |
1003 | * calling this routine! | |
1004 | */ | |
b6a758ec AK |
1005 | static noinline_for_stack |
1006 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) | |
1007 | { | |
1008 | ||
b6a758ec | 1009 | struct ext4_group_info *this_grp; |
2de8807b AG |
1010 | struct ext4_buddy e4b; |
1011 | struct page *page; | |
1012 | int ret = 0; | |
b6a758ec AK |
1013 | |
1014 | mb_debug(1, "init group %u\n", group); | |
b6a758ec AK |
1015 | this_grp = ext4_get_group_info(sb, group); |
1016 | /* | |
08c3a813 AK |
1017 | * This ensures that we don't reinit the buddy cache |
1018 | * page which map to the group from which we are already | |
1019 | * allocating. If we are looking at the buddy cache we would | |
1020 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1021 | * would have pinned buddy page to page cache. |
b6a758ec | 1022 | */ |
2de8807b AG |
1023 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b); |
1024 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { | |
b6a758ec AK |
1025 | /* |
1026 | * somebody initialized the group | |
1027 | * return without doing anything | |
1028 | */ | |
b6a758ec AK |
1029 | goto err; |
1030 | } | |
2de8807b AG |
1031 | |
1032 | page = e4b.bd_bitmap_page; | |
1033 | ret = ext4_mb_init_cache(page, NULL); | |
1034 | if (ret) | |
1035 | goto err; | |
1036 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1037 | ret = -EIO; |
1038 | goto err; | |
1039 | } | |
1040 | mark_page_accessed(page); | |
b6a758ec | 1041 | |
2de8807b | 1042 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1043 | /* |
1044 | * If both the bitmap and buddy are in | |
1045 | * the same page we don't need to force | |
1046 | * init the buddy | |
1047 | */ | |
2de8807b AG |
1048 | ret = 0; |
1049 | goto err; | |
b6a758ec | 1050 | } |
2de8807b AG |
1051 | /* init buddy cache */ |
1052 | page = e4b.bd_buddy_page; | |
1053 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap); | |
1054 | if (ret) | |
1055 | goto err; | |
1056 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1057 | ret = -EIO; |
1058 | goto err; | |
1059 | } | |
1060 | mark_page_accessed(page); | |
1061 | err: | |
2de8807b | 1062 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1063 | return ret; |
1064 | } | |
1065 | ||
8a57d9d6 CW |
1066 | /* |
1067 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1068 | * block group lock of all groups for this page; do not hold the BG lock when | |
1069 | * calling this routine! | |
1070 | */ | |
4ddfef7b ES |
1071 | static noinline_for_stack int |
1072 | ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, | |
1073 | struct ext4_buddy *e4b) | |
c9de560d | 1074 | { |
c9de560d AT |
1075 | int blocks_per_page; |
1076 | int block; | |
1077 | int pnum; | |
1078 | int poff; | |
1079 | struct page *page; | |
fdf6c7a7 | 1080 | int ret; |
920313a7 AK |
1081 | struct ext4_group_info *grp; |
1082 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1083 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1084 | |
6ba495e9 | 1085 | mb_debug(1, "load group %u\n", group); |
c9de560d AT |
1086 | |
1087 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
920313a7 | 1088 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1089 | |
1090 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1091 | e4b->bd_info = grp; |
c9de560d AT |
1092 | e4b->bd_sb = sb; |
1093 | e4b->bd_group = group; | |
1094 | e4b->bd_buddy_page = NULL; | |
1095 | e4b->bd_bitmap_page = NULL; | |
1096 | ||
f41c0750 | 1097 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1098 | /* |
1099 | * we need full data about the group | |
1100 | * to make a good selection | |
1101 | */ | |
1102 | ret = ext4_mb_init_group(sb, group); | |
1103 | if (ret) | |
1104 | return ret; | |
f41c0750 AK |
1105 | } |
1106 | ||
c9de560d AT |
1107 | /* |
1108 | * the buddy cache inode stores the block bitmap | |
1109 | * and buddy information in consecutive blocks. | |
1110 | * So for each group we need two blocks. | |
1111 | */ | |
1112 | block = group * 2; | |
1113 | pnum = block / blocks_per_page; | |
1114 | poff = block % blocks_per_page; | |
1115 | ||
1116 | /* we could use find_or_create_page(), but it locks page | |
1117 | * what we'd like to avoid in fast path ... */ | |
1118 | page = find_get_page(inode->i_mapping, pnum); | |
1119 | if (page == NULL || !PageUptodate(page)) { | |
1120 | if (page) | |
920313a7 AK |
1121 | /* |
1122 | * drop the page reference and try | |
1123 | * to get the page with lock. If we | |
1124 | * are not uptodate that implies | |
1125 | * somebody just created the page but | |
1126 | * is yet to initialize the same. So | |
1127 | * wait for it to initialize. | |
1128 | */ | |
c9de560d AT |
1129 | page_cache_release(page); |
1130 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1131 | if (page) { | |
1132 | BUG_ON(page->mapping != inode->i_mapping); | |
1133 | if (!PageUptodate(page)) { | |
fdf6c7a7 SF |
1134 | ret = ext4_mb_init_cache(page, NULL); |
1135 | if (ret) { | |
1136 | unlock_page(page); | |
1137 | goto err; | |
1138 | } | |
c9de560d AT |
1139 | mb_cmp_bitmaps(e4b, page_address(page) + |
1140 | (poff * sb->s_blocksize)); | |
1141 | } | |
1142 | unlock_page(page); | |
1143 | } | |
1144 | } | |
fdf6c7a7 SF |
1145 | if (page == NULL || !PageUptodate(page)) { |
1146 | ret = -EIO; | |
c9de560d | 1147 | goto err; |
fdf6c7a7 | 1148 | } |
c9de560d AT |
1149 | e4b->bd_bitmap_page = page; |
1150 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1151 | mark_page_accessed(page); | |
1152 | ||
1153 | block++; | |
1154 | pnum = block / blocks_per_page; | |
1155 | poff = block % blocks_per_page; | |
1156 | ||
1157 | page = find_get_page(inode->i_mapping, pnum); | |
1158 | if (page == NULL || !PageUptodate(page)) { | |
1159 | if (page) | |
1160 | page_cache_release(page); | |
1161 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1162 | if (page) { | |
1163 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 SF |
1164 | if (!PageUptodate(page)) { |
1165 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap); | |
1166 | if (ret) { | |
1167 | unlock_page(page); | |
1168 | goto err; | |
1169 | } | |
1170 | } | |
c9de560d AT |
1171 | unlock_page(page); |
1172 | } | |
1173 | } | |
fdf6c7a7 SF |
1174 | if (page == NULL || !PageUptodate(page)) { |
1175 | ret = -EIO; | |
c9de560d | 1176 | goto err; |
fdf6c7a7 | 1177 | } |
c9de560d AT |
1178 | e4b->bd_buddy_page = page; |
1179 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
1180 | mark_page_accessed(page); | |
1181 | ||
1182 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1183 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1184 | ||
1185 | return 0; | |
1186 | ||
1187 | err: | |
26626f11 YR |
1188 | if (page) |
1189 | page_cache_release(page); | |
c9de560d AT |
1190 | if (e4b->bd_bitmap_page) |
1191 | page_cache_release(e4b->bd_bitmap_page); | |
1192 | if (e4b->bd_buddy_page) | |
1193 | page_cache_release(e4b->bd_buddy_page); | |
1194 | e4b->bd_buddy = NULL; | |
1195 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1196 | return ret; |
c9de560d AT |
1197 | } |
1198 | ||
e39e07fd | 1199 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1200 | { |
1201 | if (e4b->bd_bitmap_page) | |
1202 | page_cache_release(e4b->bd_bitmap_page); | |
1203 | if (e4b->bd_buddy_page) | |
1204 | page_cache_release(e4b->bd_buddy_page); | |
1205 | } | |
1206 | ||
1207 | ||
1208 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1209 | { | |
1210 | int order = 1; | |
1211 | void *bb; | |
1212 | ||
c5e8f3f3 | 1213 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1214 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1215 | ||
c5e8f3f3 | 1216 | bb = e4b->bd_buddy; |
c9de560d AT |
1217 | while (order <= e4b->bd_blkbits + 1) { |
1218 | block = block >> 1; | |
1219 | if (!mb_test_bit(block, bb)) { | |
1220 | /* this block is part of buddy of order 'order' */ | |
1221 | return order; | |
1222 | } | |
1223 | bb += 1 << (e4b->bd_blkbits - order); | |
1224 | order++; | |
1225 | } | |
1226 | return 0; | |
1227 | } | |
1228 | ||
955ce5f5 | 1229 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1230 | { |
1231 | __u32 *addr; | |
1232 | ||
1233 | len = cur + len; | |
1234 | while (cur < len) { | |
1235 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1236 | /* fast path: clear whole word at once */ | |
1237 | addr = bm + (cur >> 3); | |
1238 | *addr = 0; | |
1239 | cur += 32; | |
1240 | continue; | |
1241 | } | |
955ce5f5 | 1242 | mb_clear_bit(cur, bm); |
c9de560d AT |
1243 | cur++; |
1244 | } | |
1245 | } | |
1246 | ||
c3e94d1d | 1247 | void ext4_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1248 | { |
1249 | __u32 *addr; | |
1250 | ||
1251 | len = cur + len; | |
1252 | while (cur < len) { | |
1253 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1254 | /* fast path: set whole word at once */ | |
1255 | addr = bm + (cur >> 3); | |
1256 | *addr = 0xffffffff; | |
1257 | cur += 32; | |
1258 | continue; | |
1259 | } | |
955ce5f5 | 1260 | mb_set_bit(cur, bm); |
c9de560d AT |
1261 | cur++; |
1262 | } | |
1263 | } | |
1264 | ||
7e5a8cdd | 1265 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
c9de560d AT |
1266 | int first, int count) |
1267 | { | |
1268 | int block = 0; | |
1269 | int max = 0; | |
1270 | int order; | |
1271 | void *buddy; | |
1272 | void *buddy2; | |
1273 | struct super_block *sb = e4b->bd_sb; | |
1274 | ||
1275 | BUG_ON(first + count > (sb->s_blocksize << 3)); | |
bc8e6740 | 1276 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
1277 | mb_check_buddy(e4b); |
1278 | mb_free_blocks_double(inode, e4b, first, count); | |
1279 | ||
1280 | e4b->bd_info->bb_free += count; | |
1281 | if (first < e4b->bd_info->bb_first_free) | |
1282 | e4b->bd_info->bb_first_free = first; | |
1283 | ||
1284 | /* let's maintain fragments counter */ | |
1285 | if (first != 0) | |
c5e8f3f3 | 1286 | block = !mb_test_bit(first - 1, e4b->bd_bitmap); |
c9de560d | 1287 | if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1288 | max = !mb_test_bit(first + count, e4b->bd_bitmap); |
c9de560d AT |
1289 | if (block && max) |
1290 | e4b->bd_info->bb_fragments--; | |
1291 | else if (!block && !max) | |
1292 | e4b->bd_info->bb_fragments++; | |
1293 | ||
1294 | /* let's maintain buddy itself */ | |
1295 | while (count-- > 0) { | |
1296 | block = first++; | |
1297 | order = 0; | |
1298 | ||
c5e8f3f3 | 1299 | if (!mb_test_bit(block, e4b->bd_bitmap)) { |
c9de560d | 1300 | ext4_fsblk_t blocknr; |
5661bd68 AM |
1301 | |
1302 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 1303 | blocknr += EXT4_C2B(EXT4_SB(sb), block); |
5d1b1b3f | 1304 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
1305 | inode ? inode->i_ino : 0, |
1306 | blocknr, | |
1307 | "freeing already freed block " | |
1308 | "(bit %u)", block); | |
c9de560d | 1309 | } |
c5e8f3f3 | 1310 | mb_clear_bit(block, e4b->bd_bitmap); |
c9de560d AT |
1311 | e4b->bd_info->bb_counters[order]++; |
1312 | ||
1313 | /* start of the buddy */ | |
1314 | buddy = mb_find_buddy(e4b, order, &max); | |
1315 | ||
1316 | do { | |
1317 | block &= ~1UL; | |
1318 | if (mb_test_bit(block, buddy) || | |
1319 | mb_test_bit(block + 1, buddy)) | |
1320 | break; | |
1321 | ||
1322 | /* both the buddies are free, try to coalesce them */ | |
1323 | buddy2 = mb_find_buddy(e4b, order + 1, &max); | |
1324 | ||
1325 | if (!buddy2) | |
1326 | break; | |
1327 | ||
1328 | if (order > 0) { | |
1329 | /* for special purposes, we don't set | |
1330 | * free bits in bitmap */ | |
1331 | mb_set_bit(block, buddy); | |
1332 | mb_set_bit(block + 1, buddy); | |
1333 | } | |
1334 | e4b->bd_info->bb_counters[order]--; | |
1335 | e4b->bd_info->bb_counters[order]--; | |
1336 | ||
1337 | block = block >> 1; | |
1338 | order++; | |
1339 | e4b->bd_info->bb_counters[order]++; | |
1340 | ||
1341 | mb_clear_bit(block, buddy2); | |
1342 | buddy = buddy2; | |
1343 | } while (1); | |
1344 | } | |
8a57d9d6 | 1345 | mb_set_largest_free_order(sb, e4b->bd_info); |
c9de560d | 1346 | mb_check_buddy(e4b); |
c9de560d AT |
1347 | } |
1348 | ||
15c006a2 | 1349 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1350 | int needed, struct ext4_free_extent *ex) |
1351 | { | |
1352 | int next = block; | |
15c006a2 | 1353 | int max, order; |
c9de560d AT |
1354 | void *buddy; |
1355 | ||
bc8e6740 | 1356 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1357 | BUG_ON(ex == NULL); |
1358 | ||
15c006a2 | 1359 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1360 | BUG_ON(buddy == NULL); |
1361 | BUG_ON(block >= max); | |
1362 | if (mb_test_bit(block, buddy)) { | |
1363 | ex->fe_len = 0; | |
1364 | ex->fe_start = 0; | |
1365 | ex->fe_group = 0; | |
1366 | return 0; | |
1367 | } | |
1368 | ||
15c006a2 RD |
1369 | /* find actual order */ |
1370 | order = mb_find_order_for_block(e4b, block); | |
1371 | block = block >> order; | |
c9de560d AT |
1372 | |
1373 | ex->fe_len = 1 << order; | |
1374 | ex->fe_start = block << order; | |
1375 | ex->fe_group = e4b->bd_group; | |
1376 | ||
1377 | /* calc difference from given start */ | |
1378 | next = next - ex->fe_start; | |
1379 | ex->fe_len -= next; | |
1380 | ex->fe_start += next; | |
1381 | ||
1382 | while (needed > ex->fe_len && | |
d8ec0c39 | 1383 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1384 | |
1385 | if (block + 1 >= max) | |
1386 | break; | |
1387 | ||
1388 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1389 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1390 | break; |
1391 | ||
b051d8dc | 1392 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1393 | |
c9de560d AT |
1394 | block = next >> order; |
1395 | ex->fe_len += 1 << order; | |
1396 | } | |
1397 | ||
1398 | BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); | |
1399 | return ex->fe_len; | |
1400 | } | |
1401 | ||
1402 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1403 | { | |
1404 | int ord; | |
1405 | int mlen = 0; | |
1406 | int max = 0; | |
1407 | int cur; | |
1408 | int start = ex->fe_start; | |
1409 | int len = ex->fe_len; | |
1410 | unsigned ret = 0; | |
1411 | int len0 = len; | |
1412 | void *buddy; | |
1413 | ||
1414 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1415 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1416 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1417 | mb_check_buddy(e4b); |
1418 | mb_mark_used_double(e4b, start, len); | |
1419 | ||
1420 | e4b->bd_info->bb_free -= len; | |
1421 | if (e4b->bd_info->bb_first_free == start) | |
1422 | e4b->bd_info->bb_first_free += len; | |
1423 | ||
1424 | /* let's maintain fragments counter */ | |
1425 | if (start != 0) | |
c5e8f3f3 | 1426 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1427 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1428 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1429 | if (mlen && max) |
1430 | e4b->bd_info->bb_fragments++; | |
1431 | else if (!mlen && !max) | |
1432 | e4b->bd_info->bb_fragments--; | |
1433 | ||
1434 | /* let's maintain buddy itself */ | |
1435 | while (len) { | |
1436 | ord = mb_find_order_for_block(e4b, start); | |
1437 | ||
1438 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1439 | /* the whole chunk may be allocated at once! */ | |
1440 | mlen = 1 << ord; | |
1441 | buddy = mb_find_buddy(e4b, ord, &max); | |
1442 | BUG_ON((start >> ord) >= max); | |
1443 | mb_set_bit(start >> ord, buddy); | |
1444 | e4b->bd_info->bb_counters[ord]--; | |
1445 | start += mlen; | |
1446 | len -= mlen; | |
1447 | BUG_ON(len < 0); | |
1448 | continue; | |
1449 | } | |
1450 | ||
1451 | /* store for history */ | |
1452 | if (ret == 0) | |
1453 | ret = len | (ord << 16); | |
1454 | ||
1455 | /* we have to split large buddy */ | |
1456 | BUG_ON(ord <= 0); | |
1457 | buddy = mb_find_buddy(e4b, ord, &max); | |
1458 | mb_set_bit(start >> ord, buddy); | |
1459 | e4b->bd_info->bb_counters[ord]--; | |
1460 | ||
1461 | ord--; | |
1462 | cur = (start >> ord) & ~1U; | |
1463 | buddy = mb_find_buddy(e4b, ord, &max); | |
1464 | mb_clear_bit(cur, buddy); | |
1465 | mb_clear_bit(cur + 1, buddy); | |
1466 | e4b->bd_info->bb_counters[ord]++; | |
1467 | e4b->bd_info->bb_counters[ord]++; | |
1468 | } | |
8a57d9d6 | 1469 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1470 | |
c5e8f3f3 | 1471 | ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
1472 | mb_check_buddy(e4b); |
1473 | ||
1474 | return ret; | |
1475 | } | |
1476 | ||
1477 | /* | |
1478 | * Must be called under group lock! | |
1479 | */ | |
1480 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1481 | struct ext4_buddy *e4b) | |
1482 | { | |
1483 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1484 | int ret; | |
1485 | ||
1486 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1487 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1488 | ||
1489 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1490 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1491 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1492 | ||
1493 | /* preallocation can change ac_b_ex, thus we store actually | |
1494 | * allocated blocks for history */ | |
1495 | ac->ac_f_ex = ac->ac_b_ex; | |
1496 | ||
1497 | ac->ac_status = AC_STATUS_FOUND; | |
1498 | ac->ac_tail = ret & 0xffff; | |
1499 | ac->ac_buddy = ret >> 16; | |
1500 | ||
c3a326a6 AK |
1501 | /* |
1502 | * take the page reference. We want the page to be pinned | |
1503 | * so that we don't get a ext4_mb_init_cache_call for this | |
1504 | * group until we update the bitmap. That would mean we | |
1505 | * double allocate blocks. The reference is dropped | |
1506 | * in ext4_mb_release_context | |
1507 | */ | |
c9de560d AT |
1508 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1509 | get_page(ac->ac_bitmap_page); | |
1510 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1511 | get_page(ac->ac_buddy_page); | |
c9de560d | 1512 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 1513 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
1514 | spin_lock(&sbi->s_md_lock); |
1515 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1516 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1517 | spin_unlock(&sbi->s_md_lock); | |
1518 | } | |
1519 | } | |
1520 | ||
1521 | /* | |
1522 | * regular allocator, for general purposes allocation | |
1523 | */ | |
1524 | ||
1525 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1526 | struct ext4_buddy *e4b, | |
1527 | int finish_group) | |
1528 | { | |
1529 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1530 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1531 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1532 | struct ext4_free_extent ex; | |
1533 | int max; | |
1534 | ||
032115fc AK |
1535 | if (ac->ac_status == AC_STATUS_FOUND) |
1536 | return; | |
c9de560d AT |
1537 | /* |
1538 | * We don't want to scan for a whole year | |
1539 | */ | |
1540 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1541 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1542 | ac->ac_status = AC_STATUS_BREAK; | |
1543 | return; | |
1544 | } | |
1545 | ||
1546 | /* | |
1547 | * Haven't found good chunk so far, let's continue | |
1548 | */ | |
1549 | if (bex->fe_len < gex->fe_len) | |
1550 | return; | |
1551 | ||
1552 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1553 | && bex->fe_group == e4b->bd_group) { | |
1554 | /* recheck chunk's availability - we don't know | |
1555 | * when it was found (within this lock-unlock | |
1556 | * period or not) */ | |
15c006a2 | 1557 | max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); |
c9de560d AT |
1558 | if (max >= gex->fe_len) { |
1559 | ext4_mb_use_best_found(ac, e4b); | |
1560 | return; | |
1561 | } | |
1562 | } | |
1563 | } | |
1564 | ||
1565 | /* | |
1566 | * The routine checks whether found extent is good enough. If it is, | |
1567 | * then the extent gets marked used and flag is set to the context | |
1568 | * to stop scanning. Otherwise, the extent is compared with the | |
1569 | * previous found extent and if new one is better, then it's stored | |
1570 | * in the context. Later, the best found extent will be used, if | |
1571 | * mballoc can't find good enough extent. | |
1572 | * | |
1573 | * FIXME: real allocation policy is to be designed yet! | |
1574 | */ | |
1575 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1576 | struct ext4_free_extent *ex, | |
1577 | struct ext4_buddy *e4b) | |
1578 | { | |
1579 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1580 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1581 | ||
1582 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
1583 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
1584 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
1585 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
1586 | ||
1587 | ac->ac_found++; | |
1588 | ||
1589 | /* | |
1590 | * The special case - take what you catch first | |
1591 | */ | |
1592 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1593 | *bex = *ex; | |
1594 | ext4_mb_use_best_found(ac, e4b); | |
1595 | return; | |
1596 | } | |
1597 | ||
1598 | /* | |
1599 | * Let's check whether the chuck is good enough | |
1600 | */ | |
1601 | if (ex->fe_len == gex->fe_len) { | |
1602 | *bex = *ex; | |
1603 | ext4_mb_use_best_found(ac, e4b); | |
1604 | return; | |
1605 | } | |
1606 | ||
1607 | /* | |
1608 | * If this is first found extent, just store it in the context | |
1609 | */ | |
1610 | if (bex->fe_len == 0) { | |
1611 | *bex = *ex; | |
1612 | return; | |
1613 | } | |
1614 | ||
1615 | /* | |
1616 | * If new found extent is better, store it in the context | |
1617 | */ | |
1618 | if (bex->fe_len < gex->fe_len) { | |
1619 | /* if the request isn't satisfied, any found extent | |
1620 | * larger than previous best one is better */ | |
1621 | if (ex->fe_len > bex->fe_len) | |
1622 | *bex = *ex; | |
1623 | } else if (ex->fe_len > gex->fe_len) { | |
1624 | /* if the request is satisfied, then we try to find | |
1625 | * an extent that still satisfy the request, but is | |
1626 | * smaller than previous one */ | |
1627 | if (ex->fe_len < bex->fe_len) | |
1628 | *bex = *ex; | |
1629 | } | |
1630 | ||
1631 | ext4_mb_check_limits(ac, e4b, 0); | |
1632 | } | |
1633 | ||
089ceecc ES |
1634 | static noinline_for_stack |
1635 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
1636 | struct ext4_buddy *e4b) |
1637 | { | |
1638 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1639 | ext4_group_t group = ex.fe_group; | |
1640 | int max; | |
1641 | int err; | |
1642 | ||
1643 | BUG_ON(ex.fe_len <= 0); | |
1644 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1645 | if (err) | |
1646 | return err; | |
1647 | ||
1648 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 1649 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
1650 | |
1651 | if (max > 0) { | |
1652 | ac->ac_b_ex = ex; | |
1653 | ext4_mb_use_best_found(ac, e4b); | |
1654 | } | |
1655 | ||
1656 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1657 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1658 | |
1659 | return 0; | |
1660 | } | |
1661 | ||
089ceecc ES |
1662 | static noinline_for_stack |
1663 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
1664 | struct ext4_buddy *e4b) |
1665 | { | |
1666 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1667 | int max; | |
1668 | int err; | |
1669 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 1670 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
1671 | struct ext4_free_extent ex; |
1672 | ||
1673 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1674 | return 0; | |
838cd0cf YY |
1675 | if (grp->bb_free == 0) |
1676 | return 0; | |
c9de560d AT |
1677 | |
1678 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1679 | if (err) | |
1680 | return err; | |
1681 | ||
1682 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 1683 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d AT |
1684 | ac->ac_g_ex.fe_len, &ex); |
1685 | ||
1686 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1687 | ext4_fsblk_t start; | |
1688 | ||
5661bd68 AM |
1689 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
1690 | ex.fe_start; | |
c9de560d AT |
1691 | /* use do_div to get remainder (would be 64-bit modulo) */ |
1692 | if (do_div(start, sbi->s_stripe) == 0) { | |
1693 | ac->ac_found++; | |
1694 | ac->ac_b_ex = ex; | |
1695 | ext4_mb_use_best_found(ac, e4b); | |
1696 | } | |
1697 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1698 | BUG_ON(ex.fe_len <= 0); | |
1699 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1700 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1701 | ac->ac_found++; | |
1702 | ac->ac_b_ex = ex; | |
1703 | ext4_mb_use_best_found(ac, e4b); | |
1704 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1705 | /* Sometimes, caller may want to merge even small | |
1706 | * number of blocks to an existing extent */ | |
1707 | BUG_ON(ex.fe_len <= 0); | |
1708 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1709 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1710 | ac->ac_found++; | |
1711 | ac->ac_b_ex = ex; | |
1712 | ext4_mb_use_best_found(ac, e4b); | |
1713 | } | |
1714 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1715 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1716 | |
1717 | return 0; | |
1718 | } | |
1719 | ||
1720 | /* | |
1721 | * The routine scans buddy structures (not bitmap!) from given order | |
1722 | * to max order and tries to find big enough chunk to satisfy the req | |
1723 | */ | |
089ceecc ES |
1724 | static noinline_for_stack |
1725 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1726 | struct ext4_buddy *e4b) |
1727 | { | |
1728 | struct super_block *sb = ac->ac_sb; | |
1729 | struct ext4_group_info *grp = e4b->bd_info; | |
1730 | void *buddy; | |
1731 | int i; | |
1732 | int k; | |
1733 | int max; | |
1734 | ||
1735 | BUG_ON(ac->ac_2order <= 0); | |
1736 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1737 | if (grp->bb_counters[i] == 0) | |
1738 | continue; | |
1739 | ||
1740 | buddy = mb_find_buddy(e4b, i, &max); | |
1741 | BUG_ON(buddy == NULL); | |
1742 | ||
ffad0a44 | 1743 | k = mb_find_next_zero_bit(buddy, max, 0); |
c9de560d AT |
1744 | BUG_ON(k >= max); |
1745 | ||
1746 | ac->ac_found++; | |
1747 | ||
1748 | ac->ac_b_ex.fe_len = 1 << i; | |
1749 | ac->ac_b_ex.fe_start = k << i; | |
1750 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1751 | ||
1752 | ext4_mb_use_best_found(ac, e4b); | |
1753 | ||
1754 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1755 | ||
1756 | if (EXT4_SB(sb)->s_mb_stats) | |
1757 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1758 | ||
1759 | break; | |
1760 | } | |
1761 | } | |
1762 | ||
1763 | /* | |
1764 | * The routine scans the group and measures all found extents. | |
1765 | * In order to optimize scanning, caller must pass number of | |
1766 | * free blocks in the group, so the routine can know upper limit. | |
1767 | */ | |
089ceecc ES |
1768 | static noinline_for_stack |
1769 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1770 | struct ext4_buddy *e4b) |
1771 | { | |
1772 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 1773 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
1774 | struct ext4_free_extent ex; |
1775 | int i; | |
1776 | int free; | |
1777 | ||
1778 | free = e4b->bd_info->bb_free; | |
1779 | BUG_ON(free <= 0); | |
1780 | ||
1781 | i = e4b->bd_info->bb_first_free; | |
1782 | ||
1783 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1784 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
1785 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
1786 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 1787 | /* |
e56eb659 | 1788 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1789 | * free blocks even though group info says we |
1790 | * we have free blocks | |
1791 | */ | |
e29136f8 | 1792 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1793 | "%d free clusters as per " |
fde4d95a | 1794 | "group info. But bitmap says 0", |
26346ff6 | 1795 | free); |
c9de560d AT |
1796 | break; |
1797 | } | |
1798 | ||
15c006a2 | 1799 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
c9de560d | 1800 | BUG_ON(ex.fe_len <= 0); |
26346ff6 | 1801 | if (free < ex.fe_len) { |
e29136f8 | 1802 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1803 | "%d free clusters as per " |
fde4d95a | 1804 | "group info. But got %d blocks", |
26346ff6 | 1805 | free, ex.fe_len); |
e56eb659 AK |
1806 | /* |
1807 | * The number of free blocks differs. This mostly | |
1808 | * indicate that the bitmap is corrupt. So exit | |
1809 | * without claiming the space. | |
1810 | */ | |
1811 | break; | |
26346ff6 | 1812 | } |
c9de560d AT |
1813 | |
1814 | ext4_mb_measure_extent(ac, &ex, e4b); | |
1815 | ||
1816 | i += ex.fe_len; | |
1817 | free -= ex.fe_len; | |
1818 | } | |
1819 | ||
1820 | ext4_mb_check_limits(ac, e4b, 1); | |
1821 | } | |
1822 | ||
1823 | /* | |
1824 | * This is a special case for storages like raid5 | |
506bf2d8 | 1825 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 1826 | */ |
089ceecc ES |
1827 | static noinline_for_stack |
1828 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
1829 | struct ext4_buddy *e4b) |
1830 | { | |
1831 | struct super_block *sb = ac->ac_sb; | |
1832 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 1833 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
1834 | struct ext4_free_extent ex; |
1835 | ext4_fsblk_t first_group_block; | |
1836 | ext4_fsblk_t a; | |
1837 | ext4_grpblk_t i; | |
1838 | int max; | |
1839 | ||
1840 | BUG_ON(sbi->s_stripe == 0); | |
1841 | ||
1842 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
1843 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
1844 | ||
c9de560d AT |
1845 | a = first_group_block + sbi->s_stripe - 1; |
1846 | do_div(a, sbi->s_stripe); | |
1847 | i = (a * sbi->s_stripe) - first_group_block; | |
1848 | ||
7137d7a4 | 1849 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 1850 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 1851 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
1852 | if (max >= sbi->s_stripe) { |
1853 | ac->ac_found++; | |
1854 | ac->ac_b_ex = ex; | |
1855 | ext4_mb_use_best_found(ac, e4b); | |
1856 | break; | |
1857 | } | |
1858 | } | |
1859 | i += sbi->s_stripe; | |
1860 | } | |
1861 | } | |
1862 | ||
8a57d9d6 | 1863 | /* This is now called BEFORE we load the buddy bitmap. */ |
c9de560d AT |
1864 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, |
1865 | ext4_group_t group, int cr) | |
1866 | { | |
1867 | unsigned free, fragments; | |
a4912123 | 1868 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
1869 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
1870 | ||
1871 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 1872 | |
01fc48e8 TT |
1873 | free = grp->bb_free; |
1874 | if (free == 0) | |
1875 | return 0; | |
1876 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
1877 | return 0; | |
1878 | ||
8a57d9d6 CW |
1879 | /* We only do this if the grp has never been initialized */ |
1880 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
1881 | int ret = ext4_mb_init_group(ac->ac_sb, group); | |
1882 | if (ret) | |
1883 | return 0; | |
1884 | } | |
c9de560d | 1885 | |
c9de560d | 1886 | fragments = grp->bb_fragments; |
c9de560d AT |
1887 | if (fragments == 0) |
1888 | return 0; | |
1889 | ||
1890 | switch (cr) { | |
1891 | case 0: | |
1892 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 1893 | |
a4912123 TT |
1894 | /* Avoid using the first bg of a flexgroup for data files */ |
1895 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
1896 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
1897 | ((group % flex_size) == 0)) | |
1898 | return 0; | |
1899 | ||
40ae3487 TT |
1900 | if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) || |
1901 | (free / fragments) >= ac->ac_g_ex.fe_len) | |
1902 | return 1; | |
1903 | ||
1904 | if (grp->bb_largest_free_order < ac->ac_2order) | |
1905 | return 0; | |
1906 | ||
8a57d9d6 | 1907 | return 1; |
c9de560d AT |
1908 | case 1: |
1909 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
1910 | return 1; | |
1911 | break; | |
1912 | case 2: | |
1913 | if (free >= ac->ac_g_ex.fe_len) | |
1914 | return 1; | |
1915 | break; | |
1916 | case 3: | |
1917 | return 1; | |
1918 | default: | |
1919 | BUG(); | |
1920 | } | |
1921 | ||
1922 | return 0; | |
1923 | } | |
1924 | ||
4ddfef7b ES |
1925 | static noinline_for_stack int |
1926 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 1927 | { |
8df9675f | 1928 | ext4_group_t ngroups, group, i; |
c9de560d AT |
1929 | int cr; |
1930 | int err = 0; | |
c9de560d AT |
1931 | struct ext4_sb_info *sbi; |
1932 | struct super_block *sb; | |
1933 | struct ext4_buddy e4b; | |
c9de560d AT |
1934 | |
1935 | sb = ac->ac_sb; | |
1936 | sbi = EXT4_SB(sb); | |
8df9675f | 1937 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 1938 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 1939 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
1940 | ngroups = sbi->s_blockfile_groups; |
1941 | ||
c9de560d AT |
1942 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
1943 | ||
1944 | /* first, try the goal */ | |
1945 | err = ext4_mb_find_by_goal(ac, &e4b); | |
1946 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
1947 | goto out; | |
1948 | ||
1949 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
1950 | goto out; | |
1951 | ||
1952 | /* | |
1953 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
1954 | * if ac2_order is set we also set criteria to 0 so that we | |
1955 | * try exact allocation using buddy. | |
1956 | */ | |
1957 | i = fls(ac->ac_g_ex.fe_len); | |
1958 | ac->ac_2order = 0; | |
1959 | /* | |
1960 | * We search using buddy data only if the order of the request | |
1961 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 1962 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
c9de560d AT |
1963 | */ |
1964 | if (i >= sbi->s_mb_order2_reqs) { | |
1965 | /* | |
1966 | * This should tell if fe_len is exactly power of 2 | |
1967 | */ | |
1968 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1969 | ac->ac_2order = i - 1; | |
1970 | } | |
1971 | ||
4ba74d00 TT |
1972 | /* if stream allocation is enabled, use global goal */ |
1973 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
1974 | /* TBD: may be hot point */ |
1975 | spin_lock(&sbi->s_md_lock); | |
1976 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
1977 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
1978 | spin_unlock(&sbi->s_md_lock); | |
1979 | } | |
4ba74d00 | 1980 | |
c9de560d AT |
1981 | /* Let's just scan groups to find more-less suitable blocks */ |
1982 | cr = ac->ac_2order ? 0 : 1; | |
1983 | /* | |
1984 | * cr == 0 try to get exact allocation, | |
1985 | * cr == 3 try to get anything | |
1986 | */ | |
1987 | repeat: | |
1988 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
1989 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
1990 | /* |
1991 | * searching for the right group start | |
1992 | * from the goal value specified | |
1993 | */ | |
1994 | group = ac->ac_g_ex.fe_group; | |
1995 | ||
8df9675f | 1996 | for (i = 0; i < ngroups; group++, i++) { |
8df9675f | 1997 | if (group == ngroups) |
c9de560d AT |
1998 | group = 0; |
1999 | ||
8a57d9d6 CW |
2000 | /* This now checks without needing the buddy page */ |
2001 | if (!ext4_mb_good_group(ac, group, cr)) | |
c9de560d AT |
2002 | continue; |
2003 | ||
c9de560d AT |
2004 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2005 | if (err) | |
2006 | goto out; | |
2007 | ||
2008 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2009 | |
2010 | /* | |
2011 | * We need to check again after locking the | |
2012 | * block group | |
2013 | */ | |
c9de560d | 2014 | if (!ext4_mb_good_group(ac, group, cr)) { |
c9de560d | 2015 | ext4_unlock_group(sb, group); |
e39e07fd | 2016 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2017 | continue; |
2018 | } | |
2019 | ||
2020 | ac->ac_groups_scanned++; | |
40ae3487 | 2021 | if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2) |
c9de560d | 2022 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2023 | else if (cr == 1 && sbi->s_stripe && |
2024 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2025 | ext4_mb_scan_aligned(ac, &e4b); |
2026 | else | |
2027 | ext4_mb_complex_scan_group(ac, &e4b); | |
2028 | ||
2029 | ext4_unlock_group(sb, group); | |
e39e07fd | 2030 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2031 | |
2032 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2033 | break; | |
2034 | } | |
2035 | } | |
2036 | ||
2037 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2038 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2039 | /* | |
2040 | * We've been searching too long. Let's try to allocate | |
2041 | * the best chunk we've found so far | |
2042 | */ | |
2043 | ||
2044 | ext4_mb_try_best_found(ac, &e4b); | |
2045 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2046 | /* | |
2047 | * Someone more lucky has already allocated it. | |
2048 | * The only thing we can do is just take first | |
2049 | * found block(s) | |
2050 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2051 | */ | |
2052 | ac->ac_b_ex.fe_group = 0; | |
2053 | ac->ac_b_ex.fe_start = 0; | |
2054 | ac->ac_b_ex.fe_len = 0; | |
2055 | ac->ac_status = AC_STATUS_CONTINUE; | |
2056 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2057 | cr = 3; | |
2058 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2059 | goto repeat; | |
2060 | } | |
2061 | } | |
2062 | out: | |
2063 | return err; | |
2064 | } | |
2065 | ||
c9de560d AT |
2066 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2067 | { | |
2068 | struct super_block *sb = seq->private; | |
c9de560d AT |
2069 | ext4_group_t group; |
2070 | ||
8df9675f | 2071 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2072 | return NULL; |
c9de560d | 2073 | group = *pos + 1; |
a9df9a49 | 2074 | return (void *) ((unsigned long) group); |
c9de560d AT |
2075 | } |
2076 | ||
2077 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2078 | { | |
2079 | struct super_block *sb = seq->private; | |
c9de560d AT |
2080 | ext4_group_t group; |
2081 | ||
2082 | ++*pos; | |
8df9675f | 2083 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2084 | return NULL; |
2085 | group = *pos + 1; | |
a9df9a49 | 2086 | return (void *) ((unsigned long) group); |
c9de560d AT |
2087 | } |
2088 | ||
2089 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2090 | { | |
2091 | struct super_block *sb = seq->private; | |
a9df9a49 | 2092 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2093 | int i; |
1c8457ca | 2094 | int err, buddy_loaded = 0; |
c9de560d | 2095 | struct ext4_buddy e4b; |
1c8457ca | 2096 | struct ext4_group_info *grinfo; |
c9de560d AT |
2097 | struct sg { |
2098 | struct ext4_group_info info; | |
a36b4498 | 2099 | ext4_grpblk_t counters[16]; |
c9de560d AT |
2100 | } sg; |
2101 | ||
2102 | group--; | |
2103 | if (group == 0) | |
2104 | seq_printf(seq, "#%-5s: %-5s %-5s %-5s " | |
2105 | "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " | |
2106 | "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", | |
2107 | "group", "free", "frags", "first", | |
2108 | "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", | |
2109 | "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); | |
2110 | ||
2111 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2112 | sizeof(struct ext4_group_info); | |
1c8457ca AK |
2113 | grinfo = ext4_get_group_info(sb, group); |
2114 | /* Load the group info in memory only if not already loaded. */ | |
2115 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2116 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2117 | if (err) { | |
2118 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2119 | return 0; | |
2120 | } | |
2121 | buddy_loaded = 1; | |
c9de560d | 2122 | } |
1c8457ca | 2123 | |
c9de560d | 2124 | memcpy(&sg, ext4_get_group_info(sb, group), i); |
1c8457ca AK |
2125 | |
2126 | if (buddy_loaded) | |
2127 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2128 | |
a9df9a49 | 2129 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2130 | sg.info.bb_fragments, sg.info.bb_first_free); |
2131 | for (i = 0; i <= 13; i++) | |
2132 | seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? | |
2133 | sg.info.bb_counters[i] : 0); | |
2134 | seq_printf(seq, " ]\n"); | |
2135 | ||
2136 | return 0; | |
2137 | } | |
2138 | ||
2139 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2140 | { | |
2141 | } | |
2142 | ||
7f1346a9 | 2143 | static const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2144 | .start = ext4_mb_seq_groups_start, |
2145 | .next = ext4_mb_seq_groups_next, | |
2146 | .stop = ext4_mb_seq_groups_stop, | |
2147 | .show = ext4_mb_seq_groups_show, | |
2148 | }; | |
2149 | ||
2150 | static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) | |
2151 | { | |
2152 | struct super_block *sb = PDE(inode)->data; | |
2153 | int rc; | |
2154 | ||
2155 | rc = seq_open(file, &ext4_mb_seq_groups_ops); | |
2156 | if (rc == 0) { | |
a271fe85 | 2157 | struct seq_file *m = file->private_data; |
c9de560d AT |
2158 | m->private = sb; |
2159 | } | |
2160 | return rc; | |
2161 | ||
2162 | } | |
2163 | ||
7f1346a9 | 2164 | static const struct file_operations ext4_mb_seq_groups_fops = { |
c9de560d AT |
2165 | .owner = THIS_MODULE, |
2166 | .open = ext4_mb_seq_groups_open, | |
2167 | .read = seq_read, | |
2168 | .llseek = seq_lseek, | |
2169 | .release = seq_release, | |
2170 | }; | |
2171 | ||
fb1813f4 CW |
2172 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
2173 | { | |
2174 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2175 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
2176 | ||
2177 | BUG_ON(!cachep); | |
2178 | return cachep; | |
2179 | } | |
5f21b0e6 | 2180 | |
28623c2f TT |
2181 | /* |
2182 | * Allocate the top-level s_group_info array for the specified number | |
2183 | * of groups | |
2184 | */ | |
2185 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
2186 | { | |
2187 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2188 | unsigned size; | |
2189 | struct ext4_group_info ***new_groupinfo; | |
2190 | ||
2191 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
2192 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2193 | if (size <= sbi->s_group_info_size) | |
2194 | return 0; | |
2195 | ||
2196 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
2197 | new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL); | |
2198 | if (!new_groupinfo) { | |
2199 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
2200 | return -ENOMEM; | |
2201 | } | |
2202 | if (sbi->s_group_info) { | |
2203 | memcpy(new_groupinfo, sbi->s_group_info, | |
2204 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); | |
2205 | ext4_kvfree(sbi->s_group_info); | |
2206 | } | |
2207 | sbi->s_group_info = new_groupinfo; | |
2208 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); | |
2209 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", | |
2210 | sbi->s_group_info_size); | |
2211 | return 0; | |
2212 | } | |
2213 | ||
5f21b0e6 | 2214 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 2215 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2216 | struct ext4_group_desc *desc) |
2217 | { | |
fb1813f4 | 2218 | int i; |
5f21b0e6 FB |
2219 | int metalen = 0; |
2220 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2221 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 2222 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
2223 | |
2224 | /* | |
2225 | * First check if this group is the first of a reserved block. | |
2226 | * If it's true, we have to allocate a new table of pointers | |
2227 | * to ext4_group_info structures | |
2228 | */ | |
2229 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2230 | metalen = sizeof(*meta_group_info) << | |
2231 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2232 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2233 | if (meta_group_info == NULL) { | |
7f6a11e7 | 2234 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 2235 | "for a buddy group"); |
5f21b0e6 FB |
2236 | goto exit_meta_group_info; |
2237 | } | |
2238 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = | |
2239 | meta_group_info; | |
2240 | } | |
2241 | ||
5f21b0e6 FB |
2242 | meta_group_info = |
2243 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; | |
2244 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); | |
2245 | ||
85556c9a | 2246 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL); |
5f21b0e6 | 2247 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 2248 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
2249 | goto exit_group_info; |
2250 | } | |
2251 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2252 | &(meta_group_info[i]->bb_state)); | |
2253 | ||
2254 | /* | |
2255 | * initialize bb_free to be able to skip | |
2256 | * empty groups without initialization | |
2257 | */ | |
2258 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2259 | meta_group_info[i]->bb_free = | |
cff1dfd7 | 2260 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
2261 | } else { |
2262 | meta_group_info[i]->bb_free = | |
021b65bb | 2263 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
2264 | } |
2265 | ||
2266 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2267 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 2268 | meta_group_info[i]->bb_free_root = RB_ROOT; |
8a57d9d6 | 2269 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
5f21b0e6 FB |
2270 | |
2271 | #ifdef DOUBLE_CHECK | |
2272 | { | |
2273 | struct buffer_head *bh; | |
2274 | meta_group_info[i]->bb_bitmap = | |
2275 | kmalloc(sb->s_blocksize, GFP_KERNEL); | |
2276 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); | |
2277 | bh = ext4_read_block_bitmap(sb, group); | |
2278 | BUG_ON(bh == NULL); | |
2279 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, | |
2280 | sb->s_blocksize); | |
2281 | put_bh(bh); | |
2282 | } | |
2283 | #endif | |
2284 | ||
2285 | return 0; | |
2286 | ||
2287 | exit_group_info: | |
2288 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 2289 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
5f21b0e6 | 2290 | kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); |
caaf7a29 TM |
2291 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL; |
2292 | } | |
5f21b0e6 FB |
2293 | exit_meta_group_info: |
2294 | return -ENOMEM; | |
2295 | } /* ext4_mb_add_groupinfo */ | |
2296 | ||
c9de560d AT |
2297 | static int ext4_mb_init_backend(struct super_block *sb) |
2298 | { | |
8df9675f | 2299 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2300 | ext4_group_t i; |
c9de560d | 2301 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 2302 | int err; |
5f21b0e6 | 2303 | struct ext4_group_desc *desc; |
fb1813f4 | 2304 | struct kmem_cache *cachep; |
5f21b0e6 | 2305 | |
28623c2f TT |
2306 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
2307 | if (err) | |
2308 | return err; | |
c9de560d | 2309 | |
c9de560d AT |
2310 | sbi->s_buddy_cache = new_inode(sb); |
2311 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 2312 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
2313 | goto err_freesgi; |
2314 | } | |
48e6061b YJ |
2315 | /* To avoid potentially colliding with an valid on-disk inode number, |
2316 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
2317 | * not in the inode hash, so it should never be found by iget(), but | |
2318 | * this will avoid confusion if it ever shows up during debugging. */ | |
2319 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 2320 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 2321 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2322 | desc = ext4_get_group_desc(sb, i, NULL); |
2323 | if (desc == NULL) { | |
9d8b9ec4 | 2324 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
2325 | goto err_freebuddy; |
2326 | } | |
5f21b0e6 FB |
2327 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2328 | goto err_freebuddy; | |
c9de560d AT |
2329 | } |
2330 | ||
2331 | return 0; | |
2332 | ||
2333 | err_freebuddy: | |
fb1813f4 | 2334 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 2335 | while (i-- > 0) |
fb1813f4 | 2336 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
28623c2f | 2337 | i = sbi->s_group_info_size; |
f1fa3342 | 2338 | while (i-- > 0) |
c9de560d AT |
2339 | kfree(sbi->s_group_info[i]); |
2340 | iput(sbi->s_buddy_cache); | |
2341 | err_freesgi: | |
f18a5f21 | 2342 | ext4_kvfree(sbi->s_group_info); |
c9de560d AT |
2343 | return -ENOMEM; |
2344 | } | |
2345 | ||
2892c15d ES |
2346 | static void ext4_groupinfo_destroy_slabs(void) |
2347 | { | |
2348 | int i; | |
2349 | ||
2350 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
2351 | if (ext4_groupinfo_caches[i]) | |
2352 | kmem_cache_destroy(ext4_groupinfo_caches[i]); | |
2353 | ext4_groupinfo_caches[i] = NULL; | |
2354 | } | |
2355 | } | |
2356 | ||
2357 | static int ext4_groupinfo_create_slab(size_t size) | |
2358 | { | |
2359 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
2360 | int slab_size; | |
2361 | int blocksize_bits = order_base_2(size); | |
2362 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2363 | struct kmem_cache *cachep; | |
2364 | ||
2365 | if (cache_index >= NR_GRPINFO_CACHES) | |
2366 | return -EINVAL; | |
2367 | ||
2368 | if (unlikely(cache_index < 0)) | |
2369 | cache_index = 0; | |
2370 | ||
2371 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
2372 | if (ext4_groupinfo_caches[cache_index]) { | |
2373 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
2374 | return 0; /* Already created */ | |
2375 | } | |
2376 | ||
2377 | slab_size = offsetof(struct ext4_group_info, | |
2378 | bb_counters[blocksize_bits + 2]); | |
2379 | ||
2380 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
2381 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
2382 | NULL); | |
2383 | ||
823ba01f TM |
2384 | ext4_groupinfo_caches[cache_index] = cachep; |
2385 | ||
2892c15d ES |
2386 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
2387 | if (!cachep) { | |
9d8b9ec4 TT |
2388 | printk(KERN_EMERG |
2389 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
2390 | return -ENOMEM; |
2391 | } | |
2392 | ||
2892c15d ES |
2393 | return 0; |
2394 | } | |
2395 | ||
9d99012f | 2396 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
2397 | { |
2398 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2399 | unsigned i, j; |
c9de560d AT |
2400 | unsigned offset; |
2401 | unsigned max; | |
74767c5a | 2402 | int ret; |
c9de560d | 2403 | |
1927805e | 2404 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
2405 | |
2406 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2407 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
2408 | ret = -ENOMEM; |
2409 | goto out; | |
c9de560d | 2410 | } |
ff7ef329 | 2411 | |
1927805e | 2412 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
2413 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2414 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
2415 | ret = -ENOMEM; |
2416 | goto out; | |
2417 | } | |
2418 | ||
2892c15d ES |
2419 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
2420 | if (ret < 0) | |
2421 | goto out; | |
c9de560d AT |
2422 | |
2423 | /* order 0 is regular bitmap */ | |
2424 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2425 | sbi->s_mb_offsets[0] = 0; | |
2426 | ||
2427 | i = 1; | |
2428 | offset = 0; | |
2429 | max = sb->s_blocksize << 2; | |
2430 | do { | |
2431 | sbi->s_mb_offsets[i] = offset; | |
2432 | sbi->s_mb_maxs[i] = max; | |
2433 | offset += 1 << (sb->s_blocksize_bits - i); | |
2434 | max = max >> 1; | |
2435 | i++; | |
2436 | } while (i <= sb->s_blocksize_bits + 1); | |
2437 | ||
c9de560d | 2438 | spin_lock_init(&sbi->s_md_lock); |
c9de560d AT |
2439 | spin_lock_init(&sbi->s_bal_lock); |
2440 | ||
2441 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2442 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2443 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2444 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2445 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27baebb8 TT |
2446 | /* |
2447 | * The default group preallocation is 512, which for 4k block | |
2448 | * sizes translates to 2 megabytes. However for bigalloc file | |
2449 | * systems, this is probably too big (i.e, if the cluster size | |
2450 | * is 1 megabyte, then group preallocation size becomes half a | |
2451 | * gigabyte!). As a default, we will keep a two megabyte | |
2452 | * group pralloc size for cluster sizes up to 64k, and after | |
2453 | * that, we will force a minimum group preallocation size of | |
2454 | * 32 clusters. This translates to 8 megs when the cluster | |
2455 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
2456 | * which seems reasonable as a default. | |
2457 | */ | |
2458 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
2459 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
2460 | /* |
2461 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
2462 | * to the lowest multiple of s_stripe which is bigger than | |
2463 | * the s_mb_group_prealloc as determined above. We want | |
2464 | * the preallocation size to be an exact multiple of the | |
2465 | * RAID stripe size so that preallocations don't fragment | |
2466 | * the stripes. | |
2467 | */ | |
2468 | if (sbi->s_stripe > 1) { | |
2469 | sbi->s_mb_group_prealloc = roundup( | |
2470 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
2471 | } | |
c9de560d | 2472 | |
730c213c | 2473 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2474 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 2475 | ret = -ENOMEM; |
7aa0baea | 2476 | goto out_free_groupinfo_slab; |
c9de560d | 2477 | } |
730c213c | 2478 | for_each_possible_cpu(i) { |
c9de560d | 2479 | struct ext4_locality_group *lg; |
730c213c | 2480 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2481 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2482 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2483 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2484 | spin_lock_init(&lg->lg_prealloc_lock); |
2485 | } | |
2486 | ||
79a77c5a YJ |
2487 | /* init file for buddy data */ |
2488 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
2489 | if (ret != 0) |
2490 | goto out_free_locality_groups; | |
79a77c5a | 2491 | |
296c355c TT |
2492 | if (sbi->s_proc) |
2493 | proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, | |
2494 | &ext4_mb_seq_groups_fops, sb); | |
c9de560d | 2495 | |
7aa0baea TM |
2496 | return 0; |
2497 | ||
2498 | out_free_locality_groups: | |
2499 | free_percpu(sbi->s_locality_groups); | |
2500 | sbi->s_locality_groups = NULL; | |
2501 | out_free_groupinfo_slab: | |
2502 | ext4_groupinfo_destroy_slabs(); | |
fb1813f4 | 2503 | out: |
7aa0baea TM |
2504 | kfree(sbi->s_mb_offsets); |
2505 | sbi->s_mb_offsets = NULL; | |
2506 | kfree(sbi->s_mb_maxs); | |
2507 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 2508 | return ret; |
c9de560d AT |
2509 | } |
2510 | ||
955ce5f5 | 2511 | /* need to called with the ext4 group lock held */ |
c9de560d AT |
2512 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
2513 | { | |
2514 | struct ext4_prealloc_space *pa; | |
2515 | struct list_head *cur, *tmp; | |
2516 | int count = 0; | |
2517 | ||
2518 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2519 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2520 | list_del(&pa->pa_group_list); | |
2521 | count++; | |
688f05a0 | 2522 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2523 | } |
2524 | if (count) | |
6ba495e9 | 2525 | mb_debug(1, "mballoc: %u PAs left\n", count); |
c9de560d AT |
2526 | |
2527 | } | |
2528 | ||
2529 | int ext4_mb_release(struct super_block *sb) | |
2530 | { | |
8df9675f | 2531 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2532 | ext4_group_t i; |
2533 | int num_meta_group_infos; | |
2534 | struct ext4_group_info *grinfo; | |
2535 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
fb1813f4 | 2536 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
c9de560d | 2537 | |
95599968 SQ |
2538 | if (sbi->s_proc) |
2539 | remove_proc_entry("mb_groups", sbi->s_proc); | |
2540 | ||
c9de560d | 2541 | if (sbi->s_group_info) { |
8df9675f | 2542 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2543 | grinfo = ext4_get_group_info(sb, i); |
2544 | #ifdef DOUBLE_CHECK | |
2545 | kfree(grinfo->bb_bitmap); | |
2546 | #endif | |
2547 | ext4_lock_group(sb, i); | |
2548 | ext4_mb_cleanup_pa(grinfo); | |
2549 | ext4_unlock_group(sb, i); | |
fb1813f4 | 2550 | kmem_cache_free(cachep, grinfo); |
c9de560d | 2551 | } |
8df9675f | 2552 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2553 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2554 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2555 | for (i = 0; i < num_meta_group_infos; i++) | |
2556 | kfree(sbi->s_group_info[i]); | |
f18a5f21 | 2557 | ext4_kvfree(sbi->s_group_info); |
c9de560d AT |
2558 | } |
2559 | kfree(sbi->s_mb_offsets); | |
2560 | kfree(sbi->s_mb_maxs); | |
2561 | if (sbi->s_buddy_cache) | |
2562 | iput(sbi->s_buddy_cache); | |
2563 | if (sbi->s_mb_stats) { | |
9d8b9ec4 TT |
2564 | ext4_msg(sb, KERN_INFO, |
2565 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
2566 | atomic_read(&sbi->s_bal_allocated), |
2567 | atomic_read(&sbi->s_bal_reqs), | |
2568 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 TT |
2569 | ext4_msg(sb, KERN_INFO, |
2570 | "mballoc: %u extents scanned, %u goal hits, " | |
2571 | "%u 2^N hits, %u breaks, %u lost", | |
c9de560d AT |
2572 | atomic_read(&sbi->s_bal_ex_scanned), |
2573 | atomic_read(&sbi->s_bal_goals), | |
2574 | atomic_read(&sbi->s_bal_2orders), | |
2575 | atomic_read(&sbi->s_bal_breaks), | |
2576 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 TT |
2577 | ext4_msg(sb, KERN_INFO, |
2578 | "mballoc: %lu generated and it took %Lu", | |
ced156e4 | 2579 | sbi->s_mb_buddies_generated, |
c9de560d | 2580 | sbi->s_mb_generation_time); |
9d8b9ec4 TT |
2581 | ext4_msg(sb, KERN_INFO, |
2582 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
2583 | atomic_read(&sbi->s_mb_preallocated), |
2584 | atomic_read(&sbi->s_mb_discarded)); | |
2585 | } | |
2586 | ||
730c213c | 2587 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
2588 | |
2589 | return 0; | |
2590 | } | |
2591 | ||
77ca6cdf | 2592 | static inline int ext4_issue_discard(struct super_block *sb, |
84130193 | 2593 | ext4_group_t block_group, ext4_grpblk_t cluster, int count) |
5c521830 | 2594 | { |
5c521830 JZ |
2595 | ext4_fsblk_t discard_block; |
2596 | ||
84130193 TT |
2597 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
2598 | ext4_group_first_block_no(sb, block_group)); | |
2599 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
2600 | trace_ext4_discard_blocks(sb, |
2601 | (unsigned long long) discard_block, count); | |
93259636 | 2602 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); |
5c521830 JZ |
2603 | } |
2604 | ||
3e624fc7 TT |
2605 | /* |
2606 | * This function is called by the jbd2 layer once the commit has finished, | |
2607 | * so we know we can free the blocks that were released with that commit. | |
2608 | */ | |
18aadd47 BJ |
2609 | static void ext4_free_data_callback(struct super_block *sb, |
2610 | struct ext4_journal_cb_entry *jce, | |
2611 | int rc) | |
c9de560d | 2612 | { |
18aadd47 | 2613 | struct ext4_free_data *entry = (struct ext4_free_data *)jce; |
c9de560d | 2614 | struct ext4_buddy e4b; |
c894058d | 2615 | struct ext4_group_info *db; |
d9f34504 | 2616 | int err, count = 0, count2 = 0; |
c9de560d | 2617 | |
18aadd47 BJ |
2618 | mb_debug(1, "gonna free %u blocks in group %u (0x%p):", |
2619 | entry->efd_count, entry->efd_group, entry); | |
c9de560d | 2620 | |
d71c1ae2 LC |
2621 | if (test_opt(sb, DISCARD)) { |
2622 | err = ext4_issue_discard(sb, entry->efd_group, | |
2623 | entry->efd_start_cluster, | |
2624 | entry->efd_count); | |
2625 | if (err && err != -EOPNOTSUPP) | |
2626 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
2627 | " group:%d block:%d count:%d failed" | |
2628 | " with %d", entry->efd_group, | |
2629 | entry->efd_start_cluster, | |
2630 | entry->efd_count, err); | |
2631 | } | |
c9de560d | 2632 | |
18aadd47 BJ |
2633 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
2634 | /* we expect to find existing buddy because it's pinned */ | |
2635 | BUG_ON(err != 0); | |
b90f6870 | 2636 | |
c9de560d | 2637 | |
18aadd47 BJ |
2638 | db = e4b.bd_info; |
2639 | /* there are blocks to put in buddy to make them really free */ | |
2640 | count += entry->efd_count; | |
2641 | count2++; | |
2642 | ext4_lock_group(sb, entry->efd_group); | |
2643 | /* Take it out of per group rb tree */ | |
2644 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
2645 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 2646 | |
18aadd47 BJ |
2647 | /* |
2648 | * Clear the trimmed flag for the group so that the next | |
2649 | * ext4_trim_fs can trim it. | |
2650 | * If the volume is mounted with -o discard, online discard | |
2651 | * is supported and the free blocks will be trimmed online. | |
2652 | */ | |
2653 | if (!test_opt(sb, DISCARD)) | |
2654 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 2655 | |
18aadd47 BJ |
2656 | if (!db->bb_free_root.rb_node) { |
2657 | /* No more items in the per group rb tree | |
2658 | * balance refcounts from ext4_mb_free_metadata() | |
2659 | */ | |
2660 | page_cache_release(e4b.bd_buddy_page); | |
2661 | page_cache_release(e4b.bd_bitmap_page); | |
3e624fc7 | 2662 | } |
18aadd47 BJ |
2663 | ext4_unlock_group(sb, entry->efd_group); |
2664 | kmem_cache_free(ext4_free_data_cachep, entry); | |
2665 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2666 | |
6ba495e9 | 2667 | mb_debug(1, "freed %u blocks in %u structures\n", count, count2); |
c9de560d AT |
2668 | } |
2669 | ||
5dabfc78 | 2670 | int __init ext4_init_mballoc(void) |
c9de560d | 2671 | { |
16828088 TT |
2672 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
2673 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d AT |
2674 | if (ext4_pspace_cachep == NULL) |
2675 | return -ENOMEM; | |
2676 | ||
16828088 TT |
2677 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
2678 | SLAB_RECLAIM_ACCOUNT); | |
256bdb49 ES |
2679 | if (ext4_ac_cachep == NULL) { |
2680 | kmem_cache_destroy(ext4_pspace_cachep); | |
2681 | return -ENOMEM; | |
2682 | } | |
c894058d | 2683 | |
18aadd47 BJ |
2684 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
2685 | SLAB_RECLAIM_ACCOUNT); | |
2686 | if (ext4_free_data_cachep == NULL) { | |
c894058d AK |
2687 | kmem_cache_destroy(ext4_pspace_cachep); |
2688 | kmem_cache_destroy(ext4_ac_cachep); | |
2689 | return -ENOMEM; | |
2690 | } | |
c9de560d AT |
2691 | return 0; |
2692 | } | |
2693 | ||
5dabfc78 | 2694 | void ext4_exit_mballoc(void) |
c9de560d | 2695 | { |
60e6679e | 2696 | /* |
3e03f9ca JDB |
2697 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
2698 | * before destroying the slab cache. | |
2699 | */ | |
2700 | rcu_barrier(); | |
c9de560d | 2701 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 2702 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 2703 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 2704 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
2705 | } |
2706 | ||
2707 | ||
2708 | /* | |
73b2c716 | 2709 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
2710 | * Returns 0 if success or error code |
2711 | */ | |
4ddfef7b ES |
2712 | static noinline_for_stack int |
2713 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 2714 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
2715 | { |
2716 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
2717 | struct ext4_group_desc *gdp; |
2718 | struct buffer_head *gdp_bh; | |
2719 | struct ext4_sb_info *sbi; | |
2720 | struct super_block *sb; | |
2721 | ext4_fsblk_t block; | |
519deca0 | 2722 | int err, len; |
c9de560d AT |
2723 | |
2724 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2725 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2726 | ||
2727 | sb = ac->ac_sb; | |
2728 | sbi = EXT4_SB(sb); | |
c9de560d AT |
2729 | |
2730 | err = -EIO; | |
574ca174 | 2731 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
c9de560d AT |
2732 | if (!bitmap_bh) |
2733 | goto out_err; | |
2734 | ||
2735 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2736 | if (err) | |
2737 | goto out_err; | |
2738 | ||
2739 | err = -EIO; | |
2740 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
2741 | if (!gdp) | |
2742 | goto out_err; | |
2743 | ||
a9df9a49 | 2744 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 2745 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 2746 | |
c9de560d AT |
2747 | err = ext4_journal_get_write_access(handle, gdp_bh); |
2748 | if (err) | |
2749 | goto out_err; | |
2750 | ||
bda00de7 | 2751 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 2752 | |
53accfa9 | 2753 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
6fd058f7 | 2754 | if (!ext4_data_block_valid(sbi, block, len)) { |
12062ddd | 2755 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 2756 | "fs metadata", block, block+len); |
519deca0 AK |
2757 | /* File system mounted not to panic on error |
2758 | * Fix the bitmap and repeat the block allocation | |
2759 | * We leak some of the blocks here. | |
2760 | */ | |
955ce5f5 | 2761 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
c3e94d1d YY |
2762 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
2763 | ac->ac_b_ex.fe_len); | |
955ce5f5 | 2764 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 2765 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 AK |
2766 | if (!err) |
2767 | err = -EAGAIN; | |
2768 | goto out_err; | |
c9de560d | 2769 | } |
955ce5f5 AK |
2770 | |
2771 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
2772 | #ifdef AGGRESSIVE_CHECK |
2773 | { | |
2774 | int i; | |
2775 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
2776 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
2777 | bitmap_bh->b_data)); | |
2778 | } | |
2779 | } | |
2780 | #endif | |
c3e94d1d YY |
2781 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
2782 | ac->ac_b_ex.fe_len); | |
c9de560d AT |
2783 | if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
2784 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
021b65bb | 2785 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 2786 | ext4_free_clusters_after_init(sb, |
021b65bb | 2787 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 2788 | } |
021b65bb TT |
2789 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
2790 | ext4_free_group_clusters_set(sb, gdp, len); | |
79f1ba49 | 2791 | ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); |
feb0ab32 | 2792 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
2793 | |
2794 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 2795 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 2796 | /* |
6bc6e63f | 2797 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 2798 | */ |
6bc6e63f AK |
2799 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
2800 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
2801 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
2802 | reserv_clstrs); | |
c9de560d | 2803 | |
772cb7c8 JS |
2804 | if (sbi->s_log_groups_per_flex) { |
2805 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
2806 | ac->ac_b_ex.fe_group); | |
9f24e420 | 2807 | atomic_sub(ac->ac_b_ex.fe_len, |
24aaa8ef | 2808 | &sbi->s_flex_groups[flex_group].free_clusters); |
772cb7c8 JS |
2809 | } |
2810 | ||
0390131b | 2811 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
2812 | if (err) |
2813 | goto out_err; | |
0390131b | 2814 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
2815 | |
2816 | out_err: | |
42a10add | 2817 | brelse(bitmap_bh); |
c9de560d AT |
2818 | return err; |
2819 | } | |
2820 | ||
2821 | /* | |
2822 | * here we normalize request for locality group | |
d7a1fee1 DE |
2823 | * Group request are normalized to s_mb_group_prealloc, which goes to |
2824 | * s_strip if we set the same via mount option. | |
2825 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 2826 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
2827 | * |
2828 | * XXX: should we try to preallocate more than the group has now? | |
2829 | */ | |
2830 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
2831 | { | |
2832 | struct super_block *sb = ac->ac_sb; | |
2833 | struct ext4_locality_group *lg = ac->ac_lg; | |
2834 | ||
2835 | BUG_ON(lg == NULL); | |
d7a1fee1 | 2836 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
6ba495e9 | 2837 | mb_debug(1, "#%u: goal %u blocks for locality group\n", |
c9de560d AT |
2838 | current->pid, ac->ac_g_ex.fe_len); |
2839 | } | |
2840 | ||
2841 | /* | |
2842 | * Normalization means making request better in terms of | |
2843 | * size and alignment | |
2844 | */ | |
4ddfef7b ES |
2845 | static noinline_for_stack void |
2846 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
2847 | struct ext4_allocation_request *ar) |
2848 | { | |
53accfa9 | 2849 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
2850 | int bsbits, max; |
2851 | ext4_lblk_t end; | |
1592d2c5 CW |
2852 | loff_t size, start_off; |
2853 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 2854 | ext4_lblk_t start; |
c9de560d | 2855 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 2856 | struct ext4_prealloc_space *pa; |
c9de560d AT |
2857 | |
2858 | /* do normalize only data requests, metadata requests | |
2859 | do not need preallocation */ | |
2860 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
2861 | return; | |
2862 | ||
2863 | /* sometime caller may want exact blocks */ | |
2864 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2865 | return; | |
2866 | ||
2867 | /* caller may indicate that preallocation isn't | |
2868 | * required (it's a tail, for example) */ | |
2869 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
2870 | return; | |
2871 | ||
2872 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
2873 | ext4_mb_normalize_group_request(ac); | |
2874 | return ; | |
2875 | } | |
2876 | ||
2877 | bsbits = ac->ac_sb->s_blocksize_bits; | |
2878 | ||
2879 | /* first, let's learn actual file size | |
2880 | * given current request is allocated */ | |
53accfa9 | 2881 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
2882 | size = size << bsbits; |
2883 | if (size < i_size_read(ac->ac_inode)) | |
2884 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 2885 | orig_size = size; |
c9de560d | 2886 | |
1930479c VC |
2887 | /* max size of free chunks */ |
2888 | max = 2 << bsbits; | |
c9de560d | 2889 | |
1930479c VC |
2890 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
2891 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
2892 | |
2893 | /* first, try to predict filesize */ | |
2894 | /* XXX: should this table be tunable? */ | |
2895 | start_off = 0; | |
2896 | if (size <= 16 * 1024) { | |
2897 | size = 16 * 1024; | |
2898 | } else if (size <= 32 * 1024) { | |
2899 | size = 32 * 1024; | |
2900 | } else if (size <= 64 * 1024) { | |
2901 | size = 64 * 1024; | |
2902 | } else if (size <= 128 * 1024) { | |
2903 | size = 128 * 1024; | |
2904 | } else if (size <= 256 * 1024) { | |
2905 | size = 256 * 1024; | |
2906 | } else if (size <= 512 * 1024) { | |
2907 | size = 512 * 1024; | |
2908 | } else if (size <= 1024 * 1024) { | |
2909 | size = 1024 * 1024; | |
1930479c | 2910 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 2911 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
2912 | (21 - bsbits)) << 21; |
2913 | size = 2 * 1024 * 1024; | |
2914 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
2915 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
2916 | (22 - bsbits)) << 22; | |
2917 | size = 4 * 1024 * 1024; | |
2918 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 2919 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
2920 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
2921 | (23 - bsbits)) << 23; | |
2922 | size = 8 * 1024 * 1024; | |
2923 | } else { | |
2924 | start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; | |
2925 | size = ac->ac_o_ex.fe_len << bsbits; | |
2926 | } | |
5a0790c2 AK |
2927 | size = size >> bsbits; |
2928 | start = start_off >> bsbits; | |
c9de560d AT |
2929 | |
2930 | /* don't cover already allocated blocks in selected range */ | |
2931 | if (ar->pleft && start <= ar->lleft) { | |
2932 | size -= ar->lleft + 1 - start; | |
2933 | start = ar->lleft + 1; | |
2934 | } | |
2935 | if (ar->pright && start + size - 1 >= ar->lright) | |
2936 | size -= start + size - ar->lright; | |
2937 | ||
2938 | end = start + size; | |
2939 | ||
2940 | /* check we don't cross already preallocated blocks */ | |
2941 | rcu_read_lock(); | |
9a0762c5 | 2942 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 2943 | ext4_lblk_t pa_end; |
c9de560d | 2944 | |
c9de560d AT |
2945 | if (pa->pa_deleted) |
2946 | continue; | |
2947 | spin_lock(&pa->pa_lock); | |
2948 | if (pa->pa_deleted) { | |
2949 | spin_unlock(&pa->pa_lock); | |
2950 | continue; | |
2951 | } | |
2952 | ||
53accfa9 TT |
2953 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
2954 | pa->pa_len); | |
c9de560d AT |
2955 | |
2956 | /* PA must not overlap original request */ | |
2957 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
2958 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
2959 | ||
38877f4e ES |
2960 | /* skip PAs this normalized request doesn't overlap with */ |
2961 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
2962 | spin_unlock(&pa->pa_lock); |
2963 | continue; | |
2964 | } | |
2965 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
2966 | ||
38877f4e | 2967 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
2968 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
2969 | BUG_ON(pa_end < start); | |
2970 | start = pa_end; | |
38877f4e | 2971 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
2972 | BUG_ON(pa->pa_lstart > end); |
2973 | end = pa->pa_lstart; | |
2974 | } | |
2975 | spin_unlock(&pa->pa_lock); | |
2976 | } | |
2977 | rcu_read_unlock(); | |
2978 | size = end - start; | |
2979 | ||
2980 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
2981 | rcu_read_lock(); | |
9a0762c5 | 2982 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 2983 | ext4_lblk_t pa_end; |
53accfa9 | 2984 | |
c9de560d AT |
2985 | spin_lock(&pa->pa_lock); |
2986 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
2987 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
2988 | pa->pa_len); | |
c9de560d AT |
2989 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
2990 | } | |
2991 | spin_unlock(&pa->pa_lock); | |
2992 | } | |
2993 | rcu_read_unlock(); | |
2994 | ||
2995 | if (start + size <= ac->ac_o_ex.fe_logical && | |
2996 | start > ac->ac_o_ex.fe_logical) { | |
9d8b9ec4 TT |
2997 | ext4_msg(ac->ac_sb, KERN_ERR, |
2998 | "start %lu, size %lu, fe_logical %lu", | |
2999 | (unsigned long) start, (unsigned long) size, | |
3000 | (unsigned long) ac->ac_o_ex.fe_logical); | |
c9de560d AT |
3001 | } |
3002 | BUG_ON(start + size <= ac->ac_o_ex.fe_logical && | |
3003 | start > ac->ac_o_ex.fe_logical); | |
7137d7a4 | 3004 | BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
3005 | |
3006 | /* now prepare goal request */ | |
3007 | ||
3008 | /* XXX: is it better to align blocks WRT to logical | |
3009 | * placement or satisfy big request as is */ | |
3010 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 3011 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
3012 | |
3013 | /* define goal start in order to merge */ | |
3014 | if (ar->pright && (ar->lright == (start + size))) { | |
3015 | /* merge to the right */ | |
3016 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3017 | &ac->ac_f_ex.fe_group, | |
3018 | &ac->ac_f_ex.fe_start); | |
3019 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3020 | } | |
3021 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3022 | /* merge to the left */ | |
3023 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3024 | &ac->ac_f_ex.fe_group, | |
3025 | &ac->ac_f_ex.fe_start); | |
3026 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3027 | } | |
3028 | ||
6ba495e9 | 3029 | mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, |
c9de560d AT |
3030 | (unsigned) orig_size, (unsigned) start); |
3031 | } | |
3032 | ||
3033 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3034 | { | |
3035 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3036 | ||
3037 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3038 | atomic_inc(&sbi->s_bal_reqs); | |
3039 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 3040 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
3041 | atomic_inc(&sbi->s_bal_success); |
3042 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3043 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3044 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3045 | atomic_inc(&sbi->s_bal_goals); | |
3046 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3047 | atomic_inc(&sbi->s_bal_breaks); | |
3048 | } | |
3049 | ||
296c355c TT |
3050 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
3051 | trace_ext4_mballoc_alloc(ac); | |
3052 | else | |
3053 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
3054 | } |
3055 | ||
b844167e CW |
3056 | /* |
3057 | * Called on failure; free up any blocks from the inode PA for this | |
3058 | * context. We don't need this for MB_GROUP_PA because we only change | |
3059 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
3060 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
3061 | */ | |
3062 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
3063 | { | |
3064 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
b844167e | 3065 | |
400db9d3 ZL |
3066 | if (pa && pa->pa_type == MB_INODE_PA) |
3067 | pa->pa_free += ac->ac_b_ex.fe_len; | |
b844167e CW |
3068 | } |
3069 | ||
c9de560d AT |
3070 | /* |
3071 | * use blocks preallocated to inode | |
3072 | */ | |
3073 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3074 | struct ext4_prealloc_space *pa) | |
3075 | { | |
53accfa9 | 3076 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3077 | ext4_fsblk_t start; |
3078 | ext4_fsblk_t end; | |
3079 | int len; | |
3080 | ||
3081 | /* found preallocated blocks, use them */ | |
3082 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
3083 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
3084 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
3085 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
3086 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
3087 | &ac->ac_b_ex.fe_start); | |
3088 | ac->ac_b_ex.fe_len = len; | |
3089 | ac->ac_status = AC_STATUS_FOUND; | |
3090 | ac->ac_pa = pa; | |
3091 | ||
3092 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 3093 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
3094 | BUG_ON(pa->pa_free < len); |
3095 | pa->pa_free -= len; | |
3096 | ||
6ba495e9 | 3097 | mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); |
c9de560d AT |
3098 | } |
3099 | ||
3100 | /* | |
3101 | * use blocks preallocated to locality group | |
3102 | */ | |
3103 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3104 | struct ext4_prealloc_space *pa) | |
3105 | { | |
03cddb80 | 3106 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3107 | |
c9de560d AT |
3108 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3109 | &ac->ac_b_ex.fe_group, | |
3110 | &ac->ac_b_ex.fe_start); | |
3111 | ac->ac_b_ex.fe_len = len; | |
3112 | ac->ac_status = AC_STATUS_FOUND; | |
3113 | ac->ac_pa = pa; | |
3114 | ||
3115 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3116 | * possible race when the group is being loaded concurrently |
c9de560d | 3117 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3118 | * in on-disk bitmap -- see ext4_mb_release_context() |
3119 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 3120 | */ |
6ba495e9 | 3121 | mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); |
c9de560d AT |
3122 | } |
3123 | ||
5e745b04 AK |
3124 | /* |
3125 | * Return the prealloc space that have minimal distance | |
3126 | * from the goal block. @cpa is the prealloc | |
3127 | * space that is having currently known minimal distance | |
3128 | * from the goal block. | |
3129 | */ | |
3130 | static struct ext4_prealloc_space * | |
3131 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3132 | struct ext4_prealloc_space *pa, | |
3133 | struct ext4_prealloc_space *cpa) | |
3134 | { | |
3135 | ext4_fsblk_t cur_distance, new_distance; | |
3136 | ||
3137 | if (cpa == NULL) { | |
3138 | atomic_inc(&pa->pa_count); | |
3139 | return pa; | |
3140 | } | |
3141 | cur_distance = abs(goal_block - cpa->pa_pstart); | |
3142 | new_distance = abs(goal_block - pa->pa_pstart); | |
3143 | ||
5a54b2f1 | 3144 | if (cur_distance <= new_distance) |
5e745b04 AK |
3145 | return cpa; |
3146 | ||
3147 | /* drop the previous reference */ | |
3148 | atomic_dec(&cpa->pa_count); | |
3149 | atomic_inc(&pa->pa_count); | |
3150 | return pa; | |
3151 | } | |
3152 | ||
c9de560d AT |
3153 | /* |
3154 | * search goal blocks in preallocated space | |
3155 | */ | |
4ddfef7b ES |
3156 | static noinline_for_stack int |
3157 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
c9de560d | 3158 | { |
53accfa9 | 3159 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 3160 | int order, i; |
c9de560d AT |
3161 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3162 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3163 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3164 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3165 | |
3166 | /* only data can be preallocated */ | |
3167 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3168 | return 0; | |
3169 | ||
3170 | /* first, try per-file preallocation */ | |
3171 | rcu_read_lock(); | |
9a0762c5 | 3172 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3173 | |
3174 | /* all fields in this condition don't change, | |
3175 | * so we can skip locking for them */ | |
3176 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
3177 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
3178 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
3179 | continue; |
3180 | ||
fb0a387d | 3181 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 3182 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
3183 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
3184 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
3185 | continue; |
3186 | ||
c9de560d AT |
3187 | /* found preallocated blocks, use them */ |
3188 | spin_lock(&pa->pa_lock); | |
3189 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3190 | atomic_inc(&pa->pa_count); | |
3191 | ext4_mb_use_inode_pa(ac, pa); | |
3192 | spin_unlock(&pa->pa_lock); | |
3193 | ac->ac_criteria = 10; | |
3194 | rcu_read_unlock(); | |
3195 | return 1; | |
3196 | } | |
3197 | spin_unlock(&pa->pa_lock); | |
3198 | } | |
3199 | rcu_read_unlock(); | |
3200 | ||
3201 | /* can we use group allocation? */ | |
3202 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3203 | return 0; | |
3204 | ||
3205 | /* inode may have no locality group for some reason */ | |
3206 | lg = ac->ac_lg; | |
3207 | if (lg == NULL) | |
3208 | return 0; | |
6be2ded1 AK |
3209 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3210 | if (order > PREALLOC_TB_SIZE - 1) | |
3211 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3212 | order = PREALLOC_TB_SIZE - 1; | |
3213 | ||
bda00de7 | 3214 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
3215 | /* |
3216 | * search for the prealloc space that is having | |
3217 | * minimal distance from the goal block. | |
3218 | */ | |
6be2ded1 AK |
3219 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3220 | rcu_read_lock(); | |
3221 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3222 | pa_inode_list) { | |
3223 | spin_lock(&pa->pa_lock); | |
3224 | if (pa->pa_deleted == 0 && | |
3225 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3226 | |
3227 | cpa = ext4_mb_check_group_pa(goal_block, | |
3228 | pa, cpa); | |
6be2ded1 | 3229 | } |
c9de560d | 3230 | spin_unlock(&pa->pa_lock); |
c9de560d | 3231 | } |
6be2ded1 | 3232 | rcu_read_unlock(); |
c9de560d | 3233 | } |
5e745b04 AK |
3234 | if (cpa) { |
3235 | ext4_mb_use_group_pa(ac, cpa); | |
3236 | ac->ac_criteria = 20; | |
3237 | return 1; | |
3238 | } | |
c9de560d AT |
3239 | return 0; |
3240 | } | |
3241 | ||
7a2fcbf7 AK |
3242 | /* |
3243 | * the function goes through all block freed in the group | |
3244 | * but not yet committed and marks them used in in-core bitmap. | |
3245 | * buddy must be generated from this bitmap | |
955ce5f5 | 3246 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3247 | */ |
3248 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3249 | ext4_group_t group) | |
3250 | { | |
3251 | struct rb_node *n; | |
3252 | struct ext4_group_info *grp; | |
3253 | struct ext4_free_data *entry; | |
3254 | ||
3255 | grp = ext4_get_group_info(sb, group); | |
3256 | n = rb_first(&(grp->bb_free_root)); | |
3257 | ||
3258 | while (n) { | |
18aadd47 BJ |
3259 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
3260 | ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); | |
7a2fcbf7 AK |
3261 | n = rb_next(n); |
3262 | } | |
3263 | return; | |
3264 | } | |
3265 | ||
c9de560d AT |
3266 | /* |
3267 | * the function goes through all preallocation in this group and marks them | |
3268 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3269 | * Need to be called with ext4 group lock held |
c9de560d | 3270 | */ |
089ceecc ES |
3271 | static noinline_for_stack |
3272 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
3273 | ext4_group_t group) |
3274 | { | |
3275 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3276 | struct ext4_prealloc_space *pa; | |
3277 | struct list_head *cur; | |
3278 | ext4_group_t groupnr; | |
3279 | ext4_grpblk_t start; | |
3280 | int preallocated = 0; | |
c9de560d AT |
3281 | int len; |
3282 | ||
3283 | /* all form of preallocation discards first load group, | |
3284 | * so the only competing code is preallocation use. | |
3285 | * we don't need any locking here | |
3286 | * notice we do NOT ignore preallocations with pa_deleted | |
3287 | * otherwise we could leave used blocks available for | |
3288 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3289 | * is dropping preallocation | |
3290 | */ | |
3291 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3292 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3293 | spin_lock(&pa->pa_lock); | |
3294 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3295 | &groupnr, &start); | |
3296 | len = pa->pa_len; | |
3297 | spin_unlock(&pa->pa_lock); | |
3298 | if (unlikely(len == 0)) | |
3299 | continue; | |
3300 | BUG_ON(groupnr != group); | |
c3e94d1d | 3301 | ext4_set_bits(bitmap, start, len); |
c9de560d | 3302 | preallocated += len; |
c9de560d | 3303 | } |
6ba495e9 | 3304 | mb_debug(1, "prellocated %u for group %u\n", preallocated, group); |
c9de560d AT |
3305 | } |
3306 | ||
3307 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3308 | { | |
3309 | struct ext4_prealloc_space *pa; | |
3310 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
3311 | kmem_cache_free(ext4_pspace_cachep, pa); | |
3312 | } | |
3313 | ||
3314 | /* | |
3315 | * drops a reference to preallocated space descriptor | |
3316 | * if this was the last reference and the space is consumed | |
3317 | */ | |
3318 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3319 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3320 | { | |
a9df9a49 | 3321 | ext4_group_t grp; |
d33a1976 | 3322 | ext4_fsblk_t grp_blk; |
c9de560d AT |
3323 | |
3324 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) | |
3325 | return; | |
3326 | ||
3327 | /* in this short window concurrent discard can set pa_deleted */ | |
3328 | spin_lock(&pa->pa_lock); | |
3329 | if (pa->pa_deleted == 1) { | |
3330 | spin_unlock(&pa->pa_lock); | |
3331 | return; | |
3332 | } | |
3333 | ||
3334 | pa->pa_deleted = 1; | |
3335 | spin_unlock(&pa->pa_lock); | |
3336 | ||
d33a1976 | 3337 | grp_blk = pa->pa_pstart; |
60e6679e | 3338 | /* |
cc0fb9ad AK |
3339 | * If doing group-based preallocation, pa_pstart may be in the |
3340 | * next group when pa is used up | |
3341 | */ | |
3342 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3343 | grp_blk--; |
3344 | ||
3345 | ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL); | |
c9de560d AT |
3346 | |
3347 | /* | |
3348 | * possible race: | |
3349 | * | |
3350 | * P1 (buddy init) P2 (regular allocation) | |
3351 | * find block B in PA | |
3352 | * copy on-disk bitmap to buddy | |
3353 | * mark B in on-disk bitmap | |
3354 | * drop PA from group | |
3355 | * mark all PAs in buddy | |
3356 | * | |
3357 | * thus, P1 initializes buddy with B available. to prevent this | |
3358 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3359 | * against that pair | |
3360 | */ | |
3361 | ext4_lock_group(sb, grp); | |
3362 | list_del(&pa->pa_group_list); | |
3363 | ext4_unlock_group(sb, grp); | |
3364 | ||
3365 | spin_lock(pa->pa_obj_lock); | |
3366 | list_del_rcu(&pa->pa_inode_list); | |
3367 | spin_unlock(pa->pa_obj_lock); | |
3368 | ||
3369 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3370 | } | |
3371 | ||
3372 | /* | |
3373 | * creates new preallocated space for given inode | |
3374 | */ | |
4ddfef7b ES |
3375 | static noinline_for_stack int |
3376 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3377 | { |
3378 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 3379 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
3380 | struct ext4_prealloc_space *pa; |
3381 | struct ext4_group_info *grp; | |
3382 | struct ext4_inode_info *ei; | |
3383 | ||
3384 | /* preallocate only when found space is larger then requested */ | |
3385 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3386 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3387 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3388 | ||
3389 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3390 | if (pa == NULL) | |
3391 | return -ENOMEM; | |
3392 | ||
3393 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3394 | int winl; | |
3395 | int wins; | |
3396 | int win; | |
3397 | int offs; | |
3398 | ||
3399 | /* we can't allocate as much as normalizer wants. | |
3400 | * so, found space must get proper lstart | |
3401 | * to cover original request */ | |
3402 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3403 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3404 | ||
3405 | /* we're limited by original request in that | |
3406 | * logical block must be covered any way | |
3407 | * winl is window we can move our chunk within */ | |
3408 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3409 | ||
3410 | /* also, we should cover whole original request */ | |
53accfa9 | 3411 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
3412 | |
3413 | /* the smallest one defines real window */ | |
3414 | win = min(winl, wins); | |
3415 | ||
53accfa9 TT |
3416 | offs = ac->ac_o_ex.fe_logical % |
3417 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
3418 | if (offs && offs < win) |
3419 | win = offs; | |
3420 | ||
53accfa9 | 3421 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
810da240 | 3422 | EXT4_NUM_B2C(sbi, win); |
c9de560d AT |
3423 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
3424 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3425 | } | |
3426 | ||
3427 | /* preallocation can change ac_b_ex, thus we store actually | |
3428 | * allocated blocks for history */ | |
3429 | ac->ac_f_ex = ac->ac_b_ex; | |
3430 | ||
3431 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3432 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3433 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3434 | pa->pa_free = pa->pa_len; | |
3435 | atomic_set(&pa->pa_count, 1); | |
3436 | spin_lock_init(&pa->pa_lock); | |
d794bf8e AK |
3437 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3438 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3439 | pa->pa_deleted = 0; |
cc0fb9ad | 3440 | pa->pa_type = MB_INODE_PA; |
c9de560d | 3441 | |
6ba495e9 | 3442 | mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, |
c9de560d | 3443 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
9bffad1e | 3444 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3445 | |
3446 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 3447 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
3448 | |
3449 | ei = EXT4_I(ac->ac_inode); | |
3450 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3451 | ||
3452 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3453 | pa->pa_inode = ac->ac_inode; | |
3454 | ||
3455 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3456 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3457 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3458 | ||
3459 | spin_lock(pa->pa_obj_lock); | |
3460 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3461 | spin_unlock(pa->pa_obj_lock); | |
3462 | ||
3463 | return 0; | |
3464 | } | |
3465 | ||
3466 | /* | |
3467 | * creates new preallocated space for locality group inodes belongs to | |
3468 | */ | |
4ddfef7b ES |
3469 | static noinline_for_stack int |
3470 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3471 | { |
3472 | struct super_block *sb = ac->ac_sb; | |
3473 | struct ext4_locality_group *lg; | |
3474 | struct ext4_prealloc_space *pa; | |
3475 | struct ext4_group_info *grp; | |
3476 | ||
3477 | /* preallocate only when found space is larger then requested */ | |
3478 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3479 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3480 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3481 | ||
3482 | BUG_ON(ext4_pspace_cachep == NULL); | |
3483 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3484 | if (pa == NULL) | |
3485 | return -ENOMEM; | |
3486 | ||
3487 | /* preallocation can change ac_b_ex, thus we store actually | |
3488 | * allocated blocks for history */ | |
3489 | ac->ac_f_ex = ac->ac_b_ex; | |
3490 | ||
3491 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3492 | pa->pa_lstart = pa->pa_pstart; | |
3493 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3494 | pa->pa_free = pa->pa_len; | |
3495 | atomic_set(&pa->pa_count, 1); | |
3496 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3497 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 3498 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 3499 | pa->pa_deleted = 0; |
cc0fb9ad | 3500 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 3501 | |
6ba495e9 | 3502 | mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, |
9bffad1e TT |
3503 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
3504 | trace_ext4_mb_new_group_pa(ac, pa); | |
c9de560d AT |
3505 | |
3506 | ext4_mb_use_group_pa(ac, pa); | |
3507 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3508 | ||
3509 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3510 | lg = ac->ac_lg; | |
3511 | BUG_ON(lg == NULL); | |
3512 | ||
3513 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3514 | pa->pa_inode = NULL; | |
3515 | ||
3516 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3517 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3518 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3519 | ||
6be2ded1 AK |
3520 | /* |
3521 | * We will later add the new pa to the right bucket | |
3522 | * after updating the pa_free in ext4_mb_release_context | |
3523 | */ | |
c9de560d AT |
3524 | return 0; |
3525 | } | |
3526 | ||
3527 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3528 | { | |
3529 | int err; | |
3530 | ||
3531 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3532 | err = ext4_mb_new_group_pa(ac); | |
3533 | else | |
3534 | err = ext4_mb_new_inode_pa(ac); | |
3535 | return err; | |
3536 | } | |
3537 | ||
3538 | /* | |
3539 | * finds all unused blocks in on-disk bitmap, frees them in | |
3540 | * in-core bitmap and buddy. | |
3541 | * @pa must be unlinked from inode and group lists, so that | |
3542 | * nobody else can find/use it. | |
3543 | * the caller MUST hold group/inode locks. | |
3544 | * TODO: optimize the case when there are no in-core structures yet | |
3545 | */ | |
4ddfef7b ES |
3546 | static noinline_for_stack int |
3547 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 3548 | struct ext4_prealloc_space *pa) |
c9de560d | 3549 | { |
c9de560d AT |
3550 | struct super_block *sb = e4b->bd_sb; |
3551 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
3552 | unsigned int end; |
3553 | unsigned int next; | |
c9de560d AT |
3554 | ext4_group_t group; |
3555 | ext4_grpblk_t bit; | |
ba80b101 | 3556 | unsigned long long grp_blk_start; |
c9de560d AT |
3557 | int err = 0; |
3558 | int free = 0; | |
3559 | ||
3560 | BUG_ON(pa->pa_deleted == 0); | |
3561 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 3562 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
3563 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
3564 | end = bit + pa->pa_len; | |
3565 | ||
c9de560d | 3566 | while (bit < end) { |
ffad0a44 | 3567 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3568 | if (bit >= end) |
3569 | break; | |
ffad0a44 | 3570 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
6ba495e9 | 3571 | mb_debug(1, " free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
3572 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
3573 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
3574 | free += next - bit; |
3575 | ||
3e1e5f50 | 3576 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
3577 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
3578 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 3579 | next - bit); |
c9de560d AT |
3580 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
3581 | bit = next + 1; | |
3582 | } | |
3583 | if (free != pa->pa_free) { | |
9d8b9ec4 TT |
3584 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
3585 | "pa %p: logic %lu, phys. %lu, len %lu", | |
3586 | pa, (unsigned long) pa->pa_lstart, | |
3587 | (unsigned long) pa->pa_pstart, | |
3588 | (unsigned long) pa->pa_len); | |
e29136f8 | 3589 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 3590 | free, pa->pa_free); |
e56eb659 AK |
3591 | /* |
3592 | * pa is already deleted so we use the value obtained | |
3593 | * from the bitmap and continue. | |
3594 | */ | |
c9de560d | 3595 | } |
c9de560d AT |
3596 | atomic_add(free, &sbi->s_mb_discarded); |
3597 | ||
3598 | return err; | |
3599 | } | |
3600 | ||
4ddfef7b ES |
3601 | static noinline_for_stack int |
3602 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 3603 | struct ext4_prealloc_space *pa) |
c9de560d | 3604 | { |
c9de560d AT |
3605 | struct super_block *sb = e4b->bd_sb; |
3606 | ext4_group_t group; | |
3607 | ext4_grpblk_t bit; | |
3608 | ||
60e07cf5 | 3609 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
3610 | BUG_ON(pa->pa_deleted == 0); |
3611 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3612 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3613 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3614 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 3615 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
3616 | |
3617 | return 0; | |
3618 | } | |
3619 | ||
3620 | /* | |
3621 | * releases all preallocations in given group | |
3622 | * | |
3623 | * first, we need to decide discard policy: | |
3624 | * - when do we discard | |
3625 | * 1) ENOSPC | |
3626 | * - how many do we discard | |
3627 | * 1) how many requested | |
3628 | */ | |
4ddfef7b ES |
3629 | static noinline_for_stack int |
3630 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3631 | ext4_group_t group, int needed) |
3632 | { | |
3633 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3634 | struct buffer_head *bitmap_bh = NULL; | |
3635 | struct ext4_prealloc_space *pa, *tmp; | |
3636 | struct list_head list; | |
3637 | struct ext4_buddy e4b; | |
3638 | int err; | |
3639 | int busy = 0; | |
3640 | int free = 0; | |
3641 | ||
6ba495e9 | 3642 | mb_debug(1, "discard preallocation for group %u\n", group); |
c9de560d AT |
3643 | |
3644 | if (list_empty(&grp->bb_prealloc_list)) | |
3645 | return 0; | |
3646 | ||
574ca174 | 3647 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3648 | if (bitmap_bh == NULL) { |
12062ddd | 3649 | ext4_error(sb, "Error reading block bitmap for %u", group); |
ce89f46c | 3650 | return 0; |
c9de560d AT |
3651 | } |
3652 | ||
3653 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3654 | if (err) { |
12062ddd | 3655 | ext4_error(sb, "Error loading buddy information for %u", group); |
ce89f46c AK |
3656 | put_bh(bitmap_bh); |
3657 | return 0; | |
3658 | } | |
c9de560d AT |
3659 | |
3660 | if (needed == 0) | |
7137d7a4 | 3661 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; |
c9de560d | 3662 | |
c9de560d | 3663 | INIT_LIST_HEAD(&list); |
c9de560d AT |
3664 | repeat: |
3665 | ext4_lock_group(sb, group); | |
3666 | list_for_each_entry_safe(pa, tmp, | |
3667 | &grp->bb_prealloc_list, pa_group_list) { | |
3668 | spin_lock(&pa->pa_lock); | |
3669 | if (atomic_read(&pa->pa_count)) { | |
3670 | spin_unlock(&pa->pa_lock); | |
3671 | busy = 1; | |
3672 | continue; | |
3673 | } | |
3674 | if (pa->pa_deleted) { | |
3675 | spin_unlock(&pa->pa_lock); | |
3676 | continue; | |
3677 | } | |
3678 | ||
3679 | /* seems this one can be freed ... */ | |
3680 | pa->pa_deleted = 1; | |
3681 | ||
3682 | /* we can trust pa_free ... */ | |
3683 | free += pa->pa_free; | |
3684 | ||
3685 | spin_unlock(&pa->pa_lock); | |
3686 | ||
3687 | list_del(&pa->pa_group_list); | |
3688 | list_add(&pa->u.pa_tmp_list, &list); | |
3689 | } | |
3690 | ||
3691 | /* if we still need more blocks and some PAs were used, try again */ | |
3692 | if (free < needed && busy) { | |
3693 | busy = 0; | |
3694 | ext4_unlock_group(sb, group); | |
3695 | /* | |
3696 | * Yield the CPU here so that we don't get soft lockup | |
3697 | * in non preempt case. | |
3698 | */ | |
3699 | yield(); | |
3700 | goto repeat; | |
3701 | } | |
3702 | ||
3703 | /* found anything to free? */ | |
3704 | if (list_empty(&list)) { | |
3705 | BUG_ON(free != 0); | |
3706 | goto out; | |
3707 | } | |
3708 | ||
3709 | /* now free all selected PAs */ | |
3710 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3711 | ||
3712 | /* remove from object (inode or locality group) */ | |
3713 | spin_lock(pa->pa_obj_lock); | |
3714 | list_del_rcu(&pa->pa_inode_list); | |
3715 | spin_unlock(pa->pa_obj_lock); | |
3716 | ||
cc0fb9ad | 3717 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 3718 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 3719 | else |
3e1e5f50 | 3720 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
3721 | |
3722 | list_del(&pa->u.pa_tmp_list); | |
3723 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3724 | } | |
3725 | ||
3726 | out: | |
3727 | ext4_unlock_group(sb, group); | |
e39e07fd | 3728 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
3729 | put_bh(bitmap_bh); |
3730 | return free; | |
3731 | } | |
3732 | ||
3733 | /* | |
3734 | * releases all non-used preallocated blocks for given inode | |
3735 | * | |
3736 | * It's important to discard preallocations under i_data_sem | |
3737 | * We don't want another block to be served from the prealloc | |
3738 | * space when we are discarding the inode prealloc space. | |
3739 | * | |
3740 | * FIXME!! Make sure it is valid at all the call sites | |
3741 | */ | |
c2ea3fde | 3742 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
3743 | { |
3744 | struct ext4_inode_info *ei = EXT4_I(inode); | |
3745 | struct super_block *sb = inode->i_sb; | |
3746 | struct buffer_head *bitmap_bh = NULL; | |
3747 | struct ext4_prealloc_space *pa, *tmp; | |
3748 | ext4_group_t group = 0; | |
3749 | struct list_head list; | |
3750 | struct ext4_buddy e4b; | |
3751 | int err; | |
3752 | ||
c2ea3fde | 3753 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
3754 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
3755 | return; | |
3756 | } | |
3757 | ||
6ba495e9 | 3758 | mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); |
9bffad1e | 3759 | trace_ext4_discard_preallocations(inode); |
c9de560d AT |
3760 | |
3761 | INIT_LIST_HEAD(&list); | |
3762 | ||
3763 | repeat: | |
3764 | /* first, collect all pa's in the inode */ | |
3765 | spin_lock(&ei->i_prealloc_lock); | |
3766 | while (!list_empty(&ei->i_prealloc_list)) { | |
3767 | pa = list_entry(ei->i_prealloc_list.next, | |
3768 | struct ext4_prealloc_space, pa_inode_list); | |
3769 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
3770 | spin_lock(&pa->pa_lock); | |
3771 | if (atomic_read(&pa->pa_count)) { | |
3772 | /* this shouldn't happen often - nobody should | |
3773 | * use preallocation while we're discarding it */ | |
3774 | spin_unlock(&pa->pa_lock); | |
3775 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
3776 | ext4_msg(sb, KERN_ERR, |
3777 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
3778 | WARN_ON(1); |
3779 | schedule_timeout_uninterruptible(HZ); | |
3780 | goto repeat; | |
3781 | ||
3782 | } | |
3783 | if (pa->pa_deleted == 0) { | |
3784 | pa->pa_deleted = 1; | |
3785 | spin_unlock(&pa->pa_lock); | |
3786 | list_del_rcu(&pa->pa_inode_list); | |
3787 | list_add(&pa->u.pa_tmp_list, &list); | |
3788 | continue; | |
3789 | } | |
3790 | ||
3791 | /* someone is deleting pa right now */ | |
3792 | spin_unlock(&pa->pa_lock); | |
3793 | spin_unlock(&ei->i_prealloc_lock); | |
3794 | ||
3795 | /* we have to wait here because pa_deleted | |
3796 | * doesn't mean pa is already unlinked from | |
3797 | * the list. as we might be called from | |
3798 | * ->clear_inode() the inode will get freed | |
3799 | * and concurrent thread which is unlinking | |
3800 | * pa from inode's list may access already | |
3801 | * freed memory, bad-bad-bad */ | |
3802 | ||
3803 | /* XXX: if this happens too often, we can | |
3804 | * add a flag to force wait only in case | |
3805 | * of ->clear_inode(), but not in case of | |
3806 | * regular truncate */ | |
3807 | schedule_timeout_uninterruptible(HZ); | |
3808 | goto repeat; | |
3809 | } | |
3810 | spin_unlock(&ei->i_prealloc_lock); | |
3811 | ||
3812 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 3813 | BUG_ON(pa->pa_type != MB_INODE_PA); |
c9de560d AT |
3814 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); |
3815 | ||
3816 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3817 | if (err) { |
12062ddd ES |
3818 | ext4_error(sb, "Error loading buddy information for %u", |
3819 | group); | |
ce89f46c AK |
3820 | continue; |
3821 | } | |
c9de560d | 3822 | |
574ca174 | 3823 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3824 | if (bitmap_bh == NULL) { |
12062ddd ES |
3825 | ext4_error(sb, "Error reading block bitmap for %u", |
3826 | group); | |
e39e07fd | 3827 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 3828 | continue; |
c9de560d AT |
3829 | } |
3830 | ||
3831 | ext4_lock_group(sb, group); | |
3832 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 3833 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
3834 | ext4_unlock_group(sb, group); |
3835 | ||
e39e07fd | 3836 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
3837 | put_bh(bitmap_bh); |
3838 | ||
3839 | list_del(&pa->u.pa_tmp_list); | |
3840 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3841 | } | |
3842 | } | |
3843 | ||
6ba495e9 | 3844 | #ifdef CONFIG_EXT4_DEBUG |
c9de560d AT |
3845 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
3846 | { | |
3847 | struct super_block *sb = ac->ac_sb; | |
8df9675f | 3848 | ext4_group_t ngroups, i; |
c9de560d | 3849 | |
a0b30c12 | 3850 | if (!ext4_mballoc_debug || |
4dd89fc6 | 3851 | (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) |
e3570639 ES |
3852 | return; |
3853 | ||
7f6a11e7 | 3854 | ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:" |
9d8b9ec4 | 3855 | " Allocation context details:"); |
7f6a11e7 | 3856 | ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d", |
c9de560d | 3857 | ac->ac_status, ac->ac_flags); |
7f6a11e7 | 3858 | ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, " |
9d8b9ec4 TT |
3859 | "goal %lu/%lu/%lu@%lu, " |
3860 | "best %lu/%lu/%lu@%lu cr %d", | |
c9de560d AT |
3861 | (unsigned long)ac->ac_o_ex.fe_group, |
3862 | (unsigned long)ac->ac_o_ex.fe_start, | |
3863 | (unsigned long)ac->ac_o_ex.fe_len, | |
3864 | (unsigned long)ac->ac_o_ex.fe_logical, | |
3865 | (unsigned long)ac->ac_g_ex.fe_group, | |
3866 | (unsigned long)ac->ac_g_ex.fe_start, | |
3867 | (unsigned long)ac->ac_g_ex.fe_len, | |
3868 | (unsigned long)ac->ac_g_ex.fe_logical, | |
3869 | (unsigned long)ac->ac_b_ex.fe_group, | |
3870 | (unsigned long)ac->ac_b_ex.fe_start, | |
3871 | (unsigned long)ac->ac_b_ex.fe_len, | |
3872 | (unsigned long)ac->ac_b_ex.fe_logical, | |
3873 | (int)ac->ac_criteria); | |
7f6a11e7 | 3874 | ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found", |
9d8b9ec4 | 3875 | ac->ac_ex_scanned, ac->ac_found); |
7f6a11e7 | 3876 | ext4_msg(ac->ac_sb, KERN_ERR, "groups: "); |
8df9675f TT |
3877 | ngroups = ext4_get_groups_count(sb); |
3878 | for (i = 0; i < ngroups; i++) { | |
c9de560d AT |
3879 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
3880 | struct ext4_prealloc_space *pa; | |
3881 | ext4_grpblk_t start; | |
3882 | struct list_head *cur; | |
3883 | ext4_lock_group(sb, i); | |
3884 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3885 | pa = list_entry(cur, struct ext4_prealloc_space, | |
3886 | pa_group_list); | |
3887 | spin_lock(&pa->pa_lock); | |
3888 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3889 | NULL, &start); | |
3890 | spin_unlock(&pa->pa_lock); | |
1c718505 AF |
3891 | printk(KERN_ERR "PA:%u:%d:%u \n", i, |
3892 | start, pa->pa_len); | |
c9de560d | 3893 | } |
60bd63d1 | 3894 | ext4_unlock_group(sb, i); |
c9de560d AT |
3895 | |
3896 | if (grp->bb_free == 0) | |
3897 | continue; | |
1c718505 | 3898 | printk(KERN_ERR "%u: %d/%d \n", |
c9de560d AT |
3899 | i, grp->bb_free, grp->bb_fragments); |
3900 | } | |
3901 | printk(KERN_ERR "\n"); | |
3902 | } | |
3903 | #else | |
3904 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
3905 | { | |
3906 | return; | |
3907 | } | |
3908 | #endif | |
3909 | ||
3910 | /* | |
3911 | * We use locality group preallocation for small size file. The size of the | |
3912 | * file is determined by the current size or the resulting size after | |
3913 | * allocation which ever is larger | |
3914 | * | |
b713a5ec | 3915 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
3916 | */ |
3917 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
3918 | { | |
3919 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3920 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
3921 | loff_t size, isize; | |
3922 | ||
3923 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3924 | return; | |
3925 | ||
4ba74d00 TT |
3926 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
3927 | return; | |
3928 | ||
53accfa9 | 3929 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
3930 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
3931 | >> bsbits; | |
c9de560d | 3932 | |
50797481 TT |
3933 | if ((size == isize) && |
3934 | !ext4_fs_is_busy(sbi) && | |
3935 | (atomic_read(&ac->ac_inode->i_writecount) == 0)) { | |
3936 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
3937 | return; | |
3938 | } | |
3939 | ||
ebbe0277 RD |
3940 | if (sbi->s_mb_group_prealloc <= 0) { |
3941 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
3942 | return; | |
3943 | } | |
3944 | ||
c9de560d | 3945 | /* don't use group allocation for large files */ |
71780577 | 3946 | size = max(size, isize); |
cc483f10 | 3947 | if (size > sbi->s_mb_stream_request) { |
4ba74d00 | 3948 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
c9de560d | 3949 | return; |
4ba74d00 | 3950 | } |
c9de560d AT |
3951 | |
3952 | BUG_ON(ac->ac_lg != NULL); | |
3953 | /* | |
3954 | * locality group prealloc space are per cpu. The reason for having | |
3955 | * per cpu locality group is to reduce the contention between block | |
3956 | * request from multiple CPUs. | |
3957 | */ | |
ca0c9584 | 3958 | ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
3959 | |
3960 | /* we're going to use group allocation */ | |
3961 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
3962 | ||
3963 | /* serialize all allocations in the group */ | |
3964 | mutex_lock(&ac->ac_lg->lg_mutex); | |
3965 | } | |
3966 | ||
4ddfef7b ES |
3967 | static noinline_for_stack int |
3968 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
3969 | struct ext4_allocation_request *ar) |
3970 | { | |
3971 | struct super_block *sb = ar->inode->i_sb; | |
3972 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3973 | struct ext4_super_block *es = sbi->s_es; | |
3974 | ext4_group_t group; | |
498e5f24 TT |
3975 | unsigned int len; |
3976 | ext4_fsblk_t goal; | |
c9de560d AT |
3977 | ext4_grpblk_t block; |
3978 | ||
3979 | /* we can't allocate > group size */ | |
3980 | len = ar->len; | |
3981 | ||
3982 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
3983 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
3984 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
3985 | |
3986 | /* start searching from the goal */ | |
3987 | goal = ar->goal; | |
3988 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
3989 | goal >= ext4_blocks_count(es)) | |
3990 | goal = le32_to_cpu(es->s_first_data_block); | |
3991 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
3992 | ||
3993 | /* set up allocation goals */ | |
53accfa9 | 3994 | ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1); |
c9de560d | 3995 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
3996 | ac->ac_sb = sb; |
3997 | ac->ac_inode = ar->inode; | |
53accfa9 | 3998 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
3999 | ac->ac_o_ex.fe_group = group; |
4000 | ac->ac_o_ex.fe_start = block; | |
4001 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 4002 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 4003 | ac->ac_flags = ar->flags; |
c9de560d AT |
4004 | |
4005 | /* we have to define context: we'll we work with a file or | |
4006 | * locality group. this is a policy, actually */ | |
4007 | ext4_mb_group_or_file(ac); | |
4008 | ||
6ba495e9 | 4009 | mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " |
c9de560d AT |
4010 | "left: %u/%u, right %u/%u to %swritable\n", |
4011 | (unsigned) ar->len, (unsigned) ar->logical, | |
4012 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4013 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4014 | (unsigned) ar->lright, (unsigned) ar->pright, | |
4015 | atomic_read(&ar->inode->i_writecount) ? "" : "non-"); | |
4016 | return 0; | |
4017 | ||
4018 | } | |
4019 | ||
6be2ded1 AK |
4020 | static noinline_for_stack void |
4021 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4022 | struct ext4_locality_group *lg, | |
4023 | int order, int total_entries) | |
4024 | { | |
4025 | ext4_group_t group = 0; | |
4026 | struct ext4_buddy e4b; | |
4027 | struct list_head discard_list; | |
4028 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 4029 | |
6ba495e9 | 4030 | mb_debug(1, "discard locality group preallocation\n"); |
6be2ded1 AK |
4031 | |
4032 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
4033 | |
4034 | spin_lock(&lg->lg_prealloc_lock); | |
4035 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
4036 | pa_inode_list) { | |
4037 | spin_lock(&pa->pa_lock); | |
4038 | if (atomic_read(&pa->pa_count)) { | |
4039 | /* | |
4040 | * This is the pa that we just used | |
4041 | * for block allocation. So don't | |
4042 | * free that | |
4043 | */ | |
4044 | spin_unlock(&pa->pa_lock); | |
4045 | continue; | |
4046 | } | |
4047 | if (pa->pa_deleted) { | |
4048 | spin_unlock(&pa->pa_lock); | |
4049 | continue; | |
4050 | } | |
4051 | /* only lg prealloc space */ | |
cc0fb9ad | 4052 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4053 | |
4054 | /* seems this one can be freed ... */ | |
4055 | pa->pa_deleted = 1; | |
4056 | spin_unlock(&pa->pa_lock); | |
4057 | ||
4058 | list_del_rcu(&pa->pa_inode_list); | |
4059 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4060 | ||
4061 | total_entries--; | |
4062 | if (total_entries <= 5) { | |
4063 | /* | |
4064 | * we want to keep only 5 entries | |
4065 | * allowing it to grow to 8. This | |
4066 | * mak sure we don't call discard | |
4067 | * soon for this list. | |
4068 | */ | |
4069 | break; | |
4070 | } | |
4071 | } | |
4072 | spin_unlock(&lg->lg_prealloc_lock); | |
4073 | ||
4074 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
4075 | ||
4076 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); | |
4077 | if (ext4_mb_load_buddy(sb, group, &e4b)) { | |
12062ddd ES |
4078 | ext4_error(sb, "Error loading buddy information for %u", |
4079 | group); | |
6be2ded1 AK |
4080 | continue; |
4081 | } | |
4082 | ext4_lock_group(sb, group); | |
4083 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4084 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
4085 | ext4_unlock_group(sb, group); |
4086 | ||
e39e07fd | 4087 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
4088 | list_del(&pa->u.pa_tmp_list); |
4089 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4090 | } | |
6be2ded1 AK |
4091 | } |
4092 | ||
4093 | /* | |
4094 | * We have incremented pa_count. So it cannot be freed at this | |
4095 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4096 | * possible from this lg. That means pa_free cannot be updated. | |
4097 | * | |
4098 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4099 | * which can cause the lg_prealloc_list to be updated. | |
4100 | */ | |
4101 | ||
4102 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4103 | { | |
4104 | int order, added = 0, lg_prealloc_count = 1; | |
4105 | struct super_block *sb = ac->ac_sb; | |
4106 | struct ext4_locality_group *lg = ac->ac_lg; | |
4107 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4108 | ||
4109 | order = fls(pa->pa_free) - 1; | |
4110 | if (order > PREALLOC_TB_SIZE - 1) | |
4111 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4112 | order = PREALLOC_TB_SIZE - 1; | |
4113 | /* Add the prealloc space to lg */ | |
f1167009 | 4114 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4115 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
4116 | pa_inode_list) { | |
4117 | spin_lock(&tmp_pa->pa_lock); | |
4118 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4119 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4120 | continue; |
4121 | } | |
4122 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4123 | /* Add to the tail of the previous entry */ | |
4124 | list_add_tail_rcu(&pa->pa_inode_list, | |
4125 | &tmp_pa->pa_inode_list); | |
4126 | added = 1; | |
4127 | /* | |
4128 | * we want to count the total | |
4129 | * number of entries in the list | |
4130 | */ | |
4131 | } | |
4132 | spin_unlock(&tmp_pa->pa_lock); | |
4133 | lg_prealloc_count++; | |
4134 | } | |
4135 | if (!added) | |
4136 | list_add_tail_rcu(&pa->pa_inode_list, | |
4137 | &lg->lg_prealloc_list[order]); | |
f1167009 | 4138 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4139 | |
4140 | /* Now trim the list to be not more than 8 elements */ | |
4141 | if (lg_prealloc_count > 8) { | |
4142 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 4143 | order, lg_prealloc_count); |
6be2ded1 AK |
4144 | return; |
4145 | } | |
4146 | return ; | |
4147 | } | |
4148 | ||
c9de560d AT |
4149 | /* |
4150 | * release all resource we used in allocation | |
4151 | */ | |
4152 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4153 | { | |
53accfa9 | 4154 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
4155 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4156 | if (pa) { | |
cc0fb9ad | 4157 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4158 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 4159 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
4160 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
4161 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
4162 | pa->pa_free -= ac->ac_b_ex.fe_len; |
4163 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4164 | spin_unlock(&pa->pa_lock); | |
c9de560d | 4165 | } |
c9de560d | 4166 | } |
ba443916 AK |
4167 | if (pa) { |
4168 | /* | |
4169 | * We want to add the pa to the right bucket. | |
4170 | * Remove it from the list and while adding | |
4171 | * make sure the list to which we are adding | |
44183d42 | 4172 | * doesn't grow big. |
ba443916 | 4173 | */ |
cc0fb9ad | 4174 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { |
ba443916 AK |
4175 | spin_lock(pa->pa_obj_lock); |
4176 | list_del_rcu(&pa->pa_inode_list); | |
4177 | spin_unlock(pa->pa_obj_lock); | |
4178 | ext4_mb_add_n_trim(ac); | |
4179 | } | |
4180 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4181 | } | |
c9de560d AT |
4182 | if (ac->ac_bitmap_page) |
4183 | page_cache_release(ac->ac_bitmap_page); | |
4184 | if (ac->ac_buddy_page) | |
4185 | page_cache_release(ac->ac_buddy_page); | |
4186 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
4187 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4188 | ext4_mb_collect_stats(ac); | |
4189 | return 0; | |
4190 | } | |
4191 | ||
4192 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4193 | { | |
8df9675f | 4194 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4195 | int ret; |
4196 | int freed = 0; | |
4197 | ||
9bffad1e | 4198 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4199 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4200 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4201 | freed += ret; | |
4202 | needed -= ret; | |
4203 | } | |
4204 | ||
4205 | return freed; | |
4206 | } | |
4207 | ||
4208 | /* | |
4209 | * Main entry point into mballoc to allocate blocks | |
4210 | * it tries to use preallocation first, then falls back | |
4211 | * to usual allocation | |
4212 | */ | |
4213 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 4214 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 4215 | { |
6bc6e63f | 4216 | int freed; |
256bdb49 | 4217 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4218 | struct ext4_sb_info *sbi; |
4219 | struct super_block *sb; | |
4220 | ext4_fsblk_t block = 0; | |
60e58e0f | 4221 | unsigned int inquota = 0; |
53accfa9 | 4222 | unsigned int reserv_clstrs = 0; |
c9de560d AT |
4223 | |
4224 | sb = ar->inode->i_sb; | |
4225 | sbi = EXT4_SB(sb); | |
4226 | ||
9bffad1e | 4227 | trace_ext4_request_blocks(ar); |
ba80b101 | 4228 | |
45dc63e7 DM |
4229 | /* Allow to use superuser reservation for quota file */ |
4230 | if (IS_NOQUOTA(ar->inode)) | |
4231 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; | |
4232 | ||
60e58e0f MC |
4233 | /* |
4234 | * For delayed allocation, we could skip the ENOSPC and | |
4235 | * EDQUOT check, as blocks and quotas have been already | |
4236 | * reserved when data being copied into pagecache. | |
4237 | */ | |
f2321097 | 4238 | if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED)) |
60e58e0f MC |
4239 | ar->flags |= EXT4_MB_DELALLOC_RESERVED; |
4240 | else { | |
4241 | /* Without delayed allocation we need to verify | |
4242 | * there is enough free blocks to do block allocation | |
4243 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4244 | */ |
55f020db | 4245 | while (ar->len && |
e7d5f315 | 4246 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 4247 | |
030ba6bc AK |
4248 | /* let others to free the space */ |
4249 | yield(); | |
4250 | ar->len = ar->len >> 1; | |
4251 | } | |
4252 | if (!ar->len) { | |
a30d542a AK |
4253 | *errp = -ENOSPC; |
4254 | return 0; | |
4255 | } | |
53accfa9 | 4256 | reserv_clstrs = ar->len; |
55f020db | 4257 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
4258 | dquot_alloc_block_nofail(ar->inode, |
4259 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
4260 | } else { |
4261 | while (ar->len && | |
53accfa9 TT |
4262 | dquot_alloc_block(ar->inode, |
4263 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
4264 | |
4265 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
4266 | ar->len--; | |
4267 | } | |
60e58e0f MC |
4268 | } |
4269 | inquota = ar->len; | |
4270 | if (ar->len == 0) { | |
4271 | *errp = -EDQUOT; | |
6c7a120a | 4272 | goto out; |
60e58e0f | 4273 | } |
07031431 | 4274 | } |
d2a17637 | 4275 | |
85556c9a | 4276 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 4277 | if (!ac) { |
363d4251 | 4278 | ar->len = 0; |
256bdb49 | 4279 | *errp = -ENOMEM; |
6c7a120a | 4280 | goto out; |
256bdb49 ES |
4281 | } |
4282 | ||
256bdb49 | 4283 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4284 | if (*errp) { |
4285 | ar->len = 0; | |
6c7a120a | 4286 | goto out; |
c9de560d AT |
4287 | } |
4288 | ||
256bdb49 ES |
4289 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4290 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4291 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4292 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4293 | repeat: |
4294 | /* allocate space in core */ | |
6c7a120a | 4295 | *errp = ext4_mb_regular_allocator(ac); |
6d138ced ES |
4296 | if (*errp) { |
4297 | ext4_discard_allocated_blocks(ac); | |
6c7a120a | 4298 | goto errout; |
6d138ced | 4299 | } |
c9de560d AT |
4300 | |
4301 | /* as we've just preallocated more space than | |
4302 | * user requested orinally, we store allocated | |
4303 | * space in a special descriptor */ | |
256bdb49 ES |
4304 | if (ac->ac_status == AC_STATUS_FOUND && |
4305 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
4306 | ext4_mb_new_preallocation(ac); | |
c9de560d | 4307 | } |
256bdb49 | 4308 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 4309 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
6c7a120a | 4310 | if (*errp == -EAGAIN) { |
8556e8f3 AK |
4311 | /* |
4312 | * drop the reference that we took | |
4313 | * in ext4_mb_use_best_found | |
4314 | */ | |
4315 | ext4_mb_release_context(ac); | |
519deca0 AK |
4316 | ac->ac_b_ex.fe_group = 0; |
4317 | ac->ac_b_ex.fe_start = 0; | |
4318 | ac->ac_b_ex.fe_len = 0; | |
4319 | ac->ac_status = AC_STATUS_CONTINUE; | |
4320 | goto repeat; | |
6d138ced | 4321 | } else if (*errp) { |
b844167e | 4322 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
4323 | goto errout; |
4324 | } else { | |
519deca0 AK |
4325 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
4326 | ar->len = ac->ac_b_ex.fe_len; | |
4327 | } | |
c9de560d | 4328 | } else { |
256bdb49 | 4329 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4330 | if (freed) |
4331 | goto repeat; | |
4332 | *errp = -ENOSPC; | |
6c7a120a AK |
4333 | } |
4334 | ||
6d138ced | 4335 | errout: |
6c7a120a | 4336 | if (*errp) { |
256bdb49 | 4337 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4338 | ar->len = 0; |
256bdb49 | 4339 | ext4_mb_show_ac(ac); |
c9de560d | 4340 | } |
256bdb49 | 4341 | ext4_mb_release_context(ac); |
6c7a120a AK |
4342 | out: |
4343 | if (ac) | |
4344 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 4345 | if (inquota && ar->len < inquota) |
53accfa9 | 4346 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 4347 | if (!ar->len) { |
f2321097 TT |
4348 | if (!ext4_test_inode_state(ar->inode, |
4349 | EXT4_STATE_DELALLOC_RESERVED)) | |
0087d9fb | 4350 | /* release all the reserved blocks if non delalloc */ |
57042651 | 4351 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 4352 | reserv_clstrs); |
0087d9fb | 4353 | } |
c9de560d | 4354 | |
9bffad1e | 4355 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4356 | |
c9de560d AT |
4357 | return block; |
4358 | } | |
c9de560d | 4359 | |
c894058d AK |
4360 | /* |
4361 | * We can merge two free data extents only if the physical blocks | |
4362 | * are contiguous, AND the extents were freed by the same transaction, | |
4363 | * AND the blocks are associated with the same group. | |
4364 | */ | |
4365 | static int can_merge(struct ext4_free_data *entry1, | |
4366 | struct ext4_free_data *entry2) | |
4367 | { | |
18aadd47 BJ |
4368 | if ((entry1->efd_tid == entry2->efd_tid) && |
4369 | (entry1->efd_group == entry2->efd_group) && | |
4370 | ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster)) | |
c894058d AK |
4371 | return 1; |
4372 | return 0; | |
4373 | } | |
4374 | ||
4ddfef7b ES |
4375 | static noinline_for_stack int |
4376 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 4377 | struct ext4_free_data *new_entry) |
c9de560d | 4378 | { |
e29136f8 | 4379 | ext4_group_t group = e4b->bd_group; |
84130193 | 4380 | ext4_grpblk_t cluster; |
7a2fcbf7 | 4381 | struct ext4_free_data *entry; |
c9de560d AT |
4382 | struct ext4_group_info *db = e4b->bd_info; |
4383 | struct super_block *sb = e4b->bd_sb; | |
4384 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4385 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
4386 | struct rb_node *parent = NULL, *new_node; | |
4387 | ||
0390131b | 4388 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4389 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4390 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4391 | ||
18aadd47 BJ |
4392 | new_node = &new_entry->efd_node; |
4393 | cluster = new_entry->efd_start_cluster; | |
c894058d | 4394 | |
c894058d AK |
4395 | if (!*n) { |
4396 | /* first free block exent. We need to | |
4397 | protect buddy cache from being freed, | |
4398 | * otherwise we'll refresh it from | |
4399 | * on-disk bitmap and lose not-yet-available | |
4400 | * blocks */ | |
4401 | page_cache_get(e4b->bd_buddy_page); | |
4402 | page_cache_get(e4b->bd_bitmap_page); | |
4403 | } | |
4404 | while (*n) { | |
4405 | parent = *n; | |
18aadd47 BJ |
4406 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
4407 | if (cluster < entry->efd_start_cluster) | |
c894058d | 4408 | n = &(*n)->rb_left; |
18aadd47 | 4409 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
4410 | n = &(*n)->rb_right; |
4411 | else { | |
e29136f8 | 4412 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
4413 | ext4_group_first_block_no(sb, group) + |
4414 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 4415 | "Block already on to-be-freed list"); |
c894058d | 4416 | return 0; |
c9de560d | 4417 | } |
c894058d | 4418 | } |
c9de560d | 4419 | |
c894058d AK |
4420 | rb_link_node(new_node, parent, n); |
4421 | rb_insert_color(new_node, &db->bb_free_root); | |
4422 | ||
4423 | /* Now try to see the extent can be merged to left and right */ | |
4424 | node = rb_prev(new_node); | |
4425 | if (node) { | |
18aadd47 | 4426 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
c894058d | 4427 | if (can_merge(entry, new_entry)) { |
18aadd47 BJ |
4428 | new_entry->efd_start_cluster = entry->efd_start_cluster; |
4429 | new_entry->efd_count += entry->efd_count; | |
c894058d | 4430 | rb_erase(node, &(db->bb_free_root)); |
18aadd47 BJ |
4431 | ext4_journal_callback_del(handle, &entry->efd_jce); |
4432 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c9de560d | 4433 | } |
c894058d | 4434 | } |
c9de560d | 4435 | |
c894058d AK |
4436 | node = rb_next(new_node); |
4437 | if (node) { | |
18aadd47 | 4438 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
c894058d | 4439 | if (can_merge(new_entry, entry)) { |
18aadd47 | 4440 | new_entry->efd_count += entry->efd_count; |
c894058d | 4441 | rb_erase(node, &(db->bb_free_root)); |
18aadd47 BJ |
4442 | ext4_journal_callback_del(handle, &entry->efd_jce); |
4443 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c9de560d AT |
4444 | } |
4445 | } | |
3e624fc7 | 4446 | /* Add the extent to transaction's private list */ |
18aadd47 BJ |
4447 | ext4_journal_callback_add(handle, ext4_free_data_callback, |
4448 | &new_entry->efd_jce); | |
c9de560d AT |
4449 | return 0; |
4450 | } | |
4451 | ||
44338711 TT |
4452 | /** |
4453 | * ext4_free_blocks() -- Free given blocks and update quota | |
4454 | * @handle: handle for this transaction | |
4455 | * @inode: inode | |
4456 | * @block: start physical block to free | |
4457 | * @count: number of blocks to count | |
5def1360 | 4458 | * @flags: flags used by ext4_free_blocks |
c9de560d | 4459 | */ |
44338711 | 4460 | void ext4_free_blocks(handle_t *handle, struct inode *inode, |
e6362609 TT |
4461 | struct buffer_head *bh, ext4_fsblk_t block, |
4462 | unsigned long count, int flags) | |
c9de560d | 4463 | { |
26346ff6 | 4464 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4465 | struct super_block *sb = inode->i_sb; |
c9de560d | 4466 | struct ext4_group_desc *gdp; |
44338711 | 4467 | unsigned long freed = 0; |
498e5f24 | 4468 | unsigned int overflow; |
c9de560d AT |
4469 | ext4_grpblk_t bit; |
4470 | struct buffer_head *gd_bh; | |
4471 | ext4_group_t block_group; | |
4472 | struct ext4_sb_info *sbi; | |
4473 | struct ext4_buddy e4b; | |
84130193 | 4474 | unsigned int count_clusters; |
c9de560d AT |
4475 | int err = 0; |
4476 | int ret; | |
4477 | ||
e6362609 TT |
4478 | if (bh) { |
4479 | if (block) | |
4480 | BUG_ON(block != bh->b_blocknr); | |
4481 | else | |
4482 | block = bh->b_blocknr; | |
4483 | } | |
c9de560d | 4484 | |
c9de560d | 4485 | sbi = EXT4_SB(sb); |
1f2acb60 TT |
4486 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
4487 | !ext4_data_block_valid(sbi, block, count)) { | |
12062ddd | 4488 | ext4_error(sb, "Freeing blocks not in datazone - " |
1f2acb60 | 4489 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
4490 | goto error_return; |
4491 | } | |
4492 | ||
0610b6e9 | 4493 | ext4_debug("freeing block %llu\n", block); |
e6362609 TT |
4494 | trace_ext4_free_blocks(inode, block, count, flags); |
4495 | ||
4496 | if (flags & EXT4_FREE_BLOCKS_FORGET) { | |
4497 | struct buffer_head *tbh = bh; | |
4498 | int i; | |
4499 | ||
4500 | BUG_ON(bh && (count > 1)); | |
4501 | ||
4502 | for (i = 0; i < count; i++) { | |
4503 | if (!bh) | |
4504 | tbh = sb_find_get_block(inode->i_sb, | |
4505 | block + i); | |
87783690 NK |
4506 | if (unlikely(!tbh)) |
4507 | continue; | |
60e6679e | 4508 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, |
e6362609 TT |
4509 | inode, tbh, block + i); |
4510 | } | |
4511 | } | |
4512 | ||
60e6679e | 4513 | /* |
e6362609 TT |
4514 | * We need to make sure we don't reuse the freed block until |
4515 | * after the transaction is committed, which we can do by | |
4516 | * treating the block as metadata, below. We make an | |
4517 | * exception if the inode is to be written in writeback mode | |
4518 | * since writeback mode has weak data consistency guarantees. | |
4519 | */ | |
4520 | if (!ext4_should_writeback_data(inode)) | |
4521 | flags |= EXT4_FREE_BLOCKS_METADATA; | |
c9de560d | 4522 | |
84130193 TT |
4523 | /* |
4524 | * If the extent to be freed does not begin on a cluster | |
4525 | * boundary, we need to deal with partial clusters at the | |
4526 | * beginning and end of the extent. Normally we will free | |
4527 | * blocks at the beginning or the end unless we are explicitly | |
4528 | * requested to avoid doing so. | |
4529 | */ | |
4530 | overflow = block & (sbi->s_cluster_ratio - 1); | |
4531 | if (overflow) { | |
4532 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
4533 | overflow = sbi->s_cluster_ratio - overflow; | |
4534 | block += overflow; | |
4535 | if (count > overflow) | |
4536 | count -= overflow; | |
4537 | else | |
4538 | return; | |
4539 | } else { | |
4540 | block -= overflow; | |
4541 | count += overflow; | |
4542 | } | |
4543 | } | |
4544 | overflow = count & (sbi->s_cluster_ratio - 1); | |
4545 | if (overflow) { | |
4546 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
4547 | if (count > overflow) | |
4548 | count -= overflow; | |
4549 | else | |
4550 | return; | |
4551 | } else | |
4552 | count += sbi->s_cluster_ratio - overflow; | |
4553 | } | |
4554 | ||
c9de560d AT |
4555 | do_more: |
4556 | overflow = 0; | |
4557 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4558 | ||
4559 | /* | |
4560 | * Check to see if we are freeing blocks across a group | |
4561 | * boundary. | |
4562 | */ | |
84130193 TT |
4563 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
4564 | overflow = EXT4_C2B(sbi, bit) + count - | |
4565 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d AT |
4566 | count -= overflow; |
4567 | } | |
810da240 | 4568 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 4569 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
ce89f46c AK |
4570 | if (!bitmap_bh) { |
4571 | err = -EIO; | |
c9de560d | 4572 | goto error_return; |
ce89f46c | 4573 | } |
c9de560d | 4574 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4575 | if (!gdp) { |
4576 | err = -EIO; | |
c9de560d | 4577 | goto error_return; |
ce89f46c | 4578 | } |
c9de560d AT |
4579 | |
4580 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4581 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4582 | in_range(block, ext4_inode_table(sb, gdp), | |
84130193 | 4583 | EXT4_SB(sb)->s_itb_per_group) || |
c9de560d | 4584 | in_range(block + count - 1, ext4_inode_table(sb, gdp), |
84130193 | 4585 | EXT4_SB(sb)->s_itb_per_group)) { |
c9de560d | 4586 | |
12062ddd | 4587 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 4588 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
4589 | /* err = 0. ext4_std_error should be a no op */ |
4590 | goto error_return; | |
c9de560d AT |
4591 | } |
4592 | ||
4593 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4594 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4595 | if (err) | |
4596 | goto error_return; | |
4597 | ||
4598 | /* | |
4599 | * We are about to modify some metadata. Call the journal APIs | |
4600 | * to unshare ->b_data if a currently-committing transaction is | |
4601 | * using it | |
4602 | */ | |
4603 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4604 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4605 | if (err) | |
4606 | goto error_return; | |
c9de560d AT |
4607 | #ifdef AGGRESSIVE_CHECK |
4608 | { | |
4609 | int i; | |
84130193 | 4610 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
4611 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
4612 | } | |
4613 | #endif | |
84130193 | 4614 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 4615 | |
920313a7 AK |
4616 | err = ext4_mb_load_buddy(sb, block_group, &e4b); |
4617 | if (err) | |
4618 | goto error_return; | |
e6362609 TT |
4619 | |
4620 | if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) { | |
7a2fcbf7 AK |
4621 | struct ext4_free_data *new_entry; |
4622 | /* | |
4623 | * blocks being freed are metadata. these blocks shouldn't | |
4624 | * be used until this transaction is committed | |
4625 | */ | |
18aadd47 | 4626 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS); |
b72143ab | 4627 | if (!new_entry) { |
02b78310 | 4628 | ext4_mb_unload_buddy(&e4b); |
b72143ab TT |
4629 | err = -ENOMEM; |
4630 | goto error_return; | |
4631 | } | |
18aadd47 BJ |
4632 | new_entry->efd_start_cluster = bit; |
4633 | new_entry->efd_group = block_group; | |
4634 | new_entry->efd_count = count_clusters; | |
4635 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 4636 | |
7a2fcbf7 | 4637 | ext4_lock_group(sb, block_group); |
84130193 | 4638 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 4639 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 4640 | } else { |
7a2fcbf7 AK |
4641 | /* need to update group_info->bb_free and bitmap |
4642 | * with group lock held. generate_buddy look at | |
4643 | * them with group lock_held | |
4644 | */ | |
d71c1ae2 LC |
4645 | if (test_opt(sb, DISCARD)) { |
4646 | err = ext4_issue_discard(sb, block_group, bit, count); | |
4647 | if (err && err != -EOPNOTSUPP) | |
4648 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
4649 | " group:%d block:%d count:%lu failed" | |
4650 | " with %d", block_group, bit, count, | |
4651 | err); | |
4652 | } | |
4653 | ||
4654 | ||
955ce5f5 | 4655 | ext4_lock_group(sb, block_group); |
84130193 TT |
4656 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
4657 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
4658 | } |
4659 | ||
021b65bb TT |
4660 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
4661 | ext4_free_group_clusters_set(sb, gdp, ret); | |
79f1ba49 | 4662 | ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); |
feb0ab32 | 4663 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 4664 | ext4_unlock_group(sb, block_group); |
57042651 | 4665 | percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters); |
c9de560d | 4666 | |
772cb7c8 JS |
4667 | if (sbi->s_log_groups_per_flex) { |
4668 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
24aaa8ef TT |
4669 | atomic_add(count_clusters, |
4670 | &sbi->s_flex_groups[flex_group].free_clusters); | |
772cb7c8 JS |
4671 | } |
4672 | ||
e39e07fd | 4673 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 4674 | |
44338711 | 4675 | freed += count; |
c9de560d | 4676 | |
7b415bf6 AK |
4677 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) |
4678 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); | |
4679 | ||
7a2fcbf7 AK |
4680 | /* We dirtied the bitmap block */ |
4681 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4682 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4683 | ||
c9de560d AT |
4684 | /* And the group descriptor block */ |
4685 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 4686 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
4687 | if (!err) |
4688 | err = ret; | |
4689 | ||
4690 | if (overflow && !err) { | |
4691 | block += count; | |
4692 | count = overflow; | |
4693 | put_bh(bitmap_bh); | |
4694 | goto do_more; | |
4695 | } | |
c9de560d AT |
4696 | error_return: |
4697 | brelse(bitmap_bh); | |
4698 | ext4_std_error(sb, err); | |
4699 | return; | |
4700 | } | |
7360d173 | 4701 | |
2846e820 | 4702 | /** |
0529155e | 4703 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
4704 | * @handle: handle to this transaction |
4705 | * @sb: super block | |
4907cb7b | 4706 | * @block: start physical block to add to the block group |
2846e820 AG |
4707 | * @count: number of blocks to free |
4708 | * | |
e73a347b | 4709 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 4710 | */ |
cc7365df | 4711 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
4712 | ext4_fsblk_t block, unsigned long count) |
4713 | { | |
4714 | struct buffer_head *bitmap_bh = NULL; | |
4715 | struct buffer_head *gd_bh; | |
4716 | ext4_group_t block_group; | |
4717 | ext4_grpblk_t bit; | |
4718 | unsigned int i; | |
4719 | struct ext4_group_desc *desc; | |
4720 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 4721 | struct ext4_buddy e4b; |
2846e820 AG |
4722 | int err = 0, ret, blk_free_count; |
4723 | ext4_grpblk_t blocks_freed; | |
2846e820 AG |
4724 | |
4725 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
4726 | ||
4740b830 YY |
4727 | if (count == 0) |
4728 | return 0; | |
4729 | ||
2846e820 | 4730 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
4731 | /* |
4732 | * Check to see if we are freeing blocks across a group | |
4733 | * boundary. | |
4734 | */ | |
cc7365df YY |
4735 | if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
4736 | ext4_warning(sb, "too much blocks added to group %u\n", | |
4737 | block_group); | |
4738 | err = -EINVAL; | |
2846e820 | 4739 | goto error_return; |
cc7365df | 4740 | } |
2cd05cc3 | 4741 | |
2846e820 | 4742 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
cc7365df YY |
4743 | if (!bitmap_bh) { |
4744 | err = -EIO; | |
2846e820 | 4745 | goto error_return; |
cc7365df YY |
4746 | } |
4747 | ||
2846e820 | 4748 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
4749 | if (!desc) { |
4750 | err = -EIO; | |
2846e820 | 4751 | goto error_return; |
cc7365df | 4752 | } |
2846e820 AG |
4753 | |
4754 | if (in_range(ext4_block_bitmap(sb, desc), block, count) || | |
4755 | in_range(ext4_inode_bitmap(sb, desc), block, count) || | |
4756 | in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || | |
4757 | in_range(block + count - 1, ext4_inode_table(sb, desc), | |
4758 | sbi->s_itb_per_group)) { | |
4759 | ext4_error(sb, "Adding blocks in system zones - " | |
4760 | "Block = %llu, count = %lu", | |
4761 | block, count); | |
cc7365df | 4762 | err = -EINVAL; |
2846e820 AG |
4763 | goto error_return; |
4764 | } | |
4765 | ||
2cd05cc3 TT |
4766 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
4767 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2846e820 AG |
4768 | if (err) |
4769 | goto error_return; | |
4770 | ||
4771 | /* | |
4772 | * We are about to modify some metadata. Call the journal APIs | |
4773 | * to unshare ->b_data if a currently-committing transaction is | |
4774 | * using it | |
4775 | */ | |
4776 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4777 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4778 | if (err) | |
4779 | goto error_return; | |
e73a347b | 4780 | |
2846e820 AG |
4781 | for (i = 0, blocks_freed = 0; i < count; i++) { |
4782 | BUFFER_TRACE(bitmap_bh, "clear bit"); | |
e73a347b | 4783 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
4784 | ext4_error(sb, "bit already cleared for block %llu", |
4785 | (ext4_fsblk_t)(block + i)); | |
4786 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
4787 | } else { | |
4788 | blocks_freed++; | |
4789 | } | |
4790 | } | |
e73a347b AG |
4791 | |
4792 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
4793 | if (err) | |
4794 | goto error_return; | |
4795 | ||
4796 | /* | |
4797 | * need to update group_info->bb_free and bitmap | |
4798 | * with group lock held. generate_buddy look at | |
4799 | * them with group lock_held | |
4800 | */ | |
2846e820 | 4801 | ext4_lock_group(sb, block_group); |
e73a347b AG |
4802 | mb_clear_bits(bitmap_bh->b_data, bit, count); |
4803 | mb_free_blocks(NULL, &e4b, bit, count); | |
021b65bb TT |
4804 | blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc); |
4805 | ext4_free_group_clusters_set(sb, desc, blk_free_count); | |
79f1ba49 | 4806 | ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); |
feb0ab32 | 4807 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 4808 | ext4_unlock_group(sb, block_group); |
57042651 | 4809 | percpu_counter_add(&sbi->s_freeclusters_counter, |
810da240 | 4810 | EXT4_NUM_B2C(sbi, blocks_freed)); |
2846e820 AG |
4811 | |
4812 | if (sbi->s_log_groups_per_flex) { | |
4813 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
810da240 | 4814 | atomic_add(EXT4_NUM_B2C(sbi, blocks_freed), |
24aaa8ef | 4815 | &sbi->s_flex_groups[flex_group].free_clusters); |
2846e820 | 4816 | } |
e73a347b AG |
4817 | |
4818 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
4819 | |
4820 | /* We dirtied the bitmap block */ | |
4821 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4822 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4823 | ||
4824 | /* And the group descriptor block */ | |
4825 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
4826 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
4827 | if (!err) | |
4828 | err = ret; | |
4829 | ||
4830 | error_return: | |
4831 | brelse(bitmap_bh); | |
4832 | ext4_std_error(sb, err); | |
cc7365df | 4833 | return err; |
2846e820 AG |
4834 | } |
4835 | ||
7360d173 LC |
4836 | /** |
4837 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
4838 | * @sb: super block for the file system | |
4839 | * @start: starting block of the free extent in the alloc. group | |
4840 | * @count: number of blocks to TRIM | |
4841 | * @group: alloc. group we are working with | |
4842 | * @e4b: ext4 buddy for the group | |
4843 | * | |
4844 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
4845 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
4846 | * be called with under the group lock. | |
4847 | */ | |
d71c1ae2 | 4848 | static int ext4_trim_extent(struct super_block *sb, int start, int count, |
d9f34504 | 4849 | ext4_group_t group, struct ext4_buddy *e4b) |
7360d173 LC |
4850 | { |
4851 | struct ext4_free_extent ex; | |
d71c1ae2 | 4852 | int ret = 0; |
7360d173 | 4853 | |
b3d4c2b1 TM |
4854 | trace_ext4_trim_extent(sb, group, start, count); |
4855 | ||
7360d173 LC |
4856 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
4857 | ||
4858 | ex.fe_start = start; | |
4859 | ex.fe_group = group; | |
4860 | ex.fe_len = count; | |
4861 | ||
4862 | /* | |
4863 | * Mark blocks used, so no one can reuse them while | |
4864 | * being trimmed. | |
4865 | */ | |
4866 | mb_mark_used(e4b, &ex); | |
4867 | ext4_unlock_group(sb, group); | |
d71c1ae2 | 4868 | ret = ext4_issue_discard(sb, group, start, count); |
7360d173 LC |
4869 | ext4_lock_group(sb, group); |
4870 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 4871 | return ret; |
7360d173 LC |
4872 | } |
4873 | ||
4874 | /** | |
4875 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
4876 | * @sb: super block for file system | |
22612283 | 4877 | * @group: group to be trimmed |
7360d173 LC |
4878 | * @start: first group block to examine |
4879 | * @max: last group block to examine | |
4880 | * @minblocks: minimum extent block count | |
4881 | * | |
4882 | * ext4_trim_all_free walks through group's buddy bitmap searching for free | |
4883 | * extents. When the free block is found, ext4_trim_extent is called to TRIM | |
4884 | * the extent. | |
4885 | * | |
4886 | * | |
4887 | * ext4_trim_all_free walks through group's block bitmap searching for free | |
4888 | * extents. When the free extent is found, mark it as used in group buddy | |
4889 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
4890 | * the group buddy bitmap. This is done until whole group is scanned. | |
4891 | */ | |
0b75a840 | 4892 | static ext4_grpblk_t |
78944086 LC |
4893 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
4894 | ext4_grpblk_t start, ext4_grpblk_t max, | |
4895 | ext4_grpblk_t minblocks) | |
7360d173 LC |
4896 | { |
4897 | void *bitmap; | |
169ddc3e | 4898 | ext4_grpblk_t next, count = 0, free_count = 0; |
78944086 | 4899 | struct ext4_buddy e4b; |
d71c1ae2 | 4900 | int ret = 0; |
7360d173 | 4901 | |
b3d4c2b1 TM |
4902 | trace_ext4_trim_all_free(sb, group, start, max); |
4903 | ||
78944086 LC |
4904 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
4905 | if (ret) { | |
4906 | ext4_error(sb, "Error in loading buddy " | |
4907 | "information for %u", group); | |
4908 | return ret; | |
4909 | } | |
78944086 | 4910 | bitmap = e4b.bd_bitmap; |
28739eea LC |
4911 | |
4912 | ext4_lock_group(sb, group); | |
3d56b8d2 TM |
4913 | if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && |
4914 | minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) | |
4915 | goto out; | |
4916 | ||
78944086 LC |
4917 | start = (e4b.bd_info->bb_first_free > start) ? |
4918 | e4b.bd_info->bb_first_free : start; | |
7360d173 | 4919 | |
913eed83 LC |
4920 | while (start <= max) { |
4921 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
4922 | if (start > max) | |
7360d173 | 4923 | break; |
913eed83 | 4924 | next = mb_find_next_bit(bitmap, max + 1, start); |
7360d173 LC |
4925 | |
4926 | if ((next - start) >= minblocks) { | |
d71c1ae2 LC |
4927 | ret = ext4_trim_extent(sb, start, |
4928 | next - start, group, &e4b); | |
4929 | if (ret && ret != -EOPNOTSUPP) | |
4930 | break; | |
4931 | ret = 0; | |
7360d173 LC |
4932 | count += next - start; |
4933 | } | |
169ddc3e | 4934 | free_count += next - start; |
7360d173 LC |
4935 | start = next + 1; |
4936 | ||
4937 | if (fatal_signal_pending(current)) { | |
4938 | count = -ERESTARTSYS; | |
4939 | break; | |
4940 | } | |
4941 | ||
4942 | if (need_resched()) { | |
4943 | ext4_unlock_group(sb, group); | |
4944 | cond_resched(); | |
4945 | ext4_lock_group(sb, group); | |
4946 | } | |
4947 | ||
169ddc3e | 4948 | if ((e4b.bd_info->bb_free - free_count) < minblocks) |
7360d173 LC |
4949 | break; |
4950 | } | |
3d56b8d2 | 4951 | |
d71c1ae2 LC |
4952 | if (!ret) { |
4953 | ret = count; | |
3d56b8d2 | 4954 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); |
d71c1ae2 | 4955 | } |
3d56b8d2 | 4956 | out: |
7360d173 | 4957 | ext4_unlock_group(sb, group); |
78944086 | 4958 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
4959 | |
4960 | ext4_debug("trimmed %d blocks in the group %d\n", | |
4961 | count, group); | |
4962 | ||
d71c1ae2 | 4963 | return ret; |
7360d173 LC |
4964 | } |
4965 | ||
4966 | /** | |
4967 | * ext4_trim_fs() -- trim ioctl handle function | |
4968 | * @sb: superblock for filesystem | |
4969 | * @range: fstrim_range structure | |
4970 | * | |
4971 | * start: First Byte to trim | |
4972 | * len: number of Bytes to trim from start | |
4973 | * minlen: minimum extent length in Bytes | |
4974 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
4975 | * start to start+len. For each such a group ext4_trim_all_free function | |
4976 | * is invoked to trim all free space. | |
4977 | */ | |
4978 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
4979 | { | |
78944086 | 4980 | struct ext4_group_info *grp; |
913eed83 | 4981 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 4982 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 4983 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
4984 | ext4_fsblk_t first_data_blk = |
4985 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 4986 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
7360d173 LC |
4987 | int ret = 0; |
4988 | ||
4989 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 4990 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
4991 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
4992 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 4993 | |
5de35e8d LC |
4994 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
4995 | start >= max_blks || | |
4996 | range->len < sb->s_blocksize) | |
7360d173 | 4997 | return -EINVAL; |
913eed83 LC |
4998 | if (end >= max_blks) |
4999 | end = max_blks - 1; | |
5000 | if (end <= first_data_blk) | |
22f10457 | 5001 | goto out; |
913eed83 | 5002 | if (start < first_data_blk) |
0f0a25bf | 5003 | start = first_data_blk; |
7360d173 | 5004 | |
913eed83 | 5005 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 5006 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 5007 | &first_group, &first_cluster); |
913eed83 | 5008 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 5009 | &last_group, &last_cluster); |
7360d173 | 5010 | |
913eed83 LC |
5011 | /* end now represents the last cluster to discard in this group */ |
5012 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
7360d173 LC |
5013 | |
5014 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
5015 | grp = ext4_get_group_info(sb, group); |
5016 | /* We only do this if the grp has never been initialized */ | |
5017 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
5018 | ret = ext4_mb_init_group(sb, group); | |
5019 | if (ret) | |
5020 | break; | |
7360d173 LC |
5021 | } |
5022 | ||
0ba08517 | 5023 | /* |
913eed83 LC |
5024 | * For all the groups except the last one, last cluster will |
5025 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
5026 | * change it for the last group, note that last_cluster is | |
5027 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 5028 | */ |
913eed83 LC |
5029 | if (group == last_group) |
5030 | end = last_cluster; | |
7360d173 | 5031 | |
78944086 | 5032 | if (grp->bb_free >= minlen) { |
7137d7a4 | 5033 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
913eed83 | 5034 | end, minlen); |
7360d173 LC |
5035 | if (cnt < 0) { |
5036 | ret = cnt; | |
7360d173 LC |
5037 | break; |
5038 | } | |
21e7fd22 | 5039 | trimmed += cnt; |
7360d173 | 5040 | } |
913eed83 LC |
5041 | |
5042 | /* | |
5043 | * For every group except the first one, we are sure | |
5044 | * that the first cluster to discard will be cluster #0. | |
5045 | */ | |
7137d7a4 | 5046 | first_cluster = 0; |
7360d173 | 5047 | } |
7360d173 | 5048 | |
3d56b8d2 TM |
5049 | if (!ret) |
5050 | atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); | |
5051 | ||
22f10457 | 5052 | out: |
aaf7d73e | 5053 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
5054 | return ret; |
5055 | } |